diff --git a/parse/train/B1hdzd5lg/B1hdzd5lg.md b/parse/train/B1hdzd5lg/B1hdzd5lg.md new file mode 100644 index 0000000000000000000000000000000000000000..61ddea527d979865c9fe00b0428436e24fef6646 --- /dev/null +++ b/parse/train/B1hdzd5lg/B1hdzd5lg.md @@ -0,0 +1,239 @@ +# WORDS OR CHARACTERS? FINE-GRAINED GATING FOR READING COMPREHENSION + +Zhilin Yang, Bhuwan Dhingra, Ye Yuan, Junjie Hu, William W. Cohen, Ruslan Salakhutdinov + +School of Computer Science +Carnegie Mellon University +{zhiliny,wcohen,rsalakhu}@cs.cmu.edu + +# ABSTRACT + +Previous work combines word-level and character-level representations using concatenation or scalar weighting, which is suboptimal for high-level tasks like reading comprehension. We present a fine-grained gating mechanism to dynamically combine word-level and character-level representations based on properties of the words. We also extend the idea of fine-grained gating to modeling the interaction between questions and paragraphs for reading comprehension. Experiments show that our approach can improve the performance on reading comprehension tasks, achieving new state-of-the-art results on the Children’s Book Test and Who Did What datasets. To demonstrate the generality of our gating mechanism, we also show improved results on a social media tag prediction task.1 + +# 1 INTRODUCTION + +Finding semantically meaningful representations of the words (also called tokens) in a document is necessary for strong performance in Natural Language Processing tasks. In neural networks, tokens are mainly represented in two ways, either using word-level representations or character-level representations. Word-level representations are obtained from a lookup table, where each unique token is represented as a vector. Character-level representations are usually obtained by applying recurrent neural networks (RNNs) or convolutional neural networks (CNNs) on the character sequence of the token, and their hidden states are combined to form the representation. Word-level representations are good at memorizing the semantics of the tokens while character-level representations are more suitable for modeling sub-word morphologies (Ling et al., 2015; Yang et al., 2016a). For example, considering “cat” and “cats”, word-level representations can only learn the similarities between the two tokens by training on a large amount of training data, while character-level representations, by design, can easily capture the similarities. Character-level representations are also used to alleviate the difficulties of modeling out-of-vocabulary (OOV) tokens (Luong & Manning, 2016). + +Hybrid word-character models have been proposed to leverage the advantages of both word-level and character-level representations. The most commonly used method is to concatenate these two representations (Yang et al., 2016a). However, concatenating word-level and character-level representations is technically problematic. For frequent tokens, the word-level representations are usually accurately estimated during the training process, and thus introducing character-level representations can potentially bias the entire representations. For infrequent tokens, the estimation of wordlevel representations have high variance, which will have negative effects when combined with the character-level representations. To address this issue, recently Miyamoto & Cho (2016) introduced a scalar gate conditioned on the word-level representations to control the ratio of the two representations. However, for the task of reading comprehension, preliminary experiments showed that this method was not able to improve the performance over concatenation. There are two possible reasons. First, word-level representations might not contain sufficient information to support the decisions of selecting between the two representations. Second, using a scalar gate means applying the same ratio for each of the dimensions, which can be suboptimal. + +In this work, we present a fine-grained gating mechanism to combine the word-level and characterlevel representations. We compute a vector gate as a linear projection of the token features followed by a sigmoid activation. We then multiplicatively apply the gate to the character-level and wordlevel representations. Each dimension of the gate controls how much information is flowed from the word-level and character-level representations respectively. We use named entity tags, part-ofspeech tags, document frequencies, and word-level representations as the features for token properties which determine the gate. More generally, our fine-grained gating mechanism can be used to model multiple levels of structure in language, including words, characters, phrases, sentences and paragraphs. In this work we focus on studying the effects on word-character gating. + +To better tackle the problem of reading comprehension, we also extend the idea of fine-grained gating for modeling the interaction between documents and queries. Previous work has shown the importance of modeling interactions between document and query tokens by introducing various attention architectures for the task (Hermann et al., 2015; Kadlec et al., 2016). Most of these use an inner product between the two representations to compute the relative importance of document tokens. The Gated-Attention Reader (Dhingra et al., 2016a) showed improved performance by replacing the inner-product with an element-wise product to allow for better matching at the semantic level. However, they use aggregated representations of the query which may lead to loss of information. In this work we use a fine-grained gating mechanism for each token in the paragraph and each token in the query. The fine-grained gating mechanism applies an element-wise multiplication of the two representations. + +We show improved performance on reading comprehension datasets, including Children’s Book Test (CBT), Who Did What, and SQuAD. On CBT, our approach achieves new state-of-the-art results without using an ensemble. Our model also improves over state-of-the-art results on the Who Did What dataset. To demonstrate the generality of our method, we apply our word-character finegrained gating mechanism to a social media tag prediction task and show improved performance over previous methods. + +Our contributions are two-fold. First, we present a fine-grained word-character gating mechanism and show improved performance on a variety of tasks including reading comprehension. Second, to better tackle the reading comprehension tasks, we extend our fine-grained gating approach to modeling the interaction between documents and queries. + +# 2 RELATED WORK + +Hybrid word-character models have been proposed to take advantages of both word-level and character-level representations. Ling et al. (2015) introduce a compositional character to word (C2W) model based on bidirectional LSTMs. Kim et al. (2016) describe a model that employs a convolutional neural network (CNN) and a highway network over characters for language modeling. Miyamoto & Cho (2016) use a gate to adaptively find the optimal mixture of the character-level and word-level inputs. Yang et al. (2016a) employ deep gated recurrent units on both character and word levels to encode morphology and context information. Concurrent to our work, Rei et al. (2016) employed a similar gating idea to combine word-level and character-level representations, but their focus is on low-level sequence tagging tasks and the gate is not conditioned on linguistic features. + +The gating mechanism is widely used in sequence modeling. Long short-term memory (LSTM) networks (Hochreiter & Schmidhuber, 1997) are designed to deal with vanishing gradients through the gating mechanism. Similar to LSTM, Gated Recurrent Unit (GRU) was proposed by Cho et al. (2014), which also uses gating units to modulate the flow of information. The gating mechanism can also be viewed as a form of attention mechanism (Bahdanau et al., 2015; Yang et al., 2016b) over two inputs. + +Similar to the idea of gating, multiplicative integration has also been shown to provide a benefit in various settings. Yang et al. (2014) find that multiplicative operations are superior to additive operations in modeling relations. Wu et al. (2016) propose to use Hadamard product to replace sum operation in recurrent networks, which gives a significant performance boost over existing RNN models. Dhingra et al. (2016a) use a multiplicative gating mechanism to achieve state-of-the-art results on question answering benchmarks. + +Reading comprehension is a challenging task for machines. A variety of models have been proposed to extract answers from given text (Hill et al., 2016; Kadlec et al., 2016; Trischler et al., 2016; Chen et al., 2016; Sordoni et al., 2016; Cui et al., 2016). Yu et al. (2016) propose a dynamic chunk reader to extract and rank a set of answer candidates from a given document to answer questions. Wang & Jiang (2016) introduce an end-to-end neural architecture which incorporates match-LSTM and pointer networks (Vinyals et al., 2015). + +# 3 FINE-GRAINED GATING + +In this section, we will describe our fine-grained gating approach in the context of reading comprehension. We first introduce the settings of reading comprehension tasks and a general neural network architecture. We will then describe our word-character gating and document-query gating approaches respectively. + +# 3.1 READING COMPREHENSION SETTING + +The reading comprehension task involves a document $P = ( p _ { 1 } , p _ { 2 } , \cdot \cdot \cdot , p _ { M } )$ and a query $Q =$ $\left( q _ { 1 } , q _ { 2 } , \cdots , q _ { N } \right)$ , where $M$ and $N$ are the lengths of the document and the query respectively. Each token $p _ { i }$ is denoted as $( \mathbf { w } _ { i } , \mathbf { C } _ { i } )$ , where $\mathbf { w } _ { i }$ is a one-hot encoding of the token in the vocabulary and $\mathbf { C } _ { i }$ is a matrix with each row representing a one-hot encoding of a character. Each token in the query $q _ { j }$ is similarly defined. We use $i$ as a subscript for documents and $j$ for queries. The output of the problem is an answer $a$ , which can either be an index or a span of indices in the document. + +Now we describe a general architecture used in this work, which is a generalization of the gated attention reader (Dhingra et al., 2016a). For each token in the document and the query, we compute a vector representation using a function $f$ . More specifically, for each token $p _ { i }$ in the document, we have $\mathbf { h } _ { i } ^ { \bar { 0 } } = f ( \mathbf { w } _ { i } , \mathbf { C } _ { i } )$ . The same function $f$ is also applied to the tokens in the query. Let $\mathbf { H } _ { p } ^ { 0 }$ and $\mathbf { H } _ { q }$ denote the vector representations computed by $f$ for tokens in documents and queries respectively. In Section 3.2, we will discuss the “word-character” fine-grained gating used to define the function $f$ . + +Suppose that we have a network of $K$ layers. At the $k$ -th layer, we apply RNNs on $\mathbf { H } _ { p } ^ { k - 1 }$ and $\mathbf { H } _ { q }$ to obtain hidden states $\mathbf { P } ^ { k }$ and $\mathbf { Q } ^ { k }$ , where $\mathbf { P } ^ { k }$ is a $M \times d$ matrix and $\mathbf { Q } ^ { k }$ is a $N \times d$ matrix with $d$ being the number of hidden units in the RNNs. Then we use a function $r$ to compute a new representation for the document $\mathbf { H } _ { p } ^ { k } = r ( \mathbf { P } ^ { k } , \mathbf { Q } ^ { k } )$ . In Section 3.3, we will introduce the “document-query” finegrained gating used to define the function $r$ . + +After going through $K$ layers, we predict the answer index $a$ using a softmax layer over hidden states $\bar { \mathbf { H } } _ { p } ^ { k }$ . For datasets where the answer is a span of text, we use two softmax layers for the start and end indices respectively. + +# 3.2 WORD-CHARACTER FINE-GRAINED GATING + +Given a one-hot encoding $\mathbf { w } _ { i }$ and a character sequence $\mathbf { C } _ { i }$ , we now describe how to compute the vector representation $\mathbf { h } _ { i } = f ( \mathbf { w } _ { i } , \mathbf { C } _ { i } )$ for the token. In the rest of the section, we will drop the subscript $i$ for notation simplicity. + +We first apply an RNN on $\mathbf { C }$ and take the hidden state in the last time step c as the character-level representation (Yang et al., 2016a). Let $\mathbf { E }$ denote the token embedding lookup table. We perform a matrix-vector multiplication Ew to obtain a word-level representation. We assume c and Ew have the same length $d _ { e }$ in this work. + +Previous methods defined $f$ using the word-level representation Ew (Collobert et al., 2011), the character-level representation c (Ling et al., 2015), or the concatenation $\left[ \mathbf { E w } ; \mathbf { c } \right]$ (Yang et al., 2016a). Unlike these methods, we propose to use a gate to dynamically choose between the word-level and character-level representations based on the properties of the token. Let $\mathbf { v }$ denote a feature vector that encodes these properties. In this work, we use the concatenation of named entity tags, partof-speech tags, binned document frequency vectors, and the word-level representations to form the feature vector $\mathbf { v }$ . Let $d _ { v }$ denote the length of $\mathbf { v }$ . + +The gate is computed as follows: + +$$ +\mathbf { g } = \sigma ( \mathbf { W } _ { g } \mathbf { v } + \mathbf { b } _ { g } ) +$$ + +![](images/f10e95e7400bc55e85299f1e966111eeaf000699ba21922a0e9aef0ca84cfda7.jpg) +Figure 1: Word-character fine-grained gating. The two lookup tables are shared. “NER”, “POS”, “frequency” refer to named entity tags, part-of-speech tags, document frequency features. + +where $\mathbf { W } _ { g }$ and ${ \bf b } _ { g }$ are the model parameters with shapes $d _ { e } \times d _ { v }$ and $d _ { e }$ , and $\sigma$ denotes an elementwise sigmoid function. + +The final representation is computed using a fine-grained gating mechanism, + +$$ +\mathbf { h } = f ( \mathbf { c } , \mathbf { w } ) = \mathbf { g } \odot \mathbf { c } + ( 1 - \mathbf { g } ) \odot ( \mathbf { E } \mathbf { w } ) +$$ + +where $\odot$ denotes element-wise product between two vectors. + +An illustration of our fine-grained gating mechanism is shown in Figure 1. Intuitively speaking, when the gate $\mathbf { g }$ has high values, more information flows from the character-level representation to the final representation; when the gate $\mathbf { g }$ has low values, the final representation is dominated by the word-level representation. + +Though Miyamoto & Cho (2016) also use a gate to choose between word-level and character-level representations, our method is different in two ways. First, we use a more fine-grained gating mechanism, i.e., vector gates rather than scalar gates. Second, we condition the gate on features that better reflect the properties of the token. For example, for noun phrases and entities, we would expect the gate to bias towards character-level representations because noun phrases and entities are usually less common and display richer morphological structure. Experiments show that these changes are key to the performance improvements for reading comprehension tasks. + +Our approach can be further generalized to a setting of multi-level networks so that we can combine multiple levels of representations using fine-grained gating mechanisms, which we leave for future work. + +# 3.3 DOCUMENT-QUERY FINE-GRAINED GATING + +Given the hidden states $\mathbf { P } ^ { k }$ and $\mathbf { Q } ^ { k }$ , we now describe how to compute a representation $\mathbf { H } ^ { k }$ that encodes the interactions between the document and the query. In this section, we drop the superscript $k$ (the layer number) for notation simplicity. Let $\mathbf { p } _ { i }$ denote the $i$ -th row of $\mathbf { P }$ and ${ \bf q } _ { j }$ denote the $j$ -row of $\mathbf { Q }$ . Let $d _ { h }$ denote the lengths of $\mathbf { p } _ { i }$ and ${ \bf q } _ { j }$ . + +Attention-over-attention (AoA) (Cui et al., 2016) defines a dot product between each pair of tokens in the document and the query, i.e., $\mathbf { p } _ { i } ^ { T } \mathbf { q } _ { j }$ , followed by row-wise and column-wise softmax nonlinearities. AoA imposes pair-wise interactions between the document and the query, but using a dot product is potentially not expressive enough and hard to generalize to multi-layer networks. The gated attention (GA) reader (Dhingra et al., 2016a) defines an element-wise product as $\mathbf { p } _ { i } \odot \mathbf { g } _ { i }$ where $\mathbf { g } _ { i }$ is a gate computed by attention mechanism on the token $p _ { i }$ and the entire query. The intuition for the gate $\mathbf { g } _ { i }$ is to attend to important information in the document. However, there is no direct pair-wise interaction between each token pair. + +![](images/284468038502d2ff8d8577366fc87421d42142474f6344ca41f53869482de701.jpg) +Figure 2: Paragraph-question fine-grained gating. + +We present a fine-grained gating method that combines the advantages of the above methods (i.e., both pairwise and element-wise). We compute the pairwise element-wise product between the hidden states in the document and the query, as shown in Figure 2. More specifically, for $p _ { i }$ and $q _ { j }$ , we have + +$$ +\mathbf { I } _ { i j } = \operatorname { t a n h } ( \mathbf { p } _ { i } \odot \mathbf { q } _ { j } ) +$$ + +where ${ \bf q } _ { j }$ can be viewed as a gate to filter the information in $\mathbf { p } _ { i }$ . We then use an attention mechanism over $\mathbf { I } _ { i j }$ to output hidden states $\mathbf { h } _ { i }$ as follows + +$$ +\mathbf { h } _ { i } = \sum _ { j } \mathrm { s o f t m a x } ( \mathbf { u } _ { h } ^ { T } \mathbf { I } _ { i j } + \mathbf { w } _ { i } ^ { T } \mathbf { w } _ { j } b _ { h 1 } + b _ { h 2 } ) \mathbf { I } _ { i j } +$$ + +where $\mathbf { u } _ { h }$ is a $d _ { v }$ -dimensional model parameter, $b _ { h 1 }$ and $b _ { h 2 }$ are scalar model parameters, $\mathbf { w } _ { i }$ and ${ \bf w } _ { j }$ are one-hot encodings for $p _ { i }$ and $q _ { j }$ respectively. We additionally use one-hot encodings in the attention mechanism to reinforce the matching between the same tokens since such information is not fully preserved in $\mathbf { I } _ { i j }$ when $k$ is large. The softmax nonlinearity is applied over all $j$ ’s. The final hidden states $\mathbf { H }$ are formed by concatenating the $\mathbf { h } _ { i }$ ’s for each token $p _ { i }$ . + +# 4 EXPERIMENTS + +We first present experimental results on the Twitter dataset where we can rule out the effects of different choices of network architectures, to demonstrate the effectiveness of our word-character fine-grained gating approach. Later we show experiments on more challenging datasets on reading comprehension to further show that our approach can be used to improve the performance on highlevel NLP tasks as well. + +# 4.1 EVALUATING WORD-CHARACTER GATING ON TWITTER + +We evaluate the effectiveness of our word-character fine-grained gating mechanism on a social media tag prediction task. We use the Twitter dataset and follow the experimental settings in Dhingra et al. (2016b). We also use the same network architecture upon the token representations, which is an LSTM layer followed by a softmax classification layer (Dhingra et al., 2016b). The Twitter dataset consists of English tweets with at least one hashtag from Twitter. Hashtags and HTML tags have been removed from the body of the tweet, and user names and URLs are replaced with special tokens. The dataset contains 2 million tweets for training, 10K for validation and 50K for testing, with a total of 2,039 distinct hashtags. The task is to predict the hashtags of each tweet. + +We compare several different methods as follows. Word char concat uses the concatenation of word-level and character-level representations as in Yang et al. (2016a); word char feat concat concatenates the word-level and character-level representations along with the features described in + +Table 1: Performance on the Twitter dataset. “word” and “char” means using word-level and character-level representations respectively. + +
ModelPrecision@1Recall@10Mean Rank
word (Dhingra et al., 2016b)0.2410.428133
char (Dhingra et al., 2016b)0.2840.485104
word char concat0.29610.4959105.8
word char feat concat0.29510.4974106.2
scalar gate0.29740.4982104.2
fine-grained gate0.30690.5119101.5
+ +Table 2: Performance on the CBT dataset. The “GA word char concat” results are extracted from Dhingra et al. (2016a). Our results on fine-grained gating are based on a single model. “CN” and “NE” are two widely used question categories. “dev” means development set, and “test” means test set. + +
ModelCN devCN testNE devNE test
GA word char concat0.7310.6960.7680.725
GA word char feat concat0.72500.69280.78150.7256
GA scalar gate0.72400.69080.78100.7260
GA fine-grained gate0.74250.70840.78900.7464
FG fine-grained gate0.75300.72040.79100.7496
Sordoni et al. (2016)0.7210.6920.7520.686
Trischler et al. (2016)0.7150.6740.7530.697
Cui et al. (2016)0.7220.6940.7780.720
Munkhdalai & Yu (2016)0.7430.7190.7820.732
Kadlec et al. (2016) ensemble0.7110.6890.7620.710
Sordoni et al. (2016) ensemble0.7410.7100.7690.720
Trischler etal. (2016) ensemble0.7360.7060.7660.718
+ +Section 3.2; scalar gate uses a scalar gate similar to Miyamoto & Cho (2016) but is conditioned on the features; fine-grained gate is our method described in Section 3.2. We include word char feat concat for a fair comparison because our fine-grained gating approach also uses the token features. + +The results are shown in Table 1. We report three evaluation metrics including precision $@ 1$ , recall $@ 1 0$ , and mean rank. Our method outperforms character-level models used in Dhingra et al. (2016b) by $2 . 2 9 \%$ , $2 . 6 9 \%$ , and 2.5 points in terms of precision, recall and mean rank respectively. We can observe that scalar gating approach (Miyamoto & Cho, 2016) can only marginally improve over the baseline methods, while fine-grained gating methods can substantially improve model performance. Note that directly concatenating the token features with the character-level and word-level representations does not boost the performance, but using the token features to compute a gate (as done in fine-grained gating) leads to better results. This indicates that the benefit of fine-grained gating mainly comes from better modeling rather than using additional features. + +# 4.2 PERFORMANCE ON READING COMPREHENSION + +After investigating the effectiveness of the word-character fine-grained gating mechanism on the Twitter dataset, we now move on to a more challenging task, reading comprehension. In this section, we experiment with two datasets, the Children’s Book Test dataset (Hill et al., 2016) and the SQuAD dataset (Rajpurkar et al., 2016). + +# 4.2.1 CLOZE-STYLE QUESTIONS + +We evaluate our model on cloze-style question answering benchmarks. + +Table 3: Performance on the Who Did What dataset. “dev” means development set, and “test” means test set. “WDW-R” is the relaxed version of WDW. + +
ModelWDW devWDW testWDW-R devWDW-R test
0.5700.590
Kadlec et al. (2016) Chen et al. (2016)0.640一0.650
Munkhdalai & Yu (2016)0.6650.6620.6700.667
Dhingra et al. (2016a)0.7160.7120.7260.726
this paper0.7230.7170.7310.726
+ +Table 4: Performance on the SQuAD dev set. Test set results are included in the brackets. + +
ModelF1Exact Match
GA word0.5492
GA word char concat0.6695 0.68570.5639
GA word char feat concat0.69040.5711
GA scalar gate0.68500.5620
GA fine-grained gate0.69830.5804
FG fine-grained gate0.71250.5995
FG fine-grained gate + ensemble0.7341 (0.733)0.6238 (0.625)
Yu et al. (2016)
Wang & Jiang (2016)0.712 (0.710) 0.700 (0.703)0.625 (0.625) 0.591 (0.595)
+ +The Children’s Book Test (CBT) dataset is built from children’s books. The whole dataset has 669,343 questions for training, 8,000 for validation and 10,000 for testing. We closely follow the setting in Dhingra et al. (2016a) and incrementally add different components to see the changes in performance. For the fine-grained gating approach, we use the same hyper-parameters as in Dhingra et al. (2016a) except that we use a character-level GRU with 100 units to be of the same size as the word lookup table. The word embeddings are updated during training. + +In addition to different ways of combining word-level and character-level representations, we also compare two different ways of integrating documents and queries: GA refers to the gated attention reader (Dhingra et al., 2016a) and FG refers to our fine-grained gating described in Section 3.3. + +The results are reported in Table 2. We report the results on common noun (CN) questions and named entity (NE) questions, which are two widely used question categories in CBT. Our finegrained gating approach achieves new state-of-the-art performance on both settings and outperforms the current state-of-the-art results by up to $1 . 7 6 \%$ without using ensembles. Our method outperforms the baseline GA reader by up to $2 . 4 \%$ , which indicates the effectiveness of the fine-grained gating mechanism. Consistent with the results on the Twitter dataset, using word-character fine-grained gating can substantially improve the performance over concatenation or scalar gating. Furthermore, we can see that document-query fine-grained gating also contributes significantly to the final results. + +We also apply our fine-grained gating model to the Who Did What (WDW) dataset (Onishi et al., 2016). As shown in Table 3, our model achieves state-of-the-art results compared to strong baselines. We fix the word embeddings during training. + +# 4.2.2 SQUAD + +The Stanford Question Answering Dataset (SQuAD) is a reading comprehension dataset collected recently (Rajpurkar et al., 2016). It contains 23,215 paragraphs come from 536 Wikipedia articles. Unlike other reading comprehension datasets such as CBT, the answers are a span of text rather than a single word. The dataset is partitioned into a training set ( $80 \%$ , 87,636 question-answer pairs), a development set ( $10 \%$ , 10,600 question-answer pairs) and a test set which is not released. + +![](images/89ddff74b55246bf812f8cbb6330db8793bb826921f94510c5ea362754ea15ba.jpg) +Figure 3: Visualization of the weight matrix $\mathbf { W } _ { g }$ . Weights for each features are averaged. Red means high and yellow means low. High weight values favor character-level representations, and low weight values favor wordlevel representations. “Organization”, “‘Person”, “Location”, and “O” are named entity tags; “DOCLEN-n” are document frequency features (larger $_ n$ means higher frequency, $n$ from 0 to 4); others are POS tags. + +![](images/e182272ce273a6e2c7869e53e7be53700145a9a8f9ecfbc54ac0980409cc5dd8.jpg) +Figure 4: Visualization of gate values in the text. Red means high and yellow means low. High gate values favor character-level representations, and low gate values favor word-level representations. + +We report our results in Table 4. “Exact match” computes the ratio of questions that are answered correctly by strict string comparison, and the F1 score is computed on the token level. We can observe that both word-character fine-grained gating and document-query fine-grained gating can substantially improve the performance, leading to state-of-the-art results among published papers. Note that at the time of submission, the best score on the leaderboard is 0.716 in exact match and 0.804 in F1 without published papers. A gap exists because our architecture described in Section 3.1 does not specifically model the answer span structure that is unique to SQuAD. In this work, we focus on this general architecture to study the effectiveness of fine-grained gating mechanisms. + +# 4.3 VISUALIZATION AND ANALYSIS + +We visualize the model parameter $\mathbf { W } _ { g }$ as described in Section 3.2. For each feature, we average the corresponding weight vector in $\mathbf { W } _ { g }$ . The results are described in Figure 3. We can see that named entities like “Organization” and noun phrases (with tags “NNP” or “NNPS”) tend to use characterlevel representations, which is consistent with human intuition because those tokens are usually infrequent or display rich morphologies. Also, DOCLEN-4, WH-adverb (“WRB”), and conjunction (“IN” and “CC”) tokens tend to use word-level representations because they appear frequently. + +We also sample random span of text from the SQuAD dataset, and visualize the average gate values in Figure 4. The results are consistent with our observations in Figure 3. Rare tokens, noun phrases, and named entities tend to use character-level representations, while others tend to use word-level representations. To further justify this argument, we also list the tokens with highest and lowest gate values in Table 5. + +Table 5: Word tokens with highest and lowest gate values. High gate values favor character-level representations, and low gate values favor word-level representations. + +
Gate valuesWord tokens
Lowestor but But These these However however among Among that when When although Although because Because until many Many than though Though this This Since since date where Where have That and And Such such number so which by By how before Before with With between Between even Even if
HighestSweetgum Untersee Jianlong Floresta Chlorella Obersee PhT Doctorin Jumonville WFTS WTSP Boven Pharm Nederrijn Otrar Rhin Magicicada WBKB Tanzler KMBC WPLG Mainau Merwede RMJM Kleitman Scheur Bodensee Kromme Horenbout Vorderrhein Chlamydomonas Scantlebury Qingshui Funchess
+ +# 5 CONCLUSIONS + +We present a fine-grained gating mechanism that dynamically combines word-level and characterlevel representations based on word properties. Experiments on the Twitter tag prediction dataset show that fine-grained gating substantially outperforms scalar gating and concatenation. Our method also improves the performance on reading comprehension and achieves new state-of-the-art results on CBT and WDW. In our future work, we plan to to apply the fine-grained gating mechanism for combining other levels of representations, such as phrases and sentences. It will also be intriguing to integrate NER and POS networks and learn the token representation in an end-to-end manner. + +# ACKNOWLEDGMENTS + +This work was funded by NVIDIA, the Office of Naval Research Scene Understanding grant N000141310721, the NSF grant IIS1250956, and Google Research. + +# REFERENCES + +Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In ICLR, 2015. + +Danqi Chen, Jason Bolton, and Christopher D Manning. A thorough examination of the cnn/daily mail reading comprehension task. In ACL, 2016. + +Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural ¨ machine translation: Encoder-decoder approaches. In Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, 2014. + +Ronan Collobert, Jason Weston, Leon Bottou, Michael Karlen, Koray Kavukcuoglu, and Pavel Kuksa. Natural ´ language processing (almost) from scratch. Journal of Machine Learning Research, 12(Aug):2493–2537, 2011. + +Yiming Cui, Zhipeng Chen, Si Wei, Shijin Wang, Ting Liu, and Guoping Hu. Attention-over-attention neural networks for reading comprehension. arXiv preprint arXiv:1607.04423, 2016. + +Bhuwan Dhingra, Hanxiao Liu, Zhilin Yang, William W Cohen, and Ruslan Salakhutdinov. Gated-attention readers for text comprehension. arXiv preprint arXiv:1606.01549, 2016a. + +Bhuwan Dhingra, Zhong Zhou, Dylan Fitzpatrick, Michael Muehl, and William W Cohen. Tweet2vec: Character-based distributed representations for social media. In ACL, 2016b. + +Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. Teaching machines to read and comprehend. In NIPS, pp. 1693–1701, 2015. + +Felix Hill, Antoine Bordes, Sumit Chopra, and Jason Weston. The goldilocks principle: Reading children’s books with explicit memory representations. In ICLR, 2016. + +Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):1735–1780, 1997. + +Rudolf Kadlec, Martin Schmid, Ondrej Bajgar, and Jan Kleindienst. Text understanding with the attention sum reader network. In ACL, 2016. + +Yoon Kim, Yacine Jernite, David Sontag, and Alexander M Rush. Character-aware neural language models. In AAAI, 2016. + +Wang Ling, Tiago Lu´ıs, Lu´ıs Marujo, Ramon Fernandez Astudillo, Silvio Amir, Chris Dyer, Alan W Black, ´ and Isabel Trancoso. Finding function in form: Compositional character models for open vocabulary word representation. In EMNLP, 2015. + +Minh-Thang Luong and Christopher D Manning. Achieving open vocabulary neural machine translation with hybrid word-character models. In ACL, 2016. + +Yasumasa Miyamoto and Kyunghyun Cho. Gated word-character recurrent language model. In EMNLP, 2016. + +Tsendsuren Munkhdalai and Hong Yu. Neural semantic encoders. arXiv preprint arXiv:1607.04315, 2016. + +Takeshi Onishi, Hai Wang, Mohit Bansal, Kevin Gimpel, and David McAllester. Who did what: A large-scale person-centered cloze dataset. arXiv preprint arXiv:1608.05457, 2016. + +Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad: $^ { 1 0 0 , 0 0 0 + }$ questions for machine comprehension of text. In EMNLP, 2016. + +Marek Rei, Gamal KO Crichton, and Sampo Pyysalo. Attending to characters in neural sequence labeling models. arXiv preprint arXiv:1611.04361, 2016. + +Alessandro Sordoni, Phillip Bachman, and Yoshua Bengio. Iterative alternating neural attention for machine reading. arXiv preprint arXiv:1606.02245, 2016. + +Adam Trischler, Zheng Ye, Xingdi Yuan, and Kaheer Suleman. Natural language comprehension with the epireader. In EMNLP, 2016. + +Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In NIPS, pp. 2692–2700, 2015. + +Shuohang Wang and Jing Jiang. Machine comprehension using match-lstm and answer pointer. arXiv preprint arXiv:1608.07905, 2016. + +Yuhuai Wu, Saizheng Zhang, Ying Zhang, Yoshua Bengio, and Ruslan Salakhutdinov. On multiplicative integration with recurrent neural networks. In NIPS, 2016. + +Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. Learning multi-relational semantics using neural-embedding models. In NIPS 2014 workshop on Learning Semantics, 2014. + +Zhilin Yang, Ruslan Salakhutdinov, and William Cohen. Multi-task cross-lingual sequence tagging from scratch. arXiv preprint arXiv:1603.06270, 2016a. + +Zhilin Yang, Ye Yuan, Yuexin Wu, Ruslan Salakhutdinov, and William W Cohen. Review networks for caption generation. In NIPS, 2016b. + +Yang Yu, Wei Zhang, Kazi Hasan, Mo Yu, Bing Xiang, and Bowen Zhou. End-to-end answer chunk extraction and ranking for reading comprehension. arXiv preprint arXiv:1610.09996, 2016. \ No newline at end of file diff --git a/parse/train/B1hdzd5lg/B1hdzd5lg_content_list.json b/parse/train/B1hdzd5lg/B1hdzd5lg_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dbf69c8b649bc810f8d886dd3efdcc4490da0e23 --- /dev/null +++ b/parse/train/B1hdzd5lg/B1hdzd5lg_content_list.json @@ -0,0 +1,1278 @@ +[ + { + "type": "text", + "text": "WORDS OR CHARACTERS? FINE-GRAINED GATING FOR READING COMPREHENSION ", + "text_level": 1, + "bbox": [ + 176, + 98, + 825, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhilin Yang, Bhuwan Dhingra, Ye Yuan, Junjie Hu, William W. Cohen, Ruslan Salakhutdinov ", + "bbox": [ + 186, + 170, + 834, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Computer Science \nCarnegie Mellon University \n{zhiliny,wcohen,rsalakhu}@cs.cmu.edu ", + "bbox": [ + 184, + 186, + 534, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 262, + 544, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous work combines word-level and character-level representations using concatenation or scalar weighting, which is suboptimal for high-level tasks like reading comprehension. We present a fine-grained gating mechanism to dynamically combine word-level and character-level representations based on properties of the words. We also extend the idea of fine-grained gating to modeling the interaction between questions and paragraphs for reading comprehension. Experiments show that our approach can improve the performance on reading comprehension tasks, achieving new state-of-the-art results on the Children’s Book Test and Who Did What datasets. To demonstrate the generality of our gating mechanism, we also show improved results on a social media tag prediction task.1 ", + "bbox": [ + 233, + 291, + 764, + 430 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 454, + 336, + 469 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Finding semantically meaningful representations of the words (also called tokens) in a document is necessary for strong performance in Natural Language Processing tasks. In neural networks, tokens are mainly represented in two ways, either using word-level representations or character-level representations. Word-level representations are obtained from a lookup table, where each unique token is represented as a vector. Character-level representations are usually obtained by applying recurrent neural networks (RNNs) or convolutional neural networks (CNNs) on the character sequence of the token, and their hidden states are combined to form the representation. Word-level representations are good at memorizing the semantics of the tokens while character-level representations are more suitable for modeling sub-word morphologies (Ling et al., 2015; Yang et al., 2016a). For example, considering “cat” and “cats”, word-level representations can only learn the similarities between the two tokens by training on a large amount of training data, while character-level representations, by design, can easily capture the similarities. Character-level representations are also used to alleviate the difficulties of modeling out-of-vocabulary (OOV) tokens (Luong & Manning, 2016). ", + "bbox": [ + 174, + 484, + 825, + 665 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hybrid word-character models have been proposed to leverage the advantages of both word-level and character-level representations. The most commonly used method is to concatenate these two representations (Yang et al., 2016a). However, concatenating word-level and character-level representations is technically problematic. For frequent tokens, the word-level representations are usually accurately estimated during the training process, and thus introducing character-level representations can potentially bias the entire representations. For infrequent tokens, the estimation of wordlevel representations have high variance, which will have negative effects when combined with the character-level representations. To address this issue, recently Miyamoto & Cho (2016) introduced a scalar gate conditioned on the word-level representations to control the ratio of the two representations. However, for the task of reading comprehension, preliminary experiments showed that this method was not able to improve the performance over concatenation. There are two possible reasons. First, word-level representations might not contain sufficient information to support the decisions of selecting between the two representations. Second, using a scalar gate means applying the same ratio for each of the dimensions, which can be suboptimal. ", + "bbox": [ + 174, + 672, + 825, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we present a fine-grained gating mechanism to combine the word-level and characterlevel representations. We compute a vector gate as a linear projection of the token features followed by a sigmoid activation. We then multiplicatively apply the gate to the character-level and wordlevel representations. Each dimension of the gate controls how much information is flowed from the word-level and character-level representations respectively. We use named entity tags, part-ofspeech tags, document frequencies, and word-level representations as the features for token properties which determine the gate. More generally, our fine-grained gating mechanism can be used to model multiple levels of structure in language, including words, characters, phrases, sentences and paragraphs. In this work we focus on studying the effects on word-character gating. ", + "bbox": [ + 176, + 875, + 823, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 823, + 202 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To better tackle the problem of reading comprehension, we also extend the idea of fine-grained gating for modeling the interaction between documents and queries. Previous work has shown the importance of modeling interactions between document and query tokens by introducing various attention architectures for the task (Hermann et al., 2015; Kadlec et al., 2016). Most of these use an inner product between the two representations to compute the relative importance of document tokens. The Gated-Attention Reader (Dhingra et al., 2016a) showed improved performance by replacing the inner-product with an element-wise product to allow for better matching at the semantic level. However, they use aggregated representations of the query which may lead to loss of information. In this work we use a fine-grained gating mechanism for each token in the paragraph and each token in the query. The fine-grained gating mechanism applies an element-wise multiplication of the two representations. ", + "bbox": [ + 174, + 208, + 825, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We show improved performance on reading comprehension datasets, including Children’s Book Test (CBT), Who Did What, and SQuAD. On CBT, our approach achieves new state-of-the-art results without using an ensemble. Our model also improves over state-of-the-art results on the Who Did What dataset. To demonstrate the generality of our method, we apply our word-character finegrained gating mechanism to a social media tag prediction task and show improved performance over previous methods. ", + "bbox": [ + 174, + 367, + 825, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are two-fold. First, we present a fine-grained word-character gating mechanism and show improved performance on a variety of tasks including reading comprehension. Second, to better tackle the reading comprehension tasks, we extend our fine-grained gating approach to modeling the interaction between documents and queries. ", + "bbox": [ + 176, + 458, + 823, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 539, + 341, + 554 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hybrid word-character models have been proposed to take advantages of both word-level and character-level representations. Ling et al. (2015) introduce a compositional character to word (C2W) model based on bidirectional LSTMs. Kim et al. (2016) describe a model that employs a convolutional neural network (CNN) and a highway network over characters for language modeling. Miyamoto & Cho (2016) use a gate to adaptively find the optimal mixture of the character-level and word-level inputs. Yang et al. (2016a) employ deep gated recurrent units on both character and word levels to encode morphology and context information. Concurrent to our work, Rei et al. (2016) employed a similar gating idea to combine word-level and character-level representations, but their focus is on low-level sequence tagging tasks and the gate is not conditioned on linguistic features. ", + "bbox": [ + 174, + 568, + 825, + 707 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The gating mechanism is widely used in sequence modeling. Long short-term memory (LSTM) networks (Hochreiter & Schmidhuber, 1997) are designed to deal with vanishing gradients through the gating mechanism. Similar to LSTM, Gated Recurrent Unit (GRU) was proposed by Cho et al. (2014), which also uses gating units to modulate the flow of information. The gating mechanism can also be viewed as a form of attention mechanism (Bahdanau et al., 2015; Yang et al., 2016b) over two inputs. ", + "bbox": [ + 174, + 714, + 825, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Similar to the idea of gating, multiplicative integration has also been shown to provide a benefit in various settings. Yang et al. (2014) find that multiplicative operations are superior to additive operations in modeling relations. Wu et al. (2016) propose to use Hadamard product to replace sum operation in recurrent networks, which gives a significant performance boost over existing RNN models. Dhingra et al. (2016a) use a multiplicative gating mechanism to achieve state-of-the-art results on question answering benchmarks. ", + "bbox": [ + 174, + 804, + 823, + 888 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reading comprehension is a challenging task for machines. A variety of models have been proposed to extract answers from given text (Hill et al., 2016; Kadlec et al., 2016; Trischler et al., 2016; Chen et al., 2016; Sordoni et al., 2016; Cui et al., 2016). Yu et al. (2016) propose a dynamic chunk reader to extract and rank a set of answer candidates from a given document to answer questions. Wang & Jiang (2016) introduce an end-to-end neural architecture which incorporates match-LSTM and pointer networks (Vinyals et al., 2015). ", + "bbox": [ + 174, + 895, + 823, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 FINE-GRAINED GATING ", + "text_level": 1, + "bbox": [ + 176, + 181, + 405, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we will describe our fine-grained gating approach in the context of reading comprehension. We first introduce the settings of reading comprehension tasks and a general neural network architecture. We will then describe our word-character gating and document-query gating approaches respectively. ", + "bbox": [ + 174, + 213, + 825, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 READING COMPREHENSION SETTING ", + "text_level": 1, + "bbox": [ + 178, + 287, + 473, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The reading comprehension task involves a document $P = ( p _ { 1 } , p _ { 2 } , \\cdot \\cdot \\cdot , p _ { M } )$ and a query $Q =$ $\\left( q _ { 1 } , q _ { 2 } , \\cdots , q _ { N } \\right)$ , where $M$ and $N$ are the lengths of the document and the query respectively. Each token $p _ { i }$ is denoted as $( \\mathbf { w } _ { i } , \\mathbf { C } _ { i } )$ , where $\\mathbf { w } _ { i }$ is a one-hot encoding of the token in the vocabulary and $\\mathbf { C } _ { i }$ is a matrix with each row representing a one-hot encoding of a character. Each token in the query $q _ { j }$ is similarly defined. We use $i$ as a subscript for documents and $j$ for queries. The output of the problem is an answer $a$ , which can either be an index or a span of indices in the document. ", + "bbox": [ + 174, + 314, + 825, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Now we describe a general architecture used in this work, which is a generalization of the gated attention reader (Dhingra et al., 2016a). For each token in the document and the query, we compute a vector representation using a function $f$ . More specifically, for each token $p _ { i }$ in the document, we have $\\mathbf { h } _ { i } ^ { \\bar { 0 } } = f ( \\mathbf { w } _ { i } , \\mathbf { C } _ { i } )$ . The same function $f$ is also applied to the tokens in the query. Let $\\mathbf { H } _ { p } ^ { 0 }$ and $\\mathbf { H } _ { q }$ denote the vector representations computed by $f$ for tokens in documents and queries respectively. In Section 3.2, we will discuss the “word-character” fine-grained gating used to define the function $f$ . ", + "bbox": [ + 173, + 405, + 825, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Suppose that we have a network of $K$ layers. At the $k$ -th layer, we apply RNNs on $\\mathbf { H } _ { p } ^ { k - 1 }$ and $\\mathbf { H } _ { q }$ to obtain hidden states $\\mathbf { P } ^ { k }$ and $\\mathbf { Q } ^ { k }$ , where $\\mathbf { P } ^ { k }$ is a $M \\times d$ matrix and $\\mathbf { Q } ^ { k }$ is a $N \\times d$ matrix with $d$ being the number of hidden units in the RNNs. Then we use a function $r$ to compute a new representation for the document $\\mathbf { H } _ { p } ^ { k } = r ( \\mathbf { P } ^ { k } , \\mathbf { Q } ^ { k } )$ . In Section 3.3, we will introduce the “document-query” finegrained gating used to define the function $r$ . ", + "bbox": [ + 174, + 510, + 825, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After going through $K$ layers, we predict the answer index $a$ using a softmax layer over hidden states $\\bar { \\mathbf { H } } _ { p } ^ { k }$ . For datasets where the answer is a span of text, we use two softmax layers for the start and end indices respectively. ", + "bbox": [ + 176, + 589, + 823, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 WORD-CHARACTER FINE-GRAINED GATING ", + "text_level": 1, + "bbox": [ + 174, + 650, + 526, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a one-hot encoding $\\mathbf { w } _ { i }$ and a character sequence $\\mathbf { C } _ { i }$ , we now describe how to compute the vector representation $\\mathbf { h } _ { i } = f ( \\mathbf { w } _ { i } , \\mathbf { C } _ { i } )$ for the token. In the rest of the section, we will drop the subscript $i$ for notation simplicity. ", + "bbox": [ + 178, + 676, + 821, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first apply an RNN on $\\mathbf { C }$ and take the hidden state in the last time step c as the character-level representation (Yang et al., 2016a). Let $\\mathbf { E }$ denote the token embedding lookup table. We perform a matrix-vector multiplication Ew to obtain a word-level representation. We assume c and Ew have the same length $d _ { e }$ in this work. ", + "bbox": [ + 174, + 726, + 825, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous methods defined $f$ using the word-level representation Ew (Collobert et al., 2011), the character-level representation c (Ling et al., 2015), or the concatenation $\\left[ \\mathbf { E w } ; \\mathbf { c } \\right]$ (Yang et al., 2016a). Unlike these methods, we propose to use a gate to dynamically choose between the word-level and character-level representations based on the properties of the token. Let $\\mathbf { v }$ denote a feature vector that encodes these properties. In this work, we use the concatenation of named entity tags, partof-speech tags, binned document frequency vectors, and the word-level representations to form the feature vector $\\mathbf { v }$ . Let $d _ { v }$ denote the length of $\\mathbf { v }$ . ", + "bbox": [ + 174, + 789, + 825, + 887 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The gate is computed as follows: ", + "bbox": [ + 174, + 893, + 390, + 907 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/1879e1834ff91ce7e7cdea9d58cd1237d894538cc8c2a492a6efa280345bea48.jpg", + "text": "$$\n\\mathbf { g } = \\sigma ( \\mathbf { W } _ { g } \\mathbf { v } + \\mathbf { b } _ { g } )\n$$", + "text_format": "latex", + "bbox": [ + 433, + 907, + 565, + 926 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f10e95e7400bc55e85299f1e966111eeaf000699ba21922a0e9aef0ca84cfda7.jpg", + "image_caption": [ + "Figure 1: Word-character fine-grained gating. The two lookup tables are shared. “NER”, “POS”, “frequency” refer to named entity tags, part-of-speech tags, document frequency features. " + ], + "image_footnote": [], + "bbox": [ + 269, + 104, + 717, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf { W } _ { g }$ and ${ \\bf b } _ { g }$ are the model parameters with shapes $d _ { e } \\times d _ { v }$ and $d _ { e }$ , and $\\sigma$ denotes an elementwise sigmoid function. ", + "bbox": [ + 174, + 386, + 820, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final representation is computed using a fine-grained gating mechanism, ", + "bbox": [ + 174, + 421, + 676, + 438 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/44026894aa630dfb83831ba155043105ba340a4961308a1db5f0aff6f127116b.jpg", + "text": "$$\n\\mathbf { h } = f ( \\mathbf { c } , \\mathbf { w } ) = \\mathbf { g } \\odot \\mathbf { c } + ( 1 - \\mathbf { g } ) \\odot ( \\mathbf { E } \\mathbf { w } )\n$$", + "text_format": "latex", + "bbox": [ + 359, + 445, + 637, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ denotes element-wise product between two vectors. ", + "bbox": [ + 174, + 472, + 573, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An illustration of our fine-grained gating mechanism is shown in Figure 1. Intuitively speaking, when the gate $\\mathbf { g }$ has high values, more information flows from the character-level representation to the final representation; when the gate $\\mathbf { g }$ has low values, the final representation is dominated by the word-level representation. ", + "bbox": [ + 174, + 493, + 825, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Though Miyamoto & Cho (2016) also use a gate to choose between word-level and character-level representations, our method is different in two ways. First, we use a more fine-grained gating mechanism, i.e., vector gates rather than scalar gates. Second, we condition the gate on features that better reflect the properties of the token. For example, for noun phrases and entities, we would expect the gate to bias towards character-level representations because noun phrases and entities are usually less common and display richer morphological structure. Experiments show that these changes are key to the performance improvements for reading comprehension tasks. ", + "bbox": [ + 174, + 556, + 825, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach can be further generalized to a setting of multi-level networks so that we can combine multiple levels of representations using fine-grained gating mechanisms, which we leave for future work. ", + "bbox": [ + 176, + 660, + 825, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 DOCUMENT-QUERY FINE-GRAINED GATING ", + "text_level": 1, + "bbox": [ + 174, + 722, + 524, + 737 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the hidden states $\\mathbf { P } ^ { k }$ and $\\mathbf { Q } ^ { k }$ , we now describe how to compute a representation $\\mathbf { H } ^ { k }$ that encodes the interactions between the document and the query. In this section, we drop the superscript $k$ (the layer number) for notation simplicity. Let $\\mathbf { p } _ { i }$ denote the $i$ -th row of $\\mathbf { P }$ and ${ \\bf q } _ { j }$ denote the $j$ -row of $\\mathbf { Q }$ . Let $d _ { h }$ denote the lengths of $\\mathbf { p } _ { i }$ and ${ \\bf q } _ { j }$ . ", + "bbox": [ + 174, + 748, + 825, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attention-over-attention (AoA) (Cui et al., 2016) defines a dot product between each pair of tokens in the document and the query, i.e., $\\mathbf { p } _ { i } ^ { T } \\mathbf { q } _ { j }$ , followed by row-wise and column-wise softmax nonlinearities. AoA imposes pair-wise interactions between the document and the query, but using a dot product is potentially not expressive enough and hard to generalize to multi-layer networks. The gated attention (GA) reader (Dhingra et al., 2016a) defines an element-wise product as $\\mathbf { p } _ { i } \\odot \\mathbf { g } _ { i }$ where $\\mathbf { g } _ { i }$ is a gate computed by attention mechanism on the token $p _ { i }$ and the entire query. The intuition for the gate $\\mathbf { g } _ { i }$ is to attend to important information in the document. However, there is no direct pair-wise interaction between each token pair. ", + "bbox": [ + 173, + 811, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/284468038502d2ff8d8577366fc87421d42142474f6344ca41f53869482de701.jpg", + "image_caption": [ + "Figure 2: Paragraph-question fine-grained gating. " + ], + "image_footnote": [], + "bbox": [ + 176, + 103, + 800, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present a fine-grained gating method that combines the advantages of the above methods (i.e., both pairwise and element-wise). We compute the pairwise element-wise product between the hidden states in the document and the query, as shown in Figure 2. More specifically, for $p _ { i }$ and $q _ { j }$ , we have ", + "bbox": [ + 173, + 377, + 825, + 431 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/2a02adcad151b1ef48daf138830045bd50697f3b2d2dbb700ca45ecefda5a53a.jpg", + "text": "$$\n\\mathbf { I } _ { i j } = \\operatorname { t a n h } ( \\mathbf { p } _ { i } \\odot \\mathbf { q } _ { j } )\n$$", + "text_format": "latex", + "bbox": [ + 428, + 431, + 570, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where ${ \\bf q } _ { j }$ can be viewed as a gate to filter the information in $\\mathbf { p } _ { i }$ . We then use an attention mechanism over $\\mathbf { I } _ { i j }$ to output hidden states $\\mathbf { h } _ { i }$ as follows ", + "bbox": [ + 173, + 452, + 823, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/684f18d9d438d0c4c302c0fdfd67203049f99927f2fa867240af1df8d84a1418.jpg", + "text": "$$\n\\mathbf { h } _ { i } = \\sum _ { j } \\mathrm { s o f t m a x } ( \\mathbf { u } _ { h } ^ { T } \\mathbf { I } _ { i j } + \\mathbf { w } _ { i } ^ { T } \\mathbf { w } _ { j } b _ { h 1 } + b _ { h 2 } ) \\mathbf { I } _ { i j }\n$$", + "text_format": "latex", + "bbox": [ + 338, + 486, + 658, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf { u } _ { h }$ is a $d _ { v }$ -dimensional model parameter, $b _ { h 1 }$ and $b _ { h 2 }$ are scalar model parameters, $\\mathbf { w } _ { i }$ and ${ \\bf w } _ { j }$ are one-hot encodings for $p _ { i }$ and $q _ { j }$ respectively. We additionally use one-hot encodings in the attention mechanism to reinforce the matching between the same tokens since such information is not fully preserved in $\\mathbf { I } _ { i j }$ when $k$ is large. The softmax nonlinearity is applied over all $j$ ’s. The final hidden states $\\mathbf { H }$ are formed by concatenating the $\\mathbf { h } _ { i }$ ’s for each token $p _ { i }$ . ", + "bbox": [ + 174, + 529, + 825, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 619, + 326, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first present experimental results on the Twitter dataset where we can rule out the effects of different choices of network architectures, to demonstrate the effectiveness of our word-character fine-grained gating approach. Later we show experiments on more challenging datasets on reading comprehension to further show that our approach can be used to improve the performance on highlevel NLP tasks as well. ", + "bbox": [ + 173, + 650, + 825, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 EVALUATING WORD-CHARACTER GATING ON TWITTER ", + "text_level": 1, + "bbox": [ + 174, + 736, + 602, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of our word-character fine-grained gating mechanism on a social media tag prediction task. We use the Twitter dataset and follow the experimental settings in Dhingra et al. (2016b). We also use the same network architecture upon the token representations, which is an LSTM layer followed by a softmax classification layer (Dhingra et al., 2016b). The Twitter dataset consists of English tweets with at least one hashtag from Twitter. Hashtags and HTML tags have been removed from the body of the tweet, and user names and URLs are replaced with special tokens. The dataset contains 2 million tweets for training, 10K for validation and 50K for testing, with a total of 2,039 distinct hashtags. The task is to predict the hashtags of each tweet. ", + "bbox": [ + 173, + 762, + 825, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare several different methods as follows. Word char concat uses the concatenation of word-level and character-level representations as in Yang et al. (2016a); word char feat concat concatenates the word-level and character-level representations along with the features described in ", + "bbox": [ + 176, + 882, + 825, + 924 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1e9948e299167411bb5473aae70b288f9ae3451b9a8e13537e41dfaa0f65e39f.jpg", + "table_caption": [ + "Table 1: Performance on the Twitter dataset. “word” and “char” means using word-level and character-level representations respectively. " + ], + "table_footnote": [], + "table_body": "
ModelPrecision@1Recall@10Mean Rank
word (Dhingra et al., 2016b)0.2410.428133
char (Dhingra et al., 2016b)0.2840.485104
word char concat0.29610.4959105.8
word char feat concat0.29510.4974106.2
scalar gate0.29740.4982104.2
fine-grained gate0.30690.5119101.5
", + "bbox": [ + 250, + 142, + 746, + 260 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4cfc865c77165ae5c5ce61c9fe9499f3ad7205d3ab3bc654bd0f449197342325.jpg", + "table_caption": [ + "Table 2: Performance on the CBT dataset. The “GA word char concat” results are extracted from Dhingra et al. (2016a). Our results on fine-grained gating are based on a single model. “CN” and “NE” are two widely used question categories. “dev” means development set, and “test” means test set. " + ], + "table_footnote": [], + "table_body": "
ModelCN devCN testNE devNE test
GA word char concat0.7310.6960.7680.725
GA word char feat concat0.72500.69280.78150.7256
GA scalar gate0.72400.69080.78100.7260
GA fine-grained gate0.74250.70840.78900.7464
FG fine-grained gate0.75300.72040.79100.7496
Sordoni et al. (2016)0.7210.6920.7520.686
Trischler et al. (2016)0.7150.6740.7530.697
Cui et al. (2016)0.7220.6940.7780.720
Munkhdalai & Yu (2016)0.7430.7190.7820.732
Kadlec et al. (2016) ensemble0.7110.6890.7620.710
Sordoni et al. (2016) ensemble0.7410.7100.7690.720
Trischler etal. (2016) ensemble0.7360.7060.7660.718
", + "bbox": [ + 245, + 333, + 753, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Section 3.2; scalar gate uses a scalar gate similar to Miyamoto & Cho (2016) but is conditioned on the features; fine-grained gate is our method described in Section 3.2. We include word char feat concat for a fair comparison because our fine-grained gating approach also uses the token features. ", + "bbox": [ + 176, + 593, + 823, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results are shown in Table 1. We report three evaluation metrics including precision $@ 1$ , recall $@ 1 0$ , and mean rank. Our method outperforms character-level models used in Dhingra et al. (2016b) by $2 . 2 9 \\%$ , $2 . 6 9 \\%$ , and 2.5 points in terms of precision, recall and mean rank respectively. We can observe that scalar gating approach (Miyamoto & Cho, 2016) can only marginally improve over the baseline methods, while fine-grained gating methods can substantially improve model performance. Note that directly concatenating the token features with the character-level and word-level representations does not boost the performance, but using the token features to compute a gate (as done in fine-grained gating) leads to better results. This indicates that the benefit of fine-grained gating mainly comes from better modeling rather than using additional features. ", + "bbox": [ + 173, + 641, + 825, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 PERFORMANCE ON READING COMPREHENSION ", + "text_level": 1, + "bbox": [ + 174, + 785, + 544, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After investigating the effectiveness of the word-character fine-grained gating mechanism on the Twitter dataset, we now move on to a more challenging task, reading comprehension. In this section, we experiment with two datasets, the Children’s Book Test dataset (Hill et al., 2016) and the SQuAD dataset (Rajpurkar et al., 2016). ", + "bbox": [ + 176, + 811, + 825, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 CLOZE-STYLE QUESTIONS ", + "text_level": 1, + "bbox": [ + 176, + 885, + 415, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our model on cloze-style question answering benchmarks. ", + "bbox": [ + 174, + 909, + 637, + 924 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bc2d44dcd815972deda2b31e09a5734abfad4815dd67c08b0fc9d6997c0946c5.jpg", + "table_caption": [ + "Table 3: Performance on the Who Did What dataset. “dev” means development set, and “test” means test set. “WDW-R” is the relaxed version of WDW. " + ], + "table_footnote": [], + "table_body": "
ModelWDW devWDW testWDW-R devWDW-R test
0.5700.590
Kadlec et al. (2016) Chen et al. (2016)0.640一0.650
Munkhdalai & Yu (2016)0.6650.6620.6700.667
Dhingra et al. (2016a)0.7160.7120.7260.726
this paper0.7230.7170.7310.726
", + "bbox": [ + 214, + 141, + 784, + 257 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d3e004cee2ede88cd486bc6c12041115e0656f94801d13ba78306e10ff2e9d0a.jpg", + "table_caption": [ + "Table 4: Performance on the SQuAD dev set. Test set results are included in the brackets. " + ], + "table_footnote": [], + "table_body": "
ModelF1Exact Match
GA word0.5492
GA word char concat0.6695 0.68570.5639
GA word char feat concat0.69040.5711
GA scalar gate0.68500.5620
GA fine-grained gate0.69830.5804
FG fine-grained gate0.71250.5995
FG fine-grained gate + ensemble0.7341 (0.733)0.6238 (0.625)
Yu et al. (2016)
Wang & Jiang (2016)0.712 (0.710) 0.700 (0.703)0.625 (0.625) 0.591 (0.595)
", + "bbox": [ + 263, + 306, + 733, + 478 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Children’s Book Test (CBT) dataset is built from children’s books. The whole dataset has 669,343 questions for training, 8,000 for validation and 10,000 for testing. We closely follow the setting in Dhingra et al. (2016a) and incrementally add different components to see the changes in performance. For the fine-grained gating approach, we use the same hyper-parameters as in Dhingra et al. (2016a) except that we use a character-level GRU with 100 units to be of the same size as the word lookup table. The word embeddings are updated during training. ", + "bbox": [ + 173, + 511, + 825, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to different ways of combining word-level and character-level representations, we also compare two different ways of integrating documents and queries: GA refers to the gated attention reader (Dhingra et al., 2016a) and FG refers to our fine-grained gating described in Section 3.3. ", + "bbox": [ + 176, + 602, + 820, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results are reported in Table 2. We report the results on common noun (CN) questions and named entity (NE) questions, which are two widely used question categories in CBT. Our finegrained gating approach achieves new state-of-the-art performance on both settings and outperforms the current state-of-the-art results by up to $1 . 7 6 \\%$ without using ensembles. Our method outperforms the baseline GA reader by up to $2 . 4 \\%$ , which indicates the effectiveness of the fine-grained gating mechanism. Consistent with the results on the Twitter dataset, using word-character fine-grained gating can substantially improve the performance over concatenation or scalar gating. Furthermore, we can see that document-query fine-grained gating also contributes significantly to the final results. ", + "bbox": [ + 174, + 650, + 825, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also apply our fine-grained gating model to the Who Did What (WDW) dataset (Onishi et al., 2016). As shown in Table 3, our model achieves state-of-the-art results compared to strong baselines. We fix the word embeddings during training. ", + "bbox": [ + 176, + 768, + 825, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 SQUAD ", + "text_level": 1, + "bbox": [ + 176, + 828, + 285, + 843 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Stanford Question Answering Dataset (SQuAD) is a reading comprehension dataset collected recently (Rajpurkar et al., 2016). It contains 23,215 paragraphs come from 536 Wikipedia articles. Unlike other reading comprehension datasets such as CBT, the answers are a span of text rather than a single word. The dataset is partitioned into a training set ( $80 \\%$ , 87,636 question-answer pairs), a development set ( $10 \\%$ , 10,600 question-answer pairs) and a test set which is not released. ", + "bbox": [ + 176, + 853, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89ddff74b55246bf812f8cbb6330db8793bb826921f94510c5ea362754ea15ba.jpg", + "image_caption": [ + "Figure 3: Visualization of the weight matrix $\\mathbf { W } _ { g }$ . Weights for each features are averaged. Red means high and yellow means low. High weight values favor character-level representations, and low weight values favor wordlevel representations. “Organization”, “‘Person”, “Location”, and “O” are named entity tags; “DOCLEN-n” are document frequency features (larger $_ n$ means higher frequency, $n$ from 0 to 4); others are POS tags. " + ], + "image_footnote": [], + "bbox": [ + 178, + 103, + 820, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e182272ce273a6e2c7869e53e7be53700145a9a8f9ecfbc54ac0980409cc5dd8.jpg", + "image_caption": [ + "Figure 4: Visualization of gate values in the text. Red means high and yellow means low. High gate values favor character-level representations, and low gate values favor word-level representations. " + ], + "image_footnote": [], + "bbox": [ + 181, + 300, + 816, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We report our results in Table 4. “Exact match” computes the ratio of questions that are answered correctly by strict string comparison, and the F1 score is computed on the token level. We can observe that both word-character fine-grained gating and document-query fine-grained gating can substantially improve the performance, leading to state-of-the-art results among published papers. Note that at the time of submission, the best score on the leaderboard is 0.716 in exact match and 0.804 in F1 without published papers. A gap exists because our architecture described in Section 3.1 does not specifically model the answer span structure that is unique to SQuAD. In this work, we focus on this general architecture to study the effectiveness of fine-grained gating mechanisms. ", + "bbox": [ + 173, + 579, + 825, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 VISUALIZATION AND ANALYSIS ", + "text_level": 1, + "bbox": [ + 176, + 729, + 434, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We visualize the model parameter $\\mathbf { W } _ { g }$ as described in Section 3.2. For each feature, we average the corresponding weight vector in $\\mathbf { W } _ { g }$ . The results are described in Figure 3. We can see that named entities like “Organization” and noun phrases (with tags “NNP” or “NNPS”) tend to use characterlevel representations, which is consistent with human intuition because those tokens are usually infrequent or display rich morphologies. Also, DOCLEN-4, WH-adverb (“WRB”), and conjunction (“IN” and “CC”) tokens tend to use word-level representations because they appear frequently. ", + "bbox": [ + 174, + 762, + 825, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also sample random span of text from the SQuAD dataset, and visualize the average gate values in Figure 4. The results are consistent with our observations in Figure 3. Rare tokens, noun phrases, and named entities tend to use character-level representations, while others tend to use word-level representations. To further justify this argument, we also list the tokens with highest and lowest gate values in Table 5. ", + "bbox": [ + 176, + 853, + 823, + 922 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3d88f65fd35053aed1c57fb91abce4362deea2ce3a4b72dcac738b31312c5589.jpg", + "table_caption": [ + "Table 5: Word tokens with highest and lowest gate values. High gate values favor character-level representations, and low gate values favor word-level representations. " + ], + "table_footnote": [], + "table_body": "
Gate valuesWord tokens
Lowestor but But These these However however among Among that when When although Although because Because until many Many than though Though this This Since since date where Where have That and And Such such number so which by By how before Before with With between Between even Even if
HighestSweetgum Untersee Jianlong Floresta Chlorella Obersee PhT Doctorin Jumonville WFTS WTSP Boven Pharm Nederrijn Otrar Rhin Magicicada WBKB Tanzler KMBC WPLG Mainau Merwede RMJM Kleitman Scheur Bodensee Kromme Horenbout Vorderrhein Chlamydomonas Scantlebury Qingshui Funchess
", + "bbox": [ + 173, + 155, + 823, + 313 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 CONCLUSIONS ", + "text_level": 1, + "bbox": [ + 176, + 342, + 328, + 358 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We present a fine-grained gating mechanism that dynamically combines word-level and characterlevel representations based on word properties. Experiments on the Twitter tag prediction dataset show that fine-grained gating substantially outperforms scalar gating and concatenation. Our method also improves the performance on reading comprehension and achieves new state-of-the-art results on CBT and WDW. In our future work, we plan to to apply the fine-grained gating mechanism for combining other levels of representations, such as phrases and sentences. It will also be intriguing to integrate NER and POS networks and learn the token representation in an end-to-end manner. ", + "bbox": [ + 174, + 373, + 825, + 472 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ACKNOWLEDGMENTS ", + "text_level": 1, + "bbox": [ + 176, + 488, + 326, + 501 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was funded by NVIDIA, the Office of Naval Research Scene Understanding grant N000141310721, the NSF grant IIS1250956, and Google Research. ", + "bbox": [ + 174, + 512, + 821, + 540 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 561, + 285, + 575 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In ICLR, 2015. ", + "bbox": [ + 174, + 583, + 823, + 608 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Danqi Chen, Jason Bolton, and Christopher D Manning. A thorough examination of the cnn/daily mail reading comprehension task. In ACL, 2016. ", + "bbox": [ + 173, + 618, + 823, + 645 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural ¨ machine translation: Encoder-decoder approaches. In Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, 2014. ", + "bbox": [ + 174, + 655, + 823, + 694 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ronan Collobert, Jason Weston, Leon Bottou, Michael Karlen, Koray Kavukcuoglu, and Pavel Kuksa. Natural ´ language processing (almost) from scratch. Journal of Machine Learning Research, 12(Aug):2493–2537, 2011. ", + "bbox": [ + 174, + 703, + 823, + 742 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Yiming Cui, Zhipeng Chen, Si Wei, Shijin Wang, Ting Liu, and Guoping Hu. Attention-over-attention neural networks for reading comprehension. arXiv preprint arXiv:1607.04423, 2016. ", + "bbox": [ + 173, + 752, + 823, + 779 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Bhuwan Dhingra, Hanxiao Liu, Zhilin Yang, William W Cohen, and Ruslan Salakhutdinov. Gated-attention readers for text comprehension. arXiv preprint arXiv:1606.01549, 2016a. ", + "bbox": [ + 176, + 789, + 823, + 815 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Bhuwan Dhingra, Zhong Zhou, Dylan Fitzpatrick, Michael Muehl, and William W Cohen. Tweet2vec: Character-based distributed representations for social media. In ACL, 2016b. ", + "bbox": [ + 176, + 825, + 823, + 852 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. Teaching machines to read and comprehend. In NIPS, pp. 1693–1701, 2015. ", + "bbox": [ + 174, + 861, + 823, + 887 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Felix Hill, Antoine Bordes, Sumit Chopra, and Jason Weston. The goldilocks principle: Reading children’s books with explicit memory representations. In ICLR, 2016. ", + "bbox": [ + 173, + 897, + 825, + 922 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):1735–1780, 1997. ", + "bbox": [ + 173, + 103, + 823, + 131 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Rudolf Kadlec, Martin Schmid, Ondrej Bajgar, and Jan Kleindienst. Text understanding with the attention sum reader network. In ACL, 2016. ", + "bbox": [ + 174, + 140, + 823, + 166 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yoon Kim, Yacine Jernite, David Sontag, and Alexander M Rush. Character-aware neural language models. In AAAI, 2016. ", + "bbox": [ + 174, + 174, + 823, + 202 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Wang Ling, Tiago Lu´ıs, Lu´ıs Marujo, Ramon Fernandez Astudillo, Silvio Amir, Chris Dyer, Alan W Black, ´ and Isabel Trancoso. Finding function in form: Compositional character models for open vocabulary word representation. In EMNLP, 2015. ", + "bbox": [ + 174, + 209, + 823, + 250 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Minh-Thang Luong and Christopher D Manning. Achieving open vocabulary neural machine translation with hybrid word-character models. In ACL, 2016. ", + "bbox": [ + 174, + 257, + 823, + 285 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yasumasa Miyamoto and Kyunghyun Cho. Gated word-character recurrent language model. In EMNLP, 2016. ", + "bbox": [ + 171, + 292, + 820, + 308 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tsendsuren Munkhdalai and Hong Yu. Neural semantic encoders. arXiv preprint arXiv:1607.04315, 2016. ", + "bbox": [ + 173, + 315, + 813, + 330 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Takeshi Onishi, Hai Wang, Mohit Bansal, Kevin Gimpel, and David McAllester. Who did what: A large-scale person-centered cloze dataset. arXiv preprint arXiv:1608.05457, 2016. ", + "bbox": [ + 176, + 338, + 823, + 364 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad: $^ { 1 0 0 , 0 0 0 + }$ questions for machine comprehension of text. In EMNLP, 2016. ", + "bbox": [ + 174, + 372, + 823, + 400 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Marek Rei, Gamal KO Crichton, and Sampo Pyysalo. Attending to characters in neural sequence labeling models. arXiv preprint arXiv:1611.04361, 2016. ", + "bbox": [ + 174, + 409, + 821, + 435 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alessandro Sordoni, Phillip Bachman, and Yoshua Bengio. Iterative alternating neural attention for machine reading. arXiv preprint arXiv:1606.02245, 2016. ", + "bbox": [ + 173, + 443, + 823, + 470 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Adam Trischler, Zheng Ye, Xingdi Yuan, and Kaheer Suleman. Natural language comprehension with the epireader. In EMNLP, 2016. ", + "bbox": [ + 171, + 478, + 821, + 506 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In NIPS, pp. 2692–2700, 2015. ", + "bbox": [ + 176, + 513, + 779, + 529 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Shuohang Wang and Jing Jiang. Machine comprehension using match-lstm and answer pointer. arXiv preprint arXiv:1608.07905, 2016. ", + "bbox": [ + 174, + 536, + 823, + 564 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yuhuai Wu, Saizheng Zhang, Ying Zhang, Yoshua Bengio, and Ruslan Salakhutdinov. On multiplicative integration with recurrent neural networks. In NIPS, 2016. ", + "bbox": [ + 174, + 571, + 821, + 598 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. Learning multi-relational semantics using neural-embedding models. In NIPS 2014 workshop on Learning Semantics, 2014. ", + "bbox": [ + 173, + 607, + 823, + 635 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Zhilin Yang, Ruslan Salakhutdinov, and William Cohen. Multi-task cross-lingual sequence tagging from scratch. arXiv preprint arXiv:1603.06270, 2016a. ", + "bbox": [ + 173, + 642, + 823, + 670 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Zhilin Yang, Ye Yuan, Yuexin Wu, Ruslan Salakhutdinov, and William W Cohen. Review networks for caption generation. In NIPS, 2016b. ", + "bbox": [ + 173, + 678, + 823, + 705 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yang Yu, Wei Zhang, Kazi Hasan, Mo Yu, Bing Xiang, and Bowen Zhou. End-to-end answer chunk extraction and ranking for reading comprehension. arXiv preprint arXiv:1610.09996, 2016. ", + "bbox": [ + 173, + 713, + 825, + 741 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/parse/train/B1hdzd5lg/B1hdzd5lg_middle.json b/parse/train/B1hdzd5lg/B1hdzd5lg_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..2bafaa81077124daee9a297d685d44c17963f62d --- /dev/null +++ b/parse/train/B1hdzd5lg/B1hdzd5lg_middle.json @@ -0,0 +1,26148 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 77, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 106, + 77, + 505, + 98 + ], + "score": 1.0, + "content": "WORDS OR CHARACTERS? FINE-GRAINED GATING", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 348, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 348, + 118 + ], + "score": 1.0, + "content": "FOR READING COMPREHENSION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 114, + 135, + 511, + 147 + ], + "lines": [ + { + "bbox": [ + 112, + 133, + 513, + 149 + ], + "spans": [ + { + "bbox": [ + 112, + 133, + 513, + 149 + ], + "score": 1.0, + "content": "Zhilin Yang, Bhuwan Dhingra, Ye Yuan, Junjie Hu, William W. Cohen, Ruslan Salakhutdinov", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 113, + 148, + 327, + 180 + ], + "lines": [ + { + "bbox": [ + 112, + 145, + 229, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 145, + 229, + 158 + ], + "score": 1.0, + "content": "School of Computer Science", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 156, + 226, + 169 + ], + "spans": [ + { + "bbox": [ + 112, + 156, + 226, + 169 + ], + "score": 1.0, + "content": "Carnegie Mellon University", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 167, + 328, + 181 + ], + "spans": [ + { + "bbox": [ + 112, + 167, + 328, + 181 + ], + "score": 1.0, + "content": "{zhiliny,wcohen,rsalakhu}@cs.cmu.edu", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 278, + 208, + 333, + 220 + ], + "lines": [ + { + "bbox": [ + 276, + 208, + 335, + 221 + ], + "spans": [ + { + "bbox": [ + 276, + 208, + 335, + 221 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 143, + 231, + 468, + 341 + ], + "lines": [ + { + "bbox": [ + 141, + 231, + 469, + 244 + ], + "spans": [ + { + "bbox": [ + 141, + 231, + 469, + 244 + ], + "score": 1.0, + "content": "Previous work combines word-level and character-level representations using con-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 242, + 470, + 255 + ], + "spans": [ + { + "bbox": [ + 141, + 242, + 470, + 255 + ], + "score": 1.0, + "content": "catenation or scalar weighting, which is suboptimal for high-level tasks like read-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 254, + 469, + 266 + ], + "spans": [ + { + "bbox": [ + 141, + 254, + 469, + 266 + ], + "score": 1.0, + "content": "ing comprehension. We present a fine-grained gating mechanism to dynamically", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 265, + 469, + 277 + ], + "spans": [ + { + "bbox": [ + 141, + 265, + 469, + 277 + ], + "score": 1.0, + "content": "combine word-level and character-level representations based on properties of the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 275, + 469, + 288 + ], + "spans": [ + { + "bbox": [ + 141, + 275, + 469, + 288 + ], + "score": 1.0, + "content": "words. We also extend the idea of fine-grained gating to modeling the interaction", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 286, + 469, + 299 + ], + "spans": [ + { + "bbox": [ + 141, + 286, + 469, + 299 + ], + "score": 1.0, + "content": "between questions and paragraphs for reading comprehension. Experiments show", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 297, + 469, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 297, + 469, + 311 + ], + "score": 1.0, + "content": "that our approach can improve the performance on reading comprehension tasks,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "score": 1.0, + "content": "achieving new state-of-the-art results on the Children’s Book Test and Who Did", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 319, + 470, + 332 + ], + "spans": [ + { + "bbox": [ + 141, + 319, + 470, + 332 + ], + "score": 1.0, + "content": "What datasets. To demonstrate the generality of our gating mechanism, we also", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 330, + 388, + 342 + ], + "spans": [ + { + "bbox": [ + 141, + 330, + 388, + 342 + ], + "score": 1.0, + "content": "show improved results on a social media tag prediction task.1", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 11.5 + }, + { + "type": "title", + "bbox": [ + 108, + 360, + 206, + 372 + ], + "lines": [ + { + "bbox": [ + 105, + 359, + 208, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 208, + 376 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 396 + ], + "score": 1.0, + "content": "Finding semantically meaningful representations of the words (also called tokens) in a document is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 396, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 407 + ], + "score": 1.0, + "content": "necessary for strong performance in Natural Language Processing tasks. In neural networks, tokens", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "are mainly represented in two ways, either using word-level representations or character-level repre-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 418, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 505, + 429 + ], + "score": 1.0, + "content": "sentations. Word-level representations are obtained from a lookup table, where each unique token is", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "represented as a vector. Character-level representations are usually obtained by applying recurrent", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 439, + 504, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 504, + 451 + ], + "score": 1.0, + "content": "neural networks (RNNs) or convolutional neural networks (CNNs) on the character sequence of the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "score": 1.0, + "content": "token, and their hidden states are combined to form the representation. Word-level representations", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 461, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 506, + 475 + ], + "score": 1.0, + "content": "are good at memorizing the semantics of the tokens while character-level representations are more", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 472, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 485 + ], + "score": 1.0, + "content": "suitable for modeling sub-word morphologies (Ling et al., 2015; Yang et al., 2016a). For example,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 483, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "considering “cat” and “cats”, word-level representations can only learn the similarities between the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "two tokens by training on a large amount of training data, while character-level representations, by", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "design, can easily capture the similarities. Character-level representations are also used to alleviate", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 516, + 461, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 461, + 529 + ], + "score": 1.0, + "content": "the difficulties of modeling out-of-vocabulary (OOV) tokens (Luong & Manning, 2016).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 686 + ], + "lines": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "Hybrid word-character models have been proposed to leverage the advantages of both word-level", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 544, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 505, + 556 + ], + "score": 1.0, + "content": "and character-level representations. The most commonly used method is to concatenate these two", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 554, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 506, + 568 + ], + "score": 1.0, + "content": "representations (Yang et al., 2016a). However, concatenating word-level and character-level repre-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "sentations is technically problematic. For frequent tokens, the word-level representations are usually", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "accurately estimated during the training process, and thus introducing character-level representa-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "tions can potentially bias the entire representations. For infrequent tokens, the estimation of word-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 598, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 610 + ], + "score": 1.0, + "content": "level representations have high variance, which will have negative effects when combined with the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 610, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 505, + 621 + ], + "score": 1.0, + "content": "character-level representations. To address this issue, recently Miyamoto & Cho (2016) introduced", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 621, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 632 + ], + "score": 1.0, + "content": "a scalar gate conditioned on the word-level representations to control the ratio of the two repre-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 631, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 506, + 645 + ], + "score": 1.0, + "content": "sentations. However, for the task of reading comprehension, preliminary experiments showed that", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 643, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 505, + 654 + ], + "score": 1.0, + "content": "this method was not able to improve the performance over concatenation. There are two possible", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 654, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 666 + ], + "score": 1.0, + "content": "reasons. First, word-level representations might not contain sufficient information to support the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "score": 1.0, + "content": "decisions of selecting between the two representations. Second, using a scalar gate means applying", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 676, + 379, + 687 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 379, + 687 + ], + "score": 1.0, + "content": "the same ratio for each of the dimensions, which can be suboptimal.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 108, + 693, + 504, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 692, + 505, + 705 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 505, + 705 + ], + "score": 1.0, + "content": "In this work, we present a fine-grained gating mechanism to combine the word-level and character-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 703, + 505, + 716 + ], + "spans": [ + { + "bbox": [ + 106, + 703, + 505, + 716 + ], + "score": 1.0, + "content": "level representations. We compute a vector gate as a linear projection of the token features followed", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 722, + 402, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 719, + 403, + 734 + ], + "spans": [ + { + "bbox": [ + 119, + 719, + 403, + 734 + ], + "score": 1.0, + "content": "1Code is available at https://github.com/kimiyoung/fg-gating", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 77, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 106, + 77, + 505, + 98 + ], + "score": 1.0, + "content": "WORDS OR CHARACTERS? FINE-GRAINED GATING", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 348, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 348, + 118 + ], + "score": 1.0, + "content": "FOR READING COMPREHENSION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 114, + 135, + 511, + 147 + ], + "lines": [ + { + "bbox": [ + 112, + 133, + 513, + 149 + ], + "spans": [ + { + "bbox": [ + 112, + 133, + 513, + 149 + ], + "score": 1.0, + "content": "Zhilin Yang, Bhuwan Dhingra, Ye Yuan, Junjie Hu, William W. Cohen, Ruslan Salakhutdinov", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 112, + 133, + 513, + 149 + ] + }, + { + "type": "list", + "bbox": [ + 113, + 148, + 327, + 180 + ], + "lines": [ + { + "bbox": [ + 112, + 145, + 229, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 145, + 229, + 158 + ], + "score": 1.0, + "content": "School of Computer Science", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 156, + 226, + 169 + ], + "spans": [ + { + "bbox": [ + 112, + 156, + 226, + 169 + ], + "score": 1.0, + "content": "Carnegie Mellon University", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 167, + 328, + 181 + ], + "spans": [ + { + "bbox": [ + 112, + 167, + 328, + 181 + ], + "score": 1.0, + "content": "{zhiliny,wcohen,rsalakhu}@cs.cmu.edu", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + } + ], + "index": 4, + "bbox_fs": [ + 112, + 145, + 328, + 181 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 208, + 333, + 220 + ], + "lines": [ + { + "bbox": [ + 276, + 208, + 335, + 221 + ], + "spans": [ + { + "bbox": [ + 276, + 208, + 335, + 221 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 143, + 231, + 468, + 341 + ], + "lines": [ + { + "bbox": [ + 141, + 231, + 469, + 244 + ], + "spans": [ + { + "bbox": [ + 141, + 231, + 469, + 244 + ], + "score": 1.0, + "content": "Previous work combines word-level and character-level representations using con-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 242, + 470, + 255 + ], + "spans": [ + { + "bbox": [ + 141, + 242, + 470, + 255 + ], + "score": 1.0, + "content": "catenation or scalar weighting, which is suboptimal for high-level tasks like read-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 254, + 469, + 266 + ], + "spans": [ + { + "bbox": [ + 141, + 254, + 469, + 266 + ], + "score": 1.0, + "content": "ing comprehension. We present a fine-grained gating mechanism to dynamically", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 265, + 469, + 277 + ], + "spans": [ + { + "bbox": [ + 141, + 265, + 469, + 277 + ], + "score": 1.0, + "content": "combine word-level and character-level representations based on properties of the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 275, + 469, + 288 + ], + "spans": [ + { + "bbox": [ + 141, + 275, + 469, + 288 + ], + "score": 1.0, + "content": "words. We also extend the idea of fine-grained gating to modeling the interaction", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 286, + 469, + 299 + ], + "spans": [ + { + "bbox": [ + 141, + 286, + 469, + 299 + ], + "score": 1.0, + "content": "between questions and paragraphs for reading comprehension. Experiments show", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 297, + 469, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 297, + 469, + 311 + ], + "score": 1.0, + "content": "that our approach can improve the performance on reading comprehension tasks,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "score": 1.0, + "content": "achieving new state-of-the-art results on the Children’s Book Test and Who Did", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 319, + 470, + 332 + ], + "spans": [ + { + "bbox": [ + 141, + 319, + 470, + 332 + ], + "score": 1.0, + "content": "What datasets. To demonstrate the generality of our gating mechanism, we also", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 330, + 388, + 342 + ], + "spans": [ + { + "bbox": [ + 141, + 330, + 388, + 342 + ], + "score": 1.0, + "content": "show improved results on a social media tag prediction task.1", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 11.5, + "bbox_fs": [ + 141, + 231, + 470, + 342 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 360, + 206, + 372 + ], + "lines": [ + { + "bbox": [ + 105, + 359, + 208, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 208, + 376 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 396 + ], + "score": 1.0, + "content": "Finding semantically meaningful representations of the words (also called tokens) in a document is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 396, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 407 + ], + "score": 1.0, + "content": "necessary for strong performance in Natural Language Processing tasks. In neural networks, tokens", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "are mainly represented in two ways, either using word-level representations or character-level repre-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 418, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 505, + 429 + ], + "score": 1.0, + "content": "sentations. Word-level representations are obtained from a lookup table, where each unique token is", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "represented as a vector. Character-level representations are usually obtained by applying recurrent", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 439, + 504, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 504, + 451 + ], + "score": 1.0, + "content": "neural networks (RNNs) or convolutional neural networks (CNNs) on the character sequence of the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "score": 1.0, + "content": "token, and their hidden states are combined to form the representation. Word-level representations", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 461, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 506, + 475 + ], + "score": 1.0, + "content": "are good at memorizing the semantics of the tokens while character-level representations are more", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 472, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 485 + ], + "score": 1.0, + "content": "suitable for modeling sub-word morphologies (Ling et al., 2015; Yang et al., 2016a). For example,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 483, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "considering “cat” and “cats”, word-level representations can only learn the similarities between the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "two tokens by training on a large amount of training data, while character-level representations, by", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "design, can easily capture the similarities. Character-level representations are also used to alleviate", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 516, + 461, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 461, + 529 + ], + "score": 1.0, + "content": "the difficulties of modeling out-of-vocabulary (OOV) tokens (Luong & Manning, 2016).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 385, + 506, + 529 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 686 + ], + "lines": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "Hybrid word-character models have been proposed to leverage the advantages of both word-level", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 544, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 505, + 556 + ], + "score": 1.0, + "content": "and character-level representations. The most commonly used method is to concatenate these two", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 554, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 506, + 568 + ], + "score": 1.0, + "content": "representations (Yang et al., 2016a). However, concatenating word-level and character-level repre-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "sentations is technically problematic. For frequent tokens, the word-level representations are usually", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "accurately estimated during the training process, and thus introducing character-level representa-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "tions can potentially bias the entire representations. For infrequent tokens, the estimation of word-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 598, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 610 + ], + "score": 1.0, + "content": "level representations have high variance, which will have negative effects when combined with the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 610, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 505, + 621 + ], + "score": 1.0, + "content": "character-level representations. To address this issue, recently Miyamoto & Cho (2016) introduced", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 621, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 632 + ], + "score": 1.0, + "content": "a scalar gate conditioned on the word-level representations to control the ratio of the two repre-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 631, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 506, + 645 + ], + "score": 1.0, + "content": "sentations. However, for the task of reading comprehension, preliminary experiments showed that", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 643, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 505, + 654 + ], + "score": 1.0, + "content": "this method was not able to improve the performance over concatenation. There are two possible", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 654, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 666 + ], + "score": 1.0, + "content": "reasons. First, word-level representations might not contain sufficient information to support the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "score": 1.0, + "content": "decisions of selecting between the two representations. Second, using a scalar gate means applying", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 676, + 379, + 687 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 379, + 687 + ], + "score": 1.0, + "content": "the same ratio for each of the dimensions, which can be suboptimal.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 532, + 506, + 687 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 693, + 504, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 692, + 505, + 705 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 505, + 705 + ], + "score": 1.0, + "content": "In this work, we present a fine-grained gating mechanism to combine the word-level and character-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 703, + 505, + 716 + ], + "spans": [ + { + "bbox": [ + 106, + 703, + 505, + 716 + ], + "score": 1.0, + "content": "level representations. We compute a vector gate as a linear projection of the token features followed", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "by a sigmoid activation. We then multiplicatively apply the gate to the character-level and word-", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 506, + 106 + ], + "score": 1.0, + "content": "level representations. Each dimension of the gate controls how much information is flowed from", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "the word-level and character-level representations respectively. We use named entity tags, part-of-", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "speech tags, document frequencies, and word-level representations as the features for token proper-", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "ties which determine the gate. More generally, our fine-grained gating mechanism can be used to", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "model multiple levels of structure in language, including words, characters, phrases, sentences and", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 149, + 442, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 442, + 162 + ], + "score": 1.0, + "content": "paragraphs. In this work we focus on studying the effects on word-character gating.", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 45.5, + "bbox_fs": [ + 106, + 692, + 505, + 716 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 160 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "by a sigmoid activation. We then multiplicatively apply the gate to the character-level and word-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 506, + 106 + ], + "score": 1.0, + "content": "level representations. Each dimension of the gate controls how much information is flowed from", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "the word-level and character-level representations respectively. We use named entity tags, part-of-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "speech tags, document frequencies, and word-level representations as the features for token proper-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "ties which determine the gate. More generally, our fine-grained gating mechanism can be used to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "model multiple levels of structure in language, including words, characters, phrases, sentences and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 149, + 442, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 442, + 162 + ], + "score": 1.0, + "content": "paragraphs. In this work we focus on studying the effects on word-character gating.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 165, + 505, + 286 + ], + "lines": [ + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "score": 1.0, + "content": "To better tackle the problem of reading comprehension, we also extend the idea of fine-grained", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "gating for modeling the interaction between documents and queries. Previous work has shown the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "importance of modeling interactions between document and query tokens by introducing various", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "attention architectures for the task (Hermann et al., 2015; Kadlec et al., 2016). Most of these use", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 221 + ], + "score": 1.0, + "content": "an inner product between the two representations to compute the relative importance of document", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "tokens. The Gated-Attention Reader (Dhingra et al., 2016a) showed improved performance by re-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "placing the inner-product with an element-wise product to allow for better matching at the semantic", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "level. However, they use aggregated representations of the query which may lead to loss of infor-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "score": 1.0, + "content": "mation. In this work we use a fine-grained gating mechanism for each token in the paragraph and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 264, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 277 + ], + "score": 1.0, + "content": "each token in the query. The fine-grained gating mechanism applies an element-wise multiplication", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 274, + 215, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 215, + 288 + ], + "score": 1.0, + "content": "of the two representations.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 291, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "We show improved performance on reading comprehension datasets, including Children’s Book Test", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "(CBT), Who Did What, and SQuAD. On CBT, our approach achieves new state-of-the-art results", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "without using an ensemble. Our model also improves over state-of-the-art results on the Who Did", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "score": 1.0, + "content": "What dataset. To demonstrate the generality of our method, we apply our word-character fine-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "grained gating mechanism to a social media tag prediction task and show improved performance", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 347, + 202, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 202, + 358 + ], + "score": 1.0, + "content": "over previous methods.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 108, + 363, + 504, + 407 + ], + "lines": [ + { + "bbox": [ + 106, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "Our contributions are two-fold. First, we present a fine-grained word-character gating mechanism", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 375, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 505, + 387 + ], + "score": 1.0, + "content": "and show improved performance on a variety of tasks including reading comprehension. Second,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "to better tackle the reading comprehension tasks, we extend our fine-grained gating approach to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 397, + 338, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 338, + 408 + ], + "score": 1.0, + "content": "modeling the interaction between documents and queries.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + }, + { + "type": "title", + "bbox": [ + 108, + 427, + 209, + 439 + ], + "lines": [ + { + "bbox": [ + 105, + 426, + 210, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 210, + 442 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "Hybrid word-character models have been proposed to take advantages of both word-level and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 461, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 506, + 474 + ], + "score": 1.0, + "content": "character-level representations. Ling et al. (2015) introduce a compositional character to word", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 471, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 506, + 485 + ], + "score": 1.0, + "content": "(C2W) model based on bidirectional LSTMs. Kim et al. (2016) describe a model that employs a", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "convolutional neural network (CNN) and a highway network over characters for language model-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "ing. Miyamoto & Cho (2016) use a gate to adaptively find the optimal mixture of the character-level", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 506, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 518 + ], + "score": 1.0, + "content": "and word-level inputs. Yang et al. (2016a) employ deep gated recurrent units on both character", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "and word levels to encode morphology and context information. Concurrent to our work, Rei et al.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "score": 1.0, + "content": "(2016) employed a similar gating idea to combine word-level and character-level representations,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 537, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 552 + ], + "score": 1.0, + "content": "but their focus is on low-level sequence tagging tasks and the gate is not conditioned on linguistic", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 548, + 144, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 144, + 562 + ], + "score": 1.0, + "content": "features.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 566, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "The gating mechanism is widely used in sequence modeling. Long short-term memory (LSTM)", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "networks (Hochreiter & Schmidhuber, 1997) are designed to deal with vanishing gradients through", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "the gating mechanism. Similar to LSTM, Gated Recurrent Unit (GRU) was proposed by Cho et al.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "(2014), which also uses gating units to modulate the flow of information. The gating mechanism", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 610, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 623 + ], + "score": 1.0, + "content": "can also be viewed as a form of attention mechanism (Bahdanau et al., 2015; Yang et al., 2016b)", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 622, + 173, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 173, + 634 + ], + "score": 1.0, + "content": "over two inputs.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 504, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "Similar to the idea of gating, multiplicative integration has also been shown to provide a benefit", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "in various settings. Yang et al. (2014) find that multiplicative operations are superior to additive", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "operations in modeling relations. Wu et al. (2016) propose to use Hadamard product to replace sum", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 671, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 683 + ], + "score": 1.0, + "content": "operation in recurrent networks, which gives a significant performance boost over existing RNN", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 681, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 694 + ], + "score": 1.0, + "content": "models. Dhingra et al. (2016a) use a multiplicative gating mechanism to achieve state-of-the-art", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 693, + 280, + 705 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 280, + 705 + ], + "score": 1.0, + "content": "results on question answering benchmarks.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 47.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Reading comprehension is a challenging task for machines. A variety of models have been proposed", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "to extract answers from given text (Hill et al., 2016; Kadlec et al., 2016; Trischler et al., 2016; Chen", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 51.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 160 + ], + "lines": [], + "index": 3, + "bbox_fs": [ + 105, + 82, + 506, + 162 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 165, + 505, + 286 + ], + "lines": [ + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "score": 1.0, + "content": "To better tackle the problem of reading comprehension, we also extend the idea of fine-grained", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "gating for modeling the interaction between documents and queries. Previous work has shown the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "importance of modeling interactions between document and query tokens by introducing various", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "attention architectures for the task (Hermann et al., 2015; Kadlec et al., 2016). Most of these use", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 221 + ], + "score": 1.0, + "content": "an inner product between the two representations to compute the relative importance of document", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "tokens. The Gated-Attention Reader (Dhingra et al., 2016a) showed improved performance by re-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "placing the inner-product with an element-wise product to allow for better matching at the semantic", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "level. However, they use aggregated representations of the query which may lead to loss of infor-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "score": 1.0, + "content": "mation. In this work we use a fine-grained gating mechanism for each token in the paragraph and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 264, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 277 + ], + "score": 1.0, + "content": "each token in the query. The fine-grained gating mechanism applies an element-wise multiplication", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 274, + 215, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 215, + 288 + ], + "score": 1.0, + "content": "of the two representations.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 164, + 506, + 288 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 291, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "We show improved performance on reading comprehension datasets, including Children’s Book Test", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "(CBT), Who Did What, and SQuAD. On CBT, our approach achieves new state-of-the-art results", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "without using an ensemble. Our model also improves over state-of-the-art results on the Who Did", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "score": 1.0, + "content": "What dataset. To demonstrate the generality of our method, we apply our word-character fine-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "grained gating mechanism to a social media tag prediction task and show improved performance", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 347, + 202, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 202, + 358 + ], + "score": 1.0, + "content": "over previous methods.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 291, + 506, + 358 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 363, + 504, + 407 + ], + "lines": [ + { + "bbox": [ + 106, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "Our contributions are two-fold. First, we present a fine-grained word-character gating mechanism", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 375, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 505, + 387 + ], + "score": 1.0, + "content": "and show improved performance on a variety of tasks including reading comprehension. Second,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "to better tackle the reading comprehension tasks, we extend our fine-grained gating approach to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 397, + 338, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 338, + 408 + ], + "score": 1.0, + "content": "modeling the interaction between documents and queries.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 363, + 505, + 408 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 427, + 209, + 439 + ], + "lines": [ + { + "bbox": [ + 105, + 426, + 210, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 210, + 442 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "Hybrid word-character models have been proposed to take advantages of both word-level and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 461, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 506, + 474 + ], + "score": 1.0, + "content": "character-level representations. Ling et al. (2015) introduce a compositional character to word", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 471, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 506, + 485 + ], + "score": 1.0, + "content": "(C2W) model based on bidirectional LSTMs. Kim et al. (2016) describe a model that employs a", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "convolutional neural network (CNN) and a highway network over characters for language model-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "ing. Miyamoto & Cho (2016) use a gate to adaptively find the optimal mixture of the character-level", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 506, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 518 + ], + "score": 1.0, + "content": "and word-level inputs. Yang et al. (2016a) employ deep gated recurrent units on both character", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "and word levels to encode morphology and context information. Concurrent to our work, Rei et al.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "score": 1.0, + "content": "(2016) employed a similar gating idea to combine word-level and character-level representations,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 537, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 552 + ], + "score": 1.0, + "content": "but their focus is on low-level sequence tagging tasks and the gate is not conditioned on linguistic", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 548, + 144, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 144, + 562 + ], + "score": 1.0, + "content": "features.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 450, + 506, + 562 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 566, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "The gating mechanism is widely used in sequence modeling. Long short-term memory (LSTM)", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "networks (Hochreiter & Schmidhuber, 1997) are designed to deal with vanishing gradients through", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "the gating mechanism. Similar to LSTM, Gated Recurrent Unit (GRU) was proposed by Cho et al.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "(2014), which also uses gating units to modulate the flow of information. The gating mechanism", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 610, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 623 + ], + "score": 1.0, + "content": "can also be viewed as a form of attention mechanism (Bahdanau et al., 2015; Yang et al., 2016b)", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 622, + 173, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 173, + 634 + ], + "score": 1.0, + "content": "over two inputs.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 565, + 505, + 634 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 504, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "Similar to the idea of gating, multiplicative integration has also been shown to provide a benefit", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "in various settings. Yang et al. (2014) find that multiplicative operations are superior to additive", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "operations in modeling relations. Wu et al. (2016) propose to use Hadamard product to replace sum", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 671, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 683 + ], + "score": 1.0, + "content": "operation in recurrent networks, which gives a significant performance boost over existing RNN", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 681, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 694 + ], + "score": 1.0, + "content": "models. Dhingra et al. (2016a) use a multiplicative gating mechanism to achieve state-of-the-art", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 693, + 280, + 705 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 280, + 705 + ], + "score": 1.0, + "content": "results on question answering benchmarks.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 47.5, + "bbox_fs": [ + 105, + 638, + 506, + 705 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Reading comprehension is a challenging task for machines. A variety of models have been proposed", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "to extract answers from given text (Hill et al., 2016; Kadlec et al., 2016; Trischler et al., 2016; Chen", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "et al., 2016; Sordoni et al., 2016; Cui et al., 2016). Yu et al. (2016) propose a dynamic chunk reader", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 92, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 104, + 92, + 506, + 107 + ], + "score": 1.0, + "content": "to extract and rank a set of answer candidates from a given document to answer questions. Wang", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "& Jiang (2016) introduce an end-to-end neural architecture which incorporates match-LSTM and", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 116, + 264, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 264, + 127 + ], + "score": 1.0, + "content": "pointer networks (Vinyals et al., 2015).", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 51.5, + "bbox_fs": [ + 105, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "et al., 2016; Sordoni et al., 2016; Cui et al., 2016). Yu et al. (2016) propose a dynamic chunk reader", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 92, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 104, + 92, + 506, + 107 + ], + "score": 1.0, + "content": "to extract and rank a set of answer candidates from a given document to answer questions. Wang", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "& Jiang (2016) introduce an end-to-end neural architecture which incorporates match-LSTM and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 116, + 264, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 264, + 127 + ], + "score": 1.0, + "content": "pointer networks (Vinyals et al., 2015).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 108, + 144, + 248, + 156 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 250, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 250, + 158 + ], + "score": 1.0, + "content": "3 FINE-GRAINED GATING", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 169, + 505, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 169, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 505, + 182 + ], + "score": 1.0, + "content": "In this section, we will describe our fine-grained gating approach in the context of reading com-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 180, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 506, + 194 + ], + "score": 1.0, + "content": "prehension. We first introduce the settings of reading comprehension tasks and a general neural", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 191, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 206 + ], + "score": 1.0, + "content": "network architecture. We will then describe our word-character gating and document-query gating", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 203, + 206, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 206, + 215 + ], + "score": 1.0, + "content": "approaches respectively.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 109, + 228, + 290, + 240 + ], + "lines": [ + { + "bbox": [ + 106, + 228, + 291, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 291, + 241 + ], + "score": 1.0, + "content": "3.1 READING COMPREHENSION SETTING", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 249, + 505, + 316 + ], + "lines": [ + { + "bbox": [ + 105, + 248, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 332, + 263 + ], + "score": 1.0, + "content": "The reading comprehension task involves a document", + "type": "text" + }, + { + "bbox": [ + 332, + 250, + 426, + 262 + ], + "score": 0.91, + "content": "P = ( p _ { 1 } , p _ { 2 } , \\cdot \\cdot \\cdot , p _ { M } )", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 248, + 482, + 263 + ], + "score": 1.0, + "content": "and a query", + "type": "text" + }, + { + "bbox": [ + 482, + 250, + 505, + 262 + ], + "score": 0.87, + "content": "Q =", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 260, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 107, + 261, + 172, + 273 + ], + "score": 0.91, + "content": "\\left( q _ { 1 } , q _ { 2 } , \\cdots , q _ { N } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 260, + 202, + 274 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 203, + 261, + 214, + 271 + ], + "score": 0.81, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 260, + 233, + 274 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 233, + 261, + 243, + 271 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 260, + 506, + 274 + ], + "score": 1.0, + "content": "are the lengths of the document and the query respectively. Each", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 131, + 284 + ], + "score": 1.0, + "content": "token", + "type": "text" + }, + { + "bbox": [ + 131, + 274, + 141, + 284 + ], + "score": 0.84, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 272, + 197, + 284 + ], + "score": 1.0, + "content": "is denoted as", + "type": "text" + }, + { + "bbox": [ + 197, + 272, + 232, + 284 + ], + "score": 0.93, + "content": "( \\mathbf { w } _ { i } , \\mathbf { C } _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 272, + 263, + 284 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 264, + 273, + 276, + 283 + ], + "score": 0.88, + "content": "\\mathbf { w } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "is a one-hot encoding of the token in the vocabulary and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 119, + 294 + ], + "score": 0.88, + "content": "\\mathbf { C } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "is a matrix with each row representing a one-hot encoding of a character. Each token in the query", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 293, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 116, + 306 + ], + "score": 0.86, + "content": "q _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 293, + 233, + 307 + ], + "score": 1.0, + "content": "is similarly defined. We use", + "type": "text" + }, + { + "bbox": [ + 234, + 295, + 239, + 304 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 293, + 375, + 307 + ], + "score": 1.0, + "content": "as a subscript for documents and", + "type": "text" + }, + { + "bbox": [ + 375, + 294, + 381, + 306 + ], + "score": 0.83, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 293, + 506, + 307 + ], + "score": 1.0, + "content": "for queries. The output of the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 305, + 470, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 194, + 317 + ], + "score": 1.0, + "content": "problem is an answer", + "type": "text" + }, + { + "bbox": [ + 194, + 307, + 200, + 315 + ], + "score": 0.71, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 305, + 470, + 317 + ], + "score": 1.0, + "content": ", which can either be an index or a span of indices in the document.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 321, + 505, + 399 + ], + "lines": [ + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "score": 1.0, + "content": "Now we describe a general architecture used in this work, which is a generalization of the gated", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 332, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 505, + 345 + ], + "score": 1.0, + "content": "attention reader (Dhingra et al., 2016a). For each token in the document and the query, we compute", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 344, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 104, + 344, + 271, + 356 + ], + "score": 1.0, + "content": "a vector representation using a function", + "type": "text" + }, + { + "bbox": [ + 271, + 344, + 278, + 355 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 344, + 422, + 356 + ], + "score": 1.0, + "content": ". More specifically, for each token", + "type": "text" + }, + { + "bbox": [ + 422, + 345, + 432, + 355 + ], + "score": 0.84, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 344, + 505, + 356 + ], + "score": 1.0, + "content": "in the document,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 144, + 367 + ], + "score": 1.0, + "content": "we have", + "type": "text" + }, + { + "bbox": [ + 144, + 354, + 214, + 367 + ], + "score": 0.93, + "content": "\\mathbf { h } _ { i } ^ { \\bar { 0 } } = f ( \\mathbf { w } _ { i } , \\mathbf { C } _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 354, + 302, + 367 + ], + "score": 1.0, + "content": ". The same function", + "type": "text" + }, + { + "bbox": [ + 302, + 355, + 309, + 366 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "is also applied to the tokens in the query. Let", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 365, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 122, + 379 + ], + "score": 0.9, + "content": "\\mathbf { H } _ { p } ^ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 365, + 140, + 378 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 140, + 366, + 155, + 378 + ], + "score": 0.87, + "content": "\\mathbf { H } _ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 365, + 347, + 378 + ], + "score": 1.0, + "content": "denote the vector representations computed by", + "type": "text" + }, + { + "bbox": [ + 347, + 366, + 354, + 377 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "for tokens in documents and queries", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "respectively. In Section 3.2, we will discuss the “word-character” fine-grained gating used to define", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 387, + 169, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 157, + 401 + ], + "score": 1.0, + "content": "the function", + "type": "text" + }, + { + "bbox": [ + 157, + 388, + 164, + 399 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 387, + 169, + 401 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 404, + 505, + 462 + ], + "lines": [ + { + "bbox": [ + 105, + 402, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 247, + 418 + ], + "score": 1.0, + "content": "Suppose that we have a network of", + "type": "text" + }, + { + "bbox": [ + 247, + 405, + 258, + 414 + ], + "score": 0.82, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 402, + 315, + 418 + ], + "score": 1.0, + "content": "layers. At the", + "type": "text" + }, + { + "bbox": [ + 315, + 405, + 321, + 414 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 402, + 435, + 418 + ], + "score": 1.0, + "content": "-th layer, we apply RNNs on", + "type": "text" + }, + { + "bbox": [ + 436, + 403, + 461, + 418 + ], + "score": 0.93, + "content": "\\mathbf { H } _ { p } ^ { k - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 402, + 479, + 418 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 479, + 404, + 494, + 417 + ], + "score": 0.88, + "content": "\\mathbf { H } _ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 402, + 506, + 418 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 414, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 104, + 414, + 187, + 432 + ], + "score": 1.0, + "content": "obtain hidden states", + "type": "text" + }, + { + "bbox": [ + 187, + 416, + 201, + 428 + ], + "score": 0.91, + "content": "\\mathbf { P } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 414, + 219, + 432 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 219, + 416, + 233, + 429 + ], + "score": 0.9, + "content": "\\mathbf { Q } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 414, + 263, + 432 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 263, + 416, + 277, + 428 + ], + "score": 0.89, + "content": "\\mathbf { P } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 414, + 294, + 432 + ], + "score": 1.0, + "content": "is a", + "type": "text" + }, + { + "bbox": [ + 294, + 417, + 321, + 428 + ], + "score": 0.91, + "content": "M \\times d", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 414, + 367, + 432 + ], + "score": 1.0, + "content": "matrix and", + "type": "text" + }, + { + "bbox": [ + 367, + 417, + 382, + 429 + ], + "score": 0.91, + "content": "\\mathbf { Q } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 414, + 398, + 432 + ], + "score": 1.0, + "content": "is a", + "type": "text" + }, + { + "bbox": [ + 398, + 417, + 424, + 428 + ], + "score": 0.9, + "content": "N \\times d", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 414, + 473, + 432 + ], + "score": 1.0, + "content": "matrix with", + "type": "text" + }, + { + "bbox": [ + 473, + 418, + 479, + 428 + ], + "score": 0.8, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 414, + 506, + 432 + ], + "score": 1.0, + "content": "being", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 366, + 441 + ], + "score": 1.0, + "content": "the number of hidden units in the RNNs. Then we use a function", + "type": "text" + }, + { + "bbox": [ + 366, + 431, + 372, + 438 + ], + "score": 0.73, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "to compute a new representation", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 437, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 178, + 453 + ], + "score": 1.0, + "content": "for the document", + "type": "text" + }, + { + "bbox": [ + 178, + 439, + 251, + 453 + ], + "score": 0.93, + "content": "\\mathbf { H } _ { p } ^ { k } = r ( \\mathbf { P } ^ { k } , \\mathbf { Q } ^ { k } )", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 437, + 506, + 453 + ], + "score": 1.0, + "content": ". In Section 3.3, we will introduce the “document-query” fine-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 450, + 285, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 274, + 464 + ], + "score": 1.0, + "content": "grained gating used to define the function", + "type": "text" + }, + { + "bbox": [ + 275, + 453, + 280, + 460 + ], + "score": 0.76, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 450, + 285, + 464 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 108, + 467, + 504, + 501 + ], + "lines": [ + { + "bbox": [ + 106, + 467, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 467, + 191, + 479 + ], + "score": 1.0, + "content": "After going through", + "type": "text" + }, + { + "bbox": [ + 192, + 468, + 202, + 477 + ], + "score": 0.81, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 467, + 353, + 479 + ], + "score": 1.0, + "content": "layers, we predict the answer index", + "type": "text" + }, + { + "bbox": [ + 354, + 470, + 360, + 477 + ], + "score": 0.63, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 467, + 505, + 479 + ], + "score": 1.0, + "content": "using a softmax layer over hidden", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 477, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 131, + 491 + ], + "score": 1.0, + "content": "states", + "type": "text" + }, + { + "bbox": [ + 131, + 478, + 146, + 492 + ], + "score": 0.89, + "content": "\\bar { \\mathbf { H } } _ { p } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 477, + 506, + 491 + ], + "score": 1.0, + "content": ". For datasets where the answer is a span of text, we use two softmax layers for the start", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 488, + 223, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 223, + 502 + ], + "score": 1.0, + "content": "and end indices respectively.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "title", + "bbox": [ + 107, + 515, + 322, + 527 + ], + "lines": [ + { + "bbox": [ + 105, + 514, + 323, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 323, + 528 + ], + "score": 1.0, + "content": "3.2 WORD-CHARACTER FINE-GRAINED GATING", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 109, + 536, + 503, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 214, + 549 + ], + "score": 1.0, + "content": "Given a one-hot encoding", + "type": "text" + }, + { + "bbox": [ + 214, + 538, + 227, + 547 + ], + "score": 0.86, + "content": "\\mathbf { w } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 536, + 333, + 549 + ], + "score": 1.0, + "content": "and a character sequence", + "type": "text" + }, + { + "bbox": [ + 333, + 537, + 346, + 548 + ], + "score": 0.88, + "content": "\\mathbf { C } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 536, + 505, + 549 + ], + "score": 1.0, + "content": ", we now describe how to compute the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 548, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 194, + 560 + ], + "score": 1.0, + "content": "vector representation", + "type": "text" + }, + { + "bbox": [ + 195, + 548, + 263, + 560 + ], + "score": 0.93, + "content": "\\mathbf { h } _ { i } = f ( \\mathbf { w } _ { i } , \\mathbf { C } _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 548, + 505, + 560 + ], + "score": 1.0, + "content": "for the token. In the rest of the section, we will drop the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 558, + 244, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 145, + 571 + ], + "score": 1.0, + "content": "subscript", + "type": "text" + }, + { + "bbox": [ + 145, + 559, + 150, + 568 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 558, + 244, + 571 + ], + "score": 1.0, + "content": "for notation simplicity.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 575, + 505, + 619 + ], + "lines": [ + { + "bbox": [ + 106, + 575, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 214, + 587 + ], + "score": 1.0, + "content": "We first apply an RNN on", + "type": "text" + }, + { + "bbox": [ + 214, + 576, + 225, + 586 + ], + "score": 0.51, + "content": "\\mathbf { C }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 575, + 505, + 587 + ], + "score": 1.0, + "content": "and take the hidden state in the last time step c as the character-level", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 266, + 599 + ], + "score": 1.0, + "content": "representation (Yang et al., 2016a). Let", + "type": "text" + }, + { + "bbox": [ + 266, + 587, + 275, + 597 + ], + "score": 0.47, + "content": "\\mathbf { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 586, + 506, + 599 + ], + "score": 1.0, + "content": "denote the token embedding lookup table. We perform a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "matrix-vector multiplication Ew to obtain a word-level representation. We assume c and Ew have", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 609, + 236, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 171, + 620 + ], + "score": 1.0, + "content": "the same length", + "type": "text" + }, + { + "bbox": [ + 172, + 609, + 182, + 619 + ], + "score": 0.88, + "content": "d _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 609, + 236, + 620 + ], + "score": 1.0, + "content": "in this work.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 625, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 625, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 214, + 637 + ], + "score": 1.0, + "content": "Previous methods defined", + "type": "text" + }, + { + "bbox": [ + 214, + 626, + 222, + 637 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 625, + 505, + 637 + ], + "score": 1.0, + "content": "using the word-level representation Ew (Collobert et al., 2011), the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 636, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 390, + 649 + ], + "score": 1.0, + "content": "character-level representation c (Ling et al., 2015), or the concatenation", + "type": "text" + }, + { + "bbox": [ + 390, + 636, + 423, + 648 + ], + "score": 0.25, + "content": "\\left[ \\mathbf { E w } ; \\mathbf { c } \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 636, + 505, + 649 + ], + "score": 1.0, + "content": "(Yang et al., 2016a).", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 647, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 506, + 660 + ], + "score": 1.0, + "content": "Unlike these methods, we propose to use a gate to dynamically choose between the word-level and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 658, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 400, + 671 + ], + "score": 1.0, + "content": "character-level representations based on the properties of the token. Let", + "type": "text" + }, + { + "bbox": [ + 401, + 660, + 408, + 668 + ], + "score": 0.36, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 658, + 506, + 671 + ], + "score": 1.0, + "content": "denote a feature vector", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 669, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 505, + 682 + ], + "score": 1.0, + "content": "that encodes these properties. In this work, we use the concatenation of named entity tags, part-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 680, + 504, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 504, + 692 + ], + "score": 1.0, + "content": "of-speech tags, binned document frequency vectors, and the word-level representations to form the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 691, + 297, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 164, + 703 + ], + "score": 1.0, + "content": "feature vector", + "type": "text" + }, + { + "bbox": [ + 164, + 693, + 171, + 701 + ], + "score": 0.26, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 691, + 191, + 703 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 191, + 691, + 203, + 702 + ], + "score": 0.89, + "content": "d _ { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 691, + 286, + 703 + ], + "score": 1.0, + "content": "denote the length of", + "type": "text" + }, + { + "bbox": [ + 286, + 693, + 293, + 701 + ], + "score": 0.61, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 691, + 297, + 703 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 708, + 239, + 719 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 240, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 240, + 721 + ], + "score": 1.0, + "content": "The gate is computed as follows:", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + }, + { + "type": "interline_equation", + "bbox": [ + 265, + 719, + 346, + 734 + ], + "lines": [ + { + "bbox": [ + 265, + 719, + 346, + 734 + ], + "spans": [ + { + "bbox": [ + 265, + 719, + 346, + 734 + ], + "score": 0.91, + "content": "\\mathbf { g } = \\sigma ( \\mathbf { W } _ { g } \\mathbf { v } + \\mathbf { b } _ { g } )", + "type": "interline_equation", + "image_path": "1879e1834ff91ce7e7cdea9d58cd1237d894538cc8c2a492a6efa280345bea48.jpg" + } + ] + } + ], + "index": 47, + "virtual_lines": [ + { + "bbox": [ + 265, + 719, + 346, + 734 + ], + "spans": [], + "index": 47 + } + ] + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 104, + 82, + 506, + 127 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 144, + 248, + 156 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 250, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 250, + 158 + ], + "score": 1.0, + "content": "3 FINE-GRAINED GATING", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 169, + 505, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 169, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 505, + 182 + ], + "score": 1.0, + "content": "In this section, we will describe our fine-grained gating approach in the context of reading com-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 180, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 506, + 194 + ], + "score": 1.0, + "content": "prehension. We first introduce the settings of reading comprehension tasks and a general neural", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 191, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 206 + ], + "score": 1.0, + "content": "network architecture. We will then describe our word-character gating and document-query gating", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 203, + 206, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 206, + 215 + ], + "score": 1.0, + "content": "approaches respectively.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 169, + 506, + 215 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 228, + 290, + 240 + ], + "lines": [ + { + "bbox": [ + 106, + 228, + 291, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 291, + 241 + ], + "score": 1.0, + "content": "3.1 READING COMPREHENSION SETTING", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 249, + 505, + 316 + ], + "lines": [ + { + "bbox": [ + 105, + 248, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 332, + 263 + ], + "score": 1.0, + "content": "The reading comprehension task involves a document", + "type": "text" + }, + { + "bbox": [ + 332, + 250, + 426, + 262 + ], + "score": 0.91, + "content": "P = ( p _ { 1 } , p _ { 2 } , \\cdot \\cdot \\cdot , p _ { M } )", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 248, + 482, + 263 + ], + "score": 1.0, + "content": "and a query", + "type": "text" + }, + { + "bbox": [ + 482, + 250, + 505, + 262 + ], + "score": 0.87, + "content": "Q =", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 260, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 107, + 261, + 172, + 273 + ], + "score": 0.91, + "content": "\\left( q _ { 1 } , q _ { 2 } , \\cdots , q _ { N } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 260, + 202, + 274 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 203, + 261, + 214, + 271 + ], + "score": 0.81, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 260, + 233, + 274 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 233, + 261, + 243, + 271 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 260, + 506, + 274 + ], + "score": 1.0, + "content": "are the lengths of the document and the query respectively. Each", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 131, + 284 + ], + "score": 1.0, + "content": "token", + "type": "text" + }, + { + "bbox": [ + 131, + 274, + 141, + 284 + ], + "score": 0.84, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 272, + 197, + 284 + ], + "score": 1.0, + "content": "is denoted as", + "type": "text" + }, + { + "bbox": [ + 197, + 272, + 232, + 284 + ], + "score": 0.93, + "content": "( \\mathbf { w } _ { i } , \\mathbf { C } _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 272, + 263, + 284 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 264, + 273, + 276, + 283 + ], + "score": 0.88, + "content": "\\mathbf { w } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "is a one-hot encoding of the token in the vocabulary and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 119, + 294 + ], + "score": 0.88, + "content": "\\mathbf { C } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "is a matrix with each row representing a one-hot encoding of a character. Each token in the query", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 293, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 116, + 306 + ], + "score": 0.86, + "content": "q _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 293, + 233, + 307 + ], + "score": 1.0, + "content": "is similarly defined. We use", + "type": "text" + }, + { + "bbox": [ + 234, + 295, + 239, + 304 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 293, + 375, + 307 + ], + "score": 1.0, + "content": "as a subscript for documents and", + "type": "text" + }, + { + "bbox": [ + 375, + 294, + 381, + 306 + ], + "score": 0.83, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 293, + 506, + 307 + ], + "score": 1.0, + "content": "for queries. The output of the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 305, + 470, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 194, + 317 + ], + "score": 1.0, + "content": "problem is an answer", + "type": "text" + }, + { + "bbox": [ + 194, + 307, + 200, + 315 + ], + "score": 0.71, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 305, + 470, + 317 + ], + "score": 1.0, + "content": ", which can either be an index or a span of indices in the document.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 248, + 506, + 317 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 321, + 505, + 399 + ], + "lines": [ + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "score": 1.0, + "content": "Now we describe a general architecture used in this work, which is a generalization of the gated", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 332, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 505, + 345 + ], + "score": 1.0, + "content": "attention reader (Dhingra et al., 2016a). For each token in the document and the query, we compute", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 344, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 104, + 344, + 271, + 356 + ], + "score": 1.0, + "content": "a vector representation using a function", + "type": "text" + }, + { + "bbox": [ + 271, + 344, + 278, + 355 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 344, + 422, + 356 + ], + "score": 1.0, + "content": ". More specifically, for each token", + "type": "text" + }, + { + "bbox": [ + 422, + 345, + 432, + 355 + ], + "score": 0.84, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 344, + 505, + 356 + ], + "score": 1.0, + "content": "in the document,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 144, + 367 + ], + "score": 1.0, + "content": "we have", + "type": "text" + }, + { + "bbox": [ + 144, + 354, + 214, + 367 + ], + "score": 0.93, + "content": "\\mathbf { h } _ { i } ^ { \\bar { 0 } } = f ( \\mathbf { w } _ { i } , \\mathbf { C } _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 354, + 302, + 367 + ], + "score": 1.0, + "content": ". The same function", + "type": "text" + }, + { + "bbox": [ + 302, + 355, + 309, + 366 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "is also applied to the tokens in the query. Let", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 365, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 122, + 379 + ], + "score": 0.9, + "content": "\\mathbf { H } _ { p } ^ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 365, + 140, + 378 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 140, + 366, + 155, + 378 + ], + "score": 0.87, + "content": "\\mathbf { H } _ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 365, + 347, + 378 + ], + "score": 1.0, + "content": "denote the vector representations computed by", + "type": "text" + }, + { + "bbox": [ + 347, + 366, + 354, + 377 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "for tokens in documents and queries", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "respectively. In Section 3.2, we will discuss the “word-character” fine-grained gating used to define", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 387, + 169, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 157, + 401 + ], + "score": 1.0, + "content": "the function", + "type": "text" + }, + { + "bbox": [ + 157, + 388, + 164, + 399 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 387, + 169, + 401 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19, + "bbox_fs": [ + 104, + 322, + 506, + 401 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 404, + 505, + 462 + ], + "lines": [ + { + "bbox": [ + 105, + 402, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 247, + 418 + ], + "score": 1.0, + "content": "Suppose that we have a network of", + "type": "text" + }, + { + "bbox": [ + 247, + 405, + 258, + 414 + ], + "score": 0.82, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 402, + 315, + 418 + ], + "score": 1.0, + "content": "layers. At the", + "type": "text" + }, + { + "bbox": [ + 315, + 405, + 321, + 414 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 402, + 435, + 418 + ], + "score": 1.0, + "content": "-th layer, we apply RNNs on", + "type": "text" + }, + { + "bbox": [ + 436, + 403, + 461, + 418 + ], + "score": 0.93, + "content": "\\mathbf { H } _ { p } ^ { k - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 402, + 479, + 418 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 479, + 404, + 494, + 417 + ], + "score": 0.88, + "content": "\\mathbf { H } _ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 402, + 506, + 418 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 414, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 104, + 414, + 187, + 432 + ], + "score": 1.0, + "content": "obtain hidden states", + "type": "text" + }, + { + "bbox": [ + 187, + 416, + 201, + 428 + ], + "score": 0.91, + "content": "\\mathbf { P } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 414, + 219, + 432 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 219, + 416, + 233, + 429 + ], + "score": 0.9, + "content": "\\mathbf { Q } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 414, + 263, + 432 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 263, + 416, + 277, + 428 + ], + "score": 0.89, + "content": "\\mathbf { P } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 414, + 294, + 432 + ], + "score": 1.0, + "content": "is a", + "type": "text" + }, + { + "bbox": [ + 294, + 417, + 321, + 428 + ], + "score": 0.91, + "content": "M \\times d", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 414, + 367, + 432 + ], + "score": 1.0, + "content": "matrix and", + "type": "text" + }, + { + "bbox": [ + 367, + 417, + 382, + 429 + ], + "score": 0.91, + "content": "\\mathbf { Q } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 414, + 398, + 432 + ], + "score": 1.0, + "content": "is a", + "type": "text" + }, + { + "bbox": [ + 398, + 417, + 424, + 428 + ], + "score": 0.9, + "content": "N \\times d", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 414, + 473, + 432 + ], + "score": 1.0, + "content": "matrix with", + "type": "text" + }, + { + "bbox": [ + 473, + 418, + 479, + 428 + ], + "score": 0.8, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 414, + 506, + 432 + ], + "score": 1.0, + "content": "being", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 366, + 441 + ], + "score": 1.0, + "content": "the number of hidden units in the RNNs. Then we use a function", + "type": "text" + }, + { + "bbox": [ + 366, + 431, + 372, + 438 + ], + "score": 0.73, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "to compute a new representation", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 437, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 178, + 453 + ], + "score": 1.0, + "content": "for the document", + "type": "text" + }, + { + "bbox": [ + 178, + 439, + 251, + 453 + ], + "score": 0.93, + "content": "\\mathbf { H } _ { p } ^ { k } = r ( \\mathbf { P } ^ { k } , \\mathbf { Q } ^ { k } )", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 437, + 506, + 453 + ], + "score": 1.0, + "content": ". In Section 3.3, we will introduce the “document-query” fine-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 450, + 285, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 274, + 464 + ], + "score": 1.0, + "content": "grained gating used to define the function", + "type": "text" + }, + { + "bbox": [ + 275, + 453, + 280, + 460 + ], + "score": 0.76, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 450, + 285, + 464 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25, + "bbox_fs": [ + 104, + 402, + 506, + 464 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 467, + 504, + 501 + ], + "lines": [ + { + "bbox": [ + 106, + 467, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 467, + 191, + 479 + ], + "score": 1.0, + "content": "After going through", + "type": "text" + }, + { + "bbox": [ + 192, + 468, + 202, + 477 + ], + "score": 0.81, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 467, + 353, + 479 + ], + "score": 1.0, + "content": "layers, we predict the answer index", + "type": "text" + }, + { + "bbox": [ + 354, + 470, + 360, + 477 + ], + "score": 0.63, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 467, + 505, + 479 + ], + "score": 1.0, + "content": "using a softmax layer over hidden", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 477, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 131, + 491 + ], + "score": 1.0, + "content": "states", + "type": "text" + }, + { + "bbox": [ + 131, + 478, + 146, + 492 + ], + "score": 0.89, + "content": "\\bar { \\mathbf { H } } _ { p } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 477, + 506, + 491 + ], + "score": 1.0, + "content": ". For datasets where the answer is a span of text, we use two softmax layers for the start", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 488, + 223, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 223, + 502 + ], + "score": 1.0, + "content": "and end indices respectively.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 106, + 467, + 506, + 502 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 515, + 322, + 527 + ], + "lines": [ + { + "bbox": [ + 105, + 514, + 323, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 323, + 528 + ], + "score": 1.0, + "content": "3.2 WORD-CHARACTER FINE-GRAINED GATING", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 109, + 536, + 503, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 214, + 549 + ], + "score": 1.0, + "content": "Given a one-hot encoding", + "type": "text" + }, + { + "bbox": [ + 214, + 538, + 227, + 547 + ], + "score": 0.86, + "content": "\\mathbf { w } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 536, + 333, + 549 + ], + "score": 1.0, + "content": "and a character sequence", + "type": "text" + }, + { + "bbox": [ + 333, + 537, + 346, + 548 + ], + "score": 0.88, + "content": "\\mathbf { C } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 536, + 505, + 549 + ], + "score": 1.0, + "content": ", we now describe how to compute the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 548, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 194, + 560 + ], + "score": 1.0, + "content": "vector representation", + "type": "text" + }, + { + "bbox": [ + 195, + 548, + 263, + 560 + ], + "score": 0.93, + "content": "\\mathbf { h } _ { i } = f ( \\mathbf { w } _ { i } , \\mathbf { C } _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 548, + 505, + 560 + ], + "score": 1.0, + "content": "for the token. In the rest of the section, we will drop the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 558, + 244, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 145, + 571 + ], + "score": 1.0, + "content": "subscript", + "type": "text" + }, + { + "bbox": [ + 145, + 559, + 150, + 568 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 558, + 244, + 571 + ], + "score": 1.0, + "content": "for notation simplicity.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 106, + 536, + 505, + 571 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 575, + 505, + 619 + ], + "lines": [ + { + "bbox": [ + 106, + 575, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 214, + 587 + ], + "score": 1.0, + "content": "We first apply an RNN on", + "type": "text" + }, + { + "bbox": [ + 214, + 576, + 225, + 586 + ], + "score": 0.51, + "content": "\\mathbf { C }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 575, + 505, + 587 + ], + "score": 1.0, + "content": "and take the hidden state in the last time step c as the character-level", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 266, + 599 + ], + "score": 1.0, + "content": "representation (Yang et al., 2016a). Let", + "type": "text" + }, + { + "bbox": [ + 266, + 587, + 275, + 597 + ], + "score": 0.47, + "content": "\\mathbf { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 586, + 506, + 599 + ], + "score": 1.0, + "content": "denote the token embedding lookup table. We perform a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "matrix-vector multiplication Ew to obtain a word-level representation. We assume c and Ew have", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 609, + 236, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 171, + 620 + ], + "score": 1.0, + "content": "the same length", + "type": "text" + }, + { + "bbox": [ + 172, + 609, + 182, + 619 + ], + "score": 0.88, + "content": "d _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 609, + 236, + 620 + ], + "score": 1.0, + "content": "in this work.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 575, + 506, + 620 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 625, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 625, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 214, + 637 + ], + "score": 1.0, + "content": "Previous methods defined", + "type": "text" + }, + { + "bbox": [ + 214, + 626, + 222, + 637 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 625, + 505, + 637 + ], + "score": 1.0, + "content": "using the word-level representation Ew (Collobert et al., 2011), the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 636, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 390, + 649 + ], + "score": 1.0, + "content": "character-level representation c (Ling et al., 2015), or the concatenation", + "type": "text" + }, + { + "bbox": [ + 390, + 636, + 423, + 648 + ], + "score": 0.25, + "content": "\\left[ \\mathbf { E w } ; \\mathbf { c } \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 636, + 505, + 649 + ], + "score": 1.0, + "content": "(Yang et al., 2016a).", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 647, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 506, + 660 + ], + "score": 1.0, + "content": "Unlike these methods, we propose to use a gate to dynamically choose between the word-level and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 658, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 400, + 671 + ], + "score": 1.0, + "content": "character-level representations based on the properties of the token. Let", + "type": "text" + }, + { + "bbox": [ + 401, + 660, + 408, + 668 + ], + "score": 0.36, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 658, + 506, + 671 + ], + "score": 1.0, + "content": "denote a feature vector", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 669, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 505, + 682 + ], + "score": 1.0, + "content": "that encodes these properties. In this work, we use the concatenation of named entity tags, part-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 680, + 504, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 504, + 692 + ], + "score": 1.0, + "content": "of-speech tags, binned document frequency vectors, and the word-level representations to form the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 691, + 297, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 164, + 703 + ], + "score": 1.0, + "content": "feature vector", + "type": "text" + }, + { + "bbox": [ + 164, + 693, + 171, + 701 + ], + "score": 0.26, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 691, + 191, + 703 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 191, + 691, + 203, + 702 + ], + "score": 0.89, + "content": "d _ { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 691, + 286, + 703 + ], + "score": 1.0, + "content": "denote the length of", + "type": "text" + }, + { + "bbox": [ + 286, + 693, + 293, + 701 + ], + "score": 0.61, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 691, + 297, + 703 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 625, + 506, + 703 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 708, + 239, + 719 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 240, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 240, + 721 + ], + "score": 1.0, + "content": "The gate is computed as follows:", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46, + "bbox_fs": [ + 106, + 708, + 240, + 721 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 265, + 719, + 346, + 734 + ], + "lines": [ + { + "bbox": [ + 265, + 719, + 346, + 734 + ], + "spans": [ + { + "bbox": [ + 265, + 719, + 346, + 734 + ], + "score": 0.91, + "content": "\\mathbf { g } = \\sigma ( \\mathbf { W } _ { g } \\mathbf { v } + \\mathbf { b } _ { g } )", + "type": "interline_equation", + "image_path": "1879e1834ff91ce7e7cdea9d58cd1237d894538cc8c2a492a6efa280345bea48.jpg" + } + ] + } + ], + "index": 47, + "virtual_lines": [ + { + "bbox": [ + 265, + 719, + 346, + 734 + ], + "spans": [], + "index": 47 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 165, + 83, + 439, + 246 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 165, + 83, + 439, + 246 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 165, + 83, + 439, + 246 + ], + "spans": [ + { + "bbox": [ + 165, + 83, + 439, + 246 + ], + "score": 0.972, + "type": "image", + "image_path": "f10e95e7400bc55e85299f1e966111eeaf000699ba21922a0e9aef0ca84cfda7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 165, + 83, + 439, + 137.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 165, + 137.33333333333334, + 439, + 191.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 165, + 191.66666666666669, + 439, + 246.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 263, + 503, + 285 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 263, + 504, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 504, + 276 + ], + "score": 1.0, + "content": "Figure 1: Word-character fine-grained gating. The two lookup tables are shared. “NER”, “POS”, “frequency”", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 274, + 384, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 384, + 285 + ], + "score": 1.0, + "content": "refer to named entity tags, part-of-speech tags, document frequency features.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 306, + 502, + 329 + ], + "lines": [ + { + "bbox": [ + 106, + 306, + 504, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 133, + 319 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 307, + 151, + 319 + ], + "score": 0.9, + "content": "\\mathbf { W } _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 306, + 168, + 319 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 169, + 307, + 181, + 319 + ], + "score": 0.89, + "content": "{ \\bf b } _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 306, + 334, + 319 + ], + "score": 1.0, + "content": "are the model parameters with shapes", + "type": "text" + }, + { + "bbox": [ + 334, + 307, + 365, + 318 + ], + "score": 0.93, + "content": "d _ { e } \\times d _ { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 306, + 383, + 319 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 384, + 307, + 394, + 318 + ], + "score": 0.89, + "content": "d _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 306, + 414, + 319 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 415, + 309, + 422, + 317 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 306, + 504, + 319 + ], + "score": 1.0, + "content": "denotes an element-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 318, + 199, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 199, + 330 + ], + "score": 1.0, + "content": "wise sigmoid function.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 334, + 414, + 347 + ], + "lines": [ + { + "bbox": [ + 105, + 333, + 414, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 414, + 349 + ], + "score": 1.0, + "content": "The final representation is computed using a fine-grained gating mechanism,", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 353, + 390, + 367 + ], + "lines": [ + { + "bbox": [ + 220, + 353, + 390, + 367 + ], + "spans": [ + { + "bbox": [ + 220, + 353, + 390, + 367 + ], + "score": 0.9, + "content": "\\mathbf { h } = f ( \\mathbf { c } , \\mathbf { w } ) = \\mathbf { g } \\odot \\mathbf { c } + ( 1 - \\mathbf { g } ) \\odot ( \\mathbf { E } \\mathbf { w } )", + "type": "interline_equation", + "image_path": "44026894aa630dfb83831ba155043105ba340a4961308a1db5f0aff6f127116b.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 220, + 353, + 390, + 367 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 374, + 351, + 386 + ], + "lines": [ + { + "bbox": [ + 106, + 374, + 352, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 374, + 133, + 387 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 376, + 142, + 385 + ], + "score": 0.81, + "content": "\\odot", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 374, + 352, + 387 + ], + "score": 1.0, + "content": "denotes element-wise product between two vectors.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 391, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 390, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 405 + ], + "score": 1.0, + "content": "An illustration of our fine-grained gating mechanism is shown in Figure 1. Intuitively speaking,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 164, + 414 + ], + "score": 1.0, + "content": "when the gate", + "type": "text" + }, + { + "bbox": [ + 165, + 404, + 173, + 414 + ], + "score": 0.36, + "content": "\\mathbf { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 402, + 505, + 414 + ], + "score": 1.0, + "content": "has high values, more information flows from the character-level representation to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 413, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 260, + 426 + ], + "score": 1.0, + "content": "the final representation; when the gate", + "type": "text" + }, + { + "bbox": [ + 260, + 415, + 268, + 425 + ], + "score": 0.39, + "content": "\\mathbf { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 413, + 505, + 426 + ], + "score": 1.0, + "content": "has low values, the final representation is dominated by the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 424, + 213, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 213, + 436 + ], + "score": 1.0, + "content": "word-level representation.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "score": 1.0, + "content": "Though Miyamoto & Cho (2016) also use a gate to choose between word-level and character-level", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "representations, our method is different in two ways. First, we use a more fine-grained gating mech-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "anism, i.e., vector gates rather than scalar gates. Second, we condition the gate on features that better", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "reflect the properties of the token. For example, for noun phrases and entities, we would expect the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 484, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 104, + 484, + 505, + 498 + ], + "score": 1.0, + "content": "gate to bias towards character-level representations because noun phrases and entities are usually", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 495, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 506, + 509 + ], + "score": 1.0, + "content": "less common and display richer morphological structure. Experiments show that these changes are", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 507, + 394, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 394, + 520 + ], + "score": 1.0, + "content": "key to the performance improvements for reading comprehension tasks.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 108, + 523, + 505, + 557 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 537 + ], + "score": 1.0, + "content": "Our approach can be further generalized to a setting of multi-level networks so that we can combine", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 535, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 505, + 547 + ], + "score": 1.0, + "content": "multiple levels of representations using fine-grained gating mechanisms, which we leave for future", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 546, + 133, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 133, + 558 + ], + "score": 1.0, + "content": "work.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 107, + 572, + 321, + 584 + ], + "lines": [ + { + "bbox": [ + 105, + 572, + 322, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 322, + 586 + ], + "score": 1.0, + "content": "3.3 DOCUMENT-QUERY FINE-GRAINED GATING", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 593, + 505, + 639 + ], + "lines": [ + { + "bbox": [ + 105, + 592, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 205, + 607 + ], + "score": 1.0, + "content": "Given the hidden states", + "type": "text" + }, + { + "bbox": [ + 206, + 593, + 220, + 604 + ], + "score": 0.89, + "content": "\\mathbf { P } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 592, + 240, + 607 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 240, + 593, + 255, + 605 + ], + "score": 0.9, + "content": "\\mathbf { Q } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 592, + 470, + 607 + ], + "score": 1.0, + "content": ", we now describe how to compute a representation", + "type": "text" + }, + { + "bbox": [ + 470, + 593, + 485, + 604 + ], + "score": 0.92, + "content": "\\mathbf { H } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 592, + 506, + 607 + ], + "score": 1.0, + "content": "that", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "encodes the interactions between the document and the query. In this section, we drop the superscript", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 107, + 616, + 113, + 626 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 615, + 298, + 628 + ], + "score": 1.0, + "content": "(the layer number) for notation simplicity. Let", + "type": "text" + }, + { + "bbox": [ + 298, + 617, + 309, + 628 + ], + "score": 0.86, + "content": "\\mathbf { p } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 615, + 353, + 628 + ], + "score": 1.0, + "content": "denote the", + "type": "text" + }, + { + "bbox": [ + 353, + 617, + 358, + 626 + ], + "score": 0.69, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 615, + 397, + 628 + ], + "score": 1.0, + "content": "-th row of", + "type": "text" + }, + { + "bbox": [ + 398, + 616, + 407, + 626 + ], + "score": 0.67, + "content": "\\mathbf { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 615, + 423, + 628 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 424, + 617, + 435, + 628 + ], + "score": 0.88, + "content": "{ \\bf q } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 615, + 479, + 628 + ], + "score": 1.0, + "content": "denote the", + "type": "text" + }, + { + "bbox": [ + 479, + 617, + 485, + 628 + ], + "score": 0.83, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "-row", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 626, + 290, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 117, + 641 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 627, + 127, + 639 + ], + "score": 0.37, + "content": "\\mathbf { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 626, + 147, + 641 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 147, + 627, + 159, + 638 + ], + "score": 0.88, + "content": "d _ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 626, + 245, + 641 + ], + "score": 1.0, + "content": "denote the lengths of", + "type": "text" + }, + { + "bbox": [ + 246, + 628, + 257, + 639 + ], + "score": 0.87, + "content": "\\mathbf { p } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 626, + 275, + 641 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 275, + 629, + 286, + 639 + ], + "score": 0.88, + "content": "{ \\bf q } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 626, + 290, + 641 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "Attention-over-attention (AoA) (Cui et al., 2016) defines a dot product between each pair of tokens", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 654, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 104, + 654, + 255, + 669 + ], + "score": 1.0, + "content": "in the document and the query, i.e.,", + "type": "text" + }, + { + "bbox": [ + 255, + 654, + 279, + 667 + ], + "score": 0.92, + "content": "\\mathbf { p } _ { i } ^ { T } \\mathbf { q } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 654, + 506, + 669 + ], + "score": 1.0, + "content": ", followed by row-wise and column-wise softmax non-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "linearities. AoA imposes pair-wise interactions between the document and the query, but using a", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "dot product is potentially not expressive enough and hard to generalize to multi-layer networks. The", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 448, + 701 + ], + "score": 1.0, + "content": "gated attention (GA) reader (Dhingra et al., 2016a) defines an element-wise product as", + "type": "text" + }, + { + "bbox": [ + 448, + 689, + 477, + 699 + ], + "score": 0.9, + "content": "\\mathbf { p } _ { i } \\odot \\mathbf { g } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 117, + 710 + ], + "score": 0.83, + "content": "\\mathbf { g } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 699, + 350, + 712 + ], + "score": 1.0, + "content": "is a gate computed by attention mechanism on the token", + "type": "text" + }, + { + "bbox": [ + 351, + 700, + 360, + 710 + ], + "score": 0.84, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "and the entire query. The intuition", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 710, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 155, + 721 + ], + "score": 1.0, + "content": "for the gate", + "type": "text" + }, + { + "bbox": [ + 156, + 711, + 167, + 721 + ], + "score": 0.85, + "content": "\\mathbf { g } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 710, + 506, + 721 + ], + "score": 1.0, + "content": "is to attend to important information in the document. However, there is no direct", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 721, + 292, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 292, + 733 + ], + "score": 1.0, + "content": "pair-wise interaction between each token pair.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 165, + 83, + 439, + 246 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 165, + 83, + 439, + 246 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 165, + 83, + 439, + 246 + ], + "spans": [ + { + "bbox": [ + 165, + 83, + 439, + 246 + ], + "score": 0.972, + "type": "image", + "image_path": "f10e95e7400bc55e85299f1e966111eeaf000699ba21922a0e9aef0ca84cfda7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 165, + 83, + 439, + 137.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 165, + 137.33333333333334, + 439, + 191.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 165, + 191.66666666666669, + 439, + 246.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 263, + 503, + 285 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 263, + 504, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 504, + 276 + ], + "score": 1.0, + "content": "Figure 1: Word-character fine-grained gating. The two lookup tables are shared. “NER”, “POS”, “frequency”", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 274, + 384, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 384, + 285 + ], + "score": 1.0, + "content": "refer to named entity tags, part-of-speech tags, document frequency features.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 306, + 502, + 329 + ], + "lines": [ + { + "bbox": [ + 106, + 306, + 504, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 133, + 319 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 307, + 151, + 319 + ], + "score": 0.9, + "content": "\\mathbf { W } _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 306, + 168, + 319 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 169, + 307, + 181, + 319 + ], + "score": 0.89, + "content": "{ \\bf b } _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 306, + 334, + 319 + ], + "score": 1.0, + "content": "are the model parameters with shapes", + "type": "text" + }, + { + "bbox": [ + 334, + 307, + 365, + 318 + ], + "score": 0.93, + "content": "d _ { e } \\times d _ { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 306, + 383, + 319 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 384, + 307, + 394, + 318 + ], + "score": 0.89, + "content": "d _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 306, + 414, + 319 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 415, + 309, + 422, + 317 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 306, + 504, + 319 + ], + "score": 1.0, + "content": "denotes an element-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 318, + 199, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 199, + 330 + ], + "score": 1.0, + "content": "wise sigmoid function.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 106, + 306, + 504, + 330 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 334, + 414, + 347 + ], + "lines": [ + { + "bbox": [ + 105, + 333, + 414, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 414, + 349 + ], + "score": 1.0, + "content": "The final representation is computed using a fine-grained gating mechanism,", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 333, + 414, + 349 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 353, + 390, + 367 + ], + "lines": [ + { + "bbox": [ + 220, + 353, + 390, + 367 + ], + "spans": [ + { + "bbox": [ + 220, + 353, + 390, + 367 + ], + "score": 0.9, + "content": "\\mathbf { h } = f ( \\mathbf { c } , \\mathbf { w } ) = \\mathbf { g } \\odot \\mathbf { c } + ( 1 - \\mathbf { g } ) \\odot ( \\mathbf { E } \\mathbf { w } )", + "type": "interline_equation", + "image_path": "44026894aa630dfb83831ba155043105ba340a4961308a1db5f0aff6f127116b.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 220, + 353, + 390, + 367 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 374, + 351, + 386 + ], + "lines": [ + { + "bbox": [ + 106, + 374, + 352, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 374, + 133, + 387 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 376, + 142, + 385 + ], + "score": 0.81, + "content": "\\odot", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 374, + 352, + 387 + ], + "score": 1.0, + "content": "denotes element-wise product between two vectors.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 374, + 352, + 387 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 391, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 390, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 405 + ], + "score": 1.0, + "content": "An illustration of our fine-grained gating mechanism is shown in Figure 1. Intuitively speaking,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 164, + 414 + ], + "score": 1.0, + "content": "when the gate", + "type": "text" + }, + { + "bbox": [ + 165, + 404, + 173, + 414 + ], + "score": 0.36, + "content": "\\mathbf { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 402, + 505, + 414 + ], + "score": 1.0, + "content": "has high values, more information flows from the character-level representation to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 413, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 260, + 426 + ], + "score": 1.0, + "content": "the final representation; when the gate", + "type": "text" + }, + { + "bbox": [ + 260, + 415, + 268, + 425 + ], + "score": 0.39, + "content": "\\mathbf { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 413, + 505, + 426 + ], + "score": 1.0, + "content": "has low values, the final representation is dominated by the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 424, + 213, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 213, + 436 + ], + "score": 1.0, + "content": "word-level representation.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 390, + 505, + 436 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "score": 1.0, + "content": "Though Miyamoto & Cho (2016) also use a gate to choose between word-level and character-level", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "representations, our method is different in two ways. First, we use a more fine-grained gating mech-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "anism, i.e., vector gates rather than scalar gates. Second, we condition the gate on features that better", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "reflect the properties of the token. For example, for noun phrases and entities, we would expect the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 484, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 104, + 484, + 505, + 498 + ], + "score": 1.0, + "content": "gate to bias towards character-level representations because noun phrases and entities are usually", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 495, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 506, + 509 + ], + "score": 1.0, + "content": "less common and display richer morphological structure. Experiments show that these changes are", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 507, + 394, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 394, + 520 + ], + "score": 1.0, + "content": "key to the performance improvements for reading comprehension tasks.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17, + "bbox_fs": [ + 104, + 440, + 506, + 520 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 523, + 505, + 557 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 537 + ], + "score": 1.0, + "content": "Our approach can be further generalized to a setting of multi-level networks so that we can combine", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 535, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 505, + 547 + ], + "score": 1.0, + "content": "multiple levels of representations using fine-grained gating mechanisms, which we leave for future", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 546, + 133, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 133, + 558 + ], + "score": 1.0, + "content": "work.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 523, + 505, + 558 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 572, + 321, + 584 + ], + "lines": [ + { + "bbox": [ + 105, + 572, + 322, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 322, + 586 + ], + "score": 1.0, + "content": "3.3 DOCUMENT-QUERY FINE-GRAINED GATING", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 593, + 505, + 639 + ], + "lines": [ + { + "bbox": [ + 105, + 592, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 205, + 607 + ], + "score": 1.0, + "content": "Given the hidden states", + "type": "text" + }, + { + "bbox": [ + 206, + 593, + 220, + 604 + ], + "score": 0.89, + "content": "\\mathbf { P } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 592, + 240, + 607 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 240, + 593, + 255, + 605 + ], + "score": 0.9, + "content": "\\mathbf { Q } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 592, + 470, + 607 + ], + "score": 1.0, + "content": ", we now describe how to compute a representation", + "type": "text" + }, + { + "bbox": [ + 470, + 593, + 485, + 604 + ], + "score": 0.92, + "content": "\\mathbf { H } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 592, + 506, + 607 + ], + "score": 1.0, + "content": "that", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "encodes the interactions between the document and the query. In this section, we drop the superscript", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 107, + 616, + 113, + 626 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 615, + 298, + 628 + ], + "score": 1.0, + "content": "(the layer number) for notation simplicity. Let", + "type": "text" + }, + { + "bbox": [ + 298, + 617, + 309, + 628 + ], + "score": 0.86, + "content": "\\mathbf { p } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 615, + 353, + 628 + ], + "score": 1.0, + "content": "denote the", + "type": "text" + }, + { + "bbox": [ + 353, + 617, + 358, + 626 + ], + "score": 0.69, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 615, + 397, + 628 + ], + "score": 1.0, + "content": "-th row of", + "type": "text" + }, + { + "bbox": [ + 398, + 616, + 407, + 626 + ], + "score": 0.67, + "content": "\\mathbf { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 615, + 423, + 628 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 424, + 617, + 435, + 628 + ], + "score": 0.88, + "content": "{ \\bf q } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 615, + 479, + 628 + ], + "score": 1.0, + "content": "denote the", + "type": "text" + }, + { + "bbox": [ + 479, + 617, + 485, + 628 + ], + "score": 0.83, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "-row", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 626, + 290, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 117, + 641 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 627, + 127, + 639 + ], + "score": 0.37, + "content": "\\mathbf { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 626, + 147, + 641 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 147, + 627, + 159, + 638 + ], + "score": 0.88, + "content": "d _ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 626, + 245, + 641 + ], + "score": 1.0, + "content": "denote the lengths of", + "type": "text" + }, + { + "bbox": [ + 246, + 628, + 257, + 639 + ], + "score": 0.87, + "content": "\\mathbf { p } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 626, + 275, + 641 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 275, + 629, + 286, + 639 + ], + "score": 0.88, + "content": "{ \\bf q } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 626, + 290, + 641 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 592, + 506, + 641 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "Attention-over-attention (AoA) (Cui et al., 2016) defines a dot product between each pair of tokens", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 654, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 104, + 654, + 255, + 669 + ], + "score": 1.0, + "content": "in the document and the query, i.e.,", + "type": "text" + }, + { + "bbox": [ + 255, + 654, + 279, + 667 + ], + "score": 0.92, + "content": "\\mathbf { p } _ { i } ^ { T } \\mathbf { q } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 654, + 506, + 669 + ], + "score": 1.0, + "content": ", followed by row-wise and column-wise softmax non-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "linearities. AoA imposes pair-wise interactions between the document and the query, but using a", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "dot product is potentially not expressive enough and hard to generalize to multi-layer networks. The", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 448, + 701 + ], + "score": 1.0, + "content": "gated attention (GA) reader (Dhingra et al., 2016a) defines an element-wise product as", + "type": "text" + }, + { + "bbox": [ + 448, + 689, + 477, + 699 + ], + "score": 0.9, + "content": "\\mathbf { p } _ { i } \\odot \\mathbf { g } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 117, + 710 + ], + "score": 0.83, + "content": "\\mathbf { g } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 699, + 350, + 712 + ], + "score": 1.0, + "content": "is a gate computed by attention mechanism on the token", + "type": "text" + }, + { + "bbox": [ + 351, + 700, + 360, + 710 + ], + "score": 0.84, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "and the entire query. The intuition", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 710, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 155, + 721 + ], + "score": 1.0, + "content": "for the gate", + "type": "text" + }, + { + "bbox": [ + 156, + 711, + 167, + 721 + ], + "score": 0.85, + "content": "\\mathbf { g } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 710, + 506, + 721 + ], + "score": 1.0, + "content": "is to attend to important information in the document. However, there is no direct", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 721, + 292, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 292, + 733 + ], + "score": 1.0, + "content": "pair-wise interaction between each token pair.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32.5, + "bbox_fs": [ + 104, + 644, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 82, + 490, + 250 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 82, + 490, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 82, + 490, + 250 + ], + "spans": [ + { + "bbox": [ + 108, + 82, + 490, + 250 + ], + "score": 0.97, + "type": "image", + "image_path": "284468038502d2ff8d8577366fc87421d42142474f6344ca41f53869482de701.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 82, + 490, + 138.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 138.0, + 490, + 194.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 194.0, + 490, + 250.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 214, + 267, + 396, + 279 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 212, + 265, + 398, + 282 + ], + "spans": [ + { + "bbox": [ + 212, + 265, + 398, + 282 + ], + "score": 1.0, + "content": "Figure 2: Paragraph-question fine-grained gating.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 106, + 299, + 505, + 342 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "We present a fine-grained gating method that combines the advantages of the above methods (i.e.,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 308, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 505, + 323 + ], + "score": 1.0, + "content": "both pairwise and element-wise). We compute the pairwise element-wise product between the hid-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 320, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 449, + 335 + ], + "score": 1.0, + "content": "den states in the document and the query, as shown in Figure 2. More specifically, for", + "type": "text" + }, + { + "bbox": [ + 450, + 323, + 459, + 333 + ], + "score": 0.84, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 320, + 477, + 335 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 477, + 323, + 487, + 334 + ], + "score": 0.85, + "content": "q _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 320, + 505, + 335 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 331, + 128, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 128, + 344 + ], + "score": 1.0, + "content": "have", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + }, + { + "type": "interline_equation", + "bbox": [ + 262, + 342, + 349, + 356 + ], + "lines": [ + { + "bbox": [ + 262, + 342, + 349, + 356 + ], + "spans": [ + { + "bbox": [ + 262, + 342, + 349, + 356 + ], + "score": 0.91, + "content": "\\mathbf { I } _ { i j } = \\operatorname { t a n h } ( \\mathbf { p } _ { i } \\odot \\mathbf { q } _ { j } )", + "type": "interline_equation", + "image_path": "2a02adcad151b1ef48daf138830045bd50697f3b2d2dbb700ca45ecefda5a53a.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 262, + 342, + 349, + 356 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 358, + 504, + 380 + ], + "lines": [ + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 133, + 370 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 360, + 144, + 370 + ], + "score": 0.86, + "content": "{ \\bf q } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 357, + 343, + 370 + ], + "score": 1.0, + "content": "can be viewed as a gate to filter the information in", + "type": "text" + }, + { + "bbox": [ + 343, + 360, + 354, + 369 + ], + "score": 0.87, + "content": "\\mathbf { p } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 357, + 505, + 370 + ], + "score": 1.0, + "content": ". We then use an attention mechanism", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 369, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 126, + 380 + ], + "score": 1.0, + "content": "over", + "type": "text" + }, + { + "bbox": [ + 126, + 370, + 139, + 381 + ], + "score": 0.89, + "content": "\\mathbf { I } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 369, + 232, + 380 + ], + "score": 1.0, + "content": "to output hidden states", + "type": "text" + }, + { + "bbox": [ + 232, + 370, + 244, + 380 + ], + "score": 0.89, + "content": "\\mathbf { h } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 369, + 288, + 380 + ], + "score": 1.0, + "content": "as follows", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 385, + 403, + 414 + ], + "lines": [ + { + "bbox": [ + 207, + 385, + 403, + 414 + ], + "spans": [ + { + "bbox": [ + 207, + 385, + 403, + 414 + ], + "score": 0.94, + "content": "\\mathbf { h } _ { i } = \\sum _ { j } \\mathrm { s o f t m a x } ( \\mathbf { u } _ { h } ^ { T } \\mathbf { I } _ { i j } + \\mathbf { w } _ { i } ^ { T } \\mathbf { w } _ { j } b _ { h 1 } + b _ { h 2 } ) \\mathbf { I } _ { i j }", + "type": "interline_equation", + "image_path": "684f18d9d438d0c4c302c0fdfd67203049f99927f2fa867240af1df8d84a1418.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 207, + 385, + 403, + 399.5 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 207, + 399.5, + 403, + 414.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 419, + 505, + 475 + ], + "lines": [ + { + "bbox": [ + 106, + 419, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 133, + 431 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 421, + 146, + 430 + ], + "score": 0.86, + "content": "\\mathbf { u } _ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 419, + 165, + 431 + ], + "score": 1.0, + "content": "is a", + "type": "text" + }, + { + "bbox": [ + 166, + 420, + 177, + 430 + ], + "score": 0.89, + "content": "d _ { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 419, + 303, + 431 + ], + "score": 1.0, + "content": "-dimensional model parameter,", + "type": "text" + }, + { + "bbox": [ + 304, + 420, + 319, + 430 + ], + "score": 0.89, + "content": "b _ { h 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 419, + 338, + 431 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 338, + 420, + 353, + 430 + ], + "score": 0.89, + "content": "b _ { h 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 419, + 473, + 431 + ], + "score": 1.0, + "content": "are scalar model parameters,", + "type": "text" + }, + { + "bbox": [ + 473, + 421, + 486, + 430 + ], + "score": 0.85, + "content": "\\mathbf { w } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 419, + 505, + 431 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 430, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 120, + 443 + ], + "score": 0.87, + "content": "{ \\bf w } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 430, + 227, + 443 + ], + "score": 1.0, + "content": "are one-hot encodings for", + "type": "text" + }, + { + "bbox": [ + 228, + 432, + 237, + 442 + ], + "score": 0.85, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 430, + 255, + 443 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 256, + 432, + 266, + 442 + ], + "score": 0.86, + "content": "q _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 430, + 505, + 443 + ], + "score": 1.0, + "content": "respectively. We additionally use one-hot encodings in the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "attention mechanism to reinforce the matching between the same tokens since such information is", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 194, + 465 + ], + "score": 1.0, + "content": "not fully preserved in", + "type": "text" + }, + { + "bbox": [ + 194, + 452, + 206, + 464 + ], + "score": 0.89, + "content": "\\mathbf { I } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 452, + 231, + 465 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 232, + 453, + 239, + 462 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 452, + 450, + 465 + ], + "score": 1.0, + "content": "is large. The softmax nonlinearity is applied over all", + "type": "text" + }, + { + "bbox": [ + 450, + 453, + 456, + 464 + ], + "score": 0.77, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 452, + 506, + 465 + ], + "score": 1.0, + "content": "’s. The final", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 461, + 395, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 160, + 477 + ], + "score": 1.0, + "content": "hidden states", + "type": "text" + }, + { + "bbox": [ + 161, + 464, + 171, + 473 + ], + "score": 0.55, + "content": "\\mathbf { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 461, + 303, + 477 + ], + "score": 1.0, + "content": "are formed by concatenating the", + "type": "text" + }, + { + "bbox": [ + 303, + 464, + 314, + 474 + ], + "score": 0.88, + "content": "\\mathbf { h } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 461, + 381, + 477 + ], + "score": 1.0, + "content": "’s for each token", + "type": "text" + }, + { + "bbox": [ + 381, + 465, + 390, + 474 + ], + "score": 0.84, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 461, + 395, + 477 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 108, + 491, + 200, + 503 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 201, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 201, + 505 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 515, + 505, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 516, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 506, + 528 + ], + "score": 1.0, + "content": "We first present experimental results on the Twitter dataset where we can rule out the effects of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 526, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 539 + ], + "score": 1.0, + "content": "different choices of network architectures, to demonstrate the effectiveness of our word-character", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 537, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 551 + ], + "score": 1.0, + "content": "fine-grained gating approach. Later we show experiments on more challenging datasets on reading", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 548, + 504, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 504, + 561 + ], + "score": 1.0, + "content": "comprehension to further show that our approach can be used to improve the performance on high-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 559, + 204, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 204, + 571 + ], + "score": 1.0, + "content": "level NLP tasks as well.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 107, + 583, + 369, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 583, + 370, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 370, + 597 + ], + "score": 1.0, + "content": "4.1 EVALUATING WORD-CHARACTER GATING ON TWITTER", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 604, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "We evaluate the effectiveness of our word-character fine-grained gating mechanism on a social media", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "tag prediction task. We use the Twitter dataset and follow the experimental settings in Dhingra et al.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "(2016b). We also use the same network architecture upon the token representations, which is an", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 638, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 650 + ], + "score": 1.0, + "content": "LSTM layer followed by a softmax classification layer (Dhingra et al., 2016b). The Twitter dataset", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "consists of English tweets with at least one hashtag from Twitter. Hashtags and HTML tags have", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "been removed from the body of the tweet, and user names and URLs are replaced with special", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 506, + 684 + ], + "score": 1.0, + "content": "tokens. The dataset contains 2 million tweets for training, 10K for validation and 50K for testing,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 682, + 457, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 457, + 694 + ], + "score": 1.0, + "content": "with a total of 2,039 distinct hashtags. The task is to predict the hashtags of each tweet.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 107, + 699, + 505, + 710 + ], + "spans": [ + { + "bbox": [ + 107, + 699, + 505, + 710 + ], + "score": 1.0, + "content": "We compare several different methods as follows. Word char concat uses the concatenation of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "word-level and character-level representations as in Yang et al. (2016a); word char feat concat", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "concatenates the word-level and character-level representations along with the features described in", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 82, + 490, + 250 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 82, + 490, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 82, + 490, + 250 + ], + "spans": [ + { + "bbox": [ + 108, + 82, + 490, + 250 + ], + "score": 0.97, + "type": "image", + "image_path": "284468038502d2ff8d8577366fc87421d42142474f6344ca41f53869482de701.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 82, + 490, + 138.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 138.0, + 490, + 194.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 194.0, + 490, + 250.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 214, + 267, + 396, + 279 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 212, + 265, + 398, + 282 + ], + "spans": [ + { + "bbox": [ + 212, + 265, + 398, + 282 + ], + "score": 1.0, + "content": "Figure 2: Paragraph-question fine-grained gating.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 106, + 299, + 505, + 342 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "We present a fine-grained gating method that combines the advantages of the above methods (i.e.,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 308, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 505, + 323 + ], + "score": 1.0, + "content": "both pairwise and element-wise). We compute the pairwise element-wise product between the hid-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 320, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 449, + 335 + ], + "score": 1.0, + "content": "den states in the document and the query, as shown in Figure 2. More specifically, for", + "type": "text" + }, + { + "bbox": [ + 450, + 323, + 459, + 333 + ], + "score": 0.84, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 320, + 477, + 335 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 477, + 323, + 487, + 334 + ], + "score": 0.85, + "content": "q _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 320, + 505, + 335 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 331, + 128, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 128, + 344 + ], + "score": 1.0, + "content": "have", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 299, + 506, + 344 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 262, + 342, + 349, + 356 + ], + "lines": [ + { + "bbox": [ + 262, + 342, + 349, + 356 + ], + "spans": [ + { + "bbox": [ + 262, + 342, + 349, + 356 + ], + "score": 0.91, + "content": "\\mathbf { I } _ { i j } = \\operatorname { t a n h } ( \\mathbf { p } _ { i } \\odot \\mathbf { q } _ { j } )", + "type": "interline_equation", + "image_path": "2a02adcad151b1ef48daf138830045bd50697f3b2d2dbb700ca45ecefda5a53a.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 262, + 342, + 349, + 356 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 358, + 504, + 380 + ], + "lines": [ + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 133, + 370 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 360, + 144, + 370 + ], + "score": 0.86, + "content": "{ \\bf q } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 357, + 343, + 370 + ], + "score": 1.0, + "content": "can be viewed as a gate to filter the information in", + "type": "text" + }, + { + "bbox": [ + 343, + 360, + 354, + 369 + ], + "score": 0.87, + "content": "\\mathbf { p } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 357, + 505, + 370 + ], + "score": 1.0, + "content": ". We then use an attention mechanism", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 369, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 126, + 380 + ], + "score": 1.0, + "content": "over", + "type": "text" + }, + { + "bbox": [ + 126, + 370, + 139, + 381 + ], + "score": 0.89, + "content": "\\mathbf { I } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 369, + 232, + 380 + ], + "score": 1.0, + "content": "to output hidden states", + "type": "text" + }, + { + "bbox": [ + 232, + 370, + 244, + 380 + ], + "score": 0.89, + "content": "\\mathbf { h } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 369, + 288, + 380 + ], + "score": 1.0, + "content": "as follows", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 357, + 505, + 381 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 385, + 403, + 414 + ], + "lines": [ + { + "bbox": [ + 207, + 385, + 403, + 414 + ], + "spans": [ + { + "bbox": [ + 207, + 385, + 403, + 414 + ], + "score": 0.94, + "content": "\\mathbf { h } _ { i } = \\sum _ { j } \\mathrm { s o f t m a x } ( \\mathbf { u } _ { h } ^ { T } \\mathbf { I } _ { i j } + \\mathbf { w } _ { i } ^ { T } \\mathbf { w } _ { j } b _ { h 1 } + b _ { h 2 } ) \\mathbf { I } _ { i j }", + "type": "interline_equation", + "image_path": "684f18d9d438d0c4c302c0fdfd67203049f99927f2fa867240af1df8d84a1418.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 207, + 385, + 403, + 399.5 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 207, + 399.5, + 403, + 414.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 419, + 505, + 475 + ], + "lines": [ + { + "bbox": [ + 106, + 419, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 133, + 431 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 421, + 146, + 430 + ], + "score": 0.86, + "content": "\\mathbf { u } _ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 419, + 165, + 431 + ], + "score": 1.0, + "content": "is a", + "type": "text" + }, + { + "bbox": [ + 166, + 420, + 177, + 430 + ], + "score": 0.89, + "content": "d _ { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 419, + 303, + 431 + ], + "score": 1.0, + "content": "-dimensional model parameter,", + "type": "text" + }, + { + "bbox": [ + 304, + 420, + 319, + 430 + ], + "score": 0.89, + "content": "b _ { h 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 419, + 338, + 431 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 338, + 420, + 353, + 430 + ], + "score": 0.89, + "content": "b _ { h 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 419, + 473, + 431 + ], + "score": 1.0, + "content": "are scalar model parameters,", + "type": "text" + }, + { + "bbox": [ + 473, + 421, + 486, + 430 + ], + "score": 0.85, + "content": "\\mathbf { w } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 419, + 505, + 431 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 430, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 120, + 443 + ], + "score": 0.87, + "content": "{ \\bf w } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 430, + 227, + 443 + ], + "score": 1.0, + "content": "are one-hot encodings for", + "type": "text" + }, + { + "bbox": [ + 228, + 432, + 237, + 442 + ], + "score": 0.85, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 430, + 255, + 443 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 256, + 432, + 266, + 442 + ], + "score": 0.86, + "content": "q _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 430, + 505, + 443 + ], + "score": 1.0, + "content": "respectively. We additionally use one-hot encodings in the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "attention mechanism to reinforce the matching between the same tokens since such information is", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 194, + 465 + ], + "score": 1.0, + "content": "not fully preserved in", + "type": "text" + }, + { + "bbox": [ + 194, + 452, + 206, + 464 + ], + "score": 0.89, + "content": "\\mathbf { I } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 452, + 231, + 465 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 232, + 453, + 239, + 462 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 452, + 450, + 465 + ], + "score": 1.0, + "content": "is large. The softmax nonlinearity is applied over all", + "type": "text" + }, + { + "bbox": [ + 450, + 453, + 456, + 464 + ], + "score": 0.77, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 452, + 506, + 465 + ], + "score": 1.0, + "content": "’s. The final", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 461, + 395, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 160, + 477 + ], + "score": 1.0, + "content": "hidden states", + "type": "text" + }, + { + "bbox": [ + 161, + 464, + 171, + 473 + ], + "score": 0.55, + "content": "\\mathbf { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 461, + 303, + 477 + ], + "score": 1.0, + "content": "are formed by concatenating the", + "type": "text" + }, + { + "bbox": [ + 303, + 464, + 314, + 474 + ], + "score": 0.88, + "content": "\\mathbf { h } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 461, + 381, + 477 + ], + "score": 1.0, + "content": "’s for each token", + "type": "text" + }, + { + "bbox": [ + 381, + 465, + 390, + 474 + ], + "score": 0.84, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 461, + 395, + 477 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 419, + 506, + 477 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 491, + 200, + 503 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 201, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 201, + 505 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 515, + 505, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 516, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 506, + 528 + ], + "score": 1.0, + "content": "We first present experimental results on the Twitter dataset where we can rule out the effects of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 526, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 539 + ], + "score": 1.0, + "content": "different choices of network architectures, to demonstrate the effectiveness of our word-character", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 537, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 551 + ], + "score": 1.0, + "content": "fine-grained gating approach. Later we show experiments on more challenging datasets on reading", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 548, + 504, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 504, + 561 + ], + "score": 1.0, + "content": "comprehension to further show that our approach can be used to improve the performance on high-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 559, + 204, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 204, + 571 + ], + "score": 1.0, + "content": "level NLP tasks as well.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 516, + 506, + 571 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 583, + 369, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 583, + 370, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 370, + 597 + ], + "score": 1.0, + "content": "4.1 EVALUATING WORD-CHARACTER GATING ON TWITTER", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 604, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "We evaluate the effectiveness of our word-character fine-grained gating mechanism on a social media", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "tag prediction task. We use the Twitter dataset and follow the experimental settings in Dhingra et al.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "(2016b). We also use the same network architecture upon the token representations, which is an", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 638, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 650 + ], + "score": 1.0, + "content": "LSTM layer followed by a softmax classification layer (Dhingra et al., 2016b). The Twitter dataset", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "consists of English tweets with at least one hashtag from Twitter. Hashtags and HTML tags have", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "been removed from the body of the tweet, and user names and URLs are replaced with special", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 506, + 684 + ], + "score": 1.0, + "content": "tokens. The dataset contains 2 million tweets for training, 10K for validation and 50K for testing,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 682, + 457, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 457, + 694 + ], + "score": 1.0, + "content": "with a total of 2,039 distinct hashtags. The task is to predict the hashtags of each tweet.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28.5, + "bbox_fs": [ + 104, + 604, + 506, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 107, + 699, + 505, + 710 + ], + "spans": [ + { + "bbox": [ + 107, + 699, + 505, + 710 + ], + "score": 1.0, + "content": "We compare several different methods as follows. Word char concat uses the concatenation of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "word-level and character-level representations as in Yang et al. (2016a); word char feat concat", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "concatenates the word-level and character-level representations along with the features described in", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 106, + 699, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 113, + 457, + 206 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 90, + 505, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 505, + 101 + ], + "score": 1.0, + "content": "Table 1: Performance on the Twitter dataset. “word” and “char” means using word-level and character-level", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 100, + 210, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 210, + 112 + ], + "score": 1.0, + "content": "representations respectively.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 153, + 113, + 457, + 206 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 153, + 113, + 457, + 206 + ], + "spans": [ + { + "bbox": [ + 153, + 113, + 457, + 206 + ], + "score": 0.975, + "html": "
ModelPrecision@1Recall@10Mean Rank
word (Dhingra et al., 2016b)0.2410.428133
char (Dhingra et al., 2016b)0.2840.485104
word char concat0.29610.4959105.8
word char feat concat0.29510.4974106.2
scalar gate0.29740.4982104.2
fine-grained gate0.30690.5119101.5
", + "type": "table", + "image_path": "1e9948e299167411bb5473aae70b288f9ae3451b9a8e13537e41dfaa0f65e39f.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 153, + 113, + 457, + 144.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 153, + 144.0, + 457, + 175.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 153, + 175.0, + 457, + 206.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 150, + 264, + 461, + 444 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 230, + 504, + 261 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 229, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 505, + 242 + ], + "score": 1.0, + "content": "Table 2: Performance on the CBT dataset. The “GA word char concat” results are extracted from Dhingra", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "et al. (2016a). Our results on fine-grained gating are based on a single model. “CN” and “NE” are two widely", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 251, + 403, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 403, + 262 + ], + "score": 1.0, + "content": "used question categories. “dev” means development set, and “test” means test set.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 150, + 264, + 461, + 444 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 150, + 264, + 461, + 444 + ], + "spans": [ + { + "bbox": [ + 150, + 264, + 461, + 444 + ], + "score": 0.98, + "html": "
ModelCN devCN testNE devNE test
GA word char concat0.7310.6960.7680.725
GA word char feat concat0.72500.69280.78150.7256
GA scalar gate0.72400.69080.78100.7260
GA fine-grained gate0.74250.70840.78900.7464
FG fine-grained gate0.75300.72040.79100.7496
Sordoni et al. (2016)0.7210.6920.7520.686
Trischler et al. (2016)0.7150.6740.7530.697
Cui et al. (2016)0.7220.6940.7780.720
Munkhdalai & Yu (2016)0.7430.7190.7820.732
Kadlec et al. (2016) ensemble0.7110.6890.7620.710
Sordoni et al. (2016) ensemble0.7410.7100.7690.720
Trischler etal. (2016) ensemble0.7360.7060.7660.718
", + "type": "table", + "image_path": "4cfc865c77165ae5c5ce61c9fe9499f3ad7205d3ab3bc654bd0f449197342325.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 150, + 264, + 461, + 324.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 150, + 324.0, + 461, + 384.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 150, + 384.0, + 461, + 444.0 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 108, + 470, + 504, + 503 + ], + "lines": [ + { + "bbox": [ + 106, + 469, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 506, + 483 + ], + "score": 1.0, + "content": "Section 3.2; scalar gate uses a scalar gate similar to Miyamoto & Cho (2016) but is conditioned on", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 481, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 481, + 505, + 493 + ], + "score": 1.0, + "content": "the features; fine-grained gate is our method described in Section 3.2. We include word char feat", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 492, + 503, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 503, + 505 + ], + "score": 1.0, + "content": "concat for a fair comparison because our fine-grained gating approach also uses the token features.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 508, + 505, + 608 + ], + "lines": [ + { + "bbox": [ + 105, + 507, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 472, + 521 + ], + "score": 1.0, + "content": "The results are shown in Table 1. We report three evaluation metrics including precision", + "type": "text" + }, + { + "bbox": [ + 472, + 509, + 486, + 519 + ], + "score": 0.64, + "content": "@ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 507, + 505, + 521 + ], + "score": 1.0, + "content": ", re-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 519, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 122, + 533 + ], + "score": 1.0, + "content": "call", + "type": "text" + }, + { + "bbox": [ + 122, + 520, + 141, + 530 + ], + "score": 0.59, + "content": "@ 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 519, + 505, + 533 + ], + "score": 1.0, + "content": ", and mean rank. Our method outperforms character-level models used in Dhingra et al.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 530, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 153, + 544 + ], + "score": 1.0, + "content": "(2016b) by", + "type": "text" + }, + { + "bbox": [ + 154, + 531, + 181, + 542 + ], + "score": 0.84, + "content": "2 . 2 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 530, + 185, + 544 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 185, + 531, + 212, + 542 + ], + "score": 0.85, + "content": "2 . 6 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 530, + 505, + 544 + ], + "score": 1.0, + "content": ", and 2.5 points in terms of precision, recall and mean rank respectively.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "We can observe that scalar gating approach (Miyamoto & Cho, 2016) can only marginally improve", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 552, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 566 + ], + "score": 1.0, + "content": "over the baseline methods, while fine-grained gating methods can substantially improve model per-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "score": 1.0, + "content": "formance. Note that directly concatenating the token features with the character-level and word-level", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "representations does not boost the performance, but using the token features to compute a gate (as", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 586, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 505, + 597 + ], + "score": 1.0, + "content": "done in fine-grained gating) leads to better results. This indicates that the benefit of fine-grained", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 597, + 428, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 428, + 609 + ], + "score": 1.0, + "content": "gating mainly comes from better modeling rather than using additional features.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 107, + 622, + 333, + 633 + ], + "lines": [ + { + "bbox": [ + 106, + 622, + 335, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 335, + 635 + ], + "score": 1.0, + "content": "4.2 PERFORMANCE ON READING COMPREHENSION", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 108, + 643, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 106, + 643, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 505, + 656 + ], + "score": 1.0, + "content": "After investigating the effectiveness of the word-character fine-grained gating mechanism on the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "Twitter dataset, we now move on to a more challenging task, reading comprehension. In this section,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "we experiment with two datasets, the Children’s Book Test dataset (Hill et al., 2016) and the SQuAD", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 676, + 235, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 235, + 689 + ], + "score": 1.0, + "content": "dataset (Rajpurkar et al., 2016).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + }, + { + "type": "title", + "bbox": [ + 108, + 701, + 254, + 712 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 256, + 715 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 256, + 715 + ], + "score": 1.0, + "content": "4.2.1 CLOZE-STYLE QUESTIONS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 720, + 390, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 719, + 391, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 719, + 391, + 734 + ], + "score": 1.0, + "content": "We evaluate our model on cloze-style question answering benchmarks.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 113, + 457, + 206 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 90, + 505, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 505, + 101 + ], + "score": 1.0, + "content": "Table 1: Performance on the Twitter dataset. “word” and “char” means using word-level and character-level", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 100, + 210, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 210, + 112 + ], + "score": 1.0, + "content": "representations respectively.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 153, + 113, + 457, + 206 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 153, + 113, + 457, + 206 + ], + "spans": [ + { + "bbox": [ + 153, + 113, + 457, + 206 + ], + "score": 0.975, + "html": "
ModelPrecision@1Recall@10Mean Rank
word (Dhingra et al., 2016b)0.2410.428133
char (Dhingra et al., 2016b)0.2840.485104
word char concat0.29610.4959105.8
word char feat concat0.29510.4974106.2
scalar gate0.29740.4982104.2
fine-grained gate0.30690.5119101.5
", + "type": "table", + "image_path": "1e9948e299167411bb5473aae70b288f9ae3451b9a8e13537e41dfaa0f65e39f.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 153, + 113, + 457, + 144.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 153, + 144.0, + 457, + 175.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 153, + 175.0, + 457, + 206.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 150, + 264, + 461, + 444 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 230, + 504, + 261 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 229, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 505, + 242 + ], + "score": 1.0, + "content": "Table 2: Performance on the CBT dataset. The “GA word char concat” results are extracted from Dhingra", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "et al. (2016a). Our results on fine-grained gating are based on a single model. “CN” and “NE” are two widely", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 251, + 403, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 403, + 262 + ], + "score": 1.0, + "content": "used question categories. “dev” means development set, and “test” means test set.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 150, + 264, + 461, + 444 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 150, + 264, + 461, + 444 + ], + "spans": [ + { + "bbox": [ + 150, + 264, + 461, + 444 + ], + "score": 0.98, + "html": "
ModelCN devCN testNE devNE test
GA word char concat0.7310.6960.7680.725
GA word char feat concat0.72500.69280.78150.7256
GA scalar gate0.72400.69080.78100.7260
GA fine-grained gate0.74250.70840.78900.7464
FG fine-grained gate0.75300.72040.79100.7496
Sordoni et al. (2016)0.7210.6920.7520.686
Trischler et al. (2016)0.7150.6740.7530.697
Cui et al. (2016)0.7220.6940.7780.720
Munkhdalai & Yu (2016)0.7430.7190.7820.732
Kadlec et al. (2016) ensemble0.7110.6890.7620.710
Sordoni et al. (2016) ensemble0.7410.7100.7690.720
Trischler etal. (2016) ensemble0.7360.7060.7660.718
", + "type": "table", + "image_path": "4cfc865c77165ae5c5ce61c9fe9499f3ad7205d3ab3bc654bd0f449197342325.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 150, + 264, + 461, + 324.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 150, + 324.0, + 461, + 384.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 150, + 384.0, + 461, + 444.0 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 108, + 470, + 504, + 503 + ], + "lines": [ + { + "bbox": [ + 106, + 469, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 506, + 483 + ], + "score": 1.0, + "content": "Section 3.2; scalar gate uses a scalar gate similar to Miyamoto & Cho (2016) but is conditioned on", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 481, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 481, + 505, + 493 + ], + "score": 1.0, + "content": "the features; fine-grained gate is our method described in Section 3.2. We include word char feat", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 492, + 503, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 503, + 505 + ], + "score": 1.0, + "content": "concat for a fair comparison because our fine-grained gating approach also uses the token features.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 469, + 506, + 505 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 508, + 505, + 608 + ], + "lines": [ + { + "bbox": [ + 105, + 507, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 472, + 521 + ], + "score": 1.0, + "content": "The results are shown in Table 1. We report three evaluation metrics including precision", + "type": "text" + }, + { + "bbox": [ + 472, + 509, + 486, + 519 + ], + "score": 0.64, + "content": "@ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 507, + 505, + 521 + ], + "score": 1.0, + "content": ", re-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 519, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 122, + 533 + ], + "score": 1.0, + "content": "call", + "type": "text" + }, + { + "bbox": [ + 122, + 520, + 141, + 530 + ], + "score": 0.59, + "content": "@ 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 519, + 505, + 533 + ], + "score": 1.0, + "content": ", and mean rank. Our method outperforms character-level models used in Dhingra et al.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 530, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 153, + 544 + ], + "score": 1.0, + "content": "(2016b) by", + "type": "text" + }, + { + "bbox": [ + 154, + 531, + 181, + 542 + ], + "score": 0.84, + "content": "2 . 2 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 530, + 185, + 544 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 185, + 531, + 212, + 542 + ], + "score": 0.85, + "content": "2 . 6 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 530, + 505, + 544 + ], + "score": 1.0, + "content": ", and 2.5 points in terms of precision, recall and mean rank respectively.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "We can observe that scalar gating approach (Miyamoto & Cho, 2016) can only marginally improve", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 552, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 566 + ], + "score": 1.0, + "content": "over the baseline methods, while fine-grained gating methods can substantially improve model per-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "score": 1.0, + "content": "formance. Note that directly concatenating the token features with the character-level and word-level", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "representations does not boost the performance, but using the token features to compute a gate (as", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 586, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 505, + 597 + ], + "score": 1.0, + "content": "done in fine-grained gating) leads to better results. This indicates that the benefit of fine-grained", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 597, + 428, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 428, + 609 + ], + "score": 1.0, + "content": "gating mainly comes from better modeling rather than using additional features.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 507, + 505, + 609 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 622, + 333, + 633 + ], + "lines": [ + { + "bbox": [ + 106, + 622, + 335, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 335, + 635 + ], + "score": 1.0, + "content": "4.2 PERFORMANCE ON READING COMPREHENSION", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 108, + 643, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 106, + 643, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 505, + 656 + ], + "score": 1.0, + "content": "After investigating the effectiveness of the word-character fine-grained gating mechanism on the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "Twitter dataset, we now move on to a more challenging task, reading comprehension. In this section,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "we experiment with two datasets, the Children’s Book Test dataset (Hill et al., 2016) and the SQuAD", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 676, + 235, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 235, + 689 + ], + "score": 1.0, + "content": "dataset (Rajpurkar et al., 2016).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 643, + 505, + 689 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 701, + 254, + 712 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 256, + 715 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 256, + 715 + ], + "score": 1.0, + "content": "4.2.1 CLOZE-STYLE QUESTIONS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 720, + 390, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 719, + 391, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 719, + 391, + 734 + ], + "score": 1.0, + "content": "We evaluate our model on cloze-style question answering benchmarks.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29, + "bbox_fs": [ + 106, + 719, + 391, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 131, + 112, + 480, + 204 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 90, + 503, + 110 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 88, + 505, + 102 + ], + "spans": [ + { + "bbox": [ + 106, + 88, + 505, + 102 + ], + "score": 1.0, + "content": "Table 3: Performance on the Who Did What dataset. “dev” means development set, and “test” means test set.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 263, + 111 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 263, + 111 + ], + "score": 1.0, + "content": "“WDW-R” is the relaxed version of WDW.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 131, + 112, + 480, + 204 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 131, + 112, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 131, + 112, + 480, + 204 + ], + "score": 0.974, + "html": "
ModelWDW devWDW testWDW-R devWDW-R test
0.5700.590
Kadlec et al. (2016) Chen et al. (2016)0.640一0.650
Munkhdalai & Yu (2016)0.6650.6620.6700.667
Dhingra et al. (2016a)0.7160.7120.7260.726
this paper0.7230.7170.7310.726
", + "type": "table", + "image_path": "bc2d44dcd815972deda2b31e09a5734abfad4815dd67c08b0fc9d6997c0946c5.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 131, + 112, + 480, + 142.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 131, + 142.66666666666666, + 480, + 173.33333333333331 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 131, + 173.33333333333331, + 480, + 203.99999999999997 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 161, + 243, + 449, + 379 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 142, + 229, + 468, + 240 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 141, + 228, + 469, + 241 + ], + "spans": [ + { + "bbox": [ + 141, + 228, + 469, + 241 + ], + "score": 1.0, + "content": "Table 4: Performance on the SQuAD dev set. Test set results are included in the brackets.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "table_body", + "bbox": [ + 161, + 243, + 449, + 379 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 161, + 243, + 449, + 379 + ], + "spans": [ + { + "bbox": [ + 161, + 243, + 449, + 379 + ], + "score": 0.975, + "html": "
ModelF1Exact Match
GA word0.5492
GA word char concat0.6695 0.68570.5639
GA word char feat concat0.69040.5711
GA scalar gate0.68500.5620
GA fine-grained gate0.69830.5804
FG fine-grained gate0.71250.5995
FG fine-grained gate + ensemble0.7341 (0.733)0.6238 (0.625)
Yu et al. (2016)
Wang & Jiang (2016)0.712 (0.710) 0.700 (0.703)0.625 (0.625) 0.591 (0.595)
", + "type": "table", + "image_path": "d3e004cee2ede88cd486bc6c12041115e0656f94801d13ba78306e10ff2e9d0a.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 161, + 243, + 449, + 288.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 161, + 288.3333333333333, + 449, + 333.66666666666663 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 161, + 333.66666666666663, + 449, + 378.99999999999994 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 106, + 405, + 505, + 471 + ], + "lines": [ + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "score": 1.0, + "content": "The Children’s Book Test (CBT) dataset is built from children’s books. The whole dataset has", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "669,343 questions for training, 8,000 for validation and 10,000 for testing. We closely follow the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 426, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 506, + 440 + ], + "score": 1.0, + "content": "setting in Dhingra et al. (2016a) and incrementally add different components to see the changes in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 438, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 506, + 452 + ], + "score": 1.0, + "content": "performance. For the fine-grained gating approach, we use the same hyper-parameters as in Dhingra", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 450, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 461 + ], + "score": 1.0, + "content": "et al. (2016a) except that we use a character-level GRU with 100 units to be of the same size as the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 459, + 389, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 389, + 474 + ], + "score": 1.0, + "content": "word lookup table. The word embeddings are updated during training.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 108, + 477, + 502, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "score": 1.0, + "content": "In addition to different ways of combining word-level and character-level representations, we also", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 488, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 505, + 500 + ], + "score": 1.0, + "content": "compare two different ways of integrating documents and queries: GA refers to the gated attention", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 499, + 490, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 490, + 512 + ], + "score": 1.0, + "content": "reader (Dhingra et al., 2016a) and FG refers to our fine-grained gating described in Section 3.3.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 515, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "The results are reported in Table 2. We report the results on common noun (CN) questions and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 527, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 539 + ], + "score": 1.0, + "content": "named entity (NE) questions, which are two widely used question categories in CBT. Our fine-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 538, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 550 + ], + "score": 1.0, + "content": "grained gating approach achieves new state-of-the-art performance on both settings and outperforms", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 549, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 274, + 561 + ], + "score": 1.0, + "content": "the current state-of-the-art results by up to", + "type": "text" + }, + { + "bbox": [ + 274, + 549, + 301, + 559 + ], + "score": 0.88, + "content": "1 . 7 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 549, + 505, + 561 + ], + "score": 1.0, + "content": "without using ensembles. Our method outperforms", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 558, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 238, + 574 + ], + "score": 1.0, + "content": "the baseline GA reader by up to", + "type": "text" + }, + { + "bbox": [ + 239, + 560, + 261, + 570 + ], + "score": 0.88, + "content": "2 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 558, + 506, + 574 + ], + "score": 1.0, + "content": ", which indicates the effectiveness of the fine-grained gating", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 571, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 505, + 582 + ], + "score": 1.0, + "content": "mechanism. Consistent with the results on the Twitter dataset, using word-character fine-grained", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "score": 1.0, + "content": "gating can substantially improve the performance over concatenation or scalar gating. Furthermore,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 593, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 605 + ], + "score": 1.0, + "content": "we can see that document-query fine-grained gating also contributes significantly to the final results.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 108, + 609, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 106, + 609, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 622 + ], + "score": 1.0, + "content": "We also apply our fine-grained gating model to the Who Did What (WDW) dataset (Onishi et al.,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 620, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 505, + 633 + ], + "score": 1.0, + "content": "2016). As shown in Table 3, our model achieves state-of-the-art results compared to strong baselines.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 631, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 287, + 645 + ], + "score": 1.0, + "content": "We fix the word embeddings during training.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 108, + 656, + 175, + 668 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 176, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 176, + 670 + ], + "score": 1.0, + "content": "4.2.2 SQUAD", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 108, + 676, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "The Stanford Question Answering Dataset (SQuAD) is a reading comprehension dataset collected", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "recently (Rajpurkar et al., 2016). It contains 23,215 paragraphs come from 536 Wikipedia articles.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "Unlike other reading comprehension datasets such as CBT, the answers are a span of text rather than", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 348, + 723 + ], + "score": 1.0, + "content": "a single word. The dataset is partitioned into a training set (", + "type": "text" + }, + { + "bbox": [ + 348, + 710, + 367, + 721 + ], + "score": 0.75, + "content": "80 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 709, + 506, + 723 + ], + "score": 1.0, + "content": ", 87,636 question-answer pairs), a", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 720, + 466, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 177, + 734 + ], + "score": 1.0, + "content": "development set (", + "type": "text" + }, + { + "bbox": [ + 177, + 721, + 196, + 731 + ], + "score": 0.69, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 720, + 466, + 734 + ], + "score": 1.0, + "content": ", 10,600 question-answer pairs) and a test set which is not released.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 131, + 112, + 480, + 204 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 90, + 503, + 110 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 88, + 505, + 102 + ], + "spans": [ + { + "bbox": [ + 106, + 88, + 505, + 102 + ], + "score": 1.0, + "content": "Table 3: Performance on the Who Did What dataset. “dev” means development set, and “test” means test set.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 263, + 111 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 263, + 111 + ], + "score": 1.0, + "content": "“WDW-R” is the relaxed version of WDW.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 131, + 112, + 480, + 204 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 131, + 112, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 131, + 112, + 480, + 204 + ], + "score": 0.974, + "html": "
ModelWDW devWDW testWDW-R devWDW-R test
0.5700.590
Kadlec et al. (2016) Chen et al. (2016)0.640一0.650
Munkhdalai & Yu (2016)0.6650.6620.6700.667
Dhingra et al. (2016a)0.7160.7120.7260.726
this paper0.7230.7170.7310.726
", + "type": "table", + "image_path": "bc2d44dcd815972deda2b31e09a5734abfad4815dd67c08b0fc9d6997c0946c5.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 131, + 112, + 480, + 142.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 131, + 142.66666666666666, + 480, + 173.33333333333331 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 131, + 173.33333333333331, + 480, + 203.99999999999997 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 161, + 243, + 449, + 379 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 142, + 229, + 468, + 240 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 141, + 228, + 469, + 241 + ], + "spans": [ + { + "bbox": [ + 141, + 228, + 469, + 241 + ], + "score": 1.0, + "content": "Table 4: Performance on the SQuAD dev set. Test set results are included in the brackets.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "table_body", + "bbox": [ + 161, + 243, + 449, + 379 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 161, + 243, + 449, + 379 + ], + "spans": [ + { + "bbox": [ + 161, + 243, + 449, + 379 + ], + "score": 0.975, + "html": "
ModelF1Exact Match
GA word0.5492
GA word char concat0.6695 0.68570.5639
GA word char feat concat0.69040.5711
GA scalar gate0.68500.5620
GA fine-grained gate0.69830.5804
FG fine-grained gate0.71250.5995
FG fine-grained gate + ensemble0.7341 (0.733)0.6238 (0.625)
Yu et al. (2016)
Wang & Jiang (2016)0.712 (0.710) 0.700 (0.703)0.625 (0.625) 0.591 (0.595)
", + "type": "table", + "image_path": "d3e004cee2ede88cd486bc6c12041115e0656f94801d13ba78306e10ff2e9d0a.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 161, + 243, + 449, + 288.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 161, + 288.3333333333333, + 449, + 333.66666666666663 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 161, + 333.66666666666663, + 449, + 378.99999999999994 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 106, + 405, + 505, + 471 + ], + "lines": [ + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "score": 1.0, + "content": "The Children’s Book Test (CBT) dataset is built from children’s books. The whole dataset has", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "669,343 questions for training, 8,000 for validation and 10,000 for testing. We closely follow the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 426, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 506, + 440 + ], + "score": 1.0, + "content": "setting in Dhingra et al. (2016a) and incrementally add different components to see the changes in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 438, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 506, + 452 + ], + "score": 1.0, + "content": "performance. For the fine-grained gating approach, we use the same hyper-parameters as in Dhingra", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 450, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 461 + ], + "score": 1.0, + "content": "et al. (2016a) except that we use a character-level GRU with 100 units to be of the same size as the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 459, + 389, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 389, + 474 + ], + "score": 1.0, + "content": "word lookup table. The word embeddings are updated during training.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 405, + 506, + 474 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 477, + 502, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "score": 1.0, + "content": "In addition to different ways of combining word-level and character-level representations, we also", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 488, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 505, + 500 + ], + "score": 1.0, + "content": "compare two different ways of integrating documents and queries: GA refers to the gated attention", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 499, + 490, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 490, + 512 + ], + "score": 1.0, + "content": "reader (Dhingra et al., 2016a) and FG refers to our fine-grained gating described in Section 3.3.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 476, + 505, + 512 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 515, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "The results are reported in Table 2. We report the results on common noun (CN) questions and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 527, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 539 + ], + "score": 1.0, + "content": "named entity (NE) questions, which are two widely used question categories in CBT. Our fine-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 538, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 550 + ], + "score": 1.0, + "content": "grained gating approach achieves new state-of-the-art performance on both settings and outperforms", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 549, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 274, + 561 + ], + "score": 1.0, + "content": "the current state-of-the-art results by up to", + "type": "text" + }, + { + "bbox": [ + 274, + 549, + 301, + 559 + ], + "score": 0.88, + "content": "1 . 7 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 549, + 505, + 561 + ], + "score": 1.0, + "content": "without using ensembles. Our method outperforms", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 558, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 238, + 574 + ], + "score": 1.0, + "content": "the baseline GA reader by up to", + "type": "text" + }, + { + "bbox": [ + 239, + 560, + 261, + 570 + ], + "score": 0.88, + "content": "2 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 558, + 506, + 574 + ], + "score": 1.0, + "content": ", which indicates the effectiveness of the fine-grained gating", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 571, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 505, + 582 + ], + "score": 1.0, + "content": "mechanism. Consistent with the results on the Twitter dataset, using word-character fine-grained", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "score": 1.0, + "content": "gating can substantially improve the performance over concatenation or scalar gating. Furthermore,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 593, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 605 + ], + "score": 1.0, + "content": "we can see that document-query fine-grained gating also contributes significantly to the final results.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 516, + 506, + 605 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 609, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 106, + 609, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 622 + ], + "score": 1.0, + "content": "We also apply our fine-grained gating model to the Who Did What (WDW) dataset (Onishi et al.,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 620, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 505, + 633 + ], + "score": 1.0, + "content": "2016). As shown in Table 3, our model achieves state-of-the-art results compared to strong baselines.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 631, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 287, + 645 + ], + "score": 1.0, + "content": "We fix the word embeddings during training.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 609, + 505, + 645 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 656, + 175, + 668 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 176, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 176, + 670 + ], + "score": 1.0, + "content": "4.2.2 SQUAD", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 108, + 676, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "The Stanford Question Answering Dataset (SQuAD) is a reading comprehension dataset collected", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "recently (Rajpurkar et al., 2016). It contains 23,215 paragraphs come from 536 Wikipedia articles.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "Unlike other reading comprehension datasets such as CBT, the answers are a span of text rather than", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 348, + 723 + ], + "score": 1.0, + "content": "a single word. The dataset is partitioned into a training set (", + "type": "text" + }, + { + "bbox": [ + 348, + 710, + 367, + 721 + ], + "score": 0.75, + "content": "80 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 709, + 506, + 723 + ], + "score": 1.0, + "content": ", 87,636 question-answer pairs), a", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 720, + 466, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 177, + 734 + ], + "score": 1.0, + "content": "development set (", + "type": "text" + }, + { + "bbox": [ + 177, + 721, + 196, + 731 + ], + "score": 0.69, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 720, + 466, + 734 + ], + "score": 1.0, + "content": ", 10,600 question-answer pairs) and a test set which is not released.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 676, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 82, + 502, + 171 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 82, + 502, + 171 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 82, + 502, + 171 + ], + "spans": [ + { + "bbox": [ + 109, + 82, + 502, + 171 + ], + "score": 0.951, + "type": "image", + "image_path": "89ddff74b55246bf812f8cbb6330db8793bb826921f94510c5ea362754ea15ba.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 82, + 502, + 111.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 111.66666666666667, + 502, + 141.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 141.33333333333334, + 502, + 171.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 184, + 505, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 184, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 268, + 198 + ], + "score": 1.0, + "content": "Figure 3: Visualization of the weight matrix", + "type": "text" + }, + { + "bbox": [ + 268, + 185, + 285, + 196 + ], + "score": 0.89, + "content": "\\mathbf { W } _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 184, + 505, + 198 + ], + "score": 1.0, + "content": ". Weights for each features are averaged. Red means high and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "yellow means low. High weight values favor character-level representations, and low weight values favor word-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "level representations. “Organization”, “‘Person”, “Location”, and “O” are named entity tags; “DOCLEN-n”", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 214, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 252, + 227 + ], + "score": 1.0, + "content": "are document frequency features (larger", + "type": "text" + }, + { + "bbox": [ + 252, + 217, + 259, + 224 + ], + "score": 0.62, + "content": "_ n", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 214, + 349, + 227 + ], + "score": 1.0, + "content": "means higher frequency,", + "type": "text" + }, + { + "bbox": [ + 350, + 217, + 356, + 224 + ], + "score": 0.72, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 214, + 481, + 227 + ], + "score": 1.0, + "content": "from 0 to 4); others are POS tags.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "image", + "bbox": [ + 111, + 238, + 500, + 384 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 238, + 500, + 384 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 111, + 238, + 500, + 384 + ], + "spans": [ + { + "bbox": [ + 111, + 238, + 500, + 384 + ], + "score": 0.972, + "type": "image", + "image_path": "e182272ce273a6e2c7869e53e7be53700145a9a8f9ecfbc54ac0980409cc5dd8.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 111, + 238, + 500, + 286.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 111, + 286.6666666666667, + 500, + 335.33333333333337 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 111, + 335.33333333333337, + 500, + 384.00000000000006 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 398, + 504, + 419 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "Figure 4: Visualization of gate values in the text. Red means high and yellow means low. High gate values", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 408, + 434, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 434, + 420 + ], + "score": 1.0, + "content": "favor character-level representations, and low gate values favor word-level representations.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + } + ], + "index": 9.25 + }, + { + "type": "text", + "bbox": [ + 106, + 459, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 458, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 506, + 472 + ], + "score": 1.0, + "content": "We report our results in Table 4. “Exact match” computes the ratio of questions that are answered", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 470, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 505, + 482 + ], + "score": 1.0, + "content": "correctly by strict string comparison, and the F1 score is computed on the token level. We can", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 480, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 495 + ], + "score": 1.0, + "content": "observe that both word-character fine-grained gating and document-query fine-grained gating can", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 493, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 505, + 505 + ], + "score": 1.0, + "content": "substantially improve the performance, leading to state-of-the-art results among published papers.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 503, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 505, + 515 + ], + "score": 1.0, + "content": "Note that at the time of submission, the best score on the leaderboard is 0.716 in exact match and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 513, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 104, + 513, + 506, + 527 + ], + "score": 1.0, + "content": "0.804 in F1 without published papers. A gap exists because our architecture described in Section", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 525, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 537 + ], + "score": 1.0, + "content": "3.1 does not specifically model the answer span structure that is unique to SQuAD. In this work, we", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 536, + 487, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 487, + 549 + ], + "score": 1.0, + "content": "focus on this general architecture to study the effectiveness of fine-grained gating mechanisms.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 108, + 578, + 266, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 576, + 269, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 269, + 591 + ], + "score": 1.0, + "content": "4.3 VISUALIZATION AND ANALYSIS", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 244, + 618 + ], + "score": 1.0, + "content": "We visualize the model parameter", + "type": "text" + }, + { + "bbox": [ + 244, + 605, + 261, + 617 + ], + "score": 0.88, + "content": "\\mathbf { W } _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "as described in Section 3.2. For each feature, we average the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 234, + 628 + ], + "score": 1.0, + "content": "corresponding weight vector in", + "type": "text" + }, + { + "bbox": [ + 234, + 616, + 252, + 628 + ], + "score": 0.88, + "content": "\\mathbf { W } _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 615, + 505, + 628 + ], + "score": 1.0, + "content": ". The results are described in Figure 3. We can see that named", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "entities like “Organization” and noun phrases (with tags “NNP” or “NNPS”) tend to use character-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "level representations, which is consistent with human intuition because those tokens are usually", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "infrequent or display rich morphologies. Also, DOCLEN-4, WH-adverb (“WRB”), and conjunction", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 659, + 486, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 486, + 673 + ], + "score": 1.0, + "content": "(“IN” and “CC”) tokens tend to use word-level representations because they appear frequently.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 108, + 676, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "We also sample random span of text from the SQuAD dataset, and visualize the average gate values", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "in Figure 4. The results are consistent with our observations in Figure 3. Rare tokens, noun phrases,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 700, + 505, + 710 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 710 + ], + "score": 1.0, + "content": "and named entities tend to use character-level representations, while others tend to use word-level", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "representations. To further justify this argument, we also list the tokens with highest and lowest gate", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 721, + 179, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 179, + 732 + ], + "score": 1.0, + "content": "values in Table 5.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 82, + 502, + 171 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 82, + 502, + 171 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 82, + 502, + 171 + ], + "spans": [ + { + "bbox": [ + 109, + 82, + 502, + 171 + ], + "score": 0.951, + "type": "image", + "image_path": "89ddff74b55246bf812f8cbb6330db8793bb826921f94510c5ea362754ea15ba.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 82, + 502, + 111.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 111.66666666666667, + 502, + 141.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 141.33333333333334, + 502, + 171.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 184, + 505, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 184, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 268, + 198 + ], + "score": 1.0, + "content": "Figure 3: Visualization of the weight matrix", + "type": "text" + }, + { + "bbox": [ + 268, + 185, + 285, + 196 + ], + "score": 0.89, + "content": "\\mathbf { W } _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 184, + 505, + 198 + ], + "score": 1.0, + "content": ". Weights for each features are averaged. Red means high and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "yellow means low. High weight values favor character-level representations, and low weight values favor word-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "level representations. “Organization”, “‘Person”, “Location”, and “O” are named entity tags; “DOCLEN-n”", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 214, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 252, + 227 + ], + "score": 1.0, + "content": "are document frequency features (larger", + "type": "text" + }, + { + "bbox": [ + 252, + 217, + 259, + 224 + ], + "score": 0.62, + "content": "_ n", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 214, + 349, + 227 + ], + "score": 1.0, + "content": "means higher frequency,", + "type": "text" + }, + { + "bbox": [ + 350, + 217, + 356, + 224 + ], + "score": 0.72, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 214, + 481, + 227 + ], + "score": 1.0, + "content": "from 0 to 4); others are POS tags.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "image", + "bbox": [ + 111, + 238, + 500, + 384 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 238, + 500, + 384 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 111, + 238, + 500, + 384 + ], + "spans": [ + { + "bbox": [ + 111, + 238, + 500, + 384 + ], + "score": 0.972, + "type": "image", + "image_path": "e182272ce273a6e2c7869e53e7be53700145a9a8f9ecfbc54ac0980409cc5dd8.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 111, + 238, + 500, + 286.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 111, + 286.6666666666667, + 500, + 335.33333333333337 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 111, + 335.33333333333337, + 500, + 384.00000000000006 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 398, + 504, + 419 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "Figure 4: Visualization of gate values in the text. Red means high and yellow means low. High gate values", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 408, + 434, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 434, + 420 + ], + "score": 1.0, + "content": "favor character-level representations, and low gate values favor word-level representations.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + } + ], + "index": 9.25 + }, + { + "type": "text", + "bbox": [ + 106, + 459, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 458, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 506, + 472 + ], + "score": 1.0, + "content": "We report our results in Table 4. “Exact match” computes the ratio of questions that are answered", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 470, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 505, + 482 + ], + "score": 1.0, + "content": "correctly by strict string comparison, and the F1 score is computed on the token level. We can", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 480, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 495 + ], + "score": 1.0, + "content": "observe that both word-character fine-grained gating and document-query fine-grained gating can", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 493, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 505, + 505 + ], + "score": 1.0, + "content": "substantially improve the performance, leading to state-of-the-art results among published papers.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 503, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 505, + 515 + ], + "score": 1.0, + "content": "Note that at the time of submission, the best score on the leaderboard is 0.716 in exact match and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 513, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 104, + 513, + 506, + 527 + ], + "score": 1.0, + "content": "0.804 in F1 without published papers. A gap exists because our architecture described in Section", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 525, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 537 + ], + "score": 1.0, + "content": "3.1 does not specifically model the answer span structure that is unique to SQuAD. In this work, we", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 536, + 487, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 487, + 549 + ], + "score": 1.0, + "content": "focus on this general architecture to study the effectiveness of fine-grained gating mechanisms.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15.5, + "bbox_fs": [ + 104, + 458, + 506, + 549 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 578, + 266, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 576, + 269, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 269, + 591 + ], + "score": 1.0, + "content": "4.3 VISUALIZATION AND ANALYSIS", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 244, + 618 + ], + "score": 1.0, + "content": "We visualize the model parameter", + "type": "text" + }, + { + "bbox": [ + 244, + 605, + 261, + 617 + ], + "score": 0.88, + "content": "\\mathbf { W } _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "as described in Section 3.2. For each feature, we average the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 234, + 628 + ], + "score": 1.0, + "content": "corresponding weight vector in", + "type": "text" + }, + { + "bbox": [ + 234, + 616, + 252, + 628 + ], + "score": 0.88, + "content": "\\mathbf { W } _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 615, + 505, + 628 + ], + "score": 1.0, + "content": ". The results are described in Figure 3. We can see that named", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "entities like “Organization” and noun phrases (with tags “NNP” or “NNPS”) tend to use character-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "level representations, which is consistent with human intuition because those tokens are usually", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "infrequent or display rich morphologies. Also, DOCLEN-4, WH-adverb (“WRB”), and conjunction", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 659, + 486, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 486, + 673 + ], + "score": 1.0, + "content": "(“IN” and “CC”) tokens tend to use word-level representations because they appear frequently.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 604, + 505, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 676, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "We also sample random span of text from the SQuAD dataset, and visualize the average gate values", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "in Figure 4. The results are consistent with our observations in Figure 3. Rare tokens, noun phrases,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 700, + 505, + 710 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 710 + ], + "score": 1.0, + "content": "and named entities tend to use character-level representations, while others tend to use word-level", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "representations. To further justify this argument, we also list the tokens with highest and lowest gate", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 721, + 179, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 179, + 732 + ], + "score": 1.0, + "content": "values in Table 5.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 676, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 123, + 504, + 248 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 90, + 504, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 504, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 504, + 102 + ], + "score": 1.0, + "content": "Table 5: Word tokens with highest and lowest gate values. High gate values favor character-level representa-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 100, + 321, + 111 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 321, + 111 + ], + "score": 1.0, + "content": "tions, and low gate values favor word-level representations.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 106, + 123, + 504, + 248 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 123, + 504, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 504, + 248 + ], + "score": 0.969, + "html": "
Gate valuesWord tokens
Lowestor but But These these However however among Among that when When although Although because Because until many Many than though Though this This Since since date where Where have That and And Such such number so which by By how before Before with With between Between even Even if
HighestSweetgum Untersee Jianlong Floresta Chlorella Obersee PhT Doctorin Jumonville WFTS WTSP Boven Pharm Nederrijn Otrar Rhin Magicicada WBKB Tanzler KMBC WPLG Mainau Merwede RMJM Kleitman Scheur Bodensee Kromme Horenbout Vorderrhein Chlamydomonas Scantlebury Qingshui Funchess
", + "type": "table", + "image_path": "3d88f65fd35053aed1c57fb91abce4362deea2ce3a4b72dcac738b31312c5589.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 106, + 123, + 504, + 164.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 106, + 164.66666666666666, + 504, + 206.33333333333331 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 206.33333333333331, + 504, + 247.99999999999997 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "title", + "bbox": [ + 108, + 271, + 201, + 284 + ], + "lines": [ + { + "bbox": [ + 104, + 269, + 203, + 287 + ], + "spans": [ + { + "bbox": [ + 104, + 269, + 203, + 287 + ], + "score": 1.0, + "content": "5 CONCLUSIONS", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 296, + 505, + 374 + ], + "lines": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "We present a fine-grained gating mechanism that dynamically combines word-level and character-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 308, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 320 + ], + "score": 1.0, + "content": "level representations based on word properties. Experiments on the Twitter tag prediction dataset", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "show that fine-grained gating substantially outperforms scalar gating and concatenation. Our method", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 330, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 342 + ], + "score": 1.0, + "content": "also improves the performance on reading comprehension and achieves new state-of-the-art results", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 354 + ], + "score": 1.0, + "content": "on CBT and WDW. In our future work, we plan to to apply the fine-grained gating mechanism for", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 350, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 506, + 365 + ], + "score": 1.0, + "content": "combining other levels of representations, such as phrases and sentences. It will also be intriguing", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 362, + 493, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 493, + 375 + ], + "score": 1.0, + "content": "to integrate NER and POS networks and learn the token representation in an end-to-end manner.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9 + }, + { + "type": "title", + "bbox": [ + 108, + 387, + 200, + 397 + ], + "lines": [ + { + "bbox": [ + 107, + 388, + 200, + 398 + ], + "spans": [ + { + "bbox": [ + 107, + 388, + 200, + 398 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 503, + 428 + ], + "lines": [ + { + "bbox": [ + 105, + 404, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 505, + 419 + ], + "score": 1.0, + "content": "This work was funded by NVIDIA, the Office of Naval Research Scene Understanding grant", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 416, + 378, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 378, + 429 + ], + "score": 1.0, + "content": "N000141310721, the NSF grant IIS1250956, and Google Research.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 107, + 445, + 175, + 456 + ], + "lines": [ + { + "bbox": [ + 106, + 444, + 176, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 176, + 458 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 504, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 472, + 244, + 483 + ], + "spans": [ + { + "bbox": [ + 116, + 472, + 244, + 483 + ], + "score": 1.0, + "content": "align and translate. In ICLR, 2015.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 490, + 504, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 504 + ], + "score": 1.0, + "content": "Danqi Chen, Jason Bolton, and Christopher D Manning. A thorough examination of the cnn/daily mail reading", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 501, + 246, + 511 + ], + "spans": [ + { + "bbox": [ + 115, + 501, + 246, + 511 + ], + "score": 1.0, + "content": "comprehension task. In ACL, 2016.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 107, + 519, + 504, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "score": 1.0, + "content": "Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural ¨", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 530, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 116, + 530, + 505, + 541 + ], + "score": 1.0, + "content": "machine translation: Encoder-decoder approaches. In Eighth Workshop on Syntax, Semantics and Structure", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 540, + 231, + 550 + ], + "spans": [ + { + "bbox": [ + 116, + 540, + 231, + 550 + ], + "score": 1.0, + "content": "in Statistical Translation, 2014.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 557, + 504, + 588 + ], + "lines": [ + { + "bbox": [ + 106, + 558, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 569 + ], + "score": 1.0, + "content": "Ronan Collobert, Jason Weston, Leon Bottou, Michael Karlen, Koray Kavukcuoglu, and Pavel Kuksa. Natural ´", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 568, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 115, + 568, + 506, + 579 + ], + "score": 1.0, + "content": "language processing (almost) from scratch. Journal of Machine Learning Research, 12(Aug):2493–2537,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 577, + 141, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 141, + 590 + ], + "score": 1.0, + "content": "2011.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 596, + 504, + 617 + ], + "lines": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "Yiming Cui, Zhipeng Chen, Si Wei, Shijin Wang, Ting Liu, and Guoping Hu. Attention-over-attention neural", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 606, + 400, + 618 + ], + "spans": [ + { + "bbox": [ + 115, + 606, + 400, + 618 + ], + "score": 1.0, + "content": "networks for reading comprehension. arXiv preprint arXiv:1607.04423, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 108, + 625, + 504, + 646 + ], + "lines": [ + { + "bbox": [ + 106, + 625, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 505, + 636 + ], + "score": 1.0, + "content": "Bhuwan Dhingra, Hanxiao Liu, Zhilin Yang, William W Cohen, and Ruslan Salakhutdinov. Gated-attention", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 636, + 383, + 647 + ], + "spans": [ + { + "bbox": [ + 115, + 636, + 383, + 647 + ], + "score": 1.0, + "content": "readers for text comprehension. arXiv preprint arXiv:1606.01549, 2016a.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 108, + 654, + 504, + 675 + ], + "lines": [ + { + "bbox": [ + 106, + 653, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 505, + 666 + ], + "score": 1.0, + "content": "Bhuwan Dhingra, Zhong Zhou, Dylan Fitzpatrick, Michael Muehl, and William W Cohen. Tweet2vec:", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 664, + 394, + 675 + ], + "spans": [ + { + "bbox": [ + 116, + 664, + 394, + 675 + ], + "score": 1.0, + "content": "Character-based distributed representations for social media. In ACL, 2016b.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 682, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 682, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 505, + 695 + ], + "score": 1.0, + "content": "Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 693, + 463, + 704 + ], + "spans": [ + { + "bbox": [ + 116, + 693, + 463, + 704 + ], + "score": 1.0, + "content": "and Phil Blunsom. Teaching machines to read and comprehend. In NIPS, pp. 1693–1701, 2015.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 711, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 712, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 712, + 505, + 723 + ], + "score": 1.0, + "content": "Felix Hill, Antoine Bordes, Sumit Chopra, and Jason Weston. The goldilocks principle: Reading children’s", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 721, + 335, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 335, + 732 + ], + "score": 1.0, + "content": "books with explicit memory representations. In ICLR, 2016.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 123, + 504, + 248 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 90, + 504, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 504, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 504, + 102 + ], + "score": 1.0, + "content": "Table 5: Word tokens with highest and lowest gate values. High gate values favor character-level representa-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 100, + 321, + 111 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 321, + 111 + ], + "score": 1.0, + "content": "tions, and low gate values favor word-level representations.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 106, + 123, + 504, + 248 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 123, + 504, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 504, + 248 + ], + "score": 0.969, + "html": "
Gate valuesWord tokens
Lowestor but But These these However however among Among that when When although Although because Because until many Many than though Though this This Since since date where Where have That and And Such such number so which by By how before Before with With between Between even Even if
HighestSweetgum Untersee Jianlong Floresta Chlorella Obersee PhT Doctorin Jumonville WFTS WTSP Boven Pharm Nederrijn Otrar Rhin Magicicada WBKB Tanzler KMBC WPLG Mainau Merwede RMJM Kleitman Scheur Bodensee Kromme Horenbout Vorderrhein Chlamydomonas Scantlebury Qingshui Funchess
", + "type": "table", + "image_path": "3d88f65fd35053aed1c57fb91abce4362deea2ce3a4b72dcac738b31312c5589.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 106, + 123, + 504, + 164.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 106, + 164.66666666666666, + 504, + 206.33333333333331 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 206.33333333333331, + 504, + 247.99999999999997 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "title", + "bbox": [ + 108, + 271, + 201, + 284 + ], + "lines": [ + { + "bbox": [ + 104, + 269, + 203, + 287 + ], + "spans": [ + { + "bbox": [ + 104, + 269, + 203, + 287 + ], + "score": 1.0, + "content": "5 CONCLUSIONS", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 296, + 505, + 374 + ], + "lines": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "We present a fine-grained gating mechanism that dynamically combines word-level and character-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 308, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 320 + ], + "score": 1.0, + "content": "level representations based on word properties. Experiments on the Twitter tag prediction dataset", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "show that fine-grained gating substantially outperforms scalar gating and concatenation. Our method", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 330, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 342 + ], + "score": 1.0, + "content": "also improves the performance on reading comprehension and achieves new state-of-the-art results", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 354 + ], + "score": 1.0, + "content": "on CBT and WDW. In our future work, we plan to to apply the fine-grained gating mechanism for", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 350, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 506, + 365 + ], + "score": 1.0, + "content": "combining other levels of representations, such as phrases and sentences. It will also be intriguing", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 362, + 493, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 493, + 375 + ], + "score": 1.0, + "content": "to integrate NER and POS networks and learn the token representation in an end-to-end manner.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 297, + 506, + 375 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 387, + 200, + 397 + ], + "lines": [ + { + "bbox": [ + 107, + 388, + 200, + 398 + ], + "spans": [ + { + "bbox": [ + 107, + 388, + 200, + 398 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 503, + 428 + ], + "lines": [ + { + "bbox": [ + 105, + 404, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 505, + 419 + ], + "score": 1.0, + "content": "This work was funded by NVIDIA, the Office of Naval Research Scene Understanding grant", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 416, + 378, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 378, + 429 + ], + "score": 1.0, + "content": "N000141310721, the NSF grant IIS1250956, and Google Research.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 404, + 505, + 429 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 445, + 175, + 456 + ], + "lines": [ + { + "bbox": [ + 106, + 444, + 176, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 176, + 458 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 504, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 472, + 244, + 483 + ], + "spans": [ + { + "bbox": [ + 116, + 472, + 244, + 483 + ], + "score": 1.0, + "content": "align and translate. In ICLR, 2015.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 461, + 505, + 483 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 490, + 504, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 504 + ], + "score": 1.0, + "content": "Danqi Chen, Jason Bolton, and Christopher D Manning. A thorough examination of the cnn/daily mail reading", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 501, + 246, + 511 + ], + "spans": [ + { + "bbox": [ + 115, + 501, + 246, + 511 + ], + "score": 1.0, + "content": "comprehension task. In ACL, 2016.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 489, + 505, + 511 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 519, + 504, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "score": 1.0, + "content": "Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural ¨", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 530, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 116, + 530, + 505, + 541 + ], + "score": 1.0, + "content": "machine translation: Encoder-decoder approaches. In Eighth Workshop on Syntax, Semantics and Structure", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 540, + 231, + 550 + ], + "spans": [ + { + "bbox": [ + 116, + 540, + 231, + 550 + ], + "score": 1.0, + "content": "in Statistical Translation, 2014.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 519, + 505, + 550 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 557, + 504, + 588 + ], + "lines": [ + { + "bbox": [ + 106, + 558, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 569 + ], + "score": 1.0, + "content": "Ronan Collobert, Jason Weston, Leon Bottou, Michael Karlen, Koray Kavukcuoglu, and Pavel Kuksa. Natural ´", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 568, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 115, + 568, + 506, + 579 + ], + "score": 1.0, + "content": "language processing (almost) from scratch. Journal of Machine Learning Research, 12(Aug):2493–2537,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 577, + 141, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 141, + 590 + ], + "score": 1.0, + "content": "2011.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 558, + 506, + 590 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 596, + 504, + 617 + ], + "lines": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "Yiming Cui, Zhipeng Chen, Si Wei, Shijin Wang, Ting Liu, and Guoping Hu. Attention-over-attention neural", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 606, + 400, + 618 + ], + "spans": [ + { + "bbox": [ + 115, + 606, + 400, + 618 + ], + "score": 1.0, + "content": "networks for reading comprehension. arXiv preprint arXiv:1607.04423, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 106, + 596, + 505, + 618 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 625, + 504, + 646 + ], + "lines": [ + { + "bbox": [ + 106, + 625, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 505, + 636 + ], + "score": 1.0, + "content": "Bhuwan Dhingra, Hanxiao Liu, Zhilin Yang, William W Cohen, and Ruslan Salakhutdinov. Gated-attention", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 636, + 383, + 647 + ], + "spans": [ + { + "bbox": [ + 115, + 636, + 383, + 647 + ], + "score": 1.0, + "content": "readers for text comprehension. arXiv preprint arXiv:1606.01549, 2016a.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 625, + 505, + 647 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 654, + 504, + 675 + ], + "lines": [ + { + "bbox": [ + 106, + 653, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 505, + 666 + ], + "score": 1.0, + "content": "Bhuwan Dhingra, Zhong Zhou, Dylan Fitzpatrick, Michael Muehl, and William W Cohen. Tweet2vec:", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 664, + 394, + 675 + ], + "spans": [ + { + "bbox": [ + 116, + 664, + 394, + 675 + ], + "score": 1.0, + "content": "Character-based distributed representations for social media. In ACL, 2016b.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 106, + 653, + 505, + 675 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 682, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 682, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 505, + 695 + ], + "score": 1.0, + "content": "Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 693, + 463, + 704 + ], + "spans": [ + { + "bbox": [ + 116, + 693, + 463, + 704 + ], + "score": 1.0, + "content": "and Phil Blunsom. Teaching machines to read and comprehend. In NIPS, pp. 1693–1701, 2015.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 682, + 505, + 704 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 711, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 712, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 712, + 505, + 723 + ], + "score": 1.0, + "content": "Felix Hill, Antoine Bordes, Sumit Chopra, and Jason Weston. The goldilocks principle: Reading children’s", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 721, + 335, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 335, + 732 + ], + "score": 1.0, + "content": "books with explicit memory representations. In ICLR, 2016.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 106, + 712, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 104 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):1735–1780,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 141, + 104 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 141, + 104 + ], + "score": 1.0, + "content": "1997.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 111, + 504, + 132 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "score": 1.0, + "content": "Rudolf Kadlec, Martin Schmid, Ondrej Bajgar, and Jan Kleindienst. Text understanding with the attention sum", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 121, + 229, + 132 + ], + "spans": [ + { + "bbox": [ + 115, + 121, + 229, + 132 + ], + "score": 1.0, + "content": "reader network. In ACL, 2016.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 138, + 504, + 160 + ], + "lines": [ + { + "bbox": [ + 106, + 138, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 151 + ], + "score": 1.0, + "content": "Yoon Kim, Yacine Jernite, David Sontag, and Alexander M Rush. Character-aware neural language models. In", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 149, + 163, + 160 + ], + "spans": [ + { + "bbox": [ + 115, + 149, + 163, + 160 + ], + "score": 1.0, + "content": "AAAI, 2016.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 166, + 504, + 198 + ], + "lines": [ + { + "bbox": [ + 107, + 167, + 504, + 178 + ], + "spans": [ + { + "bbox": [ + 107, + 167, + 504, + 178 + ], + "score": 1.0, + "content": "Wang Ling, Tiago Lu´ıs, Lu´ıs Marujo, Ramon Fernandez Astudillo, Silvio Amir, Chris Dyer, Alan W Black, ´", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 177, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 116, + 177, + 505, + 188 + ], + "score": 1.0, + "content": "and Isabel Trancoso. Finding function in form: Compositional character models for open vocabulary word", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 188, + 239, + 197 + ], + "spans": [ + { + "bbox": [ + 116, + 188, + 239, + 197 + ], + "score": 1.0, + "content": "representation. In EMNLP, 2015.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 504, + 226 + ], + "lines": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "Minh-Thang Luong and Christopher D Manning. Achieving open vocabulary neural machine translation with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 214, + 283, + 226 + ], + "spans": [ + { + "bbox": [ + 115, + 214, + 283, + 226 + ], + "score": 1.0, + "content": "hybrid word-character models. In ACL, 2016.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 105, + 232, + 502, + 244 + ], + "lines": [ + { + "bbox": [ + 106, + 231, + 504, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 504, + 245 + ], + "score": 1.0, + "content": "Yasumasa Miyamoto and Kyunghyun Cho. Gated word-character recurrent language model. In EMNLP, 2016.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 498, + 262 + ], + "lines": [ + { + "bbox": [ + 106, + 249, + 493, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 493, + 263 + ], + "score": 1.0, + "content": "Tsendsuren Munkhdalai and Hong Yu. Neural semantic encoders. arXiv preprint arXiv:1607.04315, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 108, + 268, + 504, + 289 + ], + "lines": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "Takeshi Onishi, Hai Wang, Mohit Bansal, Kevin Gimpel, and David McAllester. Who did what: A large-scale", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 278, + 374, + 290 + ], + "spans": [ + { + "bbox": [ + 115, + 278, + 374, + 290 + ], + "score": 1.0, + "content": "person-centered cloze dataset. arXiv preprint arXiv:1608.05457, 2016.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 295, + 504, + 317 + ], + "lines": [ + { + "bbox": [ + 105, + 296, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 386, + 308 + ], + "score": 1.0, + "content": "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad:", + "type": "text" + }, + { + "bbox": [ + 387, + 297, + 423, + 306 + ], + "score": 0.73, + "content": "^ { 1 0 0 , 0 0 0 + }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 296, + 505, + 308 + ], + "score": 1.0, + "content": "questions for machine", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 307, + 268, + 317 + ], + "spans": [ + { + "bbox": [ + 116, + 307, + 268, + 317 + ], + "score": 1.0, + "content": "comprehension of text. In EMNLP, 2016.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 324, + 503, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 323, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 337 + ], + "score": 1.0, + "content": "Marek Rei, Gamal KO Crichton, and Sampo Pyysalo. Attending to characters in neural sequence labeling", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 334, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 116, + 334, + 294, + 345 + ], + "score": 1.0, + "content": "models. arXiv preprint arXiv:1611.04361, 2016.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 351, + 504, + 373 + ], + "lines": [ + { + "bbox": [ + 106, + 351, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 505, + 364 + ], + "score": 1.0, + "content": "Alessandro Sordoni, Phillip Bachman, and Yoshua Bengio. Iterative alternating neural attention for machine", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 363, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 115, + 363, + 294, + 372 + ], + "score": 1.0, + "content": "reading. arXiv preprint arXiv:1606.02245, 2016.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 105, + 379, + 503, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "Adam Trischler, Zheng Ye, Xingdi Yuan, and Kaheer Suleman. Natural language comprehension with the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 390, + 222, + 401 + ], + "spans": [ + { + "bbox": [ + 115, + 390, + 222, + 401 + ], + "score": 1.0, + "content": "epireader. In EMNLP, 2016.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 108, + 407, + 477, + 419 + ], + "lines": [ + { + "bbox": [ + 106, + 407, + 476, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 476, + 420 + ], + "score": 1.0, + "content": "Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In NIPS, pp. 2692–2700, 2015.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 425, + 504, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 424, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 438 + ], + "score": 1.0, + "content": "Shuohang Wang and Jing Jiang. Machine comprehension using match-lstm and answer pointer. arXiv preprint", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 436, + 209, + 447 + ], + "spans": [ + { + "bbox": [ + 115, + 436, + 209, + 447 + ], + "score": 1.0, + "content": "arXiv:1608.07905, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 503, + 474 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "score": 1.0, + "content": "Yuhuai Wu, Saizheng Zhang, Ying Zhang, Yoshua Bengio, and Ruslan Salakhutdinov. On multiplicative inte-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 464, + 316, + 475 + ], + "spans": [ + { + "bbox": [ + 116, + 464, + 316, + 475 + ], + "score": 1.0, + "content": "gration with recurrent neural networks. In NIPS, 2016.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 481, + 504, + 503 + ], + "lines": [ + { + "bbox": [ + 106, + 482, + 504, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 504, + 493 + ], + "score": 1.0, + "content": "Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. Learning multi-relational semantics", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 492, + 436, + 503 + ], + "spans": [ + { + "bbox": [ + 115, + 492, + 436, + 503 + ], + "score": 1.0, + "content": "using neural-embedding models. In NIPS 2014 workshop on Learning Semantics, 2014.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 509, + 504, + 531 + ], + "lines": [ + { + "bbox": [ + 105, + 507, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 523 + ], + "score": 1.0, + "content": "Zhilin Yang, Ruslan Salakhutdinov, and William Cohen. Multi-task cross-lingual sequence tagging from", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 520, + 297, + 531 + ], + "spans": [ + { + "bbox": [ + 116, + 520, + 297, + 531 + ], + "score": 1.0, + "content": "scratch. arXiv preprint arXiv:1603.06270, 2016a.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 537, + 504, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "Zhilin Yang, Ye Yuan, Yuexin Wu, Ruslan Salakhutdinov, and William W Cohen. Review networks for caption", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 548, + 220, + 558 + ], + "spans": [ + { + "bbox": [ + 115, + 548, + 220, + 558 + ], + "score": 1.0, + "content": "generation. In NIPS, 2016b.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 565, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 107, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 107, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "Yang Yu, Wei Zhang, Kazi Hasan, Mo Yu, Bing Xiang, and Bowen Zhou. End-to-end answer chunk extraction", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 575, + 409, + 587 + ], + "spans": [ + { + "bbox": [ + 116, + 575, + 409, + 587 + ], + "score": 1.0, + "content": "and ranking for reading comprehension. arXiv preprint arXiv:1610.09996, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 104 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):1735–1780,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 141, + 104 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 141, + 104 + ], + "score": 1.0, + "content": "1997.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 505, + 104 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 111, + 504, + 132 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "score": 1.0, + "content": "Rudolf Kadlec, Martin Schmid, Ondrej Bajgar, and Jan Kleindienst. Text understanding with the attention sum", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 121, + 229, + 132 + ], + "spans": [ + { + "bbox": [ + 115, + 121, + 229, + 132 + ], + "score": 1.0, + "content": "reader network. In ACL, 2016.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 109, + 505, + 132 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 138, + 504, + 160 + ], + "lines": [ + { + "bbox": [ + 106, + 138, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 151 + ], + "score": 1.0, + "content": "Yoon Kim, Yacine Jernite, David Sontag, and Alexander M Rush. Character-aware neural language models. In", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 149, + 163, + 160 + ], + "spans": [ + { + "bbox": [ + 115, + 149, + 163, + 160 + ], + "score": 1.0, + "content": "AAAI, 2016.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 106, + 138, + 505, + 160 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 166, + 504, + 198 + ], + "lines": [ + { + "bbox": [ + 107, + 167, + 504, + 178 + ], + "spans": [ + { + "bbox": [ + 107, + 167, + 504, + 178 + ], + "score": 1.0, + "content": "Wang Ling, Tiago Lu´ıs, Lu´ıs Marujo, Ramon Fernandez Astudillo, Silvio Amir, Chris Dyer, Alan W Black, ´", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 177, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 116, + 177, + 505, + 188 + ], + "score": 1.0, + "content": "and Isabel Trancoso. Finding function in form: Compositional character models for open vocabulary word", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 188, + 239, + 197 + ], + "spans": [ + { + "bbox": [ + 116, + 188, + 239, + 197 + ], + "score": 1.0, + "content": "representation. In EMNLP, 2015.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 107, + 167, + 505, + 197 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 504, + 226 + ], + "lines": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "Minh-Thang Luong and Christopher D Manning. Achieving open vocabulary neural machine translation with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 214, + 283, + 226 + ], + "spans": [ + { + "bbox": [ + 115, + 214, + 283, + 226 + ], + "score": 1.0, + "content": "hybrid word-character models. In ACL, 2016.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 204, + 505, + 226 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 232, + 502, + 244 + ], + "lines": [ + { + "bbox": [ + 106, + 231, + 504, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 504, + 245 + ], + "score": 1.0, + "content": "Yasumasa Miyamoto and Kyunghyun Cho. Gated word-character recurrent language model. In EMNLP, 2016.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 231, + 504, + 245 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 498, + 262 + ], + "lines": [ + { + "bbox": [ + 106, + 249, + 493, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 493, + 263 + ], + "score": 1.0, + "content": "Tsendsuren Munkhdalai and Hong Yu. Neural semantic encoders. arXiv preprint arXiv:1607.04315, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12, + "bbox_fs": [ + 106, + 249, + 493, + 263 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 268, + 504, + 289 + ], + "lines": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "Takeshi Onishi, Hai Wang, Mohit Bansal, Kevin Gimpel, and David McAllester. Who did what: A large-scale", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 278, + 374, + 290 + ], + "spans": [ + { + "bbox": [ + 115, + 278, + 374, + 290 + ], + "score": 1.0, + "content": "person-centered cloze dataset. arXiv preprint arXiv:1608.05457, 2016.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 267, + 505, + 290 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 295, + 504, + 317 + ], + "lines": [ + { + "bbox": [ + 105, + 296, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 386, + 308 + ], + "score": 1.0, + "content": "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad:", + "type": "text" + }, + { + "bbox": [ + 387, + 297, + 423, + 306 + ], + "score": 0.73, + "content": "^ { 1 0 0 , 0 0 0 + }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 296, + 505, + 308 + ], + "score": 1.0, + "content": "questions for machine", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 307, + 268, + 317 + ], + "spans": [ + { + "bbox": [ + 116, + 307, + 268, + 317 + ], + "score": 1.0, + "content": "comprehension of text. In EMNLP, 2016.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 296, + 505, + 317 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 324, + 503, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 323, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 337 + ], + "score": 1.0, + "content": "Marek Rei, Gamal KO Crichton, and Sampo Pyysalo. Attending to characters in neural sequence labeling", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 334, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 116, + 334, + 294, + 345 + ], + "score": 1.0, + "content": "models. arXiv preprint arXiv:1611.04361, 2016.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 323, + 505, + 345 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 351, + 504, + 373 + ], + "lines": [ + { + "bbox": [ + 106, + 351, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 505, + 364 + ], + "score": 1.0, + "content": "Alessandro Sordoni, Phillip Bachman, and Yoshua Bengio. Iterative alternating neural attention for machine", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 363, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 115, + 363, + 294, + 372 + ], + "score": 1.0, + "content": "reading. arXiv preprint arXiv:1606.02245, 2016.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 351, + 505, + 372 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 379, + 503, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "Adam Trischler, Zheng Ye, Xingdi Yuan, and Kaheer Suleman. Natural language comprehension with the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 390, + 222, + 401 + ], + "spans": [ + { + "bbox": [ + 115, + 390, + 222, + 401 + ], + "score": 1.0, + "content": "epireader. In EMNLP, 2016.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 379, + 505, + 401 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 407, + 477, + 419 + ], + "lines": [ + { + "bbox": [ + 106, + 407, + 476, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 476, + 420 + ], + "score": 1.0, + "content": "Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In NIPS, pp. 2692–2700, 2015.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 106, + 407, + 476, + 420 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 425, + 504, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 424, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 438 + ], + "score": 1.0, + "content": "Shuohang Wang and Jing Jiang. Machine comprehension using match-lstm and answer pointer. arXiv preprint", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 436, + 209, + 447 + ], + "spans": [ + { + "bbox": [ + 115, + 436, + 209, + 447 + ], + "score": 1.0, + "content": "arXiv:1608.07905, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 424, + 505, + 447 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 503, + 474 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "score": 1.0, + "content": "Yuhuai Wu, Saizheng Zhang, Ying Zhang, Yoshua Bengio, and Ruslan Salakhutdinov. On multiplicative inte-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 464, + 316, + 475 + ], + "spans": [ + { + "bbox": [ + 116, + 464, + 316, + 475 + ], + "score": 1.0, + "content": "gration with recurrent neural networks. In NIPS, 2016.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 106, + 453, + 504, + 475 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 481, + 504, + 503 + ], + "lines": [ + { + "bbox": [ + 106, + 482, + 504, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 504, + 493 + ], + "score": 1.0, + "content": "Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. Learning multi-relational semantics", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 492, + 436, + 503 + ], + "spans": [ + { + "bbox": [ + 115, + 492, + 436, + 503 + ], + "score": 1.0, + "content": "using neural-embedding models. In NIPS 2014 workshop on Learning Semantics, 2014.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 482, + 504, + 503 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 509, + 504, + 531 + ], + "lines": [ + { + "bbox": [ + 105, + 507, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 523 + ], + "score": 1.0, + "content": "Zhilin Yang, Ruslan Salakhutdinov, and William Cohen. Multi-task cross-lingual sequence tagging from", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 520, + 297, + 531 + ], + "spans": [ + { + "bbox": [ + 116, + 520, + 297, + 531 + ], + "score": 1.0, + "content": "scratch. arXiv preprint arXiv:1603.06270, 2016a.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 507, + 505, + 531 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 537, + 504, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "Zhilin Yang, Ye Yuan, Yuexin Wu, Ruslan Salakhutdinov, and William W Cohen. Review networks for caption", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 548, + 220, + 558 + ], + "spans": [ + { + "bbox": [ + 115, + 548, + 220, + 558 + ], + "score": 1.0, + "content": "generation. In NIPS, 2016b.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 536, + 505, + 558 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 565, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 107, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 107, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "Yang Yu, Wei Zhang, Kazi Hasan, Mo Yu, Bing Xiang, and Bowen Zhou. End-to-end answer chunk extraction", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 575, + 409, + 587 + ], + "spans": [ + { + "bbox": [ + 116, + 575, + 409, + 587 + ], + "score": 1.0, + "content": "and ranking for reading comprehension. arXiv preprint arXiv:1610.09996, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 107, + 565, + 505, + 587 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/B1hdzd5lg/B1hdzd5lg_model.json b/parse/train/B1hdzd5lg/B1hdzd5lg_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2b3c7a3a54759bf371d3aebe19efc3a09a9ffd49 --- /dev/null +++ b/parse/train/B1hdzd5lg/B1hdzd5lg_model.json @@ -0,0 +1,13454 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1068, + 1404, + 1068, + 1404, + 1465, + 298, + 1465 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 398, + 643, + 1302, + 643, + 1302, + 949, + 398, + 949 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1481, + 1403, + 1481, + 1403, + 1908, + 298, + 1908 + ], + "score": 0.982 + }, + { + "category_id": 0, + "poly": [ + 301, + 219, + 1403, + 219, + 1403, + 324, + 301, + 324 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 301, + 1925, + 1400, + 1925, + 1400, + 1986, + 301, + 1986 + ], + "score": 0.93 + }, + { + "category_id": 2, + "poly": [ + 333, + 2006, + 1118, + 2006, + 1118, + 2034, + 333, + 2034 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 302, + 1001, + 573, + 1001, + 573, + 1036, + 302, + 1036 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.884 + }, + { + "category_id": 0, + "poly": [ + 773, + 580, + 926, + 580, + 926, + 613, + 773, + 613 + ], + "score": 0.877 + }, + { + "category_id": 1, + "poly": [ + 314, + 412, + 910, + 412, + 910, + 500, + 314, + 500 + ], + "score": 0.715 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 857, + 2089, + 857, + 2112, + 841, + 2112 + ], + "score": 0.687 + }, + { + "category_id": 1, + "poly": [ + 318, + 375, + 1421, + 375, + 1421, + 409, + 318, + 409 + ], + "score": 0.189 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 216.0, + 1405.0, + 216.0, + 1405.0, + 273.0, + 296.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 275.0, + 968.0, + 275.0, + 968.0, + 328.0, + 295.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 1998.0, + 1120.0, + 1998.0, + 1120.0, + 2041.0, + 333.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 998.0, + 579.0, + 998.0, + 579.0, + 1045.0, + 294.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 579.0, + 932.0, + 579.0, + 932.0, + 616.0, + 769.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2088.0, + 861.0, + 2088.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1070.0, + 1404.0, + 1070.0, + 1404.0, + 1102.0, + 296.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1100.0, + 1405.0, + 1100.0, + 1405.0, + 1133.0, + 293.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1129.0, + 1406.0, + 1129.0, + 1406.0, + 1166.0, + 293.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1162.0, + 1404.0, + 1162.0, + 1404.0, + 1194.0, + 296.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1189.0, + 1406.0, + 1189.0, + 1406.0, + 1228.0, + 293.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1222.0, + 1402.0, + 1222.0, + 1402.0, + 1255.0, + 296.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1250.0, + 1405.0, + 1250.0, + 1405.0, + 1287.0, + 293.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1281.0, + 1406.0, + 1281.0, + 1406.0, + 1320.0, + 292.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1313.0, + 1404.0, + 1313.0, + 1404.0, + 1349.0, + 294.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1343.0, + 1404.0, + 1343.0, + 1404.0, + 1376.0, + 293.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1374.0, + 1404.0, + 1374.0, + 1404.0, + 1410.0, + 294.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1405.0, + 1404.0, + 1405.0, + 1404.0, + 1437.0, + 296.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1435.0, + 1281.0, + 1435.0, + 1281.0, + 1471.0, + 294.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 643.0, + 1305.0, + 643.0, + 1305.0, + 679.0, + 393.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 674.0, + 1306.0, + 674.0, + 1306.0, + 710.0, + 394.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 706.0, + 1304.0, + 706.0, + 1304.0, + 741.0, + 392.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 737.0, + 1305.0, + 737.0, + 1305.0, + 770.0, + 394.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 765.0, + 1305.0, + 765.0, + 1305.0, + 801.0, + 394.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 797.0, + 1305.0, + 797.0, + 1305.0, + 833.0, + 393.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 827.0, + 1305.0, + 827.0, + 1305.0, + 866.0, + 392.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 857.0, + 1305.0, + 857.0, + 1305.0, + 890.0, + 394.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 888.0, + 1306.0, + 888.0, + 1306.0, + 924.0, + 394.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 919.0, + 1079.0, + 919.0, + 1079.0, + 951.0, + 394.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1480.0, + 1405.0, + 1480.0, + 1405.0, + 1516.0, + 294.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1513.0, + 1405.0, + 1513.0, + 1405.0, + 1545.0, + 296.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1541.0, + 1407.0, + 1541.0, + 1407.0, + 1579.0, + 292.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1572.0, + 1403.0, + 1572.0, + 1403.0, + 1608.0, + 293.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1605.0, + 1405.0, + 1605.0, + 1405.0, + 1640.0, + 293.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1634.0, + 1405.0, + 1634.0, + 1405.0, + 1669.0, + 294.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1663.0, + 1405.0, + 1663.0, + 1405.0, + 1697.0, + 292.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1696.0, + 1404.0, + 1696.0, + 1404.0, + 1727.0, + 296.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1726.0, + 1404.0, + 1726.0, + 1404.0, + 1758.0, + 293.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1754.0, + 1408.0, + 1754.0, + 1408.0, + 1793.0, + 292.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1787.0, + 1404.0, + 1787.0, + 1404.0, + 1819.0, + 296.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1819.0, + 1404.0, + 1819.0, + 1404.0, + 1850.0, + 293.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1845.0, + 1405.0, + 1845.0, + 1405.0, + 1883.0, + 293.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1879.0, + 1055.0, + 1879.0, + 1055.0, + 1911.0, + 296.0, + 1911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1924.0, + 1404.0, + 1924.0, + 1404.0, + 1959.0, + 295.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1954.0, + 1404.0, + 1954.0, + 1404.0, + 1989.0, + 296.0, + 1989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 404.0, + 638.0, + 404.0, + 638.0, + 440.0, + 312.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 435.0, + 629.0, + 435.0, + 629.0, + 472.0, + 312.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 466.0, + 913.0, + 466.0, + 913.0, + 505.0, + 313.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 372.0, + 1427.0, + 372.0, + 1427.0, + 416.0, + 312.0, + 416.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1252, + 1404, + 1252, + 1404, + 1557, + 298, + 1557 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 299, + 460, + 1403, + 460, + 1403, + 795, + 299, + 795 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1573, + 1403, + 1573, + 1403, + 1757, + 298, + 1757 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1402, + 229, + 1402, + 445, + 298, + 445 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 300, + 1010, + 1402, + 1010, + 1402, + 1133, + 300, + 1133 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 811, + 1403, + 811, + 1403, + 994, + 298, + 994 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1772, + 1402, + 1772, + 1402, + 1957, + 299, + 1957 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1972, + 1400, + 1972, + 1400, + 2034, + 298, + 2034 + ], + "score": 0.943 + }, + { + "category_id": 0, + "poly": [ + 301, + 1187, + 581, + 1187, + 581, + 1222, + 301, + 1222 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 814, + 76, + 814, + 104, + 300, + 104 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.7 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.15 + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1184.0, + 586.0, + 1184.0, + 586.0, + 1228.0, + 292.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1250.0, + 1405.0, + 1250.0, + 1405.0, + 1286.0, + 294.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1283.0, + 1406.0, + 1283.0, + 1406.0, + 1318.0, + 294.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1311.0, + 1406.0, + 1311.0, + 1406.0, + 1349.0, + 293.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1344.0, + 1404.0, + 1344.0, + 1404.0, + 1378.0, + 292.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1374.0, + 1405.0, + 1374.0, + 1405.0, + 1410.0, + 294.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1406.0, + 1406.0, + 1406.0, + 1406.0, + 1441.0, + 294.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1435.0, + 1405.0, + 1435.0, + 1405.0, + 1470.0, + 294.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1465.0, + 1405.0, + 1465.0, + 1405.0, + 1502.0, + 293.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1493.0, + 1405.0, + 1493.0, + 1405.0, + 1534.0, + 292.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1524.0, + 400.0, + 1524.0, + 400.0, + 1562.0, + 293.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 457.0, + 1405.0, + 457.0, + 1405.0, + 497.0, + 294.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 490.0, + 1405.0, + 490.0, + 1405.0, + 525.0, + 294.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 522.0, + 1405.0, + 522.0, + 1405.0, + 556.0, + 294.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 551.0, + 1405.0, + 551.0, + 1405.0, + 585.0, + 294.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 583.0, + 1405.0, + 583.0, + 1405.0, + 616.0, + 293.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 613.0, + 1405.0, + 613.0, + 1405.0, + 647.0, + 294.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 642.0, + 1404.0, + 642.0, + 1404.0, + 677.0, + 294.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 673.0, + 1405.0, + 673.0, + 1405.0, + 708.0, + 293.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 704.0, + 1405.0, + 704.0, + 1405.0, + 739.0, + 294.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 734.0, + 1407.0, + 734.0, + 1407.0, + 772.0, + 293.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 763.0, + 599.0, + 763.0, + 599.0, + 801.0, + 293.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1571.0, + 1405.0, + 1571.0, + 1405.0, + 1608.0, + 294.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1604.0, + 1404.0, + 1604.0, + 1404.0, + 1640.0, + 293.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1635.0, + 1405.0, + 1635.0, + 1405.0, + 1670.0, + 294.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1665.0, + 1405.0, + 1665.0, + 1405.0, + 1701.0, + 294.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1695.0, + 1405.0, + 1695.0, + 1405.0, + 1731.0, + 294.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1730.0, + 481.0, + 1730.0, + 481.0, + 1763.0, + 294.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 230.0, + 1404.0, + 230.0, + 1404.0, + 264.0, + 293.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 262.0, + 1406.0, + 262.0, + 1406.0, + 296.0, + 293.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 1403.0, + 291.0, + 1403.0, + 326.0, + 294.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 322.0, + 1406.0, + 322.0, + 1406.0, + 357.0, + 293.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 351.0, + 1407.0, + 351.0, + 1407.0, + 389.0, + 293.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 381.0, + 1406.0, + 381.0, + 1406.0, + 417.0, + 293.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 414.0, + 1230.0, + 414.0, + 1230.0, + 452.0, + 293.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1009.0, + 1404.0, + 1009.0, + 1404.0, + 1045.0, + 295.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1042.0, + 1404.0, + 1042.0, + 1404.0, + 1075.0, + 295.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1071.0, + 1404.0, + 1071.0, + 1404.0, + 1106.0, + 294.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1103.0, + 939.0, + 1103.0, + 939.0, + 1135.0, + 296.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 810.0, + 1405.0, + 810.0, + 1405.0, + 846.0, + 296.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 842.0, + 1405.0, + 842.0, + 1405.0, + 877.0, + 294.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 872.0, + 1405.0, + 872.0, + 1405.0, + 908.0, + 294.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 901.0, + 1405.0, + 901.0, + 1405.0, + 939.0, + 293.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 933.0, + 1407.0, + 933.0, + 1407.0, + 971.0, + 292.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 965.0, + 562.0, + 965.0, + 562.0, + 997.0, + 293.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1774.0, + 1404.0, + 1774.0, + 1404.0, + 1806.0, + 297.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1803.0, + 1403.0, + 1803.0, + 1403.0, + 1835.0, + 295.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1833.0, + 1407.0, + 1833.0, + 1407.0, + 1871.0, + 293.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1865.0, + 1406.0, + 1865.0, + 1406.0, + 1899.0, + 294.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1894.0, + 1407.0, + 1894.0, + 1407.0, + 1928.0, + 293.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1927.0, + 778.0, + 1927.0, + 778.0, + 1959.0, + 295.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1737, + 1404, + 1737, + 1404, + 1953, + 298, + 1953 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 694, + 1403, + 694, + 1403, + 879, + 298, + 879 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 893, + 1404, + 893, + 1404, + 1109, + 297, + 1109 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1123, + 1403, + 1123, + 1403, + 1285, + 298, + 1285 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 472, + 1404, + 472, + 1404, + 596, + 298, + 596 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 1403, + 228, + 1403, + 353, + 298, + 353 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1598, + 1404, + 1598, + 1404, + 1722, + 298, + 1722 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 301, + 1298, + 1400, + 1298, + 1400, + 1392, + 301, + 1392 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 303, + 1491, + 1399, + 1491, + 1399, + 1584, + 303, + 1584 + ], + "score": 0.963 + }, + { + "category_id": 8, + "poly": [ + 736, + 2000, + 961, + 2000, + 961, + 2039, + 736, + 2039 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 299, + 1967, + 664, + 1967, + 664, + 1999, + 299, + 1999 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 302, + 400, + 690, + 400, + 690, + 436, + 302, + 436 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 104, + 300, + 104 + ], + "score": 0.896 + }, + { + "category_id": 0, + "poly": [ + 299, + 1432, + 896, + 1432, + 896, + 1464, + 299, + 1464 + ], + "score": 0.889 + }, + { + "category_id": 0, + "poly": [ + 304, + 636, + 808, + 636, + 808, + 668, + 304, + 668 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.61 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2111, + 841, + 2111 + ], + "score": 0.455 + }, + { + "category_id": 13, + "poly": [ + 401, + 985, + 596, + 985, + 596, + 1020, + 401, + 1020 + ], + "score": 0.93, + "latex": "\\mathbf { h } _ { i } ^ { \\bar { 0 } } = f ( \\mathbf { w } _ { i } , \\mathbf { C } _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 542, + 1523, + 733, + 1523, + 733, + 1556, + 542, + 1556 + ], + "score": 0.93, + "latex": "\\mathbf { h } _ { i } = f ( \\mathbf { w } _ { i } , \\mathbf { C } _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 1212, + 1122, + 1282, + 1122, + 1282, + 1162, + 1212, + 1162 + ], + "score": 0.93, + "latex": "\\mathbf { H } _ { p } ^ { k - 1 }" + }, + { + "category_id": 13, + "poly": [ + 497, + 1220, + 699, + 1220, + 699, + 1259, + 497, + 1259 + ], + "score": 0.93, + "latex": "\\mathbf { H } _ { p } ^ { k } = r ( \\mathbf { P } ^ { k } , \\mathbf { Q } ^ { k } )" + }, + { + "category_id": 13, + "poly": [ + 549, + 757, + 647, + 757, + 647, + 790, + 549, + 790 + ], + "score": 0.93, + "latex": "( \\mathbf { w } _ { i } , \\mathbf { C } _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 924, + 695, + 1186, + 695, + 1186, + 729, + 924, + 729 + ], + "score": 0.91, + "latex": "P = ( p _ { 1 } , p _ { 2 } , \\cdot \\cdot \\cdot , p _ { M } )" + }, + { + "category_id": 13, + "poly": [ + 818, + 1161, + 894, + 1161, + 894, + 1191, + 818, + 1191 + ], + "score": 0.91, + "latex": "M \\times d" + }, + { + "category_id": 13, + "poly": [ + 521, + 1158, + 560, + 1158, + 560, + 1190, + 521, + 1190 + ], + "score": 0.91, + "latex": "\\mathbf { P } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 298, + 726, + 478, + 726, + 478, + 760, + 298, + 760 + ], + "score": 0.91, + "latex": "\\left( q _ { 1 } , q _ { 2 } , \\cdots , q _ { N } \\right)" + }, + { + "category_id": 14, + "poly": [ + 737, + 1999, + 962, + 1999, + 962, + 2039, + 737, + 2039 + ], + "score": 0.91, + "latex": "\\mathbf { g } = \\sigma ( \\mathbf { W } _ { g } \\mathbf { v } + \\mathbf { b } _ { g } )" + }, + { + "category_id": 13, + "poly": [ + 1022, + 1159, + 1062, + 1159, + 1062, + 1194, + 1022, + 1194 + ], + "score": 0.91, + "latex": "\\mathbf { Q } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 610, + 1158, + 649, + 1158, + 649, + 1193, + 610, + 1193 + ], + "score": 0.9, + "latex": "\\mathbf { Q } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 1108, + 1161, + 1179, + 1161, + 1179, + 1191, + 1108, + 1191 + ], + "score": 0.9, + "latex": "N \\times d" + }, + { + "category_id": 13, + "poly": [ + 297, + 1015, + 339, + 1015, + 339, + 1054, + 297, + 1054 + ], + "score": 0.9, + "latex": "\\mathbf { H } _ { p } ^ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 366, + 1329, + 408, + 1329, + 408, + 1368, + 366, + 1368 + ], + "score": 0.89, + "latex": "\\bar { \\mathbf { H } } _ { p } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 733, + 1158, + 772, + 1158, + 772, + 1190, + 733, + 1190 + ], + "score": 0.89, + "latex": "\\mathbf { P } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 533, + 1922, + 564, + 1922, + 564, + 1952, + 533, + 1952 + ], + "score": 0.89, + "latex": "d _ { v }" + }, + { + "category_id": 13, + "poly": [ + 927, + 1493, + 962, + 1493, + 962, + 1523, + 927, + 1523 + ], + "score": 0.88, + "latex": "\\mathbf { C } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1332, + 1124, + 1373, + 1124, + 1373, + 1159, + 1332, + 1159 + ], + "score": 0.88, + "latex": "\\mathbf { H } _ { q }" + }, + { + "category_id": 13, + "poly": [ + 478, + 1692, + 507, + 1692, + 507, + 1722, + 478, + 1722 + ], + "score": 0.88, + "latex": "d _ { e }" + }, + { + "category_id": 13, + "poly": [ + 734, + 761, + 769, + 761, + 769, + 787, + 734, + 787 + ], + "score": 0.88, + "latex": "\\mathbf { w } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 297, + 788, + 332, + 788, + 332, + 818, + 297, + 818 + ], + "score": 0.88, + "latex": "\\mathbf { C } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 391, + 1018, + 432, + 1018, + 432, + 1050, + 391, + 1050 + ], + "score": 0.87, + "latex": "\\mathbf { H } _ { q }" + }, + { + "category_id": 13, + "poly": [ + 1340, + 696, + 1403, + 696, + 1403, + 728, + 1340, + 728 + ], + "score": 0.87, + "latex": "Q =" + }, + { + "category_id": 13, + "poly": [ + 297, + 823, + 324, + 823, + 324, + 852, + 297, + 852 + ], + "score": 0.86, + "latex": "q _ { j }" + }, + { + "category_id": 13, + "poly": [ + 596, + 1497, + 632, + 1497, + 632, + 1522, + 596, + 1522 + ], + "score": 0.86, + "latex": "\\mathbf { w } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 966, + 1018, + 986, + 1018, + 986, + 1049, + 966, + 1049 + ], + "score": 0.86, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 597, + 1740, + 618, + 1740, + 618, + 1771, + 597, + 1771 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 840, + 987, + 861, + 987, + 861, + 1019, + 840, + 1019 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 755, + 957, + 774, + 957, + 774, + 988, + 755, + 988 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 438, + 1079, + 457, + 1079, + 457, + 1110, + 438, + 1110 + ], + "score": 0.84, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 365, + 763, + 392, + 763, + 392, + 789, + 365, + 789 + ], + "score": 0.84, + "latex": "p _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1174, + 960, + 1202, + 960, + 1202, + 987, + 1174, + 987 + ], + "score": 0.84, + "latex": "p _ { i }" + }, + { + "category_id": 13, + "poly": [ + 649, + 727, + 677, + 727, + 677, + 754, + 649, + 754 + ], + "score": 0.83, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1043, + 819, + 1060, + 819, + 1060, + 850, + 1043, + 850 + ], + "score": 0.83, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 688, + 1126, + 717, + 1126, + 717, + 1152, + 688, + 1152 + ], + "score": 0.82, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 876, + 1126, + 894, + 1126, + 894, + 1152, + 876, + 1152 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 564, + 727, + 597, + 727, + 597, + 754, + 564, + 754 + ], + "score": 0.81, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 534, + 1301, + 563, + 1301, + 563, + 1327, + 534, + 1327 + ], + "score": 0.81, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1315, + 1163, + 1333, + 1163, + 1333, + 1190, + 1315, + 1190 + ], + "score": 0.8, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 764, + 1260, + 780, + 1260, + 780, + 1280, + 764, + 1280 + ], + "score": 0.76, + "latex": "r" + }, + { + "category_id": 13, + "poly": [ + 404, + 1555, + 417, + 1555, + 417, + 1579, + 404, + 1579 + ], + "score": 0.75, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 650, + 820, + 665, + 820, + 665, + 845, + 650, + 845 + ], + "score": 0.74, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1019, + 1198, + 1036, + 1198, + 1036, + 1219, + 1019, + 1219 + ], + "score": 0.73, + "latex": "r" + }, + { + "category_id": 13, + "poly": [ + 540, + 854, + 557, + 854, + 557, + 875, + 540, + 875 + ], + "score": 0.71, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 984, + 1306, + 1002, + 1306, + 1002, + 1327, + 984, + 1327 + ], + "score": 0.63, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 796, + 1927, + 815, + 1927, + 815, + 1948, + 796, + 1948 + ], + "score": 0.61, + "latex": "\\mathbf { v }" + }, + { + "category_id": 13, + "poly": [ + 597, + 1600, + 625, + 1600, + 625, + 1628, + 597, + 1628 + ], + "score": 0.51, + "latex": "\\mathbf { C }" + }, + { + "category_id": 13, + "poly": [ + 741, + 1631, + 766, + 1631, + 766, + 1659, + 741, + 1659 + ], + "score": 0.47, + "latex": "\\mathbf { E }" + }, + { + "category_id": 13, + "poly": [ + 1114, + 1835, + 1136, + 1835, + 1136, + 1858, + 1114, + 1858 + ], + "score": 0.36, + "latex": "\\mathbf { v }" + }, + { + "category_id": 13, + "poly": [ + 457, + 1926, + 477, + 1926, + 477, + 1948, + 457, + 1948 + ], + "score": 0.26, + "latex": "\\mathbf { v }" + }, + { + "category_id": 13, + "poly": [ + 1086, + 1768, + 1175, + 1768, + 1175, + 1802, + 1086, + 1802 + ], + "score": 0.25, + "latex": "\\left[ \\mathbf { E w } ; \\mathbf { c } \\right]" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 398.0, + 696.0, + 398.0, + 696.0, + 441.0, + 293.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1429.0, + 898.0, + 1429.0, + 898.0, + 1469.0, + 294.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 636.0, + 811.0, + 636.0, + 811.0, + 672.0, + 297.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1738.0, + 596.0, + 1738.0, + 596.0, + 1772.0, + 294.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 1738.0, + 1405.0, + 1738.0, + 1405.0, + 1772.0, + 619.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1768.0, + 1085.0, + 1768.0, + 1085.0, + 1803.0, + 294.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 1768.0, + 1404.0, + 1768.0, + 1404.0, + 1803.0, + 1176.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1798.0, + 1406.0, + 1798.0, + 1406.0, + 1836.0, + 293.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1830.0, + 1113.0, + 1830.0, + 1113.0, + 1865.0, + 294.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 1830.0, + 1406.0, + 1830.0, + 1406.0, + 1865.0, + 1137.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1859.0, + 1404.0, + 1859.0, + 1404.0, + 1896.0, + 293.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1889.0, + 1402.0, + 1889.0, + 1402.0, + 1924.0, + 294.0, + 1924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1920.0, + 456.0, + 1920.0, + 456.0, + 1954.0, + 295.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 1920.0, + 532.0, + 1920.0, + 532.0, + 1954.0, + 478.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 1920.0, + 795.0, + 1920.0, + 795.0, + 1954.0, + 565.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 1920.0, + 825.0, + 1920.0, + 825.0, + 1954.0, + 816.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 691.0, + 923.0, + 691.0, + 923.0, + 733.0, + 293.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 691.0, + 1339.0, + 691.0, + 1339.0, + 733.0, + 1187.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 724.0, + 297.0, + 724.0, + 297.0, + 762.0, + 293.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 724.0, + 563.0, + 724.0, + 563.0, + 762.0, + 479.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 724.0, + 648.0, + 724.0, + 648.0, + 762.0, + 598.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 724.0, + 1406.0, + 724.0, + 1406.0, + 762.0, + 678.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 757.0, + 364.0, + 757.0, + 364.0, + 789.0, + 296.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 757.0, + 548.0, + 757.0, + 548.0, + 789.0, + 393.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 757.0, + 733.0, + 757.0, + 733.0, + 789.0, + 648.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 757.0, + 1404.0, + 757.0, + 1404.0, + 789.0, + 770.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 785.0, + 296.0, + 785.0, + 296.0, + 823.0, + 293.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 785.0, + 1404.0, + 785.0, + 1404.0, + 823.0, + 333.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 815.0, + 296.0, + 815.0, + 296.0, + 854.0, + 291.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 815.0, + 649.0, + 815.0, + 649.0, + 854.0, + 325.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 815.0, + 1042.0, + 815.0, + 1042.0, + 854.0, + 666.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 815.0, + 1408.0, + 815.0, + 1408.0, + 854.0, + 1061.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 848.0, + 539.0, + 848.0, + 539.0, + 882.0, + 292.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 848.0, + 1306.0, + 848.0, + 1306.0, + 882.0, + 558.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 895.0, + 1405.0, + 895.0, + 1405.0, + 929.0, + 295.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 924.0, + 1404.0, + 924.0, + 1404.0, + 959.0, + 295.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 957.0, + 754.0, + 957.0, + 754.0, + 990.0, + 291.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 957.0, + 1173.0, + 957.0, + 1173.0, + 990.0, + 775.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 957.0, + 1404.0, + 957.0, + 1404.0, + 990.0, + 1203.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 985.0, + 400.0, + 985.0, + 400.0, + 1022.0, + 294.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 985.0, + 839.0, + 985.0, + 839.0, + 1022.0, + 597.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 985.0, + 1406.0, + 985.0, + 1406.0, + 1022.0, + 862.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1015.0, + 390.0, + 1015.0, + 390.0, + 1052.0, + 340.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 1015.0, + 965.0, + 1015.0, + 965.0, + 1052.0, + 433.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 1015.0, + 1404.0, + 1015.0, + 1404.0, + 1052.0, + 987.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1048.0, + 1404.0, + 1048.0, + 1404.0, + 1082.0, + 294.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1075.0, + 437.0, + 1075.0, + 437.0, + 1114.0, + 293.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 1075.0, + 470.0, + 1075.0, + 470.0, + 1114.0, + 458.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1117.0, + 687.0, + 1117.0, + 687.0, + 1163.0, + 292.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 1117.0, + 875.0, + 1117.0, + 875.0, + 1163.0, + 718.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 1117.0, + 1211.0, + 1117.0, + 1211.0, + 1163.0, + 895.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 1117.0, + 1331.0, + 1117.0, + 1331.0, + 1163.0, + 1283.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1374.0, + 1117.0, + 1408.0, + 1117.0, + 1408.0, + 1163.0, + 1374.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1152.0, + 520.0, + 1152.0, + 520.0, + 1200.0, + 291.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 1152.0, + 609.0, + 1152.0, + 609.0, + 1200.0, + 561.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 1152.0, + 732.0, + 1152.0, + 732.0, + 1200.0, + 650.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 1152.0, + 817.0, + 1152.0, + 817.0, + 1200.0, + 773.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 1152.0, + 1021.0, + 1152.0, + 1021.0, + 1200.0, + 895.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 1152.0, + 1107.0, + 1152.0, + 1107.0, + 1200.0, + 1063.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 1152.0, + 1314.0, + 1152.0, + 1314.0, + 1200.0, + 1180.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1334.0, + 1152.0, + 1407.0, + 1152.0, + 1407.0, + 1200.0, + 1334.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1189.0, + 1018.0, + 1189.0, + 1018.0, + 1226.0, + 293.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 1189.0, + 1405.0, + 1189.0, + 1405.0, + 1226.0, + 1037.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1215.0, + 496.0, + 1215.0, + 496.0, + 1260.0, + 292.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 1215.0, + 1407.0, + 1215.0, + 1407.0, + 1260.0, + 700.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1250.0, + 763.0, + 1250.0, + 763.0, + 1289.0, + 293.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 1250.0, + 793.0, + 1250.0, + 793.0, + 1289.0, + 781.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 471.0, + 1403.0, + 471.0, + 1403.0, + 508.0, + 293.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 502.0, + 1406.0, + 502.0, + 1406.0, + 539.0, + 293.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 533.0, + 1405.0, + 533.0, + 1405.0, + 573.0, + 292.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 566.0, + 574.0, + 566.0, + 574.0, + 599.0, + 293.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 230.0, + 1407.0, + 230.0, + 1407.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 258.0, + 1407.0, + 258.0, + 1407.0, + 299.0, + 291.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 293.0, + 1404.0, + 293.0, + 1404.0, + 326.0, + 294.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 323.0, + 736.0, + 323.0, + 736.0, + 355.0, + 293.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1599.0, + 596.0, + 1599.0, + 596.0, + 1632.0, + 296.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 1599.0, + 1405.0, + 1599.0, + 1405.0, + 1632.0, + 626.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1630.0, + 740.0, + 1630.0, + 740.0, + 1665.0, + 292.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1630.0, + 1407.0, + 1630.0, + 1407.0, + 1665.0, + 767.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1660.0, + 1405.0, + 1660.0, + 1405.0, + 1696.0, + 293.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1692.0, + 477.0, + 1692.0, + 477.0, + 1724.0, + 296.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 1692.0, + 658.0, + 1692.0, + 658.0, + 1724.0, + 508.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1299.0, + 533.0, + 1299.0, + 533.0, + 1333.0, + 297.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1299.0, + 983.0, + 1299.0, + 983.0, + 1333.0, + 564.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 1299.0, + 1404.0, + 1299.0, + 1404.0, + 1333.0, + 1003.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1326.0, + 365.0, + 1326.0, + 365.0, + 1366.0, + 295.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 1326.0, + 1406.0, + 1326.0, + 1406.0, + 1366.0, + 409.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1358.0, + 620.0, + 1358.0, + 620.0, + 1396.0, + 295.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1489.0, + 595.0, + 1489.0, + 595.0, + 1525.0, + 295.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 1489.0, + 926.0, + 1489.0, + 926.0, + 1525.0, + 633.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1489.0, + 1404.0, + 1489.0, + 1404.0, + 1525.0, + 963.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1523.0, + 541.0, + 1523.0, + 541.0, + 1557.0, + 297.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1523.0, + 1403.0, + 1523.0, + 1403.0, + 1557.0, + 734.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1551.0, + 403.0, + 1551.0, + 403.0, + 1588.0, + 297.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 1551.0, + 679.0, + 1551.0, + 679.0, + 1588.0, + 418.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1967.0, + 667.0, + 1967.0, + 667.0, + 2003.0, + 297.0, + 2003.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1225, + 1404, + 1225, + 1404, + 1441, + 298, + 1441 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1788, + 1403, + 1788, + 1403, + 2035, + 297, + 2035 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1648, + 1403, + 1648, + 1403, + 1775, + 298, + 1775 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1087, + 1404, + 1087, + 1404, + 1210, + 299, + 1210 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 460, + 231, + 1221, + 231, + 1221, + 686, + 460, + 686 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 300, + 1455, + 1403, + 1455, + 1403, + 1548, + 300, + 1548 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 298, + 852, + 1396, + 852, + 1396, + 915, + 298, + 915 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 615, + 983, + 1085, + 983, + 1085, + 1023, + 615, + 1023 + ], + "score": 0.944 + }, + { + "category_id": 4, + "poly": [ + 291, + 733, + 1399, + 733, + 1399, + 793, + 291, + 793 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 298, + 1041, + 976, + 1041, + 976, + 1073, + 298, + 1073 + ], + "score": 0.927 + }, + { + "category_id": 1, + "poly": [ + 299, + 930, + 1150, + 930, + 1150, + 964, + 299, + 964 + ], + "score": 0.927 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 814, + 75, + 814, + 104, + 299, + 104 + ], + "score": 0.896 + }, + { + "category_id": 0, + "poly": [ + 299, + 1591, + 893, + 1591, + 893, + 1624, + 299, + 1624 + ], + "score": 0.861 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.769 + }, + { + "category_id": 13, + "poly": [ + 929, + 854, + 1016, + 854, + 1016, + 885, + 929, + 885 + ], + "score": 0.93, + "latex": "d _ { e } \\times d _ { v }" + }, + { + "category_id": 13, + "poly": [ + 710, + 1818, + 776, + 1818, + 776, + 1855, + 710, + 1855 + ], + "score": 0.92, + "latex": "\\mathbf { p } _ { i } ^ { T } \\mathbf { q } _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1307, + 1649, + 1349, + 1649, + 1349, + 1680, + 1307, + 1680 + ], + "score": 0.92, + "latex": "\\mathbf { H } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 372, + 853, + 420, + 853, + 420, + 888, + 372, + 888 + ], + "score": 0.9, + "latex": "\\mathbf { W } _ { g }" + }, + { + "category_id": 14, + "poly": [ + 613, + 983, + 1085, + 983, + 1085, + 1022, + 613, + 1022 + ], + "score": 0.9, + "latex": "\\mathbf { h } = f ( \\mathbf { c } , \\mathbf { w } ) = \\mathbf { g } \\odot \\mathbf { c } + ( 1 - \\mathbf { g } ) \\odot ( \\mathbf { E } \\mathbf { w } )" + }, + { + "category_id": 13, + "poly": [ + 1246, + 1915, + 1327, + 1915, + 1327, + 1944, + 1246, + 1944 + ], + "score": 0.9, + "latex": "\\mathbf { p } _ { i } \\odot \\mathbf { g } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 669, + 1649, + 709, + 1649, + 709, + 1683, + 669, + 1683 + ], + "score": 0.9, + "latex": "\\mathbf { Q } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 573, + 1649, + 613, + 1649, + 613, + 1680, + 573, + 1680 + ], + "score": 0.89, + "latex": "\\mathbf { P } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 1067, + 855, + 1095, + 855, + 1095, + 885, + 1067, + 885 + ], + "score": 0.89, + "latex": "d _ { e }" + }, + { + "category_id": 13, + "poly": [ + 470, + 855, + 504, + 855, + 504, + 888, + 470, + 888 + ], + "score": 0.89, + "latex": "{ \\bf b } _ { g }" + }, + { + "category_id": 13, + "poly": [ + 1178, + 1715, + 1210, + 1715, + 1210, + 1747, + 1178, + 1747 + ], + "score": 0.88, + "latex": "{ \\bf q } _ { j }" + }, + { + "category_id": 13, + "poly": [ + 410, + 1743, + 442, + 1743, + 442, + 1773, + 410, + 1773 + ], + "score": 0.88, + "latex": "d _ { h }" + }, + { + "category_id": 13, + "poly": [ + 765, + 1748, + 796, + 1748, + 796, + 1777, + 765, + 1777 + ], + "score": 0.88, + "latex": "{ \\bf q } _ { j }" + }, + { + "category_id": 13, + "poly": [ + 684, + 1746, + 714, + 1746, + 714, + 1775, + 684, + 1775 + ], + "score": 0.87, + "latex": "\\mathbf { p } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 829, + 1716, + 860, + 1716, + 860, + 1745, + 829, + 1745 + ], + "score": 0.86, + "latex": "\\mathbf { p } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 434, + 1977, + 464, + 1977, + 464, + 2004, + 434, + 2004 + ], + "score": 0.85, + "latex": "\\mathbf { g } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 975, + 1947, + 1002, + 1947, + 1002, + 1974, + 975, + 1974 + ], + "score": 0.84, + "latex": "p _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1333, + 1715, + 1349, + 1715, + 1349, + 1745, + 1333, + 1745 + ], + "score": 0.83, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 296, + 1947, + 326, + 1947, + 326, + 1974, + 296, + 1974 + ], + "score": 0.83, + "latex": "\\mathbf { g } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 372, + 1045, + 397, + 1045, + 397, + 1070, + 372, + 1070 + ], + "score": 0.81, + "latex": "\\odot" + }, + { + "category_id": 13, + "poly": [ + 298, + 1713, + 316, + 1713, + 316, + 1740, + 298, + 1740 + ], + "score": 0.79, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1153, + 859, + 1174, + 859, + 1174, + 881, + 1153, + 881 + ], + "score": 0.77, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 982, + 1715, + 995, + 1715, + 995, + 1740, + 982, + 1740 + ], + "score": 0.69, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1106, + 1713, + 1131, + 1713, + 1131, + 1741, + 1106, + 1741 + ], + "score": 0.67, + "latex": "\\mathbf { P }" + }, + { + "category_id": 13, + "poly": [ + 724, + 1153, + 745, + 1153, + 745, + 1181, + 724, + 1181 + ], + "score": 0.39, + "latex": "\\mathbf { g }" + }, + { + "category_id": 13, + "poly": [ + 328, + 1743, + 355, + 1743, + 355, + 1775, + 328, + 1775 + ], + "score": 0.37, + "latex": "\\mathbf { Q }" + }, + { + "category_id": 13, + "poly": [ + 459, + 1123, + 481, + 1123, + 481, + 1151, + 459, + 1151 + ], + "score": 0.36, + "latex": "\\mathbf { g }" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 230.0, + 648.0, + 230.0, + 648.0, + 264.0, + 533.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 256.0, + 673.0, + 256.0, + 673.0, + 292.0, + 507.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 332.0, + 601.0, + 332.0, + 601.0, + 348.0, + 582.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 349.0, + 923.0, + 349.0, + 923.0, + 397.0, + 703.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 406.0, + 816.0, + 406.0, + 816.0, + 451.0, + 701.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 409.0, + 1010.0, + 409.0, + 1010.0, + 452.0, + 904.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 461.0, + 576.0, + 461.0, + 576.0, + 508.0, + 461.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 462.0, + 716.0, + 462.0, + 716.0, + 506.0, + 602.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 461.0, + 1111.0, + 461.0, + 1111.0, + 511.0, + 963.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 506.0, + 1203.0, + 506.0, + 1203.0, + 537.0, + 1127.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 541.0, + 567.0, + 541.0, + 567.0, + 587.0, + 466.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 543.0, + 727.0, + 543.0, + 727.0, + 582.0, + 598.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 543.0, + 1217.0, + 543.0, + 1217.0, + 586.0, + 846.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 650.0, + 730.0, + 650.0, + 730.0, + 689.0, + 599.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.75, + 361.0, + 528.75, + 361.0, + 528.75, + 389.0, + 514.75, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 731.0, + 1402.0, + 731.0, + 1402.0, + 769.0, + 294.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 763.0, + 1068.0, + 763.0, + 1068.0, + 794.0, + 295.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1590.0, + 895.0, + 1590.0, + 895.0, + 1629.0, + 294.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1224.0, + 1405.0, + 1224.0, + 1405.0, + 1263.0, + 293.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1256.0, + 1405.0, + 1256.0, + 1405.0, + 1292.0, + 292.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1287.0, + 1405.0, + 1287.0, + 1405.0, + 1322.0, + 294.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1318.0, + 1405.0, + 1318.0, + 1405.0, + 1352.0, + 293.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1345.0, + 1405.0, + 1345.0, + 1405.0, + 1386.0, + 291.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1376.0, + 1406.0, + 1376.0, + 1406.0, + 1416.0, + 293.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1411.0, + 1097.0, + 1411.0, + 1097.0, + 1445.0, + 294.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1790.0, + 1405.0, + 1790.0, + 1405.0, + 1824.0, + 295.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1819.0, + 709.0, + 1819.0, + 709.0, + 1860.0, + 291.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 1819.0, + 1408.0, + 1819.0, + 1408.0, + 1860.0, + 777.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 1407.0, + 1850.0, + 1407.0, + 1886.0, + 294.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 1405.0, + 1880.0, + 1405.0, + 1916.0, + 293.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1909.0, + 1245.0, + 1909.0, + 1245.0, + 1949.0, + 291.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 1909.0, + 1408.0, + 1909.0, + 1408.0, + 1949.0, + 1328.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1942.0, + 295.0, + 1942.0, + 295.0, + 1978.0, + 291.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1942.0, + 974.0, + 1942.0, + 974.0, + 1978.0, + 327.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 1942.0, + 1408.0, + 1942.0, + 1408.0, + 1978.0, + 1003.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1974.0, + 433.0, + 1974.0, + 433.0, + 2005.0, + 292.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1974.0, + 1406.0, + 1974.0, + 1406.0, + 2005.0, + 465.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2003.0, + 812.0, + 2003.0, + 812.0, + 2037.0, + 292.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1645.0, + 572.0, + 1645.0, + 572.0, + 1687.0, + 292.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 1645.0, + 668.0, + 1645.0, + 668.0, + 1687.0, + 614.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 1645.0, + 1306.0, + 1645.0, + 1306.0, + 1687.0, + 710.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1350.0, + 1645.0, + 1408.0, + 1645.0, + 1408.0, + 1687.0, + 1350.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1678.0, + 1405.0, + 1678.0, + 1405.0, + 1717.0, + 292.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1711.0, + 297.0, + 1711.0, + 297.0, + 1747.0, + 293.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1711.0, + 828.0, + 1711.0, + 828.0, + 1747.0, + 317.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1711.0, + 981.0, + 1711.0, + 981.0, + 1747.0, + 861.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 1711.0, + 1105.0, + 1711.0, + 1105.0, + 1747.0, + 996.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 1711.0, + 1177.0, + 1711.0, + 1177.0, + 1747.0, + 1132.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1711.0, + 1332.0, + 1711.0, + 1332.0, + 1747.0, + 1211.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1350.0, + 1711.0, + 1403.0, + 1711.0, + 1403.0, + 1747.0, + 1350.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1739.0, + 327.0, + 1739.0, + 327.0, + 1781.0, + 292.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1739.0, + 409.0, + 1739.0, + 409.0, + 1781.0, + 356.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 1739.0, + 683.0, + 1739.0, + 683.0, + 1781.0, + 443.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1739.0, + 764.0, + 1739.0, + 764.0, + 1781.0, + 715.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 1739.0, + 808.0, + 1739.0, + 808.0, + 1781.0, + 797.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1085.0, + 1404.0, + 1085.0, + 1404.0, + 1125.0, + 293.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1118.0, + 458.0, + 1118.0, + 458.0, + 1151.0, + 295.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1118.0, + 1403.0, + 1118.0, + 1403.0, + 1151.0, + 482.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1149.0, + 723.0, + 1149.0, + 723.0, + 1185.0, + 294.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 1149.0, + 1405.0, + 1149.0, + 1405.0, + 1185.0, + 746.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1180.0, + 593.0, + 1180.0, + 593.0, + 1212.0, + 294.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1454.0, + 1405.0, + 1454.0, + 1405.0, + 1492.0, + 295.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1487.0, + 1405.0, + 1487.0, + 1405.0, + 1521.0, + 295.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1517.0, + 370.0, + 1517.0, + 370.0, + 1550.0, + 294.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 852.0, + 371.0, + 852.0, + 371.0, + 887.0, + 295.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 852.0, + 469.0, + 852.0, + 469.0, + 887.0, + 421.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 852.0, + 928.0, + 852.0, + 928.0, + 887.0, + 505.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 852.0, + 1066.0, + 852.0, + 1066.0, + 887.0, + 1017.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 852.0, + 1152.0, + 852.0, + 1152.0, + 887.0, + 1096.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 852.0, + 1400.0, + 852.0, + 1400.0, + 887.0, + 1175.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 886.0, + 555.0, + 886.0, + 555.0, + 918.0, + 297.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1039.0, + 371.0, + 1039.0, + 371.0, + 1077.0, + 296.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1039.0, + 978.0, + 1039.0, + 978.0, + 1077.0, + 398.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 925.0, + 1152.0, + 925.0, + 1152.0, + 971.0, + 293.0, + 971.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1680, + 1404, + 1680, + 1404, + 1927, + 297, + 1927 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1432, + 1405, + 1432, + 1405, + 1586, + 297, + 1586 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1164, + 1405, + 1164, + 1405, + 1320, + 298, + 1320 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 831, + 1404, + 831, + 1404, + 952, + 297, + 952 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 300, + 1942, + 1403, + 1942, + 1403, + 2034, + 300, + 2034 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 302, + 229, + 1363, + 229, + 1363, + 696, + 302, + 696 + ], + "score": 0.97 + }, + { + "category_id": 8, + "poly": [ + 578, + 1076, + 1119, + 1076, + 1119, + 1148, + 578, + 1148 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 297, + 995, + 1401, + 995, + 1401, + 1058, + 297, + 1058 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 728, + 953, + 969, + 953, + 969, + 989, + 728, + 989 + ], + "score": 0.933 + }, + { + "category_id": 4, + "poly": [ + 595, + 744, + 1102, + 744, + 1102, + 777, + 595, + 777 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 298, + 1622, + 1026, + 1622, + 1026, + 1656, + 298, + 1656 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 300, + 1364, + 557, + 1364, + 557, + 1398, + 300, + 1398 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 814, + 75, + 814, + 104, + 299, + 104 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.722 + }, + { + "category_id": 14, + "poly": [ + 576, + 1072, + 1120, + 1072, + 1120, + 1151, + 576, + 1151 + ], + "score": 0.94, + "latex": "\\mathbf { h } _ { i } = \\sum _ { j } \\mathrm { s o f t m a x } ( \\mathbf { u } _ { h } ^ { T } \\mathbf { I } _ { i j } + \\mathbf { w } _ { i } ^ { T } \\mathbf { w } _ { j } b _ { h 1 } + b _ { h 2 } ) \\mathbf { I } _ { i j }" + }, + { + "category_id": 14, + "poly": [ + 728, + 952, + 971, + 952, + 971, + 990, + 728, + 990 + ], + "score": 0.91, + "latex": "\\mathbf { I } _ { i j } = \\operatorname { t a n h } ( \\mathbf { p } _ { i } \\odot \\mathbf { q } _ { j } )" + }, + { + "category_id": 13, + "poly": [ + 940, + 1167, + 981, + 1167, + 981, + 1196, + 940, + 1196 + ], + "score": 0.89, + "latex": "b _ { h 2 }" + }, + { + "category_id": 13, + "poly": [ + 845, + 1167, + 887, + 1167, + 887, + 1197, + 845, + 1197 + ], + "score": 0.89, + "latex": "b _ { h 1 }" + }, + { + "category_id": 13, + "poly": [ + 540, + 1258, + 574, + 1258, + 574, + 1291, + 540, + 1291 + ], + "score": 0.89, + "latex": "\\mathbf { I } _ { i j }" + }, + { + "category_id": 13, + "poly": [ + 352, + 1029, + 387, + 1029, + 387, + 1061, + 352, + 1061 + ], + "score": 0.89, + "latex": "\\mathbf { I } _ { i j }" + }, + { + "category_id": 13, + "poly": [ + 647, + 1028, + 678, + 1028, + 678, + 1057, + 647, + 1057 + ], + "score": 0.89, + "latex": "\\mathbf { h } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 462, + 1167, + 492, + 1167, + 492, + 1196, + 462, + 1196 + ], + "score": 0.89, + "latex": "d _ { v }" + }, + { + "category_id": 13, + "poly": [ + 843, + 1289, + 873, + 1289, + 873, + 1318, + 843, + 1318 + ], + "score": 0.88, + "latex": "\\mathbf { h } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1201, + 335, + 1201, + 335, + 1231, + 297, + 1231 + ], + "score": 0.87, + "latex": "{ \\bf w } _ { j }" + }, + { + "category_id": 13, + "poly": [ + 955, + 1000, + 985, + 1000, + 985, + 1027, + 955, + 1027 + ], + "score": 0.87, + "latex": "\\mathbf { p } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 371, + 1000, + 402, + 1000, + 402, + 1029, + 371, + 1029 + ], + "score": 0.86, + "latex": "{ \\bf q } _ { j }" + }, + { + "category_id": 13, + "poly": [ + 373, + 1171, + 408, + 1171, + 408, + 1196, + 373, + 1196 + ], + "score": 0.86, + "latex": "\\mathbf { u } _ { h }" + }, + { + "category_id": 13, + "poly": [ + 712, + 1201, + 739, + 1201, + 739, + 1230, + 712, + 1230 + ], + "score": 0.86, + "latex": "q _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1316, + 1171, + 1352, + 1171, + 1352, + 1196, + 1316, + 1196 + ], + "score": 0.85, + "latex": "\\mathbf { w } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1327, + 898, + 1354, + 898, + 1354, + 928, + 1327, + 928 + ], + "score": 0.85, + "latex": "q _ { j }" + }, + { + "category_id": 13, + "poly": [ + 634, + 1201, + 660, + 1201, + 660, + 1228, + 634, + 1228 + ], + "score": 0.85, + "latex": "p _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1060, + 1293, + 1086, + 1293, + 1086, + 1319, + 1060, + 1319 + ], + "score": 0.84, + "latex": "p _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1250, + 898, + 1276, + 898, + 1276, + 925, + 1250, + 925 + ], + "score": 0.84, + "latex": "p _ { i }" + }, + { + "category_id": 13, + "poly": [ + 645, + 1259, + 664, + 1259, + 664, + 1285, + 645, + 1285 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1251, + 1259, + 1268, + 1259, + 1268, + 1290, + 1251, + 1290 + ], + "score": 0.77, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 448, + 1289, + 477, + 1289, + 477, + 1315, + 448, + 1315 + ], + "score": 0.55, + "latex": "\\mathbf { H }" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 229.0, + 841.0, + 229.0, + 841.0, + 268.0, + 727.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 291.0, + 550.0, + 291.0, + 550.0, + 330.0, + 437.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 301.0, + 1282.0, + 301.0, + 1282.0, + 318.0, + 1256.0, + 318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1304.0, + 296.0, + 1348.0, + 296.0, + 1348.0, + 322.0, + 1304.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 331.0, + 550.0, + 331.0, + 550.0, + 374.0, + 436.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 348.0, + 840.0, + 348.0, + 840.0, + 381.0, + 645.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 347.0, + 1265.0, + 347.0, + 1265.0, + 359.0, + 1257.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 340.0, + 1348.0, + 340.0, + 1348.0, + 364.0, + 1291.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 375.0, + 429.0, + 375.0, + 429.0, + 409.0, + 306.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 374.0, + 646.0, + 374.0, + 646.0, + 431.0, + 430.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 374.0, + 803.0, + 374.0, + 803.0, + 421.0, + 679.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 367.0, + 1042.0, + 367.0, + 1042.0, + 412.0, + 964.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 368.0, + 1209.0, + 368.0, + 1209.0, + 415.0, + 1084.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 375.0, + 1356.0, + 375.0, + 1356.0, + 418.0, + 1239.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 431.0, + 491.0, + 431.0, + 491.0, + 440.0, + 482.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 446.0, + 556.0, + 446.0, + 556.0, + 497.0, + 432.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 447.0, + 1360.0, + 447.0, + 1360.0, + 494.0, + 1241.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 494.0, + 582.0, + 494.0, + 582.0, + 532.0, + 415.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 506.0, + 837.0, + 506.0, + 837.0, + 537.0, + 741.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 498.0, + 1343.0, + 498.0, + 1343.0, + 535.0, + 1267.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 524.0, + 536.0, + 524.0, + 536.0, + 561.0, + 461.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 575.0, + 1058.0, + 575.0, + 1058.0, + 615.0, + 891.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 605.0, + 1011.0, + 605.0, + 1011.0, + 642.0, + 939.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 611.0, + 892.0, + 611.0, + 892.0, + 644.0, + 864.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 658.0, + 811.0, + 658.0, + 811.0, + 699.0, + 733.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1298.0, + 303.5, + 1302.0, + 303.5, + 1302.0, + 317.5, + 1298.0, + 317.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 738.0, + 1106.0, + 738.0, + 1106.0, + 785.0, + 591.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1621.0, + 1030.0, + 1621.0, + 1030.0, + 1659.0, + 294.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1361.0, + 561.0, + 1361.0, + 561.0, + 1404.0, + 293.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1679.0, + 1405.0, + 1679.0, + 1405.0, + 1716.0, + 295.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1713.0, + 1404.0, + 1713.0, + 1404.0, + 1746.0, + 294.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1743.0, + 1405.0, + 1743.0, + 1405.0, + 1777.0, + 295.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 1406.0, + 1773.0, + 1406.0, + 1807.0, + 294.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1802.0, + 1404.0, + 1802.0, + 1404.0, + 1838.0, + 292.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1835.0, + 1405.0, + 1835.0, + 1405.0, + 1869.0, + 294.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1860.0, + 1406.0, + 1860.0, + 1406.0, + 1901.0, + 291.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1895.0, + 1270.0, + 1895.0, + 1270.0, + 1929.0, + 295.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1434.0, + 1406.0, + 1434.0, + 1406.0, + 1467.0, + 297.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1463.0, + 1403.0, + 1463.0, + 1403.0, + 1498.0, + 295.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1493.0, + 1405.0, + 1493.0, + 1405.0, + 1532.0, + 292.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1524.0, + 1402.0, + 1524.0, + 1402.0, + 1561.0, + 294.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1555.0, + 569.0, + 1555.0, + 569.0, + 1588.0, + 295.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1166.0, + 372.0, + 1166.0, + 372.0, + 1199.0, + 296.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 1166.0, + 461.0, + 1166.0, + 461.0, + 1199.0, + 409.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 1166.0, + 844.0, + 1166.0, + 844.0, + 1199.0, + 493.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 1166.0, + 939.0, + 1166.0, + 939.0, + 1199.0, + 888.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 1166.0, + 1315.0, + 1166.0, + 1315.0, + 1199.0, + 982.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 1166.0, + 1403.0, + 1166.0, + 1403.0, + 1199.0, + 1353.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1195.0, + 296.0, + 1195.0, + 296.0, + 1231.0, + 293.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 1195.0, + 633.0, + 1195.0, + 633.0, + 1231.0, + 336.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1195.0, + 711.0, + 1195.0, + 711.0, + 1231.0, + 661.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1195.0, + 1405.0, + 1195.0, + 1405.0, + 1231.0, + 740.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1227.0, + 1405.0, + 1227.0, + 1405.0, + 1261.0, + 296.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1257.0, + 539.0, + 1257.0, + 539.0, + 1294.0, + 293.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1257.0, + 644.0, + 1257.0, + 644.0, + 1294.0, + 575.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 1257.0, + 1250.0, + 1257.0, + 1250.0, + 1294.0, + 665.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 1257.0, + 1406.0, + 1257.0, + 1406.0, + 1294.0, + 1269.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1283.0, + 447.0, + 1283.0, + 447.0, + 1326.0, + 293.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 1283.0, + 842.0, + 1283.0, + 842.0, + 1326.0, + 478.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 1283.0, + 1059.0, + 1283.0, + 1059.0, + 1326.0, + 874.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 1283.0, + 1098.0, + 1283.0, + 1098.0, + 1326.0, + 1087.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 831.0, + 1406.0, + 831.0, + 1406.0, + 869.0, + 296.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 858.0, + 1404.0, + 858.0, + 1404.0, + 898.0, + 295.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 889.0, + 1249.0, + 889.0, + 1249.0, + 932.0, + 294.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 889.0, + 1326.0, + 889.0, + 1326.0, + 932.0, + 1277.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 889.0, + 1405.0, + 889.0, + 1405.0, + 932.0, + 1355.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 921.0, + 358.0, + 921.0, + 358.0, + 958.0, + 293.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1944.0, + 1405.0, + 1944.0, + 1405.0, + 1974.0, + 298.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2006.0, + 296.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2005.0, + 1403.0, + 2005.0, + 1403.0, + 2035.0, + 296.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 993.0, + 370.0, + 993.0, + 370.0, + 1029.0, + 297.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 993.0, + 954.0, + 993.0, + 954.0, + 1029.0, + 403.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 993.0, + 1405.0, + 993.0, + 1405.0, + 1029.0, + 986.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1027.0, + 351.0, + 1027.0, + 351.0, + 1058.0, + 293.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1027.0, + 646.0, + 1027.0, + 646.0, + 1058.0, + 388.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 1027.0, + 802.0, + 1027.0, + 802.0, + 1058.0, + 679.0, + 1058.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1413, + 1404, + 1413, + 1404, + 1690, + 297, + 1690 + ], + "score": 0.983 + }, + { + "category_id": 5, + "poly": [ + 418, + 735, + 1281, + 735, + 1281, + 1236, + 418, + 1236 + ], + "score": 0.98, + "html": "
ModelCN devCN testNE devNE test
GA word char concat0.7310.6960.7680.725
GA word char feat concat0.72500.69280.78150.7256
GA scalar gate0.72400.69080.78100.7260
GA fine-grained gate0.74250.70840.78900.7464
FG fine-grained gate0.75300.72040.79100.7496
Sordoni et al. (2016)0.7210.6920.7520.686
Trischler et al. (2016)0.7150.6740.7530.697
Cui et al. (2016)0.7220.6940.7780.720
Munkhdalai & Yu (2016)0.7430.7190.7820.732
Kadlec et al. (2016) ensemble0.7110.6890.7620.710
Sordoni et al. (2016) ensemble0.7410.7100.7690.720
Trischler etal. (2016) ensemble0.7360.7060.7660.718
" + }, + { + "category_id": 1, + "poly": [ + 300, + 1788, + 1404, + 1788, + 1404, + 1911, + 300, + 1911 + ], + "score": 0.976 + }, + { + "category_id": 5, + "poly": [ + 426, + 316, + 1272, + 316, + 1272, + 573, + 426, + 573 + ], + "score": 0.975, + "html": "
ModelPrecision@1Recall@10Mean Rank
word (Dhingra et al., 2016b)0.2410.428133
char (Dhingra et al., 2016b)0.2840.485104
word char concat0.29610.4959105.8
word char feat concat0.29510.4974106.2
scalar gate0.29740.4982104.2
fine-grained gate0.30690.5119101.5
" + }, + { + "category_id": 6, + "poly": [ + 293, + 251, + 1405, + 251, + 1405, + 309, + 293, + 309 + ], + "score": 0.936 + }, + { + "category_id": 6, + "poly": [ + 299, + 640, + 1402, + 640, + 1402, + 727, + 299, + 727 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 300, + 1306, + 1401, + 1306, + 1401, + 1399, + 300, + 1399 + ], + "score": 0.916 + }, + { + "category_id": 1, + "poly": [ + 299, + 2002, + 1084, + 2002, + 1084, + 2035, + 299, + 2035 + ], + "score": 0.914 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 300, + 1948, + 707, + 1948, + 707, + 1979, + 300, + 1979 + ], + "score": 0.877 + }, + { + "category_id": 0, + "poly": [ + 299, + 1729, + 927, + 1729, + 927, + 1761, + 299, + 1761 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2112, + 841, + 2112 + ], + "score": 0.775 + }, + { + "category_id": 13, + "poly": [ + 515, + 1476, + 590, + 1476, + 590, + 1506, + 515, + 1506 + ], + "score": 0.85, + "latex": "2 . 6 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 428, + 1476, + 503, + 1476, + 503, + 1506, + 428, + 1506 + ], + "score": 0.84, + "latex": "2 . 2 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 1313, + 1416, + 1352, + 1416, + 1352, + 1443, + 1313, + 1443 + ], + "score": 0.64, + "latex": "@ 1" + }, + { + "category_id": 13, + "poly": [ + 340, + 1446, + 394, + 1446, + 394, + 1474, + 340, + 1474 + ], + "score": 0.59, + "latex": "@ 1 0" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 247.0, + 1405.0, + 247.0, + 1405.0, + 282.0, + 294.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 278.0, + 585.0, + 278.0, + 585.0, + 312.0, + 294.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 638.0, + 1404.0, + 638.0, + 1404.0, + 674.0, + 293.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 668.0, + 1404.0, + 668.0, + 1404.0, + 701.0, + 293.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 699.0, + 1120.0, + 699.0, + 1120.0, + 729.0, + 293.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 713.0, + 1942.0, + 713.0, + 1987.0, + 293.0, + 1987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1729.0, + 931.0, + 1729.0, + 931.0, + 1765.0, + 295.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1411.0, + 1312.0, + 1411.0, + 1312.0, + 1449.0, + 294.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 1411.0, + 1404.0, + 1411.0, + 1404.0, + 1449.0, + 1353.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1444.0, + 339.0, + 1444.0, + 339.0, + 1481.0, + 294.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1444.0, + 1404.0, + 1444.0, + 1404.0, + 1481.0, + 395.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1473.0, + 427.0, + 1473.0, + 427.0, + 1512.0, + 294.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 1473.0, + 514.0, + 1473.0, + 514.0, + 1512.0, + 504.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1473.0, + 1404.0, + 1473.0, + 1404.0, + 1512.0, + 591.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1506.0, + 1405.0, + 1506.0, + 1405.0, + 1542.0, + 295.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1534.0, + 1404.0, + 1534.0, + 1404.0, + 1574.0, + 292.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1565.0, + 1405.0, + 1565.0, + 1405.0, + 1603.0, + 292.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1596.0, + 1404.0, + 1596.0, + 1404.0, + 1633.0, + 292.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1629.0, + 1405.0, + 1629.0, + 1405.0, + 1661.0, + 295.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1659.0, + 1189.0, + 1659.0, + 1189.0, + 1694.0, + 294.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1787.0, + 1405.0, + 1787.0, + 1405.0, + 1823.0, + 295.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1818.0, + 1404.0, + 1818.0, + 1404.0, + 1853.0, + 294.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 1403.0, + 1851.0, + 1403.0, + 1883.0, + 295.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1879.0, + 654.0, + 1879.0, + 654.0, + 1915.0, + 295.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1305.0, + 1406.0, + 1305.0, + 1406.0, + 1343.0, + 295.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1338.0, + 1405.0, + 1338.0, + 1405.0, + 1371.0, + 295.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1367.0, + 1398.0, + 1367.0, + 1398.0, + 1403.0, + 294.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1998.0, + 1088.0, + 1998.0, + 1088.0, + 2039.0, + 295.0, + 2039.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1433, + 1403, + 1433, + 1403, + 1680, + 298, + 1680 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1125, + 1404, + 1125, + 1404, + 1311, + 297, + 1311 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 300, + 1880, + 1403, + 1880, + 1403, + 2034, + 300, + 2034 + ], + "score": 0.976 + }, + { + "category_id": 5, + "poly": [ + 449, + 675, + 1248, + 675, + 1248, + 1054, + 449, + 1054 + ], + "score": 0.975, + "html": "
ModelF1Exact Match
GA word0.5492
GA word char concat0.6695 0.68570.5639
GA word char feat concat0.69040.5711
GA scalar gate0.68500.5620
GA fine-grained gate0.69830.5804
FG fine-grained gate0.71250.5995
FG fine-grained gate + ensemble0.7341 (0.733)0.6238 (0.625)
Yu et al. (2016)
Wang & Jiang (2016)0.712 (0.710) 0.700 (0.703)0.625 (0.625) 0.591 (0.595)
" + }, + { + "category_id": 5, + "poly": [ + 364, + 312, + 1334, + 312, + 1334, + 569, + 364, + 569 + ], + "score": 0.974, + "html": "
ModelWDW devWDW testWDW-R devWDW-R test
0.5700.590
Kadlec et al. (2016) Chen et al. (2016)0.640一0.650
Munkhdalai & Yu (2016)0.6650.6620.6700.667
Dhingra et al. (2016a)0.7160.7120.7260.726
this paper0.7230.7170.7310.726
" + }, + { + "category_id": 1, + "poly": [ + 300, + 1694, + 1403, + 1694, + 1403, + 1787, + 300, + 1787 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 301, + 1326, + 1397, + 1326, + 1397, + 1419, + 301, + 1419 + ], + "score": 0.963 + }, + { + "category_id": 6, + "poly": [ + 296, + 250, + 1398, + 250, + 1398, + 308, + 296, + 308 + ], + "score": 0.942 + }, + { + "category_id": 0, + "poly": [ + 300, + 1824, + 487, + 1824, + 487, + 1857, + 300, + 1857 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 104, + 300, + 104 + ], + "score": 0.899 + }, + { + "category_id": 6, + "poly": [ + 395, + 638, + 1302, + 638, + 1302, + 669, + 395, + 669 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 842, + 2087, + 858, + 2087, + 858, + 2111, + 842, + 2111 + ], + "score": 0.638 + }, + { + "category_id": 2, + "poly": [ + 842, + 2087, + 859, + 2087, + 859, + 2111, + 842, + 2111 + ], + "score": 0.3 + }, + { + "category_id": 13, + "poly": [ + 763, + 1526, + 837, + 1526, + 837, + 1555, + 763, + 1555 + ], + "score": 0.88, + "latex": "1 . 7 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 664, + 1557, + 725, + 1557, + 725, + 1586, + 664, + 1586 + ], + "score": 0.88, + "latex": "2 . 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 968, + 1973, + 1022, + 1973, + 1022, + 2003, + 968, + 2003 + ], + "score": 0.75, + "latex": "80 \\%" + }, + { + "category_id": 13, + "poly": [ + 494, + 2003, + 547, + 2003, + 547, + 2032, + 494, + 2032 + ], + "score": 0.69, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 703, + 929, + 726, + 929, + 726, + 952, + 703, + 952 + ], + "score": 0.33, + "latex": "^ +" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 247.0, + 1403.0, + 247.0, + 1403.0, + 284.0, + 295.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 276.0, + 733.0, + 276.0, + 733.0, + 309.0, + 296.0, + 309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1820.0, + 491.0, + 1820.0, + 491.0, + 1863.0, + 294.0, + 1863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 636.0, + 1305.0, + 636.0, + 1305.0, + 671.0, + 394.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 861.0, + 2086.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1435.0, + 1405.0, + 1435.0, + 1405.0, + 1469.0, + 296.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1465.0, + 1405.0, + 1465.0, + 1405.0, + 1499.0, + 293.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1496.0, + 1405.0, + 1496.0, + 1405.0, + 1530.0, + 294.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1526.0, + 762.0, + 1526.0, + 762.0, + 1560.0, + 294.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 1526.0, + 1405.0, + 1526.0, + 1405.0, + 1560.0, + 838.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1552.0, + 663.0, + 1552.0, + 663.0, + 1595.0, + 292.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 1552.0, + 1407.0, + 1552.0, + 1407.0, + 1595.0, + 726.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1588.0, + 1403.0, + 1588.0, + 1403.0, + 1618.0, + 296.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1618.0, + 1407.0, + 1618.0, + 1407.0, + 1654.0, + 293.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1648.0, + 1404.0, + 1648.0, + 1404.0, + 1682.0, + 293.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1125.0, + 1405.0, + 1125.0, + 1405.0, + 1162.0, + 294.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1157.0, + 1405.0, + 1157.0, + 1405.0, + 1192.0, + 295.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1186.0, + 1406.0, + 1186.0, + 1406.0, + 1224.0, + 294.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1217.0, + 1407.0, + 1217.0, + 1407.0, + 1256.0, + 292.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1251.0, + 1405.0, + 1251.0, + 1405.0, + 1283.0, + 296.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1275.0, + 1083.0, + 1275.0, + 1083.0, + 1319.0, + 292.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1880.0, + 1405.0, + 1880.0, + 1405.0, + 1917.0, + 296.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1911.0, + 1404.0, + 1911.0, + 1404.0, + 1948.0, + 294.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1979.0, + 295.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 967.0, + 1972.0, + 967.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 1972.0, + 1408.0, + 1972.0, + 1408.0, + 2009.0, + 1023.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 493.0, + 2002.0, + 493.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 2002.0, + 1295.0, + 2002.0, + 1295.0, + 2039.0, + 548.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1692.0, + 1404.0, + 1692.0, + 1404.0, + 1730.0, + 295.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1724.0, + 1403.0, + 1724.0, + 1403.0, + 1760.0, + 295.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1754.0, + 799.0, + 1754.0, + 799.0, + 1792.0, + 295.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1323.0, + 1403.0, + 1323.0, + 1403.0, + 1363.0, + 293.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1357.0, + 1403.0, + 1357.0, + 1403.0, + 1391.0, + 295.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1387.0, + 1362.0, + 1387.0, + 1362.0, + 1424.0, + 295.0, + 1424.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1275, + 1405, + 1275, + 1405, + 1522, + 297, + 1522 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1680, + 1403, + 1680, + 1403, + 1865, + 298, + 1865 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 300, + 1880, + 1402, + 1880, + 1402, + 2033, + 300, + 2033 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 310, + 663, + 1389, + 663, + 1389, + 1069, + 310, + 1069 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 303, + 229, + 1396, + 229, + 1396, + 476, + 303, + 476 + ], + "score": 0.951 + }, + { + "category_id": 4, + "poly": [ + 296, + 513, + 1405, + 513, + 1405, + 627, + 296, + 627 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.897 + }, + { + "category_id": 4, + "poly": [ + 291, + 1106, + 1402, + 1106, + 1402, + 1165, + 291, + 1165 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 301, + 1606, + 741, + 1606, + 741, + 1636, + 301, + 1636 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.782 + }, + { + "category_id": 13, + "poly": [ + 747, + 515, + 792, + 515, + 792, + 545, + 747, + 545 + ], + "score": 0.89, + "latex": "\\mathbf { W } _ { g }" + }, + { + "category_id": 13, + "poly": [ + 679, + 1682, + 727, + 1682, + 727, + 1716, + 679, + 1716 + ], + "score": 0.88, + "latex": "\\mathbf { W } _ { g }" + }, + { + "category_id": 13, + "poly": [ + 652, + 1713, + 700, + 1713, + 700, + 1747, + 652, + 1747 + ], + "score": 0.88, + "latex": "\\mathbf { W } _ { g }" + }, + { + "category_id": 13, + "poly": [ + 973, + 603, + 991, + 603, + 991, + 623, + 973, + 623 + ], + "score": 0.72, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 701, + 603, + 720, + 603, + 720, + 623, + 701, + 623 + ], + "score": 0.62, + "latex": "_ n" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 709.0, + 512.0, + 709.0, + 512.0, + 736.0, + 487.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 719.0, + 554.0, + 719.0, + 554.0, + 730.0, + 542.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 709.0, + 826.0, + 709.0, + 826.0, + 735.0, + 795.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 709.0, + 1111.0, + 709.0, + 1111.0, + 738.0, + 1080.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1289.0, + 710.0, + 1315.0, + 710.0, + 1315.0, + 737.0, + 1289.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 847.0, + 568.0, + 847.0, + 568.0, + 876.0, + 538.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 850.0, + 772.0, + 850.0, + 772.0, + 876.0, + 744.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 862.0, + 1031.0, + 862.0, + 1031.0, + 872.0, + 1020.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 857.0, + 1087.0, + 857.0, + 1087.0, + 879.0, + 1066.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1347.0, + 856.0, + 1367.0, + 856.0, + 1367.0, + 876.0, + 1347.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 991.0, + 660.0, + 991.0, + 660.0, + 1003.0, + 649.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 989.0, + 1115.0, + 989.0, + 1115.0, + 1016.0, + 1089.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 996.0, + 1160.0, + 996.0, + 1160.0, + 1014.0, + 1134.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1349.0, + 996.0, + 1369.0, + 996.0, + 1369.0, + 1016.0, + 1349.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.25, + 706.5, + 916.25, + 706.5, + 916.25, + 733.5, + 862.25, + 733.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 702.5, + 377.0, + 702.5, + 377.0, + 724.0, + 312.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 702.5, + 426.0, + 702.5, + 426.0, + 722.0, + 363.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 698.0, + 482.0, + 698.0, + 482.0, + 714.0, + 412.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.75, + 696.0, + 642.75, + 696.0, + 642.75, + 717.0, + 562.75, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.75, + 709.5, + 729.75, + 709.5, + 729.75, + 729.5, + 680.75, + 729.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.75, + 699.0, + 791.75, + 699.0, + 791.75, + 722.5, + 721.75, + 722.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 707.5, + 967.0, + 707.5, + 967.0, + 732.5, + 913.0, + 732.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 705.5, + 1020.0, + 705.5, + 1020.0, + 726.0, + 962.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.25, + 704.0, + 1179.25, + 704.0, + 1179.25, + 725.5, + 1114.25, + 725.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 687.5, + 1250.0, + 687.5, + 1250.0, + 714.5, + 1147.0, + 714.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 699.5, + 1287.0, + 699.5, + 1287.0, + 718.5, + 1215.0, + 718.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.75, + 699.0, + 1388.75, + 699.0, + 1388.75, + 715.0, + 1318.75, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 824.0, + 649.0, + 824.0, + 649.0, + 848.5, + 554.0, + 848.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.75, + 834.5, + 382.75, + 834.5, + 382.75, + 854.0, + 307.75, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.25, + 839.0, + 429.25, + 839.0, + 429.25, + 860.0, + 364.25, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 842.5, + 478.0, + 842.5, + 478.0, + 862.5, + 418.0, + 862.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.25, + 843.5, + 682.25, + 843.5, + 682.25, + 864.5, + 626.25, + 864.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 840.0, + 738.0, + 840.0, + 738.0, + 860.0, + 674.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 845.0, + 835.0, + 845.0, + 835.0, + 866.5, + 782.0, + 866.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 845.0, + 918.0, + 845.0, + 918.0, + 867.0, + 855.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 849.5, + 965.0, + 849.5, + 965.0, + 875.0, + 907.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 831.0, + 1025.0, + 831.0, + 1025.0, + 851.0, + 935.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 846.0, + 1153.0, + 846.0, + 1153.0, + 868.0, + 1092.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 847.0, + 1290.0, + 847.0, + 1290.0, + 868.0, + 1234.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 844.5, + 1341.0, + 844.5, + 1341.0, + 866.0, + 1278.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.25, + 847.0, + 1200.25, + 847.0, + 1200.25, + 869.0, + 1137.25, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.25, + 843.5, + 1250.25, + 843.5, + 1250.25, + 864.0, + 1184.25, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 971.5, + 428.0, + 971.5, + 428.0, + 993.0, + 361.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 964.0, + 489.0, + 964.0, + 489.0, + 982.0, + 405.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 978.0, + 527.0, + 978.0, + 527.0, + 999.0, + 471.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.75, + 980.0, + 572.75, + 980.0, + 572.75, + 999.0, + 526.75, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 964.5, + 643.0, + 964.5, + 643.0, + 985.0, + 560.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.75, + 979.0, + 731.75, + 979.0, + 731.75, + 996.0, + 679.75, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 981.5, + 779.0, + 981.5, + 779.0, + 999.5, + 734.0, + 999.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.25, + 971.5, + 846.25, + 971.5, + 846.25, + 995.0, + 774.25, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.75, + 985.0, + 1040.75, + 985.0, + 1040.75, + 1014.0, + 988.75, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 983.0, + 1086.0, + 983.0, + 1086.0, + 1015.0, + 1030.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 983.0, + 916.0, + 983.0, + 916.0, + 1004.0, + 853.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 982.0, + 962.0, + 982.0, + 962.0, + 1004.0, + 894.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 987.0, + 1000.0, + 987.0, + 1000.0, + 1008.5, + 945.0, + 1008.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.75, + 976.0, + 1226.75, + 976.0, + 1226.75, + 996.5, + 1146.75, + 996.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 963.0, + 1376.0, + 963.0, + 1376.0, + 1003.5, + 1242.0, + 1003.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 981.5, + 1262.0, + 981.5, + 1262.0, + 1000.5, + 1195.0, + 1000.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 277.0, + 385.0, + 277.0, + 385.0, + 307.0, + 354.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 277.0, + 426.0, + 277.0, + 426.0, + 306.0, + 395.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 277.0, + 466.0, + 277.0, + 466.0, + 302.0, + 439.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 277.0, + 509.0, + 277.0, + 509.0, + 307.0, + 479.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 279.0, + 589.0, + 279.0, + 589.0, + 299.0, + 567.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 278.0, + 678.0, + 278.0, + 678.0, + 307.0, + 644.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 278.0, + 718.0, + 278.0, + 718.0, + 307.0, + 685.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 278.0, + 958.0, + 278.0, + 958.0, + 293.0, + 943.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 275.0, + 1050.0, + 275.0, + 1050.0, + 307.0, + 1021.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 277.0, + 1092.0, + 277.0, + 1092.0, + 307.0, + 1062.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 280.0, + 1129.0, + 280.0, + 1129.0, + 295.0, + 1114.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 389.0, + 509.0, + 389.0, + 509.0, + 421.0, + 479.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 390.0, + 671.0, + 390.0, + 671.0, + 409.0, + 651.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 396.0, + 837.0, + 396.0, + 837.0, + 407.0, + 827.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 387.0, + 883.0, + 387.0, + 883.0, + 414.0, + 858.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 389.0, + 928.0, + 389.0, + 928.0, + 419.0, + 897.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 390.0, + 963.0, + 390.0, + 963.0, + 410.0, + 944.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 390.0, + 1127.0, + 390.0, + 1127.0, + 404.0, + 1110.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 389.0, + 1176.0, + 389.0, + 1176.0, + 419.0, + 1147.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.25, + 281.5, + 644.25, + 281.5, + 644.25, + 305.0, + 592.25, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 280.5, + 771.0, + 280.5, + 771.0, + 306.5, + 718.0, + 306.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.75, + 281.0, + 1016.75, + 281.0, + 1016.75, + 301.0, + 972.75, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 279.5, + 1347.0, + 279.5, + 1347.0, + 301.5, + 1309.0, + 301.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 287.0, + 352.0, + 287.0, + 352.0, + 304.0, + 304.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 283.5, + 563.0, + 283.5, + 563.0, + 304.0, + 510.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 283.5, + 807.0, + 283.5, + 807.0, + 301.5, + 763.0, + 301.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.75, + 285.0, + 851.75, + 285.0, + 851.75, + 303.5, + 803.75, + 303.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 284.0, + 895.0, + 284.0, + 895.0, + 301.5, + 844.0, + 301.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.75, + 284.5, + 1184.75, + 284.5, + 1184.75, + 302.5, + 1137.75, + 302.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 285.5, + 1232.0, + 285.5, + 1232.0, + 306.5, + 1175.0, + 306.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 287.0, + 1270.0, + 287.0, + 1270.0, + 305.5, + 1219.0, + 305.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1347.75, + 285.5, + 1393.75, + 285.5, + 1393.75, + 302.5, + 1347.75, + 302.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 300.0, + 952.0, + 300.0, + 952.0, + 318.0, + 871.0, + 318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 297.0, + 1329.0, + 297.0, + 1329.0, + 317.5, + 1246.0, + 317.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.25, + 391.0, + 790.25, + 391.0, + 790.25, + 417.0, + 779.25, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.25, + 391.5, + 430.25, + 391.5, + 430.25, + 414.5, + 388.25, + 414.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.75, + 397.5, + 349.75, + 397.5, + 349.75, + 413.0, + 303.75, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 397.0, + 475.0, + 397.0, + 475.0, + 415.0, + 428.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 396.5, + 559.0, + 396.5, + 559.0, + 414.5, + 511.0, + 414.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.75, + 397.5, + 1017.75, + 397.5, + 1017.75, + 415.0, + 969.75, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.75, + 396.5, + 1103.75, + 396.5, + 1103.75, + 416.0, + 1052.75, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 392.5, + 1226.0, + 392.5, + 1226.0, + 411.5, + 1181.0, + 411.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.25, + 396.5, + 1268.25, + 396.5, + 1268.25, + 414.5, + 1221.25, + 414.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.25, + 399.5, + 1315.25, + 399.5, + 1315.25, + 418.0, + 1258.25, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 410.5, + 406.0, + 410.5, + 406.0, + 428.0, + 331.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 406.5, + 610.0, + 406.5, + 610.0, + 423.0, + 545.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.25, + 408.5, + 659.25, + 408.5, + 659.25, + 428.5, + 578.25, + 428.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 411.0, + 742.0, + 411.0, + 742.0, + 429.0, + 665.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 410.5, + 785.0, + 410.5, + 785.0, + 429.5, + 704.0, + 429.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1268.25, + 417.0, + 1396.25, + 417.0, + 1396.25, + 445.0, + 1268.25, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 512.0, + 746.0, + 512.0, + 746.0, + 550.0, + 295.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 512.0, + 1405.0, + 512.0, + 1405.0, + 550.0, + 793.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 539.0, + 1405.0, + 539.0, + 1405.0, + 575.0, + 294.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 565.0, + 1403.0, + 565.0, + 1403.0, + 602.0, + 294.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 595.0, + 700.0, + 595.0, + 700.0, + 633.0, + 293.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 595.0, + 972.0, + 595.0, + 972.0, + 633.0, + 721.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 595.0, + 1337.0, + 595.0, + 1337.0, + 633.0, + 992.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1106.0, + 1404.0, + 1106.0, + 1404.0, + 1139.0, + 295.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1136.0, + 1206.0, + 1136.0, + 1206.0, + 1167.0, + 295.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1602.0, + 748.0, + 1602.0, + 748.0, + 1642.0, + 293.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1273.0, + 1406.0, + 1273.0, + 1406.0, + 1312.0, + 295.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1307.0, + 1403.0, + 1307.0, + 1403.0, + 1341.0, + 295.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1335.0, + 1405.0, + 1335.0, + 1405.0, + 1376.0, + 292.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1370.0, + 1405.0, + 1370.0, + 1405.0, + 1403.0, + 295.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1399.0, + 1405.0, + 1399.0, + 1405.0, + 1432.0, + 294.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1426.0, + 1406.0, + 1426.0, + 1406.0, + 1466.0, + 291.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1460.0, + 1405.0, + 1460.0, + 1405.0, + 1494.0, + 294.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1489.0, + 1353.0, + 1489.0, + 1353.0, + 1525.0, + 294.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1678.0, + 678.0, + 1678.0, + 678.0, + 1717.0, + 294.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 1678.0, + 1405.0, + 1678.0, + 1405.0, + 1717.0, + 728.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1711.0, + 651.0, + 1711.0, + 651.0, + 1746.0, + 293.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 1711.0, + 1405.0, + 1711.0, + 1405.0, + 1746.0, + 701.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1742.0, + 1405.0, + 1742.0, + 1405.0, + 1775.0, + 293.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1771.0, + 1403.0, + 1771.0, + 1403.0, + 1808.0, + 292.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1804.0, + 1404.0, + 1804.0, + 1404.0, + 1836.0, + 296.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1832.0, + 1350.0, + 1832.0, + 1350.0, + 1870.0, + 293.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1880.0, + 1404.0, + 1880.0, + 1404.0, + 1916.0, + 295.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1909.0, + 1406.0, + 1909.0, + 1406.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1945.0, + 1403.0, + 1945.0, + 1403.0, + 1974.0, + 296.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2009.0, + 293.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 498.0, + 2003.0, + 498.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 824, + 1404, + 824, + 1404, + 1040, + 298, + 1040 + ], + "score": 0.977 + }, + { + "category_id": 5, + "poly": [ + 295, + 342, + 1402, + 342, + 1402, + 689, + 295, + 689 + ], + "score": 0.969, + "html": "
Gate valuesWord tokens
Lowestor but But These these However however among Among that when When although Although because Because until many Many than though Though this This Since since date where Where have That and And Such such number so which by By how before Before with With between Between even Even if
HighestSweetgum Untersee Jianlong Floresta Chlorella Obersee PhT Doctorin Jumonville WFTS WTSP Boven Pharm Nederrijn Otrar Rhin Magicicada WBKB Tanzler KMBC WPLG Mainau Merwede RMJM Kleitman Scheur Bodensee Kromme Horenbout Vorderrhein Chlamydomonas Scantlebury Qingshui Funchess
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1128, + 1399, + 1128, + 1399, + 1191, + 299, + 1191 + ], + "score": 0.938 + }, + { + "category_id": 0, + "poly": [ + 300, + 755, + 561, + 755, + 561, + 790, + 300, + 790 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.892 + }, + { + "category_id": 6, + "poly": [ + 295, + 251, + 1400, + 251, + 1400, + 310, + 295, + 310 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 299, + 1237, + 488, + 1237, + 488, + 1269, + 299, + 1269 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 301, + 1077, + 558, + 1077, + 558, + 1105, + 301, + 1105 + ], + "score": 0.835 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.774 + }, + { + "category_id": 1, + "poly": [ + 298, + 1285, + 1401, + 1285, + 1401, + 1341, + 298, + 1341 + ], + "score": 0.773 + }, + { + "category_id": 1, + "poly": [ + 295, + 1363, + 1400, + 1363, + 1400, + 1422, + 295, + 1422 + ], + "score": 0.683 + }, + { + "category_id": 1, + "poly": [ + 298, + 1442, + 1400, + 1442, + 1400, + 1529, + 298, + 1529 + ], + "score": 0.578 + }, + { + "category_id": 1, + "poly": [ + 298, + 1549, + 1402, + 1549, + 1402, + 1636, + 298, + 1636 + ], + "score": 0.533 + }, + { + "category_id": 1, + "poly": [ + 296, + 1977, + 1403, + 1977, + 1403, + 2033, + 296, + 2033 + ], + "score": 0.486 + }, + { + "category_id": 1, + "poly": [ + 296, + 1656, + 1401, + 1656, + 1401, + 1716, + 296, + 1716 + ], + "score": 0.469 + }, + { + "category_id": 1, + "poly": [ + 300, + 1737, + 1400, + 1737, + 1400, + 1796, + 300, + 1796 + ], + "score": 0.463 + }, + { + "category_id": 1, + "poly": [ + 298, + 1896, + 1400, + 1896, + 1400, + 1955, + 298, + 1955 + ], + "score": 0.46 + }, + { + "category_id": 1, + "poly": [ + 301, + 1817, + 1400, + 1817, + 1400, + 1875, + 301, + 1875 + ], + "score": 0.444 + }, + { + "category_id": 15, + "poly": [ + 290.0, + 748.0, + 566.0, + 748.0, + 566.0, + 799.0, + 290.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 247.0, + 1402.0, + 247.0, + 1402.0, + 284.0, + 294.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 279.0, + 894.0, + 279.0, + 894.0, + 310.0, + 296.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1234.0, + 491.0, + 1234.0, + 491.0, + 1274.0, + 296.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1079.0, + 558.0, + 1079.0, + 558.0, + 1108.0, + 299.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 826.0, + 1405.0, + 826.0, + 1405.0, + 860.0, + 296.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 857.0, + 1406.0, + 857.0, + 1406.0, + 891.0, + 294.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 885.0, + 1405.0, + 885.0, + 1405.0, + 922.0, + 293.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 917.0, + 1404.0, + 917.0, + 1404.0, + 951.0, + 293.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 946.0, + 1406.0, + 946.0, + 1406.0, + 984.0, + 293.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 974.0, + 1406.0, + 974.0, + 1406.0, + 1015.0, + 292.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1007.0, + 1372.0, + 1007.0, + 1372.0, + 1044.0, + 292.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1124.0, + 1405.0, + 1124.0, + 1405.0, + 1166.0, + 294.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1157.0, + 1052.0, + 1157.0, + 1052.0, + 1194.0, + 294.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1281.0, + 1405.0, + 1281.0, + 1405.0, + 1319.0, + 294.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1312.0, + 680.0, + 1312.0, + 680.0, + 1342.0, + 323.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1359.0, + 1405.0, + 1359.0, + 1405.0, + 1400.0, + 293.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1394.0, + 685.0, + 1394.0, + 685.0, + 1421.0, + 321.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1443.0, + 1405.0, + 1443.0, + 1405.0, + 1475.0, + 294.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1474.0, + 1404.0, + 1474.0, + 1404.0, + 1503.0, + 323.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1502.0, + 644.0, + 1502.0, + 644.0, + 1528.0, + 324.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1552.0, + 1404.0, + 1552.0, + 1404.0, + 1581.0, + 296.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1578.0, + 1406.0, + 1578.0, + 1406.0, + 1611.0, + 321.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1604.0, + 392.0, + 1604.0, + 392.0, + 1640.0, + 321.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1978.0, + 1404.0, + 1978.0, + 1404.0, + 2009.0, + 295.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 2005.0, + 932.0, + 2005.0, + 932.0, + 2035.0, + 324.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1658.0, + 1403.0, + 1658.0, + 1403.0, + 1690.0, + 297.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1686.0, + 1112.0, + 1686.0, + 1112.0, + 1718.0, + 321.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1738.0, + 1404.0, + 1738.0, + 1404.0, + 1769.0, + 297.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1767.0, + 1065.0, + 1767.0, + 1065.0, + 1798.0, + 321.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1896.0, + 1404.0, + 1896.0, + 1404.0, + 1932.0, + 294.0, + 1932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1926.0, + 1288.0, + 1926.0, + 1288.0, + 1957.0, + 323.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1816.0, + 1404.0, + 1816.0, + 1404.0, + 1851.0, + 295.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1845.0, + 1095.0, + 1845.0, + 1095.0, + 1876.0, + 323.0, + 1876.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 866, + 2088, + 866, + 2113, + 836, + 2113 + ], + "score": 0.838 + }, + { + "category_id": 1, + "poly": [ + 297, + 230, + 1401, + 230, + 1401, + 290, + 297, + 290 + ], + "score": 0.759 + }, + { + "category_id": 1, + "poly": [ + 295, + 1415, + 1400, + 1415, + 1400, + 1475, + 295, + 1475 + ], + "score": 0.713 + }, + { + "category_id": 1, + "poly": [ + 300, + 1132, + 1325, + 1132, + 1325, + 1165, + 300, + 1165 + ], + "score": 0.711 + }, + { + "category_id": 1, + "poly": [ + 294, + 1055, + 1398, + 1055, + 1398, + 1114, + 294, + 1114 + ], + "score": 0.7 + }, + { + "category_id": 1, + "poly": [ + 294, + 646, + 1397, + 646, + 1397, + 679, + 294, + 679 + ], + "score": 0.695 + }, + { + "category_id": 1, + "poly": [ + 298, + 1261, + 1398, + 1261, + 1398, + 1319, + 298, + 1319 + ], + "score": 0.693 + }, + { + "category_id": 1, + "poly": [ + 296, + 1338, + 1400, + 1338, + 1400, + 1398, + 296, + 1398 + ], + "score": 0.693 + }, + { + "category_id": 1, + "poly": [ + 295, + 977, + 1401, + 977, + 1401, + 1038, + 295, + 1038 + ], + "score": 0.681 + }, + { + "category_id": 1, + "poly": [ + 296, + 1493, + 1400, + 1493, + 1400, + 1553, + 296, + 1553 + ], + "score": 0.68 + }, + { + "category_id": 1, + "poly": [ + 299, + 309, + 1400, + 309, + 1400, + 367, + 299, + 367 + ], + "score": 0.679 + }, + { + "category_id": 1, + "poly": [ + 300, + 746, + 1400, + 746, + 1400, + 805, + 300, + 805 + ], + "score": 0.678 + }, + { + "category_id": 1, + "poly": [ + 297, + 1570, + 1403, + 1570, + 1403, + 1631, + 297, + 1631 + ], + "score": 0.67 + }, + { + "category_id": 1, + "poly": [ + 298, + 1183, + 1401, + 1183, + 1401, + 1243, + 298, + 1243 + ], + "score": 0.665 + }, + { + "category_id": 1, + "poly": [ + 299, + 385, + 1401, + 385, + 1401, + 445, + 299, + 445 + ], + "score": 0.657 + }, + { + "category_id": 1, + "poly": [ + 298, + 901, + 1398, + 901, + 1398, + 960, + 298, + 960 + ], + "score": 0.643 + }, + { + "category_id": 1, + "poly": [ + 297, + 696, + 1386, + 696, + 1386, + 729, + 297, + 729 + ], + "score": 0.632 + }, + { + "category_id": 1, + "poly": [ + 298, + 822, + 1400, + 822, + 1400, + 882, + 298, + 882 + ], + "score": 0.626 + }, + { + "category_id": 1, + "poly": [ + 299, + 463, + 1400, + 463, + 1400, + 550, + 299, + 550 + ], + "score": 0.604 + }, + { + "category_id": 1, + "poly": [ + 298, + 568, + 1401, + 568, + 1401, + 628, + 298, + 628 + ], + "score": 0.603 + }, + { + "category_id": 13, + "poly": [ + 1075, + 825, + 1175, + 825, + 1175, + 852, + 1075, + 852 + ], + "score": 0.73, + "latex": "^ { 1 0 0 , 0 0 0 + }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 1403.0, + 230.0, + 1403.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 258.0, + 392.0, + 258.0, + 392.0, + 291.0, + 320.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1411.0, + 1405.0, + 1411.0, + 1405.0, + 1454.0, + 292.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1445.0, + 827.0, + 1445.0, + 827.0, + 1475.0, + 323.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1132.0, + 1323.0, + 1132.0, + 1323.0, + 1167.0, + 296.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1053.0, + 1403.0, + 1053.0, + 1403.0, + 1091.0, + 294.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1085.0, + 617.0, + 1085.0, + 617.0, + 1114.0, + 322.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 643.0, + 1402.0, + 643.0, + 1402.0, + 682.0, + 297.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1260.0, + 1402.0, + 1260.0, + 1402.0, + 1295.0, + 297.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1289.0, + 878.0, + 1289.0, + 878.0, + 1320.0, + 323.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1339.0, + 1402.0, + 1339.0, + 1402.0, + 1370.0, + 296.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1367.0, + 1212.0, + 1367.0, + 1212.0, + 1399.0, + 321.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 977.0, + 1405.0, + 977.0, + 1405.0, + 1013.0, + 297.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1009.0, + 819.0, + 1009.0, + 819.0, + 1036.0, + 322.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1490.0, + 1404.0, + 1490.0, + 1404.0, + 1527.0, + 294.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1524.0, + 612.0, + 1524.0, + 612.0, + 1550.0, + 322.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 305.0, + 1405.0, + 305.0, + 1405.0, + 345.0, + 292.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 337.0, + 637.0, + 337.0, + 637.0, + 368.0, + 322.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 744.0, + 1404.0, + 744.0, + 1404.0, + 780.0, + 295.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 774.0, + 1040.0, + 774.0, + 1040.0, + 807.0, + 320.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1572.0, + 1404.0, + 1572.0, + 1404.0, + 1603.0, + 299.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1599.0, + 1138.0, + 1599.0, + 1138.0, + 1631.0, + 324.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1180.0, + 1405.0, + 1180.0, + 1405.0, + 1218.0, + 294.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1212.0, + 582.0, + 1212.0, + 582.0, + 1242.0, + 322.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 384.0, + 1405.0, + 384.0, + 1405.0, + 420.0, + 295.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 414.0, + 454.0, + 414.0, + 454.0, + 446.0, + 320.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 898.0, + 1403.0, + 898.0, + 1403.0, + 937.0, + 293.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 929.0, + 817.0, + 929.0, + 817.0, + 959.0, + 323.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 693.0, + 1370.0, + 693.0, + 1370.0, + 732.0, + 295.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 823.0, + 1074.0, + 823.0, + 1074.0, + 858.0, + 294.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 823.0, + 1404.0, + 823.0, + 1404.0, + 858.0, + 1176.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 853.0, + 746.0, + 853.0, + 746.0, + 883.0, + 323.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 466.0, + 1401.0, + 466.0, + 1401.0, + 495.0, + 298.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 494.0, + 1404.0, + 494.0, + 1404.0, + 523.0, + 324.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 523.0, + 664.0, + 523.0, + 664.0, + 548.0, + 323.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 568.0, + 1405.0, + 568.0, + 1405.0, + 603.0, + 294.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 597.0, + 788.0, + 597.0, + 788.0, + 628.0, + 322.0, + 628.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/ByxZX20qFQ/ByxZX20qFQ.md b/parse/train/ByxZX20qFQ/ByxZX20qFQ.md new file mode 100644 index 0000000000000000000000000000000000000000..9146d6df8b9e5e2e81adbf1da18d2654316becaf --- /dev/null +++ b/parse/train/ByxZX20qFQ/ByxZX20qFQ.md @@ -0,0 +1,236 @@ +# ADAPTIVE INPUT REPRESENTATIONS FOR NEURAL LANGUAGE MODELING + +Alexei Baevski & Michael Auli Facebook AI Research, Menlo Park, CA, USA + +# ABSTRACT + +We introduce adaptive input representations for neural language modeling which extend the adaptive softmax of Grave et al. (2017) to input representations of variable capacity. There are several choices on how to factorize the input and output layers, and whether to model words, characters or sub-word units. We perform a systematic comparison of popular choices for a self-attentional architecture. Our experiments show that models equipped with adaptive embeddings are more than twice as fast to train than the popular character input CNN while having a lower number of parameters. On the WIKITEXT-103 benchmark we achieve 18.7 perplexity, an improvement of 10.5 perplexity compared to the previously best published result and on the BILLION WORD benchmark, we achieve 23.02 perplexity.1 + +# 1 INTRODUCTION + +Language modeling is a basic task in natural language processing, with many applications such as speech recognition (Arisoy et al., 2012) and statistical machine translation (Schwenk et al., 2012; Vaswani et al., 2013; Baltescu & Blunsom, 2015). Recently, much progress has been made by neural methods (Bengio et al., 2003; Mikolov et al., 2010) based on LSTMs (Jozefowicz et al., 2016), gated ´ convolutional networks (Dauphin et al., 2017) and self-attentional networks (Al-Rfou et al., 2018). + +There are different choices for the basic unit we wish to model, including full words (Bengio et al., 2003), characters for the input (Kim et al., 2016), or also the output (Merity et al., 2018) as well as sub-words (Buckman & Neubig, 2018; Mielke & Eisner, 2018). Word-based models are particularly challenging since computing probabilities for all 800K words of the BILLION WORD benchmark is still a substantial part of the overall computation (Chen et al., 2016). + +A popular approach to lower the computational burden is to structure the output vocabulary so that not all probabilities need to be computed. The hierarchical softmax does this by introducing latent variables or clusters to simplify normalization (Goodman, 2001; Morin & Bengio, 2005; Mikolov et al., 2011). This has been further improved by the adaptive softmax which introduces a variable capacity scheme for output word embeddings, assigning more parameters to frequent words and fewer parameters to rare words (Grave et al., 2017). + +In this paper, we introduce adaptive input embeddings which extend the adaptive softmax to input word representations. This factorization assigns more capacity to frequent words and reduces the capacity for less frequent words with the benefit of reducing overfitting to rare words. For a competitive setup on the BILLION WORD benchmark, adaptive input embeddings reduce the number of parameters in the input and output layers by $23 \%$ while achieving higher accuracy over fixed size embeddings. When the adaptive input representations are tied with an adaptive softmax in the output, then the number of parameters is reduced by a total of $61 \%$ . + +Our experiments compare models based on word inputs, character inputs, as well as sub-word units using a self-attention architecture (Vaswani et al., 2017). We show that models with adaptive word representations can outperform very strong character-based models while training more than twice as fast. We also substantially improve adaptive softmax by introducing additional dropout regularization in the tail projection. On the WIKITEXT-103 benchmark we achieve a perplexity of 18.7, a reduction of 10.5 perplexity over the previously best published result. On the larger BILLION WORD benchmark our best model with adaptive input embeddings achieves 23.02 perplexity, a reduction of nearly 5 perplexity over the next best previously published result. + +![](images/52708bad0c295bb0584ec5d4f2259eabd3c9b29f269a819bd44546cef32e3818.jpg) +Figure 1: Illustration of adaptive input representations. Words are assigned to clusters $\nu _ { i }$ based on their frequency which determines the size of the representations. Embeddings are projected to a common dimension $d$ before being fed to the model. + +# 2 RELATED WORK + +Adaptive word representations are inspired by the adaptive softmax work Grave et al. (2017) which first described a GPU friendly way to construct a hierarchical softmax and showed that it performs very competitively compared to a full softmax, while offering significantly faster speed and a lower memory footprint. + +Merity et al. (2018) use a modified version of adaptive softmax which does not reduce the dimensionality of less frequent words in order to be able to share output embeddings with the input. This setup is akin to a hierarchical softmax with tied weights. We show that variable-sized input embeddings can perform better than fixed sized embeddings. Furthermore, this also enables weight sharing with an adaptive softmax output layer. + +Merity et al. (2018) evaluates both character-based and word-based factorizations but does not directly compare them to each other. We perform a direct comparison of word-based and characterbased input vocabularies and also compare to a sub-word factorization for both the input and output. Recently, Al-Rfou et al. (2018) demonstrated that self-attentional models can perform very well on language modeling tasks where the input and output is both characters. We also consider word-based benchmarks. + +# 3 ADAPTIVE INPUT REPRESENTATIONS + +The adaptive softmax exploits the fact that the distribution of word types in natural language follows a Zipfian distribution in order to improve the computation of the output probabilities. We apply the same intuition for input word embeddings with the motivation to reduce the number of parameters which frees up capacity for other parts of the model. + +We define a number of clusters that partitions the frequency ordered vocabulary $\nu ~ = ~ \nu _ { 1 } \cup$ $\mathcal { V } _ { 2 } , \ldots , \mathcal { V } _ { n - 1 } \cup \mathcal { V } _ { n }$ such that $\nu _ { i } \cap \nu _ { j } = \emptyset$ for $\forall i , j$ , and $i \neq j$ , where $\nu _ { 1 }$ contains the most frequent words and $\nu _ { n }$ the least frequent words. We will refer to $\nu _ { 1 }$ as the head and to any subsequent clusters loosely as tail. We reduce the capacity for each cluster by a factor of $k$ . That is, if words in $\nu _ { 1 }$ have dimension $d$ , then words in $\nu _ { n }$ have dimension $\frac { d } { k ^ { n - 1 } }$ . We typically set $k = 4$ following Grave et al. (2017). + +Next, we add linear projections $W _ { 1 } \in \mathbb { R } ^ { d \times d } , \ldots , W _ { n } \in \mathbb { R } ^ { d / k ^ { n - 1 } \times d }$ to map the embeddings of each cluster to dimension $d$ so that the concatenated output of the adaptive input embedding layer can be easily used by the subsequent model (Figure 1). We also project $\nu _ { 1 }$ which already has dimension $d$ . + +When presented with a number of input words, the adaptive input embedding layer partitions the words into the various clusters, performs separate lookups in the embedding tables and then projects to dimension $d$ , followed by concatenating the embeddings in the original order. + +Weight sharing. When the output layer is an adaptive softmax with the same partition of $\nu , d ,$ , and $k$ as the adaptive input layer, then we can tie the weights (Inan et al., 2016; Press & Wolf, 2017). This further reduces the number of parameters and can simultaneously improve performance (§5). We can share both the parameters for the actual words as well as the projections $W _ { 1 } , \ldots , W _ { n }$ Sharing the word embeddings is straightforward except for the head where the adaptive softmax has $n - 1$ additional embeddings for the remaining clusters which are not shared with the input. + +We share all projections, except for the head projection which is not available in the adaptive softmax since the model output is directly multiplied with the output word embeddings for the head band. Performance decreased when we added a head projection to the adaptive softmax in the output, regardless of when it was shared or not. Sharing both the word embeddings as well as the projections performed very well on WIKITEXT-103 but on BILLION WORD we only share the word embeddings as we found that this performed better on the validation set. + +# 4 EXPERIMENTAL SETUP + +# 4.1 MODEL + +We follow most of the architectural choices described in Vaswani et al. (2017) but use only a decoder network. We add sinusoidal position embeddings to the input layer and stack $N = 1 6$ blocks for both BILLION WORD and WIKITEXT-103. Each block contains two sub-blocks: the first is a multihead self-attention module with $H = 1 6$ heads. The second sub-block is a feed-forward module (FFN) of the form $R e L U ( W _ { 1 } X + b _ { 1 } ) W _ { 2 } + b _ { 2 }$ where $W _ { 1 } \in \mathbb { R } ^ { e \times e _ { f f } }$ , $W _ { 1 } \in \mathbb { R } ^ { e _ { f f } \times e }$ and $e = 1 0 2 4$ , $e _ { f f } = 4 0 9 6$ unless otherwise stated. Different to Vaswani et al. (2017) we apply layer normalization before the self-attention and FFN blocks instead of after, as we find it leads to more effective training. Sub-blocks are surrounded by a residual connection (He et al., 2015). + +We use a dropout rate of 0.1 and attention dropout of 0.1 for BILLION WORD models, and increase regularization for WIKITEXT-103 by using dropout 0.3, and 0.1 ReLU dropout as well as attention dropout 0.1. We use the same hyperparameters for all models trained on the same dataset in order to enable a like for like comparison. When the dimensionality of the input or output layer differs from $e$ , then we add a simple linear projection with no bias. + +# 4.2 DATASETS + +We experiment on the BILLION WORD benchmark and WIKITEXT-103. BILLION WORD contains 768M word tokens and has a vocabulary of about 800K word types, which corresponds to words with more than 3 occurrences in the training set (Chelba et al., 2013). + +The training data of WIKITEXT-103 comprises about 100M tokens and a vocabulary of around 260K, corresponding to types with more than 3 occurrences in the training data (Merity et al., 2016). The dataset is composed of shuffled Wikipedia articles where the context carries across sentences. + +# 4.3 BATCHING + +For BILLION WORD we batch individual sentences since the corpus does not contain document structure. For WIKITEXT-103 we partition the training data into blocks of 512 contiguous tokens ignoring document boundaries. Evaluation is the same except that we require blocks to contain complete sentences totaling up to 512 tokens.2 + +We limit the number of tokens per GPU to a maximum threshold $B$ per GPU. That is, we add examples of similar length until we reach this threshold. When we train on multiple GPUs, each GPU processes $B$ tokens using the same model parameters. This increases the effective batch size to the product of the number of GPUs and $B$ . For BILLION WORD models we use $B = 2 0 4 8$ and typically train on 32 GPUs, giving an effective batch size of 65K tokens. The smaller vocabulary of WIKITEXT-103 enables increasing $B$ to 4096 and we train on 8 GPUs. We found that large batch training is beneficial for this dataset and we therefore accumulate gradient updates over two batches before committing a parameter update (Ott et al., 2018a). This gives an effective batch size of 65K tokens for WIKITEXT-103. + +# 4.4 INPUT AND OUTPUT LAYER HYPERPARAMETERS + +Embedding sizes. For fixed size word input layers and softmax output layers we generally use embeddings of size 512 for WIKITEXT-103. When we use an adaptive softmax in the output and fixed size word embeddings for the input, then we use dimension 256 for the input embeddings for BILLION WORD and 64 for WIKITEXT-103. We tuned this choice on the validation set (Appendix A). BPE inputs and outputs have embeddings of size 1024. + +Character CNN. We model character inputs by convolving the representations of all characters in a word following Kim et al. (2015) which applies several filters, then max pooling, a number of highway layers and a projection. Character embeddings have size 128 and we apply seven filters of size 1x128, $2 \mathrm { x } 2 5 6$ , 3x384, 4x512, 5x512, 6x512, 7x512, where $3 \mathrm { x } 1 2 8$ indicates a filter processing three characters that outputs 128 features. We use a single highway layer for WIKITEXT-103, and two for BILLION WORD. We do not add start of word and end of word markers as they did not improve validation accuracy. We train on the same pre-processed data as the other models, with unknown tokens in both the inputs and outputs. + +Adaptive input representations and adaptive softmax. We use an adaptive softmax output layer to train models with large word-based vocabularies. For adaptive word inputs and adaptive softmax, we use embeddings of size $d = 1 0 2 4$ for the head and reduce the size of subsequent clusters by a factor of $k = 4$ . For WIKITEXT-103, we have three bands of size 20K $\scriptstyle ( \mathrm { d } = 1 0 2 4 )$ ), 40K $\mathrm { \Delta } \mathrm { Q } \mathrm { = } 2 5 6 )$ and 200K $\scriptstyle \mathrm { { ( d = 6 4 ) } }$ ). For BILLION WORD the bands are 60K $\scriptstyle \mathrm { d } = 1 0 2 4$ ), 100K $\mathrm { \Delta } \mathrm { Q } { = } 2 5 6 \mathrm { \Delta }$ ), and 640K $\scriptstyle \mathrm { { ( d = 6 4 ) } }$ ). + +Sub-word models. We learn a byte-pair encoding (BPE) of 32K codes on the training data of each benchmark (Sennrich et al., 2016). After applying the code to the training data we obtain a vocabulary of 33,337 tokens for WIKITEXT-103 and 32,347 tokens for BILLION WORD. BPE input/output embeddings have size 1024. The final evaluation is in terms word-level perplexity to be comparable to other models. The probability of a word is the product of the sub-word units. + +# 4.5 OPTIMIZATION + +Different to Vaswani et al. (2017) we use Nesterov’s accelerated gradient method (Sutskever et al., 2013) with a momentum of 0.99 and we renormalize gradients if their norm exceeds 0.1 (Pascanu et al., 2013). The learning rate is linearly warmed up from $1 0 ^ { - 7 }$ to 1 for 16K steps and then annealed using a cosine learning rate schedule with $C$ cycles (Loshchilov & Hutter, 2016). Each cycle runs for twice the number of updates than the previous cycle and we lower the maximum and minimum learning rates by a rate $M$ compared to the previous cycle. The initial minimum learning rate is $1 0 ^ { - 5 }$ and the maximum is 1. + +BILLION WORD models train for a total of 975K updates over $C = 3$ cycles, the first cycle takes 137K steps, and we set $M = 0 . 6$ . The WIKITEXT-103 models train for 286K steps over $C = 4$ cycles, the first cycle takes 18K setps and we set $M \ : = \ : 0 . 7 5$ . We run experiments on DGX-1 machines with 8 NVIDIA V100 GPUs and machines are interconnected by Infiniband. We also use the NCCL2 library and the torch.distributed package for inter-GPU communication. We train models with 16-bit floating point precision, following Ott et al. (2018b). + +Table 1: Test perplexity on BILLION WORD. Adaptive inputs share parameters with an adaptive softmax. Training times of Char-CNN and Adaptive input models are measured when training with 64 GPUs. †does not include embedding and softmax layers + +
TestTrain Time (hours)Parameters
Dauphin et al. (2017)31.9428M
J6zefowicz et al. (2016)30.01,040M
Shazeer et al. (2017)28.0-4,371Mt
Char-CNN25.8879366M
Adaptive inputs25.2255331M
Adaptive inputs (large)23.9172465M
Adaptive inputs (very large)23.021451026M
10 LSTMs + SNM10-SKIP (Shazeer et al.,2016)23.71=
+ +
TestTrain Time (hours)Parameters
Grave et al. (2016)40.8
Dauphin et al. (2017)37.2229M
Merity et al. (2018)33.0=151M
Rae et al. (2018)29.2-
Adaptive inputs18.767247M
+ +Table 2: Test perplexity on WIKITEXT-103 (cf. Table 1). Training time is based on 8 GPUs. + +# 5 EXPERIMENTS AND RESULTS + +# 5.1 MAIN RESULTS + +For the main results on BILLION WORD, we doubled the batch size by training on 64 GPUs instead of 32 GPUs. We also consider two larger setups, one where we added four more blocks $N = 2 0$ ) and increased the FFN dimension to $e _ { f f } = 6 1 4 4$ (large), and another where we add another four blocks $N = 2 4$ ) with $e _ { f f } = 8 1 9 2$ and $e = 1 5 3 6$ (very large). All other settings follow $\ S 4 . 4$ and all models were trained for the same number of steps. + +Table 1 compares our models to previous work on BILLION WORD. The adaptive input model outperforms the best previously reported result at an order of magnitude fewer parameters. Our large model performs nearly as well as an ensemble of over ten models and achieves a new state of the art of 24.14 perplexity. Our very large model performs as well as an ensemble of over ten models and achieves 23.02 perplexity. The Char-CNN model performs $0 . 6 ~ \mathrm { P P L }$ worse than the standard adaptive input model even though it trained for over $40 \%$ longer. + +Table 2 shows our result on WIKITEXT-103 where adaptive inputs achieve 18.7 perplexity. For this result only, we partition the training data into blocks of 3072 contiguous tokens instead of 512 tokens as for other experiments. During evaluation we require blocks to contain complete sentences totaling up to 3072 tokens of which the first 2560 tokens serve as context to score the last 512 tokens; we take care to score all tokens in the test and validation sets. We motivate this choice in $\ S 5 . 3$ . + +# .2 COMPARISON OF INPUT AND OUTPUT LAYER FACTORIZATIONS + +Next, we perform a systematic comparison of different input and output layer factorizations. We consider a word-based setup with fixed size word input embeddings and a standard word softmax (SM) where embeddings have either dimension 512 (WIKITEXT-103) or 64 (BILLION WORD). We consider tying the input and output embeddings (SM-T). Instead of words, we try less sparse subword units, both in the input and output, with embeddings of size 1024 (BPE) and shared weights (BPE-T). Next, we consider replacing the fixed size output representations by an adaptive softmax (ASM) and characters as input (CNN). Finally, we use both adaptive input word representations as well as an adaptive softmax (ADP) and a tied version (ADP-T). All models use the same selfattention architecture described in $\ S 4 . 1$ . + +Table 3: Test perplexity on WIKITEXT-103 for various input and output layer factorizations. Training speed was measured on a single 8-GPU machine. $( ^ { * } )$ indicates a modified training regime because of large memory requirements: the maximum number of tokens per GPU was lowered to 1024 from 4096 but the same number of updates were performed by processing four batches before committing a weight update. + +
InputOutputValidTestTrain Time (hours)Params
SMEmbeddingSoftmax23.8724.9257*476.8M
BPEBPE EmbeddingBPE Softmax23.1324.2530270M
BPE-TBPE EmbeddingBPE Softmax (tied)22.4623.4530235.7M
SM-TEmbeddingSoftmax (tied)22.6323.3856*339.7M
ASMEmbeddingAdaptive21.2322.1835263.1M
CNNChar-CNNAdaptive20.8621.7970266.3M
ADPAdaptiveAdaptive20.9521.7434291.3M
ADP-TAdaptiveAdaptive (tied)19.7920.5130246.9M
+ +
InputOutputValidTestTrain time (hours)Params
BPE-TBPE EmbeddingBPE Softmax (shared)27.4427.5134234.7M
BPEBPE EmbeddingBPE Softmax27.0227.1335267.8M
ASMEmbeddingAdaptive26.9727.0662532.8M
CNNChar-CNNAdaptive26.1326.2592365.8M
ADPAdaptiveAdaptive26.3826.4965458.4M
ADP-TAdaptiveAdaptive (shared)25.5125.5843330.8M
+ +Table 4: Test perplexity on BILLION WORD. Training speed measured on four 8-GPU machines. + +Table 3 shows results when training all configurations for the same number of updates. Adaptive input representations with tied input and output layers (ADP-T) achieve the highest accuracy at the same speed as the BPE models which have a very small vocabulary (33K versus 260K). CNN is 1 perplexity worse than ADP-T and requires well over twice the training time. It is the slowest approach, even though it has a fast adaptive softmax in the output. Fixed word embeddings perform least well (SM). Sub-word units are fast to train and perform better than word models with fixed sized embeddings. ASM improves over SM and greatly speeds up training. For ASM, we found that reducing the dimension of the input word embeddings to 64 on WIKITEXT-103 results in better accuracy (Appendix A). + +Table 4 shows that adaptive input representations perform equally well on BILLION WORD compared to other factorizations. ADP-T is $34 \%$ faster than ADP because there are fewer parameters to update. Similar to before, ADP-T trains more than twice as fast as CNN at higher accuracy, however, the accuracy gap is narrower than for WIKITEXT-103. Regularization is more important on WIKITEXT-103 while models for BILLION WORD benefit from additional capacity. Because of this we used input word embeddings of size 256 for ASM. + +We also trained CNN without replacing input words outside the vocabulary by an unknown symbol, however, this only improved validation perplexity by 0.16. + +![](images/4b396896fe5d317f22b0248271298f7ce8c2e521c3dc122395301b4c3ec88349.jpg) +Figure 2: Loss of models binned by word frequency on the test set of WIKITEXT-103. Bins are not cumulative. + +![](images/373e2f18c0f9c2f71d23422298631e38730245add24d65b6f615d793b9273992.jpg) +Figure 3: Loss of models when binning by the frequency of the previous word measured on WIKITEXT-103 (cf. Figure 2). + +# 5.3 ANALYSIS + +Next, we turn to the question of how well models perform on rare words compared to frequent words. We compute the average loss for each word in the test set and group words by frequency. + +Figure 2 shows results on WIKITEXT-103. Tying weights helps all models on rare words, likely because of regularization effects. Fixed size word embeddings with a word softmax (SM and SMT) do not perform well on rare words. This is likely due to underfitting on common words and we use the largest possible embedding size we could fit on 16GB GPU cards given our batch size (more experimentation in Appendix A). BPE and BPE-T perform poorly on rare words because probabilities are a product of several sub-word units. ADP-T performs best across all frequency ranges. Figure 3 bins the loss by the frequency of the previous word and shows that CNN does well when it has rare words in the context, however, ADP-T does best across all bins. + +Figure 4 shows an equivalent analysis for BILLION WORD. The largest differences between models is on rare words. CNN performs best on very rare words but is outperformed by ADP in all other settings. Similar to WIKITEXT-103, BPE and BPE-T perform poorly on rare words. Further + +![](images/d73dcb6e7e5c4af036340a7099e37fe1418673e98a1f8915cde2d051627402b8.jpg) +Figure 4: Loss of models when binning by word frequency on the test set of BILLION WORD. Bins are not cumulative. + +
Train block sizeInference context sizeValidation perplexityTest perplexity
51219.7920.51
5120 48018.3519.03
2048018.9619.53
2048153618.2318.88
3072018.6319.34
3072256017.9718.70
+ +Table 5: Perplexity on WIKITEXT-103 with different context sizes during training and inference. Training block size is the number of consecutive tokens considered during training. Inference context is the number of tokens provided at evaluation before scoring tokens. + +analysis (Appendix 5.3) binning the loss by the frequency of the previous word shows that weight sharing also helps for BILLION WORD and that CNN does very well on rare words for BILLION WORD compared to other models. + +Table 5 shows the importance of context size for WIKITEXT-103. Training block size is the number of consecutive tokens that are considered at once during training. Inference context is the number of tokens that are provided at evaluation before any tokens are scored. Simply increasing the training block size from 512 to 3072 results in an improvement of nearly 1.2 perplexity with no inference context window. Increasing the context size at inference time results in an improvement of 0.6 perplexity for the largest training block size. + +# 5.4 ADAPTIVE SOFTMAX VS. FULL SOFTMAX + +We also found that adaptive softmax can benefit from additional regularization of rare words. Adaptive softmax first projects the model output to the dimension of a particular cluster and then computes a dot product with the respective word embeddings. We add dropout to the output of the first projection for all clusters, except for the head. This change enables the adaptive softmax to outperform a standard softmax over fixed size output word embeddings on WIKITEXT-103 (Table 6). + +However, we found that adding dropout in this way is not helpful for larger datasets such as BILLION WORD. Unfortunately, a standard softmax over 800K words is not tractable and we were unable to make a comparison. It may be possible to achieve better results by tuning dropout for each band of the tail and we leave this for future work. + +Table 6: Perplexity on WIKITEXT-103 when regularizing rare words in adaptive softmax. + +
Tail dropoutValidation perplexity
Softmax (SM)N/A23.87
Adaptive (ADP)0.024.74
Adaptive (ADP)0.221.23
+ +# 6 CONCLUSION + +Adaptive input embeddings vary the size of input word embeddings which can improve accuracy while drastically reducing the number of model parameters. When sharing parameters with an adaptive softmax, the number of parameters can be further reduced which improves training speed. We presented a comparison between different input and output layer factorizations including word inputs, character inputs and sub-word units in both the input and output. + +Our experiments show that models with adaptive input embeddings train faster compared to character input CNNs while achieving higher accuracy. We achieve new state of the art results on WIKITEXT-103 and BILLION WORD. In future work, we will apply variable sized input embeddings to other tasks. + +# ACKNOWLEDGMENTS + +We thank Tom Bosc for fruitful comments and suggestions. + +# REFERENCES + +Rami Al-Rfou, Dokook Choe, Noah Constant, Mandy Guo, and Llion Jones. Character-level language modeling with deeper self-attention. arXiv, abs/1808.04444, 2018. + +Ebru Arisoy, Tara N. Sainath, Brian Kingsbury, and Bhuvana Ramabhadran. Deep Neural Network Language Models. In NAACL-HLT Workshop on the Future of Language Modeling for HLT, 2012. + +Paul Baltescu and Phil Blunsom. Pragmatic neural language modelling in machine translation. In Proc. of ACL, 2015. + +Yoshua Bengio, Rejean Ducharme, Pascal Vincent, and Christian Jauvin. A Neural Probabilistic´ Language Model. JMLR, 3:1137–1155, 2003. + +Jacob Buckman and Graham Neubig. Neural lattice language models. TACL, 6:529–541, 2018. + +Ciprian Chelba, Tomas Mikolov, Mike Schuster, Qi Ge, Thorsten Brants, Phillipp Koehn, and Tony Robinson. One billion word benchmark for measuring progress in statistical language modeling. Technical report, Google, 2013. + +Wenlin Chen, David Grangier, and Michael Auli. Strategies for training large vocabulary neural language models. In Proc. of ACL, 2016. + +Yann N. Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated convolutional networks. In Proc. of ICML, 2017. + +Joshua Goodman. Classes for Fast Maximum Entropy Training. In Proc. of ICASSP, 2001. + +Edouard Grave, Armand Joulin, and Nicolas Usunier. Improving neural language models with a continuous cache. arXiv, abs/1612.04426, 2016. + +Edouard Grave, Armand Joulin, Moustapha Cisse, David Grangier, and Herv ´ e J ´ egou. Efficient ´ softmax approximation for gpus. In Proc. of ICML, 2017. + +Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proc. of CVPR, 2015. +Hakan Inan, Khashayar Khosravi, and Richard Socher. Tying word vectors and word classifiers: A loss framework for language modeling. arXiv, abs/1611.01462, 2016. +Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring the ´ limits of language modeling. arXiv, abs/1602.02410, 2016. +Yoon Kim, Yacine Jernite, David Sontag, and Alexander M. Rush. Character-aware neural language models. arXiv, abs/1508.06615, 2015. +Yoon Kim, Yacine Jernite, David Sontag, and Alexander M Rush. Character-aware neural language models. In AAAI, pp. 2741–2749, 2016. +Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with restarts. arXiv, abs/1608.03983, 2016. +Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv, abs/1609.07843, 2016. +Stephen Merity, Nitish Shirish Keskar, and Richard Socher. An analysis of neural language modeling at multiple scales. arXiv, abs/1803.08240, 2018. +Sebastian J. Mielke and Jason Eisner. Spell once, summon anywhere: A two-level open-vocabulary language model. arXiv, abs/1804.08205, 2018. +Toma´s Mikolov, Karafi ˇ at Martin, Luk ´ a´s Burget, Jan Cernock ˇ y, and Sanjeev Khudanpur. Recurrent ´ Neural Network based Language Model. In Proc. of Interspeech, 2010. +Toma´s Mikolov, Stefan Kombrink, Luk ˇ a´s Burget, Jan Cernock ˇ y, and Sanjeev Khudanpur. Exten-´ sions of Recurrent Neural Network Language Model. In Proc. of ICASSP, pp. 5528–5531, 2011. +Frederic Morin and Yoshua Bengio. Hierarchical Probabilistic Neural Network Language Model. In Proc. of AISTATS, 2005. +Myle Ott, Michael Auli, David Grangier, and MarcAurelio Ranzato. Analyzing uncertainty in neural machine translation. In Proc. of ICML, 2018a. +Myle Ott, Sergey Edunov, David Grangier, and Michael Auli. Scaling neural machine translation. In Proc. of WMT, 2018b. +Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural networks. In Proc. of ICML, 2013. +Ofir Press and Lior Wolf. Using the output embedding to improve language models. In Proc. of EACL, 2017. +Jack W. Rae, Chris Dyer, Peter Dayan, and Timothy P. Lillicrap. Fast parametric learning with activation memorization. arXiv, abs/1803.10049, 2018. +Holger Schwenk, Anthony Rousseau, and Mohammed Attik. Large, Pruned or Continuous Space Language Models on a GPU for Statistical Machine Translation. In NAACL-HLT Workshop on the Future of Language Modeling for HLT, 2012. +Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. In Proc. of ACL, 2016. +Noam Shazeer, Joris Pelemans, and Ciprian Chelba. Sparse non-negative matrix language modeling. In Proc. of Interspeech, 2016. +Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc V. Le, Geoffrey E. Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv, abs/1701.06538, 2017. + +Ilya Sutskever, James Martens, George E. Dahl, and Geoffrey E. Hinton. On the importance of initialization and momentum in deep learning. In Proc. of ICML, 2013. + +Ashish Vaswani, Yinggong Zhao, Victoria Fossum, and David Chiang. Decoding with Large-scale Neural Language Models improves Translation. In Proc. of EMNLP, October 2013. + +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. In Proc. of NIPS, 2017. + +SUPPLEMENTARY MATERIAL + +# A ADDITIONAL EXPERIMENTS ON WIKITEXT-103 + +This appendix shows various ablation. Table 7 shows that reducing the capacity of fixed size word input embddings is beneficial on WIKITEXT-103. The next set of results in Table 7 shows results for various settings of the SM and SM-T models. We also experimented with sharing the head projection but found this to perform less well than not sharing it. Finally, Table 8 shows various band sizes for adaptive input word embbedings. + +We also show the performance of Merity et al. (2018) who use an adaptive softmax with equally sized word representations and share the input and output embeddings (no dim reduction, tied). + +Table 7: Validation perplexity of our models on WIKITEXT-103. + +
Input OutputDropout Valid PPLParameters
256d EmbeddingAdaptive0.323.39314.7M
128d EmbeddingAdaptive0.321.51280.3M
64d EmbeddingAdaptive0.321.23263.1M
32d EmbeddingAdaptive0.321.78254.5M
512d Embedding512d Softmax (tied)0.322.63339.7M
512d Embedding512d Softmax (tied)0.428.31339.7M
512d Embedding512d Softmax0.323.87476.8
512d Embedding512d Softmax0.427.64476.8
256d Embedding256d Softmax (tied)0.322.65270.6M
256d Embedding256d Softmax0.324.13339.1M
64d Embedding512d Softmax0.324.74356.3M
AdaptiveAdaptive (tied emb, not proj)0.320.06247.3M
AdaptiveAdaptive (tied emb/proj not head)0.319.79246.9M
AdaptiveAdaptive (tied emb/proj + head)0.320.06246.9M
512d Embedding512d Softmax (tied)0.322.63339.7M
512d Embedding 512d Adaptive (no dim reduction, tied)0.325.48340.2M
+ +
Softmax cutoffValid PPL
20k/40k/200k19.79
20k/140k/100k 20k/40k/60k/140k20.26
60k/100k/100k20.53 20.52
5k/155k/100k
20k/40k/200k20.06
10k/60k/190k19.99 19.79
+ +Table 8: Validation perplexity on WIKITEXT-103 with tied adaptive inputs & outputs. The bands signify the number of words belonging to each band. In every case, the first band has dimension 1024, the second band 256, the third band 64 and the fourth band (if it exists) 16. + +# B ANALYSIS + +This appendix extends the analysis in $\ S 5 . 3$ by showing a breakdown of the test loss when binning by the frequency of the previous word. + +![](images/62e4d7a047372091139f4e3c0bc67293ce7fc2b1aef9dae0b74901d294ce4ff9.jpg) +Figure 5: Loss of models when binning by the frequency of the previous word measured on BILLION WORD (cf. Figure 3). \ No newline at end of file diff --git a/parse/train/ByxZX20qFQ/ByxZX20qFQ_content_list.json b/parse/train/ByxZX20qFQ/ByxZX20qFQ_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8e48fe42390610ba6375b254f10a5a5be056d1b0 --- /dev/null +++ b/parse/train/ByxZX20qFQ/ByxZX20qFQ_content_list.json @@ -0,0 +1,1201 @@ +[ + { + "type": "text", + "text": "ADAPTIVE INPUT REPRESENTATIONS FOR NEURAL LANGUAGE MODELING ", + "text_level": 1, + "bbox": [ + 174, + 99, + 620, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alexei Baevski & Michael Auli Facebook AI Research, Menlo Park, CA, USA ", + "bbox": [ + 184, + 170, + 400, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 250, + 544, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce adaptive input representations for neural language modeling which extend the adaptive softmax of Grave et al. (2017) to input representations of variable capacity. There are several choices on how to factorize the input and output layers, and whether to model words, characters or sub-word units. We perform a systematic comparison of popular choices for a self-attentional architecture. Our experiments show that models equipped with adaptive embeddings are more than twice as fast to train than the popular character input CNN while having a lower number of parameters. On the WIKITEXT-103 benchmark we achieve 18.7 perplexity, an improvement of 10.5 perplexity compared to the previously best published result and on the BILLION WORD benchmark, we achieve 23.02 perplexity.1 ", + "bbox": [ + 233, + 281, + 764, + 420 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 446, + 336, + 463 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Language modeling is a basic task in natural language processing, with many applications such as speech recognition (Arisoy et al., 2012) and statistical machine translation (Schwenk et al., 2012; Vaswani et al., 2013; Baltescu & Blunsom, 2015). Recently, much progress has been made by neural methods (Bengio et al., 2003; Mikolov et al., 2010) based on LSTMs (Jozefowicz et al., 2016), gated ´ convolutional networks (Dauphin et al., 2017) and self-attentional networks (Al-Rfou et al., 2018). ", + "bbox": [ + 174, + 479, + 825, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are different choices for the basic unit we wish to model, including full words (Bengio et al., 2003), characters for the input (Kim et al., 2016), or also the output (Merity et al., 2018) as well as sub-words (Buckman & Neubig, 2018; Mielke & Eisner, 2018). Word-based models are particularly challenging since computing probabilities for all 800K words of the BILLION WORD benchmark is still a substantial part of the overall computation (Chen et al., 2016). ", + "bbox": [ + 174, + 556, + 825, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A popular approach to lower the computational burden is to structure the output vocabulary so that not all probabilities need to be computed. The hierarchical softmax does this by introducing latent variables or clusters to simplify normalization (Goodman, 2001; Morin & Bengio, 2005; Mikolov et al., 2011). This has been further improved by the adaptive softmax which introduces a variable capacity scheme for output word embeddings, assigning more parameters to frequent words and fewer parameters to rare words (Grave et al., 2017). ", + "bbox": [ + 174, + 633, + 823, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce adaptive input embeddings which extend the adaptive softmax to input word representations. This factorization assigns more capacity to frequent words and reduces the capacity for less frequent words with the benefit of reducing overfitting to rare words. For a competitive setup on the BILLION WORD benchmark, adaptive input embeddings reduce the number of parameters in the input and output layers by $23 \\%$ while achieving higher accuracy over fixed size embeddings. When the adaptive input representations are tied with an adaptive softmax in the output, then the number of parameters is reduced by a total of $61 \\%$ . ", + "bbox": [ + 174, + 724, + 825, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our experiments compare models based on word inputs, character inputs, as well as sub-word units using a self-attention architecture (Vaswani et al., 2017). We show that models with adaptive word representations can outperform very strong character-based models while training more than twice as fast. We also substantially improve adaptive softmax by introducing additional dropout regularization in the tail projection. On the WIKITEXT-103 benchmark we achieve a perplexity of 18.7, a reduction of 10.5 perplexity over the previously best published result. On the larger BILLION WORD benchmark our best model with adaptive input embeddings achieves 23.02 perplexity, a reduction of nearly 5 perplexity over the next best previously published result. ", + "bbox": [ + 174, + 829, + 823, + 898 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/52708bad0c295bb0584ec5d4f2259eabd3c9b29f269a819bd44546cef32e3818.jpg", + "image_caption": [ + "Figure 1: Illustration of adaptive input representations. Words are assigned to clusters $\\nu _ { i }$ based on their frequency which determines the size of the representations. Embeddings are projected to a common dimension $d$ before being fed to the model. " + ], + "image_footnote": [], + "bbox": [ + 307, + 106, + 691, + 327 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 410, + 825, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 472, + 344, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adaptive word representations are inspired by the adaptive softmax work Grave et al. (2017) which first described a GPU friendly way to construct a hierarchical softmax and showed that it performs very competitively compared to a full softmax, while offering significantly faster speed and a lower memory footprint. ", + "bbox": [ + 174, + 502, + 825, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Merity et al. (2018) use a modified version of adaptive softmax which does not reduce the dimensionality of less frequent words in order to be able to share output embeddings with the input. This setup is akin to a hierarchical softmax with tied weights. We show that variable-sized input embeddings can perform better than fixed sized embeddings. Furthermore, this also enables weight sharing with an adaptive softmax output layer. ", + "bbox": [ + 174, + 565, + 825, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Merity et al. (2018) evaluates both character-based and word-based factorizations but does not directly compare them to each other. We perform a direct comparison of word-based and characterbased input vocabularies and also compare to a sub-word factorization for both the input and output. Recently, Al-Rfou et al. (2018) demonstrated that self-attentional models can perform very well on language modeling tasks where the input and output is both characters. We also consider word-based benchmarks. ", + "bbox": [ + 174, + 642, + 825, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 ADAPTIVE INPUT REPRESENTATIONS ", + "text_level": 1, + "bbox": [ + 176, + 746, + 514, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The adaptive softmax exploits the fact that the distribution of word types in natural language follows a Zipfian distribution in order to improve the computation of the output probabilities. We apply the same intuition for input word embeddings with the motivation to reduce the number of parameters which frees up capacity for other parts of the model. ", + "bbox": [ + 174, + 776, + 825, + 833 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We define a number of clusters that partitions the frequency ordered vocabulary $\\nu ~ = ~ \\nu _ { 1 } \\cup$ $\\mathcal { V } _ { 2 } , \\ldots , \\mathcal { V } _ { n - 1 } \\cup \\mathcal { V } _ { n }$ such that $\\nu _ { i } \\cap \\nu _ { j } = \\emptyset$ for $\\forall i , j$ , and $i \\neq j$ , where $\\nu _ { 1 }$ contains the most frequent words and $\\nu _ { n }$ the least frequent words. We will refer to $\\nu _ { 1 }$ as the head and to any subsequent clusters loosely as tail. We reduce the capacity for each cluster by a factor of $k$ . That is, if words in $\\nu _ { 1 }$ have dimension $d$ , then words in $\\nu _ { n }$ have dimension $\\frac { d } { k ^ { n - 1 } }$ . We typically set $k = 4$ following Grave et al. (2017). ", + "bbox": [ + 173, + 840, + 825, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Next, we add linear projections $W _ { 1 } \\in \\mathbb { R } ^ { d \\times d } , \\ldots , W _ { n } \\in \\mathbb { R } ^ { d / k ^ { n - 1 } \\times d }$ to map the embeddings of each cluster to dimension $d$ so that the concatenated output of the adaptive input embedding layer can be easily used by the subsequent model (Figure 1). We also project $\\nu _ { 1 }$ which already has dimension $d$ . ", + "bbox": [ + 174, + 102, + 825, + 146 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When presented with a number of input words, the adaptive input embedding layer partitions the words into the various clusters, performs separate lookups in the embedding tables and then projects to dimension $d$ , followed by concatenating the embeddings in the original order. ", + "bbox": [ + 176, + 152, + 821, + 195 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Weight sharing. When the output layer is an adaptive softmax with the same partition of $\\nu , d ,$ , and $k$ as the adaptive input layer, then we can tie the weights (Inan et al., 2016; Press & Wolf, 2017). This further reduces the number of parameters and can simultaneously improve performance (§5). We can share both the parameters for the actual words as well as the projections $W _ { 1 } , \\ldots , W _ { n }$ Sharing the word embeddings is straightforward except for the head where the adaptive softmax has $n - 1$ additional embeddings for the remaining clusters which are not shared with the input. ", + "bbox": [ + 174, + 212, + 825, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We share all projections, except for the head projection which is not available in the adaptive softmax since the model output is directly multiplied with the output word embeddings for the head band. Performance decreased when we added a head projection to the adaptive softmax in the output, regardless of when it was shared or not. Sharing both the word embeddings as well as the projections performed very well on WIKITEXT-103 but on BILLION WORD we only share the word embeddings as we found that this performed better on the validation set. ", + "bbox": [ + 174, + 303, + 825, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 EXPERIMENTAL SETUP ", + "text_level": 1, + "bbox": [ + 176, + 409, + 398, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1 MODEL ", + "text_level": 1, + "bbox": [ + 174, + 441, + 267, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We follow most of the architectural choices described in Vaswani et al. (2017) but use only a decoder network. We add sinusoidal position embeddings to the input layer and stack $N = 1 6$ blocks for both BILLION WORD and WIKITEXT-103. Each block contains two sub-blocks: the first is a multihead self-attention module with $H = 1 6$ heads. The second sub-block is a feed-forward module (FFN) of the form $R e L U ( W _ { 1 } X + b _ { 1 } ) W _ { 2 } + b _ { 2 }$ where $W _ { 1 } \\in \\mathbb { R } ^ { e \\times e _ { f f } }$ , $W _ { 1 } \\in \\mathbb { R } ^ { e _ { f f } \\times e }$ and $e = 1 0 2 4$ , $e _ { f f } = 4 0 9 6$ unless otherwise stated. Different to Vaswani et al. (2017) we apply layer normalization before the self-attention and FFN blocks instead of after, as we find it leads to more effective training. Sub-blocks are surrounded by a residual connection (He et al., 2015). ", + "bbox": [ + 174, + 468, + 825, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use a dropout rate of 0.1 and attention dropout of 0.1 for BILLION WORD models, and increase regularization for WIKITEXT-103 by using dropout 0.3, and 0.1 ReLU dropout as well as attention dropout 0.1. We use the same hyperparameters for all models trained on the same dataset in order to enable a like for like comparison. When the dimensionality of the input or output layer differs from $e$ , then we add a simple linear projection with no bias. ", + "bbox": [ + 174, + 587, + 825, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2 DATASETS ", + "text_level": 1, + "bbox": [ + 174, + 675, + 287, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We experiment on the BILLION WORD benchmark and WIKITEXT-103. BILLION WORD contains 768M word tokens and has a vocabulary of about 800K word types, which corresponds to words with more than 3 occurrences in the training set (Chelba et al., 2013). ", + "bbox": [ + 176, + 703, + 821, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The training data of WIKITEXT-103 comprises about 100M tokens and a vocabulary of around 260K, corresponding to types with more than 3 occurrences in the training data (Merity et al., 2016). The dataset is composed of shuffled Wikipedia articles where the context carries across sentences. ", + "bbox": [ + 176, + 752, + 823, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.3 BATCHING ", + "text_level": 1, + "bbox": [ + 174, + 813, + 289, + 827 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For BILLION WORD we batch individual sentences since the corpus does not contain document structure. For WIKITEXT-103 we partition the training data into blocks of 512 contiguous tokens ignoring document boundaries. Evaluation is the same except that we require blocks to contain complete sentences totaling up to 512 tokens.2 ", + "bbox": [ + 174, + 840, + 825, + 895 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We limit the number of tokens per GPU to a maximum threshold $B$ per GPU. That is, we add examples of similar length until we reach this threshold. When we train on multiple GPUs, each GPU processes $B$ tokens using the same model parameters. This increases the effective batch size to the product of the number of GPUs and $B$ . For BILLION WORD models we use $B = 2 0 4 8$ and typically train on 32 GPUs, giving an effective batch size of 65K tokens. The smaller vocabulary of WIKITEXT-103 enables increasing $B$ to 4096 and we train on 8 GPUs. We found that large batch training is beneficial for this dataset and we therefore accumulate gradient updates over two batches before committing a parameter update (Ott et al., 2018a). This gives an effective batch size of 65K tokens for WIKITEXT-103. ", + "bbox": [ + 174, + 103, + 825, + 228 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.4 INPUT AND OUTPUT LAYER HYPERPARAMETERS ", + "text_level": 1, + "bbox": [ + 174, + 256, + 549, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Embedding sizes. For fixed size word input layers and softmax output layers we generally use embeddings of size 512 for WIKITEXT-103. When we use an adaptive softmax in the output and fixed size word embeddings for the input, then we use dimension 256 for the input embeddings for BILLION WORD and 64 for WIKITEXT-103. We tuned this choice on the validation set (Appendix A). BPE inputs and outputs have embeddings of size 1024. ", + "bbox": [ + 174, + 285, + 825, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Character CNN. We model character inputs by convolving the representations of all characters in a word following Kim et al. (2015) which applies several filters, then max pooling, a number of highway layers and a projection. Character embeddings have size 128 and we apply seven filters of size 1x128, $2 \\mathrm { x } 2 5 6$ , 3x384, 4x512, 5x512, 6x512, 7x512, where $3 \\mathrm { x } 1 2 8$ indicates a filter processing three characters that outputs 128 features. We use a single highway layer for WIKITEXT-103, and two for BILLION WORD. We do not add start of word and end of word markers as they did not improve validation accuracy. We train on the same pre-processed data as the other models, with unknown tokens in both the inputs and outputs. ", + "bbox": [ + 173, + 378, + 825, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive input representations and adaptive softmax. We use an adaptive softmax output layer to train models with large word-based vocabularies. For adaptive word inputs and adaptive softmax, we use embeddings of size $d = 1 0 2 4$ for the head and reduce the size of subsequent clusters by a factor of $k = 4$ . For WIKITEXT-103, we have three bands of size 20K $\\scriptstyle ( \\mathrm { d } = 1 0 2 4 )$ ), 40K $\\mathrm { \\Delta } \\mathrm { Q } \\mathrm { = } 2 5 6 )$ and 200K $\\scriptstyle \\mathrm { { ( d = 6 4 ) } }$ ). For BILLION WORD the bands are 60K $\\scriptstyle \\mathrm { d } = 1 0 2 4$ ), 100K $\\mathrm { \\Delta } \\mathrm { Q } { = } 2 5 6 \\mathrm { \\Delta }$ ), and 640K $\\scriptstyle \\mathrm { { ( d = 6 4 ) } }$ ). ", + "bbox": [ + 174, + 515, + 825, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sub-word models. We learn a byte-pair encoding (BPE) of 32K codes on the training data of each benchmark (Sennrich et al., 2016). After applying the code to the training data we obtain a vocabulary of 33,337 tokens for WIKITEXT-103 and 32,347 tokens for BILLION WORD. BPE input/output embeddings have size 1024. The final evaluation is in terms word-level perplexity to be comparable to other models. The probability of a word is the product of the sub-word units. ", + "bbox": [ + 174, + 609, + 825, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.5 OPTIMIZATION ", + "text_level": 1, + "bbox": [ + 176, + 705, + 318, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Different to Vaswani et al. (2017) we use Nesterov’s accelerated gradient method (Sutskever et al., 2013) with a momentum of 0.99 and we renormalize gradients if their norm exceeds 0.1 (Pascanu et al., 2013). The learning rate is linearly warmed up from $1 0 ^ { - 7 }$ to 1 for 16K steps and then annealed using a cosine learning rate schedule with $C$ cycles (Loshchilov & Hutter, 2016). Each cycle runs for twice the number of updates than the previous cycle and we lower the maximum and minimum learning rates by a rate $M$ compared to the previous cycle. The initial minimum learning rate is $1 0 ^ { - 5 }$ and the maximum is 1. ", + "bbox": [ + 174, + 734, + 825, + 833 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "BILLION WORD models train for a total of 975K updates over $C = 3$ cycles, the first cycle takes 137K steps, and we set $M = 0 . 6$ . The WIKITEXT-103 models train for 286K steps over $C = 4$ cycles, the first cycle takes 18K setps and we set $M \\ : = \\ : 0 . 7 5$ . We run experiments on DGX-1 machines with 8 NVIDIA V100 GPUs and machines are interconnected by Infiniband. We also use the NCCL2 library and the torch.distributed package for inter-GPU communication. We train models with 16-bit floating point precision, following Ott et al. (2018b). ", + "bbox": [ + 174, + 840, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/0c7c963dd8b2903cef8d6d8ec3f02912b76db5a52245768b59e1643f85ce4e3b.jpg", + "table_caption": [ + "Table 1: Test perplexity on BILLION WORD. Adaptive inputs share parameters with an adaptive softmax. Training times of Char-CNN and Adaptive input models are measured when training with 64 GPUs. †does not include embedding and softmax layers " + ], + "table_footnote": [], + "table_body": "
TestTrain Time (hours)Parameters
Dauphin et al. (2017)31.9428M
J6zefowicz et al. (2016)30.01,040M
Shazeer et al. (2017)28.0-4,371Mt
Char-CNN25.8879366M
Adaptive inputs25.2255331M
Adaptive inputs (large)23.9172465M
Adaptive inputs (very large)23.021451026M
10 LSTMs + SNM10-SKIP (Shazeer et al.,2016)23.71=
", + "bbox": [ + 200, + 101, + 795, + 275 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0d1b611f90a7f9ae450f1e0728a887cd39a36f52a2ceafdc4ab9a882f69e86cc.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2: Test perplexity on WIKITEXT-103 (cf. Table 1). Training time is based on 8 GPUs. " + ], + "table_body": "
TestTrain Time (hours)Parameters
Grave et al. (2016)40.8
Dauphin et al. (2017)37.2229M
Merity et al. (2018)33.0=151M
Rae et al. (2018)29.2-
Adaptive inputs18.767247M
", + "bbox": [ + 297, + 352, + 699, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 EXPERIMENTS AND RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 527, + 449, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1 MAIN RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 559, + 320, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the main results on BILLION WORD, we doubled the batch size by training on 64 GPUs instead of 32 GPUs. We also consider two larger setups, one where we added four more blocks $N = 2 0$ ) and increased the FFN dimension to $e _ { f f } = 6 1 4 4$ (large), and another where we add another four blocks $N = 2 4$ ) with $e _ { f f } = 8 1 9 2$ and $e = 1 5 3 6$ (very large). All other settings follow $\\ S 4 . 4$ and all models were trained for the same number of steps. ", + "bbox": [ + 174, + 585, + 825, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 compares our models to previous work on BILLION WORD. The adaptive input model outperforms the best previously reported result at an order of magnitude fewer parameters. Our large model performs nearly as well as an ensemble of over ten models and achieves a new state of the art of 24.14 perplexity. Our very large model performs as well as an ensemble of over ten models and achieves 23.02 perplexity. The Char-CNN model performs $0 . 6 ~ \\mathrm { P P L }$ worse than the standard adaptive input model even though it trained for over $40 \\%$ longer. ", + "bbox": [ + 174, + 662, + 825, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2 shows our result on WIKITEXT-103 where adaptive inputs achieve 18.7 perplexity. For this result only, we partition the training data into blocks of 3072 contiguous tokens instead of 512 tokens as for other experiments. During evaluation we require blocks to contain complete sentences totaling up to 3072 tokens of which the first 2560 tokens serve as context to score the last 512 tokens; we take care to score all tokens in the test and validation sets. We motivate this choice in $\\ S 5 . 3$ . ", + "bbox": [ + 174, + 753, + 825, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": ".2 COMPARISON OF INPUT AND OUTPUT LAYER FACTORIZATIONS ", + "text_level": 1, + "bbox": [ + 186, + 842, + 643, + 856 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we perform a systematic comparison of different input and output layer factorizations. We consider a word-based setup with fixed size word input embeddings and a standard word softmax (SM) where embeddings have either dimension 512 (WIKITEXT-103) or 64 (BILLION WORD). We consider tying the input and output embeddings (SM-T). Instead of words, we try less sparse subword units, both in the input and output, with embeddings of size 1024 (BPE) and shared weights (BPE-T). Next, we consider replacing the fixed size output representations by an adaptive softmax (ASM) and characters as input (CNN). Finally, we use both adaptive input word representations as well as an adaptive softmax (ADP) and a tied version (ADP-T). All models use the same selfattention architecture described in $\\ S 4 . 1$ . ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e3e37fff1071f7ce2f4fbe3586ddfabf4960159d940eb2fd51570e533afed4df.jpg", + "table_caption": [ + "Table 3: Test perplexity on WIKITEXT-103 for various input and output layer factorizations. Training speed was measured on a single 8-GPU machine. $( ^ { * } )$ indicates a modified training regime because of large memory requirements: the maximum number of tokens per GPU was lowered to 1024 from 4096 but the same number of updates were performed by processing four batches before committing a weight update. " + ], + "table_footnote": [], + "table_body": "
InputOutputValidTestTrain Time (hours)Params
SMEmbeddingSoftmax23.8724.9257*476.8M
BPEBPE EmbeddingBPE Softmax23.1324.2530270M
BPE-TBPE EmbeddingBPE Softmax (tied)22.4623.4530235.7M
SM-TEmbeddingSoftmax (tied)22.6323.3856*339.7M
ASMEmbeddingAdaptive21.2322.1835263.1M
CNNChar-CNNAdaptive20.8621.7970266.3M
ADPAdaptiveAdaptive20.9521.7434291.3M
ADP-TAdaptiveAdaptive (tied)19.7920.5130246.9M
", + "bbox": [ + 186, + 101, + 812, + 261 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/aa93371ce50bf62d0f1b00bb824bec36f123c85c9463cdc667ce6a25a44cfb43.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4: Test perplexity on BILLION WORD. Training speed measured on four 8-GPU machines. " + ], + "table_body": "
InputOutputValidTestTrain time (hours)Params
BPE-TBPE EmbeddingBPE Softmax (shared)27.4427.5134234.7M
BPEBPE EmbeddingBPE Softmax27.0227.1335267.8M
ASMEmbeddingAdaptive26.9727.0662532.8M
CNNChar-CNNAdaptive26.1326.2592365.8M
ADPAdaptiveAdaptive26.3826.4965458.4M
ADP-TAdaptiveAdaptive (shared)25.5125.5843330.8M
", + "bbox": [ + 179, + 383, + 818, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 595, + 825, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 shows results when training all configurations for the same number of updates. Adaptive input representations with tied input and output layers (ADP-T) achieve the highest accuracy at the same speed as the BPE models which have a very small vocabulary (33K versus 260K). CNN is 1 perplexity worse than ADP-T and requires well over twice the training time. It is the slowest approach, even though it has a fast adaptive softmax in the output. Fixed word embeddings perform least well (SM). Sub-word units are fast to train and perform better than word models with fixed sized embeddings. ASM improves over SM and greatly speeds up training. For ASM, we found that reducing the dimension of the input word embeddings to 64 on WIKITEXT-103 results in better accuracy (Appendix A). ", + "bbox": [ + 174, + 672, + 825, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4 shows that adaptive input representations perform equally well on BILLION WORD compared to other factorizations. ADP-T is $34 \\%$ faster than ADP because there are fewer parameters to update. Similar to before, ADP-T trains more than twice as fast as CNN at higher accuracy, however, the accuracy gap is narrower than for WIKITEXT-103. Regularization is more important on WIKITEXT-103 while models for BILLION WORD benefit from additional capacity. Because of this we used input word embeddings of size 256 for ASM. ", + "bbox": [ + 174, + 805, + 825, + 888 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also trained CNN without replacing input words outside the vocabulary by an unknown symbol, however, this only improved validation perplexity by 0.16. ", + "bbox": [ + 173, + 895, + 821, + 924 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b396896fe5d317f22b0248271298f7ce8c2e521c3dc122395301b4c3ec88349.jpg", + "image_caption": [ + "Figure 2: Loss of models binned by word frequency on the test set of WIKITEXT-103. Bins are not cumulative. " + ], + "image_footnote": [], + "bbox": [ + 176, + 98, + 823, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/373e2f18c0f9c2f71d23422298631e38730245add24d65b6f615d793b9273992.jpg", + "image_caption": [ + "Figure 3: Loss of models when binning by the frequency of the previous word measured on WIKITEXT-103 (cf. Figure 2). " + ], + "image_footnote": [], + "bbox": [ + 176, + 391, + 823, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3 ANALYSIS ", + "text_level": 1, + "bbox": [ + 174, + 702, + 285, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we turn to the question of how well models perform on rare words compared to frequent words. We compute the average loss for each word in the test set and group words by frequency. ", + "bbox": [ + 174, + 728, + 823, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 2 shows results on WIKITEXT-103. Tying weights helps all models on rare words, likely because of regularization effects. Fixed size word embeddings with a word softmax (SM and SMT) do not perform well on rare words. This is likely due to underfitting on common words and we use the largest possible embedding size we could fit on 16GB GPU cards given our batch size (more experimentation in Appendix A). BPE and BPE-T perform poorly on rare words because probabilities are a product of several sub-word units. ADP-T performs best across all frequency ranges. Figure 3 bins the loss by the frequency of the previous word and shows that CNN does well when it has rare words in the context, however, ADP-T does best across all bins. ", + "bbox": [ + 174, + 762, + 825, + 875 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 shows an equivalent analysis for BILLION WORD. The largest differences between models is on rare words. CNN performs best on very rare words but is outperformed by ADP in all other settings. Similar to WIKITEXT-103, BPE and BPE-T perform poorly on rare words. Further ", + "bbox": [ + 174, + 882, + 823, + 924 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d73dcb6e7e5c4af036340a7099e37fe1418673e98a1f8915cde2d051627402b8.jpg", + "image_caption": [ + "Figure 4: Loss of models when binning by word frequency on the test set of BILLION WORD. Bins are not cumulative. " + ], + "image_footnote": [], + "bbox": [ + 176, + 98, + 821, + 337 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/34ba39af36272851bc56541c57fa7eeaea8880ea2ca40ce648a9adb53f077da7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Train block sizeInference context sizeValidation perplexityTest perplexity
51219.7920.51
5120 48018.3519.03
2048018.9619.53
2048153618.2318.88
3072018.6319.34
3072256017.9718.70
", + "bbox": [ + 333, + 396, + 663, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5: Perplexity on WIKITEXT-103 with different context sizes during training and inference. Training block size is the number of consecutive tokens considered during training. Inference context is the number of tokens provided at evaluation before scoring tokens. ", + "bbox": [ + 174, + 539, + 823, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "analysis (Appendix 5.3) binning the loss by the frequency of the previous word shows that weight sharing also helps for BILLION WORD and that CNN does very well on rare words for BILLION WORD compared to other models. ", + "bbox": [ + 174, + 611, + 825, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5 shows the importance of context size for WIKITEXT-103. Training block size is the number of consecutive tokens that are considered at once during training. Inference context is the number of tokens that are provided at evaluation before any tokens are scored. Simply increasing the training block size from 512 to 3072 results in an improvement of nearly 1.2 perplexity with no inference context window. Increasing the context size at inference time results in an improvement of 0.6 perplexity for the largest training block size. ", + "bbox": [ + 174, + 660, + 825, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4 ADAPTIVE SOFTMAX VS. FULL SOFTMAX ", + "text_level": 1, + "bbox": [ + 174, + 763, + 504, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also found that adaptive softmax can benefit from additional regularization of rare words. Adaptive softmax first projects the model output to the dimension of a particular cluster and then computes a dot product with the respective word embeddings. We add dropout to the output of the first projection for all clusters, except for the head. This change enables the adaptive softmax to outperform a standard softmax over fixed size output word embeddings on WIKITEXT-103 (Table 6). ", + "bbox": [ + 174, + 790, + 823, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, we found that adding dropout in this way is not helpful for larger datasets such as BILLION WORD. Unfortunately, a standard softmax over 800K words is not tractable and we were unable to make a comparison. It may be possible to achieve better results by tuning dropout for each band of the tail and we leave this for future work. ", + "bbox": [ + 176, + 867, + 823, + 922 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/787763e4f22c859d1fb0dcb91dcc43e757ceb5d613ad01d21320a2ab8bb917d1.jpg", + "table_caption": [ + "Table 6: Perplexity on WIKITEXT-103 when regularizing rare words in adaptive softmax. " + ], + "table_footnote": [], + "table_body": "
Tail dropoutValidation perplexity
Softmax (SM)N/A23.87
Adaptive (ADP)0.024.74
Adaptive (ADP)0.221.23
", + "bbox": [ + 361, + 101, + 638, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 CONCLUSION ", + "text_level": 1, + "bbox": [ + 174, + 242, + 318, + 258 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Adaptive input embeddings vary the size of input word embeddings which can improve accuracy while drastically reducing the number of model parameters. When sharing parameters with an adaptive softmax, the number of parameters can be further reduced which improves training speed. We presented a comparison between different input and output layer factorizations including word inputs, character inputs and sub-word units in both the input and output. ", + "bbox": [ + 174, + 273, + 825, + 344 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our experiments show that models with adaptive input embeddings train faster compared to character input CNNs while achieving higher accuracy. We achieve new state of the art results on WIKITEXT-103 and BILLION WORD. In future work, we will apply variable sized input embeddings to other tasks. ", + "bbox": [ + 174, + 351, + 825, + 406 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ACKNOWLEDGMENTS ", + "text_level": 1, + "bbox": [ + 176, + 424, + 326, + 438 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Tom Bosc for fruitful comments and suggestions. ", + "bbox": [ + 174, + 446, + 563, + 462 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 483, + 285, + 498 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Rami Al-Rfou, Dokook Choe, Noah Constant, Mandy Guo, and Llion Jones. Character-level language modeling with deeper self-attention. arXiv, abs/1808.04444, 2018. ", + "bbox": [ + 173, + 507, + 823, + 535 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ebru Arisoy, Tara N. Sainath, Brian Kingsbury, and Bhuvana Ramabhadran. Deep Neural Network Language Models. In NAACL-HLT Workshop on the Future of Language Modeling for HLT, 2012. ", + "bbox": [ + 173, + 545, + 823, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Paul Baltescu and Phil Blunsom. Pragmatic neural language modelling in machine translation. In Proc. of ACL, 2015. ", + "bbox": [ + 174, + 598, + 823, + 627 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Yoshua Bengio, Rejean Ducharme, Pascal Vincent, and Christian Jauvin. A Neural Probabilistic´ Language Model. JMLR, 3:1137–1155, 2003. ", + "bbox": [ + 174, + 637, + 820, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jacob Buckman and Graham Neubig. Neural lattice language models. TACL, 6:529–541, 2018. ", + "bbox": [ + 169, + 675, + 799, + 691 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ciprian Chelba, Tomas Mikolov, Mike Schuster, Qi Ge, Thorsten Brants, Phillipp Koehn, and Tony Robinson. One billion word benchmark for measuring progress in statistical language modeling. Technical report, Google, 2013. ", + "bbox": [ + 176, + 700, + 820, + 743 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Wenlin Chen, David Grangier, and Michael Auli. Strategies for training large vocabulary neural language models. In Proc. of ACL, 2016. ", + "bbox": [ + 171, + 753, + 823, + 782 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Yann N. Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated convolutional networks. In Proc. of ICML, 2017. ", + "bbox": [ + 169, + 791, + 823, + 821 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Joshua Goodman. Classes for Fast Maximum Entropy Training. In Proc. of ICASSP, 2001. ", + "bbox": [ + 169, + 830, + 771, + 847 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Edouard Grave, Armand Joulin, and Nicolas Usunier. Improving neural language models with a continuous cache. arXiv, abs/1612.04426, 2016. ", + "bbox": [ + 171, + 856, + 823, + 886 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Edouard Grave, Armand Joulin, Moustapha Cisse, David Grangier, and Herv ´ e J ´ egou. Efficient ´ softmax approximation for gpus. In Proc. of ICML, 2017. ", + "bbox": [ + 173, + 895, + 820, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proc. of CVPR, 2015. \nHakan Inan, Khashayar Khosravi, and Richard Socher. Tying word vectors and word classifiers: A loss framework for language modeling. arXiv, abs/1611.01462, 2016. \nRafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring the ´ limits of language modeling. arXiv, abs/1602.02410, 2016. \nYoon Kim, Yacine Jernite, David Sontag, and Alexander M. Rush. Character-aware neural language models. arXiv, abs/1508.06615, 2015. \nYoon Kim, Yacine Jernite, David Sontag, and Alexander M Rush. Character-aware neural language models. In AAAI, pp. 2741–2749, 2016. \nIlya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with restarts. arXiv, abs/1608.03983, 2016. \nStephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv, abs/1609.07843, 2016. \nStephen Merity, Nitish Shirish Keskar, and Richard Socher. An analysis of neural language modeling at multiple scales. arXiv, abs/1803.08240, 2018. \nSebastian J. Mielke and Jason Eisner. Spell once, summon anywhere: A two-level open-vocabulary language model. arXiv, abs/1804.08205, 2018. \nToma´s Mikolov, Karafi ˇ at Martin, Luk ´ a´s Burget, Jan Cernock ˇ y, and Sanjeev Khudanpur. Recurrent ´ Neural Network based Language Model. In Proc. of Interspeech, 2010. \nToma´s Mikolov, Stefan Kombrink, Luk ˇ a´s Burget, Jan Cernock ˇ y, and Sanjeev Khudanpur. Exten-´ sions of Recurrent Neural Network Language Model. In Proc. of ICASSP, pp. 5528–5531, 2011. \nFrederic Morin and Yoshua Bengio. Hierarchical Probabilistic Neural Network Language Model. In Proc. of AISTATS, 2005. \nMyle Ott, Michael Auli, David Grangier, and MarcAurelio Ranzato. Analyzing uncertainty in neural machine translation. In Proc. of ICML, 2018a. \nMyle Ott, Sergey Edunov, David Grangier, and Michael Auli. Scaling neural machine translation. In Proc. of WMT, 2018b. \nRazvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural networks. In Proc. of ICML, 2013. \nOfir Press and Lior Wolf. Using the output embedding to improve language models. In Proc. of EACL, 2017. \nJack W. Rae, Chris Dyer, Peter Dayan, and Timothy P. Lillicrap. Fast parametric learning with activation memorization. arXiv, abs/1803.10049, 2018. \nHolger Schwenk, Anthony Rousseau, and Mohammed Attik. Large, Pruned or Continuous Space Language Models on a GPU for Statistical Machine Translation. In NAACL-HLT Workshop on the Future of Language Modeling for HLT, 2012. \nRico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. In Proc. of ACL, 2016. \nNoam Shazeer, Joris Pelemans, and Ciprian Chelba. Sparse non-negative matrix language modeling. In Proc. of Interspeech, 2016. \nNoam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc V. Le, Geoffrey E. Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv, abs/1701.06538, 2017. ", + "bbox": [ + 169, + 65, + 828, + 929 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ilya Sutskever, James Martens, George E. Dahl, and Geoffrey E. Hinton. On the importance of initialization and momentum in deep learning. In Proc. of ICML, 2013. ", + "bbox": [ + 171, + 103, + 825, + 132 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ashish Vaswani, Yinggong Zhao, Victoria Fossum, and David Chiang. Decoding with Large-scale Neural Language Models improves Translation. In Proc. of EMNLP, October 2013. ", + "bbox": [ + 173, + 140, + 823, + 170 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. In Proc. of NIPS, 2017. ", + "bbox": [ + 176, + 179, + 823, + 208 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "SUPPLEMENTARY MATERIAL ", + "bbox": [ + 176, + 102, + 416, + 117 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A ADDITIONAL EXPERIMENTS ON WIKITEXT-103 ", + "text_level": 1, + "bbox": [ + 176, + 136, + 599, + 151 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "This appendix shows various ablation. Table 7 shows that reducing the capacity of fixed size word input embddings is beneficial on WIKITEXT-103. The next set of results in Table 7 shows results for various settings of the SM and SM-T models. We also experimented with sharing the head projection but found this to perform less well than not sharing it. Finally, Table 8 shows various band sizes for adaptive input word embbedings. ", + "bbox": [ + 174, + 166, + 825, + 237 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We also show the performance of Merity et al. (2018) who use an adaptive softmax with equally sized word representations and share the input and output embeddings (no dim reduction, tied). ", + "bbox": [ + 174, + 244, + 823, + 272 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3fbfe30985bf9514dacb12e0943de5c19d13b7cd9c0a18913baf4f9cc354e365.jpg", + "table_caption": [ + "Table 7: Validation perplexity of our models on WIKITEXT-103. " + ], + "table_footnote": [], + "table_body": "
Input OutputDropout Valid PPLParameters
256d EmbeddingAdaptive0.323.39314.7M
128d EmbeddingAdaptive0.321.51280.3M
64d EmbeddingAdaptive0.321.23263.1M
32d EmbeddingAdaptive0.321.78254.5M
512d Embedding512d Softmax (tied)0.322.63339.7M
512d Embedding512d Softmax (tied)0.428.31339.7M
512d Embedding512d Softmax0.323.87476.8
512d Embedding512d Softmax0.427.64476.8
256d Embedding256d Softmax (tied)0.322.65270.6M
256d Embedding256d Softmax0.324.13339.1M
64d Embedding512d Softmax0.324.74356.3M
AdaptiveAdaptive (tied emb, not proj)0.320.06247.3M
AdaptiveAdaptive (tied emb/proj not head)0.319.79246.9M
AdaptiveAdaptive (tied emb/proj + head)0.320.06246.9M
512d Embedding512d Softmax (tied)0.322.63339.7M
512d Embedding 512d Adaptive (no dim reduction, tied)0.325.48340.2M
", + "bbox": [ + 171, + 285, + 852, + 556 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/87847578c5c5390ef40a775e10e07897a5106fcd1e9ae1dbe98b927732c8091d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Softmax cutoffValid PPL
20k/40k/200k19.79
20k/140k/100k 20k/40k/60k/140k20.26
60k/100k/100k20.53 20.52
5k/155k/100k
20k/40k/200k20.06
10k/60k/190k19.99 19.79
", + "bbox": [ + 380, + 602, + 616, + 729 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 8: Validation perplexity on WIKITEXT-103 with tied adaptive inputs & outputs. The bands signify the number of words belonging to each band. In every case, the first band has dimension 1024, the second band 256, the third band 64 and the fourth band (if it exists) 16. ", + "bbox": [ + 173, + 739, + 825, + 782 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "B ANALYSIS ", + "text_level": 1, + "bbox": [ + 174, + 810, + 295, + 827 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "This appendix extends the analysis in $\\ S 5 . 3$ by showing a breakdown of the test loss when binning by the frequency of the previous word. ", + "bbox": [ + 176, + 842, + 823, + 872 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/62e4d7a047372091139f4e3c0bc67293ce7fc2b1aef9dae0b74901d294ce4ff9.jpg", + "image_caption": [ + "Figure 5: Loss of models when binning by the frequency of the previous word measured on BILLION WORD (cf. Figure 3). " + ], + "image_footnote": [], + "bbox": [ + 176, + 367, + 823, + 607 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/parse/train/ByxZX20qFQ/ByxZX20qFQ_middle.json b/parse/train/ByxZX20qFQ/ByxZX20qFQ_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..4e36288d23ae48c6aa4956a714e518a9ffc91931 --- /dev/null +++ b/parse/train/ByxZX20qFQ/ByxZX20qFQ_middle.json @@ -0,0 +1,25612 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 79, + 380, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 381, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 381, + 97 + ], + "score": 1.0, + "content": "ADAPTIVE INPUT REPRESENTATIONS", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 97, + 383, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 383, + 118 + ], + "score": 1.0, + "content": "FOR NEURAL LANGUAGE MODELING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 135, + 245, + 168 + ], + "lines": [ + { + "bbox": [ + 113, + 135, + 245, + 146 + ], + "spans": [ + { + "bbox": [ + 113, + 135, + 245, + 146 + ], + "score": 1.0, + "content": "Alexei Baevski & Michael Auli", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 145, + 208, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 145, + 208, + 158 + ], + "score": 1.0, + "content": "Facebook AI Research,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 157, + 205, + 169 + ], + "spans": [ + { + "bbox": [ + 112, + 157, + 205, + 169 + ], + "score": 1.0, + "content": "Menlo Park, CA, USA", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 278, + 198, + 333, + 209 + ], + "lines": [ + { + "bbox": [ + 276, + 196, + 336, + 212 + ], + "spans": [ + { + "bbox": [ + 276, + 196, + 336, + 212 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 223, + 468, + 333 + ], + "lines": [ + { + "bbox": [ + 142, + 223, + 469, + 235 + ], + "spans": [ + { + "bbox": [ + 142, + 223, + 469, + 235 + ], + "score": 1.0, + "content": "We introduce adaptive input representations for neural language modeling which", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 234, + 469, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 469, + 246 + ], + "score": 1.0, + "content": "extend the adaptive softmax of Grave et al. (2017) to input representations of vari-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 245, + 470, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 470, + 258 + ], + "score": 1.0, + "content": "able capacity. There are several choices on how to factorize the input and output", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "score": 1.0, + "content": "layers, and whether to model words, characters or sub-word units. We perform a", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 267, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 470, + 279 + ], + "score": 1.0, + "content": "systematic comparison of popular choices for a self-attentional architecture. Our", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 279, + 469, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 279, + 469, + 290 + ], + "score": 1.0, + "content": "experiments show that models equipped with adaptive embeddings are more than", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "score": 1.0, + "content": "twice as fast to train than the popular character input CNN while having a lower", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 299, + 469, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 469, + 312 + ], + "score": 1.0, + "content": "number of parameters. On the WIKITEXT-103 benchmark we achieve 18.7 per-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 311, + 470, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 470, + 323 + ], + "score": 1.0, + "content": "plexity, an improvement of 10.5 perplexity compared to the previously best pub-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 321, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 321, + 469, + 335 + ], + "score": 1.0, + "content": "lished result and on the BILLION WORD benchmark, we achieve 23.02 perplexity.1", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 108, + 354, + 206, + 367 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 208, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 208, + 371 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 380, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 106, + 381, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 505, + 393 + ], + "score": 1.0, + "content": "Language modeling is a basic task in natural language processing, with many applications such as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 391, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 404 + ], + "score": 1.0, + "content": "speech recognition (Arisoy et al., 2012) and statistical machine translation (Schwenk et al., 2012;", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "score": 1.0, + "content": "Vaswani et al., 2013; Baltescu & Blunsom, 2015). Recently, much progress has been made by neural", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 413, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 505, + 426 + ], + "score": 1.0, + "content": "methods (Bengio et al., 2003; Mikolov et al., 2010) based on LSTMs (Jozefowicz et al., 2016), gated ´", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 424, + 502, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 502, + 437 + ], + "score": 1.0, + "content": "convolutional networks (Dauphin et al., 2017) and self-attentional networks (Al-Rfou et al., 2018).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 505, + 496 + ], + "lines": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "There are different choices for the basic unit we wish to model, including full words (Bengio et al.,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "2003), characters for the input (Kim et al., 2016), or also the output (Merity et al., 2018) as well as", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 462, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 477 + ], + "score": 1.0, + "content": "sub-words (Buckman & Neubig, 2018; Mielke & Eisner, 2018). Word-based models are particularly", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "challenging since computing probabilities for all 800K words of the BILLION WORD benchmark is", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 486, + 380, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 380, + 498 + ], + "score": 1.0, + "content": "still a substantial part of the overall computation (Chen et al., 2016).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 504, + 568 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 506, + 515 + ], + "score": 1.0, + "content": "A popular approach to lower the computational burden is to structure the output vocabulary so that", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 513, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 525 + ], + "score": 1.0, + "content": "not all probabilities need to be computed. The hierarchical softmax does this by introducing latent", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 524, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 536 + ], + "score": 1.0, + "content": "variables or clusters to simplify normalization (Goodman, 2001; Morin & Bengio, 2005; Mikolov", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 536, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 505, + 547 + ], + "score": 1.0, + "content": "et al., 2011). This has been further improved by the adaptive softmax which introduces a variable", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 546, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 559 + ], + "score": 1.0, + "content": "capacity scheme for output word embeddings, assigning more parameters to frequent words and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 557, + 315, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 315, + 570 + ], + "score": 1.0, + "content": "fewer parameters to rare words (Grave et al., 2017).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 574, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "In this paper, we introduce adaptive input embeddings which extend the adaptive softmax to input", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "word representations. This factorization assigns more capacity to frequent words and reduces the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 595, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 609 + ], + "score": 1.0, + "content": "capacity for less frequent words with the benefit of reducing overfitting to rare words. For a com-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 607, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 505, + 618 + ], + "score": 1.0, + "content": "petitive setup on the BILLION WORD benchmark, adaptive input embeddings reduce the number", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 304, + 630 + ], + "score": 1.0, + "content": "of parameters in the input and output layers by", + "type": "text" + }, + { + "bbox": [ + 304, + 618, + 324, + 628 + ], + "score": 0.86, + "content": "23 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "while achieving higher accuracy over fixed", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 628, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 641 + ], + "score": 1.0, + "content": "size embeddings. When the adaptive input representations are tied with an adaptive softmax in the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 640, + 378, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 354, + 652 + ], + "score": 1.0, + "content": "output, then the number of parameters is reduced by a total of", + "type": "text" + }, + { + "bbox": [ + 355, + 640, + 374, + 650 + ], + "score": 0.87, + "content": "61 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 640, + 378, + 652 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 657, + 504, + 712 + ], + "lines": [ + { + "bbox": [ + 106, + 657, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 505, + 669 + ], + "score": 1.0, + "content": "Our experiments compare models based on word inputs, character inputs, as well as sub-word units", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "score": 1.0, + "content": "using a self-attention architecture (Vaswani et al., 2017). We show that models with adaptive word", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 678, + 505, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 692 + ], + "score": 1.0, + "content": "representations can outperform very strong character-based models while training more than twice", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 689, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 703 + ], + "score": 1.0, + "content": "as fast. We also substantially improve adaptive softmax by introducing additional dropout regular-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "ization in the tail projection. On the WIKITEXT-103 benchmark we achieve a perplexity of 18.7, a", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 722, + 451, + 732 + ], + "lines": [ + { + "bbox": [ + 120, + 720, + 452, + 734 + ], + "spans": [ + { + "bbox": [ + 120, + 720, + 452, + 734 + ], + "score": 1.0, + "content": "1Code and pre-trained models available at http://github.com/pytorch/fairseq", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 79, + 380, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 381, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 381, + 97 + ], + "score": 1.0, + "content": "ADAPTIVE INPUT REPRESENTATIONS", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 97, + 383, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 383, + 118 + ], + "score": 1.0, + "content": "FOR NEURAL LANGUAGE MODELING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 135, + 245, + 168 + ], + "lines": [ + { + "bbox": [ + 113, + 135, + 245, + 146 + ], + "spans": [ + { + "bbox": [ + 113, + 135, + 245, + 146 + ], + "score": 1.0, + "content": "Alexei Baevski & Michael Auli", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 145, + 208, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 145, + 208, + 158 + ], + "score": 1.0, + "content": "Facebook AI Research,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 157, + 205, + 169 + ], + "spans": [ + { + "bbox": [ + 112, + 157, + 205, + 169 + ], + "score": 1.0, + "content": "Menlo Park, CA, USA", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 111, + 135, + 245, + 169 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 198, + 333, + 209 + ], + "lines": [ + { + "bbox": [ + 276, + 196, + 336, + 212 + ], + "spans": [ + { + "bbox": [ + 276, + 196, + 336, + 212 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 223, + 468, + 333 + ], + "lines": [ + { + "bbox": [ + 142, + 223, + 469, + 235 + ], + "spans": [ + { + "bbox": [ + 142, + 223, + 469, + 235 + ], + "score": 1.0, + "content": "We introduce adaptive input representations for neural language modeling which", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 234, + 469, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 469, + 246 + ], + "score": 1.0, + "content": "extend the adaptive softmax of Grave et al. (2017) to input representations of vari-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 245, + 470, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 470, + 258 + ], + "score": 1.0, + "content": "able capacity. There are several choices on how to factorize the input and output", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "score": 1.0, + "content": "layers, and whether to model words, characters or sub-word units. We perform a", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 267, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 470, + 279 + ], + "score": 1.0, + "content": "systematic comparison of popular choices for a self-attentional architecture. Our", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 279, + 469, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 279, + 469, + 290 + ], + "score": 1.0, + "content": "experiments show that models equipped with adaptive embeddings are more than", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "score": 1.0, + "content": "twice as fast to train than the popular character input CNN while having a lower", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 299, + 469, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 469, + 312 + ], + "score": 1.0, + "content": "number of parameters. On the WIKITEXT-103 benchmark we achieve 18.7 per-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 311, + 470, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 470, + 323 + ], + "score": 1.0, + "content": "plexity, an improvement of 10.5 perplexity compared to the previously best pub-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 321, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 321, + 469, + 335 + ], + "score": 1.0, + "content": "lished result and on the BILLION WORD benchmark, we achieve 23.02 perplexity.1", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 10.5, + "bbox_fs": [ + 141, + 223, + 470, + 335 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 354, + 206, + 367 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 208, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 208, + 371 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 380, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 106, + 381, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 505, + 393 + ], + "score": 1.0, + "content": "Language modeling is a basic task in natural language processing, with many applications such as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 391, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 404 + ], + "score": 1.0, + "content": "speech recognition (Arisoy et al., 2012) and statistical machine translation (Schwenk et al., 2012;", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "score": 1.0, + "content": "Vaswani et al., 2013; Baltescu & Blunsom, 2015). Recently, much progress has been made by neural", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 413, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 505, + 426 + ], + "score": 1.0, + "content": "methods (Bengio et al., 2003; Mikolov et al., 2010) based on LSTMs (Jozefowicz et al., 2016), gated ´", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 424, + 502, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 502, + 437 + ], + "score": 1.0, + "content": "convolutional networks (Dauphin et al., 2017) and self-attentional networks (Al-Rfou et al., 2018).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 381, + 505, + 437 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 505, + 496 + ], + "lines": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "There are different choices for the basic unit we wish to model, including full words (Bengio et al.,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "2003), characters for the input (Kim et al., 2016), or also the output (Merity et al., 2018) as well as", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 462, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 477 + ], + "score": 1.0, + "content": "sub-words (Buckman & Neubig, 2018; Mielke & Eisner, 2018). Word-based models are particularly", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "challenging since computing probabilities for all 800K words of the BILLION WORD benchmark is", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 486, + 380, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 380, + 498 + ], + "score": 1.0, + "content": "still a substantial part of the overall computation (Chen et al., 2016).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 441, + 505, + 498 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 504, + 568 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 506, + 515 + ], + "score": 1.0, + "content": "A popular approach to lower the computational burden is to structure the output vocabulary so that", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 513, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 525 + ], + "score": 1.0, + "content": "not all probabilities need to be computed. The hierarchical softmax does this by introducing latent", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 524, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 536 + ], + "score": 1.0, + "content": "variables or clusters to simplify normalization (Goodman, 2001; Morin & Bengio, 2005; Mikolov", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 536, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 505, + 547 + ], + "score": 1.0, + "content": "et al., 2011). This has been further improved by the adaptive softmax which introduces a variable", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 546, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 559 + ], + "score": 1.0, + "content": "capacity scheme for output word embeddings, assigning more parameters to frequent words and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 557, + 315, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 315, + 570 + ], + "score": 1.0, + "content": "fewer parameters to rare words (Grave et al., 2017).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 502, + 506, + 570 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 574, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "In this paper, we introduce adaptive input embeddings which extend the adaptive softmax to input", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "word representations. This factorization assigns more capacity to frequent words and reduces the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 595, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 609 + ], + "score": 1.0, + "content": "capacity for less frequent words with the benefit of reducing overfitting to rare words. For a com-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 607, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 505, + 618 + ], + "score": 1.0, + "content": "petitive setup on the BILLION WORD benchmark, adaptive input embeddings reduce the number", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 304, + 630 + ], + "score": 1.0, + "content": "of parameters in the input and output layers by", + "type": "text" + }, + { + "bbox": [ + 304, + 618, + 324, + 628 + ], + "score": 0.86, + "content": "23 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "while achieving higher accuracy over fixed", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 628, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 641 + ], + "score": 1.0, + "content": "size embeddings. When the adaptive input representations are tied with an adaptive softmax in the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 640, + 378, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 354, + 652 + ], + "score": 1.0, + "content": "output, then the number of parameters is reduced by a total of", + "type": "text" + }, + { + "bbox": [ + 355, + 640, + 374, + 650 + ], + "score": 0.87, + "content": "61 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 640, + 378, + 652 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 574, + 506, + 652 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 657, + 504, + 712 + ], + "lines": [ + { + "bbox": [ + 106, + 657, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 505, + 669 + ], + "score": 1.0, + "content": "Our experiments compare models based on word inputs, character inputs, as well as sub-word units", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "score": 1.0, + "content": "using a self-attention architecture (Vaswani et al., 2017). We show that models with adaptive word", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 678, + 505, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 692 + ], + "score": 1.0, + "content": "representations can outperform very strong character-based models while training more than twice", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 689, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 703 + ], + "score": 1.0, + "content": "as fast. We also substantially improve adaptive softmax by introducing additional dropout regular-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "ization in the tail projection. On the WIKITEXT-103 benchmark we achieve a perplexity of 18.7, a", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "reduction of 10.5 perplexity over the previously best published result. On the larger BILLION WORD", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 336, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 348 + ], + "score": 1.0, + "content": "benchmark our best model with adaptive input embeddings achieves 23.02 perplexity, a reduction", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 347, + 380, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 380, + 360 + ], + "score": 1.0, + "content": "of nearly 5 perplexity over the next best previously published result.", + "type": "text", + "cross_page": true + } + ], + "index": 18 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 657, + 506, + 713 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 188, + 84, + 423, + 259 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 188, + 84, + 423, + 259 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 188, + 84, + 423, + 259 + ], + "spans": [ + { + "bbox": [ + 188, + 84, + 423, + 259 + ], + "score": 0.971, + "type": "image", + "image_path": "52708bad0c295bb0584ec5d4f2259eabd3c9b29f269a819bd44546cef32e3818.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 188, + 84, + 423, + 97.46153846153847 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 188, + 97.46153846153847, + 423, + 110.92307692307693 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 188, + 110.92307692307693, + 423, + 124.3846153846154 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 188, + 124.3846153846154, + 423, + 137.84615384615387 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 188, + 137.84615384615387, + 423, + 151.30769230769232 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 188, + 151.30769230769232, + 423, + 164.76923076923077 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 188, + 164.76923076923077, + 423, + 178.23076923076923 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 188, + 178.23076923076923, + 423, + 191.69230769230768 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 188, + 191.69230769230768, + 423, + 205.15384615384613 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 188, + 205.15384615384613, + 423, + 218.61538461538458 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 188, + 218.61538461538458, + 423, + 232.07692307692304 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 188, + 232.07692307692304, + 423, + 245.5384615384615 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 188, + 245.5384615384615, + 423, + 258.99999999999994 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 271, + 504, + 305 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 455, + 284 + ], + "score": 1.0, + "content": "Figure 1: Illustration of adaptive input representations. Words are assigned to clusters", + "type": "text" + }, + { + "bbox": [ + 455, + 272, + 465, + 282 + ], + "score": 0.87, + "content": "\\nu _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "based on", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "their frequency which determines the size of the representations. Embeddings are projected to a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 294, + 317, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 188, + 305 + ], + "score": 1.0, + "content": "common dimension", + "type": "text" + }, + { + "bbox": [ + 188, + 294, + 195, + 303 + ], + "score": 0.76, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 294, + 317, + 305 + ], + "score": 1.0, + "content": "before being fed to the model.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + } + ], + "index": 10.0 + }, + { + "type": "text", + "bbox": [ + 107, + 325, + 505, + 358 + ], + "lines": [ + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "reduction of 10.5 perplexity over the previously best published result. On the larger BILLION WORD", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 336, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 348 + ], + "score": 1.0, + "content": "benchmark our best model with adaptive input embeddings achieves 23.02 perplexity, a reduction", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 347, + 380, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 380, + 360 + ], + "score": 1.0, + "content": "of nearly 5 perplexity over the next best previously published result.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 108, + 374, + 211, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 373, + 213, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 213, + 389 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 398, + 505, + 443 + ], + "lines": [ + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "score": 1.0, + "content": "Adaptive word representations are inspired by the adaptive softmax work Grave et al. (2017) which", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 409, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 505, + 422 + ], + "score": 1.0, + "content": "first described a GPU friendly way to construct a hierarchical softmax and showed that it performs", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 421, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 506, + 434 + ], + "score": 1.0, + "content": "very competitively compared to a full softmax, while offering significantly faster speed and a lower", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 432, + 182, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 182, + 444 + ], + "score": 1.0, + "content": "memory footprint.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 448, + 505, + 504 + ], + "lines": [ + { + "bbox": [ + 106, + 448, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 505, + 460 + ], + "score": 1.0, + "content": "Merity et al. (2018) use a modified version of adaptive softmax which does not reduce the dimen-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 460, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 505, + 472 + ], + "score": 1.0, + "content": "sionality of less frequent words in order to be able to share output embeddings with the input. This", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "setup is akin to a hierarchical softmax with tied weights. We show that variable-sized input embed-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 480, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 496 + ], + "score": 1.0, + "content": "dings can perform better than fixed sized embeddings. Furthermore, this also enables weight sharing", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 492, + 261, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 261, + 505 + ], + "score": 1.0, + "content": "with an adaptive softmax output layer.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 509, + 505, + 575 + ], + "lines": [ + { + "bbox": [ + 106, + 510, + 504, + 521 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 504, + 521 + ], + "score": 1.0, + "content": "Merity et al. (2018) evaluates both character-based and word-based factorizations but does not di-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 520, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 533 + ], + "score": 1.0, + "content": "rectly compare them to each other. We perform a direct comparison of word-based and character-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 531, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 544 + ], + "score": 1.0, + "content": "based input vocabularies and also compare to a sub-word factorization for both the input and output.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "Recently, Al-Rfou et al. (2018) demonstrated that self-attentional models can perform very well on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 553, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 565 + ], + "score": 1.0, + "content": "language modeling tasks where the input and output is both characters. We also consider word-based", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 564, + 159, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 159, + 576 + ], + "score": 1.0, + "content": "benchmarks.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31.5 + }, + { + "type": "title", + "bbox": [ + 108, + 591, + 315, + 604 + ], + "lines": [ + { + "bbox": [ + 104, + 590, + 317, + 606 + ], + "spans": [ + { + "bbox": [ + 104, + 590, + 317, + 606 + ], + "score": 1.0, + "content": "3 ADAPTIVE INPUT REPRESENTATIONS", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 615, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "The adaptive softmax exploits the fact that the distribution of word types in natural language follows", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "a Zipfian distribution in order to improve the computation of the output probabilities. We apply the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 639, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 505, + 650 + ], + "score": 1.0, + "content": "same intuition for input word embeddings with the motivation to reduce the number of parameters", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 649, + 317, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 317, + 661 + ], + "score": 1.0, + "content": "which frees up capacity for other parts of the model.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 664, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 453, + 678 + ], + "score": 1.0, + "content": "We define a number of clusters that partitions the frequency ordered vocabulary", + "type": "text" + }, + { + "bbox": [ + 453, + 666, + 505, + 677 + ], + "score": 0.91, + "content": "\\nu ~ = ~ \\nu _ { 1 } \\cup", + "type": "inline_equation" + } + ], + "index": 40 + }, + { + "bbox": [ + 107, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 107, + 677, + 186, + 689 + ], + "score": 0.92, + "content": "\\mathcal { V } _ { 2 } , \\ldots , \\mathcal { V } _ { n - 1 } \\cup \\mathcal { V } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 677, + 228, + 690 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 228, + 677, + 283, + 689 + ], + "score": 0.93, + "content": "\\nu _ { i } \\cap \\nu _ { j } = \\emptyset", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 677, + 299, + 690 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 299, + 677, + 319, + 688 + ], + "score": 0.89, + "content": "\\forall i , j", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 677, + 341, + 690 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 341, + 677, + 367, + 688 + ], + "score": 0.9, + "content": "i \\neq j", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 677, + 399, + 690 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 399, + 677, + 411, + 688 + ], + "score": 0.88, + "content": "\\nu _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "contains the most fre-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 174, + 700 + ], + "score": 1.0, + "content": "quent words and", + "type": "text" + }, + { + "bbox": [ + 174, + 689, + 187, + 699 + ], + "score": 0.87, + "content": "\\nu _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 687, + 354, + 700 + ], + "score": 1.0, + "content": "the least frequent words. We will refer to", + "type": "text" + }, + { + "bbox": [ + 354, + 689, + 366, + 699 + ], + "score": 0.88, + "content": "\\nu _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "as the head and to any subsequent", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 423, + 712 + ], + "score": 1.0, + "content": "clusters loosely as tail. We reduce the capacity for each cluster by a factor of", + "type": "text" + }, + { + "bbox": [ + 423, + 699, + 430, + 709 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 699, + 505, + 712 + ], + "score": 1.0, + "content": ". That is, if words", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 708, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 104, + 708, + 117, + 724 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 117, + 710, + 129, + 721 + ], + "score": 0.87, + "content": "\\nu _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 708, + 195, + 724 + ], + "score": 1.0, + "content": "have dimension", + "type": "text" + }, + { + "bbox": [ + 196, + 710, + 202, + 720 + ], + "score": 0.74, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 708, + 263, + 724 + ], + "score": 1.0, + "content": ", then words in", + "type": "text" + }, + { + "bbox": [ + 263, + 710, + 276, + 721 + ], + "score": 0.88, + "content": "\\nu _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 708, + 343, + 724 + ], + "score": 1.0, + "content": "have dimension", + "type": "text" + }, + { + "bbox": [ + 344, + 709, + 364, + 722 + ], + "score": 0.92, + "content": "\\frac { d } { k ^ { n - 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 708, + 436, + 724 + ], + "score": 1.0, + "content": ". We typically set", + "type": "text" + }, + { + "bbox": [ + 437, + 710, + 463, + 720 + ], + "score": 0.89, + "content": "k = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 708, + 506, + 724 + ], + "score": 1.0, + "content": "following", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 721, + 187, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 187, + 732 + ], + "score": 1.0, + "content": "Grave et al. (2017).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 188, + 84, + 423, + 259 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 188, + 84, + 423, + 259 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 188, + 84, + 423, + 259 + ], + "spans": [ + { + "bbox": [ + 188, + 84, + 423, + 259 + ], + "score": 0.971, + "type": "image", + "image_path": "52708bad0c295bb0584ec5d4f2259eabd3c9b29f269a819bd44546cef32e3818.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 188, + 84, + 423, + 97.46153846153847 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 188, + 97.46153846153847, + 423, + 110.92307692307693 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 188, + 110.92307692307693, + 423, + 124.3846153846154 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 188, + 124.3846153846154, + 423, + 137.84615384615387 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 188, + 137.84615384615387, + 423, + 151.30769230769232 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 188, + 151.30769230769232, + 423, + 164.76923076923077 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 188, + 164.76923076923077, + 423, + 178.23076923076923 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 188, + 178.23076923076923, + 423, + 191.69230769230768 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 188, + 191.69230769230768, + 423, + 205.15384615384613 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 188, + 205.15384615384613, + 423, + 218.61538461538458 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 188, + 218.61538461538458, + 423, + 232.07692307692304 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 188, + 232.07692307692304, + 423, + 245.5384615384615 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 188, + 245.5384615384615, + 423, + 258.99999999999994 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 271, + 504, + 305 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 455, + 284 + ], + "score": 1.0, + "content": "Figure 1: Illustration of adaptive input representations. Words are assigned to clusters", + "type": "text" + }, + { + "bbox": [ + 455, + 272, + 465, + 282 + ], + "score": 0.87, + "content": "\\nu _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "based on", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "their frequency which determines the size of the representations. Embeddings are projected to a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 294, + 317, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 188, + 305 + ], + "score": 1.0, + "content": "common dimension", + "type": "text" + }, + { + "bbox": [ + 188, + 294, + 195, + 303 + ], + "score": 0.76, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 294, + 317, + 305 + ], + "score": 1.0, + "content": "before being fed to the model.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + } + ], + "index": 10.0 + }, + { + "type": "text", + "bbox": [ + 107, + 325, + 505, + 358 + ], + "lines": [], + "index": 17, + "bbox_fs": [ + 105, + 325, + 505, + 360 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 374, + 211, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 373, + 213, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 213, + 389 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 398, + 505, + 443 + ], + "lines": [ + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "score": 1.0, + "content": "Adaptive word representations are inspired by the adaptive softmax work Grave et al. (2017) which", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 409, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 505, + 422 + ], + "score": 1.0, + "content": "first described a GPU friendly way to construct a hierarchical softmax and showed that it performs", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 421, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 506, + 434 + ], + "score": 1.0, + "content": "very competitively compared to a full softmax, while offering significantly faster speed and a lower", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 432, + 182, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 182, + 444 + ], + "score": 1.0, + "content": "memory footprint.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 399, + 506, + 444 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 448, + 505, + 504 + ], + "lines": [ + { + "bbox": [ + 106, + 448, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 505, + 460 + ], + "score": 1.0, + "content": "Merity et al. (2018) use a modified version of adaptive softmax which does not reduce the dimen-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 460, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 505, + 472 + ], + "score": 1.0, + "content": "sionality of less frequent words in order to be able to share output embeddings with the input. This", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "setup is akin to a hierarchical softmax with tied weights. We show that variable-sized input embed-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 480, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 496 + ], + "score": 1.0, + "content": "dings can perform better than fixed sized embeddings. Furthermore, this also enables weight sharing", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 492, + 261, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 261, + 505 + ], + "score": 1.0, + "content": "with an adaptive softmax output layer.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 448, + 505, + 505 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 509, + 505, + 575 + ], + "lines": [ + { + "bbox": [ + 106, + 510, + 504, + 521 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 504, + 521 + ], + "score": 1.0, + "content": "Merity et al. (2018) evaluates both character-based and word-based factorizations but does not di-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 520, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 533 + ], + "score": 1.0, + "content": "rectly compare them to each other. We perform a direct comparison of word-based and character-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 531, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 544 + ], + "score": 1.0, + "content": "based input vocabularies and also compare to a sub-word factorization for both the input and output.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "Recently, Al-Rfou et al. (2018) demonstrated that self-attentional models can perform very well on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 553, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 565 + ], + "score": 1.0, + "content": "language modeling tasks where the input and output is both characters. We also consider word-based", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 564, + 159, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 159, + 576 + ], + "score": 1.0, + "content": "benchmarks.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 510, + 505, + 576 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 591, + 315, + 604 + ], + "lines": [ + { + "bbox": [ + 104, + 590, + 317, + 606 + ], + "spans": [ + { + "bbox": [ + 104, + 590, + 317, + 606 + ], + "score": 1.0, + "content": "3 ADAPTIVE INPUT REPRESENTATIONS", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 615, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "The adaptive softmax exploits the fact that the distribution of word types in natural language follows", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "a Zipfian distribution in order to improve the computation of the output probabilities. We apply the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 639, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 505, + 650 + ], + "score": 1.0, + "content": "same intuition for input word embeddings with the motivation to reduce the number of parameters", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 649, + 317, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 317, + 661 + ], + "score": 1.0, + "content": "which frees up capacity for other parts of the model.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 615, + 506, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 664, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 453, + 678 + ], + "score": 1.0, + "content": "We define a number of clusters that partitions the frequency ordered vocabulary", + "type": "text" + }, + { + "bbox": [ + 453, + 666, + 505, + 677 + ], + "score": 0.91, + "content": "\\nu ~ = ~ \\nu _ { 1 } \\cup", + "type": "inline_equation" + } + ], + "index": 40 + }, + { + "bbox": [ + 107, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 107, + 677, + 186, + 689 + ], + "score": 0.92, + "content": "\\mathcal { V } _ { 2 } , \\ldots , \\mathcal { V } _ { n - 1 } \\cup \\mathcal { V } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 677, + 228, + 690 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 228, + 677, + 283, + 689 + ], + "score": 0.93, + "content": "\\nu _ { i } \\cap \\nu _ { j } = \\emptyset", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 677, + 299, + 690 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 299, + 677, + 319, + 688 + ], + "score": 0.89, + "content": "\\forall i , j", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 677, + 341, + 690 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 341, + 677, + 367, + 688 + ], + "score": 0.9, + "content": "i \\neq j", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 677, + 399, + 690 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 399, + 677, + 411, + 688 + ], + "score": 0.88, + "content": "\\nu _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "contains the most fre-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 174, + 700 + ], + "score": 1.0, + "content": "quent words and", + "type": "text" + }, + { + "bbox": [ + 174, + 689, + 187, + 699 + ], + "score": 0.87, + "content": "\\nu _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 687, + 354, + 700 + ], + "score": 1.0, + "content": "the least frequent words. We will refer to", + "type": "text" + }, + { + "bbox": [ + 354, + 689, + 366, + 699 + ], + "score": 0.88, + "content": "\\nu _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "as the head and to any subsequent", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 423, + 712 + ], + "score": 1.0, + "content": "clusters loosely as tail. We reduce the capacity for each cluster by a factor of", + "type": "text" + }, + { + "bbox": [ + 423, + 699, + 430, + 709 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 699, + 505, + 712 + ], + "score": 1.0, + "content": ". That is, if words", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 708, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 104, + 708, + 117, + 724 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 117, + 710, + 129, + 721 + ], + "score": 0.87, + "content": "\\nu _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 708, + 195, + 724 + ], + "score": 1.0, + "content": "have dimension", + "type": "text" + }, + { + "bbox": [ + 196, + 710, + 202, + 720 + ], + "score": 0.74, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 708, + 263, + 724 + ], + "score": 1.0, + "content": ", then words in", + "type": "text" + }, + { + "bbox": [ + 263, + 710, + 276, + 721 + ], + "score": 0.88, + "content": "\\nu _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 708, + 343, + 724 + ], + "score": 1.0, + "content": "have dimension", + "type": "text" + }, + { + "bbox": [ + 344, + 709, + 364, + 722 + ], + "score": 0.92, + "content": "\\frac { d } { k ^ { n - 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 708, + 436, + 724 + ], + "score": 1.0, + "content": ". We typically set", + "type": "text" + }, + { + "bbox": [ + 437, + 710, + 463, + 720 + ], + "score": 0.89, + "content": "k = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 708, + 506, + 724 + ], + "score": 1.0, + "content": "following", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 721, + 187, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 187, + 732 + ], + "score": 1.0, + "content": "Grave et al. (2017).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42.5, + "bbox_fs": [ + 104, + 664, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 81, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 104, + 78, + 507, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 78, + 233, + 96 + ], + "score": 1.0, + "content": "Next, we add linear projections", + "type": "text" + }, + { + "bbox": [ + 234, + 80, + 377, + 94 + ], + "score": 0.9, + "content": "W _ { 1 } \\in \\mathbb { R } ^ { d \\times d } , \\ldots , W _ { n } \\in \\mathbb { R } ^ { d / k ^ { n - 1 } \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 78, + 507, + 96 + ], + "score": 1.0, + "content": "to map the embeddings of each", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 190, + 106 + ], + "score": 1.0, + "content": "cluster to dimension", + "type": "text" + }, + { + "bbox": [ + 190, + 94, + 197, + 104 + ], + "score": 0.74, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "so that the concatenated output of the adaptive input embedding layer can be", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 363, + 117 + ], + "score": 1.0, + "content": "easily used by the subsequent model (Figure 1). We also project", + "type": "text" + }, + { + "bbox": [ + 363, + 105, + 375, + 116 + ], + "score": 0.88, + "content": "\\nu _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 105, + 495, + 117 + ], + "score": 1.0, + "content": "which already has dimension", + "type": "text" + }, + { + "bbox": [ + 495, + 105, + 501, + 114 + ], + "score": 0.65, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 105, + 505, + 117 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 108, + 121, + 503, + 155 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 504, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 504, + 133 + ], + "score": 1.0, + "content": "When presented with a number of input words, the adaptive input embedding layer partitions the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "words into the various clusters, performs separate lookups in the embedding tables and then projects", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 428, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 160, + 156 + ], + "score": 1.0, + "content": "to dimension", + "type": "text" + }, + { + "bbox": [ + 161, + 144, + 167, + 153 + ], + "score": 0.77, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 144, + 428, + 156 + ], + "score": 1.0, + "content": ", followed by concatenating the embeddings in the original order.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 168, + 505, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 168, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 106, + 169, + 482, + 180 + ], + "score": 1.0, + "content": "Weight sharing. When the output layer is an adaptive softmax with the same partition of", + "type": "text" + }, + { + "bbox": [ + 482, + 168, + 502, + 179 + ], + "score": 0.3, + "content": "\\nu , d ,", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 169, + 505, + 180 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 179, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 124, + 192 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 180, + 132, + 190 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 179, + 505, + 192 + ], + "score": 1.0, + "content": "as the adaptive input layer, then we can tie the weights (Inan et al., 2016; Press & Wolf,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "score": 1.0, + "content": "2017). This further reduces the number of parameters and can simultaneously improve performance", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 202, + 504, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 202, + 452, + 214 + ], + "score": 1.0, + "content": "(§5). We can share both the parameters for the actual words as well as the projections", + "type": "text" + }, + { + "bbox": [ + 452, + 202, + 504, + 213 + ], + "score": 0.91, + "content": "W _ { 1 } , \\ldots , W _ { n }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 212, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 225 + ], + "score": 1.0, + "content": "Sharing the word embeddings is straightforward except for the head where the adaptive softmax has", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 224, + 474, + 236 + ], + "spans": [ + { + "bbox": [ + 107, + 224, + 131, + 234 + ], + "score": 0.85, + "content": "n - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 224, + 474, + 236 + ], + "score": 1.0, + "content": "additional embeddings for the remaining clusters which are not shared with the input.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 240, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 505, + 253 + ], + "score": 1.0, + "content": "We share all projections, except for the head projection which is not available in the adaptive softmax", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "since the model output is directly multiplied with the output word embeddings for the head band.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 262, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 505, + 275 + ], + "score": 1.0, + "content": "Performance decreased when we added a head projection to the adaptive softmax in the output,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 273, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 505, + 286 + ], + "score": 1.0, + "content": "regardless of when it was shared or not. Sharing both the word embeddings as well as the projections", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 284, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 505, + 297 + ], + "score": 1.0, + "content": "performed very well on WIKITEXT-103 but on BILLION WORD we only share the word embeddings", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 295, + 347, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 347, + 307 + ], + "score": 1.0, + "content": "as we found that this performed better on the validation set.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 108, + 324, + 244, + 337 + ], + "lines": [ + { + "bbox": [ + 105, + 324, + 245, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 245, + 338 + ], + "score": 1.0, + "content": "4 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 107, + 350, + 164, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 166, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 166, + 362 + ], + "score": 1.0, + "content": "4.1 MODEL", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 106, + 371, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 371, + 505, + 383 + ], + "score": 1.0, + "content": "We follow most of the architectural choices described in Vaswani et al. (2017) but use only a decoder", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 383, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 424, + 394 + ], + "score": 1.0, + "content": "network. We add sinusoidal position embeddings to the input layer and stack", + "type": "text" + }, + { + "bbox": [ + 424, + 383, + 460, + 393 + ], + "score": 0.9, + "content": "N = 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 383, + 505, + 394 + ], + "score": 1.0, + "content": "blocks for", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "both BILLION WORD and WIKITEXT-103. Each block contains two sub-blocks: the first is a multi-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 405, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 238, + 415 + ], + "score": 1.0, + "content": "head self-attention module with", + "type": "text" + }, + { + "bbox": [ + 239, + 405, + 275, + 415 + ], + "score": 0.88, + "content": "H = 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 405, + 505, + 415 + ], + "score": 1.0, + "content": "heads. The second sub-block is a feed-forward module", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 415, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 180, + 428 + ], + "score": 1.0, + "content": "(FFN) of the form", + "type": "text" + }, + { + "bbox": [ + 180, + 415, + 295, + 428 + ], + "score": 0.91, + "content": "R e L U ( W _ { 1 } X + b _ { 1 } ) W _ { 2 } + b _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 415, + 324, + 428 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 324, + 415, + 381, + 426 + ], + "score": 0.91, + "content": "W _ { 1 } \\in \\mathbb { R } ^ { e \\times e _ { f f } }", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 415, + 385, + 428 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 386, + 415, + 444, + 426 + ], + "score": 0.92, + "content": "W _ { 1 } \\in \\mathbb { R } ^ { e _ { f f } \\times e }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 415, + 462, + 428 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 462, + 416, + 501, + 426 + ], + "score": 0.91, + "content": "e = 1 0 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 415, + 506, + 428 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 426, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 107, + 427, + 156, + 439 + ], + "score": 0.91, + "content": "e _ { f f } = 4 0 9 6", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 426, + 506, + 439 + ], + "score": 1.0, + "content": "unless otherwise stated. Different to Vaswani et al. (2017) we apply layer normalization", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "score": 1.0, + "content": "before the self-attention and FFN blocks instead of after, as we find it leads to more effective training.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 448, + 385, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 385, + 460 + ], + "score": 1.0, + "content": "Sub-blocks are surrounded by a residual connection (He et al., 2015).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 465, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 106, + 464, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 478 + ], + "score": 1.0, + "content": "We use a dropout rate of 0.1 and attention dropout of 0.1 for BILLION WORD models, and increase", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "score": 1.0, + "content": "regularization for WIKITEXT-103 by using dropout 0.3, and 0.1 ReLU dropout as well as attention", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "dropout 0.1. We use the same hyperparameters for all models trained on the same dataset in order to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "enable a like for like comparison. When the dimensionality of the input or output layer differs from", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 509, + 324, + 522 + ], + "spans": [ + { + "bbox": [ + 107, + 511, + 112, + 519 + ], + "score": 0.58, + "content": "e", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 509, + 324, + 522 + ], + "score": 1.0, + "content": ", then we add a simple linear projection with no bias.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30 + }, + { + "type": "title", + "bbox": [ + 107, + 535, + 176, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 177, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 177, + 549 + ], + "score": 1.0, + "content": "4.2 DATASETS", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 108, + 557, + 503, + 590 + ], + "lines": [ + { + "bbox": [ + 107, + 556, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 107, + 556, + 505, + 569 + ], + "score": 1.0, + "content": "We experiment on the BILLION WORD benchmark and WIKITEXT-103. BILLION WORD contains", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 568, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 505, + 581 + ], + "score": 1.0, + "content": "768M word tokens and has a vocabulary of about 800K word types, which corresponds to words", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 578, + 385, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 385, + 591 + ], + "score": 1.0, + "content": "with more than 3 occurrences in the training set (Chelba et al., 2013).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 108, + 596, + 504, + 629 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 609 + ], + "score": 1.0, + "content": "The training data of WIKITEXT-103 comprises about 100M tokens and a vocabulary of around 260K,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "score": 1.0, + "content": "corresponding to types with more than 3 occurrences in the training data (Merity et al., 2016). The", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 618, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 482, + 631 + ], + "score": 1.0, + "content": "dataset is composed of shuffled Wikipedia articles where the context carries across sentences.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "title", + "bbox": [ + 107, + 644, + 177, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 178, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 178, + 657 + ], + "score": 1.0, + "content": "4.3 BATCHING", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 709 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "For BILLION WORD we batch individual sentences since the corpus does not contain document", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 677, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 688 + ], + "score": 1.0, + "content": "structure. For WIKITEXT-103 we partition the training data into blocks of 512 contiguous tokens", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "ignoring document boundaries. Evaluation is the same except that we require blocks to contain", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 699, + 294, + 710 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 294, + 710 + ], + "score": 1.0, + "content": "complete sentences totaling up to 512 tokens.2", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 116, + 721, + 457, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 720, + 459, + 733 + ], + "spans": [ + { + "bbox": [ + 118, + 720, + 459, + 733 + ], + "score": 1.0, + "content": "2 Respecting document boundaries may lead to better results and we leave this to future work.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 81, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 104, + 78, + 507, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 78, + 233, + 96 + ], + "score": 1.0, + "content": "Next, we add linear projections", + "type": "text" + }, + { + "bbox": [ + 234, + 80, + 377, + 94 + ], + "score": 0.9, + "content": "W _ { 1 } \\in \\mathbb { R } ^ { d \\times d } , \\ldots , W _ { n } \\in \\mathbb { R } ^ { d / k ^ { n - 1 } \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 78, + 507, + 96 + ], + "score": 1.0, + "content": "to map the embeddings of each", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 190, + 106 + ], + "score": 1.0, + "content": "cluster to dimension", + "type": "text" + }, + { + "bbox": [ + 190, + 94, + 197, + 104 + ], + "score": 0.74, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "so that the concatenated output of the adaptive input embedding layer can be", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 363, + 117 + ], + "score": 1.0, + "content": "easily used by the subsequent model (Figure 1). We also project", + "type": "text" + }, + { + "bbox": [ + 363, + 105, + 375, + 116 + ], + "score": 0.88, + "content": "\\nu _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 105, + 495, + 117 + ], + "score": 1.0, + "content": "which already has dimension", + "type": "text" + }, + { + "bbox": [ + 495, + 105, + 501, + 114 + ], + "score": 0.65, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 105, + 505, + 117 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 104, + 78, + 507, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 121, + 503, + 155 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 504, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 504, + 133 + ], + "score": 1.0, + "content": "When presented with a number of input words, the adaptive input embedding layer partitions the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "words into the various clusters, performs separate lookups in the embedding tables and then projects", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 428, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 160, + 156 + ], + "score": 1.0, + "content": "to dimension", + "type": "text" + }, + { + "bbox": [ + 161, + 144, + 167, + 153 + ], + "score": 0.77, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 144, + 428, + 156 + ], + "score": 1.0, + "content": ", followed by concatenating the embeddings in the original order.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 120, + 505, + 156 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 168, + 505, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 168, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 106, + 169, + 482, + 180 + ], + "score": 1.0, + "content": "Weight sharing. When the output layer is an adaptive softmax with the same partition of", + "type": "text" + }, + { + "bbox": [ + 482, + 168, + 502, + 179 + ], + "score": 0.3, + "content": "\\nu , d ,", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 169, + 505, + 180 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 179, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 124, + 192 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 180, + 132, + 190 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 179, + 505, + 192 + ], + "score": 1.0, + "content": "as the adaptive input layer, then we can tie the weights (Inan et al., 2016; Press & Wolf,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "score": 1.0, + "content": "2017). This further reduces the number of parameters and can simultaneously improve performance", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 202, + 504, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 202, + 452, + 214 + ], + "score": 1.0, + "content": "(§5). We can share both the parameters for the actual words as well as the projections", + "type": "text" + }, + { + "bbox": [ + 452, + 202, + 504, + 213 + ], + "score": 0.91, + "content": "W _ { 1 } , \\ldots , W _ { n }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 212, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 225 + ], + "score": 1.0, + "content": "Sharing the word embeddings is straightforward except for the head where the adaptive softmax has", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 224, + 474, + 236 + ], + "spans": [ + { + "bbox": [ + 107, + 224, + 131, + 234 + ], + "score": 0.85, + "content": "n - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 224, + 474, + 236 + ], + "score": 1.0, + "content": "additional embeddings for the remaining clusters which are not shared with the input.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 168, + 505, + 236 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 240, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 505, + 253 + ], + "score": 1.0, + "content": "We share all projections, except for the head projection which is not available in the adaptive softmax", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "since the model output is directly multiplied with the output word embeddings for the head band.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 262, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 505, + 275 + ], + "score": 1.0, + "content": "Performance decreased when we added a head projection to the adaptive softmax in the output,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 273, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 505, + 286 + ], + "score": 1.0, + "content": "regardless of when it was shared or not. Sharing both the word embeddings as well as the projections", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 284, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 505, + 297 + ], + "score": 1.0, + "content": "performed very well on WIKITEXT-103 but on BILLION WORD we only share the word embeddings", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 295, + 347, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 347, + 307 + ], + "score": 1.0, + "content": "as we found that this performed better on the validation set.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 240, + 505, + 307 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 324, + 244, + 337 + ], + "lines": [ + { + "bbox": [ + 105, + 324, + 245, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 245, + 338 + ], + "score": 1.0, + "content": "4 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 107, + 350, + 164, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 166, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 166, + 362 + ], + "score": 1.0, + "content": "4.1 MODEL", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 106, + 371, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 371, + 505, + 383 + ], + "score": 1.0, + "content": "We follow most of the architectural choices described in Vaswani et al. (2017) but use only a decoder", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 383, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 424, + 394 + ], + "score": 1.0, + "content": "network. We add sinusoidal position embeddings to the input layer and stack", + "type": "text" + }, + { + "bbox": [ + 424, + 383, + 460, + 393 + ], + "score": 0.9, + "content": "N = 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 383, + 505, + 394 + ], + "score": 1.0, + "content": "blocks for", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "both BILLION WORD and WIKITEXT-103. Each block contains two sub-blocks: the first is a multi-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 405, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 238, + 415 + ], + "score": 1.0, + "content": "head self-attention module with", + "type": "text" + }, + { + "bbox": [ + 239, + 405, + 275, + 415 + ], + "score": 0.88, + "content": "H = 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 405, + 505, + 415 + ], + "score": 1.0, + "content": "heads. The second sub-block is a feed-forward module", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 415, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 180, + 428 + ], + "score": 1.0, + "content": "(FFN) of the form", + "type": "text" + }, + { + "bbox": [ + 180, + 415, + 295, + 428 + ], + "score": 0.91, + "content": "R e L U ( W _ { 1 } X + b _ { 1 } ) W _ { 2 } + b _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 415, + 324, + 428 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 324, + 415, + 381, + 426 + ], + "score": 0.91, + "content": "W _ { 1 } \\in \\mathbb { R } ^ { e \\times e _ { f f } }", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 415, + 385, + 428 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 386, + 415, + 444, + 426 + ], + "score": 0.92, + "content": "W _ { 1 } \\in \\mathbb { R } ^ { e _ { f f } \\times e }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 415, + 462, + 428 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 462, + 416, + 501, + 426 + ], + "score": 0.91, + "content": "e = 1 0 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 415, + 506, + 428 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 426, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 107, + 427, + 156, + 439 + ], + "score": 0.91, + "content": "e _ { f f } = 4 0 9 6", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 426, + 506, + 439 + ], + "score": 1.0, + "content": "unless otherwise stated. Different to Vaswani et al. (2017) we apply layer normalization", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "score": 1.0, + "content": "before the self-attention and FFN blocks instead of after, as we find it leads to more effective training.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 448, + 385, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 385, + 460 + ], + "score": 1.0, + "content": "Sub-blocks are surrounded by a residual connection (He et al., 2015).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 371, + 506, + 460 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 465, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 106, + 464, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 478 + ], + "score": 1.0, + "content": "We use a dropout rate of 0.1 and attention dropout of 0.1 for BILLION WORD models, and increase", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "score": 1.0, + "content": "regularization for WIKITEXT-103 by using dropout 0.3, and 0.1 ReLU dropout as well as attention", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "dropout 0.1. We use the same hyperparameters for all models trained on the same dataset in order to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "enable a like for like comparison. When the dimensionality of the input or output layer differs from", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 509, + 324, + 522 + ], + "spans": [ + { + "bbox": [ + 107, + 511, + 112, + 519 + ], + "score": 0.58, + "content": "e", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 509, + 324, + 522 + ], + "score": 1.0, + "content": ", then we add a simple linear projection with no bias.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 464, + 505, + 522 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 535, + 176, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 177, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 177, + 549 + ], + "score": 1.0, + "content": "4.2 DATASETS", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 108, + 557, + 503, + 590 + ], + "lines": [ + { + "bbox": [ + 107, + 556, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 107, + 556, + 505, + 569 + ], + "score": 1.0, + "content": "We experiment on the BILLION WORD benchmark and WIKITEXT-103. BILLION WORD contains", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 568, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 505, + 581 + ], + "score": 1.0, + "content": "768M word tokens and has a vocabulary of about 800K word types, which corresponds to words", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 578, + 385, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 385, + 591 + ], + "score": 1.0, + "content": "with more than 3 occurrences in the training set (Chelba et al., 2013).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 106, + 556, + 505, + 591 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 596, + 504, + 629 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 609 + ], + "score": 1.0, + "content": "The training data of WIKITEXT-103 comprises about 100M tokens and a vocabulary of around 260K,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "score": 1.0, + "content": "corresponding to types with more than 3 occurrences in the training data (Merity et al., 2016). The", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 618, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 482, + 631 + ], + "score": 1.0, + "content": "dataset is composed of shuffled Wikipedia articles where the context carries across sentences.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 106, + 595, + 505, + 631 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 644, + 177, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 178, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 178, + 657 + ], + "score": 1.0, + "content": "4.3 BATCHING", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 709 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "For BILLION WORD we batch individual sentences since the corpus does not contain document", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 677, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 688 + ], + "score": 1.0, + "content": "structure. For WIKITEXT-103 we partition the training data into blocks of 512 contiguous tokens", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "ignoring document boundaries. Evaluation is the same except that we require blocks to contain", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 699, + 294, + 710 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 294, + 710 + ], + "score": 1.0, + "content": "complete sentences totaling up to 512 tokens.2", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 666, + 505, + 710 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 181 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 383, + 95 + ], + "score": 1.0, + "content": "We limit the number of tokens per GPU to a maximum threshold", + "type": "text" + }, + { + "bbox": [ + 383, + 83, + 393, + 93 + ], + "score": 0.74, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "per GPU. That is, we add", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "examples of similar length until we reach this threshold. When we train on multiple GPUs, each", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 170, + 117 + ], + "score": 1.0, + "content": "GPU processes", + "type": "text" + }, + { + "bbox": [ + 171, + 105, + 180, + 114 + ], + "score": 0.77, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "tokens using the same model parameters. This increases the effective batch size", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 280, + 127 + ], + "score": 1.0, + "content": "to the product of the number of GPUs and", + "type": "text" + }, + { + "bbox": [ + 281, + 116, + 290, + 125 + ], + "score": 0.7, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 115, + 442, + 127 + ], + "score": 1.0, + "content": ". For BILLION WORD models we use", + "type": "text" + }, + { + "bbox": [ + 442, + 115, + 487, + 126 + ], + "score": 0.89, + "content": "B = 2 0 4 8", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 115, + 506, + 127 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 506, + 138 + ], + "score": 1.0, + "content": "typically train on 32 GPUs, giving an effective batch size of 65K tokens. The smaller vocabulary of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 248, + 149 + ], + "score": 1.0, + "content": "WIKITEXT-103 enables increasing", + "type": "text" + }, + { + "bbox": [ + 249, + 138, + 258, + 147 + ], + "score": 0.78, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 137, + 505, + 149 + ], + "score": 1.0, + "content": "to 4096 and we train on 8 GPUs. We found that large batch", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 149, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 160 + ], + "score": 1.0, + "content": "training is beneficial for this dataset and we therefore accumulate gradient updates over two batches", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 506, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 172 + ], + "score": 1.0, + "content": "before committing a parameter update (Ott et al., 2018a). This gives an effective batch size of 65K", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 170, + 216, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 216, + 182 + ], + "score": 1.0, + "content": "tokens for WIKITEXT-103.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 107, + 203, + 336, + 213 + ], + "lines": [ + { + "bbox": [ + 106, + 202, + 338, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 202, + 338, + 214 + ], + "score": 1.0, + "content": "4.4 INPUT AND OUTPUT LAYER HYPERPARAMETERS", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "Embedding sizes. For fixed size word input layers and softmax output layers we generally use", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "score": 1.0, + "content": "embeddings of size 512 for WIKITEXT-103. When we use an adaptive softmax in the output and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "fixed size word embeddings for the input, then we use dimension 256 for the input embeddings for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "BILLION WORD and 64 for WIKITEXT-103. We tuned this choice on the validation set (Appendix", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 269, + 344, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 344, + 282 + ], + "score": 1.0, + "content": "A). BPE inputs and outputs have embeddings of size 1024.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 300, + 505, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 300, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 506, + 314 + ], + "score": 1.0, + "content": "Character CNN. We model character inputs by convolving the representations of all characters", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 311, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 506, + 325 + ], + "score": 1.0, + "content": "in a word following Kim et al. (2015) which applies several filters, then max pooling, a number of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 323, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 506, + 335 + ], + "score": 1.0, + "content": "highway layers and a projection. Character embeddings have size 128 and we apply seven filters of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 332, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 104, + 332, + 154, + 347 + ], + "score": 1.0, + "content": "size 1x128,", + "type": "text" + }, + { + "bbox": [ + 155, + 334, + 181, + 344 + ], + "score": 0.39, + "content": "2 \\mathrm { x } 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 332, + 364, + 347 + ], + "score": 1.0, + "content": ", 3x384, 4x512, 5x512, 6x512, 7x512, where", + "type": "text" + }, + { + "bbox": [ + 365, + 334, + 391, + 344 + ], + "score": 0.26, + "content": "3 \\mathrm { x } 1 2 8", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 332, + 506, + 347 + ], + "score": 1.0, + "content": "indicates a filter processing", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 345, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 356 + ], + "score": 1.0, + "content": "three characters that outputs 128 features. We use a single highway layer for WIKITEXT-103, and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 356, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 367 + ], + "score": 1.0, + "content": "two for BILLION WORD. We do not add start of word and end of word markers as they did not", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 366, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 379 + ], + "score": 1.0, + "content": "improve validation accuracy. We train on the same pre-processed data as the other models, with", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 377, + 297, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 297, + 390 + ], + "score": 1.0, + "content": "unknown tokens in both the inputs and outputs.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 408, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 407, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 422 + ], + "score": 1.0, + "content": "Adaptive input representations and adaptive softmax. We use an adaptive softmax output layer", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 419, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 506, + 433 + ], + "score": 1.0, + "content": "to train models with large word-based vocabularies. For adaptive word inputs and adaptive softmax,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 429, + 507, + 444 + ], + "spans": [ + { + "bbox": [ + 104, + 429, + 218, + 444 + ], + "score": 1.0, + "content": "we use embeddings of size", + "type": "text" + }, + { + "bbox": [ + 218, + 431, + 260, + 441 + ], + "score": 0.89, + "content": "d = 1 0 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 429, + 507, + 444 + ], + "score": 1.0, + "content": "for the head and reduce the size of subsequent clusters by a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 442, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 143, + 453 + ], + "score": 1.0, + "content": "factor of", + "type": "text" + }, + { + "bbox": [ + 144, + 442, + 169, + 452 + ], + "score": 0.89, + "content": "k = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 442, + 393, + 453 + ], + "score": 1.0, + "content": ". For WIKITEXT-103, we have three bands of size 20K", + "type": "text" + }, + { + "bbox": [ + 393, + 442, + 428, + 452 + ], + "score": 0.57, + "content": "\\scriptstyle ( \\mathrm { d } = 1 0 2 4 )", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 442, + 456, + 453 + ], + "score": 1.0, + "content": "), 40K", + "type": "text" + }, + { + "bbox": [ + 456, + 442, + 486, + 452 + ], + "score": 0.73, + "content": "\\mathrm { \\Delta } \\mathrm { Q } \\mathrm { = } 2 5 6 )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 442, + 505, + 453 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 453, + 499, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 134, + 465 + ], + "score": 1.0, + "content": "200K", + "type": "text" + }, + { + "bbox": [ + 134, + 453, + 157, + 464 + ], + "score": 0.71, + "content": "\\scriptstyle \\mathrm { { ( d = 6 4 ) } }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 453, + 324, + 465 + ], + "score": 1.0, + "content": "). For BILLION WORD the bands are 60K", + "type": "text" + }, + { + "bbox": [ + 324, + 453, + 357, + 463 + ], + "score": 0.78, + "content": "\\scriptstyle \\mathrm { d } = 1 0 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 453, + 390, + 465 + ], + "score": 1.0, + "content": "), 100K", + "type": "text" + }, + { + "bbox": [ + 391, + 453, + 420, + 464 + ], + "score": 0.7, + "content": "\\mathrm { \\Delta } \\mathrm { Q } { = } 2 5 6 \\mathrm { \\Delta }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 453, + 469, + 465 + ], + "score": 1.0, + "content": "), and 640K", + "type": "text" + }, + { + "bbox": [ + 470, + 453, + 493, + 464 + ], + "score": 0.72, + "content": "\\scriptstyle \\mathrm { { ( d = 6 4 ) } }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 453, + 499, + 465 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 483, + 505, + 538 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "score": 1.0, + "content": "Sub-word models. We learn a byte-pair encoding (BPE) of 32K codes on the training data of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "score": 1.0, + "content": "each benchmark (Sennrich et al., 2016). After applying the code to the training data we obtain", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 505, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 104, + 505, + 506, + 518 + ], + "score": 1.0, + "content": "a vocabulary of 33,337 tokens for WIKITEXT-103 and 32,347 tokens for BILLION WORD. BPE", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "score": 1.0, + "content": "input/output embeddings have size 1024. The final evaluation is in terms word-level perplexity to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 527, + 486, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 486, + 540 + ], + "score": 1.0, + "content": "be comparable to other models. The probability of a word is the product of the sub-word units.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30 + }, + { + "type": "title", + "bbox": [ + 108, + 559, + 195, + 570 + ], + "lines": [ + { + "bbox": [ + 105, + 558, + 197, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 197, + 573 + ], + "score": 1.0, + "content": "4.5 OPTIMIZATION", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 582, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "Different to Vaswani et al. (2017) we use Nesterov’s accelerated gradient method (Sutskever et al.,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 593, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 606 + ], + "score": 1.0, + "content": "2013) with a momentum of 0.99 and we renormalize gradients if their norm exceeds 0.1 (Pascanu", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 336, + 618 + ], + "score": 1.0, + "content": "et al., 2013). The learning rate is linearly warmed up from", + "type": "text" + }, + { + "bbox": [ + 337, + 604, + 359, + 615 + ], + "score": 0.9, + "content": "1 0 ^ { - 7 }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "to 1 for 16K steps and then annealed", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 278, + 628 + ], + "score": 1.0, + "content": "using a cosine learning rate schedule with", + "type": "text" + }, + { + "bbox": [ + 278, + 616, + 288, + 626 + ], + "score": 0.78, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "cycles (Loshchilov & Hutter, 2016). Each cycle runs", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "for twice the number of updates than the previous cycle and we lower the maximum and minimum", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 204, + 651 + ], + "score": 1.0, + "content": "learning rates by a rate", + "type": "text" + }, + { + "bbox": [ + 204, + 639, + 216, + 648 + ], + "score": 0.76, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "compared to the previous cycle. The initial minimum learning rate is", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 648, + 223, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 128, + 659 + ], + "score": 0.88, + "content": "1 0 ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 648, + 223, + 660 + ], + "score": 1.0, + "content": "and the maximum is 1.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 362, + 678 + ], + "score": 1.0, + "content": "BILLION WORD models train for a total of 975K updates over", + "type": "text" + }, + { + "bbox": [ + 362, + 666, + 392, + 676 + ], + "score": 0.89, + "content": "C = 3", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "cycles, the first cycle takes", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 676, + 504, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 204, + 689 + ], + "score": 1.0, + "content": "137K steps, and we set", + "type": "text" + }, + { + "bbox": [ + 205, + 677, + 245, + 687 + ], + "score": 0.89, + "content": "M = 0 . 6", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 676, + 473, + 689 + ], + "score": 1.0, + "content": ". The WIKITEXT-103 models train for 286K steps over", + "type": "text" + }, + { + "bbox": [ + 474, + 677, + 504, + 687 + ], + "score": 0.88, + "content": "C = 4", + "type": "inline_equation" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 316, + 700 + ], + "score": 1.0, + "content": "cycles, the first cycle takes 18K setps and we set", + "type": "text" + }, + { + "bbox": [ + 316, + 688, + 363, + 698 + ], + "score": 0.89, + "content": "M \\ : = \\ : 0 . 7 5", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 687, + 505, + 700 + ], + "score": 1.0, + "content": ". We run experiments on DGX-1", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "machines with 8 NVIDIA V100 GPUs and machines are interconnected by Infiniband. We also", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "use the NCCL2 library and the torch.distributed package for inter-GPU communication. We train", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 721, + 396, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 396, + 732 + ], + "score": 1.0, + "content": "models with 16-bit floating point precision, following Ott et al. (2018b).", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 181 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 383, + 95 + ], + "score": 1.0, + "content": "We limit the number of tokens per GPU to a maximum threshold", + "type": "text" + }, + { + "bbox": [ + 383, + 83, + 393, + 93 + ], + "score": 0.74, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "per GPU. That is, we add", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "examples of similar length until we reach this threshold. When we train on multiple GPUs, each", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 170, + 117 + ], + "score": 1.0, + "content": "GPU processes", + "type": "text" + }, + { + "bbox": [ + 171, + 105, + 180, + 114 + ], + "score": 0.77, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "tokens using the same model parameters. This increases the effective batch size", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 280, + 127 + ], + "score": 1.0, + "content": "to the product of the number of GPUs and", + "type": "text" + }, + { + "bbox": [ + 281, + 116, + 290, + 125 + ], + "score": 0.7, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 115, + 442, + 127 + ], + "score": 1.0, + "content": ". For BILLION WORD models we use", + "type": "text" + }, + { + "bbox": [ + 442, + 115, + 487, + 126 + ], + "score": 0.89, + "content": "B = 2 0 4 8", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 115, + 506, + 127 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 506, + 138 + ], + "score": 1.0, + "content": "typically train on 32 GPUs, giving an effective batch size of 65K tokens. The smaller vocabulary of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 248, + 149 + ], + "score": 1.0, + "content": "WIKITEXT-103 enables increasing", + "type": "text" + }, + { + "bbox": [ + 249, + 138, + 258, + 147 + ], + "score": 0.78, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 137, + 505, + 149 + ], + "score": 1.0, + "content": "to 4096 and we train on 8 GPUs. We found that large batch", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 149, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 160 + ], + "score": 1.0, + "content": "training is beneficial for this dataset and we therefore accumulate gradient updates over two batches", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 506, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 172 + ], + "score": 1.0, + "content": "before committing a parameter update (Ott et al., 2018a). This gives an effective batch size of 65K", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 170, + 216, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 216, + 182 + ], + "score": 1.0, + "content": "tokens for WIKITEXT-103.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 82, + 506, + 182 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 203, + 336, + 213 + ], + "lines": [ + { + "bbox": [ + 106, + 202, + 338, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 202, + 338, + 214 + ], + "score": 1.0, + "content": "4.4 INPUT AND OUTPUT LAYER HYPERPARAMETERS", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "Embedding sizes. For fixed size word input layers and softmax output layers we generally use", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "score": 1.0, + "content": "embeddings of size 512 for WIKITEXT-103. When we use an adaptive softmax in the output and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "fixed size word embeddings for the input, then we use dimension 256 for the input embeddings for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "BILLION WORD and 64 for WIKITEXT-103. We tuned this choice on the validation set (Appendix", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 269, + 344, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 344, + 282 + ], + "score": 1.0, + "content": "A). BPE inputs and outputs have embeddings of size 1024.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 225, + 505, + 282 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 300, + 505, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 300, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 506, + 314 + ], + "score": 1.0, + "content": "Character CNN. We model character inputs by convolving the representations of all characters", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 311, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 506, + 325 + ], + "score": 1.0, + "content": "in a word following Kim et al. (2015) which applies several filters, then max pooling, a number of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 323, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 506, + 335 + ], + "score": 1.0, + "content": "highway layers and a projection. Character embeddings have size 128 and we apply seven filters of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 332, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 104, + 332, + 154, + 347 + ], + "score": 1.0, + "content": "size 1x128,", + "type": "text" + }, + { + "bbox": [ + 155, + 334, + 181, + 344 + ], + "score": 0.39, + "content": "2 \\mathrm { x } 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 332, + 364, + 347 + ], + "score": 1.0, + "content": ", 3x384, 4x512, 5x512, 6x512, 7x512, where", + "type": "text" + }, + { + "bbox": [ + 365, + 334, + 391, + 344 + ], + "score": 0.26, + "content": "3 \\mathrm { x } 1 2 8", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 332, + 506, + 347 + ], + "score": 1.0, + "content": "indicates a filter processing", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 345, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 356 + ], + "score": 1.0, + "content": "three characters that outputs 128 features. We use a single highway layer for WIKITEXT-103, and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 356, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 367 + ], + "score": 1.0, + "content": "two for BILLION WORD. We do not add start of word and end of word markers as they did not", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 366, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 379 + ], + "score": 1.0, + "content": "improve validation accuracy. We train on the same pre-processed data as the other models, with", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 377, + 297, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 297, + 390 + ], + "score": 1.0, + "content": "unknown tokens in both the inputs and outputs.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18.5, + "bbox_fs": [ + 104, + 300, + 506, + 390 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 408, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 407, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 422 + ], + "score": 1.0, + "content": "Adaptive input representations and adaptive softmax. We use an adaptive softmax output layer", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 419, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 506, + 433 + ], + "score": 1.0, + "content": "to train models with large word-based vocabularies. For adaptive word inputs and adaptive softmax,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 429, + 507, + 444 + ], + "spans": [ + { + "bbox": [ + 104, + 429, + 218, + 444 + ], + "score": 1.0, + "content": "we use embeddings of size", + "type": "text" + }, + { + "bbox": [ + 218, + 431, + 260, + 441 + ], + "score": 0.89, + "content": "d = 1 0 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 429, + 507, + 444 + ], + "score": 1.0, + "content": "for the head and reduce the size of subsequent clusters by a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 442, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 143, + 453 + ], + "score": 1.0, + "content": "factor of", + "type": "text" + }, + { + "bbox": [ + 144, + 442, + 169, + 452 + ], + "score": 0.89, + "content": "k = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 442, + 393, + 453 + ], + "score": 1.0, + "content": ". For WIKITEXT-103, we have three bands of size 20K", + "type": "text" + }, + { + "bbox": [ + 393, + 442, + 428, + 452 + ], + "score": 0.57, + "content": "\\scriptstyle ( \\mathrm { d } = 1 0 2 4 )", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 442, + 456, + 453 + ], + "score": 1.0, + "content": "), 40K", + "type": "text" + }, + { + "bbox": [ + 456, + 442, + 486, + 452 + ], + "score": 0.73, + "content": "\\mathrm { \\Delta } \\mathrm { Q } \\mathrm { = } 2 5 6 )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 442, + 505, + 453 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 453, + 499, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 134, + 465 + ], + "score": 1.0, + "content": "200K", + "type": "text" + }, + { + "bbox": [ + 134, + 453, + 157, + 464 + ], + "score": 0.71, + "content": "\\scriptstyle \\mathrm { { ( d = 6 4 ) } }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 453, + 324, + 465 + ], + "score": 1.0, + "content": "). For BILLION WORD the bands are 60K", + "type": "text" + }, + { + "bbox": [ + 324, + 453, + 357, + 463 + ], + "score": 0.78, + "content": "\\scriptstyle \\mathrm { d } = 1 0 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 453, + 390, + 465 + ], + "score": 1.0, + "content": "), 100K", + "type": "text" + }, + { + "bbox": [ + 391, + 453, + 420, + 464 + ], + "score": 0.7, + "content": "\\mathrm { \\Delta } \\mathrm { Q } { = } 2 5 6 \\mathrm { \\Delta }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 453, + 469, + 465 + ], + "score": 1.0, + "content": "), and 640K", + "type": "text" + }, + { + "bbox": [ + 470, + 453, + 493, + 464 + ], + "score": 0.72, + "content": "\\scriptstyle \\mathrm { { ( d = 6 4 ) } }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 453, + 499, + 465 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25, + "bbox_fs": [ + 104, + 407, + 507, + 465 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 483, + 505, + 538 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "score": 1.0, + "content": "Sub-word models. We learn a byte-pair encoding (BPE) of 32K codes on the training data of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "score": 1.0, + "content": "each benchmark (Sennrich et al., 2016). After applying the code to the training data we obtain", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 505, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 104, + 505, + 506, + 518 + ], + "score": 1.0, + "content": "a vocabulary of 33,337 tokens for WIKITEXT-103 and 32,347 tokens for BILLION WORD. BPE", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "score": 1.0, + "content": "input/output embeddings have size 1024. The final evaluation is in terms word-level perplexity to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 527, + 486, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 486, + 540 + ], + "score": 1.0, + "content": "be comparable to other models. The probability of a word is the product of the sub-word units.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30, + "bbox_fs": [ + 104, + 483, + 506, + 540 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 559, + 195, + 570 + ], + "lines": [ + { + "bbox": [ + 105, + 558, + 197, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 197, + 573 + ], + "score": 1.0, + "content": "4.5 OPTIMIZATION", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 582, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "Different to Vaswani et al. (2017) we use Nesterov’s accelerated gradient method (Sutskever et al.,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 593, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 606 + ], + "score": 1.0, + "content": "2013) with a momentum of 0.99 and we renormalize gradients if their norm exceeds 0.1 (Pascanu", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 336, + 618 + ], + "score": 1.0, + "content": "et al., 2013). The learning rate is linearly warmed up from", + "type": "text" + }, + { + "bbox": [ + 337, + 604, + 359, + 615 + ], + "score": 0.9, + "content": "1 0 ^ { - 7 }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "to 1 for 16K steps and then annealed", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 278, + 628 + ], + "score": 1.0, + "content": "using a cosine learning rate schedule with", + "type": "text" + }, + { + "bbox": [ + 278, + 616, + 288, + 626 + ], + "score": 0.78, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "cycles (Loshchilov & Hutter, 2016). Each cycle runs", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "for twice the number of updates than the previous cycle and we lower the maximum and minimum", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 204, + 651 + ], + "score": 1.0, + "content": "learning rates by a rate", + "type": "text" + }, + { + "bbox": [ + 204, + 639, + 216, + 648 + ], + "score": 0.76, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "compared to the previous cycle. The initial minimum learning rate is", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 648, + 223, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 128, + 659 + ], + "score": 0.88, + "content": "1 0 ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 648, + 223, + 660 + ], + "score": 1.0, + "content": "and the maximum is 1.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 582, + 505, + 660 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 362, + 678 + ], + "score": 1.0, + "content": "BILLION WORD models train for a total of 975K updates over", + "type": "text" + }, + { + "bbox": [ + 362, + 666, + 392, + 676 + ], + "score": 0.89, + "content": "C = 3", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "cycles, the first cycle takes", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 676, + 504, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 204, + 689 + ], + "score": 1.0, + "content": "137K steps, and we set", + "type": "text" + }, + { + "bbox": [ + 205, + 677, + 245, + 687 + ], + "score": 0.89, + "content": "M = 0 . 6", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 676, + 473, + 689 + ], + "score": 1.0, + "content": ". The WIKITEXT-103 models train for 286K steps over", + "type": "text" + }, + { + "bbox": [ + 474, + 677, + 504, + 687 + ], + "score": 0.88, + "content": "C = 4", + "type": "inline_equation" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 316, + 700 + ], + "score": 1.0, + "content": "cycles, the first cycle takes 18K setps and we set", + "type": "text" + }, + { + "bbox": [ + 316, + 688, + 363, + 698 + ], + "score": 0.89, + "content": "M \\ : = \\ : 0 . 7 5", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 687, + 505, + 700 + ], + "score": 1.0, + "content": ". We run experiments on DGX-1", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "machines with 8 NVIDIA V100 GPUs and machines are interconnected by Infiniband. We also", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "use the NCCL2 library and the torch.distributed package for inter-GPU communication. We train", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 721, + 396, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 396, + 732 + ], + "score": 1.0, + "content": "models with 16-bit floating point precision, following Ott et al. (2018b).", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 665, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 123, + 80, + 487, + 218 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 123, + 80, + 487, + 218 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 123, + 80, + 487, + 218 + ], + "spans": [ + { + "bbox": [ + 123, + 80, + 487, + 218 + ], + "score": 0.985, + "html": "
TestTrain Time (hours)Parameters
Dauphin et al. (2017)31.9428M
J6zefowicz et al. (2016)30.01,040M
Shazeer et al. (2017)28.0-4,371Mt
Char-CNN25.8879366M
Adaptive inputs25.2255331M
Adaptive inputs (large)23.9172465M
Adaptive inputs (very large)23.021451026M
10 LSTMs + SNM10-SKIP (Shazeer et al.,2016)23.71=
", + "type": "table", + "image_path": "0c7c963dd8b2903cef8d6d8ec3f02912b76db5a52245768b59e1643f85ce4e3b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 123, + 80, + 487, + 126.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 123, + 126.0, + 487, + 172.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 123, + 172.0, + 487, + 218.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 226, + 505, + 267 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 224, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 506, + 240 + ], + "score": 1.0, + "content": "Table 1: Test perplexity on BILLION WORD. Adaptive inputs share parameters with an adaptive", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 234, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 506, + 249 + ], + "score": 1.0, + "content": "softmax. Training times of Char-CNN and Adaptive input models are measured when training with", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 244, + 147, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 147, + 259 + ], + "score": 1.0, + "content": "64 GPUs.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 255, + 284, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 284, + 269 + ], + "score": 1.0, + "content": "†does not include embedding and softmax layers", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "table", + "bbox": [ + 182, + 279, + 428, + 378 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 182, + 279, + 428, + 378 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 182, + 279, + 428, + 378 + ], + "spans": [ + { + "bbox": [ + 182, + 279, + 428, + 378 + ], + "score": 0.98, + "html": "
TestTrain Time (hours)Parameters
Grave et al. (2016)40.8
Dauphin et al. (2017)37.2229M
Merity et al. (2018)33.0=151M
Rae et al. (2018)29.2-
Adaptive inputs18.767247M
", + "type": "table", + "image_path": "0d1b611f90a7f9ae450f1e0728a887cd39a36f52a2ceafdc4ab9a882f69e86cc.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 182, + 279, + 428, + 312.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 182, + 312.0, + 428, + 345.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 182, + 345.0, + 428, + 378.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 119, + 385, + 489, + 397 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 120, + 385, + 490, + 398 + ], + "spans": [ + { + "bbox": [ + 120, + 385, + 490, + 398 + ], + "score": 1.0, + "content": "Table 2: Test perplexity on WIKITEXT-103 (cf. Table 1). Training time is based on 8 GPUs.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + } + ], + "index": 9.0 + }, + { + "type": "title", + "bbox": [ + 107, + 418, + 275, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 276, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 276, + 433 + ], + "score": 1.0, + "content": "5 EXPERIMENTS AND RESULTS", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 443, + 196, + 455 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 198, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 198, + 456 + ], + "score": 1.0, + "content": "5.1 MAIN RESULTS", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 464, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "score": 1.0, + "content": "For the main results on BILLION WORD, we doubled the batch size by training on 64 GPUs instead", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 465, + 488 + ], + "score": 1.0, + "content": "of 32 GPUs. We also consider two larger setups, one where we added four more blocks", + "type": "text" + }, + { + "bbox": [ + 465, + 476, + 501, + 486 + ], + "score": 0.85, + "content": "N = 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 475, + 505, + 488 + ], + "score": 1.0, + "content": ")", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 486, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 257, + 500 + ], + "score": 1.0, + "content": "and increased the FFN dimension to", + "type": "text" + }, + { + "bbox": [ + 257, + 487, + 309, + 498 + ], + "score": 0.87, + "content": "e _ { f f } = 6 1 4 4", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 486, + 506, + 500 + ], + "score": 1.0, + "content": "(large), and another where we add another four", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 138, + 510 + ], + "score": 1.0, + "content": "blocks", + "type": "text" + }, + { + "bbox": [ + 138, + 498, + 173, + 508 + ], + "score": 0.84, + "content": "N = 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 497, + 196, + 510 + ], + "score": 1.0, + "content": ") with", + "type": "text" + }, + { + "bbox": [ + 197, + 498, + 246, + 510 + ], + "score": 0.9, + "content": "e _ { f f } = 8 1 9 2", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 497, + 264, + 510 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 264, + 498, + 304, + 508 + ], + "score": 0.85, + "content": "e = 1 5 3 6", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 497, + 456, + 510 + ], + "score": 1.0, + "content": "(very large). All other settings follow", + "type": "text" + }, + { + "bbox": [ + 456, + 498, + 475, + 509 + ], + "score": 0.82, + "content": "\\ S 4 . 4", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "and all", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 507, + 310, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 310, + 522 + ], + "score": 1.0, + "content": "models were trained for the same number of steps.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "Table 1 compares our models to previous work on BILLION WORD. The adaptive input model", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "outperforms the best previously reported result at an order of magnitude fewer parameters. Our", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 548, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 505, + 559 + ], + "score": 1.0, + "content": "large model performs nearly as well as an ensemble of over ten models and achieves a new state", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 558, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 571 + ], + "score": 1.0, + "content": "of the art of 24.14 perplexity. Our very large model performs as well as an ensemble of over ten", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 404, + 582 + ], + "score": 1.0, + "content": "models and achieves 23.02 perplexity. The Char-CNN model performs", + "type": "text" + }, + { + "bbox": [ + 405, + 569, + 440, + 580 + ], + "score": 0.38, + "content": "0 . 6 ~ \\mathrm { P P L }", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "worse than the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 580, + 402, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 352, + 593 + ], + "score": 1.0, + "content": "standard adaptive input model even though it trained for over", + "type": "text" + }, + { + "bbox": [ + 352, + 581, + 372, + 591 + ], + "score": 0.86, + "content": "40 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 580, + 402, + 593 + ], + "score": 1.0, + "content": "longer.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 597, + 505, + 652 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "Table 2 shows our result on WIKITEXT-103 where adaptive inputs achieve 18.7 perplexity. For", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "this result only, we partition the training data into blocks of 3072 contiguous tokens instead of 512", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 619, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 505, + 631 + ], + "score": 1.0, + "content": "tokens as for other experiments. During evaluation we require blocks to contain complete sentences", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 628, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 644 + ], + "score": 1.0, + "content": "totaling up to 3072 tokens of which the first 2560 tokens serve as context to score the last 512 tokens;", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 641, + 483, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 462, + 653 + ], + "score": 1.0, + "content": "we take care to score all tokens in the test and validation sets. We motivate this choice in", + "type": "text" + }, + { + "bbox": [ + 462, + 641, + 480, + 653 + ], + "score": 0.81, + "content": "\\ S 5 . 3", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 641, + 483, + 653 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26 + }, + { + "type": "title", + "bbox": [ + 114, + 667, + 394, + 678 + ], + "lines": [ + { + "bbox": [ + 112, + 666, + 396, + 680 + ], + "spans": [ + { + "bbox": [ + 112, + 666, + 396, + 680 + ], + "score": 1.0, + "content": ".2 COMPARISON OF INPUT AND OUTPUT LAYER FACTORIZATIONS", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "Next, we perform a systematic comparison of different input and output layer factorizations. We", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "consider a word-based setup with fixed size word input embeddings and a standard word softmax", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "(SM) where embeddings have either dimension 512 (WIKITEXT-103) or 64 (BILLION WORD). We", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "consider tying the input and output embeddings (SM-T). Instead of words, we try less sparse sub-", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 123, + 80, + 487, + 218 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 123, + 80, + 487, + 218 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 123, + 80, + 487, + 218 + ], + "spans": [ + { + "bbox": [ + 123, + 80, + 487, + 218 + ], + "score": 0.985, + "html": "
TestTrain Time (hours)Parameters
Dauphin et al. (2017)31.9428M
J6zefowicz et al. (2016)30.01,040M
Shazeer et al. (2017)28.0-4,371Mt
Char-CNN25.8879366M
Adaptive inputs25.2255331M
Adaptive inputs (large)23.9172465M
Adaptive inputs (very large)23.021451026M
10 LSTMs + SNM10-SKIP (Shazeer et al.,2016)23.71=
", + "type": "table", + "image_path": "0c7c963dd8b2903cef8d6d8ec3f02912b76db5a52245768b59e1643f85ce4e3b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 123, + 80, + 487, + 126.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 123, + 126.0, + 487, + 172.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 123, + 172.0, + 487, + 218.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 226, + 505, + 267 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 224, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 506, + 240 + ], + "score": 1.0, + "content": "Table 1: Test perplexity on BILLION WORD. Adaptive inputs share parameters with an adaptive", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 234, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 506, + 249 + ], + "score": 1.0, + "content": "softmax. Training times of Char-CNN and Adaptive input models are measured when training with", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 244, + 147, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 147, + 259 + ], + "score": 1.0, + "content": "64 GPUs.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 255, + 284, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 284, + 269 + ], + "score": 1.0, + "content": "†does not include embedding and softmax layers", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "table", + "bbox": [ + 182, + 279, + 428, + 378 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 182, + 279, + 428, + 378 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 182, + 279, + 428, + 378 + ], + "spans": [ + { + "bbox": [ + 182, + 279, + 428, + 378 + ], + "score": 0.98, + "html": "
TestTrain Time (hours)Parameters
Grave et al. (2016)40.8
Dauphin et al. (2017)37.2229M
Merity et al. (2018)33.0=151M
Rae et al. (2018)29.2-
Adaptive inputs18.767247M
", + "type": "table", + "image_path": "0d1b611f90a7f9ae450f1e0728a887cd39a36f52a2ceafdc4ab9a882f69e86cc.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 182, + 279, + 428, + 312.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 182, + 312.0, + 428, + 345.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 182, + 345.0, + 428, + 378.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 119, + 385, + 489, + 397 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 120, + 385, + 490, + 398 + ], + "spans": [ + { + "bbox": [ + 120, + 385, + 490, + 398 + ], + "score": 1.0, + "content": "Table 2: Test perplexity on WIKITEXT-103 (cf. Table 1). Training time is based on 8 GPUs.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + } + ], + "index": 9.0 + }, + { + "type": "title", + "bbox": [ + 107, + 418, + 275, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 276, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 276, + 433 + ], + "score": 1.0, + "content": "5 EXPERIMENTS AND RESULTS", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 443, + 196, + 455 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 198, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 198, + 456 + ], + "score": 1.0, + "content": "5.1 MAIN RESULTS", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 464, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "score": 1.0, + "content": "For the main results on BILLION WORD, we doubled the batch size by training on 64 GPUs instead", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 465, + 488 + ], + "score": 1.0, + "content": "of 32 GPUs. We also consider two larger setups, one where we added four more blocks", + "type": "text" + }, + { + "bbox": [ + 465, + 476, + 501, + 486 + ], + "score": 0.85, + "content": "N = 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 475, + 505, + 488 + ], + "score": 1.0, + "content": ")", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 486, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 257, + 500 + ], + "score": 1.0, + "content": "and increased the FFN dimension to", + "type": "text" + }, + { + "bbox": [ + 257, + 487, + 309, + 498 + ], + "score": 0.87, + "content": "e _ { f f } = 6 1 4 4", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 486, + 506, + 500 + ], + "score": 1.0, + "content": "(large), and another where we add another four", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 138, + 510 + ], + "score": 1.0, + "content": "blocks", + "type": "text" + }, + { + "bbox": [ + 138, + 498, + 173, + 508 + ], + "score": 0.84, + "content": "N = 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 497, + 196, + 510 + ], + "score": 1.0, + "content": ") with", + "type": "text" + }, + { + "bbox": [ + 197, + 498, + 246, + 510 + ], + "score": 0.9, + "content": "e _ { f f } = 8 1 9 2", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 497, + 264, + 510 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 264, + 498, + 304, + 508 + ], + "score": 0.85, + "content": "e = 1 5 3 6", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 497, + 456, + 510 + ], + "score": 1.0, + "content": "(very large). All other settings follow", + "type": "text" + }, + { + "bbox": [ + 456, + 498, + 475, + 509 + ], + "score": 0.82, + "content": "\\ S 4 . 4", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "and all", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 507, + 310, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 310, + 522 + ], + "score": 1.0, + "content": "models were trained for the same number of steps.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 464, + 506, + 522 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "Table 1 compares our models to previous work on BILLION WORD. The adaptive input model", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "outperforms the best previously reported result at an order of magnitude fewer parameters. Our", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 548, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 505, + 559 + ], + "score": 1.0, + "content": "large model performs nearly as well as an ensemble of over ten models and achieves a new state", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 558, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 571 + ], + "score": 1.0, + "content": "of the art of 24.14 perplexity. Our very large model performs as well as an ensemble of over ten", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 404, + 582 + ], + "score": 1.0, + "content": "models and achieves 23.02 perplexity. The Char-CNN model performs", + "type": "text" + }, + { + "bbox": [ + 405, + 569, + 440, + 580 + ], + "score": 0.38, + "content": "0 . 6 ~ \\mathrm { P P L }", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "worse than the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 580, + 402, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 352, + 593 + ], + "score": 1.0, + "content": "standard adaptive input model even though it trained for over", + "type": "text" + }, + { + "bbox": [ + 352, + 581, + 372, + 591 + ], + "score": 0.86, + "content": "40 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 580, + 402, + 593 + ], + "score": 1.0, + "content": "longer.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 525, + 506, + 593 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 597, + 505, + 652 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "Table 2 shows our result on WIKITEXT-103 where adaptive inputs achieve 18.7 perplexity. For", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "this result only, we partition the training data into blocks of 3072 contiguous tokens instead of 512", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 619, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 505, + 631 + ], + "score": 1.0, + "content": "tokens as for other experiments. During evaluation we require blocks to contain complete sentences", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 628, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 644 + ], + "score": 1.0, + "content": "totaling up to 3072 tokens of which the first 2560 tokens serve as context to score the last 512 tokens;", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 641, + 483, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 462, + 653 + ], + "score": 1.0, + "content": "we take care to score all tokens in the test and validation sets. We motivate this choice in", + "type": "text" + }, + { + "bbox": [ + 462, + 641, + 480, + 653 + ], + "score": 0.81, + "content": "\\ S 5 . 3", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 641, + 483, + 653 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 597, + 506, + 653 + ] + }, + { + "type": "title", + "bbox": [ + 114, + 667, + 394, + 678 + ], + "lines": [ + { + "bbox": [ + 112, + 666, + 396, + 680 + ], + "spans": [ + { + "bbox": [ + 112, + 666, + 396, + 680 + ], + "score": 1.0, + "content": ".2 COMPARISON OF INPUT AND OUTPUT LAYER FACTORIZATIONS", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "Next, we perform a systematic comparison of different input and output layer factorizations. We", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "consider a word-based setup with fixed size word input embeddings and a standard word softmax", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "(SM) where embeddings have either dimension 512 (WIKITEXT-103) or 64 (BILLION WORD). We", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "consider tying the input and output embeddings (SM-T). Instead of words, we try less sparse sub-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 505, + 484 + ], + "score": 1.0, + "content": "word units, both in the input and output, with embeddings of size 1024 (BPE) and shared weights", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "score": 1.0, + "content": "(BPE-T). Next, we consider replacing the fixed size output representations by an adaptive softmax", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "(ASM) and characters as input (CNN). Finally, we use both adaptive input word representations", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "as well as an adaptive softmax (ADP) and a tied version (ADP-T). All models use the same self-", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 516, + 266, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 244, + 528 + ], + "score": 1.0, + "content": "attention architecture described in", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 244, + 516, + 262, + 528 + ], + "score": 0.74, + "content": "\\ S 4 . 1", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 263, + 516, + 266, + 528 + ], + "score": 1.0, + "content": ".", + "type": "text", + "cross_page": true + } + ], + "index": 16 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 687, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 114, + 80, + 497, + 207 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 114, + 80, + 497, + 207 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 80, + 497, + 207 + ], + "spans": [ + { + "bbox": [ + 114, + 80, + 497, + 207 + ], + "score": 0.982, + "html": "
InputOutputValidTestTrain Time (hours)Params
SMEmbeddingSoftmax23.8724.9257*476.8M
BPEBPE EmbeddingBPE Softmax23.1324.2530270M
BPE-TBPE EmbeddingBPE Softmax (tied)22.4623.4530235.7M
SM-TEmbeddingSoftmax (tied)22.6323.3856*339.7M
ASMEmbeddingAdaptive21.2322.1835263.1M
CNNChar-CNNAdaptive20.8621.7970266.3M
ADPAdaptiveAdaptive20.9521.7434291.3M
ADP-TAdaptiveAdaptive (tied)19.7920.5130246.9M
", + "type": "table", + "image_path": "e3e37fff1071f7ce2f4fbe3586ddfabf4960159d940eb2fd51570e533afed4df.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 114, + 80, + 497, + 122.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 114, + 122.33333333333334, + 497, + 164.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 114, + 164.66666666666669, + 497, + 207.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 214, + 505, + 270 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 215, + 504, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 504, + 227 + ], + "score": 1.0, + "content": "Table 3: Test perplexity on WIKITEXT-103 for various input and output layer factorizations. Train-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 337, + 239 + ], + "score": 1.0, + "content": "ing speed was measured on a single 8-GPU machine.", + "type": "text" + }, + { + "bbox": [ + 338, + 226, + 350, + 237 + ], + "score": 0.62, + "content": "( ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "indicates a modified training regime", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "because of large memory requirements: the maximum number of tokens per GPU was lowered to", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 246, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 505, + 261 + ], + "score": 1.0, + "content": "1024 from 4096 but the same number of updates were performed by processing four batches before", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 259, + 222, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 222, + 271 + ], + "score": 1.0, + "content": "committing a weight update.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "table", + "bbox": [ + 110, + 304, + 501, + 409 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 110, + 304, + 501, + 409 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 304, + 501, + 409 + ], + "spans": [ + { + "bbox": [ + 110, + 304, + 501, + 409 + ], + "score": 0.981, + "html": "
InputOutputValidTestTrain time (hours)Params
BPE-TBPE EmbeddingBPE Softmax (shared)27.4427.5134234.7M
BPEBPE EmbeddingBPE Softmax27.0227.1335267.8M
ASMEmbeddingAdaptive26.9727.0662532.8M
CNNChar-CNNAdaptive26.1326.2592365.8M
ADPAdaptiveAdaptive26.3826.4965458.4M
ADP-TAdaptiveAdaptive (shared)25.5125.5843330.8M
", + "type": "table", + "image_path": "aa93371ce50bf62d0f1b00bb824bec36f123c85c9463cdc667ce6a25a44cfb43.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 110, + 304, + 501, + 339.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 110, + 339.0, + 501, + 374.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 110, + 374.0, + 501, + 409.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 112, + 417, + 497, + 429 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 112, + 416, + 499, + 430 + ], + "spans": [ + { + "bbox": [ + 112, + 416, + 499, + 430 + ], + "score": 1.0, + "content": "Table 4: Test perplexity on BILLION WORD. Training speed measured on four 8-GPU machines.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + } + ], + "index": 10.0 + }, + { + "type": "text", + "bbox": [ + 107, + 472, + 505, + 528 + ], + "lines": [ + { + "bbox": [ + 106, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 505, + 484 + ], + "score": 1.0, + "content": "word units, both in the input and output, with embeddings of size 1024 (BPE) and shared weights", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "score": 1.0, + "content": "(BPE-T). Next, we consider replacing the fixed size output representations by an adaptive softmax", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "(ASM) and characters as input (CNN). Finally, we use both adaptive input word representations", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "as well as an adaptive softmax (ADP) and a tied version (ADP-T). All models use the same self-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 516, + 266, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 244, + 528 + ], + "score": 1.0, + "content": "attention architecture described in", + "type": "text" + }, + { + "bbox": [ + 244, + 516, + 262, + 528 + ], + "score": 0.74, + "content": "\\ S 4 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 516, + 266, + 528 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 532, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 546 + ], + "score": 1.0, + "content": "Table 3 shows results when training all configurations for the same number of updates. Adaptive", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 545, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 557 + ], + "score": 1.0, + "content": "input representations with tied input and output layers (ADP-T) achieve the highest accuracy at the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "score": 1.0, + "content": "same speed as the BPE models which have a very small vocabulary (33K versus 260K). CNN is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 567, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 578 + ], + "score": 1.0, + "content": "1 perplexity worse than ADP-T and requires well over twice the training time. It is the slowest", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "approach, even though it has a fast adaptive softmax in the output. Fixed word embeddings perform", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "least well (SM). Sub-word units are fast to train and perform better than word models with fixed", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 600, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 611 + ], + "score": 1.0, + "content": "sized embeddings. ASM improves over SM and greatly speeds up training. For ASM, we found", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "score": 1.0, + "content": "that reducing the dimension of the input word embeddings to 64 on WIKITEXT-103 results in better", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 621, + 205, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 205, + 634 + ], + "score": 1.0, + "content": "accuracy (Appendix A).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 638, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "Table 4 shows that adaptive input representations perform equally well on BILLION WORD com-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 270, + 662 + ], + "score": 1.0, + "content": "pared to other factorizations. ADP-T is", + "type": "text" + }, + { + "bbox": [ + 271, + 649, + 290, + 659 + ], + "score": 0.87, + "content": "34 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "faster than ADP because there are fewer parameters", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 659, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 104, + 659, + 506, + 674 + ], + "score": 1.0, + "content": "to update. Similar to before, ADP-T trains more than twice as fast as CNN at higher accuracy,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "score": 1.0, + "content": "however, the accuracy gap is narrower than for WIKITEXT-103. Regularization is more important", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 681, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 694 + ], + "score": 1.0, + "content": "on WIKITEXT-103 while models for BILLION WORD benefit from additional capacity. Because of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 693, + 343, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 343, + 704 + ], + "score": 1.0, + "content": "this we used input word embeddings of size 256 for ASM.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "We also trained CNN without replacing input words outside the vocabulary by an unknown symbol,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 720, + 342, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 342, + 733 + ], + "score": 1.0, + "content": "however, this only improved validation perplexity by 0.16.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 114, + 80, + 497, + 207 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 114, + 80, + 497, + 207 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 80, + 497, + 207 + ], + "spans": [ + { + "bbox": [ + 114, + 80, + 497, + 207 + ], + "score": 0.982, + "html": "
InputOutputValidTestTrain Time (hours)Params
SMEmbeddingSoftmax23.8724.9257*476.8M
BPEBPE EmbeddingBPE Softmax23.1324.2530270M
BPE-TBPE EmbeddingBPE Softmax (tied)22.4623.4530235.7M
SM-TEmbeddingSoftmax (tied)22.6323.3856*339.7M
ASMEmbeddingAdaptive21.2322.1835263.1M
CNNChar-CNNAdaptive20.8621.7970266.3M
ADPAdaptiveAdaptive20.9521.7434291.3M
ADP-TAdaptiveAdaptive (tied)19.7920.5130246.9M
", + "type": "table", + "image_path": "e3e37fff1071f7ce2f4fbe3586ddfabf4960159d940eb2fd51570e533afed4df.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 114, + 80, + 497, + 122.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 114, + 122.33333333333334, + 497, + 164.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 114, + 164.66666666666669, + 497, + 207.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 214, + 505, + 270 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 215, + 504, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 504, + 227 + ], + "score": 1.0, + "content": "Table 3: Test perplexity on WIKITEXT-103 for various input and output layer factorizations. Train-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 337, + 239 + ], + "score": 1.0, + "content": "ing speed was measured on a single 8-GPU machine.", + "type": "text" + }, + { + "bbox": [ + 338, + 226, + 350, + 237 + ], + "score": 0.62, + "content": "( ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "indicates a modified training regime", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "because of large memory requirements: the maximum number of tokens per GPU was lowered to", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 246, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 505, + 261 + ], + "score": 1.0, + "content": "1024 from 4096 but the same number of updates were performed by processing four batches before", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 259, + 222, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 222, + 271 + ], + "score": 1.0, + "content": "committing a weight update.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "table", + "bbox": [ + 110, + 304, + 501, + 409 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 110, + 304, + 501, + 409 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 304, + 501, + 409 + ], + "spans": [ + { + "bbox": [ + 110, + 304, + 501, + 409 + ], + "score": 0.981, + "html": "
InputOutputValidTestTrain time (hours)Params
BPE-TBPE EmbeddingBPE Softmax (shared)27.4427.5134234.7M
BPEBPE EmbeddingBPE Softmax27.0227.1335267.8M
ASMEmbeddingAdaptive26.9727.0662532.8M
CNNChar-CNNAdaptive26.1326.2592365.8M
ADPAdaptiveAdaptive26.3826.4965458.4M
ADP-TAdaptiveAdaptive (shared)25.5125.5843330.8M
", + "type": "table", + "image_path": "aa93371ce50bf62d0f1b00bb824bec36f123c85c9463cdc667ce6a25a44cfb43.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 110, + 304, + 501, + 339.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 110, + 339.0, + 501, + 374.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 110, + 374.0, + 501, + 409.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 112, + 417, + 497, + 429 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 112, + 416, + 499, + 430 + ], + "spans": [ + { + "bbox": [ + 112, + 416, + 499, + 430 + ], + "score": 1.0, + "content": "Table 4: Test perplexity on BILLION WORD. Training speed measured on four 8-GPU machines.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + } + ], + "index": 10.0 + }, + { + "type": "text", + "bbox": [ + 107, + 472, + 505, + 528 + ], + "lines": [], + "index": 14, + "bbox_fs": [ + 105, + 472, + 505, + 528 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 532, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 546 + ], + "score": 1.0, + "content": "Table 3 shows results when training all configurations for the same number of updates. Adaptive", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 545, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 557 + ], + "score": 1.0, + "content": "input representations with tied input and output layers (ADP-T) achieve the highest accuracy at the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "score": 1.0, + "content": "same speed as the BPE models which have a very small vocabulary (33K versus 260K). CNN is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 567, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 578 + ], + "score": 1.0, + "content": "1 perplexity worse than ADP-T and requires well over twice the training time. It is the slowest", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "approach, even though it has a fast adaptive softmax in the output. Fixed word embeddings perform", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "least well (SM). Sub-word units are fast to train and perform better than word models with fixed", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 600, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 611 + ], + "score": 1.0, + "content": "sized embeddings. ASM improves over SM and greatly speeds up training. For ASM, we found", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "score": 1.0, + "content": "that reducing the dimension of the input word embeddings to 64 on WIKITEXT-103 results in better", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 621, + 205, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 205, + 634 + ], + "score": 1.0, + "content": "accuracy (Appendix A).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 532, + 506, + 634 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 638, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "Table 4 shows that adaptive input representations perform equally well on BILLION WORD com-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 270, + 662 + ], + "score": 1.0, + "content": "pared to other factorizations. ADP-T is", + "type": "text" + }, + { + "bbox": [ + 271, + 649, + 290, + 659 + ], + "score": 0.87, + "content": "34 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "faster than ADP because there are fewer parameters", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 659, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 104, + 659, + 506, + 674 + ], + "score": 1.0, + "content": "to update. Similar to before, ADP-T trains more than twice as fast as CNN at higher accuracy,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "score": 1.0, + "content": "however, the accuracy gap is narrower than for WIKITEXT-103. Regularization is more important", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 681, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 694 + ], + "score": 1.0, + "content": "on WIKITEXT-103 while models for BILLION WORD benefit from additional capacity. Because of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 693, + 343, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 343, + 704 + ], + "score": 1.0, + "content": "this we used input word embeddings of size 256 for ASM.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28.5, + "bbox_fs": [ + 104, + 637, + 506, + 704 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "We also trained CNN without replacing input words outside the vocabulary by an unknown symbol,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 720, + 342, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 342, + 733 + ], + "score": 1.0, + "content": "however, this only improved validation perplexity by 0.16.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 78, + 504, + 267 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 78, + 504, + 267 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 78, + 504, + 267 + ], + "spans": [ + { + "bbox": [ + 108, + 78, + 504, + 267 + ], + "score": 0.972, + "type": "image", + "image_path": "4b396896fe5d317f22b0248271298f7ce8c2e521c3dc122395301b4c3ec88349.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 78, + 504, + 141.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 141.0, + 504, + 204.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 204.0, + 504, + 267.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 278, + 504, + 300 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "score": 1.0, + "content": "Figure 2: Loss of models binned by word frequency on the test set of WIKITEXT-103. Bins are not", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 290, + 155, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 155, + 302 + ], + "score": 1.0, + "content": "cumulative.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 108, + 310, + 504, + 500 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 310, + 504, + 500 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 310, + 504, + 500 + ], + "spans": [ + { + "bbox": [ + 108, + 310, + 504, + 500 + ], + "score": 0.971, + "type": "image", + "image_path": "373e2f18c0f9c2f71d23422298631e38730245add24d65b6f615d793b9273992.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 108, + 310, + 504, + 373.3333333333333 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 108, + 373.3333333333333, + 504, + 436.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 108, + 436.66666666666663, + 504, + 499.99999999999994 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 511, + 504, + 534 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 509, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 506, + 525 + ], + "score": 1.0, + "content": "Figure 3: Loss of models when binning by the frequency of the previous word measured on", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 522, + 230, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 230, + 535 + ], + "score": 1.0, + "content": "WIKITEXT-103 (cf. Figure 2).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "title", + "bbox": [ + 107, + 556, + 175, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 555, + 177, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 177, + 569 + ], + "score": 1.0, + "content": "5.3 ANALYSIS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 577, + 504, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "Next, we turn to the question of how well models perform on rare words compared to frequent", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 587, + 493, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 493, + 602 + ], + "score": 1.0, + "content": "words. We compute the average loss for each word in the test set and group words by frequency.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "Figure 2 shows results on WIKITEXT-103. Tying weights helps all models on rare words, likely", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 616, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 627 + ], + "score": 1.0, + "content": "because of regularization effects. Fixed size word embeddings with a word softmax (SM and SM-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "T) do not perform well on rare words. This is likely due to underfitting on common words and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "we use the largest possible embedding size we could fit on 16GB GPU cards given our batch size", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "(more experimentation in Appendix A). BPE and BPE-T perform poorly on rare words because", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 659, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 674 + ], + "score": 1.0, + "content": "probabilities are a product of several sub-word units. ADP-T performs best across all frequency", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "ranges. Figure 3 bins the loss by the frequency of the previous word and shows that CNN does well", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 681, + 432, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 432, + 694 + ], + "score": 1.0, + "content": "when it has rare words in the context, however, ADP-T does best across all bins.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Figure 4 shows an equivalent analysis for BILLION WORD. The largest differences between mod-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "els is on rare words. CNN performs best on very rare words but is outperformed by ADP in all", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "other settings. Similar to WIKITEXT-103, BPE and BPE-T perform poorly on rare words. Further", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 78, + 504, + 267 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 78, + 504, + 267 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 78, + 504, + 267 + ], + "spans": [ + { + "bbox": [ + 108, + 78, + 504, + 267 + ], + "score": 0.972, + "type": "image", + "image_path": "4b396896fe5d317f22b0248271298f7ce8c2e521c3dc122395301b4c3ec88349.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 78, + 504, + 141.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 141.0, + 504, + 204.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 204.0, + 504, + 267.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 278, + 504, + 300 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "score": 1.0, + "content": "Figure 2: Loss of models binned by word frequency on the test set of WIKITEXT-103. Bins are not", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 290, + 155, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 155, + 302 + ], + "score": 1.0, + "content": "cumulative.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 108, + 310, + 504, + 500 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 310, + 504, + 500 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 310, + 504, + 500 + ], + "spans": [ + { + "bbox": [ + 108, + 310, + 504, + 500 + ], + "score": 0.971, + "type": "image", + "image_path": "373e2f18c0f9c2f71d23422298631e38730245add24d65b6f615d793b9273992.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 108, + 310, + 504, + 373.3333333333333 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 108, + 373.3333333333333, + 504, + 436.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 108, + 436.66666666666663, + 504, + 499.99999999999994 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 511, + 504, + 534 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 509, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 506, + 525 + ], + "score": 1.0, + "content": "Figure 3: Loss of models when binning by the frequency of the previous word measured on", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 522, + 230, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 230, + 535 + ], + "score": 1.0, + "content": "WIKITEXT-103 (cf. Figure 2).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "title", + "bbox": [ + 107, + 556, + 175, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 555, + 177, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 177, + 569 + ], + "score": 1.0, + "content": "5.3 ANALYSIS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 577, + 504, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "Next, we turn to the question of how well models perform on rare words compared to frequent", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 587, + 493, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 493, + 602 + ], + "score": 1.0, + "content": "words. We compute the average loss for each word in the test set and group words by frequency.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 577, + 505, + 602 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "Figure 2 shows results on WIKITEXT-103. Tying weights helps all models on rare words, likely", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 616, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 627 + ], + "score": 1.0, + "content": "because of regularization effects. Fixed size word embeddings with a word softmax (SM and SM-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "T) do not perform well on rare words. This is likely due to underfitting on common words and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "we use the largest possible embedding size we could fit on 16GB GPU cards given our batch size", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "(more experimentation in Appendix A). BPE and BPE-T perform poorly on rare words because", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 659, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 674 + ], + "score": 1.0, + "content": "probabilities are a product of several sub-word units. ADP-T performs best across all frequency", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "ranges. Figure 3 bins the loss by the frequency of the previous word and shows that CNN does well", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 681, + 432, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 432, + 694 + ], + "score": 1.0, + "content": "when it has rare words in the context, however, ADP-T does best across all bins.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 605, + 505, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Figure 4 shows an equivalent analysis for BILLION WORD. The largest differences between mod-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "els is on rare words. CNN performs best on very rare words but is outperformed by ADP in all", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "other settings. Similar to WIKITEXT-103, BPE and BPE-T perform poorly on rare words. Further", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 698, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 78, + 503, + 267 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 78, + 503, + 267 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 78, + 503, + 267 + ], + "spans": [ + { + "bbox": [ + 108, + 78, + 503, + 267 + ], + "score": 0.972, + "type": "image", + "image_path": "d73dcb6e7e5c4af036340a7099e37fe1418673e98a1f8915cde2d051627402b8.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 78, + 503, + 141.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 141.0, + 503, + 204.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 204.0, + 503, + 267.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 278, + 505, + 300 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 277, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 506, + 291 + ], + "score": 1.0, + "content": "Figure 4: Loss of models when binning by word frequency on the test set of BILLION WORD. Bins", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 290, + 185, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 185, + 301 + ], + "score": 1.0, + "content": "are not cumulative.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 204, + 314, + 406, + 419 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 204, + 314, + 406, + 419 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 204, + 314, + 406, + 419 + ], + "spans": [ + { + "bbox": [ + 204, + 314, + 406, + 419 + ], + "score": 0.95, + "html": "
Train block sizeInference context sizeValidation perplexityTest perplexity
51219.7920.51
5120 48018.3519.03
2048018.9619.53
2048153618.2318.88
3072018.6319.34
3072256017.9718.70
", + "type": "table", + "image_path": "34ba39af36272851bc56541c57fa7eeaea8880ea2ca40ce648a9adb53f077da7.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 204, + 314, + 406, + 327.125 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 204, + 327.125, + 406, + 340.25 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 204, + 340.25, + 406, + 353.375 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 204, + 353.375, + 406, + 366.5 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 204, + 366.5, + 406, + 379.625 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 204, + 379.625, + 406, + 392.75 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 204, + 392.75, + 406, + 405.875 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 204, + 405.875, + 406, + 419.0 + ], + "spans": [], + "index": 12 + } + ] + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 427, + 504, + 460 + ], + "lines": [ + { + "bbox": [ + 105, + 425, + 504, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 504, + 441 + ], + "score": 1.0, + "content": "Table 5: Perplexity on WIKITEXT-103 with different context sizes during training and inference.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "score": 1.0, + "content": "Training block size is the number of consecutive tokens considered during training. Inference con-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 449, + 402, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 402, + 461 + ], + "score": 1.0, + "content": "text is the number of tokens provided at evaluation before scoring tokens.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 484, + 505, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "score": 1.0, + "content": "analysis (Appendix 5.3) binning the loss by the frequency of the previous word shows that weight", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 496, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 507 + ], + "score": 1.0, + "content": "sharing also helps for BILLION WORD and that CNN does very well on rare words for BILLION", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 507, + 244, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 244, + 518 + ], + "score": 1.0, + "content": "WORD compared to other models.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 505, + 589 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "Table 5 shows the importance of context size for WIKITEXT-103. Training block size is the number", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 535, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 505, + 546 + ], + "score": 1.0, + "content": "of consecutive tokens that are considered at once during training. Inference context is the number of", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 543, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 559 + ], + "score": 1.0, + "content": "tokens that are provided at evaluation before any tokens are scored. Simply increasing the training", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "block size from 512 to 3072 results in an improvement of nearly 1.2 perplexity with no inference", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 567, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 579 + ], + "score": 1.0, + "content": "context window. Increasing the context size at inference time results in an improvement of 0.6", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 578, + 285, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 285, + 591 + ], + "score": 1.0, + "content": "perplexity for the largest training block size.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5 + }, + { + "type": "title", + "bbox": [ + 107, + 605, + 309, + 617 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 310, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 310, + 618 + ], + "score": 1.0, + "content": "5.4 ADAPTIVE SOFTMAX VS. FULL SOFTMAX", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 626, + 504, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "We also found that adaptive softmax can benefit from additional regularization of rare words. Adap-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "tive softmax first projects the model output to the dimension of a particular cluster and then computes", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "score": 1.0, + "content": "a dot product with the respective word embeddings. We add dropout to the output of the first pro-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 660, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 104, + 660, + 506, + 673 + ], + "score": 1.0, + "content": "jection for all clusters, except for the head. This change enables the adaptive softmax to outperform", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 671, + 464, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 464, + 684 + ], + "score": 1.0, + "content": "a standard softmax over fixed size output word embeddings on WIKITEXT-103 (Table 6).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "However, we found that adding dropout in this way is not helpful for larger datasets such as BILLION", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "WORD. Unfortunately, a standard softmax over 800K words is not tractable and we were unable to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "make a comparison. It may be possible to achieve better results by tuning dropout for each band of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 721, + 272, + 731 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 272, + 731 + ], + "score": 1.0, + "content": "the tail and we leave this for future work.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 78, + 503, + 267 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 78, + 503, + 267 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 78, + 503, + 267 + ], + "spans": [ + { + "bbox": [ + 108, + 78, + 503, + 267 + ], + "score": 0.972, + "type": "image", + "image_path": "d73dcb6e7e5c4af036340a7099e37fe1418673e98a1f8915cde2d051627402b8.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 78, + 503, + 141.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 141.0, + 503, + 204.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 204.0, + 503, + 267.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 278, + 505, + 300 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 277, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 506, + 291 + ], + "score": 1.0, + "content": "Figure 4: Loss of models when binning by word frequency on the test set of BILLION WORD. Bins", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 290, + 185, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 185, + 301 + ], + "score": 1.0, + "content": "are not cumulative.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 204, + 314, + 406, + 419 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 204, + 314, + 406, + 419 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 204, + 314, + 406, + 419 + ], + "spans": [ + { + "bbox": [ + 204, + 314, + 406, + 419 + ], + "score": 0.95, + "html": "
Train block sizeInference context sizeValidation perplexityTest perplexity
51219.7920.51
5120 48018.3519.03
2048018.9619.53
2048153618.2318.88
3072018.6319.34
3072256017.9718.70
", + "type": "table", + "image_path": "34ba39af36272851bc56541c57fa7eeaea8880ea2ca40ce648a9adb53f077da7.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 204, + 314, + 406, + 327.125 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 204, + 327.125, + 406, + 340.25 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 204, + 340.25, + 406, + 353.375 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 204, + 353.375, + 406, + 366.5 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 204, + 366.5, + 406, + 379.625 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 204, + 379.625, + 406, + 392.75 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 204, + 392.75, + 406, + 405.875 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 204, + 405.875, + 406, + 419.0 + ], + "spans": [], + "index": 12 + } + ] + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 427, + 504, + 460 + ], + "lines": [ + { + "bbox": [ + 105, + 425, + 504, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 504, + 441 + ], + "score": 1.0, + "content": "Table 5: Perplexity on WIKITEXT-103 with different context sizes during training and inference.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "score": 1.0, + "content": "Training block size is the number of consecutive tokens considered during training. Inference con-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 449, + 402, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 402, + 461 + ], + "score": 1.0, + "content": "text is the number of tokens provided at evaluation before scoring tokens.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 425, + 505, + 461 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 484, + 505, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "score": 1.0, + "content": "analysis (Appendix 5.3) binning the loss by the frequency of the previous word shows that weight", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 496, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 507 + ], + "score": 1.0, + "content": "sharing also helps for BILLION WORD and that CNN does very well on rare words for BILLION", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 507, + 244, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 244, + 518 + ], + "score": 1.0, + "content": "WORD compared to other models.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 483, + 506, + 518 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 505, + 589 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "Table 5 shows the importance of context size for WIKITEXT-103. Training block size is the number", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 535, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 505, + 546 + ], + "score": 1.0, + "content": "of consecutive tokens that are considered at once during training. Inference context is the number of", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 543, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 559 + ], + "score": 1.0, + "content": "tokens that are provided at evaluation before any tokens are scored. Simply increasing the training", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "block size from 512 to 3072 results in an improvement of nearly 1.2 perplexity with no inference", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 567, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 579 + ], + "score": 1.0, + "content": "context window. Increasing the context size at inference time results in an improvement of 0.6", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 578, + 285, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 285, + 591 + ], + "score": 1.0, + "content": "perplexity for the largest training block size.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 523, + 506, + 591 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 605, + 309, + 617 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 310, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 310, + 618 + ], + "score": 1.0, + "content": "5.4 ADAPTIVE SOFTMAX VS. FULL SOFTMAX", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 626, + 504, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "We also found that adaptive softmax can benefit from additional regularization of rare words. Adap-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "tive softmax first projects the model output to the dimension of a particular cluster and then computes", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "score": 1.0, + "content": "a dot product with the respective word embeddings. We add dropout to the output of the first pro-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 660, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 104, + 660, + 506, + 673 + ], + "score": 1.0, + "content": "jection for all clusters, except for the head. This change enables the adaptive softmax to outperform", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 671, + 464, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 464, + 684 + ], + "score": 1.0, + "content": "a standard softmax over fixed size output word embeddings on WIKITEXT-103 (Table 6).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28, + "bbox_fs": [ + 104, + 626, + 506, + 684 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "However, we found that adding dropout in this way is not helpful for larger datasets such as BILLION", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "WORD. Unfortunately, a standard softmax over 800K words is not tractable and we were unable to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "make a comparison. It may be possible to achieve better results by tuning dropout for each band of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 721, + 272, + 731 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 272, + 731 + ], + "score": 1.0, + "content": "the tail and we leave this for future work.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 688, + 506, + 731 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 221, + 80, + 391, + 152 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 221, + 80, + 391, + 152 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 221, + 80, + 391, + 152 + ], + "spans": [ + { + "bbox": [ + 221, + 80, + 391, + 152 + ], + "score": 0.974, + "html": "
Tail dropoutValidation perplexity
Softmax (SM)N/A23.87
Adaptive (ADP)0.024.74
Adaptive (ADP)0.221.23
", + "type": "table", + "image_path": "787763e4f22c859d1fb0dcb91dcc43e757ceb5d613ad01d21320a2ab8bb917d1.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 221, + 80, + 391, + 94.4 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 221, + 94.4, + 391, + 108.80000000000001 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 221, + 108.80000000000001, + 391, + 123.20000000000002 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 221, + 123.20000000000002, + 391, + 137.60000000000002 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 221, + 137.60000000000002, + 391, + 152.00000000000003 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 126, + 160, + 483, + 172 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 159, + 485, + 173 + ], + "spans": [ + { + "bbox": [ + 126, + 159, + 485, + 173 + ], + "score": 1.0, + "content": "Table 6: Perplexity on WIKITEXT-103 when regularizing rare words in adaptive softmax.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 107, + 192, + 195, + 205 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 197, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 197, + 208 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 217, + 505, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 216, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 231 + ], + "score": 1.0, + "content": "Adaptive input embeddings vary the size of input word embeddings which can improve accuracy", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 228, + 504, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 504, + 242 + ], + "score": 1.0, + "content": "while drastically reducing the number of model parameters. When sharing parameters with an adap-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 239, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 505, + 253 + ], + "score": 1.0, + "content": "tive softmax, the number of parameters can be further reduced which improves training speed. We", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 250, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 264 + ], + "score": 1.0, + "content": "presented a comparison between different input and output layer factorizations including word in-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 262, + 388, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 388, + 275 + ], + "score": 1.0, + "content": "puts, character inputs and sub-word units in both the input and output.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 278, + 505, + 322 + ], + "lines": [ + { + "bbox": [ + 106, + 279, + 504, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 504, + 290 + ], + "score": 1.0, + "content": "Our experiments show that models with adaptive input embeddings train faster compared to char-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 289, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 506, + 302 + ], + "score": 1.0, + "content": "acter input CNNs while achieving higher accuracy. We achieve new state of the art results on", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 300, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 505, + 312 + ], + "score": 1.0, + "content": "WIKITEXT-103 and BILLION WORD. In future work, we will apply variable sized input embed-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 311, + 189, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 189, + 323 + ], + "score": 1.0, + "content": "dings to other tasks.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "title", + "bbox": [ + 108, + 336, + 200, + 347 + ], + "lines": [ + { + "bbox": [ + 107, + 336, + 200, + 347 + ], + "spans": [ + { + "bbox": [ + 107, + 336, + 200, + 347 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 345, + 366 + ], + "lines": [ + { + "bbox": [ + 106, + 353, + 347, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 347, + 368 + ], + "score": 1.0, + "content": "We thank Tom Bosc for fruitful comments and suggestions.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 108, + 383, + 175, + 395 + ], + "lines": [ + { + "bbox": [ + 106, + 383, + 176, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 176, + 397 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 402, + 504, + 424 + ], + "lines": [ + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "score": 1.0, + "content": "Rami Al-Rfou, Dokook Choe, Noah Constant, Mandy Guo, and Llion Jones. Character-level lan-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 413, + 412, + 425 + ], + "spans": [ + { + "bbox": [ + 115, + 413, + 412, + 425 + ], + "score": 1.0, + "content": "guage modeling with deeper self-attention. arXiv, abs/1808.04444, 2018.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 432, + 504, + 466 + ], + "lines": [ + { + "bbox": [ + 106, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "Ebru Arisoy, Tara N. Sainath, Brian Kingsbury, and Bhuvana Ramabhadran. Deep Neural Network", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 114, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 114, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "Language Models. In NAACL-HLT Workshop on the Future of Language Modeling for HLT,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 453, + 143, + 466 + ], + "spans": [ + { + "bbox": [ + 115, + 453, + 143, + 466 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 504, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "Paul Baltescu and Phil Blunsom. Pragmatic neural language modelling in machine translation. In", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 486, + 198, + 497 + ], + "spans": [ + { + "bbox": [ + 116, + 486, + 198, + 497 + ], + "score": 1.0, + "content": "Proc. of ACL, 2015.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 505, + 502, + 528 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 504, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 504, + 518 + ], + "score": 1.0, + "content": "Yoshua Bengio, Rejean Ducharme, Pascal Vincent, and Christian Jauvin. A Neural Probabilistic´", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 516, + 302, + 529 + ], + "spans": [ + { + "bbox": [ + 116, + 516, + 302, + 529 + ], + "score": 1.0, + "content": "Language Model. JMLR, 3:1137–1155, 2003.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 104, + 535, + 489, + 548 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 490, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 490, + 550 + ], + "score": 1.0, + "content": "Jacob Buckman and Graham Neubig. Neural lattice language models. TACL, 6:529–541, 2018.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 108, + 555, + 502, + 589 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 504, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 504, + 569 + ], + "score": 1.0, + "content": "Ciprian Chelba, Tomas Mikolov, Mike Schuster, Qi Ge, Thorsten Brants, Phillipp Koehn, and Tony", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 565, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 115, + 565, + 505, + 581 + ], + "score": 1.0, + "content": "Robinson. One billion word benchmark for measuring progress in statistical language modeling.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 578, + 245, + 590 + ], + "spans": [ + { + "bbox": [ + 116, + 578, + 245, + 590 + ], + "score": 1.0, + "content": "Technical report, Google, 2013.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 105, + 597, + 504, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 596, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 610 + ], + "score": 1.0, + "content": "Wenlin Chen, David Grangier, and Michael Auli. Strategies for training large vocabulary neural", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 608, + 282, + 621 + ], + "spans": [ + { + "bbox": [ + 115, + 608, + 282, + 621 + ], + "score": 1.0, + "content": "language models. In Proc. of ACL, 2016.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 104, + 627, + 504, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 627, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 641 + ], + "score": 1.0, + "content": "Yann N. Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 640, + 314, + 650 + ], + "spans": [ + { + "bbox": [ + 116, + 640, + 314, + 650 + ], + "score": 1.0, + "content": "convolutional networks. In Proc. of ICML, 2017.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 104, + 658, + 472, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 658, + 473, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 473, + 672 + ], + "score": 1.0, + "content": "Joshua Goodman. Classes for Fast Maximum Entropy Training. In Proc. of ICASSP, 2001.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 105, + 678, + 504, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 692 + ], + "score": 1.0, + "content": "Edouard Grave, Armand Joulin, and Nicolas Usunier. Improving neural language models with a", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 690, + 312, + 702 + ], + "spans": [ + { + "bbox": [ + 116, + 690, + 312, + 702 + ], + "score": 1.0, + "content": "continuous cache. arXiv, abs/1612.04426, 2016.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "Edouard Grave, Armand Joulin, Moustapha Cisse, David Grangier, and Herv ´ e J ´ egou. Efficient ´", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 720, + 350, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 350, + 732 + ], + "score": 1.0, + "content": "softmax approximation for gpus. In Proc. of ICML, 2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 221, + 80, + 391, + 152 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 221, + 80, + 391, + 152 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 221, + 80, + 391, + 152 + ], + "spans": [ + { + "bbox": [ + 221, + 80, + 391, + 152 + ], + "score": 0.974, + "html": "
Tail dropoutValidation perplexity
Softmax (SM)N/A23.87
Adaptive (ADP)0.024.74
Adaptive (ADP)0.221.23
", + "type": "table", + "image_path": "787763e4f22c859d1fb0dcb91dcc43e757ceb5d613ad01d21320a2ab8bb917d1.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 221, + 80, + 391, + 94.4 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 221, + 94.4, + 391, + 108.80000000000001 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 221, + 108.80000000000001, + 391, + 123.20000000000002 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 221, + 123.20000000000002, + 391, + 137.60000000000002 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 221, + 137.60000000000002, + 391, + 152.00000000000003 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 126, + 160, + 483, + 172 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 159, + 485, + 173 + ], + "spans": [ + { + "bbox": [ + 126, + 159, + 485, + 173 + ], + "score": 1.0, + "content": "Table 6: Perplexity on WIKITEXT-103 when regularizing rare words in adaptive softmax.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 107, + 192, + 195, + 205 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 197, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 197, + 208 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 217, + 505, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 216, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 231 + ], + "score": 1.0, + "content": "Adaptive input embeddings vary the size of input word embeddings which can improve accuracy", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 228, + 504, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 504, + 242 + ], + "score": 1.0, + "content": "while drastically reducing the number of model parameters. When sharing parameters with an adap-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 239, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 505, + 253 + ], + "score": 1.0, + "content": "tive softmax, the number of parameters can be further reduced which improves training speed. We", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 250, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 264 + ], + "score": 1.0, + "content": "presented a comparison between different input and output layer factorizations including word in-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 262, + 388, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 388, + 275 + ], + "score": 1.0, + "content": "puts, character inputs and sub-word units in both the input and output.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 216, + 505, + 275 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 278, + 505, + 322 + ], + "lines": [ + { + "bbox": [ + 106, + 279, + 504, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 504, + 290 + ], + "score": 1.0, + "content": "Our experiments show that models with adaptive input embeddings train faster compared to char-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 289, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 506, + 302 + ], + "score": 1.0, + "content": "acter input CNNs while achieving higher accuracy. We achieve new state of the art results on", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 300, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 505, + 312 + ], + "score": 1.0, + "content": "WIKITEXT-103 and BILLION WORD. In future work, we will apply variable sized input embed-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 311, + 189, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 189, + 323 + ], + "score": 1.0, + "content": "dings to other tasks.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 279, + 506, + 323 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 336, + 200, + 347 + ], + "lines": [ + { + "bbox": [ + 107, + 336, + 200, + 347 + ], + "spans": [ + { + "bbox": [ + 107, + 336, + 200, + 347 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 345, + 366 + ], + "lines": [ + { + "bbox": [ + 106, + 353, + 347, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 347, + 368 + ], + "score": 1.0, + "content": "We thank Tom Bosc for fruitful comments and suggestions.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 353, + 347, + 368 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 383, + 175, + 395 + ], + "lines": [ + { + "bbox": [ + 106, + 383, + 176, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 176, + 397 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 402, + 504, + 424 + ], + "lines": [ + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "score": 1.0, + "content": "Rami Al-Rfou, Dokook Choe, Noah Constant, Mandy Guo, and Llion Jones. Character-level lan-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 413, + 412, + 425 + ], + "spans": [ + { + "bbox": [ + 115, + 413, + 412, + 425 + ], + "score": 1.0, + "content": "guage modeling with deeper self-attention. arXiv, abs/1808.04444, 2018.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 402, + 505, + 425 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 432, + 504, + 466 + ], + "lines": [ + { + "bbox": [ + 106, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "Ebru Arisoy, Tara N. Sainath, Brian Kingsbury, and Bhuvana Ramabhadran. Deep Neural Network", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 114, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 114, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "Language Models. In NAACL-HLT Workshop on the Future of Language Modeling for HLT,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 453, + 143, + 466 + ], + "spans": [ + { + "bbox": [ + 115, + 453, + 143, + 466 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 432, + 505, + 466 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 504, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "Paul Baltescu and Phil Blunsom. Pragmatic neural language modelling in machine translation. In", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 486, + 198, + 497 + ], + "spans": [ + { + "bbox": [ + 116, + 486, + 198, + 497 + ], + "score": 1.0, + "content": "Proc. of ACL, 2015.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 474, + 505, + 497 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 505, + 502, + 528 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 504, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 504, + 518 + ], + "score": 1.0, + "content": "Yoshua Bengio, Rejean Ducharme, Pascal Vincent, and Christian Jauvin. A Neural Probabilistic´", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 516, + 302, + 529 + ], + "spans": [ + { + "bbox": [ + 116, + 516, + 302, + 529 + ], + "score": 1.0, + "content": "Language Model. JMLR, 3:1137–1155, 2003.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 106, + 505, + 504, + 529 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 535, + 489, + 548 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 490, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 490, + 550 + ], + "score": 1.0, + "content": "Jacob Buckman and Graham Neubig. Neural lattice language models. TACL, 6:529–541, 2018.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 534, + 490, + 550 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 555, + 502, + 589 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 504, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 504, + 569 + ], + "score": 1.0, + "content": "Ciprian Chelba, Tomas Mikolov, Mike Schuster, Qi Ge, Thorsten Brants, Phillipp Koehn, and Tony", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 565, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 115, + 565, + 505, + 581 + ], + "score": 1.0, + "content": "Robinson. One billion word benchmark for measuring progress in statistical language modeling.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 578, + 245, + 590 + ], + "spans": [ + { + "bbox": [ + 116, + 578, + 245, + 590 + ], + "score": 1.0, + "content": "Technical report, Google, 2013.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 106, + 555, + 505, + 590 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 597, + 504, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 596, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 610 + ], + "score": 1.0, + "content": "Wenlin Chen, David Grangier, and Michael Auli. Strategies for training large vocabulary neural", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 608, + 282, + 621 + ], + "spans": [ + { + "bbox": [ + 115, + 608, + 282, + 621 + ], + "score": 1.0, + "content": "language models. In Proc. of ACL, 2016.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 596, + 505, + 621 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 627, + 504, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 627, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 641 + ], + "score": 1.0, + "content": "Yann N. Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 640, + 314, + 650 + ], + "spans": [ + { + "bbox": [ + 116, + 640, + 314, + 650 + ], + "score": 1.0, + "content": "convolutional networks. In Proc. of ICML, 2017.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 106, + 627, + 505, + 650 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 658, + 472, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 658, + 473, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 473, + 672 + ], + "score": 1.0, + "content": "Joshua Goodman. Classes for Fast Maximum Entropy Training. In Proc. of ICASSP, 2001.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 658, + 473, + 672 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 678, + 504, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 692 + ], + "score": 1.0, + "content": "Edouard Grave, Armand Joulin, and Nicolas Usunier. Improving neural language models with a", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 690, + 312, + 702 + ], + "spans": [ + { + "bbox": [ + 116, + 690, + 312, + 702 + ], + "score": 1.0, + "content": "continuous cache. arXiv, abs/1612.04426, 2016.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 677, + 506, + 702 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "Edouard Grave, Armand Joulin, Moustapha Cisse, David Grangier, and Herv ´ e J ´ egou. Efficient ´", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 720, + 350, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 350, + 732 + ], + "score": 1.0, + "content": "softmax approximation for gpus. In Proc. of ICML, 2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 52, + 507, + 736 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 97 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 114, + 93, + 272, + 108 + ], + "spans": [ + { + "bbox": [ + 114, + 93, + 272, + 108 + ], + "score": 1.0, + "content": "Recognition. In Proc. of CVPR, 2015.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 113, + 506, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 113, + 506, + 125 + ], + "score": 1.0, + "content": "Hakan Inan, Khashayar Khosravi, and Richard Socher. Tying word vectors and word classifiers: A", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 124, + 398, + 137 + ], + "spans": [ + { + "bbox": [ + 114, + 124, + 398, + 137 + ], + "score": 1.0, + "content": "loss framework for language modeling. arXiv, abs/1611.01462, 2016.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 141, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 104, + 141, + 506, + 157 + ], + "score": 1.0, + "content": "Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring the ´", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 154, + 356, + 168 + ], + "spans": [ + { + "bbox": [ + 114, + 154, + 356, + 168 + ], + "score": 1.0, + "content": "limits of language modeling. arXiv, abs/1602.02410, 2016.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 172, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 189 + ], + "score": 1.0, + "content": "Yoon Kim, Yacine Jernite, David Sontag, and Alexander M. Rush. Character-aware neural language", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 185, + 272, + 197 + ], + "spans": [ + { + "bbox": [ + 115, + 185, + 272, + 197 + ], + "score": 1.0, + "content": "models. arXiv, abs/1508.06615, 2015.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 201, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 506, + 219 + ], + "score": 1.0, + "content": "Yoon Kim, Yacine Jernite, David Sontag, and Alexander M Rush. Character-aware neural language", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 214, + 279, + 228 + ], + "spans": [ + { + "bbox": [ + 114, + 214, + 279, + 228 + ], + "score": 1.0, + "content": "models. In AAAI, pp. 2741–2749, 2016.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 233, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 506, + 247 + ], + "score": 1.0, + "content": "Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with restarts. arXiv,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 245, + 210, + 257 + ], + "spans": [ + { + "bbox": [ + 115, + 245, + 210, + 257 + ], + "score": 1.0, + "content": "abs/1608.03983, 2016.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 264, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 276 + ], + "score": 1.0, + "content": "Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 275, + 272, + 288 + ], + "spans": [ + { + "bbox": [ + 114, + 275, + 272, + 288 + ], + "score": 1.0, + "content": "models. arXiv, abs/1609.07843, 2016.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 293, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 506, + 308 + ], + "score": 1.0, + "content": "Stephen Merity, Nitish Shirish Keskar, and Richard Socher. An analysis of neural language modeling", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 114, + 305, + 313, + 318 + ], + "spans": [ + { + "bbox": [ + 114, + 305, + 313, + 318 + ], + "score": 1.0, + "content": "at multiple scales. arXiv, abs/1803.08240, 2018.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 323, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 104, + 323, + 506, + 339 + ], + "score": 1.0, + "content": "Sebastian J. Mielke and Jason Eisner. Spell once, summon anywhere: A two-level open-vocabulary", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 336, + 307, + 348 + ], + "spans": [ + { + "bbox": [ + 115, + 336, + 307, + 348 + ], + "score": 1.0, + "content": "language model. arXiv, abs/1804.08205, 2018.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 354, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 506, + 369 + ], + "score": 1.0, + "content": "Toma´s Mikolov, Karafi ˇ at Martin, Luk ´ a´s Burget, Jan Cernock ˇ y, and Sanjeev Khudanpur. Recurrent ´", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 366, + 403, + 378 + ], + "spans": [ + { + "bbox": [ + 115, + 366, + 403, + 378 + ], + "score": 1.0, + "content": "Neural Network based Language Model. In Proc. of Interspeech, 2010.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 384, + 504, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 504, + 399 + ], + "score": 1.0, + "content": "Toma´s Mikolov, Stefan Kombrink, Luk ˇ a´s Burget, Jan Cernock ˇ y, and Sanjeev Khudanpur. Exten-´", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 114, + 396, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 114, + 396, + 505, + 410 + ], + "score": 1.0, + "content": "sions of Recurrent Neural Network Language Model. In Proc. of ICASSP, pp. 5528–5531, 2011.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "score": 1.0, + "content": "Frederic Morin and Yoshua Bengio. Hierarchical Probabilistic Neural Network Language Model.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 114, + 427, + 227, + 439 + ], + "spans": [ + { + "bbox": [ + 114, + 427, + 227, + 439 + ], + "score": 1.0, + "content": "In Proc. of AISTATS, 2005.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "score": 1.0, + "content": "Myle Ott, Michael Auli, David Grangier, and MarcAurelio Ranzato. Analyzing uncertainty in neural", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 114, + 456, + 305, + 470 + ], + "spans": [ + { + "bbox": [ + 114, + 456, + 305, + 470 + ], + "score": 1.0, + "content": "machine translation. In Proc. of ICML, 2018a.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 474, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 104, + 474, + 505, + 490 + ], + "score": 1.0, + "content": "Myle Ott, Sergey Edunov, David Grangier, and Michael Auli. Scaling neural machine translation.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 114, + 486, + 218, + 501 + ], + "spans": [ + { + "bbox": [ + 114, + 486, + 218, + 501 + ], + "score": 1.0, + "content": "In Proc. of WMT, 2018b.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 505, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 520 + ], + "score": 1.0, + "content": "Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 114, + 517, + 259, + 531 + ], + "spans": [ + { + "bbox": [ + 114, + 517, + 259, + 531 + ], + "score": 1.0, + "content": "networks. In Proc. of ICML, 2013.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 535, + 507, + 550 + ], + "spans": [ + { + "bbox": [ + 104, + 535, + 507, + 550 + ], + "score": 1.0, + "content": "Ofir Press and Lior Wolf. Using the output embedding to improve language models. In Proc. of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 114, + 547, + 172, + 561 + ], + "spans": [ + { + "bbox": [ + 114, + 547, + 172, + 561 + ], + "score": 1.0, + "content": "EACL, 2017.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "score": 1.0, + "content": "Jack W. Rae, Chris Dyer, Peter Dayan, and Timothy P. Lillicrap. Fast parametric learning with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 114, + 577, + 340, + 591 + ], + "spans": [ + { + "bbox": [ + 114, + 577, + 340, + 591 + ], + "score": 1.0, + "content": "activation memorization. arXiv, abs/1803.10049, 2018.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "score": 1.0, + "content": "Holger Schwenk, Anthony Rousseau, and Mohammed Attik. Large, Pruned or Continuous Space", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 114, + 606, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 114, + 606, + 506, + 623 + ], + "score": 1.0, + "content": "Language Models on a GPU for Statistical Machine Translation. In NAACL-HLT Workshop on", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 619, + 315, + 632 + ], + "spans": [ + { + "bbox": [ + 115, + 619, + 315, + 632 + ], + "score": 1.0, + "content": "the Future of Language Modeling for HLT, 2012.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 114, + 650, + 272, + 662 + ], + "spans": [ + { + "bbox": [ + 114, + 650, + 272, + 662 + ], + "score": 1.0, + "content": "subword units. In Proc. of ACL, 2016.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 667, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 682 + ], + "score": 1.0, + "content": "Noam Shazeer, Joris Pelemans, and Ciprian Chelba. Sparse non-negative matrix language modeling.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 113, + 679, + 238, + 692 + ], + "spans": [ + { + "bbox": [ + 113, + 679, + 238, + 692 + ], + "score": 1.0, + "content": "In Proc. of Interspeech, 2016.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 698, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 713 + ], + "score": 1.0, + "content": "Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc V. Le, Geoffrey E. Hinton,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 114, + 709, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 114, + 709, + 506, + 725 + ], + "score": 1.0, + "content": "and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 114, + 720, + 238, + 734 + ], + "spans": [ + { + "bbox": [ + 114, + 720, + 238, + 734 + ], + "score": 1.0, + "content": "arXiv, abs/1701.06538, 2017.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 21.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 104, + 52, + 507, + 736 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 97 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 93, + 272, + 108 + ], + "spans": [ + { + "bbox": [ + 114, + 93, + 272, + 108 + ], + "score": 1.0, + "content": "Recognition. In Proc. of CVPR, 2015.", + "type": "text" + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 113, + 506, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 113, + 506, + 125 + ], + "score": 1.0, + "content": "Hakan Inan, Khashayar Khosravi, and Richard Socher. Tying word vectors and word classifiers: A", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 124, + 398, + 137 + ], + "spans": [ + { + "bbox": [ + 114, + 124, + 398, + 137 + ], + "score": 1.0, + "content": "loss framework for language modeling. arXiv, abs/1611.01462, 2016.", + "type": "text" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 141, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 104, + 141, + 506, + 157 + ], + "score": 1.0, + "content": "Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring the ´", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 154, + 356, + 168 + ], + "spans": [ + { + "bbox": [ + 114, + 154, + 356, + 168 + ], + "score": 1.0, + "content": "limits of language modeling. arXiv, abs/1602.02410, 2016.", + "type": "text" + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 172, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 189 + ], + "score": 1.0, + "content": "Yoon Kim, Yacine Jernite, David Sontag, and Alexander M. Rush. Character-aware neural language", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 185, + 272, + 197 + ], + "spans": [ + { + "bbox": [ + 115, + 185, + 272, + 197 + ], + "score": 1.0, + "content": "models. arXiv, abs/1508.06615, 2015.", + "type": "text" + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 201, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 506, + 219 + ], + "score": 1.0, + "content": "Yoon Kim, Yacine Jernite, David Sontag, and Alexander M Rush. Character-aware neural language", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 214, + 279, + 228 + ], + "spans": [ + { + "bbox": [ + 114, + 214, + 279, + 228 + ], + "score": 1.0, + "content": "models. In AAAI, pp. 2741–2749, 2016.", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 233, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 506, + 247 + ], + "score": 1.0, + "content": "Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with restarts. arXiv,", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 245, + 210, + 257 + ], + "spans": [ + { + "bbox": [ + 115, + 245, + 210, + 257 + ], + "score": 1.0, + "content": "abs/1608.03983, 2016.", + "type": "text" + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 264, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 276 + ], + "score": 1.0, + "content": "Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 275, + 272, + 288 + ], + "spans": [ + { + "bbox": [ + 114, + 275, + 272, + 288 + ], + "score": 1.0, + "content": "models. arXiv, abs/1609.07843, 2016.", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 293, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 506, + 308 + ], + "score": 1.0, + "content": "Stephen Merity, Nitish Shirish Keskar, and Richard Socher. An analysis of neural language modeling", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 305, + 313, + 318 + ], + "spans": [ + { + "bbox": [ + 114, + 305, + 313, + 318 + ], + "score": 1.0, + "content": "at multiple scales. arXiv, abs/1803.08240, 2018.", + "type": "text" + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 323, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 104, + 323, + 506, + 339 + ], + "score": 1.0, + "content": "Sebastian J. Mielke and Jason Eisner. Spell once, summon anywhere: A two-level open-vocabulary", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 336, + 307, + 348 + ], + "spans": [ + { + "bbox": [ + 115, + 336, + 307, + 348 + ], + "score": 1.0, + "content": "language model. arXiv, abs/1804.08205, 2018.", + "type": "text" + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 354, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 506, + 369 + ], + "score": 1.0, + "content": "Toma´s Mikolov, Karafi ˇ at Martin, Luk ´ a´s Burget, Jan Cernock ˇ y, and Sanjeev Khudanpur. Recurrent ´", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 366, + 403, + 378 + ], + "spans": [ + { + "bbox": [ + 115, + 366, + 403, + 378 + ], + "score": 1.0, + "content": "Neural Network based Language Model. In Proc. of Interspeech, 2010.", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 384, + 504, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 504, + 399 + ], + "score": 1.0, + "content": "Toma´s Mikolov, Stefan Kombrink, Luk ˇ a´s Burget, Jan Cernock ˇ y, and Sanjeev Khudanpur. Exten-´", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 396, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 114, + 396, + 505, + 410 + ], + "score": 1.0, + "content": "sions of Recurrent Neural Network Language Model. In Proc. of ICASSP, pp. 5528–5531, 2011.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "score": 1.0, + "content": "Frederic Morin and Yoshua Bengio. Hierarchical Probabilistic Neural Network Language Model.", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 427, + 227, + 439 + ], + "spans": [ + { + "bbox": [ + 114, + 427, + 227, + 439 + ], + "score": 1.0, + "content": "In Proc. of AISTATS, 2005.", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "score": 1.0, + "content": "Myle Ott, Michael Auli, David Grangier, and MarcAurelio Ranzato. Analyzing uncertainty in neural", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 456, + 305, + 470 + ], + "spans": [ + { + "bbox": [ + 114, + 456, + 305, + 470 + ], + "score": 1.0, + "content": "machine translation. In Proc. of ICML, 2018a.", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 474, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 104, + 474, + 505, + 490 + ], + "score": 1.0, + "content": "Myle Ott, Sergey Edunov, David Grangier, and Michael Auli. Scaling neural machine translation.", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 486, + 218, + 501 + ], + "spans": [ + { + "bbox": [ + 114, + 486, + 218, + 501 + ], + "score": 1.0, + "content": "In Proc. of WMT, 2018b.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 505, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 520 + ], + "score": 1.0, + "content": "Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 517, + 259, + 531 + ], + "spans": [ + { + "bbox": [ + 114, + 517, + 259, + 531 + ], + "score": 1.0, + "content": "networks. In Proc. of ICML, 2013.", + "type": "text" + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 535, + 507, + 550 + ], + "spans": [ + { + "bbox": [ + 104, + 535, + 507, + 550 + ], + "score": 1.0, + "content": "Ofir Press and Lior Wolf. Using the output embedding to improve language models. In Proc. of", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 547, + 172, + 561 + ], + "spans": [ + { + "bbox": [ + 114, + 547, + 172, + 561 + ], + "score": 1.0, + "content": "EACL, 2017.", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "score": 1.0, + "content": "Jack W. Rae, Chris Dyer, Peter Dayan, and Timothy P. Lillicrap. Fast parametric learning with", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 577, + 340, + 591 + ], + "spans": [ + { + "bbox": [ + 114, + 577, + 340, + 591 + ], + "score": 1.0, + "content": "activation memorization. arXiv, abs/1803.10049, 2018.", + "type": "text" + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "score": 1.0, + "content": "Holger Schwenk, Anthony Rousseau, and Mohammed Attik. Large, Pruned or Continuous Space", + "type": "text" + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 606, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 114, + 606, + 506, + 623 + ], + "score": 1.0, + "content": "Language Models on a GPU for Statistical Machine Translation. In NAACL-HLT Workshop on", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 619, + 315, + 632 + ], + "spans": [ + { + "bbox": [ + 115, + 619, + 315, + 632 + ], + "score": 1.0, + "content": "the Future of Language Modeling for HLT, 2012.", + "type": "text" + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 650, + 272, + 662 + ], + "spans": [ + { + "bbox": [ + 114, + 650, + 272, + 662 + ], + "score": 1.0, + "content": "subword units. In Proc. of ACL, 2016.", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 667, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 682 + ], + "score": 1.0, + "content": "Noam Shazeer, Joris Pelemans, and Ciprian Chelba. Sparse non-negative matrix language modeling.", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 679, + 238, + 692 + ], + "spans": [ + { + "bbox": [ + 113, + 679, + 238, + 692 + ], + "score": 1.0, + "content": "In Proc. of Interspeech, 2016.", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 698, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 713 + ], + "score": 1.0, + "content": "Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc V. Le, Geoffrey E. Hinton,", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 709, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 114, + 709, + 506, + 725 + ], + "score": 1.0, + "content": "and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 114, + 720, + 238, + 734 + ], + "spans": [ + { + "bbox": [ + 114, + 720, + 238, + 734 + ], + "score": 1.0, + "content": "arXiv, abs/1701.06538, 2017.", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + } + ], + "index": 21.5, + "bbox_fs": [ + 104, + 82, + 507, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Ilya Sutskever, James Martens, George E. Dahl, and Geoffrey E. Hinton. On the importance of", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 404, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 404, + 106 + ], + "score": 1.0, + "content": "initialization and momentum in deep learning. In Proc. of ICML, 2013.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 106, + 111, + 504, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 504, + 124 + ], + "score": 1.0, + "content": "Ashish Vaswani, Yinggong Zhao, Victoria Fossum, and David Chiang. Decoding with Large-scale", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 453, + 136 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 453, + 136 + ], + "score": 1.0, + "content": "Neural Language Models improves Translation. In Proc. of EMNLP, October 2013.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 108, + 142, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 106, + 141, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 505, + 155 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 154, + 471, + 165 + ], + "spans": [ + { + "bbox": [ + 116, + 154, + 471, + 165 + ], + "score": 1.0, + "content": "Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. In Proc. of NIPS, 2017.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Ilya Sutskever, James Martens, George E. Dahl, and Geoffrey E. Hinton. On the importance of", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 404, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 404, + 106 + ], + "score": 1.0, + "content": "initialization and momentum in deep learning. In Proc. of ICML, 2013.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 106, + 111, + 504, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 504, + 124 + ], + "score": 1.0, + "content": "Ashish Vaswani, Yinggong Zhao, Victoria Fossum, and David Chiang. Decoding with Large-scale", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 453, + 136 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 453, + 136 + ], + "score": 1.0, + "content": "Neural Language Models improves Translation. In Proc. of EMNLP, October 2013.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 106, + 111, + 504, + 136 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 142, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 106, + 141, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 505, + 155 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 154, + 471, + 165 + ], + "spans": [ + { + "bbox": [ + 116, + 154, + 471, + 165 + ], + "score": 1.0, + "content": "Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. In Proc. of NIPS, 2017.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 106, + 141, + 505, + 165 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 81, + 255, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 257, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 257, + 95 + ], + "score": 1.0, + "content": "SUPPLEMENTARY MATERIAL", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 108, + 367, + 120 + ], + "lines": [ + { + "bbox": [ + 106, + 108, + 368, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 108, + 368, + 122 + ], + "score": 1.0, + "content": "A ADDITIONAL EXPERIMENTS ON WIKITEXT-103", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 505, + 188 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "This appendix shows various ablation. Table 7 shows that reducing the capacity of fixed size word", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 144, + 504, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 504, + 155 + ], + "score": 1.0, + "content": "input embddings is beneficial on WIKITEXT-103. The next set of results in Table 7 shows results", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 155, + 504, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 504, + 167 + ], + "score": 1.0, + "content": "for various settings of the SM and SM-T models. We also experimented with sharing the head", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 164, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 506, + 179 + ], + "score": 1.0, + "content": "projection but found this to perform less well than not sharing it. Finally, Table 8 shows various", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 176, + 299, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 299, + 190 + ], + "score": 1.0, + "content": "band sizes for adaptive input word embbedings.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 194, + 504, + 216 + ], + "lines": [ + { + "bbox": [ + 106, + 194, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 505, + 206 + ], + "score": 1.0, + "content": "We also show the performance of Merity et al. (2018) who use an adaptive softmax with equally", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 204, + 487, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 487, + 217 + ], + "score": 1.0, + "content": "sized word representations and share the input and output embeddings (no dim reduction, tied).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "table", + "bbox": [ + 105, + 226, + 522, + 441 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 105, + 226, + 522, + 441 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 226, + 522, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 522, + 441 + ], + "score": 0.985, + "html": "
Input OutputDropout Valid PPLParameters
256d EmbeddingAdaptive0.323.39314.7M
128d EmbeddingAdaptive0.321.51280.3M
64d EmbeddingAdaptive0.321.23263.1M
32d EmbeddingAdaptive0.321.78254.5M
512d Embedding512d Softmax (tied)0.322.63339.7M
512d Embedding512d Softmax (tied)0.428.31339.7M
512d Embedding512d Softmax0.323.87476.8
512d Embedding512d Softmax0.427.64476.8
256d Embedding256d Softmax (tied)0.322.65270.6M
256d Embedding256d Softmax0.324.13339.1M
64d Embedding512d Softmax0.324.74356.3M
AdaptiveAdaptive (tied emb, not proj)0.320.06247.3M
AdaptiveAdaptive (tied emb/proj not head)0.319.79246.9M
AdaptiveAdaptive (tied emb/proj + head)0.320.06246.9M
512d Embedding512d Softmax (tied)0.322.63339.7M
512d Embedding 512d Adaptive (no dim reduction, tied)0.325.48340.2M
", + "type": "table", + "image_path": "3fbfe30985bf9514dacb12e0943de5c19d13b7cd9c0a18913baf4f9cc354e365.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 105, + 226, + 522, + 297.6666666666667 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 105, + 297.6666666666667, + 522, + 369.33333333333337 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 105, + 369.33333333333337, + 522, + 441.00000000000006 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 175, + 448, + 434, + 461 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 176, + 448, + 435, + 461 + ], + "spans": [ + { + "bbox": [ + 176, + 448, + 435, + 461 + ], + "score": 1.0, + "content": "Table 7: Validation perplexity of our models on WIKITEXT-103.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + } + ], + "index": 11.0 + }, + { + "type": "table", + "bbox": [ + 233, + 477, + 377, + 578 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 233, + 477, + 377, + 578 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 233, + 477, + 377, + 578 + ], + "spans": [ + { + "bbox": [ + 233, + 477, + 377, + 578 + ], + "score": 0.972, + "html": "
Softmax cutoffValid PPL
20k/40k/200k19.79
20k/140k/100k 20k/40k/60k/140k20.26
60k/100k/100k20.53 20.52
5k/155k/100k
20k/40k/200k20.06
10k/60k/190k19.99 19.79
", + "type": "table", + "image_path": "87847578c5c5390ef40a775e10e07897a5106fcd1e9ae1dbe98b927732c8091d.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 233, + 477, + 377, + 527.5 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 233, + 527.5, + 377, + 578.0 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 586, + 505, + 620 + ], + "lines": [ + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "score": 1.0, + "content": "Table 8: Validation perplexity on WIKITEXT-103 with tied adaptive inputs & outputs. The bands", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "signify the number of words belonging to each band. In every case, the first band has dimension", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 608, + 431, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 431, + 619 + ], + "score": 1.0, + "content": "1024, the second band 256, the third band 64 and the fourth band (if it exists) 16.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 107, + 642, + 181, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 641, + 182, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 182, + 658 + ], + "score": 1.0, + "content": "B ANALYSIS", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 108, + 667, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 257, + 681 + ], + "score": 1.0, + "content": "This appendix extends the analysis in", + "type": "text" + }, + { + "bbox": [ + 257, + 668, + 275, + 680 + ], + "score": 0.79, + "content": "\\ S 5 . 3", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 666, + 505, + 681 + ], + "score": 1.0, + "content": "by showing a breakdown of the test loss when binning by", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 678, + 251, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 251, + 691 + ], + "score": 1.0, + "content": "the frequency of the previous word.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 81, + 255, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 257, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 257, + 95 + ], + "score": 1.0, + "content": "SUPPLEMENTARY MATERIAL", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 106, + 80, + 257, + 95 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 108, + 367, + 120 + ], + "lines": [ + { + "bbox": [ + 106, + 108, + 368, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 108, + 368, + 122 + ], + "score": 1.0, + "content": "A ADDITIONAL EXPERIMENTS ON WIKITEXT-103", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 505, + 188 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "This appendix shows various ablation. Table 7 shows that reducing the capacity of fixed size word", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 144, + 504, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 504, + 155 + ], + "score": 1.0, + "content": "input embddings is beneficial on WIKITEXT-103. The next set of results in Table 7 shows results", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 155, + 504, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 504, + 167 + ], + "score": 1.0, + "content": "for various settings of the SM and SM-T models. We also experimented with sharing the head", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 164, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 506, + 179 + ], + "score": 1.0, + "content": "projection but found this to perform less well than not sharing it. Finally, Table 8 shows various", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 176, + 299, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 299, + 190 + ], + "score": 1.0, + "content": "band sizes for adaptive input word embbedings.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 132, + 506, + 190 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 194, + 504, + 216 + ], + "lines": [ + { + "bbox": [ + 106, + 194, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 505, + 206 + ], + "score": 1.0, + "content": "We also show the performance of Merity et al. (2018) who use an adaptive softmax with equally", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 204, + 487, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 487, + 217 + ], + "score": 1.0, + "content": "sized word representations and share the input and output embeddings (no dim reduction, tied).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 194, + 505, + 217 + ] + }, + { + "type": "table", + "bbox": [ + 105, + 226, + 522, + 441 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 105, + 226, + 522, + 441 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 226, + 522, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 522, + 441 + ], + "score": 0.985, + "html": "
Input OutputDropout Valid PPLParameters
256d EmbeddingAdaptive0.323.39314.7M
128d EmbeddingAdaptive0.321.51280.3M
64d EmbeddingAdaptive0.321.23263.1M
32d EmbeddingAdaptive0.321.78254.5M
512d Embedding512d Softmax (tied)0.322.63339.7M
512d Embedding512d Softmax (tied)0.428.31339.7M
512d Embedding512d Softmax0.323.87476.8
512d Embedding512d Softmax0.427.64476.8
256d Embedding256d Softmax (tied)0.322.65270.6M
256d Embedding256d Softmax0.324.13339.1M
64d Embedding512d Softmax0.324.74356.3M
AdaptiveAdaptive (tied emb, not proj)0.320.06247.3M
AdaptiveAdaptive (tied emb/proj not head)0.319.79246.9M
AdaptiveAdaptive (tied emb/proj + head)0.320.06246.9M
512d Embedding512d Softmax (tied)0.322.63339.7M
512d Embedding 512d Adaptive (no dim reduction, tied)0.325.48340.2M
", + "type": "table", + "image_path": "3fbfe30985bf9514dacb12e0943de5c19d13b7cd9c0a18913baf4f9cc354e365.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 105, + 226, + 522, + 297.6666666666667 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 105, + 297.6666666666667, + 522, + 369.33333333333337 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 105, + 369.33333333333337, + 522, + 441.00000000000006 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 175, + 448, + 434, + 461 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 176, + 448, + 435, + 461 + ], + "spans": [ + { + "bbox": [ + 176, + 448, + 435, + 461 + ], + "score": 1.0, + "content": "Table 7: Validation perplexity of our models on WIKITEXT-103.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + } + ], + "index": 11.0 + }, + { + "type": "table", + "bbox": [ + 233, + 477, + 377, + 578 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 233, + 477, + 377, + 578 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 233, + 477, + 377, + 578 + ], + "spans": [ + { + "bbox": [ + 233, + 477, + 377, + 578 + ], + "score": 0.972, + "html": "
Softmax cutoffValid PPL
20k/40k/200k19.79
20k/140k/100k 20k/40k/60k/140k20.26
60k/100k/100k20.53 20.52
5k/155k/100k
20k/40k/200k20.06
10k/60k/190k19.99 19.79
", + "type": "table", + "image_path": "87847578c5c5390ef40a775e10e07897a5106fcd1e9ae1dbe98b927732c8091d.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 233, + 477, + 377, + 527.5 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 233, + 527.5, + 377, + 578.0 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 586, + 505, + 620 + ], + "lines": [ + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "score": 1.0, + "content": "Table 8: Validation perplexity on WIKITEXT-103 with tied adaptive inputs & outputs. The bands", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "signify the number of words belonging to each band. In every case, the first band has dimension", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 608, + 431, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 431, + 619 + ], + "score": 1.0, + "content": "1024, the second band 256, the third band 64 and the fourth band (if it exists) 16.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 586, + 506, + 619 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 642, + 181, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 641, + 182, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 182, + 658 + ], + "score": 1.0, + "content": "B ANALYSIS", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 108, + 667, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 257, + 681 + ], + "score": 1.0, + "content": "This appendix extends the analysis in", + "type": "text" + }, + { + "bbox": [ + 257, + 668, + 275, + 680 + ], + "score": 0.79, + "content": "\\ S 5 . 3", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 666, + 505, + 681 + ], + "score": 1.0, + "content": "by showing a breakdown of the test loss when binning by", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 678, + 251, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 251, + 691 + ], + "score": 1.0, + "content": "the frequency of the previous word.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 666, + 505, + 691 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 291, + 504, + 481 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 291, + 504, + 481 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 291, + 504, + 481 + ], + "spans": [ + { + "bbox": [ + 108, + 291, + 504, + 481 + ], + "score": 0.968, + "type": "image", + "image_path": "62e4d7a047372091139f4e3c0bc67293ce7fc2b1aef9dae0b74901d294ce4ff9.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 291, + 504, + 354.3333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 354.3333333333333, + 504, + 417.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 417.66666666666663, + 504, + 480.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 492, + 506, + 515 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 492, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 505, + 506 + ], + "score": 1.0, + "content": "Figure 5: Loss of models when binning by the frequency of the previous word measured on BILLION", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 503, + 194, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 194, + 516 + ], + "score": 1.0, + "content": "WORD (cf. Figure 3).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 291, + 504, + 481 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 291, + 504, + 481 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 291, + 504, + 481 + ], + "spans": [ + { + "bbox": [ + 108, + 291, + 504, + 481 + ], + "score": 0.968, + "type": "image", + "image_path": "62e4d7a047372091139f4e3c0bc67293ce7fc2b1aef9dae0b74901d294ce4ff9.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 291, + 504, + 354.3333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 354.3333333333333, + 504, + 417.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 417.66666666666663, + 504, + 480.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 492, + 506, + 515 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 492, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 505, + 506 + ], + "score": 1.0, + "content": "Figure 5: Loss of models when binning by the frequency of the previous word measured on BILLION", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 503, + 194, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 194, + 516 + ], + "score": 1.0, + "content": "WORD (cf. Figure 3).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/ByxZX20qFQ/ByxZX20qFQ_model.json b/parse/train/ByxZX20qFQ/ByxZX20qFQ_model.json new file mode 100644 index 0000000000000000000000000000000000000000..86a1c1b347e12c03ee84c90266372fd20c288f80 --- /dev/null +++ b/parse/train/ByxZX20qFQ/ByxZX20qFQ_model.json @@ -0,0 +1,14691 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 620, + 1302, + 620, + 1302, + 925, + 398, + 925 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1595, + 1403, + 1595, + 1403, + 1809, + 298, + 1809 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1396, + 1402, + 1396, + 1402, + 1579, + 299, + 1579 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1058, + 1403, + 1058, + 1403, + 1211, + 298, + 1211 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1227, + 1403, + 1227, + 1403, + 1380, + 298, + 1380 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1825, + 1401, + 1825, + 1401, + 1978, + 299, + 1978 + ], + "score": 0.97 + }, + { + "category_id": 0, + "poly": [ + 299, + 220, + 1058, + 220, + 1058, + 324, + 299, + 324 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 314, + 376, + 681, + 376, + 681, + 468, + 314, + 468 + ], + "score": 0.943 + }, + { + "category_id": 2, + "poly": [ + 329, + 2006, + 1254, + 2006, + 1254, + 2034, + 329, + 2034 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 302, + 986, + 573, + 986, + 573, + 1022, + 302, + 1022 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 105, + 300, + 105 + ], + "score": 0.877 + }, + { + "category_id": 0, + "poly": [ + 773, + 550, + 927, + 550, + 927, + 583, + 773, + 583 + ], + "score": 0.866 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 857, + 2089, + 857, + 2112, + 841, + 2112 + ], + "score": 0.702 + }, + { + "category_id": 13, + "poly": [ + 987, + 1779, + 1040, + 1779, + 1040, + 1807, + 987, + 1807 + ], + "score": 0.87, + "latex": "61 \\%" + }, + { + "category_id": 13, + "poly": [ + 846, + 1718, + 902, + 1718, + 902, + 1747, + 846, + 1747 + ], + "score": 0.86, + "latex": "23 \\%" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 218.0, + 1060.0, + 218.0, + 1060.0, + 271.0, + 296.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 272.0, + 1064.0, + 272.0, + 1064.0, + 328.0, + 293.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 2001.0, + 1258.0, + 2001.0, + 1258.0, + 2039.0, + 334.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 983.0, + 579.0, + 983.0, + 579.0, + 1031.0, + 294.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 546.0, + 934.0, + 546.0, + 934.0, + 589.0, + 768.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2088.0, + 861.0, + 2088.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 620.0, + 1305.0, + 620.0, + 1305.0, + 654.0, + 395.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 650.0, + 1305.0, + 650.0, + 1305.0, + 685.0, + 394.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 681.0, + 1306.0, + 681.0, + 1306.0, + 717.0, + 394.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 712.0, + 1308.0, + 712.0, + 1308.0, + 747.0, + 393.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 744.0, + 1306.0, + 744.0, + 1306.0, + 776.0, + 393.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 775.0, + 1305.0, + 775.0, + 1305.0, + 807.0, + 394.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 803.0, + 1305.0, + 803.0, + 1305.0, + 838.0, + 393.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 833.0, + 1303.0, + 833.0, + 1303.0, + 867.0, + 394.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 864.0, + 1306.0, + 864.0, + 1306.0, + 899.0, + 393.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 893.0, + 1304.0, + 893.0, + 1304.0, + 932.0, + 392.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1595.0, + 1405.0, + 1595.0, + 1405.0, + 1630.0, + 294.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1626.0, + 1404.0, + 1626.0, + 1404.0, + 1660.0, + 293.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1655.0, + 1404.0, + 1655.0, + 1404.0, + 1693.0, + 292.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1687.0, + 1405.0, + 1687.0, + 1405.0, + 1719.0, + 293.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1718.0, + 845.0, + 1718.0, + 845.0, + 1752.0, + 294.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 1718.0, + 1404.0, + 1718.0, + 1404.0, + 1752.0, + 903.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1746.0, + 1407.0, + 1746.0, + 1407.0, + 1783.0, + 292.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1779.0, + 986.0, + 1779.0, + 986.0, + 1812.0, + 293.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1779.0, + 1051.0, + 1779.0, + 1051.0, + 1812.0, + 1041.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1397.0, + 1406.0, + 1397.0, + 1406.0, + 1432.0, + 295.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1427.0, + 1406.0, + 1427.0, + 1406.0, + 1461.0, + 292.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1457.0, + 1404.0, + 1457.0, + 1404.0, + 1491.0, + 294.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1489.0, + 1404.0, + 1489.0, + 1404.0, + 1520.0, + 295.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1517.0, + 1406.0, + 1517.0, + 1406.0, + 1554.0, + 293.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1548.0, + 876.0, + 1548.0, + 876.0, + 1584.0, + 294.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1059.0, + 1405.0, + 1059.0, + 1405.0, + 1092.0, + 296.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1088.0, + 1405.0, + 1088.0, + 1405.0, + 1124.0, + 293.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1119.0, + 1403.0, + 1119.0, + 1403.0, + 1152.0, + 297.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1148.0, + 1405.0, + 1148.0, + 1405.0, + 1185.0, + 294.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1179.0, + 1396.0, + 1179.0, + 1396.0, + 1215.0, + 293.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1227.0, + 1404.0, + 1227.0, + 1404.0, + 1260.0, + 296.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1255.0, + 1405.0, + 1255.0, + 1405.0, + 1293.0, + 293.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1286.0, + 1404.0, + 1286.0, + 1404.0, + 1325.0, + 293.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1319.0, + 1405.0, + 1319.0, + 1405.0, + 1352.0, + 296.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1351.0, + 1058.0, + 1351.0, + 1058.0, + 1384.0, + 296.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1826.0, + 1403.0, + 1826.0, + 1403.0, + 1859.0, + 297.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1855.0, + 1403.0, + 1855.0, + 1403.0, + 1888.0, + 295.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1886.0, + 1405.0, + 1886.0, + 1405.0, + 1923.0, + 293.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1916.0, + 1403.0, + 1916.0, + 1403.0, + 1953.0, + 294.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1944.0, + 1406.0, + 1944.0, + 1406.0, + 1983.0, + 293.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 377.0, + 681.0, + 377.0, + 681.0, + 408.0, + 314.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 405.0, + 578.0, + 405.0, + 578.0, + 440.0, + 311.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 438.0, + 572.0, + 438.0, + 572.0, + 470.0, + 312.0, + 470.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1416, + 1404, + 1416, + 1404, + 1599, + 298, + 1599 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1850, + 1403, + 1850, + 1403, + 2034, + 297, + 2034 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1246, + 1403, + 1246, + 1403, + 1401, + 298, + 1401 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1711, + 1403, + 1711, + 1403, + 1835, + 298, + 1835 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1108, + 1403, + 1108, + 1403, + 1231, + 299, + 1231 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 903, + 1404, + 903, + 1404, + 997, + 298, + 997 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 524, + 235, + 1177, + 235, + 1177, + 721, + 524, + 721 + ], + "score": 0.971 + }, + { + "category_id": 4, + "poly": [ + 297, + 754, + 1401, + 754, + 1401, + 848, + 297, + 848 + ], + "score": 0.959 + }, + { + "category_id": 0, + "poly": [ + 300, + 1643, + 877, + 1643, + 877, + 1679, + 300, + 1679 + ], + "score": 0.923 + }, + { + "category_id": 0, + "poly": [ + 301, + 1039, + 588, + 1039, + 588, + 1076, + 301, + 1076 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.715 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.135 + }, + { + "category_id": 13, + "poly": [ + 635, + 1881, + 787, + 1881, + 787, + 1916, + 635, + 1916 + ], + "score": 0.93, + "latex": "\\nu _ { i } \\cap \\nu _ { j } = \\emptyset" + }, + { + "category_id": 13, + "poly": [ + 956, + 1970, + 1013, + 1970, + 1013, + 2008, + 956, + 2008 + ], + "score": 0.92, + "latex": "\\frac { d } { k ^ { n - 1 } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1881, + 518, + 1881, + 518, + 1914, + 298, + 1914 + ], + "score": 0.92, + "latex": "\\mathcal { V } _ { 2 } , \\ldots , \\mathcal { V } _ { n - 1 } \\cup \\mathcal { V } _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1260, + 1850, + 1403, + 1850, + 1403, + 1882, + 1260, + 1882 + ], + "score": 0.91, + "latex": "\\nu ~ = ~ \\nu _ { 1 } \\cup" + }, + { + "category_id": 13, + "poly": [ + 949, + 1882, + 1020, + 1882, + 1020, + 1913, + 949, + 1913 + ], + "score": 0.9, + "latex": "i \\neq j" + }, + { + "category_id": 13, + "poly": [ + 1214, + 1973, + 1287, + 1973, + 1287, + 2001, + 1214, + 2001 + ], + "score": 0.89, + "latex": "k = 4" + }, + { + "category_id": 13, + "poly": [ + 833, + 1882, + 887, + 1882, + 887, + 1913, + 833, + 1913 + ], + "score": 0.89, + "latex": "\\forall i , j" + }, + { + "category_id": 13, + "poly": [ + 985, + 1914, + 1018, + 1914, + 1018, + 1942, + 985, + 1942 + ], + "score": 0.88, + "latex": "\\nu _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 733, + 1974, + 768, + 1974, + 768, + 2003, + 733, + 2003 + ], + "score": 0.88, + "latex": "\\nu _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1111, + 1882, + 1144, + 1882, + 1144, + 1912, + 1111, + 1912 + ], + "score": 0.88, + "latex": "\\nu _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 326, + 1973, + 359, + 1973, + 359, + 2003, + 326, + 2003 + ], + "score": 0.87, + "latex": "\\nu _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1265, + 757, + 1294, + 757, + 1294, + 786, + 1265, + 786 + ], + "score": 0.87, + "latex": "\\nu _ { i }" + }, + { + "category_id": 13, + "poly": [ + 486, + 1914, + 522, + 1914, + 522, + 1942, + 486, + 1942 + ], + "score": 0.87, + "latex": "\\nu _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1177, + 1944, + 1195, + 1944, + 1195, + 1970, + 1177, + 1970 + ], + "score": 0.8, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 524, + 818, + 542, + 818, + 542, + 844, + 524, + 844 + ], + "score": 0.76, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 545, + 1974, + 563, + 1974, + 563, + 2000, + 545, + 2000 + ], + "score": 0.74, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 640, + 648, + 665, + 648, + 665, + 668, + 640, + 668 + ], + "score": 0.74, + "latex": "\\nu _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 527, + 592, + 540, + 592, + 540, + 609, + 527, + 609 + ], + "score": 0.73, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 920, + 404, + 1008, + 404, + 1008, + 448, + 920, + 448 + ], + "score": 0.7, + "latex": "{ \\frac { d } { k ^ { n - 1 } } } \\to d" + }, + { + "category_id": 13, + "poly": [ + 1007, + 625, + 1035, + 625, + 1035, + 647, + 1007, + 647 + ], + "score": 0.69, + "latex": "\\nu _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1119, + 566, + 1175, + 566, + 1175, + 616, + 1119, + 616 + ], + "score": 0.58, + "latex": "\\frac { d } { k ^ { n - 1 } }" + }, + { + "category_id": 13, + "poly": [ + 768, + 355, + 782, + 355, + 782, + 374, + 768, + 374 + ], + "score": 0.46, + "latex": "d" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 231.0, + 877.0, + 231.0, + 877.0, + 266.0, + 809.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 351.0, + 787.0, + 351.0, + 787.0, + 376.0, + 783.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 416.0, + 784.0, + 416.0, + 784.0, + 451.0, + 707.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 414.0, + 919.0, + 414.0, + 919.0, + 448.0, + 916.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 414.0, + 1013.0, + 414.0, + 1013.0, + 448.0, + 1009.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 473.0, + 687.0, + 473.0, + 687.0, + 503.0, + 623.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 464.0, + 1052.0, + 464.0, + 1052.0, + 499.0, + 983.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 613.0, + 1006.0, + 613.0, + 1006.0, + 654.0, + 1000.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 613.0, + 1043.0, + 613.0, + 1043.0, + 654.0, + 1036.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 690.0, + 901.0, + 690.0, + 901.0, + 727.0, + 778.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 276.5, + 845.0, + 276.5, + 845.0, + 289.0, + 842.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 587.0, + 541.0, + 587.0, + 541.0, + 610.5, + 528.0, + 610.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 755.0, + 1264.0, + 755.0, + 1264.0, + 789.0, + 296.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 755.0, + 1403.0, + 755.0, + 1403.0, + 789.0, + 1295.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 785.0, + 1405.0, + 785.0, + 1405.0, + 822.0, + 295.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 818.0, + 523.0, + 818.0, + 523.0, + 848.0, + 296.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 818.0, + 883.0, + 818.0, + 883.0, + 848.0, + 543.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1640.0, + 882.0, + 1640.0, + 882.0, + 1684.0, + 291.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1037.0, + 593.0, + 1037.0, + 593.0, + 1081.0, + 292.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1417.0, + 1402.0, + 1417.0, + 1402.0, + 1448.0, + 296.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1447.0, + 1404.0, + 1447.0, + 1404.0, + 1482.0, + 294.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1476.0, + 1404.0, + 1476.0, + 1404.0, + 1513.0, + 293.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1506.0, + 1404.0, + 1506.0, + 1404.0, + 1542.0, + 294.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1537.0, + 1405.0, + 1537.0, + 1405.0, + 1571.0, + 292.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1567.0, + 444.0, + 1567.0, + 444.0, + 1602.0, + 294.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1847.0, + 1259.0, + 1847.0, + 1259.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 1847.0, + 1407.0, + 1847.0, + 1407.0, + 1885.0, + 1404.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 1881.0, + 634.0, + 1881.0, + 634.0, + 1917.0, + 519.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 1881.0, + 832.0, + 1881.0, + 832.0, + 1917.0, + 788.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 1881.0, + 948.0, + 1881.0, + 948.0, + 1917.0, + 888.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 1881.0, + 1110.0, + 1881.0, + 1110.0, + 1917.0, + 1021.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1917.0, + 1145.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1909.0, + 485.0, + 1909.0, + 485.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1909.0, + 984.0, + 1909.0, + 984.0, + 1947.0, + 523.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 1909.0, + 1408.0, + 1909.0, + 1408.0, + 1947.0, + 1019.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 1176.0, + 1942.0, + 1176.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1978.0, + 1196.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1967.0, + 325.0, + 1967.0, + 325.0, + 2013.0, + 290.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1967.0, + 544.0, + 1967.0, + 544.0, + 2013.0, + 360.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1967.0, + 732.0, + 1967.0, + 732.0, + 2013.0, + 564.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 1967.0, + 955.0, + 1967.0, + 955.0, + 2013.0, + 769.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 1967.0, + 1213.0, + 1967.0, + 1213.0, + 2013.0, + 1014.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1288.0, + 1967.0, + 1408.0, + 1967.0, + 1408.0, + 2013.0, + 1288.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2003.0, + 520.0, + 2003.0, + 520.0, + 2036.0, + 296.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1246.0, + 1403.0, + 1246.0, + 1403.0, + 1280.0, + 296.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1278.0, + 1404.0, + 1278.0, + 1404.0, + 1312.0, + 296.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1308.0, + 1405.0, + 1308.0, + 1405.0, + 1342.0, + 292.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1334.0, + 1405.0, + 1334.0, + 1405.0, + 1378.0, + 292.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1369.0, + 725.0, + 1369.0, + 725.0, + 1405.0, + 295.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1710.0, + 1407.0, + 1710.0, + 1407.0, + 1747.0, + 293.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1741.0, + 1405.0, + 1741.0, + 1405.0, + 1778.0, + 293.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1775.0, + 1404.0, + 1775.0, + 1404.0, + 1808.0, + 294.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1805.0, + 883.0, + 1805.0, + 883.0, + 1837.0, + 296.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1110.0, + 1404.0, + 1110.0, + 1404.0, + 1143.0, + 295.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1137.0, + 1405.0, + 1137.0, + 1405.0, + 1174.0, + 293.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1170.0, + 1407.0, + 1170.0, + 1407.0, + 1206.0, + 295.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1202.0, + 506.0, + 1202.0, + 506.0, + 1236.0, + 295.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 903.0, + 1405.0, + 903.0, + 1405.0, + 940.0, + 293.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 935.0, + 1404.0, + 935.0, + 1404.0, + 969.0, + 294.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 965.0, + 1057.0, + 965.0, + 1057.0, + 1002.0, + 294.0, + 1002.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1032, + 1404, + 1032, + 1404, + 1278, + 298, + 1278 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 468, + 1403, + 468, + 1403, + 653, + 298, + 653 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 669, + 1403, + 669, + 1403, + 852, + 299, + 852 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1293, + 1403, + 1293, + 1403, + 1447, + 299, + 1447 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 225, + 1403, + 225, + 1403, + 323, + 299, + 323 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 301, + 1656, + 1401, + 1656, + 1401, + 1749, + 301, + 1749 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 302, + 1548, + 1399, + 1548, + 1399, + 1641, + 302, + 1641 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 300, + 338, + 1398, + 338, + 1398, + 431, + 300, + 431 + ], + "score": 0.958 + }, + { + "category_id": 2, + "poly": [ + 324, + 2005, + 1270, + 2005, + 1270, + 2033, + 324, + 2033 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 299, + 1790, + 492, + 1790, + 492, + 1822, + 299, + 1822 + ], + "score": 0.885 + }, + { + "category_id": 0, + "poly": [ + 299, + 973, + 456, + 973, + 456, + 1004, + 299, + 1004 + ], + "score": 0.883 + }, + { + "category_id": 0, + "poly": [ + 301, + 902, + 679, + 902, + 679, + 937, + 301, + 937 + ], + "score": 0.878 + }, + { + "category_id": 0, + "poly": [ + 299, + 1488, + 489, + 1488, + 489, + 1520, + 299, + 1520 + ], + "score": 0.876 + }, + { + "category_id": 1, + "poly": [ + 299, + 1850, + 1403, + 1850, + 1403, + 1972, + 299, + 1972 + ], + "score": 0.674 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.668 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.303 + }, + { + "category_id": 13, + "poly": [ + 1073, + 1155, + 1236, + 1155, + 1236, + 1186, + 1073, + 1186 + ], + "score": 0.92, + "latex": "W _ { 1 } \\in \\mathbb { R } ^ { e _ { f f } \\times e }" + }, + { + "category_id": 13, + "poly": [ + 1257, + 562, + 1400, + 562, + 1400, + 593, + 1257, + 593 + ], + "score": 0.91, + "latex": "W _ { 1 } , \\ldots , W _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1285, + 1156, + 1394, + 1156, + 1394, + 1185, + 1285, + 1185 + ], + "score": 0.91, + "latex": "e = 1 0 2 4" + }, + { + "category_id": 13, + "poly": [ + 502, + 1155, + 822, + 1155, + 822, + 1189, + 502, + 1189 + ], + "score": 0.91, + "latex": "R e L U ( W _ { 1 } X + b _ { 1 } ) W _ { 2 } + b _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 901, + 1155, + 1061, + 1155, + 1061, + 1186, + 901, + 1186 + ], + "score": 0.91, + "latex": "W _ { 1 } \\in \\mathbb { R } ^ { e \\times e _ { f f } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1187, + 434, + 1187, + 434, + 1220, + 298, + 1220 + ], + "score": 0.91, + "latex": "e _ { f f } = 4 0 9 6" + }, + { + "category_id": 13, + "poly": [ + 650, + 224, + 1048, + 224, + 1048, + 263, + 650, + 263 + ], + "score": 0.9, + "latex": "W _ { 1 } \\in \\mathbb { R } ^ { d \\times d } , \\ldots , W _ { n } \\in \\mathbb { R } ^ { d / k ^ { n - 1 } \\times d }" + }, + { + "category_id": 13, + "poly": [ + 1179, + 1065, + 1279, + 1065, + 1279, + 1093, + 1179, + 1093 + ], + "score": 0.9, + "latex": "N = 1 6" + }, + { + "category_id": 13, + "poly": [ + 664, + 1125, + 766, + 1125, + 766, + 1153, + 664, + 1153 + ], + "score": 0.88, + "latex": "H = 1 6" + }, + { + "category_id": 13, + "poly": [ + 1011, + 293, + 1044, + 293, + 1044, + 323, + 1011, + 323 + ], + "score": 0.88, + "latex": "\\nu _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 624, + 365, + 624, + 365, + 652, + 298, + 652 + ], + "score": 0.85, + "latex": "n - 1" + }, + { + "category_id": 13, + "poly": [ + 347, + 502, + 367, + 502, + 367, + 528, + 347, + 528 + ], + "score": 0.8, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 448, + 401, + 466, + 401, + 466, + 427, + 448, + 427 + ], + "score": 0.77, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 529, + 263, + 548, + 263, + 548, + 289, + 529, + 289 + ], + "score": 0.74, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1377, + 293, + 1394, + 293, + 1394, + 319, + 1377, + 319 + ], + "score": 0.65, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 298, + 1422, + 313, + 1422, + 313, + 1443, + 298, + 1443 + ], + "score": 0.58, + "latex": "e" + }, + { + "category_id": 13, + "poly": [ + 1341, + 471, + 1363, + 471, + 1363, + 498, + 1341, + 498 + ], + "score": 0.52, + "latex": "\\nu" + }, + { + "category_id": 13, + "poly": [ + 1377, + 472, + 1394, + 472, + 1394, + 497, + 1377, + 497 + ], + "score": 0.43, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1340, + 469, + 1396, + 469, + 1396, + 499, + 1340, + 499 + ], + "score": 0.3, + "latex": "\\nu , d ," + }, + { + "category_id": 15, + "poly": [ + 330.0, + 2001.0, + 1275.0, + 2001.0, + 1275.0, + 2038.0, + 330.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1787.0, + 497.0, + 1787.0, + 497.0, + 1826.0, + 293.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 971.0, + 463.0, + 971.0, + 463.0, + 1008.0, + 294.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 901.0, + 681.0, + 901.0, + 681.0, + 941.0, + 294.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1485.0, + 494.0, + 1485.0, + 494.0, + 1525.0, + 293.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1032.0, + 1405.0, + 1032.0, + 1405.0, + 1066.0, + 296.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1064.0, + 1178.0, + 1064.0, + 1178.0, + 1096.0, + 292.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 1064.0, + 1405.0, + 1064.0, + 1405.0, + 1096.0, + 1280.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1094.0, + 1404.0, + 1094.0, + 1404.0, + 1126.0, + 293.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1125.0, + 663.0, + 1125.0, + 663.0, + 1155.0, + 294.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1125.0, + 1404.0, + 1125.0, + 1404.0, + 1155.0, + 767.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1153.0, + 501.0, + 1153.0, + 501.0, + 1190.0, + 293.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 1153.0, + 900.0, + 1153.0, + 900.0, + 1190.0, + 823.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 1153.0, + 1072.0, + 1153.0, + 1072.0, + 1190.0, + 1062.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1153.0, + 1284.0, + 1153.0, + 1284.0, + 1190.0, + 1237.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 1153.0, + 1406.0, + 1153.0, + 1406.0, + 1190.0, + 1395.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1184.0, + 297.0, + 1184.0, + 297.0, + 1221.0, + 292.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1184.0, + 1406.0, + 1184.0, + 1406.0, + 1221.0, + 435.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1215.0, + 1404.0, + 1215.0, + 1404.0, + 1251.0, + 292.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1245.0, + 1071.0, + 1245.0, + 1071.0, + 1280.0, + 293.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 470.0, + 1339.0, + 470.0, + 1339.0, + 502.0, + 297.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 470.0, + 1404.0, + 470.0, + 1404.0, + 502.0, + 1397.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 499.0, + 346.0, + 499.0, + 346.0, + 534.0, + 293.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 499.0, + 1405.0, + 499.0, + 1405.0, + 534.0, + 368.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 529.0, + 1405.0, + 529.0, + 1405.0, + 566.0, + 293.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 562.0, + 1256.0, + 562.0, + 1256.0, + 597.0, + 296.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 591.0, + 1405.0, + 591.0, + 1405.0, + 627.0, + 294.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 624.0, + 297.0, + 624.0, + 297.0, + 656.0, + 294.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 624.0, + 1319.0, + 624.0, + 1319.0, + 656.0, + 366.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 667.0, + 1405.0, + 667.0, + 1405.0, + 704.0, + 295.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 699.0, + 1405.0, + 699.0, + 1405.0, + 733.0, + 292.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 728.0, + 1405.0, + 728.0, + 1405.0, + 766.0, + 292.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 761.0, + 1405.0, + 761.0, + 1405.0, + 795.0, + 293.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 791.0, + 1404.0, + 791.0, + 1404.0, + 826.0, + 294.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 821.0, + 964.0, + 821.0, + 964.0, + 855.0, + 293.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1291.0, + 1405.0, + 1291.0, + 1405.0, + 1329.0, + 295.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1325.0, + 1404.0, + 1325.0, + 1404.0, + 1358.0, + 295.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1354.0, + 1405.0, + 1354.0, + 1405.0, + 1391.0, + 295.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1385.0, + 1405.0, + 1385.0, + 1405.0, + 1422.0, + 294.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1415.0, + 297.0, + 1415.0, + 297.0, + 1452.0, + 294.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 1415.0, + 902.0, + 1415.0, + 902.0, + 1452.0, + 314.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 217.0, + 649.0, + 217.0, + 649.0, + 269.0, + 289.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 217.0, + 1409.0, + 217.0, + 1409.0, + 269.0, + 1049.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 262.0, + 528.0, + 262.0, + 528.0, + 296.0, + 295.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 262.0, + 1403.0, + 262.0, + 1403.0, + 296.0, + 549.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 293.0, + 1010.0, + 293.0, + 1010.0, + 327.0, + 295.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 293.0, + 1376.0, + 293.0, + 1376.0, + 327.0, + 1045.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 293.0, + 1404.0, + 293.0, + 1404.0, + 327.0, + 1395.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1655.0, + 1405.0, + 1655.0, + 1405.0, + 1693.0, + 296.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1688.0, + 1405.0, + 1688.0, + 1405.0, + 1722.0, + 296.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1717.0, + 1340.0, + 1717.0, + 1340.0, + 1753.0, + 295.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1547.0, + 1403.0, + 1547.0, + 1403.0, + 1581.0, + 298.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1578.0, + 1403.0, + 1578.0, + 1403.0, + 1614.0, + 296.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1608.0, + 1072.0, + 1608.0, + 1072.0, + 1644.0, + 296.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 336.0, + 1402.0, + 336.0, + 1402.0, + 372.0, + 295.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 365.0, + 1404.0, + 365.0, + 1404.0, + 405.0, + 294.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 400.0, + 447.0, + 400.0, + 447.0, + 434.0, + 296.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 400.0, + 1190.0, + 400.0, + 1190.0, + 434.0, + 467.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 1405.0, + 1851.0, + 1405.0, + 1883.0, + 295.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1881.0, + 1404.0, + 1881.0, + 1404.0, + 1913.0, + 294.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1913.0, + 1403.0, + 1913.0, + 1403.0, + 1946.0, + 297.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 819.0, + 1942.0, + 819.0, + 1974.0, + 293.0, + 1974.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 836, + 1404, + 836, + 1404, + 1082, + 297, + 1082 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1404, + 229, + 1404, + 505, + 298, + 505 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1619, + 1403, + 1619, + 1403, + 1834, + 298, + 1834 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1850, + 1403, + 1850, + 1403, + 2034, + 299, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1343, + 1403, + 1343, + 1403, + 1497, + 299, + 1497 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1136, + 1403, + 1136, + 1403, + 1289, + 299, + 1289 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 629, + 1403, + 629, + 1403, + 782, + 298, + 782 + ], + "score": 0.976 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 300, + 1555, + 543, + 1555, + 543, + 1585, + 300, + 1585 + ], + "score": 0.887 + }, + { + "category_id": 0, + "poly": [ + 298, + 564, + 935, + 564, + 935, + 594, + 298, + 594 + ], + "score": 0.818 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.751 + }, + { + "category_id": 13, + "poly": [ + 937, + 1680, + 998, + 1680, + 998, + 1710, + 937, + 1710 + ], + "score": 0.9, + "latex": "1 0 ^ { - 7 }" + }, + { + "category_id": 13, + "poly": [ + 400, + 1228, + 471, + 1228, + 471, + 1256, + 400, + 1256 + ], + "score": 0.89, + "latex": "k = 4" + }, + { + "category_id": 13, + "poly": [ + 1230, + 322, + 1354, + 322, + 1354, + 351, + 1230, + 351 + ], + "score": 0.89, + "latex": "B = 2 0 4 8" + }, + { + "category_id": 13, + "poly": [ + 570, + 1882, + 682, + 1882, + 682, + 1910, + 570, + 1910 + ], + "score": 0.89, + "latex": "M = 0 . 6" + }, + { + "category_id": 13, + "poly": [ + 607, + 1198, + 723, + 1198, + 723, + 1226, + 607, + 1226 + ], + "score": 0.89, + "latex": "d = 1 0 2 4" + }, + { + "category_id": 13, + "poly": [ + 879, + 1912, + 1011, + 1912, + 1011, + 1941, + 879, + 1941 + ], + "score": 0.89, + "latex": "M \\ : = \\ : 0 . 7 5" + }, + { + "category_id": 13, + "poly": [ + 1007, + 1852, + 1090, + 1852, + 1090, + 1879, + 1007, + 1879 + ], + "score": 0.89, + "latex": "C = 3" + }, + { + "category_id": 13, + "poly": [ + 1317, + 1881, + 1402, + 1881, + 1402, + 1910, + 1317, + 1910 + ], + "score": 0.88, + "latex": "C = 4" + }, + { + "category_id": 13, + "poly": [ + 297, + 1802, + 357, + 1802, + 357, + 1833, + 297, + 1833 + ], + "score": 0.88, + "latex": "1 0 ^ { - 5 }" + }, + { + "category_id": 13, + "poly": [ + 774, + 1713, + 800, + 1713, + 800, + 1741, + 774, + 1741 + ], + "score": 0.78, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 692, + 384, + 718, + 384, + 718, + 410, + 692, + 410 + ], + "score": 0.78, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 902, + 1259, + 993, + 1259, + 993, + 1288, + 902, + 1288 + ], + "score": 0.78, + "latex": "\\scriptstyle \\mathrm { d } = 1 0 2 4" + }, + { + "category_id": 13, + "poly": [ + 475, + 293, + 501, + 293, + 501, + 319, + 475, + 319 + ], + "score": 0.77, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 569, + 1775, + 602, + 1775, + 602, + 1801, + 569, + 1801 + ], + "score": 0.76, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 1065, + 231, + 1092, + 231, + 1092, + 259, + 1065, + 259 + ], + "score": 0.74, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1268, + 1228, + 1350, + 1228, + 1350, + 1257, + 1268, + 1257 + ], + "score": 0.73, + "latex": "\\mathrm { \\Delta } \\mathrm { Q } \\mathrm { = } 2 5 6 )" + }, + { + "category_id": 13, + "poly": [ + 1306, + 1260, + 1371, + 1260, + 1371, + 1289, + 1306, + 1289 + ], + "score": 0.72, + "latex": "\\scriptstyle \\mathrm { { ( d = 6 4 ) } }" + }, + { + "category_id": 13, + "poly": [ + 374, + 1260, + 438, + 1260, + 438, + 1289, + 374, + 1289 + ], + "score": 0.71, + "latex": "\\scriptstyle \\mathrm { { ( d = 6 4 ) } }" + }, + { + "category_id": 13, + "poly": [ + 781, + 323, + 807, + 323, + 807, + 349, + 781, + 349 + ], + "score": 0.7, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1087, + 1260, + 1167, + 1260, + 1167, + 1289, + 1087, + 1289 + ], + "score": 0.7, + "latex": "\\mathrm { \\Delta } \\mathrm { Q } { = } 2 5 6 \\mathrm { \\Delta }" + }, + { + "category_id": 13, + "poly": [ + 1094, + 1228, + 1189, + 1228, + 1189, + 1257, + 1094, + 1257 + ], + "score": 0.57, + "latex": "\\scriptstyle ( \\mathrm { d } = 1 0 2 4 )" + }, + { + "category_id": 13, + "poly": [ + 431, + 929, + 505, + 929, + 505, + 958, + 431, + 958 + ], + "score": 0.39, + "latex": "2 \\mathrm { x } 2 5 6" + }, + { + "category_id": 13, + "poly": [ + 1014, + 929, + 1088, + 929, + 1088, + 958, + 1014, + 958 + ], + "score": 0.26, + "latex": "3 \\mathrm { x } 1 2 8" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1550.0, + 548.0, + 1550.0, + 548.0, + 1593.0, + 293.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 563.0, + 940.0, + 563.0, + 940.0, + 597.0, + 295.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 834.0, + 1406.0, + 834.0, + 1406.0, + 874.0, + 294.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 866.0, + 1407.0, + 866.0, + 1407.0, + 904.0, + 292.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 898.0, + 1406.0, + 898.0, + 1406.0, + 931.0, + 295.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 924.0, + 430.0, + 924.0, + 430.0, + 965.0, + 291.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 924.0, + 1013.0, + 924.0, + 1013.0, + 965.0, + 506.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 924.0, + 1406.0, + 924.0, + 1406.0, + 965.0, + 1089.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 960.0, + 1405.0, + 960.0, + 1405.0, + 990.0, + 296.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 989.0, + 1406.0, + 989.0, + 1406.0, + 1021.0, + 292.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1019.0, + 1405.0, + 1019.0, + 1405.0, + 1053.0, + 294.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1049.0, + 827.0, + 1049.0, + 827.0, + 1085.0, + 295.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1064.0, + 229.0, + 1064.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 264.0, + 1093.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 260.0, + 1405.0, + 260.0, + 1405.0, + 295.0, + 292.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 293.0, + 474.0, + 293.0, + 474.0, + 325.0, + 296.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 293.0, + 1405.0, + 293.0, + 1405.0, + 325.0, + 502.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 320.0, + 780.0, + 320.0, + 780.0, + 354.0, + 292.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 320.0, + 1229.0, + 320.0, + 1229.0, + 354.0, + 808.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 320.0, + 1406.0, + 320.0, + 1406.0, + 354.0, + 1355.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 353.0, + 1406.0, + 353.0, + 1406.0, + 385.0, + 294.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 383.0, + 691.0, + 383.0, + 691.0, + 415.0, + 296.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 383.0, + 1405.0, + 383.0, + 1405.0, + 415.0, + 719.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 414.0, + 1405.0, + 414.0, + 1405.0, + 447.0, + 296.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 443.0, + 1406.0, + 443.0, + 1406.0, + 479.0, + 293.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 474.0, + 602.0, + 474.0, + 602.0, + 507.0, + 294.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1617.0, + 1405.0, + 1617.0, + 1405.0, + 1655.0, + 293.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1649.0, + 1405.0, + 1649.0, + 1405.0, + 1686.0, + 293.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1679.0, + 936.0, + 1679.0, + 936.0, + 1717.0, + 292.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 1679.0, + 1405.0, + 1679.0, + 1405.0, + 1717.0, + 999.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1713.0, + 773.0, + 1713.0, + 773.0, + 1747.0, + 293.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 1713.0, + 1404.0, + 1713.0, + 1404.0, + 1747.0, + 801.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1742.0, + 1405.0, + 1742.0, + 1405.0, + 1777.0, + 293.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1772.0, + 568.0, + 1772.0, + 568.0, + 1809.0, + 293.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 1772.0, + 1405.0, + 1772.0, + 1405.0, + 1809.0, + 603.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1802.0, + 622.0, + 1802.0, + 622.0, + 1836.0, + 358.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1849.0, + 1006.0, + 1849.0, + 1006.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1849.0, + 1405.0, + 1849.0, + 1405.0, + 1885.0, + 1091.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1879.0, + 569.0, + 1879.0, + 569.0, + 1915.0, + 294.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 1879.0, + 1316.0, + 1879.0, + 1316.0, + 1915.0, + 683.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1879.0, + 1407.0, + 1879.0, + 1407.0, + 1915.0, + 1403.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 878.0, + 1910.0, + 878.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1910.0, + 1404.0, + 1910.0, + 1404.0, + 1946.0, + 1012.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1975.0, + 295.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1974.0, + 1404.0, + 1974.0, + 1404.0, + 2005.0, + 295.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 2004.0, + 1101.0, + 2004.0, + 1101.0, + 2036.0, + 297.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1343.0, + 1407.0, + 1343.0, + 1407.0, + 1380.0, + 294.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1375.0, + 1404.0, + 1375.0, + 1404.0, + 1408.0, + 295.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1404.0, + 1407.0, + 1404.0, + 1407.0, + 1440.0, + 290.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1435.0, + 1407.0, + 1435.0, + 1407.0, + 1472.0, + 295.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1465.0, + 1351.0, + 1465.0, + 1351.0, + 1502.0, + 294.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1133.0, + 1405.0, + 1133.0, + 1405.0, + 1173.0, + 294.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1164.0, + 1407.0, + 1164.0, + 1407.0, + 1203.0, + 293.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1193.0, + 606.0, + 1193.0, + 606.0, + 1236.0, + 290.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 1193.0, + 1409.0, + 1193.0, + 1409.0, + 1236.0, + 724.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1230.0, + 399.0, + 1230.0, + 399.0, + 1259.0, + 295.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 1230.0, + 1093.0, + 1230.0, + 1093.0, + 1259.0, + 472.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 1230.0, + 1267.0, + 1230.0, + 1267.0, + 1259.0, + 1190.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 1230.0, + 1403.0, + 1230.0, + 1403.0, + 1259.0, + 1351.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1259.0, + 373.0, + 1259.0, + 373.0, + 1292.0, + 295.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 1259.0, + 901.0, + 1259.0, + 901.0, + 1292.0, + 439.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 1259.0, + 1086.0, + 1259.0, + 1086.0, + 1292.0, + 994.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 1259.0, + 1305.0, + 1259.0, + 1305.0, + 1292.0, + 1168.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1259.0, + 1387.0, + 1259.0, + 1387.0, + 1292.0, + 1372.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 626.0, + 1405.0, + 626.0, + 1405.0, + 665.0, + 293.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 658.0, + 1404.0, + 658.0, + 1404.0, + 694.0, + 294.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 690.0, + 1405.0, + 690.0, + 1405.0, + 727.0, + 294.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 720.0, + 1404.0, + 720.0, + 1404.0, + 757.0, + 294.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 749.0, + 956.0, + 749.0, + 956.0, + 786.0, + 294.0, + 786.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 344, + 224, + 1354, + 224, + 1354, + 606, + 344, + 606 + ], + "score": 0.985, + "html": "
TestTrain Time (hours)Parameters
Dauphin et al. (2017)31.9428M
J6zefowicz et al. (2016)30.01,040M
Shazeer et al. (2017)28.0-4,371Mt
Char-CNN25.8879366M
Adaptive inputs25.2255331M
Adaptive inputs (large)23.9172465M
Adaptive inputs (very large)23.021451026M
10 LSTMs + SNM10-SKIP (Shazeer et al.,2016)23.71=
" + }, + { + "category_id": 5, + "poly": [ + 507, + 775, + 1191, + 775, + 1191, + 1051, + 507, + 1051 + ], + "score": 0.98, + "html": "
TestTrain Time (hours)Parameters
Grave et al. (2016)40.8
Dauphin et al. (2017)37.2229M
Merity et al. (2018)33.0=151M
Rae et al. (2018)29.2-
Adaptive inputs18.767247M
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1460, + 1403, + 1460, + 1403, + 1645, + 298, + 1645 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1290, + 1403, + 1290, + 1403, + 1445, + 298, + 1445 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1659, + 1403, + 1659, + 1403, + 1813, + 299, + 1813 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1402, + 1911, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.975 + }, + { + "category_id": 0, + "poly": [ + 299, + 1162, + 764, + 1162, + 764, + 1197, + 299, + 1197 + ], + "score": 0.913 + }, + { + "category_id": 0, + "poly": [ + 299, + 1233, + 547, + 1233, + 547, + 1265, + 299, + 1265 + ], + "score": 0.91 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.899 + }, + { + "category_id": 6, + "poly": [ + 298, + 629, + 1403, + 629, + 1403, + 743, + 298, + 743 + ], + "score": 0.882 + }, + { + "category_id": 0, + "poly": [ + 319, + 1855, + 1096, + 1855, + 1096, + 1884, + 319, + 1884 + ], + "score": 0.818 + }, + { + "category_id": 7, + "poly": [ + 331, + 1071, + 1360, + 1071, + 1360, + 1105, + 331, + 1105 + ], + "score": 0.718 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.699 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.096 + }, + { + "category_id": 13, + "poly": [ + 548, + 1384, + 685, + 1384, + 685, + 1417, + 548, + 1417 + ], + "score": 0.9, + "latex": "e _ { f f } = 8 1 9 2" + }, + { + "category_id": 13, + "poly": [ + 716, + 1353, + 859, + 1353, + 859, + 1386, + 716, + 1386 + ], + "score": 0.87, + "latex": "e _ { f f } = 6 1 4 4" + }, + { + "category_id": 13, + "poly": [ + 979, + 1614, + 1034, + 1614, + 1034, + 1643, + 979, + 1643 + ], + "score": 0.86, + "latex": "40 \\%" + }, + { + "category_id": 13, + "poly": [ + 735, + 1386, + 846, + 1386, + 846, + 1413, + 735, + 1413 + ], + "score": 0.85, + "latex": "e = 1 5 3 6" + }, + { + "category_id": 13, + "poly": [ + 1294, + 1323, + 1393, + 1323, + 1393, + 1352, + 1294, + 1352 + ], + "score": 0.85, + "latex": "N = 2 0" + }, + { + "category_id": 13, + "poly": [ + 385, + 1384, + 481, + 1384, + 481, + 1413, + 385, + 1413 + ], + "score": 0.84, + "latex": "N = 2 4" + }, + { + "category_id": 13, + "poly": [ + 1269, + 1384, + 1320, + 1384, + 1320, + 1416, + 1269, + 1416 + ], + "score": 0.82, + "latex": "\\ S 4 . 4" + }, + { + "category_id": 13, + "poly": [ + 1285, + 1782, + 1335, + 1782, + 1335, + 1814, + 1285, + 1814 + ], + "score": 0.81, + "latex": "\\ S 5 . 3" + }, + { + "category_id": 13, + "poly": [ + 1125, + 1583, + 1223, + 1583, + 1223, + 1612, + 1125, + 1612 + ], + "score": 0.38, + "latex": "0 . 6 ~ \\mathrm { P P L }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1159.0, + 768.0, + 1159.0, + 768.0, + 1203.0, + 292.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1233.0, + 551.0, + 1233.0, + 551.0, + 1269.0, + 294.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 624.0, + 1407.0, + 624.0, + 1407.0, + 669.0, + 292.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 651.0, + 1407.0, + 651.0, + 1407.0, + 694.0, + 292.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 680.0, + 411.0, + 680.0, + 411.0, + 720.0, + 294.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 709.0, + 789.0, + 709.0, + 789.0, + 749.0, + 295.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 1851.0, + 1102.0, + 1851.0, + 1102.0, + 1889.0, + 313.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 1071.0, + 1363.0, + 1071.0, + 1363.0, + 1107.0, + 336.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2119.0, + 838.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1460.0, + 1404.0, + 1460.0, + 1404.0, + 1495.0, + 296.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1493.0, + 1405.0, + 1493.0, + 1405.0, + 1525.0, + 296.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1523.0, + 1403.0, + 1523.0, + 1403.0, + 1555.0, + 296.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1550.0, + 1407.0, + 1550.0, + 1407.0, + 1588.0, + 293.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1582.0, + 1124.0, + 1582.0, + 1124.0, + 1617.0, + 294.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 1582.0, + 1405.0, + 1582.0, + 1405.0, + 1617.0, + 1224.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1612.0, + 978.0, + 1612.0, + 978.0, + 1648.0, + 294.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 1612.0, + 1119.0, + 1612.0, + 1119.0, + 1648.0, + 1035.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1289.0, + 1404.0, + 1289.0, + 1404.0, + 1325.0, + 294.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1322.0, + 1293.0, + 1322.0, + 1293.0, + 1356.0, + 296.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1322.0, + 1404.0, + 1322.0, + 1404.0, + 1356.0, + 1394.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1350.0, + 715.0, + 1350.0, + 715.0, + 1391.0, + 293.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1350.0, + 1408.0, + 1350.0, + 1408.0, + 1391.0, + 860.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1382.0, + 384.0, + 1382.0, + 384.0, + 1419.0, + 294.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1382.0, + 547.0, + 1382.0, + 547.0, + 1419.0, + 482.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1382.0, + 734.0, + 1382.0, + 734.0, + 1419.0, + 686.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 1382.0, + 1268.0, + 1382.0, + 1268.0, + 1419.0, + 847.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1321.0, + 1382.0, + 1405.0, + 1382.0, + 1405.0, + 1419.0, + 1321.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1410.0, + 863.0, + 1410.0, + 863.0, + 1451.0, + 293.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1661.0, + 1404.0, + 1661.0, + 1404.0, + 1694.0, + 297.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1690.0, + 1403.0, + 1690.0, + 1403.0, + 1723.0, + 297.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1721.0, + 1405.0, + 1721.0, + 1405.0, + 1755.0, + 297.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1747.0, + 1407.0, + 1747.0, + 1407.0, + 1790.0, + 293.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1782.0, + 1284.0, + 1782.0, + 1284.0, + 1814.0, + 294.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1336.0, + 1782.0, + 1344.0, + 1782.0, + 1344.0, + 1814.0, + 1336.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1911.0, + 1407.0, + 1911.0, + 1407.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1973.0, + 1407.0, + 1973.0, + 1407.0, + 2006.0, + 293.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2004.0, + 1403.0, + 2004.0, + 1403.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1481, + 1404, + 1481, + 1404, + 1758, + 298, + 1758 + ], + "score": 0.982 + }, + { + "category_id": 5, + "poly": [ + 317, + 223, + 1382, + 223, + 1382, + 575, + 317, + 575 + ], + "score": 0.982, + "html": "
InputOutputValidTestTrain Time (hours)Params
SMEmbeddingSoftmax23.8724.9257*476.8M
BPEBPE EmbeddingBPE Softmax23.1324.2530270M
BPE-TBPE EmbeddingBPE Softmax (tied)22.4623.4530235.7M
SM-TEmbeddingSoftmax (tied)22.6323.3856*339.7M
ASMEmbeddingAdaptive21.2322.1835263.1M
CNNChar-CNNAdaptive20.8621.7970266.3M
ADPAdaptiveAdaptive20.9521.7434291.3M
ADP-TAdaptiveAdaptive (tied)19.7920.5130246.9M
" + }, + { + "category_id": 5, + "poly": [ + 308, + 845, + 1392, + 845, + 1392, + 1137, + 308, + 1137 + ], + "score": 0.981, + "html": "
InputOutputValidTestTrain time (hours)Params
BPE-TBPE EmbeddingBPE Softmax (shared)27.4427.5134234.7M
BPEBPE EmbeddingBPE Softmax27.0227.1335267.8M
ASMEmbeddingAdaptive26.9727.0662532.8M
CNNChar-CNNAdaptive26.1326.2592365.8M
ADPAdaptiveAdaptive26.3826.4965458.4M
ADP-TAdaptiveAdaptive (shared)25.5125.5843330.8M
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1773, + 1403, + 1773, + 1403, + 1957, + 298, + 1957 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1312, + 1403, + 1312, + 1403, + 1467, + 298, + 1467 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1972, + 1399, + 1972, + 1399, + 2035, + 297, + 2035 + ], + "score": 0.94 + }, + { + "category_id": 6, + "poly": [ + 297, + 597, + 1405, + 597, + 1405, + 752, + 297, + 752 + ], + "score": 0.926 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.771 + }, + { + "category_id": 7, + "poly": [ + 312, + 1159, + 1382, + 1159, + 1382, + 1192, + 312, + 1192 + ], + "score": 0.68 + }, + { + "category_id": 6, + "poly": [ + 312, + 1159, + 1382, + 1159, + 1382, + 1192, + 312, + 1192 + ], + "score": 0.217 + }, + { + "category_id": 13, + "poly": [ + 753, + 1804, + 808, + 1804, + 808, + 1833, + 753, + 1833 + ], + "score": 0.87, + "latex": "34 \\%" + }, + { + "category_id": 13, + "poly": [ + 680, + 1436, + 730, + 1436, + 730, + 1468, + 680, + 1468 + ], + "score": 0.74, + "latex": "\\ S 4 . 1" + }, + { + "category_id": 13, + "poly": [ + 939, + 630, + 974, + 630, + 974, + 660, + 939, + 660 + ], + "score": 0.62, + "latex": "( ^ { * } )" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 599.0, + 1402.0, + 599.0, + 1402.0, + 632.0, + 296.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 627.0, + 938.0, + 627.0, + 938.0, + 664.0, + 295.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 627.0, + 1403.0, + 627.0, + 1403.0, + 664.0, + 975.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 659.0, + 1405.0, + 659.0, + 1405.0, + 696.0, + 295.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 686.0, + 1405.0, + 686.0, + 1405.0, + 726.0, + 295.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 720.0, + 618.0, + 720.0, + 618.0, + 753.0, + 296.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 1156.0, + 1387.0, + 1156.0, + 1387.0, + 1195.0, + 312.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 1156.0, + 1387.0, + 1156.0, + 1387.0, + 1195.0, + 312.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1479.0, + 1405.0, + 1479.0, + 1405.0, + 1517.0, + 295.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1515.0, + 1405.0, + 1515.0, + 1405.0, + 1548.0, + 296.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1542.0, + 1406.0, + 1542.0, + 1406.0, + 1578.0, + 292.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1575.0, + 1405.0, + 1575.0, + 1405.0, + 1608.0, + 294.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1604.0, + 1405.0, + 1604.0, + 1405.0, + 1640.0, + 293.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1631.0, + 1405.0, + 1631.0, + 1405.0, + 1668.0, + 293.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1667.0, + 1405.0, + 1667.0, + 1405.0, + 1699.0, + 294.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1695.0, + 1406.0, + 1695.0, + 1406.0, + 1732.0, + 294.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1727.0, + 570.0, + 1727.0, + 570.0, + 1763.0, + 294.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1771.0, + 1405.0, + 1771.0, + 1405.0, + 1809.0, + 293.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1802.0, + 752.0, + 1802.0, + 752.0, + 1840.0, + 293.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 1802.0, + 1405.0, + 1802.0, + 1405.0, + 1840.0, + 809.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1831.0, + 1408.0, + 1831.0, + 1408.0, + 1873.0, + 291.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1865.0, + 1405.0, + 1865.0, + 1405.0, + 1901.0, + 293.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1894.0, + 1407.0, + 1894.0, + 1407.0, + 1930.0, + 293.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1926.0, + 953.0, + 1926.0, + 953.0, + 1958.0, + 294.0, + 1958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1312.0, + 1403.0, + 1312.0, + 1403.0, + 1346.0, + 296.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1346.0, + 1404.0, + 1346.0, + 1404.0, + 1379.0, + 296.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1374.0, + 1405.0, + 1374.0, + 1405.0, + 1409.0, + 293.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1404.0, + 1405.0, + 1404.0, + 1405.0, + 1441.0, + 293.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1436.0, + 679.0, + 1436.0, + 679.0, + 1469.0, + 294.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 1436.0, + 739.0, + 1436.0, + 739.0, + 1469.0, + 731.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1972.0, + 1403.0, + 1972.0, + 1403.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2001.0, + 952.0, + 2001.0, + 952.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1680, + 1403, + 1680, + 1403, + 1926, + 298, + 1926 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 302, + 218, + 1401, + 218, + 1401, + 744, + 302, + 744 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 301, + 863, + 1401, + 863, + 1401, + 1390, + 301, + 1390 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 299, + 1942, + 1402, + 1942, + 1402, + 2035, + 299, + 2035 + ], + "score": 0.967 + }, + { + "category_id": 4, + "poly": [ + 298, + 774, + 1401, + 774, + 1401, + 836, + 298, + 836 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 298, + 1604, + 1401, + 1604, + 1401, + 1666, + 298, + 1666 + ], + "score": 0.947 + }, + { + "category_id": 4, + "poly": [ + 296, + 1420, + 1401, + 1420, + 1401, + 1485, + 296, + 1485 + ], + "score": 0.941 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 815, + 76, + 815, + 104, + 298, + 104 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 298, + 1546, + 488, + 1546, + 488, + 1577, + 298, + 1577 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 842, + 2087, + 858, + 2087, + 858, + 2111, + 842, + 2111 + ], + "score": 0.652 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.286 + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 256.0, + 1177.0, + 256.0, + 1177.0, + 300.0, + 1128.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 256.0, + 1351.0, + 256.0, + 1351.0, + 298.0, + 1257.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 258.0, + 1230.0, + 258.0, + 1230.0, + 298.0, + 1170.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1227.0, + 262.0, + 1251.0, + 262.0, + 1251.0, + 293.0, + 1227.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 262.0, + 1274.0, + 262.0, + 1274.0, + 294.0, + 1241.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 271.0, + 457.0, + 271.0, + 457.0, + 552.0, + 442.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 274.0, + 417.0, + 274.0, + 417.0, + 313.0, + 371.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 282.0, + 498.0, + 282.0, + 498.0, + 662.0, + 482.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 291.0, + 1175.0, + 291.0, + 1175.0, + 329.0, + 1130.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 291.0, + 1356.0, + 291.0, + 1356.0, + 333.0, + 1257.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 293.0, + 1234.0, + 293.0, + 1234.0, + 331.0, + 1165.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 295.0, + 1249.0, + 295.0, + 1249.0, + 327.0, + 1231.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 298.0, + 1267.0, + 298.0, + 1267.0, + 327.0, + 1245.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 307.0, + 536.0, + 307.0, + 536.0, + 326.0, + 527.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 311.0, + 524.0, + 311.0, + 524.0, + 662.0, + 509.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 320.0, + 472.0, + 320.0, + 472.0, + 662.0, + 456.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 322.0, + 536.0, + 322.0, + 536.0, + 488.0, + 523.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 322.0, + 1178.0, + 322.0, + 1178.0, + 365.0, + 1130.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 327.0, + 1237.0, + 327.0, + 1237.0, + 365.0, + 1160.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 327.0, + 1350.0, + 327.0, + 1350.0, + 370.0, + 1264.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 331.0, + 1249.0, + 331.0, + 1249.0, + 361.0, + 1231.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 332.0, + 1267.0, + 332.0, + 1267.0, + 361.0, + 1245.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 350.0, + 413.0, + 350.0, + 413.0, + 387.0, + 386.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 355.0, + 552.0, + 355.0, + 552.0, + 378.0, + 539.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 364.0, + 593.0, + 364.0, + 593.0, + 426.0, + 577.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 357.0, + 1360.0, + 357.0, + 1360.0, + 404.0, + 1130.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 380.0, + 662.0, + 380.0, + 662.0, + 662.0, + 645.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 383.0, + 632.0, + 383.0, + 632.0, + 663.0, + 615.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 383.0, + 672.0, + 383.0, + 672.0, + 661.0, + 659.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 384.0, + 645.0, + 384.0, + 645.0, + 477.0, + 633.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 389.0, + 607.0, + 389.0, + 607.0, + 662.0, + 590.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 416.0, + 339.0, + 416.0, + 339.0, + 487.0, + 299.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 425.0, + 413.0, + 425.0, + 413.0, + 463.0, + 384.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 453.0, + 739.0, + 453.0, + 739.0, + 537.0, + 727.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 454.0, + 765.0, + 454.0, + 765.0, + 472.0, + 757.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 458.0, + 780.0, + 458.0, + 780.0, + 477.0, + 768.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 458.0, + 795.0, + 458.0, + 795.0, + 502.0, + 783.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 472.0, + 780.0, + 472.0, + 780.0, + 499.0, + 768.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 483.0, + 536.0, + 483.0, + 536.0, + 601.0, + 523.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 496.0, + 864.0, + 496.0, + 864.0, + 662.0, + 848.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 499.0, + 906.0, + 499.0, + 906.0, + 664.0, + 886.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 503.0, + 412.0, + 503.0, + 412.0, + 532.0, + 388.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 503.0, + 874.0, + 503.0, + 874.0, + 516.0, + 865.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 505.0, + 915.0, + 505.0, + 915.0, + 521.0, + 905.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 507.0, + 930.0, + 507.0, + 930.0, + 661.0, + 918.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 507.0, + 943.0, + 507.0, + 943.0, + 661.0, + 931.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 513.0, + 645.0, + 513.0, + 645.0, + 585.0, + 633.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 536.0, + 998.0, + 536.0, + 998.0, + 661.0, + 986.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 536.0, + 1010.0, + 536.0, + 1010.0, + 630.0, + 999.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 536.0, + 1041.0, + 536.0, + 1041.0, + 662.0, + 1025.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 537.0, + 767.0, + 537.0, + 767.0, + 661.0, + 754.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 540.0, + 1064.0, + 540.0, + 1064.0, + 650.0, + 1051.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 541.0, + 1081.0, + 541.0, + 1081.0, + 645.0, + 1065.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 546.0, + 740.0, + 546.0, + 740.0, + 624.0, + 728.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 547.0, + 456.0, + 547.0, + 456.0, + 619.0, + 443.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 556.0, + 780.0, + 556.0, + 780.0, + 582.0, + 768.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 574.0, + 413.0, + 574.0, + 413.0, + 608.0, + 386.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 597.0, + 534.0, + 597.0, + 534.0, + 609.0, + 526.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 582.0, + 644.0, + 582.0, + 644.0, + 595.0, + 635.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 584.0, + 1133.0, + 584.0, + 1133.0, + 655.0, + 1120.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 584.0, + 1146.0, + 584.0, + 1146.0, + 660.0, + 1134.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 581.0, + 1216.0, + 581.0, + 1216.0, + 664.0, + 1159.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 626.0, + 534.0, + 626.0, + 534.0, + 649.0, + 526.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 615.0, + 645.0, + 615.0, + 645.0, + 661.0, + 633.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 625.0, + 740.0, + 625.0, + 740.0, + 661.0, + 728.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 619.0, + 780.0, + 619.0, + 780.0, + 661.0, + 768.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 629.0, + 795.0, + 629.0, + 795.0, + 661.0, + 783.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 626.0, + 805.0, + 626.0, + 805.0, + 660.0, + 797.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 631.0, + 819.0, + 631.0, + 819.0, + 647.0, + 810.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 614.0, + 1270.0, + 614.0, + 1270.0, + 641.0, + 1257.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 614.0, + 1284.0, + 614.0, + 1284.0, + 661.0, + 1271.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 613.0, + 1311.0, + 613.0, + 1311.0, + 664.0, + 1295.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1321.0, + 614.0, + 1351.0, + 614.0, + 1351.0, + 663.0, + 1321.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 647.0, + 1064.0, + 647.0, + 1064.0, + 660.0, + 1054.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 641.0, + 1322.0, + 641.0, + 1322.0, + 655.0, + 1314.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 673.0, + 523.0, + 673.0, + 523.0, + 717.0, + 474.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 674.0, + 664.0, + 674.0, + 664.0, + 716.0, + 603.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 673.0, + 797.0, + 673.0, + 797.0, + 716.0, + 742.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 674.0, + 938.0, + 674.0, + 938.0, + 715.0, + 871.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 674.0, + 1079.0, + 674.0, + 1079.0, + 716.0, + 999.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 672.0, + 1205.0, + 672.0, + 1205.0, + 717.0, + 1145.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 673.0, + 1347.0, + 673.0, + 1347.0, + 717.0, + 1274.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 701.0, + 1022.0, + 701.0, + 1022.0, + 751.0, + 786.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 900.0, + 410.0, + 900.0, + 410.0, + 922.0, + 391.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 902.0, + 1013.0, + 902.0, + 1013.0, + 951.0, + 916.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 906.0, + 832.0, + 906.0, + 832.0, + 946.0, + 787.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 907.0, + 887.0, + 907.0, + 887.0, + 943.0, + 825.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 910.0, + 908.0, + 910.0, + 908.0, + 942.0, + 886.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 910.0, + 930.0, + 910.0, + 930.0, + 943.0, + 901.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 923.0, + 1270.0, + 923.0, + 1270.0, + 973.0, + 1257.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 938.0, + 1015.0, + 938.0, + 1015.0, + 983.0, + 913.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 939.0, + 1285.0, + 939.0, + 1285.0, + 1253.0, + 1270.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 940.0, + 832.0, + 940.0, + 832.0, + 977.0, + 790.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 940.0, + 893.0, + 940.0, + 893.0, + 983.0, + 822.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 943.0, + 906.0, + 943.0, + 906.0, + 975.0, + 889.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 945.0, + 926.0, + 945.0, + 926.0, + 976.0, + 903.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 945.0, + 1311.0, + 945.0, + 1311.0, + 1310.0, + 1296.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1324.0, + 962.0, + 1336.0, + 962.0, + 1336.0, + 1036.0, + 1324.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1336.0, + 965.0, + 1351.0, + 965.0, + 1351.0, + 1308.0, + 1336.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 969.0, + 341.0, + 969.0, + 341.0, + 1227.0, + 301.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 973.0, + 898.0, + 973.0, + 898.0, + 1015.0, + 817.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 973.0, + 1010.0, + 973.0, + 1010.0, + 1016.0, + 923.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 974.0, + 834.0, + 974.0, + 834.0, + 1012.0, + 788.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 978.0, + 907.0, + 978.0, + 907.0, + 1008.0, + 889.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 980.0, + 1363.0, + 980.0, + 1363.0, + 1008.0, + 1351.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 981.0, + 923.0, + 981.0, + 923.0, + 1009.0, + 905.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 981.0, + 1135.0, + 981.0, + 1135.0, + 1002.0, + 1122.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 991.0, + 1150.0, + 991.0, + 1150.0, + 1254.0, + 1135.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 994.0, + 1176.0, + 994.0, + 1176.0, + 1310.0, + 1160.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 1001.0, + 1018.0, + 1001.0, + 1018.0, + 1055.0, + 814.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1004.0, + 419.0, + 1004.0, + 419.0, + 1052.0, + 360.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 1006.0, + 836.0, + 1006.0, + 836.0, + 1046.0, + 788.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 1021.0, + 1200.0, + 1021.0, + 1200.0, + 1261.0, + 1189.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 1025.0, + 1216.0, + 1025.0, + 1216.0, + 1259.0, + 1200.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1322.0, + 1030.0, + 1337.0, + 1030.0, + 1337.0, + 1305.0, + 1322.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 1038.0, + 460.0, + 1038.0, + 460.0, + 1309.0, + 440.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1069.0, + 498.0, + 1069.0, + 498.0, + 1310.0, + 482.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 1073.0, + 512.0, + 1073.0, + 512.0, + 1131.0, + 496.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 1080.0, + 1270.0, + 1080.0, + 1270.0, + 1248.0, + 1257.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1081.0, + 538.0, + 1081.0, + 538.0, + 1144.0, + 522.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 1091.0, + 1363.0, + 1091.0, + 1363.0, + 1174.0, + 1351.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 1092.0, + 470.0, + 1092.0, + 470.0, + 1310.0, + 454.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1125.0, + 416.0, + 1125.0, + 416.0, + 1166.0, + 387.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1125.0, + 510.0, + 1125.0, + 510.0, + 1193.0, + 497.0, + 1193.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 1128.0, + 592.0, + 1128.0, + 592.0, + 1169.0, + 580.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 1136.0, + 526.0, + 1136.0, + 526.0, + 1308.0, + 510.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1136.0, + 633.0, + 1136.0, + 633.0, + 1169.0, + 620.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1144.0, + 671.0, + 1144.0, + 671.0, + 1152.0, + 661.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1150.0, + 607.0, + 1150.0, + 607.0, + 1253.0, + 591.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 1155.0, + 551.0, + 1155.0, + 551.0, + 1178.0, + 537.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 1155.0, + 646.0, + 1155.0, + 646.0, + 1173.0, + 632.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 1174.0, + 660.0, + 1174.0, + 660.0, + 1212.0, + 647.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 1191.0, + 741.0, + 1191.0, + 741.0, + 1276.0, + 729.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1191.0, + 767.0, + 1191.0, + 767.0, + 1206.0, + 757.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1239.0, + 419.0, + 1239.0, + 419.0, + 1287.0, + 358.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 1226.0, + 795.0, + 1226.0, + 795.0, + 1305.0, + 782.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1221.0, + 807.0, + 1221.0, + 807.0, + 1308.0, + 796.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 1222.0, + 821.0, + 1222.0, + 821.0, + 1242.0, + 809.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 1232.0, + 861.0, + 1232.0, + 861.0, + 1281.0, + 852.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 1235.0, + 879.0, + 1235.0, + 879.0, + 1288.0, + 862.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1233.0, + 930.0, + 1233.0, + 930.0, + 1311.0, + 887.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 1256.0, + 945.0, + 1256.0, + 945.0, + 1307.0, + 932.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 1243.0, + 997.0, + 1243.0, + 997.0, + 1286.0, + 988.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 1247.0, + 1041.0, + 1247.0, + 1041.0, + 1310.0, + 1025.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 1258.0, + 1212.0, + 1258.0, + 1212.0, + 1280.0, + 1203.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 1283.0, + 605.0, + 1283.0, + 605.0, + 1309.0, + 592.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1292.0, + 630.0, + 1292.0, + 630.0, + 1305.0, + 621.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 1293.0, + 659.0, + 1293.0, + 659.0, + 1304.0, + 649.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 1283.0, + 861.0, + 1283.0, + 861.0, + 1305.0, + 852.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 1283.0, + 997.0, + 1283.0, + 997.0, + 1305.0, + 988.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 1283.0, + 1010.0, + 1283.0, + 1010.0, + 1305.0, + 1001.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 1277.0, + 1215.0, + 1277.0, + 1215.0, + 1307.0, + 1202.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 1321.0, + 521.0, + 1321.0, + 521.0, + 1362.0, + 477.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 1321.0, + 665.0, + 1321.0, + 665.0, + 1362.0, + 603.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 1320.0, + 796.0, + 1320.0, + 796.0, + 1363.0, + 742.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1321.0, + 937.0, + 1321.0, + 937.0, + 1362.0, + 871.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 1321.0, + 1080.0, + 1321.0, + 1080.0, + 1363.0, + 998.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1321.0, + 1205.0, + 1321.0, + 1205.0, + 1362.0, + 1147.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 1320.0, + 1347.0, + 1320.0, + 1347.0, + 1363.0, + 1273.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 1347.0, + 1024.0, + 1347.0, + 1024.0, + 1399.0, + 784.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 773.0, + 1405.0, + 773.0, + 1405.0, + 809.0, + 294.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 806.0, + 433.0, + 806.0, + 433.0, + 839.0, + 294.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1416.0, + 1406.0, + 1416.0, + 1406.0, + 1459.0, + 292.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1451.0, + 640.0, + 1451.0, + 640.0, + 1487.0, + 296.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1543.0, + 492.0, + 1543.0, + 492.0, + 1582.0, + 293.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1682.0, + 1404.0, + 1682.0, + 1404.0, + 1715.0, + 294.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1712.0, + 1405.0, + 1712.0, + 1405.0, + 1744.0, + 293.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1743.0, + 1404.0, + 1743.0, + 1404.0, + 1777.0, + 296.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 1404.0, + 1773.0, + 1404.0, + 1807.0, + 294.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1804.0, + 1405.0, + 1804.0, + 1405.0, + 1838.0, + 296.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1832.0, + 1405.0, + 1832.0, + 1405.0, + 1873.0, + 292.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1864.0, + 1405.0, + 1864.0, + 1405.0, + 1898.0, + 293.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1893.0, + 1201.0, + 1893.0, + 1201.0, + 1928.0, + 294.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1979.0, + 294.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 1403.0, + 2003.0, + 1403.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1603.0, + 1403.0, + 1603.0, + 1403.0, + 1639.0, + 296.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1632.0, + 1371.0, + 1632.0, + 1371.0, + 1673.0, + 293.0, + 1673.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1453, + 1404, + 1453, + 1404, + 1638, + 298, + 1638 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1741, + 1402, + 1741, + 1402, + 1896, + 298, + 1896 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 300, + 1911, + 1402, + 1911, + 1402, + 2033, + 300, + 2033 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 301, + 217, + 1399, + 217, + 1399, + 744, + 301, + 744 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 1345, + 1403, + 1345, + 1403, + 1438, + 299, + 1438 + ], + "score": 0.97 + }, + { + "category_id": 5, + "poly": [ + 568, + 873, + 1130, + 873, + 1130, + 1165, + 568, + 1165 + ], + "score": 0.95, + "html": "
Train block sizeInference context sizeValidation perplexityTest perplexity
51219.7920.51
5120 48018.3519.03
2048018.9619.53
2048153618.2318.88
3072018.6319.34
3072256017.9718.70
" + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.896 + }, + { + "category_id": 4, + "poly": [ + 295, + 773, + 1403, + 773, + 1403, + 836, + 295, + 836 + ], + "score": 0.884 + }, + { + "category_id": 0, + "poly": [ + 298, + 1683, + 859, + 1683, + 859, + 1714, + 298, + 1714 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2112, + 841, + 2112 + ], + "score": 0.79 + }, + { + "category_id": 1, + "poly": [ + 299, + 1187, + 1401, + 1187, + 1401, + 1280, + 299, + 1280 + ], + "score": 0.708 + }, + { + "category_id": 6, + "poly": [ + 299, + 1187, + 1401, + 1187, + 1401, + 1280, + 299, + 1280 + ], + "score": 0.191 + }, + { + "category_id": 15, + "poly": [ + 371.0, + 233.0, + 416.0, + 233.0, + 416.0, + 272.0, + 371.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 256.0, + 483.0, + 256.0, + 483.0, + 665.0, + 463.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 269.0, + 543.0, + 269.0, + 543.0, + 298.0, + 526.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 286.0, + 498.0, + 286.0, + 498.0, + 430.0, + 482.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 288.0, + 1234.0, + 288.0, + 1234.0, + 327.0, + 1163.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 288.0, + 1357.0, + 288.0, + 1357.0, + 329.0, + 1257.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 289.0, + 527.0, + 289.0, + 527.0, + 616.0, + 511.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 291.0, + 1173.0, + 291.0, + 1173.0, + 324.0, + 1133.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1228.0, + 292.0, + 1252.0, + 292.0, + 1252.0, + 324.0, + 1228.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 293.0, + 514.0, + 293.0, + 514.0, + 612.0, + 494.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 296.0, + 1266.0, + 296.0, + 1266.0, + 321.0, + 1246.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 319.0, + 1349.0, + 319.0, + 1349.0, + 364.0, + 1136.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 348.0, + 1361.0, + 348.0, + 1361.0, + 400.0, + 1216.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 353.0, + 1182.0, + 353.0, + 1182.0, + 395.0, + 1128.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 354.0, + 1241.0, + 354.0, + 1241.0, + 396.0, + 1161.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 369.0, + 416.0, + 369.0, + 416.0, + 409.0, + 371.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 370.0, + 617.0, + 370.0, + 617.0, + 664.0, + 600.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 375.0, + 633.0, + 375.0, + 633.0, + 664.0, + 613.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 376.0, + 647.0, + 376.0, + 647.0, + 665.0, + 630.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 376.0, + 664.0, + 376.0, + 664.0, + 667.0, + 644.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 384.0, + 678.0, + 384.0, + 678.0, + 662.0, + 662.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 416.0, + 340.0, + 416.0, + 340.0, + 487.0, + 300.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 443.0, + 755.0, + 443.0, + 755.0, + 667.0, + 735.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 456.0, + 769.0, + 456.0, + 769.0, + 662.0, + 754.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 456.0, + 800.0, + 456.0, + 800.0, + 667.0, + 780.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 457.0, + 784.0, + 457.0, + 784.0, + 665.0, + 768.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 495.0, + 889.0, + 495.0, + 889.0, + 665.0, + 872.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 500.0, + 901.0, + 500.0, + 901.0, + 662.0, + 889.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 500.0, + 921.0, + 500.0, + 921.0, + 668.0, + 901.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 500.0, + 932.0, + 500.0, + 932.0, + 664.0, + 916.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 511.0, + 413.0, + 511.0, + 413.0, + 545.0, + 386.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 533.0, + 1023.0, + 533.0, + 1023.0, + 663.0, + 1007.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 537.0, + 1067.0, + 537.0, + 1067.0, + 663.0, + 1055.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 539.0, + 1082.0, + 539.0, + 1082.0, + 625.0, + 1070.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1026.0, + 541.0, + 1038.0, + 541.0, + 1038.0, + 661.0, + 1026.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 541.0, + 1053.0, + 541.0, + 1053.0, + 661.0, + 1041.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 565.0, + 1159.0, + 565.0, + 1159.0, + 665.0, + 1143.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 566.0, + 1204.0, + 566.0, + 1204.0, + 665.0, + 1188.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 569.0, + 1174.0, + 569.0, + 1174.0, + 662.0, + 1158.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 569.0, + 1189.0, + 569.0, + 1189.0, + 667.0, + 1173.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 570.0, + 1218.0, + 570.0, + 1218.0, + 629.0, + 1206.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 609.0, + 527.0, + 609.0, + 527.0, + 664.0, + 496.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 648.0, + 1082.0, + 648.0, + 1082.0, + 659.0, + 1072.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 627.0, + 1216.0, + 627.0, + 1216.0, + 659.0, + 1207.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 603.0, + 1354.0, + 603.0, + 1354.0, + 674.0, + 1262.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 653.0, + 411.0, + 653.0, + 411.0, + 676.0, + 390.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 674.0, + 519.0, + 674.0, + 519.0, + 714.0, + 477.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 674.0, + 663.0, + 674.0, + 663.0, + 714.0, + 604.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 674.0, + 796.0, + 674.0, + 796.0, + 716.0, + 743.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 674.0, + 936.0, + 674.0, + 936.0, + 714.0, + 871.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 674.0, + 1078.0, + 674.0, + 1078.0, + 716.0, + 999.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 674.0, + 1203.0, + 674.0, + 1203.0, + 716.0, + 1147.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 673.0, + 1347.0, + 673.0, + 1347.0, + 716.0, + 1274.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 702.0, + 1021.0, + 702.0, + 1021.0, + 750.0, + 786.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 770.0, + 1406.0, + 770.0, + 1406.0, + 811.0, + 292.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 806.0, + 516.0, + 806.0, + 516.0, + 838.0, + 295.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1681.0, + 862.0, + 1681.0, + 862.0, + 1718.0, + 294.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1183.0, + 1402.0, + 1183.0, + 1402.0, + 1225.0, + 293.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1217.0, + 1403.0, + 1217.0, + 1403.0, + 1254.0, + 295.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1249.0, + 1118.0, + 1249.0, + 1118.0, + 1283.0, + 295.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1455.0, + 1405.0, + 1455.0, + 1405.0, + 1487.0, + 296.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1487.0, + 1405.0, + 1487.0, + 1405.0, + 1519.0, + 296.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1511.0, + 1406.0, + 1511.0, + 1406.0, + 1554.0, + 292.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1545.0, + 1405.0, + 1545.0, + 1405.0, + 1580.0, + 295.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1576.0, + 1406.0, + 1576.0, + 1406.0, + 1610.0, + 293.0, + 1610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1607.0, + 794.0, + 1607.0, + 794.0, + 1642.0, + 293.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1739.0, + 1404.0, + 1739.0, + 1404.0, + 1779.0, + 294.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1772.0, + 1406.0, + 1772.0, + 1406.0, + 1809.0, + 294.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1803.0, + 1406.0, + 1803.0, + 1406.0, + 1840.0, + 292.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1834.0, + 1406.0, + 1834.0, + 1406.0, + 1870.0, + 291.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1864.0, + 1289.0, + 1864.0, + 1289.0, + 1901.0, + 292.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1912.0, + 1404.0, + 1912.0, + 1404.0, + 1944.0, + 296.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1943.0, + 1406.0, + 1943.0, + 1406.0, + 1975.0, + 296.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 1407.0, + 1972.0, + 1407.0, + 2007.0, + 293.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2004.0, + 758.0, + 2004.0, + 758.0, + 2033.0, + 296.0, + 2033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1344.0, + 1407.0, + 1344.0, + 1407.0, + 1382.0, + 294.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1379.0, + 1403.0, + 1379.0, + 1403.0, + 1410.0, + 297.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1410.0, + 678.0, + 1410.0, + 678.0, + 1440.0, + 295.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1183.0, + 1402.0, + 1183.0, + 1402.0, + 1225.0, + 293.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1217.0, + 1403.0, + 1217.0, + 1403.0, + 1254.0, + 295.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1249.0, + 1118.0, + 1249.0, + 1118.0, + 1283.0, + 295.0, + 1283.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 604, + 1403, + 604, + 1403, + 759, + 299, + 759 + ], + "score": 0.976 + }, + { + "category_id": 5, + "poly": [ + 614, + 224, + 1087, + 224, + 1087, + 424, + 614, + 424 + ], + "score": 0.974, + "html": "
Tail dropoutValidation perplexity
Softmax (SM)N/A23.87
Adaptive (ADP)0.024.74
Adaptive (ADP)0.221.23
" + }, + { + "category_id": 1, + "poly": [ + 299, + 774, + 1403, + 774, + 1403, + 897, + 299, + 897 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 985, + 959, + 985, + 959, + 1018, + 299, + 1018 + ], + "score": 0.916 + }, + { + "category_id": 0, + "poly": [ + 299, + 535, + 543, + 535, + 543, + 570, + 299, + 570 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 301, + 935, + 558, + 935, + 558, + 964, + 301, + 964 + ], + "score": 0.836 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.758 + }, + { + "category_id": 1, + "poly": [ + 296, + 1117, + 1402, + 1117, + 1402, + 1180, + 296, + 1180 + ], + "score": 0.674 + }, + { + "category_id": 0, + "poly": [ + 300, + 1066, + 488, + 1066, + 488, + 1099, + 300, + 1099 + ], + "score": 0.658 + }, + { + "category_id": 1, + "poly": [ + 296, + 1201, + 1400, + 1201, + 1400, + 1296, + 296, + 1296 + ], + "score": 0.598 + }, + { + "category_id": 1, + "poly": [ + 291, + 1744, + 1402, + 1744, + 1402, + 1810, + 291, + 1810 + ], + "score": 0.578 + }, + { + "category_id": 1, + "poly": [ + 299, + 1404, + 1396, + 1404, + 1396, + 1468, + 299, + 1468 + ], + "score": 0.552 + }, + { + "category_id": 1, + "poly": [ + 290, + 1830, + 1312, + 1830, + 1312, + 1866, + 290, + 1866 + ], + "score": 0.549 + }, + { + "category_id": 1, + "poly": [ + 298, + 1318, + 1400, + 1318, + 1400, + 1382, + 298, + 1382 + ], + "score": 0.547 + }, + { + "category_id": 1, + "poly": [ + 293, + 1659, + 1402, + 1659, + 1402, + 1724, + 293, + 1724 + ], + "score": 0.539 + }, + { + "category_id": 1, + "poly": [ + 291, + 1488, + 1359, + 1488, + 1359, + 1524, + 291, + 1524 + ], + "score": 0.525 + }, + { + "category_id": 1, + "poly": [ + 300, + 1543, + 1397, + 1543, + 1397, + 1638, + 300, + 1638 + ], + "score": 0.524 + }, + { + "category_id": 1, + "poly": [ + 297, + 1971, + 1397, + 1971, + 1397, + 2035, + 297, + 2035 + ], + "score": 0.507 + }, + { + "category_id": 1, + "poly": [ + 292, + 1886, + 1400, + 1886, + 1400, + 1950, + 292, + 1950 + ], + "score": 0.498 + }, + { + "category_id": 6, + "poly": [ + 352, + 445, + 1343, + 445, + 1343, + 478, + 352, + 478 + ], + "score": 0.395 + }, + { + "category_id": 1, + "poly": [ + 352, + 445, + 1343, + 445, + 1343, + 478, + 352, + 478 + ], + "score": 0.333 + }, + { + "category_id": 1, + "poly": [ + 300, + 1066, + 488, + 1066, + 488, + 1099, + 300, + 1099 + ], + "score": 0.239 + }, + { + "category_id": 15, + "poly": [ + 292.0, + 531.0, + 548.0, + 531.0, + 548.0, + 579.0, + 292.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 936.0, + 557.0, + 936.0, + 557.0, + 966.0, + 299.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1065.0, + 491.0, + 1065.0, + 491.0, + 1103.0, + 295.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 443.0, + 1349.0, + 443.0, + 1349.0, + 482.0, + 351.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 602.0, + 1403.0, + 602.0, + 1403.0, + 643.0, + 294.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 634.0, + 1402.0, + 634.0, + 1402.0, + 674.0, + 294.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 666.0, + 1405.0, + 666.0, + 1405.0, + 703.0, + 294.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 696.0, + 1404.0, + 696.0, + 1404.0, + 735.0, + 292.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 728.0, + 1078.0, + 728.0, + 1078.0, + 765.0, + 294.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 775.0, + 1402.0, + 775.0, + 1402.0, + 808.0, + 295.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 805.0, + 1407.0, + 805.0, + 1407.0, + 839.0, + 293.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 834.0, + 1403.0, + 834.0, + 1403.0, + 869.0, + 294.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 866.0, + 525.0, + 866.0, + 525.0, + 899.0, + 295.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 982.0, + 964.0, + 982.0, + 964.0, + 1023.0, + 295.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1119.0, + 1403.0, + 1119.0, + 1403.0, + 1151.0, + 296.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1149.0, + 1145.0, + 1149.0, + 1145.0, + 1183.0, + 321.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1201.0, + 1405.0, + 1201.0, + 1405.0, + 1239.0, + 295.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1231.0, + 1405.0, + 1231.0, + 1405.0, + 1272.0, + 319.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1261.0, + 399.0, + 1261.0, + 399.0, + 1297.0, + 320.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1743.0, + 1405.0, + 1743.0, + 1405.0, + 1783.0, + 297.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1778.0, + 874.0, + 1778.0, + 874.0, + 1808.0, + 323.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1404.0, + 1402.0, + 1404.0, + 1402.0, + 1440.0, + 296.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1435.0, + 841.0, + 1435.0, + 841.0, + 1471.0, + 323.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1829.0, + 1315.0, + 1829.0, + 1315.0, + 1868.0, + 293.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1318.0, + 1404.0, + 1318.0, + 1404.0, + 1354.0, + 294.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1350.0, + 552.0, + 1350.0, + 552.0, + 1383.0, + 324.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1658.0, + 1404.0, + 1658.0, + 1404.0, + 1697.0, + 295.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1689.0, + 786.0, + 1689.0, + 786.0, + 1725.0, + 322.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1485.0, + 1362.0, + 1485.0, + 1362.0, + 1529.0, + 292.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1542.0, + 1402.0, + 1542.0, + 1402.0, + 1582.0, + 295.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1571.0, + 1403.0, + 1571.0, + 1403.0, + 1614.0, + 320.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1606.0, + 683.0, + 1606.0, + 683.0, + 1640.0, + 324.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1970.0, + 1405.0, + 1970.0, + 1405.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2002.0, + 973.0, + 2002.0, + 973.0, + 2036.0, + 322.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1882.0, + 1407.0, + 1882.0, + 1407.0, + 1924.0, + 292.0, + 1924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1918.0, + 867.0, + 1918.0, + 867.0, + 1950.0, + 324.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 443.0, + 1349.0, + 443.0, + 1349.0, + 482.0, + 351.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1065.0, + 491.0, + 1065.0, + 491.0, + 1103.0, + 295.0, + 1103.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 866, + 2088, + 866, + 2113, + 836, + 2113 + ], + "score": 0.838 + }, + { + "category_id": 1, + "poly": [ + 289, + 147, + 1409, + 147, + 1409, + 2045, + 289, + 2045 + ], + "score": 0.604 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 230.0, + 1408.0, + 230.0, + 1408.0, + 270.0, + 294.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 261.0, + 756.0, + 261.0, + 756.0, + 301.0, + 319.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 316.0, + 1406.0, + 316.0, + 1406.0, + 349.0, + 296.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 345.0, + 1107.0, + 345.0, + 1107.0, + 382.0, + 317.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 392.0, + 1406.0, + 392.0, + 1406.0, + 438.0, + 290.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 428.0, + 989.0, + 428.0, + 989.0, + 467.0, + 319.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 478.0, + 1406.0, + 478.0, + 1406.0, + 526.0, + 294.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 515.0, + 756.0, + 515.0, + 756.0, + 549.0, + 322.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 559.0, + 1408.0, + 559.0, + 1408.0, + 609.0, + 292.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 597.0, + 775.0, + 597.0, + 775.0, + 636.0, + 319.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 649.0, + 1408.0, + 649.0, + 1408.0, + 688.0, + 294.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 682.0, + 584.0, + 682.0, + 584.0, + 715.0, + 322.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 736.0, + 1406.0, + 736.0, + 1406.0, + 769.0, + 294.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 765.0, + 758.0, + 765.0, + 758.0, + 802.0, + 317.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 815.0, + 1406.0, + 815.0, + 1406.0, + 857.0, + 292.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 848.0, + 870.0, + 848.0, + 870.0, + 886.0, + 317.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 898.0, + 1406.0, + 898.0, + 1406.0, + 942.0, + 290.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 936.0, + 853.0, + 936.0, + 853.0, + 969.0, + 322.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 984.0, + 1406.0, + 984.0, + 1406.0, + 1025.0, + 292.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1019.0, + 1122.0, + 1019.0, + 1122.0, + 1052.0, + 322.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1067.0, + 1402.0, + 1067.0, + 1402.0, + 1109.0, + 292.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1100.0, + 1404.0, + 1100.0, + 1404.0, + 1140.0, + 319.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1154.0, + 1404.0, + 1154.0, + 1404.0, + 1194.0, + 294.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1188.0, + 633.0, + 1188.0, + 633.0, + 1221.0, + 319.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1235.0, + 1404.0, + 1235.0, + 1404.0, + 1275.0, + 292.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1267.0, + 849.0, + 1267.0, + 849.0, + 1308.0, + 317.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1319.0, + 1404.0, + 1319.0, + 1404.0, + 1362.0, + 290.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1352.0, + 608.0, + 1352.0, + 608.0, + 1392.0, + 317.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1404.0, + 1406.0, + 1404.0, + 1406.0, + 1446.0, + 292.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1437.0, + 720.0, + 1437.0, + 720.0, + 1477.0, + 319.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1487.0, + 1410.0, + 1487.0, + 1410.0, + 1529.0, + 290.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1521.0, + 478.0, + 1521.0, + 478.0, + 1560.0, + 319.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1577.0, + 1404.0, + 1577.0, + 1404.0, + 1610.0, + 294.0, + 1610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1604.0, + 946.0, + 1604.0, + 946.0, + 1643.0, + 319.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1656.0, + 1406.0, + 1656.0, + 1406.0, + 1695.0, + 292.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1685.0, + 1408.0, + 1685.0, + 1408.0, + 1731.0, + 317.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1722.0, + 877.0, + 1722.0, + 877.0, + 1756.0, + 322.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1774.0, + 1404.0, + 1774.0, + 1404.0, + 1808.0, + 296.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1806.0, + 756.0, + 1806.0, + 756.0, + 1839.0, + 319.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1854.0, + 1404.0, + 1854.0, + 1404.0, + 1897.0, + 292.0, + 1897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 1887.0, + 662.0, + 1887.0, + 662.0, + 1924.0, + 315.0, + 1924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1981.0, + 294.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1970.0, + 1406.0, + 1970.0, + 1406.0, + 2016.0, + 317.0, + 2016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 2001.0, + 662.0, + 2001.0, + 662.0, + 2039.0, + 319.0, + 2039.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 105, + 300, + 105 + ], + "score": 0.872 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.833 + }, + { + "category_id": 1, + "poly": [ + 294, + 228, + 1403, + 228, + 1403, + 294, + 294, + 294 + ], + "score": 0.741 + }, + { + "category_id": 1, + "poly": [ + 296, + 311, + 1401, + 311, + 1401, + 377, + 296, + 377 + ], + "score": 0.672 + }, + { + "category_id": 1, + "poly": [ + 300, + 395, + 1400, + 395, + 1400, + 459, + 300, + 459 + ], + "score": 0.666 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 227.0, + 1405.0, + 227.0, + 1405.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 258.0, + 1124.0, + 258.0, + 1124.0, + 296.0, + 321.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 310.0, + 1402.0, + 310.0, + 1402.0, + 347.0, + 296.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 344.0, + 1259.0, + 344.0, + 1259.0, + 378.0, + 321.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 392.0, + 1404.0, + 392.0, + 1404.0, + 432.0, + 295.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 428.0, + 1310.0, + 428.0, + 1310.0, + 460.0, + 324.0, + 460.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 294, + 628, + 1452, + 628, + 1452, + 1225, + 294, + 1225 + ], + "score": 0.985, + "html": "
Input OutputDropout Valid PPLParameters
256d EmbeddingAdaptive0.323.39314.7M
128d EmbeddingAdaptive0.321.51280.3M
64d EmbeddingAdaptive0.321.23263.1M
32d EmbeddingAdaptive0.321.78254.5M
512d Embedding512d Softmax (tied)0.322.63339.7M
512d Embedding512d Softmax (tied)0.428.31339.7M
512d Embedding512d Softmax0.323.87476.8
512d Embedding512d Softmax0.427.64476.8
256d Embedding256d Softmax (tied)0.322.65270.6M
256d Embedding256d Softmax0.324.13339.1M
64d Embedding512d Softmax0.324.74356.3M
AdaptiveAdaptive (tied emb, not proj)0.320.06247.3M
AdaptiveAdaptive (tied emb/proj not head)0.319.79246.9M
AdaptiveAdaptive (tied emb/proj + head)0.320.06246.9M
512d Embedding512d Softmax (tied)0.322.63339.7M
512d Embedding 512d Adaptive (no dim reduction, tied)0.325.48340.2M
" + }, + { + "category_id": 1, + "poly": [ + 298, + 369, + 1404, + 369, + 1404, + 524, + 298, + 524 + ], + "score": 0.978 + }, + { + "category_id": 5, + "poly": [ + 649, + 1326, + 1048, + 1326, + 1048, + 1606, + 649, + 1606 + ], + "score": 0.972, + "html": "
Softmax cutoffValid PPL
20k/40k/200k19.79
20k/140k/100k 20k/40k/60k/140k20.26
60k/100k/100k20.53 20.52
5k/155k/100k
20k/40k/200k20.06
10k/60k/190k19.99 19.79
" + }, + { + "category_id": 1, + "poly": [ + 300, + 1855, + 1401, + 1855, + 1401, + 1920, + 300, + 1920 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 299, + 539, + 1401, + 539, + 1401, + 602, + 299, + 602 + ], + "score": 0.939 + }, + { + "category_id": 0, + "poly": [ + 299, + 1786, + 503, + 1786, + 503, + 1821, + 299, + 1821 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 300, + 300, + 1020, + 300, + 1020, + 336, + 300, + 336 + ], + "score": 0.868 + }, + { + "category_id": 6, + "poly": [ + 488, + 1247, + 1207, + 1247, + 1207, + 1281, + 488, + 1281 + ], + "score": 0.862 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.851 + }, + { + "category_id": 1, + "poly": [ + 301, + 226, + 711, + 226, + 711, + 261, + 301, + 261 + ], + "score": 0.805 + }, + { + "category_id": 1, + "poly": [ + 297, + 1630, + 1403, + 1630, + 1403, + 1723, + 297, + 1723 + ], + "score": 0.743 + }, + { + "category_id": 6, + "poly": [ + 297, + 1630, + 1403, + 1630, + 1403, + 1723, + 297, + 1723 + ], + "score": 0.129 + }, + { + "category_id": 0, + "poly": [ + 301, + 226, + 711, + 226, + 711, + 261, + 301, + 261 + ], + "score": 0.104 + }, + { + "category_id": 13, + "poly": [ + 715, + 1857, + 765, + 1857, + 765, + 1889, + 715, + 1889 + ], + "score": 0.79, + "latex": "\\ S 5 . 3" + }, + { + "category_id": 13, + "poly": [ + 806, + 1111, + 830, + 1111, + 830, + 1136, + 806, + 1136 + ], + "score": 0.25, + "latex": "^ +" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1781.0, + 507.0, + 1781.0, + 507.0, + 1829.0, + 292.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 300.0, + 1023.0, + 300.0, + 1023.0, + 340.0, + 295.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 1246.0, + 1210.0, + 1246.0, + 1210.0, + 1283.0, + 491.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1628.0, + 1407.0, + 1628.0, + 1407.0, + 1666.0, + 293.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1661.0, + 1405.0, + 1661.0, + 1405.0, + 1695.0, + 295.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1690.0, + 1198.0, + 1690.0, + 1198.0, + 1722.0, + 297.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 223.0, + 714.0, + 223.0, + 714.0, + 266.0, + 296.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 367.0, + 1405.0, + 367.0, + 1405.0, + 405.0, + 295.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 400.0, + 1402.0, + 400.0, + 1402.0, + 433.0, + 294.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 431.0, + 1401.0, + 431.0, + 1401.0, + 464.0, + 294.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 458.0, + 1407.0, + 458.0, + 1407.0, + 498.0, + 292.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 489.0, + 831.0, + 489.0, + 831.0, + 530.0, + 293.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 714.0, + 1850.0, + 714.0, + 1894.0, + 294.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 1850.0, + 1403.0, + 1850.0, + 1403.0, + 1894.0, + 766.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1886.0, + 698.0, + 1886.0, + 698.0, + 1922.0, + 296.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 539.0, + 1403.0, + 539.0, + 1403.0, + 574.0, + 297.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 568.0, + 1355.0, + 568.0, + 1355.0, + 604.0, + 295.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 223.0, + 714.0, + 223.0, + 714.0, + 266.0, + 296.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1628.0, + 1407.0, + 1628.0, + 1407.0, + 1666.0, + 293.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1661.0, + 1405.0, + 1661.0, + 1405.0, + 1695.0, + 295.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1690.0, + 1198.0, + 1690.0, + 1198.0, + 1722.0, + 297.0, + 1722.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 302, + 811, + 1401, + 811, + 1401, + 1338, + 302, + 1338 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 294, + 1368, + 1406, + 1368, + 1406, + 1432, + 294, + 1432 + ], + "score": 0.938 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.843 + }, + { + "category_id": 15, + "poly": [ + 591.0, + 830.0, + 1207.0, + 830.0, + 1207.0, + 874.0, + 591.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 864.0, + 417.0, + 864.0, + 417.0, + 906.0, + 362.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 915.0, + 341.0, + 915.0, + 341.0, + 1174.0, + 301.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 954.0, + 412.0, + 954.0, + 412.0, + 985.0, + 388.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 1001.0, + 1296.0, + 1001.0, + 1296.0, + 1261.0, + 1280.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 1011.0, + 1311.0, + 1011.0, + 1311.0, + 1197.0, + 1296.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 1016.0, + 1326.0, + 1016.0, + 1326.0, + 1256.0, + 1311.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1325.0, + 1016.0, + 1341.0, + 1016.0, + 1341.0, + 1256.0, + 1325.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1344.0, + 1018.0, + 1352.0, + 1018.0, + 1352.0, + 1035.0, + 1344.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 1023.0, + 528.0, + 1023.0, + 528.0, + 1256.0, + 512.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1033.0, + 417.0, + 1033.0, + 417.0, + 1076.0, + 361.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1344.0, + 1039.0, + 1352.0, + 1039.0, + 1352.0, + 1056.0, + 1344.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 1121.0, + 414.0, + 1121.0, + 414.0, + 1159.0, + 386.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 1051.0, + 513.0, + 1051.0, + 513.0, + 1261.0, + 464.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1046.0, + 541.0, + 1046.0, + 541.0, + 1208.0, + 528.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 1109.0, + 617.0, + 1109.0, + 617.0, + 1258.0, + 600.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1128.0, + 662.0, + 1128.0, + 662.0, + 1150.0, + 648.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1342.0, + 1057.0, + 1355.0, + 1057.0, + 1355.0, + 1193.0, + 1342.0, + 1193.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1203.0, + 417.0, + 1203.0, + 417.0, + 1246.0, + 361.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 1146.0, + 629.0, + 1146.0, + 629.0, + 1256.0, + 617.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 1155.0, + 660.0, + 1155.0, + 660.0, + 1258.0, + 632.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1150.0, + 677.0, + 1150.0, + 677.0, + 1170.0, + 663.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1161.0, + 813.0, + 1161.0, + 813.0, + 1262.0, + 735.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1181.0, + 935.0, + 1181.0, + 935.0, + 1263.0, + 869.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 1199.0, + 946.0, + 1199.0, + 946.0, + 1214.0, + 936.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 1193.0, + 1026.0, + 1193.0, + 1026.0, + 1262.0, + 1006.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 1207.0, + 1036.0, + 1207.0, + 1036.0, + 1220.0, + 1028.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 1206.0, + 1071.0, + 1206.0, + 1071.0, + 1259.0, + 1039.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1208.0, + 1081.0, + 1208.0, + 1081.0, + 1224.0, + 1073.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 1155.0, + 1160.0, + 1155.0, + 1160.0, + 1261.0, + 1144.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 1161.0, + 1219.0, + 1161.0, + 1219.0, + 1262.0, + 1161.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 1243.0, + 810.0, + 1243.0, + 810.0, + 1253.0, + 800.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 1219.0, + 1040.0, + 1219.0, + 1040.0, + 1256.0, + 1024.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1299.0, + 1245.0, + 1309.0, + 1245.0, + 1309.0, + 1253.0, + 1299.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 1268.0, + 521.0, + 1268.0, + 521.0, + 1308.0, + 477.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 1269.0, + 663.0, + 1269.0, + 663.0, + 1310.0, + 604.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 1268.0, + 795.0, + 1268.0, + 795.0, + 1310.0, + 743.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 1268.0, + 936.0, + 1268.0, + 936.0, + 1308.0, + 873.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1268.0, + 1079.0, + 1268.0, + 1079.0, + 1310.0, + 1000.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 1268.0, + 1204.0, + 1268.0, + 1204.0, + 1308.0, + 1148.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 1269.0, + 1346.0, + 1269.0, + 1346.0, + 1308.0, + 1276.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1296.0, + 1022.0, + 1296.0, + 1022.0, + 1344.0, + 786.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1367.0, + 1405.0, + 1367.0, + 1405.0, + 1406.0, + 295.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1399.0, + 539.0, + 1399.0, + 539.0, + 1435.0, + 295.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/HJxB5sRcFQ/HJxB5sRcFQ.md b/parse/train/HJxB5sRcFQ/HJxB5sRcFQ.md new file mode 100644 index 0000000000000000000000000000000000000000..171dbad360816782337196f073d746ebd6aadaf8 --- /dev/null +++ b/parse/train/HJxB5sRcFQ/HJxB5sRcFQ.md @@ -0,0 +1,289 @@ +# LAYOUTGAN: GENERATING GRAPHIC LAYOUTSWITH WIREFRAME DISCRIMINATORS + +Jianan $\mathbf { L i } ^ { 1 }$ ∗, Jimei Yang2, Aaron Hertzmann2, Jianming Zhang2, Tingfa $\mathbf { X } \mathbf { u } ^ { 1 }$ + +1. Beijing Institute of Technology 2. Adobe Research + +{20090964,xutingfa}@bit.edu.cn, {jimyang,hertzman,jianmzha}@adobe.com + +# ABSTRACT + +Layout is important for graphic design and scene generation. We propose a novel Generative Adversarial Network, called LayoutGAN, that synthesizes layouts by modeling geometric relations of different types of 2D elements. The generator of LayoutGAN takes as input a set of randomly-placed 2D graphic elements and uses self-attention modules to refine their labels and geometric parameters jointly to produce a realistic layout. Accurate alignment is critical for good layouts. We thus propose a novel differentiable wireframe rendering layer that maps the generated layout to a wireframe image, upon which a CNN-based discriminator is used to optimize the layouts in image space. We validate the effectiveness of LayoutGAN in various experiments including MNIST digit generation, document layout generation, clipart abstract scene generation and tangram graphic design. + +# 1 INTRODUCTION + +Graphic design is an important visual communication tool in our modern world, encompassing everything from book covers to magazine layouts to web design. Whereas methods for generating realistic natural-looking images have made significant progress lately, particularly with Generative Adversarial Networks (GANs) (Karras et al., 2018), methods for creating designs are far more primitive. This is, in part, due to the difficulty of finding data representations suitable for learning. Graphic designs are normally composed of vector representations of primitive objects, such as polygons, curves and ellipses, instead of pixels laid on a regular lattice. The quality and content of a design depends on the presence of elements, their attributes, and their relations to other elements. The visual perception of design depends on the arrangement of these elements; misalignment of two elements of just a few millimeters can ruin the design. Training from images of designs using conventional GANs synthesizes the layouts in pixel space, and thus mixes up layout and its rendering, and thus would be unlikely to capture layout styles well. Modeling such highly-structured data using neural networks is of great interest as they usually represent human abstract knowledge about the visual world (Zitnick & Parikh, 2013; Song et al., 2017) and how this knowledge are expressed via documents and designs (Deka et al., 2017; Yang et al., 2017). + +This paper introduces LayoutGAN, a novel GAN which directly synthesizes a set of graphical elements in a design. In a given design problem, a fixed set of element classes (e.g., “title,” “figure”) is specified in advance. In our network, each element is represented by its class probabilities and its geometric parameters, i.e., bounding-box keypoints. The generator takes as input graphic elements with randomly-sampled class probabilities and geometric parameters, and arranges them in a design; the output is the refined class probabilities and geometric parameters of the design elements. The generator has the desirable property of being permutation-invariant: it will generate the same layout if we re-order the input elements. + +We propose two kinds of discriminator networks for this structured data. The first is similar in structure to the generator: it operates directly on the class probabilities and geometric parameters of the elements. Though effective, it is not sensitive enough to misalignment and occlusion between elements. The second discriminator operates in the visual domain. Like a human viewer who judges a design by looking at its rasterized image, the relations among different elements can be evaluated well by mapping them to 2D layouts. Then Convolutional Neural Networks (CNNs) can be used for layout optimization as they are specialized in distinguishing visual patterns including but not limited to misalignment and occlusion. However, the key challenge is how to map the geometric parameters to pixel-level layouts differentiably. One approach would be to render the graphic elements into bitmap masks using Spatial Transformer Networks (Jaderberg et al., 2015). But we found that filled pixels within the design elements cause occlusions and are ineffective for back-propagation, for example, when a small polygon hides behind a larger one. We experimented with bitmap mask rendering but it was not successful. In this paper, we propose a novel differentiable wireframe rendering layer that rasterizes both synthesized and real structured data of graphic elements into wireframe images, upon which a standard CNN can be used to optimize the layout across the visual and the graphic domain. The wireframe rendering discriminator has several advantages. First, convolution layers are very good at extracting spatial patterns of images so that they are more sensitive to alignment. Second, the rendered wireframes make elements visible even when they overlap and thus the network is alleviated from inferring the occlusions that may occur in other renderings such as masks. + +We evaluate the LayoutGAN for several different tasks, including a sanity test on MNIST digits, generating page layouts from labeled bounding boxes, generating clipart abstract scenes, tangram graphic design, and mobile app design layouts. In each case, our method successfully generates layouts respecting the types of elements and their relationships for the problem domain. + +In summary, the LayoutGAN comprises the following contributions: 1. A Generator that directly synthesizes structured data, represented as a resolution-independent set of labeled graphic elements in a design. 2. A differentiable wireframe rendering layer which allows the Discriminator to judge alignment from discrete element arrangements. + +# 2 RELATED WORK + +Structured data generation. Convolutional networks have been shown successful for generating data in regular lattice, such as images (Radford et al., 2015), videos (Vondrick et al., 2016) and 3D volumes (Yan et al., 2016; Wu et al., 2016). When generating highly-structured data, such as text (Donahue et al., 2015) and programs (Reed & De Freitas, 2015), recurrent networks are often the first choice (Sutskever et al., 2014), especially equipped with attention (Bahdanau et al., 2014) and memory modules (Graves et al., 2014). Recently, researchers show that convolutional networks can be also used to synthesize sequences (Oord et al., 2016; van den Oord et al., 2016) using autoregressive models. However, in many cases, an object has no sequential order (Vinyals et al., 2015), but is a set of elements, e.g. point clouds. Fan et al. (2017) propose a point set generation network for synthesizing 3D point clouds of the object shape from a single images. It is further paired with a point set classification network (Charles et al., 2017) for auto-encoding 3D point clouds (Achlioptas et al., 2017). Our work extends the set representation to more general primitive objects, i.e., labeled polygons. Meanwhile, researchers also model the structured data of connected elements using graph convolutions (Kipf & Welling, 2017). + +Data-driven graphic design. Automating layout is a classic problem in graphic design (Hurst et al., 2009). O’Donovan et al. (2014) formulate an energy function by assembling various heuristic visual cues and design principles to optimize single-page layouts, and extend this to an interactive tool (O’Donovan et al., 2015). The model parameters are learned from a small number of example designs. Pang et al. (2016) optimize layout for desired gaze direction. Deka et al. (2017) collect a mobile app design database for harnessing data-driven applications and present preliminary results of learning similarities of pixel-level textural/non-textual masks for design search, but do not learn models from this data. Swearngin et al. (2018) propose an interactive system that converts example design screenshots to vector graphics for designers to re-use and edit. Bylinskii et al. (2017) analyze the visual importance of graphic designs and use saliency map as the driving force to assist retargeting and thumbnailing. Previous methods have learned models for other graphic design elements, such as fonts (O’Donovan et al., 2014) and colors (O’Donovan et al., 2011). These are orthogonal to the layout problem, and could be combined in future work. No previous method has learned to create design or layout from large datasets, and no previous work has applied GANs to layout. + +3D scene synthesis. Interior scene synthesis and furniture layout generation draws great interest in graphics community. Early approaches focus on optimization of hand-crafted design principles (Merrell et al., 2011) and learning statistical priors of pairwise object relationships (Fisher et al., 2012) due to limited data. Wang et al. (2018a) recently proposed a sequential decision making approach to indoor scene synthesis. In each step, a CNN is trained to predict either location or category of one object by looking at the rendered top-down views. This resembles our wireframe rendering discriminator, in the sense of using convolutions to capture spatial patterns of layouts. + +![](images/6d7c4e44cbab0b386322b6d183c456233be917dfc8168c78c361d237ed97a2c0.jpg) +Figure 1: Overall architecture of LayoutGAN. The generator takes as input graphic elements with randomly sampled class probabilities and geometric parameters from Uniform and Gaussian distribution respectively. An encoder embeds the input and feeds them into the stacked relation module, which refines the embedded features of each element in a coordinative manner by considering its semantic and spatial relations with all the other elements. Finally, a decoder decodes the refined features back to class probabilities and geometric parameters. The wireframe rendering discriminator feeds the generated results to a differentiable wireframe rendering layer which raterizes the input graphic elements into 2D wireframe images, upon which a CNN is applied for layout optimization. + +# 3 LAYOUTGAN + +This section describes our data and model representations. + +# 3.1 DESIGN REPRESENTATION + +In our model, a graphic design is comprised of a set of $N$ primitive design elements $\{ ( p _ { 1 } , \pmb { \theta _ { 1 } } ) , \cdot \cdot \cdot , ( p _ { N } , \pmb { \theta } _ { N } ) \}$ . Each element has a set of geometric parameters $\pmb { \theta }$ , and a vector of class probabilities $\pmb { p }$ . The entries of these variables are problem-dependent. For example, document layouts include 6 classes, e.g. “title” and “picture” while clipart layouts include 6 classes, e.g. “boy” and “hat”. For 2D point set generation (in MNIST digits), $\mathbf { \bar { \theta } } \equiv [ \mathbf { \bar { \boldsymbol { x } } } , y ]$ , representing the coordinates of each point; for bounding-box generation in document layout, $\pmb \theta \equiv \left[ x ^ { L } , y ^ { T } , x ^ { R } , y ^ { B } \right]$ , representing the top-left and bottom-right coordinates of each bounding box; for layouts with scale and flip (clipart abstract scenes), $\pmb \theta \equiv [ x , y , s , l ]$ , representing the center coordinates, scale and flip of each element. + +# 3.2 GENERATOR ARCHITECTURE + +In the LayoutGAN, the Generator is a function $G ( z )$ that takes a layout as input, where $z =$ $\{ ( p _ { 1 } , \pmb { \theta _ { 1 } } ) , \cdot \cdot \cdot , ( p _ { N } , \pmb { \theta } _ { N } ) \}$ consisting of initial graphic elements with randomly-sampled geometric parameters $\theta _ { i }$ , and one-hot encoding of a randomly-sampled class $\mathbf { \mathit { p } } _ { i }$ . The Generator outputs a refined layout $G ( z ) = \{ ( p _ { 1 } ^ { \prime } , \theta _ { 1 } ^ { \prime } ) , \cdot \cdot \cdot , \bar { ( } p _ { N } ^ { \prime } , \theta _ { N } ^ { \prime } ) \}$ which is meant to resemble a real graphic design. Note that, unlike, conventional GANs where $_ z$ represents a low-dimensional latent variable, our $_ { z }$ represents initial random graphic layout that has the same structure as the real one. The Discriminator learns to capture the geometric relations among different types of elements for layout optimization from both the graphic and the visual domain. Next, we go into details of the generator and discriminator design. + +As illustrated in Figure 1, the generator takes as input a set of graphic elements with random class probabilities and geometric parameters sampled from Uniform and Gaussian distribution respectively. An encoder consisting of a multilayer perceptron network (implemented as multiple fully connected layers) first embeds the class one-hot vectors and geometric parameters of each graphic element. The relation module, implemented as self-attention inspired by Wang et al. (2018b), is then used to embed the feature of each graphic element as a function of its spatial context, i.e., its relations with all the other elements in the design. Denote $f ( p _ { i } , \theta _ { i } )$ as the embedded feature of the graphic element $i$ , its refined feature representation $f ^ { \prime } ( p _ { i } , \theta _ { i } )$ can be obtained through a contextual residual learning process, which is defined as: + +$$ +f ^ { \prime } ( p _ { i } , \theta _ { i } ) = W _ { r } \frac { 1 } { N } \sum _ { \forall j \neq i } H ( f ( p _ { i } , \theta _ { i } ) , f ( p _ { j } , \theta _ { j } ) ) U ( f ( p _ { j } , \theta _ { j } ) ) + f ( p _ { i } , \theta _ { i } ) , +$$ + +where an unary function $U$ computes a representation of the embedded feature $f ( p _ { j } , \theta _ { j } )$ of element $j$ and a pairwise function $\pmb { H }$ computes a scalar value representing the relation between elements $i$ and $j$ . Thus, all the other elements $j \neq i$ contribute to the feature refinement of element $i$ by summing up their relations. The response is normalized by the total number of elements in the set, $N$ . The weight matrix $W _ { r }$ computes a linear embedding, producing the contextual residual to be added to $f ( p _ { i } , \theta _ { i } )$ for feature refinement. In our experiments, we define $\pmb { H }$ as a dot-product: + +$$ +\pmb { H } ( f ( p _ { i } , \pmb { \theta } _ { i } ) , f ( p _ { j } , \pmb { \theta } _ { j } ) ) = \psi ( f ( p _ { i } , \pmb { \theta } _ { i } ) ) ^ { T } \phi ( f ( p _ { j } , \pmb { \theta } _ { j } ) ) , +$$ + +where $\psi ( f ( p _ { i } , \pmb { \theta } _ { i } ) ) = W _ { \psi } f ( p _ { i } , \pmb { \theta } _ { i } )$ and $\phi ( f ( p _ { j } , \pmb { \theta } _ { j } ) ) = W _ { \phi } f ( p _ { j } , \pmb { \theta } _ { j } )$ are two linear embeddings. We stack 4 relation modules for feature refinement in our experiments. Finally, a decoder consisting of another multilayer perceptron network followed by two branches of fully connected layer with a sigmoid activation function is used to map the refined feature of each element back to class probabilities and geometric parameters respectively. Optionally, Non-Maximum Suppression (NMS) can be applied to remove duplicated elements. + +# 3.3 DISCRIMINATOR NETWORK ARCHITECTURES + +The discriminator aims to distinguish between synthesized and real layouts. We present two types of discriminators, one based on relation modules directly built upon layout parameters, and the other based on how the layouts look like via rendering. + +# 3.3.1 RELATION-BASED DISCRIMINATOR + +The Relation-Based Discriminator takes as input a set of graphic elements represented by class probabilities and geometric parameters, and feeds them to an encoder consisting of a multilayer perceptron network for feature embedding $f ( p _ { i } , \theta _ { i } )$ . It then extracts their global graphical relations $g ( r ( p _ { 1 } , \theta _ { 1 } ) , \cdot \cdot \cdot , r ( p _ { N } , \theta _ { N } ) )$ where $r ( p _ { i } , \theta _ { i } )$ represents a simplified relation module as in equation 1 by removing the shortcut connection, and $\textbf { { g } }$ is a max-pooling function used in (Charles et al., 2017). Thus, the global relations among all the graphic elements can be modeled, upon which a classifier composed by a multilayer perception network is applied for real/fake prediction. + +# 3.3.2 WIREFRAME RENDERING DISCRIMINATOR + +The Wireframe Rendering Discriminator exploits CNNs to classify layouts, in order to learn to classify the visual properties of a layout. The Discriminator consists of a wireframe rendering layer, which produces an output image $\pmb { I }$ , which is then fed into a CNN for classification. + +The rasterization is performed as follows. Suppose there are $N$ elements in the design, with parameters $\{ ( p _ { 1 } , \pmb { \theta _ { 1 } } ) , . . . , ( p _ { N } , \pmb { \theta } _ { N } ) \}$ . Each element can be rendered into its own grayscale image $F _ { \theta } ( x , y )$ ; rendering details for specific types of elements are given below. The image dimensions for each individual $\pmb { F }$ are $W \times H$ , where $W$ and $H$ are the width and height of the design in pixels. + +The layer output is a multi-channel image $\pmb { I }$ of dimensions $W \times H \times M$ , where each channel corresponds to one of the $M$ element types. In other words, pixel $( x , y )$ of $\pmb { I }$ is a class activation vector for that pixel, and is computed as: + +$$ +\pmb { I } ( x , y , c ) = \operatorname* { m a x } _ { i \in [ 1 . . . N ] } p _ { i , c } \pmb { F } _ { \theta _ { i } } ( x , y ) +$$ + +where $p _ { i , c }$ is the probability that element $i$ is of class $c$ + +We next describe the rendering process for computing $F _ { \theta }$ , in the cases where $\pmb \theta$ represents a point, a rectangle, and a triangle. + +Point. We start with the simplest geometric form, a single keypoint $\theta _ { i } = ( x _ { i } , y _ { i } )$ for element $i$ . We implement an interpolation kernel $k$ for its rasterization. Its spatial rendering response on $( x , y )$ in the rendered image can be written as: + +$$ +F _ { \theta _ { i } } ( x , y ) = k ( x - x _ { i } ) k ( y - y _ { i } ) . +$$ + +![](images/ff33a8ebcae43f98ddef388c3817179178471890959278d5168cd6dac657d1d0.jpg) +Figure 2: Wireframe rendering of different polygons (point, rectangle and triangle). The black grids represent grids of target image. The orange dots/dotted lines represent the graphic element mapped onto the image grid. The blue solid lines represent the rasterized wireframes expressed as differentiable functions of graphic elements in terms of both class probilities and geometric parameters. + +We adopt bilinear interpretation (Johnson et al., 2016), corresponding to the kernel $\boldsymbol k ( d ) \_ =$ $\operatorname* { m a x } ( 0 , \bar { 1 } - | d | )$ (implemented as ReLU activation), as shown in Figure 2. As $\pmb { I }$ is a differentiable function of the class probabilities and the coordinates, the sub-gradients of the rasterized image can be thus propagated backward to them. We validate such rendering design for MNIST digit generation, detailed experiments can be seen in Section 4.1. + +Rectangle. We now consider more complex polygons. Assuming an element is a rectangle, or bounding box represented by its top-left and bottom-right coordinates $\pmb { \theta } = ( x ^ { L } , y ^ { T } , x ^ { R } , y ^ { B } )$ , which is very common in various designs. Specifically, considering a rectangle $_ i$ with coordinates $\theta _ { i } = $ $( x _ { i } ^ { L } , y _ { i } ^ { T } , x _ { i } ^ { R } , y _ { i } ^ { B } )$ , as shown in Figure 2, the black grids represent the locations in the rendered image and the orange dotted box represents the rectangle being rasterized in the rendered image. For a wireframe representation, only the points near the boundary of the dotted box (lie in blue solid line) are related to the rectangle, so its spatial rendering response on $( x , y )$ can be formulated as: + +$$ +\begin{array} { r } { F _ { \theta _ { i } } ( x , y ) = \operatorname* { m a x } \left( \begin{array} { l } { k ( x - x _ { i } ^ { L } ) b ( y - y _ { i } ^ { T } ) b ( y _ { i } ^ { B } - y ) , } \\ { k ( x - x _ { i } ^ { R } ) b ( y - y _ { i } ^ { T } ) b ( y _ { i } ^ { B } - y ) , } \\ { k ( y - y _ { i } ^ { T } ) b ( x - x _ { i } ^ { L } ) b ( x _ { i } ^ { R } - x ) , } \\ { k ( y - y _ { i } ^ { B } ) b ( x - x _ { i } ^ { L } ) b ( x _ { i } ^ { R } - x ) } \end{array} \right) , } \end{array} +$$ + +where $b ( d ) = \operatorname* { m i n } ( \operatorname* { m a x } ( 0 , d ) , 1 )$ constraining the rendering to nearby pixels. + +Triangle. We further describe the wireframe rendering process of another geometric form, triangle. For triangle $i$ represented by its three vertices’ coordinates $\pmb { \theta _ { i } } = ( x _ { i } ^ { 1 } , y _ { i } ^ { 1 } , x _ { i } ^ { \top } , y _ { i } ^ { 2 } , x _ { i } ^ { 3 } , y _ { i } ^ { 3 } )$ , when $x _ { i } ^ { \bar { 1 } } \neq$ $x _ { i } ^ { 2 } \neq x _ { i } ^ { 3 }$ , its spatial rendering response on $( x , y )$ in the rendered image can be calculated as: + +$$ +\begin{array} { r } { F _ { \theta _ { i } } ( x , y ) = \operatorname* { m a x } \left( \begin{array} { l } { k ( y - \frac { ( y _ { i } ^ { 2 } - y _ { i } ^ { 1 } ) \cdot ( x - x _ { i } ^ { 1 } ) } { x _ { i } ^ { 2 } - x _ { i } ^ { 1 } } - y _ { i } ^ { 1 } ) b ( x - x _ { i } ^ { 1 } ) b ( x _ { i } ^ { 2 } - x ) , } \\ { k ( y - \frac { ( y _ { i } ^ { 3 } - y _ { i } ^ { 1 } ) \cdot ( x - x _ { i } ^ { 1 } ) } { x _ { i } ^ { 3 } - x _ { i } ^ { 1 } } - y _ { i } ^ { 1 } ) b ( x - x _ { i } ^ { 3 } ) b ( x _ { i } ^ { 1 } - x ) , } \\ { k ( y - \frac { ( y _ { i } ^ { 3 } - y _ { i } ^ { 2 } ) \cdot ( x - x _ { i } ^ { 2 } ) } { x _ { i } ^ { 3 } - x _ { i } ^ { 2 } } - y _ { i } ^ { 2 } ) b ( x - x _ { i } ^ { 3 } ) b ( x _ { i } ^ { 2 } - x ) } \end{array} \right) , } \end{array} +$$ + +Through this wireframe rendering process, gradients can be propagated backward to both the class probabilities and geometric parameters of the graphic elements for joint optimization. + +A CNN consisting of 3 convolutional layers, followed by a fully connected layer with sigmoid activation, is applied to the rasterization layer $\pmb { I }$ to predict real/fake graphic layouts. + +# 4 EXPERIMENTS + +The implementation is based on TensorFlow (Abadi et al., 2016). The network parameters are initialized from zero-mean Gaussian with standard deviation of 0.02. All the networks are optimized using Adam (Kingma & Ba, 2014) with a fixed learning rate of 0.00002. Detailed architectures can be found in the appendix. + +# 4.1 MNIST DIGIT GENERATION + +As a toy problem, we generate MNIST layouts. MNIST is a handwritten digit database consisting of 60,000 training and 10,000 testing images. For each image, we extract the locations of 128 randomly-selected foreground pixels as the graphic representation so that digit generation can be formulated as generating of 2D point layouts. In Figure 3a, each image shows $8 \times 8$ digits rendered from point layouts. The left and middle images are the generated samples from the LayoutGAN with a relation-based discriminator and a wireframe rendering discriminator, respectively. The right image shows the digits rendered from ground-truth point layouts. One can see that the LayoutGAN with either discriminator captures various patterns. The wireframe rendering discriminator generates more compact, better-aligned point layouts. For quantitative evaluation, we first train a multilayer perceptron network for digit classification, which achieves an accuracy of $9 8 . 9 1 \%$ on MNIST test set, and then use the classifier to compute the inception score (Salimans et al., 2016) of 10,000 generated samples. As shown in Table 1, the LayoutGAN with wireframe rendering discriminator achieves a higher inception score than with the relation-based discriminator. To shed light on the relational refinement process of LayoutGAN, we mark each input random point with a locationspecific color, and track their refined locations in the generated layouts, as shown in Figure 3b. One can see that colors change gradually along the strokes, showing the network learns some contextually consistent local displacements of the points. + +![](images/a10d4ec280b2dd747078d880f5fbce6b8ddfb30566943603c69b4dd1b57b5bca.jpg) +Figure 3: Results on MNIST digit generation. + +Table 1: Inception scores for generated digits. + +
MethodsScore± std
Relation6.53 ± .09
Wireframe7.36 ± .07
Real data9.81 ± .08
+ +Table 2: Spatial analysis of document layout. + +
MethodsOverlap(%)Alignment(%)
Relation1.526.4
Wireframe1.173.4
Real data0.050.5
+ +# 4.2 DOCUMENT LAYOUT GENERATION + +One document page consists of a number of regions with different element types, such as heading, paragraph, table, figure, caption and list. Each region is represented by a bounding box. Modeling layouts of regions is critical for document analysis, retargeting, and synthesis. In real document data, these bounding boxes are often carefully aligned along the canonical axes, and their placement follows some particular patterns, such as heading always appearing above paragraph or table. Some example layouts are shown in the appendix. + +In this experiment, we focus on one-column layouts with no more than 9 bounding boxes that may belong to 6 possible classes as mentioned above. We are interested in how these layout patterns can be captured by our network. For training data, we collect totally around 25,000 layouts from real documents. We also implemented a baseline method that represents the layout by semantic masks as used in Yang et al. (2017); Deka et al. (2017). Specifically, we render all elements into masks and train a DCGAN (Radford et al., 2015) to generate mask layouts in pixels. Then we extract the connected mask regions of each class and output enclosing bounding boxes of all regions. Figure 4 shows some representative generation results (each row consists of 6 samples; high-resolution results can be found in the Appendix). The first row shows the results from DCGAN. DCGAN mixes a layout and its rendering in the generation process. When convolutional layers in DCGAN fails to perfectly replicate the rendering process (which is usually the case), it generates fuzzy and noisy label maps, making the extracted bounding boxes less consistent and less accurate. The third row shows the results from LayoutGAN with wireframe rendering discriminator. We retrieve the most similar real layouts from the training set in the last row as references. It can be seen that LayoutGAN can well capture different document layout patterns with clear definitions of instance-level semantic bounding boxes. + +To validate the advantage of wireframe rendering discriminator, we compare the generated results with those from the LayoutGAN with relation-based discriminator. As shown in the second row of Figure 4, the latter can also capture different layout patterns but sometimes suffers from overlapping and misalignment issues. As the bounding boxes in real document layouts are either left- or centeraligned with no overlaps with each other (except for captions with figures or tables), we propose two metrics to quantitatively measure the quality of the generated layouts. The first one is overlap index, which is the percentage of total overlapping area among any two bounding boxes inside the whole page. The second one is alignment index which is calculated by finding the minimum standard deviation of either left or center coordinates of all the bounding boxes. Table 2 provides quantitative comparisons of real layouts and synthesized layouts from LayoutGAN with two different discriminators. One can see that LayoutGAN with wireframe rendering discriminator achieves lower value in both overlapping index and alignment index than the one with relation-based discriminator, validating the superiority of the proposed wireframe rendering method for layout generation. A similar conclusion can also be drawn by analyzing the loss functions. We add shift perturbations to the bounding boxes of real layouts and feed them to both discriminators to examine their loss behaviors. Figure 6 visualizes the loss landscape of both discriminators corresponding to different extent of shift perturbation. One can see that the loss surface w.r.t. shift perturbation of the wireframe rendering discriminator is much smoother than the one of relation-based discriminator. + +![](images/1b7f0b9374988295ff6466770bc433d9b1b8193220fd8eca8a0aba65c44d495b.jpg) +Figure 4: Document layout comparison. + +![](images/eb1c35ca53cdf5b316f1c04cd6b7b8c2044d49e7dbdf44c72cdc26db71c5a679.jpg) +Figure 5: Clipart abstract scene generation. + +![](images/c03606f88f5395b716af09a8b97c3528cd6c04cee92c55e211f4c586467d2a12.jpg) +Figure 6: Discriminator loss landscapes. + +Table 3: User study results for Clipart abstract scenes. + +
MethodsExcellent (%)Fair (%)Poor (%)
DCGAN6.738.854.5
Relation17.250.332.5
Wireframe37.348.014.7
+ +# 4.3 CLIPART ABSTRACT SCENE GENERATION + +We now consider generating scenes composed of a set of specific clipart elements. We use the abstract scene dataset (Zitnick et al., 2016) including boy, girl, glasses, hat, sun, and tree elements. To synthesize a reasonable abstract scene, we first use LayoutGAN to generate the layout of inner scene elements by representing each of them as a labeled bounding box with normalized center coordinates, width and height, and flip indicator. Then, we render the corresponding clipart image of each scene element onto a background image according to the predicted position (center coordinates), scale (width and height) and flip, forming a synthesized abstract scene. It is a challenging task as accurate pairwise or higher-order relations of objects are required for synthesizing reasonable scenes, for example, the glasses/hat should be exactly on the eyes/head of a boy/girl with proper scale and orientation. In Figure 5, the first three rows shows the generated layouts and corresponding rendered scenes by DCGAN, LayoutGAN with relation-based discriminator and LayoutGAN with wireframe rendering discriminator respectively. The last row shows some samples rendered from ground-truth scene layouts. It can be seen that, compared to DCGAN and the LayoutGAN with relation-based discriminator, the one with wireframe rendering discriminator can capture pairwise relations of objects precisely, forming more meaningful scene layouts (for example, the glasses/hat accurately lies on the eyes/head of the person with varying scales and flips), validating the superiority of the proposed wireframe rendering strategy. In addition, we conduct a user study involving 20 participants for a subjective evaluation of the results. Specifically, we randomly sample 30 abstract scenes synthesized by each of the above three models. A subject is asked to rate the synthetic scenes in terms of three scores (Excellent, Fair, Poor) by following three criteria: 1) whether they have coherent overall structures, 2) whether different objects are placed in reasonable relative positions, scales and flips, and 3) whether there are duplicate objects. We compute the rating percentage of all sampled scenes for each model. As shown in Table 3, the results of LayoutGAN with the wireframe rendering discriminator receive better ratings, which suggests that they are preferred by participants. + +![](images/e72b1bce01905f2f6eebea87be8c192d62b7b669e4b3971246872afda07ec9e0.jpg) +Figure 7: Optimizing perturbed tangram. + +![](images/58a966ad6f41348dd87f9734ec90a918cef932270ec4c5e5b18dd128e1566253.jpg) + +![](images/00c0fd1c77d7eebaffc11c0b612d9811737fc843209a51ca7cd17d2b896b2ceb.jpg) +Progressive Generated Layouts +Figure 8: Baseline results of tangram design. Real Layouts +Figure 9: Training progression of tangram graphic design generation (from left to right). + +# 4.4 TANGRAM GRAPHIC DESIGN + +A tangram is a geometric puzzle aiming to form specific shapes using seven pieces of 2D shapes without overlaps. The seven pieces include two large right triangles, one medium right triangle, two small right triangles, one square and one parallelogram, which can be assembled to form a square of side one unit and having area one square unit when choosing a unit of measurement. We collect 149 tangram graphic designs including animals, people and objects. In our experiments, we consider eight rotation/reflection poses for each piece. Given the seven pieces with each initialized by random location and ground-truth pose and class, the LayoutGAN with wireframe rendering discriminator is trained to refine their configurations jointly to form reasonable layouts. Note that it is a challenging task due to the complex configurations and limited data. We perform two experiments. The first one is a perturbation recovery test, in which we first add some random spatial perturbations to seven pieces of real layouts and train a LayoutGAN to recover the original layouts. Figure 7 shows that our network is able to pull back the displaced shapes to the right locations, confirming that the network successfully figures out the relations between different graphic elements. We further train another LayoutGAN to generate tangram designs from purely random initialization. In Figure 9 (high-resolution results can be found in the appendix), each row represents the sampled progressive generated results and the retrieved similar real tangrams. We also present some generated results of DCGAN and a sequential model similar to Wang et al. (2018a) for comparison. One can see from Figure 8 that DCGAN cannot well model the spatial relations among different shapes while the sequential model suffers from accumulated errors. In contrast, the LayoutGAN can generate meaningful tangrams like fox and person, although others may be hard to interpret, as shown in the penultimate row. + +# 5 CONCLUSION + +In this paper, we have proposed a novel LayoutGAN for generating layouts of relational graphic elements. Different from traditional GANs that generate images in pixel-level, the LayoutGAN can directly output a set of relational graphic elements. A novel differentiable wireframe rendering layer was proposed to rasterize the generated graphic elements to wireframe images, making it feasible to leverage CNNs as discriminator for better layout optimization from visual domain. Future works include adding a content representation, such as text, icon and picture to each graphic element. + +# REFERENCES + +Mart´ın Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, et al. Tensorflow: a system for largescale machine learning. In Proceedings of the 12th USENIX conference on Operating Systems Design and Implementation, pp. 265–283, 2016. + +Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. arXiv preprint arXiv:1707.02392, 2017. + +Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. + +Zoya Bylinskii, Nam Wook Kim, Peter O’Donovan, Sami Alsheikh, Spandan Madan, Hanspeter Pfister, Fredo Durand, Bryan Russell, and Aaron Hertzmann. Learning visual importance for graphic designs and data visualizations. In Proceedings of the 30th Annual ACM Symposium on User Interface Software and Technology, pp. 57–69. ACM, 2017. + +R Qi Charles, Hao Su, Mo Kaichun, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 77–85, 2017. + +Biplab Deka, Zifeng Huang, Chad Franzen, Joshua Hibschman, Daniel Afergan, Yang Li, Jeffrey Nichols, and Ranjitha Kumar. Rico: A mobile app dataset for building data-driven design applications. In Proceedings of the 30th Annual ACM Symposium on User Interface Software and Technology, pp. 845–854. ACM, 2017. + +Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venugopalan, Kate Saenko, and Trevor Darrell. Long-term recurrent convolutional networks for visual recognition and description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2625–2634, 2015. + +Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, volume 2, pp. 6, 2017. + +Matthew Fisher, Daniel Ritchie, Manolis Savva, Thomas Funkhouser, and Pat Hanrahan. Examplebased synthesis of 3d object arrangements. ACM Transactions on Graphics (TOG), 31(6):135, 2012. + +Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint arXiv:1410.5401, 2014. + +Nathan Hurst, Wilmot Li, and Kim Marriott. Review of automatic document formatting. In Proceedings of the 9th ACM symposium on Document engineering, pp. 99–108. ACM, 2009. + +Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. In Advances in Neural Information Processing Systems, pp. 2017–2025, 2015. + +Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4565–4574, 2016. + +Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. + +Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. + +Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations, 2017. + +Paul Merrell, Eric Schkufza, Zeyang Li, Maneesh Agrawala, and Vladlen Koltun. Interactive furniture layout using interior design guidelines. ACM Transactions on Graphics (TOG), 30(4):87, 2011. + +Peter O’Donovan, Aseem Agarwala, and Aaron Hertzmann. Color compatibility from large datasets. In ACM Transactions on Graphics (TOG), volume 30, pp. 63. ACM, 2011. + +Peter O’Donovan, Janis L ¯ ¯ıbeks, Aseem Agarwala, and Aaron Hertzmann. Exploratory font selection using crowdsourced attributes. ACM Transactions on Graphics (TOG), 33(4):92, 2014. + +Peter O’Donovan, Aseem Agarwala, and Aaron Hertzmann. Designscape: Design with interactive layout suggestions. In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, pp. 1221–1224. ACM, 2015. + +Aaron van den Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, and Koray ¨ Kavukcuoglu. Conditional image generation with pixelcnn decoders. In Advances in Neural Information Processing Systems, pp. 4797–4805, 2016. + +Peter O’Donovan, Aseem Agarwala, and Aaron Hertzmann. Learning layouts for single-page graphic designs. IEEE Transactions on Visualization and Computer Graphics, 20(8):1200–1213, 2014. + +Xufang Pang, Ying Cao, Rynson WH Lau, and Antoni B Chan. Directing user attention via visual flow on web designs. ACM Transactions on Graphics (TOG), 35(6):240, 2016. + +Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. + +Scott Reed and Nando De Freitas. Neural programmer-interpreters. arXiv preprint arXiv:1511.06279, 2015. + +Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. In Advances in Neural Information Processing Systems, pp. 2234–2242, 2016. + +Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 190–198, 2017. + +Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems, pp. 3104–3112, 2014. + +Amanda Swearngin, Mira Dontcheva, Wilmot Li, Joel Brandt, Morgan Dixon, and Andrew J Ko. Rewire: Interface design assistance from examples. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, pp. 504. ACM, 2018. + +Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, ¨ Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. In 9th ISCA Speech Synthesis Workshop, pp. 125–125, 2016. + +Oriol Vinyals, Samy Bengio, and Manjunath Kudlur. Order matters: Sequence to sequence for sets. arXiv preprint arXiv:1511.06391, 2015. + +Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. In Advances In Neural Information Processing Systems, pp. 613–621, 2016. + +Kai Wang, Manolis Savva, Angel X Chang, and Daniel Ritchie. Deep convolutional priors for indoor scene synthesis. ACM Transactions on Graphics (TOG), 37(4):70, 2018a. + +Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, volume 1, pp. 4, 2018b. + +Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In Advances in Neural Information Processing Systems, pp. 82–90, 2016. +Xinchen Yan, Jimei Yang, Ersin Yumer, Yijie Guo, and Honglak Lee. Perspective transformer nets: Learning single-view 3d object reconstruction without 3d supervision. In Advances in Neural Information Processing Systems, pp. 1696–1704, 2016. +Xiao Yang, Ersin Yumer, Paul Asente, Mike Kraley, Daniel Kifer, and C Lee Giles. Learning to extract semantic structure from documents using multimodal fully convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 5315– 5324, 2017. +C Lawrence Zitnick and Devi Parikh. Bringing semantics into focus using visual abstraction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3009– 3016, 2013. +C Lawrence Zitnick, Ramakrishna Vedantam, and Devi Parikh. Adopting abstract images for semantic scene understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(4):627–638, 2016. + +# 6 APPENDIX + +# 6.1 REAL DOCUMENT PAGES + +For better understanding the document semantic layouts, we show some layout samples and their corresponding real document pages, as shown in Figure 1. + +![](images/48697f4442f26ffebf238f89d6c19fa0e7000dabe3381482f56c3911a0bd16cc.jpg) + +Figure 1: Visualization of document layout samples and their corresponding real document pages. + +6.2 HIGH-RESOLUTION SYNTHESIZED DOCUMENT LAYOUTS + +![](images/59a94604fbb9961f70c224562e3a9f201e0a8d28bcb1b717ef19d6b70562d4f9.jpg) +Figure 2: Document layout generation. + +# 6.3 HIGH-RESOLUTION SYNTHESIZED CLIPART ABSTRACT SCENES + +![](images/59a9b616d7637834ad1a983b6585eb88310fbbfa6b6c991135c7b94e28480936.jpg) +Figure 3: Clipart abstract scene generation. + +![](images/b00fe714a8c08369c14cb18a0e232ba0e9a8119cdf5a7ef43faa974305d6a709.jpg) +Figure 4: Tangram graphic design. + +# 6.5 DETAILED ARCHITECTURE DESIGN OF THE GENERATOR + +The generator takes as input a set of graphic elements with random class probabilities and geometric parameters sampled from Uniform and Gaussian distribution respectively. An encoder consisting of three fully connected layers first embeds the semantic and geometric parameters of each graphic element. Two cascaded relation modules implemented as self-attention module followed by a basic residual block (“bottleneck” building block) form one relation block. We cascade two such relation blocks (totally 4 relation modules) for contextual feature refinement. Finally, a decoder consisting of several fully connected layers followed by two heads (implemented as the fully connected layer) with sigmoid activation are used to decode the refined feature of each element back to class probability distributions and geometric parameters respectively. + +# 6.6 SYNTHESIZED MOBILE APP LAYOUT DESIGNS + +We further conduct a new experiment of mobile app layout design by using the RICO dataset (Deka et al., 2017). For better understanding the mobile app layouts, we first show some layout samples and their corresponding real mobile app screenshots, as shown in Figure 5. + +![](images/de4a5d5cb562c7b1454d3bf7d559d6a13f30ebf080f65f675ab122c778ffcda2.jpg) +Figure 5: Visualization of mobile app layout samples and their corresponding real mobile app screenshots. + +![](images/c2672c8eb692ba95e8eeb467cec04d726767fb2c2337614e94647e9c0b7aa8ff.jpg) +Figure 6 shows the results from LayoutGAN and the retrieved most-similar real layouts from the training set as references. Both the wireframe and the corresponding mask layouts are provided for better visualization. +Figure 6: Mobile app layout design. \ No newline at end of file diff --git a/parse/train/HJxB5sRcFQ/HJxB5sRcFQ_content_list.json b/parse/train/HJxB5sRcFQ/HJxB5sRcFQ_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1ad5585598fe3e32b6151a448db5065932accaa1 --- /dev/null +++ b/parse/train/HJxB5sRcFQ/HJxB5sRcFQ_content_list.json @@ -0,0 +1,1562 @@ +[ + { + "type": "text", + "text": "LAYOUTGAN: GENERATING GRAPHIC LAYOUTSWITH WIREFRAME DISCRIMINATORS", + "text_level": 1, + "bbox": [ + 176, + 98, + 823, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jianan $\\mathbf { L i } ^ { 1 }$ ∗, Jimei Yang2, Aaron Hertzmann2, Jianming Zhang2, Tingfa $\\mathbf { X } \\mathbf { u } ^ { 1 }$ ", + "bbox": [ + 183, + 169, + 774, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Beijing Institute of Technology 2. Adobe Research ", + "bbox": [ + 316, + 196, + 700, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{20090964,xutingfa}@bit.edu.cn, {jimyang,hertzman,jianmzha}@adobe.com ", + "bbox": [ + 176, + 224, + 839, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 276, + 544, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Layout is important for graphic design and scene generation. We propose a novel Generative Adversarial Network, called LayoutGAN, that synthesizes layouts by modeling geometric relations of different types of 2D elements. The generator of LayoutGAN takes as input a set of randomly-placed 2D graphic elements and uses self-attention modules to refine their labels and geometric parameters jointly to produce a realistic layout. Accurate alignment is critical for good layouts. We thus propose a novel differentiable wireframe rendering layer that maps the generated layout to a wireframe image, upon which a CNN-based discriminator is used to optimize the layouts in image space. We validate the effectiveness of LayoutGAN in various experiments including MNIST digit generation, document layout generation, clipart abstract scene generation and tangram graphic design. ", + "bbox": [ + 233, + 308, + 764, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 487, + 336, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Graphic design is an important visual communication tool in our modern world, encompassing everything from book covers to magazine layouts to web design. Whereas methods for generating realistic natural-looking images have made significant progress lately, particularly with Generative Adversarial Networks (GANs) (Karras et al., 2018), methods for creating designs are far more primitive. This is, in part, due to the difficulty of finding data representations suitable for learning. Graphic designs are normally composed of vector representations of primitive objects, such as polygons, curves and ellipses, instead of pixels laid on a regular lattice. The quality and content of a design depends on the presence of elements, their attributes, and their relations to other elements. The visual perception of design depends on the arrangement of these elements; misalignment of two elements of just a few millimeters can ruin the design. Training from images of designs using conventional GANs synthesizes the layouts in pixel space, and thus mixes up layout and its rendering, and thus would be unlikely to capture layout styles well. Modeling such highly-structured data using neural networks is of great interest as they usually represent human abstract knowledge about the visual world (Zitnick & Parikh, 2013; Song et al., 2017) and how this knowledge are expressed via documents and designs (Deka et al., 2017; Yang et al., 2017). ", + "bbox": [ + 173, + 512, + 825, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces LayoutGAN, a novel GAN which directly synthesizes a set of graphical elements in a design. In a given design problem, a fixed set of element classes (e.g., “title,” “figure”) is specified in advance. In our network, each element is represented by its class probabilities and its geometric parameters, i.e., bounding-box keypoints. The generator takes as input graphic elements with randomly-sampled class probabilities and geometric parameters, and arranges them in a design; the output is the refined class probabilities and geometric parameters of the design elements. The generator has the desirable property of being permutation-invariant: it will generate the same layout if we re-order the input elements. ", + "bbox": [ + 174, + 727, + 825, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose two kinds of discriminator networks for this structured data. The first is similar in structure to the generator: it operates directly on the class probabilities and geometric parameters of the elements. Though effective, it is not sensitive enough to misalignment and occlusion between elements. The second discriminator operates in the visual domain. Like a human viewer who judges a design by looking at its rasterized image, the relations among different elements can be evaluated well by mapping them to 2D layouts. Then Convolutional Neural Networks (CNNs) can be used for layout optimization as they are specialized in distinguishing visual patterns including but not limited to misalignment and occlusion. However, the key challenge is how to map the geometric parameters to pixel-level layouts differentiably. One approach would be to render the graphic elements into bitmap masks using Spatial Transformer Networks (Jaderberg et al., 2015). But we found that filled pixels within the design elements cause occlusions and are ineffective for back-propagation, for example, when a small polygon hides behind a larger one. We experimented with bitmap mask rendering but it was not successful. In this paper, we propose a novel differentiable wireframe rendering layer that rasterizes both synthesized and real structured data of graphic elements into wireframe images, upon which a standard CNN can be used to optimize the layout across the visual and the graphic domain. The wireframe rendering discriminator has several advantages. First, convolution layers are very good at extracting spatial patterns of images so that they are more sensitive to alignment. Second, the rendered wireframes make elements visible even when they overlap and thus the network is alleviated from inferring the occlusions that may occur in other renderings such as masks. ", + "bbox": [ + 174, + 845, + 823, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 104, + 825, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate the LayoutGAN for several different tasks, including a sanity test on MNIST digits, generating page layouts from labeled bounding boxes, generating clipart abstract scenes, tangram graphic design, and mobile app design layouts. In each case, our method successfully generates layouts respecting the types of elements and their relationships for the problem domain. ", + "bbox": [ + 174, + 332, + 823, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the LayoutGAN comprises the following contributions: 1. A Generator that directly synthesizes structured data, represented as a resolution-independent set of labeled graphic elements in a design. 2. A differentiable wireframe rendering layer which allows the Discriminator to judge alignment from discrete element arrangements. ", + "bbox": [ + 174, + 395, + 823, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 470, + 344, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Structured data generation. Convolutional networks have been shown successful for generating data in regular lattice, such as images (Radford et al., 2015), videos (Vondrick et al., 2016) and 3D volumes (Yan et al., 2016; Wu et al., 2016). When generating highly-structured data, such as text (Donahue et al., 2015) and programs (Reed & De Freitas, 2015), recurrent networks are often the first choice (Sutskever et al., 2014), especially equipped with attention (Bahdanau et al., 2014) and memory modules (Graves et al., 2014). Recently, researchers show that convolutional networks can be also used to synthesize sequences (Oord et al., 2016; van den Oord et al., 2016) using autoregressive models. However, in many cases, an object has no sequential order (Vinyals et al., 2015), but is a set of elements, e.g. point clouds. Fan et al. (2017) propose a point set generation network for synthesizing 3D point clouds of the object shape from a single images. It is further paired with a point set classification network (Charles et al., 2017) for auto-encoding 3D point clouds (Achlioptas et al., 2017). Our work extends the set representation to more general primitive objects, i.e., labeled polygons. Meanwhile, researchers also model the structured data of connected elements using graph convolutions (Kipf & Welling, 2017). ", + "bbox": [ + 174, + 498, + 825, + 693 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data-driven graphic design. Automating layout is a classic problem in graphic design (Hurst et al., 2009). O’Donovan et al. (2014) formulate an energy function by assembling various heuristic visual cues and design principles to optimize single-page layouts, and extend this to an interactive tool (O’Donovan et al., 2015). The model parameters are learned from a small number of example designs. Pang et al. (2016) optimize layout for desired gaze direction. Deka et al. (2017) collect a mobile app design database for harnessing data-driven applications and present preliminary results of learning similarities of pixel-level textural/non-textual masks for design search, but do not learn models from this data. Swearngin et al. (2018) propose an interactive system that converts example design screenshots to vector graphics for designers to re-use and edit. Bylinskii et al. (2017) analyze the visual importance of graphic designs and use saliency map as the driving force to assist retargeting and thumbnailing. Previous methods have learned models for other graphic design elements, such as fonts (O’Donovan et al., 2014) and colors (O’Donovan et al., 2011). These are orthogonal to the layout problem, and could be combined in future work. No previous method has learned to create design or layout from large datasets, and no previous work has applied GANs to layout. ", + "bbox": [ + 174, + 696, + 825, + 891 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D scene synthesis. Interior scene synthesis and furniture layout generation draws great interest in graphics community. Early approaches focus on optimization of hand-crafted design principles (Merrell et al., 2011) and learning statistical priors of pairwise object relationships (Fisher et al., 2012) due to limited data. Wang et al. (2018a) recently proposed a sequential decision making approach to indoor scene synthesis. In each step, a CNN is trained to predict either location or category of one object by looking at the rendered top-down views. This resembles our wireframe rendering discriminator, in the sense of using convolutions to capture spatial patterns of layouts. ", + "bbox": [ + 173, + 895, + 821, + 924 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6d7c4e44cbab0b386322b6d183c456233be917dfc8168c78c361d237ed97a2c0.jpg", + "image_caption": [ + "Figure 1: Overall architecture of LayoutGAN. The generator takes as input graphic elements with randomly sampled class probabilities and geometric parameters from Uniform and Gaussian distribution respectively. An encoder embeds the input and feeds them into the stacked relation module, which refines the embedded features of each element in a coordinative manner by considering its semantic and spatial relations with all the other elements. Finally, a decoder decodes the refined features back to class probabilities and geometric parameters. The wireframe rendering discriminator feeds the generated results to a differentiable wireframe rendering layer which raterizes the input graphic elements into 2D wireframe images, upon which a CNN is applied for layout optimization. " + ], + "image_footnote": [], + "bbox": [ + 173, + 88, + 823, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 400, + 825, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 LAYOUTGAN ", + "text_level": 1, + "bbox": [ + 174, + 484, + 321, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section describes our data and model representations. ", + "bbox": [ + 174, + 505, + 555, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 DESIGN REPRESENTATION ", + "text_level": 1, + "bbox": [ + 174, + 530, + 398, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our model, a graphic design is comprised of a set of $N$ primitive design elements $\\{ ( p _ { 1 } , \\pmb { \\theta _ { 1 } } ) , \\cdot \\cdot \\cdot , ( p _ { N } , \\pmb { \\theta } _ { N } ) \\}$ . Each element has a set of geometric parameters $\\pmb { \\theta }$ , and a vector of class probabilities $\\pmb { p }$ . The entries of these variables are problem-dependent. For example, document layouts include 6 classes, e.g. “title” and “picture” while clipart layouts include 6 classes, e.g. “boy” and “hat”. For 2D point set generation (in MNIST digits), $\\mathbf { \\bar { \\theta } } \\equiv [ \\mathbf { \\bar { \\boldsymbol { x } } } , y ]$ , representing the coordinates of each point; for bounding-box generation in document layout, $\\pmb \\theta \\equiv \\left[ x ^ { L } , y ^ { T } , x ^ { R } , y ^ { B } \\right]$ , representing the top-left and bottom-right coordinates of each bounding box; for layouts with scale and flip (clipart abstract scenes), $\\pmb \\theta \\equiv [ x , y , s , l ]$ , representing the center coordinates, scale and flip of each element. ", + "bbox": [ + 173, + 550, + 825, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 GENERATOR ARCHITECTURE ", + "text_level": 1, + "bbox": [ + 176, + 688, + 415, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the LayoutGAN, the Generator is a function $G ( z )$ that takes a layout as input, where $z =$ $\\{ ( p _ { 1 } , \\pmb { \\theta _ { 1 } } ) , \\cdot \\cdot \\cdot , ( p _ { N } , \\pmb { \\theta } _ { N } ) \\}$ consisting of initial graphic elements with randomly-sampled geometric parameters $\\theta _ { i }$ , and one-hot encoding of a randomly-sampled class $\\mathbf { \\mathit { p } } _ { i }$ . The Generator outputs a refined layout $G ( z ) = \\{ ( p _ { 1 } ^ { \\prime } , \\theta _ { 1 } ^ { \\prime } ) , \\cdot \\cdot \\cdot , \\bar { ( } p _ { N } ^ { \\prime } , \\theta _ { N } ^ { \\prime } ) \\}$ which is meant to resemble a real graphic design. Note that, unlike, conventional GANs where $_ z$ represents a low-dimensional latent variable, our $_ { z }$ represents initial random graphic layout that has the same structure as the real one. The Discriminator learns to capture the geometric relations among different types of elements for layout optimization from both the graphic and the visual domain. Next, we go into details of the generator and discriminator design. ", + "bbox": [ + 173, + 708, + 825, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Figure 1, the generator takes as input a set of graphic elements with random class probabilities and geometric parameters sampled from Uniform and Gaussian distribution respectively. An encoder consisting of a multilayer perceptron network (implemented as multiple fully connected layers) first embeds the class one-hot vectors and geometric parameters of each graphic element. The relation module, implemented as self-attention inspired by Wang et al. (2018b), is then used to embed the feature of each graphic element as a function of its spatial context, i.e., its relations with all the other elements in the design. Denote $f ( p _ { i } , \\theta _ { i } )$ as the embedded feature of the graphic element $i$ , its refined feature representation $f ^ { \\prime } ( p _ { i } , \\theta _ { i } )$ can be obtained through a contextual residual learning process, which is defined as: ", + "bbox": [ + 174, + 840, + 825, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 146 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/e3abe037053bdc2592f0cd9af25873546f3adaea0a1713a9d4cee41d5a5f7b09.jpg", + "text": "$$\nf ^ { \\prime } ( p _ { i } , \\theta _ { i } ) = W _ { r } \\frac { 1 } { N } \\sum _ { \\forall j \\neq i } H ( f ( p _ { i } , \\theta _ { i } ) , f ( p _ { j } , \\theta _ { j } ) ) U ( f ( p _ { j } , \\theta _ { j } ) ) + f ( p _ { i } , \\theta _ { i } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 245, + 148, + 751, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where an unary function $U$ computes a representation of the embedded feature $f ( p _ { j } , \\theta _ { j } )$ of element $j$ and a pairwise function $\\pmb { H }$ computes a scalar value representing the relation between elements $i$ and $j$ . Thus, all the other elements $j \\neq i$ contribute to the feature refinement of element $i$ by summing up their relations. The response is normalized by the total number of elements in the set, $N$ . The weight matrix $W _ { r }$ computes a linear embedding, producing the contextual residual to be added to $f ( p _ { i } , \\theta _ { i } )$ for feature refinement. In our experiments, we define $\\pmb { H }$ as a dot-product: ", + "bbox": [ + 173, + 189, + 825, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/29600091bf77d9f4f57eb3415ed8af20f2e144e9750f7f90c889f13fd2c5188c.jpg", + "text": "$$\n\\pmb { H } ( f ( p _ { i } , \\pmb { \\theta } _ { i } ) , f ( p _ { j } , \\pmb { \\theta } _ { j } ) ) = \\psi ( f ( p _ { i } , \\pmb { \\theta } _ { i } ) ) ^ { T } \\phi ( f ( p _ { j } , \\pmb { \\theta } _ { j } ) ) ,\n$$", + "text_format": "latex", + "bbox": [ + 310, + 277, + 686, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\psi ( f ( p _ { i } , \\pmb { \\theta } _ { i } ) ) = W _ { \\psi } f ( p _ { i } , \\pmb { \\theta } _ { i } )$ and $\\phi ( f ( p _ { j } , \\pmb { \\theta } _ { j } ) ) = W _ { \\phi } f ( p _ { j } , \\pmb { \\theta } _ { j } )$ are two linear embeddings. We stack 4 relation modules for feature refinement in our experiments. Finally, a decoder consisting of another multilayer perceptron network followed by two branches of fully connected layer with a sigmoid activation function is used to map the refined feature of each element back to class probabilities and geometric parameters respectively. Optionally, Non-Maximum Suppression (NMS) can be applied to remove duplicated elements. ", + "bbox": [ + 173, + 301, + 825, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 DISCRIMINATOR NETWORK ARCHITECTURES ", + "text_level": 1, + "bbox": [ + 174, + 398, + 526, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The discriminator aims to distinguish between synthesized and real layouts. We present two types of discriminators, one based on relation modules directly built upon layout parameters, and the other based on how the layouts look like via rendering. ", + "bbox": [ + 174, + 419, + 825, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 RELATION-BASED DISCRIMINATOR ", + "text_level": 1, + "bbox": [ + 174, + 470, + 468, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Relation-Based Discriminator takes as input a set of graphic elements represented by class probabilities and geometric parameters, and feeds them to an encoder consisting of a multilayer perceptron network for feature embedding $f ( p _ { i } , \\theta _ { i } )$ . It then extracts their global graphical relations $g ( r ( p _ { 1 } , \\theta _ { 1 } ) , \\cdot \\cdot \\cdot , r ( p _ { N } , \\theta _ { N } ) )$ where $r ( p _ { i } , \\theta _ { i } )$ represents a simplified relation module as in equation 1 by removing the shortcut connection, and $\\textbf { { g } }$ is a max-pooling function used in (Charles et al., 2017). Thus, the global relations among all the graphic elements can be modeled, upon which a classifier composed by a multilayer perception network is applied for real/fake prediction. ", + "bbox": [ + 173, + 488, + 825, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 WIREFRAME RENDERING DISCRIMINATOR ", + "text_level": 1, + "bbox": [ + 174, + 595, + 519, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Wireframe Rendering Discriminator exploits CNNs to classify layouts, in order to learn to classify the visual properties of a layout. The Discriminator consists of a wireframe rendering layer, which produces an output image $\\pmb { I }$ , which is then fed into a CNN for classification. ", + "bbox": [ + 176, + 613, + 825, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The rasterization is performed as follows. Suppose there are $N$ elements in the design, with parameters $\\{ ( p _ { 1 } , \\pmb { \\theta _ { 1 } } ) , . . . , ( p _ { N } , \\pmb { \\theta } _ { N } ) \\}$ . Each element can be rendered into its own grayscale image $F _ { \\theta } ( x , y )$ ; rendering details for specific types of elements are given below. The image dimensions for each individual $\\pmb { F }$ are $W \\times H$ , where $W$ and $H$ are the width and height of the design in pixels. ", + "bbox": [ + 174, + 662, + 825, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The layer output is a multi-channel image $\\pmb { I }$ of dimensions $W \\times H \\times M$ , where each channel corresponds to one of the $M$ element types. In other words, pixel $( x , y )$ of $\\pmb { I }$ is a class activation vector for that pixel, and is computed as: ", + "bbox": [ + 174, + 724, + 825, + 767 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/9eec4be5eb257463fad7b0f98f3de4468809c4c3a85e5171528437f211320921.jpg", + "text": "$$\n\\pmb { I } ( x , y , c ) = \\operatorname* { m a x } _ { i \\in [ 1 . . . N ] } p _ { i , c } \\pmb { F } _ { \\theta _ { i } } ( x , y )\n$$", + "text_format": "latex", + "bbox": [ + 385, + 772, + 612, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p _ { i , c }$ is the probability that element $i$ is of class $c$ ", + "bbox": [ + 174, + 801, + 534, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We next describe the rendering process for computing $F _ { \\theta }$ , in the cases where $\\pmb \\theta$ represents a point, a rectangle, and a triangle. ", + "bbox": [ + 174, + 821, + 825, + 852 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Point. We start with the simplest geometric form, a single keypoint $\\theta _ { i } = ( x _ { i } , y _ { i } )$ for element $i$ . We implement an interpolation kernel $k$ for its rasterization. Its spatial rendering response on $( x , y )$ in the rendered image can be written as: ", + "bbox": [ + 174, + 857, + 825, + 900 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/1b692adce297264250ec4674e971f0cb85cbdd64c0e350bb4e78beaa40e5daa7.jpg", + "text": "$$\nF _ { \\theta _ { i } } ( x , y ) = k ( x - x _ { i } ) k ( y - y _ { i } ) .\n$$", + "text_format": "latex", + "bbox": [ + 388, + 905, + 611, + 922 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ff33a8ebcae43f98ddef388c3817179178471890959278d5168cd6dac657d1d0.jpg", + "image_caption": [ + "Figure 2: Wireframe rendering of different polygons (point, rectangle and triangle). The black grids represent grids of target image. The orange dots/dotted lines represent the graphic element mapped onto the image grid. The blue solid lines represent the rasterized wireframes expressed as differentiable functions of graphic elements in terms of both class probilities and geometric parameters. " + ], + "image_footnote": [], + "bbox": [ + 233, + 89, + 764, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt bilinear interpretation (Johnson et al., 2016), corresponding to the kernel $\\boldsymbol k ( d ) \\_ =$ $\\operatorname* { m a x } ( 0 , \\bar { 1 } - | d | )$ (implemented as ReLU activation), as shown in Figure 2. As $\\pmb { I }$ is a differentiable function of the class probabilities and the coordinates, the sub-gradients of the rasterized image can be thus propagated backward to them. We validate such rendering design for MNIST digit generation, detailed experiments can be seen in Section 4.1. ", + "bbox": [ + 173, + 301, + 825, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rectangle. We now consider more complex polygons. Assuming an element is a rectangle, or bounding box represented by its top-left and bottom-right coordinates $\\pmb { \\theta } = ( x ^ { L } , y ^ { T } , x ^ { R } , y ^ { B } )$ , which is very common in various designs. Specifically, considering a rectangle $_ i$ with coordinates $\\theta _ { i } = $ $( x _ { i } ^ { L } , y _ { i } ^ { T } , x _ { i } ^ { R } , y _ { i } ^ { B } )$ , as shown in Figure 2, the black grids represent the locations in the rendered image and the orange dotted box represents the rectangle being rasterized in the rendered image. For a wireframe representation, only the points near the boundary of the dotted box (lie in blue solid line) are related to the rectangle, so its spatial rendering response on $( x , y )$ can be formulated as: ", + "bbox": [ + 173, + 378, + 826, + 477 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/f34ab105e89f71d89378e54a8dbe7c0957ffbd4b1af58bd0735805d869c375c6.jpg", + "text": "$$\n\\begin{array} { r } { F _ { \\theta _ { i } } ( x , y ) = \\operatorname* { m a x } \\left( \\begin{array} { l } { k ( x - x _ { i } ^ { L } ) b ( y - y _ { i } ^ { T } ) b ( y _ { i } ^ { B } - y ) , } \\\\ { k ( x - x _ { i } ^ { R } ) b ( y - y _ { i } ^ { T } ) b ( y _ { i } ^ { B } - y ) , } \\\\ { k ( y - y _ { i } ^ { T } ) b ( x - x _ { i } ^ { L } ) b ( x _ { i } ^ { R } - x ) , } \\\\ { k ( y - y _ { i } ^ { B } ) b ( x - x _ { i } ^ { L } ) b ( x _ { i } ^ { R } - x ) } \\end{array} \\right) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 315, + 482, + 679, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $b ( d ) = \\operatorname* { m i n } ( \\operatorname* { m a x } ( 0 , d ) , 1 )$ constraining the rendering to nearby pixels. ", + "bbox": [ + 174, + 553, + 679, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Triangle. We further describe the wireframe rendering process of another geometric form, triangle. For triangle $i$ represented by its three vertices’ coordinates $\\pmb { \\theta _ { i } } = ( x _ { i } ^ { 1 } , y _ { i } ^ { 1 } , x _ { i } ^ { \\top } , y _ { i } ^ { 2 } , x _ { i } ^ { 3 } , y _ { i } ^ { 3 } )$ , when $x _ { i } ^ { \\bar { 1 } } \\neq$ $x _ { i } ^ { 2 } \\neq x _ { i } ^ { 3 }$ , its spatial rendering response on $( x , y )$ in the rendered image can be calculated as: ", + "bbox": [ + 176, + 574, + 825, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/f522d3921ce50be5c1f255681b4d230a0d5ee05230cbbf8ecfa7095a00d83e34.jpg", + "text": "$$\n\\begin{array} { r } { F _ { \\theta _ { i } } ( x , y ) = \\operatorname* { m a x } \\left( \\begin{array} { l } { k ( y - \\frac { ( y _ { i } ^ { 2 } - y _ { i } ^ { 1 } ) \\cdot ( x - x _ { i } ^ { 1 } ) } { x _ { i } ^ { 2 } - x _ { i } ^ { 1 } } - y _ { i } ^ { 1 } ) b ( x - x _ { i } ^ { 1 } ) b ( x _ { i } ^ { 2 } - x ) , } \\\\ { k ( y - \\frac { ( y _ { i } ^ { 3 } - y _ { i } ^ { 1 } ) \\cdot ( x - x _ { i } ^ { 1 } ) } { x _ { i } ^ { 3 } - x _ { i } ^ { 1 } } - y _ { i } ^ { 1 } ) b ( x - x _ { i } ^ { 3 } ) b ( x _ { i } ^ { 1 } - x ) , } \\\\ { k ( y - \\frac { ( y _ { i } ^ { 3 } - y _ { i } ^ { 2 } ) \\cdot ( x - x _ { i } ^ { 2 } ) } { x _ { i } ^ { 3 } - x _ { i } ^ { 2 } } - y _ { i } ^ { 2 } ) b ( x - x _ { i } ^ { 3 } ) b ( x _ { i } ^ { 2 } - x ) } \\end{array} \\right) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 261, + 623, + 735, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through this wireframe rendering process, gradients can be propagated backward to both the class probabilities and geometric parameters of the graphic elements for joint optimization. ", + "bbox": [ + 174, + 699, + 825, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A CNN consisting of 3 convolutional layers, followed by a fully connected layer with sigmoid activation, is applied to the rasterization layer $\\pmb { I }$ to predict real/fake graphic layouts. ", + "bbox": [ + 174, + 733, + 825, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 174, + 777, + 328, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The implementation is based on TensorFlow (Abadi et al., 2016). The network parameters are initialized from zero-mean Gaussian with standard deviation of 0.02. All the networks are optimized using Adam (Kingma & Ba, 2014) with a fixed learning rate of 0.00002. Detailed architectures can be found in the appendix. ", + "bbox": [ + 173, + 796, + 825, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 MNIST DIGIT GENERATION ", + "text_level": 1, + "bbox": [ + 174, + 864, + 408, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As a toy problem, we generate MNIST layouts. MNIST is a handwritten digit database consisting of 60,000 training and 10,000 testing images. For each image, we extract the locations of 128 randomly-selected foreground pixels as the graphic representation so that digit generation can be formulated as generating of 2D point layouts. In Figure 3a, each image shows $8 \\times 8$ digits rendered from point layouts. The left and middle images are the generated samples from the LayoutGAN with a relation-based discriminator and a wireframe rendering discriminator, respectively. The right image shows the digits rendered from ground-truth point layouts. One can see that the LayoutGAN with either discriminator captures various patterns. The wireframe rendering discriminator generates more compact, better-aligned point layouts. For quantitative evaluation, we first train a multilayer perceptron network for digit classification, which achieves an accuracy of $9 8 . 9 1 \\%$ on MNIST test set, and then use the classifier to compute the inception score (Salimans et al., 2016) of 10,000 generated samples. As shown in Table 1, the LayoutGAN with wireframe rendering discriminator achieves a higher inception score than with the relation-based discriminator. To shed light on the relational refinement process of LayoutGAN, we mark each input random point with a locationspecific color, and track their refined locations in the generated layouts, as shown in Figure 3b. One can see that colors change gradually along the strokes, showing the network learns some contextually consistent local displacements of the points. ", + "bbox": [ + 174, + 882, + 825, + 924 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a10d4ec280b2dd747078d880f5fbce6b8ddfb30566943603c69b4dd1b57b5bca.jpg", + "image_caption": [ + "Figure 3: Results on MNIST digit generation. " + ], + "image_footnote": [], + "bbox": [ + 186, + 87, + 797, + 210 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8f315d45457c736ab4b69f9651c7c50eab97d2d6ccdb636e41032de59d067050.jpg", + "table_caption": [ + "Table 1: Inception scores for generated digits. " + ], + "table_footnote": [], + "table_body": "
MethodsScore± std
Relation6.53 ± .09
Wireframe7.36 ± .07
Real data9.81 ± .08
", + "bbox": [ + 241, + 261, + 433, + 333 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9c0dce76805dbad046dd504eb714599cc73ff533aab11e371b5a5c722f5b98cd.jpg", + "table_caption": [ + "Table 2: Spatial analysis of document layout. " + ], + "table_footnote": [], + "table_body": "
MethodsOverlap(%)Alignment(%)
Relation1.526.4
Wireframe1.173.4
Real data0.050.5
", + "bbox": [ + 508, + 261, + 810, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 343, + 825, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 DOCUMENT LAYOUT GENERATION ", + "text_level": 1, + "bbox": [ + 176, + 545, + 452, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "One document page consists of a number of regions with different element types, such as heading, paragraph, table, figure, caption and list. Each region is represented by a bounding box. Modeling layouts of regions is critical for document analysis, retargeting, and synthesis. In real document data, these bounding boxes are often carefully aligned along the canonical axes, and their placement follows some particular patterns, such as heading always appearing above paragraph or table. Some example layouts are shown in the appendix. ", + "bbox": [ + 174, + 563, + 825, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this experiment, we focus on one-column layouts with no more than 9 bounding boxes that may belong to 6 possible classes as mentioned above. We are interested in how these layout patterns can be captured by our network. For training data, we collect totally around 25,000 layouts from real documents. We also implemented a baseline method that represents the layout by semantic masks as used in Yang et al. (2017); Deka et al. (2017). Specifically, we render all elements into masks and train a DCGAN (Radford et al., 2015) to generate mask layouts in pixels. Then we extract the connected mask regions of each class and output enclosing bounding boxes of all regions. Figure 4 shows some representative generation results (each row consists of 6 samples; high-resolution results can be found in the Appendix). The first row shows the results from DCGAN. DCGAN mixes a layout and its rendering in the generation process. When convolutional layers in DCGAN fails to perfectly replicate the rendering process (which is usually the case), it generates fuzzy and noisy label maps, making the extracted bounding boxes less consistent and less accurate. The third row shows the results from LayoutGAN with wireframe rendering discriminator. We retrieve the most similar real layouts from the training set in the last row as references. It can be seen that LayoutGAN can well capture different document layout patterns with clear definitions of instance-level semantic bounding boxes. ", + "bbox": [ + 174, + 654, + 825, + 875 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the advantage of wireframe rendering discriminator, we compare the generated results with those from the LayoutGAN with relation-based discriminator. As shown in the second row of Figure 4, the latter can also capture different layout patterns but sometimes suffers from overlapping and misalignment issues. As the bounding boxes in real document layouts are either left- or centeraligned with no overlaps with each other (except for captions with figures or tables), we propose two metrics to quantitatively measure the quality of the generated layouts. The first one is overlap index, which is the percentage of total overlapping area among any two bounding boxes inside the whole page. The second one is alignment index which is calculated by finding the minimum standard deviation of either left or center coordinates of all the bounding boxes. Table 2 provides quantitative comparisons of real layouts and synthesized layouts from LayoutGAN with two different discriminators. One can see that LayoutGAN with wireframe rendering discriminator achieves lower value in both overlapping index and alignment index than the one with relation-based discriminator, validating the superiority of the proposed wireframe rendering method for layout generation. A similar conclusion can also be drawn by analyzing the loss functions. We add shift perturbations to the bounding boxes of real layouts and feed them to both discriminators to examine their loss behaviors. Figure 6 visualizes the loss landscape of both discriminators corresponding to different extent of shift perturbation. One can see that the loss surface w.r.t. shift perturbation of the wireframe rendering discriminator is much smoother than the one of relation-based discriminator. ", + "bbox": [ + 176, + 882, + 823, + 924 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1b7f0b9374988295ff6466770bc433d9b1b8193220fd8eca8a0aba65c44d495b.jpg", + "image_caption": [ + "Figure 4: Document layout comparison. " + ], + "image_footnote": [], + "bbox": [ + 178, + 92, + 493, + 310 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb1c35ca53cdf5b316f1c04cd6b7b8c2044d49e7dbdf44c72cdc26db71c5a679.jpg", + "image_caption": [ + "Figure 5: Clipart abstract scene generation. " + ], + "image_footnote": [], + "bbox": [ + 526, + 83, + 790, + 310 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 338, + 825, + 546 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c03606f88f5395b716af09a8b97c3528cd6c04cee92c55e211f4c586467d2a12.jpg", + "image_caption": [ + "Figure 6: Discriminator loss landscapes. " + ], + "image_footnote": [], + "bbox": [ + 200, + 553, + 437, + 636 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b7a61251481b5b2779dfd957179eec7373977ebea52cec2a995c07590f798c7c.jpg", + "table_caption": [ + "Table 3: User study results for Clipart abstract scenes. " + ], + "table_footnote": [], + "table_body": "
MethodsExcellent (%)Fair (%)Poor (%)
DCGAN6.738.854.5
Relation17.250.332.5
Wireframe37.348.014.7
", + "bbox": [ + 468, + 579, + 820, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 CLIPART ABSTRACT SCENE GENERATION ", + "text_level": 1, + "bbox": [ + 173, + 667, + 495, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now consider generating scenes composed of a set of specific clipart elements. We use the abstract scene dataset (Zitnick et al., 2016) including boy, girl, glasses, hat, sun, and tree elements. To synthesize a reasonable abstract scene, we first use LayoutGAN to generate the layout of inner scene elements by representing each of them as a labeled bounding box with normalized center coordinates, width and height, and flip indicator. Then, we render the corresponding clipart image of each scene element onto a background image according to the predicted position (center coordinates), scale (width and height) and flip, forming a synthesized abstract scene. It is a challenging task as accurate pairwise or higher-order relations of objects are required for synthesizing reasonable scenes, for example, the glasses/hat should be exactly on the eyes/head of a boy/girl with proper scale and orientation. In Figure 5, the first three rows shows the generated layouts and corresponding rendered scenes by DCGAN, LayoutGAN with relation-based discriminator and LayoutGAN with wireframe rendering discriminator respectively. The last row shows some samples rendered from ground-truth scene layouts. It can be seen that, compared to DCGAN and the LayoutGAN with relation-based discriminator, the one with wireframe rendering discriminator can capture pairwise relations of objects precisely, forming more meaningful scene layouts (for example, the glasses/hat accurately lies on the eyes/head of the person with varying scales and flips), validating the superiority of the proposed wireframe rendering strategy. In addition, we conduct a user study involving 20 participants for a subjective evaluation of the results. Specifically, we randomly sample 30 abstract scenes synthesized by each of the above three models. A subject is asked to rate the synthetic scenes in terms of three scores (Excellent, Fair, Poor) by following three criteria: 1) whether they have coherent overall structures, 2) whether different objects are placed in reasonable relative positions, scales and flips, and 3) whether there are duplicate objects. We compute the rating percentage of all sampled scenes for each model. As shown in Table 3, the results of LayoutGAN with the wireframe rendering discriminator receive better ratings, which suggests that they are preferred by participants. ", + "bbox": [ + 173, + 688, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e72b1bce01905f2f6eebea87be8c192d62b7b669e4b3971246872afda07ec9e0.jpg", + "image_caption": [ + "Figure 7: Optimizing perturbed tangram. " + ], + "image_footnote": [], + "bbox": [ + 173, + 88, + 460, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/58a966ad6f41348dd87f9734ec90a918cef932270ec4c5e5b18dd128e1566253.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 472, + 79, + 789, + 184 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/00c0fd1c77d7eebaffc11c0b612d9811737fc843209a51ca7cd17d2b896b2ceb.jpg", + "image_caption": [ + "Progressive Generated Layouts ", + "Figure 8: Baseline results of tangram design. Real Layouts ", + "Figure 9: Training progression of tangram graphic design generation (from left to right). " + ], + "image_footnote": [], + "bbox": [ + 212, + 218, + 779, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 409, + 825, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4 TANGRAM GRAPHIC DESIGN ", + "text_level": 1, + "bbox": [ + 176, + 512, + 410, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A tangram is a geometric puzzle aiming to form specific shapes using seven pieces of 2D shapes without overlaps. The seven pieces include two large right triangles, one medium right triangle, two small right triangles, one square and one parallelogram, which can be assembled to form a square of side one unit and having area one square unit when choosing a unit of measurement. We collect 149 tangram graphic designs including animals, people and objects. In our experiments, we consider eight rotation/reflection poses for each piece. Given the seven pieces with each initialized by random location and ground-truth pose and class, the LayoutGAN with wireframe rendering discriminator is trained to refine their configurations jointly to form reasonable layouts. Note that it is a challenging task due to the complex configurations and limited data. We perform two experiments. The first one is a perturbation recovery test, in which we first add some random spatial perturbations to seven pieces of real layouts and train a LayoutGAN to recover the original layouts. Figure 7 shows that our network is able to pull back the displaced shapes to the right locations, confirming that the network successfully figures out the relations between different graphic elements. We further train another LayoutGAN to generate tangram designs from purely random initialization. In Figure 9 (high-resolution results can be found in the appendix), each row represents the sampled progressive generated results and the retrieved similar real tangrams. We also present some generated results of DCGAN and a sequential model similar to Wang et al. (2018a) for comparison. One can see from Figure 8 that DCGAN cannot well model the spatial relations among different shapes while the sequential model suffers from accumulated errors. In contrast, the LayoutGAN can generate meaningful tangrams like fox and person, although others may be hard to interpret, as shown in the penultimate row. ", + "bbox": [ + 173, + 526, + 825, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 CONCLUSION ", + "text_level": 1, + "bbox": [ + 174, + 824, + 320, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have proposed a novel LayoutGAN for generating layouts of relational graphic elements. Different from traditional GANs that generate images in pixel-level, the LayoutGAN can directly output a set of relational graphic elements. A novel differentiable wireframe rendering layer was proposed to rasterize the generated graphic elements to wireframe images, making it feasible to leverage CNNs as discriminator for better layout optimization from visual domain. Future works include adding a content representation, such as text, icon and picture to each graphic element. ", + "bbox": [ + 174, + 840, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 103, + 285, + 117 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Mart´ın Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, et al. Tensorflow: a system for largescale machine learning. In Proceedings of the 12th USENIX conference on Operating Systems Design and Implementation, pp. 265–283, 2016. ", + "bbox": [ + 173, + 126, + 825, + 183 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. arXiv preprint arXiv:1707.02392, 2017. ", + "bbox": [ + 171, + 193, + 823, + 222 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. ", + "bbox": [ + 174, + 232, + 821, + 261 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Zoya Bylinskii, Nam Wook Kim, Peter O’Donovan, Sami Alsheikh, Spandan Madan, Hanspeter Pfister, Fredo Durand, Bryan Russell, and Aaron Hertzmann. Learning visual importance for graphic designs and data visualizations. In Proceedings of the 30th Annual ACM Symposium on User Interface Software and Technology, pp. 57–69. ACM, 2017. ", + "bbox": [ + 173, + 271, + 825, + 329 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "R Qi Charles, Hao Su, Mo Kaichun, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 77–85, 2017. ", + "bbox": [ + 174, + 338, + 823, + 382 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Biplab Deka, Zifeng Huang, Chad Franzen, Joshua Hibschman, Daniel Afergan, Yang Li, Jeffrey Nichols, and Ranjitha Kumar. Rico: A mobile app dataset for building data-driven design applications. In Proceedings of the 30th Annual ACM Symposium on User Interface Software and Technology, pp. 845–854. ACM, 2017. ", + "bbox": [ + 174, + 391, + 825, + 448 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venugopalan, Kate Saenko, and Trevor Darrell. Long-term recurrent convolutional networks for visual recognition and description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2625–2634, 2015. ", + "bbox": [ + 173, + 458, + 825, + 515 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, volume 2, pp. 6, 2017. ", + "bbox": [ + 173, + 525, + 823, + 569 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Matthew Fisher, Daniel Ritchie, Manolis Savva, Thomas Funkhouser, and Pat Hanrahan. Examplebased synthesis of 3d object arrangements. ACM Transactions on Graphics (TOG), 31(6):135, 2012. ", + "bbox": [ + 171, + 578, + 823, + 621 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint arXiv:1410.5401, 2014. ", + "bbox": [ + 171, + 632, + 823, + 661 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Nathan Hurst, Wilmot Li, and Kim Marriott. Review of automatic document formatting. In Proceedings of the 9th ACM symposium on Document engineering, pp. 99–108. ACM, 2009. ", + "bbox": [ + 173, + 670, + 821, + 700 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. In Advances in Neural Information Processing Systems, pp. 2017–2025, 2015. ", + "bbox": [ + 173, + 709, + 823, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4565–4574, 2016. ", + "bbox": [ + 173, + 750, + 825, + 792 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. ", + "bbox": [ + 174, + 803, + 823, + 844 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. ", + "bbox": [ + 173, + 856, + 823, + 885 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations, 2017. ", + "bbox": [ + 174, + 895, + 821, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Paul Merrell, Eric Schkufza, Zeyang Li, Maneesh Agrawala, and Vladlen Koltun. Interactive furniture layout using interior design guidelines. ACM Transactions on Graphics (TOG), 30(4):87, 2011. ", + "bbox": [ + 173, + 103, + 821, + 146 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Peter O’Donovan, Aseem Agarwala, and Aaron Hertzmann. Color compatibility from large datasets. In ACM Transactions on Graphics (TOG), volume 30, pp. 63. ACM, 2011. ", + "bbox": [ + 173, + 156, + 821, + 185 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Peter O’Donovan, Janis L ¯ ¯ıbeks, Aseem Agarwala, and Aaron Hertzmann. Exploratory font selection using crowdsourced attributes. ACM Transactions on Graphics (TOG), 33(4):92, 2014. ", + "bbox": [ + 173, + 195, + 823, + 224 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Peter O’Donovan, Aseem Agarwala, and Aaron Hertzmann. Designscape: Design with interactive layout suggestions. In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, pp. 1221–1224. ACM, 2015. ", + "bbox": [ + 174, + 234, + 825, + 277 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Aaron van den Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, and Koray ¨ Kavukcuoglu. Conditional image generation with pixelcnn decoders. In Advances in Neural Information Processing Systems, pp. 4797–4805, 2016. ", + "bbox": [ + 173, + 287, + 825, + 332 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Peter O’Donovan, Aseem Agarwala, and Aaron Hertzmann. Learning layouts for single-page graphic designs. IEEE Transactions on Visualization and Computer Graphics, 20(8):1200–1213, 2014. ", + "bbox": [ + 176, + 340, + 825, + 383 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Xufang Pang, Ying Cao, Rynson WH Lau, and Antoni B Chan. Directing user attention via visual flow on web designs. ACM Transactions on Graphics (TOG), 35(6):240, 2016. ", + "bbox": [ + 173, + 393, + 821, + 422 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. ", + "bbox": [ + 173, + 433, + 821, + 463 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Scott Reed and Nando De Freitas. Neural programmer-interpreters. arXiv preprint arXiv:1511.06279, 2015. ", + "bbox": [ + 173, + 472, + 825, + 501 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. In Advances in Neural Information Processing Systems, pp. 2234–2242, 2016. ", + "bbox": [ + 173, + 511, + 825, + 555 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 190–198, 2017. ", + "bbox": [ + 173, + 564, + 823, + 608 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems, pp. 3104–3112, 2014. ", + "bbox": [ + 169, + 617, + 823, + 647 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Amanda Swearngin, Mira Dontcheva, Wilmot Li, Joel Brandt, Morgan Dixon, and Andrew J Ko. Rewire: Interface design assistance from examples. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, pp. 504. ACM, 2018. ", + "bbox": [ + 176, + 656, + 823, + 700 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, ¨ Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. In 9th ISCA Speech Synthesis Workshop, pp. 125–125, 2016. ", + "bbox": [ + 174, + 710, + 825, + 753 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Oriol Vinyals, Samy Bengio, and Manjunath Kudlur. Order matters: Sequence to sequence for sets. arXiv preprint arXiv:1511.06391, 2015. ", + "bbox": [ + 174, + 763, + 823, + 792 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. In Advances In Neural Information Processing Systems, pp. 613–621, 2016. ", + "bbox": [ + 171, + 803, + 823, + 832 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kai Wang, Manolis Savva, Angel X Chang, and Daniel Ritchie. Deep convolutional priors for indoor scene synthesis. ACM Transactions on Graphics (TOG), 37(4):70, 2018a. ", + "bbox": [ + 171, + 842, + 823, + 872 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, volume 1, pp. 4, 2018b. ", + "bbox": [ + 174, + 882, + 825, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In Advances in Neural Information Processing Systems, pp. 82–90, 2016. \nXinchen Yan, Jimei Yang, Ersin Yumer, Yijie Guo, and Honglak Lee. Perspective transformer nets: Learning single-view 3d object reconstruction without 3d supervision. In Advances in Neural Information Processing Systems, pp. 1696–1704, 2016. \nXiao Yang, Ersin Yumer, Paul Asente, Mike Kraley, Daniel Kifer, and C Lee Giles. Learning to extract semantic structure from documents using multimodal fully convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 5315– 5324, 2017. \nC Lawrence Zitnick and Devi Parikh. Bringing semantics into focus using visual abstraction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3009– 3016, 2013. \nC Lawrence Zitnick, Ramakrishna Vedantam, and Devi Parikh. Adopting abstract images for semantic scene understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(4):627–638, 2016. ", + "bbox": [ + 171, + 102, + 826, + 366 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6 APPENDIX ", + "text_level": 1, + "bbox": [ + 174, + 102, + 294, + 117 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.1 REAL DOCUMENT PAGES ", + "text_level": 1, + "bbox": [ + 174, + 135, + 387, + 148 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For better understanding the document semantic layouts, we show some layout samples and their corresponding real document pages, as shown in Figure 1. ", + "bbox": [ + 174, + 160, + 825, + 189 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/48697f4442f26ffebf238f89d6c19fa0e7000dabe3381482f56c3911a0bd16cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 202, + 723, + 479 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Figure 1: Visualization of document layout samples and their corresponding real document pages. ", + "bbox": [ + 174, + 489, + 818, + 506 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.2 HIGH-RESOLUTION SYNTHESIZED DOCUMENT LAYOUTS ", + "bbox": [ + 173, + 517, + 604, + 531 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/59a94604fbb9961f70c224562e3a9f201e0a8d28bcb1b717ef19d6b70562d4f9.jpg", + "image_caption": [ + "Figure 2: Document layout generation. " + ], + "image_footnote": [], + "bbox": [ + 236, + 549, + 759, + 907 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.3 HIGH-RESOLUTION SYNTHESIZED CLIPART ABSTRACT SCENES ", + "text_level": 1, + "bbox": [ + 178, + 104, + 645, + 117 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/59a9b616d7637834ad1a983b6585eb88310fbbfa6b6c991135c7b94e28480936.jpg", + "image_caption": [ + "Figure 3: Clipart abstract scene generation. " + ], + "image_footnote": [], + "bbox": [ + 176, + 130, + 820, + 693 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b00fe714a8c08369c14cb18a0e232ba0e9a8119cdf5a7ef43faa974305d6a709.jpg", + "image_caption": [ + "Figure 4: Tangram graphic design. " + ], + "image_footnote": [], + "bbox": [ + 171, + 136, + 820, + 446 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6.5 DETAILED ARCHITECTURE DESIGN OF THE GENERATOR ", + "text_level": 1, + "bbox": [ + 173, + 481, + 599, + 494 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The generator takes as input a set of graphic elements with random class probabilities and geometric parameters sampled from Uniform and Gaussian distribution respectively. An encoder consisting of three fully connected layers first embeds the semantic and geometric parameters of each graphic element. Two cascaded relation modules implemented as self-attention module followed by a basic residual block (“bottleneck” building block) form one relation block. We cascade two such relation blocks (totally 4 relation modules) for contextual feature refinement. Finally, a decoder consisting of several fully connected layers followed by two heads (implemented as the fully connected layer) with sigmoid activation are used to decode the refined feature of each element back to class probability distributions and geometric parameters respectively. ", + "bbox": [ + 173, + 506, + 825, + 632 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6.6 SYNTHESIZED MOBILE APP LAYOUT DESIGNS ", + "text_level": 1, + "bbox": [ + 178, + 104, + 526, + 117 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We further conduct a new experiment of mobile app layout design by using the RICO dataset (Deka et al., 2017). For better understanding the mobile app layouts, we first show some layout samples and their corresponding real mobile app screenshots, as shown in Figure 5. ", + "bbox": [ + 174, + 128, + 825, + 172 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/de4a5d5cb562c7b1454d3bf7d559d6a13f30ebf080f65f675ab122c778ffcda2.jpg", + "image_caption": [ + "Figure 5: Visualization of mobile app layout samples and their corresponding real mobile app screenshots. " + ], + "image_footnote": [], + "bbox": [ + 174, + 186, + 820, + 568 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/c2672c8eb692ba95e8eeb467cec04d726767fb2c2337614e94647e9c0b7aa8ff.jpg", + "image_caption": [ + "Figure 6 shows the results from LayoutGAN and the retrieved most-similar real layouts from the training set as references. Both the wireframe and the corresponding mask layouts are provided for better visualization. ", + "Figure 6: Mobile app layout design. " + ], + "image_footnote": [], + "bbox": [ + 274, + 148, + 723, + 737 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/parse/train/HJxB5sRcFQ/HJxB5sRcFQ_middle.json b/parse/train/HJxB5sRcFQ/HJxB5sRcFQ_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..c5bf7494c958a36ca7a4e48c149c9857f993b77b --- /dev/null +++ b/parse/train/HJxB5sRcFQ/HJxB5sRcFQ_middle.json @@ -0,0 +1,33962 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 98 + ], + "score": 1.0, + "content": "LAYOUTGAN: GENERATING GRAPHIC LAYOUTS", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 380, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 380, + 117 + ], + "score": 1.0, + "content": "WITH WIREFRAME DISCRIMINATORS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 134, + 474, + 147 + ], + "lines": [ + { + "bbox": [ + 110, + 132, + 475, + 150 + ], + "spans": [ + { + "bbox": [ + 110, + 132, + 143, + 150 + ], + "score": 1.0, + "content": "Jianan", + "type": "text" + }, + { + "bbox": [ + 143, + 135, + 159, + 146 + ], + "score": 0.54, + "content": "\\mathbf { L i } ^ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 132, + 457, + 150 + ], + "score": 1.0, + "content": "∗, Jimei Yang2, Aaron Hertzmann2, Jianming Zhang2, Tingfa", + "type": "text" + }, + { + "bbox": [ + 457, + 135, + 475, + 146 + ], + "score": 0.26, + "content": "\\mathbf { X } \\mathbf { u } ^ { 1 }", + "type": "inline_equation" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 194, + 156, + 429, + 168 + ], + "lines": [ + { + "bbox": [ + 193, + 154, + 430, + 172 + ], + "spans": [ + { + "bbox": [ + 193, + 154, + 332, + 172 + ], + "score": 1.0, + "content": "1. Beijing Institute of Technology", + "type": "text" + }, + { + "bbox": [ + 349, + 156, + 430, + 169 + ], + "score": 1.0, + "content": "2. Adobe Research", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 108, + 178, + 514, + 191 + ], + "lines": [ + { + "bbox": [ + 105, + 177, + 516, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 516, + 193 + ], + "score": 1.0, + "content": "{20090964,xutingfa}@bit.edu.cn, {jimyang,hertzman,jianmzha}@adobe.com", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 278, + 219, + 333, + 231 + ], + "lines": [ + { + "bbox": [ + 277, + 219, + 335, + 232 + ], + "spans": [ + { + "bbox": [ + 277, + 219, + 335, + 232 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 244, + 468, + 365 + ], + "lines": [ + { + "bbox": [ + 141, + 244, + 470, + 257 + ], + "spans": [ + { + "bbox": [ + 141, + 244, + 470, + 257 + ], + "score": 1.0, + "content": "Layout is important for graphic design and scene generation. We propose a novel", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 254, + 469, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 254, + 469, + 268 + ], + "score": 1.0, + "content": "Generative Adversarial Network, called LayoutGAN, that synthesizes layouts by", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "score": 1.0, + "content": "modeling geometric relations of different types of 2D elements. The generator", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 277, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 277, + 470, + 290 + ], + "score": 1.0, + "content": "of LayoutGAN takes as input a set of randomly-placed 2D graphic elements and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 289, + 469, + 300 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 300 + ], + "score": 1.0, + "content": "uses self-attention modules to refine their labels and geometric parameters jointly", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 299, + 470, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 470, + 311 + ], + "score": 1.0, + "content": "to produce a realistic layout. Accurate alignment is critical for good layouts. We", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 310, + 469, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 310, + 469, + 323 + ], + "score": 1.0, + "content": "thus propose a novel differentiable wireframe rendering layer that maps the gener-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "score": 1.0, + "content": "ated layout to a wireframe image, upon which a CNN-based discriminator is used", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 333, + 469, + 344 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 469, + 344 + ], + "score": 1.0, + "content": "to optimize the layouts in image space. We validate the effectiveness of Layout-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 343, + 469, + 356 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 469, + 356 + ], + "score": 1.0, + "content": "GAN in various experiments including MNIST digit generation, document layout", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 353, + 435, + 367 + ], + "spans": [ + { + "bbox": [ + 141, + 353, + 435, + 367 + ], + "score": 1.0, + "content": "generation, clipart abstract scene generation and tangram graphic design.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 108, + 386, + 206, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 208, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 208, + 401 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 406, + 505, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 406, + 504, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 504, + 419 + ], + "score": 1.0, + "content": "Graphic design is an important visual communication tool in our modern world, encompassing ev-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "score": 1.0, + "content": "erything from book covers to magazine layouts to web design. Whereas methods for generating", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "score": 1.0, + "content": "realistic natural-looking images have made significant progress lately, particularly with Generative", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 438, + 504, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 504, + 452 + ], + "score": 1.0, + "content": "Adversarial Networks (GANs) (Karras et al., 2018), methods for creating designs are far more prim-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 448, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 505, + 464 + ], + "score": 1.0, + "content": "itive. This is, in part, due to the difficulty of finding data representations suitable for learning.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 460, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 475 + ], + "score": 1.0, + "content": "Graphic designs are normally composed of vector representations of primitive objects, such as poly-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 472, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 104, + 472, + 506, + 485 + ], + "score": 1.0, + "content": "gons, curves and ellipses, instead of pixels laid on a regular lattice. The quality and content of a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 483, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 495 + ], + "score": 1.0, + "content": "design depends on the presence of elements, their attributes, and their relations to other elements.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 493, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 506, + 507 + ], + "score": 1.0, + "content": "The visual perception of design depends on the arrangement of these elements; misalignment of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 504, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 519 + ], + "score": 1.0, + "content": "two elements of just a few millimeters can ruin the design. Training from images of designs using", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "conventional GANs synthesizes the layouts in pixel space, and thus mixes up layout and its render-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 527, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 539 + ], + "score": 1.0, + "content": "ing, and thus would be unlikely to capture layout styles well. Modeling such highly-structured data", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 538, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 505, + 551 + ], + "score": 1.0, + "content": "using neural networks is of great interest as they usually represent human abstract knowledge about", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 548, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 505, + 562 + ], + "score": 1.0, + "content": "the visual world (Zitnick & Parikh, 2013; Song et al., 2017) and how this knowledge are expressed", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 559, + 368, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 368, + 572 + ], + "score": 1.0, + "content": "via documents and designs (Deka et al., 2017; Yang et al., 2017).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 576, + 505, + 664 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 588 + ], + "score": 1.0, + "content": "This paper introduces LayoutGAN, a novel GAN which directly synthesizes a set of graphical ele-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "ments in a design. In a given design problem, a fixed set of element classes (e.g., “title,” “figure”)", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 599, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 610 + ], + "score": 1.0, + "content": "is specified in advance. In our network, each element is represented by its class probabilities and its", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 610, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 622 + ], + "score": 1.0, + "content": "geometric parameters, i.e., bounding-box keypoints. The generator takes as input graphic elements", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 619, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 634 + ], + "score": 1.0, + "content": "with randomly-sampled class probabilities and geometric parameters, and arranges them in a design;", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 631, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 645 + ], + "score": 1.0, + "content": "the output is the refined class probabilities and geometric parameters of the design elements. The", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 643, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 655 + ], + "score": 1.0, + "content": "generator has the desirable property of being permutation-invariant: it will generate the same layout", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 653, + 242, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 242, + 666 + ], + "score": 1.0, + "content": "if we re-order the input elements.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 670, + 504, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 670, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 506, + 682 + ], + "score": 1.0, + "content": "We propose two kinds of discriminator networks for this structured data. The first is similar in", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 681, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 506, + 693 + ], + "score": 1.0, + "content": "structure to the generator: it operates directly on the class probabilities and geometric parameters of", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 692, + 506, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 506, + 704 + ], + "score": 1.0, + "content": "the elements. Though effective, it is not sensitive enough to misalignment and occlusion between", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 702, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 506, + 716 + ], + "score": 1.0, + "content": "elements. The second discriminator operates in the visual domain. Like a human viewer who judges", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 721, + 340, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 720, + 340, + 733 + ], + "spans": [ + { + "bbox": [ + 118, + 720, + 340, + 733 + ], + "score": 1.0, + "content": "∗Work done during Jianan Li's internship at Adobe Research.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 98 + ], + "score": 1.0, + "content": "LAYOUTGAN: GENERATING GRAPHIC LAYOUTS", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 380, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 380, + 117 + ], + "score": 1.0, + "content": "WITH WIREFRAME DISCRIMINATORS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 134, + 474, + 147 + ], + "lines": [ + { + "bbox": [ + 110, + 132, + 475, + 150 + ], + "spans": [ + { + "bbox": [ + 110, + 132, + 143, + 150 + ], + "score": 1.0, + "content": "Jianan", + "type": "text" + }, + { + "bbox": [ + 143, + 135, + 159, + 146 + ], + "score": 0.54, + "content": "\\mathbf { L i } ^ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 132, + 457, + 150 + ], + "score": 1.0, + "content": "∗, Jimei Yang2, Aaron Hertzmann2, Jianming Zhang2, Tingfa", + "type": "text" + }, + { + "bbox": [ + 457, + 135, + 475, + 146 + ], + "score": 0.26, + "content": "\\mathbf { X } \\mathbf { u } ^ { 1 }", + "type": "inline_equation" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 110, + 132, + 475, + 150 + ] + }, + { + "type": "text", + "bbox": [ + 194, + 156, + 429, + 168 + ], + "lines": [ + { + "bbox": [ + 193, + 154, + 430, + 172 + ], + "spans": [ + { + "bbox": [ + 193, + 154, + 332, + 172 + ], + "score": 1.0, + "content": "1. Beijing Institute of Technology", + "type": "text" + }, + { + "bbox": [ + 349, + 156, + 430, + 169 + ], + "score": 1.0, + "content": "2. Adobe Research", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3, + "bbox_fs": [ + 193, + 154, + 430, + 172 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 178, + 514, + 191 + ], + "lines": [ + { + "bbox": [ + 105, + 177, + 516, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 516, + 193 + ], + "score": 1.0, + "content": "{20090964,xutingfa}@bit.edu.cn, {jimyang,hertzman,jianmzha}@adobe.com", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 177, + 516, + 193 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 219, + 333, + 231 + ], + "lines": [ + { + "bbox": [ + 277, + 219, + 335, + 232 + ], + "spans": [ + { + "bbox": [ + 277, + 219, + 335, + 232 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 244, + 468, + 365 + ], + "lines": [ + { + "bbox": [ + 141, + 244, + 470, + 257 + ], + "spans": [ + { + "bbox": [ + 141, + 244, + 470, + 257 + ], + "score": 1.0, + "content": "Layout is important for graphic design and scene generation. We propose a novel", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 254, + 469, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 254, + 469, + 268 + ], + "score": 1.0, + "content": "Generative Adversarial Network, called LayoutGAN, that synthesizes layouts by", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "score": 1.0, + "content": "modeling geometric relations of different types of 2D elements. The generator", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 277, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 277, + 470, + 290 + ], + "score": 1.0, + "content": "of LayoutGAN takes as input a set of randomly-placed 2D graphic elements and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 289, + 469, + 300 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 300 + ], + "score": 1.0, + "content": "uses self-attention modules to refine their labels and geometric parameters jointly", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 299, + 470, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 470, + 311 + ], + "score": 1.0, + "content": "to produce a realistic layout. Accurate alignment is critical for good layouts. We", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 310, + 469, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 310, + 469, + 323 + ], + "score": 1.0, + "content": "thus propose a novel differentiable wireframe rendering layer that maps the gener-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "score": 1.0, + "content": "ated layout to a wireframe image, upon which a CNN-based discriminator is used", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 333, + 469, + 344 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 469, + 344 + ], + "score": 1.0, + "content": "to optimize the layouts in image space. We validate the effectiveness of Layout-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 343, + 469, + 356 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 469, + 356 + ], + "score": 1.0, + "content": "GAN in various experiments including MNIST digit generation, document layout", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 353, + 435, + 367 + ], + "spans": [ + { + "bbox": [ + 141, + 353, + 435, + 367 + ], + "score": 1.0, + "content": "generation, clipart abstract scene generation and tangram graphic design.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 11, + "bbox_fs": [ + 141, + 244, + 470, + 367 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 386, + 206, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 208, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 208, + 401 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 406, + 505, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 406, + 504, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 504, + 419 + ], + "score": 1.0, + "content": "Graphic design is an important visual communication tool in our modern world, encompassing ev-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "score": 1.0, + "content": "erything from book covers to magazine layouts to web design. Whereas methods for generating", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "score": 1.0, + "content": "realistic natural-looking images have made significant progress lately, particularly with Generative", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 438, + 504, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 504, + 452 + ], + "score": 1.0, + "content": "Adversarial Networks (GANs) (Karras et al., 2018), methods for creating designs are far more prim-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 448, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 505, + 464 + ], + "score": 1.0, + "content": "itive. This is, in part, due to the difficulty of finding data representations suitable for learning.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 460, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 475 + ], + "score": 1.0, + "content": "Graphic designs are normally composed of vector representations of primitive objects, such as poly-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 472, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 104, + 472, + 506, + 485 + ], + "score": 1.0, + "content": "gons, curves and ellipses, instead of pixels laid on a regular lattice. The quality and content of a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 483, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 495 + ], + "score": 1.0, + "content": "design depends on the presence of elements, their attributes, and their relations to other elements.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 493, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 506, + 507 + ], + "score": 1.0, + "content": "The visual perception of design depends on the arrangement of these elements; misalignment of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 504, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 519 + ], + "score": 1.0, + "content": "two elements of just a few millimeters can ruin the design. Training from images of designs using", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "conventional GANs synthesizes the layouts in pixel space, and thus mixes up layout and its render-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 527, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 539 + ], + "score": 1.0, + "content": "ing, and thus would be unlikely to capture layout styles well. Modeling such highly-structured data", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 538, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 505, + 551 + ], + "score": 1.0, + "content": "using neural networks is of great interest as they usually represent human abstract knowledge about", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 548, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 505, + 562 + ], + "score": 1.0, + "content": "the visual world (Zitnick & Parikh, 2013; Song et al., 2017) and how this knowledge are expressed", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 559, + 368, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 368, + 572 + ], + "score": 1.0, + "content": "via documents and designs (Deka et al., 2017; Yang et al., 2017).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 25, + "bbox_fs": [ + 104, + 406, + 506, + 572 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 576, + 505, + 664 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 588 + ], + "score": 1.0, + "content": "This paper introduces LayoutGAN, a novel GAN which directly synthesizes a set of graphical ele-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "ments in a design. In a given design problem, a fixed set of element classes (e.g., “title,” “figure”)", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 599, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 610 + ], + "score": 1.0, + "content": "is specified in advance. In our network, each element is represented by its class probabilities and its", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 610, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 622 + ], + "score": 1.0, + "content": "geometric parameters, i.e., bounding-box keypoints. The generator takes as input graphic elements", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 619, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 634 + ], + "score": 1.0, + "content": "with randomly-sampled class probabilities and geometric parameters, and arranges them in a design;", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 631, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 645 + ], + "score": 1.0, + "content": "the output is the refined class probabilities and geometric parameters of the design elements. The", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 643, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 655 + ], + "score": 1.0, + "content": "generator has the desirable property of being permutation-invariant: it will generate the same layout", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 653, + 242, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 242, + 666 + ], + "score": 1.0, + "content": "if we re-order the input elements.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 577, + 505, + 666 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 670, + 504, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 670, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 506, + 682 + ], + "score": 1.0, + "content": "We propose two kinds of discriminator networks for this structured data. The first is similar in", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 681, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 506, + 693 + ], + "score": 1.0, + "content": "structure to the generator: it operates directly on the class probabilities and geometric parameters of", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 692, + 506, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 506, + 704 + ], + "score": 1.0, + "content": "the elements. Though effective, it is not sensitive enough to misalignment and occlusion between", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 702, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 506, + 716 + ], + "score": 1.0, + "content": "elements. The second discriminator operates in the visual domain. Like a human viewer who judges", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "a design by looking at its rasterized image, the relations among different elements can be evaluated", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "well by mapping them to 2D layouts. Then Convolutional Neural Networks (CNNs) can be used", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 506, + 118 + ], + "score": 1.0, + "content": "for layout optimization as they are specialized in distinguishing visual patterns including but not", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "limited to misalignment and occlusion. However, the key challenge is how to map the geometric pa-", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 505, + 139 + ], + "score": 1.0, + "content": "rameters to pixel-level layouts differentiably. One approach would be to render the graphic elements", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 138, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 149 + ], + "score": 1.0, + "content": "into bitmap masks using Spatial Transformer Networks (Jaderberg et al., 2015). But we found that", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 147, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 505, + 162 + ], + "score": 1.0, + "content": "filled pixels within the design elements cause occlusions and are ineffective for back-propagation,", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 160, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 505, + 171 + ], + "score": 1.0, + "content": "for example, when a small polygon hides behind a larger one. We experimented with bitmap mask", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 170, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 505, + 184 + ], + "score": 1.0, + "content": "rendering but it was not successful. In this paper, we propose a novel differentiable wireframe", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "rendering layer that rasterizes both synthesized and real structured data of graphic elements into", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 204 + ], + "score": 1.0, + "content": "wireframe images, upon which a standard CNN can be used to optimize the layout across the visual", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 216 + ], + "score": 1.0, + "content": "and the graphic domain. The wireframe rendering discriminator has several advantages. First, con-", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 214, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 226 + ], + "score": 1.0, + "content": "volution layers are very good at extracting spatial patterns of images so that they are more sensitive", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 225, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 237 + ], + "score": 1.0, + "content": "to alignment. Second, the rendered wireframes make elements visible even when they overlap and", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "score": 1.0, + "content": "thus the network is alleviated from inferring the occlusions that may occur in other renderings such", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 248, + 147, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 147, + 258 + ], + "score": 1.0, + "content": "as masks.", + "type": "text", + "cross_page": true + } + ], + "index": 15 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 670, + 506, + 716 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 83, + 505, + 258 + ], + "lines": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "a design by looking at its rasterized image, the relations among different elements can be evaluated", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "well by mapping them to 2D layouts. Then Convolutional Neural Networks (CNNs) can be used", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 506, + 118 + ], + "score": 1.0, + "content": "for layout optimization as they are specialized in distinguishing visual patterns including but not", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "limited to misalignment and occlusion. However, the key challenge is how to map the geometric pa-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 505, + 139 + ], + "score": 1.0, + "content": "rameters to pixel-level layouts differentiably. One approach would be to render the graphic elements", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 138, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 149 + ], + "score": 1.0, + "content": "into bitmap masks using Spatial Transformer Networks (Jaderberg et al., 2015). But we found that", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 147, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 505, + 162 + ], + "score": 1.0, + "content": "filled pixels within the design elements cause occlusions and are ineffective for back-propagation,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 160, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 505, + 171 + ], + "score": 1.0, + "content": "for example, when a small polygon hides behind a larger one. We experimented with bitmap mask", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 170, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 505, + 184 + ], + "score": 1.0, + "content": "rendering but it was not successful. In this paper, we propose a novel differentiable wireframe", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "rendering layer that rasterizes both synthesized and real structured data of graphic elements into", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 204 + ], + "score": 1.0, + "content": "wireframe images, upon which a standard CNN can be used to optimize the layout across the visual", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 216 + ], + "score": 1.0, + "content": "and the graphic domain. The wireframe rendering discriminator has several advantages. First, con-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 214, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 226 + ], + "score": 1.0, + "content": "volution layers are very good at extracting spatial patterns of images so that they are more sensitive", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 225, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 237 + ], + "score": 1.0, + "content": "to alignment. Second, the rendered wireframes make elements visible even when they overlap and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "score": 1.0, + "content": "thus the network is alleviated from inferring the occlusions that may occur in other renderings such", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 248, + 147, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 147, + 258 + ], + "score": 1.0, + "content": "as masks.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 504, + 308 + ], + "lines": [ + { + "bbox": [ + 106, + 263, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 505, + 276 + ], + "score": 1.0, + "content": "We evaluate the LayoutGAN for several different tasks, including a sanity test on MNIST digits,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "generating page layouts from labeled bounding boxes, generating clipart abstract scenes, tangram", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "score": 1.0, + "content": "graphic design, and mobile app design layouts. In each case, our method successfully generates", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 297, + 458, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 458, + 309 + ], + "score": 1.0, + "content": "layouts respecting the types of elements and their relationships for the problem domain.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 107, + 313, + 504, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 314, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 505, + 326 + ], + "score": 1.0, + "content": "In summary, the LayoutGAN comprises the following contributions: 1. A Generator that directly", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 325, + 506, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 506, + 337 + ], + "score": 1.0, + "content": "synthesizes structured data, represented as a resolution-independent set of labeled graphic elements", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "in a design. 2. A differentiable wireframe rendering layer which allows the Discriminator to judge", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 347, + 297, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 297, + 358 + ], + "score": 1.0, + "content": "alignment from discrete element arrangements.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "title", + "bbox": [ + 108, + 373, + 211, + 386 + ], + "lines": [ + { + "bbox": [ + 104, + 372, + 213, + 389 + ], + "spans": [ + { + "bbox": [ + 104, + 372, + 213, + 389 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 395, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 393, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 506, + 409 + ], + "score": 1.0, + "content": "Structured data generation. Convolutional networks have been shown successful for generating", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "data in regular lattice, such as images (Radford et al., 2015), videos (Vondrick et al., 2016) and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "score": 1.0, + "content": "3D volumes (Yan et al., 2016; Wu et al., 2016). When generating highly-structured data, such as", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "text (Donahue et al., 2015) and programs (Reed & De Freitas, 2015), recurrent networks are often", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "the first choice (Sutskever et al., 2014), especially equipped with attention (Bahdanau et al., 2014)", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 450, + 504, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 504, + 462 + ], + "score": 1.0, + "content": "and memory modules (Graves et al., 2014). Recently, researchers show that convolutional networks", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 460, + 504, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 504, + 473 + ], + "score": 1.0, + "content": "can be also used to synthesize sequences (Oord et al., 2016; van den Oord et al., 2016) using auto-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 484 + ], + "score": 1.0, + "content": "regressive models. However, in many cases, an object has no sequential order (Vinyals et al., 2015),", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "score": 1.0, + "content": "but is a set of elements, e.g. point clouds. Fan et al. (2017) propose a point set generation network", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 494, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 506, + 507 + ], + "score": 1.0, + "content": "for synthesizing 3D point clouds of the object shape from a single images. It is further paired with a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 504, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 506, + 518 + ], + "score": 1.0, + "content": "point set classification network (Charles et al., 2017) for auto-encoding 3D point clouds (Achlioptas", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "et al., 2017). Our work extends the set representation to more general primitive objects, i.e., labeled", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "polygons. Meanwhile, researchers also model the structured data of connected elements using graph", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 538, + 258, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 258, + 550 + ], + "score": 1.0, + "content": "convolutions (Kipf & Welling, 2017).", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 552, + 505, + 706 + ], + "lines": [ + { + "bbox": [ + 106, + 553, + 504, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 504, + 564 + ], + "score": 1.0, + "content": "Data-driven graphic design. Automating layout is a classic problem in graphic design (Hurst et al.,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "2009). O’Donovan et al. (2014) formulate an energy function by assembling various heuristic vi-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "score": 1.0, + "content": "sual cues and design principles to optimize single-page layouts, and extend this to an interactive", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 586, + 504, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 504, + 597 + ], + "score": 1.0, + "content": "tool (O’Donovan et al., 2015). The model parameters are learned from a small number of example", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 595, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 610 + ], + "score": 1.0, + "content": "designs. Pang et al. (2016) optimize layout for desired gaze direction. Deka et al. (2017) collect a", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "mobile app design database for harnessing data-driven applications and present preliminary results", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 619, + 504, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 504, + 630 + ], + "score": 1.0, + "content": "of learning similarities of pixel-level textural/non-textual masks for design search, but do not learn", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "models from this data. Swearngin et al. (2018) propose an interactive system that converts example", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 639, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 505, + 654 + ], + "score": 1.0, + "content": "design screenshots to vector graphics for designers to re-use and edit. Bylinskii et al. (2017) analyze", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 650, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 664 + ], + "score": 1.0, + "content": "the visual importance of graphic designs and use saliency map as the driving force to assist retar-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 662, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 506, + 675 + ], + "score": 1.0, + "content": "geting and thumbnailing. Previous methods have learned models for other graphic design elements,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 673, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 505, + 685 + ], + "score": 1.0, + "content": "such as fonts (O’Donovan et al., 2014) and colors (O’Donovan et al., 2011). These are orthogonal", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 685, + 505, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 505, + 696 + ], + "score": 1.0, + "content": "to the layout problem, and could be combined in future work. No previous method has learned to", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 695, + 484, + 708 + ], + "spans": [ + { + "bbox": [ + 105, + 695, + 484, + 708 + ], + "score": 1.0, + "content": "create design or layout from large datasets, and no previous work has applied GANs to layout.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 45.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "3D scene synthesis. Interior scene synthesis and furniture layout generation draws great interest", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "in graphics community. Early approaches focus on optimization of hand-crafted design princi-", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 53.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 13, + "width": 8 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 83, + 505, + 258 + ], + "lines": [], + "index": 7.5, + "bbox_fs": [ + 105, + 83, + 506, + 258 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 504, + 308 + ], + "lines": [ + { + "bbox": [ + 106, + 263, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 505, + 276 + ], + "score": 1.0, + "content": "We evaluate the LayoutGAN for several different tasks, including a sanity test on MNIST digits,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "generating page layouts from labeled bounding boxes, generating clipart abstract scenes, tangram", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "score": 1.0, + "content": "graphic design, and mobile app design layouts. In each case, our method successfully generates", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 297, + 458, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 458, + 309 + ], + "score": 1.0, + "content": "layouts respecting the types of elements and their relationships for the problem domain.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 263, + 506, + 309 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 313, + 504, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 314, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 505, + 326 + ], + "score": 1.0, + "content": "In summary, the LayoutGAN comprises the following contributions: 1. A Generator that directly", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 325, + 506, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 506, + 337 + ], + "score": 1.0, + "content": "synthesizes structured data, represented as a resolution-independent set of labeled graphic elements", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "in a design. 2. A differentiable wireframe rendering layer which allows the Discriminator to judge", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 347, + 297, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 297, + 358 + ], + "score": 1.0, + "content": "alignment from discrete element arrangements.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 314, + 506, + 358 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 373, + 211, + 386 + ], + "lines": [ + { + "bbox": [ + 104, + 372, + 213, + 389 + ], + "spans": [ + { + "bbox": [ + 104, + 372, + 213, + 389 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 395, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 393, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 506, + 409 + ], + "score": 1.0, + "content": "Structured data generation. Convolutional networks have been shown successful for generating", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "data in regular lattice, such as images (Radford et al., 2015), videos (Vondrick et al., 2016) and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "score": 1.0, + "content": "3D volumes (Yan et al., 2016; Wu et al., 2016). When generating highly-structured data, such as", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "text (Donahue et al., 2015) and programs (Reed & De Freitas, 2015), recurrent networks are often", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "the first choice (Sutskever et al., 2014), especially equipped with attention (Bahdanau et al., 2014)", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 450, + 504, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 504, + 462 + ], + "score": 1.0, + "content": "and memory modules (Graves et al., 2014). Recently, researchers show that convolutional networks", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 460, + 504, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 504, + 473 + ], + "score": 1.0, + "content": "can be also used to synthesize sequences (Oord et al., 2016; van den Oord et al., 2016) using auto-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 484 + ], + "score": 1.0, + "content": "regressive models. However, in many cases, an object has no sequential order (Vinyals et al., 2015),", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "score": 1.0, + "content": "but is a set of elements, e.g. point clouds. Fan et al. (2017) propose a point set generation network", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 494, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 506, + 507 + ], + "score": 1.0, + "content": "for synthesizing 3D point clouds of the object shape from a single images. It is further paired with a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 504, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 506, + 518 + ], + "score": 1.0, + "content": "point set classification network (Charles et al., 2017) for auto-encoding 3D point clouds (Achlioptas", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "et al., 2017). Our work extends the set representation to more general primitive objects, i.e., labeled", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "polygons. Meanwhile, researchers also model the structured data of connected elements using graph", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 538, + 258, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 258, + 550 + ], + "score": 1.0, + "content": "convolutions (Kipf & Welling, 2017).", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 393, + 506, + 550 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 552, + 505, + 706 + ], + "lines": [ + { + "bbox": [ + 106, + 553, + 504, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 504, + 564 + ], + "score": 1.0, + "content": "Data-driven graphic design. Automating layout is a classic problem in graphic design (Hurst et al.,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "2009). O’Donovan et al. (2014) formulate an energy function by assembling various heuristic vi-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "score": 1.0, + "content": "sual cues and design principles to optimize single-page layouts, and extend this to an interactive", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 586, + 504, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 504, + 597 + ], + "score": 1.0, + "content": "tool (O’Donovan et al., 2015). The model parameters are learned from a small number of example", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 595, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 610 + ], + "score": 1.0, + "content": "designs. Pang et al. (2016) optimize layout for desired gaze direction. Deka et al. (2017) collect a", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "mobile app design database for harnessing data-driven applications and present preliminary results", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 619, + 504, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 504, + 630 + ], + "score": 1.0, + "content": "of learning similarities of pixel-level textural/non-textual masks for design search, but do not learn", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "models from this data. Swearngin et al. (2018) propose an interactive system that converts example", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 639, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 505, + 654 + ], + "score": 1.0, + "content": "design screenshots to vector graphics for designers to re-use and edit. Bylinskii et al. (2017) analyze", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 650, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 664 + ], + "score": 1.0, + "content": "the visual importance of graphic designs and use saliency map as the driving force to assist retar-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 662, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 506, + 675 + ], + "score": 1.0, + "content": "geting and thumbnailing. Previous methods have learned models for other graphic design elements,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 673, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 505, + 685 + ], + "score": 1.0, + "content": "such as fonts (O’Donovan et al., 2014) and colors (O’Donovan et al., 2011). These are orthogonal", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 685, + 505, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 505, + 696 + ], + "score": 1.0, + "content": "to the layout problem, and could be combined in future work. No previous method has learned to", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 695, + 484, + 708 + ], + "spans": [ + { + "bbox": [ + 105, + 695, + 484, + 708 + ], + "score": 1.0, + "content": "create design or layout from large datasets, and no previous work has applied GANs to layout.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 553, + 506, + 708 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "3D scene synthesis. Interior scene synthesis and furniture layout generation draws great interest", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "in graphics community. Early approaches focus on optimization of hand-crafted design princi-", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "ples (Merrell et al., 2011) and learning statistical priors of pairwise object relationships (Fisher", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 328, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 342 + ], + "score": 1.0, + "content": "et al., 2012) due to limited data. Wang et al. (2018a) recently proposed a sequential decision mak-", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 339, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 505, + 352 + ], + "score": 1.0, + "content": "ing approach to indoor scene synthesis. In each step, a CNN is trained to predict either location or", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 350, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 505, + 363 + ], + "score": 1.0, + "content": "category of one object by looking at the rendered top-down views. This resembles our wireframe", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 361, + 491, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 491, + 374 + ], + "score": 1.0, + "content": "rendering discriminator, in the sense of using convolutions to capture spatial patterns of layouts.", + "type": "text", + "cross_page": true + } + ], + "index": 15 + } + ], + "index": 53.5, + "bbox_fs": [ + 105, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 70, + 504, + 212 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 70, + 504, + 212 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 70, + 504, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 504, + 212 + ], + "score": 0.968, + "type": "image", + "image_path": "6d7c4e44cbab0b386322b6d183c456233be917dfc8168c78c361d237ed97a2c0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 70, + 504, + 117.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 117.33333333333334, + 504, + 164.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 164.66666666666669, + 504, + 212.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 222, + 505, + 310 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 222, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 505, + 234 + ], + "score": 1.0, + "content": "Figure 1: Overall architecture of LayoutGAN. The generator takes as input graphic elements with", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 233, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 505, + 245 + ], + "score": 1.0, + "content": "randomly sampled class probabilities and geometric parameters from Uniform and Gaussian distri-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 505, + 257 + ], + "score": 1.0, + "content": "bution respectively. An encoder embeds the input and feeds them into the stacked relation module,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "score": 1.0, + "content": "which refines the embedded features of each element in a coordinative manner by considering its", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 266, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 505, + 278 + ], + "score": 1.0, + "content": "semantic and spatial relations with all the other elements. Finally, a decoder decodes the refined fea-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 277, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 506, + 289 + ], + "score": 1.0, + "content": "tures back to class probabilities and geometric parameters. The wireframe rendering discriminator", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 288, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 505, + 300 + ], + "score": 1.0, + "content": "feeds the generated results to a differentiable wireframe rendering layer which raterizes the input", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 299, + 504, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 504, + 312 + ], + "score": 1.0, + "content": "graphic elements into 2D wireframe images, upon which a CNN is applied for layout optimization.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5 + } + ], + "index": 3.75 + }, + { + "type": "text", + "bbox": [ + 107, + 317, + 505, + 372 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "ples (Merrell et al., 2011) and learning statistical priors of pairwise object relationships (Fisher", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 328, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 342 + ], + "score": 1.0, + "content": "et al., 2012) due to limited data. Wang et al. (2018a) recently proposed a sequential decision mak-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 339, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 505, + 352 + ], + "score": 1.0, + "content": "ing approach to indoor scene synthesis. In each step, a CNN is trained to predict either location or", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 350, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 505, + 363 + ], + "score": 1.0, + "content": "category of one object by looking at the rendered top-down views. This resembles our wireframe", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 361, + 491, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 491, + 374 + ], + "score": 1.0, + "content": "rendering discriminator, in the sense of using convolutions to capture spatial patterns of layouts.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 107, + 384, + 197, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 383, + 198, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 198, + 399 + ], + "score": 1.0, + "content": "3 LAYOUTGAN", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 340, + 411 + ], + "lines": [ + { + "bbox": [ + 106, + 398, + 342, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 342, + 413 + ], + "score": 1.0, + "content": "This section describes our data and model representations.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 107, + 420, + 244, + 431 + ], + "lines": [ + { + "bbox": [ + 106, + 420, + 245, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 245, + 433 + ], + "score": 1.0, + "content": "3.1 DESIGN REPRESENTATION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 436, + 505, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 435, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 376, + 448 + ], + "score": 1.0, + "content": "In our model, a graphic design is comprised of a set of", + "type": "text" + }, + { + "bbox": [ + 376, + 436, + 387, + 446 + ], + "score": 0.78, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 435, + 506, + 448 + ], + "score": 1.0, + "content": "primitive design elements", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 447, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 107, + 447, + 216, + 459 + ], + "score": 0.92, + "content": "\\{ ( p _ { 1 } , \\pmb { \\theta _ { 1 } } ) , \\cdot \\cdot \\cdot , ( p _ { N } , \\pmb { \\theta } _ { N } ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 447, + 426, + 460 + ], + "score": 1.0, + "content": ". Each element has a set of geometric parameters", + "type": "text" + }, + { + "bbox": [ + 427, + 448, + 434, + 457 + ], + "score": 0.68, + "content": "\\pmb { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 447, + 506, + 460 + ], + "score": 1.0, + "content": ", and a vector of", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 180, + 470 + ], + "score": 1.0, + "content": "class probabilities", + "type": "text" + }, + { + "bbox": [ + 181, + 460, + 188, + 469 + ], + "score": 0.77, + "content": "\\pmb { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 459, + 505, + 470 + ], + "score": 1.0, + "content": ". The entries of these variables are problem-dependent. For example, document", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 468, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 483 + ], + "score": 1.0, + "content": "layouts include 6 classes, e.g. “title” and “picture” while clipart layouts include 6 classes, e.g. “boy”", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 480, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 343, + 493 + ], + "score": 1.0, + "content": "and “hat”. For 2D point set generation (in MNIST digits),", + "type": "text" + }, + { + "bbox": [ + 343, + 480, + 385, + 492 + ], + "score": 0.92, + "content": "\\mathbf { \\bar { \\theta } } \\equiv [ \\mathbf { \\bar { \\boldsymbol { x } } } , y ]", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 480, + 505, + 493 + ], + "score": 1.0, + "content": ", representing the coordinates", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 103, + 488, + 507, + 507 + ], + "spans": [ + { + "bbox": [ + 103, + 488, + 367, + 507 + ], + "score": 1.0, + "content": "of each point; for bounding-box generation in document layout,", + "type": "text" + }, + { + "bbox": [ + 368, + 491, + 458, + 505 + ], + "score": 0.94, + "content": "\\pmb \\theta \\equiv \\left[ x ^ { L } , y ^ { T } , x ^ { R } , y ^ { B } \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 488, + 507, + 507 + ], + "score": 1.0, + "content": ", represent-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 502, + 504, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 504, + 515 + ], + "score": 1.0, + "content": "ing the top-left and bottom-right coordinates of each bounding box; for layouts with scale and flip", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 513, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 207, + 527 + ], + "score": 1.0, + "content": "(clipart abstract scenes),", + "type": "text" + }, + { + "bbox": [ + 207, + 514, + 266, + 526 + ], + "score": 0.93, + "content": "\\pmb \\theta \\equiv [ x , y , s , l ]", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 513, + 505, + 527 + ], + "score": 1.0, + "content": ", representing the center coordinates, scale and flip of each", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 525, + 143, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 143, + 537 + ], + "score": 1.0, + "content": "element.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 108, + 545, + 254, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 545, + 255, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 255, + 558 + ], + "score": 1.0, + "content": "3.2 GENERATOR ARCHITECTURE", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 561, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 308, + 575 + ], + "score": 1.0, + "content": "In the LayoutGAN, the Generator is a function", + "type": "text" + }, + { + "bbox": [ + 308, + 561, + 331, + 573 + ], + "score": 0.92, + "content": "G ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 560, + 483, + 575 + ], + "score": 1.0, + "content": "that takes a layout as input, where", + "type": "text" + }, + { + "bbox": [ + 483, + 563, + 505, + 573 + ], + "score": 0.81, + "content": "z =", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 572, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 107, + 572, + 216, + 585 + ], + "score": 0.89, + "content": "\\{ ( p _ { 1 } , \\pmb { \\theta _ { 1 } } ) , \\cdot \\cdot \\cdot , ( p _ { N } , \\pmb { \\theta } _ { N } ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 572, + 505, + 586 + ], + "score": 1.0, + "content": "consisting of initial graphic elements with randomly-sampled geomet-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 166, + 595 + ], + "score": 1.0, + "content": "ric parameters", + "type": "text" + }, + { + "bbox": [ + 167, + 584, + 177, + 594 + ], + "score": 0.85, + "content": "\\theta _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 582, + 388, + 595 + ], + "score": 1.0, + "content": ", and one-hot encoding of a randomly-sampled class", + "type": "text" + }, + { + "bbox": [ + 389, + 584, + 399, + 595 + ], + "score": 0.84, + "content": "\\mathbf { \\mathit { p } } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 582, + 506, + 595 + ], + "score": 1.0, + "content": ". The Generator outputs a", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 592, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 104, + 592, + 165, + 609 + ], + "score": 1.0, + "content": "refined layout", + "type": "text" + }, + { + "bbox": [ + 166, + 594, + 314, + 606 + ], + "score": 0.92, + "content": "G ( z ) = \\{ ( p _ { 1 } ^ { \\prime } , \\theta _ { 1 } ^ { \\prime } ) , \\cdot \\cdot \\cdot , \\bar { ( } p _ { N } ^ { \\prime } , \\theta _ { N } ^ { \\prime } ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 592, + 506, + 609 + ], + "score": 1.0, + "content": "which is meant to resemble a real graphic de-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 313, + 618 + ], + "score": 1.0, + "content": "sign. Note that, unlike, conventional GANs where", + "type": "text" + }, + { + "bbox": [ + 313, + 607, + 321, + 615 + ], + "score": 0.72, + "content": "_ z", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "represents a low-dimensional latent variable,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 104, + 616, + 122, + 628 + ], + "score": 1.0, + "content": "our", + "type": "text" + }, + { + "bbox": [ + 122, + 618, + 130, + 626 + ], + "score": 0.73, + "content": "_ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "represents initial random graphic layout that has the same structure as the real one. The Dis-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "criminator learns to capture the geometric relations among different types of elements for layout", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "optimization from both the graphic and the visual domain. Next, we go into details of the generator", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 648, + 210, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 210, + 662 + ], + "score": 1.0, + "content": "and discriminator design.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "As illustrated in Figure 1, the generator takes as input a set of graphic elements with random class", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "probabilities and geometric parameters sampled from Uniform and Gaussian distribution respec-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "tively. An encoder consisting of a multilayer perceptron network (implemented as multiple fully", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "connected layers) first embeds the class one-hot vectors and geometric parameters of each graphic", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "element. The relation module, implemented as self-attention inspired by Wang et al. (2018b), is", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "then used to embed the feature of each graphic element as a function of its spatial context, i.e., its", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 70, + 504, + 212 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 70, + 504, + 212 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 70, + 504, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 504, + 212 + ], + "score": 0.968, + "type": "image", + "image_path": "6d7c4e44cbab0b386322b6d183c456233be917dfc8168c78c361d237ed97a2c0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 70, + 504, + 117.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 117.33333333333334, + 504, + 164.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 164.66666666666669, + 504, + 212.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 222, + 505, + 310 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 222, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 505, + 234 + ], + "score": 1.0, + "content": "Figure 1: Overall architecture of LayoutGAN. The generator takes as input graphic elements with", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 233, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 505, + 245 + ], + "score": 1.0, + "content": "randomly sampled class probabilities and geometric parameters from Uniform and Gaussian distri-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 505, + 257 + ], + "score": 1.0, + "content": "bution respectively. An encoder embeds the input and feeds them into the stacked relation module,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "score": 1.0, + "content": "which refines the embedded features of each element in a coordinative manner by considering its", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 266, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 505, + 278 + ], + "score": 1.0, + "content": "semantic and spatial relations with all the other elements. Finally, a decoder decodes the refined fea-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 277, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 506, + 289 + ], + "score": 1.0, + "content": "tures back to class probabilities and geometric parameters. The wireframe rendering discriminator", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 288, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 505, + 300 + ], + "score": 1.0, + "content": "feeds the generated results to a differentiable wireframe rendering layer which raterizes the input", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 299, + 504, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 504, + 312 + ], + "score": 1.0, + "content": "graphic elements into 2D wireframe images, upon which a CNN is applied for layout optimization.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5 + } + ], + "index": 3.75 + }, + { + "type": "text", + "bbox": [ + 107, + 317, + 505, + 372 + ], + "lines": [], + "index": 13, + "bbox_fs": [ + 105, + 317, + 505, + 374 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 384, + 197, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 383, + 198, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 198, + 399 + ], + "score": 1.0, + "content": "3 LAYOUTGAN", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 340, + 411 + ], + "lines": [ + { + "bbox": [ + 106, + 398, + 342, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 342, + 413 + ], + "score": 1.0, + "content": "This section describes our data and model representations.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 398, + 342, + 413 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 420, + 244, + 431 + ], + "lines": [ + { + "bbox": [ + 106, + 420, + 245, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 245, + 433 + ], + "score": 1.0, + "content": "3.1 DESIGN REPRESENTATION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 436, + 505, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 435, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 376, + 448 + ], + "score": 1.0, + "content": "In our model, a graphic design is comprised of a set of", + "type": "text" + }, + { + "bbox": [ + 376, + 436, + 387, + 446 + ], + "score": 0.78, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 435, + 506, + 448 + ], + "score": 1.0, + "content": "primitive design elements", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 447, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 107, + 447, + 216, + 459 + ], + "score": 0.92, + "content": "\\{ ( p _ { 1 } , \\pmb { \\theta _ { 1 } } ) , \\cdot \\cdot \\cdot , ( p _ { N } , \\pmb { \\theta } _ { N } ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 447, + 426, + 460 + ], + "score": 1.0, + "content": ". Each element has a set of geometric parameters", + "type": "text" + }, + { + "bbox": [ + 427, + 448, + 434, + 457 + ], + "score": 0.68, + "content": "\\pmb { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 447, + 506, + 460 + ], + "score": 1.0, + "content": ", and a vector of", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 180, + 470 + ], + "score": 1.0, + "content": "class probabilities", + "type": "text" + }, + { + "bbox": [ + 181, + 460, + 188, + 469 + ], + "score": 0.77, + "content": "\\pmb { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 459, + 505, + 470 + ], + "score": 1.0, + "content": ". The entries of these variables are problem-dependent. For example, document", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 468, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 483 + ], + "score": 1.0, + "content": "layouts include 6 classes, e.g. “title” and “picture” while clipart layouts include 6 classes, e.g. “boy”", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 480, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 343, + 493 + ], + "score": 1.0, + "content": "and “hat”. For 2D point set generation (in MNIST digits),", + "type": "text" + }, + { + "bbox": [ + 343, + 480, + 385, + 492 + ], + "score": 0.92, + "content": "\\mathbf { \\bar { \\theta } } \\equiv [ \\mathbf { \\bar { \\boldsymbol { x } } } , y ]", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 480, + 505, + 493 + ], + "score": 1.0, + "content": ", representing the coordinates", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 103, + 488, + 507, + 507 + ], + "spans": [ + { + "bbox": [ + 103, + 488, + 367, + 507 + ], + "score": 1.0, + "content": "of each point; for bounding-box generation in document layout,", + "type": "text" + }, + { + "bbox": [ + 368, + 491, + 458, + 505 + ], + "score": 0.94, + "content": "\\pmb \\theta \\equiv \\left[ x ^ { L } , y ^ { T } , x ^ { R } , y ^ { B } \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 488, + 507, + 507 + ], + "score": 1.0, + "content": ", represent-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 502, + 504, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 504, + 515 + ], + "score": 1.0, + "content": "ing the top-left and bottom-right coordinates of each bounding box; for layouts with scale and flip", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 513, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 207, + 527 + ], + "score": 1.0, + "content": "(clipart abstract scenes),", + "type": "text" + }, + { + "bbox": [ + 207, + 514, + 266, + 526 + ], + "score": 0.93, + "content": "\\pmb \\theta \\equiv [ x , y , s , l ]", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 513, + 505, + 527 + ], + "score": 1.0, + "content": ", representing the center coordinates, scale and flip of each", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 525, + 143, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 143, + 537 + ], + "score": 1.0, + "content": "element.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23, + "bbox_fs": [ + 103, + 435, + 507, + 537 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 545, + 254, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 545, + 255, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 255, + 558 + ], + "score": 1.0, + "content": "3.2 GENERATOR ARCHITECTURE", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 561, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 308, + 575 + ], + "score": 1.0, + "content": "In the LayoutGAN, the Generator is a function", + "type": "text" + }, + { + "bbox": [ + 308, + 561, + 331, + 573 + ], + "score": 0.92, + "content": "G ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 560, + 483, + 575 + ], + "score": 1.0, + "content": "that takes a layout as input, where", + "type": "text" + }, + { + "bbox": [ + 483, + 563, + 505, + 573 + ], + "score": 0.81, + "content": "z =", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 572, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 107, + 572, + 216, + 585 + ], + "score": 0.89, + "content": "\\{ ( p _ { 1 } , \\pmb { \\theta _ { 1 } } ) , \\cdot \\cdot \\cdot , ( p _ { N } , \\pmb { \\theta } _ { N } ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 572, + 505, + 586 + ], + "score": 1.0, + "content": "consisting of initial graphic elements with randomly-sampled geomet-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 166, + 595 + ], + "score": 1.0, + "content": "ric parameters", + "type": "text" + }, + { + "bbox": [ + 167, + 584, + 177, + 594 + ], + "score": 0.85, + "content": "\\theta _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 582, + 388, + 595 + ], + "score": 1.0, + "content": ", and one-hot encoding of a randomly-sampled class", + "type": "text" + }, + { + "bbox": [ + 389, + 584, + 399, + 595 + ], + "score": 0.84, + "content": "\\mathbf { \\mathit { p } } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 582, + 506, + 595 + ], + "score": 1.0, + "content": ". The Generator outputs a", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 592, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 104, + 592, + 165, + 609 + ], + "score": 1.0, + "content": "refined layout", + "type": "text" + }, + { + "bbox": [ + 166, + 594, + 314, + 606 + ], + "score": 0.92, + "content": "G ( z ) = \\{ ( p _ { 1 } ^ { \\prime } , \\theta _ { 1 } ^ { \\prime } ) , \\cdot \\cdot \\cdot , \\bar { ( } p _ { N } ^ { \\prime } , \\theta _ { N } ^ { \\prime } ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 592, + 506, + 609 + ], + "score": 1.0, + "content": "which is meant to resemble a real graphic de-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 313, + 618 + ], + "score": 1.0, + "content": "sign. Note that, unlike, conventional GANs where", + "type": "text" + }, + { + "bbox": [ + 313, + 607, + 321, + 615 + ], + "score": 0.72, + "content": "_ z", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "represents a low-dimensional latent variable,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 104, + 616, + 122, + 628 + ], + "score": 1.0, + "content": "our", + "type": "text" + }, + { + "bbox": [ + 122, + 618, + 130, + 626 + ], + "score": 0.73, + "content": "_ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "represents initial random graphic layout that has the same structure as the real one. The Dis-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "criminator learns to capture the geometric relations among different types of elements for layout", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "optimization from both the graphic and the visual domain. Next, we go into details of the generator", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 648, + 210, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 210, + 662 + ], + "score": 1.0, + "content": "and discriminator design.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33, + "bbox_fs": [ + 104, + 560, + 506, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "As illustrated in Figure 1, the generator takes as input a set of graphic elements with random class", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "probabilities and geometric parameters sampled from Uniform and Gaussian distribution respec-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "tively. An encoder consisting of a multilayer perceptron network (implemented as multiple fully", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "connected layers) first embeds the class one-hot vectors and geometric parameters of each graphic", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "element. The relation module, implemented as self-attention inspired by Wang et al. (2018b), is", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "then used to embed the feature of each graphic element as a function of its spatial context, i.e., its", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 340, + 95 + ], + "score": 1.0, + "content": "relations with all the other elements in the design. Denote", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 340, + 82, + 379, + 95 + ], + "score": 0.93, + "content": "f ( p _ { i } , \\theta _ { i } )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 379, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "as the embedded feature of the", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 173, + 106 + ], + "score": 1.0, + "content": "graphic element", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 173, + 95, + 178, + 104 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 178, + 93, + 313, + 106 + ], + "score": 1.0, + "content": ", its refined feature representation", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 313, + 93, + 354, + 106 + ], + "score": 0.93, + "content": "f ^ { \\prime } ( p _ { i } , \\theta _ { i } )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 354, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "can be obtained through a contextual", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 293, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 293, + 117 + ], + "score": 1.0, + "content": "residual learning process, which is defined as:", + "type": "text", + "cross_page": true + } + ], + "index": 2 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 665, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 340, + 95 + ], + "score": 1.0, + "content": "relations with all the other elements in the design. Denote", + "type": "text" + }, + { + "bbox": [ + 340, + 82, + 379, + 95 + ], + "score": 0.93, + "content": "f ( p _ { i } , \\theta _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "as the embedded feature of the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 173, + 106 + ], + "score": 1.0, + "content": "graphic element", + "type": "text" + }, + { + "bbox": [ + 173, + 95, + 178, + 104 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 93, + 313, + 106 + ], + "score": 1.0, + "content": ", its refined feature representation", + "type": "text" + }, + { + "bbox": [ + 313, + 93, + 354, + 106 + ], + "score": 0.93, + "content": "f ^ { \\prime } ( p _ { i } , \\theta _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "can be obtained through a contextual", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 293, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 293, + 117 + ], + "score": 1.0, + "content": "residual learning process, which is defined as:", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "interline_equation", + "bbox": [ + 150, + 118, + 460, + 150 + ], + "lines": [ + { + "bbox": [ + 150, + 118, + 460, + 150 + ], + "spans": [ + { + "bbox": [ + 150, + 118, + 460, + 150 + ], + "score": 0.93, + "content": "f ^ { \\prime } ( p _ { i } , \\theta _ { i } ) = W _ { r } \\frac { 1 } { N } \\sum _ { \\forall j \\neq i } H ( f ( p _ { i } , \\theta _ { i } ) , f ( p _ { j } , \\theta _ { j } ) ) U ( f ( p _ { j } , \\theta _ { j } ) ) + f ( p _ { i } , \\theta _ { i } ) ,", + "type": "interline_equation", + "image_path": "e3abe037053bdc2592f0cd9af25873546f3adaea0a1713a9d4cee41d5a5f7b09.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 150, + 118, + 460, + 128.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 150, + 128.66666666666666, + 460, + 139.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 150, + 139.33333333333331, + 460, + 149.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 150, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 106, + 150, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 205, + 163 + ], + "score": 1.0, + "content": "where an unary function", + "type": "text" + }, + { + "bbox": [ + 205, + 151, + 215, + 160 + ], + "score": 0.82, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 150, + 420, + 163 + ], + "score": 1.0, + "content": "computes a representation of the embedded feature", + "type": "text" + }, + { + "bbox": [ + 420, + 150, + 459, + 162 + ], + "score": 0.93, + "content": "f ( p _ { j } , \\theta _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 150, + 506, + 163 + ], + "score": 1.0, + "content": "of element", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 161, + 504, + 173 + ], + "spans": [ + { + "bbox": [ + 107, + 162, + 113, + 173 + ], + "score": 0.8, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 161, + 212, + 173 + ], + "score": 1.0, + "content": "and a pairwise function", + "type": "text" + }, + { + "bbox": [ + 212, + 162, + 224, + 171 + ], + "score": 0.79, + "content": "\\pmb { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 161, + 499, + 173 + ], + "score": 1.0, + "content": "computes a scalar value representing the relation between elements", + "type": "text" + }, + { + "bbox": [ + 499, + 162, + 504, + 171 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 171, + 504, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 124, + 184 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 173, + 131, + 184 + ], + "score": 0.81, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 171, + 258, + 184 + ], + "score": 1.0, + "content": ". Thus, all the other elements", + "type": "text" + }, + { + "bbox": [ + 258, + 172, + 286, + 184 + ], + "score": 0.91, + "content": "j \\neq i", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 171, + 485, + 184 + ], + "score": 1.0, + "content": "contribute to the feature refinement of element", + "type": "text" + }, + { + "bbox": [ + 485, + 173, + 490, + 182 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 171, + 504, + 184 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 183, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 195 + ], + "score": 1.0, + "content": "summing up their relations. The response is normalized by the total number of elements in the set,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 193, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 107, + 195, + 117, + 204 + ], + "score": 0.78, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 193, + 201, + 207 + ], + "score": 1.0, + "content": ". The weight matrix", + "type": "text" + }, + { + "bbox": [ + 201, + 194, + 218, + 205 + ], + "score": 0.88, + "content": "W _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 193, + 506, + 207 + ], + "score": 1.0, + "content": "computes a linear embedding, producing the contextual residual to be", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 205, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 143, + 217 + ], + "score": 1.0, + "content": "added to", + "type": "text" + }, + { + "bbox": [ + 144, + 205, + 181, + 217 + ], + "score": 0.93, + "content": "f ( p _ { i } , \\theta _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 205, + 396, + 217 + ], + "score": 1.0, + "content": "for feature refinement. In our experiments, we define", + "type": "text" + }, + { + "bbox": [ + 397, + 205, + 408, + 215 + ], + "score": 0.81, + "content": "\\pmb { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 205, + 479, + 217 + ], + "score": 1.0, + "content": "as a dot-product:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5 + }, + { + "type": "interline_equation", + "bbox": [ + 190, + 220, + 420, + 235 + ], + "lines": [ + { + "bbox": [ + 190, + 220, + 420, + 235 + ], + "spans": [ + { + "bbox": [ + 190, + 220, + 420, + 235 + ], + "score": 0.89, + "content": "\\pmb { H } ( f ( p _ { i } , \\pmb { \\theta } _ { i } ) , f ( p _ { j } , \\pmb { \\theta } _ { j } ) ) = \\psi ( f ( p _ { i } , \\pmb { \\theta } _ { i } ) ) ^ { T } \\phi ( f ( p _ { j } , \\pmb { \\theta } _ { j } ) ) ,", + "type": "interline_equation", + "image_path": "29600091bf77d9f4f57eb3415ed8af20f2e144e9750f7f90c889f13fd2c5188c.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 190, + 220, + 420, + 235 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 135, + 253 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 239, + 261, + 252 + ], + "score": 0.93, + "content": "\\psi ( f ( p _ { i } , \\pmb { \\theta } _ { i } ) ) = W _ { \\psi } f ( p _ { i } , \\pmb { \\theta } _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 239, + 281, + 253 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 282, + 239, + 409, + 252 + ], + "score": 0.92, + "content": "\\phi ( f ( p _ { j } , \\pmb { \\theta } _ { j } ) ) = W _ { \\phi } f ( p _ { j } , \\pmb { \\theta } _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 239, + 505, + 253 + ], + "score": 1.0, + "content": "are two linear embed-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "dings. We stack 4 relation modules for feature refinement in our experiments. Finally, a decoder", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 261, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 506, + 274 + ], + "score": 1.0, + "content": "consisting of another multilayer perceptron network followed by two branches of fully connected", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 272, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 506, + 284 + ], + "score": 1.0, + "content": "layer with a sigmoid activation function is used to map the refined feature of each element back to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 283, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 505, + 296 + ], + "score": 1.0, + "content": "class probabilities and geometric parameters respectively. Optionally, Non-Maximum Suppression", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 293, + 324, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 324, + 306 + ], + "score": 1.0, + "content": "(NMS) can be applied to remove duplicated elements.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 107, + 316, + 322, + 327 + ], + "lines": [ + { + "bbox": [ + 106, + 316, + 322, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 322, + 328 + ], + "score": 1.0, + "content": "3.3 DISCRIMINATOR NETWORK ARCHITECTURES", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 332, + 505, + 365 + ], + "lines": [ + { + "bbox": [ + 105, + 329, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 506, + 347 + ], + "score": 1.0, + "content": "The discriminator aims to distinguish between synthesized and real layouts. We present two types", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "score": 1.0, + "content": "of discriminators, one based on relation modules directly built upon layout parameters, and the other", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 353, + 304, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 304, + 366 + ], + "score": 1.0, + "content": "based on how the layouts look like via rendering.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 107, + 373, + 287, + 383 + ], + "lines": [ + { + "bbox": [ + 105, + 372, + 289, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 289, + 385 + ], + "score": 1.0, + "content": "3.3.1 RELATION-BASED DISCRIMINATOR", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 387, + 505, + 465 + ], + "lines": [ + { + "bbox": [ + 105, + 386, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 506, + 400 + ], + "score": 1.0, + "content": "The Relation-Based Discriminator takes as input a set of graphic elements represented by class", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 398, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 411 + ], + "score": 1.0, + "content": "probabilities and geometric parameters, and feeds them to an encoder consisting of a multilayer", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 277, + 422 + ], + "score": 1.0, + "content": "perceptron network for feature embedding", + "type": "text" + }, + { + "bbox": [ + 277, + 409, + 315, + 421 + ], + "score": 0.92, + "content": "f ( p _ { i } , \\theta _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 408, + 506, + 422 + ], + "score": 1.0, + "content": ". It then extracts their global graphical relations", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 420, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 230, + 432 + ], + "score": 0.89, + "content": "g ( r ( p _ { 1 } , \\theta _ { 1 } ) , \\cdot \\cdot \\cdot , r ( p _ { N } , \\theta _ { N } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 420, + 259, + 434 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 260, + 421, + 298, + 432 + ], + "score": 0.92, + "content": "r ( p _ { i } , \\theta _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 420, + 506, + 434 + ], + "score": 1.0, + "content": "represents a simplified relation module as in equa-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 299, + 443 + ], + "score": 1.0, + "content": "tion 1 by removing the shortcut connection, and", + "type": "text" + }, + { + "bbox": [ + 300, + 433, + 307, + 443 + ], + "score": 0.82, + "content": "\\textbf { { g } }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 430, + 506, + 443 + ], + "score": 1.0, + "content": "is a max-pooling function used in (Charles et al.,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 441, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 506, + 455 + ], + "score": 1.0, + "content": "2017). Thus, the global relations among all the graphic elements can be modeled, upon which a", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 452, + 467, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 467, + 466 + ], + "score": 1.0, + "content": "classifier composed by a multilayer perception network is applied for real/fake prediction.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 107, + 472, + 318, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 320, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 320, + 484 + ], + "score": 1.0, + "content": "3.3.2 WIREFRAME RENDERING DISCRIMINATOR", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 108, + 486, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "The Wireframe Rendering Discriminator exploits CNNs to classify layouts, in order to learn to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 497, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 511 + ], + "score": 1.0, + "content": "classify the visual properties of a layout. The Discriminator consists of a wireframe rendering layer,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 508, + 439, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 238, + 520 + ], + "score": 1.0, + "content": "which produces an output image", + "type": "text" + }, + { + "bbox": [ + 239, + 509, + 245, + 518 + ], + "score": 0.69, + "content": "\\pmb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 508, + 439, + 520 + ], + "score": 1.0, + "content": ", which is then fed into a CNN for classification.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 505, + 569 + ], + "lines": [ + { + "bbox": [ + 106, + 524, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 358, + 538 + ], + "score": 1.0, + "content": "The rasterization is performed as follows. Suppose there are", + "type": "text" + }, + { + "bbox": [ + 358, + 526, + 369, + 535 + ], + "score": 0.77, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 524, + 505, + 538 + ], + "score": 1.0, + "content": "elements in the design, with pa-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 145, + 549 + ], + "score": 1.0, + "content": "rameters", + "type": "text" + }, + { + "bbox": [ + 145, + 536, + 247, + 548 + ], + "score": 0.92, + "content": "\\{ ( p _ { 1 } , \\pmb { \\theta _ { 1 } } ) , . . . , ( p _ { N } , \\pmb { \\theta } _ { N } ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 536, + 505, + 549 + ], + "score": 1.0, + "content": ". Each element can be rendered into its own grayscale image", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 107, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 107, + 547, + 142, + 559 + ], + "score": 0.92, + "content": "F _ { \\theta } ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "; rendering details for specific types of elements are given below. The image dimensions", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 557, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 184, + 570 + ], + "score": 1.0, + "content": "for each individual", + "type": "text" + }, + { + "bbox": [ + 184, + 559, + 194, + 568 + ], + "score": 0.84, + "content": "\\pmb { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 557, + 209, + 570 + ], + "score": 1.0, + "content": "are", + "type": "text" + }, + { + "bbox": [ + 209, + 558, + 242, + 569 + ], + "score": 0.91, + "content": "W \\times H", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 557, + 272, + 570 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 273, + 559, + 284, + 568 + ], + "score": 0.84, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 557, + 302, + 570 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 303, + 559, + 313, + 568 + ], + "score": 0.83, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 557, + 505, + 570 + ], + "score": 1.0, + "content": "are the width and height of the design in pixels.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 574, + 505, + 608 + ], + "lines": [ + { + "bbox": [ + 106, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 285, + 587 + ], + "score": 1.0, + "content": "The layer output is a multi-channel image", + "type": "text" + }, + { + "bbox": [ + 285, + 576, + 292, + 585 + ], + "score": 0.77, + "content": "\\pmb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 574, + 356, + 587 + ], + "score": 1.0, + "content": "of dimensions", + "type": "text" + }, + { + "bbox": [ + 356, + 575, + 416, + 586 + ], + "score": 0.91, + "content": "W \\times H \\times M", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 574, + 505, + 587 + ], + "score": 1.0, + "content": ", where each channel", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 213, + 599 + ], + "score": 1.0, + "content": "corresponds to one of the", + "type": "text" + }, + { + "bbox": [ + 213, + 586, + 225, + 596 + ], + "score": 0.79, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 586, + 378, + 599 + ], + "score": 1.0, + "content": "element types. In other words, pixel", + "type": "text" + }, + { + "bbox": [ + 378, + 586, + 402, + 598 + ], + "score": 0.91, + "content": "( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 586, + 415, + 599 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 415, + 587, + 423, + 596 + ], + "score": 0.79, + "content": "\\pmb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "is a class activation", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 597, + 272, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 272, + 609 + ], + "score": 1.0, + "content": "vector for that pixel, and is computed as:", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 612, + 375, + 632 + ], + "lines": [ + { + "bbox": [ + 236, + 612, + 375, + 632 + ], + "spans": [ + { + "bbox": [ + 236, + 612, + 375, + 632 + ], + "score": 0.94, + "content": "\\pmb { I } ( x , y , c ) = \\operatorname* { m a x } _ { i \\in [ 1 . . . N ] } p _ { i , c } \\pmb { F } _ { \\theta _ { i } } ( x , y )", + "type": "interline_equation", + "image_path": "9eec4be5eb257463fad7b0f98f3de4468809c4c3a85e5171528437f211320921.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 236, + 612, + 375, + 632 + ], + "spans": [], + "index": 42 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 635, + 327, + 647 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 325, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 133, + 649 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 637, + 149, + 648 + ], + "score": 0.88, + "content": "p _ { i , c }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 633, + 271, + 649 + ], + "score": 1.0, + "content": "is the probability that element", + "type": "text" + }, + { + "bbox": [ + 272, + 636, + 276, + 645 + ], + "score": 0.82, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 633, + 319, + 649 + ], + "score": 1.0, + "content": "is of class", + "type": "text" + }, + { + "bbox": [ + 320, + 638, + 325, + 645 + ], + "score": 0.75, + "content": "c", + "type": "inline_equation" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 651, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 324, + 666 + ], + "score": 1.0, + "content": "We next describe the rendering process for computing", + "type": "text" + }, + { + "bbox": [ + 324, + 653, + 336, + 663 + ], + "score": 0.88, + "content": "F _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 650, + 415, + 666 + ], + "score": 1.0, + "content": ", in the cases where", + "type": "text" + }, + { + "bbox": [ + 415, + 653, + 422, + 662 + ], + "score": 0.8, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 650, + 506, + 666 + ], + "score": 1.0, + "content": "represents a point, a", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 663, + 207, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 207, + 676 + ], + "score": 1.0, + "content": "rectangle, and a triangle.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 107, + 679, + 505, + 713 + ], + "lines": [ + { + "bbox": [ + 105, + 679, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 378, + 693 + ], + "score": 1.0, + "content": "Point. We start with the simplest geometric form, a single keypoint", + "type": "text" + }, + { + "bbox": [ + 378, + 680, + 431, + 692 + ], + "score": 0.93, + "content": "\\theta _ { i } = ( x _ { i } , y _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 679, + 480, + 693 + ], + "score": 1.0, + "content": "for element", + "type": "text" + }, + { + "bbox": [ + 480, + 681, + 485, + 690 + ], + "score": 0.53, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 679, + 505, + 693 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 690, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 245, + 704 + ], + "score": 1.0, + "content": "implement an interpolation kernel", + "type": "text" + }, + { + "bbox": [ + 246, + 692, + 252, + 700 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 690, + 469, + 704 + ], + "score": 1.0, + "content": "for its rasterization. Its spatial rendering response on", + "type": "text" + }, + { + "bbox": [ + 469, + 691, + 493, + 703 + ], + "score": 0.91, + "content": "( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 690, + 505, + 704 + ], + "score": 1.0, + "content": "in", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 702, + 258, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 258, + 714 + ], + "score": 1.0, + "content": "the rendered image can be written as:", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47 + }, + { + "type": "interline_equation", + "bbox": [ + 238, + 717, + 374, + 731 + ], + "lines": [ + { + "bbox": [ + 238, + 717, + 374, + 731 + ], + "spans": [ + { + "bbox": [ + 238, + 717, + 374, + 731 + ], + "score": 0.91, + "content": "F _ { \\theta _ { i } } ( x , y ) = k ( x - x _ { i } ) k ( y - y _ { i } ) .", + "type": "interline_equation", + "image_path": "1b692adce297264250ec4674e971f0cb85cbdd64c0e350bb4e78beaa40e5daa7.jpg" + } + ] + } + ], + "index": 49, + "virtual_lines": [ + { + "bbox": [ + 238, + 717, + 374, + 731 + ], + "spans": [], + "index": 49 + } + ] + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 8 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 116 + ], + "lines": [], + "index": 1, + "bbox_fs": [ + 105, + 82, + 505, + 117 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 150, + 118, + 460, + 150 + ], + "lines": [ + { + "bbox": [ + 150, + 118, + 460, + 150 + ], + "spans": [ + { + "bbox": [ + 150, + 118, + 460, + 150 + ], + "score": 0.93, + "content": "f ^ { \\prime } ( p _ { i } , \\theta _ { i } ) = W _ { r } \\frac { 1 } { N } \\sum _ { \\forall j \\neq i } H ( f ( p _ { i } , \\theta _ { i } ) , f ( p _ { j } , \\theta _ { j } ) ) U ( f ( p _ { j } , \\theta _ { j } ) ) + f ( p _ { i } , \\theta _ { i } ) ,", + "type": "interline_equation", + "image_path": "e3abe037053bdc2592f0cd9af25873546f3adaea0a1713a9d4cee41d5a5f7b09.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 150, + 118, + 460, + 128.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 150, + 128.66666666666666, + 460, + 139.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 150, + 139.33333333333331, + 460, + 149.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 150, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 106, + 150, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 205, + 163 + ], + "score": 1.0, + "content": "where an unary function", + "type": "text" + }, + { + "bbox": [ + 205, + 151, + 215, + 160 + ], + "score": 0.82, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 150, + 420, + 163 + ], + "score": 1.0, + "content": "computes a representation of the embedded feature", + "type": "text" + }, + { + "bbox": [ + 420, + 150, + 459, + 162 + ], + "score": 0.93, + "content": "f ( p _ { j } , \\theta _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 150, + 506, + 163 + ], + "score": 1.0, + "content": "of element", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 161, + 504, + 173 + ], + "spans": [ + { + "bbox": [ + 107, + 162, + 113, + 173 + ], + "score": 0.8, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 161, + 212, + 173 + ], + "score": 1.0, + "content": "and a pairwise function", + "type": "text" + }, + { + "bbox": [ + 212, + 162, + 224, + 171 + ], + "score": 0.79, + "content": "\\pmb { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 161, + 499, + 173 + ], + "score": 1.0, + "content": "computes a scalar value representing the relation between elements", + "type": "text" + }, + { + "bbox": [ + 499, + 162, + 504, + 171 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 171, + 504, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 124, + 184 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 173, + 131, + 184 + ], + "score": 0.81, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 171, + 258, + 184 + ], + "score": 1.0, + "content": ". Thus, all the other elements", + "type": "text" + }, + { + "bbox": [ + 258, + 172, + 286, + 184 + ], + "score": 0.91, + "content": "j \\neq i", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 171, + 485, + 184 + ], + "score": 1.0, + "content": "contribute to the feature refinement of element", + "type": "text" + }, + { + "bbox": [ + 485, + 173, + 490, + 182 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 171, + 504, + 184 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 183, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 195 + ], + "score": 1.0, + "content": "summing up their relations. The response is normalized by the total number of elements in the set,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 193, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 107, + 195, + 117, + 204 + ], + "score": 0.78, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 193, + 201, + 207 + ], + "score": 1.0, + "content": ". The weight matrix", + "type": "text" + }, + { + "bbox": [ + 201, + 194, + 218, + 205 + ], + "score": 0.88, + "content": "W _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 193, + 506, + 207 + ], + "score": 1.0, + "content": "computes a linear embedding, producing the contextual residual to be", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 205, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 143, + 217 + ], + "score": 1.0, + "content": "added to", + "type": "text" + }, + { + "bbox": [ + 144, + 205, + 181, + 217 + ], + "score": 0.93, + "content": "f ( p _ { i } , \\theta _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 205, + 396, + 217 + ], + "score": 1.0, + "content": "for feature refinement. In our experiments, we define", + "type": "text" + }, + { + "bbox": [ + 397, + 205, + 408, + 215 + ], + "score": 0.81, + "content": "\\pmb { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 205, + 479, + 217 + ], + "score": 1.0, + "content": "as a dot-product:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 150, + 506, + 217 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 190, + 220, + 420, + 235 + ], + "lines": [ + { + "bbox": [ + 190, + 220, + 420, + 235 + ], + "spans": [ + { + "bbox": [ + 190, + 220, + 420, + 235 + ], + "score": 0.89, + "content": "\\pmb { H } ( f ( p _ { i } , \\pmb { \\theta } _ { i } ) , f ( p _ { j } , \\pmb { \\theta } _ { j } ) ) = \\psi ( f ( p _ { i } , \\pmb { \\theta } _ { i } ) ) ^ { T } \\phi ( f ( p _ { j } , \\pmb { \\theta } _ { j } ) ) ,", + "type": "interline_equation", + "image_path": "29600091bf77d9f4f57eb3415ed8af20f2e144e9750f7f90c889f13fd2c5188c.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 190, + 220, + 420, + 235 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 135, + 253 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 239, + 261, + 252 + ], + "score": 0.93, + "content": "\\psi ( f ( p _ { i } , \\pmb { \\theta } _ { i } ) ) = W _ { \\psi } f ( p _ { i } , \\pmb { \\theta } _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 239, + 281, + 253 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 282, + 239, + 409, + 252 + ], + "score": 0.92, + "content": "\\phi ( f ( p _ { j } , \\pmb { \\theta } _ { j } ) ) = W _ { \\phi } f ( p _ { j } , \\pmb { \\theta } _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 239, + 505, + 253 + ], + "score": 1.0, + "content": "are two linear embed-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "dings. We stack 4 relation modules for feature refinement in our experiments. Finally, a decoder", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 261, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 506, + 274 + ], + "score": 1.0, + "content": "consisting of another multilayer perceptron network followed by two branches of fully connected", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 272, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 506, + 284 + ], + "score": 1.0, + "content": "layer with a sigmoid activation function is used to map the refined feature of each element back to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 283, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 505, + 296 + ], + "score": 1.0, + "content": "class probabilities and geometric parameters respectively. Optionally, Non-Maximum Suppression", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 293, + 324, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 324, + 306 + ], + "score": 1.0, + "content": "(NMS) can be applied to remove duplicated elements.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 239, + 506, + 306 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 316, + 322, + 327 + ], + "lines": [ + { + "bbox": [ + 106, + 316, + 322, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 322, + 328 + ], + "score": 1.0, + "content": "3.3 DISCRIMINATOR NETWORK ARCHITECTURES", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 332, + 505, + 365 + ], + "lines": [ + { + "bbox": [ + 105, + 329, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 506, + 347 + ], + "score": 1.0, + "content": "The discriminator aims to distinguish between synthesized and real layouts. We present two types", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "score": 1.0, + "content": "of discriminators, one based on relation modules directly built upon layout parameters, and the other", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 353, + 304, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 304, + 366 + ], + "score": 1.0, + "content": "based on how the layouts look like via rendering.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 329, + 506, + 366 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 373, + 287, + 383 + ], + "lines": [ + { + "bbox": [ + 105, + 372, + 289, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 289, + 385 + ], + "score": 1.0, + "content": "3.3.1 RELATION-BASED DISCRIMINATOR", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 387, + 505, + 465 + ], + "lines": [ + { + "bbox": [ + 105, + 386, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 506, + 400 + ], + "score": 1.0, + "content": "The Relation-Based Discriminator takes as input a set of graphic elements represented by class", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 398, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 411 + ], + "score": 1.0, + "content": "probabilities and geometric parameters, and feeds them to an encoder consisting of a multilayer", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 277, + 422 + ], + "score": 1.0, + "content": "perceptron network for feature embedding", + "type": "text" + }, + { + "bbox": [ + 277, + 409, + 315, + 421 + ], + "score": 0.92, + "content": "f ( p _ { i } , \\theta _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 408, + 506, + 422 + ], + "score": 1.0, + "content": ". It then extracts their global graphical relations", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 420, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 230, + 432 + ], + "score": 0.89, + "content": "g ( r ( p _ { 1 } , \\theta _ { 1 } ) , \\cdot \\cdot \\cdot , r ( p _ { N } , \\theta _ { N } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 420, + 259, + 434 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 260, + 421, + 298, + 432 + ], + "score": 0.92, + "content": "r ( p _ { i } , \\theta _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 420, + 506, + 434 + ], + "score": 1.0, + "content": "represents a simplified relation module as in equa-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 299, + 443 + ], + "score": 1.0, + "content": "tion 1 by removing the shortcut connection, and", + "type": "text" + }, + { + "bbox": [ + 300, + 433, + 307, + 443 + ], + "score": 0.82, + "content": "\\textbf { { g } }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 430, + 506, + 443 + ], + "score": 1.0, + "content": "is a max-pooling function used in (Charles et al.,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 441, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 506, + 455 + ], + "score": 1.0, + "content": "2017). Thus, the global relations among all the graphic elements can be modeled, upon which a", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 452, + 467, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 467, + 466 + ], + "score": 1.0, + "content": "classifier composed by a multilayer perception network is applied for real/fake prediction.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 386, + 506, + 466 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 472, + 318, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 320, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 320, + 484 + ], + "score": 1.0, + "content": "3.3.2 WIREFRAME RENDERING DISCRIMINATOR", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 108, + 486, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "The Wireframe Rendering Discriminator exploits CNNs to classify layouts, in order to learn to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 497, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 511 + ], + "score": 1.0, + "content": "classify the visual properties of a layout. The Discriminator consists of a wireframe rendering layer,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 508, + 439, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 238, + 520 + ], + "score": 1.0, + "content": "which produces an output image", + "type": "text" + }, + { + "bbox": [ + 239, + 509, + 245, + 518 + ], + "score": 0.69, + "content": "\\pmb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 508, + 439, + 520 + ], + "score": 1.0, + "content": ", which is then fed into a CNN for classification.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 106, + 486, + 505, + 520 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 505, + 569 + ], + "lines": [ + { + "bbox": [ + 106, + 524, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 358, + 538 + ], + "score": 1.0, + "content": "The rasterization is performed as follows. Suppose there are", + "type": "text" + }, + { + "bbox": [ + 358, + 526, + 369, + 535 + ], + "score": 0.77, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 524, + 505, + 538 + ], + "score": 1.0, + "content": "elements in the design, with pa-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 145, + 549 + ], + "score": 1.0, + "content": "rameters", + "type": "text" + }, + { + "bbox": [ + 145, + 536, + 247, + 548 + ], + "score": 0.92, + "content": "\\{ ( p _ { 1 } , \\pmb { \\theta _ { 1 } } ) , . . . , ( p _ { N } , \\pmb { \\theta } _ { N } ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 536, + 505, + 549 + ], + "score": 1.0, + "content": ". Each element can be rendered into its own grayscale image", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 107, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 107, + 547, + 142, + 559 + ], + "score": 0.92, + "content": "F _ { \\theta } ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "; rendering details for specific types of elements are given below. The image dimensions", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 557, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 184, + 570 + ], + "score": 1.0, + "content": "for each individual", + "type": "text" + }, + { + "bbox": [ + 184, + 559, + 194, + 568 + ], + "score": 0.84, + "content": "\\pmb { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 557, + 209, + 570 + ], + "score": 1.0, + "content": "are", + "type": "text" + }, + { + "bbox": [ + 209, + 558, + 242, + 569 + ], + "score": 0.91, + "content": "W \\times H", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 557, + 272, + 570 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 273, + 559, + 284, + 568 + ], + "score": 0.84, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 557, + 302, + 570 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 303, + 559, + 313, + 568 + ], + "score": 0.83, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 557, + 505, + 570 + ], + "score": 1.0, + "content": "are the width and height of the design in pixels.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 524, + 505, + 570 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 574, + 505, + 608 + ], + "lines": [ + { + "bbox": [ + 106, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 285, + 587 + ], + "score": 1.0, + "content": "The layer output is a multi-channel image", + "type": "text" + }, + { + "bbox": [ + 285, + 576, + 292, + 585 + ], + "score": 0.77, + "content": "\\pmb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 574, + 356, + 587 + ], + "score": 1.0, + "content": "of dimensions", + "type": "text" + }, + { + "bbox": [ + 356, + 575, + 416, + 586 + ], + "score": 0.91, + "content": "W \\times H \\times M", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 574, + 505, + 587 + ], + "score": 1.0, + "content": ", where each channel", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 213, + 599 + ], + "score": 1.0, + "content": "corresponds to one of the", + "type": "text" + }, + { + "bbox": [ + 213, + 586, + 225, + 596 + ], + "score": 0.79, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 586, + 378, + 599 + ], + "score": 1.0, + "content": "element types. In other words, pixel", + "type": "text" + }, + { + "bbox": [ + 378, + 586, + 402, + 598 + ], + "score": 0.91, + "content": "( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 586, + 415, + 599 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 415, + 587, + 423, + 596 + ], + "score": 0.79, + "content": "\\pmb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "is a class activation", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 597, + 272, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 272, + 609 + ], + "score": 1.0, + "content": "vector for that pixel, and is computed as:", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 574, + 505, + 609 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 612, + 375, + 632 + ], + "lines": [ + { + "bbox": [ + 236, + 612, + 375, + 632 + ], + "spans": [ + { + "bbox": [ + 236, + 612, + 375, + 632 + ], + "score": 0.94, + "content": "\\pmb { I } ( x , y , c ) = \\operatorname* { m a x } _ { i \\in [ 1 . . . N ] } p _ { i , c } \\pmb { F } _ { \\theta _ { i } } ( x , y )", + "type": "interline_equation", + "image_path": "9eec4be5eb257463fad7b0f98f3de4468809c4c3a85e5171528437f211320921.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 236, + 612, + 375, + 632 + ], + "spans": [], + "index": 42 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 635, + 327, + 647 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 325, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 133, + 649 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 637, + 149, + 648 + ], + "score": 0.88, + "content": "p _ { i , c }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 633, + 271, + 649 + ], + "score": 1.0, + "content": "is the probability that element", + "type": "text" + }, + { + "bbox": [ + 272, + 636, + 276, + 645 + ], + "score": 0.82, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 633, + 319, + 649 + ], + "score": 1.0, + "content": "is of class", + "type": "text" + }, + { + "bbox": [ + 320, + 638, + 325, + 645 + ], + "score": 0.75, + "content": "c", + "type": "inline_equation" + } + ], + "index": 43 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 633, + 325, + 649 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 651, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 324, + 666 + ], + "score": 1.0, + "content": "We next describe the rendering process for computing", + "type": "text" + }, + { + "bbox": [ + 324, + 653, + 336, + 663 + ], + "score": 0.88, + "content": "F _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 650, + 415, + 666 + ], + "score": 1.0, + "content": ", in the cases where", + "type": "text" + }, + { + "bbox": [ + 415, + 653, + 422, + 662 + ], + "score": 0.8, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 650, + 506, + 666 + ], + "score": 1.0, + "content": "represents a point, a", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 663, + 207, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 207, + 676 + ], + "score": 1.0, + "content": "rectangle, and a triangle.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 650, + 506, + 676 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 679, + 505, + 713 + ], + "lines": [ + { + "bbox": [ + 105, + 679, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 378, + 693 + ], + "score": 1.0, + "content": "Point. We start with the simplest geometric form, a single keypoint", + "type": "text" + }, + { + "bbox": [ + 378, + 680, + 431, + 692 + ], + "score": 0.93, + "content": "\\theta _ { i } = ( x _ { i } , y _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 679, + 480, + 693 + ], + "score": 1.0, + "content": "for element", + "type": "text" + }, + { + "bbox": [ + 480, + 681, + 485, + 690 + ], + "score": 0.53, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 679, + 505, + 693 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 690, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 245, + 704 + ], + "score": 1.0, + "content": "implement an interpolation kernel", + "type": "text" + }, + { + "bbox": [ + 246, + 692, + 252, + 700 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 690, + 469, + 704 + ], + "score": 1.0, + "content": "for its rasterization. Its spatial rendering response on", + "type": "text" + }, + { + "bbox": [ + 469, + 691, + 493, + 703 + ], + "score": 0.91, + "content": "( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 690, + 505, + 704 + ], + "score": 1.0, + "content": "in", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 702, + 258, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 258, + 714 + ], + "score": 1.0, + "content": "the rendered image can be written as:", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 679, + 505, + 714 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 238, + 717, + 374, + 731 + ], + "lines": [ + { + "bbox": [ + 238, + 717, + 374, + 731 + ], + "spans": [ + { + "bbox": [ + 238, + 717, + 374, + 731 + ], + "score": 0.91, + "content": "F _ { \\theta _ { i } } ( x , y ) = k ( x - x _ { i } ) k ( y - y _ { i } ) .", + "type": "interline_equation", + "image_path": "1b692adce297264250ec4674e971f0cb85cbdd64c0e350bb4e78beaa40e5daa7.jpg" + } + ] + } + ], + "index": 49, + "virtual_lines": [ + { + "bbox": [ + 238, + 717, + 374, + 731 + ], + "spans": [], + "index": 49 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 71, + 468, + 178 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 143, + 71, + 468, + 178 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 143, + 71, + 468, + 178 + ], + "spans": [ + { + "bbox": [ + 143, + 71, + 468, + 178 + ], + "score": 0.97, + "type": "image", + "image_path": "ff33a8ebcae43f98ddef388c3817179178471890959278d5168cd6dac657d1d0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 143, + 71, + 468, + 106.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 143, + 106.66666666666666, + 468, + 142.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 143, + 142.33333333333331, + 468, + 177.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 187, + 505, + 232 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 186, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 201 + ], + "score": 1.0, + "content": "Figure 2: Wireframe rendering of different polygons (point, rectangle and triangle). The black grids", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 199, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 210 + ], + "score": 1.0, + "content": "represent grids of target image. The orange dots/dotted lines represent the graphic element mapped", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 210, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 505, + 221 + ], + "score": 1.0, + "content": "onto the image grid. The blue solid lines represent the rasterized wireframes expressed as differen-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 220, + 489, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 489, + 232 + ], + "score": 1.0, + "content": "tiable functions of graphic elements in terms of both class probilities and geometric parameters.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 505, + 295 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 504, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 469, + 253 + ], + "score": 1.0, + "content": "We adopt bilinear interpretation (Johnson et al., 2016), corresponding to the kernel", + "type": "text" + }, + { + "bbox": [ + 469, + 239, + 504, + 252 + ], + "score": 0.89, + "content": "\\boldsymbol k ( d ) \\_ =", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 107, + 250, + 172, + 262 + ], + "score": 0.91, + "content": "\\operatorname* { m a x } ( 0 , \\bar { 1 } - | d | )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 250, + 423, + 262 + ], + "score": 1.0, + "content": "(implemented as ReLU activation), as shown in Figure 2. As", + "type": "text" + }, + { + "bbox": [ + 423, + 251, + 431, + 261 + ], + "score": 0.65, + "content": "\\pmb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "is a differentiable", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 260, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 275 + ], + "score": 1.0, + "content": "function of the class probabilities and the coordinates, the sub-gradients of the rasterized image can", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "score": 1.0, + "content": "be thus propagated backward to them. We validate such rendering design for MNIST digit genera-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 283, + 321, + 297 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 321, + 297 + ], + "score": 1.0, + "content": "tion, detailed experiments can be seen in Section 4.1.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 300, + 506, + 378 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 313 + ], + "score": 1.0, + "content": "Rectangle. We now consider more complex polygons. Assuming an element is a rectangle, or", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 310, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 104, + 310, + 385, + 324 + ], + "score": 1.0, + "content": "bounding box represented by its top-left and bottom-right coordinates", + "type": "text" + }, + { + "bbox": [ + 386, + 310, + 474, + 323 + ], + "score": 0.93, + "content": "\\pmb { \\theta } = ( x ^ { L } , y ^ { T } , x ^ { R } , y ^ { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 310, + 506, + 324 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 321, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 104, + 321, + 404, + 335 + ], + "score": 1.0, + "content": "is very common in various designs. Specifically, considering a rectangle", + "type": "text" + }, + { + "bbox": [ + 404, + 323, + 410, + 332 + ], + "score": 0.72, + "content": "_ i", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 321, + 481, + 335 + ], + "score": 1.0, + "content": "with coordinates", + "type": "text" + }, + { + "bbox": [ + 482, + 322, + 505, + 333 + ], + "score": 0.88, + "content": "\\theta _ { i } = ", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 107, + 329, + 507, + 348 + ], + "spans": [ + { + "bbox": [ + 107, + 333, + 176, + 345 + ], + "score": 0.91, + "content": "( x _ { i } ^ { L } , y _ { i } ^ { T } , x _ { i } ^ { R } , y _ { i } ^ { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 329, + 507, + 348 + ], + "score": 1.0, + "content": ", as shown in Figure 2, the black grids represent the locations in the rendered image", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 344, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 506, + 356 + ], + "score": 1.0, + "content": "and the orange dotted box represents the rectangle being rasterized in the rendered image. For a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 355, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 355, + 505, + 367 + ], + "score": 1.0, + "content": "wireframe representation, only the points near the boundary of the dotted box (lie in blue solid line)", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 366, + 473, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 360, + 378 + ], + "score": 1.0, + "content": "are related to the rectangle, so its spatial rendering response on", + "type": "text" + }, + { + "bbox": [ + 360, + 366, + 383, + 378 + ], + "score": 0.92, + "content": "( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 366, + 473, + 378 + ], + "score": 1.0, + "content": "can be formulated as:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 193, + 382, + 416, + 434 + ], + "lines": [ + { + "bbox": [ + 193, + 382, + 416, + 434 + ], + "spans": [ + { + "bbox": [ + 193, + 382, + 416, + 434 + ], + "score": 0.95, + "content": "\\begin{array} { r } { F _ { \\theta _ { i } } ( x , y ) = \\operatorname* { m a x } \\left( \\begin{array} { l } { k ( x - x _ { i } ^ { L } ) b ( y - y _ { i } ^ { T } ) b ( y _ { i } ^ { B } - y ) , } \\\\ { k ( x - x _ { i } ^ { R } ) b ( y - y _ { i } ^ { T } ) b ( y _ { i } ^ { B } - y ) , } \\\\ { k ( y - y _ { i } ^ { T } ) b ( x - x _ { i } ^ { L } ) b ( x _ { i } ^ { R } - x ) , } \\\\ { k ( y - y _ { i } ^ { B } ) b ( x - x _ { i } ^ { L } ) b ( x _ { i } ^ { R } - x ) } \\end{array} \\right) , } \\end{array}", + "type": "interline_equation", + "image_path": "f34ab105e89f71d89378e54a8dbe7c0957ffbd4b1af58bd0735805d869c375c6.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 193, + 382, + 416, + 395.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 193, + 395.0, + 416, + 408.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 193, + 408.0, + 416, + 421.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 193, + 421.0, + 416, + 434.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 438, + 416, + 451 + ], + "lines": [ + { + "bbox": [ + 106, + 438, + 416, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 133, + 452 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 438, + 239, + 451 + ], + "score": 0.92, + "content": "b ( d ) = \\operatorname* { m i n } ( \\operatorname* { m a x } ( 0 , d ) , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 438, + 416, + 452 + ], + "score": 1.0, + "content": "constraining the rendering to nearby pixels.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 108, + 455, + 505, + 489 + ], + "lines": [ + { + "bbox": [ + 106, + 454, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 505, + 468 + ], + "score": 1.0, + "content": "Triangle. We further describe the wireframe rendering process of another geometric form, triangle.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 464, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 155, + 480 + ], + "score": 1.0, + "content": "For triangle", + "type": "text" + }, + { + "bbox": [ + 155, + 468, + 160, + 477 + ], + "score": 0.77, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 464, + 341, + 480 + ], + "score": 1.0, + "content": "represented by its three vertices’ coordinates", + "type": "text" + }, + { + "bbox": [ + 341, + 466, + 454, + 479 + ], + "score": 0.93, + "content": "\\pmb { \\theta _ { i } } = ( x _ { i } ^ { 1 } , y _ { i } ^ { 1 } , x _ { i } ^ { \\top } , y _ { i } ^ { 2 } , x _ { i } ^ { 3 } , y _ { i } ^ { 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 464, + 482, + 480 + ], + "score": 1.0, + "content": ", when", + "type": "text" + }, + { + "bbox": [ + 482, + 466, + 506, + 479 + ], + "score": 0.89, + "content": "x _ { i } ^ { \\bar { 1 } } \\neq", + "type": "inline_equation" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 477, + 476, + 491 + ], + "spans": [ + { + "bbox": [ + 107, + 477, + 141, + 489 + ], + "score": 0.92, + "content": "x _ { i } ^ { 2 } \\neq x _ { i } ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 477, + 276, + 491 + ], + "score": 1.0, + "content": ", its spatial rendering response on", + "type": "text" + }, + { + "bbox": [ + 276, + 478, + 300, + 489 + ], + "score": 0.92, + "content": "( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 477, + 476, + 491 + ], + "score": 1.0, + "content": "in the rendered image can be calculated as:", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 160, + 494, + 450, + 550 + ], + "lines": [ + { + "bbox": [ + 160, + 494, + 450, + 550 + ], + "spans": [ + { + "bbox": [ + 160, + 494, + 450, + 550 + ], + "score": 0.95, + "content": "\\begin{array} { r } { F _ { \\theta _ { i } } ( x , y ) = \\operatorname* { m a x } \\left( \\begin{array} { l } { k ( y - \\frac { ( y _ { i } ^ { 2 } - y _ { i } ^ { 1 } ) \\cdot ( x - x _ { i } ^ { 1 } ) } { x _ { i } ^ { 2 } - x _ { i } ^ { 1 } } - y _ { i } ^ { 1 } ) b ( x - x _ { i } ^ { 1 } ) b ( x _ { i } ^ { 2 } - x ) , } \\\\ { k ( y - \\frac { ( y _ { i } ^ { 3 } - y _ { i } ^ { 1 } ) \\cdot ( x - x _ { i } ^ { 1 } ) } { x _ { i } ^ { 3 } - x _ { i } ^ { 1 } } - y _ { i } ^ { 1 } ) b ( x - x _ { i } ^ { 3 } ) b ( x _ { i } ^ { 1 } - x ) , } \\\\ { k ( y - \\frac { ( y _ { i } ^ { 3 } - y _ { i } ^ { 2 } ) \\cdot ( x - x _ { i } ^ { 2 } ) } { x _ { i } ^ { 3 } - x _ { i } ^ { 2 } } - y _ { i } ^ { 2 } ) b ( x - x _ { i } ^ { 3 } ) b ( x _ { i } ^ { 2 } - x ) } \\end{array} \\right) , } \\end{array}", + "type": "interline_equation", + "image_path": "f522d3921ce50be5c1f255681b4d230a0d5ee05230cbbf8ecfa7095a00d83e34.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 160, + 494, + 450, + 512.6666666666666 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 160, + 512.6666666666666, + 450, + 531.3333333333333 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 160, + 531.3333333333333, + 450, + 549.9999999999999 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 505, + 576 + ], + "lines": [ + { + "bbox": [ + 106, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 505, + 566 + ], + "score": 1.0, + "content": "Through this wireframe rendering process, gradients can be propagated backward to both the class", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 564, + 450, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 450, + 577 + ], + "score": 1.0, + "content": "probabilities and geometric parameters of the graphic elements for joint optimization.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "score": 1.0, + "content": "A CNN consisting of 3 convolutional layers, followed by a fully connected layer with sigmoid", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 593, + 442, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 290, + 605 + ], + "score": 1.0, + "content": "activation, is applied to the rasterization layer", + "type": "text" + }, + { + "bbox": [ + 290, + 593, + 298, + 603 + ], + "score": 0.68, + "content": "\\pmb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 593, + 442, + 605 + ], + "score": 1.0, + "content": "to predict real/fake graphic layouts.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "title", + "bbox": [ + 107, + 616, + 201, + 628 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 201, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 201, + 630 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 631, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "The implementation is based on TensorFlow (Abadi et al., 2016). The network parameters are", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 642, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 505, + 654 + ], + "score": 1.0, + "content": "initialized from zero-mean Gaussian with standard deviation of 0.02. All the networks are optimized", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 654, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 505, + 666 + ], + "score": 1.0, + "content": "using Adam (Kingma & Ba, 2014) with a fixed learning rate of 0.00002. Detailed architectures can", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 664, + 210, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 210, + 677 + ], + "score": 1.0, + "content": "be found in the appendix.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5 + }, + { + "type": "title", + "bbox": [ + 107, + 685, + 250, + 696 + ], + "lines": [ + { + "bbox": [ + 105, + 684, + 252, + 697 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 252, + 697 + ], + "score": 1.0, + "content": "4.1 MNIST DIGIT GENERATION", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "As a toy problem, we generate MNIST layouts. MNIST is a handwritten digit database consisting", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "of 60,000 training and 10,000 testing images. For each image, we extract the locations of 128", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "randomly-selected foreground pixels as the graphic representation so that digit generation can be", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 71, + 468, + 178 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 143, + 71, + 468, + 178 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 143, + 71, + 468, + 178 + ], + "spans": [ + { + "bbox": [ + 143, + 71, + 468, + 178 + ], + "score": 0.97, + "type": "image", + "image_path": "ff33a8ebcae43f98ddef388c3817179178471890959278d5168cd6dac657d1d0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 143, + 71, + 468, + 106.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 143, + 106.66666666666666, + 468, + 142.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 143, + 142.33333333333331, + 468, + 177.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 187, + 505, + 232 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 186, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 201 + ], + "score": 1.0, + "content": "Figure 2: Wireframe rendering of different polygons (point, rectangle and triangle). The black grids", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 199, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 210 + ], + "score": 1.0, + "content": "represent grids of target image. The orange dots/dotted lines represent the graphic element mapped", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 210, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 505, + 221 + ], + "score": 1.0, + "content": "onto the image grid. The blue solid lines represent the rasterized wireframes expressed as differen-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 220, + 489, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 489, + 232 + ], + "score": 1.0, + "content": "tiable functions of graphic elements in terms of both class probilities and geometric parameters.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 505, + 295 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 504, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 469, + 253 + ], + "score": 1.0, + "content": "We adopt bilinear interpretation (Johnson et al., 2016), corresponding to the kernel", + "type": "text" + }, + { + "bbox": [ + 469, + 239, + 504, + 252 + ], + "score": 0.89, + "content": "\\boldsymbol k ( d ) \\_ =", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 107, + 250, + 172, + 262 + ], + "score": 0.91, + "content": "\\operatorname* { m a x } ( 0 , \\bar { 1 } - | d | )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 250, + 423, + 262 + ], + "score": 1.0, + "content": "(implemented as ReLU activation), as shown in Figure 2. As", + "type": "text" + }, + { + "bbox": [ + 423, + 251, + 431, + 261 + ], + "score": 0.65, + "content": "\\pmb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "is a differentiable", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 260, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 275 + ], + "score": 1.0, + "content": "function of the class probabilities and the coordinates, the sub-gradients of the rasterized image can", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "score": 1.0, + "content": "be thus propagated backward to them. We validate such rendering design for MNIST digit genera-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 283, + 321, + 297 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 321, + 297 + ], + "score": 1.0, + "content": "tion, detailed experiments can be seen in Section 4.1.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 239, + 505, + 297 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 300, + 506, + 378 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 313 + ], + "score": 1.0, + "content": "Rectangle. We now consider more complex polygons. Assuming an element is a rectangle, or", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 310, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 104, + 310, + 385, + 324 + ], + "score": 1.0, + "content": "bounding box represented by its top-left and bottom-right coordinates", + "type": "text" + }, + { + "bbox": [ + 386, + 310, + 474, + 323 + ], + "score": 0.93, + "content": "\\pmb { \\theta } = ( x ^ { L } , y ^ { T } , x ^ { R } , y ^ { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 310, + 506, + 324 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 321, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 104, + 321, + 404, + 335 + ], + "score": 1.0, + "content": "is very common in various designs. Specifically, considering a rectangle", + "type": "text" + }, + { + "bbox": [ + 404, + 323, + 410, + 332 + ], + "score": 0.72, + "content": "_ i", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 321, + 481, + 335 + ], + "score": 1.0, + "content": "with coordinates", + "type": "text" + }, + { + "bbox": [ + 482, + 322, + 505, + 333 + ], + "score": 0.88, + "content": "\\theta _ { i } = ", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 107, + 329, + 507, + 348 + ], + "spans": [ + { + "bbox": [ + 107, + 333, + 176, + 345 + ], + "score": 0.91, + "content": "( x _ { i } ^ { L } , y _ { i } ^ { T } , x _ { i } ^ { R } , y _ { i } ^ { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 329, + 507, + 348 + ], + "score": 1.0, + "content": ", as shown in Figure 2, the black grids represent the locations in the rendered image", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 344, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 506, + 356 + ], + "score": 1.0, + "content": "and the orange dotted box represents the rectangle being rasterized in the rendered image. For a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 355, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 355, + 505, + 367 + ], + "score": 1.0, + "content": "wireframe representation, only the points near the boundary of the dotted box (lie in blue solid line)", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 366, + 473, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 360, + 378 + ], + "score": 1.0, + "content": "are related to the rectangle, so its spatial rendering response on", + "type": "text" + }, + { + "bbox": [ + 360, + 366, + 383, + 378 + ], + "score": 0.92, + "content": "( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 366, + 473, + 378 + ], + "score": 1.0, + "content": "can be formulated as:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15, + "bbox_fs": [ + 104, + 299, + 507, + 378 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 193, + 382, + 416, + 434 + ], + "lines": [ + { + "bbox": [ + 193, + 382, + 416, + 434 + ], + "spans": [ + { + "bbox": [ + 193, + 382, + 416, + 434 + ], + "score": 0.95, + "content": "\\begin{array} { r } { F _ { \\theta _ { i } } ( x , y ) = \\operatorname* { m a x } \\left( \\begin{array} { l } { k ( x - x _ { i } ^ { L } ) b ( y - y _ { i } ^ { T } ) b ( y _ { i } ^ { B } - y ) , } \\\\ { k ( x - x _ { i } ^ { R } ) b ( y - y _ { i } ^ { T } ) b ( y _ { i } ^ { B } - y ) , } \\\\ { k ( y - y _ { i } ^ { T } ) b ( x - x _ { i } ^ { L } ) b ( x _ { i } ^ { R } - x ) , } \\\\ { k ( y - y _ { i } ^ { B } ) b ( x - x _ { i } ^ { L } ) b ( x _ { i } ^ { R } - x ) } \\end{array} \\right) , } \\end{array}", + "type": "interline_equation", + "image_path": "f34ab105e89f71d89378e54a8dbe7c0957ffbd4b1af58bd0735805d869c375c6.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 193, + 382, + 416, + 395.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 193, + 395.0, + 416, + 408.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 193, + 408.0, + 416, + 421.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 193, + 421.0, + 416, + 434.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 438, + 416, + 451 + ], + "lines": [ + { + "bbox": [ + 106, + 438, + 416, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 133, + 452 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 438, + 239, + 451 + ], + "score": 0.92, + "content": "b ( d ) = \\operatorname* { m i n } ( \\operatorname* { m a x } ( 0 , d ) , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 438, + 416, + 452 + ], + "score": 1.0, + "content": "constraining the rendering to nearby pixels.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 106, + 438, + 416, + 452 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 455, + 505, + 489 + ], + "lines": [ + { + "bbox": [ + 106, + 454, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 505, + 468 + ], + "score": 1.0, + "content": "Triangle. We further describe the wireframe rendering process of another geometric form, triangle.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 464, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 155, + 480 + ], + "score": 1.0, + "content": "For triangle", + "type": "text" + }, + { + "bbox": [ + 155, + 468, + 160, + 477 + ], + "score": 0.77, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 464, + 341, + 480 + ], + "score": 1.0, + "content": "represented by its three vertices’ coordinates", + "type": "text" + }, + { + "bbox": [ + 341, + 466, + 454, + 479 + ], + "score": 0.93, + "content": "\\pmb { \\theta _ { i } } = ( x _ { i } ^ { 1 } , y _ { i } ^ { 1 } , x _ { i } ^ { \\top } , y _ { i } ^ { 2 } , x _ { i } ^ { 3 } , y _ { i } ^ { 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 464, + 482, + 480 + ], + "score": 1.0, + "content": ", when", + "type": "text" + }, + { + "bbox": [ + 482, + 466, + 506, + 479 + ], + "score": 0.89, + "content": "x _ { i } ^ { \\bar { 1 } } \\neq", + "type": "inline_equation" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 477, + 476, + 491 + ], + "spans": [ + { + "bbox": [ + 107, + 477, + 141, + 489 + ], + "score": 0.92, + "content": "x _ { i } ^ { 2 } \\neq x _ { i } ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 477, + 276, + 491 + ], + "score": 1.0, + "content": ", its spatial rendering response on", + "type": "text" + }, + { + "bbox": [ + 276, + 478, + 300, + 489 + ], + "score": 0.92, + "content": "( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 477, + 476, + 491 + ], + "score": 1.0, + "content": "in the rendered image can be calculated as:", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 454, + 506, + 491 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 160, + 494, + 450, + 550 + ], + "lines": [ + { + "bbox": [ + 160, + 494, + 450, + 550 + ], + "spans": [ + { + "bbox": [ + 160, + 494, + 450, + 550 + ], + "score": 0.95, + "content": "\\begin{array} { r } { F _ { \\theta _ { i } } ( x , y ) = \\operatorname* { m a x } \\left( \\begin{array} { l } { k ( y - \\frac { ( y _ { i } ^ { 2 } - y _ { i } ^ { 1 } ) \\cdot ( x - x _ { i } ^ { 1 } ) } { x _ { i } ^ { 2 } - x _ { i } ^ { 1 } } - y _ { i } ^ { 1 } ) b ( x - x _ { i } ^ { 1 } ) b ( x _ { i } ^ { 2 } - x ) , } \\\\ { k ( y - \\frac { ( y _ { i } ^ { 3 } - y _ { i } ^ { 1 } ) \\cdot ( x - x _ { i } ^ { 1 } ) } { x _ { i } ^ { 3 } - x _ { i } ^ { 1 } } - y _ { i } ^ { 1 } ) b ( x - x _ { i } ^ { 3 } ) b ( x _ { i } ^ { 1 } - x ) , } \\\\ { k ( y - \\frac { ( y _ { i } ^ { 3 } - y _ { i } ^ { 2 } ) \\cdot ( x - x _ { i } ^ { 2 } ) } { x _ { i } ^ { 3 } - x _ { i } ^ { 2 } } - y _ { i } ^ { 2 } ) b ( x - x _ { i } ^ { 3 } ) b ( x _ { i } ^ { 2 } - x ) } \\end{array} \\right) , } \\end{array}", + "type": "interline_equation", + "image_path": "f522d3921ce50be5c1f255681b4d230a0d5ee05230cbbf8ecfa7095a00d83e34.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 160, + 494, + 450, + 512.6666666666666 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 160, + 512.6666666666666, + 450, + 531.3333333333333 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 160, + 531.3333333333333, + 450, + 549.9999999999999 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 505, + 576 + ], + "lines": [ + { + "bbox": [ + 106, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 505, + 566 + ], + "score": 1.0, + "content": "Through this wireframe rendering process, gradients can be propagated backward to both the class", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 564, + 450, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 450, + 577 + ], + "score": 1.0, + "content": "probabilities and geometric parameters of the graphic elements for joint optimization.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 553, + 505, + 577 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "score": 1.0, + "content": "A CNN consisting of 3 convolutional layers, followed by a fully connected layer with sigmoid", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 593, + 442, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 290, + 605 + ], + "score": 1.0, + "content": "activation, is applied to the rasterization layer", + "type": "text" + }, + { + "bbox": [ + 290, + 593, + 298, + 603 + ], + "score": 0.68, + "content": "\\pmb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 593, + 442, + 605 + ], + "score": 1.0, + "content": "to predict real/fake graphic layouts.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 581, + 505, + 605 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 616, + 201, + 628 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 201, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 201, + 630 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 631, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "The implementation is based on TensorFlow (Abadi et al., 2016). The network parameters are", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 642, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 505, + 654 + ], + "score": 1.0, + "content": "initialized from zero-mean Gaussian with standard deviation of 0.02. All the networks are optimized", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 654, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 505, + 666 + ], + "score": 1.0, + "content": "using Adam (Kingma & Ba, 2014) with a fixed learning rate of 0.00002. Detailed architectures can", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 664, + 210, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 210, + 677 + ], + "score": 1.0, + "content": "be found in the appendix.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 631, + 505, + 677 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 685, + 250, + 696 + ], + "lines": [ + { + "bbox": [ + 105, + 684, + 252, + 697 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 252, + 697 + ], + "score": 1.0, + "content": "4.1 MNIST DIGIT GENERATION", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "As a toy problem, we generate MNIST layouts. MNIST is a handwritten digit database consisting", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "of 60,000 training and 10,000 testing images. For each image, we extract the locations of 128", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "randomly-selected foreground pixels as the graphic representation so that digit generation can be", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 272, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 419, + 285 + ], + "score": 1.0, + "content": "formulated as generating of 2D point layouts. In Figure 3a, each image shows", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 419, + 273, + 442, + 283 + ], + "score": 0.9, + "content": "8 \\times 8", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 443, + 272, + 506, + 285 + ], + "score": 1.0, + "content": "digits rendered", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 284, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 295 + ], + "score": 1.0, + "content": "from point layouts. The left and middle images are the generated samples from the LayoutGAN", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "with a relation-based discriminator and a wireframe rendering discriminator, respectively. The right", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 305, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 305, + 505, + 318 + ], + "score": 1.0, + "content": "image shows the digits rendered from ground-truth point layouts. One can see that the LayoutGAN", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 316, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 505, + 329 + ], + "score": 1.0, + "content": "with either discriminator captures various patterns. The wireframe rendering discriminator generates", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 328, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 505, + 339 + ], + "score": 1.0, + "content": "more compact, better-aligned point layouts. For quantitative evaluation, we first train a multilayer", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 407, + 351 + ], + "score": 1.0, + "content": "perceptron network for digit classification, which achieves an accuracy of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 408, + 338, + 440, + 349 + ], + "score": 0.89, + "content": "9 8 . 9 1 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 441, + 338, + 506, + 351 + ], + "score": 1.0, + "content": "on MNIST test", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "set, and then use the classifier to compute the inception score (Salimans et al., 2016) of 10,000", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 360, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 104, + 360, + 506, + 372 + ], + "score": 1.0, + "content": "generated samples. As shown in Table 1, the LayoutGAN with wireframe rendering discriminator", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 371, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 371, + 505, + 384 + ], + "score": 1.0, + "content": "achieves a higher inception score than with the relation-based discriminator. To shed light on the", + "type": "text", + "cross_page": true + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "relational refinement process of LayoutGAN, we mark each input random point with a location-", + "type": "text", + "cross_page": true + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 392, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 506, + 406 + ], + "score": 1.0, + "content": "specific color, and track their refined locations in the generated layouts, as shown in Figure 3b. One", + "type": "text", + "cross_page": true + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 405, + 504, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 504, + 416 + ], + "score": 1.0, + "content": "can see that colors change gradually along the strokes, showing the network learns some contextually", + "type": "text", + "cross_page": true + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 415, + 285, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 285, + 428 + ], + "score": 1.0, + "content": "consistent local displacements of the points.", + "type": "text", + "cross_page": true + } + ], + "index": 25 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 698, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 114, + 69, + 488, + 167 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 69, + 488, + 167 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 69, + 488, + 167 + ], + "spans": [ + { + "bbox": [ + 114, + 69, + 488, + 167 + ], + "score": 0.932, + "type": "image", + "image_path": "a10d4ec280b2dd747078d880f5fbce6b8ddfb30566943603c69b4dd1b57b5bca.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 114, + 69, + 488, + 101.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 114, + 101.66666666666666, + 488, + 134.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 114, + 134.33333333333331, + 488, + 166.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 212, + 171, + 397, + 183 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 212, + 169, + 398, + 186 + ], + "spans": [ + { + "bbox": [ + 212, + 169, + 398, + 186 + ], + "score": 1.0, + "content": "Figure 3: Results on MNIST digit generation.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "table", + "bbox": [ + 148, + 207, + 265, + 264 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 115, + 191, + 297, + 202 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 189, + 299, + 204 + ], + "spans": [ + { + "bbox": [ + 113, + 189, + 299, + 204 + ], + "score": 1.0, + "content": "Table 1: Inception scores for generated digits.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 148, + 207, + 265, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 148, + 207, + 265, + 264 + ], + "spans": [ + { + "bbox": [ + 148, + 207, + 265, + 264 + ], + "score": 0.953, + "html": "
MethodsScore± std
Relation6.53 ± .09
Wireframe7.36 ± .07
Real data9.81 ± .08
", + "type": "table", + "image_path": "8f315d45457c736ab4b69f9651c7c50eab97d2d6ccdb636e41032de59d067050.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 148, + 207, + 265, + 235.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 148, + 235.5, + 265, + 264.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 5.75 + }, + { + "type": "table", + "bbox": [ + 311, + 207, + 496, + 264 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 312, + 191, + 494, + 202 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 313, + 189, + 496, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 496, + 204 + ], + "score": 1.0, + "content": "Table 2: Spatial analysis of document layout.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "table_body", + "bbox": [ + 311, + 207, + 496, + 264 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 311, + 207, + 496, + 264 + ], + "spans": [ + { + "bbox": [ + 311, + 207, + 496, + 264 + ], + "score": 0.964, + "html": "
MethodsOverlap(%)Alignment(%)
Relation1.526.4
Wireframe1.173.4
Real data0.050.5
", + "type": "table", + "image_path": "9c0dce76805dbad046dd504eb714599cc73ff533aab11e371b5a5c722f5b98cd.jpg" + } + ] + } + ], + "index": 9.0, + "virtual_lines": [ + { + "bbox": [ + 311, + 207, + 496, + 221.25 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 311, + 221.25, + 496, + 235.5 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 311, + 235.5, + 496, + 249.75 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 311, + 249.75, + 496, + 264.0 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 7.0 + }, + { + "type": "text", + "bbox": [ + 107, + 272, + 505, + 426 + ], + "lines": [ + { + "bbox": [ + 106, + 272, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 419, + 285 + ], + "score": 1.0, + "content": "formulated as generating of 2D point layouts. In Figure 3a, each image shows", + "type": "text" + }, + { + "bbox": [ + 419, + 273, + 442, + 283 + ], + "score": 0.9, + "content": "8 \\times 8", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 272, + 506, + 285 + ], + "score": 1.0, + "content": "digits rendered", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 284, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 295 + ], + "score": 1.0, + "content": "from point layouts. The left and middle images are the generated samples from the LayoutGAN", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "with a relation-based discriminator and a wireframe rendering discriminator, respectively. The right", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 305, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 305, + 505, + 318 + ], + "score": 1.0, + "content": "image shows the digits rendered from ground-truth point layouts. One can see that the LayoutGAN", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 316, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 505, + 329 + ], + "score": 1.0, + "content": "with either discriminator captures various patterns. The wireframe rendering discriminator generates", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 328, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 505, + 339 + ], + "score": 1.0, + "content": "more compact, better-aligned point layouts. For quantitative evaluation, we first train a multilayer", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 407, + 351 + ], + "score": 1.0, + "content": "perceptron network for digit classification, which achieves an accuracy of", + "type": "text" + }, + { + "bbox": [ + 408, + 338, + 440, + 349 + ], + "score": 0.89, + "content": "9 8 . 9 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 338, + 506, + 351 + ], + "score": 1.0, + "content": "on MNIST test", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "set, and then use the classifier to compute the inception score (Salimans et al., 2016) of 10,000", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 360, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 104, + 360, + 506, + 372 + ], + "score": 1.0, + "content": "generated samples. As shown in Table 1, the LayoutGAN with wireframe rendering discriminator", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 371, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 371, + 505, + 384 + ], + "score": 1.0, + "content": "achieves a higher inception score than with the relation-based discriminator. To shed light on the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "relational refinement process of LayoutGAN, we mark each input random point with a location-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 392, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 506, + 406 + ], + "score": 1.0, + "content": "specific color, and track their refined locations in the generated layouts, as shown in Figure 3b. One", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 405, + 504, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 504, + 416 + ], + "score": 1.0, + "content": "can see that colors change gradually along the strokes, showing the network learns some contextually", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 415, + 285, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 285, + 428 + ], + "score": 1.0, + "content": "consistent local displacements of the points.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 108, + 432, + 277, + 442 + ], + "lines": [ + { + "bbox": [ + 106, + 431, + 279, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 279, + 444 + ], + "score": 1.0, + "content": "4.2 DOCUMENT LAYOUT GENERATION", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 446, + 505, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 444, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 506, + 460 + ], + "score": 1.0, + "content": "One document page consists of a number of regions with different element types, such as heading,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 456, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 470 + ], + "score": 1.0, + "content": "paragraph, table, figure, caption and list. Each region is represented by a bounding box. Modeling", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "score": 1.0, + "content": "layouts of regions is critical for document analysis, retargeting, and synthesis. In real document", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "data, these bounding boxes are often carefully aligned along the canonical axes, and their placement", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "follows some particular patterns, such as heading always appearing above paragraph or table. Some", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 501, + 282, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 282, + 513 + ], + "score": 1.0, + "content": "example layouts are shown in the appendix.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 518, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "score": 1.0, + "content": "In this experiment, we focus on one-column layouts with no more than 9 bounding boxes that may", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 528, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 542 + ], + "score": 1.0, + "content": "belong to 6 possible classes as mentioned above. We are interested in how these layout patterns can", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "be captured by our network. For training data, we collect totally around 25,000 layouts from real", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 550, + 504, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 504, + 563 + ], + "score": 1.0, + "content": "documents. We also implemented a baseline method that represents the layout by semantic masks", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "as used in Yang et al. (2017); Deka et al. (2017). Specifically, we render all elements into masks", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "and train a DCGAN (Radford et al., 2015) to generate mask layouts in pixels. Then we extract the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "connected mask regions of each class and output enclosing bounding boxes of all regions. Figure 4", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 595, + 504, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 504, + 606 + ], + "score": 1.0, + "content": "shows some representative generation results (each row consists of 6 samples; high-resolution results", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "can be found in the Appendix). The first row shows the results from DCGAN. DCGAN mixes a", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "layout and its rendering in the generation process. When convolutional layers in DCGAN fails to", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "perfectly replicate the rendering process (which is usually the case), it generates fuzzy and noisy", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "label maps, making the extracted bounding boxes less consistent and less accurate. The third row", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "shows the results from LayoutGAN with wireframe rendering discriminator. We retrieve the most", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "similar real layouts from the training set in the last row as references. It can be seen that LayoutGAN", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "score": 1.0, + "content": "can well capture different document layout patterns with clear definitions of instance-level semantic", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 681, + 175, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 175, + 695 + ], + "score": 1.0, + "content": "bounding boxes.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "To validate the advantage of wireframe rendering discriminator, we compare the generated results", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 711, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 721 + ], + "score": 1.0, + "content": "with those from the LayoutGAN with relation-based discriminator. As shown in the second row of", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 720, + 506, + 735 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 735 + ], + "score": 1.0, + "content": "Figure 4, the latter can also capture different layout patterns but sometimes suffers from overlapping", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 50 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 114, + 69, + 488, + 167 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 69, + 488, + 167 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 69, + 488, + 167 + ], + "spans": [ + { + "bbox": [ + 114, + 69, + 488, + 167 + ], + "score": 0.932, + "type": "image", + "image_path": "a10d4ec280b2dd747078d880f5fbce6b8ddfb30566943603c69b4dd1b57b5bca.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 114, + 69, + 488, + 101.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 114, + 101.66666666666666, + 488, + 134.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 114, + 134.33333333333331, + 488, + 166.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 212, + 171, + 397, + 183 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 212, + 169, + 398, + 186 + ], + "spans": [ + { + "bbox": [ + 212, + 169, + 398, + 186 + ], + "score": 1.0, + "content": "Figure 3: Results on MNIST digit generation.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "table", + "bbox": [ + 148, + 207, + 265, + 264 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 115, + 191, + 297, + 202 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 189, + 299, + 204 + ], + "spans": [ + { + "bbox": [ + 113, + 189, + 299, + 204 + ], + "score": 1.0, + "content": "Table 1: Inception scores for generated digits.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 148, + 207, + 265, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 148, + 207, + 265, + 264 + ], + "spans": [ + { + "bbox": [ + 148, + 207, + 265, + 264 + ], + "score": 0.953, + "html": "
MethodsScore± std
Relation6.53 ± .09
Wireframe7.36 ± .07
Real data9.81 ± .08
", + "type": "table", + "image_path": "8f315d45457c736ab4b69f9651c7c50eab97d2d6ccdb636e41032de59d067050.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 148, + 207, + 265, + 235.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 148, + 235.5, + 265, + 264.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 5.75 + }, + { + "type": "table", + "bbox": [ + 311, + 207, + 496, + 264 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 312, + 191, + 494, + 202 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 313, + 189, + 496, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 496, + 204 + ], + "score": 1.0, + "content": "Table 2: Spatial analysis of document layout.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "table_body", + "bbox": [ + 311, + 207, + 496, + 264 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 311, + 207, + 496, + 264 + ], + "spans": [ + { + "bbox": [ + 311, + 207, + 496, + 264 + ], + "score": 0.964, + "html": "
MethodsOverlap(%)Alignment(%)
Relation1.526.4
Wireframe1.173.4
Real data0.050.5
", + "type": "table", + "image_path": "9c0dce76805dbad046dd504eb714599cc73ff533aab11e371b5a5c722f5b98cd.jpg" + } + ] + } + ], + "index": 9.0, + "virtual_lines": [ + { + "bbox": [ + 311, + 207, + 496, + 221.25 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 311, + 221.25, + 496, + 235.5 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 311, + 235.5, + 496, + 249.75 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 311, + 249.75, + 496, + 264.0 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 7.0 + }, + { + "type": "text", + "bbox": [ + 107, + 272, + 505, + 426 + ], + "lines": [], + "index": 18.5, + "bbox_fs": [ + 104, + 272, + 506, + 428 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 432, + 277, + 442 + ], + "lines": [ + { + "bbox": [ + 106, + 431, + 279, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 279, + 444 + ], + "score": 1.0, + "content": "4.2 DOCUMENT LAYOUT GENERATION", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 446, + 505, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 444, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 506, + 460 + ], + "score": 1.0, + "content": "One document page consists of a number of regions with different element types, such as heading,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 456, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 470 + ], + "score": 1.0, + "content": "paragraph, table, figure, caption and list. Each region is represented by a bounding box. Modeling", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "score": 1.0, + "content": "layouts of regions is critical for document analysis, retargeting, and synthesis. In real document", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "data, these bounding boxes are often carefully aligned along the canonical axes, and their placement", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "follows some particular patterns, such as heading always appearing above paragraph or table. Some", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 501, + 282, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 282, + 513 + ], + "score": 1.0, + "content": "example layouts are shown in the appendix.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 444, + 506, + 513 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 518, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "score": 1.0, + "content": "In this experiment, we focus on one-column layouts with no more than 9 bounding boxes that may", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 528, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 542 + ], + "score": 1.0, + "content": "belong to 6 possible classes as mentioned above. We are interested in how these layout patterns can", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "be captured by our network. For training data, we collect totally around 25,000 layouts from real", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 550, + 504, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 504, + 563 + ], + "score": 1.0, + "content": "documents. We also implemented a baseline method that represents the layout by semantic masks", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "as used in Yang et al. (2017); Deka et al. (2017). Specifically, we render all elements into masks", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "and train a DCGAN (Radford et al., 2015) to generate mask layouts in pixels. Then we extract the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "connected mask regions of each class and output enclosing bounding boxes of all regions. Figure 4", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 595, + 504, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 504, + 606 + ], + "score": 1.0, + "content": "shows some representative generation results (each row consists of 6 samples; high-resolution results", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "can be found in the Appendix). The first row shows the results from DCGAN. DCGAN mixes a", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "layout and its rendering in the generation process. When convolutional layers in DCGAN fails to", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "perfectly replicate the rendering process (which is usually the case), it generates fuzzy and noisy", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "label maps, making the extracted bounding boxes less consistent and less accurate. The third row", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "shows the results from LayoutGAN with wireframe rendering discriminator. We retrieve the most", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "similar real layouts from the training set in the last row as references. It can be seen that LayoutGAN", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "score": 1.0, + "content": "can well capture different document layout patterns with clear definitions of instance-level semantic", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 681, + 175, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 175, + 695 + ], + "score": 1.0, + "content": "bounding boxes.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 516, + 506, + 695 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "To validate the advantage of wireframe rendering discriminator, we compare the generated results", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 711, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 721 + ], + "score": 1.0, + "content": "with those from the LayoutGAN with relation-based discriminator. As shown in the second row of", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 720, + 506, + 735 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 735 + ], + "score": 1.0, + "content": "Figure 4, the latter can also capture different layout patterns but sometimes suffers from overlapping", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "and misalignment issues. As the bounding boxes in real document layouts are either left- or center-", + "type": "text", + "cross_page": true + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 279, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 506, + 292 + ], + "score": 1.0, + "content": "aligned with no overlaps with each other (except for captions with figures or tables), we propose two", + "type": "text", + "cross_page": true + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 290, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 302 + ], + "score": 1.0, + "content": "metrics to quantitatively measure the quality of the generated layouts. The first one is overlap index,", + "type": "text", + "cross_page": true + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 300, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 505, + 315 + ], + "score": 1.0, + "content": "which is the percentage of total overlapping area among any two bounding boxes inside the whole", + "type": "text", + "cross_page": true + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 104, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "page. The second one is alignment index which is calculated by finding the minimum standard de-", + "type": "text", + "cross_page": true + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 323, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 505, + 335 + ], + "score": 1.0, + "content": "viation of either left or center coordinates of all the bounding boxes. Table 2 provides quantitative", + "type": "text", + "cross_page": true + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "comparisons of real layouts and synthesized layouts from LayoutGAN with two different discrimi-", + "type": "text", + "cross_page": true + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 345, + 504, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 504, + 357 + ], + "score": 1.0, + "content": "nators. One can see that LayoutGAN with wireframe rendering discriminator achieves lower value", + "type": "text", + "cross_page": true + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 356, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 368 + ], + "score": 1.0, + "content": "in both overlapping index and alignment index than the one with relation-based discriminator, vali-", + "type": "text", + "cross_page": true + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 366, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 379 + ], + "score": 1.0, + "content": "dating the superiority of the proposed wireframe rendering method for layout generation. A similar", + "type": "text", + "cross_page": true + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 505, + 390 + ], + "score": 1.0, + "content": "conclusion can also be drawn by analyzing the loss functions. We add shift perturbations to the", + "type": "text", + "cross_page": true + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 388, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 505, + 401 + ], + "score": 1.0, + "content": "bounding boxes of real layouts and feed them to both discriminators to examine their loss behav-", + "type": "text", + "cross_page": true + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "score": 1.0, + "content": "iors. Figure 6 visualizes the loss landscape of both discriminators corresponding to different extent", + "type": "text", + "cross_page": true + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "of shift perturbation. One can see that the loss surface w.r.t. shift perturbation of the wireframe", + "type": "text", + "cross_page": true + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 421, + 455, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 455, + 434 + ], + "score": 1.0, + "content": "rendering discriminator is much smoother than the one of relation-based discriminator.", + "type": "text", + "cross_page": true + } + ], + "index": 42 + } + ], + "index": 50, + "bbox_fs": [ + 105, + 698, + 506, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 73, + 302, + 246 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 73, + 302, + 246 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 73, + 302, + 246 + ], + "spans": [ + { + "bbox": [ + 109, + 73, + 302, + 246 + ], + "score": 0.973, + "type": "image", + "image_path": "1b7f0b9374988295ff6466770bc433d9b1b8193220fd8eca8a0aba65c44d495b.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 109, + 73, + 302, + 86.3076923076923 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 86.3076923076923, + 302, + 99.61538461538461 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 99.61538461538461, + 302, + 112.92307692307692 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 109, + 112.92307692307692, + 302, + 126.23076923076923 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 109, + 126.23076923076923, + 302, + 139.53846153846155 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 109, + 139.53846153846155, + 302, + 152.84615384615387 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 109, + 152.84615384615387, + 302, + 166.1538461538462 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 109, + 166.1538461538462, + 302, + 179.4615384615385 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 109, + 179.4615384615385, + 302, + 192.76923076923083 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 109, + 192.76923076923083, + 302, + 206.07692307692315 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 109, + 206.07692307692315, + 302, + 219.38461538461547 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 109, + 219.38461538461547, + 302, + 232.6923076923078 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 109, + 232.6923076923078, + 302, + 246.0000000000001 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 126, + 251, + 286, + 262 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 249, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 124, + 249, + 288, + 264 + ], + "score": 1.0, + "content": "Figure 4: Document layout comparison.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + } + ], + "index": 16.0 + }, + { + "type": "image", + "bbox": [ + 322, + 66, + 484, + 246 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 322, + 66, + 484, + 246 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 322, + 66, + 484, + 246 + ], + "spans": [ + { + "bbox": [ + 322, + 66, + 484, + 246 + ], + "score": 0.972, + "type": "image", + "image_path": "eb1c35ca53cdf5b316f1c04cd6b7b8c2044d49e7dbdf44c72cdc26db71c5a679.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 322, + 66, + 484, + 79.84615384615384 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 322, + 79.84615384615384, + 484, + 93.69230769230768 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 322, + 93.69230769230768, + 484, + 107.53846153846152 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 322, + 107.53846153846152, + 484, + 121.38461538461536 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 322, + 121.38461538461536, + 484, + 135.2307692307692 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 322, + 135.2307692307692, + 484, + 149.07692307692304 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 322, + 149.07692307692304, + 484, + 162.92307692307688 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 322, + 162.92307692307688, + 484, + 176.76923076923072 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 322, + 176.76923076923072, + 484, + 190.61538461538456 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 322, + 190.61538461538456, + 484, + 204.4615384615384 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 322, + 204.4615384615384, + 484, + 218.30769230769224 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 322, + 218.30769230769224, + 484, + 232.15384615384608 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 322, + 232.15384615384608, + 484, + 245.99999999999991 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 318, + 251, + 491, + 262 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 316, + 249, + 492, + 264 + ], + "spans": [ + { + "bbox": [ + 316, + 249, + 492, + 264 + ], + "score": 1.0, + "content": "Figure 5: Clipart abstract scene generation.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + } + ], + "index": 23.0 + }, + { + "type": "text", + "bbox": [ + 106, + 268, + 505, + 433 + ], + "lines": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "and misalignment issues. As the bounding boxes in real document layouts are either left- or center-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 279, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 506, + 292 + ], + "score": 1.0, + "content": "aligned with no overlaps with each other (except for captions with figures or tables), we propose two", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 290, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 302 + ], + "score": 1.0, + "content": "metrics to quantitatively measure the quality of the generated layouts. The first one is overlap index,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 300, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 505, + 315 + ], + "score": 1.0, + "content": "which is the percentage of total overlapping area among any two bounding boxes inside the whole", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 104, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "page. The second one is alignment index which is calculated by finding the minimum standard de-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 323, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 505, + 335 + ], + "score": 1.0, + "content": "viation of either left or center coordinates of all the bounding boxes. Table 2 provides quantitative", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "comparisons of real layouts and synthesized layouts from LayoutGAN with two different discrimi-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 345, + 504, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 504, + 357 + ], + "score": 1.0, + "content": "nators. One can see that LayoutGAN with wireframe rendering discriminator achieves lower value", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 356, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 368 + ], + "score": 1.0, + "content": "in both overlapping index and alignment index than the one with relation-based discriminator, vali-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 366, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 379 + ], + "score": 1.0, + "content": "dating the superiority of the proposed wireframe rendering method for layout generation. A similar", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 505, + 390 + ], + "score": 1.0, + "content": "conclusion can also be drawn by analyzing the loss functions. We add shift perturbations to the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 388, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 505, + 401 + ], + "score": 1.0, + "content": "bounding boxes of real layouts and feed them to both discriminators to examine their loss behav-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "score": 1.0, + "content": "iors. Figure 6 visualizes the loss landscape of both discriminators corresponding to different extent", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "of shift perturbation. One can see that the loss surface w.r.t. shift perturbation of the wireframe", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 421, + 455, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 455, + 434 + ], + "score": 1.0, + "content": "rendering discriminator is much smoother than the one of relation-based discriminator.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 123, + 438, + 268, + 504 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 438, + 268, + 504 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 123, + 438, + 268, + 504 + ], + "spans": [ + { + "bbox": [ + 123, + 438, + 268, + 504 + ], + "score": 0.955, + "type": "image", + "image_path": "c03606f88f5395b716af09a8b97c3528cd6c04cee92c55e211f4c586467d2a12.jpg" + } + ] + } + ], + "index": 43.5, + "virtual_lines": [ + { + "bbox": [ + 123, + 438, + 268, + 471.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 123, + 471.0, + 268, + 504.0 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 115, + 508, + 276, + 519 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 114, + 507, + 279, + 521 + ], + "spans": [ + { + "bbox": [ + 114, + 507, + 279, + 521 + ], + "score": 1.0, + "content": "Figure 6: Discriminator loss landscapes.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + } + ], + "index": 44.25 + }, + { + "type": "table", + "bbox": [ + 287, + 459, + 502, + 516 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 286, + 443, + 503, + 455 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 285, + 441, + 504, + 457 + ], + "spans": [ + { + "bbox": [ + 285, + 441, + 504, + 457 + ], + "score": 1.0, + "content": "Table 3: User study results for Clipart abstract scenes.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + }, + { + "type": "table_body", + "bbox": [ + 287, + 459, + 502, + 516 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 287, + 459, + 502, + 516 + ], + "spans": [ + { + "bbox": [ + 287, + 459, + 502, + 516 + ], + "score": 0.974, + "html": "
MethodsExcellent (%)Fair (%)Poor (%)
DCGAN6.738.854.5
Relation17.250.332.5
Wireframe37.348.014.7
", + "type": "table", + "image_path": "b7a61251481b5b2779dfd957179eec7373977ebea52cec2a995c07590f798c7c.jpg" + } + ] + } + ], + "index": 48.5, + "virtual_lines": [ + { + "bbox": [ + 287, + 459, + 502, + 473.25 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 287, + 473.25, + 502, + 487.5 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 287, + 487.5, + 502, + 501.75 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 287, + 501.75, + 502, + 516.0 + ], + "spans": [], + "index": 50 + } + ] + } + ], + "index": 47.25 + }, + { + "type": "title", + "bbox": [ + 106, + 529, + 303, + 540 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 306, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 306, + 541 + ], + "score": 1.0, + "content": "4.3 CLIPART ABSTRACT SCENE GENERATION", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 51 + }, + { + "type": "text", + "bbox": [ + 106, + 545, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 545, + 504, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 504, + 558 + ], + "score": 1.0, + "content": "We now consider generating scenes composed of a set of specific clipart elements. We use the ab-", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "score": 1.0, + "content": "stract scene dataset (Zitnick et al., 2016) including boy, girl, glasses, hat, sun, and tree elements.", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 567, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 580 + ], + "score": 1.0, + "content": "To synthesize a reasonable abstract scene, we first use LayoutGAN to generate the layout of inner", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 104, + 577, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 104, + 577, + 505, + 592 + ], + "score": 1.0, + "content": "scene elements by representing each of them as a labeled bounding box with normalized center co-", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "score": 1.0, + "content": "ordinates, width and height, and flip indicator. Then, we render the corresponding clipart image", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "of each scene element onto a background image according to the predicted position (center coordi-", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 104, + 610, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 104, + 610, + 505, + 625 + ], + "score": 1.0, + "content": "nates), scale (width and height) and flip, forming a synthesized abstract scene. It is a challenging", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "score": 1.0, + "content": "task as accurate pairwise or higher-order relations of objects are required for synthesizing reason-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 633, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 647 + ], + "score": 1.0, + "content": "able scenes, for example, the glasses/hat should be exactly on the eyes/head of a boy/girl with proper", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 104, + 643, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 104, + 643, + 505, + 658 + ], + "score": 1.0, + "content": "scale and orientation. In Figure 5, the first three rows shows the generated layouts and corresponding", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 104, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "rendered scenes by DCGAN, LayoutGAN with relation-based discriminator and LayoutGAN with", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "wireframe rendering discriminator respectively. The last row shows some samples rendered from", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "ground-truth scene layouts. It can be seen that, compared to DCGAN and the LayoutGAN with", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "relation-based discriminator, the one with wireframe rendering discriminator can capture pairwise", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "relations of objects precisely, forming more meaningful scene layouts (for example, the glasses/hat", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 105, + 710, + 504, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 504, + 723 + ], + "score": 1.0, + "content": "accurately lies on the eyes/head of the person with varying scales and flips), validating the superior-", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "ity of the proposed wireframe rendering strategy. In addition, we conduct a user study involving 20", + "type": "text" + } + ], + "index": 68 + } + ], + "index": 60 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 73, + 302, + 246 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 73, + 302, + 246 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 73, + 302, + 246 + ], + "spans": [ + { + "bbox": [ + 109, + 73, + 302, + 246 + ], + "score": 0.973, + "type": "image", + "image_path": "1b7f0b9374988295ff6466770bc433d9b1b8193220fd8eca8a0aba65c44d495b.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 109, + 73, + 302, + 86.3076923076923 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 86.3076923076923, + 302, + 99.61538461538461 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 99.61538461538461, + 302, + 112.92307692307692 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 109, + 112.92307692307692, + 302, + 126.23076923076923 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 109, + 126.23076923076923, + 302, + 139.53846153846155 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 109, + 139.53846153846155, + 302, + 152.84615384615387 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 109, + 152.84615384615387, + 302, + 166.1538461538462 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 109, + 166.1538461538462, + 302, + 179.4615384615385 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 109, + 179.4615384615385, + 302, + 192.76923076923083 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 109, + 192.76923076923083, + 302, + 206.07692307692315 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 109, + 206.07692307692315, + 302, + 219.38461538461547 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 109, + 219.38461538461547, + 302, + 232.6923076923078 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 109, + 232.6923076923078, + 302, + 246.0000000000001 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 126, + 251, + 286, + 262 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 249, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 124, + 249, + 288, + 264 + ], + "score": 1.0, + "content": "Figure 4: Document layout comparison.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + } + ], + "index": 16.0 + }, + { + "type": "image", + "bbox": [ + 322, + 66, + 484, + 246 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 322, + 66, + 484, + 246 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 322, + 66, + 484, + 246 + ], + "spans": [ + { + "bbox": [ + 322, + 66, + 484, + 246 + ], + "score": 0.972, + "type": "image", + "image_path": "eb1c35ca53cdf5b316f1c04cd6b7b8c2044d49e7dbdf44c72cdc26db71c5a679.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 322, + 66, + 484, + 79.84615384615384 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 322, + 79.84615384615384, + 484, + 93.69230769230768 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 322, + 93.69230769230768, + 484, + 107.53846153846152 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 322, + 107.53846153846152, + 484, + 121.38461538461536 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 322, + 121.38461538461536, + 484, + 135.2307692307692 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 322, + 135.2307692307692, + 484, + 149.07692307692304 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 322, + 149.07692307692304, + 484, + 162.92307692307688 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 322, + 162.92307692307688, + 484, + 176.76923076923072 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 322, + 176.76923076923072, + 484, + 190.61538461538456 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 322, + 190.61538461538456, + 484, + 204.4615384615384 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 322, + 204.4615384615384, + 484, + 218.30769230769224 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 322, + 218.30769230769224, + 484, + 232.15384615384608 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 322, + 232.15384615384608, + 484, + 245.99999999999991 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 318, + 251, + 491, + 262 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 316, + 249, + 492, + 264 + ], + "spans": [ + { + "bbox": [ + 316, + 249, + 492, + 264 + ], + "score": 1.0, + "content": "Figure 5: Clipart abstract scene generation.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + } + ], + "index": 23.0 + }, + { + "type": "text", + "bbox": [ + 106, + 268, + 505, + 433 + ], + "lines": [], + "index": 35, + "bbox_fs": [ + 104, + 268, + 506, + 434 + ], + "lines_deleted": true + }, + { + "type": "image", + "bbox": [ + 123, + 438, + 268, + 504 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 438, + 268, + 504 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 123, + 438, + 268, + 504 + ], + "spans": [ + { + "bbox": [ + 123, + 438, + 268, + 504 + ], + "score": 0.955, + "type": "image", + "image_path": "c03606f88f5395b716af09a8b97c3528cd6c04cee92c55e211f4c586467d2a12.jpg" + } + ] + } + ], + "index": 43.5, + "virtual_lines": [ + { + "bbox": [ + 123, + 438, + 268, + 471.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 123, + 471.0, + 268, + 504.0 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 115, + 508, + 276, + 519 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 114, + 507, + 279, + 521 + ], + "spans": [ + { + "bbox": [ + 114, + 507, + 279, + 521 + ], + "score": 1.0, + "content": "Figure 6: Discriminator loss landscapes.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + } + ], + "index": 44.25 + }, + { + "type": "table", + "bbox": [ + 287, + 459, + 502, + 516 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 286, + 443, + 503, + 455 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 285, + 441, + 504, + 457 + ], + "spans": [ + { + "bbox": [ + 285, + 441, + 504, + 457 + ], + "score": 1.0, + "content": "Table 3: User study results for Clipart abstract scenes.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + }, + { + "type": "table_body", + "bbox": [ + 287, + 459, + 502, + 516 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 287, + 459, + 502, + 516 + ], + "spans": [ + { + "bbox": [ + 287, + 459, + 502, + 516 + ], + "score": 0.974, + "html": "
MethodsExcellent (%)Fair (%)Poor (%)
DCGAN6.738.854.5
Relation17.250.332.5
Wireframe37.348.014.7
", + "type": "table", + "image_path": "b7a61251481b5b2779dfd957179eec7373977ebea52cec2a995c07590f798c7c.jpg" + } + ] + } + ], + "index": 48.5, + "virtual_lines": [ + { + "bbox": [ + 287, + 459, + 502, + 473.25 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 287, + 473.25, + 502, + 487.5 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 287, + 487.5, + 502, + 501.75 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 287, + 501.75, + 502, + 516.0 + ], + "spans": [], + "index": 50 + } + ] + } + ], + "index": 47.25 + }, + { + "type": "title", + "bbox": [ + 106, + 529, + 303, + 540 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 306, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 306, + 541 + ], + "score": 1.0, + "content": "4.3 CLIPART ABSTRACT SCENE GENERATION", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 51 + }, + { + "type": "text", + "bbox": [ + 106, + 545, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 545, + 504, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 504, + 558 + ], + "score": 1.0, + "content": "We now consider generating scenes composed of a set of specific clipart elements. We use the ab-", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "score": 1.0, + "content": "stract scene dataset (Zitnick et al., 2016) including boy, girl, glasses, hat, sun, and tree elements.", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 567, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 580 + ], + "score": 1.0, + "content": "To synthesize a reasonable abstract scene, we first use LayoutGAN to generate the layout of inner", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 104, + 577, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 104, + 577, + 505, + 592 + ], + "score": 1.0, + "content": "scene elements by representing each of them as a labeled bounding box with normalized center co-", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "score": 1.0, + "content": "ordinates, width and height, and flip indicator. Then, we render the corresponding clipart image", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "of each scene element onto a background image according to the predicted position (center coordi-", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 104, + 610, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 104, + 610, + 505, + 625 + ], + "score": 1.0, + "content": "nates), scale (width and height) and flip, forming a synthesized abstract scene. It is a challenging", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "score": 1.0, + "content": "task as accurate pairwise or higher-order relations of objects are required for synthesizing reason-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 633, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 647 + ], + "score": 1.0, + "content": "able scenes, for example, the glasses/hat should be exactly on the eyes/head of a boy/girl with proper", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 104, + 643, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 104, + 643, + 505, + 658 + ], + "score": 1.0, + "content": "scale and orientation. In Figure 5, the first three rows shows the generated layouts and corresponding", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 104, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "rendered scenes by DCGAN, LayoutGAN with relation-based discriminator and LayoutGAN with", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "wireframe rendering discriminator respectively. The last row shows some samples rendered from", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "ground-truth scene layouts. It can be seen that, compared to DCGAN and the LayoutGAN with", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "relation-based discriminator, the one with wireframe rendering discriminator can capture pairwise", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "relations of objects precisely, forming more meaningful scene layouts (for example, the glasses/hat", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 105, + 710, + 504, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 504, + 723 + ], + "score": 1.0, + "content": "accurately lies on the eyes/head of the person with varying scales and flips), validating the superior-", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "ity of the proposed wireframe rendering strategy. In addition, we conduct a user study involving 20", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 105, + 323, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 506, + 336 + ], + "score": 1.0, + "content": "participants for a subjective evaluation of the results. Specifically, we randomly sample 30 abstract", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "scenes synthesized by each of the above three models. A subject is asked to rate the synthetic scenes", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "in terms of three scores (Excellent, Fair, Poor) by following three criteria: 1) whether they have", + "type": "text", + "cross_page": true + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 355, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 505, + 370 + ], + "score": 1.0, + "content": "coherent overall structures, 2) whether different objects are placed in reasonable relative positions,", + "type": "text", + "cross_page": true + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 367, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 381 + ], + "score": 1.0, + "content": "scales and flips, and 3) whether there are duplicate objects. We compute the rating percentage of all", + "type": "text", + "cross_page": true + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 379, + 504, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 504, + 390 + ], + "score": 1.0, + "content": "sampled scenes for each model. As shown in Table 3, the results of LayoutGAN with the wireframe", + "type": "text", + "cross_page": true + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "rendering discriminator receive better ratings, which suggests that they are preferred by participants.", + "type": "text", + "cross_page": true + } + ], + "index": 25 + } + ], + "index": 60, + "bbox_fs": [ + 104, + 545, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 70, + 282, + 145 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 70, + 282, + 145 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 70, + 282, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 282, + 145 + ], + "score": 0.964, + "type": "image", + "image_path": "e72b1bce01905f2f6eebea87be8c192d62b7b669e4b3971246872afda07ec9e0.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 106, + 70, + 282, + 85.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 85.0, + 282, + 100.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 100.0, + 282, + 115.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 106, + 115.0, + 282, + 130.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 130.0, + 282, + 145.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 114, + 151, + 279, + 163 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 150, + 279, + 164 + ], + "spans": [ + { + "bbox": [ + 113, + 150, + 279, + 164 + ], + "score": 1.0, + "content": "Figure 7: Optimizing perturbed tangram.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "image", + "bbox": [ + 289, + 63, + 483, + 146 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 289, + 63, + 483, + 146 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 289, + 63, + 483, + 146 + ], + "spans": [ + { + "bbox": [ + 289, + 63, + 483, + 146 + ], + "score": 0.963, + "type": "image", + "image_path": "58a966ad6f41348dd87f9734ec90a918cef932270ec4c5e5b18dd128e1566253.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 289, + 63, + 483, + 76.83333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 289, + 76.83333333333333, + 483, + 90.66666666666666 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 289, + 90.66666666666666, + 483, + 104.49999999999999 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 289, + 104.49999999999999, + 483, + 118.33333333333331 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 289, + 118.33333333333331, + 483, + 132.16666666666666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 289, + 132.16666666666666, + 483, + 146.0 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 8.5 + }, + { + "type": "image", + "bbox": [ + 130, + 173, + 477, + 305 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 249, + 165, + 331, + 172 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 248, + 164, + 331, + 173 + ], + "spans": [ + { + "bbox": [ + 248, + 164, + 331, + 173 + ], + "score": 1.0, + "content": "Progressive Generated Layouts", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "image_caption", + "bbox": [ + 297, + 151, + 478, + 173 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 295, + 149, + 478, + 165 + ], + "spans": [ + { + "bbox": [ + 295, + 149, + 478, + 165 + ], + "score": 1.0, + "content": "Figure 8: Baseline results of tangram design.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 443, + 163, + 479, + 174 + ], + "spans": [ + { + "bbox": [ + 443, + 163, + 479, + 174 + ], + "score": 1.0, + "content": "Real Layouts", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.0 + }, + { + "type": "image_body", + "bbox": [ + 130, + 173, + 477, + 305 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 130, + 173, + 477, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 173, + 477, + 305 + ], + "score": 0.961, + "type": "image", + "image_path": "00c0fd1c77d7eebaffc11c0b612d9811737fc843209a51ca7cd17d2b896b2ceb.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 130, + 173, + 477, + 217.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 130, + 217.0, + 477, + 261.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 130, + 261.0, + 477, + 305.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 125, + 309, + 480, + 321 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 128, + 308, + 483, + 322 + ], + "spans": [ + { + "bbox": [ + 128, + 308, + 483, + 322 + ], + "score": 1.0, + "content": "Figure 9: Training progression of tangram graphic design generation (from left to right).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 324, + 505, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 323, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 506, + 336 + ], + "score": 1.0, + "content": "participants for a subjective evaluation of the results. Specifically, we randomly sample 30 abstract", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "scenes synthesized by each of the above three models. A subject is asked to rate the synthetic scenes", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "in terms of three scores (Excellent, Fair, Poor) by following three criteria: 1) whether they have", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 355, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 505, + 370 + ], + "score": 1.0, + "content": "coherent overall structures, 2) whether different objects are placed in reasonable relative positions,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 367, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 381 + ], + "score": 1.0, + "content": "scales and flips, and 3) whether there are duplicate objects. We compute the rating percentage of all", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 379, + 504, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 504, + 390 + ], + "score": 1.0, + "content": "sampled scenes for each model. As shown in Table 3, the results of LayoutGAN with the wireframe", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "rendering discriminator receive better ratings, which suggests that they are preferred by participants.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 108, + 406, + 251, + 416 + ], + "lines": [ + { + "bbox": [ + 106, + 405, + 253, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 253, + 418 + ], + "score": 1.0, + "content": "4.4 TANGRAM GRAPHIC DESIGN", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 417, + 505, + 647 + ], + "lines": [ + { + "bbox": [ + 106, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "A tangram is a geometric puzzle aiming to form specific shapes using seven pieces of 2D shapes", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 427, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 505, + 441 + ], + "score": 1.0, + "content": "without overlaps. The seven pieces include two large right triangles, one medium right triangle, two", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "small right triangles, one square and one parallelogram, which can be assembled to form a square", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 448, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 506, + 463 + ], + "score": 1.0, + "content": "of side one unit and having area one square unit when choosing a unit of measurement. We collect", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 461, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 506, + 474 + ], + "score": 1.0, + "content": "149 tangram graphic designs including animals, people and objects. In our experiments, we consider", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 471, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 506, + 485 + ], + "score": 1.0, + "content": "eight rotation/reflection poses for each piece. Given the seven pieces with each initialized by random", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "score": 1.0, + "content": "location and ground-truth pose and class, the LayoutGAN with wireframe rendering discriminator is", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 492, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 104, + 492, + 506, + 508 + ], + "score": 1.0, + "content": "trained to refine their configurations jointly to form reasonable layouts. Note that it is a challenging", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 505, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 518 + ], + "score": 1.0, + "content": "task due to the complex configurations and limited data. We perform two experiments. The first", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 515, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 104, + 515, + 506, + 529 + ], + "score": 1.0, + "content": "one is a perturbation recovery test, in which we first add some random spatial perturbations to seven", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 526, + 506, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 506, + 539 + ], + "score": 1.0, + "content": "pieces of real layouts and train a LayoutGAN to recover the original layouts. Figure 7 shows that", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "our network is able to pull back the displaced shapes to the right locations, confirming that the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 549, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 506, + 561 + ], + "score": 1.0, + "content": "network successfully figures out the relations between different graphic elements. We further train", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "another LayoutGAN to generate tangram designs from purely random initialization. In Figure 9", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 570, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 505, + 583 + ], + "score": 1.0, + "content": "(high-resolution results can be found in the appendix), each row represents the sampled progressive", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 581, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 104, + 581, + 506, + 594 + ], + "score": 1.0, + "content": "generated results and the retrieved similar real tangrams. We also present some generated results", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "score": 1.0, + "content": "of DCGAN and a sequential model similar to Wang et al. (2018a) for comparison. One can see", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "score": 1.0, + "content": "from Figure 8 that DCGAN cannot well model the spatial relations among different shapes while", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 614, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 505, + 627 + ], + "score": 1.0, + "content": "the sequential model suffers from accumulated errors. In contrast, the LayoutGAN can generate", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 625, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 506, + 638 + ], + "score": 1.0, + "content": "meaningful tangrams like fox and person, although others may be hard to interpret, as shown in the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 637, + 176, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 176, + 648 + ], + "score": 1.0, + "content": "penultimate row.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 37 + }, + { + "type": "title", + "bbox": [ + 107, + 653, + 196, + 664 + ], + "lines": [ + { + "bbox": [ + 104, + 650, + 198, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 650, + 198, + 668 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "In this paper, we have proposed a novel LayoutGAN for generating layouts of relational graphic", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "elements. Different from traditional GANs that generate images in pixel-level, the LayoutGAN can", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 686, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 506, + 701 + ], + "score": 1.0, + "content": "directly output a set of relational graphic elements. A novel differentiable wireframe rendering layer", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "was proposed to rasterize the generated graphic elements to wireframe images, making it feasible", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "to leverage CNNs as discriminator for better layout optimization from visual domain. Future works", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 720, + 486, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 486, + 733 + ], + "score": 1.0, + "content": "include adding a content representation, such as text, icon and picture to each graphic element.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 51.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 70, + 282, + 145 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 70, + 282, + 145 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 70, + 282, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 282, + 145 + ], + "score": 0.964, + "type": "image", + "image_path": "e72b1bce01905f2f6eebea87be8c192d62b7b669e4b3971246872afda07ec9e0.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 106, + 70, + 282, + 85.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 85.0, + 282, + 100.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 100.0, + 282, + 115.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 106, + 115.0, + 282, + 130.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 130.0, + 282, + 145.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 114, + 151, + 279, + 163 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 150, + 279, + 164 + ], + "spans": [ + { + "bbox": [ + 113, + 150, + 279, + 164 + ], + "score": 1.0, + "content": "Figure 7: Optimizing perturbed tangram.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "image", + "bbox": [ + 289, + 63, + 483, + 146 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 289, + 63, + 483, + 146 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 289, + 63, + 483, + 146 + ], + "spans": [ + { + "bbox": [ + 289, + 63, + 483, + 146 + ], + "score": 0.963, + "type": "image", + "image_path": "58a966ad6f41348dd87f9734ec90a918cef932270ec4c5e5b18dd128e1566253.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 289, + 63, + 483, + 76.83333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 289, + 76.83333333333333, + 483, + 90.66666666666666 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 289, + 90.66666666666666, + 483, + 104.49999999999999 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 289, + 104.49999999999999, + 483, + 118.33333333333331 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 289, + 118.33333333333331, + 483, + 132.16666666666666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 289, + 132.16666666666666, + 483, + 146.0 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 8.5 + }, + { + "type": "image", + "bbox": [ + 130, + 173, + 477, + 305 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 249, + 165, + 331, + 172 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 248, + 164, + 331, + 173 + ], + "spans": [ + { + "bbox": [ + 248, + 164, + 331, + 173 + ], + "score": 1.0, + "content": "Progressive Generated Layouts", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "image_caption", + "bbox": [ + 297, + 151, + 478, + 173 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 295, + 149, + 478, + 165 + ], + "spans": [ + { + "bbox": [ + 295, + 149, + 478, + 165 + ], + "score": 1.0, + "content": "Figure 8: Baseline results of tangram design.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 443, + 163, + 479, + 174 + ], + "spans": [ + { + "bbox": [ + 443, + 163, + 479, + 174 + ], + "score": 1.0, + "content": "Real Layouts", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.0 + }, + { + "type": "image_body", + "bbox": [ + 130, + 173, + 477, + 305 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 130, + 173, + 477, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 173, + 477, + 305 + ], + "score": 0.961, + "type": "image", + "image_path": "00c0fd1c77d7eebaffc11c0b612d9811737fc843209a51ca7cd17d2b896b2ceb.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 130, + 173, + 477, + 217.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 130, + 217.0, + 477, + 261.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 130, + 261.0, + 477, + 305.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 125, + 309, + 480, + 321 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 128, + 308, + 483, + 322 + ], + "spans": [ + { + "bbox": [ + 128, + 308, + 483, + 322 + ], + "score": 1.0, + "content": "Figure 9: Training progression of tangram graphic design generation (from left to right).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 324, + 505, + 401 + ], + "lines": [], + "index": 22, + "bbox_fs": [ + 105, + 323, + 506, + 403 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 406, + 251, + 416 + ], + "lines": [ + { + "bbox": [ + 106, + 405, + 253, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 253, + 418 + ], + "score": 1.0, + "content": "4.4 TANGRAM GRAPHIC DESIGN", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 417, + 505, + 647 + ], + "lines": [ + { + "bbox": [ + 106, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "A tangram is a geometric puzzle aiming to form specific shapes using seven pieces of 2D shapes", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 427, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 505, + 441 + ], + "score": 1.0, + "content": "without overlaps. The seven pieces include two large right triangles, one medium right triangle, two", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "small right triangles, one square and one parallelogram, which can be assembled to form a square", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 448, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 506, + 463 + ], + "score": 1.0, + "content": "of side one unit and having area one square unit when choosing a unit of measurement. We collect", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 461, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 506, + 474 + ], + "score": 1.0, + "content": "149 tangram graphic designs including animals, people and objects. In our experiments, we consider", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 471, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 506, + 485 + ], + "score": 1.0, + "content": "eight rotation/reflection poses for each piece. Given the seven pieces with each initialized by random", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "score": 1.0, + "content": "location and ground-truth pose and class, the LayoutGAN with wireframe rendering discriminator is", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 492, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 104, + 492, + 506, + 508 + ], + "score": 1.0, + "content": "trained to refine their configurations jointly to form reasonable layouts. Note that it is a challenging", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 505, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 518 + ], + "score": 1.0, + "content": "task due to the complex configurations and limited data. We perform two experiments. The first", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 515, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 104, + 515, + 506, + 529 + ], + "score": 1.0, + "content": "one is a perturbation recovery test, in which we first add some random spatial perturbations to seven", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 526, + 506, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 506, + 539 + ], + "score": 1.0, + "content": "pieces of real layouts and train a LayoutGAN to recover the original layouts. Figure 7 shows that", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "our network is able to pull back the displaced shapes to the right locations, confirming that the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 549, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 506, + 561 + ], + "score": 1.0, + "content": "network successfully figures out the relations between different graphic elements. We further train", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "another LayoutGAN to generate tangram designs from purely random initialization. In Figure 9", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 570, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 505, + 583 + ], + "score": 1.0, + "content": "(high-resolution results can be found in the appendix), each row represents the sampled progressive", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 581, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 104, + 581, + 506, + 594 + ], + "score": 1.0, + "content": "generated results and the retrieved similar real tangrams. We also present some generated results", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "score": 1.0, + "content": "of DCGAN and a sequential model similar to Wang et al. (2018a) for comparison. One can see", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "score": 1.0, + "content": "from Figure 8 that DCGAN cannot well model the spatial relations among different shapes while", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 614, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 505, + 627 + ], + "score": 1.0, + "content": "the sequential model suffers from accumulated errors. In contrast, the LayoutGAN can generate", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 625, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 506, + 638 + ], + "score": 1.0, + "content": "meaningful tangrams like fox and person, although others may be hard to interpret, as shown in the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 637, + 176, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 176, + 648 + ], + "score": 1.0, + "content": "penultimate row.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 37, + "bbox_fs": [ + 104, + 417, + 506, + 648 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 653, + 196, + 664 + ], + "lines": [ + { + "bbox": [ + 104, + 650, + 198, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 650, + 198, + 668 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "In this paper, we have proposed a novel LayoutGAN for generating layouts of relational graphic", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "elements. Different from traditional GANs that generate images in pixel-level, the LayoutGAN can", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 686, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 506, + 701 + ], + "score": 1.0, + "content": "directly output a set of relational graphic elements. A novel differentiable wireframe rendering layer", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "was proposed to rasterize the generated graphic elements to wireframe images, making it feasible", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "to leverage CNNs as discriminator for better layout optimization from visual domain. Future works", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 720, + 486, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 486, + 733 + ], + "score": 1.0, + "content": "include adding a content representation, such as text, icon and picture to each graphic element.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 51.5, + "bbox_fs": [ + 105, + 665, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 505, + 145 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "score": 1.0, + "content": "Mart´ın Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 505, + 125 + ], + "score": 1.0, + "content": "Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, et al. Tensorflow: a system for large-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 122, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 116, + 122, + 505, + 136 + ], + "score": 1.0, + "content": "scale machine learning. In Proceedings of the 12th USENIX conference on Operating Systems", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 312, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 312, + 146 + ], + "score": 1.0, + "content": "Design and Implementation, pp. 265–283, 2016.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 105, + 153, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representa-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 164, + 474, + 176 + ], + "spans": [ + { + "bbox": [ + 116, + 164, + 474, + 176 + ], + "score": 1.0, + "content": "tions and generative models for 3d point clouds. arXiv preprint arXiv:1707.02392, 2017.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 184, + 503, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 183, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 198 + ], + "score": 1.0, + "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 195, + 398, + 208 + ], + "spans": [ + { + "bbox": [ + 115, + 195, + 398, + 208 + ], + "score": 1.0, + "content": "learning to align and translate. arXiv preprint arXiv:1409.0473, 2014.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 215, + 505, + 261 + ], + "lines": [ + { + "bbox": [ + 105, + 215, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 506, + 228 + ], + "score": 1.0, + "content": "Zoya Bylinskii, Nam Wook Kim, Peter O’Donovan, Sami Alsheikh, Spandan Madan, Hanspeter", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "Pfister, Fredo Durand, Bryan Russell, and Aaron Hertzmann. Learning visual importance for", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 238, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 115, + 238, + 506, + 250 + ], + "score": 1.0, + "content": "graphic designs and data visualizations. In Proceedings of the 30th Annual ACM Symposium on", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 249, + 380, + 261 + ], + "spans": [ + { + "bbox": [ + 116, + 249, + 380, + 261 + ], + "score": 1.0, + "content": "User Interface Software and Technology, pp. 57–69. ACM, 2017.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 268, + 504, + 303 + ], + "lines": [ + { + "bbox": [ + 104, + 267, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 104, + 267, + 505, + 282 + ], + "score": 1.0, + "content": "R Qi Charles, Hao Su, Mo Kaichun, and Leonidas J Guibas. Pointnet: Deep learning on point", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 279, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 115, + 279, + 505, + 293 + ], + "score": 1.0, + "content": "sets for 3d classification and segmentation. In Proceedings of the IEEE Conference on Computer", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 117, + 291, + 316, + 303 + ], + "spans": [ + { + "bbox": [ + 117, + 291, + 316, + 303 + ], + "score": 1.0, + "content": "Vision and Pattern Recognition, pp. 77–85, 2017.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 310, + 505, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "Biplab Deka, Zifeng Huang, Chad Franzen, Joshua Hibschman, Daniel Afergan, Yang Li, Jeffrey", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 114, + 320, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 114, + 320, + 505, + 336 + ], + "score": 1.0, + "content": "Nichols, and Ranjitha Kumar. Rico: A mobile app dataset for building data-driven design ap-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 332, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 115, + 332, + 506, + 346 + ], + "score": 1.0, + "content": "plications. In Proceedings of the 30th Annual ACM Symposium on User Interface Software and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 117, + 344, + 273, + 356 + ], + "spans": [ + { + "bbox": [ + 117, + 344, + 273, + 356 + ], + "score": 1.0, + "content": "Technology, pp. 845–854. ACM, 2017.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 363, + 505, + 408 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venu-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 375, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 115, + 375, + 506, + 387 + ], + "score": 1.0, + "content": "gopalan, Kate Saenko, and Trevor Darrell. Long-term recurrent convolutional networks for visual", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 386, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 115, + 386, + 506, + 399 + ], + "score": 1.0, + "content": "recognition and description. In Proceedings of the IEEE Conference on Computer Vision and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 397, + 291, + 410 + ], + "spans": [ + { + "bbox": [ + 116, + 397, + 291, + 410 + ], + "score": 1.0, + "content": "Pattern Recognition, pp. 2625–2634, 2015.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 106, + 416, + 504, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 416, + 504, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 504, + 429 + ], + "score": 1.0, + "content": "Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object recon-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 428, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 116, + 428, + 505, + 440 + ], + "score": 1.0, + "content": "struction from a single image. In Proceedings of the IEEE Conference on Computer Vision and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 439, + 296, + 451 + ], + "spans": [ + { + "bbox": [ + 116, + 439, + 296, + 451 + ], + "score": 1.0, + "content": "Pattern Recognition, volume 2, pp. 6, 2017.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 105, + 458, + 504, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 458, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 505, + 471 + ], + "score": 1.0, + "content": "Matthew Fisher, Daniel Ritchie, Manolis Savva, Thomas Funkhouser, and Pat Hanrahan. Example-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 470, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 115, + 470, + 505, + 482 + ], + "score": 1.0, + "content": "based synthesis of 3d object arrangements. ACM Transactions on Graphics (TOG), 31(6):135,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 114, + 479, + 143, + 494 + ], + "spans": [ + { + "bbox": [ + 114, + 479, + 143, + 494 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 105, + 501, + 504, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 499, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 514 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 511, + 214, + 523 + ], + "spans": [ + { + "bbox": [ + 115, + 511, + 214, + 523 + ], + "score": 1.0, + "content": "arXiv:1410.5401, 2014.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 531, + 503, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 545 + ], + "score": 1.0, + "content": "Nathan Hurst, Wilmot Li, and Kim Marriott. Review of automatic document formatting. In Pro-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 543, + 474, + 555 + ], + "spans": [ + { + "bbox": [ + 115, + 543, + 474, + 555 + ], + "score": 1.0, + "content": "ceedings of the 9th ACM symposium on Document engineering, pp. 99–108. ACM, 2009.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 562, + 504, + 586 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. In Ad-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 574, + 409, + 586 + ], + "spans": [ + { + "bbox": [ + 116, + 574, + 409, + 586 + ], + "score": 1.0, + "content": "vances in Neural Information Processing Systems, pp. 2017–2025, 2015.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 594, + 505, + 628 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization net-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 116, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "works for dense captioning. In Proceedings of the IEEE Conference on Computer Vision and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 616, + 291, + 628 + ], + "spans": [ + { + "bbox": [ + 116, + 616, + 291, + 628 + ], + "score": 1.0, + "content": "Pattern Recognition, pp. 4565–4574, 2016.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 636, + 504, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 504, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 504, + 650 + ], + "score": 1.0, + "content": "Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for im-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 647, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 115, + 647, + 506, + 660 + ], + "score": 1.0, + "content": "proved quality, stability, and variation. In International Conference on Learning Representations,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 657, + 143, + 671 + ], + "spans": [ + { + "bbox": [ + 115, + 657, + 143, + 671 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 106, + 678, + 504, + 701 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 692 + ], + "score": 1.0, + "content": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 689, + 215, + 701 + ], + "spans": [ + { + "bbox": [ + 115, + 689, + 215, + 701 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional net-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 720, + 406, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 406, + 732 + ], + "score": 1.0, + "content": "works. In International Conference on Learning Representations, 2017.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 505, + 145 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "score": 1.0, + "content": "Mart´ın Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 505, + 125 + ], + "score": 1.0, + "content": "Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, et al. Tensorflow: a system for large-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 122, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 116, + 122, + 505, + 136 + ], + "score": 1.0, + "content": "scale machine learning. In Proceedings of the 12th USENIX conference on Operating Systems", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 312, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 312, + 146 + ], + "score": 1.0, + "content": "Design and Implementation, pp. 265–283, 2016.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 99, + 505, + 146 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 153, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representa-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 164, + 474, + 176 + ], + "spans": [ + { + "bbox": [ + 116, + 164, + 474, + 176 + ], + "score": 1.0, + "content": "tions and generative models for 3d point clouds. arXiv preprint arXiv:1707.02392, 2017.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 153, + 505, + 176 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 184, + 503, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 183, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 198 + ], + "score": 1.0, + "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 195, + 398, + 208 + ], + "spans": [ + { + "bbox": [ + 115, + 195, + 398, + 208 + ], + "score": 1.0, + "content": "learning to align and translate. arXiv preprint arXiv:1409.0473, 2014.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 183, + 505, + 208 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 215, + 505, + 261 + ], + "lines": [ + { + "bbox": [ + 105, + 215, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 506, + 228 + ], + "score": 1.0, + "content": "Zoya Bylinskii, Nam Wook Kim, Peter O’Donovan, Sami Alsheikh, Spandan Madan, Hanspeter", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "Pfister, Fredo Durand, Bryan Russell, and Aaron Hertzmann. Learning visual importance for", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 238, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 115, + 238, + 506, + 250 + ], + "score": 1.0, + "content": "graphic designs and data visualizations. In Proceedings of the 30th Annual ACM Symposium on", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 249, + 380, + 261 + ], + "spans": [ + { + "bbox": [ + 116, + 249, + 380, + 261 + ], + "score": 1.0, + "content": "User Interface Software and Technology, pp. 57–69. ACM, 2017.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 215, + 506, + 261 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 268, + 504, + 303 + ], + "lines": [ + { + "bbox": [ + 104, + 267, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 104, + 267, + 505, + 282 + ], + "score": 1.0, + "content": "R Qi Charles, Hao Su, Mo Kaichun, and Leonidas J Guibas. Pointnet: Deep learning on point", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 279, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 115, + 279, + 505, + 293 + ], + "score": 1.0, + "content": "sets for 3d classification and segmentation. In Proceedings of the IEEE Conference on Computer", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 117, + 291, + 316, + 303 + ], + "spans": [ + { + "bbox": [ + 117, + 291, + 316, + 303 + ], + "score": 1.0, + "content": "Vision and Pattern Recognition, pp. 77–85, 2017.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 104, + 267, + 505, + 303 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 310, + 505, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "Biplab Deka, Zifeng Huang, Chad Franzen, Joshua Hibschman, Daniel Afergan, Yang Li, Jeffrey", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 114, + 320, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 114, + 320, + 505, + 336 + ], + "score": 1.0, + "content": "Nichols, and Ranjitha Kumar. Rico: A mobile app dataset for building data-driven design ap-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 332, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 115, + 332, + 506, + 346 + ], + "score": 1.0, + "content": "plications. In Proceedings of the 30th Annual ACM Symposium on User Interface Software and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 117, + 344, + 273, + 356 + ], + "spans": [ + { + "bbox": [ + 117, + 344, + 273, + 356 + ], + "score": 1.0, + "content": "Technology, pp. 845–854. ACM, 2017.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 106, + 311, + 506, + 356 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 363, + 505, + 408 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venu-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 375, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 115, + 375, + 506, + 387 + ], + "score": 1.0, + "content": "gopalan, Kate Saenko, and Trevor Darrell. Long-term recurrent convolutional networks for visual", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 386, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 115, + 386, + 506, + 399 + ], + "score": 1.0, + "content": "recognition and description. In Proceedings of the IEEE Conference on Computer Vision and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 397, + 291, + 410 + ], + "spans": [ + { + "bbox": [ + 116, + 397, + 291, + 410 + ], + "score": 1.0, + "content": "Pattern Recognition, pp. 2625–2634, 2015.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 363, + 506, + 410 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 416, + 504, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 416, + 504, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 504, + 429 + ], + "score": 1.0, + "content": "Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object recon-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 428, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 116, + 428, + 505, + 440 + ], + "score": 1.0, + "content": "struction from a single image. In Proceedings of the IEEE Conference on Computer Vision and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 439, + 296, + 451 + ], + "spans": [ + { + "bbox": [ + 116, + 439, + 296, + 451 + ], + "score": 1.0, + "content": "Pattern Recognition, volume 2, pp. 6, 2017.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 416, + 505, + 451 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 458, + 504, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 458, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 505, + 471 + ], + "score": 1.0, + "content": "Matthew Fisher, Daniel Ritchie, Manolis Savva, Thomas Funkhouser, and Pat Hanrahan. Example-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 470, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 115, + 470, + 505, + 482 + ], + "score": 1.0, + "content": "based synthesis of 3d object arrangements. ACM Transactions on Graphics (TOG), 31(6):135,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 114, + 479, + 143, + 494 + ], + "spans": [ + { + "bbox": [ + 114, + 479, + 143, + 494 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 458, + 505, + 494 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 501, + 504, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 499, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 514 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 511, + 214, + 523 + ], + "spans": [ + { + "bbox": [ + 115, + 511, + 214, + 523 + ], + "score": 1.0, + "content": "arXiv:1410.5401, 2014.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 499, + 505, + 523 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 531, + 503, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 545 + ], + "score": 1.0, + "content": "Nathan Hurst, Wilmot Li, and Kim Marriott. Review of automatic document formatting. In Pro-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 543, + 474, + 555 + ], + "spans": [ + { + "bbox": [ + 115, + 543, + 474, + 555 + ], + "score": 1.0, + "content": "ceedings of the 9th ACM symposium on Document engineering, pp. 99–108. ACM, 2009.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 531, + 505, + 555 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 562, + 504, + 586 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. In Ad-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 574, + 409, + 586 + ], + "spans": [ + { + "bbox": [ + 116, + 574, + 409, + 586 + ], + "score": 1.0, + "content": "vances in Neural Information Processing Systems, pp. 2017–2025, 2015.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 563, + 505, + 586 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 594, + 505, + 628 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization net-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 116, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "works for dense captioning. In Proceedings of the IEEE Conference on Computer Vision and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 616, + 291, + 628 + ], + "spans": [ + { + "bbox": [ + 116, + 616, + 291, + 628 + ], + "score": 1.0, + "content": "Pattern Recognition, pp. 4565–4574, 2016.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 594, + 505, + 628 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 636, + 504, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 504, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 504, + 650 + ], + "score": 1.0, + "content": "Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for im-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 647, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 115, + 647, + 506, + 660 + ], + "score": 1.0, + "content": "proved quality, stability, and variation. In International Conference on Learning Representations,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 657, + 143, + 671 + ], + "spans": [ + { + "bbox": [ + 115, + 657, + 143, + 671 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 635, + 506, + 671 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 678, + 504, + 701 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 692 + ], + "score": 1.0, + "content": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 689, + 215, + 701 + ], + "spans": [ + { + "bbox": [ + 115, + 689, + 215, + 701 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 676, + 506, + 701 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional net-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 720, + 406, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 406, + 732 + ], + "score": 1.0, + "content": "works. In International Conference on Learning Representations, 2017.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 106, + 708, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Paul Merrell, Eric Schkufza, Zeyang Li, Maneesh Agrawala, and Vladlen Koltun. Interactive fur-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "niture layout using interior design guidelines. ACM Transactions on Graphics (TOG), 30(4):87,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 102, + 143, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 102, + 143, + 117 + ], + "score": 1.0, + "content": "2011.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 124, + 503, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 504, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 504, + 137 + ], + "score": 1.0, + "content": "Peter O’Donovan, Aseem Agarwala, and Aaron Hertzmann. Color compatibility from large datasets.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 420, + 148 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 420, + 148 + ], + "score": 1.0, + "content": "In ACM Transactions on Graphics (TOG), volume 30, pp. 63. ACM, 2011.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 155, + 504, + 178 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "score": 1.0, + "content": "Peter O’Donovan, Janis L ¯ ¯ıbeks, Aseem Agarwala, and Aaron Hertzmann. Exploratory font selection", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 167, + 466, + 179 + ], + "spans": [ + { + "bbox": [ + 116, + 167, + 466, + 179 + ], + "score": 1.0, + "content": "using crowdsourced attributes. ACM Transactions on Graphics (TOG), 33(4):92, 2014.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 186, + 505, + 220 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 504, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 504, + 199 + ], + "score": 1.0, + "content": "Peter O’Donovan, Aseem Agarwala, and Aaron Hertzmann. Designscape: Design with interactive", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 116, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "layout suggestions. In Proceedings of the 33rd Annual ACM Conference on Human Factors in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 117, + 208, + 317, + 221 + ], + "spans": [ + { + "bbox": [ + 117, + 208, + 317, + 221 + ], + "score": 1.0, + "content": "Computing Systems, pp. 1221–1224. ACM, 2015.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 228, + 505, + 263 + ], + "lines": [ + { + "bbox": [ + 105, + 227, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 505, + 242 + ], + "score": 1.0, + "content": "Aaron van den Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, and Koray ¨", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 116, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "Kavukcuoglu. Conditional image generation with pixelcnn decoders. In Advances in Neural", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 250, + 339, + 263 + ], + "spans": [ + { + "bbox": [ + 116, + 250, + 339, + 263 + ], + "score": 1.0, + "content": "Information Processing Systems, pp. 4797–4805, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 108, + 270, + 505, + 304 + ], + "lines": [ + { + "bbox": [ + 105, + 269, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 505, + 284 + ], + "score": 1.0, + "content": "Peter O’Donovan, Aseem Agarwala, and Aaron Hertzmann. Learning layouts for single-page", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 282, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 115, + 282, + 506, + 294 + ], + "score": 1.0, + "content": "graphic designs. IEEE Transactions on Visualization and Computer Graphics, 20(8):1200–1213,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 292, + 142, + 304 + ], + "spans": [ + { + "bbox": [ + 115, + 292, + 142, + 304 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 312, + 503, + 335 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "Xufang Pang, Ying Cao, Rynson WH Lau, and Antoni B Chan. Directing user attention via visual", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 323, + 434, + 336 + ], + "spans": [ + { + "bbox": [ + 115, + 323, + 434, + 336 + ], + "score": 1.0, + "content": "flow on web designs. ACM Transactions on Graphics (TOG), 35(6):240, 2016.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 343, + 503, + 367 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 357 + ], + "score": 1.0, + "content": "Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 355, + 468, + 367 + ], + "spans": [ + { + "bbox": [ + 115, + 355, + 468, + 367 + ], + "score": 1.0, + "content": "convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 374, + 505, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 373, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 279, + 388 + ], + "score": 1.0, + "content": "Scott Reed and Nando De Freitas.", + "type": "text" + }, + { + "bbox": [ + 284, + 375, + 425, + 388 + ], + "score": 1.0, + "content": "Neural programmer-interpreters.", + "type": "text" + }, + { + "bbox": [ + 439, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "arXiv preprint", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 386, + 219, + 397 + ], + "spans": [ + { + "bbox": [ + 115, + 386, + 219, + 397 + ], + "score": 1.0, + "content": "arXiv:1511.06279, 2015.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 405, + 505, + 440 + ], + "lines": [ + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "score": 1.0, + "content": "Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 116, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "Improved techniques for training gans. In Advances in Neural Information Processing Systems,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 428, + 207, + 440 + ], + "spans": [ + { + "bbox": [ + 115, + 428, + 207, + 440 + ], + "score": 1.0, + "content": "pp. 2234–2242, 2016.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 447, + 504, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 447, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 505, + 461 + ], + "score": 1.0, + "content": "Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 459, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 116, + 459, + 505, + 471 + ], + "score": 1.0, + "content": "Semantic scene completion from a single depth image. In Proceedings of the IEEE Conference", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 470, + 381, + 483 + ], + "spans": [ + { + "bbox": [ + 116, + 470, + 381, + 483 + ], + "score": 1.0, + "content": "on Computer Vision and Pattern Recognition, pp. 190–198, 2017.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 104, + 489, + 504, + 513 + ], + "lines": [ + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 500, + 431, + 514 + ], + "spans": [ + { + "bbox": [ + 115, + 500, + 431, + 514 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems, pp. 3104–3112, 2014.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 108, + 520, + 504, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 520, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 505, + 534 + ], + "score": 1.0, + "content": "Amanda Swearngin, Mira Dontcheva, Wilmot Li, Joel Brandt, Morgan Dixon, and Andrew J Ko.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 115, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "Rewire: Interface design assistance from examples. In Proceedings of the 2018 CHI Conference", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 543, + 374, + 555 + ], + "spans": [ + { + "bbox": [ + 115, + 543, + 374, + 555 + ], + "score": 1.0, + "content": "on Human Factors in Computing Systems, pp. 504. ACM, 2018.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 563, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 577 + ], + "score": 1.0, + "content": "Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, ¨", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 116, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 585, + 406, + 598 + ], + "spans": [ + { + "bbox": [ + 116, + 585, + 406, + 598 + ], + "score": 1.0, + "content": "raw audio. In 9th ISCA Speech Synthesis Workshop, pp. 125–125, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 605, + 504, + 628 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 619 + ], + "score": 1.0, + "content": "Oriol Vinyals, Samy Bengio, and Manjunath Kudlur. Order matters: Sequence to sequence for sets.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 617, + 278, + 627 + ], + "spans": [ + { + "bbox": [ + 116, + 617, + 278, + 627 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1511.06391, 2015.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 105, + 636, + 504, + 659 + ], + "lines": [ + { + "bbox": [ + 106, + 635, + 504, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 504, + 649 + ], + "score": 1.0, + "content": "Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 647, + 421, + 660 + ], + "spans": [ + { + "bbox": [ + 116, + 647, + 421, + 660 + ], + "score": 1.0, + "content": "In Advances In Neural Information Processing Systems, pp. 613–621, 2016.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 105, + 667, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "Kai Wang, Manolis Savva, Angel X Chang, and Daniel Ritchie. Deep convolutional priors for indoor", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 679, + 413, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 679, + 413, + 691 + ], + "score": 1.0, + "content": "scene synthesis. ACM Transactions on Graphics (TOG), 37(4):70, 2018a.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 114, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 114, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, volume 1, pp.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 719, + 157, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 719, + 157, + 733 + ], + "score": 1.0, + "content": "4, 2018b.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Paul Merrell, Eric Schkufza, Zeyang Li, Maneesh Agrawala, and Vladlen Koltun. Interactive fur-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "niture layout using interior design guidelines. ACM Transactions on Graphics (TOG), 30(4):87,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 102, + 143, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 102, + 143, + 117 + ], + "score": 1.0, + "content": "2011.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 505, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 124, + 503, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 504, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 504, + 137 + ], + "score": 1.0, + "content": "Peter O’Donovan, Aseem Agarwala, and Aaron Hertzmann. Color compatibility from large datasets.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 420, + 148 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 420, + 148 + ], + "score": 1.0, + "content": "In ACM Transactions on Graphics (TOG), volume 30, pp. 63. ACM, 2011.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 123, + 504, + 148 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 155, + 504, + 178 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "score": 1.0, + "content": "Peter O’Donovan, Janis L ¯ ¯ıbeks, Aseem Agarwala, and Aaron Hertzmann. Exploratory font selection", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 167, + 466, + 179 + ], + "spans": [ + { + "bbox": [ + 116, + 167, + 466, + 179 + ], + "score": 1.0, + "content": "using crowdsourced attributes. ACM Transactions on Graphics (TOG), 33(4):92, 2014.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 154, + 505, + 179 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 186, + 505, + 220 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 504, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 504, + 199 + ], + "score": 1.0, + "content": "Peter O’Donovan, Aseem Agarwala, and Aaron Hertzmann. Designscape: Design with interactive", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 116, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "layout suggestions. In Proceedings of the 33rd Annual ACM Conference on Human Factors in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 117, + 208, + 317, + 221 + ], + "spans": [ + { + "bbox": [ + 117, + 208, + 317, + 221 + ], + "score": 1.0, + "content": "Computing Systems, pp. 1221–1224. ACM, 2015.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 187, + 505, + 221 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 228, + 505, + 263 + ], + "lines": [ + { + "bbox": [ + 105, + 227, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 505, + 242 + ], + "score": 1.0, + "content": "Aaron van den Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, and Koray ¨", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 116, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "Kavukcuoglu. Conditional image generation with pixelcnn decoders. In Advances in Neural", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 250, + 339, + 263 + ], + "spans": [ + { + "bbox": [ + 116, + 250, + 339, + 263 + ], + "score": 1.0, + "content": "Information Processing Systems, pp. 4797–4805, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 227, + 505, + 263 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 270, + 505, + 304 + ], + "lines": [ + { + "bbox": [ + 105, + 269, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 505, + 284 + ], + "score": 1.0, + "content": "Peter O’Donovan, Aseem Agarwala, and Aaron Hertzmann. Learning layouts for single-page", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 282, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 115, + 282, + 506, + 294 + ], + "score": 1.0, + "content": "graphic designs. IEEE Transactions on Visualization and Computer Graphics, 20(8):1200–1213,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 292, + 142, + 304 + ], + "spans": [ + { + "bbox": [ + 115, + 292, + 142, + 304 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 269, + 506, + 304 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 312, + 503, + 335 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "Xufang Pang, Ying Cao, Rynson WH Lau, and Antoni B Chan. Directing user attention via visual", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 323, + 434, + 336 + ], + "spans": [ + { + "bbox": [ + 115, + 323, + 434, + 336 + ], + "score": 1.0, + "content": "flow on web designs. ACM Transactions on Graphics (TOG), 35(6):240, 2016.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 313, + 505, + 336 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 343, + 503, + 367 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 357 + ], + "score": 1.0, + "content": "Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 355, + 468, + 367 + ], + "spans": [ + { + "bbox": [ + 115, + 355, + 468, + 367 + ], + "score": 1.0, + "content": "convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 342, + 505, + 367 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 374, + 505, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 373, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 279, + 388 + ], + "score": 1.0, + "content": "Scott Reed and Nando De Freitas.", + "type": "text" + }, + { + "bbox": [ + 284, + 375, + 425, + 388 + ], + "score": 1.0, + "content": "Neural programmer-interpreters.", + "type": "text" + }, + { + "bbox": [ + 439, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "arXiv preprint", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 386, + 219, + 397 + ], + "spans": [ + { + "bbox": [ + 115, + 386, + 219, + 397 + ], + "score": 1.0, + "content": "arXiv:1511.06279, 2015.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 373, + 506, + 397 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 405, + 505, + 440 + ], + "lines": [ + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "score": 1.0, + "content": "Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 116, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "Improved techniques for training gans. In Advances in Neural Information Processing Systems,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 428, + 207, + 440 + ], + "spans": [ + { + "bbox": [ + 115, + 428, + 207, + 440 + ], + "score": 1.0, + "content": "pp. 2234–2242, 2016.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 106, + 406, + 505, + 440 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 447, + 504, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 447, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 505, + 461 + ], + "score": 1.0, + "content": "Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 459, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 116, + 459, + 505, + 471 + ], + "score": 1.0, + "content": "Semantic scene completion from a single depth image. In Proceedings of the IEEE Conference", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 470, + 381, + 483 + ], + "spans": [ + { + "bbox": [ + 116, + 470, + 381, + 483 + ], + "score": 1.0, + "content": "on Computer Vision and Pattern Recognition, pp. 190–198, 2017.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 106, + 447, + 505, + 483 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 489, + 504, + 513 + ], + "lines": [ + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 500, + 431, + 514 + ], + "spans": [ + { + "bbox": [ + 115, + 500, + 431, + 514 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems, pp. 3104–3112, 2014.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 490, + 505, + 514 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 520, + 504, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 520, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 505, + 534 + ], + "score": 1.0, + "content": "Amanda Swearngin, Mira Dontcheva, Wilmot Li, Joel Brandt, Morgan Dixon, and Andrew J Ko.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 115, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "Rewire: Interface design assistance from examples. In Proceedings of the 2018 CHI Conference", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 543, + 374, + 555 + ], + "spans": [ + { + "bbox": [ + 115, + 543, + 374, + 555 + ], + "score": 1.0, + "content": "on Human Factors in Computing Systems, pp. 504. ACM, 2018.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 106, + 520, + 505, + 555 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 563, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 577 + ], + "score": 1.0, + "content": "Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, ¨", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 116, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 585, + 406, + 598 + ], + "spans": [ + { + "bbox": [ + 116, + 585, + 406, + 598 + ], + "score": 1.0, + "content": "raw audio. In 9th ISCA Speech Synthesis Workshop, pp. 125–125, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 561, + 505, + 598 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 605, + 504, + 628 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 619 + ], + "score": 1.0, + "content": "Oriol Vinyals, Samy Bengio, and Manjunath Kudlur. Order matters: Sequence to sequence for sets.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 617, + 278, + 627 + ], + "spans": [ + { + "bbox": [ + 116, + 617, + 278, + 627 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1511.06391, 2015.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 603, + 505, + 627 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 636, + 504, + 659 + ], + "lines": [ + { + "bbox": [ + 106, + 635, + 504, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 504, + 649 + ], + "score": 1.0, + "content": "Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 647, + 421, + 660 + ], + "spans": [ + { + "bbox": [ + 116, + 647, + 421, + 660 + ], + "score": 1.0, + "content": "In Advances In Neural Information Processing Systems, pp. 613–621, 2016.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 106, + 635, + 504, + 660 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 667, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "Kai Wang, Manolis Savva, Angel X Chang, and Daniel Ritchie. Deep convolutional priors for indoor", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 679, + 413, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 679, + 413, + 691 + ], + "score": 1.0, + "content": "scene synthesis. ACM Transactions on Graphics (TOG), 37(4):70, 2018a.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 667, + 505, + 691 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 114, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 114, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, volume 1, pp.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 719, + 157, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 719, + 157, + 733 + ], + "score": 1.0, + "content": "4, 2018b.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 106, + 698, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 81, + 506, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a prob-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "abilistic latent space of object shapes via 3d generative-adversarial modeling. In Advances in", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 348, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 348, + 117 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 82–90, 2016.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 505, + 136 + ], + "score": 1.0, + "content": "Xinchen Yan, Jimei Yang, Ersin Yumer, Yijie Guo, and Honglak Lee. Perspective transformer nets:", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 506, + 148 + ], + "score": 1.0, + "content": "Learning single-view 3d object reconstruction without 3d supervision. In Advances in Neural", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 145, + 339, + 158 + ], + "spans": [ + { + "bbox": [ + 115, + 145, + 339, + 158 + ], + "score": 1.0, + "content": "Information Processing Systems, pp. 1696–1704, 2016.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 164, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 506, + 177 + ], + "score": 1.0, + "content": "Xiao Yang, Ersin Yumer, Paul Asente, Mike Kraley, Daniel Kifer, and C Lee Giles. Learning to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 176, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 116, + 176, + 505, + 187 + ], + "score": 1.0, + "content": "extract semantic structure from documents using multimodal fully convolutional neural networks.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 186, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 505, + 199 + ], + "score": 1.0, + "content": "In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 5315–", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 197, + 167, + 209 + ], + "spans": [ + { + "bbox": [ + 116, + 197, + 167, + 209 + ], + "score": 1.0, + "content": "5324, 2017.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 215, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 505, + 229 + ], + "score": 1.0, + "content": "C Lawrence Zitnick and Devi Parikh. Bringing semantics into focus using visual abstraction. In", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3009–", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 238, + 167, + 250 + ], + "spans": [ + { + "bbox": [ + 116, + 238, + 167, + 250 + ], + "score": 1.0, + "content": "3016, 2013.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 257, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 269 + ], + "score": 1.0, + "content": "C Lawrence Zitnick, Ramakrishna Vedantam, and Devi Parikh. Adopting abstract images for se-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 268, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 115, + 268, + 506, + 281 + ], + "score": 1.0, + "content": "mantic scene understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 279, + 206, + 291 + ], + "spans": [ + { + "bbox": [ + 116, + 279, + 206, + 291 + ], + "score": 1.0, + "content": "38(4):627–638, 2016.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 7.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 81, + 506, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a prob-", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "abilistic latent space of object shapes via 3d generative-adversarial modeling. In Advances in", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 348, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 348, + 117 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 82–90, 2016.", + "type": "text" + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 124, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 505, + 136 + ], + "score": 1.0, + "content": "Xinchen Yan, Jimei Yang, Ersin Yumer, Yijie Guo, and Honglak Lee. Perspective transformer nets:", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 134, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 506, + 148 + ], + "score": 1.0, + "content": "Learning single-view 3d object reconstruction without 3d supervision. In Advances in Neural", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 145, + 339, + 158 + ], + "spans": [ + { + "bbox": [ + 115, + 145, + 339, + 158 + ], + "score": 1.0, + "content": "Information Processing Systems, pp. 1696–1704, 2016.", + "type": "text" + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 164, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 506, + 177 + ], + "score": 1.0, + "content": "Xiao Yang, Ersin Yumer, Paul Asente, Mike Kraley, Daniel Kifer, and C Lee Giles. Learning to", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 176, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 116, + 176, + 505, + 187 + ], + "score": 1.0, + "content": "extract semantic structure from documents using multimodal fully convolutional neural networks.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 186, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 505, + 199 + ], + "score": 1.0, + "content": "In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 5315–", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 197, + 167, + 209 + ], + "spans": [ + { + "bbox": [ + 116, + 197, + 167, + 209 + ], + "score": 1.0, + "content": "5324, 2017.", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 215, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 505, + 229 + ], + "score": 1.0, + "content": "C Lawrence Zitnick and Devi Parikh. Bringing semantics into focus using visual abstraction. In", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3009–", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 238, + 167, + 250 + ], + "spans": [ + { + "bbox": [ + 116, + 238, + 167, + 250 + ], + "score": 1.0, + "content": "3016, 2013.", + "type": "text" + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 257, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 269 + ], + "score": 1.0, + "content": "C Lawrence Zitnick, Ramakrishna Vedantam, and Devi Parikh. Adopting abstract images for se-", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 268, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 115, + 268, + 506, + 281 + ], + "score": 1.0, + "content": "mantic scene understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 279, + 206, + 291 + ], + "spans": [ + { + "bbox": [ + 116, + 279, + 206, + 291 + ], + "score": 1.0, + "content": "38(4):627–638, 2016.", + "type": "text" + } + ], + "index": 15, + "is_list_end_line": true + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 82, + 506, + 291 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 180, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 182, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 182, + 96 + ], + "score": 1.0, + "content": "6 APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 107, + 237, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 107, + 237, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 107, + 237, + 118 + ], + "score": 1.0, + "content": "6.1 REAL DOCUMENT PAGES", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 127, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "For better understanding the document semantic layouts, we show some layout samples and their", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 138, + 341, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 341, + 150 + ], + "score": 1.0, + "content": "corresponding real document pages, as shown in Figure 1.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 165, + 160, + 443, + 380 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 165, + 160, + 443, + 380 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 165, + 160, + 443, + 380 + ], + "spans": [ + { + "bbox": [ + 165, + 160, + 443, + 380 + ], + "score": 0.974, + "type": "image", + "image_path": "48697f4442f26ffebf238f89d6c19fa0e7000dabe3381482f56c3911a0bd16cc.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 165, + 160, + 443, + 174.66666666666666 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 165, + 174.66666666666666, + 443, + 189.33333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 165, + 189.33333333333331, + 443, + 203.99999999999997 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 165, + 203.99999999999997, + 443, + 218.66666666666663 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 165, + 218.66666666666663, + 443, + 233.3333333333333 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 165, + 233.3333333333333, + 443, + 247.99999999999994 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 165, + 247.99999999999994, + 443, + 262.66666666666663 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 165, + 262.66666666666663, + 443, + 277.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 165, + 277.3333333333333, + 443, + 292.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 165, + 292.0, + 443, + 306.6666666666667 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 165, + 306.6666666666667, + 443, + 321.33333333333337 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 165, + 321.33333333333337, + 443, + 336.00000000000006 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 165, + 336.00000000000006, + 443, + 350.66666666666674 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 165, + 350.66666666666674, + 443, + 365.3333333333334 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 165, + 365.3333333333334, + 443, + 380.0000000000001 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 388, + 501, + 401 + ], + "lines": [ + { + "bbox": [ + 108, + 387, + 502, + 403 + ], + "spans": [ + { + "bbox": [ + 108, + 387, + 502, + 403 + ], + "score": 1.0, + "content": "Figure 1: Visualization of document layout samples and their corresponding real document pages.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 410, + 370, + 421 + ], + "lines": [ + { + "bbox": [ + 106, + 409, + 371, + 422 + ], + "spans": [ + { + "bbox": [ + 106, + 409, + 371, + 422 + ], + "score": 1.0, + "content": "6.2 HIGH-RESOLUTION SYNTHESIZED DOCUMENT LAYOUTS", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 145, + 435, + 465, + 719 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 145, + 435, + 465, + 719 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 145, + 435, + 465, + 719 + ], + "spans": [ + { + "bbox": [ + 145, + 435, + 465, + 719 + ], + "score": 0.974, + "type": "image", + "image_path": "59a94604fbb9961f70c224562e3a9f201e0a8d28bcb1b717ef19d6b70562d4f9.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 145, + 435, + 465, + 529.6666666666666 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 145, + 529.6666666666666, + 465, + 624.3333333333333 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 145, + 624.3333333333333, + 465, + 718.9999999999999 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 227, + 729, + 383, + 740 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 225, + 726, + 385, + 744 + ], + "spans": [ + { + "bbox": [ + 225, + 726, + 385, + 744 + ], + "score": 1.0, + "content": "Figure 2: Document layout generation.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + } + ], + "index": 23.0 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 761 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 180, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 182, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 182, + 96 + ], + "score": 1.0, + "content": "6 APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 107, + 237, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 107, + 237, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 107, + 237, + 118 + ], + "score": 1.0, + "content": "6.1 REAL DOCUMENT PAGES", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 127, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "For better understanding the document semantic layouts, we show some layout samples and their", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 138, + 341, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 341, + 150 + ], + "score": 1.0, + "content": "corresponding real document pages, as shown in Figure 1.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 126, + 505, + 150 + ] + }, + { + "type": "image", + "bbox": [ + 165, + 160, + 443, + 380 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 165, + 160, + 443, + 380 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 165, + 160, + 443, + 380 + ], + "spans": [ + { + "bbox": [ + 165, + 160, + 443, + 380 + ], + "score": 0.974, + "type": "image", + "image_path": "48697f4442f26ffebf238f89d6c19fa0e7000dabe3381482f56c3911a0bd16cc.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 165, + 160, + 443, + 174.66666666666666 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 165, + 174.66666666666666, + 443, + 189.33333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 165, + 189.33333333333331, + 443, + 203.99999999999997 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 165, + 203.99999999999997, + 443, + 218.66666666666663 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 165, + 218.66666666666663, + 443, + 233.3333333333333 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 165, + 233.3333333333333, + 443, + 247.99999999999994 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 165, + 247.99999999999994, + 443, + 262.66666666666663 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 165, + 262.66666666666663, + 443, + 277.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 165, + 277.3333333333333, + 443, + 292.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 165, + 292.0, + 443, + 306.6666666666667 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 165, + 306.6666666666667, + 443, + 321.33333333333337 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 165, + 321.33333333333337, + 443, + 336.00000000000006 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 165, + 336.00000000000006, + 443, + 350.66666666666674 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 165, + 350.66666666666674, + 443, + 365.3333333333334 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 165, + 365.3333333333334, + 443, + 380.0000000000001 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 388, + 501, + 401 + ], + "lines": [ + { + "bbox": [ + 108, + 387, + 502, + 403 + ], + "spans": [ + { + "bbox": [ + 108, + 387, + 502, + 403 + ], + "score": 1.0, + "content": "Figure 1: Visualization of document layout samples and their corresponding real document pages.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19, + "bbox_fs": [ + 108, + 387, + 502, + 403 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 410, + 370, + 421 + ], + "lines": [ + { + "bbox": [ + 106, + 409, + 371, + 422 + ], + "spans": [ + { + "bbox": [ + 106, + 409, + 371, + 422 + ], + "score": 1.0, + "content": "6.2 HIGH-RESOLUTION SYNTHESIZED DOCUMENT LAYOUTS", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 106, + 409, + 371, + 422 + ] + }, + { + "type": "image", + "bbox": [ + 145, + 435, + 465, + 719 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 145, + 435, + 465, + 719 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 145, + 435, + 465, + 719 + ], + "spans": [ + { + "bbox": [ + 145, + 435, + 465, + 719 + ], + "score": 0.974, + "type": "image", + "image_path": "59a94604fbb9961f70c224562e3a9f201e0a8d28bcb1b717ef19d6b70562d4f9.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 145, + 435, + 465, + 529.6666666666666 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 145, + 529.6666666666666, + 465, + 624.3333333333333 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 145, + 624.3333333333333, + 465, + 718.9999999999999 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 227, + 729, + 383, + 740 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 225, + 726, + 385, + 744 + ], + "spans": [ + { + "bbox": [ + 225, + 726, + 385, + 744 + ], + "score": 1.0, + "content": "Figure 2: Document layout generation.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + } + ], + "index": 23.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 109, + 83, + 395, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 398, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 398, + 95 + ], + "score": 1.0, + "content": "6.3 HIGH-RESOLUTION SYNTHESIZED CLIPART ABSTRACT SCENES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 108, + 103, + 502, + 549 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 103, + 502, + 549 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 103, + 502, + 549 + ], + "spans": [ + { + "bbox": [ + 108, + 103, + 502, + 549 + ], + "score": 0.969, + "type": "image", + "image_path": "59a9b616d7637834ad1a983b6585eb88310fbbfa6b6c991135c7b94e28480936.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 108, + 103, + 502, + 251.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 251.66666666666666, + 502, + 400.3333333333333 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 108, + 400.3333333333333, + 502, + 549.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 219, + 556, + 392, + 568 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 217, + 555, + 392, + 569 + ], + "spans": [ + { + "bbox": [ + 217, + 555, + 392, + 569 + ], + "score": 1.0, + "content": "Figure 3: Clipart abstract scene generation.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + } + ], + "index": 3.0 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 109, + 83, + 395, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 398, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 398, + 95 + ], + "score": 1.0, + "content": "6.3 HIGH-RESOLUTION SYNTHESIZED CLIPART ABSTRACT SCENES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 108, + 103, + 502, + 549 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 103, + 502, + 549 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 103, + 502, + 549 + ], + "spans": [ + { + "bbox": [ + 108, + 103, + 502, + 549 + ], + "score": 0.969, + "type": "image", + "image_path": "59a9b616d7637834ad1a983b6585eb88310fbbfa6b6c991135c7b94e28480936.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 108, + 103, + 502, + 251.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 251.66666666666666, + 502, + 400.3333333333333 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 108, + 400.3333333333333, + 502, + 549.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 219, + 556, + 392, + 568 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 217, + 555, + 392, + 569 + ], + "spans": [ + { + "bbox": [ + 217, + 555, + 392, + 569 + ], + "score": 1.0, + "content": "Figure 3: Clipart abstract scene generation.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + } + ], + "index": 3.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 108, + 502, + 354 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 105, + 108, + 502, + 354 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 108, + 502, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 108, + 502, + 354 + ], + "score": 0.972, + "type": "image", + "image_path": "b00fe714a8c08369c14cb18a0e232ba0e9a8119cdf5a7ef43faa974305d6a709.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 105, + 108, + 502, + 190.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 105, + 190.0, + 502, + 272.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 105, + 272.0, + 502, + 354.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 235, + 364, + 375, + 376 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 234, + 362, + 376, + 379 + ], + "spans": [ + { + "bbox": [ + 234, + 362, + 376, + 379 + ], + "score": 1.0, + "content": "Figure 4: Tangram graphic design.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 106, + 381, + 367, + 392 + ], + "lines": [ + { + "bbox": [ + 106, + 380, + 368, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 368, + 393 + ], + "score": 1.0, + "content": "6.5 DETAILED ARCHITECTURE DESIGN OF THE GENERATOR", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 401, + 505, + 501 + ], + "lines": [ + { + "bbox": [ + 106, + 402, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 413 + ], + "score": 1.0, + "content": "The generator takes as input a set of graphic elements with random class probabilities and geometric", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 411, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 426 + ], + "score": 1.0, + "content": "parameters sampled from Uniform and Gaussian distribution respectively. An encoder consisting", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 424, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 505, + 435 + ], + "score": 1.0, + "content": "of three fully connected layers first embeds the semantic and geometric parameters of each graphic", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 433, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 506, + 447 + ], + "score": 1.0, + "content": "element. Two cascaded relation modules implemented as self-attention module followed by a basic", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 444, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 104, + 444, + 506, + 458 + ], + "score": 1.0, + "content": "residual block (“bottleneck” building block) form one relation block. We cascade two such relation", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "blocks (totally 4 relation modules) for contextual feature refinement. Finally, a decoder consisting of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 467, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 479 + ], + "score": 1.0, + "content": "several fully connected layers followed by two heads (implemented as the fully connected layer) with", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "sigmoid activation are used to decode the refined feature of each element back to class probability", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 489, + 315, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 315, + 502 + ], + "score": 1.0, + "content": "distributions and geometric parameters respectively.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 112, + 83, + 398, + 93 + ], + "lines": [ + { + "bbox": [ + 110, + 83, + 400, + 95 + ], + "spans": [ + { + "bbox": [ + 110, + 83, + 400, + 95 + ], + "score": 1.0, + "content": "6.4 HIGH-RESOLUTION SYNTHESIZED TANGRAM GRAPHIC DESIGNS", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 108, + 502, + 354 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 105, + 108, + 502, + 354 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 108, + 502, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 108, + 502, + 354 + ], + "score": 0.972, + "type": "image", + "image_path": "b00fe714a8c08369c14cb18a0e232ba0e9a8119cdf5a7ef43faa974305d6a709.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 105, + 108, + 502, + 190.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 105, + 190.0, + 502, + 272.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 105, + 272.0, + 502, + 354.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 235, + 364, + 375, + 376 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 234, + 362, + 376, + 379 + ], + "spans": [ + { + "bbox": [ + 234, + 362, + 376, + 379 + ], + "score": 1.0, + "content": "Figure 4: Tangram graphic design.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 106, + 381, + 367, + 392 + ], + "lines": [ + { + "bbox": [ + 106, + 380, + 368, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 368, + 393 + ], + "score": 1.0, + "content": "6.5 DETAILED ARCHITECTURE DESIGN OF THE GENERATOR", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 401, + 505, + 501 + ], + "lines": [ + { + "bbox": [ + 106, + 402, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 413 + ], + "score": 1.0, + "content": "The generator takes as input a set of graphic elements with random class probabilities and geometric", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 411, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 426 + ], + "score": 1.0, + "content": "parameters sampled from Uniform and Gaussian distribution respectively. An encoder consisting", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 424, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 505, + 435 + ], + "score": 1.0, + "content": "of three fully connected layers first embeds the semantic and geometric parameters of each graphic", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 433, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 506, + 447 + ], + "score": 1.0, + "content": "element. Two cascaded relation modules implemented as self-attention module followed by a basic", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 444, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 104, + 444, + 506, + 458 + ], + "score": 1.0, + "content": "residual block (“bottleneck” building block) form one relation block. We cascade two such relation", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "blocks (totally 4 relation modules) for contextual feature refinement. Finally, a decoder consisting of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 467, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 479 + ], + "score": 1.0, + "content": "several fully connected layers followed by two heads (implemented as the fully connected layer) with", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "sigmoid activation are used to decode the refined feature of each element back to class probability", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 489, + 315, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 315, + 502 + ], + "score": 1.0, + "content": "distributions and geometric parameters respectively.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9, + "bbox_fs": [ + 104, + 402, + 506, + 502 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 109, + 83, + 322, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 325, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 325, + 95 + ], + "score": 1.0, + "content": "6.6 SYNTHESIZED MOBILE APP LAYOUT DESIGNS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 102, + 505, + 137 + ], + "lines": [ + { + "bbox": [ + 106, + 102, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 506, + 116 + ], + "score": 1.0, + "content": "We further conduct a new experiment of mobile app layout design by using the RICO dataset (Deka", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "et al., 2017). For better understanding the mobile app layouts, we first show some layout samples", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 125, + 407, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 125, + 407, + 137 + ], + "score": 1.0, + "content": "and their corresponding real mobile app screenshots, as shown in Figure 5.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 107, + 148, + 502, + 450 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 148, + 502, + 450 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 148, + 502, + 450 + ], + "spans": [ + { + "bbox": [ + 107, + 148, + 502, + 450 + ], + "score": 0.971, + "type": "image", + "image_path": "de4a5d5cb562c7b1454d3bf7d559d6a13f30ebf080f65f675ab122c778ffcda2.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 107, + 148, + 502, + 248.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 107, + 248.66666666666669, + 502, + 349.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 349.33333333333337, + 502, + 450.00000000000006 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 461, + 505, + 483 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 458, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 104, + 458, + 506, + 476 + ], + "score": 1.0, + "content": "Figure 5: Visualization of mobile app layout samples and their corresponding real mobile app", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 471, + 157, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 157, + 483 + ], + "score": 1.0, + "content": "screenshots.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + } + ], + "index": 6.25 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 109, + 83, + 322, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 325, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 325, + 95 + ], + "score": 1.0, + "content": "6.6 SYNTHESIZED MOBILE APP LAYOUT DESIGNS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 102, + 505, + 137 + ], + "lines": [ + { + "bbox": [ + 106, + 102, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 506, + 116 + ], + "score": 1.0, + "content": "We further conduct a new experiment of mobile app layout design by using the RICO dataset (Deka", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "et al., 2017). For better understanding the mobile app layouts, we first show some layout samples", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 125, + 407, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 125, + 407, + 137 + ], + "score": 1.0, + "content": "and their corresponding real mobile app screenshots, as shown in Figure 5.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 102, + 506, + 137 + ] + }, + { + "type": "image", + "bbox": [ + 107, + 148, + 502, + 450 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 148, + 502, + 450 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 148, + 502, + 450 + ], + "spans": [ + { + "bbox": [ + 107, + 148, + 502, + 450 + ], + "score": 0.971, + "type": "image", + "image_path": "de4a5d5cb562c7b1454d3bf7d559d6a13f30ebf080f65f675ab122c778ffcda2.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 107, + 148, + 502, + 248.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 107, + 248.66666666666669, + 502, + 349.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 349.33333333333337, + 502, + 450.00000000000006 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 461, + 505, + 483 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 458, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 104, + 458, + 506, + 476 + ], + "score": 1.0, + "content": "Figure 5: Visualization of mobile app layout samples and their corresponding real mobile app", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 471, + 157, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 157, + 483 + ], + "score": 1.0, + "content": "screenshots.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + } + ], + "index": 6.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 118, + 443, + 584 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 106, + 82, + 504, + 116 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "Figure 6 shows the results from LayoutGAN and the retrieved most-similar real layouts from the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 506, + 106 + ], + "score": 1.0, + "content": "training set as references. Both the wireframe and the corresponding mask layouts are provided for", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 187, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 187, + 116 + ], + "score": 1.0, + "content": "better visualization.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "image_body", + "bbox": [ + 168, + 118, + 443, + 584 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 168, + 118, + 443, + 584 + ], + "spans": [ + { + "bbox": [ + 168, + 118, + 443, + 584 + ], + "score": 0.972, + "type": "image", + "image_path": "c2672c8eb692ba95e8eeb467cec04d726767fb2c2337614e94647e9c0b7aa8ff.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 168, + 118, + 443, + 130.59459459459458 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 168, + 130.59459459459458, + 443, + 143.18918918918916 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 168, + 143.18918918918916, + 443, + 155.78378378378375 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 168, + 155.78378378378375, + 443, + 168.37837837837833 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 168, + 168.37837837837833, + 443, + 180.9729729729729 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 168, + 180.9729729729729, + 443, + 193.5675675675675 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 168, + 193.5675675675675, + 443, + 206.16216216216208 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 168, + 206.16216216216208, + 443, + 218.75675675675666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 168, + 218.75675675675666, + 443, + 231.35135135135124 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 168, + 231.35135135135124, + 443, + 243.94594594594582 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 168, + 243.94594594594582, + 443, + 256.5405405405404 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 168, + 256.5405405405404, + 443, + 269.135135135135 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 168, + 269.135135135135, + 443, + 281.72972972972957 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 168, + 281.72972972972957, + 443, + 294.32432432432415 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 168, + 294.32432432432415, + 443, + 306.91891891891873 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 168, + 306.91891891891873, + 443, + 319.5135135135133 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 168, + 319.5135135135133, + 443, + 332.1081081081079 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 168, + 332.1081081081079, + 443, + 344.7027027027025 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 168, + 344.7027027027025, + 443, + 357.29729729729706 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 168, + 357.29729729729706, + 443, + 369.89189189189165 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 168, + 369.89189189189165, + 443, + 382.4864864864862 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 168, + 382.4864864864862, + 443, + 395.0810810810808 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 168, + 395.0810810810808, + 443, + 407.6756756756754 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 168, + 407.6756756756754, + 443, + 420.27027027027 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 168, + 420.27027027027, + 443, + 432.86486486486456 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 168, + 432.86486486486456, + 443, + 445.45945945945914 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 168, + 445.45945945945914, + 443, + 458.0540540540537 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 168, + 458.0540540540537, + 443, + 470.6486486486483 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 168, + 470.6486486486483, + 443, + 483.2432432432429 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 168, + 483.2432432432429, + 443, + 495.83783783783747 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 168, + 495.83783783783747, + 443, + 508.43243243243205 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 168, + 508.43243243243205, + 443, + 521.0270270270266 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 168, + 521.0270270270266, + 443, + 533.6216216216212 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 168, + 533.6216216216212, + 443, + 546.2162162162158 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 168, + 546.2162162162158, + 443, + 558.8108108108104 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 168, + 558.8108108108104, + 443, + 571.405405405405 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 168, + 571.405405405405, + 443, + 583.9999999999995 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 232, + 590, + 378, + 603 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 231, + 588, + 378, + 605 + ], + "spans": [ + { + "bbox": [ + 231, + 588, + 378, + 605 + ], + "score": 1.0, + "content": "Figure 6: Mobile app layout design.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + } + ], + "index": 21 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 118, + 443, + 584 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 106, + 82, + 504, + 116 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "Figure 6 shows the results from LayoutGAN and the retrieved most-similar real layouts from the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 506, + 106 + ], + "score": 1.0, + "content": "training set as references. Both the wireframe and the corresponding mask layouts are provided for", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 187, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 187, + 116 + ], + "score": 1.0, + "content": "better visualization.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "image_body", + "bbox": [ + 168, + 118, + 443, + 584 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 168, + 118, + 443, + 584 + ], + "spans": [ + { + "bbox": [ + 168, + 118, + 443, + 584 + ], + "score": 0.972, + "type": "image", + "image_path": "c2672c8eb692ba95e8eeb467cec04d726767fb2c2337614e94647e9c0b7aa8ff.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 168, + 118, + 443, + 130.59459459459458 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 168, + 130.59459459459458, + 443, + 143.18918918918916 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 168, + 143.18918918918916, + 443, + 155.78378378378375 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 168, + 155.78378378378375, + 443, + 168.37837837837833 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 168, + 168.37837837837833, + 443, + 180.9729729729729 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 168, + 180.9729729729729, + 443, + 193.5675675675675 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 168, + 193.5675675675675, + 443, + 206.16216216216208 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 168, + 206.16216216216208, + 443, + 218.75675675675666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 168, + 218.75675675675666, + 443, + 231.35135135135124 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 168, + 231.35135135135124, + 443, + 243.94594594594582 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 168, + 243.94594594594582, + 443, + 256.5405405405404 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 168, + 256.5405405405404, + 443, + 269.135135135135 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 168, + 269.135135135135, + 443, + 281.72972972972957 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 168, + 281.72972972972957, + 443, + 294.32432432432415 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 168, + 294.32432432432415, + 443, + 306.91891891891873 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 168, + 306.91891891891873, + 443, + 319.5135135135133 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 168, + 319.5135135135133, + 443, + 332.1081081081079 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 168, + 332.1081081081079, + 443, + 344.7027027027025 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 168, + 344.7027027027025, + 443, + 357.29729729729706 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 168, + 357.29729729729706, + 443, + 369.89189189189165 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 168, + 369.89189189189165, + 443, + 382.4864864864862 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 168, + 382.4864864864862, + 443, + 395.0810810810808 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 168, + 395.0810810810808, + 443, + 407.6756756756754 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 168, + 407.6756756756754, + 443, + 420.27027027027 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 168, + 420.27027027027, + 443, + 432.86486486486456 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 168, + 432.86486486486456, + 443, + 445.45945945945914 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 168, + 445.45945945945914, + 443, + 458.0540540540537 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 168, + 458.0540540540537, + 443, + 470.6486486486483 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 168, + 470.6486486486483, + 443, + 483.2432432432429 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 168, + 483.2432432432429, + 443, + 495.83783783783747 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 168, + 495.83783783783747, + 443, + 508.43243243243205 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 168, + 508.43243243243205, + 443, + 521.0270270270266 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 168, + 521.0270270270266, + 443, + 533.6216216216212 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 168, + 533.6216216216212, + 443, + 546.2162162162158 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 168, + 546.2162162162158, + 443, + 558.8108108108104 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 168, + 558.8108108108104, + 443, + 571.405405405405 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 168, + 571.405405405405, + 443, + 583.9999999999995 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 232, + 590, + 378, + 603 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 231, + 588, + 378, + 605 + ], + "spans": [ + { + "bbox": [ + 231, + 588, + 378, + 605 + ], + "score": 1.0, + "content": "Figure 6: Mobile app layout design.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + } + ], + "index": 21 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/HJxB5sRcFQ/HJxB5sRcFQ_model.json b/parse/train/HJxB5sRcFQ/HJxB5sRcFQ_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bb93b451730f8e6853628774573db75fac20e71e --- /dev/null +++ b/parse/train/HJxB5sRcFQ/HJxB5sRcFQ_model.json @@ -0,0 +1,20213 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1130, + 1404, + 1130, + 1404, + 1586, + 297, + 1586 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 398, + 679, + 1302, + 679, + 1302, + 1015, + 398, + 1015 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1601, + 1403, + 1601, + 1403, + 1846, + 298, + 1846 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1863, + 1402, + 1863, + 1402, + 1986, + 299, + 1986 + ], + "score": 0.967 + }, + { + "category_id": 0, + "poly": [ + 300, + 219, + 1401, + 219, + 1401, + 324, + 300, + 324 + ], + "score": 0.96 + }, + { + "category_id": 2, + "poly": [ + 331, + 2005, + 945, + 2005, + 945, + 2033, + 331, + 2033 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 301, + 1073, + 573, + 1073, + 573, + 1107, + 301, + 1107 + ], + "score": 0.894 + }, + { + "category_id": 1, + "poly": [ + 539, + 436, + 1194, + 436, + 1194, + 469, + 539, + 469 + ], + "score": 0.881 + }, + { + "category_id": 1, + "poly": [ + 312, + 373, + 1319, + 373, + 1319, + 410, + 312, + 410 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.878 + }, + { + "category_id": 0, + "poly": [ + 773, + 610, + 927, + 610, + 927, + 644, + 773, + 644 + ], + "score": 0.876 + }, + { + "category_id": 1, + "poly": [ + 301, + 497, + 1430, + 497, + 1430, + 531, + 301, + 531 + ], + "score": 0.857 + }, + { + "category_id": 2, + "poly": [ + 842, + 2089, + 857, + 2089, + 857, + 2112, + 842, + 2112 + ], + "score": 0.682 + }, + { + "category_id": 13, + "poly": [ + 399, + 375, + 442, + 375, + 442, + 407, + 399, + 407 + ], + "score": 0.54, + "latex": "\\mathbf { L i } ^ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1271, + 375, + 1321, + 375, + 1321, + 407, + 1271, + 407 + ], + "score": 0.26, + "latex": "\\mathbf { X } \\mathbf { u } ^ { 1 }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 215.0, + 1405.0, + 215.0, + 1405.0, + 274.0, + 293.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 277.0, + 1056.0, + 277.0, + 1056.0, + 327.0, + 296.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 2001.0, + 947.0, + 2001.0, + 947.0, + 2037.0, + 329.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1069.0, + 579.0, + 1069.0, + 579.0, + 1115.0, + 293.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 609.0, + 932.0, + 609.0, + 932.0, + 647.0, + 770.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 860.0, + 2088.0, + 860.0, + 2118.0, + 841.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1130.0, + 1402.0, + 1130.0, + 1402.0, + 1164.0, + 296.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1158.0, + 1404.0, + 1158.0, + 1404.0, + 1197.0, + 294.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1189.0, + 1405.0, + 1189.0, + 1405.0, + 1228.0, + 292.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1219.0, + 1402.0, + 1219.0, + 1402.0, + 1258.0, + 294.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1247.0, + 1404.0, + 1247.0, + 1404.0, + 1290.0, + 291.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1280.0, + 1404.0, + 1280.0, + 1404.0, + 1320.0, + 292.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1312.0, + 1406.0, + 1312.0, + 1406.0, + 1348.0, + 291.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1343.0, + 1406.0, + 1343.0, + 1406.0, + 1377.0, + 292.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1370.0, + 1406.0, + 1370.0, + 1406.0, + 1410.0, + 294.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1400.0, + 1405.0, + 1400.0, + 1405.0, + 1442.0, + 292.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1435.0, + 1404.0, + 1435.0, + 1404.0, + 1470.0, + 295.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1465.0, + 1405.0, + 1465.0, + 1405.0, + 1499.0, + 294.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1496.0, + 1404.0, + 1496.0, + 1404.0, + 1531.0, + 295.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1523.0, + 1405.0, + 1523.0, + 1405.0, + 1562.0, + 294.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1555.0, + 1024.0, + 1555.0, + 1024.0, + 1590.0, + 295.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 679.0, + 1306.0, + 679.0, + 1306.0, + 715.0, + 394.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 708.0, + 1304.0, + 708.0, + 1304.0, + 747.0, + 393.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 740.0, + 1306.0, + 740.0, + 1306.0, + 776.0, + 394.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 770.0, + 1306.0, + 770.0, + 1306.0, + 806.0, + 394.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 803.0, + 1305.0, + 803.0, + 1305.0, + 835.0, + 394.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 832.0, + 1306.0, + 832.0, + 1306.0, + 865.0, + 393.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 863.0, + 1305.0, + 863.0, + 1305.0, + 898.0, + 393.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 892.0, + 1307.0, + 892.0, + 1307.0, + 928.0, + 393.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 925.0, + 1305.0, + 925.0, + 1305.0, + 957.0, + 394.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 953.0, + 1305.0, + 953.0, + 1305.0, + 989.0, + 394.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 983.0, + 1210.0, + 983.0, + 1210.0, + 1022.0, + 392.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1603.0, + 1403.0, + 1603.0, + 1403.0, + 1636.0, + 296.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1632.0, + 1405.0, + 1632.0, + 1405.0, + 1667.0, + 292.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1664.0, + 1405.0, + 1664.0, + 1405.0, + 1697.0, + 293.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1695.0, + 1405.0, + 1695.0, + 1405.0, + 1729.0, + 293.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1721.0, + 1405.0, + 1721.0, + 1405.0, + 1762.0, + 292.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1755.0, + 1405.0, + 1755.0, + 1405.0, + 1792.0, + 293.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1787.0, + 1405.0, + 1787.0, + 1405.0, + 1821.0, + 294.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1814.0, + 674.0, + 1814.0, + 674.0, + 1851.0, + 293.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1863.0, + 1406.0, + 1863.0, + 1406.0, + 1895.0, + 295.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1894.0, + 1406.0, + 1894.0, + 1406.0, + 1927.0, + 295.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1923.0, + 1406.0, + 1923.0, + 1406.0, + 1957.0, + 294.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1951.0, + 1406.0, + 1951.0, + 1406.0, + 1991.0, + 293.0, + 1991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 429.0, + 923.0, + 429.0, + 923.0, + 478.0, + 537.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 434.0, + 1196.0, + 434.0, + 1196.0, + 471.0, + 972.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 368.0, + 398.0, + 368.0, + 398.0, + 418.0, + 307.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 368.0, + 1270.0, + 368.0, + 1270.0, + 418.0, + 443.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1322.0, + 368.0, + 1326.0, + 368.0, + 1326.0, + 418.0, + 1322.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 492.0, + 1436.0, + 492.0, + 1436.0, + 538.0, + 292.0, + 538.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 231, + 1403, + 231, + 1403, + 717, + 298, + 717 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1099, + 1404, + 1099, + 1404, + 1527, + 298, + 1527 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1536, + 1404, + 1536, + 1404, + 1963, + 298, + 1963 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 299, + 733, + 1402, + 733, + 1402, + 857, + 299, + 857 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 872, + 1402, + 872, + 1402, + 995, + 299, + 995 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 295, + 1972, + 1398, + 1972, + 1398, + 2034, + 295, + 2034 + ], + "score": 0.935 + }, + { + "category_id": 0, + "poly": [ + 301, + 1038, + 587, + 1038, + 587, + 1074, + 301, + 1074 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2112, + 841, + 2112 + ], + "score": 0.718 + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1035.0, + 594.0, + 1035.0, + 594.0, + 1082.0, + 291.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2121.0, + 839.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 231.0, + 1405.0, + 231.0, + 1405.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 260.0, + 1407.0, + 260.0, + 1407.0, + 298.0, + 293.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 288.0, + 1407.0, + 288.0, + 1407.0, + 329.0, + 292.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 320.0, + 1404.0, + 320.0, + 1404.0, + 358.0, + 293.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 353.0, + 1405.0, + 353.0, + 1405.0, + 387.0, + 294.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 385.0, + 1404.0, + 385.0, + 1404.0, + 415.0, + 296.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 410.0, + 1404.0, + 410.0, + 1404.0, + 451.0, + 293.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 446.0, + 1404.0, + 446.0, + 1404.0, + 476.0, + 296.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 474.0, + 1405.0, + 474.0, + 1405.0, + 512.0, + 293.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 505.0, + 1405.0, + 505.0, + 1405.0, + 539.0, + 294.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 535.0, + 1405.0, + 535.0, + 1405.0, + 569.0, + 293.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 564.0, + 1407.0, + 564.0, + 1407.0, + 602.0, + 294.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 596.0, + 1405.0, + 596.0, + 1405.0, + 630.0, + 296.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 625.0, + 1404.0, + 625.0, + 1404.0, + 660.0, + 294.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 656.0, + 1405.0, + 656.0, + 1405.0, + 693.0, + 293.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 689.0, + 411.0, + 689.0, + 411.0, + 718.0, + 292.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1093.0, + 1406.0, + 1093.0, + 1406.0, + 1138.0, + 292.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1129.0, + 1406.0, + 1129.0, + 1406.0, + 1165.0, + 294.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1158.0, + 1405.0, + 1158.0, + 1405.0, + 1196.0, + 293.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1190.0, + 1405.0, + 1190.0, + 1405.0, + 1225.0, + 294.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1222.0, + 1405.0, + 1222.0, + 1405.0, + 1257.0, + 294.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1252.0, + 1401.0, + 1252.0, + 1401.0, + 1284.0, + 296.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1280.0, + 1402.0, + 1280.0, + 1402.0, + 1316.0, + 293.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1312.0, + 1404.0, + 1312.0, + 1404.0, + 1347.0, + 292.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1341.0, + 1406.0, + 1341.0, + 1406.0, + 1380.0, + 292.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1374.0, + 1406.0, + 1374.0, + 1406.0, + 1409.0, + 293.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1402.0, + 1408.0, + 1402.0, + 1408.0, + 1441.0, + 292.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1432.0, + 1405.0, + 1432.0, + 1405.0, + 1470.0, + 292.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1463.0, + 1405.0, + 1463.0, + 1405.0, + 1502.0, + 293.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1496.0, + 718.0, + 1496.0, + 718.0, + 1530.0, + 293.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1537.0, + 1402.0, + 1537.0, + 1402.0, + 1568.0, + 296.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1565.0, + 1404.0, + 1565.0, + 1404.0, + 1601.0, + 293.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1596.0, + 1406.0, + 1596.0, + 1406.0, + 1632.0, + 294.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1629.0, + 1401.0, + 1629.0, + 1401.0, + 1661.0, + 296.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1654.0, + 1406.0, + 1654.0, + 1406.0, + 1695.0, + 293.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1689.0, + 1404.0, + 1689.0, + 1404.0, + 1724.0, + 294.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1720.0, + 1402.0, + 1720.0, + 1402.0, + 1752.0, + 296.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1747.0, + 1404.0, + 1747.0, + 1404.0, + 1784.0, + 292.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1776.0, + 1404.0, + 1776.0, + 1404.0, + 1817.0, + 293.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1808.0, + 1404.0, + 1808.0, + 1404.0, + 1846.0, + 293.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1839.0, + 1406.0, + 1839.0, + 1406.0, + 1876.0, + 292.0, + 1876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1872.0, + 1404.0, + 1872.0, + 1404.0, + 1904.0, + 294.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1903.0, + 1405.0, + 1903.0, + 1405.0, + 1934.0, + 294.0, + 1934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1932.0, + 1347.0, + 1932.0, + 1347.0, + 1967.0, + 294.0, + 1967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 733.0, + 1404.0, + 733.0, + 1404.0, + 768.0, + 295.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 765.0, + 1406.0, + 765.0, + 1406.0, + 800.0, + 293.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 796.0, + 1406.0, + 796.0, + 1406.0, + 831.0, + 293.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 827.0, + 1273.0, + 827.0, + 1273.0, + 859.0, + 295.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 873.0, + 1403.0, + 873.0, + 1403.0, + 906.0, + 295.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 903.0, + 1406.0, + 903.0, + 1406.0, + 938.0, + 293.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 931.0, + 1406.0, + 931.0, + 1406.0, + 971.0, + 293.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 965.0, + 825.0, + 965.0, + 825.0, + 997.0, + 295.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1970.0, + 1403.0, + 1970.0, + 1403.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1403.0, + 2003.0, + 1403.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1212, + 1404, + 1212, + 1404, + 1490, + 297, + 1490 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1559, + 1403, + 1559, + 1403, + 1835, + 297, + 1835 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1850, + 1403, + 1850, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 882, + 1403, + 882, + 1403, + 1036, + 298, + 1036 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 297, + 196, + 1401, + 196, + 1401, + 591, + 297, + 591 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 296, + 617, + 1404, + 617, + 1404, + 863, + 296, + 863 + ], + "score": 0.95 + }, + { + "category_id": 0, + "poly": [ + 299, + 1168, + 680, + 1168, + 680, + 1199, + 299, + 1199 + ], + "score": 0.903 + }, + { + "category_id": 0, + "poly": [ + 301, + 1515, + 706, + 1515, + 706, + 1546, + 301, + 1546 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 299, + 1068, + 549, + 1068, + 549, + 1103, + 299, + 1103 + ], + "score": 0.896 + }, + { + "category_id": 1, + "poly": [ + 299, + 1113, + 946, + 1113, + 946, + 1144, + 299, + 1144 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 814, + 76, + 814, + 104, + 298, + 104 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.626 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.356 + }, + { + "category_id": 13, + "poly": [ + 1023, + 1365, + 1273, + 1365, + 1273, + 1403, + 1023, + 1403 + ], + "score": 0.94, + "latex": "\\pmb \\theta \\equiv \\left[ x ^ { L } , y ^ { T } , x ^ { R } , y ^ { B } \\right]" + }, + { + "category_id": 13, + "poly": [ + 577, + 1429, + 741, + 1429, + 741, + 1462, + 577, + 1462 + ], + "score": 0.93, + "latex": "\\pmb \\theta \\equiv [ x , y , s , l ]" + }, + { + "category_id": 13, + "poly": [ + 462, + 1650, + 873, + 1650, + 873, + 1686, + 462, + 1686 + ], + "score": 0.92, + "latex": "G ( z ) = \\{ ( p _ { 1 } ^ { \\prime } , \\theta _ { 1 } ^ { \\prime } ) , \\cdot \\cdot \\cdot , \\bar { ( } p _ { N } ^ { \\prime } , \\theta _ { N } ^ { \\prime } ) \\}" + }, + { + "category_id": 13, + "poly": [ + 954, + 1334, + 1072, + 1334, + 1072, + 1367, + 954, + 1367 + ], + "score": 0.92, + "latex": "\\mathbf { \\bar { \\theta } } \\equiv [ \\mathbf { \\bar { \\boldsymbol { x } } } , y ]" + }, + { + "category_id": 13, + "poly": [ + 857, + 1560, + 922, + 1560, + 922, + 1593, + 857, + 1593 + ], + "score": 0.92, + "latex": "G ( z )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1243, + 600, + 1243, + 600, + 1277, + 298, + 1277 + ], + "score": 0.92, + "latex": "\\{ ( p _ { 1 } , \\pmb { \\theta _ { 1 } } ) , \\cdot \\cdot \\cdot , ( p _ { N } , \\pmb { \\theta } _ { N } ) \\}" + }, + { + "category_id": 13, + "poly": [ + 299, + 1590, + 602, + 1590, + 602, + 1625, + 299, + 1625 + ], + "score": 0.89, + "latex": "\\{ ( p _ { 1 } , \\pmb { \\theta _ { 1 } } ) , \\cdot \\cdot \\cdot , ( p _ { N } , \\pmb { \\theta } _ { N } ) \\}" + }, + { + "category_id": 13, + "poly": [ + 464, + 1623, + 492, + 1623, + 492, + 1651, + 464, + 1651 + ], + "score": 0.85, + "latex": "\\theta _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1081, + 1624, + 1111, + 1624, + 1111, + 1653, + 1081, + 1653 + ], + "score": 0.84, + "latex": "\\mathbf { \\mathit { p } } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1343, + 1564, + 1403, + 1564, + 1403, + 1592, + 1343, + 1592 + ], + "score": 0.81, + "latex": "z =" + }, + { + "category_id": 13, + "poly": [ + 1046, + 1213, + 1076, + 1213, + 1076, + 1241, + 1046, + 1241 + ], + "score": 0.78, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 503, + 1280, + 523, + 1280, + 523, + 1305, + 503, + 1305 + ], + "score": 0.77, + "latex": "\\pmb { p }" + }, + { + "category_id": 13, + "poly": [ + 341, + 1718, + 363, + 1718, + 363, + 1740, + 341, + 1740 + ], + "score": 0.73, + "latex": "_ { z }" + }, + { + "category_id": 13, + "poly": [ + 872, + 1688, + 893, + 1688, + 893, + 1709, + 872, + 1709 + ], + "score": 0.72, + "latex": "_ z" + }, + { + "category_id": 13, + "poly": [ + 1187, + 1245, + 1207, + 1245, + 1207, + 1271, + 1187, + 1271 + ], + "score": 0.68, + "latex": "\\pmb { \\theta }" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 196.0, + 577.0, + 196.0, + 577.0, + 224.0, + 457.0, + 224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 204.0, + 691.0, + 204.0, + 691.0, + 230.0, + 618.0, + 230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 207.0, + 994.0, + 207.0, + 994.0, + 227.0, + 796.0, + 227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 204.0, + 1165.0, + 204.0, + 1165.0, + 230.0, + 1091.0, + 230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 220.0, + 585.0, + 220.0, + 585.0, + 290.0, + 455.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 249.0, + 650.0, + 249.0, + 650.0, + 270.0, + 628.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 249.0, + 678.0, + 249.0, + 678.0, + 270.0, + 657.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 236.0, + 1046.0, + 236.0, + 1046.0, + 296.0, + 735.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 249.0, + 1124.0, + 249.0, + 1124.0, + 270.0, + 1102.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 249.0, + 1153.0, + 249.0, + 1153.0, + 270.0, + 1131.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 259.0, + 1210.0, + 259.0, + 1210.0, + 289.0, + 1178.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 227.0, + 1232.0, + 227.0, + 1232.0, + 257.0, + 1211.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 229.0, + 1390.0, + 229.0, + 1390.0, + 288.0, + 1235.0, + 288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 299.0, + 416.0, + 299.0, + 416.0, + 324.0, + 336.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 302.0, + 504.0, + 302.0, + 504.0, + 363.0, + 483.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 332.0, + 525.0, + 332.0, + 525.0, + 354.0, + 510.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 299.0, + 553.0, + 299.0, + 553.0, + 321.0, + 532.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 311.0, + 719.0, + 311.0, + 719.0, + 317.0, + 712.0, + 317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 290.0, + 787.0, + 290.0, + 787.0, + 359.0, + 734.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 308.0, + 835.0, + 308.0, + 835.0, + 343.0, + 804.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 346.0, + 929.0, + 346.0, + 929.0, + 366.0, + 844.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 289.0, + 1045.0, + 289.0, + 1045.0, + 359.0, + 978.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 308.0, + 1085.0, + 308.0, + 1085.0, + 337.0, + 1065.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 339.0, + 1210.0, + 339.0, + 1210.0, + 367.0, + 1179.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 315.0, + 1233.0, + 315.0, + 1233.0, + 337.0, + 1213.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 342.0, + 1265.0, + 342.0, + 1265.0, + 373.0, + 1234.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 285.0, + 1329.0, + 285.0, + 1329.0, + 354.0, + 1284.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1334.0, + 287.0, + 1364.0, + 287.0, + 1364.0, + 352.0, + 1334.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 368.0, + 584.0, + 368.0, + 584.0, + 396.0, + 456.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 356.0, + 786.0, + 356.0, + 786.0, + 397.0, + 731.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 356.0, + 1048.0, + 356.0, + 1048.0, + 399.0, + 989.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 359.0, + 1398.0, + 359.0, + 1398.0, + 383.0, + 1263.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 422.0, + 1021.0, + 422.0, + 1021.0, + 446.0, + 872.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 422.0, + 1194.0, + 422.0, + 1194.0, + 445.0, + 1151.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 437.0, + 509.0, + 437.0, + 509.0, + 455.0, + 474.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 437.0, + 582.0, + 437.0, + 582.0, + 454.0, + 523.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 446.0, + 584.0, + 446.0, + 584.0, + 468.0, + 455.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 463.0, + 509.0, + 463.0, + 509.0, + 488.0, + 482.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 466.0, + 557.0, + 466.0, + 557.0, + 486.0, + 532.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 470.0, + 767.0, + 470.0, + 767.0, + 494.0, + 665.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 467.0, + 1332.0, + 467.0, + 1332.0, + 489.0, + 1312.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 480.0, + 455.0, + 480.0, + 455.0, + 525.0, + 294.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 497.0, + 508.0, + 497.0, + 508.0, + 517.0, + 483.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 494.0, + 558.0, + 494.0, + 558.0, + 519.0, + 529.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 485.0, + 794.0, + 485.0, + 794.0, + 517.0, + 639.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 496.0, + 1402.0, + 496.0, + 1402.0, + 521.0, + 1329.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 528.0, + 559.0, + 528.0, + 559.0, + 558.0, + 509.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 522.0, + 813.0, + 522.0, + 813.0, + 547.0, + 622.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 514.0, + 1400.0, + 514.0, + 1400.0, + 539.0, + 1329.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 564.0, + 593.0, + 564.0, + 593.0, + 592.0, + 454.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.75, + 339.5, + 549.75, + 339.5, + 549.75, + 358.5, + 533.75, + 358.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.75, + 532.0, + 517.75, + 532.0, + 517.75, + 560.5, + 475.75, + 560.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 619.0, + 1404.0, + 619.0, + 1404.0, + 652.0, + 296.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 649.0, + 1404.0, + 649.0, + 1404.0, + 682.0, + 293.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 680.0, + 1404.0, + 680.0, + 1404.0, + 714.0, + 295.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 709.0, + 1404.0, + 709.0, + 1404.0, + 744.0, + 294.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 741.0, + 1404.0, + 741.0, + 1404.0, + 774.0, + 293.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 771.0, + 1406.0, + 771.0, + 1406.0, + 805.0, + 294.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 802.0, + 1405.0, + 802.0, + 1405.0, + 836.0, + 295.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 831.0, + 1402.0, + 831.0, + 1402.0, + 867.0, + 294.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1167.0, + 683.0, + 1167.0, + 683.0, + 1203.0, + 295.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1514.0, + 711.0, + 1514.0, + 711.0, + 1550.0, + 294.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1065.0, + 552.0, + 1065.0, + 552.0, + 1110.0, + 293.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1209.0, + 1045.0, + 1209.0, + 1045.0, + 1247.0, + 294.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 1209.0, + 1406.0, + 1209.0, + 1406.0, + 1247.0, + 1077.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1242.0, + 1186.0, + 1242.0, + 1186.0, + 1279.0, + 601.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 1242.0, + 1406.0, + 1242.0, + 1406.0, + 1279.0, + 1208.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1275.0, + 502.0, + 1275.0, + 502.0, + 1308.0, + 295.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1275.0, + 1405.0, + 1275.0, + 1405.0, + 1308.0, + 524.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1301.0, + 1405.0, + 1301.0, + 1405.0, + 1342.0, + 292.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1334.0, + 953.0, + 1334.0, + 953.0, + 1371.0, + 294.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1334.0, + 1405.0, + 1334.0, + 1405.0, + 1371.0, + 1073.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1356.0, + 1022.0, + 1356.0, + 1022.0, + 1410.0, + 288.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 1356.0, + 1409.0, + 1356.0, + 1409.0, + 1410.0, + 1274.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1396.0, + 1402.0, + 1396.0, + 1402.0, + 1432.0, + 294.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1427.0, + 576.0, + 1427.0, + 576.0, + 1464.0, + 294.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 1427.0, + 1405.0, + 1427.0, + 1405.0, + 1464.0, + 742.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1459.0, + 398.0, + 1459.0, + 398.0, + 1493.0, + 294.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1556.0, + 856.0, + 1556.0, + 856.0, + 1598.0, + 292.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 1556.0, + 1342.0, + 1556.0, + 1342.0, + 1598.0, + 923.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1589.0, + 298.0, + 1589.0, + 298.0, + 1629.0, + 293.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 1589.0, + 1404.0, + 1589.0, + 1404.0, + 1629.0, + 603.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1619.0, + 463.0, + 1619.0, + 463.0, + 1655.0, + 292.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 1619.0, + 1080.0, + 1619.0, + 1080.0, + 1655.0, + 493.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 1619.0, + 1407.0, + 1619.0, + 1407.0, + 1655.0, + 1112.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1646.0, + 461.0, + 1646.0, + 461.0, + 1692.0, + 290.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 1646.0, + 1408.0, + 1646.0, + 1408.0, + 1692.0, + 874.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1681.0, + 871.0, + 1681.0, + 871.0, + 1717.0, + 292.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1681.0, + 1405.0, + 1681.0, + 1405.0, + 1717.0, + 894.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1713.0, + 340.0, + 1713.0, + 340.0, + 1747.0, + 291.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1713.0, + 1405.0, + 1713.0, + 1405.0, + 1747.0, + 364.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1743.0, + 1405.0, + 1743.0, + 1405.0, + 1779.0, + 293.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1771.0, + 1408.0, + 1771.0, + 1408.0, + 1811.0, + 292.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1800.0, + 584.0, + 1800.0, + 584.0, + 1841.0, + 294.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1849.0, + 1405.0, + 1849.0, + 1405.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 1404.0, + 1880.0, + 1404.0, + 1915.0, + 293.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1910.0, + 1403.0, + 1910.0, + 1403.0, + 1946.0, + 293.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1978.0, + 292.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1970.0, + 1405.0, + 1970.0, + 1405.0, + 2007.0, + 293.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2000.0, + 1405.0, + 2000.0, + 1405.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 882.0, + 1405.0, + 882.0, + 1405.0, + 919.0, + 294.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 913.0, + 1403.0, + 913.0, + 1403.0, + 950.0, + 293.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 944.0, + 1405.0, + 944.0, + 1405.0, + 980.0, + 292.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 974.0, + 1405.0, + 974.0, + 1405.0, + 1011.0, + 293.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1004.0, + 1365.0, + 1004.0, + 1365.0, + 1041.0, + 294.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1108.0, + 950.0, + 1108.0, + 950.0, + 1149.0, + 295.0, + 1149.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1075, + 1405, + 1075, + 1405, + 1292, + 297, + 1292 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 296, + 418, + 1404, + 418, + 1404, + 602, + 296, + 602 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 664, + 1404, + 664, + 1404, + 851, + 297, + 851 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 923, + 1404, + 923, + 1404, + 1016, + 298, + 1016 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1596, + 1404, + 1596, + 1404, + 1691, + 298, + 1691 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1459, + 1404, + 1459, + 1404, + 1583, + 298, + 1583 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1888, + 1404, + 1888, + 1404, + 1981, + 298, + 1981 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 228, + 1404, + 228, + 1404, + 323, + 299, + 323 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 301, + 1351, + 1404, + 1351, + 1404, + 1445, + 301, + 1445 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 661, + 1701, + 1040, + 1701, + 1040, + 1755, + 661, + 1755 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 299, + 1811, + 1403, + 1811, + 1403, + 1875, + 299, + 1875 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 663, + 1994, + 1037, + 1994, + 1037, + 2034, + 663, + 2034 + ], + "score": 0.94 + }, + { + "category_id": 8, + "poly": [ + 419, + 335, + 1281, + 335, + 1281, + 413, + 419, + 413 + ], + "score": 0.937 + }, + { + "category_id": 1, + "poly": [ + 298, + 1764, + 911, + 1764, + 911, + 1798, + 298, + 1798 + ], + "score": 0.931 + }, + { + "category_id": 8, + "poly": [ + 527, + 615, + 1170, + 615, + 1170, + 656, + 527, + 656 + ], + "score": 0.929 + }, + { + "category_id": 0, + "poly": [ + 298, + 879, + 896, + 879, + 896, + 909, + 298, + 909 + ], + "score": 0.929 + }, + { + "category_id": 0, + "poly": [ + 298, + 1312, + 886, + 1312, + 886, + 1342, + 298, + 1342 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 816, + 74, + 816, + 105, + 299, + 105 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 299, + 1037, + 799, + 1037, + 799, + 1066, + 299, + 1066 + ], + "score": 0.904 + }, + { + "category_id": 9, + "poly": [ + 1366, + 619, + 1399, + 619, + 1399, + 649, + 1366, + 649 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1705, + 1400, + 1705, + 1400, + 1734, + 1366, + 1734 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1366, + 360, + 1399, + 360, + 1399, + 389, + 1366, + 389 + ], + "score": 0.876 + }, + { + "category_id": 9, + "poly": [ + 1366, + 2000, + 1400, + 2000, + 1400, + 2028, + 1366, + 2028 + ], + "score": 0.863 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.788 + }, + { + "category_id": 14, + "poly": [ + 658, + 1700, + 1042, + 1700, + 1042, + 1756, + 658, + 1756 + ], + "score": 0.94, + "latex": "\\pmb { I } ( x , y , c ) = \\operatorname* { m a x } _ { i \\in [ 1 . . . N ] } p _ { i , c } \\pmb { F } _ { \\theta _ { i } } ( x , y )" + }, + { + "category_id": 13, + "poly": [ + 1051, + 1889, + 1198, + 1889, + 1198, + 1923, + 1051, + 1923 + ], + "score": 0.93, + "latex": "\\theta _ { i } = ( x _ { i } , y _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 1168, + 418, + 1277, + 418, + 1277, + 452, + 1168, + 452 + ], + "score": 0.93, + "latex": "f ( p _ { j } , \\theta _ { j } )" + }, + { + "category_id": 13, + "poly": [ + 400, + 571, + 504, + 571, + 504, + 604, + 400, + 604 + ], + "score": 0.93, + "latex": "f ( p _ { i } , \\theta _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 947, + 230, + 1053, + 230, + 1053, + 264, + 947, + 264 + ], + "score": 0.93, + "latex": "f ( p _ { i } , \\theta _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 376, + 665, + 726, + 665, + 726, + 701, + 376, + 701 + ], + "score": 0.93, + "latex": "\\psi ( f ( p _ { i } , \\pmb { \\theta } _ { i } ) ) = W _ { \\psi } f ( p _ { i } , \\pmb { \\theta } _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 871, + 261, + 984, + 261, + 984, + 295, + 871, + 295 + ], + "score": 0.93, + "latex": "f ^ { \\prime } ( p _ { i } , \\theta _ { i } )" + }, + { + "category_id": 14, + "poly": [ + 418, + 330, + 1279, + 330, + 1279, + 417, + 418, + 417 + ], + "score": 0.93, + "latex": "f ^ { \\prime } ( p _ { i } , \\theta _ { i } ) = W _ { r } \\frac { 1 } { N } \\sum _ { \\forall j \\neq i } H ( f ( p _ { i } , \\theta _ { i } ) , f ( p _ { j } , \\theta _ { j } ) ) U ( f ( p _ { j } , \\theta _ { j } ) ) + f ( p _ { i } , \\theta _ { i } ) ," + }, + { + "category_id": 13, + "poly": [ + 723, + 1170, + 828, + 1170, + 828, + 1202, + 723, + 1202 + ], + "score": 0.92, + "latex": "r ( p _ { i } , \\theta _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1521, + 397, + 1521, + 397, + 1554, + 298, + 1554 + ], + "score": 0.92, + "latex": "F _ { \\theta } ( x , y )" + }, + { + "category_id": 13, + "poly": [ + 771, + 1137, + 876, + 1137, + 876, + 1170, + 771, + 1170 + ], + "score": 0.92, + "latex": "f ( p _ { i } , \\theta _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 784, + 665, + 1138, + 665, + 1138, + 701, + 784, + 701 + ], + "score": 0.92, + "latex": "\\phi ( f ( p _ { j } , \\pmb { \\theta } _ { j } ) ) = W _ { \\phi } f ( p _ { j } , \\pmb { \\theta } _ { j } )" + }, + { + "category_id": 13, + "poly": [ + 404, + 1490, + 688, + 1490, + 688, + 1524, + 404, + 1524 + ], + "score": 0.92, + "latex": "\\{ ( p _ { 1 } , \\pmb { \\theta _ { 1 } } ) , . . . , ( p _ { N } , \\pmb { \\theta } _ { N } ) \\}" + }, + { + "category_id": 13, + "poly": [ + 719, + 480, + 796, + 480, + 796, + 512, + 719, + 512 + ], + "score": 0.91, + "latex": "j \\neq i" + }, + { + "category_id": 14, + "poly": [ + 662, + 1994, + 1039, + 1994, + 1039, + 2032, + 662, + 2032 + ], + "score": 0.91, + "latex": "F _ { \\theta _ { i } } ( x , y ) = k ( x - x _ { i } ) k ( y - y _ { i } ) ." + }, + { + "category_id": 13, + "poly": [ + 583, + 1552, + 673, + 1552, + 673, + 1581, + 583, + 1581 + ], + "score": 0.91, + "latex": "W \\times H" + }, + { + "category_id": 13, + "poly": [ + 1305, + 1921, + 1371, + 1921, + 1371, + 1953, + 1305, + 1953 + ], + "score": 0.91, + "latex": "( x , y )" + }, + { + "category_id": 13, + "poly": [ + 990, + 1598, + 1157, + 1598, + 1157, + 1628, + 990, + 1628 + ], + "score": 0.91, + "latex": "W \\times H \\times M" + }, + { + "category_id": 13, + "poly": [ + 1051, + 1630, + 1117, + 1630, + 1117, + 1663, + 1051, + 1663 + ], + "score": 0.91, + "latex": "( x , y )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1168, + 641, + 1168, + 641, + 1202, + 297, + 1202 + ], + "score": 0.89, + "latex": "g ( r ( p _ { 1 } , \\theta _ { 1 } ) , \\cdot \\cdot \\cdot , r ( p _ { N } , \\theta _ { N } ) )" + }, + { + "category_id": 14, + "poly": [ + 529, + 613, + 1167, + 613, + 1167, + 655, + 529, + 655 + ], + "score": 0.89, + "latex": "\\pmb { H } ( f ( p _ { i } , \\pmb { \\theta } _ { i } ) , f ( p _ { j } , \\pmb { \\theta } _ { j } ) ) = \\psi ( f ( p _ { i } , \\pmb { \\theta } _ { i } ) ) ^ { T } \\phi ( f ( p _ { j } , \\pmb { \\theta } _ { j } ) ) ," + }, + { + "category_id": 13, + "poly": [ + 560, + 541, + 607, + 541, + 607, + 571, + 560, + 571 + ], + "score": 0.88, + "latex": "W _ { r }" + }, + { + "category_id": 13, + "poly": [ + 901, + 1814, + 934, + 1814, + 934, + 1843, + 901, + 1843 + ], + "score": 0.88, + "latex": "F _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 371, + 1770, + 415, + 1770, + 415, + 1800, + 371, + 1800 + ], + "score": 0.88, + "latex": "p _ { i , c }" + }, + { + "category_id": 13, + "poly": [ + 513, + 1553, + 539, + 1553, + 539, + 1579, + 513, + 1579 + ], + "score": 0.84, + "latex": "\\pmb { F }" + }, + { + "category_id": 13, + "poly": [ + 759, + 1553, + 791, + 1553, + 791, + 1579, + 759, + 1579 + ], + "score": 0.84, + "latex": "W" + }, + { + "category_id": 13, + "poly": [ + 842, + 1553, + 870, + 1553, + 870, + 1579, + 842, + 1579 + ], + "score": 0.83, + "latex": "H" + }, + { + "category_id": 13, + "poly": [ + 571, + 421, + 599, + 421, + 599, + 446, + 571, + 446 + ], + "score": 0.82, + "latex": "U" + }, + { + "category_id": 13, + "poly": [ + 756, + 1768, + 768, + 1768, + 768, + 1793, + 756, + 1793 + ], + "score": 0.82, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 834, + 1204, + 854, + 1204, + 854, + 1231, + 834, + 1231 + ], + "score": 0.82, + "latex": "\\textbf { { g } }" + }, + { + "category_id": 13, + "poly": [ + 684, + 1923, + 702, + 1923, + 702, + 1947, + 684, + 1947 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 347, + 482, + 364, + 482, + 364, + 512, + 347, + 512 + ], + "score": 0.81, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 1103, + 571, + 1135, + 571, + 1135, + 598, + 1103, + 598 + ], + "score": 0.81, + "latex": "\\pmb { H }" + }, + { + "category_id": 13, + "poly": [ + 298, + 451, + 314, + 451, + 314, + 481, + 298, + 481 + ], + "score": 0.8, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 1154, + 1814, + 1173, + 1814, + 1173, + 1840, + 1154, + 1840 + ], + "score": 0.8, + "latex": "\\pmb \\theta" + }, + { + "category_id": 13, + "poly": [ + 593, + 1630, + 626, + 1630, + 626, + 1658, + 593, + 1658 + ], + "score": 0.79, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 590, + 450, + 623, + 450, + 623, + 477, + 590, + 477 + ], + "score": 0.79, + "latex": "\\pmb { H }" + }, + { + "category_id": 13, + "poly": [ + 1155, + 1631, + 1175, + 1631, + 1175, + 1657, + 1155, + 1657 + ], + "score": 0.79, + "latex": "\\pmb { I }" + }, + { + "category_id": 13, + "poly": [ + 298, + 542, + 325, + 542, + 325, + 568, + 298, + 568 + ], + "score": 0.78, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 997, + 1462, + 1027, + 1462, + 1027, + 1488, + 997, + 1488 + ], + "score": 0.77, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 793, + 1600, + 813, + 1600, + 813, + 1626, + 793, + 1626 + ], + "score": 0.77, + "latex": "\\pmb { I }" + }, + { + "category_id": 13, + "poly": [ + 889, + 1773, + 904, + 1773, + 904, + 1793, + 889, + 1793 + ], + "score": 0.75, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 1388, + 451, + 1400, + 451, + 1400, + 476, + 1388, + 476 + ], + "score": 0.74, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1349, + 482, + 1363, + 482, + 1363, + 507, + 1349, + 507 + ], + "score": 0.72, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 482, + 264, + 495, + 264, + 495, + 290, + 482, + 290 + ], + "score": 0.72, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 664, + 1415, + 683, + 1415, + 683, + 1441, + 664, + 1441 + ], + "score": 0.69, + "latex": "\\pmb { I }" + }, + { + "category_id": 13, + "poly": [ + 1336, + 1892, + 1348, + 1892, + 1348, + 1917, + 1336, + 1917 + ], + "score": 0.53, + "latex": "i" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 878.0, + 896.0, + 878.0, + 896.0, + 912.0, + 295.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1310.0, + 889.0, + 1310.0, + 889.0, + 1346.0, + 294.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1034.0, + 804.0, + 1034.0, + 804.0, + 1071.0, + 294.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1073.0, + 1406.0, + 1073.0, + 1406.0, + 1113.0, + 292.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1108.0, + 1405.0, + 1108.0, + 1405.0, + 1143.0, + 295.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1136.0, + 770.0, + 1136.0, + 770.0, + 1174.0, + 292.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1136.0, + 1406.0, + 1136.0, + 1406.0, + 1174.0, + 877.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1167.0, + 722.0, + 1167.0, + 722.0, + 1206.0, + 642.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 1167.0, + 1406.0, + 1167.0, + 1406.0, + 1206.0, + 829.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1197.0, + 833.0, + 1197.0, + 833.0, + 1233.0, + 292.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 1197.0, + 1406.0, + 1197.0, + 1406.0, + 1233.0, + 855.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1226.0, + 1408.0, + 1226.0, + 1408.0, + 1266.0, + 292.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1258.0, + 1298.0, + 1258.0, + 1298.0, + 1295.0, + 294.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 417.0, + 570.0, + 417.0, + 570.0, + 453.0, + 295.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 417.0, + 1167.0, + 417.0, + 1167.0, + 453.0, + 600.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 417.0, + 1406.0, + 417.0, + 1406.0, + 453.0, + 1278.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 448.0, + 297.0, + 448.0, + 297.0, + 482.0, + 291.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 448.0, + 589.0, + 448.0, + 589.0, + 482.0, + 315.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 448.0, + 1387.0, + 448.0, + 1387.0, + 482.0, + 624.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 448.0, + 1407.0, + 448.0, + 1407.0, + 482.0, + 1401.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 477.0, + 346.0, + 477.0, + 346.0, + 513.0, + 294.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 477.0, + 718.0, + 477.0, + 718.0, + 513.0, + 365.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 477.0, + 1348.0, + 477.0, + 1348.0, + 513.0, + 797.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 477.0, + 1402.0, + 477.0, + 1402.0, + 513.0, + 1364.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 510.0, + 1405.0, + 510.0, + 1405.0, + 542.0, + 294.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 538.0, + 297.0, + 538.0, + 297.0, + 575.0, + 294.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 538.0, + 559.0, + 538.0, + 559.0, + 575.0, + 326.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 538.0, + 1406.0, + 538.0, + 1406.0, + 575.0, + 608.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 570.0, + 399.0, + 570.0, + 399.0, + 605.0, + 295.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 570.0, + 1102.0, + 570.0, + 1102.0, + 605.0, + 505.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 570.0, + 1331.0, + 570.0, + 1331.0, + 605.0, + 1136.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 664.0, + 375.0, + 664.0, + 375.0, + 704.0, + 295.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 664.0, + 783.0, + 664.0, + 783.0, + 704.0, + 727.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 664.0, + 1405.0, + 664.0, + 1405.0, + 704.0, + 1139.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 699.0, + 1404.0, + 699.0, + 1404.0, + 731.0, + 296.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 727.0, + 1406.0, + 727.0, + 1406.0, + 763.0, + 295.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 756.0, + 1406.0, + 756.0, + 1406.0, + 791.0, + 294.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 788.0, + 1405.0, + 788.0, + 1405.0, + 824.0, + 295.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 816.0, + 900.0, + 816.0, + 900.0, + 852.0, + 296.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 916.0, + 1406.0, + 916.0, + 1406.0, + 964.0, + 292.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 954.0, + 1405.0, + 954.0, + 1405.0, + 988.0, + 294.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 981.0, + 847.0, + 981.0, + 847.0, + 1019.0, + 295.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1597.0, + 792.0, + 1597.0, + 792.0, + 1631.0, + 296.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 1597.0, + 989.0, + 1597.0, + 989.0, + 1631.0, + 814.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1597.0, + 1404.0, + 1597.0, + 1404.0, + 1631.0, + 1158.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1630.0, + 592.0, + 1630.0, + 592.0, + 1664.0, + 294.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 1630.0, + 1050.0, + 1630.0, + 1050.0, + 1664.0, + 627.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 1630.0, + 1154.0, + 1630.0, + 1154.0, + 1664.0, + 1118.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 1630.0, + 1404.0, + 1630.0, + 1404.0, + 1664.0, + 1176.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1659.0, + 756.0, + 1659.0, + 756.0, + 1693.0, + 296.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1457.0, + 996.0, + 1457.0, + 996.0, + 1495.0, + 295.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 1457.0, + 1403.0, + 1457.0, + 1403.0, + 1495.0, + 1028.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1489.0, + 403.0, + 1489.0, + 403.0, + 1526.0, + 293.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 1489.0, + 1405.0, + 1489.0, + 1405.0, + 1526.0, + 689.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1521.0, + 1404.0, + 1521.0, + 1404.0, + 1556.0, + 398.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1549.0, + 512.0, + 1549.0, + 512.0, + 1585.0, + 293.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 1549.0, + 582.0, + 1549.0, + 582.0, + 1585.0, + 540.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 1549.0, + 758.0, + 1549.0, + 758.0, + 1585.0, + 674.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1549.0, + 841.0, + 1549.0, + 841.0, + 1585.0, + 792.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1549.0, + 1403.0, + 1549.0, + 1403.0, + 1585.0, + 871.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1887.0, + 1050.0, + 1887.0, + 1050.0, + 1925.0, + 293.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 1887.0, + 1335.0, + 1887.0, + 1335.0, + 1925.0, + 1199.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1349.0, + 1887.0, + 1405.0, + 1887.0, + 1405.0, + 1925.0, + 1349.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1919.0, + 683.0, + 1919.0, + 683.0, + 1956.0, + 294.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1919.0, + 1304.0, + 1919.0, + 1304.0, + 1956.0, + 703.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1919.0, + 1404.0, + 1919.0, + 1404.0, + 1956.0, + 1372.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1950.0, + 719.0, + 1950.0, + 719.0, + 1985.0, + 295.0, + 1985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 228.0, + 946.0, + 228.0, + 946.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 228.0, + 1405.0, + 228.0, + 1405.0, + 266.0, + 1054.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 261.0, + 481.0, + 261.0, + 481.0, + 296.0, + 293.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 261.0, + 870.0, + 261.0, + 870.0, + 296.0, + 496.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 296.0, + 985.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 291.0, + 814.0, + 291.0, + 814.0, + 325.0, + 295.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1352.0, + 1404.0, + 1352.0, + 1404.0, + 1386.0, + 297.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1381.0, + 1405.0, + 1381.0, + 1405.0, + 1420.0, + 296.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1413.0, + 663.0, + 1413.0, + 663.0, + 1446.0, + 295.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1413.0, + 1222.0, + 1413.0, + 1222.0, + 1446.0, + 684.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1806.0, + 900.0, + 1806.0, + 900.0, + 1851.0, + 294.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 1806.0, + 1153.0, + 1806.0, + 1153.0, + 1851.0, + 935.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 1806.0, + 1407.0, + 1806.0, + 1407.0, + 1851.0, + 1174.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1843.0, + 576.0, + 1843.0, + 576.0, + 1879.0, + 294.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1761.0, + 370.0, + 1761.0, + 370.0, + 1803.0, + 294.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 1761.0, + 755.0, + 1761.0, + 755.0, + 1803.0, + 416.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 1761.0, + 888.0, + 1761.0, + 888.0, + 1803.0, + 769.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1761.0, + 916.0, + 1761.0, + 916.0, + 1803.0, + 905.0, + 1803.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 834, + 1406, + 834, + 1406, + 1050, + 297, + 1050 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1754, + 1404, + 1754, + 1404, + 1879, + 297, + 1879 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1942, + 1403, + 1942, + 1403, + 2035, + 299, + 2035 + ], + "score": 0.974 + }, + { + "category_id": 8, + "poly": [ + 444, + 1374, + 1253, + 1374, + 1253, + 1528, + 444, + 1528 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 398, + 198, + 1301, + 198, + 1301, + 497, + 398, + 497 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 302, + 1266, + 1404, + 1266, + 1404, + 1360, + 302, + 1360 + ], + "score": 0.97 + }, + { + "category_id": 8, + "poly": [ + 537, + 1064, + 1159, + 1064, + 1159, + 1207, + 537, + 1207 + ], + "score": 0.967 + }, + { + "category_id": 4, + "poly": [ + 297, + 520, + 1403, + 520, + 1403, + 645, + 297, + 645 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 298, + 1616, + 1404, + 1616, + 1404, + 1680, + 298, + 1680 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 299, + 1540, + 1403, + 1540, + 1403, + 1602, + 299, + 1602 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 298, + 1218, + 1156, + 1218, + 1156, + 1253, + 298, + 1253 + ], + "score": 0.934 + }, + { + "category_id": 0, + "poly": [ + 299, + 1712, + 559, + 1712, + 559, + 1746, + 299, + 1746 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 299, + 1903, + 695, + 1903, + 695, + 1934, + 299, + 1934 + ], + "score": 0.914 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.888 + }, + { + "category_id": 1, + "poly": [ + 297, + 665, + 1404, + 665, + 1404, + 820, + 297, + 820 + ], + "score": 0.879 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1436, + 1400, + 1436, + 1400, + 1465, + 1366, + 1465 + ], + "score": 0.876 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1119, + 1401, + 1119, + 1401, + 1149, + 1365, + 1149 + ], + "score": 0.872 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.738 + }, + { + "category_id": 14, + "poly": [ + 537, + 1062, + 1158, + 1062, + 1158, + 1207, + 537, + 1207 + ], + "score": 0.95, + "latex": "\\begin{array} { r } { F _ { \\theta _ { i } } ( x , y ) = \\operatorname* { m a x } \\left( \\begin{array} { l } { k ( x - x _ { i } ^ { L } ) b ( y - y _ { i } ^ { T } ) b ( y _ { i } ^ { B } - y ) , } \\\\ { k ( x - x _ { i } ^ { R } ) b ( y - y _ { i } ^ { T } ) b ( y _ { i } ^ { B } - y ) , } \\\\ { k ( y - y _ { i } ^ { T } ) b ( x - x _ { i } ^ { L } ) b ( x _ { i } ^ { R } - x ) , } \\\\ { k ( y - y _ { i } ^ { B } ) b ( x - x _ { i } ^ { L } ) b ( x _ { i } ^ { R } - x ) } \\end{array} \\right) , } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 445, + 1373, + 1250, + 1373, + 1250, + 1528, + 445, + 1528 + ], + "score": 0.95, + "latex": "\\begin{array} { r } { F _ { \\theta _ { i } } ( x , y ) = \\operatorname* { m a x } \\left( \\begin{array} { l } { k ( y - \\frac { ( y _ { i } ^ { 2 } - y _ { i } ^ { 1 } ) \\cdot ( x - x _ { i } ^ { 1 } ) } { x _ { i } ^ { 2 } - x _ { i } ^ { 1 } } - y _ { i } ^ { 1 } ) b ( x - x _ { i } ^ { 1 } ) b ( x _ { i } ^ { 2 } - x ) , } \\\\ { k ( y - \\frac { ( y _ { i } ^ { 3 } - y _ { i } ^ { 1 } ) \\cdot ( x - x _ { i } ^ { 1 } ) } { x _ { i } ^ { 3 } - x _ { i } ^ { 1 } } - y _ { i } ^ { 1 } ) b ( x - x _ { i } ^ { 3 } ) b ( x _ { i } ^ { 1 } - x ) , } \\\\ { k ( y - \\frac { ( y _ { i } ^ { 3 } - y _ { i } ^ { 2 } ) \\cdot ( x - x _ { i } ^ { 2 } ) } { x _ { i } ^ { 3 } - x _ { i } ^ { 2 } } - y _ { i } ^ { 2 } ) b ( x - x _ { i } ^ { 3 } ) b ( x _ { i } ^ { 2 } - x ) } \\end{array} \\right) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 949, + 1295, + 1262, + 1295, + 1262, + 1331, + 949, + 1331 + ], + "score": 0.93, + "latex": "\\pmb { \\theta _ { i } } = ( x _ { i } ^ { 1 } , y _ { i } ^ { 1 } , x _ { i } ^ { \\top } , y _ { i } ^ { 2 } , x _ { i } ^ { 3 } , y _ { i } ^ { 3 } )" + }, + { + "category_id": 13, + "poly": [ + 1073, + 863, + 1319, + 863, + 1319, + 898, + 1073, + 898 + ], + "score": 0.93, + "latex": "\\pmb { \\theta } = ( x ^ { L } , y ^ { T } , x ^ { R } , y ^ { B } )" + }, + { + "category_id": 13, + "poly": [ + 372, + 1218, + 666, + 1218, + 666, + 1254, + 372, + 1254 + ], + "score": 0.92, + "latex": "b ( d ) = \\operatorname* { m i n } ( \\operatorname* { m a x } ( 0 , d ) , 1 )" + }, + { + "category_id": 13, + "poly": [ + 769, + 1328, + 834, + 1328, + 834, + 1361, + 769, + 1361 + ], + "score": 0.92, + "latex": "( x , y )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1326, + 393, + 1326, + 393, + 1361, + 298, + 1361 + ], + "score": 0.92, + "latex": "x _ { i } ^ { 2 } \\neq x _ { i } ^ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 1001, + 1017, + 1066, + 1017, + 1066, + 1051, + 1001, + 1051 + ], + "score": 0.92, + "latex": "( x , y )" + }, + { + "category_id": 13, + "poly": [ + 299, + 696, + 478, + 696, + 478, + 730, + 299, + 730 + ], + "score": 0.91, + "latex": "\\operatorname* { m a x } ( 0 , \\bar { 1 } - | d | )" + }, + { + "category_id": 13, + "poly": [ + 299, + 927, + 490, + 927, + 490, + 960, + 299, + 960 + ], + "score": 0.91, + "latex": "( x _ { i } ^ { L } , y _ { i } ^ { T } , x _ { i } ^ { R } , y _ { i } ^ { B } )" + }, + { + "category_id": 13, + "poly": [ + 1304, + 665, + 1402, + 665, + 1402, + 700, + 1304, + 700 + ], + "score": 0.89, + "latex": "\\boldsymbol k ( d ) \\_ =" + }, + { + "category_id": 13, + "poly": [ + 1340, + 1296, + 1406, + 1296, + 1406, + 1331, + 1340, + 1331 + ], + "score": 0.89, + "latex": "x _ { i } ^ { \\bar { 1 } } \\neq" + }, + { + "category_id": 13, + "poly": [ + 1339, + 897, + 1404, + 897, + 1404, + 927, + 1339, + 927 + ], + "score": 0.88, + "latex": "\\theta _ { i } = " + }, + { + "category_id": 13, + "poly": [ + 433, + 1300, + 446, + 1300, + 446, + 1325, + 433, + 1325 + ], + "score": 0.77, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1124, + 899, + 1140, + 899, + 1140, + 923, + 1124, + 923 + ], + "score": 0.72, + "latex": "_ i" + }, + { + "category_id": 13, + "poly": [ + 808, + 1649, + 828, + 1649, + 828, + 1675, + 808, + 1675 + ], + "score": 0.68, + "latex": "\\pmb { I }" + }, + { + "category_id": 13, + "poly": [ + 1177, + 698, + 1198, + 698, + 1198, + 725, + 1177, + 725 + ], + "score": 0.65, + "latex": "\\pmb { I }" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 212.0, + 844.0, + 212.0, + 844.0, + 236.0, + 771.0, + 236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 212.0, + 908.0, + 212.0, + 908.0, + 237.0, + 847.0, + 237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 228.0, + 1141.0, + 228.0, + 1141.0, + 240.0, + 1125.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 227.0, + 1167.0, + 227.0, + 1167.0, + 239.0, + 1152.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 255.0, + 657.0, + 255.0, + 657.0, + 278.0, + 416.0, + 278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 249.0, + 729.0, + 249.0, + 729.0, + 282.0, + 705.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 254.0, + 788.0, + 254.0, + 788.0, + 277.0, + 743.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 249.0, + 957.0, + 249.0, + 957.0, + 280.0, + 934.0, + 280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 285.0, + 728.0, + 285.0, + 728.0, + 308.0, + 707.0, + 308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 284.0, + 754.0, + 284.0, + 754.0, + 296.0, + 747.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 287.0, + 955.0, + 287.0, + 955.0, + 306.0, + 937.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 282.0, + 1180.0, + 282.0, + 1180.0, + 297.0, + 1167.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 281.0, + 1225.0, + 281.0, + 1225.0, + 298.0, + 1206.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 301.0, + 1062.0, + 301.0, + 1062.0, + 317.0, + 1043.0, + 317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 304.0, + 1094.0, + 304.0, + 1094.0, + 314.0, + 1085.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 300.0, + 1193.0, + 300.0, + 1193.0, + 312.0, + 1185.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 299.0, + 1245.0, + 299.0, + 1245.0, + 318.0, + 1226.0, + 318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 311.0, + 730.0, + 311.0, + 730.0, + 387.0, + 703.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 314.0, + 956.0, + 314.0, + 956.0, + 383.0, + 932.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 319.0, + 1259.0, + 319.0, + 1259.0, + 334.0, + 1247.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 337.0, + 551.0, + 337.0, + 551.0, + 359.0, + 514.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 357.0, + 1002.0, + 357.0, + 1002.0, + 371.0, + 990.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 354.0, + 1252.0, + 354.0, + 1252.0, + 362.0, + 1241.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 358.0, + 1291.0, + 358.0, + 1291.0, + 366.0, + 1283.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 377.0, + 1023.0, + 377.0, + 1023.0, + 386.0, + 1014.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 373.0, + 1259.0, + 373.0, + 1259.0, + 386.0, + 1247.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 392.0, + 921.0, + 392.0, + 921.0, + 411.0, + 885.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 395.0, + 1146.0, + 395.0, + 1146.0, + 417.0, + 1074.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 396.0, + 1211.0, + 396.0, + 1211.0, + 418.0, + 1152.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 436.0, + 843.0, + 436.0, + 843.0, + 457.0, + 770.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 438.0, + 908.0, + 438.0, + 908.0, + 459.0, + 852.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 471.0, + 558.0, + 471.0, + 558.0, + 499.0, + 510.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 470.0, + 872.0, + 470.0, + 872.0, + 499.0, + 792.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 470.0, + 1178.0, + 470.0, + 1178.0, + 501.0, + 1110.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 240.5, + 1189.0, + 240.5, + 1189.0, + 260.0, + 1168.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 258.5, + 1202.0, + 258.5, + 1202.0, + 278.5, + 1185.0, + 278.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 261.0, + 1111.0, + 261.0, + 1111.0, + 279.0, + 1076.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 332.5, + 1042.0, + 332.5, + 1042.0, + 351.0, + 993.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 333.0, + 1285.0, + 333.0, + 1285.0, + 357.5, + 1254.0, + 357.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 519.0, + 1405.0, + 519.0, + 1405.0, + 559.0, + 293.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 553.0, + 1404.0, + 553.0, + 1404.0, + 586.0, + 295.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 584.0, + 1403.0, + 584.0, + 1403.0, + 615.0, + 293.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 613.0, + 1360.0, + 613.0, + 1360.0, + 646.0, + 295.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1709.0, + 561.0, + 1709.0, + 561.0, + 1752.0, + 292.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1901.0, + 700.0, + 1901.0, + 700.0, + 1938.0, + 294.0, + 1938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 833.0, + 1404.0, + 833.0, + 1404.0, + 872.0, + 295.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 862.0, + 1072.0, + 862.0, + 1072.0, + 901.0, + 291.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1320.0, + 862.0, + 1407.0, + 862.0, + 1407.0, + 901.0, + 1320.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 893.0, + 1123.0, + 893.0, + 1123.0, + 932.0, + 291.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 893.0, + 1338.0, + 893.0, + 1338.0, + 932.0, + 1141.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 916.0, + 298.0, + 916.0, + 298.0, + 968.0, + 289.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 916.0, + 1411.0, + 916.0, + 1411.0, + 968.0, + 491.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 956.0, + 1407.0, + 956.0, + 1407.0, + 990.0, + 295.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 988.0, + 1404.0, + 988.0, + 1404.0, + 1022.0, + 295.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1017.0, + 1000.0, + 1017.0, + 1000.0, + 1052.0, + 295.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 1017.0, + 1315.0, + 1017.0, + 1315.0, + 1052.0, + 1067.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1753.0, + 1405.0, + 1753.0, + 1405.0, + 1790.0, + 294.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1786.0, + 1405.0, + 1786.0, + 1405.0, + 1819.0, + 295.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1818.0, + 1404.0, + 1818.0, + 1404.0, + 1850.0, + 296.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1847.0, + 585.0, + 1847.0, + 585.0, + 1881.0, + 295.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1939.0, + 1405.0, + 1939.0, + 1405.0, + 1979.0, + 294.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1973.0, + 1403.0, + 1973.0, + 1403.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1262.0, + 1403.0, + 1262.0, + 1403.0, + 1302.0, + 296.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1290.0, + 432.0, + 1290.0, + 432.0, + 1336.0, + 293.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1290.0, + 948.0, + 1290.0, + 948.0, + 1336.0, + 447.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 1290.0, + 1339.0, + 1290.0, + 1339.0, + 1336.0, + 1263.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1326.0, + 768.0, + 1326.0, + 768.0, + 1364.0, + 394.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 1326.0, + 1323.0, + 1326.0, + 1323.0, + 1364.0, + 835.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1616.0, + 1404.0, + 1616.0, + 1404.0, + 1652.0, + 294.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1649.0, + 807.0, + 1649.0, + 807.0, + 1681.0, + 296.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 1649.0, + 1229.0, + 1649.0, + 1229.0, + 1681.0, + 829.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1537.0, + 1405.0, + 1537.0, + 1405.0, + 1574.0, + 295.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1569.0, + 1252.0, + 1569.0, + 1252.0, + 1605.0, + 293.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1217.0, + 371.0, + 1217.0, + 371.0, + 1257.0, + 297.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 1217.0, + 1156.0, + 1217.0, + 1156.0, + 1257.0, + 667.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 665.0, + 1303.0, + 665.0, + 1303.0, + 703.0, + 296.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 696.0, + 298.0, + 696.0, + 298.0, + 729.0, + 295.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 696.0, + 1176.0, + 696.0, + 1176.0, + 729.0, + 479.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 696.0, + 1404.0, + 696.0, + 1404.0, + 729.0, + 1199.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 724.0, + 1404.0, + 724.0, + 1404.0, + 764.0, + 294.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 754.0, + 1404.0, + 754.0, + 1404.0, + 794.0, + 294.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 788.0, + 893.0, + 788.0, + 893.0, + 825.0, + 295.0, + 825.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1440, + 1404, + 1440, + 1404, + 1925, + 298, + 1925 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 758, + 1405, + 758, + 1405, + 1185, + 298, + 1185 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1239, + 1403, + 1239, + 1403, + 1423, + 298, + 1423 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 301, + 1942, + 1402, + 1942, + 1402, + 2034, + 301, + 2034 + ], + "score": 0.969 + }, + { + "category_id": 5, + "poly": [ + 866, + 575, + 1380, + 575, + 1380, + 734, + 866, + 734 + ], + "score": 0.964, + "html": "
MethodsOverlap(%)Alignment(%)
Relation1.526.4
Wireframe1.173.4
Real data0.050.5
" + }, + { + "category_id": 5, + "poly": [ + 412, + 576, + 737, + 576, + 737, + 734, + 412, + 734 + ], + "score": 0.953, + "html": "
MethodsScore± std
Relation6.53 ± .09
Wireframe7.36 ± .07
Real data9.81 ± .08
" + }, + { + "category_id": 3, + "poly": [ + 318, + 194, + 1357, + 194, + 1357, + 465, + 318, + 465 + ], + "score": 0.932 + }, + { + "category_id": 4, + "poly": [ + 591, + 477, + 1104, + 477, + 1104, + 511, + 591, + 511 + ], + "score": 0.923 + }, + { + "category_id": 6, + "poly": [ + 868, + 532, + 1374, + 532, + 1374, + 563, + 868, + 563 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 301, + 1201, + 772, + 1201, + 772, + 1229, + 301, + 1229 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.89 + }, + { + "category_id": 6, + "poly": [ + 321, + 531, + 826, + 531, + 826, + 563, + 321, + 563 + ], + "score": 0.867 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.779 + }, + { + "category_id": 13, + "poly": [ + 1165, + 759, + 1230, + 759, + 1230, + 787, + 1165, + 787 + ], + "score": 0.9, + "latex": "8 \\times 8" + }, + { + "category_id": 13, + "poly": [ + 1134, + 940, + 1224, + 940, + 1224, + 970, + 1134, + 970 + ], + "score": 0.89, + "latex": "9 8 . 9 1 \\%" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 197.0, + 487.0, + 197.0, + 487.0, + 227.0, + 353.0, + 227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 191.0, + 726.0, + 191.0, + 726.0, + 231.0, + 536.0, + 231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 194.0, + 899.0, + 194.0, + 899.0, + 227.0, + 781.0, + 227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 197.0, + 1112.0, + 197.0, + 1112.0, + 230.0, + 978.0, + 230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 197.0, + 1325.0, + 197.0, + 1325.0, + 227.0, + 1178.0, + 227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 222.0, + 524.0, + 222.0, + 524.0, + 434.0, + 323.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 224.0, + 815.0, + 224.0, + 815.0, + 437.0, + 528.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 224.0, + 937.0, + 224.0, + 937.0, + 318.0, + 835.0, + 318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 231.0, + 1021.0, + 231.0, + 1021.0, + 255.0, + 945.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 233.0, + 1043.0, + 233.0, + 1043.0, + 252.0, + 1023.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 233.0, + 1070.0, + 233.0, + 1070.0, + 252.0, + 1046.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 233.0, + 1094.0, + 233.0, + 1094.0, + 254.0, + 1075.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 236.0, + 1116.0, + 236.0, + 1116.0, + 250.0, + 1100.0, + 250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 231.0, + 1176.0, + 231.0, + 1176.0, + 256.0, + 1152.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 233.0, + 1198.0, + 233.0, + 1198.0, + 252.0, + 1179.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 232.0, + 1223.0, + 232.0, + 1223.0, + 255.0, + 1205.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 241.0, + 1241.0, + 241.0, + 1241.0, + 247.0, + 1235.0, + 247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 234.0, + 1273.0, + 234.0, + 1273.0, + 252.0, + 1260.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 234.0, + 1300.0, + 234.0, + 1300.0, + 251.0, + 1285.0, + 251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 234.0, + 1324.0, + 234.0, + 1324.0, + 250.0, + 1307.0, + 250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 244.0, + 373.0, + 244.0, + 373.0, + 431.0, + 315.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 244.0, + 528.0, + 244.0, + 528.0, + 434.0, + 466.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 244.0, + 664.0, + 244.0, + 664.0, + 417.0, + 600.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 244.0, + 735.0, + 244.0, + 735.0, + 413.0, + 675.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 246.0, + 845.0, + 246.0, + 845.0, + 435.0, + 753.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 250.0, + 940.0, + 250.0, + 940.0, + 413.0, + 885.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 255.0, + 994.0, + 255.0, + 994.0, + 280.0, + 945.0, + 280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 259.0, + 1018.0, + 259.0, + 1018.0, + 277.0, + 999.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1026.0, + 261.0, + 1040.0, + 261.0, + 1040.0, + 276.0, + 1026.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 260.0, + 1066.0, + 260.0, + 1066.0, + 275.0, + 1052.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 261.0, + 1094.0, + 261.0, + 1094.0, + 275.0, + 1076.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 258.0, + 1122.0, + 258.0, + 1122.0, + 281.0, + 1097.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 255.0, + 1147.0, + 255.0, + 1147.0, + 284.0, + 1123.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 255.0, + 1174.0, + 255.0, + 1174.0, + 277.0, + 1154.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 252.0, + 1199.0, + 252.0, + 1199.0, + 281.0, + 1177.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 252.0, + 1229.0, + 252.0, + 1229.0, + 304.0, + 1201.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 256.0, + 1248.0, + 256.0, + 1248.0, + 283.0, + 1231.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 255.0, + 1276.0, + 255.0, + 1276.0, + 289.0, + 1255.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 258.0, + 1298.0, + 258.0, + 1298.0, + 273.0, + 1285.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 258.0, + 1324.0, + 258.0, + 1324.0, + 277.0, + 1306.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 255.0, + 1349.0, + 255.0, + 1349.0, + 277.0, + 1331.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 278.0, + 938.0, + 278.0, + 938.0, + 307.0, + 916.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 280.0, + 972.0, + 280.0, + 972.0, + 306.0, + 947.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 283.0, + 995.0, + 283.0, + 995.0, + 305.0, + 973.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 288.0, + 1066.0, + 288.0, + 1066.0, + 301.0, + 1051.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 283.0, + 1120.0, + 283.0, + 1120.0, + 305.0, + 1100.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 278.0, + 1146.0, + 278.0, + 1146.0, + 307.0, + 1125.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 278.0, + 1176.0, + 278.0, + 1176.0, + 307.0, + 1154.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 283.0, + 1199.0, + 283.0, + 1199.0, + 303.0, + 1177.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 284.0, + 1247.0, + 284.0, + 1247.0, + 301.0, + 1233.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 282.0, + 1274.0, + 282.0, + 1274.0, + 306.0, + 1258.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 283.0, + 1324.0, + 283.0, + 1324.0, + 305.0, + 1306.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 283.0, + 1350.0, + 283.0, + 1350.0, + 303.0, + 1332.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 314.0, + 823.0, + 314.0, + 823.0, + 438.0, + 772.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 299.0, + 887.0, + 299.0, + 887.0, + 413.0, + 837.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 301.0, + 943.0, + 301.0, + 943.0, + 362.0, + 910.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 305.0, + 969.0, + 305.0, + 969.0, + 331.0, + 947.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 306.0, + 1019.0, + 306.0, + 1019.0, + 331.0, + 972.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 311.0, + 1040.0, + 311.0, + 1040.0, + 326.0, + 1024.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 314.0, + 1064.0, + 314.0, + 1064.0, + 324.0, + 1053.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 304.0, + 1146.0, + 304.0, + 1146.0, + 330.0, + 1122.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 306.0, + 1174.0, + 306.0, + 1174.0, + 328.0, + 1155.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 307.0, + 1198.0, + 307.0, + 1198.0, + 326.0, + 1179.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 313.0, + 1222.0, + 313.0, + 1222.0, + 323.0, + 1210.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 312.0, + 1272.0, + 312.0, + 1272.0, + 324.0, + 1263.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1310.0, + 312.0, + 1319.0, + 312.0, + 1319.0, + 323.0, + 1310.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 306.0, + 1350.0, + 306.0, + 1350.0, + 328.0, + 1332.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 326.0, + 893.0, + 326.0, + 893.0, + 414.0, + 862.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 328.0, + 994.0, + 328.0, + 994.0, + 358.0, + 944.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 334.0, + 1018.0, + 334.0, + 1018.0, + 354.0, + 998.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 336.0, + 1040.0, + 336.0, + 1040.0, + 352.0, + 1024.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 340.0, + 1065.0, + 340.0, + 1065.0, + 349.0, + 1052.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 335.0, + 1094.0, + 335.0, + 1094.0, + 353.0, + 1074.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 331.0, + 1146.0, + 331.0, + 1146.0, + 358.0, + 1097.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 331.0, + 1177.0, + 331.0, + 1177.0, + 355.0, + 1152.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 335.0, + 1221.0, + 335.0, + 1221.0, + 352.0, + 1208.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 340.0, + 1243.0, + 340.0, + 1243.0, + 350.0, + 1234.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 341.0, + 1322.0, + 341.0, + 1322.0, + 353.0, + 1311.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 352.0, + 947.0, + 352.0, + 947.0, + 437.0, + 906.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 357.0, + 972.0, + 357.0, + 972.0, + 382.0, + 947.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 357.0, + 1147.0, + 357.0, + 1147.0, + 383.0, + 1125.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 357.0, + 1174.0, + 357.0, + 1174.0, + 378.0, + 1155.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 360.0, + 1197.0, + 360.0, + 1197.0, + 375.0, + 1182.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 361.0, + 993.0, + 361.0, + 993.0, + 381.0, + 972.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 363.0, + 1116.0, + 363.0, + 1116.0, + 374.0, + 1106.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 364.0, + 1014.0, + 364.0, + 1014.0, + 374.0, + 1001.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 364.0, + 1038.0, + 364.0, + 1038.0, + 374.0, + 1030.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 384.0, + 970.0, + 384.0, + 970.0, + 408.0, + 947.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 385.0, + 993.0, + 385.0, + 993.0, + 404.0, + 974.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 383.0, + 1147.0, + 383.0, + 1147.0, + 407.0, + 1123.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 385.0, + 1174.0, + 385.0, + 1174.0, + 404.0, + 1155.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 386.0, + 1198.0, + 386.0, + 1198.0, + 407.0, + 1180.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 393.0, + 1217.0, + 393.0, + 1217.0, + 399.0, + 1211.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 396.0, + 425.0, + 396.0, + 425.0, + 430.0, + 391.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 400.0, + 473.0, + 400.0, + 473.0, + 429.0, + 444.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 490.0, + 397.0, + 525.0, + 397.0, + 525.0, + 430.0, + 490.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 397.0, + 740.0, + 397.0, + 740.0, + 435.0, + 529.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 403.0, + 890.0, + 403.0, + 890.0, + 430.0, + 836.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 403.0, + 916.0, + 403.0, + 916.0, + 430.0, + 892.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 407.0, + 1019.0, + 407.0, + 1019.0, + 430.0, + 945.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 409.0, + 1042.0, + 409.0, + 1042.0, + 427.0, + 1021.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 409.0, + 1070.0, + 409.0, + 1070.0, + 427.0, + 1049.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 408.0, + 1146.0, + 408.0, + 1146.0, + 430.0, + 1074.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 407.0, + 1174.0, + 407.0, + 1174.0, + 429.0, + 1154.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 408.0, + 1202.0, + 408.0, + 1202.0, + 430.0, + 1179.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 415.0, + 1272.0, + 415.0, + 1272.0, + 425.0, + 1260.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1338.0, + 415.0, + 1346.0, + 415.0, + 1346.0, + 425.0, + 1338.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 430.0, + 841.0, + 430.0, + 841.0, + 466.0, + 403.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 430.0, + 1354.0, + 430.0, + 1354.0, + 466.0, + 949.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.25, + 227.5, + 754.25, + 227.5, + 754.25, + 258.0, + 741.25, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 229.5, + 842.0, + 229.5, + 842.0, + 254.5, + 809.0, + 254.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 235.0, + 1150.0, + 235.0, + 1150.0, + 254.0, + 1120.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.25, + 285.0, + 1303.25, + 285.0, + 1303.25, + 300.0, + 1280.25, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.75, + 307.5, + 1124.75, + 307.5, + 1124.75, + 327.5, + 1094.75, + 327.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.75, + 331.5, + 1204.75, + 331.5, + 1204.75, + 353.0, + 1172.75, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.25, + 332.0, + 1346.25, + 332.0, + 1346.25, + 352.5, + 1335.25, + 352.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1334.0, + 361.0, + 1351.0, + 361.0, + 1351.0, + 377.0, + 1334.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1337.0, + 385.5, + 1344.0, + 385.5, + 1344.0, + 404.5, + 1337.0, + 404.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 414.0, + 1225.0, + 414.0, + 1225.0, + 427.0, + 1205.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 471.0, + 1108.0, + 471.0, + 1108.0, + 518.0, + 590.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 526.0, + 1378.0, + 526.0, + 1378.0, + 568.0, + 871.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1199.0, + 775.0, + 1199.0, + 775.0, + 1234.0, + 295.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 526.0, + 831.0, + 526.0, + 831.0, + 568.0, + 315.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1434.0, + 1405.0, + 1434.0, + 1405.0, + 1477.0, + 292.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1467.0, + 1405.0, + 1467.0, + 1405.0, + 1506.0, + 293.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1500.0, + 1405.0, + 1500.0, + 1405.0, + 1534.0, + 294.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1530.0, + 1402.0, + 1530.0, + 1402.0, + 1564.0, + 294.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1559.0, + 1406.0, + 1559.0, + 1406.0, + 1595.0, + 292.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1591.0, + 1405.0, + 1591.0, + 1405.0, + 1625.0, + 294.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1622.0, + 1406.0, + 1622.0, + 1406.0, + 1654.0, + 293.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1653.0, + 1402.0, + 1653.0, + 1402.0, + 1684.0, + 294.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1681.0, + 1408.0, + 1681.0, + 1408.0, + 1715.0, + 294.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1713.0, + 1406.0, + 1713.0, + 1406.0, + 1747.0, + 294.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1742.0, + 1405.0, + 1742.0, + 1405.0, + 1780.0, + 293.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1774.0, + 1405.0, + 1774.0, + 1405.0, + 1808.0, + 294.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1804.0, + 1404.0, + 1804.0, + 1404.0, + 1835.0, + 296.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1835.0, + 1405.0, + 1835.0, + 1405.0, + 1869.0, + 294.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1864.0, + 1405.0, + 1864.0, + 1405.0, + 1902.0, + 293.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1893.0, + 487.0, + 1893.0, + 487.0, + 1931.0, + 294.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 757.0, + 1164.0, + 757.0, + 1164.0, + 793.0, + 295.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 757.0, + 1406.0, + 757.0, + 1406.0, + 793.0, + 1231.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 789.0, + 1403.0, + 789.0, + 1403.0, + 821.0, + 296.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 818.0, + 1405.0, + 818.0, + 1405.0, + 854.0, + 295.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 849.0, + 1405.0, + 849.0, + 1405.0, + 884.0, + 295.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 878.0, + 1405.0, + 878.0, + 1405.0, + 916.0, + 293.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 912.0, + 1403.0, + 912.0, + 1403.0, + 944.0, + 295.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 940.0, + 1133.0, + 940.0, + 1133.0, + 977.0, + 292.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 940.0, + 1406.0, + 940.0, + 1406.0, + 977.0, + 1225.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 970.0, + 1405.0, + 970.0, + 1405.0, + 1004.0, + 292.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1001.0, + 1406.0, + 1001.0, + 1406.0, + 1036.0, + 291.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1031.0, + 1405.0, + 1031.0, + 1405.0, + 1067.0, + 295.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1060.0, + 1403.0, + 1060.0, + 1403.0, + 1097.0, + 292.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1091.0, + 1406.0, + 1091.0, + 1406.0, + 1129.0, + 292.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1125.0, + 1402.0, + 1125.0, + 1402.0, + 1156.0, + 296.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1154.0, + 792.0, + 1154.0, + 792.0, + 1189.0, + 295.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1235.0, + 1407.0, + 1235.0, + 1407.0, + 1278.0, + 293.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1268.0, + 1405.0, + 1268.0, + 1405.0, + 1307.0, + 293.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1299.0, + 1405.0, + 1299.0, + 1405.0, + 1335.0, + 294.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1328.0, + 1405.0, + 1328.0, + 1405.0, + 1365.0, + 293.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1358.0, + 1405.0, + 1358.0, + 1405.0, + 1397.0, + 293.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1392.0, + 786.0, + 1392.0, + 786.0, + 1426.0, + 293.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1978.0, + 295.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1975.0, + 1404.0, + 1975.0, + 1404.0, + 2005.0, + 297.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2000.0, + 1406.0, + 2000.0, + 1406.0, + 2042.0, + 294.0, + 2042.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1516, + 1404, + 1516, + 1404, + 2035, + 297, + 2035 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 297, + 746, + 1404, + 746, + 1404, + 1203, + 297, + 1203 + ], + "score": 0.982 + }, + { + "category_id": 5, + "poly": [ + 798, + 1275, + 1397, + 1275, + 1397, + 1436, + 798, + 1436 + ], + "score": 0.974, + "html": "
MethodsExcellent (%)Fair (%)Poor (%)
DCGAN6.738.854.5
Relation17.250.332.5
Wireframe37.348.014.7
" + }, + { + "category_id": 3, + "poly": [ + 303, + 203, + 839, + 203, + 839, + 684, + 303, + 684 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 897, + 184, + 1346, + 184, + 1346, + 684, + 897, + 684 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 343, + 1217, + 746, + 1217, + 746, + 1402, + 343, + 1402 + ], + "score": 0.955 + }, + { + "category_id": 4, + "poly": [ + 886, + 699, + 1365, + 699, + 1365, + 729, + 886, + 729 + ], + "score": 0.918 + }, + { + "category_id": 6, + "poly": [ + 797, + 1232, + 1398, + 1232, + 1398, + 1264, + 797, + 1264 + ], + "score": 0.918 + }, + { + "category_id": 4, + "poly": [ + 350, + 699, + 796, + 699, + 796, + 729, + 350, + 729 + ], + "score": 0.917 + }, + { + "category_id": 4, + "poly": [ + 321, + 1413, + 769, + 1413, + 769, + 1444, + 321, + 1444 + ], + "score": 0.912 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 814, + 76, + 814, + 104, + 298, + 104 + ], + "score": 0.889 + }, + { + "category_id": 0, + "poly": [ + 297, + 1470, + 844, + 1470, + 844, + 1500, + 297, + 1500 + ], + "score": 0.87 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2111, + 842, + 2111 + ], + "score": 0.654 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2111, + 842, + 2111 + ], + "score": 0.234 + }, + { + "category_id": 15, + "poly": [ + 320.0, + 253.0, + 365.0, + 253.0, + 365.0, + 270.0, + 320.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 366.0, + 383.0, + 366.0, + 383.0, + 382.0, + 300.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 467.0, + 373.0, + 467.0, + 373.0, + 490.0, + 308.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 482.0, + 369.0, + 482.0, + 369.0, + 506.0, + 308.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 592.0, + 369.0, + 592.0, + 369.0, + 609.0, + 316.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 660.0, + 456.0, + 660.0, + 456.0, + 684.0, + 381.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 657.0, + 537.0, + 657.0, + 537.0, + 686.0, + 464.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 661.0, + 596.0, + 661.0, + 596.0, + 683.0, + 548.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 660.0, + 669.0, + 660.0, + 669.0, + 682.0, + 628.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 658.0, + 762.0, + 658.0, + 762.0, + 685.0, + 693.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 660.0, + 815.0, + 660.0, + 815.0, + 682.0, + 784.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 243.0, + 961.0, + 243.0, + 961.0, + 259.0, + 916.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 371.0, + 983.0, + 371.0, + 983.0, + 393.0, + 893.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 402.0, + 1062.0, + 402.0, + 1062.0, + 418.0, + 1050.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 404.0, + 1274.0, + 404.0, + 1274.0, + 411.0, + 1265.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 428.0, + 1182.0, + 428.0, + 1182.0, + 439.0, + 1174.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 501.0, + 970.0, + 501.0, + 970.0, + 523.0, + 903.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 515.0, + 968.0, + 515.0, + 968.0, + 542.0, + 901.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 538.0, + 1065.0, + 538.0, + 1065.0, + 560.0, + 1050.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 596.0, + 1194.0, + 596.0, + 1194.0, + 604.0, + 1184.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 617.0, + 968.0, + 617.0, + 968.0, + 636.0, + 910.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 612.0, + 1241.0, + 612.0, + 1241.0, + 627.0, + 1220.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 628.0, + 1278.0, + 628.0, + 1278.0, + 650.0, + 1267.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 665.0, + 1029.0, + 665.0, + 1029.0, + 682.0, + 1000.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 665.0, + 1089.0, + 665.0, + 1089.0, + 683.0, + 1059.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 661.0, + 1152.0, + 661.0, + 1152.0, + 685.0, + 1117.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 665.0, + 1208.0, + 665.0, + 1208.0, + 684.0, + 1179.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 665.0, + 1269.0, + 665.0, + 1269.0, + 684.0, + 1240.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1289.0, + 664.0, + 1342.0, + 664.0, + 1342.0, + 684.0, + 1289.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 523.5, + 1160.0, + 523.5, + 1160.0, + 534.5, + 1152.0, + 534.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 1218.0, + 485.0, + 1218.0, + 485.0, + 1234.0, + 401.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 1215.0, + 702.0, + 1215.0, + 702.0, + 1237.0, + 581.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 1255.0, + 544.0, + 1255.0, + 544.0, + 1266.0, + 526.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 1263.0, + 545.0, + 1263.0, + 545.0, + 1277.0, + 525.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1271.0, + 545.0, + 1271.0, + 545.0, + 1298.0, + 523.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 1284.0, + 547.0, + 1284.0, + 547.0, + 1339.0, + 518.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1322.0, + 742.0, + 1322.0, + 742.0, + 1331.0, + 734.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 1373.0, + 666.0, + 1373.0, + 666.0, + 1391.5, + 546.0, + 1391.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1362.5, + 748.0, + 1362.5, + 748.0, + 1384.5, + 654.0, + 1384.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 693.0, + 1369.0, + 693.0, + 1369.0, + 736.0, + 878.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1227.0, + 1401.0, + 1227.0, + 1401.0, + 1270.0, + 792.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 693.0, + 801.0, + 693.0, + 801.0, + 736.0, + 346.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1409.0, + 775.0, + 1409.0, + 775.0, + 1448.0, + 319.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1469.0, + 851.0, + 1469.0, + 851.0, + 1503.0, + 295.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1515.0, + 1402.0, + 1515.0, + 1402.0, + 1551.0, + 296.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1545.0, + 1405.0, + 1545.0, + 1405.0, + 1583.0, + 292.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1576.0, + 1406.0, + 1576.0, + 1406.0, + 1613.0, + 294.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1604.0, + 1405.0, + 1604.0, + 1405.0, + 1645.0, + 291.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1636.0, + 1404.0, + 1636.0, + 1404.0, + 1674.0, + 294.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1668.0, + 1404.0, + 1668.0, + 1404.0, + 1705.0, + 294.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1696.0, + 1405.0, + 1696.0, + 1405.0, + 1738.0, + 291.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1726.0, + 1404.0, + 1726.0, + 1404.0, + 1765.0, + 292.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1759.0, + 1406.0, + 1759.0, + 1406.0, + 1799.0, + 292.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1787.0, + 1405.0, + 1787.0, + 1405.0, + 1828.0, + 291.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1818.0, + 1405.0, + 1818.0, + 1405.0, + 1857.0, + 291.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1850.0, + 1406.0, + 1850.0, + 1406.0, + 1886.0, + 295.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1917.0, + 292.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1404.0, + 1910.0, + 1404.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1944.0, + 1405.0, + 1944.0, + 1405.0, + 1977.0, + 295.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1402.0, + 1973.0, + 1402.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2001.0, + 1406.0, + 2001.0, + 1406.0, + 2040.0, + 292.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 747.0, + 1404.0, + 747.0, + 1404.0, + 778.0, + 296.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 777.0, + 1406.0, + 777.0, + 1406.0, + 812.0, + 295.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 807.0, + 1405.0, + 807.0, + 1405.0, + 841.0, + 294.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 836.0, + 1405.0, + 836.0, + 1405.0, + 875.0, + 294.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 866.0, + 1405.0, + 866.0, + 1405.0, + 902.0, + 291.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 899.0, + 1405.0, + 899.0, + 1405.0, + 933.0, + 295.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 929.0, + 1404.0, + 929.0, + 1404.0, + 962.0, + 292.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 961.0, + 1402.0, + 961.0, + 1402.0, + 992.0, + 294.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 989.0, + 1404.0, + 989.0, + 1404.0, + 1024.0, + 294.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1019.0, + 1405.0, + 1019.0, + 1405.0, + 1054.0, + 295.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1050.0, + 1404.0, + 1050.0, + 1404.0, + 1085.0, + 295.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1080.0, + 1404.0, + 1080.0, + 1404.0, + 1115.0, + 295.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1111.0, + 1405.0, + 1111.0, + 1405.0, + 1146.0, + 294.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1141.0, + 1404.0, + 1141.0, + 1404.0, + 1176.0, + 295.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1172.0, + 1265.0, + 1172.0, + 1265.0, + 1207.0, + 294.0, + 1207.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1161, + 1405, + 1161, + 1405, + 1799, + 297, + 1799 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1851, + 1403, + 1851, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 901, + 1404, + 901, + 1404, + 1115, + 297, + 1115 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 297, + 197, + 785, + 197, + 785, + 405, + 297, + 405 + ], + "score": 0.964 + }, + { + "category_id": 3, + "poly": [ + 803, + 175, + 1342, + 175, + 1342, + 407, + 803, + 407 + ], + "score": 0.963 + }, + { + "category_id": 3, + "poly": [ + 363, + 481, + 1327, + 481, + 1327, + 849, + 363, + 849 + ], + "score": 0.961 + }, + { + "category_id": 4, + "poly": [ + 349, + 860, + 1336, + 860, + 1336, + 893, + 349, + 893 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 300, + 1129, + 699, + 1129, + 699, + 1157, + 300, + 1157 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 299, + 1814, + 545, + 1814, + 545, + 1846, + 299, + 1846 + ], + "score": 0.906 + }, + { + "category_id": 4, + "poly": [ + 319, + 422, + 777, + 422, + 777, + 453, + 319, + 453 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.89 + }, + { + "category_id": 4, + "poly": [ + 829, + 422, + 1323, + 422, + 1323, + 452, + 829, + 452 + ], + "score": 0.8 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.79 + }, + { + "category_id": 4, + "poly": [ + 827, + 422, + 1328, + 422, + 1328, + 481, + 827, + 481 + ], + "score": 0.307 + }, + { + "category_id": 4, + "poly": [ + 693, + 459, + 920, + 459, + 920, + 479, + 693, + 479 + ], + "score": 0.157 + }, + { + "category_id": 15, + "poly": [ + 303.0, + 225.0, + 366.0, + 225.0, + 366.0, + 244.0, + 303.0, + 244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 202.0, + 574.0, + 202.0, + 574.0, + 266.0, + 386.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 216.0, + 761.0, + 216.0, + 761.0, + 255.0, + 651.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 293.0, + 376.0, + 293.0, + 376.0, + 313.0, + 296.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 271.0, + 622.0, + 271.0, + 622.0, + 335.0, + 418.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 293.0, + 690.0, + 293.0, + 690.0, + 313.0, + 654.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 360.0, + 377.0, + 360.0, + 377.0, + 382.0, + 296.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 358.0, + 495.0, + 358.0, + 495.0, + 381.0, + 418.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 364.0, + 552.0, + 364.0, + 552.0, + 374.0, + 531.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 173.0, + 1149.0, + 173.0, + 1149.0, + 199.0, + 990.0, + 199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 175.0, + 1334.0, + 175.0, + 1334.0, + 196.0, + 1282.0, + 196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 190.0, + 1275.0, + 190.0, + 1275.0, + 269.0, + 966.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 214.0, + 971.0, + 214.0, + 971.0, + 221.0, + 959.0, + 221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 214.0, + 1102.0, + 214.0, + 1102.0, + 221.0, + 1094.0, + 221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 219.0, + 1039.0, + 219.0, + 1039.0, + 227.0, + 1031.0, + 227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 221.0, + 1315.0, + 221.0, + 1315.0, + 253.0, + 1285.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 278.0, + 1268.0, + 278.0, + 1268.0, + 331.0, + 991.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 351.0, + 1051.0, + 351.0, + 1051.0, + 358.0, + 1039.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 476.0, + 1326.0, + 476.0, + 1326.0, + 565.0, + 409.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 550.0, + 1324.0, + 550.0, + 1324.0, + 636.0, + 378.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 623.0, + 1319.0, + 623.0, + 1319.0, + 708.0, + 365.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 699.0, + 1304.0, + 699.0, + 1304.0, + 779.0, + 373.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 771.0, + 1251.0, + 771.0, + 1251.0, + 847.0, + 382.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 858.0, + 1342.0, + 858.0, + 1342.0, + 897.0, + 356.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1126.0, + 703.0, + 1126.0, + 703.0, + 1162.0, + 295.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1807.0, + 550.0, + 1807.0, + 550.0, + 1857.0, + 291.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 419.0, + 776.0, + 419.0, + 776.0, + 458.0, + 316.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 414.0, + 1328.0, + 414.0, + 1328.0, + 460.0, + 822.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 418.0, + 1330.0, + 418.0, + 1330.0, + 456.0, + 822.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 456.0, + 922.0, + 456.0, + 922.0, + 482.0, + 822.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 454.0, + 1331.0, + 454.0, + 1331.0, + 484.0, + 1232.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 457.0, + 921.0, + 457.0, + 921.0, + 483.0, + 691.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1160.0, + 1405.0, + 1160.0, + 1405.0, + 1196.0, + 295.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1188.0, + 1405.0, + 1188.0, + 1405.0, + 1225.0, + 294.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1220.0, + 1405.0, + 1220.0, + 1405.0, + 1256.0, + 294.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1247.0, + 1407.0, + 1247.0, + 1407.0, + 1288.0, + 291.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1281.0, + 1406.0, + 1281.0, + 1406.0, + 1317.0, + 295.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1310.0, + 1406.0, + 1310.0, + 1406.0, + 1348.0, + 292.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1340.0, + 1406.0, + 1340.0, + 1406.0, + 1378.0, + 292.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1367.0, + 1406.0, + 1367.0, + 1406.0, + 1412.0, + 291.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1403.0, + 1407.0, + 1403.0, + 1407.0, + 1439.0, + 294.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1431.0, + 1407.0, + 1431.0, + 1407.0, + 1472.0, + 291.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1463.0, + 1406.0, + 1463.0, + 1406.0, + 1499.0, + 292.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1494.0, + 1405.0, + 1494.0, + 1405.0, + 1530.0, + 294.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1525.0, + 1406.0, + 1525.0, + 1406.0, + 1561.0, + 294.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1554.0, + 1405.0, + 1554.0, + 1405.0, + 1590.0, + 294.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1585.0, + 1405.0, + 1585.0, + 1405.0, + 1622.0, + 295.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1615.0, + 1406.0, + 1615.0, + 1406.0, + 1652.0, + 291.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1645.0, + 1405.0, + 1645.0, + 1405.0, + 1682.0, + 292.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1677.0, + 1403.0, + 1677.0, + 1403.0, + 1709.0, + 295.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1708.0, + 1405.0, + 1708.0, + 1405.0, + 1744.0, + 295.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1738.0, + 1406.0, + 1738.0, + 1406.0, + 1773.0, + 294.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1771.0, + 490.0, + 1771.0, + 490.0, + 1801.0, + 293.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1848.0, + 1405.0, + 1848.0, + 1405.0, + 1884.0, + 292.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1916.0, + 293.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1907.0, + 1407.0, + 1907.0, + 1407.0, + 1949.0, + 293.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1944.0, + 1403.0, + 1944.0, + 1403.0, + 1975.0, + 294.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1970.0, + 1405.0, + 1970.0, + 1405.0, + 2007.0, + 292.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 1351.0, + 2001.0, + 1351.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 899.0, + 1407.0, + 899.0, + 1407.0, + 936.0, + 293.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 930.0, + 1406.0, + 930.0, + 1406.0, + 966.0, + 292.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 959.0, + 1404.0, + 959.0, + 1404.0, + 997.0, + 292.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 988.0, + 1404.0, + 988.0, + 1404.0, + 1028.0, + 294.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1021.0, + 1406.0, + 1021.0, + 1406.0, + 1060.0, + 294.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1055.0, + 1402.0, + 1055.0, + 1402.0, + 1085.0, + 296.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1085.0, + 1404.0, + 1085.0, + 1404.0, + 1120.0, + 294.0, + 1120.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 105, + 300, + 105 + ], + "score": 0.884 + }, + { + "category_id": 0, + "poly": [ + 300, + 228, + 488, + 228, + 488, + 261, + 300, + 261 + ], + "score": 0.804 + }, + { + "category_id": 1, + "poly": [ + 295, + 1563, + 1402, + 1563, + 1402, + 1629, + 295, + 1629 + ], + "score": 0.797 + }, + { + "category_id": 1, + "poly": [ + 294, + 1392, + 1402, + 1392, + 1402, + 1456, + 294, + 1456 + ], + "score": 0.791 + }, + { + "category_id": 1, + "poly": [ + 296, + 1477, + 1399, + 1477, + 1399, + 1544, + 296, + 1544 + ], + "score": 0.788 + }, + { + "category_id": 1, + "poly": [ + 294, + 1274, + 1401, + 1274, + 1401, + 1368, + 294, + 1368 + ], + "score": 0.784 + }, + { + "category_id": 1, + "poly": [ + 296, + 1158, + 1400, + 1158, + 1400, + 1254, + 296, + 1254 + ], + "score": 0.781 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.779 + }, + { + "category_id": 1, + "poly": [ + 299, + 1768, + 1401, + 1768, + 1401, + 1861, + 299, + 1861 + ], + "score": 0.774 + }, + { + "category_id": 1, + "poly": [ + 299, + 1971, + 1399, + 1971, + 1399, + 2034, + 299, + 2034 + ], + "score": 0.758 + }, + { + "category_id": 1, + "poly": [ + 298, + 862, + 1405, + 862, + 1405, + 988, + 298, + 988 + ], + "score": 0.724 + }, + { + "category_id": 1, + "poly": [ + 298, + 747, + 1400, + 747, + 1400, + 842, + 298, + 842 + ], + "score": 0.723 + }, + { + "category_id": 1, + "poly": [ + 297, + 1011, + 1405, + 1011, + 1405, + 1135, + 297, + 1135 + ], + "score": 0.722 + }, + { + "category_id": 1, + "poly": [ + 295, + 1886, + 1401, + 1886, + 1401, + 1948, + 295, + 1948 + ], + "score": 0.717 + }, + { + "category_id": 1, + "poly": [ + 296, + 1650, + 1405, + 1650, + 1405, + 1745, + 296, + 1745 + ], + "score": 0.708 + }, + { + "category_id": 1, + "poly": [ + 298, + 513, + 1399, + 513, + 1399, + 577, + 298, + 577 + ], + "score": 0.699 + }, + { + "category_id": 1, + "poly": [ + 296, + 598, + 1405, + 598, + 1405, + 725, + 296, + 725 + ], + "score": 0.689 + }, + { + "category_id": 1, + "poly": [ + 297, + 279, + 1405, + 279, + 1405, + 404, + 297, + 404 + ], + "score": 0.655 + }, + { + "category_id": 1, + "poly": [ + 292, + 426, + 1400, + 426, + 1400, + 490, + 292, + 490 + ], + "score": 0.606 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 491.0, + 227.0, + 491.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1564.0, + 1404.0, + 1564.0, + 1404.0, + 1600.0, + 294.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1596.0, + 1138.0, + 1596.0, + 1138.0, + 1630.0, + 323.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1387.0, + 1405.0, + 1387.0, + 1405.0, + 1430.0, + 293.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1422.0, + 596.0, + 1422.0, + 596.0, + 1455.0, + 322.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1475.0, + 1403.0, + 1475.0, + 1403.0, + 1514.0, + 294.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1511.0, + 1319.0, + 1511.0, + 1319.0, + 1544.0, + 322.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1274.0, + 1403.0, + 1274.0, + 1403.0, + 1310.0, + 294.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1306.0, + 1404.0, + 1306.0, + 1404.0, + 1340.0, + 322.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1331.0, + 399.0, + 1331.0, + 399.0, + 1374.0, + 319.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1157.0, + 1402.0, + 1157.0, + 1402.0, + 1194.0, + 294.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1190.0, + 1405.0, + 1190.0, + 1405.0, + 1224.0, + 323.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1221.0, + 823.0, + 1221.0, + 823.0, + 1255.0, + 324.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1765.0, + 1402.0, + 1765.0, + 1402.0, + 1806.0, + 293.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1799.0, + 1406.0, + 1799.0, + 1406.0, + 1836.0, + 322.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1825.0, + 399.0, + 1825.0, + 399.0, + 1866.0, + 320.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1969.0, + 1403.0, + 1969.0, + 1403.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2001.0, + 1129.0, + 2001.0, + 1129.0, + 2036.0, + 323.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 864.0, + 1405.0, + 864.0, + 1405.0, + 900.0, + 295.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 890.0, + 1405.0, + 890.0, + 1405.0, + 934.0, + 319.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 924.0, + 1406.0, + 924.0, + 1406.0, + 962.0, + 321.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 957.0, + 761.0, + 957.0, + 761.0, + 990.0, + 325.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 743.0, + 1405.0, + 743.0, + 1405.0, + 785.0, + 291.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 777.0, + 1405.0, + 777.0, + 1405.0, + 814.0, + 322.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 809.0, + 878.0, + 809.0, + 878.0, + 843.0, + 326.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1011.0, + 1405.0, + 1011.0, + 1405.0, + 1047.0, + 292.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1043.0, + 1406.0, + 1043.0, + 1406.0, + 1075.0, + 321.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1074.0, + 1407.0, + 1074.0, + 1407.0, + 1110.0, + 321.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1103.0, + 809.0, + 1103.0, + 809.0, + 1139.0, + 323.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 1406.0, + 1880.0, + 1406.0, + 1923.0, + 293.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1915.0, + 599.0, + 1915.0, + 599.0, + 1949.0, + 322.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1650.0, + 1403.0, + 1650.0, + 1403.0, + 1688.0, + 294.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1682.0, + 1403.0, + 1682.0, + 1403.0, + 1716.0, + 324.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1713.0, + 809.0, + 1713.0, + 809.0, + 1747.0, + 324.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 510.0, + 1404.0, + 510.0, + 1404.0, + 551.0, + 293.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 544.0, + 1107.0, + 544.0, + 1107.0, + 578.0, + 322.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 598.0, + 1406.0, + 598.0, + 1406.0, + 636.0, + 293.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 630.0, + 1405.0, + 630.0, + 1405.0, + 665.0, + 322.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 662.0, + 1406.0, + 662.0, + 1406.0, + 697.0, + 322.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 692.0, + 1056.0, + 692.0, + 1056.0, + 725.0, + 324.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 276.0, + 1405.0, + 276.0, + 1405.0, + 315.0, + 292.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 309.0, + 1403.0, + 309.0, + 1403.0, + 349.0, + 321.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 341.0, + 1405.0, + 341.0, + 1405.0, + 378.0, + 323.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 373.0, + 867.0, + 373.0, + 867.0, + 406.0, + 322.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 426.0, + 1403.0, + 426.0, + 1403.0, + 463.0, + 294.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 458.0, + 1318.0, + 458.0, + 1318.0, + 491.0, + 323.0, + 491.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 105, + 300, + 105 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.834 + }, + { + "category_id": 1, + "poly": [ + 295, + 1041, + 1404, + 1041, + 1404, + 1105, + 295, + 1105 + ], + "score": 0.795 + }, + { + "category_id": 1, + "poly": [ + 297, + 954, + 1398, + 954, + 1398, + 1020, + 297, + 1020 + ], + "score": 0.761 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1399, + 229, + 1399, + 323, + 297, + 323 + ], + "score": 0.726 + }, + { + "category_id": 1, + "poly": [ + 295, + 868, + 1399, + 868, + 1399, + 933, + 295, + 933 + ], + "score": 0.702 + }, + { + "category_id": 1, + "poly": [ + 301, + 751, + 1403, + 751, + 1403, + 846, + 301, + 846 + ], + "score": 0.68 + }, + { + "category_id": 1, + "poly": [ + 297, + 634, + 1403, + 634, + 1403, + 731, + 297, + 731 + ], + "score": 0.631 + }, + { + "category_id": 1, + "poly": [ + 297, + 346, + 1398, + 346, + 1398, + 410, + 297, + 410 + ], + "score": 0.626 + }, + { + "category_id": 1, + "poly": [ + 296, + 1244, + 1402, + 1244, + 1402, + 1341, + 296, + 1341 + ], + "score": 0.626 + }, + { + "category_id": 1, + "poly": [ + 299, + 1681, + 1402, + 1681, + 1402, + 1746, + 299, + 1746 + ], + "score": 0.625 + }, + { + "category_id": 1, + "poly": [ + 295, + 431, + 1400, + 431, + 1400, + 497, + 295, + 497 + ], + "score": 0.622 + }, + { + "category_id": 1, + "poly": [ + 298, + 1564, + 1404, + 1564, + 1404, + 1661, + 298, + 1661 + ], + "score": 0.605 + }, + { + "category_id": 1, + "poly": [ + 298, + 518, + 1404, + 518, + 1404, + 613, + 298, + 613 + ], + "score": 0.594 + }, + { + "category_id": 1, + "poly": [ + 291, + 1361, + 1401, + 1361, + 1401, + 1427, + 291, + 1427 + ], + "score": 0.586 + }, + { + "category_id": 1, + "poly": [ + 301, + 1447, + 1400, + 1447, + 1400, + 1544, + 301, + 1544 + ], + "score": 0.585 + }, + { + "category_id": 1, + "poly": [ + 297, + 1127, + 1403, + 1127, + 1403, + 1223, + 297, + 1223 + ], + "score": 0.584 + }, + { + "category_id": 1, + "poly": [ + 293, + 1853, + 1401, + 1853, + 1401, + 1920, + 293, + 1920 + ], + "score": 0.555 + }, + { + "category_id": 1, + "poly": [ + 292, + 1767, + 1400, + 1767, + 1400, + 1833, + 292, + 1833 + ], + "score": 0.55 + }, + { + "category_id": 1, + "poly": [ + 299, + 1942, + 1403, + 1942, + 1403, + 2034, + 299, + 2034 + ], + "score": 0.429 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1038.0, + 775.0, + 1038.0, + 775.0, + 1078.0, + 294.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 1042.0, + 1182.0, + 1042.0, + 1182.0, + 1078.0, + 790.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 1042.0, + 1406.0, + 1042.0, + 1406.0, + 1079.0, + 1220.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1073.0, + 610.0, + 1073.0, + 610.0, + 1104.0, + 322.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 952.0, + 1404.0, + 952.0, + 1404.0, + 994.0, + 295.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 987.0, + 1300.0, + 987.0, + 1300.0, + 1021.0, + 322.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 226.0, + 1403.0, + 226.0, + 1403.0, + 265.0, + 292.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 261.0, + 1404.0, + 261.0, + 1404.0, + 295.0, + 322.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 285.0, + 399.0, + 285.0, + 399.0, + 326.0, + 320.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 871.0, + 1403.0, + 871.0, + 1403.0, + 903.0, + 295.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 899.0, + 1206.0, + 899.0, + 1206.0, + 934.0, + 321.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 748.0, + 1405.0, + 748.0, + 1405.0, + 791.0, + 292.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 784.0, + 1407.0, + 784.0, + 1407.0, + 819.0, + 320.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 813.0, + 397.0, + 813.0, + 397.0, + 846.0, + 321.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 631.0, + 1404.0, + 631.0, + 1404.0, + 674.0, + 294.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 668.0, + 1405.0, + 668.0, + 1405.0, + 702.0, + 324.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 697.0, + 943.0, + 697.0, + 943.0, + 732.0, + 324.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 343.0, + 1402.0, + 343.0, + 1402.0, + 383.0, + 293.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 377.0, + 1169.0, + 377.0, + 1169.0, + 413.0, + 323.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1244.0, + 1404.0, + 1244.0, + 1404.0, + 1283.0, + 295.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1276.0, + 1404.0, + 1276.0, + 1404.0, + 1311.0, + 324.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1306.0, + 1059.0, + 1306.0, + 1059.0, + 1344.0, + 323.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1677.0, + 1403.0, + 1677.0, + 1403.0, + 1720.0, + 294.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1715.0, + 774.0, + 1715.0, + 774.0, + 1744.0, + 324.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 430.0, + 1405.0, + 430.0, + 1405.0, + 468.0, + 294.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 464.0, + 1297.0, + 464.0, + 1297.0, + 498.0, + 323.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1561.0, + 1405.0, + 1561.0, + 1405.0, + 1604.0, + 292.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1596.0, + 1405.0, + 1596.0, + 1405.0, + 1631.0, + 323.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1627.0, + 1129.0, + 1627.0, + 1129.0, + 1662.0, + 323.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 520.0, + 1402.0, + 520.0, + 1402.0, + 555.0, + 294.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 551.0, + 1405.0, + 551.0, + 1405.0, + 585.0, + 323.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 580.0, + 881.0, + 580.0, + 881.0, + 614.0, + 325.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1362.0, + 1403.0, + 1362.0, + 1403.0, + 1398.0, + 295.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1391.0, + 1199.0, + 1391.0, + 1199.0, + 1428.0, + 320.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1445.0, + 1404.0, + 1445.0, + 1404.0, + 1486.0, + 295.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1479.0, + 1404.0, + 1479.0, + 1404.0, + 1514.0, + 322.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1510.0, + 1039.0, + 1510.0, + 1039.0, + 1543.0, + 322.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1128.0, + 1403.0, + 1128.0, + 1403.0, + 1163.0, + 296.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1160.0, + 1405.0, + 1160.0, + 1405.0, + 1195.0, + 324.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1191.0, + 575.0, + 1191.0, + 575.0, + 1223.0, + 321.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1854.0, + 1404.0, + 1854.0, + 1404.0, + 1890.0, + 296.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1887.0, + 1148.0, + 1887.0, + 1148.0, + 1922.0, + 322.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1766.0, + 1402.0, + 1766.0, + 1402.0, + 1805.0, + 295.0, + 1805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1799.0, + 1172.0, + 1799.0, + 1172.0, + 1835.0, + 323.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1940.0, + 1405.0, + 1940.0, + 1405.0, + 1977.0, + 295.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1969.0, + 1404.0, + 1969.0, + 1404.0, + 2011.0, + 319.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1998.0, + 437.0, + 1998.0, + 437.0, + 2038.0, + 320.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 105, + 300, + 105 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.835 + }, + { + "category_id": 1, + "poly": [ + 292, + 227, + 1408, + 227, + 1408, + 808, + 292, + 808 + ], + "score": 0.574 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 261.0, + 1406.0, + 261.0, + 1406.0, + 297.0, + 323.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 292.0, + 969.0, + 292.0, + 969.0, + 327.0, + 323.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 345.0, + 1405.0, + 345.0, + 1405.0, + 380.0, + 294.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 374.0, + 1407.0, + 374.0, + 1407.0, + 412.0, + 321.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 405.0, + 944.0, + 405.0, + 944.0, + 440.0, + 322.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 456.0, + 1406.0, + 456.0, + 1406.0, + 494.0, + 293.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 490.0, + 1404.0, + 490.0, + 1404.0, + 521.0, + 323.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 519.0, + 1405.0, + 519.0, + 1405.0, + 554.0, + 322.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 549.0, + 464.0, + 549.0, + 464.0, + 581.0, + 323.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 599.0, + 1405.0, + 599.0, + 1405.0, + 638.0, + 294.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 630.0, + 1405.0, + 630.0, + 1405.0, + 666.0, + 322.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 663.0, + 464.0, + 663.0, + 464.0, + 695.0, + 324.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 714.0, + 1404.0, + 714.0, + 1404.0, + 749.0, + 294.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 745.0, + 1406.0, + 745.0, + 1406.0, + 782.0, + 322.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 775.0, + 574.0, + 775.0, + 574.0, + 810.0, + 323.0, + 810.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 460, + 447, + 1233, + 447, + 1233, + 1057, + 460, + 1057 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 405, + 1211, + 1292, + 1211, + 1292, + 1999, + 405, + 1999 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 354, + 1403, + 354, + 1403, + 418, + 298, + 418 + ], + "score": 0.95 + }, + { + "category_id": 4, + "poly": [ + 631, + 2026, + 1066, + 2026, + 1066, + 2058, + 631, + 2058 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 298, + 226, + 501, + 226, + 501, + 261, + 298, + 261 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.904 + }, + { + "category_id": 0, + "poly": [ + 299, + 298, + 660, + 298, + 660, + 328, + 299, + 328 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.771 + }, + { + "category_id": 1, + "poly": [ + 299, + 1080, + 1393, + 1080, + 1393, + 1115, + 299, + 1115 + ], + "score": 0.726 + }, + { + "category_id": 1, + "poly": [ + 297, + 1140, + 1029, + 1140, + 1029, + 1171, + 297, + 1171 + ], + "score": 0.329 + }, + { + "category_id": 4, + "poly": [ + 299, + 1080, + 1393, + 1080, + 1393, + 1115, + 299, + 1115 + ], + "score": 0.305 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2114, + 836, + 2114 + ], + "score": 0.204 + }, + { + "category_id": 0, + "poly": [ + 297, + 1140, + 1029, + 1140, + 1029, + 1171, + 297, + 1171 + ], + "score": 0.139 + }, + { + "category_id": 15, + "poly": [ + 462.0, + 560.0, + 553.0, + 560.0, + 553.0, + 590.0, + 462.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 587.0, + 549.0, + 587.0, + 549.0, + 617.0, + 470.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 752.0, + 1181.0, + 752.0, + 1181.0, + 773.0, + 1069.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 809.0, + 712.0, + 809.0, + 712.0, + 820.0, + 645.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 854.0, + 565.0, + 854.0, + 565.0, + 885.0, + 461.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 873.0, + 1080.0, + 873.0, + 1080.0, + 931.0, + 1049.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 920.0, + 1074.0, + 920.0, + 1074.0, + 937.0, + 1062.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 940.0, + 1077.0, + 940.0, + 1077.0, + 958.0, + 1052.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 1022.0, + 675.0, + 1022.0, + 675.0, + 1057.0, + 589.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 1018.0, + 789.0, + 1018.0, + 789.0, + 1057.0, + 693.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1021.0, + 889.0, + 1021.0, + 889.0, + 1058.0, + 813.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 1021.0, + 991.0, + 1021.0, + 991.0, + 1052.0, + 927.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 1021.0, + 1115.0, + 1021.0, + 1115.0, + 1056.0, + 1025.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 1021.0, + 1203.0, + 1021.0, + 1203.0, + 1054.0, + 1155.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1290.0, + 506.0, + 1290.0, + 506.0, + 1314.0, + 437.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1477.0, + 539.0, + 1477.0, + 539.0, + 1502.0, + 404.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1649.0, + 520.0, + 1649.0, + 520.0, + 1676.0, + 420.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 1671.0, + 512.0, + 1671.0, + 512.0, + 1704.0, + 419.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 1847.0, + 515.0, + 1847.0, + 515.0, + 1873.0, + 428.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 1963.0, + 653.0, + 1963.0, + 653.0, + 1997.0, + 545.0, + 1997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 1961.0, + 772.0, + 1961.0, + 772.0, + 1998.0, + 680.0, + 1998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 1963.0, + 887.0, + 1963.0, + 887.0, + 1998.0, + 815.0, + 1998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 1967.0, + 1006.0, + 1967.0, + 1006.0, + 1991.0, + 947.0, + 1991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 1964.0, + 1144.0, + 1964.0, + 1144.0, + 1995.0, + 1060.0, + 1995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 1966.0, + 1247.0, + 1966.0, + 1247.0, + 1991.0, + 1206.0, + 1991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 2018.0, + 1070.0, + 2018.0, + 1070.0, + 2068.0, + 627.0, + 2068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 220.0, + 507.0, + 220.0, + 507.0, + 268.0, + 292.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 298.0, + 661.0, + 298.0, + 661.0, + 330.0, + 296.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 1075.0, + 1397.0, + 1075.0, + 1397.0, + 1121.0, + 300.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1137.0, + 1031.0, + 1137.0, + 1031.0, + 1174.0, + 295.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 351.0, + 1405.0, + 351.0, + 1405.0, + 390.0, + 293.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 386.0, + 948.0, + 386.0, + 948.0, + 419.0, + 293.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 1075.0, + 1397.0, + 1075.0, + 1397.0, + 1121.0, + 300.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1137.0, + 1031.0, + 1137.0, + 1031.0, + 1174.0, + 295.0, + 1174.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 301, + 287, + 1397, + 287, + 1397, + 1525, + 301, + 1525 + ], + "score": 0.969 + }, + { + "category_id": 4, + "poly": [ + 609, + 1545, + 1091, + 1545, + 1091, + 1579, + 609, + 1579 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.847 + }, + { + "category_id": 0, + "poly": [ + 303, + 232, + 1099, + 232, + 1099, + 261, + 303, + 261 + ], + "score": 0.46 + }, + { + "category_id": 15, + "poly": [ + 349.0, + 440.0, + 459.0, + 440.0, + 459.0, + 482.0, + 349.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 477.0, + 554.0, + 477.0, + 554.0, + 498.0, + 526.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 471.0, + 1378.0, + 471.0, + 1378.0, + 510.0, + 1332.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 764.0, + 508.0, + 764.0, + 508.0, + 804.0, + 299.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1074.0, + 478.0, + 1074.0, + 478.0, + 1125.0, + 321.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1116.0, + 469.0, + 1116.0, + 469.0, + 1166.0, + 320.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 1354.0, + 473.0, + 1354.0, + 473.0, + 1404.0, + 335.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 1476.0, + 615.0, + 1476.0, + 615.0, + 1511.0, + 560.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 1472.0, + 769.0, + 1472.0, + 769.0, + 1516.0, + 701.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1473.0, + 919.0, + 1473.0, + 919.0, + 1516.0, + 846.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 1473.0, + 1206.0, + 1473.0, + 1206.0, + 1514.0, + 1141.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1470.0, + 1377.0, + 1470.0, + 1377.0, + 1515.0, + 1262.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.75, + 663.5, + 1135.75, + 663.5, + 1135.75, + 780.5, + 941.75, + 780.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 1473.5, + 1057.0, + 1473.5, + 1057.0, + 1515.0, + 997.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 1543.0, + 1091.0, + 1543.0, + 1091.0, + 1583.0, + 605.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 228.0, + 1106.0, + 228.0, + 1106.0, + 266.0, + 296.0, + 266.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1114, + 1405, + 1114, + 1405, + 1392, + 297, + 1392 + ], + "score": 0.98 + }, + { + "category_id": 3, + "poly": [ + 293, + 300, + 1395, + 300, + 1395, + 985, + 293, + 985 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 655, + 1013, + 1042, + 1013, + 1042, + 1047, + 655, + 1047 + ], + "score": 0.915 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.844 + }, + { + "category_id": 0, + "poly": [ + 297, + 1059, + 1022, + 1059, + 1022, + 1090, + 297, + 1090 + ], + "score": 0.684 + }, + { + "category_id": 2, + "poly": [ + 312, + 233, + 1107, + 233, + 1107, + 261, + 312, + 261 + ], + "score": 0.446 + }, + { + "category_id": 0, + "poly": [ + 312, + 233, + 1107, + 233, + 1107, + 261, + 312, + 261 + ], + "score": 0.382 + }, + { + "category_id": 1, + "poly": [ + 297, + 1059, + 1022, + 1059, + 1022, + 1090, + 297, + 1090 + ], + "score": 0.244 + }, + { + "category_id": 15, + "poly": [ + 294.0, + 348.0, + 430.0, + 348.0, + 430.0, + 387.0, + 294.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 298.0, + 1090.0, + 298.0, + 1090.0, + 471.0, + 552.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 381.0, + 410.0, + 381.0, + 410.0, + 418.0, + 316.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 508.0, + 405.0, + 508.0, + 405.0, + 543.0, + 344.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 468.0, + 1075.0, + 468.0, + 1075.0, + 623.0, + 726.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 538.0, + 420.0, + 538.0, + 420.0, + 575.0, + 328.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 709.0, + 428.0, + 709.0, + 428.0, + 748.0, + 293.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 678.0, + 1219.0, + 678.0, + 1219.0, + 805.0, + 430.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 739.0, + 407.0, + 739.0, + 407.0, + 776.0, + 313.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 872.0, + 398.0, + 872.0, + 398.0, + 906.0, + 338.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 820.0, + 1181.0, + 820.0, + 1181.0, + 974.0, + 711.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 903.0, + 415.0, + 903.0, + 415.0, + 939.0, + 323.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 904.0, + 562.0, + 904.0, + 562.0, + 923.0, + 547.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 903.0, + 1178.0, + 903.0, + 1178.0, + 916.0, + 1156.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.25, + 509.5, + 723.25, + 509.5, + 723.25, + 580.5, + 597.25, + 580.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 1008.0, + 1045.0, + 1008.0, + 1045.0, + 1053.0, + 652.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1057.0, + 1024.0, + 1057.0, + 1024.0, + 1092.0, + 296.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 231.0, + 1113.0, + 231.0, + 1113.0, + 264.0, + 306.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 231.0, + 1113.0, + 231.0, + 1113.0, + 264.0, + 306.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1117.0, + 1405.0, + 1117.0, + 1405.0, + 1149.0, + 297.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1144.0, + 1406.0, + 1144.0, + 1406.0, + 1185.0, + 292.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1178.0, + 1403.0, + 1178.0, + 1403.0, + 1211.0, + 295.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1204.0, + 1406.0, + 1204.0, + 1406.0, + 1243.0, + 292.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1235.0, + 1407.0, + 1235.0, + 1407.0, + 1273.0, + 291.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1267.0, + 1403.0, + 1267.0, + 1403.0, + 1300.0, + 295.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1299.0, + 1403.0, + 1299.0, + 1403.0, + 1332.0, + 294.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1329.0, + 1403.0, + 1329.0, + 1403.0, + 1366.0, + 294.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1359.0, + 877.0, + 1359.0, + 877.0, + 1395.0, + 295.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1057.0, + 1024.0, + 1057.0, + 1024.0, + 1092.0, + 296.0, + 1092.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 298, + 413, + 1396, + 413, + 1396, + 1251, + 298, + 1251 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 298, + 286, + 1405, + 286, + 1405, + 381, + 298, + 381 + ], + "score": 0.963 + }, + { + "category_id": 4, + "poly": [ + 295, + 1281, + 1405, + 1281, + 1405, + 1344, + 295, + 1344 + ], + "score": 0.934 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.841 + }, + { + "category_id": 0, + "poly": [ + 303, + 232, + 896, + 232, + 896, + 261, + 303, + 261 + ], + "score": 0.792 + }, + { + "category_id": 3, + "poly": [ + 298, + 413, + 1397, + 413, + 1397, + 1252, + 298, + 1252 + ], + "score": 0.187 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 597.0, + 392.0, + 597.0, + 392.0, + 627.0, + 298.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 623.0, + 387.0, + 623.0, + 387.0, + 656.0, + 304.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 831.0, + 645.0, + 831.0, + 645.0, + 847.0, + 607.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 833.0, + 772.0, + 833.0, + 772.0, + 850.0, + 660.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 831.0, + 898.0, + 831.0, + 898.0, + 851.0, + 835.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 832.0, + 1020.0, + 832.0, + 1020.0, + 849.0, + 908.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 832.0, + 1145.0, + 832.0, + 1145.0, + 849.0, + 1097.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 832.0, + 1216.0, + 832.0, + 1216.0, + 849.0, + 1161.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1223.0, + 835.0, + 1279.0, + 835.0, + 1279.0, + 847.0, + 1223.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 835.0, + 1392.0, + 835.0, + 1392.0, + 847.0, + 1312.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 846.0, + 546.0, + 846.0, + 546.0, + 880.0, + 506.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 857.0, + 931.0, + 857.0, + 931.0, + 870.0, + 919.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 854.0, + 1009.0, + 854.0, + 1009.0, + 874.0, + 939.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 860.0, + 1284.0, + 860.0, + 1284.0, + 880.0, + 1179.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 898.0, + 579.0, + 898.0, + 579.0, + 920.0, + 425.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 894.0, + 993.0, + 894.0, + 993.0, + 911.0, + 924.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 891.0, + 1355.0, + 891.0, + 1355.0, + 925.0, + 1178.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 921.0, + 1057.0, + 921.0, + 1057.0, + 934.0, + 926.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 917.0, + 1365.0, + 917.0, + 1365.0, + 934.0, + 1178.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 935.0, + 516.0, + 935.0, + 516.0, + 954.0, + 417.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 949.0, + 806.0, + 949.0, + 806.0, + 983.0, + 736.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 943.0, + 1094.0, + 943.0, + 1094.0, + 959.0, + 923.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 930.0, + 1349.0, + 930.0, + 1349.0, + 970.0, + 1179.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 961.0, + 621.0, + 961.0, + 621.0, + 977.0, + 421.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 966.0, + 1364.0, + 966.0, + 1364.0, + 982.0, + 1179.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 973.0, + 574.0, + 973.0, + 574.0, + 990.0, + 420.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 976.0, + 1370.0, + 976.0, + 1370.0, + 996.0, + 1179.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 996.0, + 614.0, + 996.0, + 614.0, + 1013.0, + 420.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 991.0, + 853.0, + 991.0, + 853.0, + 1028.0, + 705.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 989.0, + 1369.0, + 989.0, + 1369.0, + 1018.0, + 1179.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 1027.0, + 408.0, + 1027.0, + 408.0, + 1059.0, + 302.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1021.0, + 634.0, + 1021.0, + 634.0, + 1073.0, + 420.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 1055.0, + 880.0, + 1055.0, + 880.0, + 1077.0, + 672.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1018.0, + 1133.0, + 1018.0, + 1133.0, + 1068.0, + 921.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1014.0, + 1359.0, + 1014.0, + 1359.0, + 1078.0, + 1179.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1069.0, + 630.0, + 1069.0, + 630.0, + 1086.0, + 420.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 1073.0, + 811.0, + 1073.0, + 811.0, + 1093.0, + 743.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1073.0, + 1099.0, + 1073.0, + 1099.0, + 1089.0, + 921.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1073.0, + 1352.0, + 1073.0, + 1352.0, + 1089.0, + 1179.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1081.0, + 626.0, + 1081.0, + 626.0, + 1133.0, + 420.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1103.0, + 1130.0, + 1103.0, + 1130.0, + 1130.0, + 919.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1087.0, + 1366.0, + 1087.0, + 1366.0, + 1127.0, + 1178.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 1129.0, + 795.0, + 1129.0, + 795.0, + 1151.0, + 760.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 1128.0, + 1072.0, + 1128.0, + 1072.0, + 1140.0, + 923.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1133.0, + 1344.0, + 1133.0, + 1344.0, + 1149.0, + 1183.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 1140.0, + 627.0, + 1140.0, + 627.0, + 1180.0, + 418.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1148.0, + 1127.0, + 1148.0, + 1127.0, + 1186.0, + 922.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 1143.0, + 1367.0, + 1143.0, + 1367.0, + 1188.0, + 1175.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 1181.0, + 548.0, + 1181.0, + 548.0, + 1198.0, + 508.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 1182.0, + 1055.0, + 1182.0, + 1055.0, + 1195.0, + 923.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1180.0, + 1356.0, + 1180.0, + 1356.0, + 1196.0, + 1178.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 1203.0, + 564.0, + 1203.0, + 564.0, + 1219.0, + 493.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 1190.0, + 798.0, + 1190.0, + 798.0, + 1212.0, + 756.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1202.0, + 1050.0, + 1202.0, + 1050.0, + 1218.0, + 921.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 1199.0, + 1243.0, + 1199.0, + 1243.0, + 1217.0, + 1210.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 1199.0, + 1346.0, + 1199.0, + 1346.0, + 1217.0, + 1311.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 1231.0, + 470.0, + 1231.0, + 470.0, + 1246.0, + 454.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 1231.0, + 536.0, + 1231.0, + 536.0, + 1247.0, + 519.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 1231.0, + 601.0, + 1231.0, + 601.0, + 1247.0, + 585.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1231.0, + 718.0, + 1231.0, + 718.0, + 1246.0, + 703.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 1231.0, + 786.0, + 1231.0, + 786.0, + 1246.0, + 770.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 1231.0, + 851.0, + 1231.0, + 851.0, + 1246.0, + 835.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1231.0, + 968.0, + 1231.0, + 968.0, + 1246.0, + 953.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 1231.0, + 1035.0, + 1231.0, + 1035.0, + 1246.0, + 1020.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 1231.0, + 1102.0, + 1231.0, + 1102.0, + 1246.0, + 1085.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 1231.0, + 1219.0, + 1231.0, + 1219.0, + 1246.0, + 1204.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 1231.0, + 1285.0, + 1231.0, + 1285.0, + 1246.0, + 1269.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 1231.0, + 1352.0, + 1231.0, + 1352.0, + 1247.0, + 1335.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 989.5, + 1071.0, + 989.5, + 1071.0, + 1010.0, + 983.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.25, + 1083.0, + 962.25, + 1083.0, + 962.25, + 1099.5, + 921.25, + 1099.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1273.0, + 1406.0, + 1273.0, + 1406.0, + 1323.0, + 291.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1311.0, + 438.0, + 1311.0, + 438.0, + 1344.0, + 293.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 228.0, + 903.0, + 228.0, + 903.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 596.0, + 392.0, + 596.0, + 392.0, + 626.0, + 298.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 623.0, + 386.0, + 623.0, + 386.0, + 657.0, + 302.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 831.0, + 645.0, + 831.0, + 645.0, + 849.0, + 606.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 833.0, + 773.0, + 833.0, + 773.0, + 849.0, + 661.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 833.0, + 895.0, + 833.0, + 895.0, + 849.0, + 838.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 833.0, + 1020.0, + 833.0, + 1020.0, + 849.0, + 910.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 833.0, + 1145.0, + 833.0, + 1145.0, + 849.0, + 1087.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1228.0, + 833.0, + 1285.0, + 833.0, + 1285.0, + 849.0, + 1228.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1320.0, + 836.0, + 1355.0, + 836.0, + 1355.0, + 845.0, + 1320.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1357.0, + 833.0, + 1393.0, + 833.0, + 1393.0, + 849.0, + 1357.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 859.0, + 430.0, + 859.0, + 430.0, + 869.0, + 420.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 845.0, + 548.0, + 845.0, + 548.0, + 881.0, + 507.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 854.0, + 1007.0, + 854.0, + 1007.0, + 875.0, + 939.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 861.0, + 1285.0, + 861.0, + 1285.0, + 881.0, + 1180.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 900.0, + 579.0, + 900.0, + 579.0, + 921.0, + 425.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 895.0, + 991.0, + 895.0, + 991.0, + 912.0, + 924.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 894.0, + 1355.0, + 894.0, + 1355.0, + 923.0, + 1180.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 919.0, + 1059.0, + 919.0, + 1059.0, + 936.0, + 925.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 918.0, + 1366.0, + 918.0, + 1366.0, + 935.0, + 1180.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 933.0, + 517.0, + 933.0, + 517.0, + 956.0, + 416.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 940.0, + 832.0, + 940.0, + 832.0, + 992.0, + 730.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 943.0, + 1092.0, + 943.0, + 1092.0, + 960.0, + 921.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 929.0, + 1255.0, + 929.0, + 1255.0, + 946.0, + 1178.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 954.0, + 1348.0, + 954.0, + 1348.0, + 970.0, + 1181.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 961.0, + 621.0, + 961.0, + 621.0, + 978.0, + 422.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 965.0, + 1363.0, + 965.0, + 1363.0, + 982.0, + 1178.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 973.0, + 574.0, + 973.0, + 574.0, + 989.0, + 420.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 978.0, + 1368.0, + 978.0, + 1368.0, + 995.0, + 1180.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 997.0, + 614.0, + 997.0, + 614.0, + 1014.0, + 420.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 993.0, + 852.0, + 993.0, + 852.0, + 1026.0, + 705.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 992.0, + 1069.0, + 992.0, + 1069.0, + 1009.0, + 986.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 989.0, + 1370.0, + 989.0, + 1370.0, + 1017.0, + 1178.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 1027.0, + 408.0, + 1027.0, + 408.0, + 1061.0, + 303.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 1021.0, + 633.0, + 1021.0, + 633.0, + 1111.0, + 417.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 1057.0, + 879.0, + 1057.0, + 879.0, + 1094.0, + 675.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1020.0, + 1131.0, + 1020.0, + 1131.0, + 1089.0, + 917.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1014.0, + 1367.0, + 1014.0, + 1367.0, + 1113.0, + 1177.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 1116.0, + 491.0, + 1116.0, + 491.0, + 1133.0, + 416.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1104.0, + 1129.0, + 1104.0, + 1129.0, + 1131.0, + 920.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1109.0, + 1277.0, + 1109.0, + 1277.0, + 1126.0, + 1178.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 1137.0, + 628.0, + 1137.0, + 628.0, + 1168.0, + 418.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 1131.0, + 793.0, + 1131.0, + 793.0, + 1149.0, + 763.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1127.0, + 1127.0, + 1127.0, + 1127.0, + 1177.0, + 919.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1132.0, + 1367.0, + 1132.0, + 1367.0, + 1173.0, + 1177.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 1164.0, + 608.0, + 1164.0, + 608.0, + 1181.0, + 419.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1169.0, + 1119.0, + 1169.0, + 1119.0, + 1186.0, + 921.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 1167.0, + 1352.0, + 1167.0, + 1352.0, + 1187.0, + 1176.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1182.0, + 548.0, + 1182.0, + 548.0, + 1199.0, + 505.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1181.0, + 1056.0, + 1181.0, + 1056.0, + 1197.0, + 921.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 1177.0, + 1358.0, + 1177.0, + 1358.0, + 1197.0, + 1175.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 1204.0, + 564.0, + 1204.0, + 564.0, + 1220.0, + 492.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 1190.0, + 799.0, + 1190.0, + 799.0, + 1211.0, + 756.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1201.0, + 1054.0, + 1201.0, + 1054.0, + 1218.0, + 920.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 1199.0, + 1242.0, + 1199.0, + 1242.0, + 1217.0, + 1210.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 1199.0, + 1346.0, + 1199.0, + 1346.0, + 1217.0, + 1311.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 1233.0, + 468.0, + 1233.0, + 468.0, + 1245.0, + 455.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 1231.0, + 535.0, + 1231.0, + 535.0, + 1247.0, + 519.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 1231.0, + 601.0, + 1231.0, + 601.0, + 1247.0, + 585.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 1233.0, + 718.0, + 1233.0, + 718.0, + 1245.0, + 705.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 1232.0, + 785.0, + 1232.0, + 785.0, + 1247.0, + 769.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 1232.0, + 851.0, + 1232.0, + 851.0, + 1247.0, + 835.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1233.0, + 967.0, + 1233.0, + 967.0, + 1245.0, + 955.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 1232.0, + 1036.0, + 1232.0, + 1036.0, + 1247.0, + 1019.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 1232.0, + 1101.0, + 1232.0, + 1101.0, + 1246.0, + 1085.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1233.0, + 1217.0, + 1233.0, + 1217.0, + 1243.0, + 1205.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 1232.0, + 1286.0, + 1232.0, + 1286.0, + 1246.0, + 1270.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 1232.0, + 1351.0, + 1232.0, + 1351.0, + 1247.0, + 1335.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.25, + 831.0, + 1201.25, + 831.0, + 1201.25, + 848.0, + 1162.25, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 855.5, + 935.0, + 855.5, + 935.0, + 871.0, + 913.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.75, + 1082.5, + 961.75, + 1082.5, + 961.75, + 1099.0, + 920.75, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 286.0, + 1406.0, + 286.0, + 1406.0, + 324.0, + 295.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 318.0, + 1403.0, + 318.0, + 1403.0, + 353.0, + 293.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 349.0, + 1133.0, + 349.0, + 1133.0, + 383.0, + 296.0, + 383.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 468, + 329, + 1232, + 329, + 1232, + 1623, + 468, + 1623 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 646, + 1640, + 1051, + 1640, + 1051, + 1676, + 646, + 1676 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.848 + }, + { + "category_id": 4, + "poly": [ + 296, + 228, + 1402, + 228, + 1402, + 323, + 296, + 323 + ], + "score": 0.652 + }, + { + "category_id": 1, + "poly": [ + 296, + 228, + 1402, + 228, + 1402, + 323, + 296, + 323 + ], + "score": 0.392 + }, + { + "category_id": 15, + "poly": [ + 543.0, + 350.0, + 779.0, + 350.0, + 779.0, + 386.0, + 543.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 352.0, + 1120.0, + 352.0, + 1120.0, + 385.0, + 969.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 1569.0, + 589.0, + 1569.0, + 589.0, + 1604.0, + 510.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 1567.0, + 749.0, + 1567.0, + 749.0, + 1600.0, + 651.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 1567.0, + 924.0, + 1567.0, + 924.0, + 1600.0, + 776.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 1570.0, + 1030.0, + 1570.0, + 1030.0, + 1599.0, + 971.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 1570.0, + 1180.0, + 1570.0, + 1180.0, + 1599.0, + 1124.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 1636.0, + 1052.0, + 1636.0, + 1052.0, + 1681.0, + 644.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 871.0, + 2085.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1406.0, + 228.0, + 1406.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 262.0, + 1406.0, + 262.0, + 1406.0, + 296.0, + 295.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 293.0, + 520.0, + 293.0, + 520.0, + 323.0, + 295.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1406.0, + 228.0, + 1406.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 262.0, + 1406.0, + 262.0, + 1406.0, + 296.0, + 295.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 293.0, + 520.0, + 293.0, + 520.0, + 323.0, + 295.0, + 323.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/S1Euwz-Rb/S1Euwz-Rb.md b/parse/train/S1Euwz-Rb/S1Euwz-Rb.md new file mode 100644 index 0000000000000000000000000000000000000000..a7efbabd95191e3699ac52160641a40903ec17fd --- /dev/null +++ b/parse/train/S1Euwz-Rb/S1Euwz-Rb.md @@ -0,0 +1,418 @@ +# COMPOSITIONAL ATTENTION NETWORKS FOR MACHINE REASONING + +Drew A. Hudson +Department of Computer Science Stanford University +dorarad@cs.stanford.edu + +Christopher D. Manning Department of Computer Science Stanford University manning@cs.stanford.edu + +# ABSTRACT + +We present Compositional Attention Networks, a novel fully differentiable neural network architecture, designed to facilitate explicit and expressive reasoning. While many types of neural networks are effective at learning and generalizing from massive quantities of data, this model moves away from monolithic blackbox architectures towards a design that provides a strong prior for iterative reasoning, allowing it to support explainable and structured learning, as well as generalization from a modest amount of data. The model builds on the great success of existing recurrent cells such as LSTMs: It sequences a single recurrent Memory, Attention, and Control (MAC) cell, and by careful design imposes structural constraints on the operation of each cell and the interactions between them, incorporating explicit control and soft attention mechanisms into their interfaces. We demonstrate the model’s strength and robustness on the challenging CLEVR dataset for visual reasoning, achieving a new state-of-the-art $9 8 . 9 \%$ accuracy, halving the error rate of the previous best model. More importantly, we show that the new model is more computationally efficient and data-efficient, requiring an order of magnitude less time and/or data to achieve good results. + +# 1 INTRODUCTION + +This paper considers how best to design neural networks to perform the iterative reasoning necessary for complex problem solving. Putting facts and observations together to arrive at conclusions is a central necessary ability as we work to move neural networks beyond their current great success with sensory perception tasks (LeCun et al., 1998; Krizhevsky et al., 2012) towards displaying Artificial General Intelligence. + +Concretely, we develop a novel model that we apply to the CLEVR dataset (Johnson et al., 2016) for visual question answering (VQA). VQA (Antol et al., 2015; Gupta, 2017) is a challenging multimodal task that requires responding to natural language questions about images. However, Agrawal et al. (2016) show how the first generation of successful models on VQA tasks tend to acquire only superficial comprehension of both the image and the question, exploiting dataset biases rather than capturing a sound perception and reasoning process that would lead to the correct answer (Sturm, 2014). CLEVR was created to address this problem. As illustrated in figure 1, instances in the dataset consist of rendered images featuring 3D objects of several shapes, colors, materials and sizes, coupled with unbiased, compositional questions that require an array of challenging reasoning skills such as following transitive relations, counting objects and comparing their properties, without allowing any shortcuts around such reasoning. Notably, each instance in CLEVR is also accompanied by a tree-structured functional program that was both used to construct the question and reflects its reasoning procedure – a series of predefined operations – that can be composed together to answer it. + +![](images/674dcd95919d901cb7279f9b0552761c7620c6d163165031e5b1cd19087e50a7.jpg) +Figure 1: A sample image from the CLEVR dataset, with a question: “There is a purple cube behind a metal object left to a large ball; what material is it?” + +Most neural networks are essentially very large correlation engines that will hone in on any statistical, potentially spurious pattern that allows them to model the observed data more accurately. In contrast, we seek to create a model structure that requires combining sound inference steps to solve a problem instance. At the other extreme, some approaches adopt symbolic structures that resemble the expression trees of programming languages to perform reasoning (Andreas et al., 2016b; Hu et al., 2017). In particular, some approaches to CLEVR use the supplied functional programs for supervised or semi-supervised training (Andreas et al., 2016a; Johnson et al., 2017). Not only do we wish to avoid using such supervision in our work, but we in general suspect that the rigidity of these structures and the use of an inventory of operation-specific neural modules undermines robustness and generalization, and at any rate requires more complex reinforcement learning methods. + +To address these weaknesses, while still seeking to use a sound and transparent underlying reasoning process, we propose Compositional Attention Networks, a novel, fully differentiable, non-modular architecture for reasoning tasks. Our model is a straightforward recurrent neural network with attention; the novelty lies in the use of a new Memory, Attention and Composition (MAC) cell. The constrained and deliberate design of the MAC cell was developed as a kind of strong structural prior that encourages the network to solve problems by stringing together a sequence of transparent reasoning steps. MAC cells are versatile but constrained neural units. They explicitly separate out memory from control, both represented recurrently. The unit contains three sub-units: The control unit updates the control representation based on outside instructions (for VQA, the question), learning to successively attend to different parts of the instructions; the read unit gets information out of a knowledge base (for VQA, the image) based on the control signal and the previous memory; the write unit updates the memory based on soft self-attention to previous memories, controlled by the retrieved information and the control signal. A universal MAC unit with a single set of parameters is used throughout the reasoning process, but its behavior can vary widely based on the context in which it is applied – the input to the control unit and the contents of the knowledge base. With attention, our MAC network has the capacity to represent arbitrarily complex acyclic reasoning graphs in a soft manner, while having physically sequential structure. The result is a continuous counterpart to module networks that can be trained end-to-end simply by backpropagation. + +We test the behavior of our new network on CLEVR and its associated datasets. On the primary CLEVR reasoning task, we achieve an accuracy of $9 8 . 9 \%$ , halving the error rate compared to the previous state-of-the-art FiLM model (Perez et al., 2017). In particular, we show that our architecture yields better performance on questions involving counting and aggregation. In supplementary studies, we show that the MAC network learns more quickly (both in terms of number of training epochs and training time) and more effectively from limited amounts of training data. Moreover, it also achieves a new state-of-the-art performance of $8 2 . 5 \%$ on the more varied and difficult humanauthored questions of the CLEVR-Humans dataset. The careful design of our cell encourages compositionality, versatility and transparency. We achieve these properties by defining attention-based interfaces that constrict the cell’s input and output spaces, and so constrain the interactions both between and inside cells in order to guide them towards simple reasoning behaviors. Although each cell’s functionality has only a limited range of possible continuous reasoning behaviors, when chained together in a MAC network, the whole system becomes expressive and powerful. In the future, we believe that the architecture will also prove beneficial for other multi-step reasoning and inference tasks, for instance in machine comprehension and textual question answering. + +# 2 RELATED WORK + +There have been several prominent models that address the CLEVR task. By and large they can be partitioned into two groups: module networks, which in practice have all used the strong supervision provided in the form of tree-structured functional programs that accompany each data instance, and large, relatively unstructured end-to-end differentiable networks that complement a fairly standard stack of CNNs with components that aid in performing reasoning tasks. In contrast to modular approaches (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017), our model does not require additional supervision and makes use of a single computational cell chained in sequence (like an LSTM) rather than a collection of custom modules deployed in a rigid tree structure. In contrast to augmented CNN approaches (Santoro et al., 2017; Perez et al., 2017), we suggest that our approach provides an ability for relational reasoning with better generalization capacity and higher computational efficiency. These approaches and other related work are discussed and contrasted in more detail in the supplementary material in section C. + +# 3 COMPOSITIONAL ATTENTION NETWORKS + +Compositional Attention Networks is an end-to-end architecture for question-answering tasks that sequentially performs an explicit reasoning process by stringing together small building blocks, called MAC cells, each is responsible for performing one reasoning step. + +We now provide an overview of the model, and a detailed discussion of the MAC cell. The model is composed of three components: an Input unit, the core MAC network, and an output unit. A TensorFlow implementation of the network, along with pretrained models will be made publicly available. + +In this paper we explore the model in the context of VQA. However, it should be noted that while the input and output units are naturally domain-specific and should be designed to fit the task at hand, the MAC network has been designed to be generic and more broadly applicable, and may prove useful in contexts beyond those explored in the paper, such as machine comprehension or question answering over knowledge bases, which in our belief is a promising avenue for future work. + +# 3.1 THE INPUT UNIT + +The input unit processes the raw inputs given to the system into distributed vector representations. It receives a text question (or in general, a query), and an image (or in general, a Knowledge Base (KB)) and processes each of them with a matching sub-unit, for the query and the KB, here a biLSTM and a CNN. More details can be found in the supplementary material, section A. + +At the end of this stage, we get from the query sub-unit a series of biLSTM output states, which we refer to as contextual words, $[ c w _ { 1 } , . . . , c w _ { S } ]$ , where $S$ is the length of the question. In addition, we get $q = [ \overleftrightarrow { \lambda \omega _ { 1 } } , \overrightarrow { c w _ { S } } ]$ , the concatenation of the hidden states from the backward and forward LSTM passes. We refer to $q$ as the question representation. Furthermore, we get from the Knowledge-Base sub-unit a static representation of the knowledge base. For the case of VQA, it will be represented by a continuous matrix $K B _ { V }$ of dimension $H , W , d$ , where $H = W = 1 4$ are the height and width of the transformed image, corresponding to each of its regions. + +# 3.2 THE MAC CELL + +The MAC network, which is the heart of our model, chains a sequence of small building blocks, called MAC cells, each responsible for performing one reasoning step. The model is provided access to a Knowledge Base (KB), which is, for the specific case of VQA, the given image, and then upon receiving a query, i.e. a question, the model iteratively focuses, in $p$ steps, on the query’s various parts, each reflects in turn the current reasoning step, which we term the control. Consequently, guided by this control, it retrieves the relevant information from the KB, that is then passed to the next cell in a recurrent fashion. + +Drawing inspiration from the Model-View-Controller paradigm used in software design and from the commonly exercised separation between control and data paths in computer architecture, the MAC cell is composed of three units: control unit, read unit and write unit. Each has a clearly defined role and an interface through which it interacts with the other units. See figure 2. + +The careful design and imposed interfaces that constrain the interaction between the units inside the MAC cell, as described below, serve as structural prior that limits the space of hypotheses it can learn, thereby guiding it towards acquiring the intended reasoning behaviors. As such, this prior facilitates the learning process and mitigate overfitting issues. + +In particular, and similar in spirit to Perez et al. (2017), we allow the question to interact with the Knowledge Base – the image for the case of VQA, only through indirect means: by guiding the cell to attend to different elements in the KB, as well as controlling its operation through gating mechanisms. Thus, in both cases, the interaction between these mediums, visual and textual, or knowledge and query, is mediated through probability distributions, either in the form of attention maps, or as gates, further detailed below. This stands in stark contrast to many common approaches that fuse the question and image together into the same vector space through linear combinations, multiplication, or concatenation. Rather, our controlled interaction distills the influence that the query should have in processing the Knowledge Base, casting it onto discrete probability distributions instead. + +![](images/21ef6959e8f5c66e79f26b2aa043aef1bf62c1ad19d7a86079e5194974b914be.jpg) +Figure 2: The MAC cell, which is a recurrent unit comprised of a Control Unit, Read Unit, and Write Unit. Blue shows the control flow and red shows the memory flow. See section 3.2 for details. + +The MAC cell has been designed to replace the discrete and predefined “modules” used in the modular approach (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017). Rather, we create one universal and versatile cell that is applied across all the reasoning steps, sharing both its architecture as well as its parameters, across all of its instantiations. In contrast to the discrete modules, each trained to specialize to some specific elementary reasoning task, the MAC cell is capable of demonstrating a continuous range of possible reasoning behaviors conditioned on the context in which it is applied – namely, the inputs it receives from the prior cell. + +Each cell $M A C _ { i }$ maintains two dual states: control $c _ { i }$ and memory $m _ { i }$ , both are continuous vectors of dimension $d$ . The control $c _ { i }$ represents the reasoning operation the MAC cell should accomplish in the current step – focusing only on some aspect of the whole question. This is represented by a weighted-average attention-based sum of the question words. The memory $m _ { i }$ represents the current context information deemed relevant to respond to the query, or answer the question.This is represented practically by a weighted average over elements from the KB, or for the case of VQA, regions in the image. $m _ { 0 }$ and $c _ { 0 }$ are initialized each to a random vector parameter of dimension $d$ . The memory and control states are passed from one cell to the next in a recurrent fashion, and used in a way reminiscent of Key-Value memory networks (Miller et al., 2016), as discussed below. + +# 3.2.1 THE CONTROL UNIT + +The control unit determines the reasoning operation that should be applied at this step. It receives the contextual words $[ c w _ { 1 } , . . . , c w _ { S } ]$ , the question representation $q$ , and the control state from the previous MAC cell $c _ { i - 1 }$ , all of which are vectors of dimension $d$ . + +We would like to allow our MAC cell to perform continuously varied and adaptive range of behaviors, as demanded by the question. Therefore, we define the behavior of each cell to be a function of the contextual words $[ c w _ { 1 } , . . . , c w _ { S } ]$ , weighted-averaged according to the attention distribution that the control unit produces at each step. This will allow the cell to adapt its behavior – the reasoning operation it performs – to the question it receives, instead of having a fixed set of predefined behaviours as is the case in competing approaches Andreas et al. (2016a;b); Johnson et al. (2017). + +The formal specification of the control unit is shown in figure 3. The question $q$ is linearly transformed into a vector $q _ { i }$ of the same dimension, which in turn is concatenated with the previous control state $c _ { i - 1 }$ and linearly transformed again to a $d$ -dimensional vector $c q _ { i }$ . + +$$ +\begin{array} { c } { q _ { i } = W _ { i } ^ { d , d } \cdot q + b _ { i } ^ { d } } \\ { c q _ { i } = W ^ { 2 d , d } \left[ q _ { i } , c _ { i - 1 } \right] + b ^ { d } } \end{array} +$$ + +Note that in contrast to all other parameters of the cell, which are shared across its instantiations at the different steps $i = 1 , . . . , p$ , the parameters $W _ { i } ^ { d , d }$ and $b _ { i } ^ { d }$ are different for each iteration. This is done to allow each cell to attend more readily to different aspects (i.e. parts) of the questions, depending on the index of the current step – its relative stage in the context of the whole reasoning process. + +![](images/432535676eeb69a82d7848d82a8a327d906d019f07db47ee24440c80b3beab34.jpg) +Figure 3: The Control Unit (CU) of the MAC cell. See section 3.2.1 for details. Best viewed in color. + +$c q _ { i }$ represents the current reasoning operation we would like to perform in a continuous way, taking into account both the overall meaning of the question $q _ { i }$ , as well as the words the model attended to in the previous step, $c _ { i - 1 }$ . + +However, we would like to prevent the cell from diverging in the reasoning operations it tries to perform, and instead anchor it back in the question words, by using them to represent the reasoning operation of the current step. We can achieve that by computing an attention distribution $c v _ { i }$ over the contextual words $[ c w _ { 1 } , . . . , c w _ { S } ]$ based on their similarity to $c q _ { i }$ . Then, summing the contextual words according to the attention distribution $c v _ { i }$ will allow us to have a new control state, $c _ { i }$ , which is represented again in terms of words from the question. Intuitively, it is the gist of the question that is relevant to the reasoning operation we would like to perform in the current step. + +$$ +\begin{array} { c l c r } { { c v _ { i , s } = \displaystyle \mathrm { s o f t m a x } \big ( W ^ { d , 1 } \big ( c q _ { i } \circ c w _ { s } \big ) + b ^ { 1 } \big ) } } \\ { { \big . } } \\ { { c _ { i } = \displaystyle \sum _ { s = 1 } ^ { S } c v _ { i , s } \cdot c w _ { s } } } \end{array} +$$ + +Finally, the control unit returns the current control state $c _ { i }$ , along with an attention map $c v _ { i }$ over the contextual words. + +# 3.2.2 THE READ UNIT + +The Read Unit is provided with access to the knowledge base $K B _ { V }$ , along with the previous memory state $m _ { i - 1 }$ and the current control $c _ { i }$ . It is responsible for retrieving relevant content from the Knowledge Base $K B _ { V }$ for the reasoning task that the MAC cell should accomplish at this step, which is represented by the current control state $c _ { i }$ , as explained above. Figure 4 shows a diagram. + +The relevance of the new information is judged in two stages by the “relatedness” of each element in the KB (or for the case of VQA, each region in the image) to either the memory $m _ { i - 1 }$ that has accumulated relevant information from previous iterations, or to the current control $c _ { i }$ , pointing towards the next piece of information that should be taken into account. Here, relatedness is measured by trained linear transformations comparing each element to the previous memory and the current control. + +More formally, at the first stage, the interaction between each element $K B _ { h , w }$ , where $\begin{array} { r l } { h } & { { } = } \end{array}$ $1 , . . . , H , w = 1 , . . . , W$ , and the previous memory $m _ { i - 1 }$ is computed by: + +$$ +m _ { i - 1 } ^ { \prime } = W ^ { d , d } \cdot m _ { i - 1 } + b ^ { d } +$$ + +$$ +\begin{array} { r l } { \boldsymbol { K } \boldsymbol { B } _ { h , w } ^ { \prime } = \boldsymbol { W } ^ { d , d } \cdot \boldsymbol { K } \boldsymbol { B } _ { h , w } + \boldsymbol { b } ^ { d } } & { { } } \\ { \big ( \boldsymbol { I } _ { m - K B } \big ) _ { h , w } = \boldsymbol { m } _ { i - 1 } ^ { \prime } \circ \boldsymbol { K } \boldsymbol { B } _ { h , w } ^ { \prime } } & { { } } \end{array} +$$ + +![](images/a538d781a85f8151d41cc83e31d872618ae93e0b206ef08dd299f4020e8135bf.jpg) +Figure 4: The Read Unit (RU) diagram. Blue refers to control flow, purple to knowledge flow and red to memory flow. See section 3.2.2 for description. + +These memory-KB interactions measure the relatedness of each element in the KB to the memory accumulated so far, which holds information that has been deemed relevant to handle previous reasoning steps towards addressing the question. They allow the model to perform transitive inference, retrieving a new piece of information that now seems important in light of the recent memory retrieved in a prior iteration. + +However, there are cases which necessitate the model to temporarily ignore current memories, when choosing the new information to retrieve. Logical OR is a classical example: when the model has to look at two different objects at the same time, and assuming it stored one of them at the first iteration, it should briefly ignore it, considering new information that is relevant to the question but is unrelated to the memory. In order to achieve such capability, the read unit concatenates the original KB elements to each corresponding memory-KB interaction, which are then projected back to $d$ -dimensional space (equation 6a): + +$$ +\begin{array} { c } { { I _ { m - K B } , ^ { \prime } = W ^ { 2 d , d } \left[ I _ { m - K B } , K B _ { h , w } \right] + b ^ { d } } } \\ { { I _ { c m - K B } = c _ { i } \circ \left( I _ { m - K B } \right) ^ { \prime } } } \end{array} +$$ + +At the second stage, the read unit compares the current $c _ { i }$ with these memory-KB interactions, in order to focus on the information that is relevant to the current reasoning operation that the MAC cell seeks to accomplish. The result is then passed to a softmax layer yielding an attention map $m v _ { i }$ over the KB, which is used in turn to retrieve the relevant information to perform the current reasoning step. + +$$ +\begin{array} { r } { m v _ { i } = \operatorname { s o f t m a x } \left( W ^ { d , d } \cdot I _ { c m - K B } + b ^ { d } \right) } \\ { m _ { n e w } = \displaystyle \sum _ { h , w = 1 , 1 } ^ { H , W } \left( m v _ { i } \right) _ { h , w } \cdot K B _ { h , w } } \end{array} +$$ + +Finally, the read unit returns the newly retrieved information $m _ { n e w }$ , along with an attention map $m v _ { i }$ over the Knowledge Base. + +To give an example of the read unit operation, assume a given question $q$ such as “What object is located left to the blue ball?”, whose associated answer is “cube”. Initially, no cue is provided to the model to attend to that cube, since no direct information about it presents in the question. Instead, based on its comprehension of the question, the model may start by focusing on the blue ball at the first iteration, such that the memory state $m _ { 1 }$ will capture the blue ball. However, in the second iteration, the control unit, after re-examining the question, may realize it should now look left, storing the word “left” in $c _ { 2 }$ . Then, when considering both $m _ { 1 }$ and $c _ { 2 }$ , the read unit will realize it should perform a reasoning operation corresponding to the word “left” (stored in $c _ { 2 }$ ) given a memory representing the blue ball in $m _ { 1 }$ , thereby allowing it to look left to the blue ball and find the cube. + +# 3.2.3 THE WRITE UNIT + +The Write Unit is responsible for creating the new memory state $m _ { i }$ that will reflect all the information considered to be important to answer the question so far, i.e. up to the current iteration in the + +![](images/da3df6ec5443731e5ef4aaf8424acd7af4458f638b9bc19f84b7d1ed88b9267e.jpg) +Figure 5: The Write Unit (WU) diagram. Blue refers to control flow and red to memory flow. See section 3.2.3 for description. + +reasoning process. It receives the last memory state $m _ { i - 1 }$ from the previous MAC cell, along with the newly retrieved information from the read unit in the current iteration, $m _ { n e w }$ . See figure 5 for a diagram. + +In the main design we have explored, merging the new information with the previous memory state is done simply by a linear transformation. + +$$ +m _ { i } ^ { \prime } = W ^ { 2 d , d } [ m _ { n e w } , m _ { i - 1 } ] + b ^ { d } +$$ + +In addition, we have explored two variations of this design. The first, self-attention, allows considering any previous memories rather than just the last one $m _ { i - 1 }$ , thus providing the network with the capacity to model non-sequential reasoning processes. The second variation is adding gating mechanisms to the writing unit. These may allow the model to dynamically adjust the practical length of the computation to the question complexity and stabilize the memory content throughout the sequential network (similarly to GRUs and LSTMs). + +Self-Attention. The current architecture that we have presented allows the model to perform reasoning steps in a sequence, passing control and memory states from one cell to the following. However, we would like to grant the system with more flexibility. Particularly, we would like to allow it to capture more complicated reasoning processes such as trees and graphs - Directed Acyclic Graph (DAG) in particular, where several branches of reasoning sub-processes are merged together in later stages. Indeed, the CLEVR dataset includes cases where the questions embody tree-like reasoning process, rather than just sequences, which we would like to address correctly in our model. + +We achieve that by adding self-attention connections between each MAC cell and all the prior cells. Since each cell can look on all the prior reasoning steps and their corresponding memories retrieved from the Knowledge Base, it can virtually capture any directed acyclic graph, while still having physically sequential layout. + +More formally, the current MAC cell, of the $i ^ { t h }$ iteration, is granted with access to $c _ { 1 } , . . . , c _ { i - 1 }$ along with the corresponding $m _ { 1 } , . . . , m _ { i - 1 }$ , that have been computed by the prior MAC cells. It begins by computing the similarity between $c _ { i }$ and $c _ { 1 } , . . . , c _ { i - 1 }$ , and use it to derive an attention map over the prior MAC cells $s a _ { i , j }$ for $j = 1 , . . . , i - 1$ . This represents the relevance of the $j ^ { t h }$ prior reasoning step to the current one $i$ (equation 9a). + +Then, we average the previous memories according to this resulted attention map $s a _ { i j }$ . We obtain $m _ { s a }$ , representing the information from all the other reasoning steps that is relevant to the current one (equation 9b). + +This resembles the approach of Key-Value networks (Miller et al., 2016). The similarity between control states, corresponding to the reasoning operations that are performed in each prior step, allows the model to select which memories should be taken into account, when creating the new memory – namely, which branches of the reasoning process should be merged together at this point. + +$$ +\begin{array} { c } { { s a _ { i j } = \displaystyle \mathrm { s o f t m a x } \left( W ^ { d , 1 } ( c _ { i } \circ c _ { j } ) + b ^ { 1 } \right) } } \\ { { ( m _ { s a } ) _ { i } = \displaystyle \sum _ { j = 1 } ^ { i - 1 } s a _ { i j } \cdot m _ { j } } } \end{array} +$$ + +Finally, we use $m _ { s a }$ along with $m _ { i } ^ { \prime }$ to compute $m _ { i } ^ { \prime \prime }$ , the new memory content in this variation. + +$$ +m _ { i } ^ { \prime \prime } = W ^ { 2 d , d } [ m _ { n e w } , m _ { i } ^ { \prime } ] + b ^ { d } +$$ + +Memory Gate. The currently presented MAC network has some fixed number $p$ of concatenated MAC cells, representing the length of the overall reasoning process we perform. However, not all questions require reasoning sequence of the same length. Some questions are simpler while others more complex. + +Motivated by this observation, we add a gate over the new memory computed at each step, that may selectively keep content of the previous memory $m _ { i - 1 }$ unchanged. Practically, the gate functions in a similar way to a highway network (Srivastava et al., 2015), where the gate value is conditioned on the current reasoning operation, $c _ { i }$ . + +$$ +\begin{array} { c } { { { c _ { i } } ^ { \prime } = W ^ { d , d } \cdot c _ { i } + b ^ { d } } } \\ { { m _ { i } = \mathrm { s i g m o i d } \left( c _ { i } ^ { \prime } \right) \cdot m _ { i - 1 } + ( 1 - \mathrm { s i g m o i d } \left( c _ { i } ^ { \prime } \right) ) \cdot m _ { i } ^ { \prime \prime } } } \end{array} +$$ + +The write unit returns the new memory state $m _ { i }$ , that will be passed along with $c _ { i }$ to the next MAC cell. + +# 3.2.4 DISCUSSION + +Overall, when designing the MAC cell, we have attempted to formulate the inner workings of an elementary, yet generic reasoning skills: the model decomposes the problem into steps, focusing on one at a time. At each such step, it takes into account: + +• The control $c _ { i }$ : Some aspect of the task - pointing to the future work that has left to be done. +• The previous memory or memories: The partial solution or evidence the cell has acquired so far – pointing to the past work that has already been achieved. +• The newly retrieved information $m _ { n e w }$ : that is retrieved from the knowledge base $K B$ and may or may not be transitively related to that partial solution or evidence - the present, or current work. + +Considering these three sources of information together, the cell finally adds the new information up into its working memory, $m _ { i }$ , progressing one more step towards the final answer. + +# 3.3 THE OUTPUT UNIT + +The output unit receives the question representation $q$ , along with the memory state passed from the last MAC cell $m _ { p }$ , where $p$ is the number of MAC cells in the network – representing the number of reasoning steps in the whole process. It inspects both and predicts an answer based on their concatenation. Intuitively, we would like our model to consider both the question as well as the relevant information that has been progressively retrieved from the KB, deemed the necessary information to answer it. + +Note that considering both $q$ and $m _ { p }$ is critical to answer the question. While $m _ { p }$ represents the information collected from KB, we still need to recall what has been asked about it to be able to answer accordingly. This is especially true in our case, when all other interactions between the question and the KB are mediated through attention distributions, rather than being transformed into a shared continuous vector space. + +The prediction is built out of a standard 2-layers fully-connected softmax-based classifier with hidden dimension $d$ and output dimension that matches the number of possible answers in the dataset. The classifier receives $[ m _ { p } , q ]$ as input and returns a probability distribution over the answers. + +Table 1: CLEVR accuracy by baseline methods, competing methods, and our method (MAC) - for our method we show test results with GLoVE initialization and validation results with and without GLoVE. (\*) denotes use of extra supervisory information through program labels. $( ^ { \dagger } )$ denotes use of data augmentation. $( ^ { \ddagger } )$ denotes training from raw pixels. $( ^ { \circ } )$ denotes GLoVE word vector initialization. + +
ModelOverallCountExistCompare NumbersQuery AttributeCompare Attribute
Human (Johnson et al., 2017)92.686.796.686.595.096.0
Q-type baseline (Johnson et al., 2017)41.834.650.251.036.051.3
LSTM (Johnson et al., 2017)46.841.761.169.836.851.8
CNN+LSTM (Johnson et al., 2017)52.343.765.267.149.353.0
CNN+LSTM+SA (Johnson et al., 2016)76.664.482.777.482.675.4
N2NMN* (Hu et al. 2017)83.768.585.784.990.088.7
PG+EE (9K prog.)* (Johnson et al., 2017)88.679.789.779.192.696.0
PG+EE (70oK prog.)* (Johnson et al., 2017)96.992.797.198.798.198.9
CNN+LSTM+RN†‡ (Santoro et al.,2017)95.590.197.893.697.997.1
CNN+GRU+FiLM (Perez et al.,2017)97.794.399.196.899.199.1
CNN+GRU+FiLM* (Perez et al., 2017)97.694.399.393.499.399.3
CNN+GRU+FiLM (Perez et al., 2017)97.794.399.196.899.199.1
MAC (this paper)-val98.997.199.599.399.299.2
MAC-val99.097.299.599.599.599.4
MAC-test98.997.299.599.499.399.5
+ +# 4 EXPERIMENTS + +We evaluate our model on the recent CLEVR dataset (Johnson et al., 2016). CLEVR is a synthetic dataset consisting of 700K tuples; each consists of a 3D-rendered image featuring objects of various shapes, colors, materials and sizes, coupled with compositional multi-step questions that measure performance on an array of challenging reasoning skills such as following transitive relations, counting objects and comparing their properties. In addition, each question is associated with a formal program, specifying the reasoning operations that should be performed to compute the answer, among 28 possibilities. + +We first perform experiments on the original 700k CLEVR dataset (Johnson et al., 2016), comparing to prior work. As shown in table 1, our model matches or outperforms all existing models both in overall accuracy, as well as in each category, testing different reasoning skills. In particular, for the overall performance, we achieve $9 8 . 9 4 \%$ accuracy, more than halving the error rate of the prior best model, FiLM (Perez et al., 2017). + +Counting and Numerical Comparison. Remarkably, our performance on questions testing counting and numerical comparisons is significantly higher than the competing models, which consistently struggle on this question type. Again, we nearly halve the corresponding error rate. These results demonstrate the aptitude of attention mechanisms to perform counting, reduction and aggregation, in contrast to alternative, CNN-based approaches. + +Training Length and Computational-Efficiency. We examine the learning curves of our and competing models. We have trained all models on the same architecture and used the author code for the other models. Aiming at having equal settings for comparison, we ran all models including ours with learned random words vectors. In order to make sure the results are statistically significant we ran each model multiple (10) times, and plotted the averages and confidence intervals (figure 4). The results show that our model learns significantly faster than the other leading methods, FiLM (Perez et al., 2017) and $\mathrm { P G + E E }$ (Johnson et al., 2017). While we do not have learning curves for the Relational Network model, Santoro et al. (2017) report approximately 1.4 million iterations to achieve $9 5 . 5 \%$ accuracy, which are equivalent to 125 epochs approximately, whereas our model achieves a comparable accuracy after 3 epochs only, yielding 40x reduction in the length of the training process. + +Naturally, the smaller number of required training steps also translates to comparably shorter training time. Perez et al. (2017) report training time of 4 days, equivalent to 80 epochs, to reach accuracy of $9 7 . 7 \%$ . In contrast, we achieve higher accuracy in 6 epochs, taking 9.5 hours overall, leading to $1 0 \mathrm { x }$ reduction in training time. + +![](images/c74201d70aae1d3587e14044064872d252c0e00ec6d3c4fd6b70377bb7088c75.jpg) +Figure 6: Training curves and accuracies for MACs (our model), FiLM (Perez et al., 2017), $\mathrm { P G + E E }$ (Johnson et al., 2017) and stacked-attention (Yang et al., 2016; Johnson et al., 2017). (Note: $\mathrm { P G + E E }$ uses the supported CLEVR programs as strong supervision.) Left: Training curve (accuracy/epoch). Right: Learning curve: Accuracy for $1 \%$ , $2 \%$ , $5 \%$ , $10 \%$ , $2 5 \%$ , $50 \%$ and $100 \%$ of the $7 0 0 \mathrm { k }$ CLEVR samples. + +Table 2: Accuracy on CLEVR-Humans of previous methods and our method (MAC), before (left) and after (right) fine-tuning on the CLEVR-Humans training data. $\mathrm { P G + E E }$ uses supervised data. For our method (MAC) we show test results with GLoVE initialization and validation results with and without GLoVE. + +
ModelTrain CLEVRTrain CLEVR + fine-tune HUMANS
LSTM (Johnson et al.,2017)27.536.5
CNN+LSTM (Johnson et al.,2017)37.743.2
CNN+LSTM+SA+MLP (Johnson et al.,2016)50.457.6
PG+EE (18K prog.)* (Johnson et al.,2017)54.066.6
CNN+GRU+FiLM (Perez et al., 2017)56.675.9
MAC-val56.981.2
MAC-val58.582.0
MAC-test58.682.5
+ +# 4.1 DATA EFFICIENCY + +We have explored the performance of our and other leading approaches on smaller subsets of the CLEVR dataset, in order to study the ability of models to generalize from smaller amount of data. We sampled at random subsets of CLEVR, with $10 \%$ , $2 5 \%$ and $50 \%$ of its original $7 0 0 \mathrm { k }$ size, and used them to train our and other 3 proposed models for the CLEVR task: FiLM (Perez et al., 2017), the strongly-supervised $\mathrm { P G + E E }$ (Johnson et al., 2017), and stacked-attention networks (Johnson et al., 2017; Yang et al., 2016). + +As shown in figure 4, our model outperforms the other models by a wide margin for all subsets of the CLEVR dataset. For $50 \%$ of the data, equivalent to $3 5 0 \mathrm { k }$ samples, other models obtain accuracies ranging between $70 \%$ and $92 \%$ , while our model achieves $9 7 . 9 \%$ . The gap becomes larger as the dataset size reduces: for $2 5 \%$ of the data, equivalent to $1 7 5 \mathrm { k }$ samples, performance of other models is between $50 \%$ and $7 7 \%$ , while our model maintains a high $9 5 . 4 \%$ accuracy. + +Finally, for $10 \%$ of the data $- 7 0 \mathrm { k }$ samples, still a sizeable amount – our model is the only one that manages to generalize, with performance of $8 4 . 7 \%$ on average, whereas the other three models fail, achieving $4 7 . 6 \% - 5 7 . 5 \%$ . Note that as pointed out by (Johnson et al., 2016) a simple baseline that predicts the most frequent answer for each of the question types achieves already $4 2 . 1 \%$ , suggesting that answering half of the questions correctly means that the competing models barely learn to generalize from the smaller dataset. These results demonstrate the robustness of our architecture and its key role as a structural prior guiding our network to learn the intended reasoning skills. + +4.2 CLEVR HUMANS - NATURAL LANGUAGE QUESTIONS + +We analyze our model performance on the CLEVR-Humans dataset (Johnson et al., 2017), consisting of natural language questions collected through crowdsourcing. As such, the dataset has diverse vocabulary and linguistic variations, and it also demands more varied reasoning skills. + +Since the training set is relatively small, consisting of $1 8 \mathrm { k }$ samples, we use it to finetune a model pretrained on the standard CLEVR dataset. However, since most of the vocabulary in CLEVRHumans is not covered by CLEVR, we do not train the word vectors during the pre-training stage, so to prevent drift in their meaning compared to other uncovered words in CLEVR-Humans that may be semantically related. + +As shown in table 2, our model achieves state-of-the-art performance on CLEVR-Humans both before and after fine-tuning. It surpasses the next-best FiLM model, (Perez et al., 2017) by $6 . 6 \%$ percent, achieving $8 2 . 5 \%$ . + +The results substantiate the model’s robustness against linguistic variations and noise, as well as its ability to adapt to diverse vocabulary and varied reasoning skills. Arguably, the soft attention performed over the question words allows the model to focus on the words that are most critical to answer the question and translate them to corresponding reasoning operations, giving less attention to irrelevant linguistic variations. + +# 4.3 ABLATIONS + +![](images/d542dee22ea6b6971ea171ea3647286d90bc0976dc1e0a42885a1582ff9227cc.jpg) +Figure 7: Ablations and variations of the MAC network. See text for full detail. + +Based on the validation set, we have conducted an ablation study on our model to understand better the contribution of each of its component to the overall performance. We tested each setting on the standard 700K CLEVR dataset as well as on $10 \%$ subset of the dataset. See table 3 for the numerical results. In addition, figure 4.3 presents the training curves for the different settings trained on the standard dataset. Overall, the results demonstrate the robustness of the model to hyperparameter variations such as network dimension and length, and also the impact of different aspect and components of MAC on its performance. + +Table 3: Ablations Results for the full standard CLEVR dataset and $10 \%$ subset of it. See text for full details. + +
ModelStandard CLEVR10% CLEVR
MAC standard model (dimension 512)98.984.5
dimension 25698.476.3
dimension 12897.677.0
w/o control55.651.5
w/o word-attention95.363.2
full-question-as-control80.765.0
w/o positional encoding99.180.0
linear position encoding99.083.2
identity writing unit98.584.5
only-gating writing unit99.383.1
+ self-attention83.299.2
+ gating99.483.1
+ self-attention and gating99.585.5
gate bias 098.784.9
gate bias 199.468.5
gate bias -199.077.1
unshared weights97.867.5
prediction without question information97.864.7
+ +Network Length. We have tested the model performance as a function of the network’s length – the number of MAC cells that were sequenced together. The results show the positive correlation between the network length and its performance. We can see that for 1 cell the scores are relatively $1 0 \mathrm { w } - 7 5 \%$ , but adding at least one more cell leads to a significant increase in performance above $9 5 \%$ . The performance keeps improving up to lengths 8-16 that achieve $9 8 . 9 – 9 9 . 1 \%$ . The results also teach us about the complexity of the dataset, by showing the relatively significant benefits of having at least 4 cells, each modeling a reasoning step. + +Network Dimension. We have varied the state dimension to check the robustness of the model to hyperparameters. The results on the standard CLEVR dataset show the model is able to maintain high performance with dimension of 128, albeit after a longer training process, achieving $9 7 . 6 \%$ , compared to $9 8 . 9 4 \%$ achieved with dimension of 512. However, for $10 \%$ of CLEVR, the larger 512-dimension allows accuracy increase by $7 . 5 \%$ over dimension of 128. + +Weight Sharing. We have tested the impact of sharing weights between cell has on the model performance for network of length $p = 1 2$ . The results show that for the standard dataset there is only a small difference between these settings of $1 \%$ . However, for less data, we see much more significant drop $1 6 . 9 \%$ in the unshared-parameters setting compared to the shared one. Indeed, we observe that a model with less parameter is more data-efficient and has a lower tendency to overfit the data. + +Control Unit. We have performed several ablations in the control unit to understand its contribution to the overall model performance. Based on the results, first, we can see the the question information is crucial for the model to handle the questions, as can be noted by the low performance of the model when there is no use of control signal whatsoever. Second, we have tested the model performance when using the continuous control state computed by question (2) in section 3.2.1, without having word-attention, in order to understand its relative contribution. Based on the results, we can indeed see that using word-attention is useful for accelerating the training process and achieving higher accuracies both for the standard dataset as well as for the small subset, where using word-attention increases results in $2 1 . 4 \%$ . We also see that using the “contextual words” produced by the questionunit LSTM is useful in accelerating the model performance, when compared to using the wordvectors directly. + +Reading Unit. We have conducted several ablations for the reading unit to better understand its behavior and contribution to the performance of the model. The standard MAC reading unit uses the control state – which averages the question words based on attention distributions computed per each reasoning step. In this ablation experiment, we have tested using the full question representation $q$ instead across all reasoning steps to gain better understanding of the the contribution of wordattention to the model performance. Indeed, we can see that using $q$ rather then the control state $c _ { i }$ results in a significant drops in performance $- 1 9 . 4 \%$ for the full CLEVR dataset and $1 9 . 5 \%$ for $10 \%$ of the data. + +We have conducted additional ablation experiment to better understand the contribution of using the KB features directly in the first-stage information retrieval process described in section 3.2.2, compared to using only the dot-products of the KB elements with the previous memory state $m _ { i - 1 }$ . For the full CLEVR dataset, we can see that this component has only a small impact in the final performance - ultimately resulting in $0 . 0 6 \%$ performance difference. However, for the $10 \%$ of the data, we can see that the difference in performance when ablating this component is much larger - $1 1 . 2 \%$ . + +Writing Unit Ablations. In our main MAC model variant, the memory unit merges the new information $m _ { n e w }$ with the previous memory state $m _ { i - 1 }$ by combining them through a linear transformation. In this experiment, we have explored other variations, such as assigning $m _ { n e w }$ to $m _ { i }$ directly – ignoring previous memories, or doing a linear transformation based on $m _ { n e w }$ only. The results show that in fact such variant is only slightly worse than our main variant – $0 . 4 \%$ . We also conducted an experiment in which we merge the new information with the previous memory just by a having a gate that does a weighted average of them. The results show that this variant performs equivalently to our standard linear-transformation variant. + +Writing Unit Additions. We have explored the impact of the writing unit variants described in section 3.2.3 – adding self-attention, gating mechanisms, or both, compared to our standard main model that uses a linear transformation to merge the newly retrieved information $m _ { n e w }$ with the previous memory $m _ { i }$ . For the complete CLEVR dataset we can see that indeed both these variants are very helpful in increasing the model performance. Compared to our standard MAC model that achieves $9 8 . 9 4 \%$ on the validation set, self-attention yields accuracy of $9 9 . 2 3 \%$ , gating yields $9 9 . 3 6 \%$ and adding both achieves $9 9 . 4 8 \%$ . + +Output Unit. In our standard model, the final predictions made in the output unit are based on the final memory state $m _ { p }$ as well as question representation $q$ (stands for the final hidden states of the backward and forwards passes of the LSTM). We have explored the contribution of basing the model prediction on the latter, by testing the model performance when prediction is based on memory alone, for the complete and $10 \%$ datasets. We can see that in both settings basing the model’s predictions on the question representation allows faster training and higher accuracies. Notable is the gap in performance for the $10 \%$ CLEVR - $1 9 . 8 \%$ increase by using the question representation to make predictions. These results are very reasonable intuitively, since the model is structured such that the memory holds only information that was retrieved from the image. Thus, questions that may ask for instance on different aspects (such as color or shape) of the same object in the image may result in the same memory content, which is thus does not directly contain enough information to respond such questions. + +Position. In our standard model, similarly to the practice of competing models (Santoro et al., 2017; Perez et al., 2017; Hu et al., 2017), we have concatenated positional information to each region of the image, in order to increase the model capability to perform spatial reasoning. We have explored both simple linear maps at a constant $[ - 1 , 1 ]$ as well as more complex positional encoding suggested by (Vaswani et al., 2017). However, the results for both the standard dataset and the $10 \%$ version show a very negligible improvement at best when adding positional encoding information, demonstrating the capability of MAC to perform spatial reasoning without data augmentation. + +Gate Bias Initialization. For our model variant with gating mechanism (described in section 3.2.3) we have tested the effect of setting different values for the gate bias - $- 1 , 0$ and 1. for $- 1$ the model is initialized to biased for keeping the previous memory value whereas for 1 it will be biased for using the new memory instead. We can see that for the complete dataset setting the bias to 1 is optimal – apparently since the model has enough data to learn to apply each cell effectively. In contrast, for the small $10 \%$ CLEVR data, setting the bias to 0 shows better performance, biasing the model to using less cells overall which results ultimately in a theoretically-simpler model that can fit less data more effectively. + +# 4.4 INTERPRETABILITY + +We have looked into attention maps over the image and question that the model produces during its computation and provide a few examples in figure 4.4. The first example shows us how the model parses the question in steps, first focusing on the main entity that the question is about, then on + +relation of this entity to the “brown matte thing” which is then located in the image. Finally, the model correctly focuses on the small brown cube and predicts the right answer – brown. + +The second example shows a model with 4 cells instead of 6, that similarly parse the question in iterations and focuses on the relevant objects at each step, though we can see that the reasoning process looks somewhat different when the MAC network has fewer cells. + +The last example shows how how the model handles counting and OR operations. It starts from identifying the task - computing a number, and then red objects as well as the cylinder, one at a time, allowing it ultimately to respond correctly, with the answer 2. + +![](images/e280158a489b95ddeabb4461dda86460910408860df91a493fe23a53648a08b4.jpg) + +Figure 8: Three examples of attention maps for the image and the question. See text for full details. + +# 5 CONCLUSION + +We have given a first demonstration of how a sequence of Memory, Attention and Control (MAC) cells combined into a Compositional Attention Network provides a very effective tool for neural reasoning. In future work, we wish to explore this promising architecture for other tasks and domains, including real-world VQA, machine comprehension and textual question answering. + +# REFERENCES + +Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question answering models. arXiv preprint arXiv:1606.07356, 2016. + +Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 39–48, 2016a. + +Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Learning to compose neural networks for question answering. arXiv preprint arXiv:1601.01705, 2016b. + +Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International Conference on Computer Vision, pp. 2425–2433, 2015. + +Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. + +Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. + +Djork-Arne Clevert, Thomas Unterthiner, and Sepp Hochreiter. Fast and accurate deep network ´ learning by exponential linear units (elus). arXiv preprint arXiv:1511.07289, 2015. + +Alex Graves, Navdeep Jaitly, and Abdel-rahman Mohamed. Hybrid speech recognition with deep bidirectional lstm. In Automatic Speech Recognition and Understanding (ASRU), 2013 IEEE Workshop on, pp. 273–278. IEEE, 2013. + +Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint arXiv:1410.5401, 2014. + +Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka GrabskaBarwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, ´ et al. Hybrid computing using a neural network with dynamic external memory. Nature, 538 (7626):471–476, 2016. + +Akshay Kumar Gupta. Survey of visual question answering: Datasets and techniques. arXiv preprint arXiv:1705.03865, 2017. + +Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016. + +Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8): 1735–1780, 1997. + +Ronghang Hu, Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Kate Saenko. Learning to reason: End-to-end module networks for visual question answering. arXiv preprint arXiv:1704.05526, 2017. + +Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456, 2015. + +Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. arXiv preprint arXiv:1612.06890, 2016. + +Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Judy Hoffman, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Inferring and executing programs for visual reasoning. arXiv preprint arXiv:1705.03633, 2017. + +Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. + +Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012. + +Ankit Kumar, Ozan Irsoy, Peter Ondruska, Mohit Iyyer, James Bradbury, Ishaan Gulrajani, Victor Zhong, Romain Paulus, and Richard Socher. Ask me anything: Dynamic memory networks for natural language processing. In International Conference on Machine Learning, pp. 1378–1387, 2016. + +Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´ document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. + +Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention for visual question answering. In Advances In Neural Information Processing Systems, pp. 289– 297, 2016. + +Alexander Miller, Adam Fisch, Jesse Dodge, Amir-Hossein Karimi, Antoine Bordes, and Jason Weston. Key-value memory networks for directly reading documents. arXiv preprint arXiv:1606.03126, 2016. + +Jeffrey Pennington, Richard Socher, and Christopher Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pp. 1532–1543, 2014. + +Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. arXiv preprint arXiv:1709.07871, 2017. + +Adam Santoro, David Raposo, David GT Barrett, Mateusz Malinowski, Razvan Pascanu, Peter Battaglia, and Timothy Lillicrap. A simple neural network module for relational reasoning. arXiv preprint arXiv:1706.01427, 2017. + +Rupesh Kumar Srivastava, Klaus Greff, and Jurgen Schmidhuber. Highway networks. ¨ arXiv preprint arXiv:1505.00387, 2015. + +Bob L Sturm. A simple method to determine if a music information retrieval system is a horse. IEEE Transactions on Multimedia, 16(6):1636–1644, 2014. + +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017. + +Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and textual question answering. In International Conference on Machine Learning, pp. 2397–2406, 2016. + +Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks for image question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 21–29, 2016. + +SUPPLEMENTARY MATERIAL + +# A DETAILS OF INPUT UNIT + +The input unit processes the raw inputs given to the system into distributed vector representations. It receives a text question (or in general, a query), and an image (or in general, a Knowledge Base (KB)) and processes each of them with a matching sub-unit. Here we provide details of the Query Unit and the Image Unit used in this work. + +# A.0.1 THE QUERY UNIT + +We encode a query of $S$ words into a continuous representation using a bidirectional LSTM (Hochreiter & Schmidhuber, 1997; Graves et al., 2013). Each word is associated with a word embedding $w _ { s }$ , where $s = 1 , . . . , S$ . In our case, we use GloVE words embeddings (Pennington et al., 2014). Then, these embeddings are processed by a bidirectional LST M of dimension $d$ that outputs: + +• a matching sequence of $d$ -dimensional output states, which we refer to as contextual words, $[ c w _ { 1 } , . . . , c w _ { S } ]$ • $d$ -dimensional hidden state $q = [ \overleftrightarrow { \lambda \omega _ { 1 } } , \overrightarrow { c w _ { S } } ]$ , the concatenation of the hidden states from the backward and forward passes. We refer to $q$ as the question representation. + +Intuitively, each contextual word $c w _ { s }$ represents the meaning of $s ^ { t h }$ word, in the context of the question, while the hidden state $q$ represents the overall (compositional) meaning of the question. + +# A.0.2 THE IMAGE UNIT + +Given an image, and following prior work on CLEVR (?Santoro et al., 2017; Perez et al., 2017), we extract conv4 features from ResNet101 (He et al., 2016) pretrained on ImageNet (Krizhevsky et al., 2012) which we treat as a fixed initial representation of the image, $x$ of dimension $H , W , C$ where $H = W = 1 4$ are the height and width of the transformed image and $C = 1 0 2 4$ is the number of channels. Each feature $x _ { h , w }$ represents one region in the original image. + +Similar to prior work (Hu et al., 2017; Santoro et al., 2017; Perez et al., 2017), we would like to allow our model to reason explicitly about spatial locations, as required by many of the questions in CLEVR, and therefore we concatenate to this representation a spatial map that represents each of the positions in the image. However, in contrast to prior work that uses a linear meshgrid feature map with 2 features $h$ and $w$ ranging from $- 1$ to 1, and to allow better representation of the positions, we use the positional encoding scheme proposed by Vaswani et al. (2017): + +$$ +\begin{array} { l } { p _ { ( h , 2 i ) } = \sin { \Big ( h / 1 0 0 0 0 ^ { 2 i / p _ { d } } \Big ) } } \\ { p _ { ( h , 2 i + 1 ) } = \cos { \Big ( h / 1 0 0 0 0 ^ { 2 i / p _ { d } } \Big ) } } \end{array} +$$ + +And similarly for $w$ , where $p _ { d }$ is a hyperparameter. Overall, the positional encoding of a feature at position $( h , w )$ is $[ p _ { h } , p _ { w } ]$ , the concatenation of the positional encodings for $h$ and $w$ . + +This positional encoding scheme allows better correspondence between the distance of 2 positions $( x , y )$ and $( x , y )$ in the image and a vector similarity of their positional encodings, even when $p _ { d }$ is larger than two. + +We then concatenate the obtained spatial map with $x$ , receiving a spatially-aware image representation $x _ { p }$ . Then, we pass this representation through two CNN layers with $d$ output channels and obtain a final representation of the image, which we refer to as our Visual Knowledge Base $( K B _ { V }$ that is used in further components of the model. + +# B IMPLEMENTATION AND TRAINING DETAILS + +For the question processing, we use GloVE (Pennington et al., 2014) word-vectors with dimension 300. For the image processing, we extract conv4 features from ResNet101 (He et al., 2016) pretrained on ImageNet (Krizhevsky et al., 2012), with dimension $H , W , C$ where $H = W = 1 4$ and $C = 1 0 2 4$ , followed by 2 CNN layers with kernel size 2. We use MAC network with $p = 1 2$ cells, and train it using Adam (Kingma & Ba, 2014), with learning rate $1 0 ^ { - 4 }$ . We train our model for $1 0 - 2 0$ epochs, with batch size 64, and use early stopping based on validation accuracies. During training, the moving averages of all weights of the model are maintained with the exponential decay rate of 0.999. At test time, the moving averages instead of the raw weights are used. We use dropout 0.85, and ELU (Clevert et al., 2015) which in our experience has reduce the training process compared to RELU.The training process takes roughly 10-20 hours on a single Titan X GPU. + +# C FURTHER DISCUSSION OF RELATED WORK + +In this section we provide detailed discussion of related work. Several models have been applied to the CLEVR task. These can be partitioned into two groups, module networks that use the strong supervision provided as a tree-structured functional program associated with each instance, and end-to-end, fully differentiable networks that combine a fairly standard stack of CNNs with components that aid them in performing reasoning tasks. We also discuss the relation of MAC to other approaches, such as memory networks and neural computers. + +# C.1 MODULE NETWORKS + +The modular approach (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017) first translates the given question into a tree-structured action plan, aiming to imitate the ground-truth programs provided as a form of strong-supervision. Then, it constructs a tailor-made network that executes the plan on the image in multiple steps. This network is composed of discrete units selected out of a collection of predefined modules, each responsible for an elementary reasoning operation, such as identifying an objects color, filtering them for their shape, or comparing two amounts. Each module has its own set of learned parameters (Johnson et al., 2017), or even hand-crafted design (Andreas et al., 2016a) to guide it towards its intended behavior. + +Overall, this approach makes discrete choices at two levels: the identity of each module – the behavior it should learn among a fixed set of possible types of behaviors, and the network layout – the way in which these modules are wired together to compute the answer progressively. Hence, their differentiability is confined to the boundaries of a single module, disallowing end-to-end training. + +Several key differences exist between our approaches. First, our model replaces the fixed modules collection with one versatile and universal cell that shares both its architecture and parameters across all of its instantiations, and is applied across all the reasoning steps. Second, it replaces the dynamic recursive tree structures with a sequential topology, augmented by soft attention mechanisms, as done in Bahdanau et al. (2014). This confers our network with a virtual capacity to represent arbitrarily complex Directed Acyclic Graphs (DAGs) while still having efficient and readily deployed physical sequential structure. Together, both of these relaxations allow us to effectively train our model end-to-end by backpropagation alone, whereas module networks demand a more involved training scheme that relies on the strongly-supervised programs at the first stage, and on various Reinforcement Learning (RL) techniques at the second. Furthermore, while our model can be train without the strong supervisory programs, developing adaptive reasoning skills to address the task is it trained for, the modular approach reliance on questions structured and formal representation hinder its applicability to real-world tasks. + +# C.2 AUGMENTED CONVOLUTIONAL NEURAL NETWORKS + +Alternative approaches for the CLEVR task that do not rely on the provided programs as a strong supervision signal are Santoro et al. (2017) and Perez et al. (2017). Both complement standard multi-layer Convolutional Neural Networks (CNNs) with components that aid them in handling compositional and relational questions. + +Relational Networks. Santoro et al. (2017) appends a Relation Network (RN) layer to the CNN. This layer inspects all pairs of pixels in the image, thereby enhancing the network capacity to reason over binary relations between objects. While this approach is very simple and elegant conceptually, it suffers from quadratic computational complexity, in contrast to our and other leading approaches. But beyond that, closer inspection reveals that this direct pairwise comparison might be unnecessary. Based on the analogy suggested by Santoro et al. (2017), according to which pixels are equivalent to objects and their pairwise interactions to relations, a RN layer attempts to grasp the induced graph between objects all at once in one shallow and broad layer. Conversely, our attention-based model proceeds in steps. It basically compares the image to its current memory and control for this step, aggregates the attended regions into the new memory, and repeats the process. By the same analogy, it traverses a narrow and deep path, progressively following transitive relations. Consequently, our model exhibits a relational capacity while circumventing the computational inefficiency. + +FiLM. FiLM (Perez et al., 2017) is a recently proposed method that interleaves standard CNN layers that process the given image with linear layers, reminiscent of layer normalization techniques (Ba et al., 2016; Ioffe & Szegedy, 2015). Each of these layers, called FiLM, is conditioned on the question: the question words are processed by a GRU, and its output is linearly transformed into matching biases and variances for each of the CNN layers, tilting its activations to reflect the specifics of the given question and affect the computation done over the image. + +Similarly to our model, this approach features distant modulation between the question and the image, where rather than being fused together into the same vector space, the question can affect the image processing only through constrained means – for the case of FiLM – linear transformations. However, since the same transformation is applied to all the activations homogeneously, agnostic to both their spatial location as well as the features values, this approach does not allow the question to differentiate between regions in the image based on the objects or concepts they represent – on the content of the image. This stands in stark contrast to our attention-based model, which readily allows and actually encourages the question to inform the model about relevant regions to focus on. We speculate that this still distant, yet more direct interaction between the question and the data, or image, for the case of VQA, facilitates learning and increases generalizability. It may be more suitable to VQA tasks, and CLEVR in particular, where the questions demand the responder to focus on specific objects, and reason about their properties or relations, rather than respond based only on a holistic view of the image that may lead to sub-optimal results (Yang et al., 2016), as is the case of FiLM. Indeed, as demonstrated in 4, there is significant evidence showing our models better generalization capacity, allowing it to achieve high accuracies much faster, and from less data than FiLM and other competing methods. + +# C.3 MEMORY AND ATTENTION + +Our architecture draws inspiration from recent research on memory and attention (Kumar et al., 2016; Xiong et al., 2016; Graves et al., 2014; 2016). Kumar et al. (2016); Xiong et al. (2016) propose the Dynamic Memory Network model that proceeds in an iterative process, applying soft attention to retrieve relevant information from a visual or textual KB, which is in turn accumulated into memory passed from one iteration to the next. However, in contrast to our model, it views the question as an atomic unit, whereas our model decomposes it into a multi-step action plan informing each cell in our sequential network about its current objective. Another key difference is the distant interaction between the question and the KB that characterizes our model. Conversely, DMN fuses their corresponding representations together into the same vector space. + +Graves et al. (2016; 2014) complements a neural network with a memory array it can interact with, through the means of soft attention. Analogously to our model, it partitions the model into a core neural network, called controller, as well as reading and writing heads that interact with external memory array. However, a main point distinguishing our model from this approach, is the use of dynamic memory, as in Kumar et al. (2016), instead of a fixed-array memory. Each MAC cell is associated with a memory state, our reading unit inspects only the latest memory passed from the previous state, and our writing unit creates a new memory state rather than writing to multiple slots in a fixed shared external memory. Notably, our approach is much more reminiscent of the widely successful RNN structure, rather than to Graves et al. (2016; 2014) . + +Finally, our approach has potential ties to the VQA models Hu et al. (2017); Lu et al. (2016) which also attend both the to question words and the image while progressively addressing the given question. However, both of these models have distinct specialized designs for each of their attention layers or modules, and have a discrete or fixed layout in which they are composed together. In contrast, our approach relax both of these limitations, having one universal cell design and one universal self-attending sequential network layout. + +# C.4 ATTENTION VS. CONVOLUTION + +Compared to other leading methods, our model stands out by being heavily based on soft attention, whereas most competing approaches are CNN-based, surprisingly lack any attention mechanism. Since attention is commonly used in models designed for standard VQA (Antol et al., 2015; Gupta, 2017; Lu et al., 2016; Yang et al., 2016), it is reasonable to assume that it would be beneficial to incorporate such methods into visual reasoning systems for the CLEVR task as well. In fact, attention mechanisms should be especially useful for multi-step reasoning questions such as those present in CLEVR. Such questions refer to several relations between different objects in the image and feature compositional structure that may be approached one step at a time. Thus, it should be beneficial for a cogent responder to have the capacity to selectively focus on on one or some objects at each step, traversing the relevant relational links one after the other, both at the image level, and at the question level. + +Moreover, attention mechanisms enhance our model’s ability to perform reasoning skills that pertain to aggregation of information across different regions, such as counting, finding maximum value, or performing other reduction operations over information that is spread across the image. Indeed, as discussed in 4, all existing models for visual reasoning, most of which lacking any attention mechanism, struggle with the counting and numerical comparisons questions present in CLEVR. Conversely, our model proves much more capable of performing these reasoning skills, outperforming the other approaches by a wide margin. Noticeably, incorporating soft attention into our model makes it much more adept at performing such aggregation reasoning skills, successfully addressing the this type of questions. + +Finally, as pointed out by Lu et al. (2016); Yang et al. (2016), soft attention confers the model with robustness to noise introduced from irrelevant information presents in the image, and higher capacity for handling larger and more diverse vocabulary, the latter being demonstrated in 4. It allows the model to separate the wheat from the chaff, selectively attending to the relevant information only, and arguably, being more resilient to both visual and linguistic variations. \ No newline at end of file diff --git a/parse/train/S1Euwz-Rb/S1Euwz-Rb_content_list.json b/parse/train/S1Euwz-Rb/S1Euwz-Rb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e87f73dd92898ff014b140a6a964de034759ecd5 --- /dev/null +++ b/parse/train/S1Euwz-Rb/S1Euwz-Rb_content_list.json @@ -0,0 +1,2230 @@ +[ + { + "type": "text", + "text": "COMPOSITIONAL ATTENTION NETWORKS FOR MACHINE REASONING ", + "text_level": 1, + "bbox": [ + 174, + 99, + 676, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Drew A. Hudson \nDepartment of Computer Science Stanford University \ndorarad@cs.stanford.edu ", + "bbox": [ + 183, + 170, + 411, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Christopher D. Manning Department of Computer Science Stanford University manning@cs.stanford.edu ", + "bbox": [ + 506, + 171, + 733, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 262, + 544, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Compositional Attention Networks, a novel fully differentiable neural network architecture, designed to facilitate explicit and expressive reasoning. While many types of neural networks are effective at learning and generalizing from massive quantities of data, this model moves away from monolithic blackbox architectures towards a design that provides a strong prior for iterative reasoning, allowing it to support explainable and structured learning, as well as generalization from a modest amount of data. The model builds on the great success of existing recurrent cells such as LSTMs: It sequences a single recurrent Memory, Attention, and Control (MAC) cell, and by careful design imposes structural constraints on the operation of each cell and the interactions between them, incorporating explicit control and soft attention mechanisms into their interfaces. We demonstrate the model’s strength and robustness on the challenging CLEVR dataset for visual reasoning, achieving a new state-of-the-art $9 8 . 9 \\%$ accuracy, halving the error rate of the previous best model. More importantly, we show that the new model is more computationally efficient and data-efficient, requiring an order of magnitude less time and/or data to achieve good results. ", + "bbox": [ + 233, + 296, + 764, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 550, + 334, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper considers how best to design neural networks to perform the iterative reasoning necessary for complex problem solving. Putting facts and observations together to arrive at conclusions is a central necessary ability as we work to move neural networks beyond their current great success with sensory perception tasks (LeCun et al., 1998; Krizhevsky et al., 2012) towards displaying Artificial General Intelligence. ", + "bbox": [ + 174, + 583, + 823, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Concretely, we develop a novel model that we apply to the CLEVR dataset (Johnson et al., 2016) for visual question answering (VQA). VQA (Antol et al., 2015; Gupta, 2017) is a challenging multimodal task that requires responding to natural language questions about images. However, Agrawal et al. (2016) show how the first generation of successful models on VQA tasks tend to acquire only superficial comprehension of both the image and the question, exploiting dataset biases rather than capturing a sound perception and reasoning process that would lead to the correct answer (Sturm, 2014). CLEVR was created to address this problem. As illustrated in figure 1, instances in the dataset consist of rendered images featuring 3D objects of several shapes, colors, materials and sizes, coupled with unbiased, compositional questions that require an array of challenging reasoning skills such as following transitive relations, counting objects and comparing their properties, without allowing any shortcuts around such reasoning. Notably, each instance in CLEVR is also accompanied by a tree-structured functional program that was both used to construct the question and reflects its reasoning procedure – a series of predefined operations – that can be composed together to answer it. ", + "bbox": [ + 174, + 660, + 612, + 881 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/674dcd95919d901cb7279f9b0552761c7620c6d163165031e5b1cd19087e50a7.jpg", + "image_caption": [ + "Figure 1: A sample image from the CLEVR dataset, with a question: “There is a purple cube behind a metal object left to a large ball; what material is it?” " + ], + "image_footnote": [], + "bbox": [ + 632, + 678, + 818, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 883, + 823, + 922 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most neural networks are essentially very large correlation engines that will hone in on any statistical, potentially spurious pattern that allows them to model the observed data more accurately. In contrast, we seek to create a model structure that requires combining sound inference steps to solve a problem instance. At the other extreme, some approaches adopt symbolic structures that resemble the expression trees of programming languages to perform reasoning (Andreas et al., 2016b; Hu et al., 2017). In particular, some approaches to CLEVR use the supplied functional programs for supervised or semi-supervised training (Andreas et al., 2016a; Johnson et al., 2017). Not only do we wish to avoid using such supervision in our work, but we in general suspect that the rigidity of these structures and the use of an inventory of operation-specific neural modules undermines robustness and generalization, and at any rate requires more complex reinforcement learning methods. ", + "bbox": [ + 174, + 103, + 825, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these weaknesses, while still seeking to use a sound and transparent underlying reasoning process, we propose Compositional Attention Networks, a novel, fully differentiable, non-modular architecture for reasoning tasks. Our model is a straightforward recurrent neural network with attention; the novelty lies in the use of a new Memory, Attention and Composition (MAC) cell. The constrained and deliberate design of the MAC cell was developed as a kind of strong structural prior that encourages the network to solve problems by stringing together a sequence of transparent reasoning steps. MAC cells are versatile but constrained neural units. They explicitly separate out memory from control, both represented recurrently. The unit contains three sub-units: The control unit updates the control representation based on outside instructions (for VQA, the question), learning to successively attend to different parts of the instructions; the read unit gets information out of a knowledge base (for VQA, the image) based on the control signal and the previous memory; the write unit updates the memory based on soft self-attention to previous memories, controlled by the retrieved information and the control signal. A universal MAC unit with a single set of parameters is used throughout the reasoning process, but its behavior can vary widely based on the context in which it is applied – the input to the control unit and the contents of the knowledge base. With attention, our MAC network has the capacity to represent arbitrarily complex acyclic reasoning graphs in a soft manner, while having physically sequential structure. The result is a continuous counterpart to module networks that can be trained end-to-end simply by backpropagation. ", + "bbox": [ + 174, + 250, + 825, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We test the behavior of our new network on CLEVR and its associated datasets. On the primary CLEVR reasoning task, we achieve an accuracy of $9 8 . 9 \\%$ , halving the error rate compared to the previous state-of-the-art FiLM model (Perez et al., 2017). In particular, we show that our architecture yields better performance on questions involving counting and aggregation. In supplementary studies, we show that the MAC network learns more quickly (both in terms of number of training epochs and training time) and more effectively from limited amounts of training data. Moreover, it also achieves a new state-of-the-art performance of $8 2 . 5 \\%$ on the more varied and difficult humanauthored questions of the CLEVR-Humans dataset. The careful design of our cell encourages compositionality, versatility and transparency. We achieve these properties by defining attention-based interfaces that constrict the cell’s input and output spaces, and so constrain the interactions both between and inside cells in order to guide them towards simple reasoning behaviors. Although each cell’s functionality has only a limited range of possible continuous reasoning behaviors, when chained together in a MAC network, the whole system becomes expressive and powerful. In the future, we believe that the architecture will also prove beneficial for other multi-step reasoning and inference tasks, for instance in machine comprehension and textual question answering. ", + "bbox": [ + 174, + 507, + 825, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 746, + 344, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There have been several prominent models that address the CLEVR task. By and large they can be partitioned into two groups: module networks, which in practice have all used the strong supervision provided in the form of tree-structured functional programs that accompany each data instance, and large, relatively unstructured end-to-end differentiable networks that complement a fairly standard stack of CNNs with components that aid in performing reasoning tasks. In contrast to modular approaches (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017), our model does not require additional supervision and makes use of a single computational cell chained in sequence (like an LSTM) rather than a collection of custom modules deployed in a rigid tree structure. In contrast to augmented CNN approaches (Santoro et al., 2017; Perez et al., 2017), we suggest that our approach provides an ability for relational reasoning with better generalization capacity and higher computational efficiency. These approaches and other related work are discussed and contrasted in more detail in the supplementary material in section C. ", + "bbox": [ + 174, + 784, + 825, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 823, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 COMPOSITIONAL ATTENTION NETWORKS ", + "text_level": 1, + "bbox": [ + 174, + 151, + 553, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compositional Attention Networks is an end-to-end architecture for question-answering tasks that sequentially performs an explicit reasoning process by stringing together small building blocks, called MAC cells, each is responsible for performing one reasoning step. ", + "bbox": [ + 174, + 183, + 825, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now provide an overview of the model, and a detailed discussion of the MAC cell. The model is composed of three components: an Input unit, the core MAC network, and an output unit. A TensorFlow implementation of the network, along with pretrained models will be made publicly available. ", + "bbox": [ + 174, + 231, + 823, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper we explore the model in the context of VQA. However, it should be noted that while the input and output units are naturally domain-specific and should be designed to fit the task at hand, the MAC network has been designed to be generic and more broadly applicable, and may prove useful in contexts beyond those explored in the paper, such as machine comprehension or question answering over knowledge bases, which in our belief is a promising avenue for future work. ", + "bbox": [ + 174, + 294, + 825, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 THE INPUT UNIT ", + "text_level": 1, + "bbox": [ + 176, + 381, + 333, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The input unit processes the raw inputs given to the system into distributed vector representations. It receives a text question (or in general, a query), and an image (or in general, a Knowledge Base (KB)) and processes each of them with a matching sub-unit, for the query and the KB, here a biLSTM and a CNN. More details can be found in the supplementary material, section A. ", + "bbox": [ + 174, + 406, + 825, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At the end of this stage, we get from the query sub-unit a series of biLSTM output states, which we refer to as contextual words, $[ c w _ { 1 } , . . . , c w _ { S } ]$ , where $S$ is the length of the question. In addition, we get $q = [ \\overleftrightarrow { \\lambda \\omega _ { 1 } } , \\overrightarrow { c w _ { S } } ]$ , the concatenation of the hidden states from the backward and forward LSTM passes. We refer to $q$ as the question representation. Furthermore, we get from the Knowledge-Base sub-unit a static representation of the knowledge base. For the case of VQA, it will be represented by a continuous matrix $K B _ { V }$ of dimension $H , W , d$ , where $H = W = 1 4$ are the height and width of the transformed image, corresponding to each of its regions. ", + "bbox": [ + 174, + 469, + 825, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 THE MAC CELL ", + "text_level": 1, + "bbox": [ + 176, + 583, + 326, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The MAC network, which is the heart of our model, chains a sequence of small building blocks, called MAC cells, each responsible for performing one reasoning step. The model is provided access to a Knowledge Base (KB), which is, for the specific case of VQA, the given image, and then upon receiving a query, i.e. a question, the model iteratively focuses, in $p$ steps, on the query’s various parts, each reflects in turn the current reasoning step, which we term the control. Consequently, guided by this control, it retrieves the relevant information from the KB, that is then passed to the next cell in a recurrent fashion. ", + "bbox": [ + 174, + 609, + 825, + 707 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Drawing inspiration from the Model-View-Controller paradigm used in software design and from the commonly exercised separation between control and data paths in computer architecture, the MAC cell is composed of three units: control unit, read unit and write unit. Each has a clearly defined role and an interface through which it interacts with the other units. See figure 2. ", + "bbox": [ + 176, + 714, + 823, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The careful design and imposed interfaces that constrain the interaction between the units inside the MAC cell, as described below, serve as structural prior that limits the space of hypotheses it can learn, thereby guiding it towards acquiring the intended reasoning behaviors. As such, this prior facilitates the learning process and mitigate overfitting issues. ", + "bbox": [ + 174, + 777, + 823, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In particular, and similar in spirit to Perez et al. (2017), we allow the question to interact with the Knowledge Base – the image for the case of VQA, only through indirect means: by guiding the cell to attend to different elements in the KB, as well as controlling its operation through gating mechanisms. Thus, in both cases, the interaction between these mediums, visual and textual, or knowledge and query, is mediated through probability distributions, either in the form of attention maps, or as gates, further detailed below. This stands in stark contrast to many common approaches that fuse the question and image together into the same vector space through linear combinations, multiplication, or concatenation. Rather, our controlled interaction distills the influence that the query should have in processing the Knowledge Base, casting it onto discrete probability distributions instead. ", + "bbox": [ + 174, + 840, + 825, + 922 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/21ef6959e8f5c66e79f26b2aa043aef1bf62c1ad19d7a86079e5194974b914be.jpg", + "image_caption": [ + "Figure 2: The MAC cell, which is a recurrent unit comprised of a Control Unit, Read Unit, and Write Unit. Blue shows the control flow and red shows the memory flow. See section 3.2 for details. " + ], + "image_footnote": [], + "bbox": [ + 351, + 103, + 604, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 318, + 823, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The MAC cell has been designed to replace the discrete and predefined “modules” used in the modular approach (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017). Rather, we create one universal and versatile cell that is applied across all the reasoning steps, sharing both its architecture as well as its parameters, across all of its instantiations. In contrast to the discrete modules, each trained to specialize to some specific elementary reasoning task, the MAC cell is capable of demonstrating a continuous range of possible reasoning behaviors conditioned on the context in which it is applied – namely, the inputs it receives from the prior cell. ", + "bbox": [ + 174, + 366, + 825, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each cell $M A C _ { i }$ maintains two dual states: control $c _ { i }$ and memory $m _ { i }$ , both are continuous vectors of dimension $d$ . The control $c _ { i }$ represents the reasoning operation the MAC cell should accomplish in the current step – focusing only on some aspect of the whole question. This is represented by a weighted-average attention-based sum of the question words. The memory $m _ { i }$ represents the current context information deemed relevant to respond to the query, or answer the question.This is represented practically by a weighted average over elements from the KB, or for the case of VQA, regions in the image. $m _ { 0 }$ and $c _ { 0 }$ are initialized each to a random vector parameter of dimension $d$ . The memory and control states are passed from one cell to the next in a recurrent fashion, and used in a way reminiscent of Key-Value memory networks (Miller et al., 2016), as discussed below. ", + "bbox": [ + 174, + 470, + 825, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 THE CONTROL UNIT ", + "text_level": 1, + "bbox": [ + 176, + 611, + 372, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The control unit determines the reasoning operation that should be applied at this step. It receives the contextual words $[ c w _ { 1 } , . . . , c w _ { S } ]$ , the question representation $q$ , and the control state from the previous MAC cell $c _ { i - 1 }$ , all of which are vectors of dimension $d$ . ", + "bbox": [ + 174, + 635, + 825, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We would like to allow our MAC cell to perform continuously varied and adaptive range of behaviors, as demanded by the question. Therefore, we define the behavior of each cell to be a function of the contextual words $[ c w _ { 1 } , . . . , c w _ { S } ]$ , weighted-averaged according to the attention distribution that the control unit produces at each step. This will allow the cell to adapt its behavior – the reasoning operation it performs – to the question it receives, instead of having a fixed set of predefined behaviours as is the case in competing approaches Andreas et al. (2016a;b); Johnson et al. (2017). ", + "bbox": [ + 174, + 684, + 825, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The formal specification of the control unit is shown in figure 3. The question $q$ is linearly transformed into a vector $q _ { i }$ of the same dimension, which in turn is concatenated with the previous control state $c _ { i - 1 }$ and linearly transformed again to a $d$ -dimensional vector $c q _ { i }$ . ", + "bbox": [ + 174, + 775, + 825, + 818 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/629b6ac32640eb52fb3021fd3b45ba817455f27666f7d51a163756755ee964e5.jpg", + "text": "$$\n\\begin{array} { c } { q _ { i } = W _ { i } ^ { d , d } \\cdot q + b _ { i } ^ { d } } \\\\ { c q _ { i } = W ^ { 2 d , d } \\left[ q _ { i } , c _ { i - 1 } \\right] + b ^ { d } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 408, + 839, + 591, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that in contrast to all other parameters of the cell, which are shared across its instantiations at the different steps $i = 1 , . . . , p$ , the parameters $W _ { i } ^ { d , d }$ and $b _ { i } ^ { d }$ are different for each iteration. This is done to allow each cell to attend more readily to different aspects (i.e. parts) of the questions, depending on the index of the current step – its relative stage in the context of the whole reasoning process. ", + "bbox": [ + 173, + 892, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/432535676eeb69a82d7848d82a8a327d906d019f07db47ee24440c80b3beab34.jpg", + "image_caption": [ + "Figure 3: The Control Unit (CU) of the MAC cell. See section 3.2.1 for details. Best viewed in color. " + ], + "image_footnote": [], + "bbox": [ + 330, + 102, + 663, + 194 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 263, + 825, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$c q _ { i }$ represents the current reasoning operation we would like to perform in a continuous way, taking into account both the overall meaning of the question $q _ { i }$ , as well as the words the model attended to in the previous step, $c _ { i - 1 }$ . ", + "bbox": [ + 174, + 313, + 825, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, we would like to prevent the cell from diverging in the reasoning operations it tries to perform, and instead anchor it back in the question words, by using them to represent the reasoning operation of the current step. We can achieve that by computing an attention distribution $c v _ { i }$ over the contextual words $[ c w _ { 1 } , . . . , c w _ { S } ]$ based on their similarity to $c q _ { i }$ . Then, summing the contextual words according to the attention distribution $c v _ { i }$ will allow us to have a new control state, $c _ { i }$ , which is represented again in terms of words from the question. Intuitively, it is the gist of the question that is relevant to the reasoning operation we would like to perform in the current step. ", + "bbox": [ + 173, + 361, + 825, + 459 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/d43f433dbdaf70d728243b740937cc4b7c74be9d6cda3d636612e146da07d190.jpg", + "text": "$$\n\\begin{array} { c l c r } { { c v _ { i , s } = \\displaystyle \\mathrm { s o f t m a x } \\big ( W ^ { d , 1 } \\big ( c q _ { i } \\circ c w _ { s } \\big ) + b ^ { 1 } \\big ) } } \\\\ { { \\big . } } \\\\ { { c _ { i } = \\displaystyle \\sum _ { s = 1 } ^ { S } c v _ { i , s } \\cdot c w _ { s } } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 364, + 486, + 632, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the control unit returns the current control state $c _ { i }$ , along with an attention map $c v _ { i }$ over the contextual words. ", + "bbox": [ + 173, + 565, + 823, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 THE READ UNIT ", + "text_level": 1, + "bbox": [ + 176, + 609, + 344, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Read Unit is provided with access to the knowledge base $K B _ { V }$ , along with the previous memory state $m _ { i - 1 }$ and the current control $c _ { i }$ . It is responsible for retrieving relevant content from the Knowledge Base $K B _ { V }$ for the reasoning task that the MAC cell should accomplish at this step, which is represented by the current control state $c _ { i }$ , as explained above. Figure 4 shows a diagram. ", + "bbox": [ + 174, + 635, + 825, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The relevance of the new information is judged in two stages by the “relatedness” of each element in the KB (or for the case of VQA, each region in the image) to either the memory $m _ { i - 1 }$ that has accumulated relevant information from previous iterations, or to the current control $c _ { i }$ , pointing towards the next piece of information that should be taken into account. Here, relatedness is measured by trained linear transformations comparing each element to the previous memory and the current control. ", + "bbox": [ + 174, + 696, + 825, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More formally, at the first stage, the interaction between each element $K B _ { h , w }$ , where $\\begin{array} { r l } { h } & { { } = } \\end{array}$ $1 , . . . , H , w = 1 , . . . , W$ , and the previous memory $m _ { i - 1 }$ is computed by: ", + "bbox": [ + 173, + 787, + 823, + 818 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/92d034ab00623a6201883d47cd297723d599fe90d135f474e1198ca4a4622a1a.jpg", + "text": "$$\nm _ { i - 1 } ^ { \\prime } = W ^ { d , d } \\cdot m _ { i - 1 } + b ^ { d }\n$$", + "text_format": "latex", + "bbox": [ + 408, + 835, + 589, + 856 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/ba795ad9270013570f22ffe5663537aa1ee717d610072223ec7afd2128e47f00.jpg", + "text": "$$\n\\begin{array} { r l } { \\boldsymbol { K } \\boldsymbol { B } _ { h , w } ^ { \\prime } = \\boldsymbol { W } ^ { d , d } \\cdot \\boldsymbol { K } \\boldsymbol { B } _ { h , w } + \\boldsymbol { b } ^ { d } } & { { } } \\\\ { \\big ( \\boldsymbol { I } _ { m - K B } \\big ) _ { h , w } = \\boldsymbol { m } _ { i - 1 } ^ { \\prime } \\circ \\boldsymbol { K } \\boldsymbol { B } _ { h , w } ^ { \\prime } } & { { } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 392, + 878, + 604, + 924 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a538d781a85f8151d41cc83e31d872618ae93e0b206ef08dd299f4020e8135bf.jpg", + "image_caption": [ + "Figure 4: The Read Unit (RU) diagram. Blue refers to control flow, purple to knowledge flow and red to memory flow. See section 3.2.2 for description. " + ], + "image_footnote": [], + "bbox": [ + 326, + 102, + 668, + 194 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These memory-KB interactions measure the relatedness of each element in the KB to the memory accumulated so far, which holds information that has been deemed relevant to handle previous reasoning steps towards addressing the question. They allow the model to perform transitive inference, retrieving a new piece of information that now seems important in light of the recent memory retrieved in a prior iteration. ", + "bbox": [ + 174, + 261, + 823, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, there are cases which necessitate the model to temporarily ignore current memories, when choosing the new information to retrieve. Logical OR is a classical example: when the model has to look at two different objects at the same time, and assuming it stored one of them at the first iteration, it should briefly ignore it, considering new information that is relevant to the question but is unrelated to the memory. In order to achieve such capability, the read unit concatenates the original KB elements to each corresponding memory-KB interaction, which are then projected back to $d$ -dimensional space (equation 6a): ", + "bbox": [ + 174, + 338, + 825, + 436 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/ada1fba84ad0bac03e8e9a8f8cc72cbc8bba767e7414ee8b9b65dab9dee93066.jpg", + "text": "$$\n\\begin{array} { c } { { I _ { m - K B } , ^ { \\prime } = W ^ { 2 d , d } \\left[ I _ { m - K B } , K B _ { h , w } \\right] + b ^ { d } } } \\\\ { { I _ { c m - K B } = c _ { i } \\circ \\left( I _ { m - K B } \\right) ^ { \\prime } } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 357, + 458, + 640, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At the second stage, the read unit compares the current $c _ { i }$ with these memory-KB interactions, in order to focus on the information that is relevant to the current reasoning operation that the MAC cell seeks to accomplish. The result is then passed to a softmax layer yielding an attention map $m v _ { i }$ over the KB, which is used in turn to retrieve the relevant information to perform the current reasoning step. ", + "bbox": [ + 174, + 511, + 823, + 580 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/1af356ca707634550da691f9e1c5befb2540a27633c9abfa602dcd4cdf1d0ea5.jpg", + "text": "$$\n\\begin{array} { r } { m v _ { i } = \\operatorname { s o f t m a x } \\left( W ^ { d , d } \\cdot I _ { c m - K B } + b ^ { d } \\right) } \\\\ { m _ { n e w } = \\displaystyle \\sum _ { h , w = 1 , 1 } ^ { H , W } \\left( m v _ { i } \\right) _ { h , w } \\cdot K B _ { h , w } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 366, + 603, + 632, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, the read unit returns the newly retrieved information $m _ { n e w }$ , along with an attention map $m v _ { i }$ over the Knowledge Base. ", + "bbox": [ + 171, + 683, + 823, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To give an example of the read unit operation, assume a given question $q$ such as “What object is located left to the blue ball?”, whose associated answer is “cube”. Initially, no cue is provided to the model to attend to that cube, since no direct information about it presents in the question. Instead, based on its comprehension of the question, the model may start by focusing on the blue ball at the first iteration, such that the memory state $m _ { 1 }$ will capture the blue ball. However, in the second iteration, the control unit, after re-examining the question, may realize it should now look left, storing the word “left” in $c _ { 2 }$ . Then, when considering both $m _ { 1 }$ and $c _ { 2 }$ , the read unit will realize it should perform a reasoning operation corresponding to the word “left” (stored in $c _ { 2 }$ ) given a memory representing the blue ball in $m _ { 1 }$ , thereby allowing it to look left to the blue ball and find the cube. ", + "bbox": [ + 173, + 717, + 825, + 856 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2.3 THE WRITE UNIT ", + "text_level": 1, + "bbox": [ + 176, + 871, + 351, + 886 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Write Unit is responsible for creating the new memory state $m _ { i }$ that will reflect all the information considered to be important to answer the question so far, i.e. up to the current iteration in the ", + "bbox": [ + 174, + 895, + 823, + 924 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da3df6ec5443731e5ef4aaf8424acd7af4458f638b9bc19f84b7d1ed88b9267e.jpg", + "image_caption": [ + "Figure 5: The Write Unit (WU) diagram. Blue refers to control flow and red to memory flow. See section 3.2.3 for description. " + ], + "image_footnote": [], + "bbox": [ + 367, + 102, + 632, + 198 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "reasoning process. It receives the last memory state $m _ { i - 1 }$ from the previous MAC cell, along with the newly retrieved information from the read unit in the current iteration, $m _ { n e w }$ . See figure 5 for a diagram. ", + "bbox": [ + 174, + 263, + 825, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the main design we have explored, merging the new information with the previous memory state is done simply by a linear transformation. ", + "bbox": [ + 173, + 313, + 823, + 342 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/fae487b9caae60fb93edbf054820528c6b18061cf0a3ab7e592c6d4fbd281b4a.jpg", + "text": "$$\nm _ { i } ^ { \\prime } = W ^ { 2 d , d } [ m _ { n e w } , m _ { i - 1 } ] + b ^ { d }\n$$", + "text_format": "latex", + "bbox": [ + 392, + 359, + 606, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, we have explored two variations of this design. The first, self-attention, allows considering any previous memories rather than just the last one $m _ { i - 1 }$ , thus providing the network with the capacity to model non-sequential reasoning processes. The second variation is adding gating mechanisms to the writing unit. These may allow the model to dynamically adjust the practical length of the computation to the question complexity and stabilize the memory content throughout the sequential network (similarly to GRUs and LSTMs). ", + "bbox": [ + 173, + 386, + 825, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Self-Attention. The current architecture that we have presented allows the model to perform reasoning steps in a sequence, passing control and memory states from one cell to the following. However, we would like to grant the system with more flexibility. Particularly, we would like to allow it to capture more complicated reasoning processes such as trees and graphs - Directed Acyclic Graph (DAG) in particular, where several branches of reasoning sub-processes are merged together in later stages. Indeed, the CLEVR dataset includes cases where the questions embody tree-like reasoning process, rather than just sequences, which we would like to address correctly in our model. ", + "bbox": [ + 173, + 484, + 825, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We achieve that by adding self-attention connections between each MAC cell and all the prior cells. Since each cell can look on all the prior reasoning steps and their corresponding memories retrieved from the Knowledge Base, it can virtually capture any directed acyclic graph, while still having physically sequential layout. ", + "bbox": [ + 174, + 589, + 825, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "More formally, the current MAC cell, of the $i ^ { t h }$ iteration, is granted with access to $c _ { 1 } , . . . , c _ { i - 1 }$ along with the corresponding $m _ { 1 } , . . . , m _ { i - 1 }$ , that have been computed by the prior MAC cells. It begins by computing the similarity between $c _ { i }$ and $c _ { 1 } , . . . , c _ { i - 1 }$ , and use it to derive an attention map over the prior MAC cells $s a _ { i , j }$ for $j = 1 , . . . , i - 1$ . This represents the relevance of the $j ^ { t h }$ prior reasoning step to the current one $i$ (equation 9a). ", + "bbox": [ + 174, + 650, + 825, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then, we average the previous memories according to this resulted attention map $s a _ { i j }$ . We obtain $m _ { s a }$ , representing the information from all the other reasoning steps that is relevant to the current one (equation 9b). ", + "bbox": [ + 176, + 728, + 821, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This resembles the approach of Key-Value networks (Miller et al., 2016). The similarity between control states, corresponding to the reasoning operations that are performed in each prior step, allows the model to select which memories should be taken into account, when creating the new memory – namely, which branches of the reasoning process should be merged together at this point. ", + "bbox": [ + 174, + 777, + 825, + 834 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/b2917add3d0ec9deb18d84cf0919d7c999da2714964aba59d6c8b9129667c984.jpg", + "text": "$$\n\\begin{array} { c } { { s a _ { i j } = \\displaystyle \\mathrm { s o f t m a x } \\left( W ^ { d , 1 } ( c _ { i } \\circ c _ { j } ) + b ^ { 1 } \\right) } } \\\\ { { ( m _ { s a } ) _ { i } = \\displaystyle \\sum _ { j = 1 } ^ { i - 1 } s a _ { i j } \\cdot m _ { j } } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 375, + 856, + 622, + 922 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we use $m _ { s a }$ along with $m _ { i } ^ { \\prime }$ to compute $m _ { i } ^ { \\prime \\prime }$ , the new memory content in this variation. ", + "bbox": [ + 169, + 102, + 785, + 119 + ], + "page_idx": 7 + }, + { + "type": "equation", + "img_path": "images/5e36c1389b375bd38c94672bc787841067c61a64ebb4c06b118a35c4f749cf9f.jpg", + "text": "$$\nm _ { i } ^ { \\prime \\prime } = W ^ { 2 d , d } [ m _ { n e w } , m _ { i } ^ { \\prime } ] + b ^ { d }\n$$", + "text_format": "latex", + "bbox": [ + 398, + 141, + 599, + 160 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Memory Gate. The currently presented MAC network has some fixed number $p$ of concatenated MAC cells, representing the length of the overall reasoning process we perform. However, not all questions require reasoning sequence of the same length. Some questions are simpler while others more complex. ", + "bbox": [ + 173, + 176, + 825, + 233 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motivated by this observation, we add a gate over the new memory computed at each step, that may selectively keep content of the previous memory $m _ { i - 1 }$ unchanged. Practically, the gate functions in a similar way to a highway network (Srivastava et al., 2015), where the gate value is conditioned on the current reasoning operation, $c _ { i }$ . ", + "bbox": [ + 174, + 239, + 823, + 296 + ], + "page_idx": 7 + }, + { + "type": "equation", + "img_path": "images/06b7c9f2a66fd66b565bd6bf64cfa368a00407f9ea2545d0e729a32d2b2c1b1d.jpg", + "text": "$$\n\\begin{array} { c } { { { c _ { i } } ^ { \\prime } = W ^ { d , d } \\cdot c _ { i } + b ^ { d } } } \\\\ { { m _ { i } = \\mathrm { s i g m o i d } \\left( c _ { i } ^ { \\prime } \\right) \\cdot m _ { i - 1 } + ( 1 - \\mathrm { s i g m o i d } \\left( c _ { i } ^ { \\prime } \\right) ) \\cdot m _ { i } ^ { \\prime \\prime } } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 316, + 324, + 681, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The write unit returns the new memory state $m _ { i }$ , that will be passed along with $c _ { i }$ to the next MAC cell. ", + "bbox": [ + 174, + 381, + 823, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.2.4 DISCUSSION ", + "text_level": 1, + "bbox": [ + 174, + 429, + 316, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, when designing the MAC cell, we have attempted to formulate the inner workings of an elementary, yet generic reasoning skills: the model decomposes the problem into steps, focusing on one at a time. At each such step, it takes into account: ", + "bbox": [ + 176, + 454, + 823, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "• The control $c _ { i }$ : Some aspect of the task - pointing to the future work that has left to be done. \n• The previous memory or memories: The partial solution or evidence the cell has acquired so far – pointing to the past work that has already been achieved. \n• The newly retrieved information $m _ { n e w }$ : that is retrieved from the knowledge base $K B$ and may or may not be transitively related to that partial solution or evidence - the present, or current work. ", + "bbox": [ + 215, + 510, + 825, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Considering these three sources of information together, the cell finally adds the new information up into its working memory, $m _ { i }$ , progressing one more step towards the final answer. ", + "bbox": [ + 173, + 637, + 823, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 THE OUTPUT UNIT ", + "text_level": 1, + "bbox": [ + 176, + 686, + 348, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The output unit receives the question representation $q$ , along with the memory state passed from the last MAC cell $m _ { p }$ , where $p$ is the number of MAC cells in the network – representing the number of reasoning steps in the whole process. It inspects both and predicts an answer based on their concatenation. Intuitively, we would like our model to consider both the question as well as the relevant information that has been progressively retrieved from the KB, deemed the necessary information to answer it. ", + "bbox": [ + 173, + 713, + 825, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that considering both $q$ and $m _ { p }$ is critical to answer the question. While $m _ { p }$ represents the information collected from KB, we still need to recall what has been asked about it to be able to answer accordingly. This is especially true in our case, when all other interactions between the question and the KB are mediated through attention distributions, rather than being transformed into a shared continuous vector space. ", + "bbox": [ + 174, + 804, + 825, + 875 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The prediction is built out of a standard 2-layers fully-connected softmax-based classifier with hidden dimension $d$ and output dimension that matches the number of possible answers in the dataset. The classifier receives $[ m _ { p } , q ]$ as input and returns a probability distribution over the answers. ", + "bbox": [ + 176, + 881, + 823, + 924 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a8f3f0904dbfbf06b3ae89f9cd0dcd15557eabccb6cf593d1481a4565628f79b.jpg", + "table_caption": [ + "Table 1: CLEVR accuracy by baseline methods, competing methods, and our method (MAC) - for our method we show test results with GLoVE initialization and validation results with and without GLoVE. (\\*) denotes use of extra supervisory information through program labels. $( ^ { \\dagger } )$ denotes use of data augmentation. $( ^ { \\ddagger } )$ denotes training from raw pixels. $( ^ { \\circ } )$ denotes GLoVE word vector initialization. " + ], + "table_footnote": [], + "table_body": "
ModelOverallCountExistCompare NumbersQuery AttributeCompare Attribute
Human (Johnson et al., 2017)92.686.796.686.595.096.0
Q-type baseline (Johnson et al., 2017)41.834.650.251.036.051.3
LSTM (Johnson et al., 2017)46.841.761.169.836.851.8
CNN+LSTM (Johnson et al., 2017)52.343.765.267.149.353.0
CNN+LSTM+SA (Johnson et al., 2016)76.664.482.777.482.675.4
N2NMN* (Hu et al. 2017)83.768.585.784.990.088.7
PG+EE (9K prog.)* (Johnson et al., 2017)88.679.789.779.192.696.0
PG+EE (70oK prog.)* (Johnson et al., 2017)96.992.797.198.798.198.9
CNN+LSTM+RN†‡ (Santoro et al.,2017)95.590.197.893.697.997.1
CNN+GRU+FiLM (Perez et al.,2017)97.794.399.196.899.199.1
CNN+GRU+FiLM* (Perez et al., 2017)97.694.399.393.499.399.3
CNN+GRU+FiLM (Perez et al., 2017)97.794.399.196.899.199.1
MAC (this paper)-val98.997.199.599.399.299.2
MAC-val99.097.299.599.599.599.4
MAC-test98.997.299.599.499.399.5
", + "bbox": [ + 220, + 181, + 777, + 367 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 174, + 415, + 326, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate our model on the recent CLEVR dataset (Johnson et al., 2016). CLEVR is a synthetic dataset consisting of 700K tuples; each consists of a 3D-rendered image featuring objects of various shapes, colors, materials and sizes, coupled with compositional multi-step questions that measure performance on an array of challenging reasoning skills such as following transitive relations, counting objects and comparing their properties. In addition, each question is associated with a formal program, specifying the reasoning operations that should be performed to compute the answer, among 28 possibilities. ", + "bbox": [ + 174, + 463, + 825, + 561 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We first perform experiments on the original 700k CLEVR dataset (Johnson et al., 2016), comparing to prior work. As shown in table 1, our model matches or outperforms all existing models both in overall accuracy, as well as in each category, testing different reasoning skills. In particular, for the overall performance, we achieve $9 8 . 9 4 \\%$ accuracy, more than halving the error rate of the prior best model, FiLM (Perez et al., 2017). ", + "bbox": [ + 174, + 568, + 825, + 638 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Counting and Numerical Comparison. Remarkably, our performance on questions testing counting and numerical comparisons is significantly higher than the competing models, which consistently struggle on this question type. Again, we nearly halve the corresponding error rate. These results demonstrate the aptitude of attention mechanisms to perform counting, reduction and aggregation, in contrast to alternative, CNN-based approaches. ", + "bbox": [ + 174, + 645, + 823, + 715 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training Length and Computational-Efficiency. We examine the learning curves of our and competing models. We have trained all models on the same architecture and used the author code for the other models. Aiming at having equal settings for comparison, we ran all models including ours with learned random words vectors. In order to make sure the results are statistically significant we ran each model multiple (10) times, and plotted the averages and confidence intervals (figure 4). The results show that our model learns significantly faster than the other leading methods, FiLM (Perez et al., 2017) and $\\mathrm { P G + E E }$ (Johnson et al., 2017). While we do not have learning curves for the Relational Network model, Santoro et al. (2017) report approximately 1.4 million iterations to achieve $9 5 . 5 \\%$ accuracy, which are equivalent to 125 epochs approximately, whereas our model achieves a comparable accuracy after 3 epochs only, yielding 40x reduction in the length of the training process. ", + "bbox": [ + 174, + 722, + 825, + 861 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Naturally, the smaller number of required training steps also translates to comparably shorter training time. Perez et al. (2017) report training time of 4 days, equivalent to 80 epochs, to reach accuracy of $9 7 . 7 \\%$ . In contrast, we achieve higher accuracy in 6 epochs, taking 9.5 hours overall, leading to $1 0 \\mathrm { x }$ reduction in training time. ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c74201d70aae1d3587e14044064872d252c0e00ec6d3c4fd6b70377bb7088c75.jpg", + "image_caption": [ + "Figure 6: Training curves and accuracies for MACs (our model), FiLM (Perez et al., 2017), $\\mathrm { P G + E E }$ (Johnson et al., 2017) and stacked-attention (Yang et al., 2016; Johnson et al., 2017). (Note: $\\mathrm { P G + E E }$ uses the supported CLEVR programs as strong supervision.) Left: Training curve (accuracy/epoch). Right: Learning curve: Accuracy for $1 \\%$ , $2 \\%$ , $5 \\%$ , $10 \\%$ , $2 5 \\%$ , $50 \\%$ and $100 \\%$ of the $7 0 0 \\mathrm { k }$ CLEVR samples. " + ], + "image_footnote": [], + "bbox": [ + 294, + 108, + 699, + 261 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/20d0a9d88276e853b2d4290d832c31f6ed7a6d4ae845cb8a1f4754688e965048.jpg", + "table_caption": [ + "Table 2: Accuracy on CLEVR-Humans of previous methods and our method (MAC), before (left) and after (right) fine-tuning on the CLEVR-Humans training data. $\\mathrm { P G + E E }$ uses supervised data. For our method (MAC) we show test results with GLoVE initialization and validation results with and without GLoVE. " + ], + "table_footnote": [], + "table_body": "
ModelTrain CLEVRTrain CLEVR + fine-tune HUMANS
LSTM (Johnson et al.,2017)27.536.5
CNN+LSTM (Johnson et al.,2017)37.743.2
CNN+LSTM+SA+MLP (Johnson et al.,2016)50.457.6
PG+EE (18K prog.)* (Johnson et al.,2017)54.066.6
CNN+GRU+FiLM (Perez et al., 2017)56.675.9
MAC-val56.981.2
MAC-val58.582.0
MAC-test58.682.5
", + "bbox": [ + 295, + 462, + 700, + 565 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 DATA EFFICIENCY ", + "text_level": 1, + "bbox": [ + 174, + 619, + 341, + 633 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We have explored the performance of our and other leading approaches on smaller subsets of the CLEVR dataset, in order to study the ability of models to generalize from smaller amount of data. We sampled at random subsets of CLEVR, with $10 \\%$ , $2 5 \\%$ and $50 \\%$ of its original $7 0 0 \\mathrm { k }$ size, and used them to train our and other 3 proposed models for the CLEVR task: FiLM (Perez et al., 2017), the strongly-supervised $\\mathrm { P G + E E }$ (Johnson et al., 2017), and stacked-attention networks (Johnson et al., 2017; Yang et al., 2016). ", + "bbox": [ + 174, + 659, + 825, + 742 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As shown in figure 4, our model outperforms the other models by a wide margin for all subsets of the CLEVR dataset. For $50 \\%$ of the data, equivalent to $3 5 0 \\mathrm { k }$ samples, other models obtain accuracies ranging between $70 \\%$ and $92 \\%$ , while our model achieves $9 7 . 9 \\%$ . The gap becomes larger as the dataset size reduces: for $2 5 \\%$ of the data, equivalent to $1 7 5 \\mathrm { k }$ samples, performance of other models is between $50 \\%$ and $7 7 \\%$ , while our model maintains a high $9 5 . 4 \\%$ accuracy. ", + "bbox": [ + 173, + 750, + 825, + 819 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Finally, for $10 \\%$ of the data $- 7 0 \\mathrm { k }$ samples, still a sizeable amount – our model is the only one that manages to generalize, with performance of $8 4 . 7 \\%$ on average, whereas the other three models fail, achieving $4 7 . 6 \\% - 5 7 . 5 \\%$ . Note that as pointed out by (Johnson et al., 2016) a simple baseline that predicts the most frequent answer for each of the question types achieves already $4 2 . 1 \\%$ , suggesting that answering half of the questions correctly means that the competing models barely learn to generalize from the smaller dataset. These results demonstrate the robustness of our architecture and its key role as a structural prior guiding our network to learn the intended reasoning skills. ", + "bbox": [ + 173, + 825, + 825, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 CLEVR HUMANS - NATURAL LANGUAGE QUESTIONS ", + "bbox": [ + 176, + 104, + 593, + 117 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We analyze our model performance on the CLEVR-Humans dataset (Johnson et al., 2017), consisting of natural language questions collected through crowdsourcing. As such, the dataset has diverse vocabulary and linguistic variations, and it also demands more varied reasoning skills. ", + "bbox": [ + 174, + 131, + 825, + 174 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Since the training set is relatively small, consisting of $1 8 \\mathrm { k }$ samples, we use it to finetune a model pretrained on the standard CLEVR dataset. However, since most of the vocabulary in CLEVRHumans is not covered by CLEVR, we do not train the word vectors during the pre-training stage, so to prevent drift in their meaning compared to other uncovered words in CLEVR-Humans that may be semantically related. ", + "bbox": [ + 174, + 180, + 825, + 251 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As shown in table 2, our model achieves state-of-the-art performance on CLEVR-Humans both before and after fine-tuning. It surpasses the next-best FiLM model, (Perez et al., 2017) by $6 . 6 \\%$ percent, achieving $8 2 . 5 \\%$ . ", + "bbox": [ + 174, + 257, + 825, + 299 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The results substantiate the model’s robustness against linguistic variations and noise, as well as its ability to adapt to diverse vocabulary and varied reasoning skills. Arguably, the soft attention performed over the question words allows the model to focus on the words that are most critical to answer the question and translate them to corresponding reasoning operations, giving less attention to irrelevant linguistic variations. ", + "bbox": [ + 174, + 306, + 825, + 377 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 ABLATIONS ", + "text_level": 1, + "bbox": [ + 174, + 400, + 297, + 412 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d542dee22ea6b6971ea171ea3647286d90bc0976dc1e0a42885a1582ff9227cc.jpg", + "image_caption": [ + "Figure 7: Ablations and variations of the MAC network. See text for full detail. " + ], + "image_footnote": [], + "bbox": [ + 176, + 439, + 818, + 768 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Based on the validation set, we have conducted an ablation study on our model to understand better the contribution of each of its component to the overall performance. We tested each setting on the standard 700K CLEVR dataset as well as on $10 \\%$ subset of the dataset. See table 3 for the numerical results. In addition, figure 4.3 presents the training curves for the different settings trained on the standard dataset. Overall, the results demonstrate the robustness of the model to hyperparameter variations such as network dimension and length, and also the impact of different aspect and components of MAC on its performance. ", + "bbox": [ + 173, + 825, + 825, + 924 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/cbfa912b869921d24b8997ef84de01c88a55b56dcf5407ed4090070e65e15576.jpg", + "table_caption": [ + "Table 3: Ablations Results for the full standard CLEVR dataset and $10 \\%$ subset of it. See text for full details. " + ], + "table_footnote": [], + "table_body": "
ModelStandard CLEVR10% CLEVR
MAC standard model (dimension 512)98.984.5
dimension 25698.476.3
dimension 12897.677.0
w/o control55.651.5
w/o word-attention95.363.2
full-question-as-control80.765.0
w/o positional encoding99.180.0
linear position encoding99.083.2
identity writing unit98.584.5
only-gating writing unit99.383.1
+ self-attention83.299.2
+ gating99.483.1
+ self-attention and gating99.585.5
gate bias 098.784.9
gate bias 199.468.5
gate bias -199.077.1
unshared weights97.867.5
prediction without question information97.864.7
", + "bbox": [ + 305, + 140, + 691, + 334 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Network Length. We have tested the model performance as a function of the network’s length – the number of MAC cells that were sequenced together. The results show the positive correlation between the network length and its performance. We can see that for 1 cell the scores are relatively $1 0 \\mathrm { w } - 7 5 \\%$ , but adding at least one more cell leads to a significant increase in performance above $9 5 \\%$ . The performance keeps improving up to lengths 8-16 that achieve $9 8 . 9 – 9 9 . 1 \\%$ . The results also teach us about the complexity of the dataset, by showing the relatively significant benefits of having at least 4 cells, each modeling a reasoning step. ", + "bbox": [ + 173, + 381, + 825, + 479 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Network Dimension. We have varied the state dimension to check the robustness of the model to hyperparameters. The results on the standard CLEVR dataset show the model is able to maintain high performance with dimension of 128, albeit after a longer training process, achieving $9 7 . 6 \\%$ , compared to $9 8 . 9 4 \\%$ achieved with dimension of 512. However, for $10 \\%$ of CLEVR, the larger 512-dimension allows accuracy increase by $7 . 5 \\%$ over dimension of 128. ", + "bbox": [ + 174, + 486, + 825, + 555 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Weight Sharing. We have tested the impact of sharing weights between cell has on the model performance for network of length $p = 1 2$ . The results show that for the standard dataset there is only a small difference between these settings of $1 \\%$ . However, for less data, we see much more significant drop $1 6 . 9 \\%$ in the unshared-parameters setting compared to the shared one. Indeed, we observe that a model with less parameter is more data-efficient and has a lower tendency to overfit the data. ", + "bbox": [ + 174, + 561, + 825, + 645 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Control Unit. We have performed several ablations in the control unit to understand its contribution to the overall model performance. Based on the results, first, we can see the the question information is crucial for the model to handle the questions, as can be noted by the low performance of the model when there is no use of control signal whatsoever. Second, we have tested the model performance when using the continuous control state computed by question (2) in section 3.2.1, without having word-attention, in order to understand its relative contribution. Based on the results, we can indeed see that using word-attention is useful for accelerating the training process and achieving higher accuracies both for the standard dataset as well as for the small subset, where using word-attention increases results in $2 1 . 4 \\%$ . We also see that using the “contextual words” produced by the questionunit LSTM is useful in accelerating the model performance, when compared to using the wordvectors directly. ", + "bbox": [ + 173, + 652, + 825, + 806 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Reading Unit. We have conducted several ablations for the reading unit to better understand its behavior and contribution to the performance of the model. The standard MAC reading unit uses the control state – which averages the question words based on attention distributions computed per each reasoning step. In this ablation experiment, we have tested using the full question representation $q$ instead across all reasoning steps to gain better understanding of the the contribution of wordattention to the model performance. Indeed, we can see that using $q$ rather then the control state $c _ { i }$ results in a significant drops in performance $- 1 9 . 4 \\%$ for the full CLEVR dataset and $1 9 . 5 \\%$ for $10 \\%$ of the data. ", + "bbox": [ + 173, + 811, + 825, + 924 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We have conducted additional ablation experiment to better understand the contribution of using the KB features directly in the first-stage information retrieval process described in section 3.2.2, compared to using only the dot-products of the KB elements with the previous memory state $m _ { i - 1 }$ . For the full CLEVR dataset, we can see that this component has only a small impact in the final performance - ultimately resulting in $0 . 0 6 \\%$ performance difference. However, for the $10 \\%$ of the data, we can see that the difference in performance when ablating this component is much larger - $1 1 . 2 \\%$ . ", + "bbox": [ + 174, + 103, + 825, + 200 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Writing Unit Ablations. In our main MAC model variant, the memory unit merges the new information $m _ { n e w }$ with the previous memory state $m _ { i - 1 }$ by combining them through a linear transformation. In this experiment, we have explored other variations, such as assigning $m _ { n e w }$ to $m _ { i }$ directly – ignoring previous memories, or doing a linear transformation based on $m _ { n e w }$ only. The results show that in fact such variant is only slightly worse than our main variant – $0 . 4 \\%$ . We also conducted an experiment in which we merge the new information with the previous memory just by a having a gate that does a weighted average of them. The results show that this variant performs equivalently to our standard linear-transformation variant. ", + "bbox": [ + 174, + 208, + 823, + 319 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Writing Unit Additions. We have explored the impact of the writing unit variants described in section 3.2.3 – adding self-attention, gating mechanisms, or both, compared to our standard main model that uses a linear transformation to merge the newly retrieved information $m _ { n e w }$ with the previous memory $m _ { i }$ . For the complete CLEVR dataset we can see that indeed both these variants are very helpful in increasing the model performance. Compared to our standard MAC model that achieves $9 8 . 9 4 \\%$ on the validation set, self-attention yields accuracy of $9 9 . 2 3 \\%$ , gating yields $9 9 . 3 6 \\%$ and adding both achieves $9 9 . 4 8 \\%$ . ", + "bbox": [ + 174, + 325, + 823, + 424 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Output Unit. In our standard model, the final predictions made in the output unit are based on the final memory state $m _ { p }$ as well as question representation $q$ (stands for the final hidden states of the backward and forwards passes of the LSTM). We have explored the contribution of basing the model prediction on the latter, by testing the model performance when prediction is based on memory alone, for the complete and $10 \\%$ datasets. We can see that in both settings basing the model’s predictions on the question representation allows faster training and higher accuracies. Notable is the gap in performance for the $10 \\%$ CLEVR - $1 9 . 8 \\%$ increase by using the question representation to make predictions. These results are very reasonable intuitively, since the model is structured such that the memory holds only information that was retrieved from the image. Thus, questions that may ask for instance on different aspects (such as color or shape) of the same object in the image may result in the same memory content, which is thus does not directly contain enough information to respond such questions. ", + "bbox": [ + 174, + 431, + 825, + 597 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Position. In our standard model, similarly to the practice of competing models (Santoro et al., 2017; Perez et al., 2017; Hu et al., 2017), we have concatenated positional information to each region of the image, in order to increase the model capability to perform spatial reasoning. We have explored both simple linear maps at a constant $[ - 1 , 1 ]$ as well as more complex positional encoding suggested by (Vaswani et al., 2017). However, the results for both the standard dataset and the $10 \\%$ version show a very negligible improvement at best when adding positional encoding information, demonstrating the capability of MAC to perform spatial reasoning without data augmentation. ", + "bbox": [ + 173, + 604, + 825, + 702 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Gate Bias Initialization. For our model variant with gating mechanism (described in section 3.2.3) we have tested the effect of setting different values for the gate bias - $- 1 , 0$ and 1. for $- 1$ the model is initialized to biased for keeping the previous memory value whereas for 1 it will be biased for using the new memory instead. We can see that for the complete dataset setting the bias to 1 is optimal – apparently since the model has enough data to learn to apply each cell effectively. In contrast, for the small $10 \\%$ CLEVR data, setting the bias to 0 shows better performance, biasing the model to using less cells overall which results ultimately in a theoretically-simpler model that can fit less data more effectively. ", + "bbox": [ + 174, + 708, + 825, + 820 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 INTERPRETABILITY ", + "text_level": 1, + "bbox": [ + 176, + 851, + 348, + 864 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We have looked into attention maps over the image and question that the model produces during its computation and provide a few examples in figure 4.4. The first example shows us how the model parses the question in steps, first focusing on the main entity that the question is about, then on ", + "bbox": [ + 176, + 882, + 823, + 924 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "relation of this entity to the “brown matte thing” which is then located in the image. Finally, the model correctly focuses on the small brown cube and predicts the right answer – brown. ", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The second example shows a model with 4 cells instead of 6, that similarly parse the question in iterations and focuses on the relevant objects at each step, though we can see that the reasoning process looks somewhat different when the MAC network has fewer cells. ", + "bbox": [ + 176, + 138, + 823, + 180 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The last example shows how how the model handles counting and OR operations. It starts from identifying the task - computing a number, and then red objects as well as the cylinder, one at a time, allowing it ultimately to respond correctly, with the answer 2. ", + "bbox": [ + 174, + 188, + 825, + 229 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e280158a489b95ddeabb4461dda86460910408860df91a493fe23a53648a08b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 273, + 674, + 736 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Figure 8: Three examples of attention maps for the image and the question. See text for full details. ", + "bbox": [ + 171, + 779, + 823, + 795 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 823, + 318, + 839 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have given a first demonstration of how a sequence of Memory, Attention and Control (MAC) cells combined into a Compositional Attention Network provides a very effective tool for neural reasoning. In future work, we wish to explore this promising architecture for other tasks and domains, including real-world VQA, machine comprehension and textual question answering. ", + "bbox": [ + 174, + 854, + 825, + 911 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 102, + 287, + 118 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question answering models. arXiv preprint arXiv:1606.07356, 2016. ", + "bbox": [ + 173, + 126, + 823, + 155 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 39–48, 2016a. ", + "bbox": [ + 173, + 161, + 823, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Learning to compose neural networks for question answering. arXiv preprint arXiv:1601.01705, 2016b. ", + "bbox": [ + 173, + 213, + 823, + 243 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International Conference on Computer Vision, pp. 2425–2433, 2015. ", + "bbox": [ + 174, + 251, + 823, + 294 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. ", + "bbox": [ + 173, + 301, + 823, + 330 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. ", + "bbox": [ + 173, + 338, + 823, + 367 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Djork-Arne Clevert, Thomas Unterthiner, and Sepp Hochreiter. Fast and accurate deep network ´ learning by exponential linear units (elus). arXiv preprint arXiv:1511.07289, 2015. ", + "bbox": [ + 171, + 376, + 823, + 405 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Alex Graves, Navdeep Jaitly, and Abdel-rahman Mohamed. Hybrid speech recognition with deep bidirectional lstm. In Automatic Speech Recognition and Understanding (ASRU), 2013 IEEE Workshop on, pp. 273–278. IEEE, 2013. ", + "bbox": [ + 174, + 412, + 823, + 455 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint arXiv:1410.5401, 2014. ", + "bbox": [ + 171, + 463, + 825, + 493 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka GrabskaBarwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, ´ et al. Hybrid computing using a neural network with dynamic external memory. Nature, 538 (7626):471–476, 2016. ", + "bbox": [ + 173, + 501, + 825, + 558 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Akshay Kumar Gupta. Survey of visual question answering: Datasets and techniques. arXiv preprint arXiv:1705.03865, 2017. ", + "bbox": [ + 169, + 565, + 823, + 594 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016. ", + "bbox": [ + 173, + 603, + 821, + 646 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8): 1735–1780, 1997. ", + "bbox": [ + 168, + 654, + 823, + 683 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Ronghang Hu, Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Kate Saenko. Learning to reason: End-to-end module networks for visual question answering. arXiv preprint arXiv:1704.05526, 2017. ", + "bbox": [ + 173, + 690, + 825, + 733 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456, 2015. ", + "bbox": [ + 174, + 742, + 823, + 785 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. arXiv preprint arXiv:1612.06890, 2016. ", + "bbox": [ + 173, + 792, + 823, + 837 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Judy Hoffman, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Inferring and executing programs for visual reasoning. arXiv preprint arXiv:1705.03633, 2017. ", + "bbox": [ + 171, + 844, + 821, + 887 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. ", + "bbox": [ + 171, + 895, + 821, + 924 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012. ", + "bbox": [ + 174, + 103, + 823, + 146 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Ankit Kumar, Ozan Irsoy, Peter Ondruska, Mohit Iyyer, James Bradbury, Ishaan Gulrajani, Victor Zhong, Romain Paulus, and Richard Socher. Ask me anything: Dynamic memory networks for natural language processing. In International Conference on Machine Learning, pp. 1378–1387, 2016. ", + "bbox": [ + 174, + 155, + 825, + 210 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´ document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. ", + "bbox": [ + 174, + 219, + 821, + 250 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention for visual question answering. In Advances In Neural Information Processing Systems, pp. 289– 297, 2016. ", + "bbox": [ + 174, + 257, + 826, + 300 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Alexander Miller, Adam Fisch, Jesse Dodge, Amir-Hossein Karimi, Antoine Bordes, and Jason Weston. Key-value memory networks for directly reading documents. arXiv preprint arXiv:1606.03126, 2016. ", + "bbox": [ + 174, + 309, + 826, + 352 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Jeffrey Pennington, Richard Socher, and Christopher Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pp. 1532–1543, 2014. ", + "bbox": [ + 173, + 361, + 826, + 404 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. arXiv preprint arXiv:1709.07871, 2017. ", + "bbox": [ + 173, + 412, + 825, + 443 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Adam Santoro, David Raposo, David GT Barrett, Mateusz Malinowski, Razvan Pascanu, Peter Battaglia, and Timothy Lillicrap. A simple neural network module for relational reasoning. arXiv preprint arXiv:1706.01427, 2017. ", + "bbox": [ + 173, + 450, + 823, + 493 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Rupesh Kumar Srivastava, Klaus Greff, and Jurgen Schmidhuber. Highway networks. ¨ arXiv preprint arXiv:1505.00387, 2015. ", + "bbox": [ + 169, + 501, + 825, + 531 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Bob L Sturm. A simple method to determine if a music information retrieval system is a horse. IEEE Transactions on Multimedia, 16(6):1636–1644, 2014. ", + "bbox": [ + 169, + 540, + 823, + 569 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017. ", + "bbox": [ + 173, + 577, + 825, + 621 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and textual question answering. In International Conference on Machine Learning, pp. 2397–2406, 2016. ", + "bbox": [ + 174, + 628, + 823, + 671 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks for image question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 21–29, 2016. ", + "bbox": [ + 174, + 680, + 825, + 723 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "SUPPLEMENTARY MATERIAL ", + "bbox": [ + 176, + 102, + 416, + 118 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "A DETAILS OF INPUT UNIT ", + "text_level": 1, + "bbox": [ + 176, + 138, + 416, + 156 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "The input unit processes the raw inputs given to the system into distributed vector representations. It receives a text question (or in general, a query), and an image (or in general, a Knowledge Base (KB)) and processes each of them with a matching sub-unit. Here we provide details of the Query Unit and the Image Unit used in this work. ", + "bbox": [ + 174, + 174, + 825, + 231 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "A.0.1 THE QUERY UNIT ", + "text_level": 1, + "bbox": [ + 176, + 252, + 357, + 267 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "We encode a query of $S$ words into a continuous representation using a bidirectional LSTM (Hochreiter & Schmidhuber, 1997; Graves et al., 2013). Each word is associated with a word embedding $w _ { s }$ , where $s = 1 , . . . , S$ . In our case, we use GloVE words embeddings (Pennington et al., 2014). Then, these embeddings are processed by a bidirectional LST M of dimension $d$ that outputs: ", + "bbox": [ + 174, + 279, + 825, + 335 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "• a matching sequence of $d$ -dimensional output states, which we refer to as contextual words, $[ c w _ { 1 } , . . . , c w _ { S } ]$ • $d$ -dimensional hidden state $q = [ \\overleftrightarrow { \\lambda \\omega _ { 1 } } , \\overrightarrow { c w _ { S } } ]$ , the concatenation of the hidden states from the backward and forward passes. We refer to $q$ as the question representation. ", + "bbox": [ + 217, + 352, + 825, + 420 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Intuitively, each contextual word $c w _ { s }$ represents the meaning of $s ^ { t h }$ word, in the context of the question, while the hidden state $q$ represents the overall (compositional) meaning of the question. ", + "bbox": [ + 173, + 436, + 823, + 467 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "A.0.2 THE IMAGE UNIT ", + "text_level": 1, + "bbox": [ + 176, + 487, + 354, + 502 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Given an image, and following prior work on CLEVR (?Santoro et al., 2017; Perez et al., 2017), we extract conv4 features from ResNet101 (He et al., 2016) pretrained on ImageNet (Krizhevsky et al., 2012) which we treat as a fixed initial representation of the image, $x$ of dimension $H , W , C$ where $H = W = 1 4$ are the height and width of the transformed image and $C = 1 0 2 4$ is the number of channels. Each feature $x _ { h , w }$ represents one region in the original image. ", + "bbox": [ + 174, + 513, + 825, + 584 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Similar to prior work (Hu et al., 2017; Santoro et al., 2017; Perez et al., 2017), we would like to allow our model to reason explicitly about spatial locations, as required by many of the questions in CLEVR, and therefore we concatenate to this representation a spatial map that represents each of the positions in the image. However, in contrast to prior work that uses a linear meshgrid feature map with 2 features $h$ and $w$ ranging from $- 1$ to 1, and to allow better representation of the positions, we use the positional encoding scheme proposed by Vaswani et al. (2017): ", + "bbox": [ + 173, + 590, + 825, + 675 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/23f67e68bd8fe769889cb65ef6b26ad32ea8c028817ac3af21e7f4e098381e73.jpg", + "text": "$$\n\\begin{array} { l } { p _ { ( h , 2 i ) } = \\sin { \\Big ( h / 1 0 0 0 0 ^ { 2 i / p _ { d } } \\Big ) } } \\\\ { p _ { ( h , 2 i + 1 ) } = \\cos { \\Big ( h / 1 0 0 0 0 ^ { 2 i / p _ { d } } \\Big ) } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 388, + 707, + 607, + 766 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "And similarly for $w$ , where $p _ { d }$ is a hyperparameter. Overall, the positional encoding of a feature at position $( h , w )$ is $[ p _ { h } , p _ { w } ]$ , the concatenation of the positional encodings for $h$ and $w$ . ", + "bbox": [ + 174, + 784, + 823, + 813 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "This positional encoding scheme allows better correspondence between the distance of 2 positions $( x , y )$ and $( x , y )$ in the image and a vector similarity of their positional encodings, even when $p _ { d }$ is larger than two. ", + "bbox": [ + 174, + 818, + 825, + 861 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "We then concatenate the obtained spatial map with $x$ , receiving a spatially-aware image representation $x _ { p }$ . Then, we pass this representation through two CNN layers with $d$ output channels and obtain a final representation of the image, which we refer to as our Visual Knowledge Base $( K B _ { V }$ that is used in further components of the model. ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "B IMPLEMENTATION AND TRAINING DETAILS ", + "text_level": 1, + "bbox": [ + 174, + 102, + 568, + 118 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "For the question processing, we use GloVE (Pennington et al., 2014) word-vectors with dimension 300. For the image processing, we extract conv4 features from ResNet101 (He et al., 2016) pretrained on ImageNet (Krizhevsky et al., 2012), with dimension $H , W , C$ where $H = W = 1 4$ and $C = 1 0 2 4$ , followed by 2 CNN layers with kernel size 2. We use MAC network with $p = 1 2$ cells, and train it using Adam (Kingma & Ba, 2014), with learning rate $1 0 ^ { - 4 }$ . We train our model for $1 0 - 2 0$ epochs, with batch size 64, and use early stopping based on validation accuracies. During training, the moving averages of all weights of the model are maintained with the exponential decay rate of 0.999. At test time, the moving averages instead of the raw weights are used. We use dropout 0.85, and ELU (Clevert et al., 2015) which in our experience has reduce the training process compared to RELU.The training process takes roughly 10-20 hours on a single Titan X GPU. ", + "bbox": [ + 173, + 135, + 825, + 273 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "C FURTHER DISCUSSION OF RELATED WORK ", + "text_level": 1, + "bbox": [ + 174, + 297, + 558, + 313 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "In this section we provide detailed discussion of related work. Several models have been applied to the CLEVR task. These can be partitioned into two groups, module networks that use the strong supervision provided as a tree-structured functional program associated with each instance, and end-to-end, fully differentiable networks that combine a fairly standard stack of CNNs with components that aid them in performing reasoning tasks. We also discuss the relation of MAC to other approaches, such as memory networks and neural computers. ", + "bbox": [ + 174, + 329, + 825, + 414 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "C.1 MODULE NETWORKS ", + "text_level": 1, + "bbox": [ + 176, + 433, + 366, + 446 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "The modular approach (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017) first translates the given question into a tree-structured action plan, aiming to imitate the ground-truth programs provided as a form of strong-supervision. Then, it constructs a tailor-made network that executes the plan on the image in multiple steps. This network is composed of discrete units selected out of a collection of predefined modules, each responsible for an elementary reasoning operation, such as identifying an objects color, filtering them for their shape, or comparing two amounts. Each module has its own set of learned parameters (Johnson et al., 2017), or even hand-crafted design (Andreas et al., 2016a) to guide it towards its intended behavior. ", + "bbox": [ + 174, + 459, + 825, + 570 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Overall, this approach makes discrete choices at two levels: the identity of each module – the behavior it should learn among a fixed set of possible types of behaviors, and the network layout – the way in which these modules are wired together to compute the answer progressively. Hence, their differentiability is confined to the boundaries of a single module, disallowing end-to-end training. ", + "bbox": [ + 174, + 578, + 825, + 633 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Several key differences exist between our approaches. First, our model replaces the fixed modules collection with one versatile and universal cell that shares both its architecture and parameters across all of its instantiations, and is applied across all the reasoning steps. Second, it replaces the dynamic recursive tree structures with a sequential topology, augmented by soft attention mechanisms, as done in Bahdanau et al. (2014). This confers our network with a virtual capacity to represent arbitrarily complex Directed Acyclic Graphs (DAGs) while still having efficient and readily deployed physical sequential structure. Together, both of these relaxations allow us to effectively train our model end-to-end by backpropagation alone, whereas module networks demand a more involved training scheme that relies on the strongly-supervised programs at the first stage, and on various Reinforcement Learning (RL) techniques at the second. Furthermore, while our model can be train without the strong supervisory programs, developing adaptive reasoning skills to address the task is it trained for, the modular approach reliance on questions structured and formal representation hinder its applicability to real-world tasks. ", + "bbox": [ + 174, + 641, + 825, + 820 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "C.2 AUGMENTED CONVOLUTIONAL NEURAL NETWORKS ", + "text_level": 1, + "bbox": [ + 173, + 840, + 586, + 854 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Alternative approaches for the CLEVR task that do not rely on the provided programs as a strong supervision signal are Santoro et al. (2017) and Perez et al. (2017). Both complement standard multi-layer Convolutional Neural Networks (CNNs) with components that aid them in handling compositional and relational questions. ", + "bbox": [ + 176, + 867, + 823, + 922 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Relational Networks. Santoro et al. (2017) appends a Relation Network (RN) layer to the CNN. This layer inspects all pairs of pixels in the image, thereby enhancing the network capacity to reason over binary relations between objects. While this approach is very simple and elegant conceptually, it suffers from quadratic computational complexity, in contrast to our and other leading approaches. But beyond that, closer inspection reveals that this direct pairwise comparison might be unnecessary. Based on the analogy suggested by Santoro et al. (2017), according to which pixels are equivalent to objects and their pairwise interactions to relations, a RN layer attempts to grasp the induced graph between objects all at once in one shallow and broad layer. Conversely, our attention-based model proceeds in steps. It basically compares the image to its current memory and control for this step, aggregates the attended regions into the new memory, and repeats the process. By the same analogy, it traverses a narrow and deep path, progressively following transitive relations. Consequently, our model exhibits a relational capacity while circumventing the computational inefficiency. ", + "bbox": [ + 174, + 103, + 825, + 270 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "FiLM. FiLM (Perez et al., 2017) is a recently proposed method that interleaves standard CNN layers that process the given image with linear layers, reminiscent of layer normalization techniques (Ba et al., 2016; Ioffe & Szegedy, 2015). Each of these layers, called FiLM, is conditioned on the question: the question words are processed by a GRU, and its output is linearly transformed into matching biases and variances for each of the CNN layers, tilting its activations to reflect the specifics of the given question and affect the computation done over the image. ", + "bbox": [ + 174, + 276, + 825, + 361 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Similarly to our model, this approach features distant modulation between the question and the image, where rather than being fused together into the same vector space, the question can affect the image processing only through constrained means – for the case of FiLM – linear transformations. However, since the same transformation is applied to all the activations homogeneously, agnostic to both their spatial location as well as the features values, this approach does not allow the question to differentiate between regions in the image based on the objects or concepts they represent – on the content of the image. This stands in stark contrast to our attention-based model, which readily allows and actually encourages the question to inform the model about relevant regions to focus on. We speculate that this still distant, yet more direct interaction between the question and the data, or image, for the case of VQA, facilitates learning and increases generalizability. It may be more suitable to VQA tasks, and CLEVR in particular, where the questions demand the responder to focus on specific objects, and reason about their properties or relations, rather than respond based only on a holistic view of the image that may lead to sub-optimal results (Yang et al., 2016), as is the case of FiLM. Indeed, as demonstrated in 4, there is significant evidence showing our models better generalization capacity, allowing it to achieve high accuracies much faster, and from less data than FiLM and other competing methods. ", + "bbox": [ + 174, + 367, + 825, + 590 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "C.3 MEMORY AND ATTENTION ", + "text_level": 1, + "bbox": [ + 176, + 631, + 401, + 645 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Our architecture draws inspiration from recent research on memory and attention (Kumar et al., 2016; Xiong et al., 2016; Graves et al., 2014; 2016). Kumar et al. (2016); Xiong et al. (2016) propose the Dynamic Memory Network model that proceeds in an iterative process, applying soft attention to retrieve relevant information from a visual or textual KB, which is in turn accumulated into memory passed from one iteration to the next. However, in contrast to our model, it views the question as an atomic unit, whereas our model decomposes it into a multi-step action plan informing each cell in our sequential network about its current objective. Another key difference is the distant interaction between the question and the KB that characterizes our model. Conversely, DMN fuses their corresponding representations together into the same vector space. ", + "bbox": [ + 174, + 666, + 825, + 791 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Graves et al. (2016; 2014) complements a neural network with a memory array it can interact with, through the means of soft attention. Analogously to our model, it partitions the model into a core neural network, called controller, as well as reading and writing heads that interact with external memory array. However, a main point distinguishing our model from this approach, is the use of dynamic memory, as in Kumar et al. (2016), instead of a fixed-array memory. Each MAC cell is associated with a memory state, our reading unit inspects only the latest memory passed from the previous state, and our writing unit creates a new memory state rather than writing to multiple slots in a fixed shared external memory. Notably, our approach is much more reminiscent of the widely successful RNN structure, rather than to Graves et al. (2016; 2014) . ", + "bbox": [ + 174, + 799, + 825, + 922 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Finally, our approach has potential ties to the VQA models Hu et al. (2017); Lu et al. (2016) which also attend both the to question words and the image while progressively addressing the given question. However, both of these models have distinct specialized designs for each of their attention layers or modules, and have a discrete or fixed layout in which they are composed together. In contrast, our approach relax both of these limitations, having one universal cell design and one universal self-attending sequential network layout. ", + "bbox": [ + 174, + 103, + 825, + 188 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "C.4 ATTENTION VS. CONVOLUTION ", + "text_level": 1, + "bbox": [ + 176, + 204, + 434, + 218 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Compared to other leading methods, our model stands out by being heavily based on soft attention, whereas most competing approaches are CNN-based, surprisingly lack any attention mechanism. Since attention is commonly used in models designed for standard VQA (Antol et al., 2015; Gupta, 2017; Lu et al., 2016; Yang et al., 2016), it is reasonable to assume that it would be beneficial to incorporate such methods into visual reasoning systems for the CLEVR task as well. In fact, attention mechanisms should be especially useful for multi-step reasoning questions such as those present in CLEVR. Such questions refer to several relations between different objects in the image and feature compositional structure that may be approached one step at a time. Thus, it should be beneficial for a cogent responder to have the capacity to selectively focus on on one or some objects at each step, traversing the relevant relational links one after the other, both at the image level, and at the question level. ", + "bbox": [ + 174, + 231, + 825, + 382 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Moreover, attention mechanisms enhance our model’s ability to perform reasoning skills that pertain to aggregation of information across different regions, such as counting, finding maximum value, or performing other reduction operations over information that is spread across the image. Indeed, as discussed in 4, all existing models for visual reasoning, most of which lacking any attention mechanism, struggle with the counting and numerical comparisons questions present in CLEVR. Conversely, our model proves much more capable of performing these reasoning skills, outperforming the other approaches by a wide margin. Noticeably, incorporating soft attention into our model makes it much more adept at performing such aggregation reasoning skills, successfully addressing the this type of questions. ", + "bbox": [ + 174, + 390, + 825, + 515 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Finally, as pointed out by Lu et al. (2016); Yang et al. (2016), soft attention confers the model with robustness to noise introduced from irrelevant information presents in the image, and higher capacity for handling larger and more diverse vocabulary, the latter being demonstrated in 4. It allows the model to separate the wheat from the chaff, selectively attending to the relevant information only, and arguably, being more resilient to both visual and linguistic variations. ", + "bbox": [ + 174, + 522, + 823, + 592 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/parse/train/S1Euwz-Rb/S1Euwz-Rb_middle.json b/parse/train/S1Euwz-Rb/S1Euwz-Rb_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..0371da7034dfa9847daf7265338607c90b27db4c --- /dev/null +++ b/parse/train/S1Euwz-Rb/S1Euwz-Rb_middle.json @@ -0,0 +1,54670 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 79, + 414, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 415, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 415, + 97 + ], + "score": 1.0, + "content": "COMPOSITIONAL ATTENTION NETWORKS", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 309, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 309, + 117 + ], + "score": 1.0, + "content": "FOR MACHINE REASONING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 135, + 252, + 179 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 186, + 146 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 186, + 146 + ], + "score": 1.0, + "content": "Drew A. Hudson", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 145, + 249, + 159 + ], + "spans": [ + { + "bbox": [ + 111, + 145, + 249, + 159 + ], + "score": 1.0, + "content": "Department of Computer Science", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 112, + 156, + 194, + 169 + ], + "spans": [ + { + "bbox": [ + 112, + 156, + 194, + 169 + ], + "score": 1.0, + "content": "Stanford University", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 169, + 252, + 180 + ], + "spans": [ + { + "bbox": [ + 113, + 169, + 252, + 180 + ], + "score": 1.0, + "content": "dorarad@cs.stanford.edu", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 310, + 136, + 449, + 180 + ], + "lines": [ + { + "bbox": [ + 310, + 134, + 418, + 148 + ], + "spans": [ + { + "bbox": [ + 310, + 134, + 418, + 148 + ], + "score": 1.0, + "content": "Christopher D. Manning", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 309, + 145, + 447, + 159 + ], + "spans": [ + { + "bbox": [ + 309, + 145, + 447, + 159 + ], + "score": 1.0, + "content": "Department of Computer Science", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 309, + 156, + 392, + 170 + ], + "spans": [ + { + "bbox": [ + 309, + 156, + 392, + 170 + ], + "score": 1.0, + "content": "Stanford University", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 309, + 169, + 450, + 180 + ], + "spans": [ + { + "bbox": [ + 309, + 169, + 450, + 180 + ], + "score": 1.0, + "content": "manning@cs.stanford.edu", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 278, + 208, + 333, + 220 + ], + "lines": [ + { + "bbox": [ + 276, + 207, + 336, + 222 + ], + "spans": [ + { + "bbox": [ + 276, + 207, + 336, + 222 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 143, + 235, + 468, + 411 + ], + "lines": [ + { + "bbox": [ + 142, + 235, + 469, + 248 + ], + "spans": [ + { + "bbox": [ + 142, + 235, + 469, + 248 + ], + "score": 1.0, + "content": "We present Compositional Attention Networks, a novel fully differentiable neu-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 245, + 469, + 260 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 469, + 260 + ], + "score": 1.0, + "content": "ral network architecture, designed to facilitate explicit and expressive reasoning.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 256, + 470, + 271 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 470, + 271 + ], + "score": 1.0, + "content": "While many types of neural networks are effective at learning and generalizing", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 268, + 469, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 469, + 280 + ], + "score": 1.0, + "content": "from massive quantities of data, this model moves away from monolithic black-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 279, + 469, + 291 + ], + "spans": [ + { + "bbox": [ + 141, + 279, + 469, + 291 + ], + "score": 1.0, + "content": "box architectures towards a design that provides a strong prior for iterative rea-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 291, + 469, + 302 + ], + "spans": [ + { + "bbox": [ + 142, + 291, + 469, + 302 + ], + "score": 1.0, + "content": "soning, allowing it to support explainable and structured learning, as well as", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 301, + 470, + 313 + ], + "spans": [ + { + "bbox": [ + 141, + 301, + 470, + 313 + ], + "score": 1.0, + "content": "generalization from a modest amount of data. The model builds on the great", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 312, + 469, + 324 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 469, + 324 + ], + "score": 1.0, + "content": "success of existing recurrent cells such as LSTMs: It sequences a single recur-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 323, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 469, + 335 + ], + "score": 1.0, + "content": "rent Memory, Attention, and Control (MAC) cell, and by careful design imposes", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 334, + 469, + 346 + ], + "spans": [ + { + "bbox": [ + 141, + 334, + 469, + 346 + ], + "score": 1.0, + "content": "structural constraints on the operation of each cell and the interactions between", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 345, + 469, + 357 + ], + "spans": [ + { + "bbox": [ + 142, + 345, + 469, + 357 + ], + "score": 1.0, + "content": "them, incorporating explicit control and soft attention mechanisms into their in-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 354, + 470, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 354, + 470, + 369 + ], + "score": 1.0, + "content": "terfaces. We demonstrate the model’s strength and robustness on the challenging", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 142, + 366, + 469, + 379 + ], + "spans": [ + { + "bbox": [ + 142, + 366, + 417, + 379 + ], + "score": 1.0, + "content": "CLEVR dataset for visual reasoning, achieving a new state-of-the-art", + "type": "text" + }, + { + "bbox": [ + 417, + 367, + 444, + 377 + ], + "score": 0.87, + "content": "9 8 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 366, + 469, + 379 + ], + "score": 1.0, + "content": "accu-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 378, + 469, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 378, + 469, + 390 + ], + "score": 1.0, + "content": "racy, halving the error rate of the previous best model. More importantly, we show", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 387, + 470, + 402 + ], + "spans": [ + { + "bbox": [ + 141, + 387, + 470, + 402 + ], + "score": 1.0, + "content": "that the new model is more computationally efficient and data-efficient, requiring", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 400, + 414, + 412 + ], + "spans": [ + { + "bbox": [ + 141, + 400, + 414, + 412 + ], + "score": 1.0, + "content": "an order of magnitude less time and/or data to achieve good results.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 108, + 436, + 205, + 448 + ], + "lines": [ + { + "bbox": [ + 105, + 435, + 208, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 208, + 452 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 504, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "score": 1.0, + "content": "This paper considers how best to design neural networks to perform the iterative reasoning necessary", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 474, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 506, + 486 + ], + "score": 1.0, + "content": "for complex problem solving. Putting facts and observations together to arrive at conclusions is a", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "central necessary ability as we work to move neural networks beyond their current great success with", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "sensory perception tasks (LeCun et al., 1998; Krizhevsky et al., 2012) towards displaying Artificial", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 506, + 192, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 192, + 520 + ], + "score": 1.0, + "content": "General Intelligence.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 375, + 698 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 376, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 376, + 535 + ], + "score": 1.0, + "content": "Concretely, we develop a novel model that we apply to the CLEVR", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 533, + 376, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 376, + 547 + ], + "score": 1.0, + "content": "dataset (Johnson et al., 2016) for visual question answering (VQA).", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 545, + 376, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 376, + 558 + ], + "score": 1.0, + "content": "VQA (Antol et al., 2015; Gupta, 2017) is a challenging multimodal", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 556, + 376, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 376, + 569 + ], + "score": 1.0, + "content": "task that requires responding to natural language questions about", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 567, + 376, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 376, + 580 + ], + "score": 1.0, + "content": "images. However, Agrawal et al. (2016) show how the first gen-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 577, + 376, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 376, + 591 + ], + "score": 1.0, + "content": "eration of successful models on VQA tasks tend to acquire only", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 589, + 377, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 377, + 602 + ], + "score": 1.0, + "content": "superficial comprehension of both the image and the question, ex-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 600, + 377, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 377, + 613 + ], + "score": 1.0, + "content": "ploiting dataset biases rather than capturing a sound perception and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 612, + 376, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 376, + 623 + ], + "score": 1.0, + "content": "reasoning process that would lead to the correct answer (Sturm,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 621, + 376, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 376, + 633 + ], + "score": 1.0, + "content": "2014). CLEVR was created to address this problem. As illus-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 633, + 376, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 376, + 645 + ], + "score": 1.0, + "content": "trated in figure 1, instances in the dataset consist of rendered im-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 644, + 376, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 376, + 656 + ], + "score": 1.0, + "content": "ages featuring 3D objects of several shapes, colors, materials and", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 655, + 376, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 376, + 667 + ], + "score": 1.0, + "content": "sizes, coupled with unbiased, compositional questions that require", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 666, + 376, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 376, + 678 + ], + "score": 1.0, + "content": "an array of challenging reasoning skills such as following transitive", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 677, + 376, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 376, + 690 + ], + "score": 1.0, + "content": "relations, counting objects and comparing their properties, without", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 688, + 376, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 376, + 700 + ], + "score": 1.0, + "content": "allowing any shortcuts around such reasoning. Notably, each in-", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 40.5 + }, + { + "type": "image", + "bbox": [ + 387, + 537, + 501, + 613 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 387, + 537, + 501, + 613 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 387, + 537, + 501, + 613 + ], + "spans": [ + { + "bbox": [ + 387, + 537, + 501, + 613 + ], + "score": 0.962, + "type": "image", + "image_path": "674dcd95919d901cb7279f9b0552761c7620c6d163165031e5b1cd19087e50a7.jpg" + } + ] + } + ], + "index": 49.5, + "virtual_lines": [ + { + "bbox": [ + 387, + 537, + 501, + 575.0 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 387, + 575.0, + 501, + 613.0 + ], + "spans": [], + "index": 50 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 383, + 622, + 504, + 687 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 383, + 621, + 504, + 633 + ], + "spans": [ + { + "bbox": [ + 383, + 621, + 504, + 633 + ], + "score": 1.0, + "content": "Figure 1: A sample im-", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 382, + 632, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 382, + 632, + 506, + 645 + ], + "score": 1.0, + "content": "age from the CLEVR dataset,", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 383, + 643, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 383, + 643, + 506, + 655 + ], + "score": 1.0, + "content": "with a question: “There is a", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 382, + 655, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 382, + 655, + 505, + 666 + ], + "score": 1.0, + "content": "purple cube behind a metal", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 383, + 666, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 383, + 666, + 506, + 677 + ], + "score": 1.0, + "content": "object left to a large ball; what", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 383, + 676, + 444, + 687 + ], + "spans": [ + { + "bbox": [ + 383, + 676, + 444, + 687 + ], + "score": 1.0, + "content": "material is it?”", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 53.5 + } + ], + "index": 51.5 + }, + { + "type": "text", + "bbox": [ + 108, + 700, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "stance in CLEVR is also accompanied by a tree-structured functional program that was both used to", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "construct the question and reflects its reasoning procedure – a series of predefined operations – that", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 721, + 264, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 264, + 733 + ], + "score": 1.0, + "content": "can be composed together to answer it.", + "type": "text" + } + ], + "index": 59 + } + ], + "index": 58 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 79, + 414, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 415, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 415, + 97 + ], + "score": 1.0, + "content": "COMPOSITIONAL ATTENTION NETWORKS", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 309, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 309, + 117 + ], + "score": 1.0, + "content": "FOR MACHINE REASONING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "list", + "bbox": [ + 112, + 135, + 252, + 179 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 186, + 146 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 186, + 146 + ], + "score": 1.0, + "content": "Drew A. Hudson", + "type": "text" + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 145, + 249, + 159 + ], + "spans": [ + { + "bbox": [ + 111, + 145, + 249, + 159 + ], + "score": 1.0, + "content": "Department of Computer Science", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 156, + 194, + 169 + ], + "spans": [ + { + "bbox": [ + 112, + 156, + 194, + 169 + ], + "score": 1.0, + "content": "Stanford University", + "type": "text" + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 113, + 169, + 252, + 180 + ], + "spans": [ + { + "bbox": [ + 113, + 169, + 252, + 180 + ], + "score": 1.0, + "content": "dorarad@cs.stanford.edu", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + } + ], + "index": 4.5, + "bbox_fs": [ + 111, + 136, + 252, + 180 + ] + }, + { + "type": "text", + "bbox": [ + 310, + 136, + 449, + 180 + ], + "lines": [ + { + "bbox": [ + 310, + 134, + 418, + 148 + ], + "spans": [ + { + "bbox": [ + 310, + 134, + 418, + 148 + ], + "score": 1.0, + "content": "Christopher D. Manning", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 309, + 145, + 447, + 159 + ], + "spans": [ + { + "bbox": [ + 309, + 145, + 447, + 159 + ], + "score": 1.0, + "content": "Department of Computer Science", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 309, + 156, + 392, + 170 + ], + "spans": [ + { + "bbox": [ + 309, + 156, + 392, + 170 + ], + "score": 1.0, + "content": "Stanford University", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 309, + 169, + 450, + 180 + ], + "spans": [ + { + "bbox": [ + 309, + 169, + 450, + 180 + ], + "score": 1.0, + "content": "manning@cs.stanford.edu", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5, + "bbox_fs": [ + 309, + 134, + 450, + 180 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 208, + 333, + 220 + ], + "lines": [ + { + "bbox": [ + 276, + 207, + 336, + 222 + ], + "spans": [ + { + "bbox": [ + 276, + 207, + 336, + 222 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 143, + 235, + 468, + 411 + ], + "lines": [ + { + "bbox": [ + 142, + 235, + 469, + 248 + ], + "spans": [ + { + "bbox": [ + 142, + 235, + 469, + 248 + ], + "score": 1.0, + "content": "We present Compositional Attention Networks, a novel fully differentiable neu-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 245, + 469, + 260 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 469, + 260 + ], + "score": 1.0, + "content": "ral network architecture, designed to facilitate explicit and expressive reasoning.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 256, + 470, + 271 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 470, + 271 + ], + "score": 1.0, + "content": "While many types of neural networks are effective at learning and generalizing", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 268, + 469, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 469, + 280 + ], + "score": 1.0, + "content": "from massive quantities of data, this model moves away from monolithic black-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 279, + 469, + 291 + ], + "spans": [ + { + "bbox": [ + 141, + 279, + 469, + 291 + ], + "score": 1.0, + "content": "box architectures towards a design that provides a strong prior for iterative rea-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 291, + 469, + 302 + ], + "spans": [ + { + "bbox": [ + 142, + 291, + 469, + 302 + ], + "score": 1.0, + "content": "soning, allowing it to support explainable and structured learning, as well as", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 301, + 470, + 313 + ], + "spans": [ + { + "bbox": [ + 141, + 301, + 470, + 313 + ], + "score": 1.0, + "content": "generalization from a modest amount of data. The model builds on the great", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 312, + 469, + 324 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 469, + 324 + ], + "score": 1.0, + "content": "success of existing recurrent cells such as LSTMs: It sequences a single recur-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 323, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 469, + 335 + ], + "score": 1.0, + "content": "rent Memory, Attention, and Control (MAC) cell, and by careful design imposes", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 334, + 469, + 346 + ], + "spans": [ + { + "bbox": [ + 141, + 334, + 469, + 346 + ], + "score": 1.0, + "content": "structural constraints on the operation of each cell and the interactions between", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 345, + 469, + 357 + ], + "spans": [ + { + "bbox": [ + 142, + 345, + 469, + 357 + ], + "score": 1.0, + "content": "them, incorporating explicit control and soft attention mechanisms into their in-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 354, + 470, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 354, + 470, + 369 + ], + "score": 1.0, + "content": "terfaces. We demonstrate the model’s strength and robustness on the challenging", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 142, + 366, + 469, + 379 + ], + "spans": [ + { + "bbox": [ + 142, + 366, + 417, + 379 + ], + "score": 1.0, + "content": "CLEVR dataset for visual reasoning, achieving a new state-of-the-art", + "type": "text" + }, + { + "bbox": [ + 417, + 367, + 444, + 377 + ], + "score": 0.87, + "content": "9 8 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 366, + 469, + 379 + ], + "score": 1.0, + "content": "accu-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 378, + 469, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 378, + 469, + 390 + ], + "score": 1.0, + "content": "racy, halving the error rate of the previous best model. More importantly, we show", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 387, + 470, + 402 + ], + "spans": [ + { + "bbox": [ + 141, + 387, + 470, + 402 + ], + "score": 1.0, + "content": "that the new model is more computationally efficient and data-efficient, requiring", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 400, + 414, + 412 + ], + "spans": [ + { + "bbox": [ + 141, + 400, + 414, + 412 + ], + "score": 1.0, + "content": "an order of magnitude less time and/or data to achieve good results.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 18.5, + "bbox_fs": [ + 141, + 235, + 470, + 412 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 436, + 205, + 448 + ], + "lines": [ + { + "bbox": [ + 105, + 435, + 208, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 208, + 452 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 504, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "score": 1.0, + "content": "This paper considers how best to design neural networks to perform the iterative reasoning necessary", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 474, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 506, + 486 + ], + "score": 1.0, + "content": "for complex problem solving. Putting facts and observations together to arrive at conclusions is a", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "central necessary ability as we work to move neural networks beyond their current great success with", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "sensory perception tasks (LeCun et al., 1998; Krizhevsky et al., 2012) towards displaying Artificial", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 506, + 192, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 192, + 520 + ], + "score": 1.0, + "content": "General Intelligence.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 462, + 506, + 520 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 375, + 698 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 376, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 376, + 535 + ], + "score": 1.0, + "content": "Concretely, we develop a novel model that we apply to the CLEVR", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 533, + 376, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 376, + 547 + ], + "score": 1.0, + "content": "dataset (Johnson et al., 2016) for visual question answering (VQA).", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 545, + 376, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 376, + 558 + ], + "score": 1.0, + "content": "VQA (Antol et al., 2015; Gupta, 2017) is a challenging multimodal", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 556, + 376, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 376, + 569 + ], + "score": 1.0, + "content": "task that requires responding to natural language questions about", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 567, + 376, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 376, + 580 + ], + "score": 1.0, + "content": "images. However, Agrawal et al. (2016) show how the first gen-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 577, + 376, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 376, + 591 + ], + "score": 1.0, + "content": "eration of successful models on VQA tasks tend to acquire only", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 589, + 377, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 377, + 602 + ], + "score": 1.0, + "content": "superficial comprehension of both the image and the question, ex-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 600, + 377, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 377, + 613 + ], + "score": 1.0, + "content": "ploiting dataset biases rather than capturing a sound perception and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 612, + 376, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 376, + 623 + ], + "score": 1.0, + "content": "reasoning process that would lead to the correct answer (Sturm,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 621, + 376, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 376, + 633 + ], + "score": 1.0, + "content": "2014). CLEVR was created to address this problem. As illus-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 633, + 376, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 376, + 645 + ], + "score": 1.0, + "content": "trated in figure 1, instances in the dataset consist of rendered im-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 644, + 376, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 376, + 656 + ], + "score": 1.0, + "content": "ages featuring 3D objects of several shapes, colors, materials and", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 655, + 376, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 376, + 667 + ], + "score": 1.0, + "content": "sizes, coupled with unbiased, compositional questions that require", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 666, + 376, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 376, + 678 + ], + "score": 1.0, + "content": "an array of challenging reasoning skills such as following transitive", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 677, + 376, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 376, + 690 + ], + "score": 1.0, + "content": "relations, counting objects and comparing their properties, without", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 688, + 376, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 376, + 700 + ], + "score": 1.0, + "content": "allowing any shortcuts around such reasoning. Notably, each in-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "stance in CLEVR is also accompanied by a tree-structured functional program that was both used to", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "construct the question and reflects its reasoning procedure – a series of predefined operations – that", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 721, + 264, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 264, + 733 + ], + "score": 1.0, + "content": "can be composed together to answer it.", + "type": "text" + } + ], + "index": 59 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 523, + 377, + 700 + ] + }, + { + "type": "image", + "bbox": [ + 387, + 537, + 501, + 613 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 387, + 537, + 501, + 613 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 387, + 537, + 501, + 613 + ], + "spans": [ + { + "bbox": [ + 387, + 537, + 501, + 613 + ], + "score": 0.962, + "type": "image", + "image_path": "674dcd95919d901cb7279f9b0552761c7620c6d163165031e5b1cd19087e50a7.jpg" + } + ] + } + ], + "index": 49.5, + "virtual_lines": [ + { + "bbox": [ + 387, + 537, + 501, + 575.0 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 387, + 575.0, + 501, + 613.0 + ], + "spans": [], + "index": 50 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 383, + 622, + 504, + 687 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 383, + 621, + 504, + 633 + ], + "spans": [ + { + "bbox": [ + 383, + 621, + 504, + 633 + ], + "score": 1.0, + "content": "Figure 1: A sample im-", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 382, + 632, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 382, + 632, + 506, + 645 + ], + "score": 1.0, + "content": "age from the CLEVR dataset,", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 383, + 643, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 383, + 643, + 506, + 655 + ], + "score": 1.0, + "content": "with a question: “There is a", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 382, + 655, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 382, + 655, + 505, + 666 + ], + "score": 1.0, + "content": "purple cube behind a metal", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 383, + 666, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 383, + 666, + 506, + 677 + ], + "score": 1.0, + "content": "object left to a large ball; what", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 383, + 676, + 444, + 687 + ], + "spans": [ + { + "bbox": [ + 383, + 676, + 444, + 687 + ], + "score": 1.0, + "content": "material is it?”", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 53.5 + } + ], + "index": 51.5 + }, + { + "type": "text", + "bbox": [ + 108, + 700, + 504, + 731 + ], + "lines": [], + "index": 58, + "bbox_fs": [ + 106, + 699, + 505, + 733 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Most neural networks are essentially very large correlation engines that will hone in on any statis-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "tical, potentially spurious pattern that allows them to model the observed data more accurately. In", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "contrast, we seek to create a model structure that requires combining sound inference steps to solve", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "a problem instance. At the other extreme, some approaches adopt symbolic structures that resemble", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "the expression trees of programming languages to perform reasoning (Andreas et al., 2016b; Hu", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "et al., 2017). In particular, some approaches to CLEVR use the supplied functional programs for", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "supervised or semi-supervised training (Andreas et al., 2016a; Johnson et al., 2017). Not only do we", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "score": 1.0, + "content": "wish to avoid using such supervision in our work, but we in general suspect that the rigidity of these", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 170, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 506, + 183 + ], + "score": 1.0, + "content": "structures and the use of an inventory of operation-specific neural modules undermines robustness", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 181, + 471, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 471, + 194 + ], + "score": 1.0, + "content": "and generalization, and at any rate requires more complex reinforcement learning methods.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 198, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 106, + 197, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 197, + 505, + 211 + ], + "score": 1.0, + "content": "To address these weaknesses, while still seeking to use a sound and transparent underlying reasoning", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 506, + 222 + ], + "score": 1.0, + "content": "process, we propose Compositional Attention Networks, a novel, fully differentiable, non-modular", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 506, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 506, + 233 + ], + "score": 1.0, + "content": "architecture for reasoning tasks. Our model is a straightforward recurrent neural network with at-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 504, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 504, + 243 + ], + "score": 1.0, + "content": "tention; the novelty lies in the use of a new Memory, Attention and Composition (MAC) cell. The", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "score": 1.0, + "content": "constrained and deliberate design of the MAC cell was developed as a kind of strong structural", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 254, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 265 + ], + "score": 1.0, + "content": "prior that encourages the network to solve problems by stringing together a sequence of transparent", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 264, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 277 + ], + "score": 1.0, + "content": "reasoning steps. MAC cells are versatile but constrained neural units. They explicitly separate out", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "memory from control, both represented recurrently. The unit contains three sub-units: The control", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 286, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 298 + ], + "score": 1.0, + "content": "unit updates the control representation based on outside instructions (for VQA, the question), learn-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 297, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 309 + ], + "score": 1.0, + "content": "ing to successively attend to different parts of the instructions; the read unit gets information out of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "score": 1.0, + "content": "a knowledge base (for VQA, the image) based on the control signal and the previous memory; the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 319, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 506, + 331 + ], + "score": 1.0, + "content": "write unit updates the memory based on soft self-attention to previous memories, controlled by the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 329, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 343 + ], + "score": 1.0, + "content": "retrieved information and the control signal. A universal MAC unit with a single set of parameters", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 340, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 505, + 353 + ], + "score": 1.0, + "content": "is used throughout the reasoning process, but its behavior can vary widely based on the context in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 351, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 506, + 365 + ], + "score": 1.0, + "content": "which it is applied – the input to the control unit and the contents of the knowledge base. With atten-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "tion, our MAC network has the capacity to represent arbitrarily complex acyclic reasoning graphs in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "score": 1.0, + "content": "a soft manner, while having physically sequential structure. The result is a continuous counterpart", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 384, + 423, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 423, + 397 + ], + "score": 1.0, + "content": "to module networks that can be trained end-to-end simply by backpropagation.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 402, + 505, + 566 + ], + "lines": [ + { + "bbox": [ + 106, + 400, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 505, + 415 + ], + "score": 1.0, + "content": "We test the behavior of our new network on CLEVR and its associated datasets. On the primary", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 316, + 425 + ], + "score": 1.0, + "content": "CLEVR reasoning task, we achieve an accuracy of", + "type": "text" + }, + { + "bbox": [ + 316, + 412, + 343, + 423 + ], + "score": 0.87, + "content": "9 8 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 412, + 505, + 425 + ], + "score": 1.0, + "content": ", halving the error rate compared to the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 423, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 436 + ], + "score": 1.0, + "content": "previous state-of-the-art FiLM model (Perez et al., 2017). In particular, we show that our architec-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 434, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 448 + ], + "score": 1.0, + "content": "ture yields better performance on questions involving counting and aggregation. In supplementary", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 443, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 104, + 443, + 506, + 460 + ], + "score": 1.0, + "content": "studies, we show that the MAC network learns more quickly (both in terms of number of training", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "score": 1.0, + "content": "epochs and training time) and more effectively from limited amounts of training data. Moreover, it", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 466, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 314, + 480 + ], + "score": 1.0, + "content": "also achieves a new state-of-the-art performance of", + "type": "text" + }, + { + "bbox": [ + 315, + 467, + 342, + 478 + ], + "score": 0.88, + "content": "8 2 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 466, + 505, + 480 + ], + "score": 1.0, + "content": "on the more varied and difficult human-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "authored questions of the CLEVR-Humans dataset. The careful design of our cell encourages com-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "score": 1.0, + "content": "positionality, versatility and transparency. We achieve these properties by defining attention-based", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 500, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 512 + ], + "score": 1.0, + "content": "interfaces that constrict the cell’s input and output spaces, and so constrain the interactions both", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "score": 1.0, + "content": "between and inside cells in order to guide them towards simple reasoning behaviors. Although", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "score": 1.0, + "content": "each cell’s functionality has only a limited range of possible continuous reasoning behaviors, when", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "chained together in a MAC network, the whole system becomes expressive and powerful. In the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 544, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 506, + 556 + ], + "score": 1.0, + "content": "future, we believe that the architecture will also prove beneficial for other multi-step reasoning and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 554, + 458, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 458, + 568 + ], + "score": 1.0, + "content": "inference tasks, for instance in machine comprehension and textual question answering.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 35 + }, + { + "type": "title", + "bbox": [ + 108, + 591, + 211, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 590, + 213, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 213, + 606 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "There have been several prominent models that address the CLEVR task. By and large they can be", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "partitioned into two groups: module networks, which in practice have all used the strong supervision", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "provided in the form of tree-structured functional programs that accompany each data instance, and", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "large, relatively unstructured end-to-end differentiable networks that complement a fairly standard", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "stack of CNNs with components that aid in performing reasoning tasks. In contrast to modular", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "approaches (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017), our model does not", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "require additional supervision and makes use of a single computational cell chained in sequence", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "(like an LSTM) rather than a collection of custom modules deployed in a rigid tree structure. In", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "contrast to augmented CNN approaches (Santoro et al., 2017; Perez et al., 2017), we suggest that our", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "approach provides an ability for relational reasoning with better generalization capacity and higher", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 48.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Most neural networks are essentially very large correlation engines that will hone in on any statis-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "tical, potentially spurious pattern that allows them to model the observed data more accurately. In", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "contrast, we seek to create a model structure that requires combining sound inference steps to solve", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "a problem instance. At the other extreme, some approaches adopt symbolic structures that resemble", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "the expression trees of programming languages to perform reasoning (Andreas et al., 2016b; Hu", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "et al., 2017). In particular, some approaches to CLEVR use the supplied functional programs for", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "supervised or semi-supervised training (Andreas et al., 2016a; Johnson et al., 2017). Not only do we", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "score": 1.0, + "content": "wish to avoid using such supervision in our work, but we in general suspect that the rigidity of these", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 170, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 506, + 183 + ], + "score": 1.0, + "content": "structures and the use of an inventory of operation-specific neural modules undermines robustness", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 181, + 471, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 471, + 194 + ], + "score": 1.0, + "content": "and generalization, and at any rate requires more complex reinforcement learning methods.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 82, + 506, + 194 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 198, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 106, + 197, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 197, + 505, + 211 + ], + "score": 1.0, + "content": "To address these weaknesses, while still seeking to use a sound and transparent underlying reasoning", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 506, + 222 + ], + "score": 1.0, + "content": "process, we propose Compositional Attention Networks, a novel, fully differentiable, non-modular", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 506, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 506, + 233 + ], + "score": 1.0, + "content": "architecture for reasoning tasks. Our model is a straightforward recurrent neural network with at-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 504, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 504, + 243 + ], + "score": 1.0, + "content": "tention; the novelty lies in the use of a new Memory, Attention and Composition (MAC) cell. The", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "score": 1.0, + "content": "constrained and deliberate design of the MAC cell was developed as a kind of strong structural", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 254, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 265 + ], + "score": 1.0, + "content": "prior that encourages the network to solve problems by stringing together a sequence of transparent", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 264, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 277 + ], + "score": 1.0, + "content": "reasoning steps. MAC cells are versatile but constrained neural units. They explicitly separate out", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "memory from control, both represented recurrently. The unit contains three sub-units: The control", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 286, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 298 + ], + "score": 1.0, + "content": "unit updates the control representation based on outside instructions (for VQA, the question), learn-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 297, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 309 + ], + "score": 1.0, + "content": "ing to successively attend to different parts of the instructions; the read unit gets information out of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "score": 1.0, + "content": "a knowledge base (for VQA, the image) based on the control signal and the previous memory; the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 319, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 506, + 331 + ], + "score": 1.0, + "content": "write unit updates the memory based on soft self-attention to previous memories, controlled by the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 329, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 343 + ], + "score": 1.0, + "content": "retrieved information and the control signal. A universal MAC unit with a single set of parameters", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 340, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 505, + 353 + ], + "score": 1.0, + "content": "is used throughout the reasoning process, but its behavior can vary widely based on the context in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 351, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 506, + 365 + ], + "score": 1.0, + "content": "which it is applied – the input to the control unit and the contents of the knowledge base. With atten-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "tion, our MAC network has the capacity to represent arbitrarily complex acyclic reasoning graphs in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "score": 1.0, + "content": "a soft manner, while having physically sequential structure. The result is a continuous counterpart", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 384, + 423, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 423, + 397 + ], + "score": 1.0, + "content": "to module networks that can be trained end-to-end simply by backpropagation.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 197, + 506, + 397 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 402, + 505, + 566 + ], + "lines": [ + { + "bbox": [ + 106, + 400, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 505, + 415 + ], + "score": 1.0, + "content": "We test the behavior of our new network on CLEVR and its associated datasets. On the primary", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 316, + 425 + ], + "score": 1.0, + "content": "CLEVR reasoning task, we achieve an accuracy of", + "type": "text" + }, + { + "bbox": [ + 316, + 412, + 343, + 423 + ], + "score": 0.87, + "content": "9 8 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 412, + 505, + 425 + ], + "score": 1.0, + "content": ", halving the error rate compared to the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 423, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 436 + ], + "score": 1.0, + "content": "previous state-of-the-art FiLM model (Perez et al., 2017). In particular, we show that our architec-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 434, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 448 + ], + "score": 1.0, + "content": "ture yields better performance on questions involving counting and aggregation. In supplementary", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 443, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 104, + 443, + 506, + 460 + ], + "score": 1.0, + "content": "studies, we show that the MAC network learns more quickly (both in terms of number of training", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "score": 1.0, + "content": "epochs and training time) and more effectively from limited amounts of training data. Moreover, it", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 466, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 314, + 480 + ], + "score": 1.0, + "content": "also achieves a new state-of-the-art performance of", + "type": "text" + }, + { + "bbox": [ + 315, + 467, + 342, + 478 + ], + "score": 0.88, + "content": "8 2 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 466, + 505, + 480 + ], + "score": 1.0, + "content": "on the more varied and difficult human-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "authored questions of the CLEVR-Humans dataset. The careful design of our cell encourages com-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "score": 1.0, + "content": "positionality, versatility and transparency. We achieve these properties by defining attention-based", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 500, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 512 + ], + "score": 1.0, + "content": "interfaces that constrict the cell’s input and output spaces, and so constrain the interactions both", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "score": 1.0, + "content": "between and inside cells in order to guide them towards simple reasoning behaviors. Although", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "score": 1.0, + "content": "each cell’s functionality has only a limited range of possible continuous reasoning behaviors, when", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "chained together in a MAC network, the whole system becomes expressive and powerful. In the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 544, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 506, + 556 + ], + "score": 1.0, + "content": "future, we believe that the architecture will also prove beneficial for other multi-step reasoning and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 554, + 458, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 458, + 568 + ], + "score": 1.0, + "content": "inference tasks, for instance in machine comprehension and textual question answering.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 35, + "bbox_fs": [ + 104, + 400, + 506, + 568 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 591, + 211, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 590, + 213, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 213, + 606 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "There have been several prominent models that address the CLEVR task. By and large they can be", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "partitioned into two groups: module networks, which in practice have all used the strong supervision", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "provided in the form of tree-structured functional programs that accompany each data instance, and", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "large, relatively unstructured end-to-end differentiable networks that complement a fairly standard", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "stack of CNNs with components that aid in performing reasoning tasks. In contrast to modular", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "approaches (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017), our model does not", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "require additional supervision and makes use of a single computational cell chained in sequence", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "(like an LSTM) rather than a collection of custom modules deployed in a rigid tree structure. In", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "contrast to augmented CNN approaches (Santoro et al., 2017; Perez et al., 2017), we suggest that our", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "approach provides an ability for relational reasoning with better generalization capacity and higher", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "computational efficiency. These approaches and other related work are discussed and contrasted in", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 327, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 327, + 106 + ], + "score": 1.0, + "content": "more detail in the supplementary material in section C.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 48.5, + "bbox_fs": [ + 105, + 622, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "computational efficiency. These approaches and other related work are discussed and contrasted in", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 327, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 327, + 106 + ], + "score": 1.0, + "content": "more detail in the supplementary material in section C.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 107, + 120, + 339, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 119, + 340, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 340, + 135 + ], + "score": 1.0, + "content": "3 COMPOSITIONAL ATTENTION NETWORKS", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 145, + 505, + 178 + ], + "lines": [ + { + "bbox": [ + 106, + 145, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 505, + 157 + ], + "score": 1.0, + "content": "Compositional Attention Networks is an end-to-end architecture for question-answering tasks that", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 156, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 156, + 505, + 169 + ], + "score": 1.0, + "content": "sequentially performs an explicit reasoning process by stringing together small building blocks,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 165, + 399, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 399, + 181 + ], + "score": 1.0, + "content": "called MAC cells, each is responsible for performing one reasoning step.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 183, + 504, + 228 + ], + "lines": [ + { + "bbox": [ + 106, + 184, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 505, + 196 + ], + "score": 1.0, + "content": "We now provide an overview of the model, and a detailed discussion of the MAC cell. The model", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 195, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 506, + 207 + ], + "score": 1.0, + "content": "is composed of three components: an Input unit, the core MAC network, and an output unit. A", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 205, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 506, + 219 + ], + "score": 1.0, + "content": "TensorFlow implementation of the network, along with pretrained models will be made publicly", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 217, + 148, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 148, + 228 + ], + "score": 1.0, + "content": "available.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 233, + 505, + 289 + ], + "lines": [ + { + "bbox": [ + 106, + 234, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 505, + 246 + ], + "score": 1.0, + "content": "In this paper we explore the model in the context of VQA. However, it should be noted that while the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 244, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 505, + 258 + ], + "score": 1.0, + "content": "input and output units are naturally domain-specific and should be designed to fit the task at hand,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 255, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 269 + ], + "score": 1.0, + "content": "the MAC network has been designed to be generic and more broadly applicable, and may prove", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "useful in contexts beyond those explored in the paper, such as machine comprehension or question", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 278, + 475, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 278, + 475, + 290 + ], + "score": 1.0, + "content": "answering over knowledge bases, which in our belief is a promising avenue for future work.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12 + }, + { + "type": "title", + "bbox": [ + 108, + 302, + 204, + 313 + ], + "lines": [ + { + "bbox": [ + 105, + 300, + 205, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 205, + 315 + ], + "score": 1.0, + "content": "3.1 THE INPUT UNIT", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 322, + 505, + 367 + ], + "lines": [ + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "score": 1.0, + "content": "The input unit processes the raw inputs given to the system into distributed vector representations. It", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 334, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 346 + ], + "score": 1.0, + "content": "receives a text question (or in general, a query), and an image (or in general, a Knowledge Base (KB))", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "and processes each of them with a matching sub-unit, for the query and the KB, here a biLSTM and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 356, + 412, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 412, + 367 + ], + "score": 1.0, + "content": "a CNN. More details can be found in the supplementary material, section A.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 107, + 372, + 505, + 450 + ], + "lines": [ + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "At the end of this stage, we get from the query sub-unit a series of biLSTM output states, which we", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 224, + 396 + ], + "score": 1.0, + "content": "refer to as contextual words,", + "type": "text" + }, + { + "bbox": [ + 225, + 384, + 281, + 396 + ], + "score": 0.92, + "content": "[ c w _ { 1 } , . . . , c w _ { S } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 383, + 312, + 396 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 313, + 384, + 321, + 393 + ], + "score": 0.82, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 383, + 505, + 396 + ], + "score": 1.0, + "content": "is the length of the question. In addition, we", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 394, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 122, + 407 + ], + "score": 1.0, + "content": "get", + "type": "text" + }, + { + "bbox": [ + 122, + 394, + 185, + 407 + ], + "score": 0.94, + "content": "q = [ \\overleftrightarrow { \\lambda \\omega _ { 1 } } , \\overrightarrow { c w _ { S } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 394, + 506, + 407 + ], + "score": 1.0, + "content": ", the concatenation of the hidden states from the backward and forward LSTM", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 405, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 184, + 418 + ], + "score": 1.0, + "content": "passes. We refer to", + "type": "text" + }, + { + "bbox": [ + 184, + 408, + 190, + 417 + ], + "score": 0.8, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 405, + 506, + 418 + ], + "score": 1.0, + "content": "as the question representation. Furthermore, we get from the Knowledge-Base", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 417, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 428 + ], + "score": 1.0, + "content": "sub-unit a static representation of the knowledge base. For the case of VQA, it will be represented", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 427, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 201, + 440 + ], + "score": 1.0, + "content": "by a continuous matrix", + "type": "text" + }, + { + "bbox": [ + 201, + 428, + 226, + 438 + ], + "score": 0.91, + "content": "K B _ { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 427, + 282, + 440 + ], + "score": 1.0, + "content": "of dimension", + "type": "text" + }, + { + "bbox": [ + 282, + 428, + 316, + 439 + ], + "score": 0.88, + "content": "H , W , d", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 427, + 346, + 440 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 346, + 428, + 405, + 438 + ], + "score": 0.91, + "content": "H = W = 1 4", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 427, + 505, + 440 + ], + "score": 1.0, + "content": "are the height and width", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 438, + 359, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 359, + 451 + ], + "score": 1.0, + "content": "of the transformed image, corresponding to each of its regions.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 108, + 462, + 200, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 203, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 203, + 476 + ], + "score": 1.0, + "content": "3.2 THE MAC CELL", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 483, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "The MAC network, which is the heart of our model, chains a sequence of small building blocks,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "called MAC cells, each responsible for performing one reasoning step. The model is provided access", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 505, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 519 + ], + "score": 1.0, + "content": "to a Knowledge Base (KB), which is, for the specific case of VQA, the given image, and then upon", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 378, + 529 + ], + "score": 1.0, + "content": "receiving a query, i.e. a question, the model iteratively focuses, in", + "type": "text" + }, + { + "bbox": [ + 379, + 518, + 385, + 528 + ], + "score": 0.78, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "steps, on the query’s various", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 527, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 541 + ], + "score": 1.0, + "content": "parts, each reflects in turn the current reasoning step, which we term the control. Consequently,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 538, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 506, + 551 + ], + "score": 1.0, + "content": "guided by this control, it retrieves the relevant information from the KB, that is then passed to the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 550, + 232, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 232, + 561 + ], + "score": 1.0, + "content": "next cell in a recurrent fashion.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 108, + 566, + 504, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "Drawing inspiration from the Model-View-Controller paradigm used in software design and from", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 577, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 506, + 589 + ], + "score": 1.0, + "content": "the commonly exercised separation between control and data paths in computer architecture, the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "MAC cell is composed of three units: control unit, read unit and write unit. Each has a clearly", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 599, + 462, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 462, + 611 + ], + "score": 1.0, + "content": "defined role and an interface through which it interacts with the other units. See figure 2.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "The careful design and imposed interfaces that constrain the interaction between the units inside the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "MAC cell, as described below, serve as structural prior that limits the space of hypotheses it can", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "learn, thereby guiding it towards acquiring the intended reasoning behaviors. As such, this prior", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 649, + 354, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 354, + 661 + ], + "score": 1.0, + "content": "facilitates the learning process and mitigate overfitting issues.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "In particular, and similar in spirit to Perez et al. (2017), we allow the question to interact with the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "Knowledge Base – the image for the case of VQA, only through indirect means: by guiding the cell", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "to attend to different elements in the KB, as well as controlling its operation through gating mecha-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "nisms. Thus, in both cases, the interaction between these mediums, visual and textual, or knowledge", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "and query, is mediated through probability distributions, either in the form of attention maps, or as", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "gates, further detailed below. This stands in stark contrast to many common approaches that fuse the", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 120, + 339, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 119, + 340, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 340, + 135 + ], + "score": 1.0, + "content": "3 COMPOSITIONAL ATTENTION NETWORKS", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 145, + 505, + 178 + ], + "lines": [ + { + "bbox": [ + 106, + 145, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 505, + 157 + ], + "score": 1.0, + "content": "Compositional Attention Networks is an end-to-end architecture for question-answering tasks that", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 156, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 156, + 505, + 169 + ], + "score": 1.0, + "content": "sequentially performs an explicit reasoning process by stringing together small building blocks,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 165, + 399, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 399, + 181 + ], + "score": 1.0, + "content": "called MAC cells, each is responsible for performing one reasoning step.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 145, + 505, + 181 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 183, + 504, + 228 + ], + "lines": [ + { + "bbox": [ + 106, + 184, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 505, + 196 + ], + "score": 1.0, + "content": "We now provide an overview of the model, and a detailed discussion of the MAC cell. The model", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 195, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 506, + 207 + ], + "score": 1.0, + "content": "is composed of three components: an Input unit, the core MAC network, and an output unit. A", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 205, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 506, + 219 + ], + "score": 1.0, + "content": "TensorFlow implementation of the network, along with pretrained models will be made publicly", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 217, + 148, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 148, + 228 + ], + "score": 1.0, + "content": "available.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 184, + 506, + 228 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 233, + 505, + 289 + ], + "lines": [ + { + "bbox": [ + 106, + 234, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 505, + 246 + ], + "score": 1.0, + "content": "In this paper we explore the model in the context of VQA. However, it should be noted that while the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 244, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 505, + 258 + ], + "score": 1.0, + "content": "input and output units are naturally domain-specific and should be designed to fit the task at hand,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 255, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 269 + ], + "score": 1.0, + "content": "the MAC network has been designed to be generic and more broadly applicable, and may prove", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "useful in contexts beyond those explored in the paper, such as machine comprehension or question", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 278, + 475, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 278, + 475, + 290 + ], + "score": 1.0, + "content": "answering over knowledge bases, which in our belief is a promising avenue for future work.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 234, + 505, + 290 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 302, + 204, + 313 + ], + "lines": [ + { + "bbox": [ + 105, + 300, + 205, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 205, + 315 + ], + "score": 1.0, + "content": "3.1 THE INPUT UNIT", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 322, + 505, + 367 + ], + "lines": [ + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "score": 1.0, + "content": "The input unit processes the raw inputs given to the system into distributed vector representations. It", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 334, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 346 + ], + "score": 1.0, + "content": "receives a text question (or in general, a query), and an image (or in general, a Knowledge Base (KB))", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "and processes each of them with a matching sub-unit, for the query and the KB, here a biLSTM and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 356, + 412, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 412, + 367 + ], + "score": 1.0, + "content": "a CNN. More details can be found in the supplementary material, section A.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 322, + 506, + 367 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 372, + 505, + 450 + ], + "lines": [ + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "At the end of this stage, we get from the query sub-unit a series of biLSTM output states, which we", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 224, + 396 + ], + "score": 1.0, + "content": "refer to as contextual words,", + "type": "text" + }, + { + "bbox": [ + 225, + 384, + 281, + 396 + ], + "score": 0.92, + "content": "[ c w _ { 1 } , . . . , c w _ { S } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 383, + 312, + 396 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 313, + 384, + 321, + 393 + ], + "score": 0.82, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 383, + 505, + 396 + ], + "score": 1.0, + "content": "is the length of the question. In addition, we", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 394, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 122, + 407 + ], + "score": 1.0, + "content": "get", + "type": "text" + }, + { + "bbox": [ + 122, + 394, + 185, + 407 + ], + "score": 0.94, + "content": "q = [ \\overleftrightarrow { \\lambda \\omega _ { 1 } } , \\overrightarrow { c w _ { S } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 394, + 506, + 407 + ], + "score": 1.0, + "content": ", the concatenation of the hidden states from the backward and forward LSTM", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 405, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 184, + 418 + ], + "score": 1.0, + "content": "passes. We refer to", + "type": "text" + }, + { + "bbox": [ + 184, + 408, + 190, + 417 + ], + "score": 0.8, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 405, + 506, + 418 + ], + "score": 1.0, + "content": "as the question representation. Furthermore, we get from the Knowledge-Base", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 417, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 428 + ], + "score": 1.0, + "content": "sub-unit a static representation of the knowledge base. For the case of VQA, it will be represented", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 427, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 201, + 440 + ], + "score": 1.0, + "content": "by a continuous matrix", + "type": "text" + }, + { + "bbox": [ + 201, + 428, + 226, + 438 + ], + "score": 0.91, + "content": "K B _ { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 427, + 282, + 440 + ], + "score": 1.0, + "content": "of dimension", + "type": "text" + }, + { + "bbox": [ + 282, + 428, + 316, + 439 + ], + "score": 0.88, + "content": "H , W , d", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 427, + 346, + 440 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 346, + 428, + 405, + 438 + ], + "score": 0.91, + "content": "H = W = 1 4", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 427, + 505, + 440 + ], + "score": 1.0, + "content": "are the height and width", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 438, + 359, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 359, + 451 + ], + "score": 1.0, + "content": "of the transformed image, corresponding to each of its regions.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 372, + 506, + 451 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 462, + 200, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 203, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 203, + 476 + ], + "score": 1.0, + "content": "3.2 THE MAC CELL", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 483, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "The MAC network, which is the heart of our model, chains a sequence of small building blocks,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "called MAC cells, each responsible for performing one reasoning step. The model is provided access", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 505, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 519 + ], + "score": 1.0, + "content": "to a Knowledge Base (KB), which is, for the specific case of VQA, the given image, and then upon", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 378, + 529 + ], + "score": 1.0, + "content": "receiving a query, i.e. a question, the model iteratively focuses, in", + "type": "text" + }, + { + "bbox": [ + 379, + 518, + 385, + 528 + ], + "score": 0.78, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "steps, on the query’s various", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 527, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 541 + ], + "score": 1.0, + "content": "parts, each reflects in turn the current reasoning step, which we term the control. Consequently,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 538, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 506, + 551 + ], + "score": 1.0, + "content": "guided by this control, it retrieves the relevant information from the KB, that is then passed to the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 550, + 232, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 232, + 561 + ], + "score": 1.0, + "content": "next cell in a recurrent fashion.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 483, + 506, + 561 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 566, + 504, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "Drawing inspiration from the Model-View-Controller paradigm used in software design and from", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 577, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 506, + 589 + ], + "score": 1.0, + "content": "the commonly exercised separation between control and data paths in computer architecture, the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "MAC cell is composed of three units: control unit, read unit and write unit. Each has a clearly", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 599, + 462, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 462, + 611 + ], + "score": 1.0, + "content": "defined role and an interface through which it interacts with the other units. See figure 2.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 566, + 506, + 611 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "The careful design and imposed interfaces that constrain the interaction between the units inside the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "MAC cell, as described below, serve as structural prior that limits the space of hypotheses it can", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "learn, thereby guiding it towards acquiring the intended reasoning behaviors. As such, this prior", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 649, + 354, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 354, + 661 + ], + "score": 1.0, + "content": "facilitates the learning process and mitigate overfitting issues.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 615, + 505, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "In particular, and similar in spirit to Perez et al. (2017), we allow the question to interact with the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "Knowledge Base – the image for the case of VQA, only through indirect means: by guiding the cell", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "to attend to different elements in the KB, as well as controlling its operation through gating mecha-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "nisms. Thus, in both cases, the interaction between these mediums, visual and textual, or knowledge", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "and query, is mediated through probability distributions, either in the form of attention maps, or as", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "gates, further detailed below. This stands in stark contrast to many common approaches that fuse the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 251, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 505, + 265 + ], + "score": 1.0, + "content": "question and image together into the same vector space through linear combinations, multiplication,", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 262, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 505, + 275 + ], + "score": 1.0, + "content": "or concatenation. Rather, our controlled interaction distills the influence that the query should have", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 274, + 473, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 473, + 286 + ], + "score": 1.0, + "content": "in processing the Knowledge Base, casting it onto discrete probability distributions instead.", + "type": "text", + "cross_page": true + } + ], + "index": 12 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 666, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 215, + 82, + 370, + 198 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 215, + 82, + 370, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 215, + 82, + 370, + 198 + ], + "spans": [ + { + "bbox": [ + 215, + 82, + 370, + 198 + ], + "score": 0.966, + "type": "image", + "image_path": "21ef6959e8f5c66e79f26b2aa043aef1bf62c1ad19d7a86079e5194974b914be.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 215, + 82, + 370, + 96.5 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 215, + 96.5, + 370, + 111.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 215, + 111.0, + 370, + 125.5 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 215, + 125.5, + 370, + 140.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 215, + 140.0, + 370, + 154.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 215, + 154.5, + 370, + 169.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 215, + 169.0, + 370, + 183.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 215, + 183.5, + 370, + 198.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 208, + 505, + 231 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 209, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 221 + ], + "score": 1.0, + "content": "Figure 2: The MAC cell, which is a recurrent unit comprised of a Control Unit, Read Unit, and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "Write Unit. Blue shows the control flow and red shows the memory flow. See section 3.2 for details.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 108, + 252, + 504, + 285 + ], + "lines": [ + { + "bbox": [ + 106, + 251, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 505, + 265 + ], + "score": 1.0, + "content": "question and image together into the same vector space through linear combinations, multiplication,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 262, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 505, + 275 + ], + "score": 1.0, + "content": "or concatenation. Rather, our controlled interaction distills the influence that the query should have", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 274, + 473, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 473, + 286 + ], + "score": 1.0, + "content": "in processing the Knowledge Base, casting it onto discrete probability distributions instead.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 290, + 505, + 368 + ], + "lines": [ + { + "bbox": [ + 106, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "The MAC cell has been designed to replace the discrete and predefined “modules” used in the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 301, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 505, + 314 + ], + "score": 1.0, + "content": "modular approach (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017). Rather, we", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "create one universal and versatile cell that is applied across all the reasoning steps, sharing both", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "its architecture as well as its parameters, across all of its instantiations. In contrast to the discrete", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 334, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 347 + ], + "score": 1.0, + "content": "modules, each trained to specialize to some specific elementary reasoning task, the MAC cell is", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "capable of demonstrating a continuous range of possible reasoning behaviors conditioned on the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 356, + 427, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 427, + 369 + ], + "score": 1.0, + "content": "context in which it is applied – namely, the inputs it receives from the prior cell.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 373, + 505, + 472 + ], + "lines": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 145, + 385 + ], + "score": 1.0, + "content": "Each cell", + "type": "text" + }, + { + "bbox": [ + 146, + 373, + 176, + 384 + ], + "score": 0.91, + "content": "M A C _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 373, + 314, + 385 + ], + "score": 1.0, + "content": "maintains two dual states: control", + "type": "text" + }, + { + "bbox": [ + 315, + 375, + 323, + 384 + ], + "score": 0.85, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 373, + 375, + 385 + ], + "score": 1.0, + "content": "and memory", + "type": "text" + }, + { + "bbox": [ + 376, + 375, + 389, + 384 + ], + "score": 0.85, + "content": "m _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 373, + 505, + 385 + ], + "score": 1.0, + "content": ", both are continuous vectors", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 384, + 504, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 161, + 396 + ], + "score": 1.0, + "content": "of dimension", + "type": "text" + }, + { + "bbox": [ + 162, + 385, + 168, + 394 + ], + "score": 0.7, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 384, + 221, + 396 + ], + "score": 1.0, + "content": ". The control", + "type": "text" + }, + { + "bbox": [ + 222, + 386, + 231, + 395 + ], + "score": 0.84, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 384, + 504, + 396 + ], + "score": 1.0, + "content": "represents the reasoning operation the MAC cell should accomplish", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "score": 1.0, + "content": "in the current step – focusing only on some aspect of the whole question. This is represented by", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 430, + 419 + ], + "score": 1.0, + "content": "a weighted-average attention-based sum of the question words. The memory", + "type": "text" + }, + { + "bbox": [ + 430, + 408, + 443, + 417 + ], + "score": 0.85, + "content": "m _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "represents the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 416, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 506, + 431 + ], + "score": 1.0, + "content": "current context information deemed relevant to respond to the query, or answer the question.This is", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "represented practically by a weighted average over elements from the KB, or for the case of VQA,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 439, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 194, + 452 + ], + "score": 1.0, + "content": "regions in the image.", + "type": "text" + }, + { + "bbox": [ + 195, + 441, + 209, + 450 + ], + "score": 0.85, + "content": "m _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 439, + 227, + 452 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 228, + 441, + 238, + 450 + ], + "score": 0.85, + "content": "c _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 439, + 495, + 452 + ], + "score": 1.0, + "content": "are initialized each to a random vector parameter of dimension", + "type": "text" + }, + { + "bbox": [ + 495, + 439, + 501, + 449 + ], + "score": 0.73, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 439, + 506, + 452 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "score": 1.0, + "content": "The memory and control states are passed from one cell to the next in a recurrent fashion, and used", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 461, + 484, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 484, + 474 + ], + "score": 1.0, + "content": "in a way reminiscent of Key-Value memory networks (Miller et al., 2016), as discussed below.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24 + }, + { + "type": "title", + "bbox": [ + 108, + 484, + 228, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 230, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 230, + 497 + ], + "score": 1.0, + "content": "3.2.1 THE CONTROL UNIT", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 503, + 505, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "score": 1.0, + "content": "The control unit determines the reasoning operation that should be applied at this step. It receives", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 195, + 527 + ], + "score": 1.0, + "content": "the contextual words", + "type": "text" + }, + { + "bbox": [ + 195, + 514, + 252, + 526 + ], + "score": 0.92, + "content": "[ c w _ { 1 } , . . . , c w _ { S } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 514, + 369, + 527 + ], + "score": 1.0, + "content": ", the question representation", + "type": "text" + }, + { + "bbox": [ + 370, + 516, + 376, + 525 + ], + "score": 0.76, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 514, + 505, + 527 + ], + "score": 1.0, + "content": ", and the control state from the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 525, + 369, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 184, + 537 + ], + "score": 1.0, + "content": "previous MAC cell", + "type": "text" + }, + { + "bbox": [ + 185, + 527, + 204, + 537 + ], + "score": 0.88, + "content": "c _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 525, + 358, + 537 + ], + "score": 1.0, + "content": ", all of which are vectors of dimension", + "type": "text" + }, + { + "bbox": [ + 358, + 526, + 365, + 535 + ], + "score": 0.76, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 525, + 369, + 537 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 542, + 505, + 609 + ], + "lines": [ + { + "bbox": [ + 105, + 541, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 555 + ], + "score": 1.0, + "content": "We would like to allow our MAC cell to perform continuously varied and adaptive range of behav-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 553, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 565 + ], + "score": 1.0, + "content": "iors, as demanded by the question. Therefore, we define the behavior of each cell to be a function of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 564, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 193, + 577 + ], + "score": 1.0, + "content": "the contextual words", + "type": "text" + }, + { + "bbox": [ + 194, + 564, + 250, + 576 + ], + "score": 0.92, + "content": "[ c w _ { 1 } , . . . , c w _ { S } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 564, + 506, + 577 + ], + "score": 1.0, + "content": ", weighted-averaged according to the attention distribution that", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 576, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 587 + ], + "score": 1.0, + "content": "the control unit produces at each step. This will allow the cell to adapt its behavior – the reason-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "ing operation it performs – to the question it receives, instead of having a fixed set of predefined", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 596, + 500, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 500, + 610 + ], + "score": 1.0, + "content": "behaviours as is the case in competing approaches Andreas et al. (2016a;b); Johnson et al. (2017).", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 614, + 505, + 648 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 428, + 627 + ], + "score": 1.0, + "content": "The formal specification of the control unit is shown in figure 3. The question", + "type": "text" + }, + { + "bbox": [ + 429, + 616, + 435, + 626 + ], + "score": 0.8, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 613, + 505, + 627 + ], + "score": 1.0, + "content": "is linearly trans-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 626, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 194, + 637 + ], + "score": 1.0, + "content": "formed into a vector", + "type": "text" + }, + { + "bbox": [ + 195, + 627, + 204, + 636 + ], + "score": 0.84, + "content": "q _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 626, + 505, + 637 + ], + "score": 1.0, + "content": "of the same dimension, which in turn is concatenated with the previous", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 635, + 425, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 158, + 650 + ], + "score": 1.0, + "content": "control state", + "type": "text" + }, + { + "bbox": [ + 158, + 638, + 177, + 648 + ], + "score": 0.88, + "content": "c _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 635, + 320, + 650 + ], + "score": 1.0, + "content": "and linearly transformed again to a", + "type": "text" + }, + { + "bbox": [ + 320, + 636, + 326, + 646 + ], + "score": 0.79, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 635, + 407, + 650 + ], + "score": 1.0, + "content": "-dimensional vector", + "type": "text" + }, + { + "bbox": [ + 407, + 638, + 420, + 648 + ], + "score": 0.88, + "content": "c q _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 635, + 425, + 650 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 665, + 362, + 700 + ], + "lines": [ + { + "bbox": [ + 250, + 665, + 362, + 700 + ], + "spans": [ + { + "bbox": [ + 250, + 665, + 362, + 700 + ], + "score": 0.91, + "content": "\\begin{array} { c } { q _ { i } = W _ { i } ^ { d , d } \\cdot q + b _ { i } ^ { d } } \\\\ { c q _ { i } = W ^ { 2 d , d } \\left[ q _ { i } , c _ { i - 1 } \\right] + b ^ { d } } \\end{array}", + "type": "interline_equation", + "image_path": "629b6ac32640eb52fb3021fd3b45ba817455f27666f7d51a163756755ee964e5.jpg" + } + ] + } + ], + "index": 42.5, + "virtual_lines": [ + { + "bbox": [ + 250, + 665, + 362, + 682.5 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 250, + 682.5, + 362, + 700.0 + ], + "spans": [], + "index": 43 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 707, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 707, + 505, + 720 + ], + "spans": [ + { + "bbox": [ + 106, + 707, + 505, + 720 + ], + "score": 1.0, + "content": "Note that in contrast to all other parameters of the cell, which are shared across its instantiations", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 103, + 717, + 507, + 735 + ], + "spans": [ + { + "bbox": [ + 103, + 717, + 191, + 735 + ], + "score": 1.0, + "content": "at the different steps", + "type": "text" + }, + { + "bbox": [ + 192, + 721, + 237, + 732 + ], + "score": 0.91, + "content": "i = 1 , . . . , p", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 717, + 304, + 735 + ], + "score": 1.0, + "content": ", the parameters", + "type": "text" + }, + { + "bbox": [ + 304, + 720, + 327, + 733 + ], + "score": 0.93, + "content": "W _ { i } ^ { d , d }", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 717, + 346, + 735 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 346, + 721, + 356, + 733 + ], + "score": 0.89, + "content": "b _ { i } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 717, + 507, + 735 + ], + "score": 1.0, + "content": "are different for each iteration. This", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 215, + 82, + 370, + 198 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 215, + 82, + 370, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 215, + 82, + 370, + 198 + ], + "spans": [ + { + "bbox": [ + 215, + 82, + 370, + 198 + ], + "score": 0.966, + "type": "image", + "image_path": "21ef6959e8f5c66e79f26b2aa043aef1bf62c1ad19d7a86079e5194974b914be.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 215, + 82, + 370, + 96.5 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 215, + 96.5, + 370, + 111.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 215, + 111.0, + 370, + 125.5 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 215, + 125.5, + 370, + 140.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 215, + 140.0, + 370, + 154.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 215, + 154.5, + 370, + 169.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 215, + 169.0, + 370, + 183.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 215, + 183.5, + 370, + 198.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 208, + 505, + 231 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 209, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 221 + ], + "score": 1.0, + "content": "Figure 2: The MAC cell, which is a recurrent unit comprised of a Control Unit, Read Unit, and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "Write Unit. Blue shows the control flow and red shows the memory flow. See section 3.2 for details.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 108, + 252, + 504, + 285 + ], + "lines": [], + "index": 11, + "bbox_fs": [ + 105, + 251, + 505, + 286 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 290, + 505, + 368 + ], + "lines": [ + { + "bbox": [ + 106, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "The MAC cell has been designed to replace the discrete and predefined “modules” used in the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 301, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 505, + 314 + ], + "score": 1.0, + "content": "modular approach (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017). Rather, we", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "create one universal and versatile cell that is applied across all the reasoning steps, sharing both", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "its architecture as well as its parameters, across all of its instantiations. In contrast to the discrete", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 334, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 347 + ], + "score": 1.0, + "content": "modules, each trained to specialize to some specific elementary reasoning task, the MAC cell is", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "capable of demonstrating a continuous range of possible reasoning behaviors conditioned on the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 356, + 427, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 427, + 369 + ], + "score": 1.0, + "content": "context in which it is applied – namely, the inputs it receives from the prior cell.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 290, + 505, + 369 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 373, + 505, + 472 + ], + "lines": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 145, + 385 + ], + "score": 1.0, + "content": "Each cell", + "type": "text" + }, + { + "bbox": [ + 146, + 373, + 176, + 384 + ], + "score": 0.91, + "content": "M A C _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 373, + 314, + 385 + ], + "score": 1.0, + "content": "maintains two dual states: control", + "type": "text" + }, + { + "bbox": [ + 315, + 375, + 323, + 384 + ], + "score": 0.85, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 373, + 375, + 385 + ], + "score": 1.0, + "content": "and memory", + "type": "text" + }, + { + "bbox": [ + 376, + 375, + 389, + 384 + ], + "score": 0.85, + "content": "m _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 373, + 505, + 385 + ], + "score": 1.0, + "content": ", both are continuous vectors", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 384, + 504, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 161, + 396 + ], + "score": 1.0, + "content": "of dimension", + "type": "text" + }, + { + "bbox": [ + 162, + 385, + 168, + 394 + ], + "score": 0.7, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 384, + 221, + 396 + ], + "score": 1.0, + "content": ". The control", + "type": "text" + }, + { + "bbox": [ + 222, + 386, + 231, + 395 + ], + "score": 0.84, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 384, + 504, + 396 + ], + "score": 1.0, + "content": "represents the reasoning operation the MAC cell should accomplish", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "score": 1.0, + "content": "in the current step – focusing only on some aspect of the whole question. This is represented by", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 430, + 419 + ], + "score": 1.0, + "content": "a weighted-average attention-based sum of the question words. The memory", + "type": "text" + }, + { + "bbox": [ + 430, + 408, + 443, + 417 + ], + "score": 0.85, + "content": "m _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "represents the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 416, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 506, + 431 + ], + "score": 1.0, + "content": "current context information deemed relevant to respond to the query, or answer the question.This is", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "represented practically by a weighted average over elements from the KB, or for the case of VQA,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 439, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 194, + 452 + ], + "score": 1.0, + "content": "regions in the image.", + "type": "text" + }, + { + "bbox": [ + 195, + 441, + 209, + 450 + ], + "score": 0.85, + "content": "m _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 439, + 227, + 452 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 228, + 441, + 238, + 450 + ], + "score": 0.85, + "content": "c _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 439, + 495, + 452 + ], + "score": 1.0, + "content": "are initialized each to a random vector parameter of dimension", + "type": "text" + }, + { + "bbox": [ + 495, + 439, + 501, + 449 + ], + "score": 0.73, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 439, + 506, + 452 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "score": 1.0, + "content": "The memory and control states are passed from one cell to the next in a recurrent fashion, and used", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 461, + 484, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 484, + 474 + ], + "score": 1.0, + "content": "in a way reminiscent of Key-Value memory networks (Miller et al., 2016), as discussed below.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 373, + 506, + 474 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 484, + 228, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 230, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 230, + 497 + ], + "score": 1.0, + "content": "3.2.1 THE CONTROL UNIT", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 503, + 505, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "score": 1.0, + "content": "The control unit determines the reasoning operation that should be applied at this step. It receives", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 195, + 527 + ], + "score": 1.0, + "content": "the contextual words", + "type": "text" + }, + { + "bbox": [ + 195, + 514, + 252, + 526 + ], + "score": 0.92, + "content": "[ c w _ { 1 } , . . . , c w _ { S } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 514, + 369, + 527 + ], + "score": 1.0, + "content": ", the question representation", + "type": "text" + }, + { + "bbox": [ + 370, + 516, + 376, + 525 + ], + "score": 0.76, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 514, + 505, + 527 + ], + "score": 1.0, + "content": ", and the control state from the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 525, + 369, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 184, + 537 + ], + "score": 1.0, + "content": "previous MAC cell", + "type": "text" + }, + { + "bbox": [ + 185, + 527, + 204, + 537 + ], + "score": 0.88, + "content": "c _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 525, + 358, + 537 + ], + "score": 1.0, + "content": ", all of which are vectors of dimension", + "type": "text" + }, + { + "bbox": [ + 358, + 526, + 365, + 535 + ], + "score": 0.76, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 525, + 369, + 537 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 502, + 505, + 537 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 542, + 505, + 609 + ], + "lines": [ + { + "bbox": [ + 105, + 541, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 555 + ], + "score": 1.0, + "content": "We would like to allow our MAC cell to perform continuously varied and adaptive range of behav-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 553, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 565 + ], + "score": 1.0, + "content": "iors, as demanded by the question. Therefore, we define the behavior of each cell to be a function of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 564, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 193, + 577 + ], + "score": 1.0, + "content": "the contextual words", + "type": "text" + }, + { + "bbox": [ + 194, + 564, + 250, + 576 + ], + "score": 0.92, + "content": "[ c w _ { 1 } , . . . , c w _ { S } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 564, + 506, + 577 + ], + "score": 1.0, + "content": ", weighted-averaged according to the attention distribution that", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 576, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 587 + ], + "score": 1.0, + "content": "the control unit produces at each step. This will allow the cell to adapt its behavior – the reason-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "ing operation it performs – to the question it receives, instead of having a fixed set of predefined", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 596, + 500, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 500, + 610 + ], + "score": 1.0, + "content": "behaviours as is the case in competing approaches Andreas et al. (2016a;b); Johnson et al. (2017).", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 541, + 506, + 610 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 614, + 505, + 648 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 428, + 627 + ], + "score": 1.0, + "content": "The formal specification of the control unit is shown in figure 3. The question", + "type": "text" + }, + { + "bbox": [ + 429, + 616, + 435, + 626 + ], + "score": 0.8, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 613, + 505, + 627 + ], + "score": 1.0, + "content": "is linearly trans-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 626, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 194, + 637 + ], + "score": 1.0, + "content": "formed into a vector", + "type": "text" + }, + { + "bbox": [ + 195, + 627, + 204, + 636 + ], + "score": 0.84, + "content": "q _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 626, + 505, + 637 + ], + "score": 1.0, + "content": "of the same dimension, which in turn is concatenated with the previous", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 635, + 425, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 158, + 650 + ], + "score": 1.0, + "content": "control state", + "type": "text" + }, + { + "bbox": [ + 158, + 638, + 177, + 648 + ], + "score": 0.88, + "content": "c _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 635, + 320, + 650 + ], + "score": 1.0, + "content": "and linearly transformed again to a", + "type": "text" + }, + { + "bbox": [ + 320, + 636, + 326, + 646 + ], + "score": 0.79, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 635, + 407, + 650 + ], + "score": 1.0, + "content": "-dimensional vector", + "type": "text" + }, + { + "bbox": [ + 407, + 638, + 420, + 648 + ], + "score": 0.88, + "content": "c q _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 635, + 425, + 650 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 613, + 505, + 650 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 665, + 362, + 700 + ], + "lines": [ + { + "bbox": [ + 250, + 665, + 362, + 700 + ], + "spans": [ + { + "bbox": [ + 250, + 665, + 362, + 700 + ], + "score": 0.91, + "content": "\\begin{array} { c } { q _ { i } = W _ { i } ^ { d , d } \\cdot q + b _ { i } ^ { d } } \\\\ { c q _ { i } = W ^ { 2 d , d } \\left[ q _ { i } , c _ { i - 1 } \\right] + b ^ { d } } \\end{array}", + "type": "interline_equation", + "image_path": "629b6ac32640eb52fb3021fd3b45ba817455f27666f7d51a163756755ee964e5.jpg" + } + ] + } + ], + "index": 42.5, + "virtual_lines": [ + { + "bbox": [ + 250, + 665, + 362, + 682.5 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 250, + 682.5, + 362, + 700.0 + ], + "spans": [], + "index": 43 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 707, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 707, + 505, + 720 + ], + "spans": [ + { + "bbox": [ + 106, + 707, + 505, + 720 + ], + "score": 1.0, + "content": "Note that in contrast to all other parameters of the cell, which are shared across its instantiations", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 103, + 717, + 507, + 735 + ], + "spans": [ + { + "bbox": [ + 103, + 717, + 191, + 735 + ], + "score": 1.0, + "content": "at the different steps", + "type": "text" + }, + { + "bbox": [ + 192, + 721, + 237, + 732 + ], + "score": 0.91, + "content": "i = 1 , . . . , p", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 717, + 304, + 735 + ], + "score": 1.0, + "content": ", the parameters", + "type": "text" + }, + { + "bbox": [ + 304, + 720, + 327, + 733 + ], + "score": 0.93, + "content": "W _ { i } ^ { d , d }", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 717, + 346, + 735 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 346, + 721, + 356, + 733 + ], + "score": 0.89, + "content": "b _ { i } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 717, + 507, + 735 + ], + "score": 1.0, + "content": "are different for each iteration. This", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "score": 1.0, + "content": "is done to allow each cell to attend more readily to different aspects (i.e. parts) of the questions,", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 219, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 234 + ], + "score": 1.0, + "content": "depending on the index of the current step – its relative stage in the context of the whole reasoning", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 232, + 142, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 142, + 245 + ], + "score": 1.0, + "content": "process.", + "type": "text", + "cross_page": true + } + ], + "index": 9 + } + ], + "index": 44.5, + "bbox_fs": [ + 103, + 707, + 507, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 202, + 81, + 406, + 154 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 202, + 81, + 406, + 154 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 202, + 81, + 406, + 154 + ], + "spans": [ + { + "bbox": [ + 202, + 81, + 406, + 154 + ], + "score": 0.96, + "type": "image", + "image_path": "432535676eeb69a82d7848d82a8a327d906d019f07db47ee24440c80b3beab34.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 202, + 81, + 406, + 95.6 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 202, + 95.6, + 406, + 110.19999999999999 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 202, + 110.19999999999999, + 406, + 124.79999999999998 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 202, + 124.79999999999998, + 406, + 139.39999999999998 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 202, + 139.39999999999998, + 406, + 153.99999999999997 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 164, + 504, + 186 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 164, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 506, + 176 + ], + "score": 1.0, + "content": "Figure 3: The Control Unit (CU) of the MAC cell. See section 3.2.1 for details. Best viewed in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 173, + 133, + 188 + ], + "spans": [ + { + "bbox": [ + 104, + 173, + 133, + 188 + ], + "score": 1.0, + "content": "color.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + } + ], + "index": 3.75 + }, + { + "type": "text", + "bbox": [ + 107, + 209, + 505, + 243 + ], + "lines": [ + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "score": 1.0, + "content": "is done to allow each cell to attend more readily to different aspects (i.e. parts) of the questions,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 219, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 234 + ], + "score": 1.0, + "content": "depending on the index of the current step – its relative stage in the context of the whole reasoning", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 232, + 142, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 142, + 245 + ], + "score": 1.0, + "content": "process.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 248, + 505, + 282 + ], + "lines": [ + { + "bbox": [ + 107, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 107, + 250, + 120, + 260 + ], + "score": 0.84, + "content": "c q _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "represents the current reasoning operation we would like to perform in a continuous way, taking", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 259, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 321, + 271 + ], + "score": 1.0, + "content": "into account both the overall meaning of the question", + "type": "text" + }, + { + "bbox": [ + 321, + 261, + 330, + 271 + ], + "score": 0.83, + "content": "q _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 259, + 505, + 271 + ], + "score": 1.0, + "content": ", as well as the words the model attended to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 270, + 213, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 189, + 284 + ], + "score": 1.0, + "content": "in the previous step,", + "type": "text" + }, + { + "bbox": [ + 189, + 272, + 208, + 282 + ], + "score": 0.86, + "content": "c _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 270, + 213, + 284 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 286, + 505, + 364 + ], + "lines": [ + { + "bbox": [ + 105, + 287, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 506, + 300 + ], + "score": 1.0, + "content": "However, we would like to prevent the cell from diverging in the reasoning operations it tries to", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 297, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 311 + ], + "score": 1.0, + "content": "perform, and instead anchor it back in the question words, by using them to represent the reasoning", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 308, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 469, + 322 + ], + "score": 1.0, + "content": "operation of the current step. We can achieve that by computing an attention distribution", + "type": "text" + }, + { + "bbox": [ + 469, + 311, + 483, + 320 + ], + "score": 0.85, + "content": "c v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 308, + 506, + 322 + ], + "score": 1.0, + "content": "over", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 193, + 333 + ], + "score": 1.0, + "content": "the contextual words", + "type": "text" + }, + { + "bbox": [ + 193, + 320, + 250, + 332 + ], + "score": 0.93, + "content": "[ c w _ { 1 } , . . . , c w _ { S } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 320, + 362, + 333 + ], + "score": 1.0, + "content": "based on their similarity to", + "type": "text" + }, + { + "bbox": [ + 362, + 321, + 375, + 332 + ], + "score": 0.87, + "content": "c q _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 320, + 505, + 333 + ], + "score": 1.0, + "content": ". Then, summing the contextual", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 331, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 285, + 343 + ], + "score": 1.0, + "content": "words according to the attention distribution", + "type": "text" + }, + { + "bbox": [ + 286, + 332, + 300, + 342 + ], + "score": 0.87, + "content": "c v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 331, + 465, + 343 + ], + "score": 1.0, + "content": "will allow us to have a new control state,", + "type": "text" + }, + { + "bbox": [ + 465, + 333, + 474, + 342 + ], + "score": 0.83, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 331, + 505, + 343 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "is represented again in terms of words from the question. Intuitively, it is the gist of the question that", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 351, + 436, + 367 + ], + "spans": [ + { + "bbox": [ + 104, + 351, + 436, + 367 + ], + "score": 1.0, + "content": "is relevant to the reasoning operation we would like to perform in the current step.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 385, + 387, + 437 + ], + "lines": [ + { + "bbox": [ + 223, + 385, + 387, + 437 + ], + "spans": [ + { + "bbox": [ + 223, + 385, + 387, + 437 + ], + "score": 0.93, + "content": "\\begin{array} { c l c r } { { c v _ { i , s } = \\displaystyle \\mathrm { s o f t m a x } \\big ( W ^ { d , 1 } \\big ( c q _ { i } \\circ c w _ { s } \\big ) + b ^ { 1 } \\big ) } } \\\\ { { \\big . } } \\\\ { { c _ { i } = \\displaystyle \\sum _ { s = 1 } ^ { S } c v _ { i , s } \\cdot c w _ { s } } } \\end{array}", + "type": "interline_equation", + "image_path": "d43f433dbdaf70d728243b740937cc4b7c74be9d6cda3d636612e146da07d190.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 223, + 385, + 387, + 398.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 223, + 398.0, + 387, + 411.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 223, + 411.0, + 387, + 424.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 223, + 424.0, + 387, + 437.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 448, + 504, + 470 + ], + "lines": [ + { + "bbox": [ + 106, + 448, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 329, + 460 + ], + "score": 1.0, + "content": "Finally, the control unit returns the current control state", + "type": "text" + }, + { + "bbox": [ + 329, + 451, + 338, + 459 + ], + "score": 0.86, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 448, + 455, + 460 + ], + "score": 1.0, + "content": ", along with an attention map", + "type": "text" + }, + { + "bbox": [ + 455, + 451, + 469, + 459 + ], + "score": 0.85, + "content": "c v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 448, + 505, + 460 + ], + "score": 1.0, + "content": "over the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 459, + 179, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 179, + 470 + ], + "score": 1.0, + "content": "contextual words.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 108, + 483, + 211, + 495 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 213, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 213, + 497 + ], + "score": 1.0, + "content": "3.2.2 THE READ UNIT", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 503, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 347, + 516 + ], + "score": 1.0, + "content": "The Read Unit is provided with access to the knowledge base", + "type": "text" + }, + { + "bbox": [ + 348, + 504, + 372, + 514 + ], + "score": 0.9, + "content": "K B _ { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 502, + 505, + 516 + ], + "score": 1.0, + "content": ", along with the previous memory", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 128, + 527 + ], + "score": 1.0, + "content": "state", + "type": "text" + }, + { + "bbox": [ + 128, + 516, + 152, + 526 + ], + "score": 0.91, + "content": "m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 515, + 253, + 527 + ], + "score": 1.0, + "content": "and the current control", + "type": "text" + }, + { + "bbox": [ + 253, + 516, + 262, + 525 + ], + "score": 0.83, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 515, + 505, + 527 + ], + "score": 1.0, + "content": ". It is responsible for retrieving relevant content from the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 524, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 178, + 538 + ], + "score": 1.0, + "content": "Knowledge Base", + "type": "text" + }, + { + "bbox": [ + 178, + 525, + 204, + 536 + ], + "score": 0.91, + "content": "K B _ { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 524, + 505, + 538 + ], + "score": 1.0, + "content": "for the reasoning task that the MAC cell should accomplish at this step,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 536, + 502, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 300, + 549 + ], + "score": 1.0, + "content": "which is represented by the current control state", + "type": "text" + }, + { + "bbox": [ + 300, + 538, + 309, + 547 + ], + "score": 0.84, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 536, + 502, + 549 + ], + "score": 1.0, + "content": ", as explained above. Figure 4 shows a diagram.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 552, + 505, + 619 + ], + "lines": [ + { + "bbox": [ + 106, + 552, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 506, + 565 + ], + "score": 1.0, + "content": "The relevance of the new information is judged in two stages by the “relatedness” of each element", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 446, + 577 + ], + "score": 1.0, + "content": "in the KB (or for the case of VQA, each region in the image) to either the memory", + "type": "text" + }, + { + "bbox": [ + 446, + 565, + 469, + 576 + ], + "score": 0.9, + "content": "m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 563, + 505, + 577 + ], + "score": 1.0, + "content": "that has", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 575, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 442, + 588 + ], + "score": 1.0, + "content": "accumulated relevant information from previous iterations, or to the current control", + "type": "text" + }, + { + "bbox": [ + 443, + 577, + 451, + 586 + ], + "score": 0.84, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 575, + 505, + 588 + ], + "score": 1.0, + "content": ", pointing to-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "wards the next piece of information that should be taken into account. Here, relatedness is measured", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 597, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 506, + 610 + ], + "score": 1.0, + "content": "by trained linear transformations comparing each element to the previous memory and the current", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 608, + 140, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 140, + 620 + ], + "score": 1.0, + "content": "control.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 624, + 504, + 648 + ], + "lines": [ + { + "bbox": [ + 105, + 623, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 414, + 638 + ], + "score": 1.0, + "content": "More formally, at the first stage, the interaction between each element", + "type": "text" + }, + { + "bbox": [ + 414, + 626, + 446, + 637 + ], + "score": 0.92, + "content": "K B _ { h , w }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 623, + 482, + 638 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 482, + 625, + 505, + 637 + ], + "score": 0.85, + "content": "\\begin{array} { r l } { h } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 107, + 635, + 394, + 649 + ], + "spans": [ + { + "bbox": [ + 107, + 636, + 196, + 648 + ], + "score": 0.91, + "content": "1 , . . . , H , w = 1 , . . . , W", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 635, + 304, + 649 + ], + "score": 1.0, + "content": ", and the previous memory", + "type": "text" + }, + { + "bbox": [ + 304, + 637, + 327, + 648 + ], + "score": 0.89, + "content": "m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 635, + 394, + 649 + ], + "score": 1.0, + "content": "is computed by:", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 662, + 361, + 678 + ], + "lines": [ + { + "bbox": [ + 250, + 662, + 361, + 678 + ], + "spans": [ + { + "bbox": [ + 250, + 662, + 361, + 678 + ], + "score": 0.91, + "content": "m _ { i - 1 } ^ { \\prime } = W ^ { d , d } \\cdot m _ { i - 1 } + b ^ { d }", + "type": "interline_equation", + "image_path": "92d034ab00623a6201883d47cd297723d599fe90d135f474e1198ca4a4622a1a.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 250, + 662, + 361, + 678 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 696, + 370, + 732 + ], + "lines": [ + { + "bbox": [ + 240, + 696, + 370, + 732 + ], + "spans": [ + { + "bbox": [ + 240, + 696, + 370, + 732 + ], + "score": 0.91, + "content": "\\begin{array} { r l } { \\boldsymbol { K } \\boldsymbol { B } _ { h , w } ^ { \\prime } = \\boldsymbol { W } ^ { d , d } \\cdot \\boldsymbol { K } \\boldsymbol { B } _ { h , w } + \\boldsymbol { b } ^ { d } } & { { } } \\\\ { \\big ( \\boldsymbol { I } _ { m - K B } \\big ) _ { h , w } = \\boldsymbol { m } _ { i - 1 } ^ { \\prime } \\circ \\boldsymbol { K } \\boldsymbol { B } _ { h , w } ^ { \\prime } } & { { } } \\end{array}", + "type": "interline_equation", + "image_path": "ba795ad9270013570f22ffe5663537aa1ee717d610072223ec7afd2128e47f00.jpg" + } + ] + } + ], + "index": 40.5, + "virtual_lines": [ + { + "bbox": [ + 240, + 696, + 370, + 714.0 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 240, + 714.0, + 370, + 732.0 + ], + "spans": [], + "index": 41 + } + ] + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 202, + 81, + 406, + 154 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 202, + 81, + 406, + 154 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 202, + 81, + 406, + 154 + ], + "spans": [ + { + "bbox": [ + 202, + 81, + 406, + 154 + ], + "score": 0.96, + "type": "image", + "image_path": "432535676eeb69a82d7848d82a8a327d906d019f07db47ee24440c80b3beab34.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 202, + 81, + 406, + 95.6 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 202, + 95.6, + 406, + 110.19999999999999 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 202, + 110.19999999999999, + 406, + 124.79999999999998 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 202, + 124.79999999999998, + 406, + 139.39999999999998 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 202, + 139.39999999999998, + 406, + 153.99999999999997 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 164, + 504, + 186 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 164, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 506, + 176 + ], + "score": 1.0, + "content": "Figure 3: The Control Unit (CU) of the MAC cell. See section 3.2.1 for details. Best viewed in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 173, + 133, + 188 + ], + "spans": [ + { + "bbox": [ + 104, + 173, + 133, + 188 + ], + "score": 1.0, + "content": "color.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + } + ], + "index": 3.75 + }, + { + "type": "text", + "bbox": [ + 107, + 209, + 505, + 243 + ], + "lines": [], + "index": 8, + "bbox_fs": [ + 105, + 208, + 505, + 245 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 248, + 505, + 282 + ], + "lines": [ + { + "bbox": [ + 107, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 107, + 250, + 120, + 260 + ], + "score": 0.84, + "content": "c q _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "represents the current reasoning operation we would like to perform in a continuous way, taking", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 259, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 321, + 271 + ], + "score": 1.0, + "content": "into account both the overall meaning of the question", + "type": "text" + }, + { + "bbox": [ + 321, + 261, + 330, + 271 + ], + "score": 0.83, + "content": "q _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 259, + 505, + 271 + ], + "score": 1.0, + "content": ", as well as the words the model attended to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 270, + 213, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 189, + 284 + ], + "score": 1.0, + "content": "in the previous step,", + "type": "text" + }, + { + "bbox": [ + 189, + 272, + 208, + 282 + ], + "score": 0.86, + "content": "c _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 270, + 213, + 284 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 248, + 505, + 284 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 286, + 505, + 364 + ], + "lines": [ + { + "bbox": [ + 105, + 287, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 506, + 300 + ], + "score": 1.0, + "content": "However, we would like to prevent the cell from diverging in the reasoning operations it tries to", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 297, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 311 + ], + "score": 1.0, + "content": "perform, and instead anchor it back in the question words, by using them to represent the reasoning", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 308, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 469, + 322 + ], + "score": 1.0, + "content": "operation of the current step. We can achieve that by computing an attention distribution", + "type": "text" + }, + { + "bbox": [ + 469, + 311, + 483, + 320 + ], + "score": 0.85, + "content": "c v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 308, + 506, + 322 + ], + "score": 1.0, + "content": "over", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 193, + 333 + ], + "score": 1.0, + "content": "the contextual words", + "type": "text" + }, + { + "bbox": [ + 193, + 320, + 250, + 332 + ], + "score": 0.93, + "content": "[ c w _ { 1 } , . . . , c w _ { S } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 320, + 362, + 333 + ], + "score": 1.0, + "content": "based on their similarity to", + "type": "text" + }, + { + "bbox": [ + 362, + 321, + 375, + 332 + ], + "score": 0.87, + "content": "c q _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 320, + 505, + 333 + ], + "score": 1.0, + "content": ". Then, summing the contextual", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 331, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 285, + 343 + ], + "score": 1.0, + "content": "words according to the attention distribution", + "type": "text" + }, + { + "bbox": [ + 286, + 332, + 300, + 342 + ], + "score": 0.87, + "content": "c v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 331, + 465, + 343 + ], + "score": 1.0, + "content": "will allow us to have a new control state,", + "type": "text" + }, + { + "bbox": [ + 465, + 333, + 474, + 342 + ], + "score": 0.83, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 331, + 505, + 343 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "is represented again in terms of words from the question. Intuitively, it is the gist of the question that", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 351, + 436, + 367 + ], + "spans": [ + { + "bbox": [ + 104, + 351, + 436, + 367 + ], + "score": 1.0, + "content": "is relevant to the reasoning operation we would like to perform in the current step.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 104, + 287, + 506, + 367 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 385, + 387, + 437 + ], + "lines": [ + { + "bbox": [ + 223, + 385, + 387, + 437 + ], + "spans": [ + { + "bbox": [ + 223, + 385, + 387, + 437 + ], + "score": 0.93, + "content": "\\begin{array} { c l c r } { { c v _ { i , s } = \\displaystyle \\mathrm { s o f t m a x } \\big ( W ^ { d , 1 } \\big ( c q _ { i } \\circ c w _ { s } \\big ) + b ^ { 1 } \\big ) } } \\\\ { { \\big . } } \\\\ { { c _ { i } = \\displaystyle \\sum _ { s = 1 } ^ { S } c v _ { i , s } \\cdot c w _ { s } } } \\end{array}", + "type": "interline_equation", + "image_path": "d43f433dbdaf70d728243b740937cc4b7c74be9d6cda3d636612e146da07d190.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 223, + 385, + 387, + 398.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 223, + 398.0, + 387, + 411.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 223, + 411.0, + 387, + 424.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 223, + 424.0, + 387, + 437.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 448, + 504, + 470 + ], + "lines": [ + { + "bbox": [ + 106, + 448, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 329, + 460 + ], + "score": 1.0, + "content": "Finally, the control unit returns the current control state", + "type": "text" + }, + { + "bbox": [ + 329, + 451, + 338, + 459 + ], + "score": 0.86, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 448, + 455, + 460 + ], + "score": 1.0, + "content": ", along with an attention map", + "type": "text" + }, + { + "bbox": [ + 455, + 451, + 469, + 459 + ], + "score": 0.85, + "content": "c v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 448, + 505, + 460 + ], + "score": 1.0, + "content": "over the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 459, + 179, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 179, + 470 + ], + "score": 1.0, + "content": "contextual words.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 448, + 505, + 470 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 483, + 211, + 495 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 213, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 213, + 497 + ], + "score": 1.0, + "content": "3.2.2 THE READ UNIT", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 503, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 347, + 516 + ], + "score": 1.0, + "content": "The Read Unit is provided with access to the knowledge base", + "type": "text" + }, + { + "bbox": [ + 348, + 504, + 372, + 514 + ], + "score": 0.9, + "content": "K B _ { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 502, + 505, + 516 + ], + "score": 1.0, + "content": ", along with the previous memory", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 128, + 527 + ], + "score": 1.0, + "content": "state", + "type": "text" + }, + { + "bbox": [ + 128, + 516, + 152, + 526 + ], + "score": 0.91, + "content": "m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 515, + 253, + 527 + ], + "score": 1.0, + "content": "and the current control", + "type": "text" + }, + { + "bbox": [ + 253, + 516, + 262, + 525 + ], + "score": 0.83, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 515, + 505, + 527 + ], + "score": 1.0, + "content": ". It is responsible for retrieving relevant content from the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 524, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 178, + 538 + ], + "score": 1.0, + "content": "Knowledge Base", + "type": "text" + }, + { + "bbox": [ + 178, + 525, + 204, + 536 + ], + "score": 0.91, + "content": "K B _ { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 524, + 505, + 538 + ], + "score": 1.0, + "content": "for the reasoning task that the MAC cell should accomplish at this step,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 536, + 502, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 300, + 549 + ], + "score": 1.0, + "content": "which is represented by the current control state", + "type": "text" + }, + { + "bbox": [ + 300, + 538, + 309, + 547 + ], + "score": 0.84, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 536, + 502, + 549 + ], + "score": 1.0, + "content": ", as explained above. Figure 4 shows a diagram.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 502, + 505, + 549 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 552, + 505, + 619 + ], + "lines": [ + { + "bbox": [ + 106, + 552, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 506, + 565 + ], + "score": 1.0, + "content": "The relevance of the new information is judged in two stages by the “relatedness” of each element", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 446, + 577 + ], + "score": 1.0, + "content": "in the KB (or for the case of VQA, each region in the image) to either the memory", + "type": "text" + }, + { + "bbox": [ + 446, + 565, + 469, + 576 + ], + "score": 0.9, + "content": "m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 563, + 505, + 577 + ], + "score": 1.0, + "content": "that has", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 575, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 442, + 588 + ], + "score": 1.0, + "content": "accumulated relevant information from previous iterations, or to the current control", + "type": "text" + }, + { + "bbox": [ + 443, + 577, + 451, + 586 + ], + "score": 0.84, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 575, + 505, + 588 + ], + "score": 1.0, + "content": ", pointing to-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "wards the next piece of information that should be taken into account. Here, relatedness is measured", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 597, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 506, + 610 + ], + "score": 1.0, + "content": "by trained linear transformations comparing each element to the previous memory and the current", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 608, + 140, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 140, + 620 + ], + "score": 1.0, + "content": "control.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 552, + 506, + 620 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 624, + 504, + 648 + ], + "lines": [ + { + "bbox": [ + 105, + 623, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 414, + 638 + ], + "score": 1.0, + "content": "More formally, at the first stage, the interaction between each element", + "type": "text" + }, + { + "bbox": [ + 414, + 626, + 446, + 637 + ], + "score": 0.92, + "content": "K B _ { h , w }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 623, + 482, + 638 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 482, + 625, + 505, + 637 + ], + "score": 0.85, + "content": "\\begin{array} { r l } { h } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 107, + 635, + 394, + 649 + ], + "spans": [ + { + "bbox": [ + 107, + 636, + 196, + 648 + ], + "score": 0.91, + "content": "1 , . . . , H , w = 1 , . . . , W", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 635, + 304, + 649 + ], + "score": 1.0, + "content": ", and the previous memory", + "type": "text" + }, + { + "bbox": [ + 304, + 637, + 327, + 648 + ], + "score": 0.89, + "content": "m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 635, + 394, + 649 + ], + "score": 1.0, + "content": "is computed by:", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 623, + 505, + 649 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 662, + 361, + 678 + ], + "lines": [ + { + "bbox": [ + 250, + 662, + 361, + 678 + ], + "spans": [ + { + "bbox": [ + 250, + 662, + 361, + 678 + ], + "score": 0.91, + "content": "m _ { i - 1 } ^ { \\prime } = W ^ { d , d } \\cdot m _ { i - 1 } + b ^ { d }", + "type": "interline_equation", + "image_path": "92d034ab00623a6201883d47cd297723d599fe90d135f474e1198ca4a4622a1a.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 250, + 662, + 361, + 678 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 696, + 370, + 732 + ], + "lines": [ + { + "bbox": [ + 240, + 696, + 370, + 732 + ], + "spans": [ + { + "bbox": [ + 240, + 696, + 370, + 732 + ], + "score": 0.91, + "content": "\\begin{array} { r l } { \\boldsymbol { K } \\boldsymbol { B } _ { h , w } ^ { \\prime } = \\boldsymbol { W } ^ { d , d } \\cdot \\boldsymbol { K } \\boldsymbol { B } _ { h , w } + \\boldsymbol { b } ^ { d } } & { { } } \\\\ { \\big ( \\boldsymbol { I } _ { m - K B } \\big ) _ { h , w } = \\boldsymbol { m } _ { i - 1 } ^ { \\prime } \\circ \\boldsymbol { K } \\boldsymbol { B } _ { h , w } ^ { \\prime } } & { { } } \\end{array}", + "type": "interline_equation", + "image_path": "ba795ad9270013570f22ffe5663537aa1ee717d610072223ec7afd2128e47f00.jpg" + } + ] + } + ], + "index": 40.5, + "virtual_lines": [ + { + "bbox": [ + 240, + 696, + 370, + 714.0 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 240, + 714.0, + 370, + 732.0 + ], + "spans": [], + "index": 41 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 200, + 81, + 409, + 154 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 200, + 81, + 409, + 154 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 200, + 81, + 409, + 154 + ], + "spans": [ + { + "bbox": [ + 200, + 81, + 409, + 154 + ], + "score": 0.965, + "type": "image", + "image_path": "a538d781a85f8151d41cc83e31d872618ae93e0b206ef08dd299f4020e8135bf.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 200, + 81, + 409, + 95.6 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 200, + 95.6, + 409, + 110.19999999999999 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 200, + 110.19999999999999, + 409, + 124.79999999999998 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 200, + 124.79999999999998, + 409, + 139.39999999999998 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 200, + 139.39999999999998, + 409, + 153.99999999999997 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 165, + 504, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "Figure 4: The Read Unit (RU) diagram. Blue refers to control flow, purple to knowledge flow and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 176, + 323, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 323, + 189 + ], + "score": 1.0, + "content": "red to memory flow. See section 3.2.2 for description.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + } + ], + "index": 3.75 + }, + { + "type": "text", + "bbox": [ + 107, + 207, + 504, + 262 + ], + "lines": [ + { + "bbox": [ + 106, + 208, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 505, + 220 + ], + "score": 1.0, + "content": "These memory-KB interactions measure the relatedness of each element in the KB to the mem-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 218, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 506, + 231 + ], + "score": 1.0, + "content": "ory accumulated so far, which holds information that has been deemed relevant to handle previous", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 230, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 241 + ], + "score": 1.0, + "content": "reasoning steps towards addressing the question. They allow the model to perform transitive infer-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 239, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 104, + 239, + 506, + 255 + ], + "score": 1.0, + "content": "ence, retrieving a new piece of information that now seems important in light of the recent memory", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 252, + 221, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 221, + 264 + ], + "score": 1.0, + "content": "retrieved in a prior iteration.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 268, + 505, + 346 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "However, there are cases which necessitate the model to temporarily ignore current memories, when", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 280, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 505, + 291 + ], + "score": 1.0, + "content": "choosing the new information to retrieve. Logical OR is a classical example: when the model has", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "to look at two different objects at the same time, and assuming it stored one of them at the first", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 301, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 505, + 314 + ], + "score": 1.0, + "content": "iteration, it should briefly ignore it, considering new information that is relevant to the question", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "but is unrelated to the memory. In order to achieve such capability, the read unit concatenates the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "original KB elements to each corresponding memory-KB interaction, which are then projected back", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 334, + 259, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 117, + 347 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 117, + 335, + 123, + 344 + ], + "score": 0.79, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 334, + 259, + 347 + ], + "score": 1.0, + "content": "-dimensional space (equation 6a):", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 363, + 392, + 398 + ], + "lines": [ + { + "bbox": [ + 219, + 363, + 392, + 398 + ], + "spans": [ + { + "bbox": [ + 219, + 363, + 392, + 398 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { I _ { m - K B } , ^ { \\prime } = W ^ { 2 d , d } \\left[ I _ { m - K B } , K B _ { h , w } \\right] + b ^ { d } } } \\\\ { { I _ { c m - K B } = c _ { i } \\circ \\left( I _ { m - K B } \\right) ^ { \\prime } } } \\end{array}", + "type": "interline_equation", + "image_path": "ada1fba84ad0bac03e8e9a8f8cc72cbc8bba767e7414ee8b9b65dab9dee93066.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 219, + 363, + 392, + 380.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 219, + 380.5, + 392, + 398.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 405, + 504, + 460 + ], + "lines": [ + { + "bbox": [ + 105, + 405, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 333, + 418 + ], + "score": 1.0, + "content": "At the second stage, the read unit compares the current", + "type": "text" + }, + { + "bbox": [ + 334, + 407, + 343, + 416 + ], + "score": 0.85, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 405, + 506, + 418 + ], + "score": 1.0, + "content": "with these memory-KB interactions, in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 416, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 428 + ], + "score": 1.0, + "content": "order to focus on the information that is relevant to the current reasoning operation that the MAC", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 425, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 506, + 441 + ], + "score": 1.0, + "content": "cell seeks to accomplish. The result is then passed to a softmax layer yielding an attention map", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 437, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 125, + 449 + ], + "score": 0.85, + "content": "m v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 437, + 506, + 451 + ], + "score": 1.0, + "content": "over the KB, which is used in turn to retrieve the relevant information to perform the current", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 448, + 169, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 169, + 462 + ], + "score": 1.0, + "content": "reasoning step.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23 + }, + { + "type": "interline_equation", + "bbox": [ + 224, + 478, + 387, + 532 + ], + "lines": [ + { + "bbox": [ + 224, + 478, + 387, + 532 + ], + "spans": [ + { + "bbox": [ + 224, + 478, + 387, + 532 + ], + "score": 0.91, + "content": "\\begin{array} { r } { m v _ { i } = \\operatorname { s o f t m a x } \\left( W ^ { d , d } \\cdot I _ { c m - K B } + b ^ { d } \\right) } \\\\ { m _ { n e w } = \\displaystyle \\sum _ { h , w = 1 , 1 } ^ { H , W } \\left( m v _ { i } \\right) _ { h , w } \\cdot K B _ { h , w } } \\end{array}", + "type": "interline_equation", + "image_path": "1af356ca707634550da691f9e1c5befb2540a27633c9abfa602dcd4cdf1d0ea5.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 224, + 478, + 387, + 491.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 224, + 491.5, + 387, + 505.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 224, + 505.0, + 387, + 518.5 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 224, + 518.5, + 387, + 532.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 541, + 504, + 563 + ], + "lines": [ + { + "bbox": [ + 105, + 538, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 358, + 555 + ], + "score": 1.0, + "content": "Finally, the read unit returns the newly retrieved information", + "type": "text" + }, + { + "bbox": [ + 358, + 543, + 383, + 552 + ], + "score": 0.89, + "content": "m _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 538, + 505, + 555 + ], + "score": 1.0, + "content": ", along with an attention map", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 552, + 233, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 125, + 563 + ], + "score": 0.84, + "content": "m v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 552, + 233, + 564 + ], + "score": 1.0, + "content": "over the Knowledge Base.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 568, + 505, + 678 + ], + "lines": [ + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 406, + 582 + ], + "score": 1.0, + "content": "To give an example of the read unit operation, assume a given question", + "type": "text" + }, + { + "bbox": [ + 406, + 571, + 413, + 580 + ], + "score": 0.76, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "such as “What object", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "is located left to the blue ball?”, whose associated answer is “cube”. Initially, no cue is provided", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "score": 1.0, + "content": "to the model to attend to that cube, since no direct information about it presents in the question.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 600, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 614 + ], + "score": 1.0, + "content": "Instead, based on its comprehension of the question, the model may start by focusing on the blue", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 324, + 625 + ], + "score": 1.0, + "content": "ball at the first iteration, such that the memory state", + "type": "text" + }, + { + "bbox": [ + 324, + 614, + 339, + 623 + ], + "score": 0.85, + "content": "m _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "will capture the blue ball. However, in", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 622, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 636 + ], + "score": 1.0, + "content": "the second iteration, the control unit, after re-examining the question, may realize it should now", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 250, + 647 + ], + "score": 1.0, + "content": "look left, storing the word “left” in", + "type": "text" + }, + { + "bbox": [ + 250, + 636, + 260, + 645 + ], + "score": 0.84, + "content": "c _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 634, + 385, + 647 + ], + "score": 1.0, + "content": ". Then, when considering both", + "type": "text" + }, + { + "bbox": [ + 386, + 635, + 401, + 645 + ], + "score": 0.86, + "content": "m _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 634, + 419, + 647 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 419, + 636, + 429, + 645 + ], + "score": 0.84, + "content": "c _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 634, + 505, + 647 + ], + "score": 1.0, + "content": ", the read unit will", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 466, + 658 + ], + "score": 1.0, + "content": "realize it should perform a reasoning operation corresponding to the word “left” (stored in", + "type": "text" + }, + { + "bbox": [ + 466, + 646, + 476, + 656 + ], + "score": 0.79, + "content": "c _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 645, + 505, + 658 + ], + "score": 1.0, + "content": ") given", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 657, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 657, + 264, + 668 + ], + "score": 1.0, + "content": "a memory representing the blue ball in", + "type": "text" + }, + { + "bbox": [ + 264, + 658, + 278, + 667 + ], + "score": 0.87, + "content": "m _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 657, + 506, + 668 + ], + "score": 1.0, + "content": ", thereby allowing it to look left to the blue ball and find", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 666, + 146, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 146, + 679 + ], + "score": 1.0, + "content": "the cube.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 36.5 + }, + { + "type": "title", + "bbox": [ + 108, + 690, + 215, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 217, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 217, + 703 + ], + "score": 1.0, + "content": "3.2.3 THE WRITE UNIT", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 364, + 722 + ], + "score": 1.0, + "content": "The Write Unit is responsible for creating the new memory state", + "type": "text" + }, + { + "bbox": [ + 364, + 711, + 377, + 721 + ], + "score": 0.86, + "content": "m _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "that will reflect all the informa-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "tion considered to be important to answer the question so far, i.e. up to the current iteration in the", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 200, + 81, + 409, + 154 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 200, + 81, + 409, + 154 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 200, + 81, + 409, + 154 + ], + "spans": [ + { + "bbox": [ + 200, + 81, + 409, + 154 + ], + "score": 0.965, + "type": "image", + "image_path": "a538d781a85f8151d41cc83e31d872618ae93e0b206ef08dd299f4020e8135bf.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 200, + 81, + 409, + 95.6 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 200, + 95.6, + 409, + 110.19999999999999 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 200, + 110.19999999999999, + 409, + 124.79999999999998 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 200, + 124.79999999999998, + 409, + 139.39999999999998 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 200, + 139.39999999999998, + 409, + 153.99999999999997 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 165, + 504, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "Figure 4: The Read Unit (RU) diagram. Blue refers to control flow, purple to knowledge flow and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 176, + 323, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 323, + 189 + ], + "score": 1.0, + "content": "red to memory flow. See section 3.2.2 for description.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + } + ], + "index": 3.75 + }, + { + "type": "text", + "bbox": [ + 107, + 207, + 504, + 262 + ], + "lines": [ + { + "bbox": [ + 106, + 208, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 505, + 220 + ], + "score": 1.0, + "content": "These memory-KB interactions measure the relatedness of each element in the KB to the mem-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 218, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 506, + 231 + ], + "score": 1.0, + "content": "ory accumulated so far, which holds information that has been deemed relevant to handle previous", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 230, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 241 + ], + "score": 1.0, + "content": "reasoning steps towards addressing the question. They allow the model to perform transitive infer-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 239, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 104, + 239, + 506, + 255 + ], + "score": 1.0, + "content": "ence, retrieving a new piece of information that now seems important in light of the recent memory", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 252, + 221, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 221, + 264 + ], + "score": 1.0, + "content": "retrieved in a prior iteration.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 104, + 208, + 506, + 264 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 268, + 505, + 346 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "However, there are cases which necessitate the model to temporarily ignore current memories, when", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 280, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 505, + 291 + ], + "score": 1.0, + "content": "choosing the new information to retrieve. Logical OR is a classical example: when the model has", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "to look at two different objects at the same time, and assuming it stored one of them at the first", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 301, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 505, + 314 + ], + "score": 1.0, + "content": "iteration, it should briefly ignore it, considering new information that is relevant to the question", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "but is unrelated to the memory. In order to achieve such capability, the read unit concatenates the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "original KB elements to each corresponding memory-KB interaction, which are then projected back", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 334, + 259, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 117, + 347 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 117, + 335, + 123, + 344 + ], + "score": 0.79, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 334, + 259, + 347 + ], + "score": 1.0, + "content": "-dimensional space (equation 6a):", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 268, + 505, + 347 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 363, + 392, + 398 + ], + "lines": [ + { + "bbox": [ + 219, + 363, + 392, + 398 + ], + "spans": [ + { + "bbox": [ + 219, + 363, + 392, + 398 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { I _ { m - K B } , ^ { \\prime } = W ^ { 2 d , d } \\left[ I _ { m - K B } , K B _ { h , w } \\right] + b ^ { d } } } \\\\ { { I _ { c m - K B } = c _ { i } \\circ \\left( I _ { m - K B } \\right) ^ { \\prime } } } \\end{array}", + "type": "interline_equation", + "image_path": "ada1fba84ad0bac03e8e9a8f8cc72cbc8bba767e7414ee8b9b65dab9dee93066.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 219, + 363, + 392, + 380.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 219, + 380.5, + 392, + 398.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 405, + 504, + 460 + ], + "lines": [ + { + "bbox": [ + 105, + 405, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 333, + 418 + ], + "score": 1.0, + "content": "At the second stage, the read unit compares the current", + "type": "text" + }, + { + "bbox": [ + 334, + 407, + 343, + 416 + ], + "score": 0.85, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 405, + 506, + 418 + ], + "score": 1.0, + "content": "with these memory-KB interactions, in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 416, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 428 + ], + "score": 1.0, + "content": "order to focus on the information that is relevant to the current reasoning operation that the MAC", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 425, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 506, + 441 + ], + "score": 1.0, + "content": "cell seeks to accomplish. The result is then passed to a softmax layer yielding an attention map", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 437, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 125, + 449 + ], + "score": 0.85, + "content": "m v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 437, + 506, + 451 + ], + "score": 1.0, + "content": "over the KB, which is used in turn to retrieve the relevant information to perform the current", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 448, + 169, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 169, + 462 + ], + "score": 1.0, + "content": "reasoning step.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 405, + 506, + 462 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 224, + 478, + 387, + 532 + ], + "lines": [ + { + "bbox": [ + 224, + 478, + 387, + 532 + ], + "spans": [ + { + "bbox": [ + 224, + 478, + 387, + 532 + ], + "score": 0.91, + "content": "\\begin{array} { r } { m v _ { i } = \\operatorname { s o f t m a x } \\left( W ^ { d , d } \\cdot I _ { c m - K B } + b ^ { d } \\right) } \\\\ { m _ { n e w } = \\displaystyle \\sum _ { h , w = 1 , 1 } ^ { H , W } \\left( m v _ { i } \\right) _ { h , w } \\cdot K B _ { h , w } } \\end{array}", + "type": "interline_equation", + "image_path": "1af356ca707634550da691f9e1c5befb2540a27633c9abfa602dcd4cdf1d0ea5.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 224, + 478, + 387, + 491.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 224, + 491.5, + 387, + 505.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 224, + 505.0, + 387, + 518.5 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 224, + 518.5, + 387, + 532.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 541, + 504, + 563 + ], + "lines": [ + { + "bbox": [ + 105, + 538, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 358, + 555 + ], + "score": 1.0, + "content": "Finally, the read unit returns the newly retrieved information", + "type": "text" + }, + { + "bbox": [ + 358, + 543, + 383, + 552 + ], + "score": 0.89, + "content": "m _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 538, + 505, + 555 + ], + "score": 1.0, + "content": ", along with an attention map", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 552, + 233, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 125, + 563 + ], + "score": 0.84, + "content": "m v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 552, + 233, + 564 + ], + "score": 1.0, + "content": "over the Knowledge Base.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 538, + 505, + 564 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 568, + 505, + 678 + ], + "lines": [ + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 406, + 582 + ], + "score": 1.0, + "content": "To give an example of the read unit operation, assume a given question", + "type": "text" + }, + { + "bbox": [ + 406, + 571, + 413, + 580 + ], + "score": 0.76, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "such as “What object", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "is located left to the blue ball?”, whose associated answer is “cube”. Initially, no cue is provided", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "score": 1.0, + "content": "to the model to attend to that cube, since no direct information about it presents in the question.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 600, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 614 + ], + "score": 1.0, + "content": "Instead, based on its comprehension of the question, the model may start by focusing on the blue", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 324, + 625 + ], + "score": 1.0, + "content": "ball at the first iteration, such that the memory state", + "type": "text" + }, + { + "bbox": [ + 324, + 614, + 339, + 623 + ], + "score": 0.85, + "content": "m _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "will capture the blue ball. However, in", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 622, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 636 + ], + "score": 1.0, + "content": "the second iteration, the control unit, after re-examining the question, may realize it should now", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 250, + 647 + ], + "score": 1.0, + "content": "look left, storing the word “left” in", + "type": "text" + }, + { + "bbox": [ + 250, + 636, + 260, + 645 + ], + "score": 0.84, + "content": "c _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 634, + 385, + 647 + ], + "score": 1.0, + "content": ". Then, when considering both", + "type": "text" + }, + { + "bbox": [ + 386, + 635, + 401, + 645 + ], + "score": 0.86, + "content": "m _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 634, + 419, + 647 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 419, + 636, + 429, + 645 + ], + "score": 0.84, + "content": "c _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 634, + 505, + 647 + ], + "score": 1.0, + "content": ", the read unit will", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 466, + 658 + ], + "score": 1.0, + "content": "realize it should perform a reasoning operation corresponding to the word “left” (stored in", + "type": "text" + }, + { + "bbox": [ + 466, + 646, + 476, + 656 + ], + "score": 0.79, + "content": "c _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 645, + 505, + 658 + ], + "score": 1.0, + "content": ") given", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 657, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 657, + 264, + 668 + ], + "score": 1.0, + "content": "a memory representing the blue ball in", + "type": "text" + }, + { + "bbox": [ + 264, + 658, + 278, + 667 + ], + "score": 0.87, + "content": "m _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 657, + 506, + 668 + ], + "score": 1.0, + "content": ", thereby allowing it to look left to the blue ball and find", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 666, + 146, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 146, + 679 + ], + "score": 1.0, + "content": "the cube.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 36.5, + "bbox_fs": [ + 104, + 568, + 506, + 679 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 690, + 215, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 217, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 217, + 703 + ], + "score": 1.0, + "content": "3.2.3 THE WRITE UNIT", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 364, + 722 + ], + "score": 1.0, + "content": "The Write Unit is responsible for creating the new memory state", + "type": "text" + }, + { + "bbox": [ + 364, + 711, + 377, + 721 + ], + "score": 0.86, + "content": "m _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "that will reflect all the informa-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "tion considered to be important to answer the question so far, i.e. up to the current iteration in the", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 225, + 81, + 387, + 157 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 225, + 81, + 387, + 157 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 225, + 81, + 387, + 157 + ], + "spans": [ + { + "bbox": [ + 225, + 81, + 387, + 157 + ], + "score": 0.96, + "type": "image", + "image_path": "da3df6ec5443731e5ef4aaf8424acd7af4458f638b9bc19f84b7d1ed88b9267e.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 225, + 81, + 387, + 96.2 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 225, + 96.2, + 387, + 111.4 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 225, + 111.4, + 387, + 126.60000000000001 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 225, + 126.60000000000001, + 387, + 141.8 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 225, + 141.8, + 387, + 157.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 168, + 504, + 191 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 167, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 505, + 181 + ], + "score": 1.0, + "content": "Figure 5: The Write Unit (WU) diagram. Blue refers to control flow and red to memory flow. See", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 177, + 223, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 223, + 192 + ], + "score": 1.0, + "content": "section 3.2.3 for description.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + } + ], + "index": 3.75 + }, + { + "type": "text", + "bbox": [ + 107, + 209, + 505, + 243 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 316, + 222 + ], + "score": 1.0, + "content": "reasoning process. It receives the last memory state", + "type": "text" + }, + { + "bbox": [ + 316, + 212, + 340, + 222 + ], + "score": 0.89, + "content": "m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "from the previous MAC cell, along with", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 220, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 403, + 234 + ], + "score": 1.0, + "content": "the newly retrieved information from the read unit in the current iteration,", + "type": "text" + }, + { + "bbox": [ + 404, + 222, + 429, + 232 + ], + "score": 0.89, + "content": "m _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 220, + 506, + 234 + ], + "score": 1.0, + "content": ". See figure 5 for a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 230, + 145, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 145, + 245 + ], + "score": 1.0, + "content": "diagram.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 248, + 504, + 271 + ], + "lines": [ + { + "bbox": [ + 105, + 247, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 261 + ], + "score": 1.0, + "content": "In the main design we have explored, merging the new information with the previous memory state", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 259, + 275, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 275, + 272 + ], + "score": 1.0, + "content": "is done simply by a linear transformation.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 285, + 371, + 300 + ], + "lines": [ + { + "bbox": [ + 240, + 285, + 371, + 300 + ], + "spans": [ + { + "bbox": [ + 240, + 285, + 371, + 300 + ], + "score": 0.93, + "content": "m _ { i } ^ { \\prime } = W ^ { 2 d , d } [ m _ { n e w } , m _ { i - 1 } ] + b ^ { d }", + "type": "interline_equation", + "image_path": "fae487b9caae60fb93edbf054820528c6b18061cf0a3ab7e592c6d4fbd281b4a.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 240, + 285, + 371, + 300 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 306, + 505, + 372 + ], + "lines": [ + { + "bbox": [ + 106, + 306, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 505, + 319 + ], + "score": 1.0, + "content": "In addition, we have explored two variations of this design. The first, self-attention, allows consid-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 343, + 330 + ], + "score": 1.0, + "content": "ering any previous memories rather than just the last one", + "type": "text" + }, + { + "bbox": [ + 343, + 319, + 366, + 329 + ], + "score": 0.9, + "content": "m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 317, + 505, + 330 + ], + "score": 1.0, + "content": ", thus providing the network with", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 329, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 505, + 341 + ], + "score": 1.0, + "content": "the capacity to model non-sequential reasoning processes. The second variation is adding gating", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 340, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 351 + ], + "score": 1.0, + "content": "mechanisms to the writing unit. These may allow the model to dynamically adjust the practical", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 349, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 363 + ], + "score": 1.0, + "content": "length of the computation to the question complexity and stabilize the memory content throughout", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 362, + 333, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 333, + 373 + ], + "score": 1.0, + "content": "the sequential network (similarly to GRUs and LSTMs).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 384, + 505, + 462 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "Self-Attention. The current architecture that we have presented allows the model to perform rea-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 396, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 407 + ], + "score": 1.0, + "content": "soning steps in a sequence, passing control and memory states from one cell to the following. How-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "ever, we would like to grant the system with more flexibility. Particularly, we would like to allow it", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "to capture more complicated reasoning processes such as trees and graphs - Directed Acyclic Graph", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 506, + 441 + ], + "score": 1.0, + "content": "(DAG) in particular, where several branches of reasoning sub-processes are merged together in later", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 438, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 104, + 438, + 506, + 453 + ], + "score": 1.0, + "content": "stages. Indeed, the CLEVR dataset includes cases where the questions embody tree-like reasoning", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 450, + 471, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 471, + 462 + ], + "score": 1.0, + "content": "process, rather than just sequences, which we would like to address correctly in our model.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 467, + 505, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 466, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 478 + ], + "score": 1.0, + "content": "We achieve that by adding self-attention connections between each MAC cell and all the prior cells.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 478, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 505, + 490 + ], + "score": 1.0, + "content": "Since each cell can look on all the prior reasoning steps and their corresponding memories retrieved", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "from the Knowledge Base, it can virtually capture any directed acyclic graph, while still having", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 500, + 223, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 223, + 513 + ], + "score": 1.0, + "content": "physically sequential layout.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 515, + 505, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 514, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 283, + 531 + ], + "score": 1.0, + "content": "More formally, the current MAC cell, of the", + "type": "text" + }, + { + "bbox": [ + 283, + 516, + 296, + 527 + ], + "score": 0.89, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 514, + 434, + 531 + ], + "score": 1.0, + "content": "iteration, is granted with access to", + "type": "text" + }, + { + "bbox": [ + 434, + 519, + 479, + 528 + ], + "score": 0.9, + "content": "c _ { 1 } , . . . , c _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 514, + 506, + 531 + ], + "score": 1.0, + "content": "along", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 527, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 200, + 541 + ], + "score": 1.0, + "content": "with the corresponding", + "type": "text" + }, + { + "bbox": [ + 200, + 529, + 254, + 540 + ], + "score": 0.89, + "content": "m _ { 1 } , . . . , m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 527, + 505, + 541 + ], + "score": 1.0, + "content": ", that have been computed by the prior MAC cells. It begins by", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 538, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 244, + 551 + ], + "score": 1.0, + "content": "computing the similarity between", + "type": "text" + }, + { + "bbox": [ + 244, + 541, + 253, + 550 + ], + "score": 0.84, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 538, + 271, + 551 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 272, + 540, + 316, + 550 + ], + "score": 0.89, + "content": "c _ { 1 } , . . . , c _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 538, + 505, + 551 + ], + "score": 1.0, + "content": ", and use it to derive an attention map over the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 546, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 104, + 546, + 174, + 565 + ], + "score": 1.0, + "content": "prior MAC cells", + "type": "text" + }, + { + "bbox": [ + 175, + 551, + 195, + 562 + ], + "score": 0.88, + "content": "s a _ { i , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 546, + 212, + 565 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 212, + 550, + 275, + 561 + ], + "score": 0.91, + "content": "j = 1 , . . . , i - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 546, + 425, + 565 + ], + "score": 1.0, + "content": ". This represents the relevance of the", + "type": "text" + }, + { + "bbox": [ + 426, + 549, + 440, + 561 + ], + "score": 0.9, + "content": "j ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 546, + 506, + 565 + ], + "score": 1.0, + "content": "prior reasoning", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 560, + 261, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 197, + 573 + ], + "score": 1.0, + "content": "step to the current one", + "type": "text" + }, + { + "bbox": [ + 198, + 562, + 202, + 570 + ], + "score": 0.73, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 560, + 261, + 573 + ], + "score": 1.0, + "content": "(equation 9a).", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 108, + 577, + 503, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 437, + 590 + ], + "score": 1.0, + "content": "Then, we average the previous memories according to this resulted attention map", + "type": "text" + }, + { + "bbox": [ + 438, + 579, + 456, + 590 + ], + "score": 0.91, + "content": "s a _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 577, + 505, + 590 + ], + "score": 1.0, + "content": ". We obtain", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 588, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 125, + 600 + ], + "score": 0.85, + "content": "m _ { s a }", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 588, + 506, + 601 + ], + "score": 1.0, + "content": ", representing the information from all the other reasoning steps that is relevant to the current", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 600, + 182, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 182, + 612 + ], + "score": 1.0, + "content": "one (equation 9b).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 505, + 661 + ], + "lines": [ + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "This resembles the approach of Key-Value networks (Miller et al., 2016). The similarity between", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "control states, corresponding to the reasoning operations that are performed in each prior step, allows", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "the model to select which memories should be taken into account, when creating the new memory –", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 649, + 464, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 464, + 662 + ], + "score": 1.0, + "content": "namely, which branches of the reasoning process should be merged together at this point.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5 + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 678, + 381, + 731 + ], + "lines": [ + { + "bbox": [ + 230, + 678, + 381, + 731 + ], + "spans": [ + { + "bbox": [ + 230, + 678, + 381, + 731 + ], + "score": 0.88, + "content": "\\begin{array} { c } { { s a _ { i j } = \\displaystyle \\mathrm { s o f t m a x } \\left( W ^ { d , 1 } ( c _ { i } \\circ c _ { j } ) + b ^ { 1 } \\right) } } \\\\ { { ( m _ { s a } ) _ { i } = \\displaystyle \\sum _ { j = 1 } ^ { i - 1 } s a _ { i j } \\cdot m _ { j } } } \\end{array}", + "type": "interline_equation", + "image_path": "b2917add3d0ec9deb18d84cf0919d7c999da2714964aba59d6c8b9129667c984.jpg" + } + ] + } + ], + "index": 42.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 678, + 381, + 704.5 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 230, + 704.5, + 381, + 731.0 + ], + "spans": [], + "index": 43 + } + ] + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 7 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 225, + 81, + 387, + 157 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 225, + 81, + 387, + 157 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 225, + 81, + 387, + 157 + ], + "spans": [ + { + "bbox": [ + 225, + 81, + 387, + 157 + ], + "score": 0.96, + "type": "image", + "image_path": "da3df6ec5443731e5ef4aaf8424acd7af4458f638b9bc19f84b7d1ed88b9267e.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 225, + 81, + 387, + 96.2 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 225, + 96.2, + 387, + 111.4 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 225, + 111.4, + 387, + 126.60000000000001 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 225, + 126.60000000000001, + 387, + 141.8 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 225, + 141.8, + 387, + 157.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 168, + 504, + 191 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 167, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 505, + 181 + ], + "score": 1.0, + "content": "Figure 5: The Write Unit (WU) diagram. Blue refers to control flow and red to memory flow. See", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 177, + 223, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 223, + 192 + ], + "score": 1.0, + "content": "section 3.2.3 for description.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + } + ], + "index": 3.75 + }, + { + "type": "text", + "bbox": [ + 107, + 209, + 505, + 243 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 316, + 222 + ], + "score": 1.0, + "content": "reasoning process. It receives the last memory state", + "type": "text" + }, + { + "bbox": [ + 316, + 212, + 340, + 222 + ], + "score": 0.89, + "content": "m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "from the previous MAC cell, along with", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 220, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 403, + 234 + ], + "score": 1.0, + "content": "the newly retrieved information from the read unit in the current iteration,", + "type": "text" + }, + { + "bbox": [ + 404, + 222, + 429, + 232 + ], + "score": 0.89, + "content": "m _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 220, + 506, + 234 + ], + "score": 1.0, + "content": ". See figure 5 for a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 230, + 145, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 145, + 245 + ], + "score": 1.0, + "content": "diagram.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 209, + 506, + 245 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 248, + 504, + 271 + ], + "lines": [ + { + "bbox": [ + 105, + 247, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 261 + ], + "score": 1.0, + "content": "In the main design we have explored, merging the new information with the previous memory state", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 259, + 275, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 275, + 272 + ], + "score": 1.0, + "content": "is done simply by a linear transformation.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 247, + 505, + 272 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 285, + 371, + 300 + ], + "lines": [ + { + "bbox": [ + 240, + 285, + 371, + 300 + ], + "spans": [ + { + "bbox": [ + 240, + 285, + 371, + 300 + ], + "score": 0.93, + "content": "m _ { i } ^ { \\prime } = W ^ { 2 d , d } [ m _ { n e w } , m _ { i - 1 } ] + b ^ { d }", + "type": "interline_equation", + "image_path": "fae487b9caae60fb93edbf054820528c6b18061cf0a3ab7e592c6d4fbd281b4a.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 240, + 285, + 371, + 300 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 306, + 505, + 372 + ], + "lines": [ + { + "bbox": [ + 106, + 306, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 505, + 319 + ], + "score": 1.0, + "content": "In addition, we have explored two variations of this design. The first, self-attention, allows consid-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 343, + 330 + ], + "score": 1.0, + "content": "ering any previous memories rather than just the last one", + "type": "text" + }, + { + "bbox": [ + 343, + 319, + 366, + 329 + ], + "score": 0.9, + "content": "m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 317, + 505, + 330 + ], + "score": 1.0, + "content": ", thus providing the network with", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 329, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 505, + 341 + ], + "score": 1.0, + "content": "the capacity to model non-sequential reasoning processes. The second variation is adding gating", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 340, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 351 + ], + "score": 1.0, + "content": "mechanisms to the writing unit. These may allow the model to dynamically adjust the practical", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 349, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 363 + ], + "score": 1.0, + "content": "length of the computation to the question complexity and stabilize the memory content throughout", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 362, + 333, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 333, + 373 + ], + "score": 1.0, + "content": "the sequential network (similarly to GRUs and LSTMs).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 306, + 505, + 373 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 384, + 505, + 462 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "Self-Attention. The current architecture that we have presented allows the model to perform rea-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 396, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 407 + ], + "score": 1.0, + "content": "soning steps in a sequence, passing control and memory states from one cell to the following. How-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "ever, we would like to grant the system with more flexibility. Particularly, we would like to allow it", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "to capture more complicated reasoning processes such as trees and graphs - Directed Acyclic Graph", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 506, + 441 + ], + "score": 1.0, + "content": "(DAG) in particular, where several branches of reasoning sub-processes are merged together in later", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 438, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 104, + 438, + 506, + 453 + ], + "score": 1.0, + "content": "stages. Indeed, the CLEVR dataset includes cases where the questions embody tree-like reasoning", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 450, + 471, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 471, + 462 + ], + "score": 1.0, + "content": "process, rather than just sequences, which we would like to address correctly in our model.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22, + "bbox_fs": [ + 104, + 384, + 506, + 462 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 467, + 505, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 466, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 478 + ], + "score": 1.0, + "content": "We achieve that by adding self-attention connections between each MAC cell and all the prior cells.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 478, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 505, + 490 + ], + "score": 1.0, + "content": "Since each cell can look on all the prior reasoning steps and their corresponding memories retrieved", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "from the Knowledge Base, it can virtually capture any directed acyclic graph, while still having", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 500, + 223, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 223, + 513 + ], + "score": 1.0, + "content": "physically sequential layout.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 466, + 505, + 513 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 515, + 505, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 514, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 283, + 531 + ], + "score": 1.0, + "content": "More formally, the current MAC cell, of the", + "type": "text" + }, + { + "bbox": [ + 283, + 516, + 296, + 527 + ], + "score": 0.89, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 514, + 434, + 531 + ], + "score": 1.0, + "content": "iteration, is granted with access to", + "type": "text" + }, + { + "bbox": [ + 434, + 519, + 479, + 528 + ], + "score": 0.9, + "content": "c _ { 1 } , . . . , c _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 514, + 506, + 531 + ], + "score": 1.0, + "content": "along", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 527, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 200, + 541 + ], + "score": 1.0, + "content": "with the corresponding", + "type": "text" + }, + { + "bbox": [ + 200, + 529, + 254, + 540 + ], + "score": 0.89, + "content": "m _ { 1 } , . . . , m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 527, + 505, + 541 + ], + "score": 1.0, + "content": ", that have been computed by the prior MAC cells. It begins by", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 538, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 244, + 551 + ], + "score": 1.0, + "content": "computing the similarity between", + "type": "text" + }, + { + "bbox": [ + 244, + 541, + 253, + 550 + ], + "score": 0.84, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 538, + 271, + 551 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 272, + 540, + 316, + 550 + ], + "score": 0.89, + "content": "c _ { 1 } , . . . , c _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 538, + 505, + 551 + ], + "score": 1.0, + "content": ", and use it to derive an attention map over the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 546, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 104, + 546, + 174, + 565 + ], + "score": 1.0, + "content": "prior MAC cells", + "type": "text" + }, + { + "bbox": [ + 175, + 551, + 195, + 562 + ], + "score": 0.88, + "content": "s a _ { i , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 546, + 212, + 565 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 212, + 550, + 275, + 561 + ], + "score": 0.91, + "content": "j = 1 , . . . , i - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 546, + 425, + 565 + ], + "score": 1.0, + "content": ". This represents the relevance of the", + "type": "text" + }, + { + "bbox": [ + 426, + 549, + 440, + 561 + ], + "score": 0.9, + "content": "j ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 546, + 506, + 565 + ], + "score": 1.0, + "content": "prior reasoning", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 560, + 261, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 197, + 573 + ], + "score": 1.0, + "content": "step to the current one", + "type": "text" + }, + { + "bbox": [ + 198, + 562, + 202, + 570 + ], + "score": 0.73, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 560, + 261, + 573 + ], + "score": 1.0, + "content": "(equation 9a).", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32, + "bbox_fs": [ + 104, + 514, + 506, + 573 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 577, + 503, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 437, + 590 + ], + "score": 1.0, + "content": "Then, we average the previous memories according to this resulted attention map", + "type": "text" + }, + { + "bbox": [ + 438, + 579, + 456, + 590 + ], + "score": 0.91, + "content": "s a _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 577, + 505, + 590 + ], + "score": 1.0, + "content": ". We obtain", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 588, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 125, + 600 + ], + "score": 0.85, + "content": "m _ { s a }", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 588, + 506, + 601 + ], + "score": 1.0, + "content": ", representing the information from all the other reasoning steps that is relevant to the current", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 600, + 182, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 182, + 612 + ], + "score": 1.0, + "content": "one (equation 9b).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 577, + 506, + 612 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 505, + 661 + ], + "lines": [ + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "This resembles the approach of Key-Value networks (Miller et al., 2016). The similarity between", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "control states, corresponding to the reasoning operations that are performed in each prior step, allows", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "the model to select which memories should be taken into account, when creating the new memory –", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 649, + 464, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 464, + 662 + ], + "score": 1.0, + "content": "namely, which branches of the reasoning process should be merged together at this point.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 616, + 506, + 662 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 678, + 381, + 731 + ], + "lines": [ + { + "bbox": [ + 230, + 678, + 381, + 731 + ], + "spans": [ + { + "bbox": [ + 230, + 678, + 381, + 731 + ], + "score": 0.88, + "content": "\\begin{array} { c } { { s a _ { i j } = \\displaystyle \\mathrm { s o f t m a x } \\left( W ^ { d , 1 } ( c _ { i } \\circ c _ { j } ) + b ^ { 1 } \\right) } } \\\\ { { ( m _ { s a } ) _ { i } = \\displaystyle \\sum _ { j = 1 } ^ { i - 1 } s a _ { i j } \\cdot m _ { j } } } \\end{array}", + "type": "interline_equation", + "image_path": "b2917add3d0ec9deb18d84cf0919d7c999da2714964aba59d6c8b9129667c984.jpg" + } + ] + } + ], + "index": 42.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 678, + 381, + 704.5 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 230, + 704.5, + 381, + 731.0 + ], + "spans": [], + "index": 43 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 81, + 481, + 95 + ], + "lines": [ + { + "bbox": [ + 104, + 81, + 483, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 81, + 169, + 96 + ], + "score": 1.0, + "content": "Finally, we use", + "type": "text" + }, + { + "bbox": [ + 169, + 84, + 187, + 93 + ], + "score": 0.87, + "content": "m _ { s a }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 81, + 234, + 96 + ], + "score": 1.0, + "content": "along with", + "type": "text" + }, + { + "bbox": [ + 234, + 83, + 247, + 95 + ], + "score": 0.9, + "content": "m _ { i } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 81, + 295, + 96 + ], + "score": 1.0, + "content": "to compute", + "type": "text" + }, + { + "bbox": [ + 295, + 83, + 310, + 95 + ], + "score": 0.9, + "content": "m _ { i } ^ { \\prime \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 81, + 483, + 96 + ], + "score": 1.0, + "content": ", the new memory content in this variation.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 244, + 112, + 367, + 127 + ], + "lines": [ + { + "bbox": [ + 244, + 112, + 367, + 127 + ], + "spans": [ + { + "bbox": [ + 244, + 112, + 367, + 127 + ], + "score": 0.92, + "content": "m _ { i } ^ { \\prime \\prime } = W ^ { 2 d , d } [ m _ { n e w } , m _ { i } ^ { \\prime } ] + b ^ { d }", + "type": "interline_equation", + "image_path": "5e36c1389b375bd38c94672bc787841067c61a64ebb4c06b118a35c4f749cf9f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 244, + 112, + 367, + 127 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 140, + 505, + 185 + ], + "lines": [ + { + "bbox": [ + 106, + 141, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 431, + 153 + ], + "score": 1.0, + "content": "Memory Gate. The currently presented MAC network has some fixed number", + "type": "text" + }, + { + "bbox": [ + 432, + 143, + 438, + 152 + ], + "score": 0.82, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 141, + 505, + 153 + ], + "score": 1.0, + "content": "of concatenated", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 151, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 505, + 164 + ], + "score": 1.0, + "content": "MAC cells, representing the length of the overall reasoning process we perform. However, not all", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 163, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 505, + 176 + ], + "score": 1.0, + "content": "questions require reasoning sequence of the same length. Some questions are simpler while others", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 174, + 169, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 169, + 186 + ], + "score": 1.0, + "content": "more complex.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 190, + 504, + 235 + ], + "lines": [ + { + "bbox": [ + 105, + 190, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 204 + ], + "score": 1.0, + "content": "Motivated by this observation, we add a gate over the new memory computed at each step, that may", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 301, + 214 + ], + "score": 1.0, + "content": "selectively keep content of the previous memory", + "type": "text" + }, + { + "bbox": [ + 302, + 203, + 325, + 213 + ], + "score": 0.89, + "content": "m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 201, + 505, + 214 + ], + "score": 1.0, + "content": "unchanged. Practically, the gate functions in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 213, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 505, + 225 + ], + "score": 1.0, + "content": "a similar way to a highway network (Srivastava et al., 2015), where the gate value is conditioned on", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 223, + 249, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 235, + 237 + ], + "score": 1.0, + "content": "the current reasoning operation,", + "type": "text" + }, + { + "bbox": [ + 236, + 226, + 244, + 235 + ], + "score": 0.84, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 223, + 249, + 237 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "interline_equation", + "bbox": [ + 194, + 257, + 417, + 290 + ], + "lines": [ + { + "bbox": [ + 194, + 257, + 417, + 290 + ], + "spans": [ + { + "bbox": [ + 194, + 257, + 417, + 290 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { { c _ { i } } ^ { \\prime } = W ^ { d , d } \\cdot c _ { i } + b ^ { d } } } \\\\ { { m _ { i } = \\mathrm { s i g m o i d } \\left( c _ { i } ^ { \\prime } \\right) \\cdot m _ { i - 1 } + ( 1 - \\mathrm { s i g m o i d } \\left( c _ { i } ^ { \\prime } \\right) ) \\cdot m _ { i } ^ { \\prime \\prime } } } \\end{array}", + "type": "interline_equation", + "image_path": "06b7c9f2a66fd66b565bd6bf64cfa368a00407f9ea2545d0e729a32d2b2c1b1d.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 194, + 257, + 417, + 273.5 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 194, + 273.5, + 417, + 290.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 504, + 325 + ], + "lines": [ + { + "bbox": [ + 105, + 302, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 286, + 316 + ], + "score": 1.0, + "content": "The write unit returns the new memory state", + "type": "text" + }, + { + "bbox": [ + 286, + 305, + 299, + 314 + ], + "score": 0.87, + "content": "m _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 302, + 425, + 316 + ], + "score": 1.0, + "content": ", that will be passed along with", + "type": "text" + }, + { + "bbox": [ + 425, + 305, + 434, + 314 + ], + "score": 0.85, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 302, + 506, + 316 + ], + "score": 1.0, + "content": "to the next MAC", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 312, + 127, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 127, + 326 + ], + "score": 1.0, + "content": "cell.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 107, + 340, + 194, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 195, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 195, + 353 + ], + "score": 1.0, + "content": "3.2.4 DISCUSSION", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 108, + 360, + 504, + 394 + ], + "lines": [ + { + "bbox": [ + 106, + 360, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 373 + ], + "score": 1.0, + "content": "Overall, when designing the MAC cell, we have attempted to formulate the inner workings of an", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 372, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 505, + 384 + ], + "score": 1.0, + "content": "elementary, yet generic reasoning skills: the model decomposes the problem into steps, focusing on", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 383, + 323, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 323, + 394 + ], + "score": 1.0, + "content": "one at a time. At each such step, it takes into account:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 132, + 404, + 505, + 495 + ], + "lines": [ + { + "bbox": [ + 131, + 404, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 131, + 404, + 195, + 417 + ], + "score": 1.0, + "content": "• The control", + "type": "text" + }, + { + "bbox": [ + 195, + 406, + 204, + 416 + ], + "score": 0.82, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 404, + 505, + 417 + ], + "score": 1.0, + "content": ": Some aspect of the task - pointing to the future work that has left to be", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 415, + 168, + 428 + ], + "spans": [ + { + "bbox": [ + 141, + 415, + 168, + 428 + ], + "score": 1.0, + "content": "done.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 433, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 433, + 505, + 445 + ], + "score": 1.0, + "content": "• The previous memory or memories: The partial solution or evidence the cell has acquired", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 444, + 405, + 456 + ], + "spans": [ + { + "bbox": [ + 141, + 444, + 405, + 456 + ], + "score": 1.0, + "content": "so far – pointing to the past work that has already been achieved.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 461, + 504, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 284, + 474 + ], + "score": 1.0, + "content": "• The newly retrieved information", + "type": "text" + }, + { + "bbox": [ + 285, + 463, + 309, + 473 + ], + "score": 0.89, + "content": "m _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 461, + 485, + 474 + ], + "score": 1.0, + "content": ": that is retrieved from the knowledge base", + "type": "text" + }, + { + "bbox": [ + 485, + 461, + 504, + 472 + ], + "score": 0.62, + "content": "K B", + "type": "inline_equation" + } + ], + "index": 22 + }, + { + "bbox": [ + 142, + 473, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 142, + 473, + 505, + 485 + ], + "score": 1.0, + "content": "and may or may not be transitively related to that partial solution or evidence - the present,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 485, + 209, + 495 + ], + "spans": [ + { + "bbox": [ + 141, + 485, + 209, + 495 + ], + "score": 1.0, + "content": "or current work.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 505, + 504, + 528 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "Considering these three sources of information together, the cell finally adds the new information", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 517, + 447, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 222, + 530 + ], + "score": 1.0, + "content": "up into its working memory,", + "type": "text" + }, + { + "bbox": [ + 222, + 519, + 235, + 528 + ], + "score": 0.84, + "content": "m _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 517, + 447, + 530 + ], + "score": 1.0, + "content": ", progressing one more step towards the final answer.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "title", + "bbox": [ + 108, + 544, + 213, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 215, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 215, + 557 + ], + "score": 1.0, + "content": "3.3 THE OUTPUT UNIT", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 565, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 321, + 579 + ], + "score": 1.0, + "content": "The output unit receives the question representation", + "type": "text" + }, + { + "bbox": [ + 322, + 568, + 327, + 577 + ], + "score": 0.78, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 566, + 505, + 579 + ], + "score": 1.0, + "content": ", along with the memory state passed from", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 185, + 590 + ], + "score": 1.0, + "content": "the last MAC cell", + "type": "text" + }, + { + "bbox": [ + 185, + 578, + 199, + 589 + ], + "score": 0.86, + "content": "m _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 577, + 232, + 590 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 233, + 579, + 240, + 588 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "is the number of MAC cells in the network – representing the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "score": 1.0, + "content": "number of reasoning steps in the whole process. It inspects both and predicts an answer based on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "their concatenation. Intuitively, we would like our model to consider both the question as well as", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 608, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 104, + 608, + 505, + 624 + ], + "score": 1.0, + "content": "the relevant information that has been progressively retrieved from the KB, deemed the necessary", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 621, + 208, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 208, + 633 + ], + "score": 1.0, + "content": "information to answer it.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 218, + 650 + ], + "score": 1.0, + "content": "Note that considering both", + "type": "text" + }, + { + "bbox": [ + 218, + 640, + 225, + 649 + ], + "score": 0.78, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 638, + 244, + 650 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 244, + 640, + 259, + 650 + ], + "score": 0.87, + "content": "m _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 638, + 429, + 650 + ], + "score": 1.0, + "content": "is critical to answer the question. While", + "type": "text" + }, + { + "bbox": [ + 430, + 640, + 444, + 650 + ], + "score": 0.86, + "content": "m _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "represents the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "information collected from KB, we still need to recall what has been asked about it to be able to", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "answer accordingly. This is especially true in our case, when all other interactions between the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 684 + ], + "score": 1.0, + "content": "question and the KB are mediated through attention distributions, rather than being transformed into", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 682, + 242, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 242, + 694 + ], + "score": 1.0, + "content": "a shared continuous vector space.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "The prediction is built out of a standard 2-layers fully-connected softmax-based classifier with hid-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 168, + 722 + ], + "score": 1.0, + "content": "den dimension", + "type": "text" + }, + { + "bbox": [ + 168, + 710, + 174, + 720 + ], + "score": 0.79, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "and output dimension that matches the number of possible answers in the dataset.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 719, + 482, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 197, + 734 + ], + "score": 1.0, + "content": "The classifier receives", + "type": "text" + }, + { + "bbox": [ + 198, + 721, + 226, + 733 + ], + "score": 0.93, + "content": "[ m _ { p } , q ]", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 719, + 482, + 734 + ], + "score": 1.0, + "content": "as input and returns a probability distribution over the answers.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 81, + 481, + 95 + ], + "lines": [ + { + "bbox": [ + 104, + 81, + 483, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 81, + 169, + 96 + ], + "score": 1.0, + "content": "Finally, we use", + "type": "text" + }, + { + "bbox": [ + 169, + 84, + 187, + 93 + ], + "score": 0.87, + "content": "m _ { s a }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 81, + 234, + 96 + ], + "score": 1.0, + "content": "along with", + "type": "text" + }, + { + "bbox": [ + 234, + 83, + 247, + 95 + ], + "score": 0.9, + "content": "m _ { i } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 81, + 295, + 96 + ], + "score": 1.0, + "content": "to compute", + "type": "text" + }, + { + "bbox": [ + 295, + 83, + 310, + 95 + ], + "score": 0.9, + "content": "m _ { i } ^ { \\prime \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 81, + 483, + 96 + ], + "score": 1.0, + "content": ", the new memory content in this variation.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 104, + 81, + 483, + 96 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 244, + 112, + 367, + 127 + ], + "lines": [ + { + "bbox": [ + 244, + 112, + 367, + 127 + ], + "spans": [ + { + "bbox": [ + 244, + 112, + 367, + 127 + ], + "score": 0.92, + "content": "m _ { i } ^ { \\prime \\prime } = W ^ { 2 d , d } [ m _ { n e w } , m _ { i } ^ { \\prime } ] + b ^ { d }", + "type": "interline_equation", + "image_path": "5e36c1389b375bd38c94672bc787841067c61a64ebb4c06b118a35c4f749cf9f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 244, + 112, + 367, + 127 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 140, + 505, + 185 + ], + "lines": [ + { + "bbox": [ + 106, + 141, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 431, + 153 + ], + "score": 1.0, + "content": "Memory Gate. The currently presented MAC network has some fixed number", + "type": "text" + }, + { + "bbox": [ + 432, + 143, + 438, + 152 + ], + "score": 0.82, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 141, + 505, + 153 + ], + "score": 1.0, + "content": "of concatenated", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 151, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 505, + 164 + ], + "score": 1.0, + "content": "MAC cells, representing the length of the overall reasoning process we perform. However, not all", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 163, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 505, + 176 + ], + "score": 1.0, + "content": "questions require reasoning sequence of the same length. Some questions are simpler while others", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 174, + 169, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 169, + 186 + ], + "score": 1.0, + "content": "more complex.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 141, + 505, + 186 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 190, + 504, + 235 + ], + "lines": [ + { + "bbox": [ + 105, + 190, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 204 + ], + "score": 1.0, + "content": "Motivated by this observation, we add a gate over the new memory computed at each step, that may", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 301, + 214 + ], + "score": 1.0, + "content": "selectively keep content of the previous memory", + "type": "text" + }, + { + "bbox": [ + 302, + 203, + 325, + 213 + ], + "score": 0.89, + "content": "m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 201, + 505, + 214 + ], + "score": 1.0, + "content": "unchanged. Practically, the gate functions in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 213, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 505, + 225 + ], + "score": 1.0, + "content": "a similar way to a highway network (Srivastava et al., 2015), where the gate value is conditioned on", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 223, + 249, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 235, + 237 + ], + "score": 1.0, + "content": "the current reasoning operation,", + "type": "text" + }, + { + "bbox": [ + 236, + 226, + 244, + 235 + ], + "score": 0.84, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 223, + 249, + 237 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 190, + 505, + 237 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 194, + 257, + 417, + 290 + ], + "lines": [ + { + "bbox": [ + 194, + 257, + 417, + 290 + ], + "spans": [ + { + "bbox": [ + 194, + 257, + 417, + 290 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { { c _ { i } } ^ { \\prime } = W ^ { d , d } \\cdot c _ { i } + b ^ { d } } } \\\\ { { m _ { i } = \\mathrm { s i g m o i d } \\left( c _ { i } ^ { \\prime } \\right) \\cdot m _ { i - 1 } + ( 1 - \\mathrm { s i g m o i d } \\left( c _ { i } ^ { \\prime } \\right) ) \\cdot m _ { i } ^ { \\prime \\prime } } } \\end{array}", + "type": "interline_equation", + "image_path": "06b7c9f2a66fd66b565bd6bf64cfa368a00407f9ea2545d0e729a32d2b2c1b1d.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 194, + 257, + 417, + 273.5 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 194, + 273.5, + 417, + 290.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 504, + 325 + ], + "lines": [ + { + "bbox": [ + 105, + 302, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 286, + 316 + ], + "score": 1.0, + "content": "The write unit returns the new memory state", + "type": "text" + }, + { + "bbox": [ + 286, + 305, + 299, + 314 + ], + "score": 0.87, + "content": "m _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 302, + 425, + 316 + ], + "score": 1.0, + "content": ", that will be passed along with", + "type": "text" + }, + { + "bbox": [ + 425, + 305, + 434, + 314 + ], + "score": 0.85, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 302, + 506, + 316 + ], + "score": 1.0, + "content": "to the next MAC", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 312, + 127, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 127, + 326 + ], + "score": 1.0, + "content": "cell.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 302, + 506, + 326 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 340, + 194, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 195, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 195, + 353 + ], + "score": 1.0, + "content": "3.2.4 DISCUSSION", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 108, + 360, + 504, + 394 + ], + "lines": [ + { + "bbox": [ + 106, + 360, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 373 + ], + "score": 1.0, + "content": "Overall, when designing the MAC cell, we have attempted to formulate the inner workings of an", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 372, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 505, + 384 + ], + "score": 1.0, + "content": "elementary, yet generic reasoning skills: the model decomposes the problem into steps, focusing on", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 383, + 323, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 323, + 394 + ], + "score": 1.0, + "content": "one at a time. At each such step, it takes into account:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 106, + 360, + 505, + 394 + ] + }, + { + "type": "list", + "bbox": [ + 132, + 404, + 505, + 495 + ], + "lines": [ + { + "bbox": [ + 131, + 404, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 131, + 404, + 195, + 417 + ], + "score": 1.0, + "content": "• The control", + "type": "text" + }, + { + "bbox": [ + 195, + 406, + 204, + 416 + ], + "score": 0.82, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 404, + 505, + 417 + ], + "score": 1.0, + "content": ": Some aspect of the task - pointing to the future work that has left to be", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 415, + 168, + 428 + ], + "spans": [ + { + "bbox": [ + 141, + 415, + 168, + 428 + ], + "score": 1.0, + "content": "done.", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 433, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 433, + 505, + 445 + ], + "score": 1.0, + "content": "• The previous memory or memories: The partial solution or evidence the cell has acquired", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 444, + 405, + 456 + ], + "spans": [ + { + "bbox": [ + 141, + 444, + 405, + 456 + ], + "score": 1.0, + "content": "so far – pointing to the past work that has already been achieved.", + "type": "text" + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 461, + 504, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 284, + 474 + ], + "score": 1.0, + "content": "• The newly retrieved information", + "type": "text" + }, + { + "bbox": [ + 285, + 463, + 309, + 473 + ], + "score": 0.89, + "content": "m _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 461, + 485, + 474 + ], + "score": 1.0, + "content": ": that is retrieved from the knowledge base", + "type": "text" + }, + { + "bbox": [ + 485, + 461, + 504, + 472 + ], + "score": 0.62, + "content": "K B", + "type": "inline_equation" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 473, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 142, + 473, + 505, + 485 + ], + "score": 1.0, + "content": "and may or may not be transitively related to that partial solution or evidence - the present,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 485, + 209, + 495 + ], + "spans": [ + { + "bbox": [ + 141, + 485, + 209, + 495 + ], + "score": 1.0, + "content": "or current work.", + "type": "text" + } + ], + "index": 24, + "is_list_end_line": true + } + ], + "index": 21, + "bbox_fs": [ + 131, + 404, + 505, + 495 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 505, + 504, + 528 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "Considering these three sources of information together, the cell finally adds the new information", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 517, + 447, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 222, + 530 + ], + "score": 1.0, + "content": "up into its working memory,", + "type": "text" + }, + { + "bbox": [ + 222, + 519, + 235, + 528 + ], + "score": 0.84, + "content": "m _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 517, + 447, + 530 + ], + "score": 1.0, + "content": ", progressing one more step towards the final answer.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 505, + 505, + 530 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 544, + 213, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 215, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 215, + 557 + ], + "score": 1.0, + "content": "3.3 THE OUTPUT UNIT", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 565, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 321, + 579 + ], + "score": 1.0, + "content": "The output unit receives the question representation", + "type": "text" + }, + { + "bbox": [ + 322, + 568, + 327, + 577 + ], + "score": 0.78, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 566, + 505, + 579 + ], + "score": 1.0, + "content": ", along with the memory state passed from", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 185, + 590 + ], + "score": 1.0, + "content": "the last MAC cell", + "type": "text" + }, + { + "bbox": [ + 185, + 578, + 199, + 589 + ], + "score": 0.86, + "content": "m _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 577, + 232, + 590 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 233, + 579, + 240, + 588 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "is the number of MAC cells in the network – representing the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "score": 1.0, + "content": "number of reasoning steps in the whole process. It inspects both and predicts an answer based on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "their concatenation. Intuitively, we would like our model to consider both the question as well as", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 608, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 104, + 608, + 505, + 624 + ], + "score": 1.0, + "content": "the relevant information that has been progressively retrieved from the KB, deemed the necessary", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 621, + 208, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 208, + 633 + ], + "score": 1.0, + "content": "information to answer it.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5, + "bbox_fs": [ + 104, + 566, + 506, + 633 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 218, + 650 + ], + "score": 1.0, + "content": "Note that considering both", + "type": "text" + }, + { + "bbox": [ + 218, + 640, + 225, + 649 + ], + "score": 0.78, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 638, + 244, + 650 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 244, + 640, + 259, + 650 + ], + "score": 0.87, + "content": "m _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 638, + 429, + 650 + ], + "score": 1.0, + "content": "is critical to answer the question. While", + "type": "text" + }, + { + "bbox": [ + 430, + 640, + 444, + 650 + ], + "score": 0.86, + "content": "m _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "represents the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "information collected from KB, we still need to recall what has been asked about it to be able to", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "answer accordingly. This is especially true in our case, when all other interactions between the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 684 + ], + "score": 1.0, + "content": "question and the KB are mediated through attention distributions, rather than being transformed into", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 682, + 242, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 242, + 694 + ], + "score": 1.0, + "content": "a shared continuous vector space.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 638, + 506, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "The prediction is built out of a standard 2-layers fully-connected softmax-based classifier with hid-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 168, + 722 + ], + "score": 1.0, + "content": "den dimension", + "type": "text" + }, + { + "bbox": [ + 168, + 710, + 174, + 720 + ], + "score": 0.79, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "and output dimension that matches the number of possible answers in the dataset.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 719, + 482, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 197, + 734 + ], + "score": 1.0, + "content": "The classifier receives", + "type": "text" + }, + { + "bbox": [ + 198, + 721, + 226, + 733 + ], + "score": 0.93, + "content": "[ m _ { p } , q ]", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 719, + 482, + 734 + ], + "score": 1.0, + "content": "as input and returns a probability distribution over the answers.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 698, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 144, + 476, + 291 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 80, + 505, + 135 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 80, + 505, + 92 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 505, + 92 + ], + "score": 1.0, + "content": "Table 1: CLEVR accuracy by baseline methods, competing methods, and our method (MAC) - for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 91, + 504, + 103 + ], + "spans": [ + { + "bbox": [ + 106, + 91, + 504, + 103 + ], + "score": 1.0, + "content": "our method we show test results with GLoVE initialization and validation results with and with-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 101, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 458, + 116 + ], + "score": 1.0, + "content": "out GLoVE. (*) denotes use of extra supervisory information through program labels.", + "type": "text" + }, + { + "bbox": [ + 459, + 101, + 469, + 114 + ], + "score": 0.28, + "content": "( ^ { \\dagger } )", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 101, + 506, + 116 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 113, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 217, + 126 + ], + "score": 1.0, + "content": "use of data augmentation.", + "type": "text" + }, + { + "bbox": [ + 217, + 113, + 228, + 124 + ], + "score": 0.52, + "content": "( ^ { \\ddagger } )", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 113, + 371, + 126 + ], + "score": 1.0, + "content": "denotes training from raw pixels.", + "type": "text" + }, + { + "bbox": [ + 371, + 113, + 383, + 124 + ], + "score": 0.42, + "content": "( ^ { \\circ } )", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 113, + 505, + 126 + ], + "score": 1.0, + "content": "denotes GLoVE word vector", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 123, + 162, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 162, + 136 + ], + "score": 1.0, + "content": "initialization.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "table_body", + "bbox": [ + 135, + 144, + 476, + 291 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 144, + 476, + 291 + ], + "spans": [ + { + "bbox": [ + 135, + 144, + 476, + 291 + ], + "score": 0.981, + "html": "
ModelOverallCountExistCompare NumbersQuery AttributeCompare Attribute
Human (Johnson et al., 2017)92.686.796.686.595.096.0
Q-type baseline (Johnson et al., 2017)41.834.650.251.036.051.3
LSTM (Johnson et al., 2017)46.841.761.169.836.851.8
CNN+LSTM (Johnson et al., 2017)52.343.765.267.149.353.0
CNN+LSTM+SA (Johnson et al., 2016)76.664.482.777.482.675.4
N2NMN* (Hu et al. 2017)83.768.585.784.990.088.7
PG+EE (9K prog.)* (Johnson et al., 2017)88.679.789.779.192.696.0
PG+EE (70oK prog.)* (Johnson et al., 2017)96.992.797.198.798.198.9
CNN+LSTM+RN†‡ (Santoro et al.,2017)95.590.197.893.697.997.1
CNN+GRU+FiLM (Perez et al.,2017)97.794.399.196.899.199.1
CNN+GRU+FiLM* (Perez et al., 2017)97.694.399.393.499.399.3
CNN+GRU+FiLM (Perez et al., 2017)97.794.399.196.899.199.1
MAC (this paper)-val98.997.199.599.399.299.2
MAC-val99.097.299.599.599.599.4
MAC-test98.997.299.599.499.399.5
", + "type": "table", + "image_path": "a8f3f0904dbfbf06b3ae89f9cd0dcd15557eabccb6cf593d1481a4565628f79b.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 135, + 144, + 476, + 193.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 135, + 193.0, + 476, + 242.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 135, + 242.0, + 476, + 291.0 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 4.0 + }, + { + "type": "title", + "bbox": [ + 107, + 329, + 200, + 342 + ], + "lines": [ + { + "bbox": [ + 105, + 329, + 201, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 201, + 344 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 367, + 505, + 445 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 381 + ], + "score": 1.0, + "content": "We evaluate our model on the recent CLEVR dataset (Johnson et al., 2016). CLEVR is a syn-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 378, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 506, + 392 + ], + "score": 1.0, + "content": "thetic dataset consisting of 700K tuples; each consists of a 3D-rendered image featuring objects of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "various shapes, colors, materials and sizes, coupled with compositional multi-step questions that", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "score": 1.0, + "content": "measure performance on an array of challenging reasoning skills such as following transitive re-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "score": 1.0, + "content": "lations, counting objects and comparing their properties. In addition, each question is associated", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 422, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 505, + 436 + ], + "score": 1.0, + "content": "with a formal program, specifying the reasoning operations that should be performed to compute", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 434, + 248, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 248, + 446 + ], + "score": 1.0, + "content": "the answer, among 28 possibilities.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 505, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 464 + ], + "score": 1.0, + "content": "We first perform experiments on the original 700k CLEVR dataset (Johnson et al., 2016), comparing", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "to prior work. As shown in table 1, our model matches or outperforms all existing models both in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 472, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 485 + ], + "score": 1.0, + "content": "overall accuracy, as well as in each category, testing different reasoning skills. In particular, for the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 237, + 497 + ], + "score": 1.0, + "content": "overall performance, we achieve", + "type": "text" + }, + { + "bbox": [ + 238, + 484, + 270, + 494 + ], + "score": 0.86, + "content": "9 8 . 9 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 483, + 505, + 497 + ], + "score": 1.0, + "content": "accuracy, more than halving the error rate of the prior best", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 494, + 242, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 242, + 507 + ], + "score": 1.0, + "content": "model, FiLM (Perez et al., 2017).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 511, + 504, + 567 + ], + "lines": [ + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "Counting and Numerical Comparison. Remarkably, our performance on questions testing count-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 522, + 504, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 504, + 536 + ], + "score": 1.0, + "content": "ing and numerical comparisons is significantly higher than the competing models, which consis-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "tently struggle on this question type. Again, we nearly halve the corresponding error rate. These", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 543, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 559 + ], + "score": 1.0, + "content": "results demonstrate the aptitude of attention mechanisms to perform counting, reduction and aggre-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 555, + 336, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 336, + 567 + ], + "score": 1.0, + "content": "gation, in contrast to alternative, CNN-based approaches.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 504, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 504, + 584 + ], + "score": 1.0, + "content": "Training Length and Computational-Efficiency. We examine the learning curves of our and com-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "score": 1.0, + "content": "peting models. We have trained all models on the same architecture and used the author code for", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "score": 1.0, + "content": "the other models. Aiming at having equal settings for comparison, we ran all models including ours", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "with learned random words vectors. In order to make sure the results are statistically significant we", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "score": 1.0, + "content": "ran each model multiple (10) times, and plotted the averages and confidence intervals (figure 4). The", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "results show that our model learns significantly faster than the other leading methods, FiLM (Perez", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 174, + 650 + ], + "score": 1.0, + "content": "et al., 2017) and", + "type": "text" + }, + { + "bbox": [ + 175, + 638, + 208, + 649 + ], + "score": 0.75, + "content": "\\mathrm { P G + E E }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "(Johnson et al., 2017). While we do not have learning curves for the Re-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "score": 1.0, + "content": "lational Network model, Santoro et al. (2017) report approximately 1.4 million iterations to achieve", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 134, + 671 + ], + "score": 0.86, + "content": "9 5 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "accuracy, which are equivalent to 125 epochs approximately, whereas our model achieves a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "score": 1.0, + "content": "comparable accuracy after 3 epochs only, yielding 40x reduction in the length of the training process.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "Naturally, the smaller number of required training steps also translates to comparably shorter training", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "time. Perez et al. (2017) report training time of 4 days, equivalent to 80 epochs, to reach accuracy", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 117, + 722 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 710, + 144, + 720 + ], + "score": 0.87, + "content": "9 7 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 709, + 505, + 722 + ], + "score": 1.0, + "content": ". In contrast, we achieve higher accuracy in 6 epochs, taking 9.5 hours overall, leading to", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 721, + 230, + 732 + ], + "spans": [ + { + "bbox": [ + 107, + 721, + 123, + 731 + ], + "score": 0.49, + "content": "1 0 \\mathrm { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 123, + 721, + 230, + 732 + ], + "score": 1.0, + "content": "reduction in training time.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 144, + 476, + 291 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 80, + 505, + 135 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 80, + 505, + 92 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 505, + 92 + ], + "score": 1.0, + "content": "Table 1: CLEVR accuracy by baseline methods, competing methods, and our method (MAC) - for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 91, + 504, + 103 + ], + "spans": [ + { + "bbox": [ + 106, + 91, + 504, + 103 + ], + "score": 1.0, + "content": "our method we show test results with GLoVE initialization and validation results with and with-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 101, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 458, + 116 + ], + "score": 1.0, + "content": "out GLoVE. (*) denotes use of extra supervisory information through program labels.", + "type": "text" + }, + { + "bbox": [ + 459, + 101, + 469, + 114 + ], + "score": 0.28, + "content": "( ^ { \\dagger } )", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 101, + 506, + 116 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 113, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 217, + 126 + ], + "score": 1.0, + "content": "use of data augmentation.", + "type": "text" + }, + { + "bbox": [ + 217, + 113, + 228, + 124 + ], + "score": 0.52, + "content": "( ^ { \\ddagger } )", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 113, + 371, + 126 + ], + "score": 1.0, + "content": "denotes training from raw pixels.", + "type": "text" + }, + { + "bbox": [ + 371, + 113, + 383, + 124 + ], + "score": 0.42, + "content": "( ^ { \\circ } )", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 113, + 505, + 126 + ], + "score": 1.0, + "content": "denotes GLoVE word vector", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 123, + 162, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 162, + 136 + ], + "score": 1.0, + "content": "initialization.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "table_body", + "bbox": [ + 135, + 144, + 476, + 291 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 144, + 476, + 291 + ], + "spans": [ + { + "bbox": [ + 135, + 144, + 476, + 291 + ], + "score": 0.981, + "html": "
ModelOverallCountExistCompare NumbersQuery AttributeCompare Attribute
Human (Johnson et al., 2017)92.686.796.686.595.096.0
Q-type baseline (Johnson et al., 2017)41.834.650.251.036.051.3
LSTM (Johnson et al., 2017)46.841.761.169.836.851.8
CNN+LSTM (Johnson et al., 2017)52.343.765.267.149.353.0
CNN+LSTM+SA (Johnson et al., 2016)76.664.482.777.482.675.4
N2NMN* (Hu et al. 2017)83.768.585.784.990.088.7
PG+EE (9K prog.)* (Johnson et al., 2017)88.679.789.779.192.696.0
PG+EE (70oK prog.)* (Johnson et al., 2017)96.992.797.198.798.198.9
CNN+LSTM+RN†‡ (Santoro et al.,2017)95.590.197.893.697.997.1
CNN+GRU+FiLM (Perez et al.,2017)97.794.399.196.899.199.1
CNN+GRU+FiLM* (Perez et al., 2017)97.694.399.393.499.399.3
CNN+GRU+FiLM (Perez et al., 2017)97.794.399.196.899.199.1
MAC (this paper)-val98.997.199.599.399.299.2
MAC-val99.097.299.599.599.599.4
MAC-test98.997.299.599.499.399.5
", + "type": "table", + "image_path": "a8f3f0904dbfbf06b3ae89f9cd0dcd15557eabccb6cf593d1481a4565628f79b.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 135, + 144, + 476, + 193.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 135, + 193.0, + 476, + 242.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 135, + 242.0, + 476, + 291.0 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 4.0 + }, + { + "type": "title", + "bbox": [ + 107, + 329, + 200, + 342 + ], + "lines": [ + { + "bbox": [ + 105, + 329, + 201, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 201, + 344 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 367, + 505, + 445 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 381 + ], + "score": 1.0, + "content": "We evaluate our model on the recent CLEVR dataset (Johnson et al., 2016). CLEVR is a syn-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 378, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 506, + 392 + ], + "score": 1.0, + "content": "thetic dataset consisting of 700K tuples; each consists of a 3D-rendered image featuring objects of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "various shapes, colors, materials and sizes, coupled with compositional multi-step questions that", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "score": 1.0, + "content": "measure performance on an array of challenging reasoning skills such as following transitive re-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "score": 1.0, + "content": "lations, counting objects and comparing their properties. In addition, each question is associated", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 422, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 505, + 436 + ], + "score": 1.0, + "content": "with a formal program, specifying the reasoning operations that should be performed to compute", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 434, + 248, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 248, + 446 + ], + "score": 1.0, + "content": "the answer, among 28 possibilities.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 367, + 506, + 446 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 505, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 464 + ], + "score": 1.0, + "content": "We first perform experiments on the original 700k CLEVR dataset (Johnson et al., 2016), comparing", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "to prior work. As shown in table 1, our model matches or outperforms all existing models both in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 472, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 485 + ], + "score": 1.0, + "content": "overall accuracy, as well as in each category, testing different reasoning skills. In particular, for the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 237, + 497 + ], + "score": 1.0, + "content": "overall performance, we achieve", + "type": "text" + }, + { + "bbox": [ + 238, + 484, + 270, + 494 + ], + "score": 0.86, + "content": "9 8 . 9 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 483, + 505, + 497 + ], + "score": 1.0, + "content": "accuracy, more than halving the error rate of the prior best", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 494, + 242, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 242, + 507 + ], + "score": 1.0, + "content": "model, FiLM (Perez et al., 2017).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 450, + 505, + 507 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 511, + 504, + 567 + ], + "lines": [ + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "Counting and Numerical Comparison. Remarkably, our performance on questions testing count-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 522, + 504, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 504, + 536 + ], + "score": 1.0, + "content": "ing and numerical comparisons is significantly higher than the competing models, which consis-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "tently struggle on this question type. Again, we nearly halve the corresponding error rate. These", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 543, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 559 + ], + "score": 1.0, + "content": "results demonstrate the aptitude of attention mechanisms to perform counting, reduction and aggre-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 555, + 336, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 336, + 567 + ], + "score": 1.0, + "content": "gation, in contrast to alternative, CNN-based approaches.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 511, + 505, + 567 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 504, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 504, + 584 + ], + "score": 1.0, + "content": "Training Length and Computational-Efficiency. We examine the learning curves of our and com-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "score": 1.0, + "content": "peting models. We have trained all models on the same architecture and used the author code for", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "score": 1.0, + "content": "the other models. Aiming at having equal settings for comparison, we ran all models including ours", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "with learned random words vectors. In order to make sure the results are statistically significant we", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "score": 1.0, + "content": "ran each model multiple (10) times, and plotted the averages and confidence intervals (figure 4). The", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "results show that our model learns significantly faster than the other leading methods, FiLM (Perez", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 174, + 650 + ], + "score": 1.0, + "content": "et al., 2017) and", + "type": "text" + }, + { + "bbox": [ + 175, + 638, + 208, + 649 + ], + "score": 0.75, + "content": "\\mathrm { P G + E E }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "(Johnson et al., 2017). While we do not have learning curves for the Re-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "score": 1.0, + "content": "lational Network model, Santoro et al. (2017) report approximately 1.4 million iterations to achieve", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 134, + 671 + ], + "score": 0.86, + "content": "9 5 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "accuracy, which are equivalent to 125 epochs approximately, whereas our model achieves a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "score": 1.0, + "content": "comparable accuracy after 3 epochs only, yielding 40x reduction in the length of the training process.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 572, + 506, + 684 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "Naturally, the smaller number of required training steps also translates to comparably shorter training", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "time. Perez et al. (2017) report training time of 4 days, equivalent to 80 epochs, to reach accuracy", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 117, + 722 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 710, + 144, + 720 + ], + "score": 0.87, + "content": "9 7 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 709, + 505, + 722 + ], + "score": 1.0, + "content": ". In contrast, we achieve higher accuracy in 6 epochs, taking 9.5 hours overall, leading to", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 721, + 230, + 732 + ], + "spans": [ + { + "bbox": [ + 107, + 721, + 123, + 731 + ], + "score": 0.49, + "content": "1 0 \\mathrm { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 123, + 721, + 230, + 732 + ], + "score": 1.0, + "content": "reduction in training time.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 687, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 180, + 86, + 428, + 207 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 180, + 86, + 428, + 207 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 180, + 86, + 428, + 207 + ], + "spans": [ + { + "bbox": [ + 180, + 86, + 428, + 207 + ], + "score": 0.969, + "type": "image", + "image_path": "c74201d70aae1d3587e14044064872d252c0e00ec6d3c4fd6b70377bb7088c75.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 180, + 86, + 428, + 126.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 180, + 126.33333333333334, + 428, + 166.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 180, + 166.66666666666669, + 428, + 207.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 222, + 505, + 278 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 223, + 504, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 471, + 235 + ], + "score": 1.0, + "content": "Figure 6: Training curves and accuracies for MACs (our model), FiLM (Perez et al., 2017),", + "type": "text" + }, + { + "bbox": [ + 472, + 223, + 504, + 233 + ], + "score": 0.53, + "content": "\\mathrm { P G + E E }", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 234, + 504, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 471, + 246 + ], + "score": 1.0, + "content": "(Johnson et al., 2017) and stacked-attention (Yang et al., 2016; Johnson et al., 2017). (Note:", + "type": "text" + }, + { + "bbox": [ + 472, + 234, + 504, + 244 + ], + "score": 0.54, + "content": "\\mathrm { P G + E E }", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "uses the supported CLEVR programs as strong supervision.) Left: Training curve (accuracy/epoch).", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 255, + 506, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 256, + 270 + ], + "score": 1.0, + "content": "Right: Learning curve: Accuracy for", + "type": "text" + }, + { + "bbox": [ + 256, + 256, + 271, + 266 + ], + "score": 0.82, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 255, + 273, + 270 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 274, + 256, + 289, + 266 + ], + "score": 0.79, + "content": "2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 255, + 292, + 270 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 293, + 256, + 307, + 266 + ], + "score": 0.82, + "content": "5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 255, + 311, + 270 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 312, + 256, + 331, + 266 + ], + "score": 0.84, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 255, + 334, + 270 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 334, + 256, + 354, + 266 + ], + "score": 0.84, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 255, + 357, + 270 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 358, + 256, + 378, + 266 + ], + "score": 0.85, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 255, + 396, + 270 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 397, + 256, + 420, + 266 + ], + "score": 0.87, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 255, + 447, + 270 + ], + "score": 1.0, + "content": "of the", + "type": "text" + }, + { + "bbox": [ + 447, + 256, + 469, + 266 + ], + "score": 0.68, + "content": "7 0 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 255, + 506, + 270 + ], + "score": 1.0, + "content": "CLEVR", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 266, + 144, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 144, + 280 + ], + "score": 1.0, + "content": "samples.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "table", + "bbox": [ + 181, + 366, + 429, + 448 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 312, + 505, + 356 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 312, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 506, + 325 + ], + "score": 1.0, + "content": "Table 2: Accuracy on CLEVR-Humans of previous methods and our method (MAC), before (left)", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 323, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 369, + 336 + ], + "score": 1.0, + "content": "and after (right) fine-tuning on the CLEVR-Humans training data.", + "type": "text" + }, + { + "bbox": [ + 370, + 324, + 402, + 334 + ], + "score": 0.75, + "content": "\\mathrm { P G + E E }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 323, + 506, + 336 + ], + "score": 1.0, + "content": "uses supervised data. For", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "our method (MAC) we show test results with GLoVE initialization and validation results with and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 345, + 176, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 176, + 356 + ], + "score": 1.0, + "content": "without GLoVE.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "table_body", + "bbox": [ + 181, + 366, + 429, + 448 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 181, + 366, + 429, + 448 + ], + "spans": [ + { + "bbox": [ + 181, + 366, + 429, + 448 + ], + "score": 0.979, + "html": "
ModelTrain CLEVRTrain CLEVR + fine-tune HUMANS
LSTM (Johnson et al.,2017)27.536.5
CNN+LSTM (Johnson et al.,2017)37.743.2
CNN+LSTM+SA+MLP (Johnson et al.,2016)50.457.6
PG+EE (18K prog.)* (Johnson et al.,2017)54.066.6
CNN+GRU+FiLM (Perez et al., 2017)56.675.9
MAC-val56.981.2
MAC-val58.582.0
MAC-test58.682.5
", + "type": "table", + "image_path": "20d0a9d88276e853b2d4290d832c31f6ed7a6d4ae845cb8a1f4754688e965048.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 181, + 366, + 429, + 393.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 181, + 393.3333333333333, + 429, + 420.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 181, + 420.66666666666663, + 429, + 447.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 11.25 + }, + { + "type": "title", + "bbox": [ + 107, + 491, + 209, + 502 + ], + "lines": [ + { + "bbox": [ + 105, + 490, + 210, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 210, + 504 + ], + "score": 1.0, + "content": "4.1 DATA EFFICIENCY", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 522, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 534 + ], + "score": 1.0, + "content": "We have explored the performance of our and other leading approaches on smaller subsets of the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "CLEVR dataset, in order to study the ability of models to generalize from smaller amount of data.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 303, + 556 + ], + "score": 1.0, + "content": "We sampled at random subsets of CLEVR, with", + "type": "text" + }, + { + "bbox": [ + 304, + 545, + 323, + 555 + ], + "score": 0.84, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 545, + 326, + 556 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 327, + 545, + 347, + 555 + ], + "score": 0.84, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 545, + 365, + 556 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 366, + 545, + 385, + 555 + ], + "score": 0.87, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 545, + 444, + 556 + ], + "score": 1.0, + "content": "of its original", + "type": "text" + }, + { + "bbox": [ + 444, + 545, + 466, + 555 + ], + "score": 0.44, + "content": "7 0 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 545, + 505, + 556 + ], + "score": 1.0, + "content": "size, and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "used them to train our and other 3 proposed models for the CLEVR task: FiLM (Perez et al., 2017),", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 204, + 578 + ], + "score": 1.0, + "content": "the strongly-supervised", + "type": "text" + }, + { + "bbox": [ + 205, + 566, + 237, + 577 + ], + "score": 0.67, + "content": "\\mathrm { P G + E E }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "(Johnson et al., 2017), and stacked-attention networks (Johnson", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 577, + 231, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 231, + 590 + ], + "score": 1.0, + "content": "et al., 2017; Yang et al., 2016).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 594, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 504, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 504, + 605 + ], + "score": 1.0, + "content": "As shown in figure 4, our model outperforms the other models by a wide margin for all subsets of the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 193, + 617 + ], + "score": 1.0, + "content": "CLEVR dataset. For", + "type": "text" + }, + { + "bbox": [ + 193, + 605, + 213, + 615 + ], + "score": 0.87, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 604, + 317, + 617 + ], + "score": 1.0, + "content": "of the data, equivalent to", + "type": "text" + }, + { + "bbox": [ + 317, + 605, + 339, + 615 + ], + "score": 0.63, + "content": "3 5 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "samples, other models obtain accuracies", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 104, + 615, + 176, + 629 + ], + "score": 1.0, + "content": "ranging between", + "type": "text" + }, + { + "bbox": [ + 176, + 616, + 196, + 627 + ], + "score": 0.88, + "content": "70 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 615, + 215, + 629 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 215, + 616, + 235, + 626 + ], + "score": 0.88, + "content": "92 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 615, + 345, + 629 + ], + "score": 1.0, + "content": ", while our model achieves", + "type": "text" + }, + { + "bbox": [ + 345, + 616, + 372, + 627 + ], + "score": 0.87, + "content": "9 7 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 615, + 506, + 629 + ], + "score": 1.0, + "content": ". The gap becomes larger as the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 205, + 640 + ], + "score": 1.0, + "content": "dataset size reduces: for", + "type": "text" + }, + { + "bbox": [ + 205, + 627, + 225, + 637 + ], + "score": 0.88, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 626, + 327, + 640 + ], + "score": 1.0, + "content": "of the data, equivalent to", + "type": "text" + }, + { + "bbox": [ + 327, + 627, + 349, + 637 + ], + "score": 0.5, + "content": "1 7 5 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "samples, performance of other models", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 636, + 416, + 651 + ], + "spans": [ + { + "bbox": [ + 104, + 636, + 151, + 651 + ], + "score": 1.0, + "content": "is between", + "type": "text" + }, + { + "bbox": [ + 151, + 638, + 171, + 649 + ], + "score": 0.86, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 636, + 189, + 651 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 189, + 638, + 208, + 649 + ], + "score": 0.87, + "content": "7 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 636, + 348, + 651 + ], + "score": 1.0, + "content": ", while our model maintains a high", + "type": "text" + }, + { + "bbox": [ + 348, + 638, + 375, + 649 + ], + "score": 0.83, + "content": "9 5 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 636, + 416, + 651 + ], + "score": 1.0, + "content": "accuracy.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 154, + 667 + ], + "score": 1.0, + "content": "Finally, for", + "type": "text" + }, + { + "bbox": [ + 154, + 655, + 173, + 666 + ], + "score": 0.87, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 655, + 219, + 667 + ], + "score": 1.0, + "content": "of the data", + "type": "text" + }, + { + "bbox": [ + 219, + 655, + 244, + 666 + ], + "score": 0.29, + "content": "- 7 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "samples, still a sizeable amount – our model is the only one that", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 284, + 678 + ], + "score": 1.0, + "content": "manages to generalize, with performance of", + "type": "text" + }, + { + "bbox": [ + 284, + 666, + 311, + 676 + ], + "score": 0.87, + "content": "8 4 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "on average, whereas the other three models fail,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 147, + 689 + ], + "score": 1.0, + "content": "achieving", + "type": "text" + }, + { + "bbox": [ + 147, + 677, + 204, + 687 + ], + "score": 0.89, + "content": "4 7 . 6 \\% - 5 7 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 677, + 505, + 689 + ], + "score": 1.0, + "content": ". Note that as pointed out by (Johnson et al., 2016) a simple baseline that", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 686, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 104, + 686, + 429, + 702 + ], + "score": 1.0, + "content": "predicts the most frequent answer for each of the question types achieves already", + "type": "text" + }, + { + "bbox": [ + 429, + 688, + 456, + 699 + ], + "score": 0.87, + "content": "4 2 . 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 686, + 506, + 702 + ], + "score": 1.0, + "content": ", suggesting", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "that answering half of the questions correctly means that the competing models barely learn to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "generalize from the smaller dataset. These results demonstrate the robustness of our architecture", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 720, + 484, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 484, + 732 + ], + "score": 1.0, + "content": "and its key role as a structural prior guiding our network to learn the intended reasoning skills.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 180, + 86, + 428, + 207 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 180, + 86, + 428, + 207 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 180, + 86, + 428, + 207 + ], + "spans": [ + { + "bbox": [ + 180, + 86, + 428, + 207 + ], + "score": 0.969, + "type": "image", + "image_path": "c74201d70aae1d3587e14044064872d252c0e00ec6d3c4fd6b70377bb7088c75.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 180, + 86, + 428, + 126.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 180, + 126.33333333333334, + 428, + 166.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 180, + 166.66666666666669, + 428, + 207.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 222, + 505, + 278 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 223, + 504, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 471, + 235 + ], + "score": 1.0, + "content": "Figure 6: Training curves and accuracies for MACs (our model), FiLM (Perez et al., 2017),", + "type": "text" + }, + { + "bbox": [ + 472, + 223, + 504, + 233 + ], + "score": 0.53, + "content": "\\mathrm { P G + E E }", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 234, + 504, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 471, + 246 + ], + "score": 1.0, + "content": "(Johnson et al., 2017) and stacked-attention (Yang et al., 2016; Johnson et al., 2017). (Note:", + "type": "text" + }, + { + "bbox": [ + 472, + 234, + 504, + 244 + ], + "score": 0.54, + "content": "\\mathrm { P G + E E }", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "uses the supported CLEVR programs as strong supervision.) Left: Training curve (accuracy/epoch).", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 255, + 506, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 256, + 270 + ], + "score": 1.0, + "content": "Right: Learning curve: Accuracy for", + "type": "text" + }, + { + "bbox": [ + 256, + 256, + 271, + 266 + ], + "score": 0.82, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 255, + 273, + 270 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 274, + 256, + 289, + 266 + ], + "score": 0.79, + "content": "2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 255, + 292, + 270 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 293, + 256, + 307, + 266 + ], + "score": 0.82, + "content": "5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 255, + 311, + 270 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 312, + 256, + 331, + 266 + ], + "score": 0.84, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 255, + 334, + 270 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 334, + 256, + 354, + 266 + ], + "score": 0.84, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 255, + 357, + 270 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 358, + 256, + 378, + 266 + ], + "score": 0.85, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 255, + 396, + 270 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 397, + 256, + 420, + 266 + ], + "score": 0.87, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 255, + 447, + 270 + ], + "score": 1.0, + "content": "of the", + "type": "text" + }, + { + "bbox": [ + 447, + 256, + 469, + 266 + ], + "score": 0.68, + "content": "7 0 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 255, + 506, + 270 + ], + "score": 1.0, + "content": "CLEVR", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 266, + 144, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 144, + 280 + ], + "score": 1.0, + "content": "samples.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "table", + "bbox": [ + 181, + 366, + 429, + 448 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 312, + 505, + 356 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 312, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 506, + 325 + ], + "score": 1.0, + "content": "Table 2: Accuracy on CLEVR-Humans of previous methods and our method (MAC), before (left)", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 323, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 369, + 336 + ], + "score": 1.0, + "content": "and after (right) fine-tuning on the CLEVR-Humans training data.", + "type": "text" + }, + { + "bbox": [ + 370, + 324, + 402, + 334 + ], + "score": 0.75, + "content": "\\mathrm { P G + E E }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 323, + 506, + 336 + ], + "score": 1.0, + "content": "uses supervised data. For", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "our method (MAC) we show test results with GLoVE initialization and validation results with and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 345, + 176, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 176, + 356 + ], + "score": 1.0, + "content": "without GLoVE.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "table_body", + "bbox": [ + 181, + 366, + 429, + 448 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 181, + 366, + 429, + 448 + ], + "spans": [ + { + "bbox": [ + 181, + 366, + 429, + 448 + ], + "score": 0.979, + "html": "
ModelTrain CLEVRTrain CLEVR + fine-tune HUMANS
LSTM (Johnson et al.,2017)27.536.5
CNN+LSTM (Johnson et al.,2017)37.743.2
CNN+LSTM+SA+MLP (Johnson et al.,2016)50.457.6
PG+EE (18K prog.)* (Johnson et al.,2017)54.066.6
CNN+GRU+FiLM (Perez et al., 2017)56.675.9
MAC-val56.981.2
MAC-val58.582.0
MAC-test58.682.5
", + "type": "table", + "image_path": "20d0a9d88276e853b2d4290d832c31f6ed7a6d4ae845cb8a1f4754688e965048.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 181, + 366, + 429, + 393.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 181, + 393.3333333333333, + 429, + 420.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 181, + 420.66666666666663, + 429, + 447.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 11.25 + }, + { + "type": "title", + "bbox": [ + 107, + 491, + 209, + 502 + ], + "lines": [ + { + "bbox": [ + 105, + 490, + 210, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 210, + 504 + ], + "score": 1.0, + "content": "4.1 DATA EFFICIENCY", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 522, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 534 + ], + "score": 1.0, + "content": "We have explored the performance of our and other leading approaches on smaller subsets of the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "CLEVR dataset, in order to study the ability of models to generalize from smaller amount of data.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 303, + 556 + ], + "score": 1.0, + "content": "We sampled at random subsets of CLEVR, with", + "type": "text" + }, + { + "bbox": [ + 304, + 545, + 323, + 555 + ], + "score": 0.84, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 545, + 326, + 556 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 327, + 545, + 347, + 555 + ], + "score": 0.84, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 545, + 365, + 556 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 366, + 545, + 385, + 555 + ], + "score": 0.87, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 545, + 444, + 556 + ], + "score": 1.0, + "content": "of its original", + "type": "text" + }, + { + "bbox": [ + 444, + 545, + 466, + 555 + ], + "score": 0.44, + "content": "7 0 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 545, + 505, + 556 + ], + "score": 1.0, + "content": "size, and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "used them to train our and other 3 proposed models for the CLEVR task: FiLM (Perez et al., 2017),", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 204, + 578 + ], + "score": 1.0, + "content": "the strongly-supervised", + "type": "text" + }, + { + "bbox": [ + 205, + 566, + 237, + 577 + ], + "score": 0.67, + "content": "\\mathrm { P G + E E }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "(Johnson et al., 2017), and stacked-attention networks (Johnson", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 577, + 231, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 231, + 590 + ], + "score": 1.0, + "content": "et al., 2017; Yang et al., 2016).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 523, + 506, + 590 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 594, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 504, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 504, + 605 + ], + "score": 1.0, + "content": "As shown in figure 4, our model outperforms the other models by a wide margin for all subsets of the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 193, + 617 + ], + "score": 1.0, + "content": "CLEVR dataset. For", + "type": "text" + }, + { + "bbox": [ + 193, + 605, + 213, + 615 + ], + "score": 0.87, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 604, + 317, + 617 + ], + "score": 1.0, + "content": "of the data, equivalent to", + "type": "text" + }, + { + "bbox": [ + 317, + 605, + 339, + 615 + ], + "score": 0.63, + "content": "3 5 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "samples, other models obtain accuracies", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 104, + 615, + 176, + 629 + ], + "score": 1.0, + "content": "ranging between", + "type": "text" + }, + { + "bbox": [ + 176, + 616, + 196, + 627 + ], + "score": 0.88, + "content": "70 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 615, + 215, + 629 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 215, + 616, + 235, + 626 + ], + "score": 0.88, + "content": "92 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 615, + 345, + 629 + ], + "score": 1.0, + "content": ", while our model achieves", + "type": "text" + }, + { + "bbox": [ + 345, + 616, + 372, + 627 + ], + "score": 0.87, + "content": "9 7 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 615, + 506, + 629 + ], + "score": 1.0, + "content": ". The gap becomes larger as the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 205, + 640 + ], + "score": 1.0, + "content": "dataset size reduces: for", + "type": "text" + }, + { + "bbox": [ + 205, + 627, + 225, + 637 + ], + "score": 0.88, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 626, + 327, + 640 + ], + "score": 1.0, + "content": "of the data, equivalent to", + "type": "text" + }, + { + "bbox": [ + 327, + 627, + 349, + 637 + ], + "score": 0.5, + "content": "1 7 5 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "samples, performance of other models", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 636, + 416, + 651 + ], + "spans": [ + { + "bbox": [ + 104, + 636, + 151, + 651 + ], + "score": 1.0, + "content": "is between", + "type": "text" + }, + { + "bbox": [ + 151, + 638, + 171, + 649 + ], + "score": 0.86, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 636, + 189, + 651 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 189, + 638, + 208, + 649 + ], + "score": 0.87, + "content": "7 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 636, + 348, + 651 + ], + "score": 1.0, + "content": ", while our model maintains a high", + "type": "text" + }, + { + "bbox": [ + 348, + 638, + 375, + 649 + ], + "score": 0.83, + "content": "9 5 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 636, + 416, + 651 + ], + "score": 1.0, + "content": "accuracy.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24, + "bbox_fs": [ + 104, + 594, + 506, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 154, + 667 + ], + "score": 1.0, + "content": "Finally, for", + "type": "text" + }, + { + "bbox": [ + 154, + 655, + 173, + 666 + ], + "score": 0.87, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 655, + 219, + 667 + ], + "score": 1.0, + "content": "of the data", + "type": "text" + }, + { + "bbox": [ + 219, + 655, + 244, + 666 + ], + "score": 0.29, + "content": "- 7 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "samples, still a sizeable amount – our model is the only one that", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 284, + 678 + ], + "score": 1.0, + "content": "manages to generalize, with performance of", + "type": "text" + }, + { + "bbox": [ + 284, + 666, + 311, + 676 + ], + "score": 0.87, + "content": "8 4 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "on average, whereas the other three models fail,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 147, + 689 + ], + "score": 1.0, + "content": "achieving", + "type": "text" + }, + { + "bbox": [ + 147, + 677, + 204, + 687 + ], + "score": 0.89, + "content": "4 7 . 6 \\% - 5 7 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 677, + 505, + 689 + ], + "score": 1.0, + "content": ". Note that as pointed out by (Johnson et al., 2016) a simple baseline that", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 686, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 104, + 686, + 429, + 702 + ], + "score": 1.0, + "content": "predicts the most frequent answer for each of the question types achieves already", + "type": "text" + }, + { + "bbox": [ + 429, + 688, + 456, + 699 + ], + "score": 0.87, + "content": "4 2 . 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 686, + 506, + 702 + ], + "score": 1.0, + "content": ", suggesting", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "that answering half of the questions correctly means that the competing models barely learn to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "generalize from the smaller dataset. These results demonstrate the robustness of our architecture", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 720, + 484, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 484, + 732 + ], + "score": 1.0, + "content": "and its key role as a structural prior guiding our network to learn the intended reasoning skills.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30, + "bbox_fs": [ + 104, + 655, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 83, + 363, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 364, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 364, + 96 + ], + "score": 1.0, + "content": "4.2 CLEVR HUMANS - NATURAL LANGUAGE QUESTIONS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 104, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "We analyze our model performance on the CLEVR-Humans dataset (Johnson et al., 2017), consist-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 128 + ], + "score": 1.0, + "content": "ing of natural language questions collected through crowdsourcing. As such, the dataset has diverse", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 127, + 452, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 452, + 139 + ], + "score": 1.0, + "content": "vocabulary and linguistic variations, and it also demands more varied reasoning skills.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 328, + 156 + ], + "score": 1.0, + "content": "Since the training set is relatively small, consisting of", + "type": "text" + }, + { + "bbox": [ + 328, + 144, + 345, + 154 + ], + "score": 0.33, + "content": "1 8 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "samples, we use it to finetune a model", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "score": 1.0, + "content": "pretrained on the standard CLEVR dataset. However, since most of the vocabulary in CLEVR-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 164, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 180 + ], + "score": 1.0, + "content": "Humans is not covered by CLEVR, we do not train the word vectors during the pre-training stage,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 176, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 190 + ], + "score": 1.0, + "content": "so to prevent drift in their meaning compared to other uncovered words in CLEVR-Humans that", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 188, + 222, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 222, + 200 + ], + "score": 1.0, + "content": "may be semantically related.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 204, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 505, + 216 + ], + "score": 1.0, + "content": "As shown in table 2, our model achieves state-of-the-art performance on CLEVR-Humans both", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 215, + 504, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 481, + 227 + ], + "score": 1.0, + "content": "before and after fine-tuning. It surpasses the next-best FiLM model, (Perez et al., 2017) by", + "type": "text" + }, + { + "bbox": [ + 482, + 215, + 504, + 226 + ], + "score": 0.79, + "content": "6 . 6 \\%", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 227, + 213, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 182, + 239 + ], + "score": 1.0, + "content": "percent, achieving", + "type": "text" + }, + { + "bbox": [ + 182, + 227, + 209, + 237 + ], + "score": 0.85, + "content": "8 2 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 227, + 213, + 239 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 243, + 505, + 299 + ], + "lines": [ + { + "bbox": [ + 105, + 243, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 256 + ], + "score": 1.0, + "content": "The results substantiate the model’s robustness against linguistic variations and noise, as well as", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "its ability to adapt to diverse vocabulary and varied reasoning skills. Arguably, the soft attention", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 265, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 104, + 265, + 506, + 279 + ], + "score": 1.0, + "content": "performed over the question words allows the model to focus on the words that are most critical to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 276, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 290 + ], + "score": 1.0, + "content": "answer the question and translate them to corresponding reasoning operations, giving less attention", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 287, + 240, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 240, + 300 + ], + "score": 1.0, + "content": "to irrelevant linguistic variations.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + }, + { + "type": "title", + "bbox": [ + 107, + 317, + 182, + 327 + ], + "lines": [ + { + "bbox": [ + 105, + 315, + 183, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 183, + 330 + ], + "score": 1.0, + "content": "4.3 ABLATIONS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 108, + 348, + 501, + 609 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 348, + 501, + 609 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 348, + 501, + 609 + ], + "spans": [ + { + "bbox": [ + 108, + 348, + 501, + 609 + ], + "score": 0.972, + "type": "image", + "image_path": "d542dee22ea6b6971ea171ea3647286d90bc0976dc1e0a42885a1582ff9227cc.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 108, + 348, + 501, + 435.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 108, + 435.0, + 501, + 522.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 108, + 522.0, + 501, + 609.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 145, + 624, + 464, + 636 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 145, + 623, + 465, + 637 + ], + "spans": [ + { + "bbox": [ + 145, + 623, + 465, + 637 + ], + "score": 1.0, + "content": "Figure 7: Ablations and variations of the MAC network. See text for full detail.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + } + ], + "index": 20.0 + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "Based on the validation set, we have conducted an ablation study on our model to understand better", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "the contribution of each of its component to the overall performance. We tested each setting on the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 289, + 690 + ], + "score": 1.0, + "content": "standard 700K CLEVR dataset as well as on", + "type": "text" + }, + { + "bbox": [ + 289, + 677, + 308, + 687 + ], + "score": 0.87, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "subset of the dataset. See table 3 for the numer-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "ical results. In addition, figure 4.3 presents the training curves for the different settings trained on", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 697, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 505, + 712 + ], + "score": 1.0, + "content": "the standard dataset. Overall, the results demonstrate the robustness of the model to hyperparam-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "eter variations such as network dimension and length, and also the impact of different aspect and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 721, + 272, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 272, + 734 + ], + "score": 1.0, + "content": "components of MAC on its performance.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 83, + 363, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 364, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 364, + 96 + ], + "score": 1.0, + "content": "4.2 CLEVR HUMANS - NATURAL LANGUAGE QUESTIONS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 81, + 364, + 96 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 104, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "We analyze our model performance on the CLEVR-Humans dataset (Johnson et al., 2017), consist-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 128 + ], + "score": 1.0, + "content": "ing of natural language questions collected through crowdsourcing. As such, the dataset has diverse", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 127, + 452, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 452, + 139 + ], + "score": 1.0, + "content": "vocabulary and linguistic variations, and it also demands more varied reasoning skills.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 104, + 505, + 139 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 328, + 156 + ], + "score": 1.0, + "content": "Since the training set is relatively small, consisting of", + "type": "text" + }, + { + "bbox": [ + 328, + 144, + 345, + 154 + ], + "score": 0.33, + "content": "1 8 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "samples, we use it to finetune a model", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "score": 1.0, + "content": "pretrained on the standard CLEVR dataset. However, since most of the vocabulary in CLEVR-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 164, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 180 + ], + "score": 1.0, + "content": "Humans is not covered by CLEVR, we do not train the word vectors during the pre-training stage,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 176, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 190 + ], + "score": 1.0, + "content": "so to prevent drift in their meaning compared to other uncovered words in CLEVR-Humans that", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 188, + 222, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 222, + 200 + ], + "score": 1.0, + "content": "may be semantically related.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 143, + 505, + 200 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 204, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 505, + 216 + ], + "score": 1.0, + "content": "As shown in table 2, our model achieves state-of-the-art performance on CLEVR-Humans both", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 215, + 504, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 481, + 227 + ], + "score": 1.0, + "content": "before and after fine-tuning. It surpasses the next-best FiLM model, (Perez et al., 2017) by", + "type": "text" + }, + { + "bbox": [ + 482, + 215, + 504, + 226 + ], + "score": 0.79, + "content": "6 . 6 \\%", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 227, + 213, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 182, + 239 + ], + "score": 1.0, + "content": "percent, achieving", + "type": "text" + }, + { + "bbox": [ + 182, + 227, + 209, + 237 + ], + "score": 0.85, + "content": "8 2 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 227, + 213, + 239 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 204, + 505, + 239 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 243, + 505, + 299 + ], + "lines": [ + { + "bbox": [ + 105, + 243, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 256 + ], + "score": 1.0, + "content": "The results substantiate the model’s robustness against linguistic variations and noise, as well as", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "its ability to adapt to diverse vocabulary and varied reasoning skills. Arguably, the soft attention", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 265, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 104, + 265, + 506, + 279 + ], + "score": 1.0, + "content": "performed over the question words allows the model to focus on the words that are most critical to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 276, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 290 + ], + "score": 1.0, + "content": "answer the question and translate them to corresponding reasoning operations, giving less attention", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 287, + 240, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 240, + 300 + ], + "score": 1.0, + "content": "to irrelevant linguistic variations.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14, + "bbox_fs": [ + 104, + 243, + 506, + 300 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 317, + 182, + 327 + ], + "lines": [ + { + "bbox": [ + 105, + 315, + 183, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 183, + 330 + ], + "score": 1.0, + "content": "4.3 ABLATIONS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 108, + 348, + 501, + 609 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 348, + 501, + 609 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 348, + 501, + 609 + ], + "spans": [ + { + "bbox": [ + 108, + 348, + 501, + 609 + ], + "score": 0.972, + "type": "image", + "image_path": "d542dee22ea6b6971ea171ea3647286d90bc0976dc1e0a42885a1582ff9227cc.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 108, + 348, + 501, + 435.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 108, + 435.0, + 501, + 522.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 108, + 522.0, + 501, + 609.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 145, + 624, + 464, + 636 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 145, + 623, + 465, + 637 + ], + "spans": [ + { + "bbox": [ + 145, + 623, + 465, + 637 + ], + "score": 1.0, + "content": "Figure 7: Ablations and variations of the MAC network. See text for full detail.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + } + ], + "index": 20.0 + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "Based on the validation set, we have conducted an ablation study on our model to understand better", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "the contribution of each of its component to the overall performance. We tested each setting on the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 289, + 690 + ], + "score": 1.0, + "content": "standard 700K CLEVR dataset as well as on", + "type": "text" + }, + { + "bbox": [ + 289, + 677, + 308, + 687 + ], + "score": 0.87, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "subset of the dataset. See table 3 for the numer-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "ical results. In addition, figure 4.3 presents the training curves for the different settings trained on", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 697, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 505, + 712 + ], + "score": 1.0, + "content": "the standard dataset. Overall, the results demonstrate the robustness of the model to hyperparam-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "eter variations such as network dimension and length, and also the impact of different aspect and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 721, + 272, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 272, + 734 + ], + "score": 1.0, + "content": "components of MAC on its performance.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 655, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 111, + 423, + 265 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 80, + 504, + 102 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 79, + 505, + 92 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 383, + 92 + ], + "score": 1.0, + "content": "Table 3: Ablations Results for the full standard CLEVR dataset and", + "type": "text" + }, + { + "bbox": [ + 384, + 81, + 403, + 91 + ], + "score": 0.85, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 79, + 505, + 92 + ], + "score": 1.0, + "content": "subset of it. See text for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 89, + 155, + 104 + ], + "spans": [ + { + "bbox": [ + 105, + 89, + 155, + 104 + ], + "score": 1.0, + "content": "full details.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 187, + 111, + 423, + 265 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 187, + 111, + 423, + 265 + ], + "spans": [ + { + "bbox": [ + 187, + 111, + 423, + 265 + ], + "score": 0.976, + "html": "
ModelStandard CLEVR10% CLEVR
MAC standard model (dimension 512)98.984.5
dimension 25698.476.3
dimension 12897.677.0
w/o control55.651.5
w/o word-attention95.363.2
full-question-as-control80.765.0
w/o positional encoding99.180.0
linear position encoding99.083.2
identity writing unit98.584.5
only-gating writing unit99.383.1
+ self-attention83.299.2
+ gating99.483.1
+ self-attention and gating99.585.5
gate bias 098.784.9
gate bias 199.468.5
gate bias -199.077.1
unshared weights97.867.5
prediction without question information97.864.7
", + "type": "table", + "image_path": "cbfa912b869921d24b8997ef84de01c88a55b56dcf5407ed4090070e65e15576.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 187, + 111, + 423, + 125.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 187, + 125.0, + 423, + 139.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 187, + 139.0, + 423, + 153.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 187, + 153.0, + 423, + 167.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 187, + 167.0, + 423, + 181.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 187, + 181.0, + 423, + 195.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 187, + 195.0, + 423, + 209.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 187, + 209.0, + 423, + 223.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 187, + 223.0, + 423, + 237.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 187, + 237.0, + 423, + 251.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 187, + 251.0, + 423, + 265.0 + ], + "spans": [], + "index": 12 + } + ] + } + ], + "index": 3.75 + }, + { + "type": "text", + "bbox": [ + 106, + 302, + 505, + 380 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 505, + 314 + ], + "score": 1.0, + "content": "Network Length. We have tested the model performance as a function of the network’s length –", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "the number of MAC cells that were sequenced together. The results show the positive correlation", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "between the network length and its performance. We can see that for 1 cell the scores are relatively", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 335, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 152, + 345 + ], + "score": 0.4, + "content": "1 0 \\mathrm { w } - 7 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 335, + 506, + 347 + ], + "score": 1.0, + "content": ", but adding at least one more cell leads to a significant increase in performance above", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 345, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 126, + 357 + ], + "score": 0.85, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 345, + 403, + 359 + ], + "score": 1.0, + "content": ". The performance keeps improving up to lengths 8-16 that achieve", + "type": "text" + }, + { + "bbox": [ + 404, + 346, + 451, + 357 + ], + "score": 0.87, + "content": "9 8 . 9 – 9 9 . 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 345, + 505, + 359 + ], + "score": 1.0, + "content": ". The results", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "score": 1.0, + "content": "also teach us about the complexity of the dataset, by showing the relatively significant benefits of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 367, + 327, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 327, + 381 + ], + "score": 1.0, + "content": "having at least 4 cells, each modeling a reasoning step.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 440 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "Network Dimension. We have varied the state dimension to check the robustness of the model to", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "score": 1.0, + "content": "hyperparameters. The results on the standard CLEVR dataset show the model is able to maintain", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 474, + 420 + ], + "score": 1.0, + "content": "high performance with dimension of 128, albeit after a longer training process, achieving", + "type": "text" + }, + { + "bbox": [ + 474, + 407, + 501, + 418 + ], + "score": 0.84, + "content": "9 7 . 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 406, + 505, + 420 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 416, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 160, + 432 + ], + "score": 1.0, + "content": "compared to", + "type": "text" + }, + { + "bbox": [ + 160, + 418, + 193, + 428 + ], + "score": 0.88, + "content": "9 8 . 9 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 416, + 392, + 432 + ], + "score": 1.0, + "content": "achieved with dimension of 512. However, for", + "type": "text" + }, + { + "bbox": [ + 392, + 418, + 411, + 428 + ], + "score": 0.87, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 416, + 506, + 432 + ], + "score": 1.0, + "content": "of CLEVR, the larger", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 428, + 400, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 282, + 442 + ], + "score": 1.0, + "content": "512-dimension allows accuracy increase by", + "type": "text" + }, + { + "bbox": [ + 282, + 429, + 304, + 439 + ], + "score": 0.86, + "content": "7 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 428, + 400, + 442 + ], + "score": 1.0, + "content": "over dimension of 128.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 445, + 505, + 511 + ], + "lines": [ + { + "bbox": [ + 106, + 446, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 457 + ], + "score": 1.0, + "content": "Weight Sharing. We have tested the impact of sharing weights between cell has on the model", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 249, + 469 + ], + "score": 1.0, + "content": "performance for network of length", + "type": "text" + }, + { + "bbox": [ + 249, + 457, + 281, + 468 + ], + "score": 0.91, + "content": "p = 1 2", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 456, + 506, + 469 + ], + "score": 1.0, + "content": ". The results show that for the standard dataset there is", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 308, + 480 + ], + "score": 1.0, + "content": "only a small difference between these settings of", + "type": "text" + }, + { + "bbox": [ + 308, + 468, + 323, + 478 + ], + "score": 0.86, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 467, + 506, + 480 + ], + "score": 1.0, + "content": ". However, for less data, we see much more", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 172, + 491 + ], + "score": 1.0, + "content": "significant drop", + "type": "text" + }, + { + "bbox": [ + 172, + 478, + 199, + 489 + ], + "score": 0.86, + "content": "1 6 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "in the unshared-parameters setting compared to the shared one. Indeed, we", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "observe that a model with less parameter is more data-efficient and has a lower tendency to overfit", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 499, + 143, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 143, + 513 + ], + "score": 1.0, + "content": "the data.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 517, + 505, + 639 + ], + "lines": [ + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "Control Unit. We have performed several ablations in the control unit to understand its contribution", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "to the overall model performance. Based on the results, first, we can see the the question information", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "is crucial for the model to handle the questions, as can be noted by the low performance of the model", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "when there is no use of control signal whatsoever. Second, we have tested the model performance", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 560, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 506, + 575 + ], + "score": 1.0, + "content": "when using the continuous control state computed by question (2) in section 3.2.1, without having", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 573, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 505, + 584 + ], + "score": 1.0, + "content": "word-attention, in order to understand its relative contribution. Based on the results, we can indeed", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "see that using word-attention is useful for accelerating the training process and achieving higher", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "accuracies both for the standard dataset as well as for the small subset, where using word-attention", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 604, + 504, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 183, + 617 + ], + "score": 1.0, + "content": "increases results in", + "type": "text" + }, + { + "bbox": [ + 183, + 605, + 211, + 615 + ], + "score": 0.87, + "content": "2 1 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 604, + 504, + 617 + ], + "score": 1.0, + "content": ". We also see that using the “contextual words” produced by the question-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "unit LSTM is useful in accelerating the model performance, when compared to using the word-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 627, + 172, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 172, + 639 + ], + "score": 1.0, + "content": "vectors directly.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 655 + ], + "score": 1.0, + "content": "Reading Unit. We have conducted several ablations for the reading unit to better understand its", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "behavior and contribution to the performance of the model. The standard MAC reading unit uses the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "control state – which averages the question words based on attention distributions computed per each", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "reasoning step. In this ablation experiment, we have tested using the full question representation", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 113, + 699 + ], + "score": 0.74, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "instead across all reasoning steps to gain better understanding of the the contribution of word-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 698, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 104, + 698, + 374, + 712 + ], + "score": 1.0, + "content": "attention to the model performance. Indeed, we can see that using", + "type": "text" + }, + { + "bbox": [ + 375, + 701, + 381, + 710 + ], + "score": 0.77, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 698, + 495, + 712 + ], + "score": 1.0, + "content": "rather then the control state", + "type": "text" + }, + { + "bbox": [ + 495, + 701, + 504, + 710 + ], + "score": 0.8, + "content": "c _ { i }", + "type": "inline_equation" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 710, + 504, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 282, + 721 + ], + "score": 1.0, + "content": "results in a significant drops in performance", + "type": "text" + }, + { + "bbox": [ + 282, + 710, + 316, + 720 + ], + "score": 0.8, + "content": "- 1 9 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 710, + 443, + 721 + ], + "score": 1.0, + "content": "for the full CLEVR dataset and", + "type": "text" + }, + { + "bbox": [ + 443, + 710, + 470, + 720 + ], + "score": 0.86, + "content": "1 9 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 710, + 484, + 721 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 485, + 710, + 504, + 720 + ], + "score": 0.85, + "content": "10 \\%", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 154, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 154, + 732 + ], + "score": 1.0, + "content": "of the data.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 45.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 111, + 423, + 265 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 80, + 504, + 102 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 79, + 505, + 92 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 383, + 92 + ], + "score": 1.0, + "content": "Table 3: Ablations Results for the full standard CLEVR dataset and", + "type": "text" + }, + { + "bbox": [ + 384, + 81, + 403, + 91 + ], + "score": 0.85, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 79, + 505, + 92 + ], + "score": 1.0, + "content": "subset of it. See text for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 89, + 155, + 104 + ], + "spans": [ + { + "bbox": [ + 105, + 89, + 155, + 104 + ], + "score": 1.0, + "content": "full details.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 187, + 111, + 423, + 265 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 187, + 111, + 423, + 265 + ], + "spans": [ + { + "bbox": [ + 187, + 111, + 423, + 265 + ], + "score": 0.976, + "html": "
ModelStandard CLEVR10% CLEVR
MAC standard model (dimension 512)98.984.5
dimension 25698.476.3
dimension 12897.677.0
w/o control55.651.5
w/o word-attention95.363.2
full-question-as-control80.765.0
w/o positional encoding99.180.0
linear position encoding99.083.2
identity writing unit98.584.5
only-gating writing unit99.383.1
+ self-attention83.299.2
+ gating99.483.1
+ self-attention and gating99.585.5
gate bias 098.784.9
gate bias 199.468.5
gate bias -199.077.1
unshared weights97.867.5
prediction without question information97.864.7
", + "type": "table", + "image_path": "cbfa912b869921d24b8997ef84de01c88a55b56dcf5407ed4090070e65e15576.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 187, + 111, + 423, + 125.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 187, + 125.0, + 423, + 139.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 187, + 139.0, + 423, + 153.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 187, + 153.0, + 423, + 167.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 187, + 167.0, + 423, + 181.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 187, + 181.0, + 423, + 195.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 187, + 195.0, + 423, + 209.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 187, + 209.0, + 423, + 223.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 187, + 223.0, + 423, + 237.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 187, + 237.0, + 423, + 251.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 187, + 251.0, + 423, + 265.0 + ], + "spans": [], + "index": 12 + } + ] + } + ], + "index": 3.75 + }, + { + "type": "text", + "bbox": [ + 106, + 302, + 505, + 380 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 505, + 314 + ], + "score": 1.0, + "content": "Network Length. We have tested the model performance as a function of the network’s length –", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "the number of MAC cells that were sequenced together. The results show the positive correlation", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "between the network length and its performance. We can see that for 1 cell the scores are relatively", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 335, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 152, + 345 + ], + "score": 0.4, + "content": "1 0 \\mathrm { w } - 7 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 335, + 506, + 347 + ], + "score": 1.0, + "content": ", but adding at least one more cell leads to a significant increase in performance above", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 345, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 126, + 357 + ], + "score": 0.85, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 345, + 403, + 359 + ], + "score": 1.0, + "content": ". The performance keeps improving up to lengths 8-16 that achieve", + "type": "text" + }, + { + "bbox": [ + 404, + 346, + 451, + 357 + ], + "score": 0.87, + "content": "9 8 . 9 – 9 9 . 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 345, + 505, + 359 + ], + "score": 1.0, + "content": ". The results", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "score": 1.0, + "content": "also teach us about the complexity of the dataset, by showing the relatively significant benefits of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 367, + 327, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 327, + 381 + ], + "score": 1.0, + "content": "having at least 4 cells, each modeling a reasoning step.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 302, + 506, + 381 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 440 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "Network Dimension. We have varied the state dimension to check the robustness of the model to", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "score": 1.0, + "content": "hyperparameters. The results on the standard CLEVR dataset show the model is able to maintain", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 474, + 420 + ], + "score": 1.0, + "content": "high performance with dimension of 128, albeit after a longer training process, achieving", + "type": "text" + }, + { + "bbox": [ + 474, + 407, + 501, + 418 + ], + "score": 0.84, + "content": "9 7 . 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 406, + 505, + 420 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 416, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 160, + 432 + ], + "score": 1.0, + "content": "compared to", + "type": "text" + }, + { + "bbox": [ + 160, + 418, + 193, + 428 + ], + "score": 0.88, + "content": "9 8 . 9 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 416, + 392, + 432 + ], + "score": 1.0, + "content": "achieved with dimension of 512. However, for", + "type": "text" + }, + { + "bbox": [ + 392, + 418, + 411, + 428 + ], + "score": 0.87, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 416, + 506, + 432 + ], + "score": 1.0, + "content": "of CLEVR, the larger", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 428, + 400, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 282, + 442 + ], + "score": 1.0, + "content": "512-dimension allows accuracy increase by", + "type": "text" + }, + { + "bbox": [ + 282, + 429, + 304, + 439 + ], + "score": 0.86, + "content": "7 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 428, + 400, + 442 + ], + "score": 1.0, + "content": "over dimension of 128.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 385, + 506, + 442 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 445, + 505, + 511 + ], + "lines": [ + { + "bbox": [ + 106, + 446, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 457 + ], + "score": 1.0, + "content": "Weight Sharing. We have tested the impact of sharing weights between cell has on the model", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 249, + 469 + ], + "score": 1.0, + "content": "performance for network of length", + "type": "text" + }, + { + "bbox": [ + 249, + 457, + 281, + 468 + ], + "score": 0.91, + "content": "p = 1 2", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 456, + 506, + 469 + ], + "score": 1.0, + "content": ". The results show that for the standard dataset there is", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 308, + 480 + ], + "score": 1.0, + "content": "only a small difference between these settings of", + "type": "text" + }, + { + "bbox": [ + 308, + 468, + 323, + 478 + ], + "score": 0.86, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 467, + 506, + 480 + ], + "score": 1.0, + "content": ". However, for less data, we see much more", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 172, + 491 + ], + "score": 1.0, + "content": "significant drop", + "type": "text" + }, + { + "bbox": [ + 172, + 478, + 199, + 489 + ], + "score": 0.86, + "content": "1 6 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "in the unshared-parameters setting compared to the shared one. Indeed, we", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "observe that a model with less parameter is more data-efficient and has a lower tendency to overfit", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 499, + 143, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 143, + 513 + ], + "score": 1.0, + "content": "the data.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 446, + 506, + 513 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 517, + 505, + 639 + ], + "lines": [ + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "Control Unit. We have performed several ablations in the control unit to understand its contribution", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "to the overall model performance. Based on the results, first, we can see the the question information", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "is crucial for the model to handle the questions, as can be noted by the low performance of the model", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "when there is no use of control signal whatsoever. Second, we have tested the model performance", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 560, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 506, + 575 + ], + "score": 1.0, + "content": "when using the continuous control state computed by question (2) in section 3.2.1, without having", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 573, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 505, + 584 + ], + "score": 1.0, + "content": "word-attention, in order to understand its relative contribution. Based on the results, we can indeed", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "see that using word-attention is useful for accelerating the training process and achieving higher", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "accuracies both for the standard dataset as well as for the small subset, where using word-attention", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 604, + 504, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 183, + 617 + ], + "score": 1.0, + "content": "increases results in", + "type": "text" + }, + { + "bbox": [ + 183, + 605, + 211, + 615 + ], + "score": 0.87, + "content": "2 1 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 604, + 504, + 617 + ], + "score": 1.0, + "content": ". We also see that using the “contextual words” produced by the question-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "unit LSTM is useful in accelerating the model performance, when compared to using the word-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 627, + 172, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 172, + 639 + ], + "score": 1.0, + "content": "vectors directly.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 517, + 506, + 639 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 655 + ], + "score": 1.0, + "content": "Reading Unit. We have conducted several ablations for the reading unit to better understand its", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "behavior and contribution to the performance of the model. The standard MAC reading unit uses the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "control state – which averages the question words based on attention distributions computed per each", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "reasoning step. In this ablation experiment, we have tested using the full question representation", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 113, + 699 + ], + "score": 0.74, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "instead across all reasoning steps to gain better understanding of the the contribution of word-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 698, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 104, + 698, + 374, + 712 + ], + "score": 1.0, + "content": "attention to the model performance. Indeed, we can see that using", + "type": "text" + }, + { + "bbox": [ + 375, + 701, + 381, + 710 + ], + "score": 0.77, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 698, + 495, + 712 + ], + "score": 1.0, + "content": "rather then the control state", + "type": "text" + }, + { + "bbox": [ + 495, + 701, + 504, + 710 + ], + "score": 0.8, + "content": "c _ { i }", + "type": "inline_equation" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 710, + 504, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 282, + 721 + ], + "score": 1.0, + "content": "results in a significant drops in performance", + "type": "text" + }, + { + "bbox": [ + 282, + 710, + 316, + 720 + ], + "score": 0.8, + "content": "- 1 9 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 710, + 443, + 721 + ], + "score": 1.0, + "content": "for the full CLEVR dataset and", + "type": "text" + }, + { + "bbox": [ + 443, + 710, + 470, + 720 + ], + "score": 0.86, + "content": "1 9 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 710, + 484, + 721 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 485, + 710, + 504, + 720 + ], + "score": 0.85, + "content": "10 \\%", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 154, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 154, + 732 + ], + "score": 1.0, + "content": "of the data.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 45.5, + "bbox_fs": [ + 104, + 643, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "We have conducted additional ablation experiment to better understand the contribution of using", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "the KB features directly in the first-stage information retrieval process described in section 3.2.2,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 102, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 104, + 102, + 477, + 119 + ], + "score": 1.0, + "content": "compared to using only the dot-products of the KB elements with the previous memory state", + "type": "text" + }, + { + "bbox": [ + 478, + 106, + 501, + 116 + ], + "score": 0.88, + "content": "m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 102, + 506, + 119 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "For the full CLEVR dataset, we can see that this component has only a small impact in the final", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 258, + 138 + ], + "score": 1.0, + "content": "performance - ultimately resulting in", + "type": "text" + }, + { + "bbox": [ + 258, + 126, + 285, + 137 + ], + "score": 0.87, + "content": "0 . 0 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 127, + 458, + 138 + ], + "score": 1.0, + "content": "performance difference. However, for the", + "type": "text" + }, + { + "bbox": [ + 458, + 127, + 478, + 137 + ], + "score": 0.86, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 127, + 505, + 138 + ], + "score": 1.0, + "content": "of the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "data, we can see that the difference in performance when ablating this component is much larger -", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 148, + 138, + 160 + ], + "spans": [ + { + "bbox": [ + 107, + 148, + 133, + 159 + ], + "score": 0.83, + "content": "1 1 . 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 148, + 138, + 160 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 165, + 504, + 253 + ], + "lines": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "Writing Unit Ablations. In our main MAC model variant, the memory unit merges the new in-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 149, + 190 + ], + "score": 1.0, + "content": "formation", + "type": "text" + }, + { + "bbox": [ + 149, + 178, + 174, + 187 + ], + "score": 0.88, + "content": "m _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 176, + 306, + 190 + ], + "score": 1.0, + "content": "with the previous memory state", + "type": "text" + }, + { + "bbox": [ + 306, + 178, + 329, + 188 + ], + "score": 0.89, + "content": "m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 176, + 506, + 190 + ], + "score": 1.0, + "content": "by combining them through a linear trans-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 504, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 452, + 201 + ], + "score": 1.0, + "content": "formation. In this experiment, we have explored other variations, such as assigning", + "type": "text" + }, + { + "bbox": [ + 452, + 189, + 477, + 199 + ], + "score": 0.88, + "content": "m _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 187, + 490, + 201 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 490, + 189, + 504, + 198 + ], + "score": 0.83, + "content": "m _ { i }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 199, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 436, + 210 + ], + "score": 1.0, + "content": "directly – ignoring previous memories, or doing a linear transformation based on", + "type": "text" + }, + { + "bbox": [ + 436, + 200, + 461, + 209 + ], + "score": 0.88, + "content": "m _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 199, + 506, + 210 + ], + "score": 1.0, + "content": "only. The", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 443, + 221 + ], + "score": 1.0, + "content": "results show that in fact such variant is only slightly worse than our main variant –", + "type": "text" + }, + { + "bbox": [ + 443, + 210, + 465, + 220 + ], + "score": 0.73, + "content": "0 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 208, + 506, + 221 + ], + "score": 1.0, + "content": ". We also", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "score": 1.0, + "content": "conducted an experiment in which we merge the new information with the previous memory just by", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 506, + 243 + ], + "score": 1.0, + "content": "a having a gate that does a weighted average of them. The results show that this variant performs", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 243, + 339, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 339, + 253 + ], + "score": 1.0, + "content": "equivalently to our standard linear-transformation variant.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 258, + 504, + 336 + ], + "lines": [ + { + "bbox": [ + 106, + 259, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 505, + 271 + ], + "score": 1.0, + "content": "Writing Unit Additions. We have explored the impact of the writing unit variants described in sec-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 269, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 506, + 283 + ], + "score": 1.0, + "content": "tion 3.2.3 – adding self-attention, gating mechanisms, or both, compared to our standard main model", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 280, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 406, + 294 + ], + "score": 1.0, + "content": "that uses a linear transformation to merge the newly retrieved information", + "type": "text" + }, + { + "bbox": [ + 406, + 282, + 431, + 292 + ], + "score": 0.89, + "content": "m _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 280, + 506, + 294 + ], + "score": 1.0, + "content": "with the previous", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 290, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 104, + 290, + 142, + 306 + ], + "score": 1.0, + "content": "memory", + "type": "text" + }, + { + "bbox": [ + 142, + 294, + 155, + 303 + ], + "score": 0.86, + "content": "m _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 290, + 506, + 306 + ], + "score": 1.0, + "content": ". For the complete CLEVR dataset we can see that indeed both these variants are very", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 303, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 315 + ], + "score": 1.0, + "content": "helpful in increasing the model performance. Compared to our standard MAC model that achieves", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 313, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 138, + 324 + ], + "score": 0.87, + "content": "9 8 . 9 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 313, + 362, + 326 + ], + "score": 1.0, + "content": "on the validation set, self-attention yields accuracy of", + "type": "text" + }, + { + "bbox": [ + 362, + 314, + 394, + 325 + ], + "score": 0.87, + "content": "9 9 . 2 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 313, + 453, + 326 + ], + "score": 1.0, + "content": ", gating yields", + "type": "text" + }, + { + "bbox": [ + 454, + 314, + 486, + 324 + ], + "score": 0.87, + "content": "9 9 . 3 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 313, + 506, + 326 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 325, + 228, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 192, + 337 + ], + "score": 1.0, + "content": "adding both achieves", + "type": "text" + }, + { + "bbox": [ + 192, + 325, + 225, + 335 + ], + "score": 0.87, + "content": "9 9 . 4 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 325, + 228, + 337 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 342, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "Output Unit. In our standard model, the final predictions made in the output unit are based on the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 183, + 365 + ], + "score": 1.0, + "content": "final memory state", + "type": "text" + }, + { + "bbox": [ + 184, + 354, + 198, + 365 + ], + "score": 0.87, + "content": "m _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 353, + 336, + 365 + ], + "score": 1.0, + "content": "as well as question representation", + "type": "text" + }, + { + "bbox": [ + 337, + 354, + 343, + 364 + ], + "score": 0.77, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "(stands for the final hidden states of the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "backward and forwards passes of the LSTM). We have explored the contribution of basing the model", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 374, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 388 + ], + "score": 1.0, + "content": "prediction on the latter, by testing the model performance when prediction is based on memory", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 225, + 398 + ], + "score": 1.0, + "content": "alone, for the complete and", + "type": "text" + }, + { + "bbox": [ + 225, + 385, + 245, + 396 + ], + "score": 0.87, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "datasets. We can see that in both settings basing the model’s", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 397, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 409 + ], + "score": 1.0, + "content": "predictions on the question representation allows faster training and higher accuracies. Notable is", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 231, + 420 + ], + "score": 1.0, + "content": "the gap in performance for the", + "type": "text" + }, + { + "bbox": [ + 232, + 407, + 251, + 418 + ], + "score": 0.85, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 408, + 294, + 420 + ], + "score": 1.0, + "content": "CLEVR -", + "type": "text" + }, + { + "bbox": [ + 294, + 407, + 321, + 418 + ], + "score": 0.82, + "content": "1 9 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 408, + 505, + 420 + ], + "score": 1.0, + "content": "increase by using the question representation", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 419, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 431 + ], + "score": 1.0, + "content": "to make predictions. These results are very reasonable intuitively, since the model is structured such", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 428, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 443 + ], + "score": 1.0, + "content": "that the memory holds only information that was retrieved from the image. Thus, questions that may", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 439, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 454 + ], + "score": 1.0, + "content": "ask for instance on different aspects (such as color or shape) of the same object in the image may", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 451, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 506, + 464 + ], + "score": 1.0, + "content": "result in the same memory content, which is thus does not directly contain enough information to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 462, + 204, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 204, + 474 + ], + "score": 1.0, + "content": "respond such questions.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 479, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "Position. In our standard model, similarly to the practice of competing models (Santoro et al., 2017;", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "Perez et al., 2017; Hu et al., 2017), we have concatenated positional information to each region of the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 501, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 513 + ], + "score": 1.0, + "content": "image, in order to increase the model capability to perform spatial reasoning. We have explored both", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 237, + 525 + ], + "score": 1.0, + "content": "simple linear maps at a constant", + "type": "text" + }, + { + "bbox": [ + 238, + 512, + 266, + 524 + ], + "score": 0.9, + "content": "[ - 1 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 511, + 505, + 525 + ], + "score": 1.0, + "content": "as well as more complex positional encoding suggested by", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 429, + 534 + ], + "score": 1.0, + "content": "(Vaswani et al., 2017). However, the results for both the standard dataset and the", + "type": "text" + }, + { + "bbox": [ + 430, + 523, + 449, + 533 + ], + "score": 0.87, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 522, + 505, + 534 + ], + "score": 1.0, + "content": "version show", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 533, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 104, + 533, + 506, + 547 + ], + "score": 1.0, + "content": "a very negligible improvement at best when adding positional encoding information, demonstrating", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 546, + 424, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 424, + 557 + ], + "score": 1.0, + "content": "the capability of MAC to perform spatial reasoning without data augmentation.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "Gate Bias Initialization. For our model variant with gating mechanism (described in section 3.2.3)", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 573, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 382, + 585 + ], + "score": 1.0, + "content": "we have tested the effect of setting different values for the gate bias -", + "type": "text" + }, + { + "bbox": [ + 382, + 573, + 406, + 584 + ], + "score": 0.85, + "content": "- 1 , 0", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 573, + 448, + 585 + ], + "score": 1.0, + "content": "and 1. for", + "type": "text" + }, + { + "bbox": [ + 448, + 573, + 462, + 583 + ], + "score": 0.66, + "content": "- 1", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 573, + 505, + 585 + ], + "score": 1.0, + "content": "the model", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 584, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 505, + 596 + ], + "score": 1.0, + "content": "is initialized to biased for keeping the previous memory value whereas for 1 it will be biased for", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 595, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 607 + ], + "score": 1.0, + "content": "using the new memory instead. We can see that for the complete dataset setting the bias to 1 is", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "score": 1.0, + "content": "optimal – apparently since the model has enough data to learn to apply each cell effectively. In", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 195, + 629 + ], + "score": 1.0, + "content": "contrast, for the small", + "type": "text" + }, + { + "bbox": [ + 196, + 617, + 215, + 627 + ], + "score": 0.85, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 616, + 506, + 629 + ], + "score": 1.0, + "content": "CLEVR data, setting the bias to 0 shows better performance, biasing the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "model to using less cells overall which results ultimately in a theoretically-simpler model that can", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 638, + 223, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 223, + 651 + ], + "score": 1.0, + "content": "fit less data more effectively.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 44.5 + }, + { + "type": "title", + "bbox": [ + 108, + 674, + 213, + 685 + ], + "lines": [ + { + "bbox": [ + 106, + 674, + 215, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 215, + 686 + ], + "score": 1.0, + "content": "4.4 INTERPRETABILITY", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 49 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "We have looked into attention maps over the image and question that the model produces during its", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "computation and provide a few examples in figure 4.4. The first example shows us how the model", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "parses the question in steps, first focusing on the main entity that the question is about, then on", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 51 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 763 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 13, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "We have conducted additional ablation experiment to better understand the contribution of using", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "the KB features directly in the first-stage information retrieval process described in section 3.2.2,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 102, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 104, + 102, + 477, + 119 + ], + "score": 1.0, + "content": "compared to using only the dot-products of the KB elements with the previous memory state", + "type": "text" + }, + { + "bbox": [ + 478, + 106, + 501, + 116 + ], + "score": 0.88, + "content": "m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 102, + 506, + 119 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "For the full CLEVR dataset, we can see that this component has only a small impact in the final", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 258, + 138 + ], + "score": 1.0, + "content": "performance - ultimately resulting in", + "type": "text" + }, + { + "bbox": [ + 258, + 126, + 285, + 137 + ], + "score": 0.87, + "content": "0 . 0 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 127, + 458, + 138 + ], + "score": 1.0, + "content": "performance difference. However, for the", + "type": "text" + }, + { + "bbox": [ + 458, + 127, + 478, + 137 + ], + "score": 0.86, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 127, + 505, + 138 + ], + "score": 1.0, + "content": "of the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "data, we can see that the difference in performance when ablating this component is much larger -", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 148, + 138, + 160 + ], + "spans": [ + { + "bbox": [ + 107, + 148, + 133, + 159 + ], + "score": 0.83, + "content": "1 1 . 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 148, + 138, + 160 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3, + "bbox_fs": [ + 104, + 81, + 506, + 160 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 165, + 504, + 253 + ], + "lines": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "Writing Unit Ablations. In our main MAC model variant, the memory unit merges the new in-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 149, + 190 + ], + "score": 1.0, + "content": "formation", + "type": "text" + }, + { + "bbox": [ + 149, + 178, + 174, + 187 + ], + "score": 0.88, + "content": "m _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 176, + 306, + 190 + ], + "score": 1.0, + "content": "with the previous memory state", + "type": "text" + }, + { + "bbox": [ + 306, + 178, + 329, + 188 + ], + "score": 0.89, + "content": "m _ { i - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 176, + 506, + 190 + ], + "score": 1.0, + "content": "by combining them through a linear trans-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 504, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 452, + 201 + ], + "score": 1.0, + "content": "formation. In this experiment, we have explored other variations, such as assigning", + "type": "text" + }, + { + "bbox": [ + 452, + 189, + 477, + 199 + ], + "score": 0.88, + "content": "m _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 187, + 490, + 201 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 490, + 189, + 504, + 198 + ], + "score": 0.83, + "content": "m _ { i }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 199, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 436, + 210 + ], + "score": 1.0, + "content": "directly – ignoring previous memories, or doing a linear transformation based on", + "type": "text" + }, + { + "bbox": [ + 436, + 200, + 461, + 209 + ], + "score": 0.88, + "content": "m _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 199, + 506, + 210 + ], + "score": 1.0, + "content": "only. The", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 443, + 221 + ], + "score": 1.0, + "content": "results show that in fact such variant is only slightly worse than our main variant –", + "type": "text" + }, + { + "bbox": [ + 443, + 210, + 465, + 220 + ], + "score": 0.73, + "content": "0 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 208, + 506, + 221 + ], + "score": 1.0, + "content": ". We also", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "score": 1.0, + "content": "conducted an experiment in which we merge the new information with the previous memory just by", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 506, + 243 + ], + "score": 1.0, + "content": "a having a gate that does a weighted average of them. The results show that this variant performs", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 243, + 339, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 339, + 253 + ], + "score": 1.0, + "content": "equivalently to our standard linear-transformation variant.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 165, + 506, + 253 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 258, + 504, + 336 + ], + "lines": [ + { + "bbox": [ + 106, + 259, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 505, + 271 + ], + "score": 1.0, + "content": "Writing Unit Additions. We have explored the impact of the writing unit variants described in sec-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 269, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 506, + 283 + ], + "score": 1.0, + "content": "tion 3.2.3 – adding self-attention, gating mechanisms, or both, compared to our standard main model", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 280, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 406, + 294 + ], + "score": 1.0, + "content": "that uses a linear transformation to merge the newly retrieved information", + "type": "text" + }, + { + "bbox": [ + 406, + 282, + 431, + 292 + ], + "score": 0.89, + "content": "m _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 280, + 506, + 294 + ], + "score": 1.0, + "content": "with the previous", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 290, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 104, + 290, + 142, + 306 + ], + "score": 1.0, + "content": "memory", + "type": "text" + }, + { + "bbox": [ + 142, + 294, + 155, + 303 + ], + "score": 0.86, + "content": "m _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 290, + 506, + 306 + ], + "score": 1.0, + "content": ". For the complete CLEVR dataset we can see that indeed both these variants are very", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 303, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 315 + ], + "score": 1.0, + "content": "helpful in increasing the model performance. Compared to our standard MAC model that achieves", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 313, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 138, + 324 + ], + "score": 0.87, + "content": "9 8 . 9 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 313, + 362, + 326 + ], + "score": 1.0, + "content": "on the validation set, self-attention yields accuracy of", + "type": "text" + }, + { + "bbox": [ + 362, + 314, + 394, + 325 + ], + "score": 0.87, + "content": "9 9 . 2 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 313, + 453, + 326 + ], + "score": 1.0, + "content": ", gating yields", + "type": "text" + }, + { + "bbox": [ + 454, + 314, + 486, + 324 + ], + "score": 0.87, + "content": "9 9 . 3 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 313, + 506, + 326 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 325, + 228, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 192, + 337 + ], + "score": 1.0, + "content": "adding both achieves", + "type": "text" + }, + { + "bbox": [ + 192, + 325, + 225, + 335 + ], + "score": 0.87, + "content": "9 9 . 4 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 325, + 228, + 337 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18, + "bbox_fs": [ + 104, + 259, + 506, + 337 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 342, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "Output Unit. In our standard model, the final predictions made in the output unit are based on the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 183, + 365 + ], + "score": 1.0, + "content": "final memory state", + "type": "text" + }, + { + "bbox": [ + 184, + 354, + 198, + 365 + ], + "score": 0.87, + "content": "m _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 353, + 336, + 365 + ], + "score": 1.0, + "content": "as well as question representation", + "type": "text" + }, + { + "bbox": [ + 337, + 354, + 343, + 364 + ], + "score": 0.77, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "(stands for the final hidden states of the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "backward and forwards passes of the LSTM). We have explored the contribution of basing the model", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 374, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 388 + ], + "score": 1.0, + "content": "prediction on the latter, by testing the model performance when prediction is based on memory", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 225, + 398 + ], + "score": 1.0, + "content": "alone, for the complete and", + "type": "text" + }, + { + "bbox": [ + 225, + 385, + 245, + 396 + ], + "score": 0.87, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "datasets. We can see that in both settings basing the model’s", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 397, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 409 + ], + "score": 1.0, + "content": "predictions on the question representation allows faster training and higher accuracies. Notable is", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 231, + 420 + ], + "score": 1.0, + "content": "the gap in performance for the", + "type": "text" + }, + { + "bbox": [ + 232, + 407, + 251, + 418 + ], + "score": 0.85, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 408, + 294, + 420 + ], + "score": 1.0, + "content": "CLEVR -", + "type": "text" + }, + { + "bbox": [ + 294, + 407, + 321, + 418 + ], + "score": 0.82, + "content": "1 9 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 408, + 505, + 420 + ], + "score": 1.0, + "content": "increase by using the question representation", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 419, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 431 + ], + "score": 1.0, + "content": "to make predictions. These results are very reasonable intuitively, since the model is structured such", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 428, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 443 + ], + "score": 1.0, + "content": "that the memory holds only information that was retrieved from the image. Thus, questions that may", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 439, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 454 + ], + "score": 1.0, + "content": "ask for instance on different aspects (such as color or shape) of the same object in the image may", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 451, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 506, + 464 + ], + "score": 1.0, + "content": "result in the same memory content, which is thus does not directly contain enough information to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 462, + 204, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 204, + 474 + ], + "score": 1.0, + "content": "respond such questions.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 342, + 506, + 474 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 479, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "Position. In our standard model, similarly to the practice of competing models (Santoro et al., 2017;", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "Perez et al., 2017; Hu et al., 2017), we have concatenated positional information to each region of the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 501, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 513 + ], + "score": 1.0, + "content": "image, in order to increase the model capability to perform spatial reasoning. We have explored both", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 237, + 525 + ], + "score": 1.0, + "content": "simple linear maps at a constant", + "type": "text" + }, + { + "bbox": [ + 238, + 512, + 266, + 524 + ], + "score": 0.9, + "content": "[ - 1 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 511, + 505, + 525 + ], + "score": 1.0, + "content": "as well as more complex positional encoding suggested by", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 429, + 534 + ], + "score": 1.0, + "content": "(Vaswani et al., 2017). However, the results for both the standard dataset and the", + "type": "text" + }, + { + "bbox": [ + 430, + 523, + 449, + 533 + ], + "score": 0.87, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 522, + 505, + 534 + ], + "score": 1.0, + "content": "version show", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 533, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 104, + 533, + 506, + 547 + ], + "score": 1.0, + "content": "a very negligible improvement at best when adding positional encoding information, demonstrating", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 546, + 424, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 424, + 557 + ], + "score": 1.0, + "content": "the capability of MAC to perform spatial reasoning without data augmentation.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37, + "bbox_fs": [ + 104, + 478, + 506, + 557 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "Gate Bias Initialization. For our model variant with gating mechanism (described in section 3.2.3)", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 573, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 382, + 585 + ], + "score": 1.0, + "content": "we have tested the effect of setting different values for the gate bias -", + "type": "text" + }, + { + "bbox": [ + 382, + 573, + 406, + 584 + ], + "score": 0.85, + "content": "- 1 , 0", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 573, + 448, + 585 + ], + "score": 1.0, + "content": "and 1. for", + "type": "text" + }, + { + "bbox": [ + 448, + 573, + 462, + 583 + ], + "score": 0.66, + "content": "- 1", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 573, + 505, + 585 + ], + "score": 1.0, + "content": "the model", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 584, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 505, + 596 + ], + "score": 1.0, + "content": "is initialized to biased for keeping the previous memory value whereas for 1 it will be biased for", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 595, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 607 + ], + "score": 1.0, + "content": "using the new memory instead. We can see that for the complete dataset setting the bias to 1 is", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "score": 1.0, + "content": "optimal – apparently since the model has enough data to learn to apply each cell effectively. In", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 195, + 629 + ], + "score": 1.0, + "content": "contrast, for the small", + "type": "text" + }, + { + "bbox": [ + 196, + 617, + 215, + 627 + ], + "score": 0.85, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 616, + 506, + 629 + ], + "score": 1.0, + "content": "CLEVR data, setting the bias to 0 shows better performance, biasing the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "model to using less cells overall which results ultimately in a theoretically-simpler model that can", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 638, + 223, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 223, + 651 + ], + "score": 1.0, + "content": "fit less data more effectively.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 561, + 506, + 651 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 674, + 213, + 685 + ], + "lines": [ + { + "bbox": [ + 106, + 674, + 215, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 215, + 686 + ], + "score": 1.0, + "content": "4.4 INTERPRETABILITY", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 49 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "We have looked into attention maps over the image and question that the model produces during its", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "computation and provide a few examples in figure 4.4. The first example shows us how the model", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "parses the question in steps, first focusing on the main entity that the question is about, then on", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 51, + "bbox_fs": [ + 105, + 698, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "relation of this entity to the “brown matte thing” which is then located in the image. Finally, the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 460, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 460, + 106 + ], + "score": 1.0, + "content": "model correctly focuses on the small brown cube and predicts the right answer – brown.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 108, + 110, + 504, + 143 + ], + "lines": [ + { + "bbox": [ + 106, + 109, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 109, + 506, + 123 + ], + "score": 1.0, + "content": "The second example shows a model with 4 cells instead of 6, that similarly parse the question in", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 506, + 135 + ], + "score": 1.0, + "content": "iterations and focuses on the relevant objects at each step, though we can see that the reasoning", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 405, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 405, + 144 + ], + "score": 1.0, + "content": "process looks somewhat different when the MAC network has fewer cells.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 149, + 505, + 182 + ], + "lines": [ + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "score": 1.0, + "content": "The last example shows how how the model handles counting and OR operations. It starts from", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 158, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 506, + 175 + ], + "score": 1.0, + "content": "identifying the task - computing a number, and then red objects as well as the cylinder, one at a time,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 171, + 354, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 354, + 183 + ], + "score": 1.0, + "content": "allowing it ultimately to respond correctly, with the answer 2.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 200, + 217, + 413, + 583 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 200, + 217, + 413, + 583 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 200, + 217, + 413, + 583 + ], + "spans": [ + { + "bbox": [ + 200, + 217, + 413, + 583 + ], + "score": 0.971, + "type": "image", + "image_path": "e280158a489b95ddeabb4461dda86460910408860df91a493fe23a53648a08b4.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 200, + 217, + 413, + 230.55555555555554 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 200, + 230.55555555555554, + 413, + 244.1111111111111 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 200, + 244.1111111111111, + 413, + 257.66666666666663 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 200, + 257.66666666666663, + 413, + 271.2222222222222 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 200, + 271.2222222222222, + 413, + 284.7777777777777 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 200, + 284.7777777777777, + 413, + 298.33333333333326 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 200, + 298.33333333333326, + 413, + 311.8888888888888 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 200, + 311.8888888888888, + 413, + 325.44444444444434 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 200, + 325.44444444444434, + 413, + 338.9999999999999 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 200, + 338.9999999999999, + 413, + 352.55555555555543 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 200, + 352.55555555555543, + 413, + 366.111111111111 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 200, + 366.111111111111, + 413, + 379.6666666666665 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 200, + 379.6666666666665, + 413, + 393.22222222222206 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 200, + 393.22222222222206, + 413, + 406.7777777777776 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 200, + 406.7777777777776, + 413, + 420.33333333333314 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 200, + 420.33333333333314, + 413, + 433.8888888888887 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 200, + 433.8888888888887, + 413, + 447.44444444444423 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 200, + 447.44444444444423, + 413, + 460.9999999999998 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 200, + 460.9999999999998, + 413, + 474.5555555555553 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 200, + 474.5555555555553, + 413, + 488.11111111111086 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 200, + 488.11111111111086, + 413, + 501.6666666666664 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 200, + 501.6666666666664, + 413, + 515.222222222222 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 200, + 515.222222222222, + 413, + 528.7777777777775 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 200, + 528.7777777777775, + 413, + 542.333333333333 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 200, + 542.333333333333, + 413, + 555.8888888888886 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 200, + 555.8888888888886, + 413, + 569.4444444444441 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 200, + 569.4444444444441, + 413, + 582.9999999999997 + ], + "spans": [], + "index": 34 + } + ] + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 105, + 617, + 504, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 616, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 631 + ], + "score": 1.0, + "content": "Figure 8: Three examples of attention maps for the image and the question. See text for full details.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "title", + "bbox": [ + 108, + 652, + 195, + 665 + ], + "lines": [ + { + "bbox": [ + 104, + 651, + 197, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 651, + 197, + 668 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "We have given a first demonstration of how a sequence of Memory, Attention and Control (MAC)", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "cells combined into a Compositional Attention Network provides a very effective tool for neural rea-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "soning. In future work, we wish to explore this promising architecture for other tasks and domains,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 710, + 445, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 445, + 724 + ], + "score": 1.0, + "content": "including real-world VQA, machine comprehension and textual question answering.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "relation of this entity to the “brown matte thing” which is then located in the image. Finally, the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 460, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 460, + 106 + ], + "score": 1.0, + "content": "model correctly focuses on the small brown cube and predicts the right answer – brown.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 110, + 504, + 143 + ], + "lines": [ + { + "bbox": [ + 106, + 109, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 109, + 506, + 123 + ], + "score": 1.0, + "content": "The second example shows a model with 4 cells instead of 6, that similarly parse the question in", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 506, + 135 + ], + "score": 1.0, + "content": "iterations and focuses on the relevant objects at each step, though we can see that the reasoning", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 405, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 405, + 144 + ], + "score": 1.0, + "content": "process looks somewhat different when the MAC network has fewer cells.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 109, + 506, + 144 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 149, + 505, + 182 + ], + "lines": [ + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "score": 1.0, + "content": "The last example shows how how the model handles counting and OR operations. It starts from", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 158, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 506, + 175 + ], + "score": 1.0, + "content": "identifying the task - computing a number, and then red objects as well as the cylinder, one at a time,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 171, + 354, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 354, + 183 + ], + "score": 1.0, + "content": "allowing it ultimately to respond correctly, with the answer 2.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 148, + 506, + 183 + ] + }, + { + "type": "image", + "bbox": [ + 200, + 217, + 413, + 583 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 200, + 217, + 413, + 583 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 200, + 217, + 413, + 583 + ], + "spans": [ + { + "bbox": [ + 200, + 217, + 413, + 583 + ], + "score": 0.971, + "type": "image", + "image_path": "e280158a489b95ddeabb4461dda86460910408860df91a493fe23a53648a08b4.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 200, + 217, + 413, + 230.55555555555554 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 200, + 230.55555555555554, + 413, + 244.1111111111111 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 200, + 244.1111111111111, + 413, + 257.66666666666663 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 200, + 257.66666666666663, + 413, + 271.2222222222222 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 200, + 271.2222222222222, + 413, + 284.7777777777777 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 200, + 284.7777777777777, + 413, + 298.33333333333326 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 200, + 298.33333333333326, + 413, + 311.8888888888888 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 200, + 311.8888888888888, + 413, + 325.44444444444434 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 200, + 325.44444444444434, + 413, + 338.9999999999999 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 200, + 338.9999999999999, + 413, + 352.55555555555543 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 200, + 352.55555555555543, + 413, + 366.111111111111 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 200, + 366.111111111111, + 413, + 379.6666666666665 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 200, + 379.6666666666665, + 413, + 393.22222222222206 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 200, + 393.22222222222206, + 413, + 406.7777777777776 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 200, + 406.7777777777776, + 413, + 420.33333333333314 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 200, + 420.33333333333314, + 413, + 433.8888888888887 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 200, + 433.8888888888887, + 413, + 447.44444444444423 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 200, + 447.44444444444423, + 413, + 460.9999999999998 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 200, + 460.9999999999998, + 413, + 474.5555555555553 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 200, + 474.5555555555553, + 413, + 488.11111111111086 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 200, + 488.11111111111086, + 413, + 501.6666666666664 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 200, + 501.6666666666664, + 413, + 515.222222222222 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 200, + 515.222222222222, + 413, + 528.7777777777775 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 200, + 528.7777777777775, + 413, + 542.333333333333 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 200, + 542.333333333333, + 413, + 555.8888888888886 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 200, + 555.8888888888886, + 413, + 569.4444444444441 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 200, + 569.4444444444441, + 413, + 582.9999999999997 + ], + "spans": [], + "index": 34 + } + ] + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 105, + 617, + 504, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 616, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 631 + ], + "score": 1.0, + "content": "Figure 8: Three examples of attention maps for the image and the question. See text for full details.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 616, + 505, + 631 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 652, + 195, + 665 + ], + "lines": [ + { + "bbox": [ + 104, + 651, + 197, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 651, + 197, + 668 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "We have given a first demonstration of how a sequence of Memory, Attention and Control (MAC)", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "cells combined into a Compositional Attention Network provides a very effective tool for neural rea-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "soning. In future work, we wish to explore this promising architecture for other tasks and domains,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 710, + 445, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 445, + 724 + ], + "score": 1.0, + "content": "including real-world VQA, machine comprehension and textual question answering.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 677, + 505, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 176, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 504, + 123 + ], + "lines": [ + { + "bbox": [ + 106, + 100, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 505, + 113 + ], + "score": 1.0, + "content": "Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 111, + 356, + 122 + ], + "spans": [ + { + "bbox": [ + 116, + 111, + 356, + 122 + ], + "score": 1.0, + "content": "answering models. arXiv preprint arXiv:1606.07356, 2016.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 106, + 128, + 504, + 163 + ], + "lines": [ + { + "bbox": [ + 106, + 129, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 505, + 141 + ], + "score": 1.0, + "content": "Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 139, + 506, + 154 + ], + "spans": [ + { + "bbox": [ + 115, + 139, + 506, + 154 + ], + "score": 1.0, + "content": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 39–48,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 149, + 147, + 164 + ], + "spans": [ + { + "bbox": [ + 115, + 149, + 147, + 164 + ], + "score": 1.0, + "content": "2016a.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 169, + 504, + 193 + ], + "lines": [ + { + "bbox": [ + 105, + 168, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 505, + 183 + ], + "score": 1.0, + "content": "Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Learning to compose neural", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 181, + 420, + 193 + ], + "spans": [ + { + "bbox": [ + 115, + 181, + 420, + 193 + ], + "score": 1.0, + "content": "networks for question answering. arXiv preprint arXiv:1601.01705, 2016b.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 199, + 504, + 233 + ], + "lines": [ + { + "bbox": [ + 106, + 199, + 504, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 504, + 212 + ], + "score": 1.0, + "content": "Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zit-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 210, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 116, + 210, + 505, + 223 + ], + "score": 1.0, + "content": "nick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 117, + 221, + 339, + 234 + ], + "spans": [ + { + "bbox": [ + 117, + 221, + 339, + 234 + ], + "score": 1.0, + "content": "Conference on Computer Vision, pp. 2425–2433, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 504, + 262 + ], + "lines": [ + { + "bbox": [ + 105, + 238, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 253 + ], + "score": 1.0, + "content": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 250, + 219, + 262 + ], + "spans": [ + { + "bbox": [ + 115, + 250, + 219, + 262 + ], + "score": 1.0, + "content": "arXiv:1607.06450, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 268, + 504, + 291 + ], + "lines": [ + { + "bbox": [ + 106, + 269, + 504, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 504, + 282 + ], + "score": 1.0, + "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 280, + 398, + 292 + ], + "spans": [ + { + "bbox": [ + 115, + 280, + 398, + 292 + ], + "score": 1.0, + "content": "learning to align and translate. arXiv preprint arXiv:1409.0473, 2014.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 105, + 298, + 504, + 321 + ], + "lines": [ + { + "bbox": [ + 106, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "Djork-Arne Clevert, Thomas Unterthiner, and Sepp Hochreiter. Fast and accurate deep network ´", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 308, + 451, + 322 + ], + "spans": [ + { + "bbox": [ + 115, + 308, + 451, + 322 + ], + "score": 1.0, + "content": "learning by exponential linear units (elus). arXiv preprint arXiv:1511.07289, 2015.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 327, + 504, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 505, + 341 + ], + "score": 1.0, + "content": "Alex Graves, Navdeep Jaitly, and Abdel-rahman Mohamed. Hybrid speech recognition with deep", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 339, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 116, + 339, + 505, + 351 + ], + "score": 1.0, + "content": "bidirectional lstm. In Automatic Speech Recognition and Understanding (ASRU), 2013 IEEE", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 117, + 349, + 281, + 363 + ], + "spans": [ + { + "bbox": [ + 117, + 349, + 281, + 363 + ], + "score": 1.0, + "content": "Workshop on, pp. 273–278. IEEE, 2013.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 105, + 367, + 505, + 391 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 506, + 381 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 379, + 214, + 390 + ], + "spans": [ + { + "bbox": [ + 115, + 379, + 214, + 390 + ], + "score": 1.0, + "content": "arXiv:1410.5401, 2014.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 505, + 442 + ], + "lines": [ + { + "bbox": [ + 106, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka Grabska-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 115, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "Barwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, ´", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 419, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 115, + 419, + 506, + 432 + ], + "score": 1.0, + "content": "et al. Hybrid computing using a neural network with dynamic external memory. Nature, 538", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 431, + 210, + 443 + ], + "spans": [ + { + "bbox": [ + 116, + 431, + 210, + 443 + ], + "score": 1.0, + "content": "(7626):471–476, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 104, + 448, + 504, + 471 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "score": 1.0, + "content": "Akshay Kumar Gupta. Survey of visual question answering: Datasets and techniques. arXiv preprint", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 460, + 220, + 470 + ], + "spans": [ + { + "bbox": [ + 115, + 460, + 220, + 470 + ], + "score": 1.0, + "content": "arXiv:1705.03865, 2017.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 478, + 503, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 493 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 488, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 115, + 488, + 505, + 504 + ], + "score": 1.0, + "content": "nition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 500, + 181, + 511 + ], + "spans": [ + { + "bbox": [ + 116, + 500, + 181, + 511 + ], + "score": 1.0, + "content": "770–778, 2016.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 103, + 518, + 504, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 529, + 192, + 541 + ], + "spans": [ + { + "bbox": [ + 116, + 529, + 192, + 541 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 547, + 505, + 581 + ], + "lines": [ + { + "bbox": [ + 106, + 547, + 504, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 504, + 560 + ], + "score": 1.0, + "content": "Ronghang Hu, Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Kate Saenko. Learn-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 115, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "ing to reason: End-to-end module networks for visual question answering. arXiv preprint", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 569, + 220, + 581 + ], + "spans": [ + { + "bbox": [ + 115, + 569, + 220, + 581 + ], + "score": 1.0, + "content": "arXiv:1704.05526, 2017.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 588, + 504, + 622 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 602 + ], + "score": 1.0, + "content": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 115, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 114, + 609, + 142, + 622 + ], + "spans": [ + { + "bbox": [ + 114, + 609, + 142, + 622 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 628, + 504, + 663 + ], + "lines": [ + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C Lawrence Zitnick, and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 505, + 653 + ], + "score": 1.0, + "content": "Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 651, + 324, + 663 + ], + "spans": [ + { + "bbox": [ + 115, + 651, + 324, + 663 + ], + "score": 1.0, + "content": "reasoning. arXiv preprint arXiv:1612.06890, 2016.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 105, + 669, + 503, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 669, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 505, + 681 + ], + "score": 1.0, + "content": "Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Judy Hoffman, Li Fei-Fei, C Lawrence", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 679, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 115, + 679, + 505, + 693 + ], + "score": 1.0, + "content": "Zitnick, and Ross Girshick. Inferring and executing programs for visual reasoning. arXiv preprint", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 691, + 220, + 702 + ], + "spans": [ + { + "bbox": [ + 116, + 691, + 220, + 702 + ], + "score": 1.0, + "content": "arXiv:1705.03633, 2017.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 105, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 721, + 214, + 731 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 214, + 731 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 176, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 504, + 123 + ], + "lines": [ + { + "bbox": [ + 106, + 100, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 505, + 113 + ], + "score": 1.0, + "content": "Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 111, + 356, + 122 + ], + "spans": [ + { + "bbox": [ + 116, + 111, + 356, + 122 + ], + "score": 1.0, + "content": "answering models. arXiv preprint arXiv:1606.07356, 2016.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 106, + 100, + 505, + 122 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 128, + 504, + 163 + ], + "lines": [ + { + "bbox": [ + 106, + 129, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 505, + 141 + ], + "score": 1.0, + "content": "Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 139, + 506, + 154 + ], + "spans": [ + { + "bbox": [ + 115, + 139, + 506, + 154 + ], + "score": 1.0, + "content": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 39–48,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 149, + 147, + 164 + ], + "spans": [ + { + "bbox": [ + 115, + 149, + 147, + 164 + ], + "score": 1.0, + "content": "2016a.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 129, + 506, + 164 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 169, + 504, + 193 + ], + "lines": [ + { + "bbox": [ + 105, + 168, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 505, + 183 + ], + "score": 1.0, + "content": "Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Learning to compose neural", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 181, + 420, + 193 + ], + "spans": [ + { + "bbox": [ + 115, + 181, + 420, + 193 + ], + "score": 1.0, + "content": "networks for question answering. arXiv preprint arXiv:1601.01705, 2016b.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 168, + 505, + 193 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 199, + 504, + 233 + ], + "lines": [ + { + "bbox": [ + 106, + 199, + 504, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 504, + 212 + ], + "score": 1.0, + "content": "Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zit-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 210, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 116, + 210, + 505, + 223 + ], + "score": 1.0, + "content": "nick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 117, + 221, + 339, + 234 + ], + "spans": [ + { + "bbox": [ + 117, + 221, + 339, + 234 + ], + "score": 1.0, + "content": "Conference on Computer Vision, pp. 2425–2433, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 199, + 505, + 234 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 504, + 262 + ], + "lines": [ + { + "bbox": [ + 105, + 238, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 253 + ], + "score": 1.0, + "content": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 250, + 219, + 262 + ], + "spans": [ + { + "bbox": [ + 115, + 250, + 219, + 262 + ], + "score": 1.0, + "content": "arXiv:1607.06450, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 238, + 505, + 262 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 268, + 504, + 291 + ], + "lines": [ + { + "bbox": [ + 106, + 269, + 504, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 504, + 282 + ], + "score": 1.0, + "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 280, + 398, + 292 + ], + "spans": [ + { + "bbox": [ + 115, + 280, + 398, + 292 + ], + "score": 1.0, + "content": "learning to align and translate. arXiv preprint arXiv:1409.0473, 2014.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 269, + 504, + 292 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 298, + 504, + 321 + ], + "lines": [ + { + "bbox": [ + 106, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "Djork-Arne Clevert, Thomas Unterthiner, and Sepp Hochreiter. Fast and accurate deep network ´", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 308, + 451, + 322 + ], + "spans": [ + { + "bbox": [ + 115, + 308, + 451, + 322 + ], + "score": 1.0, + "content": "learning by exponential linear units (elus). arXiv preprint arXiv:1511.07289, 2015.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 106, + 298, + 505, + 322 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 327, + 504, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 505, + 341 + ], + "score": 1.0, + "content": "Alex Graves, Navdeep Jaitly, and Abdel-rahman Mohamed. Hybrid speech recognition with deep", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 339, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 116, + 339, + 505, + 351 + ], + "score": 1.0, + "content": "bidirectional lstm. In Automatic Speech Recognition and Understanding (ASRU), 2013 IEEE", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 117, + 349, + 281, + 363 + ], + "spans": [ + { + "bbox": [ + 117, + 349, + 281, + 363 + ], + "score": 1.0, + "content": "Workshop on, pp. 273–278. IEEE, 2013.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 326, + 505, + 363 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 367, + 505, + 391 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 506, + 381 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 379, + 214, + 390 + ], + "spans": [ + { + "bbox": [ + 115, + 379, + 214, + 390 + ], + "score": 1.0, + "content": "arXiv:1410.5401, 2014.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 366, + 506, + 390 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 505, + 442 + ], + "lines": [ + { + "bbox": [ + 106, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka Grabska-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 115, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "Barwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, ´", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 419, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 115, + 419, + 506, + 432 + ], + "score": 1.0, + "content": "et al. Hybrid computing using a neural network with dynamic external memory. Nature, 538", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 431, + 210, + 443 + ], + "spans": [ + { + "bbox": [ + 116, + 431, + 210, + 443 + ], + "score": 1.0, + "content": "(7626):471–476, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 106, + 397, + 506, + 443 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 448, + 504, + 471 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "score": 1.0, + "content": "Akshay Kumar Gupta. Survey of visual question answering: Datasets and techniques. arXiv preprint", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 460, + 220, + 470 + ], + "spans": [ + { + "bbox": [ + 115, + 460, + 220, + 470 + ], + "score": 1.0, + "content": "arXiv:1705.03865, 2017.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 448, + 505, + 470 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 478, + 503, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 493 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 488, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 115, + 488, + 505, + 504 + ], + "score": 1.0, + "content": "nition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 500, + 181, + 511 + ], + "spans": [ + { + "bbox": [ + 116, + 500, + 181, + 511 + ], + "score": 1.0, + "content": "770–778, 2016.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 476, + 505, + 511 + ] + }, + { + "type": "text", + "bbox": [ + 103, + 518, + 504, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 529, + 192, + 541 + ], + "spans": [ + { + "bbox": [ + 116, + 529, + 192, + 541 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 518, + 505, + 541 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 547, + 505, + 581 + ], + "lines": [ + { + "bbox": [ + 106, + 547, + 504, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 504, + 560 + ], + "score": 1.0, + "content": "Ronghang Hu, Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Kate Saenko. Learn-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 115, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "ing to reason: End-to-end module networks for visual question answering. arXiv preprint", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 569, + 220, + 581 + ], + "spans": [ + { + "bbox": [ + 115, + 569, + 220, + 581 + ], + "score": 1.0, + "content": "arXiv:1704.05526, 2017.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 106, + 547, + 505, + 581 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 588, + 504, + 622 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 602 + ], + "score": 1.0, + "content": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 115, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 114, + 609, + 142, + 622 + ], + "spans": [ + { + "bbox": [ + 114, + 609, + 142, + 622 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 587, + 505, + 622 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 628, + 504, + 663 + ], + "lines": [ + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C Lawrence Zitnick, and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 505, + 653 + ], + "score": 1.0, + "content": "Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 651, + 324, + 663 + ], + "spans": [ + { + "bbox": [ + 115, + 651, + 324, + 663 + ], + "score": 1.0, + "content": "reasoning. arXiv preprint arXiv:1612.06890, 2016.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 106, + 628, + 505, + 663 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 669, + 503, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 669, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 505, + 681 + ], + "score": 1.0, + "content": "Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Judy Hoffman, Li Fei-Fei, C Lawrence", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 679, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 115, + 679, + 505, + 693 + ], + "score": 1.0, + "content": "Zitnick, and Ross Girshick. Inferring and executing programs for visual reasoning. arXiv preprint", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 691, + 220, + 702 + ], + "spans": [ + { + "bbox": [ + 116, + 691, + 220, + 702 + ], + "score": 1.0, + "content": "arXiv:1705.03633, 2017.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 106, + 669, + 505, + 702 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 721, + 214, + 731 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 214, + 731 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 708, + 505, + 731 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "lutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 103, + 143, + 116 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 143, + 116 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 505, + 167 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 506, + 136 + ], + "score": 1.0, + "content": "Ankit Kumar, Ozan Irsoy, Peter Ondruska, Mohit Iyyer, James Bradbury, Ishaan Gulrajani, Victor", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 506, + 147 + ], + "score": 1.0, + "content": "Zhong, Romain Paulus, and Richard Socher. Ask me anything: Dynamic memory networks for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 145, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 114, + 145, + 506, + 158 + ], + "score": 1.0, + "content": "natural language processing. In International Conference on Machine Learning, pp. 1378–1387,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 155, + 143, + 169 + ], + "spans": [ + { + "bbox": [ + 114, + 155, + 143, + 169 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 174, + 503, + 198 + ], + "lines": [ + { + "bbox": [ + 106, + 174, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 505, + 187 + ], + "score": 1.0, + "content": "Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 186, + 417, + 198 + ], + "spans": [ + { + "bbox": [ + 116, + 186, + 417, + 198 + ], + "score": 1.0, + "content": "document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 506, + 238 + ], + "lines": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 216, + 506, + 229 + ], + "spans": [ + { + "bbox": [ + 115, + 216, + 506, + 229 + ], + "score": 1.0, + "content": "for visual question answering. In Advances In Neural Information Processing Systems, pp. 289–", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 226, + 162, + 239 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 162, + 239 + ], + "score": 1.0, + "content": "297, 2016.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 245, + 506, + 279 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "Alexander Miller, Adam Fisch, Jesse Dodge, Amir-Hossein Karimi, Antoine Bordes, and Ja-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 256, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 115, + 256, + 505, + 269 + ], + "score": 1.0, + "content": "son Weston. Key-value memory networks for directly reading documents. arXiv preprint", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 268, + 219, + 279 + ], + "spans": [ + { + "bbox": [ + 116, + 268, + 219, + 279 + ], + "score": 1.0, + "content": "arXiv:1606.03126, 2016.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 286, + 506, + 320 + ], + "lines": [ + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "score": 1.0, + "content": "Jeffrey Pennington, Richard Socher, and Christopher Manning. Glove: Global vectors for word", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 115, + 297, + 506, + 311 + ], + "score": 1.0, + "content": "representation. In Proceedings of the 2014 conference on empirical methods in natural language", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 309, + 296, + 321 + ], + "spans": [ + { + "bbox": [ + 115, + 309, + 296, + 321 + ], + "score": 1.0, + "content": "processing (EMNLP), pp. 1532–1543, 2014.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 327, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 106, + 327, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 505, + 340 + ], + "score": 1.0, + "content": "Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 339, + 457, + 351 + ], + "spans": [ + { + "bbox": [ + 115, + 339, + 457, + 351 + ], + "score": 1.0, + "content": "reasoning with a general conditioning layer. arXiv preprint arXiv:1709.07871, 2017.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 504, + 391 + ], + "lines": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "Adam Santoro, David Raposo, David GT Barrett, Mateusz Malinowski, Razvan Pascanu, Peter", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 115, + 368, + 505, + 381 + ], + "score": 1.0, + "content": "Battaglia, and Timothy Lillicrap. A simple neural network module for relational reasoning. arXiv", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 380, + 254, + 391 + ], + "spans": [ + { + "bbox": [ + 115, + 380, + 254, + 391 + ], + "score": 1.0, + "content": "preprint arXiv:1706.01427, 2017.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 104, + 397, + 505, + 421 + ], + "lines": [ + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "score": 1.0, + "content": "Rupesh Kumar Srivastava, Klaus Greff, and Jurgen Schmidhuber. Highway networks. ¨ arXiv preprint", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 409, + 219, + 420 + ], + "spans": [ + { + "bbox": [ + 115, + 409, + 219, + 420 + ], + "score": 1.0, + "content": "arXiv:1505.00387, 2015.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 104, + 428, + 504, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 426, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 441 + ], + "score": 1.0, + "content": "Bob L Sturm. A simple method to determine if a music information retrieval system is a horse.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 117, + 439, + 356, + 450 + ], + "spans": [ + { + "bbox": [ + 117, + 439, + 356, + 450 + ], + "score": 1.0, + "content": "IEEE Transactions on Multimedia, 16(6):1636–1644, 2014.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 457, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 469, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 115, + 469, + 505, + 482 + ], + "score": 1.0, + "content": "Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 479, + 144, + 492 + ], + "spans": [ + { + "bbox": [ + 115, + 479, + 144, + 492 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 498, + 504, + 532 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 506, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 506, + 512 + ], + "score": 1.0, + "content": "Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 510, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 505, + 522 + ], + "score": 1.0, + "content": "textual question answering. In International Conference on Machine Learning, pp. 2397–2406,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 519, + 143, + 533 + ], + "spans": [ + { + "bbox": [ + 114, + 519, + 143, + 533 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 539, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 551, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 116, + 551, + 505, + 563 + ], + "score": 1.0, + "content": "for image question answering. In Proceedings of the IEEE Conference on Computer Vision and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 561, + 271, + 574 + ], + "spans": [ + { + "bbox": [ + 115, + 561, + 271, + 574 + ], + "score": 1.0, + "content": "Pattern Recognition, pp. 21–29, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "lutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 103, + 143, + 116 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 143, + 116 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 505, + 116 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 505, + 167 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 506, + 136 + ], + "score": 1.0, + "content": "Ankit Kumar, Ozan Irsoy, Peter Ondruska, Mohit Iyyer, James Bradbury, Ishaan Gulrajani, Victor", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 506, + 147 + ], + "score": 1.0, + "content": "Zhong, Romain Paulus, and Richard Socher. Ask me anything: Dynamic memory networks for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 145, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 114, + 145, + 506, + 158 + ], + "score": 1.0, + "content": "natural language processing. In International Conference on Machine Learning, pp. 1378–1387,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 155, + 143, + 169 + ], + "spans": [ + { + "bbox": [ + 114, + 155, + 143, + 169 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 123, + 506, + 169 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 174, + 503, + 198 + ], + "lines": [ + { + "bbox": [ + 106, + 174, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 505, + 187 + ], + "score": 1.0, + "content": "Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 186, + 417, + 198 + ], + "spans": [ + { + "bbox": [ + 116, + 186, + 417, + 198 + ], + "score": 1.0, + "content": "document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 174, + 505, + 198 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 506, + 238 + ], + "lines": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 216, + 506, + 229 + ], + "spans": [ + { + "bbox": [ + 115, + 216, + 506, + 229 + ], + "score": 1.0, + "content": "for visual question answering. In Advances In Neural Information Processing Systems, pp. 289–", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 226, + 162, + 239 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 162, + 239 + ], + "score": 1.0, + "content": "297, 2016.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 204, + 506, + 239 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 245, + 506, + 279 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "Alexander Miller, Adam Fisch, Jesse Dodge, Amir-Hossein Karimi, Antoine Bordes, and Ja-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 256, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 115, + 256, + 505, + 269 + ], + "score": 1.0, + "content": "son Weston. Key-value memory networks for directly reading documents. arXiv preprint", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 268, + 219, + 279 + ], + "spans": [ + { + "bbox": [ + 116, + 268, + 219, + 279 + ], + "score": 1.0, + "content": "arXiv:1606.03126, 2016.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 245, + 505, + 279 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 286, + 506, + 320 + ], + "lines": [ + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "score": 1.0, + "content": "Jeffrey Pennington, Richard Socher, and Christopher Manning. Glove: Global vectors for word", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 115, + 297, + 506, + 311 + ], + "score": 1.0, + "content": "representation. In Proceedings of the 2014 conference on empirical methods in natural language", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 309, + 296, + 321 + ], + "spans": [ + { + "bbox": [ + 115, + 309, + 296, + 321 + ], + "score": 1.0, + "content": "processing (EMNLP), pp. 1532–1543, 2014.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 286, + 506, + 321 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 327, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 106, + 327, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 505, + 340 + ], + "score": 1.0, + "content": "Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 339, + 457, + 351 + ], + "spans": [ + { + "bbox": [ + 115, + 339, + 457, + 351 + ], + "score": 1.0, + "content": "reasoning with a general conditioning layer. arXiv preprint arXiv:1709.07871, 2017.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 327, + 505, + 351 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 504, + 391 + ], + "lines": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "Adam Santoro, David Raposo, David GT Barrett, Mateusz Malinowski, Razvan Pascanu, Peter", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 115, + 368, + 505, + 381 + ], + "score": 1.0, + "content": "Battaglia, and Timothy Lillicrap. A simple neural network module for relational reasoning. arXiv", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 380, + 254, + 391 + ], + "spans": [ + { + "bbox": [ + 115, + 380, + 254, + 391 + ], + "score": 1.0, + "content": "preprint arXiv:1706.01427, 2017.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 357, + 505, + 391 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 397, + 505, + 421 + ], + "lines": [ + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "score": 1.0, + "content": "Rupesh Kumar Srivastava, Klaus Greff, and Jurgen Schmidhuber. Highway networks. ¨ arXiv preprint", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 409, + 219, + 420 + ], + "spans": [ + { + "bbox": [ + 115, + 409, + 219, + 420 + ], + "score": 1.0, + "content": "arXiv:1505.00387, 2015.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 397, + 505, + 420 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 428, + 504, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 426, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 441 + ], + "score": 1.0, + "content": "Bob L Sturm. A simple method to determine if a music information retrieval system is a horse.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 117, + 439, + 356, + 450 + ], + "spans": [ + { + "bbox": [ + 117, + 439, + 356, + 450 + ], + "score": 1.0, + "content": "IEEE Transactions on Multimedia, 16(6):1636–1644, 2014.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 426, + 505, + 450 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 457, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 469, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 115, + 469, + 505, + 482 + ], + "score": 1.0, + "content": "Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 479, + 144, + 492 + ], + "spans": [ + { + "bbox": [ + 115, + 479, + 144, + 492 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 456, + 506, + 492 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 498, + 504, + 532 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 506, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 506, + 512 + ], + "score": 1.0, + "content": "Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 510, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 505, + 522 + ], + "score": 1.0, + "content": "textual question answering. In International Conference on Machine Learning, pp. 2397–2406,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 519, + 143, + 533 + ], + "spans": [ + { + "bbox": [ + 114, + 519, + 143, + 533 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 106, + 498, + 506, + 533 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 539, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 551, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 116, + 551, + 505, + 563 + ], + "score": 1.0, + "content": "for image question answering. In Proceedings of the IEEE Conference on Computer Vision and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 561, + 271, + 574 + ], + "spans": [ + { + "bbox": [ + 115, + 561, + 271, + 574 + ], + "score": 1.0, + "content": "Pattern Recognition, pp. 21–29, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 106, + 540, + 505, + 574 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 81, + 255, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 257, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 257, + 95 + ], + "score": 1.0, + "content": "SUPPLEMENTARY MATERIAL", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 110, + 255, + 124 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 257, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 257, + 126 + ], + "score": 1.0, + "content": "A DETAILS OF INPUT UNIT", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 138, + 505, + 183 + ], + "lines": [ + { + "bbox": [ + 105, + 138, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 151 + ], + "score": 1.0, + "content": "The input unit processes the raw inputs given to the system into distributed vector representations.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 149, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 104, + 149, + 506, + 163 + ], + "score": 1.0, + "content": "It receives a text question (or in general, a query), and an image (or in general, a Knowledge Base", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 160, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 505, + 174 + ], + "score": 1.0, + "content": "(KB)) and processes each of them with a matching sub-unit. Here we provide details of the Query", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 172, + 279, + 184 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 279, + 184 + ], + "score": 1.0, + "content": "Unit and the Image Unit used in this work.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 108, + 200, + 219, + 212 + ], + "lines": [ + { + "bbox": [ + 106, + 199, + 220, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 220, + 214 + ], + "score": 1.0, + "content": "A.0.1 THE QUERY UNIT", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 221, + 505, + 266 + ], + "lines": [ + { + "bbox": [ + 106, + 221, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 193, + 234 + ], + "score": 1.0, + "content": "We encode a query of", + "type": "text" + }, + { + "bbox": [ + 193, + 222, + 201, + 232 + ], + "score": 0.83, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 221, + 505, + 234 + ], + "score": 1.0, + "content": "words into a continuous representation using a bidirectional LSTM (Hochre-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 232, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 505, + 246 + ], + "score": 1.0, + "content": "iter & Schmidhuber, 1997; Graves et al., 2013). Each word is associated with a word embedding", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 107, + 245, + 119, + 255 + ], + "score": 0.85, + "content": "w _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 244, + 151, + 257 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 151, + 244, + 201, + 255 + ], + "score": 0.92, + "content": "s = 1 , . . . , S", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 244, + 505, + 257 + ], + "score": 1.0, + "content": ". In our case, we use GloVE words embeddings (Pennington et al., 2014).", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 255, + 482, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 423, + 267 + ], + "score": 1.0, + "content": "Then, these embeddings are processed by a bidirectional LST M of dimension", + "type": "text" + }, + { + "bbox": [ + 423, + 255, + 430, + 264 + ], + "score": 0.77, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 255, + 482, + 267 + ], + "score": 1.0, + "content": "that outputs:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 133, + 279, + 505, + 333 + ], + "lines": [ + { + "bbox": [ + 131, + 279, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 131, + 279, + 237, + 293 + ], + "score": 1.0, + "content": "• a matching sequence of", + "type": "text" + }, + { + "bbox": [ + 238, + 281, + 244, + 290 + ], + "score": 0.83, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 279, + 506, + 293 + ], + "score": 1.0, + "content": "-dimensional output states, which we refer to as contextual words,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 291, + 200, + 303 + ], + "spans": [ + { + "bbox": [ + 143, + 291, + 200, + 303 + ], + "score": 0.79, + "content": "[ c w _ { 1 } , . . . , c w _ { S } ]", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 309, + 142, + 322 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 142, + 311, + 149, + 320 + ], + "score": 0.78, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 309, + 252, + 322 + ], + "score": 1.0, + "content": "-dimensional hidden state", + "type": "text" + }, + { + "bbox": [ + 252, + 309, + 314, + 322 + ], + "score": 0.93, + "content": "q = [ \\overleftrightarrow { \\lambda \\omega _ { 1 } } , \\overrightarrow { c w _ { S } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 309, + 505, + 322 + ], + "score": 1.0, + "content": ", the concatenation of the hidden states from the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 321, + 444, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 321, + 312, + 334 + ], + "score": 1.0, + "content": "backward and forward passes. We refer to", + "type": "text" + }, + { + "bbox": [ + 312, + 324, + 318, + 333 + ], + "score": 0.8, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 321, + 444, + 334 + ], + "score": 1.0, + "content": "as the question representation.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 346, + 504, + 370 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 245, + 359 + ], + "score": 1.0, + "content": "Intuitively, each contextual word", + "type": "text" + }, + { + "bbox": [ + 245, + 349, + 262, + 358 + ], + "score": 0.87, + "content": "c w _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 345, + 374, + 359 + ], + "score": 1.0, + "content": "represents the meaning of", + "type": "text" + }, + { + "bbox": [ + 375, + 346, + 389, + 357 + ], + "score": 0.89, + "content": "s ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 345, + 506, + 359 + ], + "score": 1.0, + "content": "word, in the context of the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 357, + 496, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 235, + 371 + ], + "score": 1.0, + "content": "question, while the hidden state", + "type": "text" + }, + { + "bbox": [ + 235, + 360, + 241, + 370 + ], + "score": 0.8, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 357, + 496, + 371 + ], + "score": 1.0, + "content": "represents the overall (compositional) meaning of the question.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 108, + 386, + 217, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 218, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 218, + 399 + ], + "score": 1.0, + "content": "A.0.2 THE IMAGE UNIT", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 505, + 463 + ], + "lines": [ + { + "bbox": [ + 106, + 408, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 505, + 420 + ], + "score": 1.0, + "content": "Given an image, and following prior work on CLEVR (?Santoro et al., 2017; Perez et al., 2017), we", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 419, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 430 + ], + "score": 1.0, + "content": "extract conv4 features from ResNet101 (He et al., 2016) pretrained on ImageNet (Krizhevsky et al.,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 377, + 442 + ], + "score": 1.0, + "content": "2012) which we treat as a fixed initial representation of the image,", + "type": "text" + }, + { + "bbox": [ + 377, + 432, + 384, + 439 + ], + "score": 0.69, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 428, + 441, + 442 + ], + "score": 1.0, + "content": "of dimension", + "type": "text" + }, + { + "bbox": [ + 441, + 430, + 477, + 441 + ], + "score": 0.88, + "content": "H , W , C", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 441, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 107, + 441, + 167, + 451 + ], + "score": 0.9, + "content": "H = W = 1 4", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 441, + 390, + 452 + ], + "score": 1.0, + "content": "are the height and width of the transformed image and", + "type": "text" + }, + { + "bbox": [ + 391, + 441, + 435, + 451 + ], + "score": 0.89, + "content": "C = 1 0 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 441, + 506, + 452 + ], + "score": 1.0, + "content": "is the number of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 450, + 396, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 199, + 466 + ], + "score": 1.0, + "content": "channels. Each feature", + "type": "text" + }, + { + "bbox": [ + 199, + 453, + 220, + 464 + ], + "score": 0.9, + "content": "x _ { h , w }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 450, + 396, + 466 + ], + "score": 1.0, + "content": "represents one region in the original image.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 468, + 505, + 535 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 481 + ], + "score": 1.0, + "content": "Similar to prior work (Hu et al., 2017; Santoro et al., 2017; Perez et al., 2017), we would like to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 480, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 491 + ], + "score": 1.0, + "content": "allow our model to reason explicitly about spatial locations, as required by many of the questions in", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 490, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 506, + 503 + ], + "score": 1.0, + "content": "CLEVR, and therefore we concatenate to this representation a spatial map that represents each of the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 500, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 104, + 500, + 506, + 515 + ], + "score": 1.0, + "content": "positions in the image. However, in contrast to prior work that uses a linear meshgrid feature map", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 511, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 169, + 525 + ], + "score": 1.0, + "content": "with 2 features", + "type": "text" + }, + { + "bbox": [ + 170, + 513, + 177, + 522 + ], + "score": 0.8, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 511, + 196, + 525 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 196, + 514, + 205, + 523 + ], + "score": 0.73, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 511, + 262, + 525 + ], + "score": 1.0, + "content": "ranging from", + "type": "text" + }, + { + "bbox": [ + 262, + 513, + 276, + 523 + ], + "score": 0.79, + "content": "- 1", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 511, + 506, + 525 + ], + "score": 1.0, + "content": "to 1, and to allow better representation of the positions,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 524, + 406, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 406, + 536 + ], + "score": 1.0, + "content": "we use the positional encoding scheme proposed by Vaswani et al. (2017):", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25.5 + }, + { + "type": "interline_equation", + "bbox": [ + 238, + 560, + 372, + 607 + ], + "lines": [ + { + "bbox": [ + 238, + 560, + 372, + 607 + ], + "spans": [ + { + "bbox": [ + 238, + 560, + 372, + 607 + ], + "score": 0.92, + "content": "\\begin{array} { l } { p _ { ( h , 2 i ) } = \\sin { \\Big ( h / 1 0 0 0 0 ^ { 2 i / p _ { d } } \\Big ) } } \\\\ { p _ { ( h , 2 i + 1 ) } = \\cos { \\Big ( h / 1 0 0 0 0 ^ { 2 i / p _ { d } } \\Big ) } } \\end{array}", + "type": "interline_equation", + "image_path": "23f67e68bd8fe769889cb65ef6b26ad32ea8c028817ac3af21e7f4e098381e73.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 238, + 560, + 372, + 583.5 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 238, + 583.5, + 372, + 607.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 504, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 620, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 178, + 633 + ], + "score": 1.0, + "content": "And similarly for", + "type": "text" + }, + { + "bbox": [ + 179, + 623, + 187, + 631 + ], + "score": 0.77, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 620, + 218, + 633 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 218, + 623, + 230, + 632 + ], + "score": 0.86, + "content": "p _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 620, + 505, + 633 + ], + "score": 1.0, + "content": "is a hyperparameter. Overall, the positional encoding of a feature at", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 632, + 449, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 141, + 645 + ], + "score": 1.0, + "content": "position", + "type": "text" + }, + { + "bbox": [ + 142, + 632, + 167, + 644 + ], + "score": 0.92, + "content": "( h , w )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 632, + 178, + 645 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 178, + 632, + 210, + 644 + ], + "score": 0.92, + "content": "[ p _ { h } , p _ { w } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 632, + 411, + 645 + ], + "score": 1.0, + "content": ", the concatenation of the positional encodings for", + "type": "text" + }, + { + "bbox": [ + 411, + 632, + 419, + 642 + ], + "score": 0.84, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 632, + 436, + 645 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 437, + 633, + 445, + 642 + ], + "score": 0.8, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 632, + 449, + 645 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 648, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "This positional encoding scheme allows better correspondence between the distance of 2 positions", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 107, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 107, + 660, + 131, + 672 + ], + "score": 0.91, + "content": "( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 659, + 149, + 673 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 150, + 660, + 173, + 672 + ], + "score": 0.92, + "content": "( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 659, + 483, + 673 + ], + "score": 1.0, + "content": "in the image and a vector similarity of their positional encodings, even when", + "type": "text" + }, + { + "bbox": [ + 483, + 661, + 495, + 672 + ], + "score": 0.86, + "content": "p _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 670, + 172, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 172, + 684 + ], + "score": 1.0, + "content": "larger than two.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 316, + 701 + ], + "score": 1.0, + "content": "We then concatenate the obtained spatial map with", + "type": "text" + }, + { + "bbox": [ + 316, + 691, + 322, + 698 + ], + "score": 0.67, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 687, + 506, + 701 + ], + "score": 1.0, + "content": ", receiving a spatially-aware image represen-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 132, + 711 + ], + "score": 1.0, + "content": "tation", + "type": "text" + }, + { + "bbox": [ + 132, + 700, + 144, + 711 + ], + "score": 0.86, + "content": "x _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 699, + 413, + 711 + ], + "score": 1.0, + "content": ". Then, we pass this representation through two CNN layers with", + "type": "text" + }, + { + "bbox": [ + 413, + 699, + 420, + 709 + ], + "score": 0.76, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "output channels and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 477, + 722 + ], + "score": 1.0, + "content": "obtain a final representation of the image, which we refer to as our Visual Knowledge Base", + "type": "text" + }, + { + "bbox": [ + 477, + 710, + 504, + 721 + ], + "score": 0.85, + "content": "( K B _ { V }", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 721, + 299, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 299, + 732 + ], + "score": 1.0, + "content": "that is used in further components of the model.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "17", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 81, + 255, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 257, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 257, + 95 + ], + "score": 1.0, + "content": "SUPPLEMENTARY MATERIAL", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 106, + 80, + 257, + 95 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 110, + 255, + 124 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 257, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 257, + 126 + ], + "score": 1.0, + "content": "A DETAILS OF INPUT UNIT", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 138, + 505, + 183 + ], + "lines": [ + { + "bbox": [ + 105, + 138, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 151 + ], + "score": 1.0, + "content": "The input unit processes the raw inputs given to the system into distributed vector representations.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 149, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 104, + 149, + 506, + 163 + ], + "score": 1.0, + "content": "It receives a text question (or in general, a query), and an image (or in general, a Knowledge Base", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 160, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 505, + 174 + ], + "score": 1.0, + "content": "(KB)) and processes each of them with a matching sub-unit. Here we provide details of the Query", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 172, + 279, + 184 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 279, + 184 + ], + "score": 1.0, + "content": "Unit and the Image Unit used in this work.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 104, + 138, + 506, + 184 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 200, + 219, + 212 + ], + "lines": [ + { + "bbox": [ + 106, + 199, + 220, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 220, + 214 + ], + "score": 1.0, + "content": "A.0.1 THE QUERY UNIT", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 221, + 505, + 266 + ], + "lines": [ + { + "bbox": [ + 106, + 221, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 193, + 234 + ], + "score": 1.0, + "content": "We encode a query of", + "type": "text" + }, + { + "bbox": [ + 193, + 222, + 201, + 232 + ], + "score": 0.83, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 221, + 505, + 234 + ], + "score": 1.0, + "content": "words into a continuous representation using a bidirectional LSTM (Hochre-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 232, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 505, + 246 + ], + "score": 1.0, + "content": "iter & Schmidhuber, 1997; Graves et al., 2013). Each word is associated with a word embedding", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 107, + 245, + 119, + 255 + ], + "score": 0.85, + "content": "w _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 244, + 151, + 257 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 151, + 244, + 201, + 255 + ], + "score": 0.92, + "content": "s = 1 , . . . , S", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 244, + 505, + 257 + ], + "score": 1.0, + "content": ". In our case, we use GloVE words embeddings (Pennington et al., 2014).", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 255, + 482, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 423, + 267 + ], + "score": 1.0, + "content": "Then, these embeddings are processed by a bidirectional LST M of dimension", + "type": "text" + }, + { + "bbox": [ + 423, + 255, + 430, + 264 + ], + "score": 0.77, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 255, + 482, + 267 + ], + "score": 1.0, + "content": "that outputs:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 221, + 505, + 267 + ] + }, + { + "type": "text", + "bbox": [ + 133, + 279, + 505, + 333 + ], + "lines": [ + { + "bbox": [ + 131, + 279, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 131, + 279, + 237, + 293 + ], + "score": 1.0, + "content": "• a matching sequence of", + "type": "text" + }, + { + "bbox": [ + 238, + 281, + 244, + 290 + ], + "score": 0.83, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 279, + 506, + 293 + ], + "score": 1.0, + "content": "-dimensional output states, which we refer to as contextual words,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 291, + 200, + 303 + ], + "spans": [ + { + "bbox": [ + 143, + 291, + 200, + 303 + ], + "score": 0.79, + "content": "[ c w _ { 1 } , . . . , c w _ { S } ]", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 309, + 142, + 322 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 142, + 311, + 149, + 320 + ], + "score": 0.78, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 309, + 252, + 322 + ], + "score": 1.0, + "content": "-dimensional hidden state", + "type": "text" + }, + { + "bbox": [ + 252, + 309, + 314, + 322 + ], + "score": 0.93, + "content": "q = [ \\overleftrightarrow { \\lambda \\omega _ { 1 } } , \\overrightarrow { c w _ { S } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 309, + 505, + 322 + ], + "score": 1.0, + "content": ", the concatenation of the hidden states from the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 321, + 444, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 321, + 312, + 334 + ], + "score": 1.0, + "content": "backward and forward passes. We refer to", + "type": "text" + }, + { + "bbox": [ + 312, + 324, + 318, + 333 + ], + "score": 0.8, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 321, + 444, + 334 + ], + "score": 1.0, + "content": "as the question representation.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 131, + 279, + 506, + 334 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 346, + 504, + 370 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 245, + 359 + ], + "score": 1.0, + "content": "Intuitively, each contextual word", + "type": "text" + }, + { + "bbox": [ + 245, + 349, + 262, + 358 + ], + "score": 0.87, + "content": "c w _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 345, + 374, + 359 + ], + "score": 1.0, + "content": "represents the meaning of", + "type": "text" + }, + { + "bbox": [ + 375, + 346, + 389, + 357 + ], + "score": 0.89, + "content": "s ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 345, + 506, + 359 + ], + "score": 1.0, + "content": "word, in the context of the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 357, + 496, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 235, + 371 + ], + "score": 1.0, + "content": "question, while the hidden state", + "type": "text" + }, + { + "bbox": [ + 235, + 360, + 241, + 370 + ], + "score": 0.8, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 357, + 496, + 371 + ], + "score": 1.0, + "content": "represents the overall (compositional) meaning of the question.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 345, + 506, + 371 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 386, + 217, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 218, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 218, + 399 + ], + "score": 1.0, + "content": "A.0.2 THE IMAGE UNIT", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 505, + 463 + ], + "lines": [ + { + "bbox": [ + 106, + 408, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 505, + 420 + ], + "score": 1.0, + "content": "Given an image, and following prior work on CLEVR (?Santoro et al., 2017; Perez et al., 2017), we", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 419, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 430 + ], + "score": 1.0, + "content": "extract conv4 features from ResNet101 (He et al., 2016) pretrained on ImageNet (Krizhevsky et al.,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 377, + 442 + ], + "score": 1.0, + "content": "2012) which we treat as a fixed initial representation of the image,", + "type": "text" + }, + { + "bbox": [ + 377, + 432, + 384, + 439 + ], + "score": 0.69, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 428, + 441, + 442 + ], + "score": 1.0, + "content": "of dimension", + "type": "text" + }, + { + "bbox": [ + 441, + 430, + 477, + 441 + ], + "score": 0.88, + "content": "H , W , C", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 441, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 107, + 441, + 167, + 451 + ], + "score": 0.9, + "content": "H = W = 1 4", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 441, + 390, + 452 + ], + "score": 1.0, + "content": "are the height and width of the transformed image and", + "type": "text" + }, + { + "bbox": [ + 391, + 441, + 435, + 451 + ], + "score": 0.89, + "content": "C = 1 0 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 441, + 506, + 452 + ], + "score": 1.0, + "content": "is the number of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 450, + 396, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 199, + 466 + ], + "score": 1.0, + "content": "channels. Each feature", + "type": "text" + }, + { + "bbox": [ + 199, + 453, + 220, + 464 + ], + "score": 0.9, + "content": "x _ { h , w }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 450, + 396, + 466 + ], + "score": 1.0, + "content": "represents one region in the original image.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 408, + 506, + 466 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 468, + 505, + 535 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 481 + ], + "score": 1.0, + "content": "Similar to prior work (Hu et al., 2017; Santoro et al., 2017; Perez et al., 2017), we would like to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 480, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 491 + ], + "score": 1.0, + "content": "allow our model to reason explicitly about spatial locations, as required by many of the questions in", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 490, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 506, + 503 + ], + "score": 1.0, + "content": "CLEVR, and therefore we concatenate to this representation a spatial map that represents each of the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 500, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 104, + 500, + 506, + 515 + ], + "score": 1.0, + "content": "positions in the image. However, in contrast to prior work that uses a linear meshgrid feature map", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 511, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 169, + 525 + ], + "score": 1.0, + "content": "with 2 features", + "type": "text" + }, + { + "bbox": [ + 170, + 513, + 177, + 522 + ], + "score": 0.8, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 511, + 196, + 525 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 196, + 514, + 205, + 523 + ], + "score": 0.73, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 511, + 262, + 525 + ], + "score": 1.0, + "content": "ranging from", + "type": "text" + }, + { + "bbox": [ + 262, + 513, + 276, + 523 + ], + "score": 0.79, + "content": "- 1", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 511, + 506, + 525 + ], + "score": 1.0, + "content": "to 1, and to allow better representation of the positions,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 524, + 406, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 406, + 536 + ], + "score": 1.0, + "content": "we use the positional encoding scheme proposed by Vaswani et al. (2017):", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25.5, + "bbox_fs": [ + 104, + 468, + 506, + 536 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 238, + 560, + 372, + 607 + ], + "lines": [ + { + "bbox": [ + 238, + 560, + 372, + 607 + ], + "spans": [ + { + "bbox": [ + 238, + 560, + 372, + 607 + ], + "score": 0.92, + "content": "\\begin{array} { l } { p _ { ( h , 2 i ) } = \\sin { \\Big ( h / 1 0 0 0 0 ^ { 2 i / p _ { d } } \\Big ) } } \\\\ { p _ { ( h , 2 i + 1 ) } = \\cos { \\Big ( h / 1 0 0 0 0 ^ { 2 i / p _ { d } } \\Big ) } } \\end{array}", + "type": "interline_equation", + "image_path": "23f67e68bd8fe769889cb65ef6b26ad32ea8c028817ac3af21e7f4e098381e73.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 238, + 560, + 372, + 583.5 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 238, + 583.5, + 372, + 607.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 504, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 620, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 178, + 633 + ], + "score": 1.0, + "content": "And similarly for", + "type": "text" + }, + { + "bbox": [ + 179, + 623, + 187, + 631 + ], + "score": 0.77, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 620, + 218, + 633 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 218, + 623, + 230, + 632 + ], + "score": 0.86, + "content": "p _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 620, + 505, + 633 + ], + "score": 1.0, + "content": "is a hyperparameter. Overall, the positional encoding of a feature at", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 632, + 449, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 141, + 645 + ], + "score": 1.0, + "content": "position", + "type": "text" + }, + { + "bbox": [ + 142, + 632, + 167, + 644 + ], + "score": 0.92, + "content": "( h , w )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 632, + 178, + 645 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 178, + 632, + 210, + 644 + ], + "score": 0.92, + "content": "[ p _ { h } , p _ { w } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 632, + 411, + 645 + ], + "score": 1.0, + "content": ", the concatenation of the positional encodings for", + "type": "text" + }, + { + "bbox": [ + 411, + 632, + 419, + 642 + ], + "score": 0.84, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 632, + 436, + 645 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 437, + 633, + 445, + 642 + ], + "score": 0.8, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 632, + 449, + 645 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 620, + 505, + 645 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 648, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "This positional encoding scheme allows better correspondence between the distance of 2 positions", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 107, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 107, + 660, + 131, + 672 + ], + "score": 0.91, + "content": "( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 659, + 149, + 673 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 150, + 660, + 173, + 672 + ], + "score": 0.92, + "content": "( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 659, + 483, + 673 + ], + "score": 1.0, + "content": "in the image and a vector similarity of their positional encodings, even when", + "type": "text" + }, + { + "bbox": [ + 483, + 661, + 495, + 672 + ], + "score": 0.86, + "content": "p _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 670, + 172, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 172, + 684 + ], + "score": 1.0, + "content": "larger than two.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 648, + 505, + 684 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 316, + 701 + ], + "score": 1.0, + "content": "We then concatenate the obtained spatial map with", + "type": "text" + }, + { + "bbox": [ + 316, + 691, + 322, + 698 + ], + "score": 0.67, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 687, + 506, + 701 + ], + "score": 1.0, + "content": ", receiving a spatially-aware image represen-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 132, + 711 + ], + "score": 1.0, + "content": "tation", + "type": "text" + }, + { + "bbox": [ + 132, + 700, + 144, + 711 + ], + "score": 0.86, + "content": "x _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 699, + 413, + 711 + ], + "score": 1.0, + "content": ". Then, we pass this representation through two CNN layers with", + "type": "text" + }, + { + "bbox": [ + 413, + 699, + 420, + 709 + ], + "score": 0.76, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "output channels and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 477, + 722 + ], + "score": 1.0, + "content": "obtain a final representation of the image, which we refer to as our Visual Knowledge Base", + "type": "text" + }, + { + "bbox": [ + 477, + 710, + 504, + 721 + ], + "score": 0.85, + "content": "( K B _ { V }", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 721, + 299, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 299, + 732 + ], + "score": 1.0, + "content": "that is used in further components of the model.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 687, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 348, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 349, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 349, + 95 + ], + "score": 1.0, + "content": "B IMPLEMENTATION AND TRAINING DETAILS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 107, + 505, + 217 + ], + "lines": [ + { + "bbox": [ + 106, + 107, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 106, + 107, + 505, + 120 + ], + "score": 1.0, + "content": "For the question processing, we use GloVE (Pennington et al., 2014) word-vectors with dimension", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 505, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 131 + ], + "score": 1.0, + "content": "300. For the image processing, we extract conv4 features from ResNet101 (He et al., 2016) pre-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 129, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 362, + 141 + ], + "score": 1.0, + "content": "trained on ImageNet (Krizhevsky et al., 2012), with dimension", + "type": "text" + }, + { + "bbox": [ + 362, + 129, + 398, + 141 + ], + "score": 0.78, + "content": "H , W , C", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 129, + 426, + 141 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 426, + 129, + 487, + 140 + ], + "score": 0.9, + "content": "H = W = 1 4", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 129, + 505, + 141 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 139, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 149, + 151 + ], + "score": 0.9, + "content": "C = 1 0 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 139, + 450, + 153 + ], + "score": 1.0, + "content": ", followed by 2 CNN layers with kernel size 2. We use MAC network with", + "type": "text" + }, + { + "bbox": [ + 450, + 141, + 481, + 151 + ], + "score": 0.89, + "content": "p = 1 2", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 139, + 506, + 153 + ], + "score": 1.0, + "content": "cells,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 150, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 104, + 151, + 379, + 163 + ], + "score": 1.0, + "content": "and train it using Adam (Kingma & Ba, 2014), with learning rate", + "type": "text" + }, + { + "bbox": [ + 379, + 150, + 401, + 162 + ], + "score": 0.9, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 151, + 506, + 163 + ], + "score": 1.0, + "content": ". We train our model for", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 161, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 141, + 173 + ], + "score": 0.76, + "content": "1 0 - 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 161, + 506, + 175 + ], + "score": 1.0, + "content": "epochs, with batch size 64, and use early stopping based on validation accuracies. During", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 173, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 506, + 186 + ], + "score": 1.0, + "content": "training, the moving averages of all weights of the model are maintained with the exponential de-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 184, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 505, + 197 + ], + "score": 1.0, + "content": "cay rate of 0.999. At test time, the moving averages instead of the raw weights are used. We use", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 195, + 506, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 506, + 208 + ], + "score": 1.0, + "content": "dropout 0.85, and ELU (Clevert et al., 2015) which in our experience has reduce the training process", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 206, + 481, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 481, + 218 + ], + "score": 1.0, + "content": "compared to RELU.The training process takes roughly 10-20 hours on a single Titan X GPU.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5.5 + }, + { + "type": "title", + "bbox": [ + 107, + 236, + 342, + 248 + ], + "lines": [ + { + "bbox": [ + 106, + 234, + 344, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 344, + 250 + ], + "score": 1.0, + "content": "C FURTHER DISCUSSION OF RELATED WORK", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 328 + ], + "lines": [ + { + "bbox": [ + 105, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "In this section we provide detailed discussion of related work. Several models have been applied", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 271, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 286 + ], + "score": 1.0, + "content": "to the CLEVR task. These can be partitioned into two groups, module networks that use the strong", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 283, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 505, + 296 + ], + "score": 1.0, + "content": "supervision provided as a tree-structured functional program associated with each instance, and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 295, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 505, + 306 + ], + "score": 1.0, + "content": "end-to-end, fully differentiable networks that combine a fairly standard stack of CNNs with com-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 306, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 506, + 317 + ], + "score": 1.0, + "content": "ponents that aid them in performing reasoning tasks. We also discuss the relation of MAC to other", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 317, + 353, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 353, + 330 + ], + "score": 1.0, + "content": "approaches, such as memory networks and neural computers.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 108, + 343, + 224, + 354 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 226, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 226, + 356 + ], + "score": 1.0, + "content": "C.1 MODULE NETWORKS", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 364, + 505, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 377 + ], + "score": 1.0, + "content": "The modular approach (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017) first translates", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "the given question into a tree-structured action plan, aiming to imitate the ground-truth programs", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 387, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 398 + ], + "score": 1.0, + "content": "provided as a form of strong-supervision. Then, it constructs a tailor-made network that executes", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "the plan on the image in multiple steps. This network is composed of discrete units selected out of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 407, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 104, + 407, + 506, + 421 + ], + "score": 1.0, + "content": "a collection of predefined modules, each responsible for an elementary reasoning operation, such as", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 419, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 506, + 433 + ], + "score": 1.0, + "content": "identifying an objects color, filtering them for their shape, or comparing two amounts. Each module", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 431, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 505, + 443 + ], + "score": 1.0, + "content": "has its own set of learned parameters (Johnson et al., 2017), or even hand-crafted design (Andreas", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 442, + 325, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 325, + 453 + ], + "score": 1.0, + "content": "et al., 2016a) to guide it towards its intended behavior.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 458, + 505, + 502 + ], + "lines": [ + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "score": 1.0, + "content": "Overall, this approach makes discrete choices at two levels: the identity of each module – the be-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 469, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 505, + 481 + ], + "score": 1.0, + "content": "havior it should learn among a fixed set of possible types of behaviors, and the network layout – the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 480, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 493 + ], + "score": 1.0, + "content": "way in which these modules are wired together to compute the answer progressively. Hence, their", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 490, + 498, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 498, + 505 + ], + "score": 1.0, + "content": "differentiability is confined to the boundaries of a single module, disallowing end-to-end training.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 508, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 509, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 505, + 520 + ], + "score": 1.0, + "content": "Several key differences exist between our approaches. First, our model replaces the fixed modules", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "score": 1.0, + "content": "collection with one versatile and universal cell that shares both its architecture and parameters across", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "all of its instantiations, and is applied across all the reasoning steps. Second, it replaces the dynamic", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "score": 1.0, + "content": "recursive tree structures with a sequential topology, augmented by soft attention mechanisms, as", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 552, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 564 + ], + "score": 1.0, + "content": "done in Bahdanau et al. (2014). This confers our network with a virtual capacity to represent arbi-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 563, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 505, + 575 + ], + "score": 1.0, + "content": "trarily complex Directed Acyclic Graphs (DAGs) while still having efficient and readily deployed", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "score": 1.0, + "content": "physical sequential structure. Together, both of these relaxations allow us to effectively train our", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "model end-to-end by backpropagation alone, whereas module networks demand a more involved", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "score": 1.0, + "content": "training scheme that relies on the strongly-supervised programs at the first stage, and on various", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 607, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 505, + 619 + ], + "score": 1.0, + "content": "Reinforcement Learning (RL) techniques at the second. Furthermore, while our model can be train", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 617, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 631 + ], + "score": 1.0, + "content": "without the strong supervisory programs, developing adaptive reasoning skills to address the task", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "is it trained for, the modular approach reliance on questions structured and formal representation", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 639, + 277, + 652 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 277, + 652 + ], + "score": 1.0, + "content": "hinder its applicability to real-world tasks.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 37 + }, + { + "type": "title", + "bbox": [ + 106, + 666, + 359, + 677 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 360, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 360, + 679 + ], + "score": 1.0, + "content": "C.2 AUGMENTED CONVOLUTIONAL NEURAL NETWORKS", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 506, + 701 + ], + "score": 1.0, + "content": "Alternative approaches for the CLEVR task that do not rely on the provided programs as a strong", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "supervision signal are Santoro et al. (2017) and Perez et al. (2017). Both complement standard", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "multi-layer Convolutional Neural Networks (CNNs) with components that aid them in handling", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 721, + 265, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 265, + 733 + ], + "score": 1.0, + "content": "compositional and relational questions.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 348, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 349, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 349, + 95 + ], + "score": 1.0, + "content": "B IMPLEMENTATION AND TRAINING DETAILS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 107, + 505, + 217 + ], + "lines": [ + { + "bbox": [ + 106, + 107, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 106, + 107, + 505, + 120 + ], + "score": 1.0, + "content": "For the question processing, we use GloVE (Pennington et al., 2014) word-vectors with dimension", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 505, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 131 + ], + "score": 1.0, + "content": "300. For the image processing, we extract conv4 features from ResNet101 (He et al., 2016) pre-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 129, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 362, + 141 + ], + "score": 1.0, + "content": "trained on ImageNet (Krizhevsky et al., 2012), with dimension", + "type": "text" + }, + { + "bbox": [ + 362, + 129, + 398, + 141 + ], + "score": 0.78, + "content": "H , W , C", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 129, + 426, + 141 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 426, + 129, + 487, + 140 + ], + "score": 0.9, + "content": "H = W = 1 4", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 129, + 505, + 141 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 139, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 149, + 151 + ], + "score": 0.9, + "content": "C = 1 0 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 139, + 450, + 153 + ], + "score": 1.0, + "content": ", followed by 2 CNN layers with kernel size 2. We use MAC network with", + "type": "text" + }, + { + "bbox": [ + 450, + 141, + 481, + 151 + ], + "score": 0.89, + "content": "p = 1 2", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 139, + 506, + 153 + ], + "score": 1.0, + "content": "cells,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 150, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 104, + 151, + 379, + 163 + ], + "score": 1.0, + "content": "and train it using Adam (Kingma & Ba, 2014), with learning rate", + "type": "text" + }, + { + "bbox": [ + 379, + 150, + 401, + 162 + ], + "score": 0.9, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 151, + 506, + 163 + ], + "score": 1.0, + "content": ". We train our model for", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 161, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 141, + 173 + ], + "score": 0.76, + "content": "1 0 - 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 161, + 506, + 175 + ], + "score": 1.0, + "content": "epochs, with batch size 64, and use early stopping based on validation accuracies. During", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 173, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 506, + 186 + ], + "score": 1.0, + "content": "training, the moving averages of all weights of the model are maintained with the exponential de-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 184, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 505, + 197 + ], + "score": 1.0, + "content": "cay rate of 0.999. At test time, the moving averages instead of the raw weights are used. We use", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 195, + 506, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 506, + 208 + ], + "score": 1.0, + "content": "dropout 0.85, and ELU (Clevert et al., 2015) which in our experience has reduce the training process", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 206, + 481, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 481, + 218 + ], + "score": 1.0, + "content": "compared to RELU.The training process takes roughly 10-20 hours on a single Titan X GPU.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5.5, + "bbox_fs": [ + 104, + 107, + 506, + 218 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 236, + 342, + 248 + ], + "lines": [ + { + "bbox": [ + 106, + 234, + 344, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 344, + 250 + ], + "score": 1.0, + "content": "C FURTHER DISCUSSION OF RELATED WORK", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 328 + ], + "lines": [ + { + "bbox": [ + 105, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "In this section we provide detailed discussion of related work. Several models have been applied", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 271, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 286 + ], + "score": 1.0, + "content": "to the CLEVR task. These can be partitioned into two groups, module networks that use the strong", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 283, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 505, + 296 + ], + "score": 1.0, + "content": "supervision provided as a tree-structured functional program associated with each instance, and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 295, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 505, + 306 + ], + "score": 1.0, + "content": "end-to-end, fully differentiable networks that combine a fairly standard stack of CNNs with com-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 306, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 506, + 317 + ], + "score": 1.0, + "content": "ponents that aid them in performing reasoning tasks. We also discuss the relation of MAC to other", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 317, + 353, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 353, + 330 + ], + "score": 1.0, + "content": "approaches, such as memory networks and neural computers.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 261, + 506, + 330 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 343, + 224, + 354 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 226, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 226, + 356 + ], + "score": 1.0, + "content": "C.1 MODULE NETWORKS", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 364, + 505, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 377 + ], + "score": 1.0, + "content": "The modular approach (Andreas et al., 2016a;b; Hu et al., 2017; Johnson et al., 2017) first translates", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "the given question into a tree-structured action plan, aiming to imitate the ground-truth programs", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 387, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 398 + ], + "score": 1.0, + "content": "provided as a form of strong-supervision. Then, it constructs a tailor-made network that executes", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "the plan on the image in multiple steps. This network is composed of discrete units selected out of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 407, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 104, + 407, + 506, + 421 + ], + "score": 1.0, + "content": "a collection of predefined modules, each responsible for an elementary reasoning operation, such as", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 419, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 506, + 433 + ], + "score": 1.0, + "content": "identifying an objects color, filtering them for their shape, or comparing two amounts. Each module", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 431, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 505, + 443 + ], + "score": 1.0, + "content": "has its own set of learned parameters (Johnson et al., 2017), or even hand-crafted design (Andreas", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 442, + 325, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 325, + 453 + ], + "score": 1.0, + "content": "et al., 2016a) to guide it towards its intended behavior.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5, + "bbox_fs": [ + 104, + 364, + 506, + 453 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 458, + 505, + 502 + ], + "lines": [ + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "score": 1.0, + "content": "Overall, this approach makes discrete choices at two levels: the identity of each module – the be-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 469, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 505, + 481 + ], + "score": 1.0, + "content": "havior it should learn among a fixed set of possible types of behaviors, and the network layout – the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 480, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 493 + ], + "score": 1.0, + "content": "way in which these modules are wired together to compute the answer progressively. Hence, their", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 490, + 498, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 498, + 505 + ], + "score": 1.0, + "content": "differentiability is confined to the boundaries of a single module, disallowing end-to-end training.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 459, + 505, + 505 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 508, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 509, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 505, + 520 + ], + "score": 1.0, + "content": "Several key differences exist between our approaches. First, our model replaces the fixed modules", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "score": 1.0, + "content": "collection with one versatile and universal cell that shares both its architecture and parameters across", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "all of its instantiations, and is applied across all the reasoning steps. Second, it replaces the dynamic", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "score": 1.0, + "content": "recursive tree structures with a sequential topology, augmented by soft attention mechanisms, as", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 552, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 564 + ], + "score": 1.0, + "content": "done in Bahdanau et al. (2014). This confers our network with a virtual capacity to represent arbi-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 563, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 505, + 575 + ], + "score": 1.0, + "content": "trarily complex Directed Acyclic Graphs (DAGs) while still having efficient and readily deployed", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "score": 1.0, + "content": "physical sequential structure. Together, both of these relaxations allow us to effectively train our", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "model end-to-end by backpropagation alone, whereas module networks demand a more involved", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "score": 1.0, + "content": "training scheme that relies on the strongly-supervised programs at the first stage, and on various", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 607, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 505, + 619 + ], + "score": 1.0, + "content": "Reinforcement Learning (RL) techniques at the second. Furthermore, while our model can be train", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 617, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 631 + ], + "score": 1.0, + "content": "without the strong supervisory programs, developing adaptive reasoning skills to address the task", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "is it trained for, the modular approach reliance on questions structured and formal representation", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 639, + 277, + 652 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 277, + 652 + ], + "score": 1.0, + "content": "hinder its applicability to real-world tasks.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 509, + 506, + 652 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 666, + 359, + 677 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 360, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 360, + 679 + ], + "score": 1.0, + "content": "C.2 AUGMENTED CONVOLUTIONAL NEURAL NETWORKS", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 506, + 701 + ], + "score": 1.0, + "content": "Alternative approaches for the CLEVR task that do not rely on the provided programs as a strong", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "supervision signal are Santoro et al. (2017) and Perez et al. (2017). Both complement standard", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "multi-layer Convolutional Neural Networks (CNNs) with components that aid them in handling", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 721, + 265, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 265, + 733 + ], + "score": 1.0, + "content": "compositional and relational questions.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 686, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "Relational Networks. Santoro et al. (2017) appends a Relation Network (RN) layer to the CNN.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "This layer inspects all pairs of pixels in the image, thereby enhancing the network capacity to reason", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "over binary relations between objects. While this approach is very simple and elegant conceptually,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 104, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "it suffers from quadratic computational complexity, in contrast to our and other leading approaches.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "But beyond that, closer inspection reveals that this direct pairwise comparison might be unnecessary.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 506, + 151 + ], + "score": 1.0, + "content": "Based on the analogy suggested by Santoro et al. (2017), according to which pixels are equivalent to", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "objects and their pairwise interactions to relations, a RN layer attempts to grasp the induced graph", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 171 + ], + "score": 1.0, + "content": "between objects all at once in one shallow and broad layer. Conversely, our attention-based model", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 169, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 506, + 185 + ], + "score": 1.0, + "content": "proceeds in steps. It basically compares the image to its current memory and control for this step,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 195 + ], + "score": 1.0, + "content": "aggregates the attended regions into the new memory, and repeats the process. By the same analogy,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "it traverses a narrow and deep path, progressively following transitive relations. Consequently, our", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 202, + 460, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 460, + 217 + ], + "score": 1.0, + "content": "model exhibits a relational capacity while circumventing the computational inefficiency.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 219, + 505, + 286 + ], + "lines": [ + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "score": 1.0, + "content": "FiLM. FiLM (Perez et al., 2017) is a recently proposed method that interleaves standard CNN", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "score": 1.0, + "content": "layers that process the given image with linear layers, reminiscent of layer normalization techniques", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "(Ba et al., 2016; Ioffe & Szegedy, 2015). Each of these layers, called FiLM, is conditioned on", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "score": 1.0, + "content": "the question: the question words are processed by a GRU, and its output is linearly transformed", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "score": 1.0, + "content": "into matching biases and variances for each of the CNN layers, tilting its activations to reflect the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 274, + 424, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 424, + 289 + ], + "score": 1.0, + "content": "specifics of the given question and affect the computation done over the image.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 291, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "Similarly to our model, this approach features distant modulation between the question and the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "score": 1.0, + "content": "image, where rather than being fused together into the same vector space, the question can affect the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 314, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 505, + 326 + ], + "score": 1.0, + "content": "image processing only through constrained means – for the case of FiLM – linear transformations.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 324, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 337 + ], + "score": 1.0, + "content": "However, since the same transformation is applied to all the activations homogeneously, agnostic to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 336, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 348 + ], + "score": 1.0, + "content": "both their spatial location as well as the features values, this approach does not allow the question", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 346, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 506, + 360 + ], + "score": 1.0, + "content": "to differentiate between regions in the image based on the objects or concepts they represent – on", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "the content of the image. This stands in stark contrast to our attention-based model, which readily", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "score": 1.0, + "content": "allows and actually encourages the question to inform the model about relevant regions to focus", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "on. We speculate that this still distant, yet more direct interaction between the question and the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "data, or image, for the case of VQA, facilitates learning and increases generalizability. It may be", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 414 + ], + "score": 1.0, + "content": "more suitable to VQA tasks, and CLEVR in particular, where the questions demand the responder", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "to focus on specific objects, and reason about their properties or relations, rather than respond based", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 424, + 506, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 436 + ], + "score": 1.0, + "content": "only on a holistic view of the image that may lead to sub-optimal results (Yang et al., 2016), as is", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "the case of FiLM. Indeed, as demonstrated in 4, there is significant evidence showing our models", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "score": 1.0, + "content": "better generalization capacity, allowing it to achieve high accuracies much faster, and from less data", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 456, + 275, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 275, + 469 + ], + "score": 1.0, + "content": "than FiLM and other competing methods.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 25.5 + }, + { + "type": "title", + "bbox": [ + 108, + 500, + 246, + 511 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 249, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 249, + 512 + ], + "score": 1.0, + "content": "C.3 MEMORY AND ATTENTION", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 528, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 504, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 504, + 540 + ], + "score": 1.0, + "content": "Our architecture draws inspiration from recent research on memory and attention (Kumar et al.,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "2016; Xiong et al., 2016; Graves et al., 2014; 2016). Kumar et al. (2016); Xiong et al. (2016)", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 104, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "propose the Dynamic Memory Network model that proceeds in an iterative process, applying soft", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "attention to retrieve relevant information from a visual or textual KB, which is in turn accumulated", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "into memory passed from one iteration to the next. However, in contrast to our model, it views the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 582, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 597 + ], + "score": 1.0, + "content": "question as an atomic unit, whereas our model decomposes it into a multi-step action plan informing", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "each cell in our sequential network about its current objective. Another key difference is the distant", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "interaction between the question and the KB that characterizes our model. Conversely, DMN fuses", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 615, + 394, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 394, + 630 + ], + "score": 1.0, + "content": "their corresponding representations together into the same vector space.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 633, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 504, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 504, + 645 + ], + "score": 1.0, + "content": "Graves et al. (2016; 2014) complements a neural network with a memory array it can interact with,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "through the means of soft attention. Analogously to our model, it partitions the model into a core", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "neural network, called controller, as well as reading and writing heads that interact with external", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "memory array. However, a main point distinguishing our model from this approach, is the use of", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "dynamic memory, as in Kumar et al. (2016), instead of a fixed-array memory. Each MAC cell is", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "associated with a memory state, our reading unit inspects only the latest memory passed from the", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "previous state, and our writing unit creates a new memory state rather than writing to multiple slots", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "in a fixed shared external memory. Notably, our approach is much more reminiscent of the widely", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 721, + 380, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 380, + 732 + ], + "score": 1.0, + "content": "successful RNN structure, rather than to Graves et al. (2016; 2014) .", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 48 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 752, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "Relational Networks. Santoro et al. (2017) appends a Relation Network (RN) layer to the CNN.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "This layer inspects all pairs of pixels in the image, thereby enhancing the network capacity to reason", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "over binary relations between objects. While this approach is very simple and elegant conceptually,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 104, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "it suffers from quadratic computational complexity, in contrast to our and other leading approaches.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "But beyond that, closer inspection reveals that this direct pairwise comparison might be unnecessary.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 506, + 151 + ], + "score": 1.0, + "content": "Based on the analogy suggested by Santoro et al. (2017), according to which pixels are equivalent to", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "objects and their pairwise interactions to relations, a RN layer attempts to grasp the induced graph", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 171 + ], + "score": 1.0, + "content": "between objects all at once in one shallow and broad layer. Conversely, our attention-based model", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 169, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 506, + 185 + ], + "score": 1.0, + "content": "proceeds in steps. It basically compares the image to its current memory and control for this step,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 195 + ], + "score": 1.0, + "content": "aggregates the attended regions into the new memory, and repeats the process. By the same analogy,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "it traverses a narrow and deep path, progressively following transitive relations. Consequently, our", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 202, + 460, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 460, + 217 + ], + "score": 1.0, + "content": "model exhibits a relational capacity while circumventing the computational inefficiency.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 5.5, + "bbox_fs": [ + 104, + 83, + 506, + 217 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 219, + 505, + 286 + ], + "lines": [ + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "score": 1.0, + "content": "FiLM. FiLM (Perez et al., 2017) is a recently proposed method that interleaves standard CNN", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "score": 1.0, + "content": "layers that process the given image with linear layers, reminiscent of layer normalization techniques", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "(Ba et al., 2016; Ioffe & Szegedy, 2015). Each of these layers, called FiLM, is conditioned on", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "score": 1.0, + "content": "the question: the question words are processed by a GRU, and its output is linearly transformed", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "score": 1.0, + "content": "into matching biases and variances for each of the CNN layers, tilting its activations to reflect the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 274, + 424, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 424, + 289 + ], + "score": 1.0, + "content": "specifics of the given question and affect the computation done over the image.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 220, + 505, + 289 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 291, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "Similarly to our model, this approach features distant modulation between the question and the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "score": 1.0, + "content": "image, where rather than being fused together into the same vector space, the question can affect the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 314, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 505, + 326 + ], + "score": 1.0, + "content": "image processing only through constrained means – for the case of FiLM – linear transformations.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 324, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 337 + ], + "score": 1.0, + "content": "However, since the same transformation is applied to all the activations homogeneously, agnostic to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 336, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 348 + ], + "score": 1.0, + "content": "both their spatial location as well as the features values, this approach does not allow the question", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 346, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 506, + 360 + ], + "score": 1.0, + "content": "to differentiate between regions in the image based on the objects or concepts they represent – on", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "the content of the image. This stands in stark contrast to our attention-based model, which readily", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "score": 1.0, + "content": "allows and actually encourages the question to inform the model about relevant regions to focus", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "on. We speculate that this still distant, yet more direct interaction between the question and the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "data, or image, for the case of VQA, facilitates learning and increases generalizability. It may be", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 414 + ], + "score": 1.0, + "content": "more suitable to VQA tasks, and CLEVR in particular, where the questions demand the responder", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "to focus on specific objects, and reason about their properties or relations, rather than respond based", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 424, + 506, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 436 + ], + "score": 1.0, + "content": "only on a holistic view of the image that may lead to sub-optimal results (Yang et al., 2016), as is", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "the case of FiLM. Indeed, as demonstrated in 4, there is significant evidence showing our models", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "score": 1.0, + "content": "better generalization capacity, allowing it to achieve high accuracies much faster, and from less data", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 456, + 275, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 275, + 469 + ], + "score": 1.0, + "content": "than FiLM and other competing methods.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 291, + 506, + 469 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 500, + 246, + 511 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 249, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 249, + 512 + ], + "score": 1.0, + "content": "C.3 MEMORY AND ATTENTION", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 528, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 504, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 504, + 540 + ], + "score": 1.0, + "content": "Our architecture draws inspiration from recent research on memory and attention (Kumar et al.,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "2016; Xiong et al., 2016; Graves et al., 2014; 2016). Kumar et al. (2016); Xiong et al. (2016)", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 104, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "propose the Dynamic Memory Network model that proceeds in an iterative process, applying soft", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "attention to retrieve relevant information from a visual or textual KB, which is in turn accumulated", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "into memory passed from one iteration to the next. However, in contrast to our model, it views the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 582, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 597 + ], + "score": 1.0, + "content": "question as an atomic unit, whereas our model decomposes it into a multi-step action plan informing", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "each cell in our sequential network about its current objective. Another key difference is the distant", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "interaction between the question and the KB that characterizes our model. Conversely, DMN fuses", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 615, + 394, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 394, + 630 + ], + "score": 1.0, + "content": "their corresponding representations together into the same vector space.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 39, + "bbox_fs": [ + 104, + 528, + 506, + 630 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 633, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 504, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 504, + 645 + ], + "score": 1.0, + "content": "Graves et al. (2016; 2014) complements a neural network with a memory array it can interact with,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "through the means of soft attention. Analogously to our model, it partitions the model into a core", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "neural network, called controller, as well as reading and writing heads that interact with external", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "memory array. However, a main point distinguishing our model from this approach, is the use of", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "dynamic memory, as in Kumar et al. (2016), instead of a fixed-array memory. Each MAC cell is", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "associated with a memory state, our reading unit inspects only the latest memory passed from the", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "previous state, and our writing unit creates a new memory state rather than writing to multiple slots", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "in a fixed shared external memory. Notably, our approach is much more reminiscent of the widely", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 721, + 380, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 380, + 732 + ], + "score": 1.0, + "content": "successful RNN structure, rather than to Graves et al. (2016; 2014) .", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 48, + "bbox_fs": [ + 105, + 633, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "score": 1.0, + "content": "Finally, our approach has potential ties to the VQA models Hu et al. (2017); Lu et al. (2016) which", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "also attend both the to question words and the image while progressively addressing the given ques-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "tion. However, both of these models have distinct specialized designs for each of their attention", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "score": 1.0, + "content": "layers or modules, and have a discrete or fixed layout in which they are composed together. In con-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "trast, our approach relax both of these limitations, having one universal cell design and one universal", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 271, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 271, + 150 + ], + "score": 1.0, + "content": "self-attending sequential network layout.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 108, + 162, + 266, + 173 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 268, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 268, + 175 + ], + "score": 1.0, + "content": "C.4 ATTENTION VS. CONVOLUTION", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 183, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 182, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 195 + ], + "score": 1.0, + "content": "Compared to other leading methods, our model stands out by being heavily based on soft attention,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "whereas most competing approaches are CNN-based, surprisingly lack any attention mechanism.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 204, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 505, + 218 + ], + "score": 1.0, + "content": "Since attention is commonly used in models designed for standard VQA (Antol et al., 2015; Gupta,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "2017; Lu et al., 2016; Yang et al., 2016), it is reasonable to assume that it would be beneficial", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "to incorporate such methods into visual reasoning systems for the CLEVR task as well. In fact,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 238, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 506, + 250 + ], + "score": 1.0, + "content": "attention mechanisms should be especially useful for multi-step reasoning questions such as those", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 248, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 506, + 262 + ], + "score": 1.0, + "content": "present in CLEVR. Such questions refer to several relations between different objects in the image", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 260, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 506, + 272 + ], + "score": 1.0, + "content": "and feature compositional structure that may be approached one step at a time. Thus, it should be", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "score": 1.0, + "content": "beneficial for a cogent responder to have the capacity to selectively focus on on one or some objects", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "score": 1.0, + "content": "at each step, traversing the relevant relational links one after the other, both at the image level, and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 293, + 191, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 191, + 304 + ], + "score": 1.0, + "content": "at the question level.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 309, + 505, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 309, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 505, + 321 + ], + "score": 1.0, + "content": "Moreover, attention mechanisms enhance our model’s ability to perform reasoning skills that pertain", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "to aggregation of information across different regions, such as counting, finding maximum value,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "or performing other reduction operations over information that is spread across the image. Indeed,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 355 + ], + "score": 1.0, + "content": "as discussed in 4, all existing models for visual reasoning, most of which lacking any attention", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 353, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 366 + ], + "score": 1.0, + "content": "mechanism, struggle with the counting and numerical comparisons questions present in CLEVR.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 363, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 377 + ], + "score": 1.0, + "content": "Conversely, our model proves much more capable of performing these reasoning skills, outperform-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 375, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 388 + ], + "score": 1.0, + "content": "ing the other approaches by a wide margin. Noticeably, incorporating soft attention into our model", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 385, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 400 + ], + "score": 1.0, + "content": "makes it much more adept at performing such aggregation reasoning skills, successfully addressing", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 397, + 211, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 211, + 410 + ], + "score": 1.0, + "content": "the this type of questions.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 504, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "score": 1.0, + "content": "Finally, as pointed out by Lu et al. (2016); Yang et al. (2016), soft attention confers the model with", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 424, + 504, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 504, + 438 + ], + "score": 1.0, + "content": "robustness to noise introduced from irrelevant information presents in the image, and higher capacity", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 436, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 505, + 448 + ], + "score": 1.0, + "content": "for handling larger and more diverse vocabulary, the latter being demonstrated in 4. It allows the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 446, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 460 + ], + "score": 1.0, + "content": "model to separate the wheat from the chaff, selectively attending to the relevant information only,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 458, + 402, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 402, + 471 + ], + "score": 1.0, + "content": "and arguably, being more resilient to both visual and linguistic variations.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "score": 1.0, + "content": "Finally, our approach has potential ties to the VQA models Hu et al. (2017); Lu et al. (2016) which", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "also attend both the to question words and the image while progressively addressing the given ques-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "tion. However, both of these models have distinct specialized designs for each of their attention", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "score": 1.0, + "content": "layers or modules, and have a discrete or fixed layout in which they are composed together. In con-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "trast, our approach relax both of these limitations, having one universal cell design and one universal", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 271, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 271, + 150 + ], + "score": 1.0, + "content": "self-attending sequential network layout.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 83, + 505, + 150 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 162, + 266, + 173 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 268, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 268, + 175 + ], + "score": 1.0, + "content": "C.4 ATTENTION VS. CONVOLUTION", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 183, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 182, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 195 + ], + "score": 1.0, + "content": "Compared to other leading methods, our model stands out by being heavily based on soft attention,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "whereas most competing approaches are CNN-based, surprisingly lack any attention mechanism.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 204, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 505, + 218 + ], + "score": 1.0, + "content": "Since attention is commonly used in models designed for standard VQA (Antol et al., 2015; Gupta,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "2017; Lu et al., 2016; Yang et al., 2016), it is reasonable to assume that it would be beneficial", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "to incorporate such methods into visual reasoning systems for the CLEVR task as well. In fact,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 238, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 506, + 250 + ], + "score": 1.0, + "content": "attention mechanisms should be especially useful for multi-step reasoning questions such as those", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 248, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 506, + 262 + ], + "score": 1.0, + "content": "present in CLEVR. Such questions refer to several relations between different objects in the image", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 260, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 506, + 272 + ], + "score": 1.0, + "content": "and feature compositional structure that may be approached one step at a time. Thus, it should be", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "score": 1.0, + "content": "beneficial for a cogent responder to have the capacity to selectively focus on on one or some objects", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "score": 1.0, + "content": "at each step, traversing the relevant relational links one after the other, both at the image level, and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 293, + 191, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 191, + 304 + ], + "score": 1.0, + "content": "at the question level.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 182, + 506, + 304 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 309, + 505, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 309, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 505, + 321 + ], + "score": 1.0, + "content": "Moreover, attention mechanisms enhance our model’s ability to perform reasoning skills that pertain", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "to aggregation of information across different regions, such as counting, finding maximum value,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "or performing other reduction operations over information that is spread across the image. Indeed,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 355 + ], + "score": 1.0, + "content": "as discussed in 4, all existing models for visual reasoning, most of which lacking any attention", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 353, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 366 + ], + "score": 1.0, + "content": "mechanism, struggle with the counting and numerical comparisons questions present in CLEVR.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 363, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 377 + ], + "score": 1.0, + "content": "Conversely, our model proves much more capable of performing these reasoning skills, outperform-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 375, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 388 + ], + "score": 1.0, + "content": "ing the other approaches by a wide margin. Noticeably, incorporating soft attention into our model", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 385, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 400 + ], + "score": 1.0, + "content": "makes it much more adept at performing such aggregation reasoning skills, successfully addressing", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 397, + 211, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 211, + 410 + ], + "score": 1.0, + "content": "the this type of questions.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 309, + 505, + 410 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 504, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "score": 1.0, + "content": "Finally, as pointed out by Lu et al. (2016); Yang et al. (2016), soft attention confers the model with", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 424, + 504, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 504, + 438 + ], + "score": 1.0, + "content": "robustness to noise introduced from irrelevant information presents in the image, and higher capacity", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 436, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 505, + 448 + ], + "score": 1.0, + "content": "for handling larger and more diverse vocabulary, the latter being demonstrated in 4. It allows the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 446, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 460 + ], + "score": 1.0, + "content": "model to separate the wheat from the chaff, selectively attending to the relevant information only,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 458, + 402, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 402, + 471 + ], + "score": 1.0, + "content": "and arguably, being more resilient to both visual and linguistic variations.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 414, + 506, + 471 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/S1Euwz-Rb/S1Euwz-Rb_model.json b/parse/train/S1Euwz-Rb/S1Euwz-Rb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dddd9a5da16ce8f61938d083275fff71d36482c9 --- /dev/null +++ b/parse/train/S1Euwz-Rb/S1Euwz-Rb_model.json @@ -0,0 +1,28722 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 654, + 1301, + 654, + 1301, + 1143, + 398, + 1143 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1455, + 1044, + 1455, + 1044, + 1941, + 298, + 1941 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1286, + 1401, + 1286, + 1401, + 1438, + 298, + 1438 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 1076, + 1493, + 1392, + 1493, + 1392, + 1705, + 1076, + 1705 + ], + "score": 0.962 + }, + { + "category_id": 4, + "poly": [ + 1066, + 1729, + 1402, + 1729, + 1402, + 1910, + 1066, + 1910 + ], + "score": 0.96 + }, + { + "category_id": 0, + "poly": [ + 298, + 221, + 1152, + 221, + 1152, + 324, + 298, + 324 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 301, + 1945, + 1400, + 1945, + 1400, + 2033, + 301, + 2033 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 313, + 377, + 700, + 377, + 700, + 499, + 313, + 499 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 863, + 378, + 1249, + 378, + 1249, + 500, + 863, + 500 + ], + "score": 0.924 + }, + { + "category_id": 0, + "poly": [ + 302, + 1212, + 572, + 1212, + 572, + 1247, + 302, + 1247 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 104, + 299, + 104 + ], + "score": 0.877 + }, + { + "category_id": 0, + "poly": [ + 773, + 580, + 927, + 580, + 927, + 613, + 773, + 613 + ], + "score": 0.87 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 857, + 2088, + 857, + 2112, + 841, + 2112 + ], + "score": 0.695 + }, + { + "category_id": 13, + "poly": [ + 1160, + 1020, + 1236, + 1020, + 1236, + 1049, + 1160, + 1049 + ], + "score": 0.87, + "latex": "9 8 . 9 \\%" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 1636.0, + 1241.0, + 1636.0, + 1241.0, + 1693.0, + 1187.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 1727.0, + 1402.0, + 1727.0, + 1402.0, + 1761.0, + 1066.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 1758.0, + 1406.0, + 1758.0, + 1406.0, + 1792.0, + 1063.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 1788.0, + 1407.0, + 1788.0, + 1407.0, + 1820.0, + 1065.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1852.0, + 1063.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 1851.0, + 1406.0, + 1851.0, + 1406.0, + 1883.0, + 1064.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 1880.0, + 1236.0, + 1880.0, + 1236.0, + 1910.0, + 1064.0, + 1910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 219.0, + 1153.0, + 219.0, + 1153.0, + 271.0, + 297.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 275.0, + 860.0, + 275.0, + 860.0, + 326.0, + 294.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1209.0, + 579.0, + 1209.0, + 579.0, + 1256.0, + 293.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 576.0, + 934.0, + 576.0, + 934.0, + 619.0, + 768.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 840.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 653.0, + 1303.0, + 653.0, + 1303.0, + 690.0, + 395.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 683.0, + 1305.0, + 683.0, + 1305.0, + 723.0, + 393.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 712.0, + 1307.0, + 712.0, + 1307.0, + 753.0, + 393.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 746.0, + 1304.0, + 746.0, + 1304.0, + 778.0, + 393.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 777.0, + 1304.0, + 777.0, + 1304.0, + 811.0, + 393.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 809.0, + 1305.0, + 809.0, + 1305.0, + 840.0, + 395.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 838.0, + 1306.0, + 838.0, + 1306.0, + 872.0, + 393.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 869.0, + 1305.0, + 869.0, + 1305.0, + 901.0, + 393.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 898.0, + 1305.0, + 898.0, + 1305.0, + 932.0, + 393.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 930.0, + 1305.0, + 930.0, + 1305.0, + 962.0, + 394.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 960.0, + 1305.0, + 960.0, + 1305.0, + 992.0, + 395.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 985.0, + 1306.0, + 985.0, + 1306.0, + 1026.0, + 393.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1019.0, + 1159.0, + 1019.0, + 1159.0, + 1054.0, + 395.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1019.0, + 1305.0, + 1019.0, + 1305.0, + 1054.0, + 1237.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1052.0, + 1304.0, + 1052.0, + 1304.0, + 1084.0, + 393.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1075.0, + 1306.0, + 1075.0, + 1306.0, + 1119.0, + 392.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1112.0, + 1150.0, + 1112.0, + 1150.0, + 1146.0, + 393.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1454.0, + 1046.0, + 1454.0, + 1046.0, + 1487.0, + 296.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1483.0, + 1045.0, + 1483.0, + 1045.0, + 1520.0, + 295.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1514.0, + 1046.0, + 1514.0, + 1046.0, + 1550.0, + 295.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1546.0, + 1047.0, + 1546.0, + 1047.0, + 1582.0, + 295.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1577.0, + 1047.0, + 1577.0, + 1047.0, + 1612.0, + 295.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1605.0, + 1047.0, + 1605.0, + 1047.0, + 1643.0, + 295.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1638.0, + 1049.0, + 1638.0, + 1049.0, + 1673.0, + 296.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1668.0, + 1048.0, + 1668.0, + 1048.0, + 1704.0, + 295.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1700.0, + 1046.0, + 1700.0, + 1046.0, + 1732.0, + 293.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1727.0, + 1046.0, + 1727.0, + 1046.0, + 1761.0, + 295.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1760.0, + 1046.0, + 1760.0, + 1046.0, + 1792.0, + 295.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1791.0, + 1047.0, + 1791.0, + 1047.0, + 1823.0, + 295.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1821.0, + 1045.0, + 1821.0, + 1045.0, + 1855.0, + 295.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 1046.0, + 1851.0, + 1046.0, + 1885.0, + 295.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1882.0, + 1046.0, + 1882.0, + 1046.0, + 1917.0, + 295.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1912.0, + 1045.0, + 1912.0, + 1045.0, + 1946.0, + 296.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1285.0, + 1405.0, + 1285.0, + 1405.0, + 1323.0, + 294.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1317.0, + 1406.0, + 1317.0, + 1406.0, + 1350.0, + 296.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1348.0, + 1405.0, + 1348.0, + 1405.0, + 1383.0, + 293.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1376.0, + 1405.0, + 1376.0, + 1405.0, + 1413.0, + 292.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1408.0, + 535.0, + 1408.0, + 535.0, + 1445.0, + 294.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1974.0, + 1404.0, + 1974.0, + 1404.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 2004.0, + 736.0, + 2004.0, + 736.0, + 2037.0, + 297.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 379.0, + 517.0, + 379.0, + 517.0, + 407.0, + 314.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 403.0, + 693.0, + 403.0, + 693.0, + 444.0, + 310.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 435.0, + 539.0, + 435.0, + 539.0, + 471.0, + 312.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 470.0, + 701.0, + 470.0, + 701.0, + 500.0, + 314.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 374.0, + 1163.0, + 374.0, + 1163.0, + 412.0, + 862.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 405.0, + 1242.0, + 405.0, + 1242.0, + 442.0, + 861.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 434.0, + 1090.0, + 434.0, + 1090.0, + 473.0, + 861.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 470.0, + 1251.0, + 470.0, + 1251.0, + 500.0, + 861.0, + 500.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 552, + 1404, + 552, + 1404, + 1102, + 298, + 1102 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 298, + 1117, + 1404, + 1117, + 1404, + 1573, + 298, + 1573 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1727, + 1404, + 1727, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 230, + 1403, + 230, + 1403, + 536, + 298, + 536 + ], + "score": 0.982 + }, + { + "category_id": 0, + "poly": [ + 301, + 1643, + 587, + 1643, + 587, + 1678, + 301, + 1678 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2112, + 841, + 2112 + ], + "score": 0.701 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.162 + }, + { + "category_id": 13, + "poly": [ + 875, + 1299, + 950, + 1299, + 950, + 1328, + 875, + 1328 + ], + "score": 0.88, + "latex": "8 2 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 880, + 1147, + 955, + 1147, + 955, + 1176, + 880, + 1176 + ], + "score": 0.87, + "latex": "9 8 . 9 \\%" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1640.0, + 593.0, + 1640.0, + 593.0, + 1684.0, + 292.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2121.0, + 839.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 548.0, + 1405.0, + 548.0, + 1405.0, + 588.0, + 295.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 583.0, + 1406.0, + 583.0, + 1406.0, + 618.0, + 292.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 612.0, + 1406.0, + 612.0, + 1406.0, + 648.0, + 294.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 644.0, + 1402.0, + 644.0, + 1402.0, + 676.0, + 296.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 672.0, + 1406.0, + 672.0, + 1406.0, + 709.0, + 293.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 706.0, + 1405.0, + 706.0, + 1405.0, + 738.0, + 296.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 734.0, + 1406.0, + 734.0, + 1406.0, + 770.0, + 293.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 765.0, + 1405.0, + 765.0, + 1405.0, + 799.0, + 293.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 795.0, + 1406.0, + 795.0, + 1406.0, + 830.0, + 293.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 826.0, + 1406.0, + 826.0, + 1406.0, + 861.0, + 294.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 855.0, + 1406.0, + 855.0, + 1406.0, + 892.0, + 292.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 887.0, + 1406.0, + 887.0, + 1406.0, + 922.0, + 294.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 916.0, + 1405.0, + 916.0, + 1405.0, + 953.0, + 292.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 946.0, + 1405.0, + 946.0, + 1405.0, + 982.0, + 293.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 976.0, + 1406.0, + 976.0, + 1406.0, + 1014.0, + 293.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1007.0, + 1404.0, + 1007.0, + 1404.0, + 1044.0, + 293.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1038.0, + 1406.0, + 1038.0, + 1406.0, + 1075.0, + 292.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1068.0, + 1175.0, + 1068.0, + 1175.0, + 1105.0, + 292.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1113.0, + 1403.0, + 1113.0, + 1403.0, + 1153.0, + 295.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1146.0, + 879.0, + 1146.0, + 879.0, + 1181.0, + 294.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1146.0, + 1404.0, + 1146.0, + 1404.0, + 1181.0, + 956.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1176.0, + 1405.0, + 1176.0, + 1405.0, + 1212.0, + 292.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1206.0, + 1405.0, + 1206.0, + 1405.0, + 1246.0, + 292.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1231.0, + 1408.0, + 1231.0, + 1408.0, + 1278.0, + 291.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1268.0, + 1407.0, + 1268.0, + 1407.0, + 1304.0, + 292.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1297.0, + 874.0, + 1297.0, + 874.0, + 1334.0, + 293.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1297.0, + 1404.0, + 1297.0, + 1404.0, + 1334.0, + 951.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1328.0, + 1404.0, + 1328.0, + 1404.0, + 1365.0, + 293.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1359.0, + 1406.0, + 1359.0, + 1406.0, + 1395.0, + 292.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1390.0, + 1405.0, + 1390.0, + 1405.0, + 1424.0, + 293.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1421.0, + 1404.0, + 1421.0, + 1404.0, + 1455.0, + 294.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1450.0, + 1405.0, + 1450.0, + 1405.0, + 1485.0, + 294.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1482.0, + 1404.0, + 1482.0, + 1404.0, + 1516.0, + 294.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1513.0, + 1406.0, + 1513.0, + 1406.0, + 1547.0, + 294.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1540.0, + 1274.0, + 1540.0, + 1274.0, + 1580.0, + 293.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1728.0, + 1404.0, + 1728.0, + 1404.0, + 1763.0, + 296.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1760.0, + 1406.0, + 1760.0, + 1406.0, + 1795.0, + 294.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1788.0, + 1407.0, + 1788.0, + 1407.0, + 1825.0, + 292.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1856.0, + 294.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1849.0, + 1406.0, + 1849.0, + 1406.0, + 1886.0, + 293.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1882.0, + 1406.0, + 1882.0, + 1406.0, + 1916.0, + 293.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1911.0, + 1406.0, + 1911.0, + 1406.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1944.0, + 1404.0, + 1944.0, + 1404.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 1406.0, + 2001.0, + 1406.0, + 2039.0, + 293.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 265.0, + 293.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 297.0, + 294.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 292.0, + 1405.0, + 292.0, + 1405.0, + 326.0, + 292.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 321.0, + 1405.0, + 321.0, + 1405.0, + 356.0, + 292.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 350.0, + 1407.0, + 350.0, + 1407.0, + 389.0, + 292.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 382.0, + 1404.0, + 382.0, + 1404.0, + 417.0, + 293.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 413.0, + 1405.0, + 413.0, + 1405.0, + 449.0, + 293.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 441.0, + 1405.0, + 441.0, + 1405.0, + 479.0, + 293.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 474.0, + 1407.0, + 474.0, + 1407.0, + 510.0, + 294.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 505.0, + 1311.0, + 505.0, + 1311.0, + 540.0, + 296.0, + 540.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1343, + 1403, + 1343, + 1403, + 1557, + 298, + 1557 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1036, + 1404, + 1036, + 1404, + 1250, + 298, + 1250 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 897, + 1403, + 897, + 1403, + 1020, + 299, + 1020 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1850, + 1403, + 1850, + 1403, + 2033, + 298, + 2033 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 649, + 1403, + 649, + 1403, + 804, + 299, + 804 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 300, + 1573, + 1402, + 1573, + 1402, + 1696, + 300, + 1696 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 511, + 1402, + 511, + 1402, + 634, + 299, + 634 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 1712, + 1401, + 1712, + 1401, + 1835, + 299, + 1835 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 403, + 1403, + 403, + 1403, + 497, + 299, + 497 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 296, + 229, + 1401, + 229, + 1401, + 292, + 296, + 292 + ], + "score": 0.95 + }, + { + "category_id": 0, + "poly": [ + 299, + 336, + 943, + 336, + 943, + 372, + 299, + 372 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 300, + 839, + 568, + 839, + 568, + 871, + 300, + 871 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 104, + 300, + 104 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 300, + 1286, + 557, + 1286, + 557, + 1318, + 300, + 1318 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2111, + 842, + 2111 + ], + "score": 0.602 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2111, + 842, + 2111 + ], + "score": 0.403 + }, + { + "category_id": 13, + "poly": [ + 340, + 1095, + 516, + 1095, + 516, + 1131, + 340, + 1131 + ], + "score": 0.94, + "latex": "q = [ \\overleftrightarrow { \\lambda \\omega _ { 1 } } , \\overrightarrow { c w _ { S } } ]" + }, + { + "category_id": 13, + "poly": [ + 625, + 1067, + 782, + 1067, + 782, + 1100, + 625, + 1100 + ], + "score": 0.92, + "latex": "[ c w _ { 1 } , . . . , c w _ { S } ]" + }, + { + "category_id": 13, + "poly": [ + 560, + 1189, + 629, + 1189, + 629, + 1219, + 560, + 1219 + ], + "score": 0.91, + "latex": "K B _ { V }" + }, + { + "category_id": 13, + "poly": [ + 963, + 1189, + 1126, + 1189, + 1126, + 1217, + 963, + 1217 + ], + "score": 0.91, + "latex": "H = W = 1 4" + }, + { + "category_id": 13, + "poly": [ + 786, + 1189, + 878, + 1189, + 878, + 1220, + 786, + 1220 + ], + "score": 0.88, + "latex": "H , W , d" + }, + { + "category_id": 13, + "poly": [ + 870, + 1068, + 892, + 1068, + 892, + 1094, + 870, + 1094 + ], + "score": 0.82, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 513, + 1134, + 530, + 1134, + 530, + 1160, + 513, + 1160 + ], + "score": 0.8, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 1053, + 1441, + 1072, + 1441, + 1072, + 1467, + 1053, + 1467 + ], + "score": 0.78, + "latex": "p" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 332.0, + 947.0, + 332.0, + 947.0, + 377.0, + 292.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 836.0, + 570.0, + 836.0, + 570.0, + 876.0, + 293.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1281.0, + 565.0, + 1281.0, + 565.0, + 1324.0, + 293.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1342.0, + 1403.0, + 1342.0, + 1403.0, + 1379.0, + 293.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1375.0, + 1405.0, + 1375.0, + 1405.0, + 1410.0, + 294.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1403.0, + 1404.0, + 1403.0, + 1404.0, + 1443.0, + 293.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1437.0, + 1052.0, + 1437.0, + 1052.0, + 1471.0, + 293.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1437.0, + 1405.0, + 1437.0, + 1405.0, + 1471.0, + 1073.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1466.0, + 1405.0, + 1466.0, + 1405.0, + 1504.0, + 292.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1497.0, + 1407.0, + 1497.0, + 1407.0, + 1532.0, + 292.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1529.0, + 647.0, + 1529.0, + 647.0, + 1559.0, + 294.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1036.0, + 1404.0, + 1036.0, + 1404.0, + 1071.0, + 294.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1066.0, + 624.0, + 1066.0, + 624.0, + 1100.0, + 293.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 1066.0, + 869.0, + 1066.0, + 869.0, + 1100.0, + 783.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 1066.0, + 1404.0, + 1066.0, + 1404.0, + 1100.0, + 893.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1096.0, + 339.0, + 1096.0, + 339.0, + 1131.0, + 292.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1096.0, + 1406.0, + 1096.0, + 1406.0, + 1131.0, + 517.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1127.0, + 512.0, + 1127.0, + 512.0, + 1163.0, + 292.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 1127.0, + 1406.0, + 1127.0, + 1406.0, + 1163.0, + 531.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1160.0, + 1404.0, + 1160.0, + 1404.0, + 1191.0, + 296.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1188.0, + 559.0, + 1188.0, + 559.0, + 1223.0, + 294.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 1188.0, + 785.0, + 1188.0, + 785.0, + 1223.0, + 630.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 1188.0, + 962.0, + 1188.0, + 962.0, + 1223.0, + 879.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 1188.0, + 1404.0, + 1188.0, + 1404.0, + 1223.0, + 1127.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1218.0, + 999.0, + 1218.0, + 999.0, + 1255.0, + 293.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 897.0, + 1407.0, + 897.0, + 1407.0, + 933.0, + 294.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 928.0, + 1407.0, + 928.0, + 1407.0, + 963.0, + 292.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 960.0, + 1405.0, + 960.0, + 1405.0, + 992.0, + 295.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 989.0, + 1147.0, + 989.0, + 1147.0, + 1022.0, + 293.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1852.0, + 1404.0, + 1852.0, + 1404.0, + 1883.0, + 296.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1883.0, + 1404.0, + 1883.0, + 1404.0, + 1915.0, + 296.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1974.0, + 1405.0, + 1974.0, + 1405.0, + 2006.0, + 296.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2001.0, + 1407.0, + 2001.0, + 1407.0, + 2039.0, + 292.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 651.0, + 1404.0, + 651.0, + 1404.0, + 684.0, + 295.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 680.0, + 1404.0, + 680.0, + 1404.0, + 717.0, + 294.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 709.0, + 1405.0, + 709.0, + 1405.0, + 749.0, + 293.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 742.0, + 1405.0, + 742.0, + 1405.0, + 780.0, + 295.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 774.0, + 1320.0, + 774.0, + 1320.0, + 808.0, + 295.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1574.0, + 1404.0, + 1574.0, + 1404.0, + 1607.0, + 295.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1605.0, + 1406.0, + 1605.0, + 1406.0, + 1638.0, + 295.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1633.0, + 1404.0, + 1633.0, + 1404.0, + 1668.0, + 293.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1666.0, + 1285.0, + 1666.0, + 1285.0, + 1698.0, + 295.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 512.0, + 1404.0, + 512.0, + 1404.0, + 545.0, + 297.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 543.0, + 1406.0, + 543.0, + 1406.0, + 576.0, + 294.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 571.0, + 1406.0, + 571.0, + 1406.0, + 609.0, + 293.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 603.0, + 412.0, + 603.0, + 412.0, + 636.0, + 293.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1711.0, + 1405.0, + 1711.0, + 1405.0, + 1746.0, + 294.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1743.0, + 1405.0, + 1743.0, + 1405.0, + 1776.0, + 295.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1774.0, + 1403.0, + 1774.0, + 1403.0, + 1810.0, + 294.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1805.0, + 985.0, + 1805.0, + 985.0, + 1837.0, + 294.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 404.0, + 1405.0, + 404.0, + 1405.0, + 438.0, + 297.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 434.0, + 1405.0, + 434.0, + 1405.0, + 472.0, + 294.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 461.0, + 1111.0, + 461.0, + 1111.0, + 505.0, + 293.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 261.0, + 911.0, + 261.0, + 911.0, + 296.0, + 296.0, + 296.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1037, + 1404, + 1037, + 1404, + 1313, + 298, + 1313 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 808, + 1403, + 808, + 1403, + 1023, + 298, + 1023 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1506, + 1403, + 1506, + 1403, + 1692, + 298, + 1692 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1706, + 1403, + 1706, + 1403, + 1800, + 298, + 1800 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1399, + 1403, + 1399, + 1403, + 1491, + 299, + 1491 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 598, + 228, + 1030, + 228, + 1030, + 552, + 598, + 552 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 301, + 700, + 1400, + 700, + 1400, + 793, + 301, + 793 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 295, + 1966, + 1405, + 1966, + 1405, + 2036, + 295, + 2036 + ], + "score": 0.951 + }, + { + "category_id": 4, + "poly": [ + 297, + 580, + 1403, + 580, + 1403, + 643, + 297, + 643 + ], + "score": 0.942 + }, + { + "category_id": 0, + "poly": [ + 300, + 1345, + 634, + 1345, + 634, + 1378, + 300, + 1378 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 104, + 299, + 104 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1367, + 1908, + 1399, + 1908, + 1399, + 1936, + 1367, + 1936 + ], + "score": 0.88 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1862, + 1399, + 1862, + 1399, + 1891, + 1366, + 1891 + ], + "score": 0.87 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.782 + }, + { + "category_id": 8, + "poly": [ + 740, + 1853, + 959, + 1853, + 959, + 1895, + 740, + 1895 + ], + "score": 0.766 + }, + { + "category_id": 8, + "poly": [ + 694, + 1901, + 1004, + 1901, + 1004, + 1942, + 694, + 1942 + ], + "score": 0.752 + }, + { + "category_id": 8, + "poly": [ + 693, + 1852, + 1006, + 1852, + 1006, + 1942, + 693, + 1942 + ], + "score": 0.231 + }, + { + "category_id": 13, + "poly": [ + 846, + 2001, + 909, + 2001, + 909, + 2037, + 846, + 2037 + ], + "score": 0.93, + "latex": "W _ { i } ^ { d , d }" + }, + { + "category_id": 13, + "poly": [ + 539, + 1569, + 697, + 1569, + 697, + 1601, + 539, + 1601 + ], + "score": 0.92, + "latex": "[ c w _ { 1 } , . . . , c w _ { S } ]" + }, + { + "category_id": 13, + "poly": [ + 543, + 1430, + 701, + 1430, + 701, + 1463, + 543, + 1463 + ], + "score": 0.92, + "latex": "[ c w _ { 1 } , . . . , c w _ { S } ]" + }, + { + "category_id": 13, + "poly": [ + 534, + 2003, + 661, + 2003, + 661, + 2035, + 534, + 2035 + ], + "score": 0.91, + "latex": "i = 1 , . . . , p" + }, + { + "category_id": 14, + "poly": [ + 695, + 1849, + 1006, + 1849, + 1006, + 1947, + 695, + 1947 + ], + "score": 0.91, + "latex": "\\begin{array} { c } { q _ { i } = W _ { i } ^ { d , d } \\cdot q + b _ { i } ^ { d } } \\\\ { c q _ { i } = W ^ { 2 d , d } \\left[ q _ { i } , c _ { i - 1 } \\right] + b ^ { d } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 406, + 1037, + 490, + 1037, + 490, + 1068, + 406, + 1068 + ], + "score": 0.91, + "latex": "M A C _ { i }" + }, + { + "category_id": 13, + "poly": [ + 963, + 2003, + 989, + 2003, + 989, + 2037, + 963, + 2037 + ], + "score": 0.89, + "latex": "b _ { i } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 440, + 1773, + 492, + 1773, + 492, + 1800, + 440, + 1800 + ], + "score": 0.88, + "latex": "c _ { i - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1132, + 1773, + 1169, + 1773, + 1169, + 1801, + 1132, + 1801 + ], + "score": 0.88, + "latex": "c q _ { i }" + }, + { + "category_id": 13, + "poly": [ + 514, + 1465, + 567, + 1465, + 567, + 1493, + 514, + 1493 + ], + "score": 0.88, + "latex": "c _ { i - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1045, + 1044, + 1081, + 1044, + 1081, + 1068, + 1045, + 1068 + ], + "score": 0.85, + "latex": "m _ { i }" + }, + { + "category_id": 13, + "poly": [ + 634, + 1225, + 663, + 1225, + 663, + 1252, + 634, + 1252 + ], + "score": 0.85, + "latex": "c _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 875, + 1042, + 899, + 1042, + 899, + 1068, + 875, + 1068 + ], + "score": 0.85, + "latex": "c _ { i }" + }, + { + "category_id": 13, + "poly": [ + 542, + 1226, + 583, + 1226, + 583, + 1252, + 542, + 1252 + ], + "score": 0.85, + "latex": "m _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1196, + 1135, + 1233, + 1135, + 1233, + 1160, + 1196, + 1160 + ], + "score": 0.85, + "latex": "m _ { i }" + }, + { + "category_id": 13, + "poly": [ + 617, + 1073, + 642, + 1073, + 642, + 1099, + 617, + 1099 + ], + "score": 0.84, + "latex": "c _ { i }" + }, + { + "category_id": 13, + "poly": [ + 542, + 1742, + 567, + 1742, + 567, + 1769, + 542, + 1769 + ], + "score": 0.84, + "latex": "q _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1192, + 1713, + 1210, + 1713, + 1210, + 1739, + 1192, + 1739 + ], + "score": 0.8, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 890, + 1769, + 908, + 1769, + 908, + 1795, + 890, + 1795 + ], + "score": 0.79, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1028, + 1435, + 1045, + 1435, + 1045, + 1461, + 1028, + 1461 + ], + "score": 0.76, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 997, + 1462, + 1014, + 1462, + 1014, + 1488, + 997, + 1488 + ], + "score": 0.76, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1377, + 1222, + 1394, + 1222, + 1394, + 1248, + 1377, + 1248 + ], + "score": 0.73, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 450, + 1071, + 468, + 1071, + 468, + 1096, + 450, + 1096 + ], + "score": 0.7, + "latex": "d" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 238.0, + 753.0, + 238.0, + 753.0, + 275.0, + 659.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 287.0, + 636.0, + 287.0, + 636.0, + 320.0, + 599.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 303.0, + 765.0, + 303.0, + 765.0, + 326.0, + 747.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 291.0, + 840.0, + 291.0, + 840.0, + 317.0, + 805.0, + 317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 287.0, + 1021.0, + 287.0, + 1021.0, + 317.0, + 989.0, + 317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 327.0, + 831.0, + 327.0, + 831.0, + 343.0, + 814.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 415.0, + 1001.0, + 415.0, + 1001.0, + 430.0, + 989.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 437.0, + 706.0, + 437.0, + 706.0, + 470.0, + 668.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 425.0, + 764.0, + 425.0, + 764.0, + 456.0, + 724.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 424.0, + 927.0, + 424.0, + 927.0, + 457.0, + 878.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 453.0, + 866.0, + 453.0, + 866.0, + 487.0, + 817.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 393.0, + 861.0, + 393.0, + 861.0, + 424.0, + 807.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 416.0, + 641.0, + 416.0, + 641.0, + 445.5, + 601.0, + 445.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 418.0, + 1030.0, + 418.0, + 1030.0, + 443.5, + 1002.0, + 443.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 582.0, + 1404.0, + 582.0, + 1404.0, + 614.0, + 296.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 609.0, + 1405.0, + 609.0, + 1405.0, + 646.0, + 295.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1341.0, + 639.0, + 1341.0, + 639.0, + 1382.0, + 294.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1037.0, + 405.0, + 1037.0, + 405.0, + 1072.0, + 295.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 1037.0, + 874.0, + 1037.0, + 874.0, + 1072.0, + 491.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1037.0, + 1044.0, + 1037.0, + 1044.0, + 1072.0, + 900.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 1037.0, + 1405.0, + 1037.0, + 1405.0, + 1072.0, + 1082.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1069.0, + 449.0, + 1069.0, + 449.0, + 1102.0, + 296.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 1069.0, + 616.0, + 1069.0, + 616.0, + 1102.0, + 469.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 1069.0, + 1402.0, + 1069.0, + 1402.0, + 1102.0, + 643.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1098.0, + 1404.0, + 1098.0, + 1404.0, + 1135.0, + 293.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1129.0, + 1195.0, + 1129.0, + 1195.0, + 1166.0, + 293.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1129.0, + 1406.0, + 1129.0, + 1406.0, + 1166.0, + 1234.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1158.0, + 1408.0, + 1158.0, + 1408.0, + 1199.0, + 292.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1191.0, + 1405.0, + 1191.0, + 1405.0, + 1226.0, + 292.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1221.0, + 541.0, + 1221.0, + 541.0, + 1256.0, + 292.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 1221.0, + 633.0, + 1221.0, + 633.0, + 1256.0, + 584.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 1221.0, + 1376.0, + 1221.0, + 1376.0, + 1256.0, + 664.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 1221.0, + 1406.0, + 1221.0, + 1406.0, + 1256.0, + 1395.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1249.0, + 1406.0, + 1249.0, + 1406.0, + 1286.0, + 294.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1281.0, + 1347.0, + 1281.0, + 1347.0, + 1317.0, + 294.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 807.0, + 1405.0, + 807.0, + 1405.0, + 842.0, + 296.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 838.0, + 1405.0, + 838.0, + 1405.0, + 874.0, + 293.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 869.0, + 1405.0, + 869.0, + 1405.0, + 903.0, + 294.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 900.0, + 1405.0, + 900.0, + 1405.0, + 934.0, + 294.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 930.0, + 1405.0, + 930.0, + 1405.0, + 965.0, + 294.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 960.0, + 1404.0, + 960.0, + 1404.0, + 994.0, + 296.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 991.0, + 1188.0, + 991.0, + 1188.0, + 1025.0, + 294.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1505.0, + 1405.0, + 1505.0, + 1405.0, + 1542.0, + 294.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1538.0, + 1405.0, + 1538.0, + 1405.0, + 1570.0, + 293.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1568.0, + 538.0, + 1568.0, + 538.0, + 1604.0, + 294.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 1568.0, + 1407.0, + 1568.0, + 1407.0, + 1604.0, + 698.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1600.0, + 1404.0, + 1600.0, + 1404.0, + 1632.0, + 296.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1630.0, + 1405.0, + 1630.0, + 1405.0, + 1664.0, + 293.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1658.0, + 1389.0, + 1658.0, + 1389.0, + 1695.0, + 293.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1704.0, + 1191.0, + 1704.0, + 1191.0, + 1742.0, + 294.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1704.0, + 1403.0, + 1704.0, + 1403.0, + 1742.0, + 1211.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1740.0, + 541.0, + 1740.0, + 541.0, + 1770.0, + 296.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1740.0, + 1403.0, + 1740.0, + 1403.0, + 1770.0, + 568.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1765.0, + 439.0, + 1765.0, + 439.0, + 1806.0, + 292.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 1765.0, + 889.0, + 1765.0, + 889.0, + 1806.0, + 493.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 1765.0, + 1131.0, + 1765.0, + 1131.0, + 1806.0, + 909.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 1765.0, + 1181.0, + 1765.0, + 1181.0, + 1806.0, + 1170.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1396.0, + 1405.0, + 1396.0, + 1405.0, + 1434.0, + 293.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1428.0, + 542.0, + 1428.0, + 542.0, + 1465.0, + 294.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1428.0, + 1027.0, + 1428.0, + 1027.0, + 1465.0, + 702.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 1428.0, + 1405.0, + 1428.0, + 1405.0, + 1465.0, + 1046.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1459.0, + 513.0, + 1459.0, + 513.0, + 1494.0, + 293.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1459.0, + 996.0, + 1459.0, + 996.0, + 1494.0, + 568.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 1459.0, + 1026.0, + 1459.0, + 1026.0, + 1494.0, + 1015.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 699.0, + 1404.0, + 699.0, + 1404.0, + 737.0, + 295.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 730.0, + 1405.0, + 730.0, + 1405.0, + 766.0, + 293.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 762.0, + 1314.0, + 762.0, + 1314.0, + 796.0, + 295.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1966.0, + 1403.0, + 1966.0, + 1403.0, + 2000.0, + 295.0, + 2000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1993.0, + 533.0, + 1993.0, + 533.0, + 2042.0, + 288.0, + 2042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1993.0, + 845.0, + 1993.0, + 845.0, + 2042.0, + 662.0, + 2042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 1993.0, + 962.0, + 1993.0, + 962.0, + 2042.0, + 910.0, + 2042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1993.0, + 1411.0, + 1993.0, + 1411.0, + 2042.0, + 990.0, + 2042.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 797, + 1405, + 797, + 1405, + 1013, + 297, + 1013 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1398, + 1403, + 1398, + 1403, + 1522, + 298, + 1522 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 581, + 1405, + 581, + 1405, + 675, + 298, + 675 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 689, + 1404, + 689, + 1404, + 784, + 299, + 784 + ], + "score": 0.971 + }, + { + "category_id": 3, + "poly": [ + 563, + 225, + 1130, + 225, + 1130, + 428, + 563, + 428 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 296, + 1246, + 1402, + 1246, + 1402, + 1307, + 296, + 1307 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 295, + 1736, + 1400, + 1736, + 1400, + 1800, + 295, + 1800 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 736, + 1126, + 961, + 1126, + 961, + 1213, + 736, + 1213 + ], + "score": 0.939 + }, + { + "category_id": 4, + "poly": [ + 297, + 456, + 1402, + 456, + 1402, + 519, + 297, + 519 + ], + "score": 0.931 + }, + { + "category_id": 0, + "poly": [ + 300, + 1344, + 588, + 1344, + 588, + 1376, + 300, + 1376 + ], + "score": 0.925 + }, + { + "category_id": 8, + "poly": [ + 696, + 1842, + 1003, + 1842, + 1003, + 1884, + 696, + 1884 + ], + "score": 0.924 + }, + { + "category_id": 8, + "poly": [ + 674, + 1940, + 1028, + 1940, + 1028, + 1981, + 674, + 1981 + ], + "score": 0.915 + }, + { + "category_id": 8, + "poly": [ + 670, + 1988, + 1031, + 1988, + 1031, + 2029, + 670, + 2029 + ], + "score": 0.915 + }, + { + "category_id": 8, + "poly": [ + 626, + 1075, + 1074, + 1075, + 1074, + 1115, + 626, + 1115 + ], + "score": 0.91 + }, + { + "category_id": 1, + "poly": [ + 298, + 1536, + 1404, + 1536, + 1404, + 1722, + 298, + 1722 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 104, + 299, + 104 + ], + "score": 0.898 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1153, + 1400, + 1153, + 1400, + 1182, + 1352, + 1182 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1354, + 1082, + 1400, + 1082, + 1400, + 1111, + 1354, + 1111 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1992, + 1400, + 1992, + 1400, + 2021, + 1352, + 2021 + ], + "score": 0.884 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1850, + 1400, + 1850, + 1400, + 1879, + 1366, + 1879 + ], + "score": 0.88 + }, + { + "category_id": 9, + "poly": [ + 1354, + 1948, + 1400, + 1948, + 1400, + 1977, + 1354, + 1977 + ], + "score": 0.769 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.757 + }, + { + "category_id": 9, + "poly": [ + 1354, + 1948, + 1400, + 1948, + 1400, + 1977, + 1354, + 1977 + ], + "score": 0.116 + }, + { + "category_id": 14, + "poly": [ + 622, + 1071, + 1076, + 1071, + 1076, + 1216, + 622, + 1216 + ], + "score": 0.93, + "latex": "\\begin{array} { c l c r } { { c v _ { i , s } = \\displaystyle \\mathrm { s o f t m a x } \\big ( W ^ { d , 1 } \\big ( c q _ { i } \\circ c w _ { s } \\big ) + b ^ { 1 } \\big ) } } \\\\ { { \\big . } } \\\\ { { c _ { i } = \\displaystyle \\sum _ { s = 1 } ^ { S } c v _ { i , s } \\cdot c w _ { s } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 538, + 890, + 695, + 890, + 695, + 923, + 538, + 923 + ], + "score": 0.93, + "latex": "[ c w _ { 1 } , . . . , c w _ { S } ]" + }, + { + "category_id": 13, + "poly": [ + 1152, + 1739, + 1239, + 1739, + 1239, + 1771, + 1152, + 1771 + ], + "score": 0.92, + "latex": "K B _ { h , w }" + }, + { + "category_id": 14, + "poly": [ + 695, + 1841, + 1005, + 1841, + 1005, + 1885, + 695, + 1885 + ], + "score": 0.91, + "latex": "m _ { i - 1 } ^ { \\prime } = W ^ { d , d } \\cdot m _ { i - 1 } + b ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1768, + 547, + 1768, + 547, + 1800, + 298, + 1800 + ], + "score": 0.91, + "latex": "1 , . . . , H , w = 1 , . . . , W" + }, + { + "category_id": 13, + "poly": [ + 497, + 1461, + 567, + 1461, + 567, + 1491, + 497, + 1491 + ], + "score": 0.91, + "latex": "K B _ { V }" + }, + { + "category_id": 14, + "poly": [ + 669, + 1936, + 1029, + 1936, + 1029, + 2035, + 669, + 2035 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } { \\boldsymbol { K } \\boldsymbol { B } _ { h , w } ^ { \\prime } = \\boldsymbol { W } ^ { d , d } \\cdot \\boldsymbol { K } \\boldsymbol { B } _ { h , w } + \\boldsymbol { b } ^ { d } } & { { } } \\\\ { \\big ( \\boldsymbol { I } _ { m - K B } \\big ) _ { h , w } = \\boldsymbol { m } _ { i - 1 } ^ { \\prime } \\circ \\boldsymbol { K } \\boldsymbol { B } _ { h , w } ^ { \\prime } } & { { } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 358, + 1435, + 424, + 1435, + 424, + 1462, + 358, + 1462 + ], + "score": 0.91, + "latex": "m _ { i - 1 }" + }, + { + "category_id": 13, + "poly": [ + 967, + 1401, + 1034, + 1401, + 1034, + 1430, + 967, + 1430 + ], + "score": 0.9, + "latex": "K B _ { V }" + }, + { + "category_id": 13, + "poly": [ + 1240, + 1572, + 1305, + 1572, + 1305, + 1601, + 1240, + 1601 + ], + "score": 0.9, + "latex": "m _ { i - 1 }" + }, + { + "category_id": 13, + "poly": [ + 846, + 1772, + 911, + 1772, + 911, + 1800, + 846, + 1800 + ], + "score": 0.89, + "latex": "m _ { i - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1007, + 894, + 1044, + 894, + 1044, + 923, + 1007, + 923 + ], + "score": 0.87, + "latex": "c q _ { i }" + }, + { + "category_id": 13, + "poly": [ + 795, + 924, + 834, + 924, + 834, + 951, + 795, + 951 + ], + "score": 0.87, + "latex": "c v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 527, + 757, + 578, + 757, + 578, + 784, + 527, + 784 + ], + "score": 0.86, + "latex": "c _ { i - 1 }" + }, + { + "category_id": 13, + "poly": [ + 916, + 1253, + 939, + 1253, + 939, + 1277, + 916, + 1277 + ], + "score": 0.86, + "latex": "c _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1266, + 1253, + 1305, + 1253, + 1305, + 1277, + 1266, + 1277 + ], + "score": 0.85, + "latex": "c v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1305, + 864, + 1344, + 864, + 1344, + 890, + 1305, + 890 + ], + "score": 0.85, + "latex": "c v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1340, + 1738, + 1403, + 1738, + 1403, + 1770, + 1340, + 1770 + ], + "score": 0.85, + "latex": "\\begin{array} { r l } { h } & { { } = } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 298, + 697, + 334, + 697, + 334, + 724, + 298, + 724 + ], + "score": 0.84, + "latex": "c q _ { i }" + }, + { + "category_id": 13, + "poly": [ + 835, + 1496, + 859, + 1496, + 859, + 1522, + 835, + 1522 + ], + "score": 0.84, + "latex": "c _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1231, + 1604, + 1255, + 1604, + 1255, + 1629, + 1231, + 1629 + ], + "score": 0.84, + "latex": "c _ { i }" + }, + { + "category_id": 13, + "poly": [ + 893, + 726, + 917, + 726, + 917, + 753, + 893, + 753 + ], + "score": 0.83, + "latex": "q _ { i }" + }, + { + "category_id": 13, + "poly": [ + 705, + 1435, + 729, + 1435, + 729, + 1461, + 705, + 1461 + ], + "score": 0.83, + "latex": "c _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1294, + 925, + 1319, + 925, + 1319, + 951, + 1294, + 951 + ], + "score": 0.83, + "latex": "c _ { i }" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 237.0, + 756.0, + 237.0, + 756.0, + 257.0, + 659.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 248.0, + 648.0, + 248.0, + 648.0, + 299.0, + 589.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 265.0, + 769.0, + 265.0, + 769.0, + 275.0, + 760.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 269.0, + 862.0, + 269.0, + 862.0, + 289.0, + 836.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 267.0, + 973.0, + 267.0, + 973.0, + 277.0, + 964.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 288.0, + 781.0, + 288.0, + 781.0, + 309.0, + 757.0, + 309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 295.0, + 893.0, + 295.0, + 893.0, + 304.0, + 884.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 294.0, + 960.0, + 294.0, + 960.0, + 303.0, + 953.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 293.0, + 985.0, + 293.0, + 985.0, + 309.0, + 962.0, + 309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 290.0, + 1132.0, + 290.0, + 1132.0, + 309.0, + 1115.0, + 309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 300.0, + 651.0, + 300.0, + 651.0, + 325.0, + 633.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 302.0, + 1115.0, + 302.0, + 1115.0, + 315.0, + 1102.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 312.0, + 624.0, + 312.0, + 624.0, + 332.0, + 605.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 330.0, + 656.0, + 330.0, + 656.0, + 361.0, + 583.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 325.0, + 1089.0, + 325.0, + 1089.0, + 359.0, + 1018.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 356.0, + 1079.0, + 356.0, + 1079.0, + 374.0, + 1029.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 369.0, + 630.0, + 369.0, + 630.0, + 396.0, + 561.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 370.0, + 1077.0, + 370.0, + 1077.0, + 388.0, + 1031.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 388.0, + 655.0, + 388.0, + 655.0, + 408.0, + 572.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 407.0, + 639.0, + 407.0, + 639.0, + 428.0, + 587.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 305.0, + 754.0, + 305.0, + 754.0, + 322.5, + 742.0, + 322.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 456.0, + 1406.0, + 456.0, + 1406.0, + 491.0, + 295.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 481.0, + 371.0, + 481.0, + 371.0, + 524.0, + 291.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1341.0, + 592.0, + 1341.0, + 592.0, + 1381.0, + 294.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 798.0, + 1406.0, + 798.0, + 1406.0, + 836.0, + 294.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 827.0, + 1405.0, + 827.0, + 1405.0, + 866.0, + 294.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 858.0, + 1304.0, + 858.0, + 1304.0, + 895.0, + 294.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 858.0, + 1406.0, + 858.0, + 1406.0, + 895.0, + 1345.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 890.0, + 537.0, + 890.0, + 537.0, + 925.0, + 295.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 890.0, + 1006.0, + 890.0, + 1006.0, + 925.0, + 696.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 890.0, + 1405.0, + 890.0, + 1405.0, + 925.0, + 1045.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 922.0, + 794.0, + 922.0, + 794.0, + 953.0, + 296.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 922.0, + 1293.0, + 922.0, + 1293.0, + 953.0, + 835.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1320.0, + 922.0, + 1403.0, + 922.0, + 1403.0, + 953.0, + 1320.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 952.0, + 1405.0, + 952.0, + 1405.0, + 986.0, + 294.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 977.0, + 1213.0, + 977.0, + 1213.0, + 1020.0, + 291.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1396.0, + 966.0, + 1396.0, + 966.0, + 1436.0, + 293.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 1396.0, + 1404.0, + 1396.0, + 1404.0, + 1436.0, + 1035.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1432.0, + 357.0, + 1432.0, + 357.0, + 1465.0, + 294.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 1432.0, + 704.0, + 1432.0, + 704.0, + 1465.0, + 425.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 1432.0, + 1405.0, + 1432.0, + 1405.0, + 1465.0, + 730.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1458.0, + 496.0, + 1458.0, + 496.0, + 1497.0, + 293.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1458.0, + 1404.0, + 1458.0, + 1404.0, + 1497.0, + 568.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1489.0, + 834.0, + 1489.0, + 834.0, + 1527.0, + 293.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1489.0, + 1395.0, + 1489.0, + 1395.0, + 1527.0, + 860.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 578.0, + 1405.0, + 578.0, + 1405.0, + 618.0, + 292.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 610.0, + 1405.0, + 610.0, + 1405.0, + 650.0, + 292.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 645.0, + 395.0, + 645.0, + 395.0, + 681.0, + 292.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 689.0, + 297.0, + 689.0, + 297.0, + 727.0, + 292.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 689.0, + 1405.0, + 689.0, + 1405.0, + 727.0, + 335.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 721.0, + 892.0, + 721.0, + 892.0, + 755.0, + 294.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 721.0, + 1405.0, + 721.0, + 1405.0, + 755.0, + 918.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 750.0, + 526.0, + 750.0, + 526.0, + 790.0, + 293.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 750.0, + 592.0, + 750.0, + 592.0, + 790.0, + 579.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1245.0, + 915.0, + 1245.0, + 915.0, + 1280.0, + 295.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 1245.0, + 1265.0, + 1245.0, + 1265.0, + 1280.0, + 940.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 1245.0, + 1404.0, + 1245.0, + 1404.0, + 1280.0, + 1306.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1277.0, + 499.0, + 1277.0, + 499.0, + 1307.0, + 294.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1732.0, + 1151.0, + 1732.0, + 1151.0, + 1773.0, + 293.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 1732.0, + 1339.0, + 1732.0, + 1339.0, + 1773.0, + 1240.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1765.0, + 297.0, + 1765.0, + 297.0, + 1804.0, + 294.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 1765.0, + 845.0, + 1765.0, + 845.0, + 1804.0, + 548.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1765.0, + 1097.0, + 1765.0, + 1097.0, + 1804.0, + 912.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1535.0, + 1406.0, + 1535.0, + 1406.0, + 1572.0, + 295.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1566.0, + 1239.0, + 1566.0, + 1239.0, + 1603.0, + 293.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 1566.0, + 1405.0, + 1566.0, + 1405.0, + 1603.0, + 1306.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1598.0, + 1230.0, + 1598.0, + 1230.0, + 1634.0, + 294.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1598.0, + 1405.0, + 1598.0, + 1405.0, + 1634.0, + 1256.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1630.0, + 1404.0, + 1630.0, + 1404.0, + 1662.0, + 296.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1660.0, + 1406.0, + 1660.0, + 1406.0, + 1695.0, + 294.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1690.0, + 391.0, + 1690.0, + 391.0, + 1723.0, + 293.0, + 1723.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1580, + 1405, + 1580, + 1405, + 1885, + 297, + 1885 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 298, + 745, + 1403, + 745, + 1403, + 962, + 298, + 962 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1125, + 1402, + 1125, + 1402, + 1280, + 298, + 1280 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 577, + 1402, + 577, + 1402, + 730, + 298, + 730 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 558, + 225, + 1137, + 225, + 1137, + 428, + 558, + 428 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 292, + 1503, + 1400, + 1503, + 1400, + 1565, + 292, + 1565 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 298, + 1972, + 1400, + 1972, + 1400, + 2034, + 298, + 2034 + ], + "score": 0.947 + }, + { + "category_id": 8, + "poly": [ + 645, + 1382, + 1052, + 1382, + 1052, + 1476, + 645, + 1476 + ], + "score": 0.939 + }, + { + "category_id": 4, + "poly": [ + 296, + 460, + 1402, + 460, + 1402, + 523, + 296, + 523 + ], + "score": 0.93 + }, + { + "category_id": 0, + "poly": [ + 300, + 1918, + 599, + 1918, + 599, + 1950, + 300, + 1950 + ], + "score": 0.918 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1412, + 1399, + 1412, + 1399, + 1443, + 1352, + 1443 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 104, + 299, + 104 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1353, + 1067, + 1399, + 1067, + 1399, + 1096, + 1353, + 1096 + ], + "score": 0.886 + }, + { + "category_id": 9, + "poly": [ + 1354, + 1022, + 1400, + 1022, + 1400, + 1051, + 1354, + 1051 + ], + "score": 0.884 + }, + { + "category_id": 9, + "poly": [ + 1354, + 1339, + 1400, + 1339, + 1400, + 1369, + 1354, + 1369 + ], + "score": 0.881 + }, + { + "category_id": 8, + "poly": [ + 626, + 1333, + 1070, + 1333, + 1070, + 1375, + 626, + 1375 + ], + "score": 0.874 + }, + { + "category_id": 8, + "poly": [ + 614, + 1015, + 1087, + 1015, + 1087, + 1056, + 614, + 1056 + ], + "score": 0.848 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.783 + }, + { + "category_id": 8, + "poly": [ + 698, + 1063, + 998, + 1063, + 998, + 1101, + 698, + 1101 + ], + "score": 0.743 + }, + { + "category_id": 8, + "poly": [ + 613, + 1014, + 1086, + 1014, + 1086, + 1102, + 613, + 1102 + ], + "score": 0.277 + }, + { + "category_id": 14, + "poly": [ + 624, + 1328, + 1075, + 1328, + 1075, + 1480, + 624, + 1480 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { m v _ { i } = \\operatorname { s o f t m a x } \\left( W ^ { d , d } \\cdot I _ { c m - K B } + b ^ { d } \\right) } \\\\ { m _ { n e w } = \\displaystyle \\sum _ { h , w = 1 , 1 } ^ { H , W } \\left( m v _ { i } \\right) _ { h , w } \\cdot K B _ { h , w } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 610, + 1010, + 1089, + 1010, + 1089, + 1106, + 610, + 1106 + ], + "score": 0.9, + "latex": "\\begin{array} { c } { { I _ { m - K B } , ^ { \\prime } = W ^ { 2 d , d } \\left[ I _ { m - K B } , K B _ { h , w } \\right] + b ^ { d } } } \\\\ { { I _ { c m - K B } = c _ { i } \\circ \\left( I _ { m - K B } \\right) ^ { \\prime } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 996, + 1510, + 1064, + 1510, + 1064, + 1534, + 996, + 1534 + ], + "score": 0.89, + "latex": "m _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 735, + 1828, + 774, + 1828, + 774, + 1855, + 735, + 1855 + ], + "score": 0.87, + "latex": "m _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1073, + 1766, + 1114, + 1766, + 1114, + 1794, + 1073, + 1794 + ], + "score": 0.86, + "latex": "m _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1013, + 1977, + 1049, + 1977, + 1049, + 2003, + 1013, + 2003 + ], + "score": 0.86, + "latex": "m _ { i }" + }, + { + "category_id": 13, + "poly": [ + 928, + 1132, + 954, + 1132, + 954, + 1157, + 928, + 1157 + ], + "score": 0.85, + "latex": "c _ { i }" + }, + { + "category_id": 13, + "poly": [ + 902, + 1707, + 943, + 1707, + 943, + 1733, + 902, + 1733 + ], + "score": 0.85, + "latex": "m _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1221, + 349, + 1221, + 349, + 1249, + 297, + 1249 + ], + "score": 0.85, + "latex": "m v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1538, + 348, + 1538, + 348, + 1564, + 297, + 1564 + ], + "score": 0.84, + "latex": "m v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1166, + 1768, + 1194, + 1768, + 1194, + 1794, + 1166, + 1794 + ], + "score": 0.84, + "latex": "c _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 696, + 1768, + 723, + 1768, + 723, + 1794, + 696, + 1794 + ], + "score": 0.84, + "latex": "c _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 326, + 931, + 344, + 931, + 344, + 957, + 326, + 957 + ], + "score": 0.79, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1296, + 1797, + 1324, + 1797, + 1324, + 1824, + 1296, + 1824 + ], + "score": 0.79, + "latex": "c _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1130, + 1587, + 1148, + 1587, + 1148, + 1612, + 1130, + 1612 + ], + "score": 0.76, + "latex": "q" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 226.0, + 858.0, + 226.0, + 858.0, + 249.0, + 842.0, + 249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 257.0, + 627.0, + 257.0, + 627.0, + 279.0, + 561.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 255.0, + 717.0, + 255.0, + 717.0, + 276.0, + 634.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 276.0, + 627.0, + 276.0, + 627.0, + 300.0, + 563.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 278.0, + 720.0, + 278.0, + 720.0, + 362.0, + 673.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 313.0, + 756.0, + 313.0, + 756.0, + 321.0, + 748.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 300.0, + 804.0, + 300.0, + 804.0, + 319.0, + 772.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 312.0, + 896.0, + 312.0, + 896.0, + 320.0, + 888.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 299.0, + 964.0, + 299.0, + 964.0, + 324.0, + 908.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 322.0, + 826.0, + 322.0, + 826.0, + 343.0, + 803.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 319.0, + 1101.0, + 319.0, + 1101.0, + 332.0, + 1089.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 344.0, + 1064.0, + 344.0, + 1064.0, + 365.0, + 996.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 353.0, + 635.0, + 353.0, + 635.0, + 379.0, + 555.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 375.0, + 616.0, + 375.0, + 616.0, + 397.0, + 576.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.75, + 297.0, + 609.75, + 297.0, + 609.75, + 320.0, + 577.75, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.25, + 304.5, + 1137.25, + 304.5, + 1137.25, + 326.5, + 1099.25, + 326.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 393.0, + 620.0, + 393.0, + 620.0, + 422.0, + 578.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 458.0, + 1403.0, + 458.0, + 1403.0, + 494.0, + 295.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 489.0, + 899.0, + 489.0, + 899.0, + 525.0, + 295.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1913.0, + 604.0, + 1913.0, + 604.0, + 1955.0, + 293.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1578.0, + 1129.0, + 1578.0, + 1129.0, + 1617.0, + 292.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 1578.0, + 1407.0, + 1578.0, + 1407.0, + 1617.0, + 1149.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1610.0, + 1405.0, + 1610.0, + 1405.0, + 1645.0, + 292.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1639.0, + 1403.0, + 1639.0, + 1403.0, + 1677.0, + 292.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1669.0, + 1405.0, + 1669.0, + 1405.0, + 1706.0, + 292.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1700.0, + 901.0, + 1700.0, + 901.0, + 1738.0, + 292.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1700.0, + 1406.0, + 1700.0, + 1406.0, + 1738.0, + 944.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1730.0, + 1406.0, + 1730.0, + 1406.0, + 1768.0, + 292.0, + 1768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1763.0, + 695.0, + 1763.0, + 695.0, + 1798.0, + 294.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 1763.0, + 1072.0, + 1763.0, + 1072.0, + 1798.0, + 724.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 1763.0, + 1165.0, + 1763.0, + 1165.0, + 1798.0, + 1115.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 1763.0, + 1405.0, + 1763.0, + 1405.0, + 1798.0, + 1195.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1793.0, + 1295.0, + 1793.0, + 1295.0, + 1829.0, + 294.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1325.0, + 1793.0, + 1405.0, + 1793.0, + 1405.0, + 1829.0, + 1325.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1825.0, + 734.0, + 1825.0, + 734.0, + 1858.0, + 290.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 1825.0, + 1406.0, + 1825.0, + 1406.0, + 1858.0, + 775.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1851.0, + 406.0, + 1851.0, + 406.0, + 1888.0, + 293.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 745.0, + 1405.0, + 745.0, + 1405.0, + 782.0, + 293.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 779.0, + 1404.0, + 779.0, + 1404.0, + 810.0, + 296.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 808.0, + 1405.0, + 808.0, + 1405.0, + 842.0, + 294.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 837.0, + 1405.0, + 837.0, + 1405.0, + 873.0, + 293.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 868.0, + 1405.0, + 868.0, + 1405.0, + 903.0, + 294.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 899.0, + 1405.0, + 899.0, + 1405.0, + 934.0, + 294.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 930.0, + 325.0, + 930.0, + 325.0, + 965.0, + 294.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 930.0, + 721.0, + 930.0, + 721.0, + 965.0, + 345.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1125.0, + 927.0, + 1125.0, + 927.0, + 1163.0, + 294.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1125.0, + 1406.0, + 1125.0, + 1406.0, + 1163.0, + 955.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1157.0, + 1404.0, + 1157.0, + 1404.0, + 1191.0, + 296.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1183.0, + 1406.0, + 1183.0, + 1406.0, + 1225.0, + 292.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1214.0, + 296.0, + 1214.0, + 296.0, + 1254.0, + 291.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1214.0, + 1407.0, + 1214.0, + 1407.0, + 1254.0, + 350.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1247.0, + 470.0, + 1247.0, + 470.0, + 1285.0, + 293.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 580.0, + 1404.0, + 580.0, + 1404.0, + 613.0, + 297.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 607.0, + 1406.0, + 607.0, + 1406.0, + 644.0, + 293.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 640.0, + 1404.0, + 640.0, + 1404.0, + 672.0, + 293.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 665.0, + 1406.0, + 665.0, + 1406.0, + 710.0, + 291.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 701.0, + 614.0, + 701.0, + 614.0, + 734.0, + 294.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1497.0, + 995.0, + 1497.0, + 995.0, + 1543.0, + 292.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 1497.0, + 1405.0, + 1497.0, + 1405.0, + 1543.0, + 1065.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1535.0, + 648.0, + 1535.0, + 648.0, + 1567.0, + 349.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1971.0, + 1012.0, + 1971.0, + 1012.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1971.0, + 1402.0, + 1971.0, + 1402.0, + 2007.0, + 1050.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1404.0, + 2002.0, + 1404.0, + 2036.0, + 293.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1067, + 1404, + 1067, + 1404, + 1284, + 297, + 1284 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 850, + 1403, + 850, + 1403, + 1036, + 297, + 1036 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1433, + 1403, + 1433, + 1403, + 1591, + 298, + 1591 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1712, + 1403, + 1712, + 1403, + 1837, + 298, + 1837 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1298, + 1404, + 1298, + 1404, + 1421, + 299, + 1421 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 582, + 1405, + 582, + 1405, + 676, + 298, + 676 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 300, + 1604, + 1398, + 1604, + 1398, + 1698, + 300, + 1698 + ], + "score": 0.962 + }, + { + "category_id": 3, + "poly": [ + 625, + 226, + 1076, + 226, + 1076, + 438, + 625, + 438 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 297, + 691, + 1401, + 691, + 1401, + 755, + 297, + 755 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 667, + 794, + 1033, + 794, + 1033, + 836, + 667, + 836 + ], + "score": 0.942 + }, + { + "category_id": 4, + "poly": [ + 296, + 468, + 1401, + 468, + 1401, + 531, + 296, + 531 + ], + "score": 0.941 + }, + { + "category_id": 8, + "poly": [ + 715, + 1938, + 983, + 1938, + 983, + 2027, + 715, + 2027 + ], + "score": 0.939 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.902 + }, + { + "category_id": 8, + "poly": [ + 642, + 1888, + 1058, + 1888, + 1058, + 1928, + 642, + 1928 + ], + "score": 0.89 + }, + { + "category_id": 9, + "poly": [ + 1354, + 1894, + 1400, + 1894, + 1400, + 1923, + 1354, + 1923 + ], + "score": 0.889 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1965, + 1400, + 1965, + 1400, + 1994, + 1351, + 1994 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1366, + 800, + 1399, + 800, + 1399, + 830, + 1366, + 830 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 842, + 2087, + 858, + 2087, + 858, + 2111, + 842, + 2111 + ], + "score": 0.695 + }, + { + "category_id": 2, + "poly": [ + 842, + 2087, + 858, + 2087, + 858, + 2111, + 842, + 2111 + ], + "score": 0.167 + }, + { + "category_id": 14, + "poly": [ + 668, + 793, + 1032, + 793, + 1032, + 834, + 668, + 834 + ], + "score": 0.93, + "latex": "m _ { i } ^ { \\prime } = W ^ { 2 d , d } [ m _ { n e w } , m _ { i - 1 } ] + b ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 1217, + 1611, + 1267, + 1611, + 1267, + 1640, + 1217, + 1640 + ], + "score": 0.91, + "latex": "s a _ { i j }" + }, + { + "category_id": 13, + "poly": [ + 590, + 1530, + 764, + 1530, + 764, + 1560, + 590, + 1560 + ], + "score": 0.91, + "latex": "j = 1 , . . . , i - 1" + }, + { + "category_id": 13, + "poly": [ + 955, + 888, + 1019, + 888, + 1019, + 915, + 955, + 915 + ], + "score": 0.9, + "latex": "m _ { i - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1184, + 1527, + 1223, + 1527, + 1223, + 1561, + 1184, + 1561 + ], + "score": 0.9, + "latex": "j ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 1208, + 1443, + 1332, + 1443, + 1332, + 1469, + 1208, + 1469 + ], + "score": 0.9, + "latex": "c _ { 1 } , . . . , c _ { i - 1 }" + }, + { + "category_id": 13, + "poly": [ + 880, + 589, + 945, + 589, + 945, + 617, + 880, + 617 + ], + "score": 0.89, + "latex": "m _ { i - 1 }" + }, + { + "category_id": 13, + "poly": [ + 756, + 1502, + 880, + 1502, + 880, + 1530, + 756, + 1530 + ], + "score": 0.89, + "latex": "c _ { 1 } , . . . , c _ { i - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1123, + 619, + 1192, + 619, + 1192, + 646, + 1123, + 646 + ], + "score": 0.89, + "latex": "m _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 788, + 1434, + 823, + 1434, + 823, + 1465, + 788, + 1465 + ], + "score": 0.89, + "latex": "i ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 558, + 1472, + 707, + 1472, + 707, + 1500, + 558, + 1500 + ], + "score": 0.89, + "latex": "m _ { 1 } , . . . , m _ { i - 1 }" + }, + { + "category_id": 14, + "poly": [ + 639, + 1884, + 1059, + 1884, + 1059, + 2032, + 639, + 2032 + ], + "score": 0.88, + "latex": "\\begin{array} { c } { { s a _ { i j } = \\displaystyle \\mathrm { s o f t m a x } \\left( W ^ { d , 1 } ( c _ { i } \\circ c _ { j } ) + b ^ { 1 } \\right) } } \\\\ { { ( m _ { s a } ) _ { i } = \\displaystyle \\sum _ { j = 1 } ^ { i - 1 } s a _ { i j } \\cdot m _ { j } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 487, + 1533, + 544, + 1533, + 544, + 1562, + 487, + 1562 + ], + "score": 0.88, + "latex": "s a _ { i , j }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1642, + 348, + 1642, + 348, + 1668, + 297, + 1668 + ], + "score": 0.85, + "latex": "m _ { s a }" + }, + { + "category_id": 13, + "poly": [ + 679, + 1503, + 704, + 1503, + 704, + 1528, + 679, + 1528 + ], + "score": 0.84, + "latex": "c _ { i }" + }, + { + "category_id": 13, + "poly": [ + 550, + 1562, + 563, + 1562, + 563, + 1586, + 550, + 1586 + ], + "score": 0.73, + "latex": "i" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 229.0, + 968.0, + 229.0, + 968.0, + 246.0, + 952.0, + 246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 264.0, + 700.0, + 264.0, + 700.0, + 286.0, + 631.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 257.0, + 804.0, + 257.0, + 804.0, + 279.0, + 715.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 284.0, + 706.0, + 284.0, + 706.0, + 304.0, + 624.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 291.0, + 813.0, + 291.0, + 813.0, + 301.0, + 804.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 302.0, + 897.0, + 302.0, + 897.0, + 319.0, + 870.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 315.0, + 880.0, + 315.0, + 880.0, + 325.0, + 872.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 323.0, + 829.0, + 323.0, + 829.0, + 343.0, + 804.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 335.0, + 870.0, + 335.0, + 870.0, + 357.0, + 841.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 332.0, + 1079.0, + 332.0, + 1079.0, + 354.0, + 1053.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 369.0, + 697.0, + 369.0, + 697.0, + 410.0, + 632.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 372.0, + 994.0, + 372.0, + 994.0, + 415.0, + 929.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 280.5, + 861.0, + 280.5, + 861.0, + 301.5, + 828.0, + 301.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 304.0, + 690.0, + 304.0, + 690.0, + 328.0, + 651.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.75, + 306.5, + 946.75, + 306.5, + 946.75, + 336.0, + 898.75, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 347.0, + 688.0, + 347.0, + 688.0, + 372.5, + 650.0, + 372.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 465.0, + 1405.0, + 465.0, + 1405.0, + 505.0, + 294.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 494.0, + 621.0, + 494.0, + 621.0, + 534.0, + 293.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1067.0, + 1404.0, + 1067.0, + 1404.0, + 1104.0, + 294.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1100.0, + 1404.0, + 1100.0, + 1404.0, + 1133.0, + 294.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1130.0, + 1406.0, + 1130.0, + 1406.0, + 1164.0, + 294.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1161.0, + 1405.0, + 1161.0, + 1405.0, + 1195.0, + 292.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1190.0, + 1406.0, + 1190.0, + 1406.0, + 1225.0, + 295.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1219.0, + 1406.0, + 1219.0, + 1406.0, + 1260.0, + 291.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1252.0, + 1309.0, + 1252.0, + 1309.0, + 1284.0, + 293.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 851.0, + 1403.0, + 851.0, + 1403.0, + 887.0, + 295.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 882.0, + 954.0, + 882.0, + 954.0, + 917.0, + 293.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 882.0, + 1405.0, + 882.0, + 1405.0, + 917.0, + 1020.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 914.0, + 1404.0, + 914.0, + 1404.0, + 949.0, + 295.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 946.0, + 1404.0, + 946.0, + 1404.0, + 977.0, + 295.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 972.0, + 1405.0, + 972.0, + 1405.0, + 1011.0, + 294.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1006.0, + 925.0, + 1006.0, + 925.0, + 1037.0, + 296.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1429.0, + 787.0, + 1429.0, + 787.0, + 1476.0, + 292.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 1429.0, + 1207.0, + 1429.0, + 1207.0, + 1476.0, + 824.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1333.0, + 1429.0, + 1407.0, + 1429.0, + 1407.0, + 1476.0, + 1333.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1466.0, + 557.0, + 1466.0, + 557.0, + 1504.0, + 294.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 1466.0, + 1405.0, + 1466.0, + 1405.0, + 1504.0, + 708.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1496.0, + 678.0, + 1496.0, + 678.0, + 1533.0, + 293.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 1496.0, + 755.0, + 1496.0, + 755.0, + 1533.0, + 705.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 1496.0, + 1404.0, + 1496.0, + 1404.0, + 1533.0, + 881.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1519.0, + 486.0, + 1519.0, + 486.0, + 1571.0, + 289.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 1519.0, + 589.0, + 1519.0, + 589.0, + 1571.0, + 545.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 1519.0, + 1183.0, + 1519.0, + 1183.0, + 1571.0, + 765.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 1519.0, + 1408.0, + 1519.0, + 1408.0, + 1571.0, + 1224.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1558.0, + 549.0, + 1558.0, + 549.0, + 1594.0, + 294.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1558.0, + 726.0, + 1558.0, + 726.0, + 1594.0, + 564.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1712.0, + 1404.0, + 1712.0, + 1404.0, + 1749.0, + 294.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1744.0, + 1407.0, + 1744.0, + 1407.0, + 1780.0, + 293.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1772.0, + 1405.0, + 1772.0, + 1405.0, + 1811.0, + 293.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1805.0, + 1291.0, + 1805.0, + 1291.0, + 1841.0, + 293.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1297.0, + 1403.0, + 1297.0, + 1403.0, + 1330.0, + 294.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1329.0, + 1405.0, + 1329.0, + 1405.0, + 1362.0, + 295.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1357.0, + 1404.0, + 1357.0, + 1404.0, + 1397.0, + 292.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1390.0, + 620.0, + 1390.0, + 620.0, + 1426.0, + 293.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 583.0, + 879.0, + 583.0, + 879.0, + 618.0, + 292.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 583.0, + 1405.0, + 583.0, + 1405.0, + 618.0, + 946.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 613.0, + 1122.0, + 613.0, + 1122.0, + 651.0, + 293.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 613.0, + 1407.0, + 613.0, + 1407.0, + 651.0, + 1193.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 641.0, + 404.0, + 641.0, + 404.0, + 683.0, + 293.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1604.0, + 1216.0, + 1604.0, + 1216.0, + 1641.0, + 295.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1268.0, + 1604.0, + 1404.0, + 1604.0, + 1404.0, + 1641.0, + 1268.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1636.0, + 296.0, + 1636.0, + 296.0, + 1672.0, + 292.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1636.0, + 1406.0, + 1636.0, + 1406.0, + 1672.0, + 349.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1667.0, + 507.0, + 1667.0, + 507.0, + 1701.0, + 294.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 687.0, + 1405.0, + 687.0, + 1405.0, + 727.0, + 292.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 722.0, + 765.0, + 722.0, + 765.0, + 756.0, + 293.0, + 756.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1572, + 1404, + 1572, + 1404, + 1757, + 297, + 1757 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1772, + 1403, + 1772, + 1403, + 1927, + 299, + 1927 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 390, + 1403, + 390, + 1403, + 515, + 297, + 515 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 529, + 1401, + 529, + 1401, + 655, + 298, + 655 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 300, + 1941, + 1401, + 1941, + 1401, + 2036, + 300, + 2036 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 300, + 1001, + 1401, + 1001, + 1401, + 1095, + 300, + 1095 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 369, + 1123, + 1403, + 1123, + 1403, + 1376, + 369, + 1376 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 297, + 1404, + 1400, + 1404, + 1400, + 1469, + 297, + 1469 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 299, + 841, + 1401, + 841, + 1401, + 903, + 299, + 903 + ], + "score": 0.947 + }, + { + "category_id": 8, + "poly": [ + 680, + 311, + 1019, + 311, + 1019, + 355, + 680, + 355 + ], + "score": 0.937 + }, + { + "category_id": 1, + "poly": [ + 289, + 227, + 1338, + 227, + 1338, + 264, + 289, + 264 + ], + "score": 0.92 + }, + { + "category_id": 0, + "poly": [ + 299, + 945, + 539, + 945, + 539, + 977, + 299, + 977 + ], + "score": 0.92 + }, + { + "category_id": 0, + "poly": [ + 300, + 1512, + 594, + 1512, + 594, + 1544, + 300, + 1544 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 816, + 74, + 816, + 105, + 298, + 105 + ], + "score": 0.911 + }, + { + "category_id": 9, + "poly": [ + 1353, + 318, + 1399, + 318, + 1399, + 349, + 1353, + 349 + ], + "score": 0.902 + }, + { + "category_id": 8, + "poly": [ + 539, + 766, + 1160, + 766, + 1160, + 804, + 539, + 804 + ], + "score": 0.901 + }, + { + "category_id": 9, + "poly": [ + 1340, + 727, + 1400, + 727, + 1400, + 758, + 1340, + 758 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1339, + 770, + 1399, + 770, + 1399, + 799, + 1339, + 799 + ], + "score": 0.892 + }, + { + "category_id": 8, + "poly": [ + 733, + 720, + 965, + 720, + 965, + 758, + 733, + 758 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.799 + }, + { + "category_id": 13, + "poly": [ + 550, + 2003, + 630, + 2003, + 630, + 2037, + 550, + 2037 + ], + "score": 0.93, + "latex": "[ m _ { p } , q ]" + }, + { + "category_id": 14, + "poly": [ + 680, + 312, + 1021, + 312, + 1021, + 353, + 680, + 353 + ], + "score": 0.92, + "latex": "m _ { i } ^ { \\prime \\prime } = W ^ { 2 d , d } [ m _ { n e w } , m _ { i } ^ { \\prime } ] + b ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 822, + 231, + 862, + 231, + 862, + 264, + 822, + 264 + ], + "score": 0.9, + "latex": "m _ { i } ^ { \\prime \\prime }" + }, + { + "category_id": 13, + "poly": [ + 651, + 231, + 687, + 231, + 687, + 264, + 651, + 264 + ], + "score": 0.9, + "latex": "m _ { i } ^ { \\prime }" + }, + { + "category_id": 14, + "poly": [ + 539, + 716, + 1160, + 716, + 1160, + 806, + 539, + 806 + ], + "score": 0.9, + "latex": "\\begin{array} { c } { { { c _ { i } } ^ { \\prime } = W ^ { d , d } \\cdot c _ { i } + b ^ { d } } } \\\\ { { m _ { i } = \\mathrm { s i g m o i d } \\left( c _ { i } ^ { \\prime } \\right) \\cdot m _ { i - 1 } + ( 1 - \\mathrm { s i g m o i d } \\left( c _ { i } ^ { \\prime } \\right) ) \\cdot m _ { i } ^ { \\prime \\prime } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 792, + 1287, + 861, + 1287, + 861, + 1314, + 792, + 1314 + ], + "score": 0.89, + "latex": "m _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 839, + 566, + 903, + 566, + 903, + 593, + 839, + 593 + ], + "score": 0.89, + "latex": "m _ { i - 1 }" + }, + { + "category_id": 13, + "poly": [ + 797, + 848, + 833, + 848, + 833, + 873, + 797, + 873 + ], + "score": 0.87, + "latex": "m _ { i }" + }, + { + "category_id": 13, + "poly": [ + 471, + 236, + 522, + 236, + 522, + 261, + 471, + 261 + ], + "score": 0.87, + "latex": "m _ { s a }" + }, + { + "category_id": 13, + "poly": [ + 680, + 1778, + 720, + 1778, + 720, + 1807, + 680, + 1807 + ], + "score": 0.87, + "latex": "m _ { p }" + }, + { + "category_id": 13, + "poly": [ + 515, + 1608, + 555, + 1608, + 555, + 1638, + 515, + 1638 + ], + "score": 0.86, + "latex": "m _ { p }" + }, + { + "category_id": 13, + "poly": [ + 1195, + 1778, + 1236, + 1778, + 1236, + 1807, + 1195, + 1807 + ], + "score": 0.86, + "latex": "m _ { p }" + }, + { + "category_id": 13, + "poly": [ + 1183, + 848, + 1207, + 848, + 1207, + 873, + 1183, + 873 + ], + "score": 0.85, + "latex": "c _ { i }" + }, + { + "category_id": 13, + "poly": [ + 656, + 629, + 680, + 629, + 680, + 653, + 656, + 653 + ], + "score": 0.84, + "latex": "c _ { i }" + }, + { + "category_id": 13, + "poly": [ + 618, + 1442, + 654, + 1442, + 654, + 1467, + 618, + 1467 + ], + "score": 0.84, + "latex": "m _ { i }" + }, + { + "category_id": 13, + "poly": [ + 544, + 1130, + 568, + 1130, + 568, + 1156, + 544, + 1156 + ], + "score": 0.82, + "latex": "c _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1200, + 399, + 1219, + 399, + 1219, + 424, + 1200, + 424 + ], + "score": 0.82, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 648, + 1610, + 668, + 1610, + 668, + 1636, + 648, + 1636 + ], + "score": 0.8, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 468, + 1974, + 486, + 1974, + 486, + 2000, + 468, + 2000 + ], + "score": 0.79, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 895, + 1580, + 911, + 1580, + 911, + 1605, + 895, + 1605 + ], + "score": 0.78, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 608, + 1779, + 626, + 1779, + 626, + 1805, + 608, + 1805 + ], + "score": 0.78, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 1349, + 1283, + 1401, + 1283, + 1401, + 1312, + 1349, + 1312 + ], + "score": 0.62, + "latex": "K B" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 941.0, + 544.0, + 941.0, + 544.0, + 983.0, + 294.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1509.0, + 598.0, + 1509.0, + 598.0, + 1549.0, + 293.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 818.0, + 72.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1573.0, + 894.0, + 1573.0, + 894.0, + 1609.0, + 295.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1573.0, + 1405.0, + 1573.0, + 1405.0, + 1609.0, + 912.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1605.0, + 514.0, + 1605.0, + 514.0, + 1640.0, + 295.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 1605.0, + 647.0, + 1605.0, + 647.0, + 1640.0, + 556.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1605.0, + 1405.0, + 1605.0, + 1405.0, + 1640.0, + 669.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1633.0, + 1406.0, + 1633.0, + 1406.0, + 1670.0, + 292.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1664.0, + 1405.0, + 1664.0, + 1405.0, + 1699.0, + 295.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1690.0, + 1405.0, + 1690.0, + 1405.0, + 1735.0, + 291.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1727.0, + 578.0, + 1727.0, + 578.0, + 1759.0, + 295.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1774.0, + 607.0, + 1774.0, + 607.0, + 1807.0, + 295.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 1774.0, + 679.0, + 1774.0, + 679.0, + 1807.0, + 627.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 1774.0, + 1194.0, + 1774.0, + 1194.0, + 1807.0, + 721.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1774.0, + 1404.0, + 1774.0, + 1404.0, + 1807.0, + 1237.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1803.0, + 1405.0, + 1803.0, + 1405.0, + 1836.0, + 295.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1834.0, + 1405.0, + 1834.0, + 1405.0, + 1869.0, + 293.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1861.0, + 1407.0, + 1861.0, + 1407.0, + 1900.0, + 293.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1895.0, + 673.0, + 1895.0, + 673.0, + 1929.0, + 292.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 393.0, + 1199.0, + 393.0, + 1199.0, + 426.0, + 295.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 393.0, + 1405.0, + 393.0, + 1405.0, + 426.0, + 1220.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 422.0, + 1405.0, + 422.0, + 1405.0, + 458.0, + 292.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 455.0, + 1405.0, + 455.0, + 1405.0, + 489.0, + 292.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 484.0, + 471.0, + 484.0, + 471.0, + 518.0, + 293.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 528.0, + 1405.0, + 528.0, + 1405.0, + 567.0, + 292.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 560.0, + 838.0, + 560.0, + 838.0, + 597.0, + 293.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 560.0, + 1405.0, + 560.0, + 1405.0, + 597.0, + 904.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 593.0, + 1403.0, + 593.0, + 1403.0, + 626.0, + 293.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 621.0, + 655.0, + 621.0, + 655.0, + 660.0, + 293.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 621.0, + 693.0, + 621.0, + 693.0, + 660.0, + 681.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1939.0, + 1403.0, + 1939.0, + 1403.0, + 1978.0, + 295.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 467.0, + 1971.0, + 467.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2007.0, + 487.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1999.0, + 549.0, + 1999.0, + 549.0, + 2040.0, + 294.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 1999.0, + 1340.0, + 1999.0, + 1340.0, + 2040.0, + 631.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1001.0, + 1405.0, + 1001.0, + 1405.0, + 1038.0, + 295.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1035.0, + 1405.0, + 1035.0, + 1405.0, + 1069.0, + 296.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1064.0, + 898.0, + 1064.0, + 898.0, + 1097.0, + 295.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1123.0, + 543.0, + 1123.0, + 543.0, + 1161.0, + 366.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 1123.0, + 1405.0, + 1123.0, + 1405.0, + 1161.0, + 569.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1153.0, + 467.0, + 1153.0, + 467.0, + 1191.0, + 394.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1204.0, + 1403.0, + 1204.0, + 1403.0, + 1238.0, + 369.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1235.0, + 1125.0, + 1235.0, + 1125.0, + 1268.0, + 393.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1281.0, + 791.0, + 1281.0, + 791.0, + 1319.0, + 369.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 1281.0, + 1348.0, + 1281.0, + 1348.0, + 1319.0, + 862.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1281.0, + 1405.0, + 1281.0, + 1405.0, + 1319.0, + 1402.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1314.0, + 1403.0, + 1314.0, + 1403.0, + 1349.0, + 395.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1348.0, + 582.0, + 1348.0, + 582.0, + 1376.0, + 394.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1404.0, + 1404.0, + 1404.0, + 1404.0, + 1440.0, + 296.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1437.0, + 617.0, + 1437.0, + 617.0, + 1473.0, + 293.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 1437.0, + 1244.0, + 1437.0, + 1244.0, + 1473.0, + 655.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 839.0, + 796.0, + 839.0, + 796.0, + 878.0, + 294.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 839.0, + 1182.0, + 839.0, + 1182.0, + 878.0, + 834.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 839.0, + 1406.0, + 839.0, + 1406.0, + 878.0, + 1208.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 869.0, + 354.0, + 869.0, + 354.0, + 906.0, + 292.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 225.0, + 470.0, + 225.0, + 470.0, + 268.0, + 290.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 225.0, + 650.0, + 225.0, + 650.0, + 268.0, + 523.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 225.0, + 821.0, + 225.0, + 821.0, + 268.0, + 688.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 225.0, + 1342.0, + 225.0, + 1342.0, + 268.0, + 863.0, + 268.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1589, + 1404, + 1589, + 1404, + 1897, + 298, + 1897 + ], + "score": 0.983 + }, + { + "category_id": 5, + "poly": [ + 376, + 402, + 1323, + 402, + 1323, + 809, + 376, + 809 + ], + "score": 0.981, + "html": "
ModelOverallCountExistCompare NumbersQuery AttributeCompare Attribute
Human (Johnson et al., 2017)92.686.796.686.595.096.0
Q-type baseline (Johnson et al., 2017)41.834.650.251.036.051.3
LSTM (Johnson et al., 2017)46.841.761.169.836.851.8
CNN+LSTM (Johnson et al., 2017)52.343.765.267.149.353.0
CNN+LSTM+SA (Johnson et al., 2016)76.664.482.777.482.675.4
N2NMN* (Hu et al. 2017)83.768.585.784.990.088.7
PG+EE (9K prog.)* (Johnson et al., 2017)88.679.789.779.192.696.0
PG+EE (70oK prog.)* (Johnson et al., 2017)96.992.797.198.798.198.9
CNN+LSTM+RN†‡ (Santoro et al.,2017)95.590.197.893.697.997.1
CNN+GRU+FiLM (Perez et al.,2017)97.794.399.196.899.199.1
CNN+GRU+FiLM* (Perez et al., 2017)97.694.399.393.499.399.3
CNN+GRU+FiLM (Perez et al., 2017)97.794.399.196.899.199.1
MAC (this paper)-val98.997.199.599.399.299.2
MAC-val99.097.299.599.599.599.4
MAC-test98.997.299.599.499.399.5
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1022, + 1404, + 1022, + 1404, + 1237, + 298, + 1237 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1421, + 1402, + 1421, + 1402, + 1575, + 298, + 1575 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1252, + 1403, + 1252, + 1403, + 1406, + 298, + 1406 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1402, + 1911, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.974 + }, + { + "category_id": 0, + "poly": [ + 299, + 916, + 557, + 916, + 557, + 952, + 299, + 952 + ], + "score": 0.913 + }, + { + "category_id": 6, + "poly": [ + 296, + 223, + 1404, + 223, + 1404, + 375, + 296, + 375 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.792 + }, + { + "category_id": 1, + "poly": [ + 296, + 223, + 1404, + 223, + 1404, + 375, + 296, + 375 + ], + "score": 0.153 + }, + { + "category_id": 13, + "poly": [ + 328, + 1973, + 402, + 1973, + 402, + 2002, + 328, + 2002 + ], + "score": 0.87, + "latex": "9 7 . 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 296, + 1834, + 373, + 1834, + 373, + 1864, + 296, + 1864 + ], + "score": 0.86, + "latex": "9 5 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 662, + 1345, + 752, + 1345, + 752, + 1374, + 662, + 1374 + ], + "score": 0.86, + "latex": "9 8 . 9 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 487, + 1774, + 578, + 1774, + 578, + 1803, + 487, + 1803 + ], + "score": 0.75, + "latex": "\\mathrm { P G + E E }" + }, + { + "category_id": 13, + "poly": [ + 396, + 635, + 559, + 635, + 559, + 658, + 396, + 658 + ], + "score": 0.7, + "latex": "\\mathrm { { C N N + L S T M + R N } ^ { \\dagger \\ddagger } }" + }, + { + "category_id": 13, + "poly": [ + 604, + 314, + 635, + 314, + 635, + 347, + 604, + 347 + ], + "score": 0.52, + "latex": "( ^ { \\ddagger } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 2003, + 343, + 2003, + 343, + 2032, + 298, + 2032 + ], + "score": 0.49, + "latex": "1 0 \\mathrm { x }" + }, + { + "category_id": 13, + "poly": [ + 490, + 718, + 501, + 718, + 501, + 734, + 490, + 734 + ], + "score": 0.44, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 1032, + 316, + 1064, + 316, + 1064, + 347, + 1032, + 347 + ], + "score": 0.42, + "latex": "( ^ { \\circ } )" + }, + { + "category_id": 13, + "poly": [ + 396, + 682, + 557, + 682, + 557, + 706, + 396, + 706 + ], + "score": 0.42, + "latex": "\\scriptstyle \\mathbf { C N N + G R U + F i L M ^ { \\ddag } }" + }, + { + "category_id": 13, + "poly": [ + 395, + 591, + 458, + 591, + 458, + 612, + 395, + 612 + ], + "score": 0.36, + "latex": "\\mathrm { P G + E E }" + }, + { + "category_id": 13, + "poly": [ + 1275, + 283, + 1305, + 283, + 1305, + 317, + 1275, + 317 + ], + "score": 0.28, + "latex": "( ^ { \\dagger } )" + }, + { + "category_id": 13, + "poly": [ + 396, + 547, + 538, + 547, + 538, + 567, + 396, + 567 + ], + "score": 0.26, + "latex": "\\mathbf { C N N + L S T M + S A }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 914.0, + 561.0, + 914.0, + 561.0, + 957.0, + 293.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 224.0, + 1404.0, + 224.0, + 1404.0, + 257.0, + 296.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 253.0, + 1402.0, + 253.0, + 1402.0, + 288.0, + 295.0, + 288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 282.0, + 1274.0, + 282.0, + 1274.0, + 323.0, + 294.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 282.0, + 1407.0, + 282.0, + 1407.0, + 323.0, + 1306.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 315.0, + 603.0, + 315.0, + 603.0, + 351.0, + 294.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 315.0, + 1031.0, + 315.0, + 1031.0, + 351.0, + 636.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 315.0, + 1405.0, + 315.0, + 1405.0, + 351.0, + 1065.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 342.0, + 450.0, + 342.0, + 450.0, + 379.0, + 294.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1591.0, + 1402.0, + 1591.0, + 1402.0, + 1623.0, + 297.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1619.0, + 1407.0, + 1619.0, + 1407.0, + 1655.0, + 292.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1650.0, + 1406.0, + 1650.0, + 1406.0, + 1687.0, + 293.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1681.0, + 1405.0, + 1681.0, + 1405.0, + 1715.0, + 295.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1712.0, + 1406.0, + 1712.0, + 1406.0, + 1748.0, + 293.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1740.0, + 1406.0, + 1740.0, + 1406.0, + 1779.0, + 292.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1771.0, + 486.0, + 1771.0, + 486.0, + 1808.0, + 292.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 1771.0, + 1404.0, + 1771.0, + 1404.0, + 1808.0, + 579.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1803.0, + 1406.0, + 1803.0, + 1406.0, + 1839.0, + 293.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 1832.0, + 1408.0, + 1832.0, + 1408.0, + 1871.0, + 374.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1865.0, + 1406.0, + 1865.0, + 1406.0, + 1901.0, + 294.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1020.0, + 1405.0, + 1020.0, + 1405.0, + 1061.0, + 293.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1052.0, + 1407.0, + 1052.0, + 1407.0, + 1089.0, + 292.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1085.0, + 1404.0, + 1085.0, + 1404.0, + 1120.0, + 294.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1116.0, + 1406.0, + 1116.0, + 1406.0, + 1150.0, + 293.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1145.0, + 1405.0, + 1145.0, + 1405.0, + 1180.0, + 293.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1174.0, + 1405.0, + 1174.0, + 1405.0, + 1213.0, + 292.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1208.0, + 691.0, + 1208.0, + 691.0, + 1239.0, + 296.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1421.0, + 1404.0, + 1421.0, + 1404.0, + 1458.0, + 297.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1452.0, + 1402.0, + 1452.0, + 1402.0, + 1489.0, + 294.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1482.0, + 1404.0, + 1482.0, + 1404.0, + 1519.0, + 294.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1510.0, + 1403.0, + 1510.0, + 1403.0, + 1554.0, + 292.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1544.0, + 935.0, + 1544.0, + 935.0, + 1577.0, + 293.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1250.0, + 1405.0, + 1250.0, + 1405.0, + 1289.0, + 294.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1283.0, + 1405.0, + 1283.0, + 1405.0, + 1318.0, + 293.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1313.0, + 1405.0, + 1313.0, + 1405.0, + 1349.0, + 292.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1344.0, + 661.0, + 1344.0, + 661.0, + 1381.0, + 294.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 1344.0, + 1405.0, + 1344.0, + 1405.0, + 1381.0, + 753.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1373.0, + 674.0, + 1373.0, + 674.0, + 1410.0, + 294.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1909.0, + 1406.0, + 1909.0, + 1406.0, + 1948.0, + 293.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1403.0, + 1941.0, + 1403.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 327.0, + 1972.0, + 327.0, + 2007.0, + 293.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2007.0, + 403.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 2004.0, + 641.0, + 2004.0, + 641.0, + 2036.0, + 344.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 224.0, + 1404.0, + 224.0, + 1404.0, + 257.0, + 296.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 253.0, + 1402.0, + 253.0, + 1402.0, + 288.0, + 295.0, + 288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 282.0, + 1274.0, + 282.0, + 1274.0, + 323.0, + 294.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 282.0, + 1407.0, + 282.0, + 1407.0, + 323.0, + 1306.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 315.0, + 603.0, + 315.0, + 603.0, + 351.0, + 294.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 315.0, + 1031.0, + 315.0, + 1031.0, + 351.0, + 636.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 315.0, + 1405.0, + 315.0, + 1405.0, + 351.0, + 1065.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 342.0, + 450.0, + 342.0, + 450.0, + 379.0, + 294.0, + 379.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1819, + 1403, + 1819, + 1403, + 2035, + 297, + 2035 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1451, + 1404, + 1451, + 1404, + 1636, + 298, + 1636 + ], + "score": 0.98 + }, + { + "category_id": 5, + "poly": [ + 503, + 1018, + 1194, + 1018, + 1194, + 1246, + 503, + 1246 + ], + "score": 0.979, + "html": "
ModelTrain CLEVRTrain CLEVR + fine-tune HUMANS
LSTM (Johnson et al.,2017)27.536.5
CNN+LSTM (Johnson et al.,2017)37.743.2
CNN+LSTM+SA+MLP (Johnson et al.,2016)50.457.6
PG+EE (18K prog.)* (Johnson et al.,2017)54.066.6
CNN+GRU+FiLM (Perez et al., 2017)56.675.9
MAC-val56.981.2
MAC-val58.582.0
MAC-test58.682.5
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1650, + 1404, + 1650, + 1404, + 1805, + 297, + 1805 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 500, + 239, + 1191, + 239, + 1191, + 575, + 500, + 575 + ], + "score": 0.969 + }, + { + "category_id": 4, + "poly": [ + 296, + 619, + 1405, + 619, + 1405, + 773, + 296, + 773 + ], + "score": 0.957 + }, + { + "category_id": 0, + "poly": [ + 298, + 1366, + 581, + 1366, + 581, + 1397, + 298, + 1397 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 865, + 2088, + 865, + 2112, + 837, + 2112 + ], + "score": 0.849 + }, + { + "category_id": 6, + "poly": [ + 299, + 868, + 1403, + 868, + 1403, + 991, + 299, + 991 + ], + "score": 0.802 + }, + { + "category_id": 1, + "poly": [ + 299, + 868, + 1403, + 868, + 1403, + 991, + 299, + 991 + ], + "score": 0.21 + }, + { + "category_id": 13, + "poly": [ + 411, + 1881, + 569, + 1881, + 569, + 1911, + 411, + 1911 + ], + "score": 0.89, + "latex": "4 7 . 6 \\% - 5 7 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 571, + 1743, + 626, + 1743, + 626, + 1771, + 571, + 1771 + ], + "score": 0.88, + "latex": "2 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 599, + 1713, + 653, + 1713, + 653, + 1741, + 599, + 1741 + ], + "score": 0.88, + "latex": "92 \\%" + }, + { + "category_id": 13, + "poly": [ + 490, + 1713, + 545, + 1713, + 545, + 1742, + 490, + 1742 + ], + "score": 0.88, + "latex": "70 \\%" + }, + { + "category_id": 13, + "poly": [ + 1017, + 1514, + 1072, + 1514, + 1072, + 1543, + 1017, + 1543 + ], + "score": 0.87, + "latex": "50 \\%" + }, + { + "category_id": 13, + "poly": [ + 538, + 1682, + 593, + 1682, + 593, + 1710, + 538, + 1710 + ], + "score": 0.87, + "latex": "50 \\%" + }, + { + "category_id": 13, + "poly": [ + 791, + 1851, + 865, + 1851, + 865, + 1880, + 791, + 1880 + ], + "score": 0.87, + "latex": "8 4 . 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 429, + 1821, + 483, + 1821, + 483, + 1850, + 429, + 1850 + ], + "score": 0.87, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 961, + 1713, + 1036, + 1713, + 1036, + 1742, + 961, + 1742 + ], + "score": 0.87, + "latex": "9 7 . 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 1194, + 1912, + 1268, + 1912, + 1268, + 1942, + 1194, + 1942 + ], + "score": 0.87, + "latex": "4 2 . 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 1103, + 712, + 1169, + 712, + 1169, + 741, + 1103, + 741 + ], + "score": 0.87, + "latex": "100 \\%" + }, + { + "category_id": 13, + "poly": [ + 526, + 1774, + 580, + 1774, + 580, + 1803, + 526, + 1803 + ], + "score": 0.87, + "latex": "7 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 421, + 1774, + 476, + 1774, + 476, + 1803, + 421, + 1803 + ], + "score": 0.86, + "latex": "50 \\%" + }, + { + "category_id": 13, + "poly": [ + 995, + 712, + 1050, + 712, + 1050, + 741, + 995, + 741 + ], + "score": 0.85, + "latex": "50 \\%" + }, + { + "category_id": 13, + "poly": [ + 930, + 712, + 985, + 712, + 985, + 741, + 930, + 741 + ], + "score": 0.84, + "latex": "2 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 909, + 1514, + 965, + 1514, + 965, + 1543, + 909, + 1543 + ], + "score": 0.84, + "latex": "2 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 845, + 1514, + 899, + 1514, + 899, + 1543, + 845, + 1543 + ], + "score": 0.84, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 867, + 712, + 920, + 712, + 920, + 741, + 867, + 741 + ], + "score": 0.84, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 969, + 1774, + 1044, + 1774, + 1044, + 1803, + 969, + 1803 + ], + "score": 0.83, + "latex": "9 5 . 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 814, + 712, + 855, + 712, + 855, + 741, + 814, + 741 + ], + "score": 0.82, + "latex": "5 \\%" + }, + { + "category_id": 13, + "poly": [ + 713, + 712, + 753, + 712, + 753, + 741, + 713, + 741 + ], + "score": 0.82, + "latex": "1 \\%" + }, + { + "category_id": 13, + "poly": [ + 762, + 712, + 803, + 712, + 803, + 741, + 762, + 741 + ], + "score": 0.79, + "latex": "2 \\%" + }, + { + "category_id": 13, + "poly": [ + 1028, + 901, + 1118, + 901, + 1118, + 930, + 1028, + 930 + ], + "score": 0.75, + "latex": "\\mathrm { P G + E E }" + }, + { + "category_id": 13, + "poly": [ + 1243, + 712, + 1304, + 712, + 1304, + 741, + 1243, + 741 + ], + "score": 0.68, + "latex": "7 0 0 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 570, + 1574, + 660, + 1574, + 660, + 1603, + 570, + 1603 + ], + "score": 0.67, + "latex": "\\mathrm { P G + E E }" + }, + { + "category_id": 13, + "poly": [ + 883, + 1682, + 944, + 1682, + 944, + 1711, + 883, + 1711 + ], + "score": 0.63, + "latex": "3 5 0 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 1312, + 651, + 1401, + 651, + 1401, + 680, + 1312, + 680 + ], + "score": 0.54, + "latex": "\\mathrm { P G + E E }" + }, + { + "category_id": 13, + "poly": [ + 1312, + 620, + 1401, + 620, + 1401, + 649, + 1312, + 649 + ], + "score": 0.53, + "latex": "\\mathrm { P G + E E }" + }, + { + "category_id": 13, + "poly": [ + 911, + 1743, + 970, + 1743, + 970, + 1772, + 911, + 1772 + ], + "score": 0.5, + "latex": "1 7 5 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 1235, + 1514, + 1296, + 1514, + 1296, + 1543, + 1235, + 1543 + ], + "score": 0.44, + "latex": "7 0 0 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 1145, + 1023, + 1160, + 1023, + 1160, + 1040, + 1145, + 1040 + ], + "score": 0.38, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 610, + 1821, + 679, + 1821, + 679, + 1850, + 610, + 1850 + ], + "score": 0.29, + "latex": "- 7 0 \\mathrm { k }" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 245.0, + 719.0, + 245.0, + 719.0, + 268.0, + 617.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 243.0, + 1150.0, + 243.0, + 1150.0, + 267.0, + 904.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 271.0, + 899.0, + 271.0, + 899.0, + 279.0, + 892.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 309.0, + 904.0, + 309.0, + 904.0, + 326.0, + 880.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 316.0, + 540.0, + 316.0, + 540.0, + 334.0, + 516.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 355.0, + 542.0, + 355.0, + 542.0, + 446.0, + 499.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 351.0, + 904.0, + 351.0, + 904.0, + 446.0, + 862.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 392.0, + 903.0, + 392.0, + 903.0, + 412.0, + 880.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 418.0, + 540.0, + 418.0, + 540.0, + 437.0, + 515.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 444.0, + 774.0, + 444.0, + 774.0, + 464.0, + 728.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 436.0, + 905.0, + 436.0, + 905.0, + 455.0, + 881.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 443.0, + 1127.0, + 443.0, + 1127.0, + 463.0, + 1081.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 469.0, + 541.0, + 469.0, + 541.0, + 489.0, + 516.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 460.0, + 797.0, + 460.0, + 797.0, + 485.0, + 727.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 457.0, + 1150.0, + 457.0, + 1150.0, + 483.0, + 1079.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 479.0, + 774.0, + 479.0, + 774.0, + 502.0, + 727.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 478.0, + 905.0, + 478.0, + 905.0, + 497.0, + 880.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 477.0, + 1129.0, + 477.0, + 1129.0, + 504.0, + 1080.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 498.0, + 763.0, + 498.0, + 763.0, + 522.0, + 727.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 498.0, + 1116.0, + 498.0, + 1116.0, + 520.0, + 1081.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 522.0, + 540.0, + 522.0, + 540.0, + 539.0, + 516.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 522.0, + 904.0, + 522.0, + 904.0, + 539.0, + 880.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 536.0, + 552.0, + 536.0, + 552.0, + 549.0, + 540.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.0, + 535.0, + 582.0, + 535.0, + 582.0, + 551.0, + 566.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 535.0, + 610.0, + 535.0, + 610.0, + 551.0, + 594.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 536.0, + 637.0, + 536.0, + 637.0, + 548.0, + 625.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 535.0, + 668.0, + 535.0, + 668.0, + 551.0, + 652.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 534.0, + 700.0, + 534.0, + 700.0, + 552.0, + 676.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 534.0, + 728.0, + 534.0, + 728.0, + 552.0, + 705.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 534.0, + 757.0, + 534.0, + 757.0, + 552.0, + 733.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 535.0, + 782.0, + 535.0, + 782.0, + 551.0, + 761.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 535.0, + 810.0, + 535.0, + 810.0, + 551.0, + 791.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 535.0, + 914.0, + 535.0, + 914.0, + 548.0, + 902.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 534.0, + 974.0, + 534.0, + 974.0, + 549.0, + 953.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 533.0, + 1031.0, + 533.0, + 1031.0, + 550.0, + 1008.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 533.0, + 1087.0, + 533.0, + 1087.0, + 551.0, + 1063.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 533.0, + 1142.0, + 533.0, + 1142.0, + 551.0, + 1119.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 548.0, + 703.0, + 548.0, + 703.0, + 571.0, + 657.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 548.0, + 1118.0, + 548.0, + 1118.0, + 567.0, + 976.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 621.0, + 1311.0, + 621.0, + 1311.0, + 654.0, + 295.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 621.0, + 1405.0, + 621.0, + 1405.0, + 654.0, + 1402.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 651.0, + 1311.0, + 651.0, + 1311.0, + 684.0, + 296.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 651.0, + 1405.0, + 651.0, + 1405.0, + 684.0, + 1402.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 681.0, + 1403.0, + 681.0, + 1403.0, + 718.0, + 294.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 709.0, + 712.0, + 709.0, + 712.0, + 750.0, + 293.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 709.0, + 761.0, + 709.0, + 761.0, + 750.0, + 754.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 709.0, + 813.0, + 709.0, + 813.0, + 750.0, + 804.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 709.0, + 866.0, + 709.0, + 866.0, + 750.0, + 856.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 709.0, + 929.0, + 709.0, + 929.0, + 750.0, + 921.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 709.0, + 994.0, + 709.0, + 994.0, + 750.0, + 986.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 709.0, + 1102.0, + 709.0, + 1102.0, + 750.0, + 1051.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 709.0, + 1242.0, + 709.0, + 1242.0, + 750.0, + 1170.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 709.0, + 1406.0, + 709.0, + 1406.0, + 750.0, + 1305.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 741.0, + 402.0, + 741.0, + 402.0, + 779.0, + 293.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1363.0, + 585.0, + 1363.0, + 585.0, + 1402.0, + 294.0, + 1402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 868.0, + 1407.0, + 868.0, + 1407.0, + 904.0, + 294.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 899.0, + 1027.0, + 899.0, + 1027.0, + 934.0, + 293.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 899.0, + 1407.0, + 899.0, + 1407.0, + 934.0, + 1119.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 930.0, + 1407.0, + 930.0, + 1407.0, + 966.0, + 293.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 961.0, + 490.0, + 961.0, + 490.0, + 991.0, + 297.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1822.0, + 428.0, + 1822.0, + 428.0, + 1853.0, + 296.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 1822.0, + 609.0, + 1822.0, + 609.0, + 1853.0, + 484.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 1822.0, + 1404.0, + 1822.0, + 1404.0, + 1853.0, + 680.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1850.0, + 790.0, + 1850.0, + 790.0, + 1884.0, + 292.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 1850.0, + 1404.0, + 1850.0, + 1404.0, + 1884.0, + 866.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1881.0, + 410.0, + 1881.0, + 410.0, + 1916.0, + 295.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1916.0, + 570.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1908.0, + 1193.0, + 1908.0, + 1193.0, + 1952.0, + 290.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 1908.0, + 1408.0, + 1908.0, + 1408.0, + 1952.0, + 1269.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 1407.0, + 1943.0, + 1407.0, + 1978.0, + 295.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2007.0, + 292.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1345.0, + 2002.0, + 1345.0, + 2036.0, + 293.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1453.0, + 1405.0, + 1453.0, + 1405.0, + 1485.0, + 297.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1481.0, + 1406.0, + 1481.0, + 1406.0, + 1517.0, + 293.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1514.0, + 844.0, + 1514.0, + 844.0, + 1546.0, + 296.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1514.0, + 908.0, + 1514.0, + 908.0, + 1546.0, + 900.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 1514.0, + 1016.0, + 1514.0, + 1016.0, + 1546.0, + 966.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1514.0, + 1234.0, + 1514.0, + 1234.0, + 1546.0, + 1073.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 1514.0, + 1405.0, + 1514.0, + 1405.0, + 1546.0, + 1297.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1542.0, + 1404.0, + 1542.0, + 1404.0, + 1578.0, + 293.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1573.0, + 569.0, + 1573.0, + 569.0, + 1607.0, + 293.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1573.0, + 1405.0, + 1573.0, + 1405.0, + 1607.0, + 661.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1603.0, + 644.0, + 1603.0, + 644.0, + 1639.0, + 293.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1652.0, + 1401.0, + 1652.0, + 1401.0, + 1681.0, + 297.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1680.0, + 537.0, + 1680.0, + 537.0, + 1716.0, + 295.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 1680.0, + 882.0, + 1680.0, + 882.0, + 1716.0, + 594.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 1680.0, + 1404.0, + 1680.0, + 1404.0, + 1716.0, + 945.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1709.0, + 489.0, + 1709.0, + 489.0, + 1749.0, + 290.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 1709.0, + 598.0, + 1709.0, + 598.0, + 1749.0, + 546.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1709.0, + 960.0, + 1709.0, + 960.0, + 1749.0, + 654.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 1709.0, + 1406.0, + 1709.0, + 1406.0, + 1749.0, + 1037.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1739.0, + 570.0, + 1739.0, + 570.0, + 1778.0, + 294.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 1739.0, + 910.0, + 1739.0, + 910.0, + 1778.0, + 627.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 1739.0, + 1405.0, + 1739.0, + 1405.0, + 1778.0, + 971.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1769.0, + 420.0, + 1769.0, + 420.0, + 1811.0, + 291.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 1769.0, + 525.0, + 1769.0, + 525.0, + 1811.0, + 477.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 1769.0, + 968.0, + 1769.0, + 968.0, + 1811.0, + 581.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1769.0, + 1156.0, + 1769.0, + 1156.0, + 1811.0, + 1045.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 868.0, + 1407.0, + 868.0, + 1407.0, + 904.0, + 294.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 899.0, + 1027.0, + 899.0, + 1027.0, + 934.0, + 293.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 899.0, + 1407.0, + 899.0, + 1407.0, + 934.0, + 1119.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 930.0, + 1407.0, + 930.0, + 1407.0, + 966.0, + 293.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 961.0, + 490.0, + 961.0, + 490.0, + 991.0, + 297.0, + 991.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1819, + 1403, + 1819, + 1403, + 2034, + 297, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 677, + 1403, + 677, + 1403, + 831, + 298, + 831 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 399, + 1403, + 399, + 1403, + 553, + 298, + 553 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 569, + 1403, + 569, + 1403, + 661, + 298, + 661 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 301, + 967, + 1393, + 967, + 1393, + 1693, + 301, + 1693 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 298, + 291, + 1403, + 291, + 1403, + 384, + 298, + 384 + ], + "score": 0.971 + }, + { + "category_id": 4, + "poly": [ + 404, + 1736, + 1290, + 1736, + 1290, + 1768, + 404, + 1768 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.886 + }, + { + "category_id": 0, + "poly": [ + 299, + 883, + 506, + 883, + 506, + 910, + 299, + 910 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 862, + 2088, + 862, + 2112, + 836, + 2112 + ], + "score": 0.84 + }, + { + "category_id": 1, + "poly": [ + 300, + 231, + 1009, + 231, + 1009, + 261, + 300, + 261 + ], + "score": 0.245 + }, + { + "category_id": 0, + "poly": [ + 300, + 231, + 1009, + 231, + 1009, + 261, + 300, + 261 + ], + "score": 0.224 + }, + { + "category_id": 13, + "poly": [ + 804, + 1882, + 858, + 1882, + 858, + 1911, + 804, + 1911 + ], + "score": 0.87, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 507, + 631, + 581, + 631, + 581, + 660, + 507, + 660 + ], + "score": 0.85, + "latex": "8 2 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1339, + 599, + 1401, + 599, + 1401, + 630, + 1339, + 630 + ], + "score": 0.79, + "latex": "6 . 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 913, + 401, + 959, + 401, + 959, + 430, + 913, + 430 + ], + "score": 0.33, + "latex": "1 8 \\mathrm { k }" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 975.0, + 483.0, + 975.0, + 483.0, + 994.0, + 390.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 976.0, + 754.0, + 976.0, + 754.0, + 996.0, + 660.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 976.0, + 1039.0, + 976.0, + 1039.0, + 996.0, + 925.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 973.0, + 1299.0, + 973.0, + 1299.0, + 997.0, + 1209.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 992.0, + 344.0, + 992.0, + 344.0, + 1003.0, + 333.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 992.0, + 615.0, + 992.0, + 615.0, + 1004.0, + 604.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 991.0, + 889.0, + 991.0, + 889.0, + 1007.0, + 873.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 993.0, + 1157.0, + 993.0, + 1157.0, + 1003.0, + 1147.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1018.0, + 347.0, + 1018.0, + 347.0, + 1036.0, + 321.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 1010.0, + 617.0, + 1010.0, + 617.0, + 1029.0, + 592.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 1012.0, + 889.0, + 1012.0, + 889.0, + 1030.0, + 862.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1033.0, + 617.0, + 1033.0, + 617.0, + 1051.0, + 596.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1034.0, + 889.0, + 1034.0, + 889.0, + 1052.0, + 867.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 1034.0, + 1160.0, + 1034.0, + 1160.0, + 1052.0, + 1135.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 1049.0, + 347.0, + 1049.0, + 347.0, + 1130.0, + 302.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1053.0, + 617.0, + 1053.0, + 617.0, + 1130.0, + 573.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1055.0, + 888.0, + 1055.0, + 888.0, + 1131.0, + 846.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 1055.0, + 1137.0, + 1055.0, + 1137.0, + 1128.0, + 1116.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1078.0, + 345.0, + 1078.0, + 345.0, + 1096.0, + 321.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1077.0, + 617.0, + 1077.0, + 617.0, + 1099.0, + 596.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1078.0, + 889.0, + 1078.0, + 889.0, + 1098.0, + 868.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 1079.0, + 1159.0, + 1079.0, + 1159.0, + 1098.0, + 1138.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1109.0, + 345.0, + 1109.0, + 345.0, + 1126.0, + 321.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 1099.0, + 617.0, + 1099.0, + 617.0, + 1119.0, + 592.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1101.0, + 888.0, + 1101.0, + 888.0, + 1120.0, + 864.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1122.0, + 617.0, + 1122.0, + 617.0, + 1141.0, + 596.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1123.0, + 889.0, + 1123.0, + 889.0, + 1142.0, + 868.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 1117.0, + 1094.0, + 1117.0, + 1094.0, + 1135.0, + 1061.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 1123.0, + 1160.0, + 1123.0, + 1160.0, + 1141.0, + 1135.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1138.0, + 345.0, + 1138.0, + 345.0, + 1155.0, + 321.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 1160.0, + 566.0, + 1160.0, + 566.0, + 1176.0, + 507.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1144.0, + 617.0, + 1144.0, + 617.0, + 1163.0, + 591.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 1151.0, + 840.0, + 1151.0, + 840.0, + 1176.0, + 777.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 1147.0, + 889.0, + 1147.0, + 889.0, + 1165.0, + 863.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 1135.0, + 1095.0, + 1135.0, + 1095.0, + 1170.0, + 1063.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1299.0, + 1139.0, + 1365.0, + 1139.0, + 1365.0, + 1170.0, + 1299.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1169.0, + 347.0, + 1169.0, + 347.0, + 1186.0, + 321.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1170.0, + 617.0, + 1170.0, + 617.0, + 1185.0, + 596.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1170.0, + 888.0, + 1170.0, + 888.0, + 1185.0, + 868.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 1170.0, + 1159.0, + 1170.0, + 1159.0, + 1185.0, + 1139.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 1182.0, + 354.0, + 1182.0, + 354.0, + 1195.0, + 343.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1184.0, + 407.0, + 1184.0, + 407.0, + 1195.0, + 398.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1181.0, + 468.0, + 1181.0, + 468.0, + 1196.0, + 450.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 1181.0, + 522.0, + 1181.0, + 522.0, + 1196.0, + 504.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 1181.0, + 576.0, + 1181.0, + 576.0, + 1196.0, + 560.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 1180.0, + 627.0, + 1180.0, + 627.0, + 1196.0, + 612.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1182.0, + 681.0, + 1182.0, + 681.0, + 1195.0, + 669.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 1179.0, + 740.0, + 1179.0, + 740.0, + 1197.0, + 719.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 1180.0, + 794.0, + 1180.0, + 794.0, + 1196.0, + 776.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 1180.0, + 848.0, + 1180.0, + 848.0, + 1196.0, + 831.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1182.0, + 897.0, + 1182.0, + 897.0, + 1195.0, + 887.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1182.0, + 951.0, + 1182.0, + 951.0, + 1195.0, + 941.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 1181.0, + 1010.0, + 1181.0, + 1010.0, + 1196.0, + 992.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 1181.0, + 1064.0, + 1181.0, + 1064.0, + 1197.0, + 1047.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 1181.0, + 1118.0, + 1181.0, + 1118.0, + 1196.0, + 1102.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1182.0, + 1167.0, + 1182.0, + 1167.0, + 1195.0, + 1157.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 1184.0, + 1222.0, + 1184.0, + 1222.0, + 1195.0, + 1212.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 1180.0, + 1280.0, + 1180.0, + 1280.0, + 1196.0, + 1263.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 1180.0, + 1335.0, + 1180.0, + 1335.0, + 1196.0, + 1318.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 1181.0, + 1390.0, + 1181.0, + 1390.0, + 1196.0, + 1373.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 1192.0, + 468.0, + 1192.0, + 468.0, + 1213.0, + 429.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1191.0, + 739.0, + 1191.0, + 739.0, + 1212.0, + 699.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 1192.0, + 1011.0, + 1192.0, + 1011.0, + 1213.0, + 971.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 1192.0, + 1282.0, + 1192.0, + 1282.0, + 1213.0, + 1242.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1216.0, + 520.0, + 1216.0, + 520.0, + 1235.0, + 393.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1216.0, + 789.0, + 1216.0, + 789.0, + 1235.0, + 662.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1215.0, + 1022.0, + 1215.0, + 1022.0, + 1234.0, + 944.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 1214.0, + 1292.0, + 1214.0, + 1292.0, + 1234.0, + 1217.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 1229.0, + 348.0, + 1229.0, + 348.0, + 1245.0, + 334.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1229.0, + 617.0, + 1229.0, + 617.0, + 1244.0, + 602.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1229.0, + 889.0, + 1229.0, + 889.0, + 1245.0, + 876.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1232.0, + 1157.0, + 1232.0, + 1157.0, + 1243.0, + 1147.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 1257.0, + 1160.0, + 1257.0, + 1160.0, + 1276.0, + 1139.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1272.0, + 890.0, + 1272.0, + 890.0, + 1291.0, + 868.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1288.0, + 348.0, + 1288.0, + 348.0, + 1369.0, + 304.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1293.0, + 616.0, + 1293.0, + 616.0, + 1368.0, + 573.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1292.0, + 867.0, + 1292.0, + 867.0, + 1367.0, + 846.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1316.0, + 889.0, + 1316.0, + 889.0, + 1335.0, + 868.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 1288.0, + 1160.0, + 1288.0, + 1160.0, + 1367.0, + 1119.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 1323.0, + 1351.0, + 1323.0, + 1351.0, + 1335.0, + 1278.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 1332.0, + 1373.0, + 1332.0, + 1373.0, + 1348.0, + 1278.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1347.0, + 349.0, + 1347.0, + 349.0, + 1366.0, + 322.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 1347.0, + 1160.0, + 1347.0, + 1160.0, + 1366.0, + 1139.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 1344.0, + 1349.0, + 1344.0, + 1349.0, + 1359.0, + 1274.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 1362.0, + 554.0, + 1362.0, + 554.0, + 1378.0, + 479.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 1362.0, + 815.0, + 1362.0, + 815.0, + 1374.0, + 739.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1362.0, + 889.0, + 1362.0, + 889.0, + 1380.0, + 868.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 1362.0, + 1078.0, + 1362.0, + 1078.0, + 1374.0, + 1003.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1275.0, + 1357.0, + 1324.0, + 1357.0, + 1324.0, + 1373.0, + 1275.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 1372.0, + 532.0, + 1372.0, + 532.0, + 1403.0, + 478.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1373.0, + 816.0, + 1373.0, + 816.0, + 1401.0, + 740.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 1373.0, + 1100.0, + 1373.0, + 1100.0, + 1399.0, + 1003.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 1378.0, + 1160.0, + 1378.0, + 1160.0, + 1396.0, + 1138.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 1398.0, + 559.0, + 1398.0, + 559.0, + 1414.0, + 483.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 1398.0, + 785.0, + 1398.0, + 785.0, + 1414.0, + 743.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 1398.0, + 1079.0, + 1398.0, + 1079.0, + 1414.0, + 1005.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1409.0, + 349.0, + 1409.0, + 349.0, + 1426.0, + 323.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1411.0, + 617.0, + 1411.0, + 617.0, + 1426.0, + 597.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1406.0, + 899.0, + 1406.0, + 899.0, + 1425.0, + 868.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 1410.0, + 1159.0, + 1410.0, + 1159.0, + 1425.0, + 1139.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 1421.0, + 358.0, + 1421.0, + 358.0, + 1437.0, + 344.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 1421.0, + 414.0, + 1421.0, + 414.0, + 1436.0, + 399.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1420.0, + 472.0, + 1420.0, + 472.0, + 1438.0, + 451.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 1419.0, + 527.0, + 1419.0, + 527.0, + 1438.0, + 506.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 1421.0, + 579.0, + 1421.0, + 579.0, + 1436.0, + 561.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 1422.0, + 627.0, + 1422.0, + 627.0, + 1437.0, + 612.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1423.0, + 681.0, + 1423.0, + 681.0, + 1436.0, + 669.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 1420.0, + 740.0, + 1420.0, + 740.0, + 1438.0, + 719.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 1420.0, + 795.0, + 1420.0, + 795.0, + 1438.0, + 774.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 1421.0, + 848.0, + 1421.0, + 848.0, + 1437.0, + 831.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1421.0, + 898.0, + 1421.0, + 898.0, + 1433.0, + 887.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1421.0, + 951.0, + 1421.0, + 951.0, + 1432.0, + 941.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 1419.0, + 1010.0, + 1419.0, + 1010.0, + 1434.0, + 992.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 1419.0, + 1064.0, + 1419.0, + 1064.0, + 1434.0, + 1047.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 1419.0, + 1119.0, + 1419.0, + 1119.0, + 1434.0, + 1102.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1422.0, + 1169.0, + 1422.0, + 1169.0, + 1434.0, + 1157.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 1422.0, + 1222.0, + 1422.0, + 1222.0, + 1434.0, + 1212.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1419.0, + 1283.0, + 1419.0, + 1283.0, + 1437.0, + 1262.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 1421.0, + 1335.0, + 1421.0, + 1335.0, + 1436.0, + 1318.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1420.0, + 1390.0, + 1420.0, + 1390.0, + 1436.0, + 1372.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 431.0, + 1432.0, + 471.0, + 1432.0, + 471.0, + 1453.0, + 431.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1433.0, + 739.0, + 1433.0, + 739.0, + 1454.0, + 699.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 1431.0, + 1011.0, + 1431.0, + 1011.0, + 1452.0, + 971.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 1432.0, + 1282.0, + 1432.0, + 1282.0, + 1453.0, + 1242.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1453.0, + 791.0, + 1453.0, + 791.0, + 1474.0, + 661.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1450.0, + 1011.0, + 1450.0, + 1011.0, + 1474.0, + 956.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 1454.0, + 1290.0, + 1454.0, + 1290.0, + 1474.0, + 1218.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1470.0, + 616.0, + 1470.0, + 616.0, + 1480.0, + 606.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1470.0, + 885.0, + 1470.0, + 885.0, + 1480.0, + 877.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1470.0, + 1157.0, + 1470.0, + 1157.0, + 1481.0, + 1147.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 1512.0, + 1160.0, + 1512.0, + 1160.0, + 1529.0, + 1135.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1532.0, + 619.0, + 1532.0, + 619.0, + 1608.0, + 575.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1532.0, + 889.0, + 1532.0, + 889.0, + 1608.0, + 845.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 1533.0, + 1137.0, + 1533.0, + 1137.0, + 1606.0, + 1116.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 1556.0, + 1160.0, + 1556.0, + 1160.0, + 1576.0, + 1139.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 1604.0, + 1098.0, + 1604.0, + 1098.0, + 1620.0, + 1016.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 1603.0, + 1160.0, + 1603.0, + 1160.0, + 1620.0, + 1135.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1619.0, + 1056.0, + 1619.0, + 1056.0, + 1635.0, + 1012.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 1618.0, + 1288.0, + 1618.0, + 1288.0, + 1634.0, + 1215.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 1636.0, + 825.0, + 1636.0, + 825.0, + 1646.0, + 810.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 1634.0, + 1077.0, + 1634.0, + 1077.0, + 1650.0, + 1018.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1635.0, + 1375.0, + 1635.0, + 1375.0, + 1647.0, + 1216.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 1648.0, + 619.0, + 1648.0, + 619.0, + 1663.0, + 599.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1647.0, + 889.0, + 1647.0, + 889.0, + 1666.0, + 867.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 1648.0, + 1160.0, + 1648.0, + 1160.0, + 1663.0, + 1140.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 1661.0, + 627.0, + 1661.0, + 627.0, + 1673.0, + 616.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1659.0, + 684.0, + 1659.0, + 684.0, + 1674.0, + 669.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 1657.0, + 741.0, + 1657.0, + 741.0, + 1677.0, + 722.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 1659.0, + 795.0, + 1659.0, + 795.0, + 1675.0, + 777.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 1659.0, + 849.0, + 1659.0, + 849.0, + 1675.0, + 833.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 1659.0, + 898.0, + 1659.0, + 898.0, + 1674.0, + 883.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 1659.0, + 953.0, + 1659.0, + 953.0, + 1674.0, + 938.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 1657.0, + 1011.0, + 1657.0, + 1011.0, + 1677.0, + 991.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 1657.0, + 1066.0, + 1657.0, + 1066.0, + 1677.0, + 1046.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 1659.0, + 1119.0, + 1659.0, + 1119.0, + 1674.0, + 1102.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1661.0, + 1169.0, + 1661.0, + 1169.0, + 1673.0, + 1157.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 1661.0, + 1223.0, + 1661.0, + 1223.0, + 1673.0, + 1212.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1657.0, + 1283.0, + 1657.0, + 1283.0, + 1677.0, + 1262.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 1659.0, + 1335.0, + 1659.0, + 1335.0, + 1675.0, + 1318.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 1659.0, + 1390.0, + 1659.0, + 1390.0, + 1674.0, + 1373.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1670.0, + 740.0, + 1670.0, + 740.0, + 1691.0, + 702.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 1670.0, + 1011.0, + 1670.0, + 1011.0, + 1691.0, + 971.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 1670.0, + 1282.0, + 1670.0, + 1282.0, + 1691.0, + 1242.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 1617.0, + 825.0, + 1617.0, + 825.0, + 1633.0, + 812.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 1733.0, + 1294.0, + 1733.0, + 1294.0, + 1771.0, + 403.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 877.0, + 511.0, + 877.0, + 511.0, + 917.0, + 293.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 227.0, + 1013.0, + 227.0, + 1013.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1822.0, + 1405.0, + 1822.0, + 1405.0, + 1853.0, + 296.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1849.0, + 1405.0, + 1849.0, + 1405.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1880.0, + 803.0, + 1880.0, + 803.0, + 1917.0, + 292.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 1880.0, + 1405.0, + 1880.0, + 1405.0, + 1917.0, + 859.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1946.0, + 293.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1938.0, + 1403.0, + 1938.0, + 1403.0, + 1979.0, + 294.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1973.0, + 1403.0, + 1973.0, + 1403.0, + 2004.0, + 296.0, + 2004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 758.0, + 2003.0, + 758.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 676.0, + 1404.0, + 676.0, + 1404.0, + 712.0, + 294.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 708.0, + 1404.0, + 708.0, + 1404.0, + 741.0, + 296.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 737.0, + 1407.0, + 737.0, + 1407.0, + 775.0, + 291.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 769.0, + 1405.0, + 769.0, + 1405.0, + 806.0, + 294.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 799.0, + 667.0, + 799.0, + 667.0, + 835.0, + 293.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 399.0, + 912.0, + 399.0, + 912.0, + 436.0, + 294.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 399.0, + 1405.0, + 399.0, + 1405.0, + 436.0, + 960.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 430.0, + 1403.0, + 430.0, + 1403.0, + 467.0, + 293.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 456.0, + 1405.0, + 456.0, + 1405.0, + 501.0, + 293.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 491.0, + 1405.0, + 491.0, + 1405.0, + 528.0, + 293.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 524.0, + 619.0, + 524.0, + 619.0, + 557.0, + 294.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 568.0, + 1404.0, + 568.0, + 1404.0, + 602.0, + 294.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 599.0, + 1338.0, + 599.0, + 1338.0, + 633.0, + 294.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 631.0, + 506.0, + 631.0, + 506.0, + 664.0, + 293.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 631.0, + 593.0, + 631.0, + 593.0, + 664.0, + 582.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 290.0, + 1404.0, + 290.0, + 1404.0, + 327.0, + 293.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 324.0, + 1405.0, + 324.0, + 1405.0, + 358.0, + 296.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 354.0, + 1256.0, + 354.0, + 1256.0, + 388.0, + 296.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 227.0, + 1013.0, + 227.0, + 1013.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1438, + 1404, + 1438, + 1404, + 1775, + 297, + 1775 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 839, + 1404, + 839, + 1404, + 1056, + 297, + 1056 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1238, + 1403, + 1238, + 1403, + 1421, + 298, + 1421 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1070, + 1403, + 1070, + 1403, + 1224, + 298, + 1224 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1788, + 1403, + 1788, + 1403, + 2034, + 297, + 2034 + ], + "score": 0.977 + }, + { + "category_id": 5, + "poly": [ + 522, + 311, + 1177, + 311, + 1177, + 738, + 522, + 738 + ], + "score": 0.976, + "html": "
ModelStandard CLEVR10% CLEVR
MAC standard model (dimension 512)98.984.5
dimension 25698.476.3
dimension 12897.677.0
w/o control55.651.5
w/o word-attention95.363.2
full-question-as-control80.765.0
w/o positional encoding99.180.0
linear position encoding99.083.2
identity writing unit98.584.5
only-gating writing unit99.383.1
+ self-attention83.299.2
+ gating99.483.1
+ self-attention and gating99.585.5
gate bias 098.784.9
gate bias 199.468.5
gate bias -199.077.1
unshared weights97.867.5
prediction without question information97.864.7
" + }, + { + "category_id": 6, + "poly": [ + 297, + 223, + 1400, + 223, + 1400, + 284, + 297, + 284 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.849 + }, + { + "category_id": 13, + "poly": [ + 693, + 1270, + 781, + 1270, + 781, + 1302, + 693, + 1302 + ], + "score": 0.91, + "latex": "p = 1 2" + }, + { + "category_id": 13, + "poly": [ + 447, + 1162, + 537, + 1162, + 537, + 1191, + 447, + 1191 + ], + "score": 0.88, + "latex": "9 8 . 9 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 511, + 1682, + 587, + 1682, + 587, + 1711, + 511, + 1711 + ], + "score": 0.87, + "latex": "2 1 . 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 1123, + 962, + 1255, + 962, + 1255, + 992, + 1123, + 992 + ], + "score": 0.87, + "latex": "9 8 . 9 – 9 9 . 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 1090, + 1162, + 1143, + 1162, + 1143, + 1191, + 1090, + 1191 + ], + "score": 0.87, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 479, + 1330, + 554, + 1330, + 554, + 1360, + 479, + 1360 + ], + "score": 0.86, + "latex": "1 6 . 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 1232, + 1973, + 1306, + 1973, + 1306, + 2002, + 1232, + 2002 + ], + "score": 0.86, + "latex": "1 9 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 858, + 1300, + 898, + 1300, + 898, + 1328, + 858, + 1328 + ], + "score": 0.86, + "latex": "1 \\%" + }, + { + "category_id": 13, + "poly": [ + 785, + 1193, + 847, + 1193, + 847, + 1222, + 785, + 1222 + ], + "score": 0.86, + "latex": "7 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1348, + 1974, + 1401, + 1974, + 1401, + 2002, + 1348, + 2002 + ], + "score": 0.85, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 296, + 963, + 351, + 963, + 351, + 992, + 296, + 992 + ], + "score": 0.85, + "latex": "9 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1067, + 225, + 1121, + 225, + 1121, + 254, + 1067, + 254 + ], + "score": 0.85, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 1318, + 1131, + 1393, + 1131, + 1393, + 1162, + 1318, + 1162 + ], + "score": 0.84, + "latex": "9 7 . 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 1376, + 1948, + 1400, + 1948, + 1400, + 1973, + 1376, + 1973 + ], + "score": 0.8, + "latex": "c _ { i }" + }, + { + "category_id": 13, + "poly": [ + 785, + 1973, + 879, + 1973, + 879, + 2002, + 785, + 2002 + ], + "score": 0.8, + "latex": "- 1 9 . 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 1042, + 1948, + 1060, + 1948, + 1060, + 1974, + 1042, + 1974 + ], + "score": 0.77, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 297, + 1918, + 315, + 1918, + 315, + 1944, + 297, + 1944 + ], + "score": 0.74, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 368, + 932, + 423, + 932, + 423, + 961, + 368, + 961 + ], + "score": 0.71, + "latex": "7 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 538, + 561, + 554, + 561, + 554, + 577, + 538, + 577 + ], + "score": 0.63, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 538, + 583, + 554, + 583, + 554, + 600, + 538, + 600 + ], + "score": 0.63, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 538, + 605, + 554, + 605, + 554, + 621, + 538, + 621 + ], + "score": 0.51, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 296, + 932, + 423, + 932, + 423, + 961, + 296, + 961 + ], + "score": 0.4, + "latex": "1 0 \\mathrm { w } - 7 5 \\%" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 222.0, + 1066.0, + 222.0, + 1066.0, + 257.0, + 295.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 222.0, + 1405.0, + 222.0, + 1405.0, + 257.0, + 1122.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 249.0, + 431.0, + 249.0, + 431.0, + 289.0, + 293.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1438.0, + 1404.0, + 1438.0, + 1404.0, + 1472.0, + 296.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1469.0, + 1404.0, + 1469.0, + 1404.0, + 1503.0, + 294.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1500.0, + 1405.0, + 1500.0, + 1405.0, + 1534.0, + 294.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1530.0, + 1405.0, + 1530.0, + 1405.0, + 1564.0, + 295.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1556.0, + 1406.0, + 1556.0, + 1406.0, + 1599.0, + 292.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1593.0, + 1404.0, + 1593.0, + 1404.0, + 1624.0, + 296.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1621.0, + 1405.0, + 1621.0, + 1405.0, + 1656.0, + 294.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1651.0, + 1404.0, + 1651.0, + 1404.0, + 1685.0, + 294.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1680.0, + 510.0, + 1680.0, + 510.0, + 1716.0, + 292.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 1680.0, + 1402.0, + 1680.0, + 1402.0, + 1716.0, + 588.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1711.0, + 1404.0, + 1711.0, + 1404.0, + 1748.0, + 294.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1744.0, + 480.0, + 1744.0, + 480.0, + 1777.0, + 296.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 840.0, + 1404.0, + 840.0, + 1404.0, + 874.0, + 295.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 869.0, + 1405.0, + 869.0, + 1405.0, + 905.0, + 294.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 899.0, + 1404.0, + 899.0, + 1404.0, + 936.0, + 292.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 931.0, + 1406.0, + 931.0, + 1406.0, + 966.0, + 424.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 960.0, + 1122.0, + 960.0, + 1122.0, + 998.0, + 352.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 960.0, + 1405.0, + 960.0, + 1405.0, + 998.0, + 1256.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 991.0, + 1407.0, + 991.0, + 1407.0, + 1029.0, + 294.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1022.0, + 910.0, + 1022.0, + 910.0, + 1060.0, + 292.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1240.0, + 1404.0, + 1240.0, + 1404.0, + 1271.0, + 297.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1269.0, + 692.0, + 1269.0, + 692.0, + 1303.0, + 293.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 1269.0, + 1407.0, + 1269.0, + 1407.0, + 1303.0, + 782.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1298.0, + 857.0, + 1298.0, + 857.0, + 1335.0, + 293.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1298.0, + 1407.0, + 1298.0, + 1407.0, + 1335.0, + 899.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1331.0, + 478.0, + 1331.0, + 478.0, + 1366.0, + 294.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 1331.0, + 1405.0, + 1331.0, + 1405.0, + 1366.0, + 555.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1359.0, + 1405.0, + 1359.0, + 1405.0, + 1395.0, + 292.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1387.0, + 399.0, + 1387.0, + 399.0, + 1425.0, + 293.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1070.0, + 1405.0, + 1070.0, + 1405.0, + 1103.0, + 296.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1099.0, + 1407.0, + 1099.0, + 1407.0, + 1135.0, + 293.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1130.0, + 1317.0, + 1130.0, + 1317.0, + 1167.0, + 294.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1130.0, + 1405.0, + 1130.0, + 1405.0, + 1167.0, + 1394.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1158.0, + 446.0, + 1158.0, + 446.0, + 1200.0, + 292.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 1158.0, + 1089.0, + 1158.0, + 1089.0, + 1200.0, + 538.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 1158.0, + 1407.0, + 1158.0, + 1407.0, + 1200.0, + 1144.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1191.0, + 784.0, + 1191.0, + 784.0, + 1228.0, + 294.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 1191.0, + 1113.0, + 1191.0, + 1113.0, + 1228.0, + 848.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1787.0, + 1405.0, + 1787.0, + 1405.0, + 1822.0, + 294.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1820.0, + 1405.0, + 1820.0, + 1405.0, + 1853.0, + 295.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1851.0, + 1405.0, + 1851.0, + 1405.0, + 1883.0, + 293.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1916.0, + 293.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1913.0, + 296.0, + 1913.0, + 296.0, + 1945.0, + 292.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1913.0, + 1405.0, + 1913.0, + 1405.0, + 1945.0, + 316.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1940.0, + 1041.0, + 1940.0, + 1041.0, + 1980.0, + 291.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 1940.0, + 1375.0, + 1940.0, + 1375.0, + 1980.0, + 1061.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1940.0, + 1407.0, + 1940.0, + 1407.0, + 1980.0, + 1401.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1973.0, + 784.0, + 1973.0, + 784.0, + 2005.0, + 292.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 1973.0, + 1231.0, + 1973.0, + 1231.0, + 2005.0, + 880.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 1973.0, + 1347.0, + 1973.0, + 1347.0, + 2005.0, + 1307.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2000.0, + 428.0, + 2000.0, + 428.0, + 2035.0, + 294.0, + 2035.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 459, + 1402, + 459, + 1402, + 704, + 299, + 704 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 298, + 950, + 1404, + 950, + 1404, + 1315, + 298, + 1315 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1331, + 1403, + 1331, + 1403, + 1546, + 297, + 1546 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 719, + 1402, + 719, + 1402, + 934, + 298, + 934 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1403, + 229, + 1403, + 444, + 298, + 444 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 300, + 1942, + 1401, + 1942, + 1401, + 2034, + 300, + 2034 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 1561, + 1403, + 1561, + 1403, + 1807, + 298, + 1807 + ], + "score": 0.963 + }, + { + "category_id": 0, + "poly": [ + 300, + 1874, + 594, + 1874, + 594, + 1903, + 300, + 1903 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 864, + 2088, + 864, + 2112, + 837, + 2112 + ], + "score": 0.843 + }, + { + "category_id": 13, + "poly": [ + 662, + 1423, + 741, + 1423, + 741, + 1457, + 662, + 1457 + ], + "score": 0.9, + "latex": "[ - 1 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 851, + 496, + 916, + 496, + 916, + 523, + 851, + 523 + ], + "score": 0.89, + "latex": "m _ { i - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1130, + 786, + 1199, + 786, + 1199, + 813, + 1130, + 813 + ], + "score": 0.89, + "latex": "m _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 1257, + 526, + 1327, + 526, + 1327, + 553, + 1257, + 553 + ], + "score": 0.88, + "latex": "m _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 415, + 495, + 484, + 495, + 484, + 522, + 415, + 522 + ], + "score": 0.88, + "latex": "m _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 1328, + 297, + 1393, + 297, + 1393, + 323, + 1328, + 323 + ], + "score": 0.88, + "latex": "m _ { i - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1213, + 557, + 1283, + 557, + 1283, + 582, + 1213, + 582 + ], + "score": 0.88, + "latex": "m _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 626, + 1072, + 681, + 1072, + 681, + 1101, + 626, + 1101 + ], + "score": 0.87, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 1195, + 1454, + 1248, + 1454, + 1248, + 1483, + 1195, + 1483 + ], + "score": 0.87, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 512, + 986, + 552, + 986, + 552, + 1014, + 512, + 1014 + ], + "score": 0.87, + "latex": "m _ { p }" + }, + { + "category_id": 13, + "poly": [ + 297, + 873, + 386, + 873, + 386, + 902, + 297, + 902 + ], + "score": 0.87, + "latex": "9 8 . 9 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 1262, + 873, + 1352, + 873, + 1352, + 902, + 1262, + 902 + ], + "score": 0.87, + "latex": "9 9 . 3 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 1007, + 873, + 1095, + 873, + 1095, + 903, + 1007, + 903 + ], + "score": 0.87, + "latex": "9 9 . 2 3 \\%" + }, + { + "category_id": 13, + "poly": [ + 718, + 352, + 793, + 352, + 793, + 383, + 718, + 383 + ], + "score": 0.87, + "latex": "0 . 0 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 536, + 904, + 625, + 904, + 625, + 933, + 536, + 933 + ], + "score": 0.87, + "latex": "9 9 . 4 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 1274, + 353, + 1328, + 353, + 1328, + 382, + 1274, + 382 + ], + "score": 0.86, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 397, + 817, + 433, + 817, + 433, + 842, + 397, + 842 + ], + "score": 0.86, + "latex": "m _ { i }" + }, + { + "category_id": 13, + "poly": [ + 645, + 1133, + 699, + 1133, + 699, + 1162, + 645, + 1162 + ], + "score": 0.85, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 1063, + 1594, + 1128, + 1594, + 1128, + 1623, + 1063, + 1623 + ], + "score": 0.85, + "latex": "- 1 , 0" + }, + { + "category_id": 13, + "poly": [ + 545, + 1714, + 599, + 1714, + 599, + 1743, + 545, + 1743 + ], + "score": 0.85, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 298, + 413, + 372, + 413, + 372, + 443, + 298, + 443 + ], + "score": 0.83, + "latex": "1 1 . 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 1363, + 527, + 1400, + 527, + 1400, + 552, + 1363, + 552 + ], + "score": 0.83, + "latex": "m _ { i }" + }, + { + "category_id": 13, + "poly": [ + 819, + 1133, + 893, + 1133, + 893, + 1162, + 819, + 1162 + ], + "score": 0.82, + "latex": "1 9 . 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 937, + 986, + 955, + 986, + 955, + 1013, + 937, + 1013 + ], + "score": 0.77, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 1232, + 584, + 1293, + 584, + 1293, + 612, + 1232, + 612 + ], + "score": 0.73, + "latex": "0 . 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 1246, + 1594, + 1285, + 1594, + 1285, + 1621, + 1246, + 1621 + ], + "score": 0.66, + "latex": "- 1" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1873.0, + 599.0, + 1873.0, + 599.0, + 1906.0, + 296.0, + 1906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2086.0, + 868.0, + 2086.0, + 868.0, + 2121.0, + 830.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 459.0, + 1403.0, + 459.0, + 1403.0, + 495.0, + 295.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 490.0, + 414.0, + 490.0, + 414.0, + 528.0, + 293.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 490.0, + 850.0, + 490.0, + 850.0, + 528.0, + 485.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 490.0, + 1407.0, + 490.0, + 1407.0, + 528.0, + 917.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 520.0, + 1256.0, + 520.0, + 1256.0, + 559.0, + 294.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 520.0, + 1362.0, + 520.0, + 1362.0, + 559.0, + 1328.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 520.0, + 1406.0, + 520.0, + 1406.0, + 559.0, + 1401.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 553.0, + 1212.0, + 553.0, + 1212.0, + 586.0, + 295.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 553.0, + 1406.0, + 553.0, + 1406.0, + 586.0, + 1284.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 579.0, + 1231.0, + 579.0, + 1231.0, + 616.0, + 294.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 579.0, + 1406.0, + 579.0, + 1406.0, + 616.0, + 1294.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 611.0, + 1403.0, + 611.0, + 1403.0, + 649.0, + 293.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 644.0, + 1406.0, + 644.0, + 1406.0, + 677.0, + 293.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 675.0, + 944.0, + 675.0, + 944.0, + 705.0, + 294.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 951.0, + 1405.0, + 951.0, + 1405.0, + 985.0, + 296.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 981.0, + 511.0, + 981.0, + 511.0, + 1015.0, + 294.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 981.0, + 936.0, + 981.0, + 936.0, + 1015.0, + 553.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 981.0, + 1405.0, + 981.0, + 1405.0, + 1015.0, + 956.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1011.0, + 1404.0, + 1011.0, + 1404.0, + 1045.0, + 294.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1039.0, + 1405.0, + 1039.0, + 1405.0, + 1079.0, + 292.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1072.0, + 625.0, + 1072.0, + 625.0, + 1106.0, + 294.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 1072.0, + 1405.0, + 1072.0, + 1405.0, + 1106.0, + 682.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1103.0, + 1405.0, + 1103.0, + 1405.0, + 1137.0, + 294.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1134.0, + 644.0, + 1134.0, + 644.0, + 1168.0, + 294.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 1134.0, + 818.0, + 1134.0, + 818.0, + 1168.0, + 700.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1134.0, + 1405.0, + 1134.0, + 1405.0, + 1168.0, + 894.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1164.0, + 1404.0, + 1164.0, + 1404.0, + 1198.0, + 294.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1191.0, + 1405.0, + 1191.0, + 1405.0, + 1232.0, + 292.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1221.0, + 1405.0, + 1221.0, + 1405.0, + 1262.0, + 292.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1254.0, + 1406.0, + 1254.0, + 1406.0, + 1291.0, + 292.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1286.0, + 568.0, + 1286.0, + 568.0, + 1319.0, + 293.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1329.0, + 1405.0, + 1329.0, + 1405.0, + 1366.0, + 294.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1361.0, + 1405.0, + 1361.0, + 1405.0, + 1397.0, + 292.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1393.0, + 1405.0, + 1393.0, + 1405.0, + 1427.0, + 295.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1422.0, + 661.0, + 1422.0, + 661.0, + 1459.0, + 294.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 1422.0, + 1403.0, + 1422.0, + 1403.0, + 1459.0, + 742.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1452.0, + 1194.0, + 1452.0, + 1194.0, + 1486.0, + 295.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 1452.0, + 1403.0, + 1452.0, + 1403.0, + 1486.0, + 1249.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1483.0, + 1407.0, + 1483.0, + 1407.0, + 1522.0, + 291.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1517.0, + 1178.0, + 1517.0, + 1178.0, + 1548.0, + 296.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 721.0, + 1404.0, + 721.0, + 1404.0, + 755.0, + 297.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 749.0, + 1406.0, + 749.0, + 1406.0, + 787.0, + 293.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 778.0, + 1129.0, + 778.0, + 1129.0, + 818.0, + 293.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 778.0, + 1406.0, + 778.0, + 1406.0, + 818.0, + 1200.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 808.0, + 396.0, + 808.0, + 396.0, + 850.0, + 291.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 808.0, + 1406.0, + 808.0, + 1406.0, + 850.0, + 434.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 842.0, + 1406.0, + 842.0, + 1406.0, + 877.0, + 294.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 871.0, + 296.0, + 871.0, + 296.0, + 908.0, + 293.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 871.0, + 1006.0, + 871.0, + 1006.0, + 908.0, + 387.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 871.0, + 1261.0, + 871.0, + 1261.0, + 908.0, + 1096.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 871.0, + 1406.0, + 871.0, + 1406.0, + 908.0, + 1353.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 904.0, + 535.0, + 904.0, + 535.0, + 938.0, + 294.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 904.0, + 636.0, + 904.0, + 636.0, + 938.0, + 626.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 1405.0, + 227.0, + 1405.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 296.0, + 294.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 286.0, + 1327.0, + 286.0, + 1327.0, + 333.0, + 291.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 286.0, + 1408.0, + 286.0, + 1408.0, + 333.0, + 1394.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 320.0, + 1404.0, + 320.0, + 1404.0, + 356.0, + 293.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 353.0, + 717.0, + 353.0, + 717.0, + 386.0, + 293.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 353.0, + 1273.0, + 353.0, + 1273.0, + 386.0, + 794.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 353.0, + 1405.0, + 353.0, + 1405.0, + 386.0, + 1329.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 381.0, + 1405.0, + 381.0, + 1405.0, + 419.0, + 292.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 413.0, + 297.0, + 413.0, + 297.0, + 446.0, + 294.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 413.0, + 385.0, + 413.0, + 385.0, + 446.0, + 373.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1978.0, + 295.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1403.0, + 1972.0, + 1403.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2004.0, + 1405.0, + 2004.0, + 1405.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1559.0, + 1405.0, + 1559.0, + 1405.0, + 1595.0, + 294.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1592.0, + 1062.0, + 1592.0, + 1062.0, + 1625.0, + 293.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 1592.0, + 1245.0, + 1592.0, + 1245.0, + 1625.0, + 1129.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 1592.0, + 1405.0, + 1592.0, + 1405.0, + 1625.0, + 1286.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1623.0, + 1405.0, + 1623.0, + 1405.0, + 1656.0, + 293.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1654.0, + 1404.0, + 1654.0, + 1404.0, + 1688.0, + 294.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1684.0, + 1405.0, + 1684.0, + 1405.0, + 1718.0, + 296.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1712.0, + 544.0, + 1712.0, + 544.0, + 1749.0, + 293.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 1712.0, + 1407.0, + 1712.0, + 1407.0, + 1749.0, + 600.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1745.0, + 1405.0, + 1745.0, + 1405.0, + 1779.0, + 294.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1773.0, + 621.0, + 1773.0, + 621.0, + 1810.0, + 293.0, + 1810.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1883, + 1404, + 1883, + 1404, + 2006, + 299, + 2006 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 556, + 604, + 1148, + 604, + 1148, + 1621, + 556, + 1621 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 299, + 415, + 1404, + 415, + 1404, + 508, + 299, + 508 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 301, + 307, + 1401, + 307, + 1401, + 399, + 301, + 399 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 293, + 229, + 1400, + 229, + 1400, + 293, + 293, + 293 + ], + "score": 0.953 + }, + { + "category_id": 0, + "poly": [ + 300, + 1813, + 544, + 1813, + 544, + 1848, + 300, + 1848 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.851 + }, + { + "category_id": 1, + "poly": [ + 294, + 1716, + 1400, + 1716, + 1400, + 1751, + 294, + 1751 + ], + "score": 0.543 + }, + { + "category_id": 4, + "poly": [ + 294, + 1716, + 1400, + 1716, + 1400, + 1751, + 294, + 1751 + ], + "score": 0.542 + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 606.0, + 1050.0, + 606.0, + 1050.0, + 615.0, + 1041.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 611.0, + 663.0, + 611.0, + 663.0, + 638.0, + 605.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 611.0, + 802.0, + 611.0, + 802.0, + 638.0, + 747.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 611.0, + 944.0, + 611.0, + 944.0, + 638.0, + 886.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 610.0, + 1043.0, + 610.0, + 1043.0, + 642.0, + 1016.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 665.0, + 1041.0, + 665.0, + 1041.0, + 678.0, + 1024.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 681.0, + 1041.0, + 681.0, + 1041.0, + 697.0, + 1015.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 697.0, + 1044.0, + 697.0, + 1044.0, + 717.0, + 1010.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 716.0, + 1044.0, + 716.0, + 1044.0, + 732.0, + 1019.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 736.0, + 1043.0, + 736.0, + 1043.0, + 749.0, + 1028.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 751.0, + 1045.0, + 751.0, + 1045.0, + 768.0, + 1022.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 772.0, + 662.0, + 772.0, + 662.0, + 797.0, + 603.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 770.0, + 804.0, + 770.0, + 804.0, + 798.0, + 746.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 772.0, + 945.0, + 772.0, + 945.0, + 798.0, + 886.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 769.0, + 1044.0, + 769.0, + 1044.0, + 805.0, + 1010.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1026.0, + 806.0, + 1043.0, + 806.0, + 1043.0, + 820.0, + 1026.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 820.0, + 1045.0, + 820.0, + 1045.0, + 839.0, + 1022.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 836.0, + 1046.0, + 836.0, + 1046.0, + 861.0, + 1013.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 857.0, + 1044.0, + 857.0, + 1044.0, + 876.0, + 1009.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 875.0, + 1042.0, + 875.0, + 1042.0, + 891.0, + 1011.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 894.0, + 1038.0, + 894.0, + 1038.0, + 907.0, + 1016.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 924.0, + 688.0, + 924.0, + 688.0, + 955.0, + 618.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 923.0, + 826.0, + 923.0, + 826.0, + 958.0, + 752.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 922.0, + 964.0, + 922.0, + 964.0, + 958.0, + 888.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 924.0, + 1098.0, + 924.0, + 1098.0, + 955.0, + 1028.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 1059.0, + 590.0, + 1059.0, + 590.0, + 1074.0, + 578.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 1092.0, + 590.0, + 1092.0, + 590.0, + 1107.0, + 578.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 1157.0, + 591.0, + 1157.0, + 591.0, + 1173.0, + 578.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1171.0, + 611.0, + 1171.0, + 611.0, + 1217.0, + 587.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1171.0, + 644.0, + 1171.0, + 644.0, + 1218.0, + 621.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 1172.0, + 674.0, + 1172.0, + 674.0, + 1192.0, + 655.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1171.0, + 707.0, + 1171.0, + 707.0, + 1206.0, + 687.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 1167.0, + 745.0, + 1167.0, + 745.0, + 1224.0, + 718.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 1167.0, + 778.0, + 1167.0, + 778.0, + 1228.0, + 751.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 1168.0, + 809.0, + 1168.0, + 809.0, + 1213.0, + 785.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 1172.0, + 842.0, + 1172.0, + 842.0, + 1194.0, + 820.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 1170.0, + 875.0, + 1170.0, + 875.0, + 1207.0, + 851.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 1170.0, + 909.0, + 1170.0, + 909.0, + 1222.0, + 886.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 1167.0, + 946.0, + 1167.0, + 946.0, + 1228.0, + 918.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1171.0, + 975.0, + 1171.0, + 975.0, + 1199.0, + 953.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 1168.0, + 1009.0, + 1168.0, + 1009.0, + 1207.0, + 984.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 1167.0, + 1046.0, + 1167.0, + 1046.0, + 1221.0, + 1019.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1170.0, + 1076.0, + 1170.0, + 1076.0, + 1229.0, + 1050.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 1170.0, + 1105.0, + 1170.0, + 1105.0, + 1226.0, + 1083.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 1166.0, + 1144.0, + 1166.0, + 1144.0, + 1222.0, + 1117.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 1249.0, + 689.0, + 1249.0, + 689.0, + 1280.0, + 619.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 1249.0, + 825.0, + 1249.0, + 825.0, + 1280.0, + 754.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1246.0, + 964.0, + 1246.0, + 964.0, + 1281.0, + 891.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 1248.0, + 1098.0, + 1248.0, + 1098.0, + 1280.0, + 1027.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 1380.0, + 606.0, + 1380.0, + 606.0, + 1395.0, + 592.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 1430.0, + 605.0, + 1430.0, + 605.0, + 1446.0, + 592.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1529.0, + 610.0, + 1529.0, + 610.0, + 1549.0, + 591.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1541.0, + 626.0, + 1541.0, + 626.0, + 1590.0, + 602.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1544.0, + 673.0, + 1544.0, + 673.0, + 1609.0, + 654.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1543.0, + 725.0, + 1543.0, + 725.0, + 1570.0, + 703.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 1545.0, + 776.0, + 1545.0, + 776.0, + 1579.0, + 754.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 1540.0, + 830.0, + 1540.0, + 830.0, + 1599.0, + 806.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 1543.0, + 878.0, + 1543.0, + 878.0, + 1580.0, + 854.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1542.0, + 931.0, + 1542.0, + 931.0, + 1592.0, + 906.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1543.0, + 979.0, + 1543.0, + 979.0, + 1604.0, + 953.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 1541.0, + 1031.0, + 1541.0, + 1031.0, + 1607.0, + 1005.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 1543.0, + 1078.0, + 1543.0, + 1078.0, + 1571.0, + 1057.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 1542.0, + 1128.0, + 1542.0, + 1128.0, + 1618.0, + 1102.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1809.0, + 548.0, + 1809.0, + 548.0, + 1857.0, + 291.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1712.0, + 1405.0, + 1712.0, + 1405.0, + 1753.0, + 293.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1883.0, + 1405.0, + 1883.0, + 1405.0, + 1919.0, + 294.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1913.0, + 1403.0, + 1913.0, + 1403.0, + 1946.0, + 295.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1947.0, + 1405.0, + 1947.0, + 1405.0, + 1979.0, + 295.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1974.0, + 1237.0, + 1974.0, + 1237.0, + 2012.0, + 293.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 412.0, + 1404.0, + 412.0, + 1404.0, + 452.0, + 293.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 440.0, + 1406.0, + 440.0, + 1406.0, + 487.0, + 292.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 476.0, + 986.0, + 476.0, + 986.0, + 510.0, + 294.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 304.0, + 1406.0, + 304.0, + 1406.0, + 343.0, + 295.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 334.0, + 1406.0, + 334.0, + 1406.0, + 376.0, + 294.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 369.0, + 1125.0, + 369.0, + 1125.0, + 401.0, + 295.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 260.0, + 1278.0, + 260.0, + 1278.0, + 297.0, + 293.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1712.0, + 1405.0, + 1712.0, + 1405.0, + 1753.0, + 293.0, + 1753.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 105, + 300, + 105 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.828 + }, + { + "category_id": 0, + "poly": [ + 300, + 227, + 489, + 227, + 489, + 262, + 300, + 262 + ], + "score": 0.794 + }, + { + "category_id": 1, + "poly": [ + 288, + 1441, + 1401, + 1441, + 1401, + 1505, + 288, + 1505 + ], + "score": 0.761 + }, + { + "category_id": 1, + "poly": [ + 293, + 1022, + 1404, + 1022, + 1404, + 1087, + 293, + 1087 + ], + "score": 0.732 + }, + { + "category_id": 1, + "poly": [ + 296, + 1328, + 1399, + 1328, + 1399, + 1423, + 296, + 1423 + ], + "score": 0.723 + }, + { + "category_id": 1, + "poly": [ + 295, + 1521, + 1403, + 1521, + 1403, + 1616, + 295, + 1616 + ], + "score": 0.722 + }, + { + "category_id": 1, + "poly": [ + 290, + 1246, + 1402, + 1246, + 1402, + 1311, + 290, + 1311 + ], + "score": 0.717 + }, + { + "category_id": 1, + "poly": [ + 295, + 664, + 1402, + 664, + 1402, + 729, + 295, + 729 + ], + "score": 0.678 + }, + { + "category_id": 1, + "poly": [ + 296, + 746, + 1401, + 746, + 1401, + 811, + 296, + 811 + ], + "score": 0.667 + }, + { + "category_id": 1, + "poly": [ + 295, + 278, + 1401, + 278, + 1401, + 342, + 295, + 342 + ], + "score": 0.66 + }, + { + "category_id": 1, + "poly": [ + 293, + 828, + 1401, + 828, + 1401, + 894, + 293, + 894 + ], + "score": 0.655 + }, + { + "category_id": 1, + "poly": [ + 298, + 1634, + 1402, + 1634, + 1402, + 1728, + 298, + 1728 + ], + "score": 0.647 + }, + { + "category_id": 1, + "poly": [ + 298, + 909, + 1401, + 909, + 1401, + 1004, + 298, + 1004 + ], + "score": 0.641 + }, + { + "category_id": 1, + "poly": [ + 299, + 553, + 1400, + 553, + 1400, + 649, + 299, + 649 + ], + "score": 0.61 + }, + { + "category_id": 1, + "poly": [ + 296, + 358, + 1401, + 358, + 1401, + 453, + 296, + 453 + ], + "score": 0.596 + }, + { + "category_id": 1, + "poly": [ + 294, + 1971, + 1398, + 1971, + 1398, + 2034, + 294, + 2034 + ], + "score": 0.59 + }, + { + "category_id": 1, + "poly": [ + 295, + 470, + 1402, + 470, + 1402, + 537, + 295, + 537 + ], + "score": 0.554 + }, + { + "category_id": 1, + "poly": [ + 296, + 1103, + 1404, + 1103, + 1404, + 1229, + 296, + 1229 + ], + "score": 0.545 + }, + { + "category_id": 1, + "poly": [ + 297, + 1746, + 1400, + 1746, + 1400, + 1842, + 297, + 1842 + ], + "score": 0.515 + }, + { + "category_id": 1, + "poly": [ + 294, + 1859, + 1399, + 1859, + 1399, + 1954, + 294, + 1954 + ], + "score": 0.507 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 490.0, + 230.0, + 490.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1440.0, + 1404.0, + 1440.0, + 1404.0, + 1479.0, + 294.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1472.0, + 534.0, + 1472.0, + 534.0, + 1504.0, + 324.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1017.0, + 1406.0, + 1017.0, + 1406.0, + 1061.0, + 294.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1053.0, + 595.0, + 1053.0, + 595.0, + 1084.0, + 321.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1324.0, + 1404.0, + 1324.0, + 1404.0, + 1371.0, + 292.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1356.0, + 1404.0, + 1356.0, + 1404.0, + 1400.0, + 320.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1390.0, + 504.0, + 1390.0, + 504.0, + 1422.0, + 324.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1522.0, + 1401.0, + 1522.0, + 1401.0, + 1556.0, + 295.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1553.0, + 1405.0, + 1553.0, + 1405.0, + 1590.0, + 322.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1583.0, + 612.0, + 1583.0, + 612.0, + 1615.0, + 322.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1245.0, + 1405.0, + 1245.0, + 1405.0, + 1285.0, + 294.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1279.0, + 612.0, + 1279.0, + 612.0, + 1308.0, + 322.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 663.0, + 1405.0, + 663.0, + 1405.0, + 703.0, + 294.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 697.0, + 611.0, + 697.0, + 611.0, + 729.0, + 322.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 748.0, + 1402.0, + 748.0, + 1402.0, + 784.0, + 295.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 779.0, + 1107.0, + 779.0, + 1107.0, + 812.0, + 321.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 278.0, + 1405.0, + 278.0, + 1405.0, + 314.0, + 297.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 311.0, + 991.0, + 311.0, + 991.0, + 341.0, + 323.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 829.0, + 1404.0, + 829.0, + 1404.0, + 865.0, + 295.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 858.0, + 1253.0, + 858.0, + 1253.0, + 895.0, + 320.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1632.0, + 1403.0, + 1632.0, + 1403.0, + 1674.0, + 293.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1666.0, + 1404.0, + 1666.0, + 1404.0, + 1700.0, + 321.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1693.0, + 397.0, + 1693.0, + 397.0, + 1730.0, + 318.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 906.0, + 1405.0, + 906.0, + 1405.0, + 949.0, + 293.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 943.0, + 1405.0, + 943.0, + 1405.0, + 977.0, + 324.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 971.0, + 781.0, + 971.0, + 781.0, + 1009.0, + 326.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 554.0, + 1402.0, + 554.0, + 1402.0, + 589.0, + 295.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 586.0, + 1405.0, + 586.0, + 1405.0, + 621.0, + 324.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 616.0, + 942.0, + 616.0, + 942.0, + 650.0, + 325.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 359.0, + 1405.0, + 359.0, + 1405.0, + 393.0, + 295.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 388.0, + 1406.0, + 388.0, + 1406.0, + 429.0, + 320.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 416.0, + 410.0, + 416.0, + 410.0, + 456.0, + 320.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1967.0, + 1404.0, + 1967.0, + 1404.0, + 2009.0, + 293.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2003.0, + 597.0, + 2003.0, + 597.0, + 2033.0, + 322.0, + 2033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 468.0, + 1404.0, + 468.0, + 1404.0, + 509.0, + 293.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 504.0, + 1169.0, + 504.0, + 1169.0, + 537.0, + 322.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1104.0, + 1404.0, + 1104.0, + 1404.0, + 1140.0, + 295.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1134.0, + 1405.0, + 1134.0, + 1405.0, + 1172.0, + 320.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1166.0, + 1406.0, + 1166.0, + 1406.0, + 1202.0, + 320.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1198.0, + 586.0, + 1198.0, + 586.0, + 1232.0, + 323.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1747.0, + 1405.0, + 1747.0, + 1405.0, + 1782.0, + 295.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1777.0, + 1405.0, + 1777.0, + 1405.0, + 1815.0, + 322.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1810.0, + 900.0, + 1810.0, + 900.0, + 1842.0, + 322.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1860.0, + 1403.0, + 1860.0, + 1403.0, + 1894.0, + 295.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1887.0, + 1404.0, + 1887.0, + 1404.0, + 1927.0, + 322.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1922.0, + 612.0, + 1922.0, + 612.0, + 1952.0, + 323.0, + 1952.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 228, + 1400, + 228, + 1400, + 323, + 299, + 323 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 299, + 343, + 1405, + 343, + 1405, + 465, + 299, + 465 + ], + "score": 0.915 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.891 + }, + { + "category_id": 1, + "poly": [ + 297, + 1271, + 1403, + 1271, + 1403, + 1367, + 297, + 1367 + ], + "score": 0.874 + }, + { + "category_id": 1, + "poly": [ + 298, + 1385, + 1402, + 1385, + 1402, + 1480, + 298, + 1480 + ], + "score": 0.863 + }, + { + "category_id": 1, + "poly": [ + 299, + 1499, + 1405, + 1499, + 1405, + 1594, + 299, + 1594 + ], + "score": 0.854 + }, + { + "category_id": 1, + "poly": [ + 298, + 568, + 1406, + 568, + 1406, + 663, + 298, + 663 + ], + "score": 0.852 + }, + { + "category_id": 1, + "poly": [ + 289, + 1189, + 1400, + 1189, + 1400, + 1254, + 289, + 1254 + ], + "score": 0.851 + }, + { + "category_id": 1, + "poly": [ + 299, + 485, + 1399, + 485, + 1399, + 551, + 299, + 551 + ], + "score": 0.842 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.836 + }, + { + "category_id": 1, + "poly": [ + 298, + 683, + 1406, + 683, + 1406, + 777, + 298, + 777 + ], + "score": 0.83 + }, + { + "category_id": 1, + "poly": [ + 291, + 1105, + 1403, + 1105, + 1403, + 1170, + 291, + 1170 + ], + "score": 0.817 + }, + { + "category_id": 1, + "poly": [ + 295, + 909, + 1403, + 909, + 1403, + 975, + 295, + 975 + ], + "score": 0.799 + }, + { + "category_id": 1, + "poly": [ + 297, + 795, + 1406, + 795, + 1406, + 891, + 297, + 891 + ], + "score": 0.793 + }, + { + "category_id": 1, + "poly": [ + 297, + 992, + 1401, + 992, + 1401, + 1087, + 297, + 1087 + ], + "score": 0.764 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 871.0, + 2085.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 260.0, + 1404.0, + 260.0, + 1404.0, + 296.0, + 322.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 288.0, + 398.0, + 288.0, + 398.0, + 324.0, + 319.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 344.0, + 1406.0, + 344.0, + 1406.0, + 380.0, + 294.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 373.0, + 1406.0, + 373.0, + 1406.0, + 410.0, + 321.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 404.0, + 1406.0, + 404.0, + 1406.0, + 440.0, + 319.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 431.0, + 398.0, + 431.0, + 398.0, + 470.0, + 319.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1268.0, + 1407.0, + 1268.0, + 1407.0, + 1310.0, + 294.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1304.0, + 1405.0, + 1304.0, + 1405.0, + 1339.0, + 321.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1331.0, + 400.0, + 1331.0, + 400.0, + 1368.0, + 320.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1385.0, + 1406.0, + 1385.0, + 1406.0, + 1423.0, + 296.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1418.0, + 1404.0, + 1418.0, + 1404.0, + 1452.0, + 322.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1443.0, + 399.0, + 1443.0, + 399.0, + 1483.0, + 319.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1500.0, + 1403.0, + 1500.0, + 1403.0, + 1534.0, + 295.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1532.0, + 1405.0, + 1532.0, + 1405.0, + 1566.0, + 323.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1561.0, + 754.0, + 1561.0, + 754.0, + 1595.0, + 322.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 568.0, + 1403.0, + 568.0, + 1403.0, + 605.0, + 293.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 600.0, + 1406.0, + 600.0, + 1406.0, + 637.0, + 321.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 629.0, + 451.0, + 629.0, + 451.0, + 664.0, + 320.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1185.0, + 1403.0, + 1185.0, + 1403.0, + 1227.0, + 293.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1220.0, + 991.0, + 1220.0, + 991.0, + 1252.0, + 325.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 484.0, + 1404.0, + 484.0, + 1404.0, + 522.0, + 297.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 518.0, + 1161.0, + 518.0, + 1161.0, + 552.0, + 323.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 681.0, + 1404.0, + 681.0, + 1404.0, + 719.0, + 295.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 713.0, + 1404.0, + 713.0, + 1404.0, + 749.0, + 322.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 746.0, + 609.0, + 746.0, + 609.0, + 775.0, + 323.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1103.0, + 1405.0, + 1103.0, + 1405.0, + 1144.0, + 293.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1138.0, + 611.0, + 1138.0, + 611.0, + 1168.0, + 322.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 910.0, + 1405.0, + 910.0, + 1405.0, + 946.0, + 295.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 942.0, + 1272.0, + 942.0, + 1272.0, + 976.0, + 322.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 795.0, + 1406.0, + 795.0, + 1406.0, + 833.0, + 294.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 827.0, + 1406.0, + 827.0, + 1406.0, + 865.0, + 321.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 860.0, + 824.0, + 860.0, + 824.0, + 892.0, + 320.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 993.0, + 1405.0, + 993.0, + 1405.0, + 1027.0, + 296.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1024.0, + 1405.0, + 1024.0, + 1405.0, + 1059.0, + 322.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1057.0, + 708.0, + 1057.0, + 708.0, + 1087.0, + 321.0, + 1087.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1300, + 1405, + 1300, + 1405, + 1488, + 297, + 1488 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 616, + 1403, + 616, + 1403, + 740, + 299, + 740 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1133, + 1404, + 1133, + 1404, + 1288, + 298, + 1288 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 386, + 1404, + 386, + 1404, + 510, + 298, + 510 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1910, + 1402, + 1910, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 298, + 1802, + 1404, + 1802, + 1404, + 1896, + 298, + 1896 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 370, + 777, + 1403, + 777, + 1403, + 927, + 370, + 927 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 297, + 963, + 1402, + 963, + 1402, + 1028, + 297, + 1028 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 661, + 1557, + 1033, + 1557, + 1033, + 1685, + 661, + 1685 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 298, + 1725, + 1400, + 1725, + 1400, + 1789, + 298, + 1789 + ], + "score": 0.947 + }, + { + "category_id": 0, + "poly": [ + 300, + 1074, + 603, + 1074, + 603, + 1106, + 300, + 1106 + ], + "score": 0.92 + }, + { + "category_id": 0, + "poly": [ + 301, + 557, + 609, + 557, + 609, + 589, + 301, + 589 + ], + "score": 0.909 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 302, + 308, + 709, + 308, + 709, + 345, + 302, + 345 + ], + "score": 0.874 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.852 + }, + { + "category_id": 1, + "poly": [ + 301, + 225, + 711, + 225, + 711, + 262, + 301, + 262 + ], + "score": 0.821 + }, + { + "category_id": 13, + "poly": [ + 702, + 860, + 874, + 860, + 874, + 897, + 702, + 897 + ], + "score": 0.93, + "latex": "q = [ \\overleftrightarrow { \\lambda \\omega _ { 1 } } , \\overrightarrow { c w _ { S } } ]" + }, + { + "category_id": 13, + "poly": [ + 395, + 1757, + 466, + 1757, + 466, + 1790, + 395, + 1790 + ], + "score": 0.92, + "latex": "( h , w )" + }, + { + "category_id": 13, + "poly": [ + 497, + 1757, + 586, + 1757, + 586, + 1790, + 497, + 1790 + ], + "score": 0.92, + "latex": "[ p _ { h } , p _ { w } ]" + }, + { + "category_id": 13, + "poly": [ + 422, + 680, + 560, + 680, + 560, + 710, + 422, + 710 + ], + "score": 0.92, + "latex": "s = 1 , . . . , S" + }, + { + "category_id": 14, + "poly": [ + 663, + 1556, + 1035, + 1556, + 1035, + 1687, + 663, + 1687 + ], + "score": 0.92, + "latex": "\\begin{array} { l } { p _ { ( h , 2 i ) } = \\sin { \\Big ( h / 1 0 0 0 0 ^ { 2 i / p _ { d } } \\Big ) } } \\\\ { p _ { ( h , 2 i + 1 ) } = \\cos { \\Big ( h / 1 0 0 0 0 ^ { 2 i / p _ { d } } \\Big ) } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 417, + 1835, + 483, + 1835, + 483, + 1868, + 417, + 1868 + ], + "score": 0.92, + "latex": "( x , y )" + }, + { + "category_id": 13, + "poly": [ + 299, + 1834, + 364, + 1834, + 364, + 1868, + 299, + 1868 + ], + "score": 0.91, + "latex": "( x , y )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1226, + 466, + 1226, + 466, + 1254, + 298, + 1254 + ], + "score": 0.9, + "latex": "H = W = 1 4" + }, + { + "category_id": 13, + "poly": [ + 555, + 1260, + 612, + 1260, + 612, + 1290, + 555, + 1290 + ], + "score": 0.9, + "latex": "x _ { h , w }" + }, + { + "category_id": 13, + "poly": [ + 1087, + 1226, + 1210, + 1226, + 1210, + 1254, + 1087, + 1254 + ], + "score": 0.89, + "latex": "C = 1 0 2 4" + }, + { + "category_id": 13, + "poly": [ + 1042, + 962, + 1081, + 962, + 1081, + 992, + 1042, + 992 + ], + "score": 0.89, + "latex": "s ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 1227, + 1195, + 1325, + 1195, + 1325, + 1226, + 1227, + 1226 + ], + "score": 0.88, + "latex": "H , W , C" + }, + { + "category_id": 13, + "poly": [ + 682, + 970, + 728, + 970, + 728, + 995, + 682, + 995 + ], + "score": 0.87, + "latex": "c w _ { s }" + }, + { + "category_id": 13, + "poly": [ + 608, + 1732, + 639, + 1732, + 639, + 1758, + 608, + 1758 + ], + "score": 0.86, + "latex": "p _ { d }" + }, + { + "category_id": 13, + "poly": [ + 1344, + 1838, + 1375, + 1838, + 1375, + 1867, + 1344, + 1867 + ], + "score": 0.86, + "latex": "p _ { d }" + }, + { + "category_id": 13, + "poly": [ + 369, + 1947, + 400, + 1947, + 400, + 1977, + 369, + 1977 + ], + "score": 0.86, + "latex": "x _ { p }" + }, + { + "category_id": 13, + "poly": [ + 1326, + 1973, + 1400, + 1973, + 1400, + 2004, + 1326, + 2004 + ], + "score": 0.85, + "latex": "( K B _ { V }" + }, + { + "category_id": 13, + "poly": [ + 298, + 683, + 331, + 683, + 331, + 709, + 298, + 709 + ], + "score": 0.85, + "latex": "w _ { s }" + }, + { + "category_id": 13, + "poly": [ + 1144, + 1758, + 1164, + 1758, + 1164, + 1784, + 1144, + 1784 + ], + "score": 0.84, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 662, + 782, + 679, + 782, + 679, + 807, + 662, + 807 + ], + "score": 0.83, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 538, + 619, + 559, + 619, + 559, + 645, + 538, + 645 + ], + "score": 0.83, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 869, + 901, + 886, + 901, + 886, + 927, + 869, + 927 + ], + "score": 0.8, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 1214, + 1761, + 1237, + 1761, + 1237, + 1784, + 1214, + 1784 + ], + "score": 0.8, + "latex": "w" + }, + { + "category_id": 13, + "poly": [ + 654, + 1002, + 671, + 1002, + 671, + 1028, + 654, + 1028 + ], + "score": 0.8, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 473, + 1426, + 493, + 1426, + 493, + 1452, + 473, + 1452 + ], + "score": 0.8, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 730, + 1426, + 769, + 1426, + 769, + 1454, + 730, + 1454 + ], + "score": 0.79, + "latex": "- 1" + }, + { + "category_id": 13, + "poly": [ + 398, + 810, + 557, + 810, + 557, + 844, + 398, + 844 + ], + "score": 0.79, + "latex": "[ c w _ { 1 } , . . . , c w _ { S } ]" + }, + { + "category_id": 13, + "poly": [ + 397, + 865, + 415, + 865, + 415, + 891, + 397, + 891 + ], + "score": 0.78, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 498, + 1733, + 521, + 1733, + 521, + 1754, + 498, + 1754 + ], + "score": 0.77, + "latex": "w" + }, + { + "category_id": 13, + "poly": [ + 1176, + 710, + 1195, + 710, + 1195, + 736, + 1176, + 736 + ], + "score": 0.77, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1149, + 1944, + 1168, + 1944, + 1168, + 1970, + 1149, + 1970 + ], + "score": 0.76, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 546, + 1430, + 570, + 1430, + 570, + 1453, + 546, + 1453 + ], + "score": 0.73, + "latex": "w" + }, + { + "category_id": 13, + "poly": [ + 1049, + 1201, + 1068, + 1201, + 1068, + 1222, + 1049, + 1222 + ], + "score": 0.69, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 879, + 1922, + 896, + 1922, + 896, + 1939, + 879, + 1939 + ], + "score": 0.67, + "latex": "x" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1071.0, + 607.0, + 1071.0, + 607.0, + 1111.0, + 296.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 555.0, + 613.0, + 555.0, + 613.0, + 595.0, + 295.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 306.0, + 714.0, + 306.0, + 714.0, + 350.0, + 295.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1300.0, + 1406.0, + 1300.0, + 1406.0, + 1337.0, + 294.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1334.0, + 1403.0, + 1334.0, + 1403.0, + 1366.0, + 296.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1363.0, + 1406.0, + 1363.0, + 1406.0, + 1399.0, + 295.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1391.0, + 1406.0, + 1391.0, + 1406.0, + 1433.0, + 291.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1422.0, + 472.0, + 1422.0, + 472.0, + 1461.0, + 294.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 1422.0, + 545.0, + 1422.0, + 545.0, + 1461.0, + 494.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 1422.0, + 729.0, + 1422.0, + 729.0, + 1461.0, + 571.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 1422.0, + 1406.0, + 1422.0, + 1406.0, + 1461.0, + 770.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1456.0, + 1128.0, + 1456.0, + 1128.0, + 1490.0, + 294.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 615.0, + 537.0, + 615.0, + 537.0, + 652.0, + 295.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 615.0, + 1403.0, + 615.0, + 1403.0, + 652.0, + 560.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 645.0, + 1404.0, + 645.0, + 1404.0, + 685.0, + 293.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 678.0, + 297.0, + 678.0, + 297.0, + 714.0, + 292.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 678.0, + 421.0, + 678.0, + 421.0, + 714.0, + 332.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 678.0, + 1405.0, + 678.0, + 1405.0, + 714.0, + 561.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 710.0, + 1175.0, + 710.0, + 1175.0, + 742.0, + 295.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 710.0, + 1341.0, + 710.0, + 1341.0, + 742.0, + 1196.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1135.0, + 1405.0, + 1135.0, + 1405.0, + 1168.0, + 297.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1164.0, + 1404.0, + 1164.0, + 1404.0, + 1197.0, + 294.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1191.0, + 1048.0, + 1191.0, + 1048.0, + 1229.0, + 293.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 1191.0, + 1226.0, + 1191.0, + 1226.0, + 1229.0, + 1069.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 1191.0, + 1406.0, + 1191.0, + 1406.0, + 1229.0, + 1326.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 1225.0, + 1086.0, + 1225.0, + 1086.0, + 1258.0, + 467.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1225.0, + 1406.0, + 1225.0, + 1406.0, + 1258.0, + 1211.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1252.0, + 554.0, + 1252.0, + 554.0, + 1296.0, + 292.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 1252.0, + 1102.0, + 1252.0, + 1102.0, + 1296.0, + 613.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 385.0, + 1403.0, + 385.0, + 1403.0, + 422.0, + 293.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 415.0, + 1406.0, + 415.0, + 1406.0, + 453.0, + 291.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 447.0, + 1404.0, + 447.0, + 1404.0, + 485.0, + 293.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 480.0, + 777.0, + 480.0, + 777.0, + 512.0, + 296.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1909.0, + 878.0, + 1909.0, + 878.0, + 1948.0, + 294.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1909.0, + 1406.0, + 1909.0, + 1406.0, + 1948.0, + 897.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 368.0, + 1943.0, + 368.0, + 1975.0, + 295.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 1943.0, + 1148.0, + 1943.0, + 1148.0, + 1975.0, + 401.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1975.0, + 1169.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1325.0, + 1972.0, + 1325.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2005.0, + 833.0, + 2005.0, + 833.0, + 2034.0, + 295.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1802.0, + 1404.0, + 1802.0, + 1404.0, + 1839.0, + 294.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 1833.0, + 416.0, + 1833.0, + 416.0, + 1871.0, + 365.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 1833.0, + 1343.0, + 1833.0, + 1343.0, + 1871.0, + 484.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 1833.0, + 1405.0, + 1833.0, + 1405.0, + 1871.0, + 1376.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1862.0, + 479.0, + 1862.0, + 479.0, + 1900.0, + 294.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 777.0, + 661.0, + 777.0, + 661.0, + 814.0, + 365.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 777.0, + 1406.0, + 777.0, + 1406.0, + 814.0, + 680.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 807.0, + 397.0, + 807.0, + 397.0, + 848.0, + 394.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 807.0, + 561.0, + 807.0, + 561.0, + 848.0, + 558.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 860.0, + 396.0, + 860.0, + 396.0, + 897.0, + 367.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 860.0, + 701.0, + 860.0, + 701.0, + 897.0, + 416.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 860.0, + 1405.0, + 860.0, + 1405.0, + 897.0, + 875.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 893.0, + 868.0, + 893.0, + 868.0, + 929.0, + 392.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 893.0, + 1235.0, + 893.0, + 1235.0, + 929.0, + 887.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 959.0, + 681.0, + 959.0, + 681.0, + 999.0, + 293.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 959.0, + 1041.0, + 959.0, + 1041.0, + 999.0, + 729.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 959.0, + 1407.0, + 959.0, + 1407.0, + 999.0, + 1082.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 993.0, + 653.0, + 993.0, + 653.0, + 1033.0, + 295.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 993.0, + 1380.0, + 993.0, + 1380.0, + 1033.0, + 672.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1724.0, + 497.0, + 1724.0, + 497.0, + 1760.0, + 296.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1724.0, + 607.0, + 1724.0, + 607.0, + 1760.0, + 522.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 1724.0, + 1404.0, + 1724.0, + 1404.0, + 1760.0, + 640.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1756.0, + 394.0, + 1756.0, + 394.0, + 1792.0, + 293.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 1756.0, + 496.0, + 1756.0, + 496.0, + 1792.0, + 467.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1756.0, + 1143.0, + 1756.0, + 1143.0, + 1792.0, + 587.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 1756.0, + 1213.0, + 1756.0, + 1213.0, + 1792.0, + 1165.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 1756.0, + 1248.0, + 1756.0, + 1248.0, + 1792.0, + 1238.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 223.0, + 714.0, + 223.0, + 714.0, + 266.0, + 296.0, + 266.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1412, + 1404, + 1412, + 1404, + 1807, + 298, + 1807 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 298, + 1404, + 298, + 1404, + 605, + 297, + 605 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1013, + 1404, + 1013, + 1404, + 1258, + 298, + 1258 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 726, + 1403, + 726, + 1403, + 912, + 299, + 912 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1274, + 1403, + 1274, + 1403, + 1396, + 299, + 1396 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 300, + 1911, + 1402, + 1911, + 1402, + 2033, + 300, + 2033 + ], + "score": 0.972 + }, + { + "category_id": 0, + "poly": [ + 297, + 1852, + 998, + 1852, + 998, + 1883, + 297, + 1883 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.88 + }, + { + "category_id": 0, + "poly": [ + 300, + 954, + 624, + 954, + 624, + 985, + 300, + 985 + ], + "score": 0.857 + }, + { + "category_id": 0, + "poly": [ + 298, + 657, + 952, + 657, + 952, + 689, + 298, + 689 + ], + "score": 0.846 + }, + { + "category_id": 0, + "poly": [ + 298, + 226, + 968, + 226, + 968, + 262, + 298, + 262 + ], + "score": 0.844 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 864, + 2088, + 864, + 2112, + 837, + 2112 + ], + "score": 0.834 + }, + { + "category_id": 13, + "poly": [ + 1054, + 419, + 1114, + 419, + 1114, + 450, + 1054, + 450 + ], + "score": 0.9, + "latex": "1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 1186, + 361, + 1354, + 361, + 1354, + 389, + 1186, + 389 + ], + "score": 0.9, + "latex": "H = W = 1 4" + }, + { + "category_id": 13, + "poly": [ + 297, + 391, + 415, + 391, + 415, + 420, + 297, + 420 + ], + "score": 0.9, + "latex": "C = 1 0 2 4" + }, + { + "category_id": 13, + "poly": [ + 1252, + 392, + 1337, + 392, + 1337, + 422, + 1252, + 422 + ], + "score": 0.89, + "latex": "p = 1 2" + }, + { + "category_id": 13, + "poly": [ + 1008, + 361, + 1106, + 361, + 1106, + 392, + 1008, + 392 + ], + "score": 0.78, + "latex": "H , W , C" + }, + { + "category_id": 13, + "poly": [ + 297, + 452, + 392, + 452, + 392, + 482, + 297, + 482 + ], + "score": 0.76, + "latex": "1 0 - 2 0" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1849.0, + 1002.0, + 1849.0, + 1002.0, + 1888.0, + 295.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 951.0, + 628.0, + 951.0, + 628.0, + 990.0, + 295.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 651.0, + 957.0, + 651.0, + 957.0, + 696.0, + 295.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 224.0, + 971.0, + 224.0, + 971.0, + 266.0, + 292.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1414.0, + 1405.0, + 1414.0, + 1405.0, + 1446.0, + 297.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1441.0, + 1406.0, + 1441.0, + 1406.0, + 1478.0, + 293.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1473.0, + 1405.0, + 1473.0, + 1405.0, + 1508.0, + 293.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1504.0, + 1405.0, + 1504.0, + 1405.0, + 1540.0, + 293.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1536.0, + 1405.0, + 1536.0, + 1405.0, + 1568.0, + 294.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1566.0, + 1404.0, + 1566.0, + 1404.0, + 1598.0, + 296.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1595.0, + 1406.0, + 1595.0, + 1406.0, + 1631.0, + 293.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1625.0, + 1405.0, + 1625.0, + 1405.0, + 1661.0, + 293.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1655.0, + 1406.0, + 1655.0, + 1406.0, + 1692.0, + 293.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1688.0, + 1405.0, + 1688.0, + 1405.0, + 1720.0, + 294.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1715.0, + 1408.0, + 1715.0, + 1408.0, + 1755.0, + 293.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1747.0, + 1405.0, + 1747.0, + 1405.0, + 1784.0, + 293.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1777.0, + 772.0, + 1777.0, + 772.0, + 1812.0, + 295.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 299.0, + 1405.0, + 299.0, + 1405.0, + 334.0, + 295.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 326.0, + 1405.0, + 326.0, + 1405.0, + 366.0, + 292.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 359.0, + 1007.0, + 359.0, + 1007.0, + 394.0, + 294.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 359.0, + 1185.0, + 359.0, + 1185.0, + 394.0, + 1107.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 359.0, + 1405.0, + 359.0, + 1405.0, + 394.0, + 1355.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 388.0, + 1251.0, + 388.0, + 1251.0, + 425.0, + 416.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1338.0, + 388.0, + 1406.0, + 388.0, + 1406.0, + 425.0, + 1338.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 420.0, + 1053.0, + 420.0, + 1053.0, + 455.0, + 291.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 420.0, + 1407.0, + 420.0, + 1407.0, + 455.0, + 1115.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 449.0, + 1406.0, + 449.0, + 1406.0, + 488.0, + 393.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 482.0, + 1406.0, + 482.0, + 1406.0, + 518.0, + 294.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 513.0, + 1404.0, + 513.0, + 1404.0, + 548.0, + 295.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 542.0, + 1406.0, + 542.0, + 1406.0, + 579.0, + 294.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 574.0, + 1338.0, + 574.0, + 1338.0, + 608.0, + 294.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1012.0, + 1405.0, + 1012.0, + 1405.0, + 1049.0, + 294.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1046.0, + 1404.0, + 1046.0, + 1404.0, + 1079.0, + 294.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1076.0, + 1406.0, + 1076.0, + 1406.0, + 1108.0, + 293.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1106.0, + 1405.0, + 1106.0, + 1405.0, + 1139.0, + 294.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1133.0, + 1406.0, + 1133.0, + 1406.0, + 1171.0, + 291.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1165.0, + 1406.0, + 1165.0, + 1406.0, + 1203.0, + 292.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1198.0, + 1404.0, + 1198.0, + 1404.0, + 1231.0, + 293.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1228.0, + 905.0, + 1228.0, + 905.0, + 1260.0, + 293.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 725.0, + 1404.0, + 725.0, + 1404.0, + 760.0, + 294.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 755.0, + 1405.0, + 755.0, + 1405.0, + 795.0, + 293.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 788.0, + 1405.0, + 788.0, + 1405.0, + 823.0, + 293.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 820.0, + 1404.0, + 820.0, + 1404.0, + 852.0, + 295.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 850.0, + 1407.0, + 850.0, + 1407.0, + 883.0, + 293.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 882.0, + 983.0, + 882.0, + 983.0, + 917.0, + 295.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1275.0, + 1403.0, + 1275.0, + 1403.0, + 1307.0, + 295.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1304.0, + 1404.0, + 1304.0, + 1404.0, + 1337.0, + 293.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1335.0, + 1405.0, + 1335.0, + 1405.0, + 1371.0, + 294.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1362.0, + 1384.0, + 1362.0, + 1384.0, + 1403.0, + 293.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1908.0, + 1406.0, + 1908.0, + 1406.0, + 1949.0, + 294.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1977.0, + 293.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2009.0, + 293.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2004.0, + 738.0, + 2004.0, + 738.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 811, + 1404, + 811, + 1404, + 1300, + 298, + 1300 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1467, + 1404, + 1467, + 1404, + 1744, + 298, + 1744 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1759, + 1404, + 1759, + 1404, + 2033, + 298, + 2033 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 230, + 1403, + 230, + 1403, + 596, + 298, + 596 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 299, + 611, + 1403, + 611, + 1403, + 796, + 299, + 796 + ], + "score": 0.976 + }, + { + "category_id": 0, + "poly": [ + 301, + 1390, + 686, + 1390, + 686, + 1420, + 301, + 1420 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 837, + 2089, + 864, + 2089, + 864, + 2112, + 837, + 2112 + ], + "score": 0.844 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1389.0, + 692.0, + 1389.0, + 692.0, + 1424.0, + 296.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 2086.0, + 869.0, + 2086.0, + 869.0, + 2125.0, + 833.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 811.0, + 1405.0, + 811.0, + 1405.0, + 846.0, + 296.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 843.0, + 1405.0, + 843.0, + 1405.0, + 878.0, + 294.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 873.0, + 1404.0, + 873.0, + 1404.0, + 907.0, + 293.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 902.0, + 1404.0, + 902.0, + 1404.0, + 938.0, + 293.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 934.0, + 1404.0, + 934.0, + 1404.0, + 968.0, + 294.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 962.0, + 1406.0, + 962.0, + 1406.0, + 1000.0, + 292.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 993.0, + 1404.0, + 993.0, + 1404.0, + 1030.0, + 293.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1024.0, + 1405.0, + 1024.0, + 1405.0, + 1060.0, + 293.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1055.0, + 1404.0, + 1055.0, + 1404.0, + 1090.0, + 294.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1086.0, + 1404.0, + 1086.0, + 1404.0, + 1121.0, + 294.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1117.0, + 1404.0, + 1117.0, + 1404.0, + 1151.0, + 294.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1146.0, + 1404.0, + 1146.0, + 1404.0, + 1181.0, + 294.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1178.0, + 1406.0, + 1178.0, + 1406.0, + 1213.0, + 294.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1208.0, + 1404.0, + 1208.0, + 1404.0, + 1242.0, + 294.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1238.0, + 1406.0, + 1238.0, + 1406.0, + 1273.0, + 294.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1269.0, + 764.0, + 1269.0, + 764.0, + 1303.0, + 294.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1469.0, + 1402.0, + 1469.0, + 1402.0, + 1502.0, + 297.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1497.0, + 1405.0, + 1497.0, + 1405.0, + 1535.0, + 293.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1528.0, + 1407.0, + 1528.0, + 1407.0, + 1566.0, + 291.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1561.0, + 1405.0, + 1561.0, + 1405.0, + 1594.0, + 294.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1590.0, + 1406.0, + 1590.0, + 1406.0, + 1626.0, + 292.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1617.0, + 1405.0, + 1617.0, + 1405.0, + 1660.0, + 292.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1653.0, + 1404.0, + 1653.0, + 1404.0, + 1685.0, + 296.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1680.0, + 1405.0, + 1680.0, + 1405.0, + 1718.0, + 293.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1710.0, + 1096.0, + 1710.0, + 1096.0, + 1750.0, + 293.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1760.0, + 1401.0, + 1760.0, + 1401.0, + 1792.0, + 297.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1791.0, + 1405.0, + 1791.0, + 1405.0, + 1823.0, + 297.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1822.0, + 1405.0, + 1822.0, + 1405.0, + 1854.0, + 296.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1852.0, + 1405.0, + 1852.0, + 1405.0, + 1884.0, + 294.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 1406.0, + 1882.0, + 1406.0, + 1918.0, + 294.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1910.0, + 1404.0, + 1910.0, + 1404.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1943.0, + 1405.0, + 1943.0, + 1405.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2009.0, + 293.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2004.0, + 1056.0, + 2004.0, + 1056.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 231.0, + 1403.0, + 231.0, + 1403.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 260.0, + 1405.0, + 260.0, + 1405.0, + 298.0, + 294.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 290.0, + 1403.0, + 290.0, + 1403.0, + 327.0, + 293.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 318.0, + 1405.0, + 318.0, + 1405.0, + 358.0, + 291.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 350.0, + 1407.0, + 350.0, + 1407.0, + 390.0, + 292.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 380.0, + 1407.0, + 380.0, + 1407.0, + 420.0, + 292.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 413.0, + 1405.0, + 413.0, + 1405.0, + 449.0, + 293.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 443.0, + 1405.0, + 443.0, + 1405.0, + 476.0, + 293.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 472.0, + 1407.0, + 472.0, + 1407.0, + 514.0, + 292.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 503.0, + 1405.0, + 503.0, + 1405.0, + 544.0, + 292.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 534.0, + 1407.0, + 534.0, + 1407.0, + 572.0, + 292.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 562.0, + 1280.0, + 562.0, + 1280.0, + 604.0, + 292.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 612.0, + 1405.0, + 612.0, + 1405.0, + 648.0, + 294.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 642.0, + 1405.0, + 642.0, + 1405.0, + 678.0, + 294.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 674.0, + 1404.0, + 674.0, + 1404.0, + 706.0, + 295.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 703.0, + 1405.0, + 703.0, + 1405.0, + 739.0, + 294.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 733.0, + 1405.0, + 733.0, + 1405.0, + 770.0, + 293.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 763.0, + 1178.0, + 763.0, + 1178.0, + 803.0, + 294.0, + 803.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 18, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 509, + 1404, + 509, + 1404, + 843, + 298, + 843 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 859, + 1404, + 859, + 1404, + 1135, + 298, + 1135 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 1151, + 1402, + 1151, + 1402, + 1305, + 299, + 1305 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1403, + 229, + 1403, + 414, + 299, + 414 + ], + "score": 0.977 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 302, + 452, + 741, + 452, + 741, + 483, + 302, + 483 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2112, + 835, + 2112 + ], + "score": 0.842 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 451.0, + 746.0, + 451.0, + 746.0, + 487.0, + 297.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2123.0, + 829.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 508.0, + 1404.0, + 508.0, + 1404.0, + 544.0, + 295.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 540.0, + 1404.0, + 540.0, + 1404.0, + 575.0, + 296.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 569.0, + 1404.0, + 569.0, + 1404.0, + 606.0, + 295.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 601.0, + 1405.0, + 601.0, + 1405.0, + 635.0, + 294.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 630.0, + 1404.0, + 630.0, + 1404.0, + 665.0, + 294.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 662.0, + 1406.0, + 662.0, + 1406.0, + 697.0, + 293.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 691.0, + 1406.0, + 691.0, + 1406.0, + 729.0, + 293.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 723.0, + 1406.0, + 723.0, + 1406.0, + 756.0, + 293.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 751.0, + 1405.0, + 751.0, + 1405.0, + 788.0, + 293.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 783.0, + 1406.0, + 783.0, + 1406.0, + 817.0, + 293.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 815.0, + 532.0, + 815.0, + 532.0, + 846.0, + 296.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 861.0, + 1404.0, + 861.0, + 1404.0, + 894.0, + 296.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 890.0, + 1404.0, + 890.0, + 1404.0, + 927.0, + 294.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 921.0, + 1405.0, + 921.0, + 1405.0, + 958.0, + 294.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 951.0, + 1404.0, + 951.0, + 1404.0, + 988.0, + 293.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 983.0, + 1404.0, + 983.0, + 1404.0, + 1019.0, + 293.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1010.0, + 1404.0, + 1010.0, + 1404.0, + 1048.0, + 293.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1044.0, + 1405.0, + 1044.0, + 1405.0, + 1078.0, + 292.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1070.0, + 1405.0, + 1070.0, + 1405.0, + 1113.0, + 292.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1104.0, + 587.0, + 1104.0, + 587.0, + 1141.0, + 294.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1150.0, + 1404.0, + 1150.0, + 1404.0, + 1187.0, + 294.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1179.0, + 1401.0, + 1179.0, + 1401.0, + 1217.0, + 294.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1212.0, + 1403.0, + 1212.0, + 1403.0, + 1245.0, + 295.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1240.0, + 1406.0, + 1240.0, + 1406.0, + 1280.0, + 294.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1273.0, + 1117.0, + 1273.0, + 1117.0, + 1310.0, + 295.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 231.0, + 1404.0, + 231.0, + 1404.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 262.0, + 1403.0, + 262.0, + 1403.0, + 297.0, + 294.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 290.0, + 1405.0, + 290.0, + 1405.0, + 327.0, + 293.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 323.0, + 1404.0, + 323.0, + 1404.0, + 355.0, + 295.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 352.0, + 1405.0, + 352.0, + 1405.0, + 388.0, + 294.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 383.0, + 754.0, + 383.0, + 754.0, + 418.0, + 294.0, + 418.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 19, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/S1sqHMZCb/S1sqHMZCb.md b/parse/train/S1sqHMZCb/S1sqHMZCb.md new file mode 100644 index 0000000000000000000000000000000000000000..cc3b948bfdd6d087df3080d72125192a6ee52565 --- /dev/null +++ b/parse/train/S1sqHMZCb/S1sqHMZCb.md @@ -0,0 +1,490 @@ +# NERVENET: LEARNING STRUCTURED POLICY WITH GRAPH NEURAL NETWORKS + +Tingwu Wang∗, Renjie Liao∗, Jimmy Ba & Sanja Fidler + +Department of Computer Science +University of Toronto +Vector Institute +{tingwuwang,rjliao}@cs.toronto.edu, +jimmy@psi.toronto.edu, fidler@cs.toronto.edu + +# ABSTRACT + +We address the problem of learning structured policies for continuous control. In traditional reinforcement learning, policies of agents are learned by multi-layer perceptrons (MLPs) which take the concatenation of all observations from the environment as input for predicting actions. In this work, we propose NerveNet to explicitly model the structure of an agent, which naturally takes the form of a graph. Specifically, serving as the agent’s policy network, NerveNet first propagates information over the structure of the agent and then predict actions for different parts of the agent. In the experiments, we first show that our NerveNet is comparable to state-of-the-art methods on standard MuJoCo environments. We further propose our customized reinforcement learning environments for benchmarking two types of structure transfer learning tasks, i.e., size and disability transfer, as well as multi-task learning. We demonstrate that policies learned by NerveNet are significantly more transferable and generalizable than policies learned by other models and are able to transfer even in a zero-shot setting. + +# 1 INTRODUCTION + +Deep reinforcement learning (RL) has received increasing attention over the past few years, with the recent success of applications such as playing Atari Games, Mnih et al. (2015), and Go, (Silver et al., 2016; 2017). Significant advances have also been made in robotics using the latest RL techniques, e.g., Levine et al. (2016); Gu et al. (2017). + +Many RL problems feature agents with multiple dependent controllers. For example, humanoid robots consist of multiple physically linked joints. Action to be taken by each joint or the body should thus not only depend on its own observations but also on actions of other joints. + +Previous approaches in RL typically use MLP to learn the agent’s policy. In particular, MLP takes the concatenation of observations from the environment as input, which may be measurements like positions, velocities of body and joints in the current time instance. The MLP policy then predicts actions to be taken by every joint and body. Thus the task of the MLP policy is to discover the latent relationships between observations. This typically leads to longer training times, requiring more exposure of the agent to the environment. In our work, we aim to exploit the body structure of an agent, and physical dependencies that naturally exist in such agents. + +We rely on the fact that bodies of most robots and animals have a discrete graph structure. Nodes of the graph may represent the joints, and edges represent the (physical) dependencies between them. In particular, we define the agent’s policy using a Graph Neural Network, Scarselli et al. (2009), which is a neural network that operates over graph structures. We refer to our model as NerveNet due to the resemblance of the neural nervous system to a graph. NerveNet propagates information between different parts of the body based on the underlying graph structure before outputting the action for each part. By doing so, NerveNet can leverage the structure information encoded by the agent’s body which is advantageous in learning the correct inductive bias, and thus is less prone to overfitting. Moreover, NerveNet is naturally suitable for structure transferring tasks as most of the model weights are shared across the nodes and edges, respectively. + +![](images/ec7ba80f49404fcff9389c3659fcdd445c72eb87234ed2686b8a6262e8e85922.jpg) +Figure 1: Visualization of the graph structure of CentipedeEight in our environment. We use this agent for testing the ability of transfer learning of our model. Since for this agent, each body node is paired with at least one joint node, we omit the body nodes and fill up the position with the corresponding joint nodes. By omitting the body nodes, a more compact graph is constructed, the details of which are illustrated in the experimental section. + +We first evaluate our NerveNet on standard RL benchmarks such as the OpenAI Gym, Brockman et al. (2016) which stem from MuJoCo. We show that our model achieves comparable results to state-of-the-art MLP based methods. To verify our claim regarding the structure transfer, we further introduce our customized RL environments which are based on the ones of Gym. Two types of structure transfer tasks are designed, size transfer and disability transfer. In particular, size transfer focuses on the scenario in which policies are learned for small-sized agents (simpler body structure) and applied directly to large-sized agents which are composed by some repetitive components shared with the small-sized agent. Secondly, disability transfer investigates scenarios in which policies are learned for one agent and applied to the same agent with some components disabled. Our experiments demonstrate that for structure transfer tasks our NerveNet is significantly better than all other competitors, and can even achieve zero-shot learning for some agents. For the multi-task learning tasks, NerveNet is also able to learn policies that are more robust with better efficiency. + +The main contribution of this paper is the following: We explore the problem of learning transferable and generalized features by incorporating a prior on the structure via graph neural networks. NerveNet permits powerful transfer learning from one structure to another, which goes well beyond the ability of previous models. NerveNet is also more robust and has more potential in performing multi-task learning. The demo and code for this project are released, under the project page of http://www.cs.toronto.edu/˜tingwuwang/nervenet.html. + +# 2 NERVENET + +In this section, we first introduce the notation. We then explain how to construct the graph for each of our agents, followed by the description of the NerveNet. Finally, we describe the learning algorithm for our model. + +We formulate the locomotion control problems as an infinite-horizon discounted Markov decision process (MDP). To fully describe the MDP for continuous control problems which include locomotion control, we define the state space or observation space as $s$ and action space as $\mathcal { A }$ . To interact with the environments, the agent generates its stochastic policy $\pi _ { \boldsymbol { \theta } } ( \boldsymbol { a } ^ { \tau } | \boldsymbol { s } ^ { \tau } )$ based on the current state $s ^ { \tau } \in \boldsymbol { S }$ , where $a ^ { \tau } \in \mathcal { A }$ is the action and $\theta$ are the parameters of the policy function. The environment on the other hand, produces a reward $r ( s ^ { \tau } , a ^ { \tau } )$ for the agent, and the agent’s objective is to find a policy that maximizes the expected reward. + +![](images/4387bb37b2b6edf82cb3edff6c91e9f09012e99f1638fa5f05208b0a537fe83a.jpg) +Figure 2: In this figure, we use Walker-Ostrich as an example of NerveNet. In the input model, for each node, NerveNet fetches the corresponding elements from the observation vector. NerveNet then computes the messages between neighbors in the graph, and updates the hidden state of each node. This process is repeated for a certain number of propagation steps. In the output model, the policy is produced by collecting the output from each controller. + +# 2.1 GRAPH CONSTRUCTION + +In real life, skeletons of most robots and animals have a discrete graph structure, and are most often trees. Simulators such as the MuJoCo engine by Todorov et al. (2012), organize the agents using an XML-based kinematic tree. In our experiments, we will use the tree graphs as per MuJoCo. Note that our model can be directly applied to arbitrary graphs. In particular, we assume two types of nodes in our tree: body and joint. The body nodes are abstract nodes used to construct the kinematic tree via nesting, which is similar to the coordinate frame system used in robotics, Spong et al. (2006). The joint node represents the degrees of freedom of motion between the two body nodes. Take a simple humanoid as an example; the body nodes Thigh and Shin are connected via the Knee, where Knee is a hinge joint. We further add a root node which observes additional information about the agent. For example, in the Reacher agent in MuJoCo, the root node has access to the target position of the agent. We build edges to form a tree graph. Fig. 1 illustrates the graph structure of an example agent, CentipedeEight. We sketch the agent and its corresponding graph in the left and right part of the figure, respectively. Note that for better visualization, we omit the joint nodes and use edges to represent the physical connections of joint nodes. Different elements of the agent are parsed into nodes with different colors. Further details are provided in the experimental section. + +# 2.2 NERVENET AS POLICY NETWORK + +We now turn to NerveNet which parametrizes the policy with a Graph Neural Network. Before delving into details, we first introduce our notation. We then specify the input model which helps to initialize the hidden state of each node. We further introduce the propagation model that updates these hidden states. Finally, we describe the output model. + +We denote the graph structure of the agent as $G = ( V , E )$ where $V$ and $E$ are the sets of nodes and edges, respectively. We focus on the directed graphs as the undirected case can be easily addressed by splitting one undirected edge into two directed edges. We denote the out-going neighborhood of node $u$ as $\mathcal { N } _ { o u t } ( u )$ which contains all endpoints $v$ with $( u , v )$ being an edge in the graph. Similarly, we denote the in-coming neighborhood of node $u$ as $\mathcal { N } _ { i n } ( u )$ . Every node $u$ has an associated node type $p _ { u } \in \{ 1 , 2 , \ldots , P \}$ , which in our case corresponds to body, joint and root. We also associate each edge $( u , v )$ with an edge type $c _ { ( u , v ) } \in \{ 1 , 2 , \ldots , C \}$ . Node type can help in capturing different importances across nodes. Edge type can be used to describe different relationships between nodes, and thus propagate information between them differently. One can also add more than one edge type to the same edge which results in a multi-graph. We stick to simple graphs for simplicity. One interesting fact is that we have two notions of “time” in our model. One is the time step in the environment which is the typical time coordinate for RL problems. The other corresponds to the internal propagation step of NerveNet. These two coordinates work as follows. At each time step of the environment, NerveNet receives observation from the environment and performs a few internal propagation steps in order to decide on the action to be taken by each node. To avoid confusion, throughout this paper, we use $\tau$ to describe the time step in the environment and $t$ for the propagation step. + +# 2.2.1 INPUT MODEL + +For each time step $\tau$ in the environment, the agent receives an observation $s ^ { \tau } \in \boldsymbol { S }$ . The observation vector $s ^ { \tau }$ is the concatenation of observations of each node. We denote the elements of observation vector $s ^ { \tau }$ corresponding to node $u$ with $x _ { u }$ . From now on, we drop the time step in the environment to derive the model for simplicity. The observation vector goes through an input network to obtain a fixed-size state vector as follows: + +$$ +h _ { u } ^ { 0 } = F _ { \mathrm { i n } } ( x _ { u } ) , +$$ + +where the subscript and superscript denote the node index and propagation step, respectively. Here, $F _ { \mathrm { i n } }$ may be a MLP and $h _ { u } ^ { 0 }$ is the state vector of node $u$ at propagation step 0. Note that we may need to pad zeros to the observation vectors if different nodes have observations of different sizes. + +# 2.2.2 PROPAGATION MODEL + +We now describe the propagation model of our NerveNet which mimics a synchronous message passing system studied in distributed computing, Attiya & Welch (2004). We will show how the state vector of each node is updated from one propagation step to the next. This update process is recurrently applied during the whole propagation. We leave the details to the appendix. + +Message Computation In particular, at propagation step $t$ , for every node $u$ , we have access to a state vector $h _ { u } ^ { t }$ . For every edge $( u , v ) \in \bar { \mathcal { N } } _ { o u t } \bar { ( u ) }$ , node $u$ computes a message vector as below, + +$$ +m _ { ( u , v ) } ^ { t } = M _ { c _ { ( u , v ) } } ( h _ { u } ^ { t } ) , +$$ + +where Mc(u,v) i s the message function which may be an identity mapping or a MLP. Note that the subscript $c _ { ( u , v ) }$ indicates that edges of the same edge type share the same instance of the message function. For example, the second torso in Fig. 1 sends a message to the first and third torso, as well as the LeftHip and RightHip. + +Message Aggregation Once every node finishes computing messages, we aggregate messages sent from all in-coming neighbors of each node. Specifically, for every node $u$ , we perform the following aggregation: + +$$ +\bar { m } _ { u } ^ { t } = A ( \{ h _ { v } ^ { t } | v \in \mathcal { N } _ { i n } ( u ) \} ) , +$$ + +where $A$ is the aggregation function which may be a summation, average or max-pooling function. Here, $\hat { m } _ { u } ^ { t }$ is the aggregated message vector which contains the information sent from the node’s neighborhood. + +States Update We now update every node’s state vector based on both the aggregated message and its current state vector. In particular, for every node $u$ , we perform the following update: + +$$ +h _ { u } ^ { t + 1 } = U _ { p _ { u } } ( h _ { u } ^ { t } , \bar { m } _ { u } ^ { t } ) , +$$ + +where $U$ is the update function which may be a gated recurrent unit (GRU), a long short term memory (LSTM) unit or a MLP. From the subscript $p _ { u }$ of $U$ , we can see that nodes of the same node type share the same instance of the update function. The above propagation model is then recurrently applied for a fixed number of time steps $T$ to get the final state vectors of all nodes, i.e., $\{ h _ { u } ^ { T } | u \in \bar { V } \}$ . + +# 2.2.3 OUTPUT MODEL + +In RL, agents typically use a MLP policy, where the network outputs the mean of the Gaussian distribution for each of the actions, while the standard deviation is a trainable vector, Schulman et al. (2017). In our output model, we also treat standard deviation in the same way. + +However, instead of predicting the action distribution of all nodes by a single network, we make predictions for each individual node. We denote the set of nodes which are assigned controllers for the actuators as $\mathcal { O }$ . For each such node, a MLP takes its final state vectors $h _ { u \in \mathcal { O } } ^ { T }$ as input and produces the mean of the action of the Gaussian policy for the corresponding actuator. For each output node $u \in \mathcal { O }$ , we define its output type as $q _ { u }$ . Different sharing schemes are available for the instance of MLP $O _ { q _ { u } }$ , for example, we can force the nodes with similar physical structure to share the instance of MLP. For example, in Fig. 1, two LeftHip nodes have a shared controller. Therefore, we have the following output model: + +$$ +\mu _ { u \in \mathcal { O } } = O _ { q _ { u } } ( h _ { u } ^ { T } ) , +$$ + +where $\mu _ { u \in \mathcal { O } }$ is the mean value for action applied on each actuator. In practice, we found that we can force controllers of different output types to share one unified controller, while not hurting the performance. By integrating the produced Gaussian policy for each action, the probability density of the stochastic policy is calculated as + +$$ +\pi _ { \boldsymbol { \theta } } ( a ^ { \tau } \vert s ^ { \tau } ) = \prod _ { \boldsymbol { u } \in \mathcal { O } } \pi _ { \boldsymbol { \theta } , \boldsymbol { u } } ( a _ { \boldsymbol { u } } ^ { \tau } \vert s ^ { \tau } ) = \prod _ { \boldsymbol { u } \in \mathcal { O } } \frac { 1 } { \sqrt { 2 \pi \sigma _ { \boldsymbol { u } } ^ { 2 } } } e ^ { ( a _ { \boldsymbol { u } } ^ { \tau } - \mu _ { \boldsymbol { u } } ) ^ { 2 } / ( 2 \sigma _ { \boldsymbol { u } } ^ { 2 } ) } , +$$ + +where $a ^ { \tau } \in { \mathcal { A } }$ is the output action, and $\sigma _ { u }$ is the variable standard deviation for each action. Here, $\theta$ represents the parameters of the policy function. + +# 2.3 LEARNING ALGORITHM + +To interact with the environments, the agent generates its stochastic policy $\pi _ { \boldsymbol { \theta } } ( \boldsymbol { a } ^ { \tau } | \boldsymbol { s } ^ { \tau } )$ after several propagation steps. The environment on the other hand, produces a reward $r ( s ^ { \tau } , a ^ { \tau } )$ for the agent, and transits to the next state with transition probability $\mathsf { \bar { P } } ( s ^ { \tau + 1 } | s ^ { \tau } )$ . The target of the agent is to maximize its cumulative return + +$$ +J ( \theta ) = \mathbb { E } _ { \pi } \left[ \sum _ { \tau = 0 } ^ { \infty } \gamma ^ { \tau } r ( s ^ { \tau } , a ^ { \tau } ) \right] . +$$ + +To optimize the expected reward, we use the proximal policy optimization (PPO) by Schulman et al. (2017). In PPO, the agents alternate between sampling trajectories with the latest policy and performing optimization on surrogate objective using the sampled trajectories. The algorithm tries to keep the KL-divergence of the new policy and the old policy within the trust region. To achieve that, PPO clips the probability ratio and adds an KL-divergence penalty term to the loss. The likelihood ratio is defined as $r ^ { \tau } ( \boldsymbol { \theta } ; \mathbf { \dot { \theta } } _ { o l d } ) = \pi _ { \boldsymbol { \theta } } ( a ^ { \tau } | s ^ { \tau } ) / \pi _ { \theta _ { o l d } } ( a ^ { \tau } | \mathbf { \bar { s } } ^ { \tau } )$ . Following the notation and the algorithm of PPO, our NerveNet tries to minimize the summation of the original loss in Eq. (7), KL-penalty and the value function loss which is defined as: + +$$ +\begin{array} { l } { \displaystyle \tilde { J } ( \theta ) = J ( \theta ) - \beta L _ { K L } ( \theta ) - \alpha L _ { V } ( \theta ) } \\ { \displaystyle \qquad = \mathbb { E } _ { \pi _ { \theta } } \left[ \sum _ { \tau = 0 } ^ { \infty } \operatorname* { m i n } \left( \hat { A } ^ { \tau } r ^ { \tau } ( \theta ) , \hat { A } ^ { \tau } \operatorname { c l i p } \left( r ^ { \tau } ( \theta ) , 1 - \epsilon , 1 + \epsilon \right) \right) \right] } \\ { \displaystyle \qquad - \beta \mathbb { E } _ { \pi _ { \theta } } \left[ \sum _ { \tau = 0 } ^ { \infty } \mathrm { K L } \left[ \pi _ { \theta } ( \cdot | s ^ { \tau } ) | \pi _ { \theta _ { o l d } } ( \cdot | s ^ { \tau } ) \right] \right] - \alpha \mathbb { E } _ { \pi _ { \theta } } \left[ \sum _ { \tau = 0 } ^ { \infty } \left( V _ { \theta } ( s ^ { \tau } ) - V ( s ^ { \tau } ) ^ { \mathrm { t a r g e t } } \right) ^ { 2 } \right] , } \end{array} +$$ + +where $\hat { A } _ { t }$ is the generalized advantage estimation (GAE) calculated using algorithm from Schulman et al. (2015b), and $\epsilon$ is the clip value, which we choose to be 0.2. Here, $\beta$ is a dynamical coefficient adjusted to keep the KL-divergence constraints, and $\alpha$ is used to balance the value loss. Note that in Eq. (8), $V ( s _ { t } ) ^ { \mathrm { t a r g e t } }$ is the target state value in accordance with the GAE method. To optimize the $\tilde { J } ( \theta )$ , PPO make use of the policy gradient in Sutton et al. (2000) to do first-order gradient descent optimization. + +Value Network To produce the state value $V _ { \theta } ( s ^ { \tau } )$ for given observation $s ^ { \tau }$ , we have several alternatives: (1) using one GNN as the policy network and using one MLP as the value network (NerveNet-MLP); (2) using one GNN as policy network and using another GNN as value network (NerveNet-2) (without sharing the parameters of the two GNNs); (3) using one GNN as both policy network and value network (NerveNet-1). The GNN for value network is very similar to the GNN for policy network. The output for value GNN is a scalar instead of a vector of mean action. We will compare these variants in the experimental section. + +# 3 RELATED WORK + +Reinforcement Learning Reinforcement learning (RL) has recently achieved huge success in a variety of applications. Powered by the progress of deep neural networks, Krizhevsky et al. (2012), agents are now able to successfully play Atari Games and beat the world’s best (human) players in the game of Go (Mnih et al., 2015; Silver et al., 2016; 2017). Based on simulation engines like MuJoCo, Todorov et al. (2012), numerous algorithms have been proposed to train agents also in continuous control problems (Schulman et al., 2017; 2015a; Heess et al., 2017; Metz et al., 2017). + +Structure in RL Most approaches that exploit priors on structure of the problem fall in the domain of hierarchical RL, (Kulkarni et al., 2016; Vezhnevets et al., 2017), which mainly focus on modeling intrinsic motivation of agents. In Hausknecht & Stone (2015), the authors extend the deep RL algorithms to MDPS with parameterized action space by exploiting the structure of action space and bounding the action space gradients. Graphs have been used in RL problems prior to our work. In Metzen (2013); Mabu et al. (2007); Shoeleh & Asadpour (2017); Mahadevan & Maggioni (2007), the authors use graphs to learn a representation of the environment. However, these methods are limited to problems with simple dynamical models like for example the task of 2d-navigation, and thus these problems are usually solved via model-based RL. However, for complex multi-joint agents, learning the dynamical model as well as predicting the transition of states is time consuming and biased. For problems of training model-free multi-joint agents in complex physical environments, relatively little attention has been devoted to modeling the physical structure of the agents. + +Graph Neural Networks There have been many efforts to generalize neural networks to graphstructured data. One line of work is based on convolutional neural networks (CNNs). In (Bruna et al., 2014; Defferrard et al., 2016; Kipf & Welling, 2017), CNNs are employed in the spectral domain relying on the graph Laplacian matrix. (Goller & Kuchler, 1996; Duvenaud et al., 2015) used hash functions in order to apply CNNs to graphs. + +Another popular direction is based on recurrent neural networks (RNNs) (Goller & Kuchler, 1996; Gori et al., 2005; Scarselli et al., 2009; Socher et al., 2011; Li et al., 2015; Tai et al., 2015). Among RNN based methods, many are only applicable to special structured graph, e.g., sequences or trees, (Socher et al., 2011; Tai et al., 2015). One class of models which are applicable to general graphs are so-called graph neural networks (GNNs), Scarselli et al. (2009). The inference procedure is a forward pass that exploits a fixed-length propagation process which resembles synchronous message passing system in the theory of distributed computing, Attiya & Welch (2004). Nodes in the graph have state vectors which are recurrently updated based on their history and received messages. One of the representative work of GNNs, i.e., gated graph neural networks (GGNNs) by Li et al. (2015), uses gated recurrent unit to update the state vectors. Learning such a model can be achieved by the back-propagation through time (BPTT) algorithm or recurrent back-propagation, Chauvin & Rumelhart (1995). It has been shown that GNNs, (Li et al., 2015; 2017; Qi et al., 2017; Gilmer et al., 2017) have a high capacity and achieve state-of-the-art performance in many applications which involve graph-structured data. In this paper, we model the structure of the reinforcement learning agents using GNNs. + +Transfer and Multi-task Learning in RL Recently, there has been increased interest in transfer learning tasks for RL, Taylor & Stone (2009), which mainly focus on transferring the policy learned from one environment to another. In Rajeswaran et al. (2017b;a), the authors show that agents in reinforcement learning are prone to over-fitting, and that the learned policies generalize poorly across environments. In model-based RL, traditional control has been well studied for generalization properties, Coros et al. (2010). Gupta et al. (2017) try to increase the transferability via learning invariant visual features. Efforts have also been made from the meta-learning perspective (Duan et al., 2016; Finn et al., 2017b;a). In Wulfmeier et al. (2017), the authors propose a method of transfer learning by using imitation learning. Transferability comes naturally in our model by exploiting the (shared) graph structure of the agents. + +Multi-task learning has also received a lot of attention, Wilson et al. (2007). In Teh et al. (2017), the authors use a distilled policy that captures common behaviour across tasks. Wilson et al. (2007); Oh et al. (2017); Andreas et al. (2016) use a hierarchical approach, where multiple sub-policies are learnt. In Yang et al. (2017); Parisotto et al. (2015), the authors exploited shared visual features to tackle multi-task learning. In Ammar et al. (2014), a multi-task policy gradient is proposed, while, Calandriello et al. (2014) propose multi-task extensions of the fitted Q-iteration algorithm. Successor features in Barreto et al. (2017) are used to boost the performance of multiple tasks. Unlike ours, these methods do not exploit the physical graph structures of the agents. + +![](images/31914ef34c44e5cb9c9bdb09331fb9917e4b0134cd1f05a2719e5c2b4c19287a.jpg) +Figure 3: Results of MLP, TreeNet and NerveNet on 8 continuous control benchmarks from the Gym. + +# 4 EXPERIMENTS + +In this section, we first verify the effectiveness of NerveNet on standard MuJoCo environments in OpenAI Gym. We then investigate the transfer abilities of NerveNet and other competitors by customizing some of those environments, as well as the multi-task learning ability and robustness. + +# 4.1 COMPARISON ON STANDARD BENCHMARKS OF MUJOCO + +Baselines We compare NerveNet with the standard MLP models utilized by Schulman et al. (2017) and another baseline which is constructed as follows. We first remove the physical graph structure and introduce an additional super node which connects to all nodes in the graph. This results in a singly rooted depth-1 tree. We refer to this baseline as TreeNet. The propagation model of TreeNet is similar to NerveNet where, however, the policy first aggregates the information from all children and then feeds the state vector of the root to the output model. This simpler model serves as a baseline to verify the importance of the graph structure. + +We run experiments on 8 simulated continuous control benchmarks from the Gym, Brockman et al. (2016), which is based on MuJoCo, Todorov et al. (2012). In particular, we use Reacher, InvertedPendulum, InvertedDoublePendulum, Swimmer, and four walking or running tasks: HalfCheetah, Hopper, Walker2d, Ant. We set the maximum number of training steps to be 1 million for all environments as it is enough to solve them. Note that for InvertedPendulum, different from the original one in Gym, we add the distance penalty of the cart and velocity penalty so that the reward is more consistent to the InvertedDoublePendulum. This change of design also makes the task more challenging. + +Table 1: Performance of the Pre-trained models on CentipedeFour and CentipedeSix. + +
C_FourReward AvgStdMaxC_SixReward AvgStdMax
NerveNet2799.91247.23833.9NerveNet2507.1738.42979.2
MLP2398.51390.43936.3MLP2793.01005.23529.5
TreeNet1429.6957.73021.7TreeNet2229.71109.43727.4
+ +Results We do grid search to find the best hyperparameters and leave the details in the Appendix 6.3. As the randomness might have a big impact on the performance, for each environment, we run 3 experiments with different random seeds and plot the average curves and the standard deviations. We show the results in Figure 3. From the figures, we can see that MLP with the same setup as in Schulman et al. (2017) works the best in most of tasks.1 NerveNet basically matches the performance of MLP in terms of sample efficiency as well as the performance after it converges. In most cases, the TreeNet is worse than NerveNet which highlights the importance of keeping the physical graph structure. + +# 4.2 STRUCTURE TRANSFER LEARNING + +We now benchmark our model in the task of structure transfer learning by creating customized environments based on the existing ones from MuJoCo. We mainly investigate two types of structure transfer learning tasks. The first one is to train a model with an agent of small size (small graph) and apply the learned model to an agent with a larger size, i.e., size transfer. When increasing the size of the agent, observation and action space also increase which makes learning more challenging. Another type of structure transfer learning is disability transfer where we first learn a model for the original agent and then apply it to the same agent with some components disabled. If one model overfits the environment, disabling some components of the agent might bring catastrophic performance degradation. Note that for both transfer tasks, all factors of environments do not change except the structure of the agent. + +Centipede We create the first environment in which the agent has a similar structure to a centipede. The goal of the agent is to run as fast as possible along the $y$ -direction in the MuJoCo environment. The agent consists of repetitive torso bodies where each one has two legs attached. For two consecutive bodies, we add two actuators which control the rotation between them. Furthermore, each leg consists of a thigh and shin, which are controlled by two hinge actuators. By linking copies of torso bodies and corresponding legs, we create agents with different lengths. Specifically, the shortest Centipede is CentipedeFour and the longest one is CentipedeFourty due to the limit of supported resource of MuJoCo. For each time step, the total reward is the speed reward minus the energy cost and force feedback from the ground. Note that in practice, we found that training a CentipedeEight from scratch is already very difficult. For size transfer experiments, we create many instances which are listed in Figure 4, like “4to06”, “6to10”. For disability transfer, we create CrippleCentipede agents of which two back legs are disabled. In Figure 4, CrippleCentipede is specified as “Cp”. + +Snakes We also create a snake-like agent which is common in robotics, Crespi & Ijspeert (2008). We design the Snake environment based on the Swimmer model in Gym. The goal of the agent is to move as fast as possible. For details of the environment, please see the schematic figure 16. + +# 4.2.1 EXPERIMENTAL SETTINGS + +To fully investigate the performance of NerveNet, we build several baseline models for structure transfer learning which are explained below. + +NerveNet For the NerveNet, since all the weights are exactly the same for the small and the largeagent models, we directly use the old weights trained on the small-agent model. When the large agent has repetitive structure, we further re-use the weights of the corresponding joints from the small-agent model. + +![](images/ded2354d5c50c0404c254caa82cabadd97bd1f063a74fb66cd5313b24ebd4bd0.jpg) +Figure 4: Performance of zero-shot learning on centipedes. For each task, we run the policy for 100 episodes and record the average reward and average length the agent runs before falling down. + +![](images/1b4f2a5fa06b4d78a1b4be689e625ce9271bf372ee76f157e6a19ba72f40e380.jpg) +Figure 5: (a), (b): Results of fine-tuning for size transfer experiments. (c), (d) Results of fine-tuning for disability transfer experiments. + +MLP Pre-trained (MLPP) For the MLP based model, while transferring from one structure to another, the size of the input layer changes since the size of the observation changes. One straightforward idea is to reuse the weights from the first hidden layer to the output layer and randomly initialize the weights of the new input layer. + +![](images/84ea5296a5d035d9829cab325bf3532db657dcf4811e266ee0b813b65808af7f.jpg) +Figure 6: Results on zero-shot transfer learning on snake agents. Each tasks are simulated for 100 episodes. + +![](images/75955e4c44eb72c36a84acb11321fa2796dcfc7110c692596195fe9c0701f3c7.jpg) +Figure 7: Results of finetuning on snake environments. + +MLP Activation Assigning (MLPAA) Another way of making MLP transferable is assigning the weights of the small-agent model to the corresponding partial weights of the large-agent model and setting the remaining weights to be zero. Note that we do not add or remove any layers from the small-agent model to the large-agent except for changing the size of the layers. By doing so, we can keep the output of the large-agent model to be same as the small-agent in the beginning, i.e., keeping the same initial policy. + +TreeNet TreeNet is similar as the model described before. We apply the same way of assigning weights as MLPAA to TreeNet for the transfer learning task. + +Random We also include the random policy which is uniformly sampled from the action space. + +# 4.2.2 RESULTS + +Centipedes For the Centipedes environment, we first run experiments of all models on CentipedeSix and CentipedeFour to get the pre-trained models for transfer learning. We train different models until these agents run equally well as possible, which is reported in Table 1. Note that, in practice, we train TreeNet on CentipedeFour for more than 8 million time steps. However, due to the difficulty of optimizing TreeNet on CentipedeFour, the performance is still lower. But visually, the TreeNet agent is able to run in CentipedeFour. + +We then examine the zero-shot performance where zero-shot means directly applying the model trained with one setting to the other without any fine-tuning. To better visualize the results, we linearly normalize the performance to get a performance score, and color the results accordingly. The normalization scheme is recorded in Appendix 11. The performance score is less than 1, and is shown in the parentheses behind the original results. As we can see from Figure 4 (full chart in Appendix 6.5), NerveNet outperforms all competitors on all settings, except in the 4toCp06 scenario. Note that transferring from CentipedeFour is more difficult than from CentipedeSix since the situation where one torso connects to two neighboring torsos only happens beyond 4 bodies. + +TreeNet has a surprisingly good performance on tasks from CentipedeFour. However, by checking the videos, the learned agent is actually not able to “move” as good as other methods. The high reward is mainly due to the fact that TreeNet policy is better at standing still and gaining alive bonus. We argue that the average running-length in each episode is also a very important metric. + +By including the results of running-length, we notice that NerveNet is the only model able to walk in the zero-shot setting. In fact, the performance of NerveNet is orders-of-magnitude better, and most of the time, agents from other methods cannot even move forward. We also notice that if transferred from CentipedeSix, NerveNet is able to provide walkable pre-trained models on all new agents. + +We fine-tune for both size transfer and disability transfer experiments and show the training curves in Figure 5. From the figure, we can see that by using the pre-trained model, NerveNet significantly decreases the number of episodes required to reach the level of reward which is considered as solved. By looking at the videos, we notice that the bottleneck of learning for the agent is “how to stand”. When training from scratch, it can be seen that almost 0.5 million time steps are spent on a very flat reward surface. Therefore, the MLPAA agents, which copy the learned policy, are able to stand and bypass this time-consuming process and reach to a good performance in the end. + +Moreover, by examining the result videos, we noticed that the “walk-cycle” behavior is observed for NerveNet but is not common for others. Walk-cycle are adopted for many insects in the world, Biewener (2003). For example, six-legged ants use a tripedal gait, where the legs are used in two separate triangles alternatively touching the ground. We give more details of walk-cycle in Section 4.5. + +One possible reason is that the agent of MLP based method (MLPAA, MLPP) learns a policy that does not utilize all legs. From CentipedeEight and up, we do not observe any MLP agents to be able to coordinate all legs whereas almost all policies learned by NerveNet use all legs. Therefore, NerveNet is better at utilizing structure information and not over-fitting the environments. + +Snakes The zero-shot performance for snakes is summarized in Figure 6. As we can see, NerveNet has the best performance on all transfer learning tasks. In most cases, NerveNet has a starting reward value of more than 300, which is a pretty good policy since 350 is considered as solved for snakeThree. By looking at the videos, we found that agents of other competitors are not able to control the new actuators in the zero-shot setting. They either overfit to the original models, where the policy is completely useless in the new setting (e.g., the MLPAA is worse than random policy in SnakeThree2SnakeFour), or the new actuators are not able to coordinate with the old actuators trained before. While for NerveNet, the actuators are able to coordinate to its neighbors, regardless of whether they are new to the agents. + +We also summarize the training curves of fine-tuning in Fig. 7. We can observe that NerveNet has a very good initialization with the pre-trained model, and the performance increases with fine-tuning. When training from scratch, NerveNet is less sample efficient compared to the MLP model which might be caused by the fact that optimizing our model is more challenging than MLP. Fine-tuning helps to improve the sample efficiency of our model by a large margin. At the same time, although the MLPAA has a very good initialization, its performance progresses slowly with the increasing number of episodes. In most experiments, the MLPAA and TreeNet did not match the performance of its non-pretrained MLP baseline. + +![](images/3709a850e281bcae1eeed9ee523b904fff989358725e2076c690a230315f97b0.jpg) +Figure 8: Results of Multi-task learning. We train the networks simultaneously on five different tasks from Walker. + +Table 2: Results of Multi-task learning, with comparison to the single-task baselines. For three models, the first row is the mean reward of each model of the last 40 iterations. The second row indicates the percentage of the performance of the multi-task model compared with the single-task baseline of each model. + +
Model HalfHumanoidHopperOstrichWolfHorseAverage
MLPRewardRatio1775.7557.7%1369.5962.0%1198.8848.2%1249.2354.5%2084.0769.7%/58.6%
TreeNetRewardRatio237.8179.3%417.2798.0%224.0757.4%247.03141.2%223.3499.2%/94.8%
NerveNetRewardRatio2536.5296.3%2113.56101.8%1714.6398.8%2054.54105.9%2343.62106.4%1101.8%
+ +# 4.3 MULTI-TASK LEARNING + +In this section, we show that NerveNet has a good potential of multi-task learning by incorporating structure prior into the network structure. It is important to point out that multi-task learning represents a very difficult, and more often the case, unsolved problem in RL. Most multi-task learning algorithms, Teh et al. (2017); Andreas et al. (2016); Oh et al. (2017); Yang et al. (2017) have not been applied to domains as difficult as locomotion for complex physical models, not to mention multi-task learning among different agents with different dynamics. + +In this work, we constrain our problem domain, and design the Walker multi-task learning taskset, which contains five 2d-walkers. We aim to test the model’s ability of multi-task learning, in particular, the ability to control multiple agents using one unified network. The walkers are very different in terms of their dynamics, since they have very distinct structures, different types and numbers of controllers. Walker-HalfHumanoid and Walker-Hopper are variants of Walker2d and Hopper from the original MuJoCo Benchmarks, respectively. Walker-Horse, Walker-Ostrich, Walker-Wolf on the other hand, are agents mimicking the natural animals. Just like the real animals, some of the agents have tails or necks to help them to balance. The detailed schematic figures are shown in the Appendix 6.8. + +
ModelHalfhumanoid HopperWolfOstrich HorseAverage
MassMLPNerveNet33.28%95.87%74.04%93.24%94.68%90.13%59.23% 40.61%80.2% 69.23%60.37%85.73%
StrengthMLPNerveNet25.96%31.11%21.77%42.20%27.32%42.84%30.08% 19.80%31.41% 36.54%24.99%36.82%
+ +Table 3: Results of robustness evaluations. Note that we show the average results for each type of parameters after perturbation. And the results are columned by the agent type. The ratio of the average performance of perturbed agents and the original performance is shown in the figure. Details are listed in 6.6. + +# 4.3.1 EXPERIMENTAL SETTINGS + +To show the ability of multi-task learning of NerveNet, we design several baselines. We use a vanilla multi-task policy update for all models. More specifically, for each sub-task in the multi-task learning task-set, we use an equal number of time steps for each policy’s update and calculate the gradients separately. Gradients are then aggregated and the mean value of gradients is applied to update the network. To compensate for the additional difficulty in training more agents and tasks, we linearly increase the number of update epochs during each update in training, as well as the total number of time steps generated before the training is terminated. The hyper-parameter setting is summarized in Appendix 6.7. + +NerveNet For NerveNet, the weights are naturally shared among different agents. More specifically, for different agents, the weight matrices for propagation and output are shared. + +MLP Sharing For the MLP method, we shared the weight matrices between hidden layers. + +MLP Aggregation In the MLP Sharing approach, the total size of the weight matrices grows with the number of tasks. For different agents, whose dimension of the observations are usually different, weights from observation to the first hidden layer cannot be reused in the MLP Sharing approach. Therefore in the MLP Aggregation method, we multiply each element of the observation vector separately by one matrix, and aggregate the resulting vectors from each element. The size of this multiplying matrix is (1, dimension of the first hidden layer). + +TreeNet Similarly, TreeNet also has the benefits that its weights are naturally shared among different agents. However, TreeNet has no knowledge of the agents’ physical structure, where the information of each node is aggregated into the root node. + +We also include the baseline of training single-task MLP for each agent. We train the single-task MLP baselines for 1 million time steps per agent. In the Figure 8, we align the results of single-task MLP baseline and the results of multi-task models by the number of episodes of one task. + +As can be seen from Figure 8, NerveNet achieves the best performance in all the sub-tasks. In Walker-HalfHumanoid, Walker-Hopper, Walker-Ostrich, Walker-Wolf our NerveNet is able to out-perform other agents by a large margin. In Walker-Horse, the performance of NerveNet and MLP Sharing are relatively similar. For MLP Sharing, the performance on other four agents are relatively limited, while for Walker-Hopper, the improvement of performance is limited from half of the experiment onwards. The MLP Aggregation and TreeNet methods are not able to solve the multi-task learning problem, with both of them stuck at a very low reward level. In the vanilla optimization setting, we show that NerveNet has a bigger potential than the baselines. + +From Table 2, one can observe that the performance of MLP drops drastically $42 \%$ performance drop) when switching from single-task to multi-task learning, while for NerveNet, there is no obvious drop in performance. Our intuition is that NerveNet is better at learning generalized features, and learning of different agents can help in training other agents, while for MLP methods, the performance decreases due the competition of different agents. + +![](images/892025ae723031d5978ddef92b55588d72f89132d24844c0cfb5e117cf25e298.jpg) +Figure 9: Diagram of the walk cycle. In the left figure, legs within the same triangle are used simultaneously. For each leg, we use the same color for their diagram on the left and their curves on the right. + +![](images/652d63005f4eea342561cb0a44ebd4dd31b388fd4be3b385a751118a06287c04.jpg) +Figure 10: Results of visualization of feature distribution and trajectory density. As can be seen from the figure, NerveNet agent is able to learn shareable features for its legs, and certain walk-cycle is learnt during training. + +# 4.4 ROBUSTNESS OF LEARNT POLICIES + +In this section, we also report the robustness of our policy by perturbing the agent parameters. In reality, the parameters simulated might be different from the actual parameters of the agents. Therefore, it is important that the agent is robust to parameters perturbation. The model that has the better ability to learn generalized features are likely more robust. + +We perturb the mass of the geometries (rigid bodies) in MuJoCo as well as the scale of the forces of the joints. We use the pre-trained models with similar performance on the original task for both the MLP and NerveNet. The performance is tested in five agents from Walker task set. The average performance is recorded in Table 3, and the specific details are summarized in Appendix 6.6. The robustness of NerveNet’ policy is likely due to the structure prior of the agent instilled in the network, which facilitates overfitting. + +![](images/26128649b00315a80503ddc2f5af6927c5694b169d26b2066dfb4ce234144b70.jpg) +Figure 11: Results of several variants of NerveNet for the reinforcement learning agents. + +# 4.5 INTERPRETING THE LEARNED REPRESENTATIONS + +In this section, we try to visualize and interpret the learned representations. We extract the final state vectors of nodes of NerveNet trained on CentipedeEight. We then apply $1 - D$ and 2- $D$ PCA on the node representations. In Figure 10, we notice that each pair of legs is able to learn invariant representations, despite their different position in the agent. We further plot the trajectory density map in the feature map. By recording the period of the walk-cycle, we plot the transformed features of the 6 legs on Figure 9. As we can see, there is a clear periodic behavior of our hidden representations learned by our model. Furthermore, the representations of adjacent left legs and the adjacent right legs demonstrate a phase shift, which further proves that our agents are able to learn the walk-cycle without any additional supervision. + +# 4.6 COMPARISON OF MODEL VARIANTS + +We have several variants of NerveNet, based on the type of network we use for the policy/value representation. We here compare all variants. Again, we run experiments for each task three times. The details of hyper-parameters are given in the Appendix. For each environment, we train the network for one million time steps, with batch size 2050 for one update. + +As we can see from Figure 11, the NerveNet-MLP and NerveNet-2 variants perform better than NerveNet-1. One potential reason is that sharing the weights of the value and policy networks makes the trust-region based optimization methods, like PPO, more sensitive to the weight $\alpha$ of the value function in equation 8. Based on the figure, choosing $\alpha$ to be 1 is not giving good performance on the tasks we experimented on. + +# 5 CONCLUSION + +In this paper, we aimed to exploit the body structure of Reinforcement Learning agents in the form of graphs. We introduced a novel model called NerveNet which uses a Graph Neural Network to represent the agent’s policy. At each time instance of the environment, NerveNet takes observations for each of the body joints, and propagates information between them using non-linear messages computed with a neural network. Propagation is done through the edges which represent natural dependencies between joints, such as physical connectivity. We experimentally showed that our NerveNet achieves comparable performance to state-of-the-art methods on standard MuJoCo environments. We further propose our customized reinforcement learning environments for benchmarking two types of structure transfer learning tasks, i.e., size and disability transfer. We demonstrate that policies learned by NerveNet are significantly better than policies learned by other models and are able to transfer even in a zero-shot setting. + +# REFERENCES + +Haitham B Ammar, Eric Eaton, Paul Ruvolo, and Matthew Taylor. Online multi-task learning for policy gradient methods. In Proceedings of the 31st International Conference on Machine Learning (ICML-14), pp. 1206–1214, 2014. + +Jacob Andreas, Dan Klein, and Sergey Levine. Modular multitask reinforcement learning with policy sketches. arXiv preprint arXiv:1611.01796, 2016. + +Hagit Attiya and Jennifer Welch. Distributed computing: fundamentals, simulations, and advanced topics, volume 19. John Wiley & Sons, 2004. + +Andre Barreto, Will Dabney, R ´ emi Munos, Jonathan J Hunt, Tom Schaul, David Silver, and Hado P ´ van Hasselt. Successor features for transfer in reinforcement learning. In Advances in Neural Information Processing Systems, pp. 4056–4066, 2017. + +Andrew A Biewener. Animal locomotion. Oxford University Press, 2003. + +Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016. + +Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs. ICLR, 2014. + +Daniele Calandriello, Alessandro Lazaric, and Marcello Restelli. Sparse multi-task reinforcement learning. In Advances in Neural Information Processing Systems, pp. 819–827, 2014. + +Yves Chauvin and David E Rumelhart. Backpropagation: theory, architectures, and applications. Psychology Press, 1995. + +Stelian Coros, Philippe Beaudoin, and Michiel Van de Panne. Generalized biped walking control. ACM Transactions on Graphics (TOG), 29(4):130, 2010. + +Alessandro Crespi and Auke Jan Ijspeert. Online optimization of swimming and crawling in an amphibious snake robot. IEEE Transactions on Robotics, 24(1):75–87, 2008. + +Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on ¨ graphs with fast localized spectral filtering. In NIPS, 2016. + +Yan Duan, John Schulman, Xi Chen, Peter L Bartlett, Ilya Sutskever, and Pieter Abbeel. Rl 2: Fast reinforcement learning via slow reinforcement learning. arXiv preprint arXiv:1611.02779, 2016. + +David K Duvenaud, Dougal Maclaurin, Jorge Iparraguirre, Rafael Bombarell, Timothy Hirzel, Alan´ Aspuru-Guzik, and Ryan P Adams. Convolutional networks on graphs for learning molecular fingerprints. In NIPS, 2015. + +Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. arXiv preprint arXiv:1703.03400, 2017a. + +Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imitation learning via meta-learning. arXiv preprint arXiv:1709.04905, 2017b. + +Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. arXiv preprint arXiv:1704.01212, 2017. + +Christoph Goller and Andreas Kuchler. Learning task-dependent distributed representations by backpropagation through structure. In Neural Networks, 1996., IEEE International Conference on, volume 1, pp. 347–352. IEEE, 1996. + +Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. In IJCNN, 2005. + +Shixiang Gu, Ethan Holly, Timothy Lillicrap, and Sergey Levine. Deep reinforcement learning for robotic manipulation with asynchronous off-policy updates. In ICRA, 2017. + +Abhishek Gupta, Coline Devin, YuXuan Liu, Pieter Abbeel, and Sergey Levine. Learning invariant feature spaces to transfer skills with reinforcement learning. arXiv preprint arXiv:1703.02949, 2017. + +Matthew Hausknecht and Peter Stone. Deep reinforcement learning in parameterized action space. arXiv preprint arXiv:1511.04143, 2015. + +Nicolas Heess, Srinivasan Sriram, Jay Lemmon, Josh Merel, Greg Wayne, Yuval Tassa, Tom Erez, Ziyu Wang, Ali Eslami, Martin Riedmiller, et al. Emergence of locomotion behaviours in rich environments. arXiv preprint arXiv:1707.02286, 2017. + +Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. ICLR, 2017. + +Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012. + +Tejas D Kulkarni, Karthik Narasimhan, Ardavan Saeedi, and Josh Tenenbaum. Hierarchical deep reinforcement learning: Integrating temporal abstraction and intrinsic motivation. In Advances in Neural Information Processing Systems, pp. 3675–3683, 2016. + +Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. J. Mach. Learn. Res., 17(1):1334–1373, January 2016. ISSN 1532-4435. URL http://dl.acm.org/citation.cfm?id=2946645.2946684. + +Ruiyu Li, Makarand Tapaswi, Renjie Liao, Jiaya Jia, Raquel Urtasun, and Sanja Fidler. Situation recognition with graph neural networks. 2017. + +Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural networks. arXiv preprint arXiv:1511.05493, 2015. + +Shingo Mabu, Kotaro Hirasawa, and Jinglu Hu. A graph-based evolutionary algorithm: Genetic network programming (gnp) and its extension using reinforcement learning. Evolutionary Computation, 15(3):369–398, 2007. + +Sridhar Mahadevan and Mauro Maggioni. Proto-value functions: A laplacian framework for learning representation and control in markov decision processes. Journal of Machine Learning Research, 8(Oct):2169–2231, 2007. + +Luke Metz, Julian Ibarz, Navdeep Jaitly, and James Davidson. Discrete sequential prediction of continuous actions for deep rl. arXiv preprint arXiv:1705.05035, 2017. + +Jan Hendrik Metzen. Learning graph-based representations for continuous reinforcement learning domains. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pp. 81–96. Springer, 2013. + +Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015. + +Junhyuk Oh, Satinder Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization with multi-task deep reinforcement learning. arXiv preprint arXiv:1706.05064, 2017. + +Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Actor-mimic: Deep multitask and transfer reinforcement learning. arXiv preprint arXiv:1511.06342, 2015. + +Xiaojuan Qi, Renjie Liao, Jiaya Jia, Sanja Fidler, and Raquel Urtasun. 3d graph neural networks for rgbd semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 5199–5208, 2017. + +Aravind Rajeswaran, Kendall Lowrey, Emanuel Todorov, and Sham Kakade. Towards generalization and simplicity in continuous control. arXiv preprint arXiv:1703.02660, 2017a. + +Aravind Rajeswaran, Kendall Lowrey, Emanuel Todorov, and Sham Kakade. Towards generalization and simplicity in continuous control. arXiv preprint arXiv:1703.02660, 2017b. + +Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE TNN, 2009. + +John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In Proceedings of the 32nd International Conference on Machine Learning (ICML-15), pp. 1889–1897, 2015a. + +John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. Highdimensional continuous control using generalized advantage estimation. arXiv preprint arXiv:1506.02438, 2015b. + +John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. + +Farzaneh Shoeleh and Masoud Asadpour. Graph based skill acquisition and transfer learning for continuous reinforcement learning domains. Pattern Recognition Letters, 87:104–116, 2017. + +David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016. + +David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go without human knowledge. Nature, 550(7676):354–359, 2017. + +Richard Socher, Cliff C Lin, Chris Manning, and Andrew Y Ng. Parsing natural scenes and natural language with recursive neural networks. In Proceedings of the 28th international conference on machine learning (ICML-11), pp. 129–136, 2011. + +Mark W Spong, Seth Hutchinson, Mathukumalli Vidyasagar, et al. Robot modeling and control, volume 3. Wiley New York, 2006. + +Richard S Sutton, David A McAllester, Satinder P Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. In Advances in neural information processing systems, pp. 1057–1063, 2000. + +Kai Sheng Tai, Richard Socher, and Christopher D Manning. Improved semantic representations from tree-structured long short-term memory networks. ACL, 2015. + +Matthew E Taylor and Peter Stone. Transfer learning for reinforcement learning domains: A survey. Journal of Machine Learning Research, 10(Jul):1633–1685, 2009. + +Yee Teh, Victor Bapst, Razvan Pascanu, Nicolas Heess, John Quan, James Kirkpatrick, Wojciech M Czarnecki, and Raia Hadsell. Distral: Robust multitask reinforcement learning. In Advances in Neural Information Processing Systems, pp. 4497–4507, 2017. + +Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ International Conference on, pp. 5026– 5033. IEEE, 2012. + +Alexander Sasha Vezhnevets, Simon Osindero, Tom Schaul, Nicolas Heess, Max Jaderberg, David Silver, and Koray Kavukcuoglu. Feudal networks for hierarchical reinforcement learning. arXiv preprint arXiv:1703.01161, 2017. + +Aaron Wilson, Alan Fern, Soumya Ray, and Prasad Tadepalli. Multi-task reinforcement learning: a hierarchical bayesian approach. In Proceedings of the 24th international conference on Machine learning, pp. 1015–1022. ACM, 2007. + +Markus Wulfmeier, Ingmar Posner, and Pieter Abbeel. Mutual alignment transfer learning. arXiv preprint arXiv:1707.07907, 2017. + +Zhaoyang Yang, Kathryn Merrick, Hussein Abbass, and Lianwen Jin. Multi-task deep reinforcement learning for continuous action control. In Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, IJCAI-17, pp. 3301–3307, 2017. + +# 6 APPENDIX + +# 6.1 DETAILS OF NERVENET + +We use MLP to compute the messages which uses tanh nonlinearities as the activation function. We do a grid search on the size of the MLP to compute the messages, the details of which are listed in Table 5, 4. + +Throughout all of our experiments, we use average aggregation and GRU as the update function. + +Table 4: Parameters used during training. + +
ParametersValue SetParameters Value Set
Value Discount Factor y.99GAE入.95
PPO Clip Value0.2Starting Learning Rate3e-4
Gradient Clip Value5.0Target KL0.01
+ +# 6.2 GRAPH OF AGENT + +In MuJoCo, we observe that most body nodes are paired with one and only one joint node. Thus, we simply merge the two paired nodes into one. We point out that this model is very compact, and is the standard graph we use in our experiments. + +In the Gym environments, observation for the joint nodes normally includes the angular velocity, twist angle and optionally the torque for the hinge joint, and position information for the positional joint. For the body nodes, velocity, inertia, and force are common observations. For example in the centipede environment 1, the LeftHip node will receive the angular velocity $\varpi _ { j }$ , and the twist angle $\theta _ { j }$ . + +# 6.3 HYPERPARAMETER SEARCH + +For MLP, we run grid search with the hidden size from two layers to three layers, and with hidden size from 32 to 256. For NerveNet, to reduce the time spent on grid search, we constrain the propagation network and output network to be the same shape. Similarly, we run grid search with the network’s hidden size, and at the same time, we run a grid search on the size of node’s hidden states from 32 to 64. For the TreeNet, we run similar grid search on the node’s hidden states and output network’s shape. + +For details of hyperparameter search, please see Table 5, 7, 6. + +Table 5: Hyperparameter grid search options for MLP. + +
MLPValue Tried
Network Shape[64,64],[128,128], [256,256],[64,64,64]
Number of Iteration Per Update10,20
Use KL PenaltyYes, No
Learning Rate SchedulerLinear Decay, Adaptive, Constant
+ +Table 6: Hyperparameter grid search options for TreeNet. + +
TreeNetValue Tried
Network Shape[64,64], [128,128], [256,256]
Number of Iteration Per Update10,20
Use KL PenaltyYes,No
Learning Rate SchedulerLinear Decay, Adaptive, Constant
+ +Table 7: Hyperparameter grid search options for NerveNet. + +
NerveNetValue Tried
Network Shape[64,64],[128,128],[256, 256]
Number of Iteration Per Update10,20
Use KL PenaltyYes,No
Learning Rate SchedulerLinear Decay, Adaptive, Constant
Number of Propogation Steps3,4,5,6
NerveNetVariantsNerveNet-1,NerveNet-2,NerveNet-MLP
Size ofNodes'Hidden State32,64,128
Merge joint and body NodeYes,No
Output NetworkShared, Separate
Disable Edge TypeYes, No
Add Skip-connection from / to rootYes,No
+ +# 6.4 SCHEMATIC FIGURES OF THE PARSED AGENTS + +In this section, we also plot the schematic figures of the agents for readers’ reference. Graph structures are automatically parsed from the MuJoCo XML configuration files. + +![](images/e7bb923f2aed9a6363a3a02049afe2ed3ae65c849658c6e2bd754b853c8d9f19.jpg) +Figure 12: Schematic diagrams and auto-parsed graph structures of the InvertedPendulum and InvertedDoublePendulum. + +![](images/3ecaa892fd0c5dde6ad0cbe96e739a085bbe439b8b996b0c352579ba4ec09a38.jpg) +Figure 13: Schematic diagrams and auto-parsed graph structures of the Walker2d and Reacher. + +![](images/cf363b16eec569ed2a1a9999406d9bf9b77be3fcd51151252e7b3aa593148326.jpg) +Figure 14: Schematic diagrams and auto-parsed graph structures of the SnakeSix and Swimmer. + +![](images/dff29462eb705451d33123a58042e9f900230477745841d1761508e310e2368a.jpg) +Figure 15: Schematic diagrams and auto-parsed graph structures of the Ant. + +![](images/24f34f799745d2698fe80b8df160044a5a1c445eeb646dbd1136e757d410d30f.jpg) +Figure 16: Schematic diagrams and auto-parsed graph structures of the Hopper and HalfCheetah. + +# 6.5 DETAILS OF ZERO-SHOT LEARNING RESULTS + +
539396449484264448423252454福
91-101-101855 8004195593351 264 L'L9-
11.971-6'061-9'S61- 52586872111475113 461-14411
46018三99 5814 4444-1103435587-848 8'I
91552929-854福F44
2233344886335929'L6163289464
2806I3品188282
19453m0222222337494988881044
983841 9491 448888
4 4164216152
4199919189LI62312441111444449五T1LI1046101224494466
4004484854469413559 54944135583封9'$9937
1 667-中458334514
5334542283348
57611-46'0-059S01-61-409m6n6
eeeeeeeeeeeeeeeeeeeeeieeeegeeeeeeeeeeeeegeeeeeteeeeeeereeneeteeteereeeeeicreeeeeereeeeeeegeeer erneeeeeeeeeeeeereeeeeeeeeeeeeeeeeercreeeneeeeeeeeeerergereeeeereeteeereeieeieneeeeeeeeeeereeeeereeeeeeeeeeeeeeeeeeeeneeeeeneeeeeeeeeeeicrrereeerreteetgereeeneeeeeeeeeetereenetrectreeeetrerreeeeeerrcegeeeeeerereeeeeeecrgepeeeeelepeeeeeeseeneeeeeeeereeeeeeaeeeeeeeeeeeeeeeeteeea
+ +
971900933554ISL9816 48124111111
福848-R488845118101888636-
13-808-668-69L-498-486-29'88-8£6L-728-4155-58888-764156-
55446 8314448F4364学3999京08329
9'SL-0388756-6'1-6499L-28756-7I1-3678-896-8678-$'96-
8386'091590556699496007844481115446548814福683515554
3314541S2453 34251863
4568833853494
L094169 48248
518588408584548443244334
23225496970829449$'0998552455413331255551835454TOIIOI03134111441455514694445918041887551957138050675545
S66L339835369558635-8三61
14444431148B6953 国5058818
2533773828361949'101326.30059296078
993打99111
eeeereeneeeeeeeeet eereereeerreeeeegereeeeereraaergeeeedeereeeergereeeeeieeeereeneereergereeereeeeeeeeeeeeeceeeeeeeeeeeeeeeeeceeeereeeeeeeeeecreareeereseeeeeeeereeeetGeepeeeeeecegepeeeeetCeeereneerecegeeeeeeeeeerereeeeeneeeereeeeeeelereeeeeeeeeeneeeeneeeeeeeeeeeeeeereeeeeeeeeeereeeecreeeeererrereeneerectgereeetireeeeeeeseeetneeeeieeeeereretreresesaeeeeeeeeeeti
+ +
1.55290.00668744444435339LL35.55759'8116.44695511.441114£99L1143423360.6778'8311.15
555418.3554318.0044.5511915678IL'LS40.64488189011117118951511.351L0116410338130
66.4441442555511.09533.35154.31435.5514445198'0L0153.59£9.09.044/.6156.559.55S1855438.5053816.7916.1460.00434
111152585-56.95-6'L[-4441-9L'9S-1135911.51555331460'8L7'01-£6.33-11.04410'61-1571-89.97-L6'67-353335331531-88'L7-40.05-
15:114353555.815531966986.69355555.3511.333335.21448.1135.3845.55335.713233219'8466.55370.3380.5535.7128.85
10028B28281055131641621328388084660
ererreresdererertrreresdeeiereeerereeeeeeertrrereeeeer erraeraseeereerteeeeeer crrrrreereeretrrereeeeerrereedreeeeerrereeeeerCereresdeeeeerrtrreresdeeererrreeereereeeeetrereedeercrrereeeerreerrereeeetrCgrrreesdrerecrraedticerseerrretrrereeeirereeeereecrreaeeeeereereeseirreeeeeeeeeeareeeeereeeiceerreeeeteseseeaEreereedeereeretrrereseeereedeeeeeeeseeeeterreresdeeeeerereeseseeeeeeresdreeseeereseeererarrgeaeedeereeeereeseeeereeeeserrreerrareseeercarereeeerreetrrreeeeerreeeeteeeeseeeeercaaeesdeteeerteseeeear
+ +In the MLP-Bind method, we bind the weights of MLP. By doing this, the weights of the agent from the similar structures will be shared. For example, in the centipede environment, the weights from observation to action of all the LeftHips are constrained to be same. + +Table 10: Results of the zero-shot learning of Snake +6.5.1 LINEAR NORMALIZATION ZERO-SHOT RESULTS AND COLORIZATION. + +
Reward Min
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v1-31.719-43.328-64.958-58.811308.994301.4
SnakeFive-v1-56.38229.262-55.562-39.346
SnakeSix-v1-44.888-5.306-42.744-76.42266.066
SnakeSeven-v1-52.764-25.511-49.765-59.52231.557329.456333.237253.828
SnakeFive-v1-56.382-30.466-49.953-47.339
SnakeSix-v1SnakeSeven-v1-44.888-35.655-52.18-57.334-90.837-52.015
-52.76489.358
Reward Std
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v19.5445.79210.76910.1465.5135.788
SnakeFive-v113.3186.19111.7157.883
SnakeSix-v111.51513.53510.30416.8697.3957.1795.856.60414.084
SnakeSeven-v111.2112.13110.9816.953
SnakeFive-v113.31852.7111.5448.5198.51923.10913.529
SnakeSix-v1SnakeSeven-v111.51511.2164.54910.68310.52811.615
Reward Max
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v13.863-19.40910.1180.918338.374326.974
SnakeFive-v120.58563.8663.136-2.329
SnakeSix-v17.658.7793.51-2.139305.011
SnakeSeven-v1-2.2851.2651.325-28.087275.536
SnakeFive-v120.585139.6175.3051.487356.366
SnakeSix-v1SnakeSeven-v17.6-2.28156.771143.8242.3162.596-4.1598.394367.768336.561
Reward Average
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v1-13.531-30.185-12.363-21.824325.476314.919
SnakeFive-v1-17.82951.411-16.481-21.256
SnakeSix-v1-19.8932.2-19.674-42.461282.426
SnakeSeven-v1-22.09923.533-22.378-42.742256.293
SnakeFive-v1-17.82989.823-16.54-15.975342.881
SnakeSix-v1-19.89105.632-20.454-34.343351.85313.149
SnakeSeven-v1-22.099115.123-22.81-22.383
+ +As the scale of zero-shot results is very different, we normalize the results across different models for each transfer learning task. For each task, we record the worst value of results from different models and the pre-set worst value $V _ { \mathrm { m i n } }$ . we set the normalization minimun value as this worst value. We calculate the normalization maximum value by $\lfloor \mathrm { m a x } ( V ) / \mathrm { I n t L e n } \rfloor$ $^ *$ IntLen. + +Table 11: Parameters for linear normalization during zero-shot results. Results are shown in the order of Centipedes’ reward, Centipedes’ running-length, Snakes’ reward + +
parametersVminIntLen
value-100,-100, -2030,30,30
+ +6.6 DETAILS OF ROBUSTNESS RESULTS +Table 12: Detail results of Robustness testing. + +
ModelMass
Hopper Halfhumanoid HorseWolfOstrich
NerveNet1981.592047.921968.872062.361852.762519.62392.052282.842161.611804.522304.251900.431862.211955.16922.12932.48983.11910.29791.84
1703.471648.281576.211354.51
1239.72946.08
MLP2268.641932.981637.471591.511073.1421.82115.75278.921224.98353.561692.081083.621932.641754.94878.75945.94
690.22538.14443.921459.931039.9814.55879.66805.99802.41
ModelStrength
HopperHalfhumanoidHorseWolfOstrich
NerveNet1961.621117.91492.33468.4446.141528.89940.96594.53308.08249.62316.82960.341827.73862.19274.77189.49133.84794.56539.45371.19255.12197.84
488.51305.16
305.16201.42
MLP800.15431.97429.76426.27412.391439.891040.341236.881730.881108.32230.85483.36
7209.54
765.7306.08519.73163.36
117.0565.29228.26187.7288.28108.97232.8133.99
+ +# 6.7 HYPERPARAMETERS FOR MULTI-TASK LEARNING + +Table 13: Hyperparameter settings for multi-task learning. + +
MLP, TreeNetand NerveNetValue Tried
Network Shape[64, 64]
Number of Iteration Per Update10
Use KL PenaltyNo
Learning Rate SchedulerAdaptive
NerveNetValue Tried
Number of Propogation Steps4
NerveNetVariants Size of Nodes'Hidden StateNerveNet-1
64
Merge joint and body NodeYes
Output Network Disable Edge Type Separate
Add Skip-connection from /to rootYes No
+ +# 6.8 SCHEMATIC FIGURES OF WA L K E R S AGENTS + +We design Walker task-set, which contains five 2d walkers in the MuJoCo engine. + +![](images/e4db0554bf395b640e337ab9dd90cfa8eba43566ee9855ed7bab06708c5c8236.jpg) +Figure 17: Schematic diagrams and auto-parsed graph structures of Walker-Ostrich and Walker-Wolf. + +![](images/24d53bd50058ae9dce42cfb0dbe3b1f6127ebefc151bad848cb175c05a53e2d4.jpg) +Figure 18: Schematic diagrams and auto-parsed graph structures of Walker-Hopper and Walker-HalfHumanoid. + +![](images/9de69acaa7f53ccf56bad5e7e49119e3e4eb8f81606d03c69dfcfd1a0fa3d1be.jpg) +Figure 19: Schematic diagrams and auto-parsed graph structures of Walker-Horse. \ No newline at end of file diff --git a/parse/train/S1sqHMZCb/S1sqHMZCb_content_list.json b/parse/train/S1sqHMZCb/S1sqHMZCb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..995a70db93c510a0ea0a4b3b805d85800cfba45f --- /dev/null +++ b/parse/train/S1sqHMZCb/S1sqHMZCb_content_list.json @@ -0,0 +1,2628 @@ +[ + { + "type": "text", + "text": "NERVENET: LEARNING STRUCTURED POLICY WITH GRAPH NEURAL NETWORKS ", + "text_level": 1, + "bbox": [ + 174, + 98, + 821, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tingwu Wang∗, Renjie Liao∗, Jimmy Ba & Sanja Fidler ", + "bbox": [ + 184, + 170, + 570, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computer Science \nUniversity of Toronto \nVector Institute \n{tingwuwang,rjliao}@cs.toronto.edu, \njimmy@psi.toronto.edu, fidler@cs.toronto.edu ", + "bbox": [ + 184, + 188, + 617, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 290, + 544, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We address the problem of learning structured policies for continuous control. In traditional reinforcement learning, policies of agents are learned by multi-layer perceptrons (MLPs) which take the concatenation of all observations from the environment as input for predicting actions. In this work, we propose NerveNet to explicitly model the structure of an agent, which naturally takes the form of a graph. Specifically, serving as the agent’s policy network, NerveNet first propagates information over the structure of the agent and then predict actions for different parts of the agent. In the experiments, we first show that our NerveNet is comparable to state-of-the-art methods on standard MuJoCo environments. We further propose our customized reinforcement learning environments for benchmarking two types of structure transfer learning tasks, i.e., size and disability transfer, as well as multi-task learning. We demonstrate that policies learned by NerveNet are significantly more transferable and generalizable than policies learned by other models and are able to transfer even in a zero-shot setting. ", + "bbox": [ + 233, + 321, + 764, + 516 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 542, + 334, + 558 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep reinforcement learning (RL) has received increasing attention over the past few years, with the recent success of applications such as playing Atari Games, Mnih et al. (2015), and Go, (Silver et al., 2016; 2017). Significant advances have also been made in robotics using the latest RL techniques, e.g., Levine et al. (2016); Gu et al. (2017). ", + "bbox": [ + 174, + 574, + 825, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many RL problems feature agents with multiple dependent controllers. For example, humanoid robots consist of multiple physically linked joints. Action to be taken by each joint or the body should thus not only depend on its own observations but also on actions of other joints. ", + "bbox": [ + 174, + 637, + 823, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous approaches in RL typically use MLP to learn the agent’s policy. In particular, MLP takes the concatenation of observations from the environment as input, which may be measurements like positions, velocities of body and joints in the current time instance. The MLP policy then predicts actions to be taken by every joint and body. Thus the task of the MLP policy is to discover the latent relationships between observations. This typically leads to longer training times, requiring more exposure of the agent to the environment. In our work, we aim to exploit the body structure of an agent, and physical dependencies that naturally exist in such agents. ", + "bbox": [ + 174, + 685, + 825, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We rely on the fact that bodies of most robots and animals have a discrete graph structure. Nodes of the graph may represent the joints, and edges represent the (physical) dependencies between them. In particular, we define the agent’s policy using a Graph Neural Network, Scarselli et al. (2009), which is a neural network that operates over graph structures. We refer to our model as NerveNet due to the resemblance of the neural nervous system to a graph. NerveNet propagates information between different parts of the body based on the underlying graph structure before outputting the action for each part. By doing so, NerveNet can leverage the structure information encoded by the agent’s body which is advantageous in learning the correct inductive bias, and thus is less prone to overfitting. Moreover, NerveNet is naturally suitable for structure transferring tasks as most of the model weights are shared across the nodes and edges, respectively. ", + "bbox": [ + 174, + 790, + 825, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ec7ba80f49404fcff9389c3659fcdd445c72eb87234ed2686b8a6262e8e85922.jpg", + "image_caption": [ + "Figure 1: Visualization of the graph structure of CentipedeEight in our environment. We use this agent for testing the ability of transfer learning of our model. Since for this agent, each body node is paired with at least one joint node, we omit the body nodes and fill up the position with the corresponding joint nodes. By omitting the body nodes, a more compact graph is constructed, the details of which are illustrated in the experimental section. " + ], + "image_footnote": [], + "bbox": [ + 187, + 113, + 802, + 279 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 404, + 823, + 433 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first evaluate our NerveNet on standard RL benchmarks such as the OpenAI Gym, Brockman et al. (2016) which stem from MuJoCo. We show that our model achieves comparable results to state-of-the-art MLP based methods. To verify our claim regarding the structure transfer, we further introduce our customized RL environments which are based on the ones of Gym. Two types of structure transfer tasks are designed, size transfer and disability transfer. In particular, size transfer focuses on the scenario in which policies are learned for small-sized agents (simpler body structure) and applied directly to large-sized agents which are composed by some repetitive components shared with the small-sized agent. Secondly, disability transfer investigates scenarios in which policies are learned for one agent and applied to the same agent with some components disabled. Our experiments demonstrate that for structure transfer tasks our NerveNet is significantly better than all other competitors, and can even achieve zero-shot learning for some agents. For the multi-task learning tasks, NerveNet is also able to learn policies that are more robust with better efficiency. ", + "bbox": [ + 174, + 439, + 825, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contribution of this paper is the following: We explore the problem of learning transferable and generalized features by incorporating a prior on the structure via graph neural networks. NerveNet permits powerful transfer learning from one structure to another, which goes well beyond the ability of previous models. NerveNet is also more robust and has more potential in performing multi-task learning. The demo and code for this project are released, under the project page of http://www.cs.toronto.edu/˜tingwuwang/nervenet.html. ", + "bbox": [ + 174, + 612, + 825, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 NERVENET ", + "text_level": 1, + "bbox": [ + 174, + 734, + 300, + 751 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first introduce the notation. We then explain how to construct the graph for each of our agents, followed by the description of the NerveNet. Finally, we describe the learning algorithm for our model. ", + "bbox": [ + 176, + 776, + 823, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We formulate the locomotion control problems as an infinite-horizon discounted Markov decision process (MDP). To fully describe the MDP for continuous control problems which include locomotion control, we define the state space or observation space as $s$ and action space as $\\mathcal { A }$ . To interact with the environments, the agent generates its stochastic policy $\\pi _ { \\boldsymbol { \\theta } } ( \\boldsymbol { a } ^ { \\tau } | \\boldsymbol { s } ^ { \\tau } )$ based on the current state $s ^ { \\tau } \\in \\boldsymbol { S }$ , where $a ^ { \\tau } \\in \\mathcal { A }$ is the action and $\\theta$ are the parameters of the policy function. The environment on the other hand, produces a reward $r ( s ^ { \\tau } , a ^ { \\tau } )$ for the agent, and the agent’s objective is to find a policy that maximizes the expected reward. ", + "bbox": [ + 174, + 825, + 825, + 924 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4387bb37b2b6edf82cb3edff6c91e9f09012e99f1638fa5f05208b0a537fe83a.jpg", + "image_caption": [ + "Figure 2: In this figure, we use Walker-Ostrich as an example of NerveNet. In the input model, for each node, NerveNet fetches the corresponding elements from the observation vector. NerveNet then computes the messages between neighbors in the graph, and updates the hidden state of each node. This process is repeated for a certain number of propagation steps. In the output model, the policy is produced by collecting the output from each controller. " + ], + "image_footnote": [], + "bbox": [ + 179, + 109, + 820, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 GRAPH CONSTRUCTION ", + "text_level": 1, + "bbox": [ + 174, + 401, + 382, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In real life, skeletons of most robots and animals have a discrete graph structure, and are most often trees. Simulators such as the MuJoCo engine by Todorov et al. (2012), organize the agents using an XML-based kinematic tree. In our experiments, we will use the tree graphs as per MuJoCo. Note that our model can be directly applied to arbitrary graphs. In particular, we assume two types of nodes in our tree: body and joint. The body nodes are abstract nodes used to construct the kinematic tree via nesting, which is similar to the coordinate frame system used in robotics, Spong et al. (2006). The joint node represents the degrees of freedom of motion between the two body nodes. Take a simple humanoid as an example; the body nodes Thigh and Shin are connected via the Knee, where Knee is a hinge joint. We further add a root node which observes additional information about the agent. For example, in the Reacher agent in MuJoCo, the root node has access to the target position of the agent. We build edges to form a tree graph. Fig. 1 illustrates the graph structure of an example agent, CentipedeEight. We sketch the agent and its corresponding graph in the left and right part of the figure, respectively. Note that for better visualization, we omit the joint nodes and use edges to represent the physical connections of joint nodes. Different elements of the agent are parsed into nodes with different colors. Further details are provided in the experimental section. ", + "bbox": [ + 173, + 428, + 825, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 NERVENET AS POLICY NETWORK ", + "text_level": 1, + "bbox": [ + 178, + 667, + 449, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now turn to NerveNet which parametrizes the policy with a Graph Neural Network. Before delving into details, we first introduce our notation. We then specify the input model which helps to initialize the hidden state of each node. We further introduce the propagation model that updates these hidden states. Finally, we describe the output model. ", + "bbox": [ + 174, + 694, + 825, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We denote the graph structure of the agent as $G = ( V , E )$ where $V$ and $E$ are the sets of nodes and edges, respectively. We focus on the directed graphs as the undirected case can be easily addressed by splitting one undirected edge into two directed edges. We denote the out-going neighborhood of node $u$ as $\\mathcal { N } _ { o u t } ( u )$ which contains all endpoints $v$ with $( u , v )$ being an edge in the graph. Similarly, we denote the in-coming neighborhood of node $u$ as $\\mathcal { N } _ { i n } ( u )$ . Every node $u$ has an associated node type $p _ { u } \\in \\{ 1 , 2 , \\ldots , P \\}$ , which in our case corresponds to body, joint and root. We also associate each edge $( u , v )$ with an edge type $c _ { ( u , v ) } \\in \\{ 1 , 2 , \\ldots , C \\}$ . Node type can help in capturing different importances across nodes. Edge type can be used to describe different relationships between nodes, and thus propagate information between them differently. One can also add more than one edge type to the same edge which results in a multi-graph. We stick to simple graphs for simplicity. One interesting fact is that we have two notions of “time” in our model. One is the time step in the environment which is the typical time coordinate for RL problems. The other corresponds to the internal propagation step of NerveNet. These two coordinates work as follows. At each time step of the environment, NerveNet receives observation from the environment and performs a few internal propagation steps in order to decide on the action to be taken by each node. To avoid confusion, throughout this paper, we use $\\tau$ to describe the time step in the environment and $t$ for the propagation step. ", + "bbox": [ + 173, + 757, + 825, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2.1 INPUT MODEL ", + "text_level": 1, + "bbox": [ + 174, + 188, + 328, + 203 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each time step $\\tau$ in the environment, the agent receives an observation $s ^ { \\tau } \\in \\boldsymbol { S }$ . The observation vector $s ^ { \\tau }$ is the concatenation of observations of each node. We denote the elements of observation vector $s ^ { \\tau }$ corresponding to node $u$ with $x _ { u }$ . From now on, we drop the time step in the environment to derive the model for simplicity. The observation vector goes through an input network to obtain a fixed-size state vector as follows: ", + "bbox": [ + 173, + 212, + 825, + 282 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/4d3d1bea999308a51e89d71addd9f2cddc61e7d0b475a53154f5b75a5f56b894.jpg", + "text": "$$\nh _ { u } ^ { 0 } = F _ { \\mathrm { i n } } ( x _ { u } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 449, + 286, + 547, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the subscript and superscript denote the node index and propagation step, respectively. Here, $F _ { \\mathrm { i n } }$ may be a MLP and $h _ { u } ^ { 0 }$ is the state vector of node $u$ at propagation step 0. Note that we may need to pad zeros to the observation vectors if different nodes have observations of different sizes. ", + "bbox": [ + 174, + 309, + 825, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2.2 PROPAGATION MODEL ", + "text_level": 1, + "bbox": [ + 174, + 366, + 385, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now describe the propagation model of our NerveNet which mimics a synchronous message passing system studied in distributed computing, Attiya & Welch (2004). We will show how the state vector of each node is updated from one propagation step to the next. This update process is recurrently applied during the whole propagation. We leave the details to the appendix. ", + "bbox": [ + 173, + 391, + 825, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Message Computation In particular, at propagation step $t$ , for every node $u$ , we have access to a state vector $h _ { u } ^ { t }$ . For every edge $( u , v ) \\in \\bar { \\mathcal { N } } _ { o u t } \\bar { ( u ) }$ , node $u$ computes a message vector as below, ", + "bbox": [ + 174, + 462, + 825, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/9a00b23aa896ebed9b90bc2eab833d6bda16d764875375b3e56d783dd56d6394.jpg", + "text": "$$\nm _ { ( u , v ) } ^ { t } = M _ { c _ { ( u , v ) } } ( h _ { u } ^ { t } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 421, + 494, + 575, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where Mc(u,v) i s the message function which may be an identity mapping or a MLP. Note that the subscript $c _ { ( u , v ) }$ indicates that edges of the same edge type share the same instance of the message function. For example, the second torso in Fig. 1 sends a message to the first and third torso, as well as the LeftHip and RightHip. ", + "bbox": [ + 174, + 518, + 825, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Message Aggregation Once every node finishes computing messages, we aggregate messages sent from all in-coming neighbors of each node. Specifically, for every node $u$ , we perform the following aggregation: ", + "bbox": [ + 176, + 590, + 825, + 632 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/6da0a44d915afbd5d8e27938b4c7ead1a6c9099a095d41838359532a55d3c41d.jpg", + "text": "$$\n\\bar { m } _ { u } ^ { t } = A ( \\{ h _ { v } ^ { t } | v \\in \\mathcal { N } _ { i n } ( u ) \\} ) ,\n$$", + "text_format": "latex", + "bbox": [ + 401, + 636, + 594, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $A$ is the aggregation function which may be a summation, average or max-pooling function. Here, $\\hat { m } _ { u } ^ { t }$ is the aggregated message vector which contains the information sent from the node’s neighborhood. ", + "bbox": [ + 174, + 660, + 825, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "States Update We now update every node’s state vector based on both the aggregated message and its current state vector. In particular, for every node $u$ , we perform the following update: ", + "bbox": [ + 171, + 717, + 821, + 746 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/5730230f07c18c79391a9f1b4a21e7d335ee5509bbe644efaaa609de68afa92a.jpg", + "text": "$$\nh _ { u } ^ { t + 1 } = U _ { p _ { u } } ( h _ { u } ^ { t } , \\bar { m } _ { u } ^ { t } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 424, + 750, + 571, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $U$ is the update function which may be a gated recurrent unit (GRU), a long short term memory (LSTM) unit or a MLP. From the subscript $p _ { u }$ of $U$ , we can see that nodes of the same node type share the same instance of the update function. The above propagation model is then recurrently applied for a fixed number of time steps $T$ to get the final state vectors of all nodes, i.e., $\\{ h _ { u } ^ { T } | u \\in \\bar { V } \\}$ . ", + "bbox": [ + 174, + 772, + 825, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2.3 OUTPUT MODEL ", + "text_level": 1, + "bbox": [ + 174, + 857, + 344, + 871 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In RL, agents typically use a MLP policy, where the network outputs the mean of the Gaussian distribution for each of the actions, while the standard deviation is a trainable vector, Schulman et al. (2017). In our output model, we also treat standard deviation in the same way. ", + "bbox": [ + 174, + 881, + 823, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, instead of predicting the action distribution of all nodes by a single network, we make predictions for each individual node. We denote the set of nodes which are assigned controllers for the actuators as $\\mathcal { O }$ . For each such node, a MLP takes its final state vectors $h _ { u \\in \\mathcal { O } } ^ { T }$ as input and produces the mean of the action of the Gaussian policy for the corresponding actuator. For each output node $u \\in \\mathcal { O }$ , we define its output type as $q _ { u }$ . Different sharing schemes are available for the instance of MLP $O _ { q _ { u } }$ , for example, we can force the nodes with similar physical structure to share the instance of MLP. For example, in Fig. 1, two LeftHip nodes have a shared controller. Therefore, we have the following output model: ", + "bbox": [ + 173, + 102, + 825, + 215 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/c1ca0583550ae71a4838cf1240500c3fb1ccac16546004d7251de3096fd83fd4.jpg", + "text": "$$\n\\mu _ { u \\in \\mathcal { O } } = O _ { q _ { u } } ( h _ { u } ^ { T } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 436, + 215, + 562, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu _ { u \\in \\mathcal { O } }$ is the mean value for action applied on each actuator. In practice, we found that we can force controllers of different output types to share one unified controller, while not hurting the performance. By integrating the produced Gaussian policy for each action, the probability density of the stochastic policy is calculated as ", + "bbox": [ + 174, + 236, + 825, + 291 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/d8dcb57eac4cd1d7463487642c75c2932c5bdd93450154d9ad19f794116027ac.jpg", + "text": "$$\n\\pi _ { \\boldsymbol { \\theta } } ( a ^ { \\tau } \\vert s ^ { \\tau } ) = \\prod _ { \\boldsymbol { u } \\in \\mathcal { O } } \\pi _ { \\boldsymbol { \\theta } , \\boldsymbol { u } } ( a _ { \\boldsymbol { u } } ^ { \\tau } \\vert s ^ { \\tau } ) = \\prod _ { \\boldsymbol { u } \\in \\mathcal { O } } \\frac { 1 } { \\sqrt { 2 \\pi \\sigma _ { \\boldsymbol { u } } ^ { 2 } } } e ^ { ( a _ { \\boldsymbol { u } } ^ { \\tau } - \\mu _ { \\boldsymbol { u } } ) ^ { 2 } / ( 2 \\sigma _ { \\boldsymbol { u } } ^ { 2 } ) } ,\n$$", + "text_format": "latex", + "bbox": [ + 290, + 292, + 705, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $a ^ { \\tau } \\in { \\mathcal { A } }$ is the output action, and $\\sigma _ { u }$ is the variable standard deviation for each action. Here, $\\theta$ represents the parameters of the policy function. ", + "bbox": [ + 174, + 332, + 823, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 LEARNING ALGORITHM ", + "text_level": 1, + "bbox": [ + 174, + 376, + 382, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To interact with the environments, the agent generates its stochastic policy $\\pi _ { \\boldsymbol { \\theta } } ( \\boldsymbol { a } ^ { \\tau } | \\boldsymbol { s } ^ { \\tau } )$ after several propagation steps. The environment on the other hand, produces a reward $r ( s ^ { \\tau } , a ^ { \\tau } )$ for the agent, and transits to the next state with transition probability $\\mathsf { \\bar { P } } ( s ^ { \\tau + 1 } | s ^ { \\tau } )$ . The target of the agent is to maximize its cumulative return ", + "bbox": [ + 173, + 401, + 825, + 457 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/c85c65cb21c8b32e0696ec38231bd699b0c5d41bc53454e1f3598e7ad47274a1.jpg", + "text": "$$\nJ ( \\theta ) = \\mathbb { E } _ { \\pi } \\left[ \\sum _ { \\tau = 0 } ^ { \\infty } \\gamma ^ { \\tau } r ( s ^ { \\tau } , a ^ { \\tau } ) \\right] .\n$$", + "text_format": "latex", + "bbox": [ + 393, + 455, + 604, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To optimize the expected reward, we use the proximal policy optimization (PPO) by Schulman et al. (2017). In PPO, the agents alternate between sampling trajectories with the latest policy and performing optimization on surrogate objective using the sampled trajectories. The algorithm tries to keep the KL-divergence of the new policy and the old policy within the trust region. To achieve that, PPO clips the probability ratio and adds an KL-divergence penalty term to the loss. The likelihood ratio is defined as $r ^ { \\tau } ( \\boldsymbol { \\theta } ; \\mathbf { \\dot { \\theta } } _ { o l d } ) = \\pi _ { \\boldsymbol { \\theta } } ( a ^ { \\tau } | s ^ { \\tau } ) / \\pi _ { \\theta _ { o l d } } ( a ^ { \\tau } | \\mathbf { \\bar { s } } ^ { \\tau } )$ . Following the notation and the algorithm of PPO, our NerveNet tries to minimize the summation of the original loss in Eq. (7), KL-penalty and the value function loss which is defined as: ", + "bbox": [ + 173, + 506, + 825, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/41caae5f10177c97ef816aa61d192768e07967153ed851eeedad01034bed3999.jpg", + "text": "$$\n\\begin{array} { l } { \\displaystyle \\tilde { J } ( \\theta ) = J ( \\theta ) - \\beta L _ { K L } ( \\theta ) - \\alpha L _ { V } ( \\theta ) } \\\\ { \\displaystyle \\qquad = \\mathbb { E } _ { \\pi _ { \\theta } } \\left[ \\sum _ { \\tau = 0 } ^ { \\infty } \\operatorname* { m i n } \\left( \\hat { A } ^ { \\tau } r ^ { \\tau } ( \\theta ) , \\hat { A } ^ { \\tau } \\operatorname { c l i p } \\left( r ^ { \\tau } ( \\theta ) , 1 - \\epsilon , 1 + \\epsilon \\right) \\right) \\right] } \\\\ { \\displaystyle \\qquad - \\beta \\mathbb { E } _ { \\pi _ { \\theta } } \\left[ \\sum _ { \\tau = 0 } ^ { \\infty } \\mathrm { K L } \\left[ \\pi _ { \\theta } ( \\cdot | s ^ { \\tau } ) | \\pi _ { \\theta _ { o l d } } ( \\cdot | s ^ { \\tau } ) \\right] \\right] - \\alpha \\mathbb { E } _ { \\pi _ { \\theta } } \\left[ \\sum _ { \\tau = 0 } ^ { \\infty } \\left( V _ { \\theta } ( s ^ { \\tau } ) - V ( s ^ { \\tau } ) ^ { \\mathrm { t a r g e t } } \\right) ^ { 2 } \\right] , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 189, + 616, + 787, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat { A } _ { t }$ is the generalized advantage estimation (GAE) calculated using algorithm from Schulman et al. (2015b), and $\\epsilon$ is the clip value, which we choose to be 0.2. Here, $\\beta$ is a dynamical coefficient adjusted to keep the KL-divergence constraints, and $\\alpha$ is used to balance the value loss. Note that in Eq. (8), $V ( s _ { t } ) ^ { \\mathrm { t a r g e t } }$ is the target state value in accordance with the GAE method. To optimize the $\\tilde { J } ( \\theta )$ , PPO make use of the policy gradient in Sutton et al. (2000) to do first-order gradient descent optimization. ", + "bbox": [ + 173, + 726, + 825, + 813 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Value Network To produce the state value $V _ { \\theta } ( s ^ { \\tau } )$ for given observation $s ^ { \\tau }$ , we have several alternatives: (1) using one GNN as the policy network and using one MLP as the value network (NerveNet-MLP); (2) using one GNN as policy network and using another GNN as value network (NerveNet-2) (without sharing the parameters of the two GNNs); (3) using one GNN as both policy network and value network (NerveNet-1). The GNN for value network is very similar to the GNN for policy network. The output for value GNN is a scalar instead of a vector of mean action. We will compare these variants in the experimental section. ", + "bbox": [ + 173, + 825, + 825, + 924 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 102, + 344, + 118 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reinforcement Learning Reinforcement learning (RL) has recently achieved huge success in a variety of applications. Powered by the progress of deep neural networks, Krizhevsky et al. (2012), agents are now able to successfully play Atari Games and beat the world’s best (human) players in the game of Go (Mnih et al., 2015; Silver et al., 2016; 2017). Based on simulation engines like MuJoCo, Todorov et al. (2012), numerous algorithms have been proposed to train agents also in continuous control problems (Schulman et al., 2017; 2015a; Heess et al., 2017; Metz et al., 2017). ", + "bbox": [ + 174, + 135, + 825, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Structure in RL Most approaches that exploit priors on structure of the problem fall in the domain of hierarchical RL, (Kulkarni et al., 2016; Vezhnevets et al., 2017), which mainly focus on modeling intrinsic motivation of agents. In Hausknecht & Stone (2015), the authors extend the deep RL algorithms to MDPS with parameterized action space by exploiting the structure of action space and bounding the action space gradients. Graphs have been used in RL problems prior to our work. In Metzen (2013); Mabu et al. (2007); Shoeleh & Asadpour (2017); Mahadevan & Maggioni (2007), the authors use graphs to learn a representation of the environment. However, these methods are limited to problems with simple dynamical models like for example the task of 2d-navigation, and thus these problems are usually solved via model-based RL. However, for complex multi-joint agents, learning the dynamical model as well as predicting the transition of states is time consuming and biased. For problems of training model-free multi-joint agents in complex physical environments, relatively little attention has been devoted to modeling the physical structure of the agents. ", + "bbox": [ + 174, + 234, + 825, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Graph Neural Networks There have been many efforts to generalize neural networks to graphstructured data. One line of work is based on convolutional neural networks (CNNs). In (Bruna et al., 2014; Defferrard et al., 2016; Kipf & Welling, 2017), CNNs are employed in the spectral domain relying on the graph Laplacian matrix. (Goller & Kuchler, 1996; Duvenaud et al., 2015) used hash functions in order to apply CNNs to graphs. ", + "bbox": [ + 174, + 420, + 825, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Another popular direction is based on recurrent neural networks (RNNs) (Goller & Kuchler, 1996; Gori et al., 2005; Scarselli et al., 2009; Socher et al., 2011; Li et al., 2015; Tai et al., 2015). Among RNN based methods, many are only applicable to special structured graph, e.g., sequences or trees, (Socher et al., 2011; Tai et al., 2015). One class of models which are applicable to general graphs are so-called graph neural networks (GNNs), Scarselli et al. (2009). The inference procedure is a forward pass that exploits a fixed-length propagation process which resembles synchronous message passing system in the theory of distributed computing, Attiya & Welch (2004). Nodes in the graph have state vectors which are recurrently updated based on their history and received messages. One of the representative work of GNNs, i.e., gated graph neural networks (GGNNs) by Li et al. (2015), uses gated recurrent unit to update the state vectors. Learning such a model can be achieved by the back-propagation through time (BPTT) algorithm or recurrent back-propagation, Chauvin & Rumelhart (1995). It has been shown that GNNs, (Li et al., 2015; 2017; Qi et al., 2017; Gilmer et al., 2017) have a high capacity and achieve state-of-the-art performance in many applications which involve graph-structured data. In this paper, we model the structure of the reinforcement learning agents using GNNs. ", + "bbox": [ + 174, + 496, + 825, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Transfer and Multi-task Learning in RL Recently, there has been increased interest in transfer learning tasks for RL, Taylor & Stone (2009), which mainly focus on transferring the policy learned from one environment to another. In Rajeswaran et al. (2017b;a), the authors show that agents in reinforcement learning are prone to over-fitting, and that the learned policies generalize poorly across environments. In model-based RL, traditional control has been well studied for generalization properties, Coros et al. (2010). Gupta et al. (2017) try to increase the transferability via learning invariant visual features. Efforts have also been made from the meta-learning perspective (Duan et al., 2016; Finn et al., 2017b;a). In Wulfmeier et al. (2017), the authors propose a method of transfer learning by using imitation learning. Transferability comes naturally in our model by exploiting the (shared) graph structure of the agents. ", + "bbox": [ + 174, + 722, + 825, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-task learning has also received a lot of attention, Wilson et al. (2007). In Teh et al. (2017), the authors use a distilled policy that captures common behaviour across tasks. Wilson et al. (2007); Oh et al. (2017); Andreas et al. (2016) use a hierarchical approach, where multiple sub-policies are learnt. In Yang et al. (2017); Parisotto et al. (2015), the authors exploited shared visual features to tackle multi-task learning. In Ammar et al. (2014), a multi-task policy gradient is proposed, while, Calandriello et al. (2014) propose multi-task extensions of the fitted Q-iteration algorithm. Successor features in Barreto et al. (2017) are used to boost the performance of multiple tasks. Unlike ours, these methods do not exploit the physical graph structures of the agents. ", + "bbox": [ + 176, + 867, + 823, + 924 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/31914ef34c44e5cb9c9bdb09331fb9917e4b0134cd1f05a2719e5c2b4c19287a.jpg", + "image_caption": [ + "Figure 3: Results of MLP, TreeNet and NerveNet on 8 continuous control benchmarks from the Gym. " + ], + "image_footnote": [], + "bbox": [ + 184, + 106, + 808, + 443 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 512, + 825, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 590, + 326, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we first verify the effectiveness of NerveNet on standard MuJoCo environments in OpenAI Gym. We then investigate the transfer abilities of NerveNet and other competitors by customizing some of those environments, as well as the multi-task learning ability and robustness. ", + "bbox": [ + 174, + 622, + 825, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 COMPARISON ON STANDARD BENCHMARKS OF MUJOCO ", + "text_level": 1, + "bbox": [ + 173, + 681, + 611, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines We compare NerveNet with the standard MLP models utilized by Schulman et al. (2017) and another baseline which is constructed as follows. We first remove the physical graph structure and introduce an additional super node which connects to all nodes in the graph. This results in a singly rooted depth-1 tree. We refer to this baseline as TreeNet. The propagation model of TreeNet is similar to NerveNet where, however, the policy first aggregates the information from all children and then feeds the state vector of the root to the output model. This simpler model serves as a baseline to verify the importance of the graph structure. ", + "bbox": [ + 174, + 708, + 825, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We run experiments on 8 simulated continuous control benchmarks from the Gym, Brockman et al. (2016), which is based on MuJoCo, Todorov et al. (2012). In particular, we use Reacher, InvertedPendulum, InvertedDoublePendulum, Swimmer, and four walking or running tasks: HalfCheetah, Hopper, Walker2d, Ant. We set the maximum number of training steps to be 1 million for all environments as it is enough to solve them. Note that for InvertedPendulum, different from the original one in Gym, we add the distance penalty of the cart and velocity penalty so that the reward is more consistent to the InvertedDoublePendulum. This change of design also makes the task more challenging. ", + "bbox": [ + 174, + 813, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/670ab56a6ac4211ba1a235ac8454c4b3f3398ff13994de22b67b272be84cc5d7.jpg", + "table_caption": [ + "Table 1: Performance of the Pre-trained models on CentipedeFour and CentipedeSix. " + ], + "table_footnote": [], + "table_body": "
C_FourReward AvgStdMaxC_SixReward AvgStdMax
NerveNet2799.91247.23833.9NerveNet2507.1738.42979.2
MLP2398.51390.43936.3MLP2793.01005.23529.5
TreeNet1429.6957.73021.7TreeNet2229.71109.43727.4
", + "bbox": [ + 184, + 127, + 813, + 199 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results We do grid search to find the best hyperparameters and leave the details in the Appendix 6.3. As the randomness might have a big impact on the performance, for each environment, we run 3 experiments with different random seeds and plot the average curves and the standard deviations. We show the results in Figure 3. From the figures, we can see that MLP with the same setup as in Schulman et al. (2017) works the best in most of tasks.1 NerveNet basically matches the performance of MLP in terms of sample efficiency as well as the performance after it converges. In most cases, the TreeNet is worse than NerveNet which highlights the importance of keeping the physical graph structure. ", + "bbox": [ + 173, + 224, + 825, + 337 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 STRUCTURE TRANSFER LEARNING ", + "text_level": 1, + "bbox": [ + 176, + 354, + 459, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We now benchmark our model in the task of structure transfer learning by creating customized environments based on the existing ones from MuJoCo. We mainly investigate two types of structure transfer learning tasks. The first one is to train a model with an agent of small size (small graph) and apply the learned model to an agent with a larger size, i.e., size transfer. When increasing the size of the agent, observation and action space also increase which makes learning more challenging. Another type of structure transfer learning is disability transfer where we first learn a model for the original agent and then apply it to the same agent with some components disabled. If one model overfits the environment, disabling some components of the agent might bring catastrophic performance degradation. Note that for both transfer tasks, all factors of environments do not change except the structure of the agent. ", + "bbox": [ + 174, + 380, + 825, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Centipede We create the first environment in which the agent has a similar structure to a centipede. The goal of the agent is to run as fast as possible along the $y$ -direction in the MuJoCo environment. The agent consists of repetitive torso bodies where each one has two legs attached. For two consecutive bodies, we add two actuators which control the rotation between them. Furthermore, each leg consists of a thigh and shin, which are controlled by two hinge actuators. By linking copies of torso bodies and corresponding legs, we create agents with different lengths. Specifically, the shortest Centipede is CentipedeFour and the longest one is CentipedeFourty due to the limit of supported resource of MuJoCo. For each time step, the total reward is the speed reward minus the energy cost and force feedback from the ground. Note that in practice, we found that training a CentipedeEight from scratch is already very difficult. For size transfer experiments, we create many instances which are listed in Figure 4, like “4to06”, “6to10”. For disability transfer, we create CrippleCentipede agents of which two back legs are disabled. In Figure 4, CrippleCentipede is specified as “Cp”. ", + "bbox": [ + 173, + 535, + 825, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Snakes We also create a snake-like agent which is common in robotics, Crespi & Ijspeert (2008). We design the Snake environment based on the Swimmer model in Gym. The goal of the agent is to move as fast as possible. For details of the environment, please see the schematic figure 16. ", + "bbox": [ + 176, + 732, + 825, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.1 EXPERIMENTAL SETTINGS ", + "text_level": 1, + "bbox": [ + 176, + 790, + 413, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To fully investigate the performance of NerveNet, we build several baseline models for structure transfer learning which are explained below. ", + "bbox": [ + 174, + 814, + 823, + 842 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "NerveNet For the NerveNet, since all the weights are exactly the same for the small and the largeagent models, we directly use the old weights trained on the small-agent model. When the large agent has repetitive structure, we further re-use the weights of the corresponding joints from the small-agent model. ", + "bbox": [ + 176, + 858, + 821, + 886 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ded2354d5c50c0404c254caa82cabadd97bd1f063a74fb66cd5313b24ebd4bd0.jpg", + "image_caption": [ + "Figure 4: Performance of zero-shot learning on centipedes. For each task, we run the policy for 100 episodes and record the average reward and average length the agent runs before falling down. " + ], + "image_footnote": [], + "bbox": [ + 173, + 101, + 834, + 347 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1b4f2a5fa06b4d78a1b4be689e625ce9271bf372ee76f157e6a19ba72f40e380.jpg", + "image_caption": [ + "Figure 5: (a), (b): Results of fine-tuning for size transfer experiments. (c), (d) Results of fine-tuning for disability transfer experiments. " + ], + "image_footnote": [], + "bbox": [ + 186, + 414, + 792, + 776 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 845, + 823, + 875 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "MLP Pre-trained (MLPP) For the MLP based model, while transferring from one structure to another, the size of the input layer changes since the size of the observation changes. One straightforward idea is to reuse the weights from the first hidden layer to the output layer and randomly initialize the weights of the new input layer. ", + "bbox": [ + 173, + 895, + 823, + 924 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/84ea5296a5d035d9829cab325bf3532db657dcf4811e266ee0b813b65808af7f.jpg", + "image_caption": [ + "Figure 6: Results on zero-shot transfer learning on snake agents. Each tasks are simulated for 100 episodes. " + ], + "image_footnote": [], + "bbox": [ + 236, + 103, + 761, + 217 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/75955e4c44eb72c36a84acb11321fa2796dcfc7110c692596195fe9c0701f3c7.jpg", + "image_caption": [ + "Figure 7: Results of finetuning on snake environments. " + ], + "image_footnote": [], + "bbox": [ + 186, + 284, + 805, + 652 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 708, + 823, + 737 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "MLP Activation Assigning (MLPAA) Another way of making MLP transferable is assigning the weights of the small-agent model to the corresponding partial weights of the large-agent model and setting the remaining weights to be zero. Note that we do not add or remove any layers from the small-agent model to the large-agent except for changing the size of the layers. By doing so, we can keep the output of the large-agent model to be same as the small-agent in the beginning, i.e., keeping the same initial policy. ", + "bbox": [ + 174, + 756, + 825, + 842 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "TreeNet TreeNet is similar as the model described before. We apply the same way of assigning weights as MLPAA to TreeNet for the transfer learning task. ", + "bbox": [ + 174, + 861, + 821, + 888 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Random We also include the random policy which is uniformly sampled from the action space. ", + "bbox": [ + 171, + 909, + 810, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2.2 RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 103, + 292, + 117 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Centipedes For the Centipedes environment, we first run experiments of all models on CentipedeSix and CentipedeFour to get the pre-trained models for transfer learning. We train different models until these agents run equally well as possible, which is reported in Table 1. Note that, in practice, we train TreeNet on CentipedeFour for more than 8 million time steps. However, due to the difficulty of optimizing TreeNet on CentipedeFour, the performance is still lower. But visually, the TreeNet agent is able to run in CentipedeFour. ", + "bbox": [ + 174, + 130, + 825, + 213 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We then examine the zero-shot performance where zero-shot means directly applying the model trained with one setting to the other without any fine-tuning. To better visualize the results, we linearly normalize the performance to get a performance score, and color the results accordingly. The normalization scheme is recorded in Appendix 11. The performance score is less than 1, and is shown in the parentheses behind the original results. As we can see from Figure 4 (full chart in Appendix 6.5), NerveNet outperforms all competitors on all settings, except in the 4toCp06 scenario. Note that transferring from CentipedeFour is more difficult than from CentipedeSix since the situation where one torso connects to two neighboring torsos only happens beyond 4 bodies. ", + "bbox": [ + 174, + 220, + 825, + 332 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "TreeNet has a surprisingly good performance on tasks from CentipedeFour. However, by checking the videos, the learned agent is actually not able to “move” as good as other methods. The high reward is mainly due to the fact that TreeNet policy is better at standing still and gaining alive bonus. We argue that the average running-length in each episode is also a very important metric. ", + "bbox": [ + 176, + 338, + 825, + 395 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "By including the results of running-length, we notice that NerveNet is the only model able to walk in the zero-shot setting. In fact, the performance of NerveNet is orders-of-magnitude better, and most of the time, agents from other methods cannot even move forward. We also notice that if transferred from CentipedeSix, NerveNet is able to provide walkable pre-trained models on all new agents. ", + "bbox": [ + 174, + 401, + 825, + 457 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We fine-tune for both size transfer and disability transfer experiments and show the training curves in Figure 5. From the figure, we can see that by using the pre-trained model, NerveNet significantly decreases the number of episodes required to reach the level of reward which is considered as solved. By looking at the videos, we notice that the bottleneck of learning for the agent is “how to stand”. When training from scratch, it can be seen that almost 0.5 million time steps are spent on a very flat reward surface. Therefore, the MLPAA agents, which copy the learned policy, are able to stand and bypass this time-consuming process and reach to a good performance in the end. ", + "bbox": [ + 174, + 464, + 825, + 561 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Moreover, by examining the result videos, we noticed that the “walk-cycle” behavior is observed for NerveNet but is not common for others. Walk-cycle are adopted for many insects in the world, Biewener (2003). For example, six-legged ants use a tripedal gait, where the legs are used in two separate triangles alternatively touching the ground. We give more details of walk-cycle in Section 4.5. ", + "bbox": [ + 174, + 569, + 823, + 638 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "One possible reason is that the agent of MLP based method (MLPAA, MLPP) learns a policy that does not utilize all legs. From CentipedeEight and up, we do not observe any MLP agents to be able to coordinate all legs whereas almost all policies learned by NerveNet use all legs. Therefore, NerveNet is better at utilizing structure information and not over-fitting the environments. ", + "bbox": [ + 174, + 645, + 825, + 702 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Snakes The zero-shot performance for snakes is summarized in Figure 6. As we can see, NerveNet has the best performance on all transfer learning tasks. In most cases, NerveNet has a starting reward value of more than 300, which is a pretty good policy since 350 is considered as solved for snakeThree. By looking at the videos, we found that agents of other competitors are not able to control the new actuators in the zero-shot setting. They either overfit to the original models, where the policy is completely useless in the new setting (e.g., the MLPAA is worse than random policy in SnakeThree2SnakeFour), or the new actuators are not able to coordinate with the old actuators trained before. While for NerveNet, the actuators are able to coordinate to its neighbors, regardless of whether they are new to the agents. ", + "bbox": [ + 173, + 722, + 825, + 847 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We also summarize the training curves of fine-tuning in Fig. 7. We can observe that NerveNet has a very good initialization with the pre-trained model, and the performance increases with fine-tuning. When training from scratch, NerveNet is less sample efficient compared to the MLP model which might be caused by the fact that optimizing our model is more challenging than MLP. Fine-tuning helps to improve the sample efficiency of our model by a large margin. At the same time, although the MLPAA has a very good initialization, its performance progresses slowly with the increasing number of episodes. In most experiments, the MLPAA and TreeNet did not match the performance of its non-pretrained MLP baseline. ", + "bbox": [ + 176, + 854, + 823, + 924 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3709a850e281bcae1eeed9ee523b904fff989358725e2076c690a230315f97b0.jpg", + "image_caption": [ + "Figure 8: Results of Multi-task learning. We train the networks simultaneously on five different tasks from Walker. " + ], + "image_footnote": [], + "bbox": [ + 189, + 106, + 797, + 323 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7bceb7a5394f06e3831463e6553b8f644ac3399036bf2990a8f84efe73728009.jpg", + "table_caption": [ + "Table 2: Results of Multi-task learning, with comparison to the single-task baselines. For three models, the first row is the mean reward of each model of the last 40 iterations. The second row indicates the percentage of the performance of the multi-task model compared with the single-task baseline of each model. " + ], + "table_footnote": [], + "table_body": "
Model HalfHumanoidHopperOstrichWolfHorseAverage
MLPRewardRatio1775.7557.7%1369.5962.0%1198.8848.2%1249.2354.5%2084.0769.7%/58.6%
TreeNetRewardRatio237.8179.3%417.2798.0%224.0757.4%247.03141.2%223.3499.2%/94.8%
NerveNetRewardRatio2536.5296.3%2113.56101.8%1714.6398.8%2054.54105.9%2343.62106.4%1101.8%
", + "bbox": [ + 178, + 385, + 820, + 511 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 611, + 825, + 654 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 MULTI-TASK LEARNING ", + "text_level": 1, + "bbox": [ + 174, + 679, + 383, + 693 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we show that NerveNet has a good potential of multi-task learning by incorporating structure prior into the network structure. It is important to point out that multi-task learning represents a very difficult, and more often the case, unsolved problem in RL. Most multi-task learning algorithms, Teh et al. (2017); Andreas et al. (2016); Oh et al. (2017); Yang et al. (2017) have not been applied to domains as difficult as locomotion for complex physical models, not to mention multi-task learning among different agents with different dynamics. ", + "bbox": [ + 174, + 708, + 825, + 791 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this work, we constrain our problem domain, and design the Walker multi-task learning taskset, which contains five 2d-walkers. We aim to test the model’s ability of multi-task learning, in particular, the ability to control multiple agents using one unified network. The walkers are very different in terms of their dynamics, since they have very distinct structures, different types and numbers of controllers. Walker-HalfHumanoid and Walker-Hopper are variants of Walker2d and Hopper from the original MuJoCo Benchmarks, respectively. Walker-Horse, Walker-Ostrich, Walker-Wolf on the other hand, are agents mimicking the natural animals. Just like the real animals, some of the agents have tails or necks to help them to balance. The detailed schematic figures are shown in the Appendix 6.8. ", + "bbox": [ + 173, + 797, + 825, + 924 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/024ef3d5511d4e580856abdd8bfa2049dc00cf405bbe5b20dc053da6a84fa458.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelHalfhumanoid HopperWolfOstrich HorseAverage
MassMLPNerveNet33.28%95.87%74.04%93.24%94.68%90.13%59.23% 40.61%80.2% 69.23%60.37%85.73%
StrengthMLPNerveNet25.96%31.11%21.77%42.20%27.32%42.84%30.08% 19.80%31.41% 36.54%24.99%36.82%
", + "bbox": [ + 181, + 102, + 815, + 193 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 3: Results of robustness evaluations. Note that we show the average results for each type of parameters after perturbation. And the results are columned by the agent type. The ratio of the average performance of perturbed agents and the original performance is shown in the figure. Details are listed in 6.6. ", + "bbox": [ + 174, + 203, + 825, + 258 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3.1 EXPERIMENTAL SETTINGS ", + "text_level": 1, + "bbox": [ + 176, + 290, + 413, + 305 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To show the ability of multi-task learning of NerveNet, we design several baselines. We use a vanilla multi-task policy update for all models. More specifically, for each sub-task in the multi-task learning task-set, we use an equal number of time steps for each policy’s update and calculate the gradients separately. Gradients are then aggregated and the mean value of gradients is applied to update the network. To compensate for the additional difficulty in training more agents and tasks, we linearly increase the number of update epochs during each update in training, as well as the total number of time steps generated before the training is terminated. The hyper-parameter setting is summarized in Appendix 6.7. ", + "bbox": [ + 174, + 316, + 825, + 429 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "NerveNet For NerveNet, the weights are naturally shared among different agents. More specifically, for different agents, the weight matrices for propagation and output are shared. ", + "bbox": [ + 173, + 449, + 821, + 477 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "MLP Sharing For the MLP method, we shared the weight matrices between hidden layers. ", + "bbox": [ + 173, + 497, + 782, + 512 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "MLP Aggregation In the MLP Sharing approach, the total size of the weight matrices grows with the number of tasks. For different agents, whose dimension of the observations are usually different, weights from observation to the first hidden layer cannot be reused in the MLP Sharing approach. Therefore in the MLP Aggregation method, we multiply each element of the observation vector separately by one matrix, and aggregate the resulting vectors from each element. The size of this multiplying matrix is (1, dimension of the first hidden layer). ", + "bbox": [ + 174, + 532, + 825, + 616 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "TreeNet Similarly, TreeNet also has the benefits that its weights are naturally shared among different agents. However, TreeNet has no knowledge of the agents’ physical structure, where the information of each node is aggregated into the root node. ", + "bbox": [ + 176, + 637, + 825, + 679 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We also include the baseline of training single-task MLP for each agent. We train the single-task MLP baselines for 1 million time steps per agent. In the Figure 8, we align the results of single-task MLP baseline and the results of multi-task models by the number of episodes of one task. ", + "bbox": [ + 176, + 686, + 825, + 728 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As can be seen from Figure 8, NerveNet achieves the best performance in all the sub-tasks. In Walker-HalfHumanoid, Walker-Hopper, Walker-Ostrich, Walker-Wolf our NerveNet is able to out-perform other agents by a large margin. In Walker-Horse, the performance of NerveNet and MLP Sharing are relatively similar. For MLP Sharing, the performance on other four agents are relatively limited, while for Walker-Hopper, the improvement of performance is limited from half of the experiment onwards. The MLP Aggregation and TreeNet methods are not able to solve the multi-task learning problem, with both of them stuck at a very low reward level. In the vanilla optimization setting, we show that NerveNet has a bigger potential than the baselines. ", + "bbox": [ + 174, + 734, + 825, + 847 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "From Table 2, one can observe that the performance of MLP drops drastically $42 \\%$ performance drop) when switching from single-task to multi-task learning, while for NerveNet, there is no obvious drop in performance. Our intuition is that NerveNet is better at learning generalized features, and learning of different agents can help in training other agents, while for MLP methods, the performance decreases due the competition of different agents. ", + "bbox": [ + 174, + 854, + 823, + 924 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/892025ae723031d5978ddef92b55588d72f89132d24844c0cfb5e117cf25e298.jpg", + "image_caption": [ + "Figure 9: Diagram of the walk cycle. In the left figure, legs within the same triangle are used simultaneously. For each leg, we use the same color for their diagram on the left and their curves on the right. " + ], + "image_footnote": [], + "bbox": [ + 197, + 116, + 799, + 236 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/652d63005f4eea342561cb0a44ebd4dd31b388fd4be3b385a751118a06287c04.jpg", + "image_caption": [ + "Figure 10: Results of visualization of feature distribution and trajectory density. As can be seen from the figure, NerveNet agent is able to learn shareable features for its legs, and certain walk-cycle is learnt during training. " + ], + "image_footnote": [], + "bbox": [ + 187, + 335, + 803, + 643 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 ROBUSTNESS OF LEARNT POLICIES ", + "text_level": 1, + "bbox": [ + 176, + 743, + 460, + 757 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this section, we also report the robustness of our policy by perturbing the agent parameters. In reality, the parameters simulated might be different from the actual parameters of the agents. Therefore, it is important that the agent is robust to parameters perturbation. The model that has the better ability to learn generalized features are likely more robust. ", + "bbox": [ + 174, + 776, + 825, + 833 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We perturb the mass of the geometries (rigid bodies) in MuJoCo as well as the scale of the forces of the joints. We use the pre-trained models with similar performance on the original task for both the MLP and NerveNet. The performance is tested in five agents from Walker task set. The average performance is recorded in Table 3, and the specific details are summarized in Appendix 6.6. The robustness of NerveNet’ policy is likely due to the structure prior of the agent instilled in the network, which facilitates overfitting. ", + "bbox": [ + 174, + 840, + 825, + 924 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/26128649b00315a80503ddc2f5af6927c5694b169d26b2066dfb4ce234144b70.jpg", + "image_caption": [ + "Figure 11: Results of several variants of NerveNet for the reinforcement learning agents. " + ], + "image_footnote": [], + "bbox": [ + 192, + 106, + 799, + 212 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "4.5 INTERPRETING THE LEARNED REPRESENTATIONS ", + "text_level": 1, + "bbox": [ + 174, + 265, + 560, + 280 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "In this section, we try to visualize and interpret the learned representations. We extract the final state vectors of nodes of NerveNet trained on CentipedeEight. We then apply $1 - D$ and 2- $D$ PCA on the node representations. In Figure 10, we notice that each pair of legs is able to learn invariant representations, despite their different position in the agent. We further plot the trajectory density map in the feature map. By recording the period of the walk-cycle, we plot the transformed features of the 6 legs on Figure 9. As we can see, there is a clear periodic behavior of our hidden representations learned by our model. Furthermore, the representations of adjacent left legs and the adjacent right legs demonstrate a phase shift, which further proves that our agents are able to learn the walk-cycle without any additional supervision. ", + "bbox": [ + 174, + 291, + 825, + 416 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "4.6 COMPARISON OF MODEL VARIANTS ", + "text_level": 1, + "bbox": [ + 178, + 434, + 464, + 448 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We have several variants of NerveNet, based on the type of network we use for the policy/value representation. We here compare all variants. Again, we run experiments for each task three times. The details of hyper-parameters are given in the Appendix. For each environment, we train the network for one million time steps, with batch size 2050 for one update. ", + "bbox": [ + 176, + 460, + 823, + 515 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "As we can see from Figure 11, the NerveNet-MLP and NerveNet-2 variants perform better than NerveNet-1. One potential reason is that sharing the weights of the value and policy networks makes the trust-region based optimization methods, like PPO, more sensitive to the weight $\\alpha$ of the value function in equation 8. Based on the figure, choosing $\\alpha$ to be 1 is not giving good performance on the tasks we experimented on. ", + "bbox": [ + 174, + 523, + 825, + 593 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "5 CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 613, + 318, + 630 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "In this paper, we aimed to exploit the body structure of Reinforcement Learning agents in the form of graphs. We introduced a novel model called NerveNet which uses a Graph Neural Network to represent the agent’s policy. At each time instance of the environment, NerveNet takes observations for each of the body joints, and propagates information between them using non-linear messages computed with a neural network. Propagation is done through the edges which represent natural dependencies between joints, such as physical connectivity. We experimentally showed that our NerveNet achieves comparable performance to state-of-the-art methods on standard MuJoCo environments. We further propose our customized reinforcement learning environments for benchmarking two types of structure transfer learning tasks, i.e., size and disability transfer. We demonstrate that policies learned by NerveNet are significantly better than policies learned by other models and are able to transfer even in a zero-shot setting. ", + "bbox": [ + 174, + 645, + 825, + 797 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 819, + 285, + 834 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Haitham B Ammar, Eric Eaton, Paul Ruvolo, and Matthew Taylor. Online multi-task learning for policy gradient methods. In Proceedings of the 31st International Conference on Machine Learning (ICML-14), pp. 1206–1214, 2014. ", + "bbox": [ + 176, + 843, + 823, + 886 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Jacob Andreas, Dan Klein, and Sergey Levine. Modular multitask reinforcement learning with policy sketches. arXiv preprint arXiv:1611.01796, 2016. ", + "bbox": [ + 173, + 895, + 820, + 924 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Hagit Attiya and Jennifer Welch. Distributed computing: fundamentals, simulations, and advanced topics, volume 19. John Wiley & Sons, 2004. ", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Andre Barreto, Will Dabney, R ´ emi Munos, Jonathan J Hunt, Tom Schaul, David Silver, and Hado P ´ van Hasselt. Successor features for transfer in reinforcement learning. In Advances in Neural Information Processing Systems, pp. 4056–4066, 2017. ", + "bbox": [ + 176, + 142, + 821, + 185 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Andrew A Biewener. Animal locomotion. Oxford University Press, 2003. ", + "bbox": [ + 173, + 195, + 656, + 212 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016. ", + "bbox": [ + 176, + 222, + 825, + 251 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs. ICLR, 2014. ", + "bbox": [ + 174, + 261, + 823, + 290 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Daniele Calandriello, Alessandro Lazaric, and Marcello Restelli. Sparse multi-task reinforcement learning. In Advances in Neural Information Processing Systems, pp. 819–827, 2014. ", + "bbox": [ + 173, + 300, + 823, + 330 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Yves Chauvin and David E Rumelhart. Backpropagation: theory, architectures, and applications. Psychology Press, 1995. ", + "bbox": [ + 173, + 339, + 821, + 369 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Stelian Coros, Philippe Beaudoin, and Michiel Van de Panne. Generalized biped walking control. ACM Transactions on Graphics (TOG), 29(4):130, 2010. ", + "bbox": [ + 171, + 378, + 823, + 409 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Alessandro Crespi and Auke Jan Ijspeert. Online optimization of swimming and crawling in an amphibious snake robot. IEEE Transactions on Robotics, 24(1):75–87, 2008. ", + "bbox": [ + 174, + 419, + 821, + 448 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on ¨ graphs with fast localized spectral filtering. In NIPS, 2016. ", + "bbox": [ + 174, + 458, + 821, + 488 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Yan Duan, John Schulman, Xi Chen, Peter L Bartlett, Ilya Sutskever, and Pieter Abbeel. Rl 2: Fast reinforcement learning via slow reinforcement learning. arXiv preprint arXiv:1611.02779, 2016. ", + "bbox": [ + 174, + 497, + 823, + 527 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "David K Duvenaud, Dougal Maclaurin, Jorge Iparraguirre, Rafael Bombarell, Timothy Hirzel, Alan´ Aspuru-Guzik, and Ryan P Adams. Convolutional networks on graphs for learning molecular fingerprints. In NIPS, 2015. ", + "bbox": [ + 176, + 537, + 823, + 580 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. arXiv preprint arXiv:1703.03400, 2017a. ", + "bbox": [ + 169, + 590, + 823, + 619 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imitation learning via meta-learning. arXiv preprint arXiv:1709.04905, 2017b. ", + "bbox": [ + 173, + 630, + 823, + 660 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. arXiv preprint arXiv:1704.01212, 2017. ", + "bbox": [ + 173, + 670, + 821, + 699 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Christoph Goller and Andreas Kuchler. Learning task-dependent distributed representations by backpropagation through structure. In Neural Networks, 1996., IEEE International Conference on, volume 1, pp. 347–352. IEEE, 1996. ", + "bbox": [ + 174, + 709, + 823, + 752 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. In IJCNN, 2005. ", + "bbox": [ + 173, + 762, + 821, + 791 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Shixiang Gu, Ethan Holly, Timothy Lillicrap, and Sergey Levine. Deep reinforcement learning for robotic manipulation with asynchronous off-policy updates. In ICRA, 2017. ", + "bbox": [ + 169, + 803, + 823, + 832 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Abhishek Gupta, Coline Devin, YuXuan Liu, Pieter Abbeel, and Sergey Levine. Learning invariant feature spaces to transfer skills with reinforcement learning. arXiv preprint arXiv:1703.02949, 2017. ", + "bbox": [ + 173, + 842, + 823, + 883 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Matthew Hausknecht and Peter Stone. Deep reinforcement learning in parameterized action space. arXiv preprint arXiv:1511.04143, 2015. ", + "bbox": [ + 174, + 895, + 821, + 922 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Nicolas Heess, Srinivasan Sriram, Jay Lemmon, Josh Merel, Greg Wayne, Yuval Tassa, Tom Erez, Ziyu Wang, Ali Eslami, Martin Riedmiller, et al. Emergence of locomotion behaviours in rich environments. arXiv preprint arXiv:1707.02286, 2017. ", + "bbox": [ + 176, + 103, + 821, + 146 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. ICLR, 2017. ", + "bbox": [ + 173, + 156, + 821, + 185 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012. ", + "bbox": [ + 174, + 195, + 825, + 238 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Tejas D Kulkarni, Karthik Narasimhan, Ardavan Saeedi, and Josh Tenenbaum. Hierarchical deep reinforcement learning: Integrating temporal abstraction and intrinsic motivation. In Advances in Neural Information Processing Systems, pp. 3675–3683, 2016. ", + "bbox": [ + 174, + 248, + 826, + 291 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. J. Mach. Learn. Res., 17(1):1334–1373, January 2016. ISSN 1532-4435. URL http://dl.acm.org/citation.cfm?id=2946645.2946684. ", + "bbox": [ + 178, + 301, + 821, + 344 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Ruiyu Li, Makarand Tapaswi, Renjie Liao, Jiaya Jia, Raquel Urtasun, and Sanja Fidler. Situation recognition with graph neural networks. 2017. ", + "bbox": [ + 171, + 354, + 823, + 383 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural networks. arXiv preprint arXiv:1511.05493, 2015. ", + "bbox": [ + 174, + 393, + 823, + 422 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Shingo Mabu, Kotaro Hirasawa, and Jinglu Hu. A graph-based evolutionary algorithm: Genetic network programming (gnp) and its extension using reinforcement learning. Evolutionary Computation, 15(3):369–398, 2007. ", + "bbox": [ + 174, + 433, + 825, + 476 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Sridhar Mahadevan and Mauro Maggioni. Proto-value functions: A laplacian framework for learning representation and control in markov decision processes. Journal of Machine Learning Research, 8(Oct):2169–2231, 2007. ", + "bbox": [ + 174, + 486, + 825, + 530 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Luke Metz, Julian Ibarz, Navdeep Jaitly, and James Davidson. Discrete sequential prediction of continuous actions for deep rl. arXiv preprint arXiv:1705.05035, 2017. ", + "bbox": [ + 169, + 540, + 823, + 569 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Jan Hendrik Metzen. Learning graph-based representations for continuous reinforcement learning domains. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pp. 81–96. Springer, 2013. ", + "bbox": [ + 176, + 578, + 823, + 622 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015. ", + "bbox": [ + 174, + 631, + 825, + 675 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Junhyuk Oh, Satinder Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization with multi-task deep reinforcement learning. arXiv preprint arXiv:1706.05064, 2017. ", + "bbox": [ + 173, + 684, + 821, + 714 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Actor-mimic: Deep multitask and transfer reinforcement learning. arXiv preprint arXiv:1511.06342, 2015. ", + "bbox": [ + 173, + 723, + 825, + 753 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Xiaojuan Qi, Renjie Liao, Jiaya Jia, Sanja Fidler, and Raquel Urtasun. 3d graph neural networks for rgbd semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 5199–5208, 2017. ", + "bbox": [ + 176, + 762, + 825, + 806 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Aravind Rajeswaran, Kendall Lowrey, Emanuel Todorov, and Sham Kakade. Towards generalization and simplicity in continuous control. arXiv preprint arXiv:1703.02660, 2017a. ", + "bbox": [ + 171, + 815, + 823, + 845 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Aravind Rajeswaran, Kendall Lowrey, Emanuel Todorov, and Sham Kakade. Towards generalization and simplicity in continuous control. arXiv preprint arXiv:1703.02660, 2017b. ", + "bbox": [ + 171, + 856, + 825, + 885 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE TNN, 2009. ", + "bbox": [ + 176, + 895, + 821, + 924 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In Proceedings of the 32nd International Conference on Machine Learning (ICML-15), pp. 1889–1897, 2015a. ", + "bbox": [ + 174, + 103, + 823, + 146 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. Highdimensional continuous control using generalized advantage estimation. arXiv preprint arXiv:1506.02438, 2015b. ", + "bbox": [ + 174, + 155, + 821, + 196 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. ", + "bbox": [ + 171, + 205, + 823, + 234 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Farzaneh Shoeleh and Masoud Asadpour. Graph based skill acquisition and transfer learning for continuous reinforcement learning domains. Pattern Recognition Letters, 87:104–116, 2017. ", + "bbox": [ + 173, + 244, + 821, + 273 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016. ", + "bbox": [ + 174, + 281, + 825, + 325 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go without human knowledge. Nature, 550(7676):354–359, 2017. ", + "bbox": [ + 178, + 333, + 823, + 376 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Richard Socher, Cliff C Lin, Chris Manning, and Andrew Y Ng. Parsing natural scenes and natural language with recursive neural networks. In Proceedings of the 28th international conference on machine learning (ICML-11), pp. 129–136, 2011. ", + "bbox": [ + 173, + 385, + 823, + 429 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Mark W Spong, Seth Hutchinson, Mathukumalli Vidyasagar, et al. Robot modeling and control, volume 3. Wiley New York, 2006. ", + "bbox": [ + 173, + 436, + 823, + 465 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Richard S Sutton, David A McAllester, Satinder P Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. In Advances in neural information processing systems, pp. 1057–1063, 2000. ", + "bbox": [ + 173, + 473, + 823, + 517 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Kai Sheng Tai, Richard Socher, and Christopher D Manning. Improved semantic representations from tree-structured long short-term memory networks. ACL, 2015. ", + "bbox": [ + 171, + 526, + 823, + 555 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Matthew E Taylor and Peter Stone. Transfer learning for reinforcement learning domains: A survey. Journal of Machine Learning Research, 10(Jul):1633–1685, 2009. ", + "bbox": [ + 169, + 564, + 823, + 593 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Yee Teh, Victor Bapst, Razvan Pascanu, Nicolas Heess, John Quan, James Kirkpatrick, Wojciech M Czarnecki, and Raia Hadsell. Distral: Robust multitask reinforcement learning. In Advances in Neural Information Processing Systems, pp. 4497–4507, 2017. ", + "bbox": [ + 178, + 602, + 823, + 645 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ International Conference on, pp. 5026– 5033. IEEE, 2012. ", + "bbox": [ + 174, + 652, + 825, + 695 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Alexander Sasha Vezhnevets, Simon Osindero, Tom Schaul, Nicolas Heess, Max Jaderberg, David Silver, and Koray Kavukcuoglu. Feudal networks for hierarchical reinforcement learning. arXiv preprint arXiv:1703.01161, 2017. ", + "bbox": [ + 174, + 704, + 825, + 747 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Aaron Wilson, Alan Fern, Soumya Ray, and Prasad Tadepalli. Multi-task reinforcement learning: a hierarchical bayesian approach. In Proceedings of the 24th international conference on Machine learning, pp. 1015–1022. ACM, 2007. ", + "bbox": [ + 173, + 756, + 823, + 799 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Markus Wulfmeier, Ingmar Posner, and Pieter Abbeel. Mutual alignment transfer learning. arXiv preprint arXiv:1707.07907, 2017. ", + "bbox": [ + 173, + 808, + 821, + 837 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Zhaoyang Yang, Kathryn Merrick, Hussein Abbass, and Lianwen Jin. Multi-task deep reinforcement learning for continuous action control. In Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, IJCAI-17, pp. 3301–3307, 2017. ", + "bbox": [ + 174, + 845, + 825, + 888 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "6 APPENDIX ", + "text_level": 1, + "bbox": [ + 174, + 102, + 294, + 118 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "6.1 DETAILS OF NERVENET ", + "text_level": 1, + "bbox": [ + 174, + 132, + 382, + 147 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "We use MLP to compute the messages which uses tanh nonlinearities as the activation function. We do a grid search on the size of the MLP to compute the messages, the details of which are listed in Table 5, 4. ", + "bbox": [ + 174, + 159, + 825, + 202 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Throughout all of our experiments, we use average aggregation and GRU as the update function. ", + "bbox": [ + 169, + 208, + 803, + 223 + ], + "page_idx": 18 + }, + { + "type": "table", + "img_path": "images/cd0f977cb6b65e68b1f84e18ebc157e379e0f5aa01cfbb48a18959ac79d15233.jpg", + "table_caption": [ + "Table 4: Parameters used during training. " + ], + "table_footnote": [], + "table_body": "
ParametersValue SetParameters Value Set
Value Discount Factor y.99GAE入.95
PPO Clip Value0.2Starting Learning Rate3e-4
Gradient Clip Value5.0Target KL0.01
", + "bbox": [ + 238, + 261, + 758, + 334 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "6.2 GRAPH OF AGENT ", + "text_level": 1, + "bbox": [ + 174, + 364, + 341, + 378 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "In MuJoCo, we observe that most body nodes are paired with one and only one joint node. Thus, we simply merge the two paired nodes into one. We point out that this model is very compact, and is the standard graph we use in our experiments. ", + "bbox": [ + 176, + 390, + 825, + 433 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "In the Gym environments, observation for the joint nodes normally includes the angular velocity, twist angle and optionally the torque for the hinge joint, and position information for the positional joint. For the body nodes, velocity, inertia, and force are common observations. For example in the centipede environment 1, the LeftHip node will receive the angular velocity $\\varpi _ { j }$ , and the twist angle $\\theta _ { j }$ . ", + "bbox": [ + 174, + 440, + 825, + 511 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "6.3 HYPERPARAMETER SEARCH ", + "text_level": 1, + "bbox": [ + 176, + 526, + 411, + 541 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "For MLP, we run grid search with the hidden size from two layers to three layers, and with hidden size from 32 to 256. For NerveNet, to reduce the time spent on grid search, we constrain the propagation network and output network to be the same shape. Similarly, we run grid search with the network’s hidden size, and at the same time, we run a grid search on the size of node’s hidden states from 32 to 64. For the TreeNet, we run similar grid search on the node’s hidden states and output network’s shape. ", + "bbox": [ + 174, + 551, + 825, + 637 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "For details of hyperparameter search, please see Table 5, 7, 6. ", + "bbox": [ + 173, + 643, + 576, + 659 + ], + "page_idx": 18 + }, + { + "type": "table", + "img_path": "images/aadeeb33b22fb2a35899a91dd3d3759eb93db7322b47f9e4482bc890a3f03f0f.jpg", + "table_caption": [ + "Table 5: Hyperparameter grid search options for MLP. " + ], + "table_footnote": [], + "table_body": "
MLPValue Tried
Network Shape[64,64],[128,128], [256,256],[64,64,64]
Number of Iteration Per Update10,20
Use KL PenaltyYes, No
Learning Rate SchedulerLinear Decay, Adaptive, Constant
", + "bbox": [ + 236, + 696, + 761, + 784 + ], + "page_idx": 18 + }, + { + "type": "table", + "img_path": "images/54379f7c00f8dfbccabd71c68389ee3e2d2fd69249803c8b5cb5dad8f723c3e8.jpg", + "table_caption": [ + "Table 6: Hyperparameter grid search options for TreeNet. " + ], + "table_footnote": [], + "table_body": "
TreeNetValue Tried
Network Shape[64,64], [128,128], [256,256]
Number of Iteration Per Update10,20
Use KL PenaltyYes,No
Learning Rate SchedulerLinear Decay, Adaptive, Constant
", + "bbox": [ + 263, + 829, + 733, + 915 + ], + "page_idx": 18 + }, + { + "type": "table", + "img_path": "images/be8567db64be48d464b8a43cd7cecfa1a64df314fa0bf6ed3e82da8181a39196.jpg", + "table_caption": [ + "Table 7: Hyperparameter grid search options for NerveNet. " + ], + "table_footnote": [], + "table_body": "
NerveNetValue Tried
Network Shape[64,64],[128,128],[256, 256]
Number of Iteration Per Update10,20
Use KL PenaltyYes,No
Learning Rate SchedulerLinear Decay, Adaptive, Constant
Number of Propogation Steps3,4,5,6
NerveNetVariantsNerveNet-1,NerveNet-2,NerveNet-MLP
Size ofNodes'Hidden State32,64,128
Merge joint and body NodeYes,No
Output NetworkShared, Separate
Disable Edge TypeYes, No
Add Skip-connection from / to rootYes,No
", + "bbox": [ + 220, + 127, + 777, + 310 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "6.4 SCHEMATIC FIGURES OF THE PARSED AGENTS ", + "text_level": 1, + "bbox": [ + 173, + 335, + 540, + 349 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "In this section, we also plot the schematic figures of the agents for readers’ reference. Graph structures are automatically parsed from the MuJoCo XML configuration files. ", + "bbox": [ + 174, + 361, + 823, + 390 + ], + "page_idx": 19 + }, + { + "type": "image", + "img_path": "images/e7bb923f2aed9a6363a3a02049afe2ed3ae65c849658c6e2bd754b853c8d9f19.jpg", + "image_caption": [ + "Figure 12: Schematic diagrams and auto-parsed graph structures of the InvertedPendulum and InvertedDoublePendulum. " + ], + "image_footnote": [], + "bbox": [ + 210, + 416, + 790, + 582 + ], + "page_idx": 19 + }, + { + "type": "image", + "img_path": "images/3ecaa892fd0c5dde6ad0cbe96e739a085bbe439b8b996b0c352579ba4ec09a38.jpg", + "image_caption": [ + "Figure 13: Schematic diagrams and auto-parsed graph structures of the Walker2d and Reacher. " + ], + "image_footnote": [], + "bbox": [ + 194, + 661, + 802, + 797 + ], + "page_idx": 19 + }, + { + "type": "image", + "img_path": "images/cf363b16eec569ed2a1a9999406d9bf9b77be3fcd51151252e7b3aa593148326.jpg", + "image_caption": [ + "Figure 14: Schematic diagrams and auto-parsed graph structures of the SnakeSix and Swimmer. " + ], + "image_footnote": [], + "bbox": [ + 196, + 123, + 795, + 250 + ], + "page_idx": 20 + }, + { + "type": "image", + "img_path": "images/dff29462eb705451d33123a58042e9f900230477745841d1761508e310e2368a.jpg", + "image_caption": [ + "Figure 15: Schematic diagrams and auto-parsed graph structures of the Ant. " + ], + "image_footnote": [], + "bbox": [ + 220, + 320, + 785, + 559 + ], + "page_idx": 20 + }, + { + "type": "image", + "img_path": "images/24f34f799745d2698fe80b8df160044a5a1c445eeb646dbd1136e757d410d30f.jpg", + "image_caption": [ + "Figure 16: Schematic diagrams and auto-parsed graph structures of the Hopper and HalfCheetah. " + ], + "image_footnote": [], + "bbox": [ + 200, + 625, + 799, + 824 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "6.5 DETAILS OF ZERO-SHOT LEARNING RESULTS ", + "text_level": 1, + "bbox": [ + 176, + 104, + 529, + 117 + ], + "page_idx": 21 + }, + { + "type": "table", + "img_path": "images/ff9980b0373dbcf1a9519042fb2fe2eb47d30158b75f9c8499c1e844a1ce6ac9.jpg", + "table_caption": [ + "", + "" + ], + "table_footnote": [], + "table_body": "
539396449484264448423252454福
91-101-101855 8004195593351 264 L'L9-
11.971-6'061-9'S61- 52586872111475113 461-14411
46018三99 5814 4444-1103435587-848 8'I
91552929-854福F44
2233344886335929'L6163289464
2806I3品188282
19453m0222222337494988881044
983841 9491 448888
4 4164216152
4199919189LI62312441111444449五T1LI1046101224494466
4004484854469413559 54944135583封9'$9937
1 667-中458334514
5334542283348
57611-46'0-059S01-61-409m6n6
eeeeeeeeeeeeeeeeeeeeeieeeegeeeeeeeeeeeeegeeeeeteeeeeeereeneeteeteereeeeeicreeeeeereeeeeeegeeer erneeeeeeeeeeeeereeeeeeeeeeeeeeeeeercreeeneeeeeeeeeerergereeeeereeteeereeieeieneeeeeeeeeeereeeeereeeeeeeeeeeeeeeeeeeeneeeeeneeeeeeeeeeeicrrereeerreteetgereeeneeeeeeeeeetereenetrectreeeetrerreeeeeerrcegeeeeeerereeeeeeecrgepeeeeelepeeeeeeseeneeeeeeeereeeeeeaeeeeeeeeeeeeeeeeteeea
", + "bbox": [ + 256, + 179, + 527, + 877 + ], + "page_idx": 21 + }, + { + "type": "table", + "img_path": "images/dfb5a3b4eb73c9332db6beb500365a48a459fe578cfa003f8e5708e40128a51b.jpg", + "table_caption": [ + "" + ], + "table_footnote": [], + "table_body": "
971900933554ISL9816 48124111111
福848-R488845118101888636-
13-808-668-69L-498-486-29'88-8£6L-728-4155-58888-764156-
55446 8314448F4364学3999京08329
9'SL-0388756-6'1-6499L-28756-7I1-3678-896-8678-$'96-
8386'091590556699496007844481115446548814福683515554
3314541S2453 34251863
4568833853494
L094169 48248
518588408584548443244334
23225496970829449$'0998552455413331255551835454TOIIOI03134111441455514694445918041887551957138050675545
S66L339835369558635-8三61
14444431148B6953 国5058818
2533773828361949'101326.30059296078
993打99111
eeeereeneeeeeeeeet eereereeerreeeeegereeeeereraaergeeeedeereeeergereeeeeieeeereeneereergereeereeeeeeeeeeeeeceeeeeeeeeeeeeeeeeceeeereeeeeeeeeecreareeereseeeeeeeereeeetGeepeeeeeecegepeeeeetCeeereneerecegeeeeeeeeeerereeeeeneeeereeeeeeelereeeeeeeeeeneeeeneeeeeeeeeeeeeeereeeeeeeeeeereeeecreeeeererrereeneerectgereeetireeeeeeeseeetneeeeieeeeereretreresesaeeeeeeeeeeti
", + "bbox": [ + 550, + 176, + 813, + 876 + ], + "page_idx": 21 + }, + { + "type": "table", + "img_path": "images/b0b769b050b71f164316d38f75aa2fa2beb1010592b70fe5e038c3a45826a78f.jpg", + "table_caption": [ + "", + "" + ], + "table_footnote": [], + "table_body": "
1.55290.00668744444435339LL35.55759'8116.44695511.441114£99L1143423360.6778'8311.15
555418.3554318.0044.5511915678IL'LS40.64488189011117118951511.351L0116410338130
66.4441442555511.09533.35154.31435.5514445198'0L0153.59£9.09.044/.6156.559.55S1855438.5053816.7916.1460.00434
111152585-56.95-6'L[-4441-9L'9S-1135911.51555331460'8L7'01-£6.33-11.04410'61-1571-89.97-L6'67-353335331531-88'L7-40.05-
15:114353555.815531966986.69355555.3511.333335.21448.1135.3845.55335.713233219'8466.55370.3380.5535.7128.85
10028B28281055131641621328388084660
ererreresdererertrreresdeeiereeerereeeeeeertrrereeeeer erraeraseeereerteeeeeer crrrrreereeretrrereeeeerrereedreeeeerrereeeeerCereresdeeeeerrtrreresdeeererrreeereereeeeetrereedeercrrereeeerreerrereeeetrCgrrreesdrerecrraedticerseerrretrrereeeirereeeereecrreaeeeeereereeseirreeeeeeeeeeareeeeereeeiceerreeeeteseseeaEreereedeereeretrrereseeereedeeeeeeeseeeeterreresdeeeeerereeseseeeeeeresdreeseeereseeererarrgeaeedeereeeereeseeeereeeeserrreerrareseeercarereeeerreetrrreeeeerreeeeteeeeseeeeercaaeesdeteeerteseeeear
", + "bbox": [ + 343, + 181, + 704, + 672 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "In the MLP-Bind method, we bind the weights of MLP. By doing this, the weights of the agent from the similar structures will be shared. For example, in the centipede environment, the weights from observation to action of all the LeftHips are constrained to be same. ", + "bbox": [ + 174, + 782, + 825, + 825 + ], + "page_idx": 22 + }, + { + "type": "table", + "img_path": "images/6e10e55c33e7a4e84bf33169da0a7c4acbe289137d4dbd8961f161c9a4f46c38.jpg", + "table_caption": [ + "Table 10: Results of the zero-shot learning of Snake ", + "6.5.1 LINEAR NORMALIZATION ZERO-SHOT RESULTS AND COLORIZATION. " + ], + "table_footnote": [], + "table_body": "
Reward Min
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v1-31.719-43.328-64.958-58.811308.994301.4
SnakeFive-v1-56.38229.262-55.562-39.346
SnakeSix-v1-44.888-5.306-42.744-76.42266.066
SnakeSeven-v1-52.764-25.511-49.765-59.52231.557329.456333.237253.828
SnakeFive-v1-56.382-30.466-49.953-47.339
SnakeSix-v1SnakeSeven-v1-44.888-35.655-52.18-57.334-90.837-52.015
-52.76489.358
Reward Std
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v19.5445.79210.76910.1465.5135.788
SnakeFive-v113.3186.19111.7157.883
SnakeSix-v111.51513.53510.30416.8697.3957.1795.856.60414.084
SnakeSeven-v111.2112.13110.9816.953
SnakeFive-v113.31852.7111.5448.5198.51923.10913.529
SnakeSix-v1SnakeSeven-v111.51511.2164.54910.68310.52811.615
Reward Max
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v13.863-19.40910.1180.918338.374326.974
SnakeFive-v120.58563.8663.136-2.329
SnakeSix-v17.658.7793.51-2.139305.011
SnakeSeven-v1-2.2851.2651.325-28.087275.536
SnakeFive-v120.585139.6175.3051.487356.366
SnakeSix-v1SnakeSeven-v17.6-2.28156.771143.8242.3162.596-4.1598.394367.768336.561
Reward Average
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v1-13.531-30.185-12.363-21.824325.476314.919
SnakeFive-v1-17.82951.411-16.481-21.256
SnakeSix-v1-19.8932.2-19.674-42.461282.426
SnakeSeven-v1-22.09923.533-22.378-42.742256.293
SnakeFive-v1-17.82989.823-16.54-15.975342.881
SnakeSix-v1-19.89105.632-20.454-34.343351.85313.149
SnakeSeven-v1-22.099115.123-22.81-22.383
", + "bbox": [ + 199, + 122, + 795, + 707 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "As the scale of zero-shot results is very different, we normalize the results across different models for each transfer learning task. For each task, we record the worst value of results from different models and the pre-set worst value $V _ { \\mathrm { m i n } }$ . we set the normalization minimun value as this worst value. We calculate the normalization maximum value by $\\lfloor \\mathrm { m a x } ( V ) / \\mathrm { I n t L e n } \\rfloor$ $^ *$ IntLen. ", + "bbox": [ + 174, + 752, + 825, + 809 + ], + "page_idx": 23 + }, + { + "type": "table", + "img_path": "images/dd9260e1ab2daad1bfcea3b6f8621cfccf78d2760864779b625029b35c1995e4.jpg", + "table_caption": [ + "Table 11: Parameters for linear normalization during zero-shot results. Results are shown in the order of Centipedes’ reward, Centipedes’ running-length, Snakes’ reward " + ], + "table_footnote": [], + "table_body": "
parametersVminIntLen
value-100,-100, -2030,30,30
", + "bbox": [ + 349, + 863, + 648, + 906 + ], + "page_idx": 23 + }, + { + "type": "table", + "img_path": "images/149be23ac6501df5ffee2b243d8730e45c0e416d58559db5e15b40927b122fac.jpg", + "table_caption": [ + "6.6 DETAILS OF ROBUSTNESS RESULTS ", + "Table 12: Detail results of Robustness testing. " + ], + "table_footnote": [], + "table_body": "
ModelMass
Hopper Halfhumanoid HorseWolfOstrich
NerveNet1981.592047.921968.872062.361852.762519.62392.052282.842161.611804.522304.251900.431862.211955.16922.12932.48983.11910.29791.84
1703.471648.281576.211354.51
1239.72946.08
MLP2268.641932.981637.471591.511073.1421.82115.75278.921224.98353.561692.081083.621932.641754.94878.75945.94
690.22538.14443.921459.931039.9814.55879.66805.99802.41
ModelStrength
HopperHalfhumanoidHorseWolfOstrich
NerveNet1961.621117.91492.33468.4446.141528.89940.96594.53308.08249.62316.82960.341827.73862.19274.77189.49133.84794.56539.45371.19255.12197.84
488.51305.16
305.16201.42
MLP800.15431.97429.76426.27412.391439.891040.341236.881730.881108.32230.85483.36
7209.54
765.7306.08519.73163.36
117.0565.29228.26187.7288.28108.97232.8133.99
", + "bbox": [ + 253, + 135, + 743, + 518 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "6.7 HYPERPARAMETERS FOR MULTI-TASK LEARNING ", + "text_level": 1, + "bbox": [ + 173, + 571, + 562, + 588 + ], + "page_idx": 24 + }, + { + "type": "table", + "img_path": "images/9d8959eb0e6443be88c8602880ed2d737116266dfa7e25399e10eb71bc123454.jpg", + "table_caption": [ + "Table 13: Hyperparameter settings for multi-task learning. " + ], + "table_footnote": [], + "table_body": "
MLP, TreeNetand NerveNetValue Tried
Network Shape[64, 64]
Number of Iteration Per Update10
Use KL PenaltyNo
Learning Rate SchedulerAdaptive
NerveNetValue Tried
Number of Propogation Steps4
NerveNetVariants Size of Nodes'Hidden StateNerveNet-1
64
Merge joint and body NodeYes
Output Network Disable Edge Type Separate
Add Skip-connection from /to rootYes No
", + "bbox": [ + 316, + 628, + 681, + 838 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "6.8 SCHEMATIC FIGURES OF WA L K E R S AGENTS ", + "text_level": 1, + "bbox": [ + 173, + 864, + 521, + 878 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "We design Walker task-set, which contains five 2d walkers in the MuJoCo engine. ", + "bbox": [ + 173, + 890, + 722, + 905 + ], + "page_idx": 24 + }, + { + "type": "image", + "img_path": "images/e4db0554bf395b640e337ab9dd90cfa8eba43566ee9855ed7bab06708c5c8236.jpg", + "image_caption": [ + "Figure 17: Schematic diagrams and auto-parsed graph structures of Walker-Ostrich and Walker-Wolf. " + ], + "image_footnote": [], + "bbox": [ + 184, + 142, + 818, + 284 + ], + "page_idx": 25 + }, + { + "type": "image", + "img_path": "images/24d53bd50058ae9dce42cfb0dbe3b1f6127ebefc151bad848cb175c05a53e2d4.jpg", + "image_caption": [ + "Figure 18: Schematic diagrams and auto-parsed graph structures of Walker-Hopper and Walker-HalfHumanoid. " + ], + "image_footnote": [], + "bbox": [ + 251, + 404, + 750, + 587 + ], + "page_idx": 25 + }, + { + "type": "image", + "img_path": "images/9de69acaa7f53ccf56bad5e7e49119e3e4eb8f81606d03c69dfcfd1a0fa3d1be.jpg", + "image_caption": [ + "Figure 19: Schematic diagrams and auto-parsed graph structures of Walker-Horse. " + ], + "image_footnote": [], + "bbox": [ + 281, + 710, + 714, + 854 + ], + "page_idx": 25 + } +] \ No newline at end of file diff --git a/parse/train/S1sqHMZCb/S1sqHMZCb_middle.json b/parse/train/S1sqHMZCb/S1sqHMZCb_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..b884413c5bf217834cb2c03a5fb250a5c8d78143 --- /dev/null +++ b/parse/train/S1sqHMZCb/S1sqHMZCb_middle.json @@ -0,0 +1,50178 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 97 + ], + "score": 1.0, + "content": "NERVENET: LEARNING STRUCTURED POLICY WITH", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 97, + 320, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 320, + 118 + ], + "score": 1.0, + "content": "GRAPH NEURAL NETWORKS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 135, + 349, + 147 + ], + "lines": [ + { + "bbox": [ + 111, + 134, + 351, + 150 + ], + "spans": [ + { + "bbox": [ + 111, + 134, + 351, + 150 + ], + "score": 1.0, + "content": "Tingwu Wang∗, Renjie Liao∗, Jimmy Ba & Sanja Fidler", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 113, + 149, + 378, + 201 + ], + "lines": [ + { + "bbox": [ + 112, + 146, + 248, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 248, + 158 + ], + "score": 1.0, + "content": "Department of Computer Science", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 156, + 202, + 169 + ], + "spans": [ + { + "bbox": [ + 111, + 156, + 202, + 169 + ], + "score": 1.0, + "content": "University of Toronto", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 168, + 176, + 180 + ], + "spans": [ + { + "bbox": [ + 112, + 168, + 176, + 180 + ], + "score": 1.0, + "content": "Vector Institute", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 112, + 179, + 320, + 191 + ], + "spans": [ + { + "bbox": [ + 112, + 179, + 320, + 191 + ], + "score": 1.0, + "content": "{tingwuwang,rjliao}@cs.toronto.edu,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 113, + 191, + 378, + 201 + ], + "spans": [ + { + "bbox": [ + 113, + 191, + 378, + 201 + ], + "score": 1.0, + "content": "jimmy@psi.toronto.edu, fidler@cs.toronto.edu", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 278, + 230, + 333, + 242 + ], + "lines": [ + { + "bbox": [ + 276, + 229, + 336, + 244 + ], + "spans": [ + { + "bbox": [ + 276, + 229, + 336, + 244 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 143, + 255, + 468, + 409 + ], + "lines": [ + { + "bbox": [ + 142, + 255, + 469, + 267 + ], + "spans": [ + { + "bbox": [ + 142, + 255, + 469, + 267 + ], + "score": 1.0, + "content": "We address the problem of learning structured policies for continuous control. In", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "score": 1.0, + "content": "traditional reinforcement learning, policies of agents are learned by multi-layer", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 277, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 277, + 470, + 290 + ], + "score": 1.0, + "content": "perceptrons (MLPs) which take the concatenation of all observations from the en-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 289, + 469, + 300 + ], + "spans": [ + { + "bbox": [ + 142, + 289, + 469, + 300 + ], + "score": 1.0, + "content": "vironment as input for predicting actions. In this work, we propose NerveNet to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 299, + 470, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 470, + 311 + ], + "score": 1.0, + "content": "explicitly model the structure of an agent, which naturally takes the form of a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "score": 1.0, + "content": "graph. Specifically, serving as the agent’s policy network, NerveNet first propa-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 322, + 470, + 333 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 470, + 333 + ], + "score": 1.0, + "content": "gates information over the structure of the agent and then predict actions for differ-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 333, + 469, + 344 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 469, + 344 + ], + "score": 1.0, + "content": "ent parts of the agent. In the experiments, we first show that our NerveNet is com-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 343, + 470, + 354 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 470, + 354 + ], + "score": 1.0, + "content": "parable to state-of-the-art methods on standard MuJoCo environments. We further", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 140, + 354, + 470, + 367 + ], + "spans": [ + { + "bbox": [ + 140, + 354, + 470, + 367 + ], + "score": 1.0, + "content": "propose our customized reinforcement learning environments for benchmarking", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 365, + 470, + 378 + ], + "spans": [ + { + "bbox": [ + 141, + 365, + 470, + 378 + ], + "score": 1.0, + "content": "two types of structure transfer learning tasks, i.e., size and disability transfer, as", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 375, + 470, + 389 + ], + "spans": [ + { + "bbox": [ + 141, + 375, + 470, + 389 + ], + "score": 1.0, + "content": "well as multi-task learning. We demonstrate that policies learned by NerveNet are", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 387, + 470, + 399 + ], + "spans": [ + { + "bbox": [ + 141, + 387, + 470, + 399 + ], + "score": 1.0, + "content": "significantly more transferable and generalizable than policies learned by other", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 397, + 376, + 411 + ], + "spans": [ + { + "bbox": [ + 141, + 397, + 376, + 411 + ], + "score": 1.0, + "content": "models and are able to transfer even in a zero-shot setting.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 108, + 430, + 205, + 442 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 208, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 208, + 446 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 455, + 505, + 499 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "Deep reinforcement learning (RL) has received increasing attention over the past few years, with the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 466, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 478 + ], + "score": 1.0, + "content": "recent success of applications such as playing Atari Games, Mnih et al. (2015), and Go, (Silver et al.,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 476, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 506, + 490 + ], + "score": 1.0, + "content": "2016; 2017). Significant advances have also been made in robotics using the latest RL techniques,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 488, + 278, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 278, + 501 + ], + "score": 1.0, + "content": "e.g., Levine et al. (2016); Gu et al. (2017).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 505, + 504, + 537 + ], + "lines": [ + { + "bbox": [ + 105, + 504, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 506, + 518 + ], + "score": 1.0, + "content": "Many RL problems feature agents with multiple dependent controllers. For example, humanoid", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "robots consist of multiple physically linked joints. Action to be taken by each joint or the body", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 527, + 456, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 456, + 539 + ], + "score": 1.0, + "content": "should thus not only depend on its own observations but also on actions of other joints.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 543, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "score": 1.0, + "content": "Previous approaches in RL typically use MLP to learn the agent’s policy. In particular, MLP takes", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "the concatenation of observations from the environment as input, which may be measurements like", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "positions, velocities of body and joints in the current time instance. The MLP policy then predicts", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "actions to be taken by every joint and body. Thus the task of the MLP policy is to discover the latent", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 586, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 601 + ], + "score": 1.0, + "content": "relationships between observations. This typically leads to longer training times, requiring more", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "exposure of the agent to the environment. In our work, we aim to exploit the body structure of an", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 609, + 380, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 380, + 623 + ], + "score": 1.0, + "content": "agent, and physical dependencies that naturally exist in such agents.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 626, + 505, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 625, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 506, + 638 + ], + "score": 1.0, + "content": "We rely on the fact that bodies of most robots and animals have a discrete graph structure. Nodes of", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 636, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 650 + ], + "score": 1.0, + "content": "the graph may represent the joints, and edges represent the (physical) dependencies between them.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 648, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "In particular, we define the agent’s policy using a Graph Neural Network, Scarselli et al. (2009),", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "which is a neural network that operates over graph structures. We refer to our model as NerveNet", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "due to the resemblance of the neural nervous system to a graph. NerveNet propagates information", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 680, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 505, + 694 + ], + "score": 1.0, + "content": "between different parts of the body based on the underlying graph structure before outputting the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 692, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 505, + 704 + ], + "score": 1.0, + "content": "action for each part. By doing so, NerveNet can leverage the structure information encoded by the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 703, + 505, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 703, + 505, + 716 + ], + "score": 1.0, + "content": "agent’s body which is advantageous in learning the correct inductive bias, and thus is less prone to", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 41.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 120, + 722, + 237, + 731 + ], + "lines": [ + { + "bbox": [ + 119, + 719, + 238, + 734 + ], + "spans": [ + { + "bbox": [ + 119, + 719, + 238, + 734 + ], + "score": 1.0, + "content": "∗Two authors contribute equally.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 97 + ], + "score": 1.0, + "content": "NERVENET: LEARNING STRUCTURED POLICY WITH", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 97, + 320, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 320, + 118 + ], + "score": 1.0, + "content": "GRAPH NEURAL NETWORKS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 135, + 349, + 147 + ], + "lines": [ + { + "bbox": [ + 111, + 134, + 351, + 150 + ], + "spans": [ + { + "bbox": [ + 111, + 134, + 351, + 150 + ], + "score": 1.0, + "content": "Tingwu Wang∗, Renjie Liao∗, Jimmy Ba & Sanja Fidler", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 111, + 134, + 351, + 150 + ] + }, + { + "type": "list", + "bbox": [ + 113, + 149, + 378, + 201 + ], + "lines": [ + { + "bbox": [ + 112, + 146, + 248, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 248, + 158 + ], + "score": 1.0, + "content": "Department of Computer Science", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 156, + 202, + 169 + ], + "spans": [ + { + "bbox": [ + 111, + 156, + 202, + 169 + ], + "score": 1.0, + "content": "University of Toronto", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 168, + 176, + 180 + ], + "spans": [ + { + "bbox": [ + 112, + 168, + 176, + 180 + ], + "score": 1.0, + "content": "Vector Institute", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 179, + 320, + 191 + ], + "spans": [ + { + "bbox": [ + 112, + 179, + 320, + 191 + ], + "score": 1.0, + "content": "{tingwuwang,rjliao}@cs.toronto.edu,", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 191, + 378, + 201 + ], + "spans": [ + { + "bbox": [ + 113, + 191, + 378, + 201 + ], + "score": 1.0, + "content": "jimmy@psi.toronto.edu, fidler@cs.toronto.edu", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + } + ], + "index": 5, + "bbox_fs": [ + 111, + 146, + 378, + 201 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 230, + 333, + 242 + ], + "lines": [ + { + "bbox": [ + 276, + 229, + 336, + 244 + ], + "spans": [ + { + "bbox": [ + 276, + 229, + 336, + 244 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 143, + 255, + 468, + 409 + ], + "lines": [ + { + "bbox": [ + 142, + 255, + 469, + 267 + ], + "spans": [ + { + "bbox": [ + 142, + 255, + 469, + 267 + ], + "score": 1.0, + "content": "We address the problem of learning structured policies for continuous control. In", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "score": 1.0, + "content": "traditional reinforcement learning, policies of agents are learned by multi-layer", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 277, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 277, + 470, + 290 + ], + "score": 1.0, + "content": "perceptrons (MLPs) which take the concatenation of all observations from the en-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 289, + 469, + 300 + ], + "spans": [ + { + "bbox": [ + 142, + 289, + 469, + 300 + ], + "score": 1.0, + "content": "vironment as input for predicting actions. In this work, we propose NerveNet to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 299, + 470, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 470, + 311 + ], + "score": 1.0, + "content": "explicitly model the structure of an agent, which naturally takes the form of a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "score": 1.0, + "content": "graph. Specifically, serving as the agent’s policy network, NerveNet first propa-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 322, + 470, + 333 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 470, + 333 + ], + "score": 1.0, + "content": "gates information over the structure of the agent and then predict actions for differ-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 333, + 469, + 344 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 469, + 344 + ], + "score": 1.0, + "content": "ent parts of the agent. In the experiments, we first show that our NerveNet is com-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 343, + 470, + 354 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 470, + 354 + ], + "score": 1.0, + "content": "parable to state-of-the-art methods on standard MuJoCo environments. We further", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 140, + 354, + 470, + 367 + ], + "spans": [ + { + "bbox": [ + 140, + 354, + 470, + 367 + ], + "score": 1.0, + "content": "propose our customized reinforcement learning environments for benchmarking", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 365, + 470, + 378 + ], + "spans": [ + { + "bbox": [ + 141, + 365, + 470, + 378 + ], + "score": 1.0, + "content": "two types of structure transfer learning tasks, i.e., size and disability transfer, as", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 375, + 470, + 389 + ], + "spans": [ + { + "bbox": [ + 141, + 375, + 470, + 389 + ], + "score": 1.0, + "content": "well as multi-task learning. We demonstrate that policies learned by NerveNet are", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 387, + 470, + 399 + ], + "spans": [ + { + "bbox": [ + 141, + 387, + 470, + 399 + ], + "score": 1.0, + "content": "significantly more transferable and generalizable than policies learned by other", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 397, + 376, + 411 + ], + "spans": [ + { + "bbox": [ + 141, + 397, + 376, + 411 + ], + "score": 1.0, + "content": "models and are able to transfer even in a zero-shot setting.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 15.5, + "bbox_fs": [ + 140, + 255, + 470, + 411 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 430, + 205, + 442 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 208, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 208, + 446 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 455, + 505, + 499 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "Deep reinforcement learning (RL) has received increasing attention over the past few years, with the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 466, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 478 + ], + "score": 1.0, + "content": "recent success of applications such as playing Atari Games, Mnih et al. (2015), and Go, (Silver et al.,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 476, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 506, + 490 + ], + "score": 1.0, + "content": "2016; 2017). Significant advances have also been made in robotics using the latest RL techniques,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 488, + 278, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 278, + 501 + ], + "score": 1.0, + "content": "e.g., Levine et al. (2016); Gu et al. (2017).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 455, + 506, + 501 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 505, + 504, + 537 + ], + "lines": [ + { + "bbox": [ + 105, + 504, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 506, + 518 + ], + "score": 1.0, + "content": "Many RL problems feature agents with multiple dependent controllers. For example, humanoid", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "robots consist of multiple physically linked joints. Action to be taken by each joint or the body", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 527, + 456, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 456, + 539 + ], + "score": 1.0, + "content": "should thus not only depend on its own observations but also on actions of other joints.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 504, + 506, + 539 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 543, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "score": 1.0, + "content": "Previous approaches in RL typically use MLP to learn the agent’s policy. In particular, MLP takes", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "the concatenation of observations from the environment as input, which may be measurements like", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "positions, velocities of body and joints in the current time instance. The MLP policy then predicts", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "actions to be taken by every joint and body. Thus the task of the MLP policy is to discover the latent", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 586, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 601 + ], + "score": 1.0, + "content": "relationships between observations. This typically leads to longer training times, requiring more", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "exposure of the agent to the environment. In our work, we aim to exploit the body structure of an", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 609, + 380, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 380, + 623 + ], + "score": 1.0, + "content": "agent, and physical dependencies that naturally exist in such agents.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 543, + 506, + 623 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 626, + 505, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 625, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 506, + 638 + ], + "score": 1.0, + "content": "We rely on the fact that bodies of most robots and animals have a discrete graph structure. Nodes of", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 636, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 650 + ], + "score": 1.0, + "content": "the graph may represent the joints, and edges represent the (physical) dependencies between them.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 648, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "In particular, we define the agent’s policy using a Graph Neural Network, Scarselli et al. (2009),", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "which is a neural network that operates over graph structures. We refer to our model as NerveNet", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "due to the resemblance of the neural nervous system to a graph. NerveNet propagates information", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 680, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 505, + 694 + ], + "score": 1.0, + "content": "between different parts of the body based on the underlying graph structure before outputting the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 692, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 505, + 704 + ], + "score": 1.0, + "content": "action for each part. By doing so, NerveNet can leverage the structure information encoded by the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 703, + 505, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 703, + 505, + 716 + ], + "score": 1.0, + "content": "agent’s body which is advantageous in learning the correct inductive bias, and thus is less prone to", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 320, + 504, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 504, + 333 + ], + "score": 1.0, + "content": "overfitting. Moreover, NerveNet is naturally suitable for structure transferring tasks as most of the", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 331, + 374, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 374, + 344 + ], + "score": 1.0, + "content": "model weights are shared across the nodes and edges, respectively.", + "type": "text", + "cross_page": true + } + ], + "index": 9 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 625, + 506, + 716 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 90, + 491, + 221 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 90, + 491, + 221 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 90, + 491, + 221 + ], + "spans": [ + { + "bbox": [ + 115, + 90, + 491, + 221 + ], + "score": 0.973, + "type": "image", + "image_path": "ec7ba80f49404fcff9389c3659fcdd445c72eb87234ed2686b8a6262e8e85922.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 90, + 491, + 133.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 133.66666666666666, + 491, + 177.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 177.33333333333331, + 491, + 220.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 231, + 505, + 286 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 230, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 244 + ], + "score": 1.0, + "content": "Figure 1: Visualization of the graph structure of CentipedeEight in our environment. We use", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "this agent for testing the ability of transfer learning of our model. Since for this agent, each body", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "node is paired with at least one joint node, we omit the body nodes and fill up the position", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 263, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 506, + 277 + ], + "score": 1.0, + "content": "with the corresponding joint nodes. By omitting the body nodes, a more compact graph is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 273, + 407, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 407, + 288 + ], + "score": 1.0, + "content": "constructed, the details of which are illustrated in the experimental section.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 106, + 320, + 504, + 343 + ], + "lines": [ + { + "bbox": [ + 106, + 320, + 504, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 504, + 333 + ], + "score": 1.0, + "content": "overfitting. Moreover, NerveNet is naturally suitable for structure transferring tasks as most of the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 331, + 374, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 374, + 344 + ], + "score": 1.0, + "content": "model weights are shared across the nodes and edges, respectively.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 348, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 348, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 505, + 361 + ], + "score": 1.0, + "content": "We first evaluate our NerveNet on standard RL benchmarks such as the OpenAI Gym, Brockman", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 359, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 506, + 371 + ], + "score": 1.0, + "content": "et al. (2016) which stem from MuJoCo. We show that our model achieves comparable results to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "state-of-the-art MLP based methods. To verify our claim regarding the structure transfer, we further", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 380, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 395 + ], + "score": 1.0, + "content": "introduce our customized RL environments which are based on the ones of Gym. Two types of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "structure transfer tasks are designed, size transfer and disability transfer. In particular, size transfer", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "focuses on the scenario in which policies are learned for small-sized agents (simpler body structure)", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 414, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 428 + ], + "score": 1.0, + "content": "and applied directly to large-sized agents which are composed by some repetitive components shared", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "with the small-sized agent. Secondly, disability transfer investigates scenarios in which policies", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "score": 1.0, + "content": "are learned for one agent and applied to the same agent with some components disabled. Our", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "experiments demonstrate that for structure transfer tasks our NerveNet is significantly better than", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 458, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 505, + 470 + ], + "score": 1.0, + "content": "all other competitors, and can even achieve zero-shot learning for some agents. For the multi-task", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 469, + 491, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 491, + 482 + ], + "score": 1.0, + "content": "learning tasks, NerveNet is also able to learn policies that are more robust with better efficiency.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 485, + 505, + 552 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "The main contribution of this paper is the following: We explore the problem of learning transfer-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "able and generalized features by incorporating a prior on the structure via graph neural networks.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 506, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 520 + ], + "score": 1.0, + "content": "NerveNet permits powerful transfer learning from one structure to another, which goes well beyond", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "score": 1.0, + "content": "the ability of previous models. NerveNet is also more robust and has more potential in perform-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "score": 1.0, + "content": "ing multi-task learning. The demo and code for this project are released, under the project page of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 540, + 416, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 416, + 553 + ], + "score": 1.0, + "content": "http://www.cs.toronto.edu/˜tingwuwang/nervenet.html.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 107, + 582, + 184, + 595 + ], + "lines": [ + { + "bbox": [ + 104, + 580, + 186, + 597 + ], + "spans": [ + { + "bbox": [ + 104, + 580, + 186, + 597 + ], + "score": 1.0, + "content": "2 NERVENET", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 108, + 615, + 504, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "In this section, we first introduce the notation. We then explain how to construct the graph for each of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "our agents, followed by the description of the NerveNet. Finally, we describe the learning algorithm", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 638, + 166, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 166, + 649 + ], + "score": 1.0, + "content": "for our model.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "We formulate the locomotion control problems as an infinite-horizon discounted Markov decision", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "process (MDP). To fully describe the MDP for continuous control problems which include locomo-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 356, + 690 + ], + "score": 1.0, + "content": "tion control, we define the state space or observation space as", + "type": "text" + }, + { + "bbox": [ + 356, + 677, + 364, + 687 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 677, + 445, + 690 + ], + "score": 1.0, + "content": "and action space as", + "type": "text" + }, + { + "bbox": [ + 446, + 677, + 455, + 687 + ], + "score": 0.75, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 677, + 505, + 690 + ], + "score": 1.0, + "content": ". To interact", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 359, + 700 + ], + "score": 1.0, + "content": "with the environments, the agent generates its stochastic policy", + "type": "text" + }, + { + "bbox": [ + 359, + 687, + 400, + 700 + ], + "score": 0.93, + "content": "\\pi _ { \\boldsymbol { \\theta } } ( \\boldsymbol { a } ^ { \\tau } | \\boldsymbol { s } ^ { \\tau } )", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "based on the current state", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 107, + 699, + 138, + 709 + ], + "score": 0.9, + "content": "s ^ { \\tau } \\in \\boldsymbol { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 698, + 169, + 711 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 169, + 699, + 202, + 709 + ], + "score": 0.91, + "content": "a ^ { \\tau } \\in \\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 698, + 272, + 711 + ], + "score": 1.0, + "content": "is the action and", + "type": "text" + }, + { + "bbox": [ + 273, + 699, + 279, + 709 + ], + "score": 0.81, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "are the parameters of the policy function. The environ-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 284, + 722 + ], + "score": 1.0, + "content": "ment on the other hand, produces a reward", + "type": "text" + }, + { + "bbox": [ + 284, + 710, + 322, + 722 + ], + "score": 0.93, + "content": "r ( s ^ { \\tau } , a ^ { \\tau } )", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "for the agent, and the agent’s objective is to", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 721, + 307, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 307, + 733 + ], + "score": 1.0, + "content": "find a policy that maximizes the expected reward.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 90, + 491, + 221 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 90, + 491, + 221 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 90, + 491, + 221 + ], + "spans": [ + { + "bbox": [ + 115, + 90, + 491, + 221 + ], + "score": 0.973, + "type": "image", + "image_path": "ec7ba80f49404fcff9389c3659fcdd445c72eb87234ed2686b8a6262e8e85922.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 90, + 491, + 133.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 133.66666666666666, + 491, + 177.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 177.33333333333331, + 491, + 220.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 231, + 505, + 286 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 230, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 244 + ], + "score": 1.0, + "content": "Figure 1: Visualization of the graph structure of CentipedeEight in our environment. We use", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "this agent for testing the ability of transfer learning of our model. Since for this agent, each body", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "node is paired with at least one joint node, we omit the body nodes and fill up the position", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 263, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 506, + 277 + ], + "score": 1.0, + "content": "with the corresponding joint nodes. By omitting the body nodes, a more compact graph is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 273, + 407, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 407, + 288 + ], + "score": 1.0, + "content": "constructed, the details of which are illustrated in the experimental section.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 106, + 320, + 504, + 343 + ], + "lines": [], + "index": 8.5, + "bbox_fs": [ + 105, + 320, + 504, + 344 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 348, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 348, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 505, + 361 + ], + "score": 1.0, + "content": "We first evaluate our NerveNet on standard RL benchmarks such as the OpenAI Gym, Brockman", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 359, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 506, + 371 + ], + "score": 1.0, + "content": "et al. (2016) which stem from MuJoCo. We show that our model achieves comparable results to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "state-of-the-art MLP based methods. To verify our claim regarding the structure transfer, we further", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 380, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 395 + ], + "score": 1.0, + "content": "introduce our customized RL environments which are based on the ones of Gym. Two types of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "structure transfer tasks are designed, size transfer and disability transfer. In particular, size transfer", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "focuses on the scenario in which policies are learned for small-sized agents (simpler body structure)", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 414, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 428 + ], + "score": 1.0, + "content": "and applied directly to large-sized agents which are composed by some repetitive components shared", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "with the small-sized agent. Secondly, disability transfer investigates scenarios in which policies", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "score": 1.0, + "content": "are learned for one agent and applied to the same agent with some components disabled. Our", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "experiments demonstrate that for structure transfer tasks our NerveNet is significantly better than", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 458, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 505, + 470 + ], + "score": 1.0, + "content": "all other competitors, and can even achieve zero-shot learning for some agents. For the multi-task", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 469, + 491, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 491, + 482 + ], + "score": 1.0, + "content": "learning tasks, NerveNet is also able to learn policies that are more robust with better efficiency.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 348, + 506, + 482 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 485, + 505, + 552 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "The main contribution of this paper is the following: We explore the problem of learning transfer-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "able and generalized features by incorporating a prior on the structure via graph neural networks.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 506, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 520 + ], + "score": 1.0, + "content": "NerveNet permits powerful transfer learning from one structure to another, which goes well beyond", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "score": 1.0, + "content": "the ability of previous models. NerveNet is also more robust and has more potential in perform-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "score": 1.0, + "content": "ing multi-task learning. The demo and code for this project are released, under the project page of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 540, + 416, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 416, + 553 + ], + "score": 1.0, + "content": "http://www.cs.toronto.edu/˜tingwuwang/nervenet.html.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 485, + 506, + 553 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 582, + 184, + 595 + ], + "lines": [ + { + "bbox": [ + 104, + 580, + 186, + 597 + ], + "spans": [ + { + "bbox": [ + 104, + 580, + 186, + 597 + ], + "score": 1.0, + "content": "2 NERVENET", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 108, + 615, + 504, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "In this section, we first introduce the notation. We then explain how to construct the graph for each of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "our agents, followed by the description of the NerveNet. Finally, we describe the learning algorithm", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 638, + 166, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 166, + 649 + ], + "score": 1.0, + "content": "for our model.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 106, + 616, + 506, + 649 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "We formulate the locomotion control problems as an infinite-horizon discounted Markov decision", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "process (MDP). To fully describe the MDP for continuous control problems which include locomo-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 356, + 690 + ], + "score": 1.0, + "content": "tion control, we define the state space or observation space as", + "type": "text" + }, + { + "bbox": [ + 356, + 677, + 364, + 687 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 677, + 445, + 690 + ], + "score": 1.0, + "content": "and action space as", + "type": "text" + }, + { + "bbox": [ + 446, + 677, + 455, + 687 + ], + "score": 0.75, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 677, + 505, + 690 + ], + "score": 1.0, + "content": ". To interact", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 359, + 700 + ], + "score": 1.0, + "content": "with the environments, the agent generates its stochastic policy", + "type": "text" + }, + { + "bbox": [ + 359, + 687, + 400, + 700 + ], + "score": 0.93, + "content": "\\pi _ { \\boldsymbol { \\theta } } ( \\boldsymbol { a } ^ { \\tau } | \\boldsymbol { s } ^ { \\tau } )", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "based on the current state", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 107, + 699, + 138, + 709 + ], + "score": 0.9, + "content": "s ^ { \\tau } \\in \\boldsymbol { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 698, + 169, + 711 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 169, + 699, + 202, + 709 + ], + "score": 0.91, + "content": "a ^ { \\tau } \\in \\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 698, + 272, + 711 + ], + "score": 1.0, + "content": "is the action and", + "type": "text" + }, + { + "bbox": [ + 273, + 699, + 279, + 709 + ], + "score": 0.81, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "are the parameters of the policy function. The environ-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 284, + 722 + ], + "score": 1.0, + "content": "ment on the other hand, produces a reward", + "type": "text" + }, + { + "bbox": [ + 284, + 710, + 322, + 722 + ], + "score": 0.93, + "content": "r ( s ^ { \\tau } , a ^ { \\tau } )", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "for the agent, and the agent’s objective is to", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 721, + 307, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 307, + 733 + ], + "score": 1.0, + "content": "find a policy that maximizes the expected reward.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 655, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 87, + 502, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 87, + 502, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 87, + 502, + 225 + ], + "spans": [ + { + "bbox": [ + 110, + 87, + 502, + 225 + ], + "score": 0.971, + "type": "image", + "image_path": "4387bb37b2b6edf82cb3edff6c91e9f09012e99f1638fa5f05208b0a537fe83a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 87, + 502, + 133.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 133.0, + 502, + 179.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 179.0, + 502, + 225.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 240, + 505, + 295 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 239, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 505, + 253 + ], + "score": 1.0, + "content": "Figure 2: In this figure, we use Walker-Ostrich as an example of NerveNet. In the input model,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "for each node, NerveNet fetches the corresponding elements from the observation vector. NerveNet", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 261, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 275 + ], + "score": 1.0, + "content": "then computes the messages between neighbors in the graph, and updates the hidden state of each", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 273, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 505, + 286 + ], + "score": 1.0, + "content": "node. This process is repeated for a certain number of propagation steps. In the output model, the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 284, + 365, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 365, + 297 + ], + "score": 1.0, + "content": "policy is produced by collecting the output from each controller.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "title", + "bbox": [ + 107, + 318, + 234, + 329 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 235, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 235, + 331 + ], + "score": 1.0, + "content": "2.1 GRAPH CONSTRUCTION", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 339, + 505, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "In real life, skeletons of most robots and animals have a discrete graph structure, and are most often", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 348, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 506, + 364 + ], + "score": 1.0, + "content": "trees. Simulators such as the MuJoCo engine by Todorov et al. (2012), organize the agents using an", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 360, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 505, + 374 + ], + "score": 1.0, + "content": "XML-based kinematic tree. In our experiments, we will use the tree graphs as per MuJoCo. Note", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 371, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 506, + 385 + ], + "score": 1.0, + "content": "that our model can be directly applied to arbitrary graphs. In particular, we assume two types of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 383, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 506, + 395 + ], + "score": 1.0, + "content": "nodes in our tree: body and joint. The body nodes are abstract nodes used to construct the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 393, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 506, + 406 + ], + "score": 1.0, + "content": "kinematic tree via nesting, which is similar to the coordinate frame system used in robotics, Spong", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 403, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 104, + 403, + 505, + 418 + ], + "score": 1.0, + "content": "et al. (2006). The joint node represents the degrees of freedom of motion between the two body", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 416, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 506, + 428 + ], + "score": 1.0, + "content": "nodes. Take a simple humanoid as an example; the body nodes Thigh and Shin are connected via", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "score": 1.0, + "content": "the Knee, where Knee is a hinge joint. We further add a root node which observes additional", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 438, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 450 + ], + "score": 1.0, + "content": "information about the agent. For example, in the Reacher agent in MuJoCo, the root node has", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 449, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 505, + 461 + ], + "score": 1.0, + "content": "access to the target position of the agent. We build edges to form a tree graph. Fig. 1 illustrates the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 505, + 472 + ], + "score": 1.0, + "content": "graph structure of an example agent, CentipedeEight. We sketch the agent and its correspond-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 470, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 104, + 470, + 506, + 483 + ], + "score": 1.0, + "content": "ing graph in the left and right part of the figure, respectively. Note that for better visualization, we", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 481, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 505, + 493 + ], + "score": 1.0, + "content": "omit the joint nodes and use edges to represent the physical connections of joint nodes. Dif-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "ferent elements of the agent are parsed into nodes with different colors. Further details are provided", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 503, + 219, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 219, + 516 + ], + "score": 1.0, + "content": "in the experimental section.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 109, + 529, + 275, + 540 + ], + "lines": [ + { + "bbox": [ + 106, + 529, + 277, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 277, + 542 + ], + "score": 1.0, + "content": "2.2 NERVENET AS POLICY NETWORK", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 550, + 505, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 562 + ], + "score": 1.0, + "content": "We now turn to NerveNet which parametrizes the policy with a Graph Neural Network. Before", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "delving into details, we first introduce our notation. We then specify the input model which helps", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "to initialize the hidden state of each node. We further introduce the propagation model that updates", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 583, + 342, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 342, + 595 + ], + "score": 1.0, + "content": "these hidden states. Finally, we describe the output model.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 600, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 600, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 289, + 612 + ], + "score": 1.0, + "content": "We denote the graph structure of the agent as", + "type": "text" + }, + { + "bbox": [ + 289, + 600, + 338, + 612 + ], + "score": 0.93, + "content": "G = ( V , E )", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 600, + 366, + 612 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 366, + 600, + 376, + 610 + ], + "score": 0.76, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 600, + 394, + 612 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 394, + 601, + 403, + 610 + ], + "score": 0.82, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 600, + 505, + 612 + ], + "score": 1.0, + "content": "are the sets of nodes and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "score": 1.0, + "content": "edges, respectively. We focus on the directed graphs as the undirected case can be easily addressed", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "by splitting one undirected edge into two directed edges. We denote the out-going neighborhood", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 139, + 646 + ], + "score": 1.0, + "content": "of node", + "type": "text" + }, + { + "bbox": [ + 140, + 635, + 147, + 643 + ], + "score": 0.7, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 632, + 159, + 646 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 159, + 633, + 194, + 645 + ], + "score": 0.93, + "content": "\\mathcal { N } _ { o u t } ( u )", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 632, + 313, + 646 + ], + "score": 1.0, + "content": "which contains all endpoints", + "type": "text" + }, + { + "bbox": [ + 313, + 635, + 320, + 643 + ], + "score": 0.75, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 632, + 342, + 646 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 342, + 633, + 365, + 644 + ], + "score": 0.9, + "content": "( u , v )", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 632, + 505, + 646 + ], + "score": 1.0, + "content": "being an edge in the graph. Simi-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 321, + 657 + ], + "score": 1.0, + "content": "larly, we denote the in-coming neighborhood of node", + "type": "text" + }, + { + "bbox": [ + 322, + 646, + 329, + 654 + ], + "score": 0.75, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 643, + 341, + 657 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 341, + 644, + 372, + 656 + ], + "score": 0.91, + "content": "\\mathcal { N } _ { i n } ( u )", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 643, + 425, + 657 + ], + "score": 1.0, + "content": ". Every node", + "type": "text" + }, + { + "bbox": [ + 425, + 646, + 432, + 654 + ], + "score": 0.72, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "has an associated", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 148, + 668 + ], + "score": 1.0, + "content": "node type", + "type": "text" + }, + { + "bbox": [ + 148, + 655, + 226, + 667 + ], + "score": 0.92, + "content": "p _ { u } \\in \\{ 1 , 2 , \\ldots , P \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 654, + 505, + 668 + ], + "score": 1.0, + "content": ", which in our case corresponds to body, joint and root. We also", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 665, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 104, + 665, + 190, + 680 + ], + "score": 1.0, + "content": "associate each edge", + "type": "text" + }, + { + "bbox": [ + 190, + 667, + 213, + 678 + ], + "score": 0.89, + "content": "( u , v )", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 665, + 292, + 680 + ], + "score": 1.0, + "content": "with an edge type", + "type": "text" + }, + { + "bbox": [ + 292, + 666, + 384, + 679 + ], + "score": 0.91, + "content": "c _ { ( u , v ) } \\in \\{ 1 , 2 , \\ldots , C \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 665, + 505, + 680 + ], + "score": 1.0, + "content": ". Node type can help in cap-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "turing different importances across nodes. Edge type can be used to describe different relationships", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "between nodes, and thus propagate information between them differently. One can also add more", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "than one edge type to the same edge which results in a multi-graph. We stick to simple graphs for", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "simplicity. One interesting fact is that we have two notions of “time” in our model. One is the time", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "step in the environment which is the typical time coordinate for RL problems. The other corresponds", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 35.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 87, + 502, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 87, + 502, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 87, + 502, + 225 + ], + "spans": [ + { + "bbox": [ + 110, + 87, + 502, + 225 + ], + "score": 0.971, + "type": "image", + "image_path": "4387bb37b2b6edf82cb3edff6c91e9f09012e99f1638fa5f05208b0a537fe83a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 87, + 502, + 133.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 133.0, + 502, + 179.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 179.0, + 502, + 225.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 240, + 505, + 295 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 239, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 505, + 253 + ], + "score": 1.0, + "content": "Figure 2: In this figure, we use Walker-Ostrich as an example of NerveNet. In the input model,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "for each node, NerveNet fetches the corresponding elements from the observation vector. NerveNet", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 261, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 275 + ], + "score": 1.0, + "content": "then computes the messages between neighbors in the graph, and updates the hidden state of each", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 273, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 505, + 286 + ], + "score": 1.0, + "content": "node. This process is repeated for a certain number of propagation steps. In the output model, the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 284, + 365, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 365, + 297 + ], + "score": 1.0, + "content": "policy is produced by collecting the output from each controller.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "title", + "bbox": [ + 107, + 318, + 234, + 329 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 235, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 235, + 331 + ], + "score": 1.0, + "content": "2.1 GRAPH CONSTRUCTION", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 339, + 505, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "In real life, skeletons of most robots and animals have a discrete graph structure, and are most often", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 348, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 506, + 364 + ], + "score": 1.0, + "content": "trees. Simulators such as the MuJoCo engine by Todorov et al. (2012), organize the agents using an", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 360, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 505, + 374 + ], + "score": 1.0, + "content": "XML-based kinematic tree. In our experiments, we will use the tree graphs as per MuJoCo. Note", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 371, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 506, + 385 + ], + "score": 1.0, + "content": "that our model can be directly applied to arbitrary graphs. In particular, we assume two types of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 383, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 506, + 395 + ], + "score": 1.0, + "content": "nodes in our tree: body and joint. The body nodes are abstract nodes used to construct the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 393, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 506, + 406 + ], + "score": 1.0, + "content": "kinematic tree via nesting, which is similar to the coordinate frame system used in robotics, Spong", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 403, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 104, + 403, + 505, + 418 + ], + "score": 1.0, + "content": "et al. (2006). The joint node represents the degrees of freedom of motion between the two body", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 416, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 506, + 428 + ], + "score": 1.0, + "content": "nodes. Take a simple humanoid as an example; the body nodes Thigh and Shin are connected via", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "score": 1.0, + "content": "the Knee, where Knee is a hinge joint. We further add a root node which observes additional", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 438, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 450 + ], + "score": 1.0, + "content": "information about the agent. For example, in the Reacher agent in MuJoCo, the root node has", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 449, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 505, + 461 + ], + "score": 1.0, + "content": "access to the target position of the agent. We build edges to form a tree graph. Fig. 1 illustrates the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 505, + 472 + ], + "score": 1.0, + "content": "graph structure of an example agent, CentipedeEight. We sketch the agent and its correspond-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 470, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 104, + 470, + 506, + 483 + ], + "score": 1.0, + "content": "ing graph in the left and right part of the figure, respectively. Note that for better visualization, we", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 481, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 505, + 493 + ], + "score": 1.0, + "content": "omit the joint nodes and use edges to represent the physical connections of joint nodes. Dif-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "ferent elements of the agent are parsed into nodes with different colors. Further details are provided", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 503, + 219, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 219, + 516 + ], + "score": 1.0, + "content": "in the experimental section.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 16.5, + "bbox_fs": [ + 104, + 338, + 506, + 516 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 529, + 275, + 540 + ], + "lines": [ + { + "bbox": [ + 106, + 529, + 277, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 277, + 542 + ], + "score": 1.0, + "content": "2.2 NERVENET AS POLICY NETWORK", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 550, + 505, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 562 + ], + "score": 1.0, + "content": "We now turn to NerveNet which parametrizes the policy with a Graph Neural Network. Before", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "delving into details, we first introduce our notation. We then specify the input model which helps", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "to initialize the hidden state of each node. We further introduce the propagation model that updates", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 583, + 342, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 342, + 595 + ], + "score": 1.0, + "content": "these hidden states. Finally, we describe the output model.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 550, + 505, + 595 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 600, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 600, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 289, + 612 + ], + "score": 1.0, + "content": "We denote the graph structure of the agent as", + "type": "text" + }, + { + "bbox": [ + 289, + 600, + 338, + 612 + ], + "score": 0.93, + "content": "G = ( V , E )", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 600, + 366, + 612 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 366, + 600, + 376, + 610 + ], + "score": 0.76, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 600, + 394, + 612 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 394, + 601, + 403, + 610 + ], + "score": 0.82, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 600, + 505, + 612 + ], + "score": 1.0, + "content": "are the sets of nodes and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "score": 1.0, + "content": "edges, respectively. We focus on the directed graphs as the undirected case can be easily addressed", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "by splitting one undirected edge into two directed edges. We denote the out-going neighborhood", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 139, + 646 + ], + "score": 1.0, + "content": "of node", + "type": "text" + }, + { + "bbox": [ + 140, + 635, + 147, + 643 + ], + "score": 0.7, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 632, + 159, + 646 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 159, + 633, + 194, + 645 + ], + "score": 0.93, + "content": "\\mathcal { N } _ { o u t } ( u )", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 632, + 313, + 646 + ], + "score": 1.0, + "content": "which contains all endpoints", + "type": "text" + }, + { + "bbox": [ + 313, + 635, + 320, + 643 + ], + "score": 0.75, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 632, + 342, + 646 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 342, + 633, + 365, + 644 + ], + "score": 0.9, + "content": "( u , v )", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 632, + 505, + 646 + ], + "score": 1.0, + "content": "being an edge in the graph. Simi-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 321, + 657 + ], + "score": 1.0, + "content": "larly, we denote the in-coming neighborhood of node", + "type": "text" + }, + { + "bbox": [ + 322, + 646, + 329, + 654 + ], + "score": 0.75, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 643, + 341, + 657 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 341, + 644, + 372, + 656 + ], + "score": 0.91, + "content": "\\mathcal { N } _ { i n } ( u )", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 643, + 425, + 657 + ], + "score": 1.0, + "content": ". Every node", + "type": "text" + }, + { + "bbox": [ + 425, + 646, + 432, + 654 + ], + "score": 0.72, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "has an associated", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 148, + 668 + ], + "score": 1.0, + "content": "node type", + "type": "text" + }, + { + "bbox": [ + 148, + 655, + 226, + 667 + ], + "score": 0.92, + "content": "p _ { u } \\in \\{ 1 , 2 , \\ldots , P \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 654, + 505, + 668 + ], + "score": 1.0, + "content": ", which in our case corresponds to body, joint and root. We also", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 665, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 104, + 665, + 190, + 680 + ], + "score": 1.0, + "content": "associate each edge", + "type": "text" + }, + { + "bbox": [ + 190, + 667, + 213, + 678 + ], + "score": 0.89, + "content": "( u , v )", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 665, + 292, + 680 + ], + "score": 1.0, + "content": "with an edge type", + "type": "text" + }, + { + "bbox": [ + 292, + 666, + 384, + 679 + ], + "score": 0.91, + "content": "c _ { ( u , v ) } \\in \\{ 1 , 2 , \\ldots , C \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 665, + 505, + 680 + ], + "score": 1.0, + "content": ". Node type can help in cap-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "turing different importances across nodes. Edge type can be used to describe different relationships", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "between nodes, and thus propagate information between them differently. One can also add more", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "than one edge type to the same edge which results in a multi-graph. We stick to simple graphs for", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "simplicity. One interesting fact is that we have two notions of “time” in our model. One is the time", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "step in the environment which is the typical time coordinate for RL problems. The other corresponds", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "to the internal propagation step of NerveNet. These two coordinates work as follows. At each time", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "step of the environment, NerveNet receives observation from the environment and performs a few", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "internal propagation steps in order to decide on the action to be taken by each node. To avoid con-", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 260, + 128 + ], + "score": 1.0, + "content": "fusion, throughout this paper, we use", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 261, + 118, + 267, + 125 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 268, + 114, + 468, + 128 + ], + "score": 1.0, + "content": "to describe the time step in the environment and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 469, + 116, + 474, + 125 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 474, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "for the", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 177, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 177, + 140 + ], + "score": 1.0, + "content": "propagation step.", + "type": "text", + "cross_page": true + } + ], + "index": 4 + } + ], + "index": 35.5, + "bbox_fs": [ + 104, + 600, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "to the internal propagation step of NerveNet. These two coordinates work as follows. At each time", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "step of the environment, NerveNet receives observation from the environment and performs a few", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "internal propagation steps in order to decide on the action to be taken by each node. To avoid con-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 260, + 128 + ], + "score": 1.0, + "content": "fusion, throughout this paper, we use", + "type": "text" + }, + { + "bbox": [ + 261, + 118, + 267, + 125 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 114, + 468, + 128 + ], + "score": 1.0, + "content": "to describe the time step in the environment and", + "type": "text" + }, + { + "bbox": [ + 469, + 116, + 474, + 125 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "for the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 177, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 177, + 140 + ], + "score": 1.0, + "content": "propagation step.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 107, + 149, + 201, + 161 + ], + "lines": [ + { + "bbox": [ + 106, + 149, + 203, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 203, + 162 + ], + "score": 1.0, + "content": "2.2.1 INPUT MODEL", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 168, + 505, + 224 + ], + "lines": [ + { + "bbox": [ + 106, + 169, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 106, + 169, + 182, + 180 + ], + "score": 1.0, + "content": "For each time step", + "type": "text" + }, + { + "bbox": [ + 182, + 171, + 189, + 179 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 169, + 404, + 180 + ], + "score": 1.0, + "content": "in the environment, the agent receives an observation", + "type": "text" + }, + { + "bbox": [ + 404, + 169, + 434, + 179 + ], + "score": 0.9, + "content": "s ^ { \\tau } \\in \\boldsymbol { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 169, + 505, + 180 + ], + "score": 1.0, + "content": ". The observation", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 180, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 134, + 192 + ], + "score": 1.0, + "content": "vector", + "type": "text" + }, + { + "bbox": [ + 134, + 181, + 144, + 190 + ], + "score": 0.86, + "content": "s ^ { \\tau }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 180, + 505, + 192 + ], + "score": 1.0, + "content": "is the concatenation of observations of each node. We denote the elements of observation", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 190, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 133, + 204 + ], + "score": 1.0, + "content": "vector", + "type": "text" + }, + { + "bbox": [ + 134, + 192, + 145, + 201 + ], + "score": 0.86, + "content": "s ^ { \\tau }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 190, + 237, + 204 + ], + "score": 1.0, + "content": "corresponding to node", + "type": "text" + }, + { + "bbox": [ + 237, + 193, + 244, + 201 + ], + "score": 0.75, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 190, + 265, + 204 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 266, + 192, + 277, + 202 + ], + "score": 0.85, + "content": "x _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 190, + 505, + 204 + ], + "score": 1.0, + "content": ". From now on, we drop the time step in the environment", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 201, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 506, + 214 + ], + "score": 1.0, + "content": "to derive the model for simplicity. The observation vector goes through an input network to obtain", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 212, + 249, + 226 + ], + "spans": [ + { + "bbox": [ + 104, + 212, + 249, + 226 + ], + "score": 1.0, + "content": "a fixed-size state vector as follows:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8 + }, + { + "type": "interline_equation", + "bbox": [ + 275, + 227, + 335, + 242 + ], + "lines": [ + { + "bbox": [ + 275, + 227, + 335, + 242 + ], + "spans": [ + { + "bbox": [ + 275, + 227, + 335, + 242 + ], + "score": 0.93, + "content": "h _ { u } ^ { 0 } = F _ { \\mathrm { i n } } ( x _ { u } ) ,", + "type": "interline_equation", + "image_path": "4d3d1bea999308a51e89d71addd9f2cddc61e7d0b475a53154f5b75a5f56b894.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 275, + 227, + 335, + 242 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 245, + 505, + 279 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 505, + 259 + ], + "score": 1.0, + "content": "where the subscript and superscript denote the node index and propagation step, respectively. Here,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 107, + 256, + 506, + 270 + ], + "spans": [ + { + "bbox": [ + 107, + 257, + 120, + 268 + ], + "score": 0.88, + "content": "F _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 256, + 198, + 270 + ], + "score": 1.0, + "content": "may be a MLP and", + "type": "text" + }, + { + "bbox": [ + 199, + 256, + 210, + 268 + ], + "score": 0.91, + "content": "h _ { u } ^ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 256, + 315, + 270 + ], + "score": 1.0, + "content": "is the state vector of node", + "type": "text" + }, + { + "bbox": [ + 315, + 259, + 322, + 267 + ], + "score": 0.75, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 256, + 506, + 270 + ], + "score": 1.0, + "content": "at propagation step 0. Note that we may need", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 268, + 478, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 478, + 280 + ], + "score": 1.0, + "content": "to pad zeros to the observation vectors if different nodes have observations of different sizes.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 107, + 290, + 236, + 302 + ], + "lines": [ + { + "bbox": [ + 105, + 289, + 237, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 237, + 303 + ], + "score": 1.0, + "content": "2.2.2 PROPAGATION MODEL", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 310, + 505, + 355 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 324 + ], + "score": 1.0, + "content": "We now describe the propagation model of our NerveNet which mimics a synchronous message", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "score": 1.0, + "content": "passing system studied in distributed computing, Attiya & Welch (2004). We will show how the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 332, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 505, + 345 + ], + "score": 1.0, + "content": "state vector of each node is updated from one propagation step to the next. This update process is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 343, + 456, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 456, + 356 + ], + "score": 1.0, + "content": "recurrently applied during the whole propagation. We leave the details to the appendix.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 107, + 366, + 505, + 389 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 344, + 379 + ], + "score": 1.0, + "content": "Message Computation In particular, at propagation step", + "type": "text" + }, + { + "bbox": [ + 344, + 368, + 349, + 376 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 366, + 413, + 379 + ], + "score": 1.0, + "content": ", for every node", + "type": "text" + }, + { + "bbox": [ + 414, + 369, + 420, + 376 + ], + "score": 0.71, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 366, + 506, + 379 + ], + "score": 1.0, + "content": ", we have access to a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 376, + 488, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 154, + 390 + ], + "score": 1.0, + "content": "state vector", + "type": "text" + }, + { + "bbox": [ + 155, + 378, + 167, + 389 + ], + "score": 0.9, + "content": "h _ { u } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 376, + 233, + 390 + ], + "score": 1.0, + "content": ". For every edge", + "type": "text" + }, + { + "bbox": [ + 233, + 377, + 303, + 390 + ], + "score": 0.94, + "content": "( u , v ) \\in \\bar { \\mathcal { N } } _ { o u t } \\bar { ( u ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 376, + 328, + 390 + ], + "score": 1.0, + "content": ", node", + "type": "text" + }, + { + "bbox": [ + 329, + 379, + 336, + 387 + ], + "score": 0.75, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 376, + 488, + 390 + ], + "score": 1.0, + "content": "computes a message vector as below,", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "interline_equation", + "bbox": [ + 258, + 392, + 352, + 408 + ], + "lines": [ + { + "bbox": [ + 258, + 392, + 352, + 408 + ], + "spans": [ + { + "bbox": [ + 258, + 392, + 352, + 408 + ], + "score": 0.92, + "content": "m _ { ( u , v ) } ^ { t } = M _ { c _ { ( u , v ) } } ( h _ { u } ^ { t } ) ,", + "type": "interline_equation", + "image_path": "9a00b23aa896ebed9b90bc2eab833d6bda16d764875375b3e56d783dd56d6394.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 258, + 392, + 352, + 408 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 411, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 108, + 410, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 108, + 410, + 172, + 426 + ], + "score": 1.0, + "content": "where Mc(u,v) i", + "type": "text" + }, + { + "bbox": [ + 167, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "s the message function which may be an identity mapping or a MLP. Note that the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 422, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 145, + 437 + ], + "score": 1.0, + "content": "subscript", + "type": "text" + }, + { + "bbox": [ + 146, + 426, + 168, + 435 + ], + "score": 0.89, + "content": "c _ { ( u , v ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 422, + 506, + 437 + ], + "score": 1.0, + "content": "indicates that edges of the same edge type share the same instance of the message", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "function. For example, the second torso in Fig. 1 sends a message to the first and third torso, as", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 444, + 266, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 266, + 457 + ], + "score": 1.0, + "content": "well as the LeftHip and RightHip.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 108, + 468, + 505, + 501 + ], + "lines": [ + { + "bbox": [ + 105, + 466, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 482 + ], + "score": 1.0, + "content": "Message Aggregation Once every node finishes computing messages, we aggregate messages", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 478, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 427, + 492 + ], + "score": 1.0, + "content": "sent from all in-coming neighbors of each node. Specifically, for every node", + "type": "text" + }, + { + "bbox": [ + 428, + 481, + 434, + 489 + ], + "score": 0.72, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 478, + 505, + 492 + ], + "score": 1.0, + "content": ", we perform the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 490, + 200, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 200, + 502 + ], + "score": 1.0, + "content": "following aggregation:", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 504, + 364, + 519 + ], + "lines": [ + { + "bbox": [ + 246, + 504, + 364, + 519 + ], + "spans": [ + { + "bbox": [ + 246, + 504, + 364, + 519 + ], + "score": 0.93, + "content": "\\bar { m } _ { u } ^ { t } = A ( \\{ h _ { v } ^ { t } | v \\in \\mathcal { N } _ { i n } ( u ) \\} ) ,", + "type": "interline_equation", + "image_path": "6da0a44d915afbd5d8e27938b4c7ead1a6c9099a095d41838359532a55d3c41d.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 246, + 504, + 364, + 519 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 108, + 522, + 504, + 536 + ], + "spans": [ + { + "bbox": [ + 108, + 522, + 136, + 536 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 136, + 524, + 145, + 533 + ], + "score": 0.8, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 522, + 504, + 536 + ], + "score": 1.0, + "content": "is the aggregation function which may be a summation, average or max-pooling function.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 532, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 132, + 547 + ], + "score": 1.0, + "content": "Here,", + "type": "text" + }, + { + "bbox": [ + 132, + 534, + 147, + 546 + ], + "score": 0.91, + "content": "\\hat { m } _ { u } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 532, + 506, + 547 + ], + "score": 1.0, + "content": "is the aggregated message vector which contains the information sent from the node’s", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 544, + 167, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 167, + 557 + ], + "score": 1.0, + "content": "neighborhood.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 105, + 568, + 503, + 591 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 582 + ], + "score": 1.0, + "content": "States Update We now update every node’s state vector based on both the aggregated message", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 579, + 477, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 330, + 592 + ], + "score": 1.0, + "content": "and its current state vector. In particular, for every node", + "type": "text" + }, + { + "bbox": [ + 330, + 582, + 337, + 589 + ], + "score": 0.77, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 579, + 477, + 592 + ], + "score": 1.0, + "content": ", we perform the following update:", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "interline_equation", + "bbox": [ + 260, + 594, + 350, + 609 + ], + "lines": [ + { + "bbox": [ + 260, + 594, + 350, + 609 + ], + "spans": [ + { + "bbox": [ + 260, + 594, + 350, + 609 + ], + "score": 0.93, + "content": "h _ { u } ^ { t + 1 } = U _ { p _ { u } } ( h _ { u } ^ { t } , \\bar { m } _ { u } ^ { t } ) ,", + "type": "interline_equation", + "image_path": "5730230f07c18c79391a9f1b4a21e7d335ee5509bbe644efaaa609de68afa92a.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 260, + 594, + 350, + 609 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 612, + 505, + 668 + ], + "lines": [ + { + "bbox": [ + 110, + 611, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 110, + 611, + 138, + 624 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 138, + 613, + 147, + 622 + ], + "score": 0.79, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 611, + 505, + 624 + ], + "score": 1.0, + "content": "is the update function which may be a gated recurrent unit (GRU), a long short term", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 623, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 322, + 635 + ], + "score": 1.0, + "content": "memory (LSTM) unit or a MLP. From the subscript", + "type": "text" + }, + { + "bbox": [ + 322, + 625, + 334, + 635 + ], + "score": 0.87, + "content": "p _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 623, + 348, + 635 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 348, + 624, + 357, + 633 + ], + "score": 0.82, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 623, + 505, + 635 + ], + "score": 1.0, + "content": ", we can see that nodes of the same", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 633, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 647 + ], + "score": 1.0, + "content": "node type share the same instance of the update function. The above propagation model is then", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 644, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 314, + 658 + ], + "score": 1.0, + "content": "recurrently applied for a fixed number of time steps", + "type": "text" + }, + { + "bbox": [ + 315, + 646, + 323, + 655 + ], + "score": 0.84, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 644, + 506, + 658 + ], + "score": 1.0, + "content": "to get the final state vectors of all nodes, i.e.,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 107, + 654, + 163, + 670 + ], + "spans": [ + { + "bbox": [ + 107, + 655, + 158, + 668 + ], + "score": 0.92, + "content": "\\{ h _ { u } ^ { T } | u \\in \\bar { V } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 654, + 163, + 670 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39 + }, + { + "type": "title", + "bbox": [ + 107, + 679, + 211, + 690 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 213, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 213, + 692 + ], + "score": 1.0, + "content": "2.2.3 OUTPUT MODEL", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "In RL, agents typically use a MLP policy, where the network outputs the mean of the Gaussian", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 710, + 504, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 504, + 721 + ], + "score": 1.0, + "content": "distribution for each of the actions, while the standard deviation is a trainable vector, Schulman", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 720, + 442, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 442, + 734 + ], + "score": 1.0, + "content": "et al. (2017). In our output model, we also treat standard deviation in the same way.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 8 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [], + "index": 2, + "bbox_fs": [ + 105, + 82, + 505, + 140 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 149, + 201, + 161 + ], + "lines": [ + { + "bbox": [ + 106, + 149, + 203, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 203, + 162 + ], + "score": 1.0, + "content": "2.2.1 INPUT MODEL", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 168, + 505, + 224 + ], + "lines": [ + { + "bbox": [ + 106, + 169, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 106, + 169, + 182, + 180 + ], + "score": 1.0, + "content": "For each time step", + "type": "text" + }, + { + "bbox": [ + 182, + 171, + 189, + 179 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 169, + 404, + 180 + ], + "score": 1.0, + "content": "in the environment, the agent receives an observation", + "type": "text" + }, + { + "bbox": [ + 404, + 169, + 434, + 179 + ], + "score": 0.9, + "content": "s ^ { \\tau } \\in \\boldsymbol { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 169, + 505, + 180 + ], + "score": 1.0, + "content": ". The observation", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 180, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 134, + 192 + ], + "score": 1.0, + "content": "vector", + "type": "text" + }, + { + "bbox": [ + 134, + 181, + 144, + 190 + ], + "score": 0.86, + "content": "s ^ { \\tau }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 180, + 505, + 192 + ], + "score": 1.0, + "content": "is the concatenation of observations of each node. We denote the elements of observation", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 190, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 133, + 204 + ], + "score": 1.0, + "content": "vector", + "type": "text" + }, + { + "bbox": [ + 134, + 192, + 145, + 201 + ], + "score": 0.86, + "content": "s ^ { \\tau }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 190, + 237, + 204 + ], + "score": 1.0, + "content": "corresponding to node", + "type": "text" + }, + { + "bbox": [ + 237, + 193, + 244, + 201 + ], + "score": 0.75, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 190, + 265, + 204 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 266, + 192, + 277, + 202 + ], + "score": 0.85, + "content": "x _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 190, + 505, + 204 + ], + "score": 1.0, + "content": ". From now on, we drop the time step in the environment", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 201, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 506, + 214 + ], + "score": 1.0, + "content": "to derive the model for simplicity. The observation vector goes through an input network to obtain", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 212, + 249, + 226 + ], + "spans": [ + { + "bbox": [ + 104, + 212, + 249, + 226 + ], + "score": 1.0, + "content": "a fixed-size state vector as follows:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8, + "bbox_fs": [ + 104, + 169, + 506, + 226 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 275, + 227, + 335, + 242 + ], + "lines": [ + { + "bbox": [ + 275, + 227, + 335, + 242 + ], + "spans": [ + { + "bbox": [ + 275, + 227, + 335, + 242 + ], + "score": 0.93, + "content": "h _ { u } ^ { 0 } = F _ { \\mathrm { i n } } ( x _ { u } ) ,", + "type": "interline_equation", + "image_path": "4d3d1bea999308a51e89d71addd9f2cddc61e7d0b475a53154f5b75a5f56b894.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 275, + 227, + 335, + 242 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 245, + 505, + 279 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 505, + 259 + ], + "score": 1.0, + "content": "where the subscript and superscript denote the node index and propagation step, respectively. Here,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 107, + 256, + 506, + 270 + ], + "spans": [ + { + "bbox": [ + 107, + 257, + 120, + 268 + ], + "score": 0.88, + "content": "F _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 256, + 198, + 270 + ], + "score": 1.0, + "content": "may be a MLP and", + "type": "text" + }, + { + "bbox": [ + 199, + 256, + 210, + 268 + ], + "score": 0.91, + "content": "h _ { u } ^ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 256, + 315, + 270 + ], + "score": 1.0, + "content": "is the state vector of node", + "type": "text" + }, + { + "bbox": [ + 315, + 259, + 322, + 267 + ], + "score": 0.75, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 256, + 506, + 270 + ], + "score": 1.0, + "content": "at propagation step 0. Note that we may need", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 268, + 478, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 478, + 280 + ], + "score": 1.0, + "content": "to pad zeros to the observation vectors if different nodes have observations of different sizes.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 245, + 506, + 280 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 290, + 236, + 302 + ], + "lines": [ + { + "bbox": [ + 105, + 289, + 237, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 237, + 303 + ], + "score": 1.0, + "content": "2.2.2 PROPAGATION MODEL", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 310, + 505, + 355 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 324 + ], + "score": 1.0, + "content": "We now describe the propagation model of our NerveNet which mimics a synchronous message", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "score": 1.0, + "content": "passing system studied in distributed computing, Attiya & Welch (2004). We will show how the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 332, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 505, + 345 + ], + "score": 1.0, + "content": "state vector of each node is updated from one propagation step to the next. This update process is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 343, + 456, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 456, + 356 + ], + "score": 1.0, + "content": "recurrently applied during the whole propagation. We leave the details to the appendix.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 309, + 505, + 356 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 366, + 505, + 389 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 344, + 379 + ], + "score": 1.0, + "content": "Message Computation In particular, at propagation step", + "type": "text" + }, + { + "bbox": [ + 344, + 368, + 349, + 376 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 366, + 413, + 379 + ], + "score": 1.0, + "content": ", for every node", + "type": "text" + }, + { + "bbox": [ + 414, + 369, + 420, + 376 + ], + "score": 0.71, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 366, + 506, + 379 + ], + "score": 1.0, + "content": ", we have access to a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 376, + 488, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 154, + 390 + ], + "score": 1.0, + "content": "state vector", + "type": "text" + }, + { + "bbox": [ + 155, + 378, + 167, + 389 + ], + "score": 0.9, + "content": "h _ { u } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 376, + 233, + 390 + ], + "score": 1.0, + "content": ". For every edge", + "type": "text" + }, + { + "bbox": [ + 233, + 377, + 303, + 390 + ], + "score": 0.94, + "content": "( u , v ) \\in \\bar { \\mathcal { N } } _ { o u t } \\bar { ( u ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 376, + 328, + 390 + ], + "score": 1.0, + "content": ", node", + "type": "text" + }, + { + "bbox": [ + 329, + 379, + 336, + 387 + ], + "score": 0.75, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 376, + 488, + 390 + ], + "score": 1.0, + "content": "computes a message vector as below,", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 366, + 506, + 390 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 258, + 392, + 352, + 408 + ], + "lines": [ + { + "bbox": [ + 258, + 392, + 352, + 408 + ], + "spans": [ + { + "bbox": [ + 258, + 392, + 352, + 408 + ], + "score": 0.92, + "content": "m _ { ( u , v ) } ^ { t } = M _ { c _ { ( u , v ) } } ( h _ { u } ^ { t } ) ,", + "type": "interline_equation", + "image_path": "9a00b23aa896ebed9b90bc2eab833d6bda16d764875375b3e56d783dd56d6394.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 258, + 392, + 352, + 408 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 411, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 108, + 410, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 108, + 410, + 172, + 426 + ], + "score": 1.0, + "content": "where Mc(u,v) i", + "type": "text" + }, + { + "bbox": [ + 167, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "s the message function which may be an identity mapping or a MLP. Note that the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 422, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 145, + 437 + ], + "score": 1.0, + "content": "subscript", + "type": "text" + }, + { + "bbox": [ + 146, + 426, + 168, + 435 + ], + "score": 0.89, + "content": "c _ { ( u , v ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 422, + 506, + 437 + ], + "score": 1.0, + "content": "indicates that edges of the same edge type share the same instance of the message", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "function. For example, the second torso in Fig. 1 sends a message to the first and third torso, as", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 444, + 266, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 266, + 457 + ], + "score": 1.0, + "content": "well as the LeftHip and RightHip.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 410, + 506, + 457 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 468, + 505, + 501 + ], + "lines": [ + { + "bbox": [ + 105, + 466, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 482 + ], + "score": 1.0, + "content": "Message Aggregation Once every node finishes computing messages, we aggregate messages", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 478, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 427, + 492 + ], + "score": 1.0, + "content": "sent from all in-coming neighbors of each node. Specifically, for every node", + "type": "text" + }, + { + "bbox": [ + 428, + 481, + 434, + 489 + ], + "score": 0.72, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 478, + 505, + 492 + ], + "score": 1.0, + "content": ", we perform the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 490, + 200, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 200, + 502 + ], + "score": 1.0, + "content": "following aggregation:", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 466, + 505, + 502 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 504, + 364, + 519 + ], + "lines": [ + { + "bbox": [ + 246, + 504, + 364, + 519 + ], + "spans": [ + { + "bbox": [ + 246, + 504, + 364, + 519 + ], + "score": 0.93, + "content": "\\bar { m } _ { u } ^ { t } = A ( \\{ h _ { v } ^ { t } | v \\in \\mathcal { N } _ { i n } ( u ) \\} ) ,", + "type": "interline_equation", + "image_path": "6da0a44d915afbd5d8e27938b4c7ead1a6c9099a095d41838359532a55d3c41d.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 246, + 504, + 364, + 519 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 108, + 522, + 504, + 536 + ], + "spans": [ + { + "bbox": [ + 108, + 522, + 136, + 536 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 136, + 524, + 145, + 533 + ], + "score": 0.8, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 522, + 504, + 536 + ], + "score": 1.0, + "content": "is the aggregation function which may be a summation, average or max-pooling function.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 532, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 132, + 547 + ], + "score": 1.0, + "content": "Here,", + "type": "text" + }, + { + "bbox": [ + 132, + 534, + 147, + 546 + ], + "score": 0.91, + "content": "\\hat { m } _ { u } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 532, + 506, + 547 + ], + "score": 1.0, + "content": "is the aggregated message vector which contains the information sent from the node’s", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 544, + 167, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 167, + 557 + ], + "score": 1.0, + "content": "neighborhood.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 522, + 506, + 557 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 568, + 503, + 591 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 582 + ], + "score": 1.0, + "content": "States Update We now update every node’s state vector based on both the aggregated message", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 579, + 477, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 330, + 592 + ], + "score": 1.0, + "content": "and its current state vector. In particular, for every node", + "type": "text" + }, + { + "bbox": [ + 330, + 582, + 337, + 589 + ], + "score": 0.77, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 579, + 477, + 592 + ], + "score": 1.0, + "content": ", we perform the following update:", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 566, + 505, + 592 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 260, + 594, + 350, + 609 + ], + "lines": [ + { + "bbox": [ + 260, + 594, + 350, + 609 + ], + "spans": [ + { + "bbox": [ + 260, + 594, + 350, + 609 + ], + "score": 0.93, + "content": "h _ { u } ^ { t + 1 } = U _ { p _ { u } } ( h _ { u } ^ { t } , \\bar { m } _ { u } ^ { t } ) ,", + "type": "interline_equation", + "image_path": "5730230f07c18c79391a9f1b4a21e7d335ee5509bbe644efaaa609de68afa92a.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 260, + 594, + 350, + 609 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 612, + 505, + 668 + ], + "lines": [ + { + "bbox": [ + 110, + 611, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 110, + 611, + 138, + 624 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 138, + 613, + 147, + 622 + ], + "score": 0.79, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 611, + 505, + 624 + ], + "score": 1.0, + "content": "is the update function which may be a gated recurrent unit (GRU), a long short term", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 623, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 322, + 635 + ], + "score": 1.0, + "content": "memory (LSTM) unit or a MLP. From the subscript", + "type": "text" + }, + { + "bbox": [ + 322, + 625, + 334, + 635 + ], + "score": 0.87, + "content": "p _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 623, + 348, + 635 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 348, + 624, + 357, + 633 + ], + "score": 0.82, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 623, + 505, + 635 + ], + "score": 1.0, + "content": ", we can see that nodes of the same", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 633, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 647 + ], + "score": 1.0, + "content": "node type share the same instance of the update function. The above propagation model is then", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 644, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 314, + 658 + ], + "score": 1.0, + "content": "recurrently applied for a fixed number of time steps", + "type": "text" + }, + { + "bbox": [ + 315, + 646, + 323, + 655 + ], + "score": 0.84, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 644, + 506, + 658 + ], + "score": 1.0, + "content": "to get the final state vectors of all nodes, i.e.,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 107, + 654, + 163, + 670 + ], + "spans": [ + { + "bbox": [ + 107, + 655, + 158, + 668 + ], + "score": 0.92, + "content": "\\{ h _ { u } ^ { T } | u \\in \\bar { V } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 654, + 163, + 670 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 611, + 506, + 670 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 679, + 211, + 690 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 213, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 213, + 692 + ], + "score": 1.0, + "content": "2.2.3 OUTPUT MODEL", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "In RL, agents typically use a MLP policy, where the network outputs the mean of the Gaussian", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 710, + 504, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 504, + 721 + ], + "score": 1.0, + "content": "distribution for each of the actions, while the standard deviation is a trainable vector, Schulman", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 720, + 442, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 442, + 734 + ], + "score": 1.0, + "content": "et al. (2017). In our output model, we also treat standard deviation in the same way.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 698, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 171 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "However, instead of predicting the action distribution of all nodes by a single network, we make", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "predictions for each individual node. We denote the set of nodes which are assigned controllers", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 103, + 102, + 507, + 120 + ], + "spans": [ + { + "bbox": [ + 103, + 102, + 187, + 120 + ], + "score": 1.0, + "content": "for the actuators as", + "type": "text" + }, + { + "bbox": [ + 187, + 105, + 196, + 114 + ], + "score": 0.63, + "content": "\\mathcal { O }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 102, + 427, + 120 + ], + "score": 1.0, + "content": ". For each such node, a MLP takes its final state vectors", + "type": "text" + }, + { + "bbox": [ + 427, + 104, + 452, + 117 + ], + "score": 0.92, + "content": "h _ { u \\in \\mathcal { O } } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 102, + 507, + 120 + ], + "score": 1.0, + "content": "as input and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "produces the mean of the action of the Gaussian policy for the corresponding actuator. For each", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 158, + 139 + ], + "score": 1.0, + "content": "output node", + "type": "text" + }, + { + "bbox": [ + 158, + 127, + 189, + 137 + ], + "score": 0.9, + "content": "u \\in \\mathcal { O }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 127, + 309, + 139 + ], + "score": 1.0, + "content": ", we define its output type as", + "type": "text" + }, + { + "bbox": [ + 309, + 128, + 320, + 138 + ], + "score": 0.86, + "content": "q _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 127, + 505, + 139 + ], + "score": 1.0, + "content": ". Different sharing schemes are available for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 193, + 150 + ], + "score": 1.0, + "content": "the instance of MLP", + "type": "text" + }, + { + "bbox": [ + 193, + 137, + 210, + 149 + ], + "score": 0.93, + "content": "O _ { q _ { u } }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 137, + 506, + 150 + ], + "score": 1.0, + "content": ", for example, we can force the nodes with similar physical structure to", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 147, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 506, + 162 + ], + "score": 1.0, + "content": "share the instance of MLP. For example, in Fig. 1, two LeftHip nodes have a shared controller.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 300, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 300, + 171 + ], + "score": 1.0, + "content": "Therefore, we have the following output model:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5 + }, + { + "type": "interline_equation", + "bbox": [ + 267, + 171, + 344, + 187 + ], + "lines": [ + { + "bbox": [ + 267, + 171, + 344, + 187 + ], + "spans": [ + { + "bbox": [ + 267, + 171, + 344, + 187 + ], + "score": 0.92, + "content": "\\mu _ { u \\in \\mathcal { O } } = O _ { q _ { u } } ( h _ { u } ^ { T } ) ,", + "type": "interline_equation", + "image_path": "c1ca0583550ae71a4838cf1240500c3fb1ccac16546004d7251de3096fd83fd4.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 267, + 171, + 344, + 187 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 187, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 109, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 109, + 187, + 136, + 200 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 137, + 189, + 161, + 199 + ], + "score": 0.88, + "content": "\\mu _ { u \\in \\mathcal { O } }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "is the mean value for action applied on each actuator. In practice, we found that we", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "can force controllers of different output types to share one unified controller, while not hurting the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 223 + ], + "score": 1.0, + "content": "performance. By integrating the produced Gaussian policy for each action, the probability density", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 221, + 264, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 264, + 232 + ], + "score": 1.0, + "content": "of the stochastic policy is calculated as", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "interline_equation", + "bbox": [ + 178, + 232, + 432, + 262 + ], + "lines": [ + { + "bbox": [ + 178, + 232, + 432, + 262 + ], + "spans": [ + { + "bbox": [ + 178, + 232, + 432, + 262 + ], + "score": 0.94, + "content": "\\pi _ { \\boldsymbol { \\theta } } ( a ^ { \\tau } \\vert s ^ { \\tau } ) = \\prod _ { \\boldsymbol { u } \\in \\mathcal { O } } \\pi _ { \\boldsymbol { \\theta } , \\boldsymbol { u } } ( a _ { \\boldsymbol { u } } ^ { \\tau } \\vert s ^ { \\tau } ) = \\prod _ { \\boldsymbol { u } \\in \\mathcal { O } } \\frac { 1 } { \\sqrt { 2 \\pi \\sigma _ { \\boldsymbol { u } } ^ { 2 } } } e ^ { ( a _ { \\boldsymbol { u } } ^ { \\tau } - \\mu _ { \\boldsymbol { u } } ) ^ { 2 } / ( 2 \\sigma _ { \\boldsymbol { u } } ^ { 2 } ) } ,", + "type": "interline_equation", + "image_path": "d8dcb57eac4cd1d7463487642c75c2932c5bdd93450154d9ad19f794116027ac.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 178, + 232, + 432, + 242.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 178, + 242.0, + 432, + 252.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 178, + 252.0, + 432, + 262.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 504, + 285 + ], + "lines": [ + { + "bbox": [ + 108, + 261, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 108, + 261, + 136, + 275 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 136, + 263, + 168, + 273 + ], + "score": 0.91, + "content": "a ^ { \\tau } \\in { \\mathcal { A } }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 261, + 267, + 275 + ], + "score": 1.0, + "content": "is the output action, and", + "type": "text" + }, + { + "bbox": [ + 268, + 264, + 280, + 273 + ], + "score": 0.86, + "content": "\\sigma _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 261, + 505, + 275 + ], + "score": 1.0, + "content": "is the variable standard deviation for each action. Here,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 273, + 308, + 286 + ], + "spans": [ + { + "bbox": [ + 107, + 274, + 113, + 284 + ], + "score": 0.8, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 273, + 308, + 286 + ], + "score": 1.0, + "content": "represents the parameters of the policy function.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 107, + 298, + 234, + 309 + ], + "lines": [ + { + "bbox": [ + 105, + 297, + 235, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 235, + 310 + ], + "score": 1.0, + "content": "2.3 LEARNING ALGORITHM", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 318, + 505, + 362 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 410, + 331 + ], + "score": 1.0, + "content": "To interact with the environments, the agent generates its stochastic policy", + "type": "text" + }, + { + "bbox": [ + 410, + 318, + 451, + 330 + ], + "score": 0.9, + "content": "\\pi _ { \\boldsymbol { \\theta } } ( \\boldsymbol { a } ^ { \\tau } | \\boldsymbol { s } ^ { \\tau } )", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "after several", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 409, + 342 + ], + "score": 1.0, + "content": "propagation steps. The environment on the other hand, produces a reward", + "type": "text" + }, + { + "bbox": [ + 409, + 330, + 447, + 341 + ], + "score": 0.92, + "content": "r ( s ^ { \\tau } , a ^ { \\tau } )", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "for the agent,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 339, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 333, + 353 + ], + "score": 1.0, + "content": "and transits to the next state with transition probability", + "type": "text" + }, + { + "bbox": [ + 333, + 340, + 381, + 352 + ], + "score": 0.93, + "content": "\\mathsf { \\bar { P } } ( s ^ { \\tau + 1 } | s ^ { \\tau } )", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 339, + 506, + 353 + ], + "score": 1.0, + "content": ". The target of the agent is to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 351, + 232, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 232, + 363 + ], + "score": 1.0, + "content": "maximize its cumulative return", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 361, + 370, + 395 + ], + "lines": [ + { + "bbox": [ + 241, + 361, + 370, + 395 + ], + "spans": [ + { + "bbox": [ + 241, + 361, + 370, + 395 + ], + "score": 0.94, + "content": "J ( \\theta ) = \\mathbb { E } _ { \\pi } \\left[ \\sum _ { \\tau = 0 } ^ { \\infty } \\gamma ^ { \\tau } r ( s ^ { \\tau } , a ^ { \\tau } ) \\right] .", + "type": "interline_equation", + "image_path": "c85c65cb21c8b32e0696ec38231bd699b0c5d41bc53454e1f3598e7ad47274a1.jpg" + } + ] + } + ], + "index": 23.5, + "virtual_lines": [ + { + "bbox": [ + 241, + 361, + 370, + 378.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 241, + 378.0, + 370, + 395.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 401, + 505, + 489 + ], + "lines": [ + { + "bbox": [ + 106, + 401, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 505, + 413 + ], + "score": 1.0, + "content": "To optimize the expected reward, we use the proximal policy optimization (PPO) by Schulman", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "score": 1.0, + "content": "et al. (2017). In PPO, the agents alternate between sampling trajectories with the latest policy and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 424, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 435 + ], + "score": 1.0, + "content": "performing optimization on surrogate objective using the sampled trajectories. The algorithm tries to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 434, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 446 + ], + "score": 1.0, + "content": "keep the KL-divergence of the new policy and the old policy within the trust region. To achieve that,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 444, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 104, + 444, + 506, + 457 + ], + "score": 1.0, + "content": "PPO clips the probability ratio and adds an KL-divergence penalty term to the loss. The likelihood", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 455, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 104, + 455, + 180, + 470 + ], + "score": 1.0, + "content": "ratio is defined as", + "type": "text" + }, + { + "bbox": [ + 180, + 456, + 334, + 468 + ], + "score": 0.91, + "content": "r ^ { \\tau } ( \\boldsymbol { \\theta } ; \\mathbf { \\dot { \\theta } } _ { o l d } ) = \\pi _ { \\boldsymbol { \\theta } } ( a ^ { \\tau } | s ^ { \\tau } ) / \\pi _ { \\theta _ { o l d } } ( a ^ { \\tau } | \\mathbf { \\bar { s } } ^ { \\tau } )", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 455, + 506, + 470 + ], + "score": 1.0, + "content": ". Following the notation and the algorithm", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 465, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 104, + 465, + 505, + 480 + ], + "score": 1.0, + "content": "of PPO, our NerveNet tries to minimize the summation of the original loss in Eq. (7), KL-penalty", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 478, + 297, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 297, + 489 + ], + "score": 1.0, + "content": "and the value function loss which is defined as:", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28.5 + }, + { + "type": "interline_equation", + "bbox": [ + 116, + 488, + 482, + 574 + ], + "lines": [ + { + "bbox": [ + 116, + 488, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 116, + 488, + 482, + 574 + ], + "score": 0.94, + "content": "\\begin{array} { l } { \\displaystyle \\tilde { J } ( \\theta ) = J ( \\theta ) - \\beta L _ { K L } ( \\theta ) - \\alpha L _ { V } ( \\theta ) } \\\\ { \\displaystyle \\qquad = \\mathbb { E } _ { \\pi _ { \\theta } } \\left[ \\sum _ { \\tau = 0 } ^ { \\infty } \\operatorname* { m i n } \\left( \\hat { A } ^ { \\tau } r ^ { \\tau } ( \\theta ) , \\hat { A } ^ { \\tau } \\operatorname { c l i p } \\left( r ^ { \\tau } ( \\theta ) , 1 - \\epsilon , 1 + \\epsilon \\right) \\right) \\right] } \\\\ { \\displaystyle \\qquad - \\beta \\mathbb { E } _ { \\pi _ { \\theta } } \\left[ \\sum _ { \\tau = 0 } ^ { \\infty } \\mathrm { K L } \\left[ \\pi _ { \\theta } ( \\cdot | s ^ { \\tau } ) | \\pi _ { \\theta _ { o l d } } ( \\cdot | s ^ { \\tau } ) \\right] \\right] - \\alpha \\mathbb { E } _ { \\pi _ { \\theta } } \\left[ \\sum _ { \\tau = 0 } ^ { \\infty } \\left( V _ { \\theta } ( s ^ { \\tau } ) - V ( s ^ { \\tau } ) ^ { \\mathrm { t a r g e t } } \\right) ^ { 2 } \\right] , } \\end{array}", + "type": "interline_equation", + "image_path": "41caae5f10177c97ef816aa61d192768e07967153ed851eeedad01034bed3999.jpg" + } + ] + } + ], + "index": 34, + "virtual_lines": [ + { + "bbox": [ + 116, + 488, + 482, + 516.6666666666666 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 116, + 516.6666666666666, + 482, + 545.3333333333333 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 116, + 545.3333333333333, + 482, + 573.9999999999999 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 575, + 505, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 574, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 133, + 588 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 574, + 145, + 586 + ], + "score": 0.9, + "content": "\\hat { A } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 574, + 506, + 588 + ], + "score": 1.0, + "content": "is the generalized advantage estimation (GAE) calculated using algorithm from Schulman", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 181, + 599 + ], + "score": 1.0, + "content": "et al. (2015b), and", + "type": "text" + }, + { + "bbox": [ + 182, + 588, + 187, + 596 + ], + "score": 0.75, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 586, + 392, + 599 + ], + "score": 1.0, + "content": "is the clip value, which we choose to be 0.2. Here,", + "type": "text" + }, + { + "bbox": [ + 392, + 587, + 399, + 598 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "is a dynamical coefficient", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 318, + 609 + ], + "score": 1.0, + "content": "adjusted to keep the KL-divergence constraints, and", + "type": "text" + }, + { + "bbox": [ + 319, + 599, + 327, + 607 + ], + "score": 0.8, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "is used to balance the value loss. Note that", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 607, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 150, + 622 + ], + "score": 1.0, + "content": "in Eq. (8),", + "type": "text" + }, + { + "bbox": [ + 150, + 608, + 191, + 621 + ], + "score": 0.91, + "content": "V ( s _ { t } ) ^ { \\mathrm { t a r g e t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 607, + 506, + 622 + ], + "score": 1.0, + "content": "is the target state value in accordance with the GAE method. To optimize the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 619, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 107, + 619, + 127, + 633 + ], + "score": 0.91, + "content": "\\tilde { J } ( \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 619, + 506, + 635 + ], + "score": 1.0, + "content": ", PPO make use of the policy gradient in Sutton et al. (2000) to do first-order gradient descent", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 632, + 162, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 162, + 645 + ], + "score": 1.0, + "content": "optimization.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 290, + 668 + ], + "score": 1.0, + "content": "Value Network To produce the state value", + "type": "text" + }, + { + "bbox": [ + 290, + 655, + 319, + 667 + ], + "score": 0.93, + "content": "V _ { \\theta } ( s ^ { \\tau } )", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 654, + 409, + 668 + ], + "score": 1.0, + "content": "for given observation", + "type": "text" + }, + { + "bbox": [ + 410, + 656, + 420, + 665 + ], + "score": 0.85, + "content": "s ^ { \\tau }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 654, + 505, + 668 + ], + "score": 1.0, + "content": ", we have several al-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "score": 1.0, + "content": "ternatives: (1) using one GNN as the policy network and using one MLP as the value network", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "(NerveNet-MLP); (2) using one GNN as policy network and using another GNN as value network", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 702 + ], + "score": 1.0, + "content": "(NerveNet-2) (without sharing the parameters of the two GNNs); (3) using one GNN as both policy", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "network and value network (NerveNet-1). The GNN for value network is very similar to the GNN", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 721 + ], + "score": 1.0, + "content": "for policy network. The output for value GNN is a scalar instead of a vector of mean action. We", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 721, + 331, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 331, + 733 + ], + "score": 1.0, + "content": "will compare these variants in the experimental section.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 171 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "However, instead of predicting the action distribution of all nodes by a single network, we make", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "predictions for each individual node. We denote the set of nodes which are assigned controllers", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 103, + 102, + 507, + 120 + ], + "spans": [ + { + "bbox": [ + 103, + 102, + 187, + 120 + ], + "score": 1.0, + "content": "for the actuators as", + "type": "text" + }, + { + "bbox": [ + 187, + 105, + 196, + 114 + ], + "score": 0.63, + "content": "\\mathcal { O }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 102, + 427, + 120 + ], + "score": 1.0, + "content": ". For each such node, a MLP takes its final state vectors", + "type": "text" + }, + { + "bbox": [ + 427, + 104, + 452, + 117 + ], + "score": 0.92, + "content": "h _ { u \\in \\mathcal { O } } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 102, + 507, + 120 + ], + "score": 1.0, + "content": "as input and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "produces the mean of the action of the Gaussian policy for the corresponding actuator. For each", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 158, + 139 + ], + "score": 1.0, + "content": "output node", + "type": "text" + }, + { + "bbox": [ + 158, + 127, + 189, + 137 + ], + "score": 0.9, + "content": "u \\in \\mathcal { O }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 127, + 309, + 139 + ], + "score": 1.0, + "content": ", we define its output type as", + "type": "text" + }, + { + "bbox": [ + 309, + 128, + 320, + 138 + ], + "score": 0.86, + "content": "q _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 127, + 505, + 139 + ], + "score": 1.0, + "content": ". Different sharing schemes are available for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 193, + 150 + ], + "score": 1.0, + "content": "the instance of MLP", + "type": "text" + }, + { + "bbox": [ + 193, + 137, + 210, + 149 + ], + "score": 0.93, + "content": "O _ { q _ { u } }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 137, + 506, + 150 + ], + "score": 1.0, + "content": ", for example, we can force the nodes with similar physical structure to", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 147, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 506, + 162 + ], + "score": 1.0, + "content": "share the instance of MLP. For example, in Fig. 1, two LeftHip nodes have a shared controller.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 300, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 300, + 171 + ], + "score": 1.0, + "content": "Therefore, we have the following output model:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5, + "bbox_fs": [ + 103, + 82, + 507, + 171 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 267, + 171, + 344, + 187 + ], + "lines": [ + { + "bbox": [ + 267, + 171, + 344, + 187 + ], + "spans": [ + { + "bbox": [ + 267, + 171, + 344, + 187 + ], + "score": 0.92, + "content": "\\mu _ { u \\in \\mathcal { O } } = O _ { q _ { u } } ( h _ { u } ^ { T } ) ,", + "type": "interline_equation", + "image_path": "c1ca0583550ae71a4838cf1240500c3fb1ccac16546004d7251de3096fd83fd4.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 267, + 171, + 344, + 187 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 187, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 109, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 109, + 187, + 136, + 200 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 137, + 189, + 161, + 199 + ], + "score": 0.88, + "content": "\\mu _ { u \\in \\mathcal { O } }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "is the mean value for action applied on each actuator. In practice, we found that we", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "can force controllers of different output types to share one unified controller, while not hurting the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 223 + ], + "score": 1.0, + "content": "performance. By integrating the produced Gaussian policy for each action, the probability density", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 221, + 264, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 264, + 232 + ], + "score": 1.0, + "content": "of the stochastic policy is calculated as", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 187, + 506, + 232 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 178, + 232, + 432, + 262 + ], + "lines": [ + { + "bbox": [ + 178, + 232, + 432, + 262 + ], + "spans": [ + { + "bbox": [ + 178, + 232, + 432, + 262 + ], + "score": 0.94, + "content": "\\pi _ { \\boldsymbol { \\theta } } ( a ^ { \\tau } \\vert s ^ { \\tau } ) = \\prod _ { \\boldsymbol { u } \\in \\mathcal { O } } \\pi _ { \\boldsymbol { \\theta } , \\boldsymbol { u } } ( a _ { \\boldsymbol { u } } ^ { \\tau } \\vert s ^ { \\tau } ) = \\prod _ { \\boldsymbol { u } \\in \\mathcal { O } } \\frac { 1 } { \\sqrt { 2 \\pi \\sigma _ { \\boldsymbol { u } } ^ { 2 } } } e ^ { ( a _ { \\boldsymbol { u } } ^ { \\tau } - \\mu _ { \\boldsymbol { u } } ) ^ { 2 } / ( 2 \\sigma _ { \\boldsymbol { u } } ^ { 2 } ) } ,", + "type": "interline_equation", + "image_path": "d8dcb57eac4cd1d7463487642c75c2932c5bdd93450154d9ad19f794116027ac.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 178, + 232, + 432, + 242.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 178, + 242.0, + 432, + 252.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 178, + 252.0, + 432, + 262.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 504, + 285 + ], + "lines": [ + { + "bbox": [ + 108, + 261, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 108, + 261, + 136, + 275 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 136, + 263, + 168, + 273 + ], + "score": 0.91, + "content": "a ^ { \\tau } \\in { \\mathcal { A } }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 261, + 267, + 275 + ], + "score": 1.0, + "content": "is the output action, and", + "type": "text" + }, + { + "bbox": [ + 268, + 264, + 280, + 273 + ], + "score": 0.86, + "content": "\\sigma _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 261, + 505, + 275 + ], + "score": 1.0, + "content": "is the variable standard deviation for each action. Here,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 273, + 308, + 286 + ], + "spans": [ + { + "bbox": [ + 107, + 274, + 113, + 284 + ], + "score": 0.8, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 273, + 308, + 286 + ], + "score": 1.0, + "content": "represents the parameters of the policy function.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 107, + 261, + 505, + 286 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 298, + 234, + 309 + ], + "lines": [ + { + "bbox": [ + 105, + 297, + 235, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 235, + 310 + ], + "score": 1.0, + "content": "2.3 LEARNING ALGORITHM", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 318, + 505, + 362 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 410, + 331 + ], + "score": 1.0, + "content": "To interact with the environments, the agent generates its stochastic policy", + "type": "text" + }, + { + "bbox": [ + 410, + 318, + 451, + 330 + ], + "score": 0.9, + "content": "\\pi _ { \\boldsymbol { \\theta } } ( \\boldsymbol { a } ^ { \\tau } | \\boldsymbol { s } ^ { \\tau } )", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "after several", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 409, + 342 + ], + "score": 1.0, + "content": "propagation steps. The environment on the other hand, produces a reward", + "type": "text" + }, + { + "bbox": [ + 409, + 330, + 447, + 341 + ], + "score": 0.92, + "content": "r ( s ^ { \\tau } , a ^ { \\tau } )", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "for the agent,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 339, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 333, + 353 + ], + "score": 1.0, + "content": "and transits to the next state with transition probability", + "type": "text" + }, + { + "bbox": [ + 333, + 340, + 381, + 352 + ], + "score": 0.93, + "content": "\\mathsf { \\bar { P } } ( s ^ { \\tau + 1 } | s ^ { \\tau } )", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 339, + 506, + 353 + ], + "score": 1.0, + "content": ". The target of the agent is to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 351, + 232, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 232, + 363 + ], + "score": 1.0, + "content": "maximize its cumulative return", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 318, + 506, + 363 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 361, + 370, + 395 + ], + "lines": [ + { + "bbox": [ + 241, + 361, + 370, + 395 + ], + "spans": [ + { + "bbox": [ + 241, + 361, + 370, + 395 + ], + "score": 0.94, + "content": "J ( \\theta ) = \\mathbb { E } _ { \\pi } \\left[ \\sum _ { \\tau = 0 } ^ { \\infty } \\gamma ^ { \\tau } r ( s ^ { \\tau } , a ^ { \\tau } ) \\right] .", + "type": "interline_equation", + "image_path": "c85c65cb21c8b32e0696ec38231bd699b0c5d41bc53454e1f3598e7ad47274a1.jpg" + } + ] + } + ], + "index": 23.5, + "virtual_lines": [ + { + "bbox": [ + 241, + 361, + 370, + 378.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 241, + 378.0, + 370, + 395.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 401, + 505, + 489 + ], + "lines": [ + { + "bbox": [ + 106, + 401, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 505, + 413 + ], + "score": 1.0, + "content": "To optimize the expected reward, we use the proximal policy optimization (PPO) by Schulman", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "score": 1.0, + "content": "et al. (2017). In PPO, the agents alternate between sampling trajectories with the latest policy and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 424, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 435 + ], + "score": 1.0, + "content": "performing optimization on surrogate objective using the sampled trajectories. The algorithm tries to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 434, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 446 + ], + "score": 1.0, + "content": "keep the KL-divergence of the new policy and the old policy within the trust region. To achieve that,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 444, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 104, + 444, + 506, + 457 + ], + "score": 1.0, + "content": "PPO clips the probability ratio and adds an KL-divergence penalty term to the loss. The likelihood", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 455, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 104, + 455, + 180, + 470 + ], + "score": 1.0, + "content": "ratio is defined as", + "type": "text" + }, + { + "bbox": [ + 180, + 456, + 334, + 468 + ], + "score": 0.91, + "content": "r ^ { \\tau } ( \\boldsymbol { \\theta } ; \\mathbf { \\dot { \\theta } } _ { o l d } ) = \\pi _ { \\boldsymbol { \\theta } } ( a ^ { \\tau } | s ^ { \\tau } ) / \\pi _ { \\theta _ { o l d } } ( a ^ { \\tau } | \\mathbf { \\bar { s } } ^ { \\tau } )", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 455, + 506, + 470 + ], + "score": 1.0, + "content": ". Following the notation and the algorithm", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 465, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 104, + 465, + 505, + 480 + ], + "score": 1.0, + "content": "of PPO, our NerveNet tries to minimize the summation of the original loss in Eq. (7), KL-penalty", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 478, + 297, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 297, + 489 + ], + "score": 1.0, + "content": "and the value function loss which is defined as:", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28.5, + "bbox_fs": [ + 104, + 401, + 506, + 489 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 116, + 488, + 482, + 574 + ], + "lines": [ + { + "bbox": [ + 116, + 488, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 116, + 488, + 482, + 574 + ], + "score": 0.94, + "content": "\\begin{array} { l } { \\displaystyle \\tilde { J } ( \\theta ) = J ( \\theta ) - \\beta L _ { K L } ( \\theta ) - \\alpha L _ { V } ( \\theta ) } \\\\ { \\displaystyle \\qquad = \\mathbb { E } _ { \\pi _ { \\theta } } \\left[ \\sum _ { \\tau = 0 } ^ { \\infty } \\operatorname* { m i n } \\left( \\hat { A } ^ { \\tau } r ^ { \\tau } ( \\theta ) , \\hat { A } ^ { \\tau } \\operatorname { c l i p } \\left( r ^ { \\tau } ( \\theta ) , 1 - \\epsilon , 1 + \\epsilon \\right) \\right) \\right] } \\\\ { \\displaystyle \\qquad - \\beta \\mathbb { E } _ { \\pi _ { \\theta } } \\left[ \\sum _ { \\tau = 0 } ^ { \\infty } \\mathrm { K L } \\left[ \\pi _ { \\theta } ( \\cdot | s ^ { \\tau } ) | \\pi _ { \\theta _ { o l d } } ( \\cdot | s ^ { \\tau } ) \\right] \\right] - \\alpha \\mathbb { E } _ { \\pi _ { \\theta } } \\left[ \\sum _ { \\tau = 0 } ^ { \\infty } \\left( V _ { \\theta } ( s ^ { \\tau } ) - V ( s ^ { \\tau } ) ^ { \\mathrm { t a r g e t } } \\right) ^ { 2 } \\right] , } \\end{array}", + "type": "interline_equation", + "image_path": "41caae5f10177c97ef816aa61d192768e07967153ed851eeedad01034bed3999.jpg" + } + ] + } + ], + "index": 34, + "virtual_lines": [ + { + "bbox": [ + 116, + 488, + 482, + 516.6666666666666 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 116, + 516.6666666666666, + 482, + 545.3333333333333 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 116, + 545.3333333333333, + 482, + 573.9999999999999 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 575, + 505, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 574, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 133, + 588 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 574, + 145, + 586 + ], + "score": 0.9, + "content": "\\hat { A } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 574, + 506, + 588 + ], + "score": 1.0, + "content": "is the generalized advantage estimation (GAE) calculated using algorithm from Schulman", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 181, + 599 + ], + "score": 1.0, + "content": "et al. (2015b), and", + "type": "text" + }, + { + "bbox": [ + 182, + 588, + 187, + 596 + ], + "score": 0.75, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 586, + 392, + 599 + ], + "score": 1.0, + "content": "is the clip value, which we choose to be 0.2. Here,", + "type": "text" + }, + { + "bbox": [ + 392, + 587, + 399, + 598 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "is a dynamical coefficient", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 318, + 609 + ], + "score": 1.0, + "content": "adjusted to keep the KL-divergence constraints, and", + "type": "text" + }, + { + "bbox": [ + 319, + 599, + 327, + 607 + ], + "score": 0.8, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "is used to balance the value loss. Note that", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 607, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 150, + 622 + ], + "score": 1.0, + "content": "in Eq. (8),", + "type": "text" + }, + { + "bbox": [ + 150, + 608, + 191, + 621 + ], + "score": 0.91, + "content": "V ( s _ { t } ) ^ { \\mathrm { t a r g e t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 607, + 506, + 622 + ], + "score": 1.0, + "content": "is the target state value in accordance with the GAE method. To optimize the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 619, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 107, + 619, + 127, + 633 + ], + "score": 0.91, + "content": "\\tilde { J } ( \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 619, + 506, + 635 + ], + "score": 1.0, + "content": ", PPO make use of the policy gradient in Sutton et al. (2000) to do first-order gradient descent", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 632, + 162, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 162, + 645 + ], + "score": 1.0, + "content": "optimization.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 574, + 506, + 645 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 290, + 668 + ], + "score": 1.0, + "content": "Value Network To produce the state value", + "type": "text" + }, + { + "bbox": [ + 290, + 655, + 319, + 667 + ], + "score": 0.93, + "content": "V _ { \\theta } ( s ^ { \\tau } )", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 654, + 409, + 668 + ], + "score": 1.0, + "content": "for given observation", + "type": "text" + }, + { + "bbox": [ + 410, + 656, + 420, + 665 + ], + "score": 0.85, + "content": "s ^ { \\tau }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 654, + 505, + 668 + ], + "score": 1.0, + "content": ", we have several al-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "score": 1.0, + "content": "ternatives: (1) using one GNN as the policy network and using one MLP as the value network", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "(NerveNet-MLP); (2) using one GNN as policy network and using another GNN as value network", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 702 + ], + "score": 1.0, + "content": "(NerveNet-2) (without sharing the parameters of the two GNNs); (3) using one GNN as both policy", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "network and value network (NerveNet-1). The GNN for value network is very similar to the GNN", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 721 + ], + "score": 1.0, + "content": "for policy network. The output for value GNN is a scalar instead of a vector of mean action. We", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 721, + 331, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 331, + 733 + ], + "score": 1.0, + "content": "will compare these variants in the experimental section.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 654, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 211, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 213, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 213, + 96 + ], + "score": 1.0, + "content": "3 RELATED WORK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 107, + 505, + 173 + ], + "lines": [ + { + "bbox": [ + 105, + 106, + 506, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 506, + 120 + ], + "score": 1.0, + "content": "Reinforcement Learning Reinforcement learning (RL) has recently achieved huge success in a", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 118, + 505, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 505, + 131 + ], + "score": 1.0, + "content": "variety of applications. Powered by the progress of deep neural networks, Krizhevsky et al. (2012),", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 129, + 506, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 506, + 142 + ], + "score": 1.0, + "content": "agents are now able to successfully play Atari Games and beat the world’s best (human) players", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 141, + 504, + 152 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 504, + 152 + ], + "score": 1.0, + "content": "in the game of Go (Mnih et al., 2015; Silver et al., 2016; 2017). Based on simulation engines like", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 150, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 505, + 164 + ], + "score": 1.0, + "content": "MuJoCo, Todorov et al. (2012), numerous algorithms have been proposed to train agents also in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 162, + 500, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 500, + 174 + ], + "score": 1.0, + "content": "continuous control problems (Schulman et al., 2017; 2015a; Heess et al., 2017; Metz et al., 2017).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 186, + 505, + 319 + ], + "lines": [ + { + "bbox": [ + 106, + 187, + 504, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 504, + 199 + ], + "score": 1.0, + "content": "Structure in RL Most approaches that exploit priors on structure of the problem fall in the do-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "main of hierarchical RL, (Kulkarni et al., 2016; Vezhnevets et al., 2017), which mainly focus on", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "score": 1.0, + "content": "modeling intrinsic motivation of agents. In Hausknecht & Stone (2015), the authors extend the deep", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 220, + 506, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 506, + 233 + ], + "score": 1.0, + "content": "RL algorithms to MDPS with parameterized action space by exploiting the structure of action space", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 231, + 504, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 504, + 243 + ], + "score": 1.0, + "content": "and bounding the action space gradients. Graphs have been used in RL problems prior to our work.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "In Metzen (2013); Mabu et al. (2007); Shoeleh & Asadpour (2017); Mahadevan & Maggioni (2007),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "the authors use graphs to learn a representation of the environment. However, these methods are lim-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "score": 1.0, + "content": "ited to problems with simple dynamical models like for example the task of 2d-navigation, and thus", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 274, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 506, + 288 + ], + "score": 1.0, + "content": "these problems are usually solved via model-based RL. However, for complex multi-joint agents,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 285, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 298 + ], + "score": 1.0, + "content": "learning the dynamical model as well as predicting the transition of states is time consuming and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 297, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 309 + ], + "score": 1.0, + "content": "biased. For problems of training model-free multi-joint agents in complex physical environments,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 307, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 469, + 320 + ], + "score": 1.0, + "content": "relatively little attention has been devoted to modeling the physical structure of the agents.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 333, + 505, + 388 + ], + "lines": [ + { + "bbox": [ + 106, + 332, + 504, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 504, + 345 + ], + "score": 1.0, + "content": "Graph Neural Networks There have been many efforts to generalize neural networks to graph-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 344, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 505, + 356 + ], + "score": 1.0, + "content": "structured data. One line of work is based on convolutional neural networks (CNNs). In (Bruna", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "et al., 2014; Defferrard et al., 2016; Kipf & Welling, 2017), CNNs are employed in the spectral", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "score": 1.0, + "content": "domain relying on the graph Laplacian matrix. (Goller & Kuchler, 1996; Duvenaud et al., 2015)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 376, + 325, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 325, + 390 + ], + "score": 1.0, + "content": "used hash functions in order to apply CNNs to graphs.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 393, + 505, + 558 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 505, + 406 + ], + "score": 1.0, + "content": "Another popular direction is based on recurrent neural networks (RNNs) (Goller & Kuchler, 1996;", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 404, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 505, + 417 + ], + "score": 1.0, + "content": "Gori et al., 2005; Scarselli et al., 2009; Socher et al., 2011; Li et al., 2015; Tai et al., 2015).", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "score": 1.0, + "content": "Among RNN based methods, many are only applicable to special structured graph, e.g., sequences", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "or trees, (Socher et al., 2011; Tai et al., 2015). One class of models which are applicable to general", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "score": 1.0, + "content": "graphs are so-called graph neural networks (GNNs), Scarselli et al. (2009). The inference procedure", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 449, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 505, + 461 + ], + "score": 1.0, + "content": "is a forward pass that exploits a fixed-length propagation process which resembles synchronous mes-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 460, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 472 + ], + "score": 1.0, + "content": "sage passing system in the theory of distributed computing, Attiya & Welch (2004). Nodes in the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "score": 1.0, + "content": "graph have state vectors which are recurrently updated based on their history and received messages.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 482, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 494 + ], + "score": 1.0, + "content": "One of the representative work of GNNs, i.e., gated graph neural networks (GGNNs) by Li et al.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 492, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 505 + ], + "score": 1.0, + "content": "(2015), uses gated recurrent unit to update the state vectors. Learning such a model can be achieved", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 503, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 505, + 515 + ], + "score": 1.0, + "content": "by the back-propagation through time (BPTT) algorithm or recurrent back-propagation, Chauvin &", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 512, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 505, + 528 + ], + "score": 1.0, + "content": "Rumelhart (1995). It has been shown that GNNs, (Li et al., 2015; 2017; Qi et al., 2017; Gilmer et al.,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 524, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 538 + ], + "score": 1.0, + "content": "2017) have a high capacity and achieve state-of-the-art performance in many applications which in-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 535, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 506, + 550 + ], + "score": 1.0, + "content": "volve graph-structured data. In this paper, we model the structure of the reinforcement learning", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 548, + 189, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 189, + 559 + ], + "score": 1.0, + "content": "agents using GNNs.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "Transfer and Multi-task Learning in RL Recently, there has been increased interest in transfer", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 583, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 594 + ], + "score": 1.0, + "content": "learning tasks for RL, Taylor & Stone (2009), which mainly focus on transferring the policy learned", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "from one environment to another. In Rajeswaran et al. (2017b;a), the authors show that agents", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "in reinforcement learning are prone to over-fitting, and that the learned policies generalize poorly", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "across environments. In model-based RL, traditional control has been well studied for generalization", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 626, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 641 + ], + "score": 1.0, + "content": "properties, Coros et al. (2010). Gupta et al. (2017) try to increase the transferability via learning", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "invariant visual features. Efforts have also been made from the meta-learning perspective (Duan", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "et al., 2016; Finn et al., 2017b;a). In Wulfmeier et al. (2017), the authors propose a method of transfer", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "learning by using imitation learning. Transferability comes naturally in our model by exploiting the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 670, + 260, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 260, + 683 + ], + "score": 1.0, + "content": "(shared) graph structure of the agents.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "Multi-task learning has also received a lot of attention, Wilson et al. (2007). In Teh et al. (2017),", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "the authors use a distilled policy that captures common behaviour across tasks. Wilson et al. (2007);", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "Oh et al. (2017); Andreas et al. (2016) use a hierarchical approach, where multiple sub-policies are", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 721, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 506, + 732 + ], + "score": 1.0, + "content": "learnt. In Yang et al. (2017); Parisotto et al. (2015), the authors exploited shared visual features", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 50.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 211, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 213, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 213, + 96 + ], + "score": 1.0, + "content": "3 RELATED WORK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 107, + 505, + 173 + ], + "lines": [ + { + "bbox": [ + 105, + 106, + 506, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 506, + 120 + ], + "score": 1.0, + "content": "Reinforcement Learning Reinforcement learning (RL) has recently achieved huge success in a", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 118, + 505, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 505, + 131 + ], + "score": 1.0, + "content": "variety of applications. Powered by the progress of deep neural networks, Krizhevsky et al. (2012),", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 129, + 506, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 506, + 142 + ], + "score": 1.0, + "content": "agents are now able to successfully play Atari Games and beat the world’s best (human) players", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 141, + 504, + 152 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 504, + 152 + ], + "score": 1.0, + "content": "in the game of Go (Mnih et al., 2015; Silver et al., 2016; 2017). Based on simulation engines like", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 150, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 505, + 164 + ], + "score": 1.0, + "content": "MuJoCo, Todorov et al. (2012), numerous algorithms have been proposed to train agents also in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 162, + 500, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 500, + 174 + ], + "score": 1.0, + "content": "continuous control problems (Schulman et al., 2017; 2015a; Heess et al., 2017; Metz et al., 2017).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 106, + 506, + 174 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 186, + 505, + 319 + ], + "lines": [ + { + "bbox": [ + 106, + 187, + 504, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 504, + 199 + ], + "score": 1.0, + "content": "Structure in RL Most approaches that exploit priors on structure of the problem fall in the do-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "main of hierarchical RL, (Kulkarni et al., 2016; Vezhnevets et al., 2017), which mainly focus on", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "score": 1.0, + "content": "modeling intrinsic motivation of agents. In Hausknecht & Stone (2015), the authors extend the deep", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 220, + 506, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 506, + 233 + ], + "score": 1.0, + "content": "RL algorithms to MDPS with parameterized action space by exploiting the structure of action space", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 231, + 504, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 504, + 243 + ], + "score": 1.0, + "content": "and bounding the action space gradients. Graphs have been used in RL problems prior to our work.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "In Metzen (2013); Mabu et al. (2007); Shoeleh & Asadpour (2017); Mahadevan & Maggioni (2007),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "the authors use graphs to learn a representation of the environment. However, these methods are lim-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "score": 1.0, + "content": "ited to problems with simple dynamical models like for example the task of 2d-navigation, and thus", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 274, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 506, + 288 + ], + "score": 1.0, + "content": "these problems are usually solved via model-based RL. However, for complex multi-joint agents,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 285, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 298 + ], + "score": 1.0, + "content": "learning the dynamical model as well as predicting the transition of states is time consuming and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 297, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 309 + ], + "score": 1.0, + "content": "biased. For problems of training model-free multi-joint agents in complex physical environments,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 307, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 469, + 320 + ], + "score": 1.0, + "content": "relatively little attention has been devoted to modeling the physical structure of the agents.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 187, + 506, + 320 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 333, + 505, + 388 + ], + "lines": [ + { + "bbox": [ + 106, + 332, + 504, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 504, + 345 + ], + "score": 1.0, + "content": "Graph Neural Networks There have been many efforts to generalize neural networks to graph-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 344, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 505, + 356 + ], + "score": 1.0, + "content": "structured data. One line of work is based on convolutional neural networks (CNNs). In (Bruna", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "et al., 2014; Defferrard et al., 2016; Kipf & Welling, 2017), CNNs are employed in the spectral", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "score": 1.0, + "content": "domain relying on the graph Laplacian matrix. (Goller & Kuchler, 1996; Duvenaud et al., 2015)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 376, + 325, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 325, + 390 + ], + "score": 1.0, + "content": "used hash functions in order to apply CNNs to graphs.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 332, + 505, + 390 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 393, + 505, + 558 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 505, + 406 + ], + "score": 1.0, + "content": "Another popular direction is based on recurrent neural networks (RNNs) (Goller & Kuchler, 1996;", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 404, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 505, + 417 + ], + "score": 1.0, + "content": "Gori et al., 2005; Scarselli et al., 2009; Socher et al., 2011; Li et al., 2015; Tai et al., 2015).", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "score": 1.0, + "content": "Among RNN based methods, many are only applicable to special structured graph, e.g., sequences", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "or trees, (Socher et al., 2011; Tai et al., 2015). One class of models which are applicable to general", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "score": 1.0, + "content": "graphs are so-called graph neural networks (GNNs), Scarselli et al. (2009). The inference procedure", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 449, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 505, + 461 + ], + "score": 1.0, + "content": "is a forward pass that exploits a fixed-length propagation process which resembles synchronous mes-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 460, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 472 + ], + "score": 1.0, + "content": "sage passing system in the theory of distributed computing, Attiya & Welch (2004). Nodes in the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "score": 1.0, + "content": "graph have state vectors which are recurrently updated based on their history and received messages.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 482, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 494 + ], + "score": 1.0, + "content": "One of the representative work of GNNs, i.e., gated graph neural networks (GGNNs) by Li et al.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 492, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 505 + ], + "score": 1.0, + "content": "(2015), uses gated recurrent unit to update the state vectors. Learning such a model can be achieved", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 503, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 505, + 515 + ], + "score": 1.0, + "content": "by the back-propagation through time (BPTT) algorithm or recurrent back-propagation, Chauvin &", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 512, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 505, + 528 + ], + "score": 1.0, + "content": "Rumelhart (1995). It has been shown that GNNs, (Li et al., 2015; 2017; Qi et al., 2017; Gilmer et al.,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 524, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 538 + ], + "score": 1.0, + "content": "2017) have a high capacity and achieve state-of-the-art performance in many applications which in-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 535, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 506, + 550 + ], + "score": 1.0, + "content": "volve graph-structured data. In this paper, we model the structure of the reinforcement learning", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 548, + 189, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 189, + 559 + ], + "score": 1.0, + "content": "agents using GNNs.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 393, + 506, + 559 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "Transfer and Multi-task Learning in RL Recently, there has been increased interest in transfer", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 583, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 594 + ], + "score": 1.0, + "content": "learning tasks for RL, Taylor & Stone (2009), which mainly focus on transferring the policy learned", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "from one environment to another. In Rajeswaran et al. (2017b;a), the authors show that agents", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "in reinforcement learning are prone to over-fitting, and that the learned policies generalize poorly", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "across environments. In model-based RL, traditional control has been well studied for generalization", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 626, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 641 + ], + "score": 1.0, + "content": "properties, Coros et al. (2010). Gupta et al. (2017) try to increase the transferability via learning", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "invariant visual features. Efforts have also been made from the meta-learning perspective (Duan", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "et al., 2016; Finn et al., 2017b;a). In Wulfmeier et al. (2017), the authors propose a method of transfer", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "learning by using imitation learning. Transferability comes naturally in our model by exploiting the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 670, + 260, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 260, + 683 + ], + "score": 1.0, + "content": "(shared) graph structure of the agents.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 572, + 505, + 683 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "Multi-task learning has also received a lot of attention, Wilson et al. (2007). In Teh et al. (2017),", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "the authors use a distilled policy that captures common behaviour across tasks. Wilson et al. (2007);", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "Oh et al. (2017); Andreas et al. (2016) use a hierarchical approach, where multiple sub-policies are", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 721, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 506, + 732 + ], + "score": 1.0, + "content": "learnt. In Yang et al. (2017); Parisotto et al. (2015), the authors exploited shared visual features", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "to tackle multi-task learning. In Ammar et al. (2014), a multi-task policy gradient is proposed,", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "score": 1.0, + "content": "while, Calandriello et al. (2014) propose multi-task extensions of the fitted Q-iteration algorithm.", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "Successor features in Barreto et al. (2017) are used to boost the performance of multiple tasks.", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 439, + 447, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 447, + 451 + ], + "score": 1.0, + "content": "Unlike ours, these methods do not exploit the physical graph structures of the agents.", + "type": "text", + "cross_page": true + } + ], + "index": 8 + } + ], + "index": 50.5, + "bbox_fs": [ + 106, + 688, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 113, + 84, + 495, + 351 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 84, + 495, + 351 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 84, + 495, + 351 + ], + "spans": [ + { + "bbox": [ + 113, + 84, + 495, + 351 + ], + "score": 0.977, + "type": "image", + "image_path": "31914ef34c44e5cb9c9bdb09331fb9917e4b0134cd1f05a2719e5c2b4c19287a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 113, + 84, + 495, + 173.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 113, + 173.0, + 495, + 262.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 113, + 262.0, + 495, + 351.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 361, + 504, + 384 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "Figure 3: Results of MLP, TreeNet and NerveNet on 8 continuous control benchmarks from the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 371, + 134, + 387 + ], + "spans": [ + { + "bbox": [ + 104, + 371, + 134, + 387 + ], + "score": 1.0, + "content": "Gym.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 505, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "to tackle multi-task learning. In Ammar et al. (2014), a multi-task policy gradient is proposed,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "score": 1.0, + "content": "while, Calandriello et al. (2014) propose multi-task extensions of the fitted Q-iteration algorithm.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "Successor features in Barreto et al. (2017) are used to boost the performance of multiple tasks.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 439, + 447, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 447, + 451 + ], + "score": 1.0, + "content": "Unlike ours, these methods do not exploit the physical graph structures of the agents.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 108, + 468, + 200, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 466, + 201, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 201, + 482 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 493, + 505, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 505 + ], + "score": 1.0, + "content": "In this section, we first verify the effectiveness of NerveNet on standard MuJoCo environments", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 504, + 504, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 504, + 516 + ], + "score": 1.0, + "content": "in OpenAI Gym. We then investigate the transfer abilities of NerveNet and other competitors by", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 515, + 501, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 501, + 528 + ], + "score": 1.0, + "content": "customizing some of those environments, as well as the multi-task learning ability and robustness.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 106, + 540, + 374, + 551 + ], + "lines": [ + { + "bbox": [ + 106, + 540, + 375, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 375, + 553 + ], + "score": 1.0, + "content": "4.1 COMPARISON ON STANDARD BENCHMARKS OF MUJOCO", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "Baselines We compare NerveNet with the standard MLP models utilized by Schulman et al. (2017)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 571, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 506, + 585 + ], + "score": 1.0, + "content": "and another baseline which is constructed as follows. We first remove the physical graph structure", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "and introduce an additional super node which connects to all nodes in the graph. This results in a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "singly rooted depth-1 tree. We refer to this baseline as TreeNet. The propagation model of TreeNet", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 604, + 504, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 504, + 617 + ], + "score": 1.0, + "content": "is similar to NerveNet where, however, the policy first aggregates the information from all children", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "and then feeds the state vector of the root to the output model. This simpler model serves as a", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 627, + 331, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 331, + 640 + ], + "score": 1.0, + "content": "baseline to verify the importance of the graph structure.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 655 + ], + "score": 1.0, + "content": "We run experiments on 8 simulated continuous control benchmarks from the Gym, Brock-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "man et al. (2016), which is based on MuJoCo, Todorov et al. (2012). In particular, we use", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "Reacher, InvertedPendulum, InvertedDoublePendulum, Swimmer, and four walk-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "ing or running tasks: HalfCheetah, Hopper, Walker2d, Ant. We set the maximum", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "number of training steps to be 1 million for all environments as it is enough to solve them.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Note that for InvertedPendulum, different from the original one in Gym, we add the dis-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "tance penalty of the cart and velocity penalty so that the reward is more consistent to the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 719, + 488, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 488, + 734 + ], + "score": 1.0, + "content": "InvertedDoublePendulum. This change of design also makes the task more challenging.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 113, + 84, + 495, + 351 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 84, + 495, + 351 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 84, + 495, + 351 + ], + "spans": [ + { + "bbox": [ + 113, + 84, + 495, + 351 + ], + "score": 0.977, + "type": "image", + "image_path": "31914ef34c44e5cb9c9bdb09331fb9917e4b0134cd1f05a2719e5c2b4c19287a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 113, + 84, + 495, + 173.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 113, + 173.0, + 495, + 262.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 113, + 262.0, + 495, + 351.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 361, + 504, + 384 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "Figure 3: Results of MLP, TreeNet and NerveNet on 8 continuous control benchmarks from the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 371, + 134, + 387 + ], + "spans": [ + { + "bbox": [ + 104, + 371, + 134, + 387 + ], + "score": 1.0, + "content": "Gym.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 505, + 451 + ], + "lines": [], + "index": 6.5, + "bbox_fs": [ + 105, + 407, + 505, + 451 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 468, + 200, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 466, + 201, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 201, + 482 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 493, + 505, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 505 + ], + "score": 1.0, + "content": "In this section, we first verify the effectiveness of NerveNet on standard MuJoCo environments", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 504, + 504, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 504, + 516 + ], + "score": 1.0, + "content": "in OpenAI Gym. We then investigate the transfer abilities of NerveNet and other competitors by", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 515, + 501, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 501, + 528 + ], + "score": 1.0, + "content": "customizing some of those environments, as well as the multi-task learning ability and robustness.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 493, + 505, + 528 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 540, + 374, + 551 + ], + "lines": [ + { + "bbox": [ + 106, + 540, + 375, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 375, + 553 + ], + "score": 1.0, + "content": "4.1 COMPARISON ON STANDARD BENCHMARKS OF MUJOCO", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "Baselines We compare NerveNet with the standard MLP models utilized by Schulman et al. (2017)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 571, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 506, + 585 + ], + "score": 1.0, + "content": "and another baseline which is constructed as follows. We first remove the physical graph structure", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "and introduce an additional super node which connects to all nodes in the graph. This results in a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "singly rooted depth-1 tree. We refer to this baseline as TreeNet. The propagation model of TreeNet", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 604, + 504, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 504, + 617 + ], + "score": 1.0, + "content": "is similar to NerveNet where, however, the policy first aggregates the information from all children", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "and then feeds the state vector of the root to the output model. This simpler model serves as a", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 627, + 331, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 331, + 640 + ], + "score": 1.0, + "content": "baseline to verify the importance of the graph structure.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 561, + 506, + 640 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 655 + ], + "score": 1.0, + "content": "We run experiments on 8 simulated continuous control benchmarks from the Gym, Brock-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "man et al. (2016), which is based on MuJoCo, Todorov et al. (2012). In particular, we use", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "Reacher, InvertedPendulum, InvertedDoublePendulum, Swimmer, and four walk-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "ing or running tasks: HalfCheetah, Hopper, Walker2d, Ant. We set the maximum", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "number of training steps to be 1 million for all environments as it is enough to solve them.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Note that for InvertedPendulum, different from the original one in Gym, we add the dis-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "tance penalty of the cart and velocity penalty so that the reward is more consistent to the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 719, + 488, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 488, + 734 + ], + "score": 1.0, + "content": "InvertedDoublePendulum. This change of design also makes the task more challenging.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 643, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 113, + 101, + 498, + 158 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 111, + 80, + 493, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 78, + 495, + 93 + ], + "spans": [ + { + "bbox": [ + 115, + 78, + 495, + 93 + ], + "score": 1.0, + "content": "Table 1: Performance of the Pre-trained models on CentipedeFour and CentipedeSix.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 113, + 101, + 498, + 158 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 101, + 498, + 158 + ], + "spans": [ + { + "bbox": [ + 113, + 101, + 498, + 158 + ], + "score": 0.976, + "html": "
C_FourReward AvgStdMaxC_SixReward AvgStdMax
NerveNet2799.91247.23833.9NerveNet2507.1738.42979.2
MLP2398.51390.43936.3MLP2793.01005.23529.5
TreeNet1429.6957.73021.7TreeNet2229.71109.43727.4
", + "type": "table", + "image_path": "670ab56a6ac4211ba1a235ac8454c4b3f3398ff13994de22b67b272be84cc5d7.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 113, + 101, + 498, + 120.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 113, + 120.0, + 498, + 139.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 113, + 139.0, + 498, + 158.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 106, + 178, + 505, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 177, + 504, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 504, + 192 + ], + "score": 1.0, + "content": "Results We do grid search to find the best hyperparameters and leave the details in the Ap-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 190, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 202 + ], + "score": 1.0, + "content": "pendix 6.3. As the randomness might have a big impact on the performance, for each environment,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 201, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 505, + 212 + ], + "score": 1.0, + "content": "we run 3 experiments with different random seeds and plot the average curves and the standard de-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 212, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 505, + 224 + ], + "score": 1.0, + "content": "viations. We show the results in Figure 3. From the figures, we can see that MLP with the same", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "setup as in Schulman et al. (2017) works the best in most of tasks.1 NerveNet basically matches the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 232, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 505, + 247 + ], + "score": 1.0, + "content": "performance of MLP in terms of sample efficiency as well as the performance after it converges.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 245, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 257 + ], + "score": 1.0, + "content": "In most cases, the TreeNet is worse than NerveNet which highlights the importance of keeping the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 256, + 208, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 256, + 208, + 268 + ], + "score": 1.0, + "content": "physical graph structure.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7.5 + }, + { + "type": "title", + "bbox": [ + 108, + 281, + 281, + 292 + ], + "lines": [ + { + "bbox": [ + 105, + 280, + 282, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 282, + 293 + ], + "score": 1.0, + "content": "4.2 STRUCTURE TRANSFER LEARNING", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 301, + 505, + 411 + ], + "lines": [ + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "score": 1.0, + "content": "We now benchmark our model in the task of structure transfer learning by creating customized", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 312, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 506, + 325 + ], + "score": 1.0, + "content": "environments based on the existing ones from MuJoCo. We mainly investigate two types of structure", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 324, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 335 + ], + "score": 1.0, + "content": "transfer learning tasks. The first one is to train a model with an agent of small size (small graph) and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "apply the learned model to an agent with a larger size, i.e., size transfer. When increasing the size", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 344, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 359 + ], + "score": 1.0, + "content": "of the agent, observation and action space also increase which makes learning more challenging.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 356, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 369 + ], + "score": 1.0, + "content": "Another type of structure transfer learning is disability transfer where we first learn a model for", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "score": 1.0, + "content": "the original agent and then apply it to the same agent with some components disabled. If one", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 379, + 504, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 504, + 390 + ], + "score": 1.0, + "content": "model overfits the environment, disabling some components of the agent might bring catastrophic", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 388, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 505, + 403 + ], + "score": 1.0, + "content": "performance degradation. Note that for both transfer tasks, all factors of environments do not change", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 400, + 239, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 239, + 413 + ], + "score": 1.0, + "content": "except the structure of the agent.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 424, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 106, + 424, + 504, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 504, + 435 + ], + "score": 1.0, + "content": "Centipede We create the first environment in which the agent has a similar structure to a centipede.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 342, + 448 + ], + "score": 1.0, + "content": "The goal of the agent is to run as fast as possible along the", + "type": "text" + }, + { + "bbox": [ + 342, + 437, + 348, + 446 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "-direction in the MuJoCo environment.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "The agent consists of repetitive torso bodies where each one has two legs attached. For two con-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "secutive bodies, we add two actuators which control the rotation between them. Furthermore, each", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 468, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 481 + ], + "score": 1.0, + "content": "leg consists of a thigh and shin, which are controlled by two hinge actuators. By linking copies", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "of torso bodies and corresponding legs, we create agents with different lengths. Specifically, the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "shortest Centipede is CentipedeFour and the longest one is CentipedeFourty due to the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "score": 1.0, + "content": "limit of supported resource of MuJoCo. For each time step, the total reward is the speed reward", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 512, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 506, + 525 + ], + "score": 1.0, + "content": "minus the energy cost and force feedback from the ground. Note that in practice, we found that", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 522, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 537 + ], + "score": 1.0, + "content": "training a CentipedeEight from scratch is already very difficult. For size transfer experiments,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "we create many instances which are listed in Figure 4, like “4to06”, “6to10”. For disability trans-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "score": 1.0, + "content": "fer, we create CrippleCentipede agents of which two back legs are disabled. In Figure 4,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 556, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 287, + 569 + ], + "score": 1.0, + "content": "CrippleCentipede is specified as “Cp”.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 108, + 580, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "Snakes We also create a snake-like agent which is common in robotics, Crespi & Ijspeert (2008).", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 590, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 505, + 603 + ], + "score": 1.0, + "content": "We design the Snake environment based on the Swimmer model in Gym. The goal of the agent is", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 601, + 483, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 483, + 614 + ], + "score": 1.0, + "content": "to move as fast as possible. For details of the environment, please see the schematic figure 16.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "title", + "bbox": [ + 108, + 626, + 253, + 637 + ], + "lines": [ + { + "bbox": [ + 106, + 625, + 254, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 254, + 638 + ], + "score": 1.0, + "content": "4.2.1 EXPERIMENTAL SETTINGS", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 645, + 504, + 667 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "To fully investigate the performance of NerveNet, we build several baseline models for structure", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 656, + 285, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 285, + 668 + ], + "score": 1.0, + "content": "transfer learning which are explained below.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 108, + 680, + 503, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 679, + 504, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 504, + 693 + ], + "score": 1.0, + "content": "NerveNet For the NerveNet, since all the weights are exactly the same for the small and the large-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 690, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 505, + 704 + ], + "score": 1.0, + "content": "agent models, we directly use the old weights trained on the small-agent model. When the large", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 712, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 119, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "1By applying the adaptive learning rate schedule from Schulman et al. (2017), we obtained better perfor-", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 720, + 280, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 280, + 733 + ], + "score": 1.0, + "content": "mances than the ones reported in original paper.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 113, + 101, + 498, + 158 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 111, + 80, + 493, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 78, + 495, + 93 + ], + "spans": [ + { + "bbox": [ + 115, + 78, + 495, + 93 + ], + "score": 1.0, + "content": "Table 1: Performance of the Pre-trained models on CentipedeFour and CentipedeSix.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 113, + 101, + 498, + 158 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 101, + 498, + 158 + ], + "spans": [ + { + "bbox": [ + 113, + 101, + 498, + 158 + ], + "score": 0.976, + "html": "
C_FourReward AvgStdMaxC_SixReward AvgStdMax
NerveNet2799.91247.23833.9NerveNet2507.1738.42979.2
MLP2398.51390.43936.3MLP2793.01005.23529.5
TreeNet1429.6957.73021.7TreeNet2229.71109.43727.4
", + "type": "table", + "image_path": "670ab56a6ac4211ba1a235ac8454c4b3f3398ff13994de22b67b272be84cc5d7.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 113, + 101, + 498, + 120.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 113, + 120.0, + 498, + 139.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 113, + 139.0, + 498, + 158.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 106, + 178, + 505, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 177, + 504, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 504, + 192 + ], + "score": 1.0, + "content": "Results We do grid search to find the best hyperparameters and leave the details in the Ap-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 190, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 202 + ], + "score": 1.0, + "content": "pendix 6.3. As the randomness might have a big impact on the performance, for each environment,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 201, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 505, + 212 + ], + "score": 1.0, + "content": "we run 3 experiments with different random seeds and plot the average curves and the standard de-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 212, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 505, + 224 + ], + "score": 1.0, + "content": "viations. We show the results in Figure 3. From the figures, we can see that MLP with the same", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "setup as in Schulman et al. (2017) works the best in most of tasks.1 NerveNet basically matches the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 232, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 505, + 247 + ], + "score": 1.0, + "content": "performance of MLP in terms of sample efficiency as well as the performance after it converges.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 245, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 257 + ], + "score": 1.0, + "content": "In most cases, the TreeNet is worse than NerveNet which highlights the importance of keeping the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 256, + 208, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 256, + 208, + 268 + ], + "score": 1.0, + "content": "physical graph structure.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 177, + 505, + 268 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 281, + 281, + 292 + ], + "lines": [ + { + "bbox": [ + 105, + 280, + 282, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 282, + 293 + ], + "score": 1.0, + "content": "4.2 STRUCTURE TRANSFER LEARNING", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 301, + 505, + 411 + ], + "lines": [ + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "score": 1.0, + "content": "We now benchmark our model in the task of structure transfer learning by creating customized", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 312, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 506, + 325 + ], + "score": 1.0, + "content": "environments based on the existing ones from MuJoCo. We mainly investigate two types of structure", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 324, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 335 + ], + "score": 1.0, + "content": "transfer learning tasks. The first one is to train a model with an agent of small size (small graph) and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "apply the learned model to an agent with a larger size, i.e., size transfer. When increasing the size", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 344, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 359 + ], + "score": 1.0, + "content": "of the agent, observation and action space also increase which makes learning more challenging.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 356, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 369 + ], + "score": 1.0, + "content": "Another type of structure transfer learning is disability transfer where we first learn a model for", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "score": 1.0, + "content": "the original agent and then apply it to the same agent with some components disabled. If one", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 379, + 504, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 504, + 390 + ], + "score": 1.0, + "content": "model overfits the environment, disabling some components of the agent might bring catastrophic", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 388, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 505, + 403 + ], + "score": 1.0, + "content": "performance degradation. Note that for both transfer tasks, all factors of environments do not change", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 400, + 239, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 239, + 413 + ], + "score": 1.0, + "content": "except the structure of the agent.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 301, + 506, + 413 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 424, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 106, + 424, + 504, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 504, + 435 + ], + "score": 1.0, + "content": "Centipede We create the first environment in which the agent has a similar structure to a centipede.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 342, + 448 + ], + "score": 1.0, + "content": "The goal of the agent is to run as fast as possible along the", + "type": "text" + }, + { + "bbox": [ + 342, + 437, + 348, + 446 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "-direction in the MuJoCo environment.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "The agent consists of repetitive torso bodies where each one has two legs attached. For two con-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "secutive bodies, we add two actuators which control the rotation between them. Furthermore, each", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 468, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 481 + ], + "score": 1.0, + "content": "leg consists of a thigh and shin, which are controlled by two hinge actuators. By linking copies", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "of torso bodies and corresponding legs, we create agents with different lengths. Specifically, the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "shortest Centipede is CentipedeFour and the longest one is CentipedeFourty due to the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "score": 1.0, + "content": "limit of supported resource of MuJoCo. For each time step, the total reward is the speed reward", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 512, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 506, + 525 + ], + "score": 1.0, + "content": "minus the energy cost and force feedback from the ground. Note that in practice, we found that", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 522, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 537 + ], + "score": 1.0, + "content": "training a CentipedeEight from scratch is already very difficult. For size transfer experiments,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "we create many instances which are listed in Figure 4, like “4to06”, “6to10”. For disability trans-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "score": 1.0, + "content": "fer, we create CrippleCentipede agents of which two back legs are disabled. In Figure 4,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 556, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 287, + 569 + ], + "score": 1.0, + "content": "CrippleCentipede is specified as “Cp”.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 424, + 506, + 569 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 580, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "Snakes We also create a snake-like agent which is common in robotics, Crespi & Ijspeert (2008).", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 590, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 505, + 603 + ], + "score": 1.0, + "content": "We design the Snake environment based on the Swimmer model in Gym. The goal of the agent is", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 601, + 483, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 483, + 614 + ], + "score": 1.0, + "content": "to move as fast as possible. For details of the environment, please see the schematic figure 16.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 106, + 579, + 505, + 614 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 626, + 253, + 637 + ], + "lines": [ + { + "bbox": [ + 106, + 625, + 254, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 254, + 638 + ], + "score": 1.0, + "content": "4.2.1 EXPERIMENTAL SETTINGS", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 645, + 504, + 667 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "To fully investigate the performance of NerveNet, we build several baseline models for structure", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 656, + 285, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 285, + 668 + ], + "score": 1.0, + "content": "transfer learning which are explained below.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 644, + 506, + 668 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 680, + 503, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 679, + 504, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 504, + 693 + ], + "score": 1.0, + "content": "NerveNet For the NerveNet, since all the weights are exactly the same for the small and the large-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 690, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 505, + 704 + ], + "score": 1.0, + "content": "agent models, we directly use the old weights trained on the small-agent model. When the large", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 669, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 506, + 685 + ], + "score": 1.0, + "content": "agent has repetitive structure, we further re-use the weights of the corresponding joints from the", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 682, + 185, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 185, + 694 + ], + "score": 1.0, + "content": "small-agent model.", + "type": "text", + "cross_page": true + } + ], + "index": 11 + } + ], + "index": 42.5, + "bbox_fs": [ + 106, + 679, + 505, + 704 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 80, + 511, + 275 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 80, + 511, + 275 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 80, + 511, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 511, + 275 + ], + "score": 0.345, + "type": "image", + "image_path": "ded2354d5c50c0404c254caa82cabadd97bd1f063a74fb66cd5313b24ebd4bd0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 80, + 511, + 145.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 145.0, + 511, + 210.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 210.0, + 511, + 275.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 282, + 506, + 306 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "Figure 4: Performance of zero-shot learning on centipedes. For each task, we run the policy for 100", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 293, + 484, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 484, + 306 + ], + "score": 1.0, + "content": "episodes and record the average reward and average length the agent runs before falling down.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 114, + 328, + 485, + 615 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 328, + 485, + 615 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 114, + 328, + 485, + 615 + ], + "spans": [ + { + "bbox": [ + 114, + 328, + 485, + 615 + ], + "score": 0.976, + "type": "image", + "image_path": "1b4f2a5fa06b4d78a1b4be689e625ce9271bf372ee76f157e6a19ba72f40e380.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 114, + 328, + 485, + 423.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 114, + 423.6666666666667, + 485, + 519.3333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 114, + 519.3333333333334, + 485, + 615.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 623, + 504, + 645 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 622, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 637 + ], + "score": 1.0, + "content": "Figure 5: (a), (b): Results of fine-tuning for size transfer experiments. (c), (d) Results of fine-tuning", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 635, + 248, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 248, + 646 + ], + "score": 1.0, + "content": "for disability transfer experiments.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "text", + "bbox": [ + 106, + 670, + 504, + 693 + ], + "lines": [ + { + "bbox": [ + 104, + 669, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 506, + 685 + ], + "score": 1.0, + "content": "agent has repetitive structure, we further re-use the weights of the corresponding joints from the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 682, + 185, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 185, + 694 + ], + "score": 1.0, + "content": "small-agent model.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "MLP Pre-trained (MLPP) For the MLP based model, while transferring from one structure to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 721, + 504, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 504, + 732 + ], + "score": 1.0, + "content": "another, the size of the input layer changes since the size of the observation changes. One straight-", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 80, + 511, + 275 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 80, + 511, + 275 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 80, + 511, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 511, + 275 + ], + "score": 0.345, + "type": "image", + "image_path": "ded2354d5c50c0404c254caa82cabadd97bd1f063a74fb66cd5313b24ebd4bd0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 80, + 511, + 145.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 145.0, + 511, + 210.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 210.0, + 511, + 275.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 282, + 506, + 306 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "Figure 4: Performance of zero-shot learning on centipedes. For each task, we run the policy for 100", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 293, + 484, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 484, + 306 + ], + "score": 1.0, + "content": "episodes and record the average reward and average length the agent runs before falling down.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 114, + 328, + 485, + 615 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 328, + 485, + 615 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 114, + 328, + 485, + 615 + ], + "spans": [ + { + "bbox": [ + 114, + 328, + 485, + 615 + ], + "score": 0.976, + "type": "image", + "image_path": "1b4f2a5fa06b4d78a1b4be689e625ce9271bf372ee76f157e6a19ba72f40e380.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 114, + 328, + 485, + 423.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 114, + 423.6666666666667, + 485, + 519.3333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 114, + 519.3333333333334, + 485, + 615.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 623, + 504, + 645 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 622, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 637 + ], + "score": 1.0, + "content": "Figure 5: (a), (b): Results of fine-tuning for size transfer experiments. (c), (d) Results of fine-tuning", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 635, + 248, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 248, + 646 + ], + "score": 1.0, + "content": "for disability transfer experiments.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "text", + "bbox": [ + 106, + 670, + 504, + 693 + ], + "lines": [], + "index": 10.5, + "bbox_fs": [ + 104, + 669, + 506, + 694 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "MLP Pre-trained (MLPP) For the MLP based model, while transferring from one structure to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 721, + 504, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 504, + 732 + ], + "score": 1.0, + "content": "another, the size of the input layer changes since the size of the observation changes. One straight-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 559, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 576 + ], + "score": 1.0, + "content": "forward idea is to reuse the weights from the first hidden layer to the output layer and randomly", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 572, + 284, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 284, + 586 + ], + "score": 1.0, + "content": "initialize the weights of the new input layer.", + "type": "text", + "cross_page": true + } + ], + "index": 10 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 82, + 466, + 172 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 145, + 82, + 466, + 172 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 145, + 82, + 466, + 172 + ], + "spans": [ + { + "bbox": [ + 145, + 82, + 466, + 172 + ], + "score": 0.933, + "type": "image", + "image_path": "84ea5296a5d035d9829cab325bf3532db657dcf4811e266ee0b813b65808af7f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 145, + 82, + 466, + 112.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 145, + 112.0, + 466, + 142.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 145, + 142.0, + 466, + 172.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 179, + 505, + 201 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 178, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 505, + 191 + ], + "score": 1.0, + "content": "Figure 6: Results on zero-shot transfer learning on snake agents. Each tasks are simulated for 100", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 189, + 146, + 203 + ], + "spans": [ + { + "bbox": [ + 104, + 189, + 146, + 203 + ], + "score": 1.0, + "content": "episodes.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 114, + 225, + 493, + 517 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 225, + 493, + 517 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 114, + 225, + 493, + 517 + ], + "spans": [ + { + "bbox": [ + 114, + 225, + 493, + 517 + ], + "score": 0.977, + "type": "image", + "image_path": "75955e4c44eb72c36a84acb11321fa2796dcfc7110c692596195fe9c0701f3c7.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 114, + 225, + 493, + 322.3333333333333 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 114, + 322.3333333333333, + 493, + 419.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 114, + 419.66666666666663, + 493, + 517.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 195, + 525, + 415, + 537 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 195, + 524, + 416, + 538 + ], + "spans": [ + { + "bbox": [ + 195, + 524, + 416, + 538 + ], + "score": 1.0, + "content": "Figure 7: Results of finetuning on snake environments.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 7.0 + }, + { + "type": "text", + "bbox": [ + 108, + 561, + 504, + 584 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 576 + ], + "score": 1.0, + "content": "forward idea is to reuse the weights from the first hidden layer to the output layer and randomly", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 572, + 284, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 284, + 586 + ], + "score": 1.0, + "content": "initialize the weights of the new input layer.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 599, + 505, + 667 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 504, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 504, + 612 + ], + "score": 1.0, + "content": "MLP Activation Assigning (MLPAA) Another way of making MLP transferable is assigning the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "weights of the small-agent model to the corresponding partial weights of the large-agent model and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "setting the remaining weights to be zero. Note that we do not add or remove any layers from the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "small-agent model to the large-agent except for changing the size of the layers. By doing so, we can", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 642, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 506, + 658 + ], + "score": 1.0, + "content": "keep the output of the large-agent model to be same as the small-agent in the beginning, i.e., keeping", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 655, + 199, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 199, + 668 + ], + "score": 1.0, + "content": "the same initial policy.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 682, + 503, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 680, + 505, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 505, + 696 + ], + "score": 1.0, + "content": "TreeNet TreeNet is similar as the model described before. We apply the same way of assigning", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 692, + 351, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 351, + 706 + ], + "score": 1.0, + "content": "weights as MLPAA to TreeNet for the transfer learning task.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 105, + 720, + 496, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 718, + 498, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 498, + 734 + ], + "score": 1.0, + "content": "Random We also include the random policy which is uniformly sampled from the action space.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 82, + 466, + 172 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 145, + 82, + 466, + 172 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 145, + 82, + 466, + 172 + ], + "spans": [ + { + "bbox": [ + 145, + 82, + 466, + 172 + ], + "score": 0.933, + "type": "image", + "image_path": "84ea5296a5d035d9829cab325bf3532db657dcf4811e266ee0b813b65808af7f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 145, + 82, + 466, + 112.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 145, + 112.0, + 466, + 142.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 145, + 142.0, + 466, + 172.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 179, + 505, + 201 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 178, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 505, + 191 + ], + "score": 1.0, + "content": "Figure 6: Results on zero-shot transfer learning on snake agents. Each tasks are simulated for 100", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 189, + 146, + 203 + ], + "spans": [ + { + "bbox": [ + 104, + 189, + 146, + 203 + ], + "score": 1.0, + "content": "episodes.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 114, + 225, + 493, + 517 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 225, + 493, + 517 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 114, + 225, + 493, + 517 + ], + "spans": [ + { + "bbox": [ + 114, + 225, + 493, + 517 + ], + "score": 0.977, + "type": "image", + "image_path": "75955e4c44eb72c36a84acb11321fa2796dcfc7110c692596195fe9c0701f3c7.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 114, + 225, + 493, + 322.3333333333333 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 114, + 322.3333333333333, + 493, + 419.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 114, + 419.66666666666663, + 493, + 517.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 195, + 525, + 415, + 537 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 195, + 524, + 416, + 538 + ], + "spans": [ + { + "bbox": [ + 195, + 524, + 416, + 538 + ], + "score": 1.0, + "content": "Figure 7: Results of finetuning on snake environments.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 7.0 + }, + { + "type": "text", + "bbox": [ + 108, + 561, + 504, + 584 + ], + "lines": [], + "index": 9.5, + "bbox_fs": [ + 105, + 559, + 505, + 586 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 599, + 505, + 667 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 504, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 504, + 612 + ], + "score": 1.0, + "content": "MLP Activation Assigning (MLPAA) Another way of making MLP transferable is assigning the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "weights of the small-agent model to the corresponding partial weights of the large-agent model and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "setting the remaining weights to be zero. Note that we do not add or remove any layers from the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "small-agent model to the large-agent except for changing the size of the layers. By doing so, we can", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 642, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 506, + 658 + ], + "score": 1.0, + "content": "keep the output of the large-agent model to be same as the small-agent in the beginning, i.e., keeping", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 655, + 199, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 199, + 668 + ], + "score": 1.0, + "content": "the same initial policy.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 599, + 506, + 668 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 682, + 503, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 680, + 505, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 505, + 696 + ], + "score": 1.0, + "content": "TreeNet TreeNet is similar as the model described before. We apply the same way of assigning", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 692, + 351, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 351, + 706 + ], + "score": 1.0, + "content": "weights as MLPAA to TreeNet for the transfer learning task.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 680, + 505, + 706 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 720, + 496, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 718, + 498, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 498, + 734 + ], + "score": 1.0, + "content": "Random We also include the random policy which is uniformly sampled from the action space.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 718, + 498, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 179, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 180, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 180, + 95 + ], + "score": 1.0, + "content": "4.2.2 RESULTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 103, + 505, + 169 + ], + "lines": [ + { + "bbox": [ + 106, + 103, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 505, + 116 + ], + "score": 1.0, + "content": "Centipedes For the Centipedes environment, we first run experiments of all models on", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "score": 1.0, + "content": "CentipedeSix and CentipedeFour to get the pre-trained models for transfer learning. We", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 506, + 138 + ], + "score": 1.0, + "content": "train different models until these agents run equally well as possible, which is reported in Table 1.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 135, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 505, + 150 + ], + "score": 1.0, + "content": "Note that, in practice, we train TreeNet on CentipedeFour for more than 8 million time steps.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 147, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 505, + 160 + ], + "score": 1.0, + "content": "However, due to the difficulty of optimizing TreeNet on CentipedeFour, the performance is still", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 159, + 408, + 170 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 408, + 170 + ], + "score": 1.0, + "content": "lower. But visually, the TreeNet agent is able to run in CentipedeFour.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 175, + 505, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 175, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 175, + 505, + 187 + ], + "score": 1.0, + "content": "We then examine the zero-shot performance where zero-shot means directly applying the model", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 186, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 198 + ], + "score": 1.0, + "content": "trained with one setting to the other without any fine-tuning. To better visualize the results, we", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "score": 1.0, + "content": "linearly normalize the performance to get a performance score, and color the results accordingly.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 209, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 219 + ], + "score": 1.0, + "content": "The normalization scheme is recorded in Appendix 11. The performance score is less than 1, and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 220, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 231 + ], + "score": 1.0, + "content": "is shown in the parentheses behind the original results. As we can see from Figure 4 (full chart in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 229, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 505, + 242 + ], + "score": 1.0, + "content": "Appendix 6.5), NerveNet outperforms all competitors on all settings, except in the 4toCp06 scenario.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 240, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 505, + 253 + ], + "score": 1.0, + "content": "Note that transferring from CentipedeFour is more difficult than from CentipedeSix since", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 251, + 491, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 491, + 264 + ], + "score": 1.0, + "content": "the situation where one torso connects to two neighboring torsos only happens beyond 4 bodies.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 108, + 268, + 505, + 313 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "TreeNet has a surprisingly good performance on tasks from CentipedeFour. However, by check-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 280, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 505, + 291 + ], + "score": 1.0, + "content": "ing the videos, the learned agent is actually not able to “move” as good as other methods. The high", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "reward is mainly due to the fact that TreeNet policy is better at standing still and gaining alive bonus.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 300, + 464, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 464, + 315 + ], + "score": 1.0, + "content": "We argue that the average running-length in each episode is also a very important metric.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 318, + 505, + 362 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "By including the results of running-length, we notice that NerveNet is the only model able to walk in", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 330, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 505, + 341 + ], + "score": 1.0, + "content": "the zero-shot setting. In fact, the performance of NerveNet is orders-of-magnitude better, and most", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 340, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 505, + 352 + ], + "score": 1.0, + "content": "of the time, agents from other methods cannot even move forward. We also notice that if transferred", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 352, + 504, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 504, + 363 + ], + "score": 1.0, + "content": "from CentipedeSix, NerveNet is able to provide walkable pre-trained models on all new agents.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 368, + 505, + 445 + ], + "lines": [ + { + "bbox": [ + 106, + 367, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 505, + 381 + ], + "score": 1.0, + "content": "We fine-tune for both size transfer and disability transfer experiments and show the training curves", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 379, + 504, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 504, + 392 + ], + "score": 1.0, + "content": "in Figure 5. From the figure, we can see that by using the pre-trained model, NerveNet significantly", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "decreases the number of episodes required to reach the level of reward which is considered as solved.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 401, + 504, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 504, + 414 + ], + "score": 1.0, + "content": "By looking at the videos, we notice that the bottleneck of learning for the agent is “how to stand”.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 411, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 425 + ], + "score": 1.0, + "content": "When training from scratch, it can be seen that almost 0.5 million time steps are spent on a very flat", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 423, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 436 + ], + "score": 1.0, + "content": "reward surface. Therefore, the MLPAA agents, which copy the learned policy, are able to stand and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 434, + 431, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 431, + 447 + ], + "score": 1.0, + "content": "bypass this time-consuming process and reach to a good performance in the end.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 451, + 504, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 506, + 463 + ], + "score": 1.0, + "content": "Moreover, by examining the result videos, we noticed that the “walk-cycle” behavior is observed", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "for NerveNet but is not common for others. Walk-cycle are adopted for many insects in the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 473, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 505, + 485 + ], + "score": 1.0, + "content": "world, Biewener (2003). For example, six-legged ants use a tripedal gait, where the legs are used", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "score": 1.0, + "content": "in two separate triangles alternatively touching the ground. We give more details of walk-cycle in", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 494, + 155, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 155, + 505 + ], + "score": 1.0, + "content": "Section 4.5.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 511, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 511, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 506, + 524 + ], + "score": 1.0, + "content": "One possible reason is that the agent of MLP based method (MLPAA, MLPP) learns a policy that", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "does not utilize all legs. From CentipedeEight and up, we do not observe any MLP agents to be", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "score": 1.0, + "content": "able to coordinate all legs whereas almost all policies learned by NerveNet use all legs. Therefore,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 545, + 466, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 466, + 557 + ], + "score": 1.0, + "content": "NerveNet is better at utilizing structure information and not over-fitting the environments.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 106, + 572, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 571, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 506, + 585 + ], + "score": 1.0, + "content": "Snakes The zero-shot performance for snakes is summarized in Figure 6. As we can see, NerveNet", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "score": 1.0, + "content": "has the best performance on all transfer learning tasks. In most cases, NerveNet has a starting", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "reward value of more than 300, which is a pretty good policy since 350 is considered as solved for", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "snakeThree. By looking at the videos, we found that agents of other competitors are not able to", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 615, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 104, + 615, + 506, + 630 + ], + "score": 1.0, + "content": "control the new actuators in the zero-shot setting. They either overfit to the original models, where", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "the policy is completely useless in the new setting (e.g., the MLPAA is worse than random policy in", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 639, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 650 + ], + "score": 1.0, + "content": "SnakeThree2SnakeFour), or the new actuators are not able to coordinate with the old actuators", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "trained before. While for NerveNet, the actuators are able to coordinate to its neighbors, regardless", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 660, + 259, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 259, + 672 + ], + "score": 1.0, + "content": "of whether they are new to the agents.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 108, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "We also summarize the training curves of fine-tuning in Fig. 7. We can observe that NerveNet has a", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "very good initialization with the pre-trained model, and the performance increases with fine-tuning.", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 107, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 107, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "When training from scratch, NerveNet is less sample efficient compared to the MLP model which", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 708, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 724 + ], + "score": 1.0, + "content": "might be caused by the fact that optimizing our model is more challenging than MLP. Fine-tuning", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "helps to improve the sample efficiency of our model by a large margin. At the same time, although", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 50 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 179, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 180, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 180, + 95 + ], + "score": 1.0, + "content": "4.2.2 RESULTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 103, + 505, + 169 + ], + "lines": [ + { + "bbox": [ + 106, + 103, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 505, + 116 + ], + "score": 1.0, + "content": "Centipedes For the Centipedes environment, we first run experiments of all models on", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "score": 1.0, + "content": "CentipedeSix and CentipedeFour to get the pre-trained models for transfer learning. We", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 506, + 138 + ], + "score": 1.0, + "content": "train different models until these agents run equally well as possible, which is reported in Table 1.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 135, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 505, + 150 + ], + "score": 1.0, + "content": "Note that, in practice, we train TreeNet on CentipedeFour for more than 8 million time steps.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 147, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 505, + 160 + ], + "score": 1.0, + "content": "However, due to the difficulty of optimizing TreeNet on CentipedeFour, the performance is still", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 159, + 408, + 170 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 408, + 170 + ], + "score": 1.0, + "content": "lower. But visually, the TreeNet agent is able to run in CentipedeFour.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 103, + 506, + 170 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 175, + 505, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 175, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 175, + 505, + 187 + ], + "score": 1.0, + "content": "We then examine the zero-shot performance where zero-shot means directly applying the model", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 186, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 198 + ], + "score": 1.0, + "content": "trained with one setting to the other without any fine-tuning. To better visualize the results, we", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "score": 1.0, + "content": "linearly normalize the performance to get a performance score, and color the results accordingly.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 209, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 219 + ], + "score": 1.0, + "content": "The normalization scheme is recorded in Appendix 11. The performance score is less than 1, and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 220, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 231 + ], + "score": 1.0, + "content": "is shown in the parentheses behind the original results. As we can see from Figure 4 (full chart in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 229, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 505, + 242 + ], + "score": 1.0, + "content": "Appendix 6.5), NerveNet outperforms all competitors on all settings, except in the 4toCp06 scenario.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 240, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 505, + 253 + ], + "score": 1.0, + "content": "Note that transferring from CentipedeFour is more difficult than from CentipedeSix since", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 251, + 491, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 491, + 264 + ], + "score": 1.0, + "content": "the situation where one torso connects to two neighboring torsos only happens beyond 4 bodies.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 175, + 505, + 264 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 268, + 505, + 313 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "TreeNet has a surprisingly good performance on tasks from CentipedeFour. However, by check-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 280, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 505, + 291 + ], + "score": 1.0, + "content": "ing the videos, the learned agent is actually not able to “move” as good as other methods. The high", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "reward is mainly due to the fact that TreeNet policy is better at standing still and gaining alive bonus.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 300, + 464, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 464, + 315 + ], + "score": 1.0, + "content": "We argue that the average running-length in each episode is also a very important metric.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 268, + 505, + 315 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 318, + 505, + 362 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "By including the results of running-length, we notice that NerveNet is the only model able to walk in", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 330, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 505, + 341 + ], + "score": 1.0, + "content": "the zero-shot setting. In fact, the performance of NerveNet is orders-of-magnitude better, and most", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 340, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 505, + 352 + ], + "score": 1.0, + "content": "of the time, agents from other methods cannot even move forward. We also notice that if transferred", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 352, + 504, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 504, + 363 + ], + "score": 1.0, + "content": "from CentipedeSix, NerveNet is able to provide walkable pre-trained models on all new agents.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 318, + 505, + 363 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 368, + 505, + 445 + ], + "lines": [ + { + "bbox": [ + 106, + 367, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 505, + 381 + ], + "score": 1.0, + "content": "We fine-tune for both size transfer and disability transfer experiments and show the training curves", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 379, + 504, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 504, + 392 + ], + "score": 1.0, + "content": "in Figure 5. From the figure, we can see that by using the pre-trained model, NerveNet significantly", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "decreases the number of episodes required to reach the level of reward which is considered as solved.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 401, + 504, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 504, + 414 + ], + "score": 1.0, + "content": "By looking at the videos, we notice that the bottleneck of learning for the agent is “how to stand”.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 411, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 425 + ], + "score": 1.0, + "content": "When training from scratch, it can be seen that almost 0.5 million time steps are spent on a very flat", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 423, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 436 + ], + "score": 1.0, + "content": "reward surface. Therefore, the MLPAA agents, which copy the learned policy, are able to stand and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 434, + 431, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 431, + 447 + ], + "score": 1.0, + "content": "bypass this time-consuming process and reach to a good performance in the end.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 367, + 505, + 447 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 451, + 504, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 506, + 463 + ], + "score": 1.0, + "content": "Moreover, by examining the result videos, we noticed that the “walk-cycle” behavior is observed", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "for NerveNet but is not common for others. Walk-cycle are adopted for many insects in the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 473, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 505, + 485 + ], + "score": 1.0, + "content": "world, Biewener (2003). For example, six-legged ants use a tripedal gait, where the legs are used", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "score": 1.0, + "content": "in two separate triangles alternatively touching the ground. We give more details of walk-cycle in", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 494, + 155, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 155, + 505 + ], + "score": 1.0, + "content": "Section 4.5.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 450, + 506, + 505 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 511, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 511, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 506, + 524 + ], + "score": 1.0, + "content": "One possible reason is that the agent of MLP based method (MLPAA, MLPP) learns a policy that", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "does not utilize all legs. From CentipedeEight and up, we do not observe any MLP agents to be", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "score": 1.0, + "content": "able to coordinate all legs whereas almost all policies learned by NerveNet use all legs. Therefore,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 545, + 466, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 466, + 557 + ], + "score": 1.0, + "content": "NerveNet is better at utilizing structure information and not over-fitting the environments.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 511, + 506, + 557 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 572, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 571, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 506, + 585 + ], + "score": 1.0, + "content": "Snakes The zero-shot performance for snakes is summarized in Figure 6. As we can see, NerveNet", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "score": 1.0, + "content": "has the best performance on all transfer learning tasks. In most cases, NerveNet has a starting", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "reward value of more than 300, which is a pretty good policy since 350 is considered as solved for", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "snakeThree. By looking at the videos, we found that agents of other competitors are not able to", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 615, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 104, + 615, + 506, + 630 + ], + "score": 1.0, + "content": "control the new actuators in the zero-shot setting. They either overfit to the original models, where", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "the policy is completely useless in the new setting (e.g., the MLPAA is worse than random policy in", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 639, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 650 + ], + "score": 1.0, + "content": "SnakeThree2SnakeFour), or the new actuators are not able to coordinate with the old actuators", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "trained before. While for NerveNet, the actuators are able to coordinate to its neighbors, regardless", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 660, + 259, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 259, + 672 + ], + "score": 1.0, + "content": "of whether they are new to the agents.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43, + "bbox_fs": [ + 104, + 571, + 506, + 672 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "We also summarize the training curves of fine-tuning in Fig. 7. We can observe that NerveNet has a", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "very good initialization with the pre-trained model, and the performance increases with fine-tuning.", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 107, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 107, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "When training from scratch, NerveNet is less sample efficient compared to the MLP model which", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 708, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 724 + ], + "score": 1.0, + "content": "might be caused by the fact that optimizing our model is more challenging than MLP. Fine-tuning", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "helps to improve the sample efficiency of our model by a large margin. At the same time, although", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 483, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 498 + ], + "score": 1.0, + "content": "the MLPAA has a very good initialization, its performance progresses slowly with the increasing", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "number of episodes. In most experiments, the MLPAA and TreeNet did not match the performance", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 507, + 251, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 251, + 519 + ], + "score": 1.0, + "content": "of its non-pretrained MLP baseline.", + "type": "text", + "cross_page": true + } + ], + "index": 14 + } + ], + "index": 50, + "bbox_fs": [ + 105, + 676, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 84, + 488, + 256 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 84, + 488, + 256 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 84, + 488, + 256 + ], + "spans": [ + { + "bbox": [ + 116, + 84, + 488, + 256 + ], + "score": 0.971, + "type": "image", + "image_path": "3709a850e281bcae1eeed9ee523b904fff989358725e2076c690a230315f97b0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 116, + 84, + 488, + 141.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 116, + 141.33333333333334, + 488, + 198.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 116, + 198.66666666666669, + 488, + 256.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 265, + 504, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 263, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 279 + ], + "score": 1.0, + "content": "Figure 8: Results of Multi-task learning. We train the networks simultaneously on five different", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 276, + 192, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 192, + 289 + ], + "score": 1.0, + "content": "tasks from Walker.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 109, + 305, + 502, + 405 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 109, + 305, + 502, + 405 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 305, + 502, + 405 + ], + "spans": [ + { + "bbox": [ + 109, + 305, + 502, + 405 + ], + "score": 0.978, + "html": "
Model HalfHumanoidHopperOstrichWolfHorseAverage
MLPRewardRatio1775.7557.7%1369.5962.0%1198.8848.2%1249.2354.5%2084.0769.7%/58.6%
TreeNetRewardRatio237.8179.3%417.2798.0%224.0757.4%247.03141.2%223.3499.2%/94.8%
NerveNetRewardRatio2536.5296.3%2113.56101.8%1714.6398.8%2054.54105.9%2343.62106.4%1101.8%
", + "type": "table", + "image_path": "7bceb7a5394f06e3831463e6553b8f644ac3399036bf2990a8f84efe73728009.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 109, + 305, + 502, + 338.3333333333333 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 109, + 338.3333333333333, + 502, + 371.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 109, + 371.66666666666663, + 502, + 404.99999999999994 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 412, + 505, + 457 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "score": 1.0, + "content": "Table 2: Results of Multi-task learning, with comparison to the single-task baselines. For three", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "models, the first row is the mean reward of each model of the last 40 iterations. The second row", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "indicates the percentage of the performance of the multi-task model compared with the single-task", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 446, + 203, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 203, + 456 + ], + "score": 1.0, + "content": "baseline of each model.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + } + ], + "index": 7.75 + }, + { + "type": "text", + "bbox": [ + 107, + 484, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 498 + ], + "score": 1.0, + "content": "the MLPAA has a very good initialization, its performance progresses slowly with the increasing", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "number of episodes. In most experiments, the MLPAA and TreeNet did not match the performance", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 507, + 251, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 251, + 519 + ], + "score": 1.0, + "content": "of its non-pretrained MLP baseline.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 107, + 538, + 235, + 549 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 236, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 236, + 550 + ], + "score": 1.0, + "content": "4.3 MULTI-TASK LEARNING", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "In this section, we show that NerveNet has a good potential of multi-task learning by incorporating", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "structure prior into the network structure. It is important to point out that multi-task learning repre-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 582, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 104, + 582, + 506, + 597 + ], + "score": 1.0, + "content": "sents a very difficult, and more often the case, unsolved problem in RL. Most multi-task learning", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "algorithms, Teh et al. (2017); Andreas et al. (2016); Oh et al. (2017); Yang et al. (2017) have not", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "been applied to domains as difficult as locomotion for complex physical models, not to mention", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 616, + 379, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 379, + 629 + ], + "score": 1.0, + "content": "multi-task learning among different agents with different dynamics.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 504, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 504, + 645 + ], + "score": 1.0, + "content": "In this work, we constrain our problem domain, and design the Walker multi-task learning task-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 642, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 104, + 642, + 506, + 658 + ], + "score": 1.0, + "content": "set, which contains five 2d-walkers. We aim to test the model’s ability of multi-task learning,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "in particular, the ability to control multiple agents using one unified network. The walkers are", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "very different in terms of their dynamics, since they have very distinct structures, different types", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "and numbers of controllers. Walker-HalfHumanoid and Walker-Hopper are variants of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 686, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 104, + 686, + 506, + 702 + ], + "score": 1.0, + "content": "Walker2d and Hopper from the original MuJoCo Benchmarks, respectively. Walker-Horse,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 699, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 504, + 711 + ], + "score": 1.0, + "content": "Walker-Ostrich, Walker-Wolf on the other hand, are agents mimicking the natural animals.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "Just like the real animals, some of the agents have tails or necks to help them to balance. The detailed", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 720, + 306, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 306, + 733 + ], + "score": 1.0, + "content": "schematic figures are shown in the Appendix 6.8.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 84, + 488, + 256 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 84, + 488, + 256 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 84, + 488, + 256 + ], + "spans": [ + { + "bbox": [ + 116, + 84, + 488, + 256 + ], + "score": 0.971, + "type": "image", + "image_path": "3709a850e281bcae1eeed9ee523b904fff989358725e2076c690a230315f97b0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 116, + 84, + 488, + 141.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 116, + 141.33333333333334, + 488, + 198.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 116, + 198.66666666666669, + 488, + 256.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 265, + 504, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 263, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 279 + ], + "score": 1.0, + "content": "Figure 8: Results of Multi-task learning. We train the networks simultaneously on five different", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 276, + 192, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 192, + 289 + ], + "score": 1.0, + "content": "tasks from Walker.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 109, + 305, + 502, + 405 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 109, + 305, + 502, + 405 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 305, + 502, + 405 + ], + "spans": [ + { + "bbox": [ + 109, + 305, + 502, + 405 + ], + "score": 0.978, + "html": "
Model HalfHumanoidHopperOstrichWolfHorseAverage
MLPRewardRatio1775.7557.7%1369.5962.0%1198.8848.2%1249.2354.5%2084.0769.7%/58.6%
TreeNetRewardRatio237.8179.3%417.2798.0%224.0757.4%247.03141.2%223.3499.2%/94.8%
NerveNetRewardRatio2536.5296.3%2113.56101.8%1714.6398.8%2054.54105.9%2343.62106.4%1101.8%
", + "type": "table", + "image_path": "7bceb7a5394f06e3831463e6553b8f644ac3399036bf2990a8f84efe73728009.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 109, + 305, + 502, + 338.3333333333333 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 109, + 338.3333333333333, + 502, + 371.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 109, + 371.66666666666663, + 502, + 404.99999999999994 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 412, + 505, + 457 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "score": 1.0, + "content": "Table 2: Results of Multi-task learning, with comparison to the single-task baselines. For three", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "models, the first row is the mean reward of each model of the last 40 iterations. The second row", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "indicates the percentage of the performance of the multi-task model compared with the single-task", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 446, + 203, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 203, + 456 + ], + "score": 1.0, + "content": "baseline of each model.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + } + ], + "index": 7.75 + }, + { + "type": "text", + "bbox": [ + 107, + 484, + 505, + 518 + ], + "lines": [], + "index": 13, + "bbox_fs": [ + 105, + 483, + 505, + 519 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 538, + 235, + 549 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 236, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 236, + 550 + ], + "score": 1.0, + "content": "4.3 MULTI-TASK LEARNING", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "In this section, we show that NerveNet has a good potential of multi-task learning by incorporating", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "structure prior into the network structure. It is important to point out that multi-task learning repre-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 582, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 104, + 582, + 506, + 597 + ], + "score": 1.0, + "content": "sents a very difficult, and more often the case, unsolved problem in RL. Most multi-task learning", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "algorithms, Teh et al. (2017); Andreas et al. (2016); Oh et al. (2017); Yang et al. (2017) have not", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "been applied to domains as difficult as locomotion for complex physical models, not to mention", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 616, + 379, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 379, + 629 + ], + "score": 1.0, + "content": "multi-task learning among different agents with different dynamics.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5, + "bbox_fs": [ + 104, + 560, + 506, + 629 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 504, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 504, + 645 + ], + "score": 1.0, + "content": "In this work, we constrain our problem domain, and design the Walker multi-task learning task-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 642, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 104, + 642, + 506, + 658 + ], + "score": 1.0, + "content": "set, which contains five 2d-walkers. We aim to test the model’s ability of multi-task learning,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "in particular, the ability to control multiple agents using one unified network. The walkers are", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "very different in terms of their dynamics, since they have very distinct structures, different types", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "and numbers of controllers. Walker-HalfHumanoid and Walker-Hopper are variants of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 686, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 104, + 686, + 506, + 702 + ], + "score": 1.0, + "content": "Walker2d and Hopper from the original MuJoCo Benchmarks, respectively. Walker-Horse,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 699, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 504, + 711 + ], + "score": 1.0, + "content": "Walker-Ostrich, Walker-Wolf on the other hand, are agents mimicking the natural animals.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "Just like the real animals, some of the agents have tails or necks to help them to balance. The detailed", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 720, + 306, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 306, + 733 + ], + "score": 1.0, + "content": "schematic figures are shown in the Appendix 6.8.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26, + "bbox_fs": [ + 104, + 633, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 81, + 499, + 153 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 111, + 81, + 499, + 153 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 81, + 499, + 153 + ], + "spans": [ + { + "bbox": [ + 111, + 81, + 499, + 153 + ], + "score": 0.979, + "html": "
ModelHalfhumanoid HopperWolfOstrich HorseAverage
MassMLPNerveNet33.28%95.87%74.04%93.24%94.68%90.13%59.23% 40.61%80.2% 69.23%60.37%85.73%
StrengthMLPNerveNet25.96%31.11%21.77%42.20%27.32%42.84%30.08% 19.80%31.41% 36.54%24.99%36.82%
", + "type": "table", + "image_path": "024ef3d5511d4e580856abdd8bfa2049dc00cf405bbe5b20dc053da6a84fa458.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 81, + 499, + 105.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 105.0, + 499, + 129.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 129.0, + 499, + 153.0 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 161, + 505, + 205 + ], + "lines": [ + { + "bbox": [ + 105, + 159, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 174 + ], + "score": 1.0, + "content": "Table 3: Results of robustness evaluations. Note that we show the average results for each type", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 172, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 184 + ], + "score": 1.0, + "content": "of parameters after perturbation. And the results are columned by the agent type. The ratio of the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 183, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 195 + ], + "score": 1.0, + "content": "average performance of perturbed agents and the original performance is shown in the figure. Details", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 194, + 172, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 172, + 205 + ], + "score": 1.0, + "content": "are listed in 6.6.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 108, + 230, + 253, + 242 + ], + "lines": [ + { + "bbox": [ + 105, + 229, + 254, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 254, + 243 + ], + "score": 1.0, + "content": "4.3.1 EXPERIMENTAL SETTINGS", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 251, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 252, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 506, + 264 + ], + "score": 1.0, + "content": "To show the ability of multi-task learning of NerveNet, we design several baselines. We use a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 263, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 274 + ], + "score": 1.0, + "content": "vanilla multi-task policy update for all models. More specifically, for each sub-task in the multi-task", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 274, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 505, + 286 + ], + "score": 1.0, + "content": "learning task-set, we use an equal number of time steps for each policy’s update and calculate the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 284, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 505, + 297 + ], + "score": 1.0, + "content": "gradients separately. Gradients are then aggregated and the mean value of gradients is applied to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "update the network. To compensate for the additional difficulty in training more agents and tasks,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 306, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 318 + ], + "score": 1.0, + "content": "we linearly increase the number of update epochs during each update in training, as well as the total", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "number of time steps generated before the training is terminated. The hyper-parameter setting is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 329, + 228, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 228, + 341 + ], + "score": 1.0, + "content": "summarized in Appendix 6.7.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 356, + 503, + 378 + ], + "lines": [ + { + "bbox": [ + 106, + 356, + 504, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 504, + 369 + ], + "score": 1.0, + "content": "NerveNet For NerveNet, the weights are naturally shared among different agents. More specifi-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 367, + 447, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 447, + 381 + ], + "score": 1.0, + "content": "cally, for different agents, the weight matrices for propagation and output are shared.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 394, + 479, + 406 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 480, + 408 + ], + "score": 1.0, + "content": "MLP Sharing For the MLP method, we shared the weight matrices between hidden layers.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 422, + 505, + 488 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "MLP Aggregation In the MLP Sharing approach, the total size of the weight matrices grows with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "the number of tasks. For different agents, whose dimension of the observations are usually different,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "weights from observation to the first hidden layer cannot be reused in the MLP Sharing approach.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "Therefore in the MLP Aggregation method, we multiply each element of the observation vector", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 467, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 479 + ], + "score": 1.0, + "content": "separately by one matrix, and aggregate the resulting vectors from each element. The size of this", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 478, + 351, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 351, + 489 + ], + "score": 1.0, + "content": "multiplying matrix is (1, dimension of the first hidden layer).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 108, + 505, + 505, + 538 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 504, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 504, + 518 + ], + "score": 1.0, + "content": "TreeNet Similarly, TreeNet also has the benefits that its weights are naturally shared among dif-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "ferent agents. However, TreeNet has no knowledge of the agents’ physical structure, where the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 527, + 339, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 339, + 540 + ], + "score": 1.0, + "content": "information of each node is aggregated into the root node.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 108, + 544, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 106, + 544, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 505, + 556 + ], + "score": 1.0, + "content": "We also include the baseline of training single-task MLP for each agent. We train the single-task", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "MLP baselines for 1 million time steps per agent. In the Figure 8, we align the results of single-task", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 565, + 466, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 466, + 578 + ], + "score": 1.0, + "content": "MLP baseline and the results of multi-task models by the number of episodes of one task.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 582, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 583, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 594 + ], + "score": 1.0, + "content": "As can be seen from Figure 8, NerveNet achieves the best performance in all the sub-tasks. In", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "Walker-HalfHumanoid, Walker-Hopper, Walker-Ostrich, Walker-Wolf our Ner-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "veNet is able to out-perform other agents by a large margin. In Walker-Horse, the performance", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "of NerveNet and MLP Sharing are relatively similar. For MLP Sharing, the performance on other", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "four agents are relatively limited, while for Walker-Hopper, the improvement of performance is", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "limited from half of the experiment onwards. The MLP Aggregation and TreeNet methods are not", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "able to solve the multi-task learning problem, with both of them stuck at a very low reward level. In", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 660, + 495, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 495, + 672 + ], + "score": 1.0, + "content": "the vanilla optimization setting, we show that NerveNet has a bigger potential than the baselines.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 430, + 690 + ], + "score": 1.0, + "content": "From Table 2, one can observe that the performance of MLP drops drastically", + "type": "text" + }, + { + "bbox": [ + 430, + 677, + 451, + 688 + ], + "score": 0.85, + "content": "42 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "performance", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "drop) when switching from single-task to multi-task learning, while for NerveNet, there is no obvi-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "ous drop in performance. Our intuition is that NerveNet is better at learning generalized features,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "and learning of different agents can help in training other agents, while for MLP methods, the per-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 346, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 346, + 733 + ], + "score": 1.0, + "content": "formance decreases due the competition of different agents.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 763 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 13, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 81, + 499, + 153 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 111, + 81, + 499, + 153 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 81, + 499, + 153 + ], + "spans": [ + { + "bbox": [ + 111, + 81, + 499, + 153 + ], + "score": 0.979, + "html": "
ModelHalfhumanoid HopperWolfOstrich HorseAverage
MassMLPNerveNet33.28%95.87%74.04%93.24%94.68%90.13%59.23% 40.61%80.2% 69.23%60.37%85.73%
StrengthMLPNerveNet25.96%31.11%21.77%42.20%27.32%42.84%30.08% 19.80%31.41% 36.54%24.99%36.82%
", + "type": "table", + "image_path": "024ef3d5511d4e580856abdd8bfa2049dc00cf405bbe5b20dc053da6a84fa458.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 81, + 499, + 105.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 105.0, + 499, + 129.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 129.0, + 499, + 153.0 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 161, + 505, + 205 + ], + "lines": [ + { + "bbox": [ + 105, + 159, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 174 + ], + "score": 1.0, + "content": "Table 3: Results of robustness evaluations. Note that we show the average results for each type", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 172, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 184 + ], + "score": 1.0, + "content": "of parameters after perturbation. And the results are columned by the agent type. The ratio of the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 183, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 195 + ], + "score": 1.0, + "content": "average performance of perturbed agents and the original performance is shown in the figure. Details", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 194, + 172, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 172, + 205 + ], + "score": 1.0, + "content": "are listed in 6.6.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 159, + 506, + 205 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 230, + 253, + 242 + ], + "lines": [ + { + "bbox": [ + 105, + 229, + 254, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 254, + 243 + ], + "score": 1.0, + "content": "4.3.1 EXPERIMENTAL SETTINGS", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 251, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 252, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 506, + 264 + ], + "score": 1.0, + "content": "To show the ability of multi-task learning of NerveNet, we design several baselines. We use a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 263, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 274 + ], + "score": 1.0, + "content": "vanilla multi-task policy update for all models. More specifically, for each sub-task in the multi-task", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 274, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 505, + 286 + ], + "score": 1.0, + "content": "learning task-set, we use an equal number of time steps for each policy’s update and calculate the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 284, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 505, + 297 + ], + "score": 1.0, + "content": "gradients separately. Gradients are then aggregated and the mean value of gradients is applied to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "update the network. To compensate for the additional difficulty in training more agents and tasks,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 306, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 318 + ], + "score": 1.0, + "content": "we linearly increase the number of update epochs during each update in training, as well as the total", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "number of time steps generated before the training is terminated. The hyper-parameter setting is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 329, + 228, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 228, + 341 + ], + "score": 1.0, + "content": "summarized in Appendix 6.7.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 252, + 506, + 341 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 356, + 503, + 378 + ], + "lines": [ + { + "bbox": [ + 106, + 356, + 504, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 504, + 369 + ], + "score": 1.0, + "content": "NerveNet For NerveNet, the weights are naturally shared among different agents. More specifi-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 367, + 447, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 447, + 381 + ], + "score": 1.0, + "content": "cally, for different agents, the weight matrices for propagation and output are shared.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 356, + 504, + 381 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 394, + 479, + 406 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 480, + 408 + ], + "score": 1.0, + "content": "MLP Sharing For the MLP method, we shared the weight matrices between hidden layers.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18, + "bbox_fs": [ + 106, + 393, + 480, + 408 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 422, + 505, + 488 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "MLP Aggregation In the MLP Sharing approach, the total size of the weight matrices grows with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "the number of tasks. For different agents, whose dimension of the observations are usually different,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "weights from observation to the first hidden layer cannot be reused in the MLP Sharing approach.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "Therefore in the MLP Aggregation method, we multiply each element of the observation vector", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 467, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 479 + ], + "score": 1.0, + "content": "separately by one matrix, and aggregate the resulting vectors from each element. The size of this", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 478, + 351, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 351, + 489 + ], + "score": 1.0, + "content": "multiplying matrix is (1, dimension of the first hidden layer).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 423, + 505, + 489 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 505, + 505, + 538 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 504, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 504, + 518 + ], + "score": 1.0, + "content": "TreeNet Similarly, TreeNet also has the benefits that its weights are naturally shared among dif-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "ferent agents. However, TreeNet has no knowledge of the agents’ physical structure, where the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 527, + 339, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 339, + 540 + ], + "score": 1.0, + "content": "information of each node is aggregated into the root node.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 505, + 505, + 540 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 544, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 106, + 544, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 505, + 556 + ], + "score": 1.0, + "content": "We also include the baseline of training single-task MLP for each agent. We train the single-task", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "MLP baselines for 1 million time steps per agent. In the Figure 8, we align the results of single-task", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 565, + 466, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 466, + 578 + ], + "score": 1.0, + "content": "MLP baseline and the results of multi-task models by the number of episodes of one task.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 544, + 505, + 578 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 582, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 583, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 594 + ], + "score": 1.0, + "content": "As can be seen from Figure 8, NerveNet achieves the best performance in all the sub-tasks. In", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "Walker-HalfHumanoid, Walker-Hopper, Walker-Ostrich, Walker-Wolf our Ner-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "veNet is able to out-perform other agents by a large margin. In Walker-Horse, the performance", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "of NerveNet and MLP Sharing are relatively similar. For MLP Sharing, the performance on other", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "four agents are relatively limited, while for Walker-Hopper, the improvement of performance is", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "limited from half of the experiment onwards. The MLP Aggregation and TreeNet methods are not", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "able to solve the multi-task learning problem, with both of them stuck at a very low reward level. In", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 660, + 495, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 495, + 672 + ], + "score": 1.0, + "content": "the vanilla optimization setting, we show that NerveNet has a bigger potential than the baselines.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 583, + 505, + 672 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 430, + 690 + ], + "score": 1.0, + "content": "From Table 2, one can observe that the performance of MLP drops drastically", + "type": "text" + }, + { + "bbox": [ + 430, + 677, + 451, + 688 + ], + "score": 0.85, + "content": "42 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "performance", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "drop) when switching from single-task to multi-task learning, while for NerveNet, there is no obvi-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "ous drop in performance. Our intuition is that NerveNet is better at learning generalized features,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "and learning of different agents can help in training other agents, while for MLP methods, the per-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 346, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 346, + 733 + ], + "score": 1.0, + "content": "formance decreases due the competition of different agents.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 676, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 121, + 92, + 489, + 187 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 121, + 92, + 489, + 187 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 121, + 92, + 489, + 187 + ], + "spans": [ + { + "bbox": [ + 121, + 92, + 489, + 187 + ], + "score": 0.951, + "type": "image", + "image_path": "892025ae723031d5978ddef92b55588d72f89132d24844c0cfb5e117cf25e298.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 121, + 92, + 489, + 123.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 121, + 123.66666666666667, + 489, + 155.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 121, + 155.33333333333334, + 489, + 187.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 201, + 506, + 236 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 201, + 506, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 506, + 215 + ], + "score": 1.0, + "content": "Figure 9: Diagram of the walk cycle. In the left figure, legs within the same triangle are used", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 213, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 505, + 226 + ], + "score": 1.0, + "content": "simultaneously. For each leg, we use the same color for their diagram on the left and their curves on", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 222, + 146, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 146, + 238 + ], + "score": 1.0, + "content": "the right.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 115, + 266, + 492, + 510 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 266, + 492, + 510 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 115, + 266, + 492, + 510 + ], + "spans": [ + { + "bbox": [ + 115, + 266, + 492, + 510 + ], + "score": 0.971, + "type": "image", + "image_path": "652d63005f4eea342561cb0a44ebd4dd31b388fd4be3b385a751118a06287c04.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 115, + 266, + 492, + 347.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 115, + 347.3333333333333, + 492, + 428.66666666666663 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 115, + 428.66666666666663, + 492, + 509.99999999999994 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 520, + 505, + 554 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 520, + 504, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 504, + 532 + ], + "score": 1.0, + "content": "Figure 10: Results of visualization of feature distribution and trajectory density. As can be seen from", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 530, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 505, + 544 + ], + "score": 1.0, + "content": "the figure, NerveNet agent is able to learn shareable features for its legs, and certain walk-cycle is", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 541, + 196, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 196, + 556 + ], + "score": 1.0, + "content": "learnt during training.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 108, + 589, + 282, + 600 + ], + "lines": [ + { + "bbox": [ + 106, + 588, + 284, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 284, + 601 + ], + "score": 1.0, + "content": "4.4 ROBUSTNESS OF LEARNT POLICIES", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 615, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "In this section, we also report the robustness of our policy by perturbing the agent parameters.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "In reality, the parameters simulated might be different from the actual parameters of the agents.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "Therefore, it is important that the agent is robust to parameters perturbation. The model that has the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 649, + 367, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 367, + 661 + ], + "score": 1.0, + "content": "better ability to learn generalized features are likely more robust.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "We perturb the mass of the geometries (rigid bodies) in MuJoCo as well as the scale of the forces of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "the joints. We use the pre-trained models with similar performance on the original task for both the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 686, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 505, + 702 + ], + "score": 1.0, + "content": "MLP and NerveNet. The performance is tested in five agents from Walker task set. The average", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "performance is recorded in Table 3, and the specific details are summarized in Appendix 6.6. The", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "robustness of NerveNet’ policy is likely due to the structure prior of the agent instilled in the network,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 719, + 221, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 221, + 734 + ], + "score": 1.0, + "content": "which facilitates overfitting.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 121, + 92, + 489, + 187 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 121, + 92, + 489, + 187 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 121, + 92, + 489, + 187 + ], + "spans": [ + { + "bbox": [ + 121, + 92, + 489, + 187 + ], + "score": 0.951, + "type": "image", + "image_path": "892025ae723031d5978ddef92b55588d72f89132d24844c0cfb5e117cf25e298.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 121, + 92, + 489, + 123.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 121, + 123.66666666666667, + 489, + 155.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 121, + 155.33333333333334, + 489, + 187.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 201, + 506, + 236 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 201, + 506, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 506, + 215 + ], + "score": 1.0, + "content": "Figure 9: Diagram of the walk cycle. In the left figure, legs within the same triangle are used", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 213, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 505, + 226 + ], + "score": 1.0, + "content": "simultaneously. For each leg, we use the same color for their diagram on the left and their curves on", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 222, + 146, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 146, + 238 + ], + "score": 1.0, + "content": "the right.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 115, + 266, + 492, + 510 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 266, + 492, + 510 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 115, + 266, + 492, + 510 + ], + "spans": [ + { + "bbox": [ + 115, + 266, + 492, + 510 + ], + "score": 0.971, + "type": "image", + "image_path": "652d63005f4eea342561cb0a44ebd4dd31b388fd4be3b385a751118a06287c04.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 115, + 266, + 492, + 347.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 115, + 347.3333333333333, + 492, + 428.66666666666663 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 115, + 428.66666666666663, + 492, + 509.99999999999994 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 520, + 505, + 554 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 520, + 504, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 504, + 532 + ], + "score": 1.0, + "content": "Figure 10: Results of visualization of feature distribution and trajectory density. As can be seen from", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 530, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 505, + 544 + ], + "score": 1.0, + "content": "the figure, NerveNet agent is able to learn shareable features for its legs, and certain walk-cycle is", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 541, + 196, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 196, + 556 + ], + "score": 1.0, + "content": "learnt during training.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 108, + 589, + 282, + 600 + ], + "lines": [ + { + "bbox": [ + 106, + 588, + 284, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 284, + 601 + ], + "score": 1.0, + "content": "4.4 ROBUSTNESS OF LEARNT POLICIES", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 615, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "In this section, we also report the robustness of our policy by perturbing the agent parameters.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "In reality, the parameters simulated might be different from the actual parameters of the agents.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "Therefore, it is important that the agent is robust to parameters perturbation. The model that has the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 649, + 367, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 367, + 661 + ], + "score": 1.0, + "content": "better ability to learn generalized features are likely more robust.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 615, + 505, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "We perturb the mass of the geometries (rigid bodies) in MuJoCo as well as the scale of the forces of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "the joints. We use the pre-trained models with similar performance on the original task for both the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 686, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 505, + 702 + ], + "score": 1.0, + "content": "MLP and NerveNet. The performance is tested in five agents from Walker task set. The average", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "performance is recorded in Table 3, and the specific details are summarized in Appendix 6.6. The", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "robustness of NerveNet’ policy is likely due to the structure prior of the agent instilled in the network,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 719, + 221, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 221, + 734 + ], + "score": 1.0, + "content": "which facilitates overfitting.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 665, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 118, + 84, + 489, + 168 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 84, + 489, + 168 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 84, + 489, + 168 + ], + "spans": [ + { + "bbox": [ + 118, + 84, + 489, + 168 + ], + "score": 0.961, + "type": "image", + "image_path": "26128649b00315a80503ddc2f5af6927c5694b169d26b2066dfb4ce234144b70.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 118, + 84, + 489, + 112.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 118, + 112.0, + 489, + 140.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 118, + 140.0, + 489, + 168.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 128, + 177, + 482, + 189 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 176, + 484, + 190 + ], + "spans": [ + { + "bbox": [ + 126, + 176, + 484, + 190 + ], + "score": 1.0, + "content": "Figure 11: Results of several variants of NerveNet for the reinforcement learning agents.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 107, + 210, + 343, + 222 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 344, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 344, + 223 + ], + "score": 1.0, + "content": "4.5 INTERPRETING THE LEARNED REPRESENTATIONS", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 231, + 505, + 330 + ], + "lines": [ + { + "bbox": [ + 106, + 232, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 505, + 243 + ], + "score": 1.0, + "content": "In this section, we try to visualize and interpret the learned representations. We extract the final", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 242, + 504, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 448, + 255 + ], + "score": 1.0, + "content": "state vectors of nodes of NerveNet trained on CentipedeEight. We then apply", + "type": "text" + }, + { + "bbox": [ + 448, + 243, + 466, + 253 + ], + "score": 0.37, + "content": "1 - D", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 242, + 494, + 255 + ], + "score": 1.0, + "content": "and 2-", + "type": "text" + }, + { + "bbox": [ + 495, + 243, + 504, + 252 + ], + "score": 0.33, + "content": "D", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "PCA on the node representations. In Figure 10, we notice that each pair of legs is able to learn", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 262, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 505, + 279 + ], + "score": 1.0, + "content": "invariant representations, despite their different position in the agent. We further plot the trajectory", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "density map in the feature map. By recording the period of the walk-cycle, we plot the transformed", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 287, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 505, + 298 + ], + "score": 1.0, + "content": "features of the 6 legs on Figure 9. As we can see, there is a clear periodic behavior of our hidden", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 297, + 504, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 504, + 310 + ], + "score": 1.0, + "content": "representations learned by our model. Furthermore, the representations of adjacent left legs and the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "adjacent right legs demonstrate a phase shift, which further proves that our agents are able to learn", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 320, + 309, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 309, + 331 + ], + "score": 1.0, + "content": "the walk-cycle without any additional supervision.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9 + }, + { + "type": "title", + "bbox": [ + 109, + 344, + 284, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 286, + 357 + ], + "score": 1.0, + "content": "4.6 COMPARISON OF MODEL VARIANTS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 108, + 365, + 504, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "score": 1.0, + "content": "We have several variants of NerveNet, based on the type of network we use for the policy/value", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "representation. We here compare all variants. Again, we run experiments for each task three times.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 386, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 506, + 399 + ], + "score": 1.0, + "content": "The details of hyper-parameters are given in the Appendix. For each environment, we train the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 398, + 396, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 396, + 410 + ], + "score": 1.0, + "content": "network for one million time steps, with batch size 2050 for one update.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 415, + 505, + 470 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 428 + ], + "score": 1.0, + "content": "As we can see from Figure 11, the NerveNet-MLP and NerveNet-2 variants perform better than", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 426, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 505, + 438 + ], + "score": 1.0, + "content": "NerveNet-1. One potential reason is that sharing the weights of the value and policy networks", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 470, + 450 + ], + "score": 1.0, + "content": "makes the trust-region based optimization methods, like PPO, more sensitive to the weight", + "type": "text" + }, + { + "bbox": [ + 471, + 438, + 478, + 447 + ], + "score": 0.79, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 437, + 505, + 450 + ], + "score": 1.0, + "content": "of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 446, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 341, + 462 + ], + "score": 1.0, + "content": "value function in equation 8. Based on the figure, choosing", + "type": "text" + }, + { + "bbox": [ + 342, + 450, + 349, + 457 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 446, + 506, + 462 + ], + "score": 1.0, + "content": "to be 1 is not giving good performance", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 458, + 242, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 242, + 472 + ], + "score": 1.0, + "content": "on the tasks we experimented on.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 108, + 486, + 195, + 499 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 197, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 197, + 502 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 511, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 525 + ], + "score": 1.0, + "content": "In this paper, we aimed to exploit the body structure of Reinforcement Learning agents in the form", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 522, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 535 + ], + "score": 1.0, + "content": "of graphs. We introduced a novel model called NerveNet which uses a Graph Neural Network to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "represent the agent’s policy. At each time instance of the environment, NerveNet takes observations", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "score": 1.0, + "content": "for each of the body joints, and propagates information between them using non-linear messages", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 555, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 569 + ], + "score": 1.0, + "content": "computed with a neural network. Propagation is done through the edges which represent natural", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 566, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 580 + ], + "score": 1.0, + "content": "dependencies between joints, such as physical connectivity. We experimentally showed that our", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "NerveNet achieves comparable performance to state-of-the-art methods on standard MuJoCo envi-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 588, + 504, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 504, + 601 + ], + "score": 1.0, + "content": "ronments. We further propose our customized reinforcement learning environments for benchmark-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 600, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 612 + ], + "score": 1.0, + "content": "ing two types of structure transfer learning tasks, i.e., size and disability transfer. We demonstrate", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "that policies learned by NerveNet are significantly better than policies learned by other models and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 621, + 292, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 292, + 635 + ], + "score": 1.0, + "content": "are able to transfer even in a zero-shot setting.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30 + }, + { + "type": "title", + "bbox": [ + 107, + 649, + 175, + 661 + ], + "lines": [ + { + "bbox": [ + 106, + 649, + 176, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 176, + 663 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 108, + 668, + 504, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 681 + ], + "score": 1.0, + "content": "Haitham B Ammar, Eric Eaton, Paul Ruvolo, and Matthew Taylor. Online multi-task learning for", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 680, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 116, + 680, + 505, + 691 + ], + "score": 1.0, + "content": "policy gradient methods. In Proceedings of the 31st International Conference on Machine Learn-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 690, + 271, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 271, + 702 + ], + "score": 1.0, + "content": "ing (ICML-14), pp. 1206–1214, 2014.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "Jacob Andreas, Dan Klein, and Sergey Levine. Modular multitask reinforcement learning with", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 721, + 345, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 345, + 732 + ], + "score": 1.0, + "content": "policy sketches. arXiv preprint arXiv:1611.01796, 2016.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 118, + 84, + 489, + 168 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 84, + 489, + 168 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 84, + 489, + 168 + ], + "spans": [ + { + "bbox": [ + 118, + 84, + 489, + 168 + ], + "score": 0.961, + "type": "image", + "image_path": "26128649b00315a80503ddc2f5af6927c5694b169d26b2066dfb4ce234144b70.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 118, + 84, + 489, + 112.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 118, + 112.0, + 489, + 140.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 118, + 140.0, + 489, + 168.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 128, + 177, + 482, + 189 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 176, + 484, + 190 + ], + "spans": [ + { + "bbox": [ + 126, + 176, + 484, + 190 + ], + "score": 1.0, + "content": "Figure 11: Results of several variants of NerveNet for the reinforcement learning agents.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 107, + 210, + 343, + 222 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 344, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 344, + 223 + ], + "score": 1.0, + "content": "4.5 INTERPRETING THE LEARNED REPRESENTATIONS", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 231, + 505, + 330 + ], + "lines": [ + { + "bbox": [ + 106, + 232, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 505, + 243 + ], + "score": 1.0, + "content": "In this section, we try to visualize and interpret the learned representations. We extract the final", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 242, + 504, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 448, + 255 + ], + "score": 1.0, + "content": "state vectors of nodes of NerveNet trained on CentipedeEight. We then apply", + "type": "text" + }, + { + "bbox": [ + 448, + 243, + 466, + 253 + ], + "score": 0.37, + "content": "1 - D", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 242, + 494, + 255 + ], + "score": 1.0, + "content": "and 2-", + "type": "text" + }, + { + "bbox": [ + 495, + 243, + 504, + 252 + ], + "score": 0.33, + "content": "D", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "PCA on the node representations. In Figure 10, we notice that each pair of legs is able to learn", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 262, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 505, + 279 + ], + "score": 1.0, + "content": "invariant representations, despite their different position in the agent. We further plot the trajectory", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "density map in the feature map. By recording the period of the walk-cycle, we plot the transformed", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 287, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 505, + 298 + ], + "score": 1.0, + "content": "features of the 6 legs on Figure 9. As we can see, there is a clear periodic behavior of our hidden", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 297, + 504, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 504, + 310 + ], + "score": 1.0, + "content": "representations learned by our model. Furthermore, the representations of adjacent left legs and the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "adjacent right legs demonstrate a phase shift, which further proves that our agents are able to learn", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 320, + 309, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 309, + 331 + ], + "score": 1.0, + "content": "the walk-cycle without any additional supervision.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 232, + 506, + 331 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 344, + 284, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 286, + 357 + ], + "score": 1.0, + "content": "4.6 COMPARISON OF MODEL VARIANTS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 108, + 365, + 504, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "score": 1.0, + "content": "We have several variants of NerveNet, based on the type of network we use for the policy/value", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "representation. We here compare all variants. Again, we run experiments for each task three times.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 386, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 506, + 399 + ], + "score": 1.0, + "content": "The details of hyper-parameters are given in the Appendix. For each environment, we train the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 398, + 396, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 396, + 410 + ], + "score": 1.0, + "content": "network for one million time steps, with batch size 2050 for one update.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 364, + 506, + 410 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 415, + 505, + 470 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 428 + ], + "score": 1.0, + "content": "As we can see from Figure 11, the NerveNet-MLP and NerveNet-2 variants perform better than", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 426, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 505, + 438 + ], + "score": 1.0, + "content": "NerveNet-1. One potential reason is that sharing the weights of the value and policy networks", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 470, + 450 + ], + "score": 1.0, + "content": "makes the trust-region based optimization methods, like PPO, more sensitive to the weight", + "type": "text" + }, + { + "bbox": [ + 471, + 438, + 478, + 447 + ], + "score": 0.79, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 437, + 505, + 450 + ], + "score": 1.0, + "content": "of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 446, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 341, + 462 + ], + "score": 1.0, + "content": "value function in equation 8. Based on the figure, choosing", + "type": "text" + }, + { + "bbox": [ + 342, + 450, + 349, + 457 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 446, + 506, + 462 + ], + "score": 1.0, + "content": "to be 1 is not giving good performance", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 458, + 242, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 242, + 472 + ], + "score": 1.0, + "content": "on the tasks we experimented on.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 414, + 506, + 472 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 486, + 195, + 499 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 197, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 197, + 502 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 511, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 525 + ], + "score": 1.0, + "content": "In this paper, we aimed to exploit the body structure of Reinforcement Learning agents in the form", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 522, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 535 + ], + "score": 1.0, + "content": "of graphs. We introduced a novel model called NerveNet which uses a Graph Neural Network to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "represent the agent’s policy. At each time instance of the environment, NerveNet takes observations", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "score": 1.0, + "content": "for each of the body joints, and propagates information between them using non-linear messages", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 555, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 569 + ], + "score": 1.0, + "content": "computed with a neural network. Propagation is done through the edges which represent natural", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 566, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 580 + ], + "score": 1.0, + "content": "dependencies between joints, such as physical connectivity. We experimentally showed that our", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "NerveNet achieves comparable performance to state-of-the-art methods on standard MuJoCo envi-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 588, + 504, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 504, + 601 + ], + "score": 1.0, + "content": "ronments. We further propose our customized reinforcement learning environments for benchmark-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 600, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 612 + ], + "score": 1.0, + "content": "ing two types of structure transfer learning tasks, i.e., size and disability transfer. We demonstrate", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "that policies learned by NerveNet are significantly better than policies learned by other models and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 621, + 292, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 292, + 635 + ], + "score": 1.0, + "content": "are able to transfer even in a zero-shot setting.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 510, + 506, + 635 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 649, + 175, + 661 + ], + "lines": [ + { + "bbox": [ + 106, + 649, + 176, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 176, + 663 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 108, + 668, + 504, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 681 + ], + "score": 1.0, + "content": "Haitham B Ammar, Eric Eaton, Paul Ruvolo, and Matthew Taylor. Online multi-task learning for", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 680, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 116, + 680, + 505, + 691 + ], + "score": 1.0, + "content": "policy gradient methods. In Proceedings of the 31st International Conference on Machine Learn-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 690, + 271, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 271, + 702 + ], + "score": 1.0, + "content": "ing (ICML-14), pp. 1206–1214, 2014.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 667, + 505, + 702 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "Jacob Andreas, Dan Klein, and Sergey Levine. Modular multitask reinforcement learning with", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 721, + 345, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 345, + 732 + ], + "score": 1.0, + "content": "policy sketches. arXiv preprint arXiv:1611.01796, 2016.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 709, + 504, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Hagit Attiya and Jennifer Welch. Distributed computing: fundamentals, simulations, and advanced", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 94, + 301, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 301, + 106 + ], + "score": 1.0, + "content": "topics, volume 19. John Wiley & Sons, 2004.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 108, + 113, + 503, + 147 + ], + "lines": [ + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "score": 1.0, + "content": "Andre Barreto, Will Dabney, R ´ emi Munos, Jonathan J Hunt, Tom Schaul, David Silver, and Hado P ´", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 125, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 117, + 125, + 505, + 137 + ], + "score": 1.0, + "content": "van Hasselt. Successor features for transfer in reinforcement learning. In Advances in Neural", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 339, + 149 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 339, + 149 + ], + "score": 1.0, + "content": "Information Processing Systems, pp. 4056–4066, 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 155, + 402, + 168 + ], + "lines": [ + { + "bbox": [ + 106, + 155, + 402, + 169 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 402, + 169 + ], + "score": 1.0, + "content": "Andrew A Biewener. Animal locomotion. Oxford University Press, 2003.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 108, + 176, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 106, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 186, + 416, + 201 + ], + "spans": [ + { + "bbox": [ + 116, + 186, + 416, + 201 + ], + "score": 1.0, + "content": "Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 207, + 504, + 230 + ], + "lines": [ + { + "bbox": [ + 105, + 206, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 505, + 221 + ], + "score": 1.0, + "content": "Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 219, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 116, + 219, + 295, + 231 + ], + "score": 1.0, + "content": "connected networks on graphs. ICLR, 2014.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 238, + 504, + 262 + ], + "lines": [ + { + "bbox": [ + 106, + 238, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 506, + 251 + ], + "score": 1.0, + "content": "Daniele Calandriello, Alessandro Lazaric, and Marcello Restelli. Sparse multi-task reinforcement", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 249, + 460, + 262 + ], + "spans": [ + { + "bbox": [ + 115, + 249, + 460, + 262 + ], + "score": 1.0, + "content": "learning. In Advances in Neural Information Processing Systems, pp. 819–827, 2014.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 269, + 503, + 293 + ], + "lines": [ + { + "bbox": [ + 106, + 269, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 505, + 283 + ], + "score": 1.0, + "content": "Yves Chauvin and David E Rumelhart. Backpropagation: theory, architectures, and applications.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 280, + 216, + 293 + ], + "spans": [ + { + "bbox": [ + 115, + 280, + 216, + 293 + ], + "score": 1.0, + "content": "Psychology Press, 1995.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 105, + 300, + 504, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 300, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 505, + 315 + ], + "score": 1.0, + "content": "Stelian Coros, Philippe Beaudoin, and Michiel Van de Panne. Generalized biped walking control.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 117, + 312, + 344, + 324 + ], + "spans": [ + { + "bbox": [ + 117, + 312, + 344, + 324 + ], + "score": 1.0, + "content": "ACM Transactions on Graphics (TOG), 29(4):130, 2010.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 332, + 503, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 345 + ], + "score": 1.0, + "content": "Alessandro Crespi and Auke Jan Ijspeert. Online optimization of swimming and crawling in an", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 343, + 427, + 355 + ], + "spans": [ + { + "bbox": [ + 116, + 343, + 427, + 355 + ], + "score": 1.0, + "content": "amphibious snake robot. IEEE Transactions on Robotics, 24(1):75–87, 2008.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 363, + 503, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on ¨", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 375, + 353, + 387 + ], + "spans": [ + { + "bbox": [ + 116, + 375, + 353, + 387 + ], + "score": 1.0, + "content": "graphs with fast localized spectral filtering. In NIPS, 2016.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 394, + 504, + 418 + ], + "lines": [ + { + "bbox": [ + 107, + 396, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 107, + 396, + 505, + 407 + ], + "score": 1.0, + "content": "Yan Duan, John Schulman, Xi Chen, Peter L Bartlett, Ilya Sutskever, and Pieter Abbeel. Rl 2: Fast", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 115, + 406, + 505, + 418 + ], + "score": 1.0, + "content": "reinforcement learning via slow reinforcement learning. arXiv preprint arXiv:1611.02779, 2016.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 108, + 426, + 504, + 460 + ], + "lines": [ + { + "bbox": [ + 106, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "David K Duvenaud, Dougal Maclaurin, Jorge Iparraguirre, Rafael Bombarell, Timothy Hirzel, Alan´", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 438, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 116, + 438, + 505, + 450 + ], + "score": 1.0, + "content": "Aspuru-Guzik, and Ryan P Adams. Convolutional networks on graphs for learning molecular", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 448, + 230, + 460 + ], + "spans": [ + { + "bbox": [ + 116, + 448, + 230, + 460 + ], + "score": 1.0, + "content": "fingerprints. In NIPS, 2015.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 104, + 468, + 504, + 491 + ], + "lines": [ + { + "bbox": [ + 106, + 468, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 481 + ], + "score": 1.0, + "content": "Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 480, + 357, + 491 + ], + "spans": [ + { + "bbox": [ + 116, + 480, + 357, + 491 + ], + "score": 1.0, + "content": "of deep networks. arXiv preprint arXiv:1703.03400, 2017a.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 499, + 504, + 523 + ], + "lines": [ + { + "bbox": [ + 107, + 500, + 504, + 512 + ], + "spans": [ + { + "bbox": [ + 107, + 500, + 504, + 512 + ], + "score": 1.0, + "content": "Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imita-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 510, + 412, + 523 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 412, + 523 + ], + "score": 1.0, + "content": "tion learning via meta-learning. arXiv preprint arXiv:1709.04905, 2017b.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 531, + 503, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 531, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 544 + ], + "score": 1.0, + "content": "Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 542, + 444, + 554 + ], + "spans": [ + { + "bbox": [ + 115, + 542, + 444, + 554 + ], + "score": 1.0, + "content": "message passing for quantum chemistry. arXiv preprint arXiv:1704.01212, 2017.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 562, + 504, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 576 + ], + "score": 1.0, + "content": "Christoph Goller and Andreas Kuchler. Learning task-dependent distributed representations by", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 117, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 117, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "backpropagation through structure. In Neural Networks, 1996., IEEE International Conference", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 585, + 280, + 596 + ], + "spans": [ + { + "bbox": [ + 116, + 585, + 280, + 596 + ], + "score": 1.0, + "content": "on, volume 1, pp. 347–352. IEEE, 1996.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 604, + 503, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 618 + ], + "score": 1.0, + "content": "Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 615, + 185, + 627 + ], + "spans": [ + { + "bbox": [ + 116, + 615, + 185, + 627 + ], + "score": 1.0, + "content": "In IJCNN, 2005.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 104, + 636, + 504, + 659 + ], + "lines": [ + { + "bbox": [ + 106, + 636, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 505, + 649 + ], + "score": 1.0, + "content": "Shixiang Gu, Ethan Holly, Timothy Lillicrap, and Sergey Levine. Deep reinforcement learning for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 647, + 422, + 659 + ], + "spans": [ + { + "bbox": [ + 115, + 647, + 422, + 659 + ], + "score": 1.0, + "content": "robotic manipulation with asynchronous off-policy updates. In ICRA, 2017.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 106, + 667, + 504, + 700 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 680 + ], + "score": 1.0, + "content": "Abhishek Gupta, Coline Devin, YuXuan Liu, Pieter Abbeel, and Sergey Levine. Learning invariant", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "feature spaces to transfer skills with reinforcement learning. arXiv preprint arXiv:1703.02949,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 687, + 144, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 687, + 144, + 702 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "Matthew Hausknecht and Peter Stone. Deep reinforcement learning in parameterized action space.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 720, + 279, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 279, + 732 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1511.04143, 2015.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Hagit Attiya and Jennifer Welch. Distributed computing: fundamentals, simulations, and advanced", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 94, + 301, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 301, + 106 + ], + "score": 1.0, + "content": "topics, volume 19. John Wiley & Sons, 2004.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 113, + 503, + 147 + ], + "lines": [ + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "score": 1.0, + "content": "Andre Barreto, Will Dabney, R ´ emi Munos, Jonathan J Hunt, Tom Schaul, David Silver, and Hado P ´", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 125, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 117, + 125, + 505, + 137 + ], + "score": 1.0, + "content": "van Hasselt. Successor features for transfer in reinforcement learning. In Advances in Neural", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 339, + 149 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 339, + 149 + ], + "score": 1.0, + "content": "Information Processing Systems, pp. 4056–4066, 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 106, + 114, + 505, + 149 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 155, + 402, + 168 + ], + "lines": [ + { + "bbox": [ + 106, + 155, + 402, + 169 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 402, + 169 + ], + "score": 1.0, + "content": "Andrew A Biewener. Animal locomotion. Oxford University Press, 2003.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5, + "bbox_fs": [ + 106, + 155, + 402, + 169 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 176, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 106, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 186, + 416, + 201 + ], + "spans": [ + { + "bbox": [ + 116, + 186, + 416, + 201 + ], + "score": 1.0, + "content": "Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 106, + 176, + 505, + 201 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 207, + 504, + 230 + ], + "lines": [ + { + "bbox": [ + 105, + 206, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 505, + 221 + ], + "score": 1.0, + "content": "Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 219, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 116, + 219, + 295, + 231 + ], + "score": 1.0, + "content": "connected networks on graphs. ICLR, 2014.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 206, + 505, + 231 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 238, + 504, + 262 + ], + "lines": [ + { + "bbox": [ + 106, + 238, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 506, + 251 + ], + "score": 1.0, + "content": "Daniele Calandriello, Alessandro Lazaric, and Marcello Restelli. Sparse multi-task reinforcement", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 249, + 460, + 262 + ], + "spans": [ + { + "bbox": [ + 115, + 249, + 460, + 262 + ], + "score": 1.0, + "content": "learning. In Advances in Neural Information Processing Systems, pp. 819–827, 2014.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 106, + 238, + 506, + 262 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 269, + 503, + 293 + ], + "lines": [ + { + "bbox": [ + 106, + 269, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 505, + 283 + ], + "score": 1.0, + "content": "Yves Chauvin and David E Rumelhart. Backpropagation: theory, architectures, and applications.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 280, + 216, + 293 + ], + "spans": [ + { + "bbox": [ + 115, + 280, + 216, + 293 + ], + "score": 1.0, + "content": "Psychology Press, 1995.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 106, + 269, + 505, + 293 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 300, + 504, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 300, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 505, + 315 + ], + "score": 1.0, + "content": "Stelian Coros, Philippe Beaudoin, and Michiel Van de Panne. Generalized biped walking control.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 117, + 312, + 344, + 324 + ], + "spans": [ + { + "bbox": [ + 117, + 312, + 344, + 324 + ], + "score": 1.0, + "content": "ACM Transactions on Graphics (TOG), 29(4):130, 2010.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 300, + 505, + 324 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 332, + 503, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 345 + ], + "score": 1.0, + "content": "Alessandro Crespi and Auke Jan Ijspeert. Online optimization of swimming and crawling in an", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 343, + 427, + 355 + ], + "spans": [ + { + "bbox": [ + 116, + 343, + 427, + 355 + ], + "score": 1.0, + "content": "amphibious snake robot. IEEE Transactions on Robotics, 24(1):75–87, 2008.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 331, + 505, + 355 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 363, + 503, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on ¨", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 375, + 353, + 387 + ], + "spans": [ + { + "bbox": [ + 116, + 375, + 353, + 387 + ], + "score": 1.0, + "content": "graphs with fast localized spectral filtering. In NIPS, 2016.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 363, + 505, + 387 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 394, + 504, + 418 + ], + "lines": [ + { + "bbox": [ + 107, + 396, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 107, + 396, + 505, + 407 + ], + "score": 1.0, + "content": "Yan Duan, John Schulman, Xi Chen, Peter L Bartlett, Ilya Sutskever, and Pieter Abbeel. Rl 2: Fast", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 115, + 406, + 505, + 418 + ], + "score": 1.0, + "content": "reinforcement learning via slow reinforcement learning. arXiv preprint arXiv:1611.02779, 2016.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 107, + 396, + 505, + 418 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 426, + 504, + 460 + ], + "lines": [ + { + "bbox": [ + 106, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "David K Duvenaud, Dougal Maclaurin, Jorge Iparraguirre, Rafael Bombarell, Timothy Hirzel, Alan´", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 438, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 116, + 438, + 505, + 450 + ], + "score": 1.0, + "content": "Aspuru-Guzik, and Ryan P Adams. Convolutional networks on graphs for learning molecular", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 448, + 230, + 460 + ], + "spans": [ + { + "bbox": [ + 116, + 448, + 230, + 460 + ], + "score": 1.0, + "content": "fingerprints. In NIPS, 2015.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 106, + 426, + 505, + 460 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 468, + 504, + 491 + ], + "lines": [ + { + "bbox": [ + 106, + 468, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 481 + ], + "score": 1.0, + "content": "Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 480, + 357, + 491 + ], + "spans": [ + { + "bbox": [ + 116, + 480, + 357, + 491 + ], + "score": 1.0, + "content": "of deep networks. arXiv preprint arXiv:1703.03400, 2017a.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 468, + 505, + 491 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 499, + 504, + 523 + ], + "lines": [ + { + "bbox": [ + 107, + 500, + 504, + 512 + ], + "spans": [ + { + "bbox": [ + 107, + 500, + 504, + 512 + ], + "score": 1.0, + "content": "Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imita-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 510, + 412, + 523 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 412, + 523 + ], + "score": 1.0, + "content": "tion learning via meta-learning. arXiv preprint arXiv:1709.04905, 2017b.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 107, + 500, + 504, + 523 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 531, + 503, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 531, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 544 + ], + "score": 1.0, + "content": "Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 542, + 444, + 554 + ], + "spans": [ + { + "bbox": [ + 115, + 542, + 444, + 554 + ], + "score": 1.0, + "content": "message passing for quantum chemistry. arXiv preprint arXiv:1704.01212, 2017.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 531, + 505, + 554 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 562, + 504, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 576 + ], + "score": 1.0, + "content": "Christoph Goller and Andreas Kuchler. Learning task-dependent distributed representations by", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 117, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 117, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "backpropagation through structure. In Neural Networks, 1996., IEEE International Conference", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 585, + 280, + 596 + ], + "spans": [ + { + "bbox": [ + 116, + 585, + 280, + 596 + ], + "score": 1.0, + "content": "on, volume 1, pp. 347–352. IEEE, 1996.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 561, + 505, + 596 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 604, + 503, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 618 + ], + "score": 1.0, + "content": "Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 615, + 185, + 627 + ], + "spans": [ + { + "bbox": [ + 116, + 615, + 185, + 627 + ], + "score": 1.0, + "content": "In IJCNN, 2005.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 603, + 505, + 627 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 636, + 504, + 659 + ], + "lines": [ + { + "bbox": [ + 106, + 636, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 505, + 649 + ], + "score": 1.0, + "content": "Shixiang Gu, Ethan Holly, Timothy Lillicrap, and Sergey Levine. Deep reinforcement learning for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 647, + 422, + 659 + ], + "spans": [ + { + "bbox": [ + 115, + 647, + 422, + 659 + ], + "score": 1.0, + "content": "robotic manipulation with asynchronous off-policy updates. In ICRA, 2017.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 106, + 636, + 505, + 659 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 667, + 504, + 700 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 680 + ], + "score": 1.0, + "content": "Abhishek Gupta, Coline Devin, YuXuan Liu, Pieter Abbeel, and Sergey Levine. Learning invariant", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "feature spaces to transfer skills with reinforcement learning. arXiv preprint arXiv:1703.02949,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 687, + 144, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 687, + 144, + 702 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 106, + 666, + 505, + 702 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "Matthew Hausknecht and Peter Stone. Deep reinforcement learning in parameterized action space.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 720, + 279, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 279, + 732 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1511.04143, 2015.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Nicolas Heess, Srinivasan Sriram, Jay Lemmon, Josh Merel, Greg Wayne, Yuval Tassa, Tom Erez,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "Ziyu Wang, Ali Eslami, Martin Riedmiller, et al. Emergence of locomotion behaviours in rich", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 339, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 339, + 117 + ], + "score": 1.0, + "content": "environments. arXiv preprint arXiv:1707.02286, 2017.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 124, + 503, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 137 + ], + "score": 1.0, + "content": "Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional net-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 199, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 199, + 146 + ], + "score": 1.0, + "content": "works. ICLR, 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 155, + 505, + 189 + ], + "lines": [ + { + "bbox": [ + 105, + 155, + 504, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 504, + 168 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 165, + 506, + 180 + ], + "spans": [ + { + "bbox": [ + 115, + 165, + 506, + 180 + ], + "score": 1.0, + "content": "lutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 176, + 143, + 190 + ], + "spans": [ + { + "bbox": [ + 114, + 176, + 143, + 190 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 197, + 506, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 196, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 506, + 212 + ], + "score": 1.0, + "content": "Tejas D Kulkarni, Karthik Narasimhan, Ardavan Saeedi, and Josh Tenenbaum. Hierarchical deep", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 209, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 115, + 209, + 506, + 221 + ], + "score": 1.0, + "content": "reinforcement learning: Integrating temporal abstraction and intrinsic motivation. In Advances in", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 219, + 369, + 233 + ], + "spans": [ + { + "bbox": [ + 115, + 219, + 369, + 233 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 3675–3683, 2016.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 109, + 239, + 503, + 273 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 505, + 253 + ], + "score": 1.0, + "content": "Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuo-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 251, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 115, + 251, + 505, + 262 + ], + "score": 1.0, + "content": "motor policies. J. Mach. Learn. Res., 17(1):1334–1373, January 2016. ISSN 1532-4435. URL", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 117, + 263, + 415, + 273 + ], + "spans": [ + { + "bbox": [ + 117, + 263, + 415, + 273 + ], + "score": 1.0, + "content": "http://dl.acm.org/citation.cfm?id=2946645.2946684.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 105, + 281, + 504, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 281, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 506, + 295 + ], + "score": 1.0, + "content": "Ruiyu Li, Makarand Tapaswi, Renjie Liao, Jiaya Jia, Raquel Urtasun, and Sanja Fidler. Situation", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 292, + 305, + 305 + ], + "spans": [ + { + "bbox": [ + 115, + 292, + 305, + 305 + ], + "score": 1.0, + "content": "recognition with graph neural networks. 2017.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 312, + 504, + 335 + ], + "lines": [ + { + "bbox": [ + 106, + 312, + 504, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 504, + 326 + ], + "score": 1.0, + "content": "Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 324, + 321, + 336 + ], + "spans": [ + { + "bbox": [ + 115, + 324, + 321, + 336 + ], + "score": 1.0, + "content": "networks. arXiv preprint arXiv:1511.05493, 2015.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 505, + 377 + ], + "lines": [ + { + "bbox": [ + 105, + 343, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 505, + 357 + ], + "score": 1.0, + "content": "Shingo Mabu, Kotaro Hirasawa, and Jinglu Hu. A graph-based evolutionary algorithm: Genetic", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 356, + 504, + 368 + ], + "spans": [ + { + "bbox": [ + 115, + 356, + 504, + 368 + ], + "score": 1.0, + "content": "network programming (gnp) and its extension using reinforcement learning. Evolutionary Com-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 366, + 244, + 377 + ], + "spans": [ + { + "bbox": [ + 114, + 366, + 244, + 377 + ], + "score": 1.0, + "content": "putation, 15(3):369–398, 2007.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 399 + ], + "score": 1.0, + "content": "Sridhar Mahadevan and Mauro Maggioni. Proto-value functions: A laplacian framework for learn-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 396, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 115, + 396, + 505, + 410 + ], + "score": 1.0, + "content": "ing representation and control in markov decision processes. Journal of Machine Learning Re-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 408, + 251, + 419 + ], + "spans": [ + { + "bbox": [ + 115, + 408, + 251, + 419 + ], + "score": 1.0, + "content": "search, 8(Oct):2169–2231, 2007.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 104, + 428, + 504, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "score": 1.0, + "content": "Luke Metz, Julian Ibarz, Navdeep Jaitly, and James Davidson. Discrete sequential prediction of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 438, + 404, + 451 + ], + "spans": [ + { + "bbox": [ + 115, + 438, + 404, + 451 + ], + "score": 1.0, + "content": "continuous actions for deep rl. arXiv preprint arXiv:1705.05035, 2017.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 108, + 458, + 504, + 493 + ], + "lines": [ + { + "bbox": [ + 106, + 459, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 471 + ], + "score": 1.0, + "content": "Jan Hendrik Metzen. Learning graph-based representations for continuous reinforcement learn-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 115, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "ing domains. In Joint European Conference on Machine Learning and Knowledge Discovery in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 481, + 272, + 493 + ], + "spans": [ + { + "bbox": [ + 116, + 481, + 272, + 493 + ], + "score": 1.0, + "content": "Databases, pp. 81–96. Springer, 2013.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 505, + 535 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 504, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 504, + 514 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Belle-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 116, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "mare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 523, + 441, + 536 + ], + "spans": [ + { + "bbox": [ + 115, + 523, + 441, + 536 + ], + "score": 1.0, + "content": "control through deep reinforcement learning. Nature, 518(7540):529–533, 2015.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 542, + 503, + 566 + ], + "lines": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "Junhyuk Oh, Satinder Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 554, + 440, + 566 + ], + "spans": [ + { + "bbox": [ + 115, + 554, + 440, + 566 + ], + "score": 1.0, + "content": "multi-task deep reinforcement learning. arXiv preprint arXiv:1706.05064, 2017.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 573, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 106, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Actor-mimic: Deep multitask and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 585, + 409, + 597 + ], + "spans": [ + { + "bbox": [ + 115, + 585, + 409, + 597 + ], + "score": 1.0, + "content": "transfer reinforcement learning. arXiv preprint arXiv:1511.06342, 2015.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 108, + 604, + 505, + 639 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "Xiaojuan Qi, Renjie Liao, Jiaya Jia, Sanja Fidler, and Raquel Urtasun. 3d graph neural networks", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 115, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "for rgbd semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 627, + 291, + 640 + ], + "spans": [ + { + "bbox": [ + 116, + 627, + 291, + 640 + ], + "score": 1.0, + "content": "Pattern Recognition, pp. 5199–5208, 2017.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 105, + 646, + 504, + 670 + ], + "lines": [ + { + "bbox": [ + 106, + 647, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 647, + 505, + 660 + ], + "score": 1.0, + "content": "Aravind Rajeswaran, Kendall Lowrey, Emanuel Todorov, and Sham Kakade. Towards generalization", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 658, + 433, + 671 + ], + "spans": [ + { + "bbox": [ + 115, + 658, + 433, + 671 + ], + "score": 1.0, + "content": "and simplicity in continuous control. arXiv preprint arXiv:1703.02660, 2017a.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 105, + 678, + 505, + 701 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "Aravind Rajeswaran, Kendall Lowrey, Emanuel Todorov, and Sham Kakade. Towards generalization", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 690, + 433, + 702 + ], + "spans": [ + { + "bbox": [ + 116, + 690, + 433, + 702 + ], + "score": 1.0, + "content": "and simplicity in continuous control. arXiv preprint arXiv:1703.02660, 2017b.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 720, + 326, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 326, + 733 + ], + "score": 1.0, + "content": "The graph neural network model. IEEE TNN, 2009.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "17", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Nicolas Heess, Srinivasan Sriram, Jay Lemmon, Josh Merel, Greg Wayne, Yuval Tassa, Tom Erez,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "Ziyu Wang, Ali Eslami, Martin Riedmiller, et al. Emergence of locomotion behaviours in rich", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 339, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 339, + 117 + ], + "score": 1.0, + "content": "environments. arXiv preprint arXiv:1707.02286, 2017.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 82, + 505, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 124, + 503, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 137 + ], + "score": 1.0, + "content": "Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional net-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 199, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 199, + 146 + ], + "score": 1.0, + "content": "works. ICLR, 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 123, + 505, + 146 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 155, + 505, + 189 + ], + "lines": [ + { + "bbox": [ + 105, + 155, + 504, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 504, + 168 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 165, + 506, + 180 + ], + "spans": [ + { + "bbox": [ + 115, + 165, + 506, + 180 + ], + "score": 1.0, + "content": "lutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 176, + 143, + 190 + ], + "spans": [ + { + "bbox": [ + 114, + 176, + 143, + 190 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 155, + 506, + 190 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 197, + 506, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 196, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 506, + 212 + ], + "score": 1.0, + "content": "Tejas D Kulkarni, Karthik Narasimhan, Ardavan Saeedi, and Josh Tenenbaum. Hierarchical deep", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 209, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 115, + 209, + 506, + 221 + ], + "score": 1.0, + "content": "reinforcement learning: Integrating temporal abstraction and intrinsic motivation. In Advances in", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 219, + 369, + 233 + ], + "spans": [ + { + "bbox": [ + 115, + 219, + 369, + 233 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 3675–3683, 2016.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 196, + 506, + 233 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 239, + 503, + 273 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 505, + 253 + ], + "score": 1.0, + "content": "Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuo-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 251, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 115, + 251, + 505, + 262 + ], + "score": 1.0, + "content": "motor policies. J. Mach. Learn. Res., 17(1):1334–1373, January 2016. ISSN 1532-4435. URL", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 117, + 263, + 415, + 273 + ], + "spans": [ + { + "bbox": [ + 117, + 263, + 415, + 273 + ], + "score": 1.0, + "content": "http://dl.acm.org/citation.cfm?id=2946645.2946684.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 106, + 239, + 505, + 273 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 281, + 504, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 281, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 506, + 295 + ], + "score": 1.0, + "content": "Ruiyu Li, Makarand Tapaswi, Renjie Liao, Jiaya Jia, Raquel Urtasun, and Sanja Fidler. Situation", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 292, + 305, + 305 + ], + "spans": [ + { + "bbox": [ + 115, + 292, + 305, + 305 + ], + "score": 1.0, + "content": "recognition with graph neural networks. 2017.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 281, + 506, + 305 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 312, + 504, + 335 + ], + "lines": [ + { + "bbox": [ + 106, + 312, + 504, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 504, + 326 + ], + "score": 1.0, + "content": "Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 324, + 321, + 336 + ], + "spans": [ + { + "bbox": [ + 115, + 324, + 321, + 336 + ], + "score": 1.0, + "content": "networks. arXiv preprint arXiv:1511.05493, 2015.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 312, + 504, + 336 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 505, + 377 + ], + "lines": [ + { + "bbox": [ + 105, + 343, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 505, + 357 + ], + "score": 1.0, + "content": "Shingo Mabu, Kotaro Hirasawa, and Jinglu Hu. A graph-based evolutionary algorithm: Genetic", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 356, + 504, + 368 + ], + "spans": [ + { + "bbox": [ + 115, + 356, + 504, + 368 + ], + "score": 1.0, + "content": "network programming (gnp) and its extension using reinforcement learning. Evolutionary Com-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 366, + 244, + 377 + ], + "spans": [ + { + "bbox": [ + 114, + 366, + 244, + 377 + ], + "score": 1.0, + "content": "putation, 15(3):369–398, 2007.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 343, + 505, + 377 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 399 + ], + "score": 1.0, + "content": "Sridhar Mahadevan and Mauro Maggioni. Proto-value functions: A laplacian framework for learn-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 396, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 115, + 396, + 505, + 410 + ], + "score": 1.0, + "content": "ing representation and control in markov decision processes. Journal of Machine Learning Re-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 408, + 251, + 419 + ], + "spans": [ + { + "bbox": [ + 115, + 408, + 251, + 419 + ], + "score": 1.0, + "content": "search, 8(Oct):2169–2231, 2007.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 385, + 505, + 419 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 428, + 504, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "score": 1.0, + "content": "Luke Metz, Julian Ibarz, Navdeep Jaitly, and James Davidson. Discrete sequential prediction of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 438, + 404, + 451 + ], + "spans": [ + { + "bbox": [ + 115, + 438, + 404, + 451 + ], + "score": 1.0, + "content": "continuous actions for deep rl. arXiv preprint arXiv:1705.05035, 2017.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 428, + 506, + 451 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 458, + 504, + 493 + ], + "lines": [ + { + "bbox": [ + 106, + 459, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 471 + ], + "score": 1.0, + "content": "Jan Hendrik Metzen. Learning graph-based representations for continuous reinforcement learn-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 115, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "ing domains. In Joint European Conference on Machine Learning and Knowledge Discovery in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 481, + 272, + 493 + ], + "spans": [ + { + "bbox": [ + 116, + 481, + 272, + 493 + ], + "score": 1.0, + "content": "Databases, pp. 81–96. Springer, 2013.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 459, + 505, + 493 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 505, + 535 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 504, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 504, + 514 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Belle-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 116, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "mare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 523, + 441, + 536 + ], + "spans": [ + { + "bbox": [ + 115, + 523, + 441, + 536 + ], + "score": 1.0, + "content": "control through deep reinforcement learning. Nature, 518(7540):529–533, 2015.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 106, + 500, + 505, + 536 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 542, + 503, + 566 + ], + "lines": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "Junhyuk Oh, Satinder Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 554, + 440, + 566 + ], + "spans": [ + { + "bbox": [ + 115, + 554, + 440, + 566 + ], + "score": 1.0, + "content": "multi-task deep reinforcement learning. arXiv preprint arXiv:1706.05064, 2017.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 542, + 505, + 566 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 573, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 106, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Actor-mimic: Deep multitask and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 585, + 409, + 597 + ], + "spans": [ + { + "bbox": [ + 115, + 585, + 409, + 597 + ], + "score": 1.0, + "content": "transfer reinforcement learning. arXiv preprint arXiv:1511.06342, 2015.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 106, + 574, + 505, + 597 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 604, + 505, + 639 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "Xiaojuan Qi, Renjie Liao, Jiaya Jia, Sanja Fidler, and Raquel Urtasun. 3d graph neural networks", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 115, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "for rgbd semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 627, + 291, + 640 + ], + "spans": [ + { + "bbox": [ + 116, + 627, + 291, + 640 + ], + "score": 1.0, + "content": "Pattern Recognition, pp. 5199–5208, 2017.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 106, + 604, + 505, + 640 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 646, + 504, + 670 + ], + "lines": [ + { + "bbox": [ + 106, + 647, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 647, + 505, + 660 + ], + "score": 1.0, + "content": "Aravind Rajeswaran, Kendall Lowrey, Emanuel Todorov, and Sham Kakade. Towards generalization", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 658, + 433, + 671 + ], + "spans": [ + { + "bbox": [ + 115, + 658, + 433, + 671 + ], + "score": 1.0, + "content": "and simplicity in continuous control. arXiv preprint arXiv:1703.02660, 2017a.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 106, + 647, + 505, + 671 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 678, + 505, + 701 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "Aravind Rajeswaran, Kendall Lowrey, Emanuel Todorov, and Sham Kakade. Towards generalization", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 690, + 433, + 702 + ], + "spans": [ + { + "bbox": [ + 116, + 690, + 433, + 702 + ], + "score": 1.0, + "content": "and simplicity in continuous control. arXiv preprint arXiv:1703.02660, 2017b.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 106, + 678, + 505, + 702 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 720, + 326, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 326, + 733 + ], + "score": 1.0, + "content": "The graph neural network model. IEEE TNN, 2009.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 106, + 709, + 504, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "policy optimization. In Proceedings of the 32nd International Conference on Machine Learning", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 259, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 259, + 117 + ], + "score": 1.0, + "content": "(ICML-15), pp. 1889–1897, 2015a.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 503, + 156 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 135 + ], + "score": 1.0, + "content": "John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 505, + 147 + ], + "score": 1.0, + "content": "dimensional continuous control using generalized advantage estimation. arXiv preprint", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 145, + 224, + 156 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 224, + 156 + ], + "score": 1.0, + "content": "arXiv:1506.02438, 2015b.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 105, + 163, + 504, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 505, + 178 + ], + "score": 1.0, + "content": "John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 175, + 380, + 187 + ], + "spans": [ + { + "bbox": [ + 116, + 175, + 380, + 187 + ], + "score": 1.0, + "content": "optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 503, + 217 + ], + "lines": [ + { + "bbox": [ + 106, + 193, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 505, + 207 + ], + "score": 1.0, + "content": "Farzaneh Shoeleh and Masoud Asadpour. Graph based skill acquisition and transfer learning for", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 205, + 489, + 217 + ], + "spans": [ + { + "bbox": [ + 116, + 205, + 489, + 217 + ], + "score": 1.0, + "content": "continuous reinforcement learning domains. Pattern Recognition Letters, 87:104–116, 2017.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 223, + 505, + 258 + ], + "lines": [ + { + "bbox": [ + 105, + 223, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 237 + ], + "score": 1.0, + "content": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 234, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 234, + 505, + 248 + ], + "score": 1.0, + "content": "Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 246, + 493, + 258 + ], + "spans": [ + { + "bbox": [ + 116, + 246, + 493, + 258 + ], + "score": 1.0, + "content": "the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 109, + 264, + 504, + 298 + ], + "lines": [ + { + "bbox": [ + 106, + 263, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 505, + 279 + ], + "score": 1.0, + "content": "David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 274, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 116, + 274, + 506, + 290 + ], + "score": 1.0, + "content": "Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 117, + 287, + 370, + 299 + ], + "spans": [ + { + "bbox": [ + 117, + 287, + 370, + 299 + ], + "score": 1.0, + "content": "without human knowledge. Nature, 550(7676):354–359, 2017.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 305, + 504, + 340 + ], + "lines": [ + { + "bbox": [ + 105, + 304, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 506, + 318 + ], + "score": 1.0, + "content": "Richard Socher, Cliff C Lin, Chris Manning, and Andrew Y Ng. Parsing natural scenes and natural", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 317, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 116, + 317, + 505, + 329 + ], + "score": 1.0, + "content": "language with recursive neural networks. In Proceedings of the 28th international conference on", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 328, + 317, + 339 + ], + "spans": [ + { + "bbox": [ + 116, + 328, + 317, + 339 + ], + "score": 1.0, + "content": "machine learning (ICML-11), pp. 129–136, 2011.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 346, + 504, + 369 + ], + "lines": [ + { + "bbox": [ + 105, + 344, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 360 + ], + "score": 1.0, + "content": "Mark W Spong, Seth Hutchinson, Mathukumalli Vidyasagar, et al. Robot modeling and control,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 357, + 256, + 370 + ], + "spans": [ + { + "bbox": [ + 116, + 357, + 256, + 370 + ], + "score": 1.0, + "content": "volume 3. Wiley New York, 2006.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 375, + 504, + 410 + ], + "lines": [ + { + "bbox": [ + 106, + 375, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 505, + 389 + ], + "score": 1.0, + "content": "Richard S Sutton, David A McAllester, Satinder P Singh, and Yishay Mansour. Policy gradient", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 387, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 115, + 387, + 505, + 400 + ], + "score": 1.0, + "content": "methods for reinforcement learning with function approximation. In Advances in neural informa-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 398, + 305, + 411 + ], + "spans": [ + { + "bbox": [ + 116, + 398, + 305, + 411 + ], + "score": 1.0, + "content": "tion processing systems, pp. 1057–1063, 2000.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 105, + 417, + 504, + 440 + ], + "lines": [ + { + "bbox": [ + 106, + 416, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 430 + ], + "score": 1.0, + "content": "Kai Sheng Tai, Richard Socher, and Christopher D Manning. Improved semantic representations", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 428, + 389, + 440 + ], + "spans": [ + { + "bbox": [ + 115, + 428, + 389, + 440 + ], + "score": 1.0, + "content": "from tree-structured long short-term memory networks. ACL, 2015.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 104, + 447, + 504, + 470 + ], + "lines": [ + { + "bbox": [ + 105, + 445, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 461 + ], + "score": 1.0, + "content": "Matthew E Taylor and Peter Stone. Transfer learning for reinforcement learning domains: A survey.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 458, + 384, + 470 + ], + "spans": [ + { + "bbox": [ + 115, + 458, + 384, + 470 + ], + "score": 1.0, + "content": "Journal of Machine Learning Research, 10(Jul):1633–1685, 2009.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 109, + 477, + 504, + 511 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 490 + ], + "score": 1.0, + "content": "Yee Teh, Victor Bapst, Razvan Pascanu, Nicolas Heess, John Quan, James Kirkpatrick, Wojciech M", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 487, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 115, + 487, + 506, + 501 + ], + "score": 1.0, + "content": "Czarnecki, and Raia Hadsell. Distral: Robust multitask reinforcement learning. In Advances in", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 499, + 369, + 511 + ], + "spans": [ + { + "bbox": [ + 116, + 499, + 369, + 511 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 4497–4507, 2017.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "score": 1.0, + "content": "Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 115, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "In Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ International Conference on, pp. 5026–", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 540, + 193, + 551 + ], + "spans": [ + { + "bbox": [ + 116, + 540, + 193, + 551 + ], + "score": 1.0, + "content": "5033. IEEE, 2012.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 558, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 106, + 558, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 571 + ], + "score": 1.0, + "content": "Alexander Sasha Vezhnevets, Simon Osindero, Tom Schaul, Nicolas Heess, Max Jaderberg, David", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 116, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "Silver, and Koray Kavukcuoglu. Feudal networks for hierarchical reinforcement learning. arXiv", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 581, + 254, + 592 + ], + "spans": [ + { + "bbox": [ + 115, + 581, + 254, + 592 + ], + "score": 1.0, + "content": "preprint arXiv:1703.01161, 2017.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 599, + 504, + 633 + ], + "lines": [ + { + "bbox": [ + 105, + 598, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 506, + 613 + ], + "score": 1.0, + "content": "Aaron Wilson, Alan Fern, Soumya Ray, and Prasad Tadepalli. Multi-task reinforcement learning: a", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 116, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "hierarchical bayesian approach. In Proceedings of the 24th international conference on Machine", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 621, + 272, + 633 + ], + "spans": [ + { + "bbox": [ + 116, + 621, + 272, + 633 + ], + "score": 1.0, + "content": "learning, pp. 1015–1022. ACM, 2007.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 640, + 503, + 663 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "score": 1.0, + "content": "Markus Wulfmeier, Ingmar Posner, and Pieter Abbeel. Mutual alignment transfer learning. arXiv", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 651, + 254, + 663 + ], + "spans": [ + { + "bbox": [ + 115, + 651, + 254, + 663 + ], + "score": 1.0, + "content": "preprint arXiv:1707.07907, 2017.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 670, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 684 + ], + "score": 1.0, + "content": "Zhaoyang Yang, Kathryn Merrick, Hussein Abbass, and Lianwen Jin. Multi-task deep reinforcement", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 117, + 682, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 117, + 682, + 505, + 693 + ], + "score": 1.0, + "content": "learning for continuous action control. In Proceedings of the Twenty-Sixth International Joint", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 692, + 399, + 705 + ], + "spans": [ + { + "bbox": [ + 116, + 692, + 399, + 705 + ], + "score": 1.0, + "content": "Conference on Artificial Intelligence, IJCAI-17, pp. 3301–3307, 2017.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "policy optimization. In Proceedings of the 32nd International Conference on Machine Learning", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 259, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 259, + 117 + ], + "score": 1.0, + "content": "(ICML-15), pp. 1889–1897, 2015a.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 506, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 503, + 156 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 135 + ], + "score": 1.0, + "content": "John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 505, + 147 + ], + "score": 1.0, + "content": "dimensional continuous control using generalized advantage estimation. arXiv preprint", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 145, + 224, + 156 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 224, + 156 + ], + "score": 1.0, + "content": "arXiv:1506.02438, 2015b.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 123, + 505, + 156 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 163, + 504, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 505, + 178 + ], + "score": 1.0, + "content": "John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 175, + 380, + 187 + ], + "spans": [ + { + "bbox": [ + 116, + 175, + 380, + 187 + ], + "score": 1.0, + "content": "optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 162, + 505, + 187 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 503, + 217 + ], + "lines": [ + { + "bbox": [ + 106, + 193, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 505, + 207 + ], + "score": 1.0, + "content": "Farzaneh Shoeleh and Masoud Asadpour. Graph based skill acquisition and transfer learning for", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 205, + 489, + 217 + ], + "spans": [ + { + "bbox": [ + 116, + 205, + 489, + 217 + ], + "score": 1.0, + "content": "continuous reinforcement learning domains. Pattern Recognition Letters, 87:104–116, 2017.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 106, + 193, + 505, + 217 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 223, + 505, + 258 + ], + "lines": [ + { + "bbox": [ + 105, + 223, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 237 + ], + "score": 1.0, + "content": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 234, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 234, + 505, + 248 + ], + "score": 1.0, + "content": "Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 246, + 493, + 258 + ], + "spans": [ + { + "bbox": [ + 116, + 246, + 493, + 258 + ], + "score": 1.0, + "content": "the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 223, + 505, + 258 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 264, + 504, + 298 + ], + "lines": [ + { + "bbox": [ + 106, + 263, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 505, + 279 + ], + "score": 1.0, + "content": "David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 274, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 116, + 274, + 506, + 290 + ], + "score": 1.0, + "content": "Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 117, + 287, + 370, + 299 + ], + "spans": [ + { + "bbox": [ + 117, + 287, + 370, + 299 + ], + "score": 1.0, + "content": "without human knowledge. Nature, 550(7676):354–359, 2017.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 106, + 263, + 506, + 299 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 305, + 504, + 340 + ], + "lines": [ + { + "bbox": [ + 105, + 304, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 506, + 318 + ], + "score": 1.0, + "content": "Richard Socher, Cliff C Lin, Chris Manning, and Andrew Y Ng. Parsing natural scenes and natural", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 317, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 116, + 317, + 505, + 329 + ], + "score": 1.0, + "content": "language with recursive neural networks. In Proceedings of the 28th international conference on", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 328, + 317, + 339 + ], + "spans": [ + { + "bbox": [ + 116, + 328, + 317, + 339 + ], + "score": 1.0, + "content": "machine learning (ICML-11), pp. 129–136, 2011.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 304, + 506, + 339 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 346, + 504, + 369 + ], + "lines": [ + { + "bbox": [ + 105, + 344, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 360 + ], + "score": 1.0, + "content": "Mark W Spong, Seth Hutchinson, Mathukumalli Vidyasagar, et al. Robot modeling and control,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 357, + 256, + 370 + ], + "spans": [ + { + "bbox": [ + 116, + 357, + 256, + 370 + ], + "score": 1.0, + "content": "volume 3. Wiley New York, 2006.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 344, + 505, + 370 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 375, + 504, + 410 + ], + "lines": [ + { + "bbox": [ + 106, + 375, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 505, + 389 + ], + "score": 1.0, + "content": "Richard S Sutton, David A McAllester, Satinder P Singh, and Yishay Mansour. Policy gradient", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 387, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 115, + 387, + 505, + 400 + ], + "score": 1.0, + "content": "methods for reinforcement learning with function approximation. In Advances in neural informa-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 398, + 305, + 411 + ], + "spans": [ + { + "bbox": [ + 116, + 398, + 305, + 411 + ], + "score": 1.0, + "content": "tion processing systems, pp. 1057–1063, 2000.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 375, + 505, + 411 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 417, + 504, + 440 + ], + "lines": [ + { + "bbox": [ + 106, + 416, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 430 + ], + "score": 1.0, + "content": "Kai Sheng Tai, Richard Socher, and Christopher D Manning. Improved semantic representations", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 428, + 389, + 440 + ], + "spans": [ + { + "bbox": [ + 115, + 428, + 389, + 440 + ], + "score": 1.0, + "content": "from tree-structured long short-term memory networks. ACL, 2015.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 106, + 416, + 505, + 440 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 447, + 504, + 470 + ], + "lines": [ + { + "bbox": [ + 105, + 445, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 461 + ], + "score": 1.0, + "content": "Matthew E Taylor and Peter Stone. Transfer learning for reinforcement learning domains: A survey.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 458, + 384, + 470 + ], + "spans": [ + { + "bbox": [ + 115, + 458, + 384, + 470 + ], + "score": 1.0, + "content": "Journal of Machine Learning Research, 10(Jul):1633–1685, 2009.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 445, + 505, + 470 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 477, + 504, + 511 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 490 + ], + "score": 1.0, + "content": "Yee Teh, Victor Bapst, Razvan Pascanu, Nicolas Heess, John Quan, James Kirkpatrick, Wojciech M", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 487, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 115, + 487, + 506, + 501 + ], + "score": 1.0, + "content": "Czarnecki, and Raia Hadsell. Distral: Robust multitask reinforcement learning. In Advances in", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 499, + 369, + 511 + ], + "spans": [ + { + "bbox": [ + 116, + 499, + 369, + 511 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 4497–4507, 2017.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 106, + 476, + 506, + 511 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "score": 1.0, + "content": "Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 115, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "In Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ International Conference on, pp. 5026–", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 540, + 193, + 551 + ], + "spans": [ + { + "bbox": [ + 116, + 540, + 193, + 551 + ], + "score": 1.0, + "content": "5033. IEEE, 2012.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 516, + 505, + 551 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 558, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 106, + 558, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 571 + ], + "score": 1.0, + "content": "Alexander Sasha Vezhnevets, Simon Osindero, Tom Schaul, Nicolas Heess, Max Jaderberg, David", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 116, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "Silver, and Koray Kavukcuoglu. Feudal networks for hierarchical reinforcement learning. arXiv", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 581, + 254, + 592 + ], + "spans": [ + { + "bbox": [ + 115, + 581, + 254, + 592 + ], + "score": 1.0, + "content": "preprint arXiv:1703.01161, 2017.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 106, + 558, + 505, + 592 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 599, + 504, + 633 + ], + "lines": [ + { + "bbox": [ + 105, + 598, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 506, + 613 + ], + "score": 1.0, + "content": "Aaron Wilson, Alan Fern, Soumya Ray, and Prasad Tadepalli. Multi-task reinforcement learning: a", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 116, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "hierarchical bayesian approach. In Proceedings of the 24th international conference on Machine", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 621, + 272, + 633 + ], + "spans": [ + { + "bbox": [ + 116, + 621, + 272, + 633 + ], + "score": 1.0, + "content": "learning, pp. 1015–1022. ACM, 2007.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 598, + 506, + 633 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 640, + 503, + 663 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "score": 1.0, + "content": "Markus Wulfmeier, Ingmar Posner, and Pieter Abbeel. Mutual alignment transfer learning. arXiv", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 651, + 254, + 663 + ], + "spans": [ + { + "bbox": [ + 115, + 651, + 254, + 663 + ], + "score": 1.0, + "content": "preprint arXiv:1707.07907, 2017.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 640, + 505, + 663 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 670, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 684 + ], + "score": 1.0, + "content": "Zhaoyang Yang, Kathryn Merrick, Hussein Abbass, and Lianwen Jin. Multi-task deep reinforcement", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 117, + 682, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 117, + 682, + 505, + 693 + ], + "score": 1.0, + "content": "learning for continuous action control. In Proceedings of the Twenty-Sixth International Joint", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 692, + 399, + 705 + ], + "spans": [ + { + "bbox": [ + 116, + 692, + 399, + 705 + ], + "score": 1.0, + "content": "Conference on Artificial Intelligence, IJCAI-17, pp. 3301–3307, 2017.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 669, + 506, + 705 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 180, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 182, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 182, + 96 + ], + "score": 1.0, + "content": "6 APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 105, + 234, + 117 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 235, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 235, + 119 + ], + "score": 1.0, + "content": "6.1 DETAILS OF NERVENET", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 126, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 106, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "We use MLP to compute the messages which uses tanh nonlinearities as the activation function.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "We do a grid search on the size of the MLP to compute the messages, the details of which are listed", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 147, + 162, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 162, + 160 + ], + "score": 1.0, + "content": "in Table 5, 4.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 104, + 165, + 492, + 177 + ], + "lines": [ + { + "bbox": [ + 106, + 163, + 493, + 179 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 493, + 179 + ], + "score": 1.0, + "content": "Throughout all of our experiments, we use average aggregation and GRU as the update function.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 146, + 207, + 464, + 265 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 222, + 187, + 388, + 199 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 221, + 185, + 389, + 201 + ], + "spans": [ + { + "bbox": [ + 221, + 185, + 389, + 201 + ], + "score": 1.0, + "content": "Table 4: Parameters used during training.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 146, + 207, + 464, + 265 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 146, + 207, + 464, + 265 + ], + "spans": [ + { + "bbox": [ + 146, + 207, + 464, + 265 + ], + "score": 0.981, + "html": "
ParametersValue SetParameters Value Set
Value Discount Factor y.99GAE入.95
PPO Clip Value0.2Starting Learning Rate3e-4
Gradient Clip Value5.0Target KL0.01
", + "type": "table", + "image_path": "cd0f977cb6b65e68b1f84e18ebc157e379e0f5aa01cfbb48a18959ac79d15233.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 146, + 207, + 464, + 226.33333333333334 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 146, + 226.33333333333334, + 464, + 245.66666666666669 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 146, + 245.66666666666669, + 464, + 265.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 7.0 + }, + { + "type": "title", + "bbox": [ + 107, + 289, + 209, + 300 + ], + "lines": [ + { + "bbox": [ + 106, + 288, + 211, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 211, + 302 + ], + "score": 1.0, + "content": "6.2 GRAPH OF AGENT", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 108, + 309, + 505, + 343 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "In MuJoCo, we observe that most body nodes are paired with one and only one joint node. Thus,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "score": 1.0, + "content": "we simply merge the two paired nodes into one. We point out that this model is very compact, and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 332, + 302, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 302, + 344 + ], + "score": 1.0, + "content": "is the standard graph we use in our experiments.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 349, + 505, + 405 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 504, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 504, + 361 + ], + "score": 1.0, + "content": "In the Gym environments, observation for the joint nodes normally includes the angular veloc-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 360, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 506, + 372 + ], + "score": 1.0, + "content": "ity, twist angle and optionally the torque for the hinge joint, and position information for the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "score": 1.0, + "content": "positional joint. For the body nodes, velocity, inertia, and force are common observations. For", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 379, + 507, + 397 + ], + "spans": [ + { + "bbox": [ + 104, + 379, + 487, + 397 + ], + "score": 1.0, + "content": "example in the centipede environment 1, the LeftHip node will receive the angular velocity", + "type": "text" + }, + { + "bbox": [ + 488, + 384, + 501, + 394 + ], + "score": 0.86, + "content": "\\varpi _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 379, + 507, + 397 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 392, + 198, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 183, + 407 + ], + "score": 1.0, + "content": "and the twist angle", + "type": "text" + }, + { + "bbox": [ + 184, + 393, + 194, + 405 + ], + "score": 0.88, + "content": "\\theta _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 392, + 198, + 407 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 108, + 417, + 252, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 417, + 253, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 253, + 430 + ], + "score": 1.0, + "content": "6.3 HYPERPARAMETER SEARCH", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 437, + 505, + 505 + ], + "lines": [ + { + "bbox": [ + 106, + 438, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 450 + ], + "score": 1.0, + "content": "For MLP, we run grid search with the hidden size from two layers to three layers, and with hidden", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "score": 1.0, + "content": "size from 32 to 256. For NerveNet, to reduce the time spent on grid search, we constrain the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 460, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 472 + ], + "score": 1.0, + "content": "propagation network and output network to be the same shape. Similarly, we run grid search with", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "the network’s hidden size, and at the same time, we run a grid search on the size of node’s hidden", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 483, + 506, + 494 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 506, + 494 + ], + "score": 1.0, + "content": "states from 32 to 64. For the TreeNet, we run similar grid search on the node’s hidden states and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 493, + 204, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 204, + 506 + ], + "score": 1.0, + "content": "output network’s shape.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 510, + 353, + 522 + ], + "lines": [ + { + "bbox": [ + 106, + 509, + 354, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 354, + 522 + ], + "score": 1.0, + "content": "For details of hyperparameter search, please see Table 5, 7, 6.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "table", + "bbox": [ + 145, + 552, + 466, + 621 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 196, + 531, + 415, + 543 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 195, + 531, + 415, + 544 + ], + "spans": [ + { + "bbox": [ + 195, + 531, + 415, + 544 + ], + "score": 1.0, + "content": "Table 5: Hyperparameter grid search options for MLP.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "table_body", + "bbox": [ + 145, + 552, + 466, + 621 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 145, + 552, + 466, + 621 + ], + "spans": [ + { + "bbox": [ + 145, + 552, + 466, + 621 + ], + "score": 0.977, + "html": "
MLPValue Tried
Network Shape[64,64],[128,128], [256,256],[64,64,64]
Number of Iteration Per Update10,20
Use KL PenaltyYes, No
Learning Rate SchedulerLinear Decay, Adaptive, Constant
", + "type": "table", + "image_path": "aadeeb33b22fb2a35899a91dd3d3759eb93db7322b47f9e4482bc890a3f03f0f.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 145, + 552, + 466, + 575.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 145, + 575.0, + 466, + 598.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 145, + 598.0, + 466, + 621.0 + ], + "spans": [], + "index": 30 + } + ] + } + ], + "index": 28.0 + }, + { + "type": "table", + "bbox": [ + 161, + 657, + 449, + 725 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 189, + 636, + 421, + 648 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 189, + 635, + 422, + 649 + ], + "spans": [ + { + "bbox": [ + 189, + 635, + 422, + 649 + ], + "score": 1.0, + "content": "Table 6: Hyperparameter grid search options for TreeNet.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "table_body", + "bbox": [ + 161, + 657, + 449, + 725 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 161, + 657, + 449, + 725 + ], + "spans": [ + { + "bbox": [ + 161, + 657, + 449, + 725 + ], + "score": 0.983, + "html": "
TreeNetValue Tried
Network Shape[64,64], [128,128], [256,256]
Number of Iteration Per Update10,20
Use KL PenaltyYes,No
Learning Rate SchedulerLinear Decay, Adaptive, Constant
", + "type": "table", + "image_path": "54379f7c00f8dfbccabd71c68389ee3e2d2fd69249803c8b5cb5dad8f723c3e8.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 161, + 657, + 449, + 679.6666666666666 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 161, + 679.6666666666666, + 449, + 702.3333333333333 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 161, + 702.3333333333333, + 449, + 724.9999999999999 + ], + "spans": [], + "index": 34 + } + ] + } + ], + "index": 32.0 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 180, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 182, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 182, + 96 + ], + "score": 1.0, + "content": "6 APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 105, + 234, + 117 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 235, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 235, + 119 + ], + "score": 1.0, + "content": "6.1 DETAILS OF NERVENET", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 126, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 106, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "We use MLP to compute the messages which uses tanh nonlinearities as the activation function.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "We do a grid search on the size of the MLP to compute the messages, the details of which are listed", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 147, + 162, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 162, + 160 + ], + "score": 1.0, + "content": "in Table 5, 4.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 126, + 505, + 160 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 165, + 492, + 177 + ], + "lines": [ + { + "bbox": [ + 106, + 163, + 493, + 179 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 493, + 179 + ], + "score": 1.0, + "content": "Throughout all of our experiments, we use average aggregation and GRU as the update function.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5, + "bbox_fs": [ + 106, + 163, + 493, + 179 + ] + }, + { + "type": "table", + "bbox": [ + 146, + 207, + 464, + 265 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 222, + 187, + 388, + 199 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 221, + 185, + 389, + 201 + ], + "spans": [ + { + "bbox": [ + 221, + 185, + 389, + 201 + ], + "score": 1.0, + "content": "Table 4: Parameters used during training.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 146, + 207, + 464, + 265 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 146, + 207, + 464, + 265 + ], + "spans": [ + { + "bbox": [ + 146, + 207, + 464, + 265 + ], + "score": 0.981, + "html": "
ParametersValue SetParameters Value Set
Value Discount Factor y.99GAE入.95
PPO Clip Value0.2Starting Learning Rate3e-4
Gradient Clip Value5.0Target KL0.01
", + "type": "table", + "image_path": "cd0f977cb6b65e68b1f84e18ebc157e379e0f5aa01cfbb48a18959ac79d15233.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 146, + 207, + 464, + 226.33333333333334 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 146, + 226.33333333333334, + 464, + 245.66666666666669 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 146, + 245.66666666666669, + 464, + 265.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 7.0 + }, + { + "type": "title", + "bbox": [ + 107, + 289, + 209, + 300 + ], + "lines": [ + { + "bbox": [ + 106, + 288, + 211, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 211, + 302 + ], + "score": 1.0, + "content": "6.2 GRAPH OF AGENT", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 108, + 309, + 505, + 343 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "In MuJoCo, we observe that most body nodes are paired with one and only one joint node. Thus,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "score": 1.0, + "content": "we simply merge the two paired nodes into one. We point out that this model is very compact, and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 332, + 302, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 302, + 344 + ], + "score": 1.0, + "content": "is the standard graph we use in our experiments.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 309, + 505, + 344 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 349, + 505, + 405 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 504, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 504, + 361 + ], + "score": 1.0, + "content": "In the Gym environments, observation for the joint nodes normally includes the angular veloc-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 360, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 506, + 372 + ], + "score": 1.0, + "content": "ity, twist angle and optionally the torque for the hinge joint, and position information for the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "score": 1.0, + "content": "positional joint. For the body nodes, velocity, inertia, and force are common observations. For", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 379, + 507, + 397 + ], + "spans": [ + { + "bbox": [ + 104, + 379, + 487, + 397 + ], + "score": 1.0, + "content": "example in the centipede environment 1, the LeftHip node will receive the angular velocity", + "type": "text" + }, + { + "bbox": [ + 488, + 384, + 501, + 394 + ], + "score": 0.86, + "content": "\\varpi _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 379, + 507, + 397 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 392, + 198, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 183, + 407 + ], + "score": 1.0, + "content": "and the twist angle", + "type": "text" + }, + { + "bbox": [ + 184, + 393, + 194, + 405 + ], + "score": 0.88, + "content": "\\theta _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 392, + 198, + 407 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16, + "bbox_fs": [ + 104, + 349, + 507, + 407 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 417, + 252, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 417, + 253, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 253, + 430 + ], + "score": 1.0, + "content": "6.3 HYPERPARAMETER SEARCH", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 437, + 505, + 505 + ], + "lines": [ + { + "bbox": [ + 106, + 438, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 450 + ], + "score": 1.0, + "content": "For MLP, we run grid search with the hidden size from two layers to three layers, and with hidden", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "score": 1.0, + "content": "size from 32 to 256. For NerveNet, to reduce the time spent on grid search, we constrain the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 460, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 472 + ], + "score": 1.0, + "content": "propagation network and output network to be the same shape. Similarly, we run grid search with", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "the network’s hidden size, and at the same time, we run a grid search on the size of node’s hidden", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 483, + 506, + 494 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 506, + 494 + ], + "score": 1.0, + "content": "states from 32 to 64. For the TreeNet, we run similar grid search on the node’s hidden states and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 493, + 204, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 204, + 506 + ], + "score": 1.0, + "content": "output network’s shape.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 438, + 506, + 506 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 510, + 353, + 522 + ], + "lines": [ + { + "bbox": [ + 106, + 509, + 354, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 354, + 522 + ], + "score": 1.0, + "content": "For details of hyperparameter search, please see Table 5, 7, 6.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 106, + 509, + 354, + 522 + ] + }, + { + "type": "table", + "bbox": [ + 145, + 552, + 466, + 621 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 196, + 531, + 415, + 543 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 195, + 531, + 415, + 544 + ], + "spans": [ + { + "bbox": [ + 195, + 531, + 415, + 544 + ], + "score": 1.0, + "content": "Table 5: Hyperparameter grid search options for MLP.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "table_body", + "bbox": [ + 145, + 552, + 466, + 621 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 145, + 552, + 466, + 621 + ], + "spans": [ + { + "bbox": [ + 145, + 552, + 466, + 621 + ], + "score": 0.977, + "html": "
MLPValue Tried
Network Shape[64,64],[128,128], [256,256],[64,64,64]
Number of Iteration Per Update10,20
Use KL PenaltyYes, No
Learning Rate SchedulerLinear Decay, Adaptive, Constant
", + "type": "table", + "image_path": "aadeeb33b22fb2a35899a91dd3d3759eb93db7322b47f9e4482bc890a3f03f0f.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 145, + 552, + 466, + 575.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 145, + 575.0, + 466, + 598.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 145, + 598.0, + 466, + 621.0 + ], + "spans": [], + "index": 30 + } + ] + } + ], + "index": 28.0 + }, + { + "type": "table", + "bbox": [ + 161, + 657, + 449, + 725 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 189, + 636, + 421, + 648 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 189, + 635, + 422, + 649 + ], + "spans": [ + { + "bbox": [ + 189, + 635, + 422, + 649 + ], + "score": 1.0, + "content": "Table 6: Hyperparameter grid search options for TreeNet.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "table_body", + "bbox": [ + 161, + 657, + 449, + 725 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 161, + 657, + 449, + 725 + ], + "spans": [ + { + "bbox": [ + 161, + 657, + 449, + 725 + ], + "score": 0.983, + "html": "
TreeNetValue Tried
Network Shape[64,64], [128,128], [256,256]
Number of Iteration Per Update10,20
Use KL PenaltyYes,No
Learning Rate SchedulerLinear Decay, Adaptive, Constant
", + "type": "table", + "image_path": "54379f7c00f8dfbccabd71c68389ee3e2d2fd69249803c8b5cb5dad8f723c3e8.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 161, + 657, + 449, + 679.6666666666666 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 161, + 679.6666666666666, + 449, + 702.3333333333333 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 161, + 702.3333333333333, + 449, + 724.9999999999999 + ], + "spans": [], + "index": 34 + } + ] + } + ], + "index": 32.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 101, + 476, + 246 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 186, + 80, + 424, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 186, + 79, + 424, + 94 + ], + "spans": [ + { + "bbox": [ + 186, + 79, + 424, + 94 + ], + "score": 1.0, + "content": "Table 7: Hyperparameter grid search options for NerveNet.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 135, + 101, + 476, + 246 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 101, + 476, + 246 + ], + "spans": [ + { + "bbox": [ + 135, + 101, + 476, + 246 + ], + "score": 0.984, + "html": "
NerveNetValue Tried
Network Shape[64,64],[128,128],[256, 256]
Number of Iteration Per Update10,20
Use KL PenaltyYes,No
Learning Rate SchedulerLinear Decay, Adaptive, Constant
Number of Propogation Steps3,4,5,6
NerveNetVariantsNerveNet-1,NerveNet-2,NerveNet-MLP
Size ofNodes'Hidden State32,64,128
Merge joint and body NodeYes,No
Output NetworkShared, Separate
Disable Edge TypeYes, No
Add Skip-connection from / to rootYes,No
", + "type": "table", + "image_path": "be8567db64be48d464b8a43cd7cecfa1a64df314fa0bf6ed3e82da8181a39196.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 135, + 101, + 476, + 149.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 135, + 149.33333333333334, + 476, + 197.66666666666669 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 135, + 197.66666666666669, + 476, + 246.00000000000003 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "title", + "bbox": [ + 106, + 266, + 331, + 277 + ], + "lines": [ + { + "bbox": [ + 106, + 266, + 331, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 331, + 278 + ], + "score": 1.0, + "content": "6.4 SCHEMATIC FIGURES OF THE PARSED AGENTS", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 286, + 504, + 309 + ], + "lines": [ + { + "bbox": [ + 105, + 286, + 504, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 504, + 299 + ], + "score": 1.0, + "content": "In this section, we also plot the schematic figures of the agents for readers’ reference. Graph struc-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 297, + 403, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 403, + 310 + ], + "score": 1.0, + "content": "tures are automatically parsed from the MuJoCo XML configuration files.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "image", + "bbox": [ + 129, + 330, + 484, + 461 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 129, + 330, + 484, + 461 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 330, + 484, + 461 + ], + "spans": [ + { + "bbox": [ + 129, + 330, + 484, + 461 + ], + "score": 0.967, + "type": "image", + "image_path": "e7bb923f2aed9a6363a3a02049afe2ed3ae65c849658c6e2bd754b853c8d9f19.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 129, + 330, + 484, + 373.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 129, + 373.6666666666667, + 484, + 417.33333333333337 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 129, + 417.33333333333337, + 484, + 461.00000000000006 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 471, + 505, + 493 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 471, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 505, + 484 + ], + "score": 1.0, + "content": "Figure 12: Schematic diagrams and auto-parsed graph structures of the InvertedPendulum and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 483, + 243, + 494 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 243, + 494 + ], + "score": 1.0, + "content": "InvertedDoublePendulum.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + } + ], + "index": 9.25 + }, + { + "type": "image", + "bbox": [ + 119, + 524, + 491, + 632 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 524, + 491, + 632 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 119, + 524, + 491, + 632 + ], + "spans": [ + { + "bbox": [ + 119, + 524, + 491, + 632 + ], + "score": 0.966, + "type": "image", + "image_path": "3ecaa892fd0c5dde6ad0cbe96e739a085bbe439b8b996b0c352579ba4ec09a38.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 119, + 524, + 491, + 560.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 119, + 560.0, + 491, + 596.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 119, + 596.0, + 491, + 632.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 641, + 504, + 653 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 653 + ], + "score": 1.0, + "content": "Figure 13: Schematic diagrams and auto-parsed graph structures of the Walker2d and Reacher.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + } + ], + "index": 14.0 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 101, + 476, + 246 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 186, + 80, + 424, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 186, + 79, + 424, + 94 + ], + "spans": [ + { + "bbox": [ + 186, + 79, + 424, + 94 + ], + "score": 1.0, + "content": "Table 7: Hyperparameter grid search options for NerveNet.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 135, + 101, + 476, + 246 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 101, + 476, + 246 + ], + "spans": [ + { + "bbox": [ + 135, + 101, + 476, + 246 + ], + "score": 0.984, + "html": "
NerveNetValue Tried
Network Shape[64,64],[128,128],[256, 256]
Number of Iteration Per Update10,20
Use KL PenaltyYes,No
Learning Rate SchedulerLinear Decay, Adaptive, Constant
Number of Propogation Steps3,4,5,6
NerveNetVariantsNerveNet-1,NerveNet-2,NerveNet-MLP
Size ofNodes'Hidden State32,64,128
Merge joint and body NodeYes,No
Output NetworkShared, Separate
Disable Edge TypeYes, No
Add Skip-connection from / to rootYes,No
", + "type": "table", + "image_path": "be8567db64be48d464b8a43cd7cecfa1a64df314fa0bf6ed3e82da8181a39196.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 135, + 101, + 476, + 149.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 135, + 149.33333333333334, + 476, + 197.66666666666669 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 135, + 197.66666666666669, + 476, + 246.00000000000003 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "title", + "bbox": [ + 106, + 266, + 331, + 277 + ], + "lines": [ + { + "bbox": [ + 106, + 266, + 331, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 331, + 278 + ], + "score": 1.0, + "content": "6.4 SCHEMATIC FIGURES OF THE PARSED AGENTS", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 286, + 504, + 309 + ], + "lines": [ + { + "bbox": [ + 105, + 286, + 504, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 504, + 299 + ], + "score": 1.0, + "content": "In this section, we also plot the schematic figures of the agents for readers’ reference. Graph struc-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 297, + 403, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 403, + 310 + ], + "score": 1.0, + "content": "tures are automatically parsed from the MuJoCo XML configuration files.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 286, + 504, + 310 + ] + }, + { + "type": "image", + "bbox": [ + 129, + 330, + 484, + 461 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 129, + 330, + 484, + 461 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 330, + 484, + 461 + ], + "spans": [ + { + "bbox": [ + 129, + 330, + 484, + 461 + ], + "score": 0.967, + "type": "image", + "image_path": "e7bb923f2aed9a6363a3a02049afe2ed3ae65c849658c6e2bd754b853c8d9f19.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 129, + 330, + 484, + 373.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 129, + 373.6666666666667, + 484, + 417.33333333333337 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 129, + 417.33333333333337, + 484, + 461.00000000000006 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 471, + 505, + 493 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 471, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 505, + 484 + ], + "score": 1.0, + "content": "Figure 12: Schematic diagrams and auto-parsed graph structures of the InvertedPendulum and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 483, + 243, + 494 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 243, + 494 + ], + "score": 1.0, + "content": "InvertedDoublePendulum.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + } + ], + "index": 9.25 + }, + { + "type": "image", + "bbox": [ + 119, + 524, + 491, + 632 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 524, + 491, + 632 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 119, + 524, + 491, + 632 + ], + "spans": [ + { + "bbox": [ + 119, + 524, + 491, + 632 + ], + "score": 0.966, + "type": "image", + "image_path": "3ecaa892fd0c5dde6ad0cbe96e739a085bbe439b8b996b0c352579ba4ec09a38.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 119, + 524, + 491, + 560.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 119, + 560.0, + 491, + 596.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 119, + 596.0, + 491, + 632.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 641, + 504, + 653 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 653 + ], + "score": 1.0, + "content": "Figure 13: Schematic diagrams and auto-parsed graph structures of the Walker2d and Reacher.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + } + ], + "index": 14.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 120, + 98, + 487, + 198 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 120, + 98, + 487, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 98, + 487, + 198 + ], + "spans": [ + { + "bbox": [ + 120, + 98, + 487, + 198 + ], + "score": 0.964, + "type": "image", + "image_path": "cf363b16eec569ed2a1a9999406d9bf9b77be3fcd51151252e7b3aa593148326.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 120, + 98, + 487, + 131.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 120, + 131.33333333333334, + 487, + 164.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 120, + 164.66666666666669, + 487, + 198.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 208, + 504, + 220 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 207, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 506, + 221 + ], + "score": 1.0, + "content": "Figure 14: Schematic diagrams and auto-parsed graph structures of the SnakeSix and Swimmer.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 135, + 254, + 481, + 443 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 135, + 254, + 481, + 443 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 135, + 254, + 481, + 443 + ], + "spans": [ + { + "bbox": [ + 135, + 254, + 481, + 443 + ], + "score": 0.968, + "type": "image", + "image_path": "dff29462eb705451d33123a58042e9f900230477745841d1761508e310e2368a.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 135, + 254, + 481, + 317.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 135, + 317.0, + 481, + 380.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 135, + 380.0, + 481, + 443.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 151, + 461, + 459, + 474 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 151, + 461, + 460, + 475 + ], + "spans": [ + { + "bbox": [ + 151, + 461, + 460, + 475 + ], + "score": 1.0, + "content": "Figure 15: Schematic diagrams and auto-parsed graph structures of the Ant.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "image", + "bbox": [ + 123, + 495, + 489, + 653 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 495, + 489, + 653 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 123, + 495, + 489, + 653 + ], + "spans": [ + { + "bbox": [ + 123, + 495, + 489, + 653 + ], + "score": 0.963, + "type": "image", + "image_path": "24f34f799745d2698fe80b8df160044a5a1c445eeb646dbd1136e757d410d30f.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 123, + 495, + 489, + 547.6666666666666 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 123, + 547.6666666666666, + 489, + 600.3333333333333 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 123, + 600.3333333333333, + 489, + 652.9999999999999 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 674, + 505, + 696 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 672, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 506, + 687 + ], + "score": 1.0, + "content": "Figure 16: Schematic diagrams and auto-parsed graph structures of the Hopper and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 685, + 177, + 696 + ], + "spans": [ + { + "bbox": [ + 106, + 685, + 177, + 696 + ], + "score": 1.0, + "content": "HalfCheetah.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + } + ], + "index": 10.25 + } + ], + "page_idx": 20, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 120, + 98, + 487, + 198 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 120, + 98, + 487, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 98, + 487, + 198 + ], + "spans": [ + { + "bbox": [ + 120, + 98, + 487, + 198 + ], + "score": 0.964, + "type": "image", + "image_path": "cf363b16eec569ed2a1a9999406d9bf9b77be3fcd51151252e7b3aa593148326.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 120, + 98, + 487, + 131.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 120, + 131.33333333333334, + 487, + 164.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 120, + 164.66666666666669, + 487, + 198.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 208, + 504, + 220 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 207, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 506, + 221 + ], + "score": 1.0, + "content": "Figure 14: Schematic diagrams and auto-parsed graph structures of the SnakeSix and Swimmer.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 135, + 254, + 481, + 443 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 135, + 254, + 481, + 443 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 135, + 254, + 481, + 443 + ], + "spans": [ + { + "bbox": [ + 135, + 254, + 481, + 443 + ], + "score": 0.968, + "type": "image", + "image_path": "dff29462eb705451d33123a58042e9f900230477745841d1761508e310e2368a.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 135, + 254, + 481, + 317.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 135, + 317.0, + 481, + 380.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 135, + 380.0, + 481, + 443.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 151, + 461, + 459, + 474 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 151, + 461, + 460, + 475 + ], + "spans": [ + { + "bbox": [ + 151, + 461, + 460, + 475 + ], + "score": 1.0, + "content": "Figure 15: Schematic diagrams and auto-parsed graph structures of the Ant.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "image", + "bbox": [ + 123, + 495, + 489, + 653 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 495, + 489, + 653 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 123, + 495, + 489, + 653 + ], + "spans": [ + { + "bbox": [ + 123, + 495, + 489, + 653 + ], + "score": 0.963, + "type": "image", + "image_path": "24f34f799745d2698fe80b8df160044a5a1c445eeb646dbd1136e757d410d30f.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 123, + 495, + 489, + 547.6666666666666 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 123, + 547.6666666666666, + 489, + 600.3333333333333 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 123, + 600.3333333333333, + 489, + 652.9999999999999 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 674, + 505, + 696 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 672, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 506, + 687 + ], + "score": 1.0, + "content": "Figure 16: Schematic diagrams and auto-parsed graph structures of the Hopper and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 685, + 177, + 696 + ], + "spans": [ + { + "bbox": [ + 106, + 685, + 177, + 696 + ], + "score": 1.0, + "content": "HalfCheetah.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + } + ], + "index": 10.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 83, + 324, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 326, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 326, + 95 + ], + "score": 1.0, + "content": "6.5 DETAILS OF ZERO-SHOT LEARNING RESULTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 157, + 142, + 323, + 695 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 111, + 295, + 123, + 533 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 295, + 123, + 533 + ], + "spans": [], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "table_caption", + "bbox": [ + 136, + 133, + 156, + 582 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 134, + 187, + 146, + 526 + ], + "spans": [ + { + "bbox": [ + 134, + 187, + 146, + 240 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 134, + 340, + 145, + 378 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 134, + 484, + 145, + 526 + ], + "score": 0.0, + "content": "", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 157, + 142, + 323, + 695 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 157, + 142, + 323, + 695 + ], + "spans": [ + { + "bbox": [ + 157, + 142, + 323, + 695 + ], + "score": 0.823, + "html": "
539396449484264448423252454福
91-101-101855 8004195593351 264 L'L9-
11.971-6'061-9'S61- 52586872111475113 461-14411
46018三99 5814 4444-1103435587-848 8'I
91552929-854福F44
2233344886335929'L6163289464
2806I3品188282
19453m0222222337494988881044
983841 9491 448888
4 4164216152
4199919189LI62312441111444449五T1LI1046101224494466
4004484854469413559 54944135583封9'$9937
1 667-中458334514
5334542283348
57611-46'0-059S01-61-409m6n6
eeeeeeeeeeeeeeeeeeeeeieeeegeeeeeeeeeeeeegeeeeeteeeeeeereeneeteeteereeeeeicreeeeeereeeeeeegeeer erneeeeeeeeeeeeereeeeeeeeeeeeeeeeeercreeeneeeeeeeeeerergereeeeereeteeereeieeieneeeeeeeeeeereeeeereeeeeeeeeeeeeeeeeeeeneeeeeneeeeeeeeeeeicrrereeerreteetgereeeneeeeeeeeeetereenetrectreeeetrerreeeeeerrcegeeeeeerereeeeeeecrgepeeeeelepeeeeeeseeneeeeeeeereeeeeeaeeeeeeeeeeeeeeeeteeea
", + "type": "table", + "image_path": "ff9980b0373dbcf1a9519042fb2fe2eb47d30158b75f9c8499c1e844a1ce6ac9.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 157, + 142, + 323, + 326.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 157, + 326.33333333333337, + 323, + 510.66666666666674 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 157, + 510.66666666666674, + 323, + 695.0000000000001 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 337, + 140, + 498, + 694 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 325, + 288, + 338, + 623 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 325, + 288, + 338, + 623 + ], + "spans": [], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 337, + 140, + 498, + 694 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 337, + 140, + 498, + 694 + ], + "spans": [ + { + "bbox": [ + 337, + 140, + 498, + 694 + ], + "score": 0.871, + "html": "
971900933554ISL9816 48124111111
福848-R488845118101888636-
13-808-668-69L-498-486-29'88-8£6L-728-4155-58888-764156-
55446 8314448F4364学3999京08329
9'SL-0388756-6'1-6499L-28756-7I1-3678-896-8678-$'96-
8386'091590556699496007844481115446548814福683515554
3314541S2453 34251863
4568833853494
L094169 48248
518588408584548443244334
23225496970829449$'0998552455413331255551835454TOIIOI03134111441455514694445918041887551957138050675545
S66L339835369558635-8三61
14444431148B6953 国5058818
2533773828361949'101326.30059296078
993打99111
eeeereeneeeeeeeeet eereereeerreeeeegereeeeereraaergeeeedeereeeergereeeeeieeeereeneereergereeereeeeeeeeeeeeeceeeeeeeeeeeeeeeeeceeeereeeeeeeeeecreareeereseeeeeeeereeeetGeepeeeeeecegepeeeeetCeeereneerecegeeeeeeeeeerereeeeeneeeereeeeeeelereeeeeeeeeeneeeeneeeeeeeeeeeeeeereeeeeeeeeeereeeecreeeeererrereeneerectgereeetireeeeeeeseeetneeeeieeeeereretreresesaeeeeeeeeeeti
", + "type": "table", + "image_path": "dfb5a3b4eb73c9332db6beb500365a48a459fe578cfa003f8e5708e40128a51b.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 337, + 140, + 498, + 324.66666666666663 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 337, + 324.66666666666663, + 498, + 509.33333333333326 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 337, + 509.33333333333326, + 498, + 693.9999999999999 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 6.5 + } + ], + "page_idx": 21, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "22", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 83, + 324, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 326, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 326, + 95 + ], + "score": 1.0, + "content": "6.5 DETAILS OF ZERO-SHOT LEARNING RESULTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 157, + 142, + 323, + 695 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 111, + 295, + 123, + 533 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 295, + 123, + 533 + ], + "spans": [], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "table_caption", + "bbox": [ + 136, + 133, + 156, + 582 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 134, + 187, + 146, + 526 + ], + "spans": [ + { + "bbox": [ + 134, + 187, + 146, + 240 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 134, + 340, + 145, + 378 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 134, + 484, + 145, + 526 + ], + "score": 0.0, + "content": "", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 157, + 142, + 323, + 695 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 157, + 142, + 323, + 695 + ], + "spans": [ + { + "bbox": [ + 157, + 142, + 323, + 695 + ], + "score": 0.823, + "html": "
539396449484264448423252454福
91-101-101855 8004195593351 264 L'L9-
11.971-6'061-9'S61- 52586872111475113 461-14411
46018三99 5814 4444-1103435587-848 8'I
91552929-854福F44
2233344886335929'L6163289464
2806I3品188282
19453m0222222337494988881044
983841 9491 448888
4 4164216152
4199919189LI62312441111444449五T1LI1046101224494466
4004484854469413559 54944135583封9'$9937
1 667-中458334514
5334542283348
57611-46'0-059S01-61-409m6n6
eeeeeeeeeeeeeeeeeeeeeieeeegeeeeeeeeeeeeegeeeeeteeeeeeereeneeteeteereeeeeicreeeeeereeeeeeegeeer erneeeeeeeeeeeeereeeeeeeeeeeeeeeeeercreeeneeeeeeeeeerergereeeeereeteeereeieeieneeeeeeeeeeereeeeereeeeeeeeeeeeeeeeeeeeneeeeeneeeeeeeeeeeicrrereeerreteetgereeeneeeeeeeeeetereenetrectreeeetrerreeeeeerrcegeeeeeerereeeeeeecrgepeeeeelepeeeeeeseeneeeeeeeereeeeeeaeeeeeeeeeeeeeeeeteeea
", + "type": "table", + "image_path": "ff9980b0373dbcf1a9519042fb2fe2eb47d30158b75f9c8499c1e844a1ce6ac9.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 157, + 142, + 323, + 326.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 157, + 326.33333333333337, + 323, + 510.66666666666674 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 157, + 510.66666666666674, + 323, + 695.0000000000001 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 337, + 140, + 498, + 694 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 325, + 288, + 338, + 623 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 325, + 288, + 338, + 623 + ], + "spans": [], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 337, + 140, + 498, + 694 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 337, + 140, + 498, + 694 + ], + "spans": [ + { + "bbox": [ + 337, + 140, + 498, + 694 + ], + "score": 0.871, + "html": "
971900933554ISL9816 48124111111
福848-R488845118101888636-
13-808-668-69L-498-486-29'88-8£6L-728-4155-58888-764156-
55446 8314448F4364学3999京08329
9'SL-0388756-6'1-6499L-28756-7I1-3678-896-8678-$'96-
8386'091590556699496007844481115446548814福683515554
3314541S2453 34251863
4568833853494
L094169 48248
518588408584548443244334
23225496970829449$'0998552455413331255551835454TOIIOI03134111441455514694445918041887551957138050675545
S66L339835369558635-8三61
14444431148B6953 国5058818
2533773828361949'101326.30059296078
993打99111
eeeereeneeeeeeeeet eereereeerreeeeegereeeeereraaergeeeedeereeeergereeeeeieeeereeneereergereeereeeeeeeeeeeeeceeeeeeeeeeeeeeeeeceeeereeeeeeeeeecreareeereseeeeeeeereeeetGeepeeeeeecegepeeeeetCeeereneerecegeeeeeeeeeerereeeeeneeeereeeeeeelereeeeeeeeeeneeeeneeeeeeeeeeeeeeereeeeeeeeeeereeeecreeeeererrereeneerectgereeetireeeeeeeseeetneeeeieeeeereretreresesaeeeeeeeeeeti
", + "type": "table", + "image_path": "dfb5a3b4eb73c9332db6beb500365a48a459fe578cfa003f8e5708e40128a51b.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 337, + 140, + 498, + 324.66666666666663 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 337, + 324.66666666666663, + 498, + 509.33333333333326 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 337, + 509.33333333333326, + 498, + 693.9999999999999 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 6.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 210, + 144, + 431, + 533 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 210, + 144, + 431, + 533 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 210, + 144, + 431, + 533 + ], + "spans": [ + { + "bbox": [ + 210, + 144, + 431, + 533 + ], + "score": 0.974, + "html": "
1.55290.00668744444435339LL35.55759'8116.44695511.441114£99L1143423360.6778'8311.15
555418.3554318.0044.5511915678IL'LS40.64488189011117118951511.351L0116410338130
66.4441442555511.09533.35154.31435.5514445198'0L0153.59£9.09.044/.6156.559.55S1855438.5053816.7916.1460.00434
111152585-56.95-6'L[-4441-9L'9S-1135911.51555331460'8L7'01-£6.33-11.04410'61-1571-89.97-L6'67-353335331531-88'L7-40.05-
15:114353555.815531966986.69355555.3511.333335.21448.1135.3845.55335.713233219'8466.55370.3380.5535.7128.85
10028B28281055131641621328388084660
ererreresdererertrreresdeeiereeerereeeeeeertrrereeeeer erraeraseeereerteeeeeer crrrrreereeretrrereeeeerrereedreeeeerrereeeeerCereresdeeeeerrtrreresdeeererrreeereereeeeetrereedeercrrereeeerreerrereeeetrCgrrreesdrerecrraedticerseerrretrrereeeirereeeereecrreaeeeeereereeseirreeeeeeeeeeareeeeereeeiceerreeeeteseseeaEreereedeereeretrrereseeereedeeeeeeeseeeeterreresdeeeeerereeseseeeeeeresdreeseeereseeererarrgeaeedeereeeereeseeeereeeeserrreerrareseeercarereeeerreetrrreeeeerreeeeteeeeseeeeercaaeesdeteeerteseeeear
", + "type": "table", + "image_path": "b0b769b050b71f164316d38f75aa2fa2beb1010592b70fe5e038c3a45826a78f.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 210, + 144, + 431, + 157.89285714285714 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 210, + 157.89285714285714, + 431, + 171.78571428571428 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 210, + 171.78571428571428, + 431, + 185.67857142857142 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 210, + 185.67857142857142, + 431, + 199.57142857142856 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 210, + 199.57142857142856, + 431, + 213.4642857142857 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 210, + 213.4642857142857, + 431, + 227.35714285714283 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 210, + 227.35714285714283, + 431, + 241.24999999999997 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 210, + 241.24999999999997, + 431, + 255.1428571428571 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 210, + 255.1428571428571, + 431, + 269.0357142857143 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 210, + 269.0357142857143, + 431, + 282.92857142857144 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 210, + 282.92857142857144, + 431, + 296.8214285714286 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 210, + 296.8214285714286, + 431, + 310.7142857142858 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 210, + 310.7142857142858, + 431, + 324.60714285714295 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 210, + 324.60714285714295, + 431, + 338.5000000000001 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 210, + 338.5000000000001, + 431, + 352.3928571428573 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 210, + 352.3928571428573, + 431, + 366.28571428571445 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 210, + 366.28571428571445, + 431, + 380.1785714285716 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 210, + 380.1785714285716, + 431, + 394.0714285714288 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 210, + 394.0714285714288, + 431, + 407.96428571428595 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 210, + 407.96428571428595, + 431, + 421.8571428571431 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 210, + 421.8571428571431, + 431, + 435.7500000000003 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 210, + 435.7500000000003, + 431, + 449.64285714285745 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 210, + 449.64285714285745, + 431, + 463.5357142857146 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 210, + 463.5357142857146, + 431, + 477.4285714285718 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 210, + 477.4285714285718, + 431, + 491.32142857142895 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 210, + 491.32142857142895, + 431, + 505.2142857142861 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 210, + 505.2142857142861, + 431, + 519.1071428571432 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 210, + 519.1071428571432, + 431, + 533.0000000000003 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 202, + 176, + 215, + 501 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 204, + 469, + 213, + 487 + ], + "spans": [ + { + "bbox": [ + 204, + 469, + 213, + 487 + ], + "score": 0.0, + "content": "", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "table_caption", + "bbox": [ + 176, + 153, + 192, + 524 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 176, + 153, + 192, + 524 + ], + "spans": [], + "index": 29 + } + ], + "index": 29 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 620, + 505, + 654 + ], + "lines": [ + { + "bbox": [ + 105, + 619, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 633 + ], + "score": 1.0, + "content": "In the MLP-Bind method, we bind the weights of MLP. By doing this, the weights of the agent from", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "the similar structures will be shared. For example, in the centipede environment, the weights", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 642, + 401, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 401, + 655 + ], + "score": 1.0, + "content": "from observation to action of all the LeftHips are constrained to be same.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + } + ], + "page_idx": 22, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 763 + ], + "score": 1.0, + "content": "23", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 210, + 144, + 431, + 533 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 210, + 144, + 431, + 533 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 210, + 144, + 431, + 533 + ], + "spans": [ + { + "bbox": [ + 210, + 144, + 431, + 533 + ], + "score": 0.974, + "html": "
1.55290.00668744444435339LL35.55759'8116.44695511.441114£99L1143423360.6778'8311.15
555418.3554318.0044.5511915678IL'LS40.64488189011117118951511.351L0116410338130
66.4441442555511.09533.35154.31435.5514445198'0L0153.59£9.09.044/.6156.559.55S1855438.5053816.7916.1460.00434
111152585-56.95-6'L[-4441-9L'9S-1135911.51555331460'8L7'01-£6.33-11.04410'61-1571-89.97-L6'67-353335331531-88'L7-40.05-
15:114353555.815531966986.69355555.3511.333335.21448.1135.3845.55335.713233219'8466.55370.3380.5535.7128.85
10028B28281055131641621328388084660
ererreresdererertrreresdeeiereeerereeeeeeertrrereeeeer erraeraseeereerteeeeeer crrrrreereeretrrereeeeerrereedreeeeerrereeeeerCereresdeeeeerrtrreresdeeererrreeereereeeeetrereedeercrrereeeerreerrereeeetrCgrrreesdrerecrraedticerseerrretrrereeeirereeeereecrreaeeeeereereeseirreeeeeeeeeeareeeeereeeiceerreeeeteseseeaEreereedeereeretrrereseeereedeeeeeeeseeeeterreresdeeeeerereeseseeeeeeresdreeseeereseeererarrgeaeedeereeeereeseeeereeeeserrreerrareseeercarereeeerreetrrreeeeerreeeeteeeeseeeeercaaeesdeteeerteseeeear
", + "type": "table", + "image_path": "b0b769b050b71f164316d38f75aa2fa2beb1010592b70fe5e038c3a45826a78f.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 210, + 144, + 431, + 157.89285714285714 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 210, + 157.89285714285714, + 431, + 171.78571428571428 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 210, + 171.78571428571428, + 431, + 185.67857142857142 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 210, + 185.67857142857142, + 431, + 199.57142857142856 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 210, + 199.57142857142856, + 431, + 213.4642857142857 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 210, + 213.4642857142857, + 431, + 227.35714285714283 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 210, + 227.35714285714283, + 431, + 241.24999999999997 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 210, + 241.24999999999997, + 431, + 255.1428571428571 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 210, + 255.1428571428571, + 431, + 269.0357142857143 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 210, + 269.0357142857143, + 431, + 282.92857142857144 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 210, + 282.92857142857144, + 431, + 296.8214285714286 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 210, + 296.8214285714286, + 431, + 310.7142857142858 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 210, + 310.7142857142858, + 431, + 324.60714285714295 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 210, + 324.60714285714295, + 431, + 338.5000000000001 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 210, + 338.5000000000001, + 431, + 352.3928571428573 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 210, + 352.3928571428573, + 431, + 366.28571428571445 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 210, + 366.28571428571445, + 431, + 380.1785714285716 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 210, + 380.1785714285716, + 431, + 394.0714285714288 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 210, + 394.0714285714288, + 431, + 407.96428571428595 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 210, + 407.96428571428595, + 431, + 421.8571428571431 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 210, + 421.8571428571431, + 431, + 435.7500000000003 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 210, + 435.7500000000003, + 431, + 449.64285714285745 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 210, + 449.64285714285745, + 431, + 463.5357142857146 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 210, + 463.5357142857146, + 431, + 477.4285714285718 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 210, + 477.4285714285718, + 431, + 491.32142857142895 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 210, + 491.32142857142895, + 431, + 505.2142857142861 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 210, + 505.2142857142861, + 431, + 519.1071428571432 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 210, + 519.1071428571432, + 431, + 533.0000000000003 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 202, + 176, + 215, + 501 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 204, + 469, + 213, + 487 + ], + "spans": [ + { + "bbox": [ + 204, + 469, + 213, + 487 + ], + "score": 0.0, + "content": "", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "table_caption", + "bbox": [ + 176, + 153, + 192, + 524 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 176, + 153, + 192, + 524 + ], + "spans": [], + "index": 29 + } + ], + "index": 29 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 620, + 505, + 654 + ], + "lines": [ + { + "bbox": [ + 105, + 619, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 633 + ], + "score": 1.0, + "content": "In the MLP-Bind method, we bind the weights of MLP. By doing this, the weights of the agent from", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "the similar structures will be shared. For example, in the centipede environment, the weights", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 642, + 401, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 401, + 655 + ], + "score": 1.0, + "content": "from observation to action of all the LeftHips are constrained to be same.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 619, + 505, + 655 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 122, + 97, + 487, + 560 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 198, + 80, + 413, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 197, + 78, + 413, + 94 + ], + "spans": [ + { + "bbox": [ + 197, + 78, + 413, + 94 + ], + "score": 1.0, + "content": "Table 10: Results of the zero-shot learning of Snake", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 122, + 97, + 487, + 560 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 97, + 487, + 560 + ], + "spans": [ + { + "bbox": [ + 122, + 97, + 487, + 560 + ], + "score": 0.982, + "html": "
Reward Min
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v1-31.719-43.328-64.958-58.811308.994301.4
SnakeFive-v1-56.38229.262-55.562-39.346
SnakeSix-v1-44.888-5.306-42.744-76.42266.066
SnakeSeven-v1-52.764-25.511-49.765-59.52231.557329.456333.237253.828
SnakeFive-v1-56.382-30.466-49.953-47.339
SnakeSix-v1SnakeSeven-v1-44.888-35.655-52.18-57.334-90.837-52.015
-52.76489.358
Reward Std
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v19.5445.79210.76910.1465.5135.788
SnakeFive-v113.3186.19111.7157.883
SnakeSix-v111.51513.53510.30416.8697.3957.1795.856.60414.084
SnakeSeven-v111.2112.13110.9816.953
SnakeFive-v113.31852.7111.5448.5198.51923.10913.529
SnakeSix-v1SnakeSeven-v111.51511.2164.54910.68310.52811.615
Reward Max
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v13.863-19.40910.1180.918338.374326.974
SnakeFive-v120.58563.8663.136-2.329
SnakeSix-v17.658.7793.51-2.139305.011
SnakeSeven-v1-2.2851.2651.325-28.087275.536
SnakeFive-v120.585139.6175.3051.487356.366
SnakeSix-v1SnakeSeven-v17.6-2.28156.771143.8242.3162.596-4.1598.394367.768336.561
Reward Average
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v1-13.531-30.185-12.363-21.824325.476314.919
SnakeFive-v1-17.82951.411-16.481-21.256
SnakeSix-v1-19.8932.2-19.674-42.461282.426
SnakeSeven-v1-22.09923.533-22.378-42.742256.293
SnakeFive-v1-17.82989.823-16.54-15.975342.881
SnakeSix-v1-19.89105.632-20.454-34.343351.85313.149
SnakeSeven-v1-22.099115.123-22.81-22.383
", + "type": "table", + "image_path": "6e10e55c33e7a4e84bf33169da0a7c4acbe289137d4dbd8961f161c9a4f46c38.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 122, + 97, + 487, + 251.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 122, + 251.33333333333334, + 487, + 405.6666666666667 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 122, + 405.6666666666667, + 487, + 560.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 576, + 433, + 588 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 576, + 437, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 437, + 590 + ], + "score": 1.0, + "content": "6.5.1 LINEAR NORMALIZATION ZERO-SHOT RESULTS AND COLORIZATION.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 505, + 641 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 607 + ], + "score": 1.0, + "content": "As the scale of zero-shot results is very different, we normalize the results across different models", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 607, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 506, + 619 + ], + "score": 1.0, + "content": "for each transfer learning task. For each task, we record the worst value of results from different", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 617, + 507, + 634 + ], + "spans": [ + { + "bbox": [ + 104, + 617, + 252, + 634 + ], + "score": 1.0, + "content": "models and the pre-set worst value", + "type": "text" + }, + { + "bbox": [ + 253, + 618, + 276, + 631 + ], + "score": 0.9, + "content": "V _ { \\mathrm { m i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 617, + 507, + 634 + ], + "score": 1.0, + "content": ". we set the normalization minimun value as this worst", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 628, + 454, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 339, + 641 + ], + "score": 1.0, + "content": "value. We calculate the normalization maximum value by", + "type": "text" + }, + { + "bbox": [ + 340, + 628, + 412, + 641 + ], + "score": 0.33, + "content": "\\lfloor \\mathrm { m a x } ( V ) / \\mathrm { I n t L e n } \\rfloor", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 630, + 422, + 639 + ], + "score": 0.61, + "content": "^ *", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 628, + 454, + 641 + ], + "score": 1.0, + "content": "IntLen.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "table", + "bbox": [ + 214, + 684, + 397, + 718 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 651, + 504, + 674 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 651, + 504, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 504, + 662 + ], + "score": 1.0, + "content": "Table 11: Parameters for linear normalization during zero-shot results. Results are shown in the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 662, + 401, + 674 + ], + "spans": [ + { + "bbox": [ + 107, + 662, + 401, + 674 + ], + "score": 1.0, + "content": "order of Centipedes’ reward, Centipedes’ running-length, Snakes’ reward", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "table_body", + "bbox": [ + 214, + 684, + 397, + 718 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 214, + 684, + 397, + 718 + ], + "spans": [ + { + "bbox": [ + 214, + 684, + 397, + 718 + ], + "score": 0.967, + "html": "
parametersVminIntLen
value-100,-100, -2030,30,30
", + "type": "table", + "image_path": "dd9260e1ab2daad1bfcea3b6f8621cfccf78d2760864779b625029b35c1995e4.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 214, + 684, + 397, + 701.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 214, + 701.0, + 397, + 718.0 + ], + "spans": [], + "index": 12 + } + ] + } + ], + "index": 10.5 + } + ], + "page_idx": 23, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 122, + 97, + 487, + 560 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 198, + 80, + 413, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 197, + 78, + 413, + 94 + ], + "spans": [ + { + "bbox": [ + 197, + 78, + 413, + 94 + ], + "score": 1.0, + "content": "Table 10: Results of the zero-shot learning of Snake", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 122, + 97, + 487, + 560 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 97, + 487, + 560 + ], + "spans": [ + { + "bbox": [ + 122, + 97, + 487, + 560 + ], + "score": 0.982, + "html": "
Reward Min
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v1-31.719-43.328-64.958-58.811308.994301.4
SnakeFive-v1-56.38229.262-55.562-39.346
SnakeSix-v1-44.888-5.306-42.744-76.42266.066
SnakeSeven-v1-52.764-25.511-49.765-59.52231.557329.456333.237253.828
SnakeFive-v1-56.382-30.466-49.953-47.339
SnakeSix-v1SnakeSeven-v1-44.888-35.655-52.18-57.334-90.837-52.015
-52.76489.358
Reward Std
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v19.5445.79210.76910.1465.5135.788
SnakeFive-v113.3186.19111.7157.883
SnakeSix-v111.51513.53510.30416.8697.3957.1795.856.60414.084
SnakeSeven-v111.2112.13110.9816.953
SnakeFive-v113.31852.7111.5448.5198.51923.10913.529
SnakeSix-v1SnakeSeven-v111.51511.2164.54910.68310.52811.615
Reward Max
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v13.863-19.40910.1180.918338.374326.974
SnakeFive-v120.58563.8663.136-2.329
SnakeSix-v17.658.7793.51-2.139305.011
SnakeSeven-v1-2.2851.2651.325-28.087275.536
SnakeFive-v120.585139.6175.3051.487356.366
SnakeSix-v1SnakeSeven-v17.6-2.28156.771143.8242.3162.596-4.1598.394367.768336.561
Reward Average
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v1-13.531-30.185-12.363-21.824325.476314.919
SnakeFive-v1-17.82951.411-16.481-21.256
SnakeSix-v1-19.8932.2-19.674-42.461282.426
SnakeSeven-v1-22.09923.533-22.378-42.742256.293
SnakeFive-v1-17.82989.823-16.54-15.975342.881
SnakeSix-v1-19.89105.632-20.454-34.343351.85313.149
SnakeSeven-v1-22.099115.123-22.81-22.383
", + "type": "table", + "image_path": "6e10e55c33e7a4e84bf33169da0a7c4acbe289137d4dbd8961f161c9a4f46c38.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 122, + 97, + 487, + 251.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 122, + 251.33333333333334, + 487, + 405.6666666666667 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 122, + 405.6666666666667, + 487, + 560.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 576, + 433, + 588 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 576, + 437, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 437, + 590 + ], + "score": 1.0, + "content": "6.5.1 LINEAR NORMALIZATION ZERO-SHOT RESULTS AND COLORIZATION.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 505, + 641 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 607 + ], + "score": 1.0, + "content": "As the scale of zero-shot results is very different, we normalize the results across different models", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 607, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 506, + 619 + ], + "score": 1.0, + "content": "for each transfer learning task. For each task, we record the worst value of results from different", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 617, + 507, + 634 + ], + "spans": [ + { + "bbox": [ + 104, + 617, + 252, + 634 + ], + "score": 1.0, + "content": "models and the pre-set worst value", + "type": "text" + }, + { + "bbox": [ + 253, + 618, + 276, + 631 + ], + "score": 0.9, + "content": "V _ { \\mathrm { m i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 617, + 507, + 634 + ], + "score": 1.0, + "content": ". we set the normalization minimun value as this worst", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 628, + 454, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 339, + 641 + ], + "score": 1.0, + "content": "value. We calculate the normalization maximum value by", + "type": "text" + }, + { + "bbox": [ + 340, + 628, + 412, + 641 + ], + "score": 0.33, + "content": "\\lfloor \\mathrm { m a x } ( V ) / \\mathrm { I n t L e n } \\rfloor", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 630, + 422, + 639 + ], + "score": 0.61, + "content": "^ *", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 628, + 454, + 641 + ], + "score": 1.0, + "content": "IntLen.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 104, + 597, + 507, + 641 + ] + }, + { + "type": "table", + "bbox": [ + 214, + 684, + 397, + 718 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 651, + 504, + 674 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 651, + 504, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 504, + 662 + ], + "score": 1.0, + "content": "Table 11: Parameters for linear normalization during zero-shot results. Results are shown in the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 662, + 401, + 674 + ], + "spans": [ + { + "bbox": [ + 107, + 662, + 401, + 674 + ], + "score": 1.0, + "content": "order of Centipedes’ reward, Centipedes’ running-length, Snakes’ reward", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "table_body", + "bbox": [ + 214, + 684, + 397, + 718 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 214, + 684, + 397, + 718 + ], + "spans": [ + { + "bbox": [ + 214, + 684, + 397, + 718 + ], + "score": 0.967, + "html": "
parametersVminIntLen
value-100,-100, -2030,30,30
", + "type": "table", + "image_path": "dd9260e1ab2daad1bfcea3b6f8621cfccf78d2760864779b625029b35c1995e4.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 214, + 684, + 397, + 701.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 214, + 701.0, + 397, + 718.0 + ], + "spans": [], + "index": 12 + } + ] + } + ], + "index": 10.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 155, + 107, + 455, + 411 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 82, + 282, + 94 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 82, + 283, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 283, + 95 + ], + "score": 1.0, + "content": "6.6 DETAILS OF ROBUSTNESS RESULTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 155, + 107, + 455, + 411 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 155, + 107, + 455, + 411 + ], + "spans": [ + { + "bbox": [ + 155, + 107, + 455, + 411 + ], + "score": 0.984, + "html": "
ModelMass
Hopper Halfhumanoid HorseWolfOstrich
NerveNet1981.592047.921968.872062.361852.762519.62392.052282.842161.611804.522304.251900.431862.211955.16922.12932.48983.11910.29791.84
1703.471648.281576.211354.51
1239.72946.08
MLP2268.641932.981637.471591.511073.1421.82115.75278.921224.98353.561692.081083.621932.641754.94878.75945.94
690.22538.14443.921459.931039.9814.55879.66805.99802.41
ModelStrength
HopperHalfhumanoidHorseWolfOstrich
NerveNet1961.621117.91492.33468.4446.141528.89940.96594.53308.08249.62316.82960.341827.73862.19274.77189.49133.84794.56539.45371.19255.12197.84
488.51305.16
305.16201.42
MLP800.15431.97429.76426.27412.391439.891040.341236.881730.881108.32230.85483.36
7209.54
765.7306.08519.73163.36
117.0565.29228.26187.7288.28108.97232.8133.99
", + "type": "table", + "image_path": "149be23ac6501df5ffee2b243d8730e45c0e416d58559db5e15b40927b122fac.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 155, + 107, + 455, + 121.47619047619048 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 155, + 121.47619047619048, + 455, + 135.95238095238096 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 155, + 135.95238095238096, + 455, + 150.42857142857144 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 155, + 150.42857142857144, + 455, + 164.90476190476193 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 155, + 164.90476190476193, + 455, + 179.3809523809524 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 155, + 179.3809523809524, + 455, + 193.8571428571429 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 155, + 193.8571428571429, + 455, + 208.33333333333337 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 155, + 208.33333333333337, + 455, + 222.80952380952385 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 155, + 222.80952380952385, + 455, + 237.28571428571433 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 155, + 237.28571428571433, + 455, + 251.76190476190482 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 155, + 251.76190476190482, + 455, + 266.2380952380953 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 155, + 266.2380952380953, + 455, + 280.7142857142858 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 155, + 280.7142857142858, + 455, + 295.19047619047626 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 155, + 295.19047619047626, + 455, + 309.66666666666674 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 155, + 309.66666666666674, + 455, + 324.1428571428572 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 155, + 324.1428571428572, + 455, + 338.6190476190477 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 155, + 338.6190476190477, + 455, + 353.0952380952382 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 155, + 353.0952380952382, + 455, + 367.57142857142867 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 155, + 367.57142857142867, + 455, + 382.04761904761915 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 155, + 382.04761904761915, + 455, + 396.52380952380963 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 155, + 396.52380952380963, + 455, + 411.0000000000001 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 213, + 419, + 398, + 432 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 212, + 417, + 399, + 434 + ], + "spans": [ + { + "bbox": [ + 212, + 417, + 399, + 434 + ], + "score": 1.0, + "content": "Table 12: Detail results of Robustness testing.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 106, + 453, + 344, + 466 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 344, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 344, + 466 + ], + "score": 1.0, + "content": "6.7 HYPERPARAMETERS FOR MULTI-TASK LEARNING", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 194, + 498, + 417, + 664 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 188, + 477, + 423, + 489 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 187, + 475, + 423, + 492 + ], + "spans": [ + { + "bbox": [ + 187, + 475, + 423, + 492 + ], + "score": 1.0, + "content": "Table 13: Hyperparameter settings for multi-task learning.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "table_body", + "bbox": [ + 194, + 498, + 417, + 664 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 194, + 498, + 417, + 664 + ], + "spans": [ + { + "bbox": [ + 194, + 498, + 417, + 664 + ], + "score": 0.981, + "html": "
MLP, TreeNetand NerveNetValue Tried
Network Shape[64, 64]
Number of Iteration Per Update10
Use KL PenaltyNo
Learning Rate SchedulerAdaptive
NerveNetValue Tried
Number of Propogation Steps4
NerveNetVariants Size of Nodes'Hidden StateNerveNet-1
64
Merge joint and body NodeYes
Output Network Disable Edge Type Separate
Add Skip-connection from /to rootYes No
", + "type": "table", + "image_path": "9d8959eb0e6443be88c8602880ed2d737116266dfa7e25399e10eb71bc123454.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 194, + 498, + 417, + 513.0909090909091 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 194, + 513.0909090909091, + 417, + 528.1818181818182 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 194, + 528.1818181818182, + 417, + 543.2727272727274 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 194, + 543.2727272727274, + 417, + 558.3636363636365 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 194, + 558.3636363636365, + 417, + 573.4545454545456 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 194, + 573.4545454545456, + 417, + 588.5454545454547 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 194, + 588.5454545454547, + 417, + 603.6363636363639 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 194, + 603.6363636363639, + 417, + 618.727272727273 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 194, + 618.727272727273, + 417, + 633.8181818181821 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 194, + 633.8181818181821, + 417, + 648.9090909090912 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 194, + 648.9090909090912, + 417, + 664.0000000000003 + ], + "spans": [], + "index": 35 + } + ] + } + ], + "index": 27.0 + }, + { + "type": "title", + "bbox": [ + 106, + 685, + 319, + 696 + ], + "lines": [ + { + "bbox": [ + 106, + 685, + 320, + 697 + ], + "spans": [ + { + "bbox": [ + 106, + 685, + 320, + 697 + ], + "score": 1.0, + "content": "6.8 SCHEMATIC FIGURES OF WA L K E R S AGENTS", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 705, + 442, + 717 + ], + "lines": [ + { + "bbox": [ + 105, + 704, + 443, + 719 + ], + "spans": [ + { + "bbox": [ + 105, + 704, + 443, + 719 + ], + "score": 1.0, + "content": "We design Walker task-set, which contains five 2d walkers in the MuJoCo engine.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + } + ], + "page_idx": 24, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 155, + 107, + 455, + 411 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 82, + 282, + 94 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 82, + 283, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 283, + 95 + ], + "score": 1.0, + "content": "6.6 DETAILS OF ROBUSTNESS RESULTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 155, + 107, + 455, + 411 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 155, + 107, + 455, + 411 + ], + "spans": [ + { + "bbox": [ + 155, + 107, + 455, + 411 + ], + "score": 0.984, + "html": "
ModelMass
Hopper Halfhumanoid HorseWolfOstrich
NerveNet1981.592047.921968.872062.361852.762519.62392.052282.842161.611804.522304.251900.431862.211955.16922.12932.48983.11910.29791.84
1703.471648.281576.211354.51
1239.72946.08
MLP2268.641932.981637.471591.511073.1421.82115.75278.921224.98353.561692.081083.621932.641754.94878.75945.94
690.22538.14443.921459.931039.9814.55879.66805.99802.41
ModelStrength
HopperHalfhumanoidHorseWolfOstrich
NerveNet1961.621117.91492.33468.4446.141528.89940.96594.53308.08249.62316.82960.341827.73862.19274.77189.49133.84794.56539.45371.19255.12197.84
488.51305.16
305.16201.42
MLP800.15431.97429.76426.27412.391439.891040.341236.881730.881108.32230.85483.36
7209.54
765.7306.08519.73163.36
117.0565.29228.26187.7288.28108.97232.8133.99
", + "type": "table", + "image_path": "149be23ac6501df5ffee2b243d8730e45c0e416d58559db5e15b40927b122fac.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 155, + 107, + 455, + 121.47619047619048 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 155, + 121.47619047619048, + 455, + 135.95238095238096 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 155, + 135.95238095238096, + 455, + 150.42857142857144 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 155, + 150.42857142857144, + 455, + 164.90476190476193 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 155, + 164.90476190476193, + 455, + 179.3809523809524 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 155, + 179.3809523809524, + 455, + 193.8571428571429 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 155, + 193.8571428571429, + 455, + 208.33333333333337 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 155, + 208.33333333333337, + 455, + 222.80952380952385 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 155, + 222.80952380952385, + 455, + 237.28571428571433 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 155, + 237.28571428571433, + 455, + 251.76190476190482 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 155, + 251.76190476190482, + 455, + 266.2380952380953 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 155, + 266.2380952380953, + 455, + 280.7142857142858 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 155, + 280.7142857142858, + 455, + 295.19047619047626 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 155, + 295.19047619047626, + 455, + 309.66666666666674 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 155, + 309.66666666666674, + 455, + 324.1428571428572 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 155, + 324.1428571428572, + 455, + 338.6190476190477 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 155, + 338.6190476190477, + 455, + 353.0952380952382 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 155, + 353.0952380952382, + 455, + 367.57142857142867 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 155, + 367.57142857142867, + 455, + 382.04761904761915 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 155, + 382.04761904761915, + 455, + 396.52380952380963 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 155, + 396.52380952380963, + 455, + 411.0000000000001 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 213, + 419, + 398, + 432 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 212, + 417, + 399, + 434 + ], + "spans": [ + { + "bbox": [ + 212, + 417, + 399, + 434 + ], + "score": 1.0, + "content": "Table 12: Detail results of Robustness testing.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 106, + 453, + 344, + 466 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 344, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 344, + 466 + ], + "score": 1.0, + "content": "6.7 HYPERPARAMETERS FOR MULTI-TASK LEARNING", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 194, + 498, + 417, + 664 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 188, + 477, + 423, + 489 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 187, + 475, + 423, + 492 + ], + "spans": [ + { + "bbox": [ + 187, + 475, + 423, + 492 + ], + "score": 1.0, + "content": "Table 13: Hyperparameter settings for multi-task learning.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "table_body", + "bbox": [ + 194, + 498, + 417, + 664 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 194, + 498, + 417, + 664 + ], + "spans": [ + { + "bbox": [ + 194, + 498, + 417, + 664 + ], + "score": 0.981, + "html": "
MLP, TreeNetand NerveNetValue Tried
Network Shape[64, 64]
Number of Iteration Per Update10
Use KL PenaltyNo
Learning Rate SchedulerAdaptive
NerveNetValue Tried
Number of Propogation Steps4
NerveNetVariants Size of Nodes'Hidden StateNerveNet-1
64
Merge joint and body NodeYes
Output Network Disable Edge Type Separate
Add Skip-connection from /to rootYes No
", + "type": "table", + "image_path": "9d8959eb0e6443be88c8602880ed2d737116266dfa7e25399e10eb71bc123454.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 194, + 498, + 417, + 513.0909090909091 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 194, + 513.0909090909091, + 417, + 528.1818181818182 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 194, + 528.1818181818182, + 417, + 543.2727272727274 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 194, + 543.2727272727274, + 417, + 558.3636363636365 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 194, + 558.3636363636365, + 417, + 573.4545454545456 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 194, + 573.4545454545456, + 417, + 588.5454545454547 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 194, + 588.5454545454547, + 417, + 603.6363636363639 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 194, + 603.6363636363639, + 417, + 618.727272727273 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 194, + 618.727272727273, + 417, + 633.8181818181821 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 194, + 633.8181818181821, + 417, + 648.9090909090912 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 194, + 648.9090909090912, + 417, + 664.0000000000003 + ], + "spans": [], + "index": 35 + } + ] + } + ], + "index": 27.0 + }, + { + "type": "title", + "bbox": [ + 106, + 685, + 319, + 696 + ], + "lines": [ + { + "bbox": [ + 106, + 685, + 320, + 697 + ], + "spans": [ + { + "bbox": [ + 106, + 685, + 320, + 697 + ], + "score": 1.0, + "content": "6.8 SCHEMATIC FIGURES OF WA L K E R S AGENTS", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 705, + 442, + 717 + ], + "lines": [ + { + "bbox": [ + 105, + 704, + 443, + 719 + ], + "spans": [ + { + "bbox": [ + 105, + 704, + 443, + 719 + ], + "score": 1.0, + "content": "We design Walker task-set, which contains five 2d walkers in the MuJoCo engine.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 704, + 443, + 719 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 113, + 113, + 501, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 113, + 501, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 113, + 501, + 225 + ], + "spans": [ + { + "bbox": [ + 113, + 113, + 501, + 225 + ], + "score": 0.953, + "type": "image", + "image_path": "e4db0554bf395b640e337ab9dd90cfa8eba43566ee9855ed7bab06708c5c8236.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 113, + 113, + 501, + 150.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 113, + 150.33333333333334, + 501, + 187.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 113, + 187.66666666666669, + 501, + 225.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 237, + 505, + 260 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 236, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 506, + 251 + ], + "score": 1.0, + "content": "Figure 17: Schematic diagrams and auto-parsed graph structures of Walker-Ostrich and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 249, + 178, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 178, + 260 + ], + "score": 1.0, + "content": "Walker-Wolf.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 154, + 320, + 459, + 465 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 154, + 320, + 459, + 465 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 154, + 320, + 459, + 465 + ], + "spans": [ + { + "bbox": [ + 154, + 320, + 459, + 465 + ], + "score": 0.964, + "type": "image", + "image_path": "24d53bd50058ae9dce42cfb0dbe3b1f6127ebefc151bad848cb175c05a53e2d4.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 154, + 320, + 459, + 368.3333333333333 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 154, + 368.3333333333333, + 459, + 416.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 154, + 416.66666666666663, + 459, + 464.99999999999994 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 477, + 505, + 501 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 477, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 505, + 490 + ], + "score": 1.0, + "content": "Figure 18: Schematic diagrams and auto-parsed graph structures of Walker-Hopper and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 489, + 225, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 225, + 501 + ], + "score": 1.0, + "content": "Walker-HalfHumanoid.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "image", + "bbox": [ + 172, + 563, + 437, + 677 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 172, + 563, + 437, + 677 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 172, + 563, + 437, + 677 + ], + "spans": [ + { + "bbox": [ + 172, + 563, + 437, + 677 + ], + "score": 0.959, + "type": "image", + "image_path": "9de69acaa7f53ccf56bad5e7e49119e3e4eb8f81606d03c69dfcfd1a0fa3d1be.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 172, + 563, + 437, + 601.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 172, + 601.0, + 437, + 639.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 172, + 639.0, + 437, + 677.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 131, + 694, + 478, + 706 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 131, + 693, + 480, + 707 + ], + "spans": [ + { + "bbox": [ + 131, + 693, + 480, + 707 + ], + "score": 1.0, + "content": "Figure 19: Schematic diagrams and auto-parsed graph structures of Walker-Horse.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + } + ], + "index": 12.0 + } + ], + "page_idx": 25, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "26", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 182, + 113, + 215, + 120 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 113, + 113, + 501, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 113, + 501, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 113, + 501, + 225 + ], + "spans": [ + { + "bbox": [ + 113, + 113, + 501, + 225 + ], + "score": 0.953, + "type": "image", + "image_path": "e4db0554bf395b640e337ab9dd90cfa8eba43566ee9855ed7bab06708c5c8236.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 113, + 113, + 501, + 150.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 113, + 150.33333333333334, + 501, + 187.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 113, + 187.66666666666669, + 501, + 225.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 237, + 505, + 260 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 236, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 506, + 251 + ], + "score": 1.0, + "content": "Figure 17: Schematic diagrams and auto-parsed graph structures of Walker-Ostrich and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 249, + 178, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 178, + 260 + ], + "score": 1.0, + "content": "Walker-Wolf.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 154, + 320, + 459, + 465 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 154, + 320, + 459, + 465 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 154, + 320, + 459, + 465 + ], + "spans": [ + { + "bbox": [ + 154, + 320, + 459, + 465 + ], + "score": 0.964, + "type": "image", + "image_path": "24d53bd50058ae9dce42cfb0dbe3b1f6127ebefc151bad848cb175c05a53e2d4.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 154, + 320, + 459, + 368.3333333333333 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 154, + 368.3333333333333, + 459, + 416.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 154, + 416.66666666666663, + 459, + 464.99999999999994 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 477, + 505, + 501 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 477, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 505, + 490 + ], + "score": 1.0, + "content": "Figure 18: Schematic diagrams and auto-parsed graph structures of Walker-Hopper and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 489, + 225, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 225, + 501 + ], + "score": 1.0, + "content": "Walker-HalfHumanoid.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "image", + "bbox": [ + 172, + 563, + 437, + 677 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 172, + 563, + 437, + 677 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 172, + 563, + 437, + 677 + ], + "spans": [ + { + "bbox": [ + 172, + 563, + 437, + 677 + ], + "score": 0.959, + "type": "image", + "image_path": "9de69acaa7f53ccf56bad5e7e49119e3e4eb8f81606d03c69dfcfd1a0fa3d1be.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 172, + 563, + 437, + 601.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 172, + 601.0, + 437, + 639.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 172, + 639.0, + 437, + 677.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 131, + 694, + 478, + 706 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 131, + 693, + 480, + 707 + ], + "spans": [ + { + "bbox": [ + 131, + 693, + 480, + 707 + ], + "score": 1.0, + "content": "Figure 19: Schematic diagrams and auto-parsed graph structures of Walker-Horse.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + } + ], + "index": 12.0 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/S1sqHMZCb/S1sqHMZCb_model.json b/parse/train/S1sqHMZCb/S1sqHMZCb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..42487bf73a0cac9af52b93e9a65303b25c34d134 --- /dev/null +++ b/parse/train/S1sqHMZCb/S1sqHMZCb_model.json @@ -0,0 +1,35361 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 710, + 1302, + 710, + 1302, + 1137, + 398, + 1137 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 299, + 1511, + 1403, + 1511, + 1403, + 1725, + 299, + 1725 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1740, + 1404, + 1740, + 1404, + 1986, + 299, + 1986 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1264, + 1403, + 1264, + 1403, + 1387, + 299, + 1387 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 1403, + 1402, + 1403, + 1402, + 1494, + 299, + 1494 + ], + "score": 0.968 + }, + { + "category_id": 0, + "poly": [ + 299, + 219, + 1398, + 219, + 1398, + 323, + 299, + 323 + ], + "score": 0.96 + }, + { + "category_id": 2, + "poly": [ + 335, + 2007, + 659, + 2007, + 659, + 2033, + 335, + 2033 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 302, + 1195, + 572, + 1195, + 572, + 1230, + 302, + 1230 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 773, + 641, + 927, + 641, + 927, + 674, + 773, + 674 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 104, + 299, + 104 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 857, + 2088, + 857, + 2112, + 841, + 2112 + ], + "score": 0.713 + }, + { + "category_id": 1, + "poly": [ + 315, + 414, + 1050, + 414, + 1050, + 560, + 315, + 560 + ], + "score": 0.626 + }, + { + "category_id": 1, + "poly": [ + 314, + 376, + 971, + 376, + 971, + 409, + 314, + 409 + ], + "score": 0.617 + }, + { + "category_id": 15, + "poly": [ + 294.0, + 216.0, + 1404.0, + 216.0, + 1404.0, + 272.0, + 294.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 272.0, + 891.0, + 272.0, + 891.0, + 328.0, + 297.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1998.0, + 662.0, + 1998.0, + 662.0, + 2040.0, + 331.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1192.0, + 579.0, + 1192.0, + 579.0, + 1239.0, + 293.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 637.0, + 934.0, + 637.0, + 934.0, + 680.0, + 768.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 840.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 711.0, + 1304.0, + 711.0, + 1304.0, + 743.0, + 397.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 739.0, + 1307.0, + 739.0, + 1307.0, + 777.0, + 393.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 771.0, + 1306.0, + 771.0, + 1306.0, + 806.0, + 392.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 803.0, + 1305.0, + 803.0, + 1305.0, + 835.0, + 395.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 833.0, + 1307.0, + 833.0, + 1307.0, + 866.0, + 394.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 864.0, + 1305.0, + 864.0, + 1305.0, + 899.0, + 394.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 895.0, + 1306.0, + 895.0, + 1306.0, + 927.0, + 393.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 925.0, + 1305.0, + 925.0, + 1305.0, + 956.0, + 394.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 954.0, + 1307.0, + 954.0, + 1307.0, + 986.0, + 392.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 984.0, + 1307.0, + 984.0, + 1307.0, + 1020.0, + 391.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1014.0, + 1306.0, + 1014.0, + 1306.0, + 1050.0, + 393.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1044.0, + 1306.0, + 1044.0, + 1306.0, + 1081.0, + 393.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1076.0, + 1306.0, + 1076.0, + 1306.0, + 1111.0, + 394.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1104.0, + 1047.0, + 1104.0, + 1047.0, + 1144.0, + 393.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1510.0, + 1407.0, + 1510.0, + 1407.0, + 1548.0, + 294.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1541.0, + 1404.0, + 1541.0, + 1404.0, + 1578.0, + 294.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1571.0, + 1405.0, + 1571.0, + 1405.0, + 1608.0, + 293.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1603.0, + 1405.0, + 1603.0, + 1405.0, + 1638.0, + 294.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1630.0, + 1405.0, + 1630.0, + 1405.0, + 1671.0, + 293.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1665.0, + 1405.0, + 1665.0, + 1405.0, + 1699.0, + 295.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1694.0, + 1057.0, + 1694.0, + 1057.0, + 1731.0, + 293.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1738.0, + 1406.0, + 1738.0, + 1406.0, + 1774.0, + 295.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1769.0, + 1405.0, + 1769.0, + 1405.0, + 1808.0, + 292.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1802.0, + 1404.0, + 1802.0, + 1404.0, + 1835.0, + 294.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1833.0, + 1405.0, + 1833.0, + 1405.0, + 1867.0, + 295.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1863.0, + 1404.0, + 1863.0, + 1404.0, + 1897.0, + 295.0, + 1897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1891.0, + 1405.0, + 1891.0, + 1405.0, + 1930.0, + 292.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1924.0, + 1404.0, + 1924.0, + 1404.0, + 1958.0, + 294.0, + 1958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1954.0, + 1405.0, + 1954.0, + 1405.0, + 1990.0, + 293.0, + 1990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1264.0, + 1404.0, + 1264.0, + 1404.0, + 1300.0, + 295.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1296.0, + 1405.0, + 1296.0, + 1405.0, + 1329.0, + 294.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1323.0, + 1407.0, + 1323.0, + 1407.0, + 1363.0, + 293.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1357.0, + 773.0, + 1357.0, + 773.0, + 1392.0, + 293.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1402.0, + 1406.0, + 1402.0, + 1406.0, + 1439.0, + 294.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1431.0, + 1404.0, + 1431.0, + 1404.0, + 1470.0, + 293.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1465.0, + 1267.0, + 1465.0, + 1267.0, + 1499.0, + 295.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 408.0, + 691.0, + 408.0, + 691.0, + 440.0, + 312.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 436.0, + 562.0, + 436.0, + 562.0, + 472.0, + 311.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 467.0, + 491.0, + 467.0, + 491.0, + 500.0, + 312.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 498.0, + 891.0, + 498.0, + 891.0, + 533.0, + 312.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 531.0, + 1050.0, + 531.0, + 1050.0, + 561.0, + 314.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 373.0, + 977.0, + 373.0, + 977.0, + 417.0, + 310.0, + 417.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 968, + 1404, + 968, + 1404, + 1334, + 298, + 1334 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 298, + 1349, + 1403, + 1349, + 1403, + 1534, + 298, + 1534 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1819, + 1403, + 1819, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 321, + 251, + 1364, + 251, + 1364, + 615, + 321, + 615 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 301, + 1711, + 1401, + 1711, + 1401, + 1804, + 301, + 1804 + ], + "score": 0.961 + }, + { + "category_id": 4, + "poly": [ + 296, + 642, + 1404, + 642, + 1404, + 796, + 296, + 796 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 295, + 890, + 1400, + 890, + 1400, + 954, + 295, + 954 + ], + "score": 0.95 + }, + { + "category_id": 0, + "poly": [ + 299, + 1617, + 512, + 1617, + 512, + 1653, + 299, + 1653 + ], + "score": 0.909 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.697 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.185 + }, + { + "category_id": 13, + "poly": [ + 790, + 1973, + 895, + 1973, + 895, + 2006, + 790, + 2006 + ], + "score": 0.93, + "latex": "r ( s ^ { \\tau } , a ^ { \\tau } )" + }, + { + "category_id": 13, + "poly": [ + 999, + 1911, + 1113, + 1911, + 1113, + 1945, + 999, + 1945 + ], + "score": 0.93, + "latex": "\\pi _ { \\boldsymbol { \\theta } } ( \\boldsymbol { a } ^ { \\tau } | \\boldsymbol { s } ^ { \\tau } )" + }, + { + "category_id": 13, + "poly": [ + 471, + 1943, + 563, + 1943, + 563, + 1971, + 471, + 1971 + ], + "score": 0.91, + "latex": "a ^ { \\tau } \\in \\mathcal { A }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1943, + 384, + 1943, + 384, + 1972, + 298, + 1972 + ], + "score": 0.9, + "latex": "s ^ { \\tau } \\in \\boldsymbol { S }" + }, + { + "category_id": 13, + "poly": [ + 759, + 1944, + 776, + 1944, + 776, + 1970, + 759, + 1970 + ], + "score": 0.81, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 991, + 1883, + 1013, + 1883, + 1013, + 1909, + 991, + 1909 + ], + "score": 0.81, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1239, + 1882, + 1265, + 1882, + 1265, + 1909, + 1239, + 1909 + ], + "score": 0.75, + "latex": "\\mathcal { A }" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 292.0, + 898.0, + 292.0, + 898.0, + 321.0, + 825.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 292.0, + 1035.0, + 292.0, + 1035.0, + 321.0, + 960.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 292.0, + 1171.0, + 292.0, + 1171.0, + 321.0, + 1095.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 293.0, + 1304.0, + 293.0, + 1304.0, + 319.0, + 1231.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 351.0, + 898.0, + 351.0, + 898.0, + 380.0, + 869.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 353.0, + 1033.0, + 353.0, + 1033.0, + 379.0, + 1008.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 352.0, + 1169.0, + 352.0, + 1169.0, + 379.0, + 1144.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 353.0, + 1303.0, + 353.0, + 1303.0, + 379.0, + 1220.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 409.0, + 1088.0, + 409.0, + 1088.0, + 440.0, + 1058.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 409.0, + 1223.0, + 409.0, + 1223.0, + 440.0, + 1193.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1327.0, + 410.0, + 1358.0, + 410.0, + 1358.0, + 440.0, + 1327.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 442.0, + 961.0, + 442.0, + 961.0, + 466.0, + 914.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 442.0, + 1095.0, + 442.0, + 1095.0, + 466.0, + 1049.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 443.0, + 1231.0, + 443.0, + 1231.0, + 466.0, + 1185.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1320.0, + 445.0, + 1359.0, + 445.0, + 1359.0, + 466.0, + 1320.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 468.0, + 897.0, + 468.0, + 897.0, + 496.0, + 804.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 469.0, + 1034.0, + 469.0, + 1034.0, + 496.0, + 940.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 468.0, + 1171.0, + 468.0, + 1171.0, + 497.0, + 1075.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 468.0, + 1303.0, + 468.0, + 1303.0, + 496.0, + 1210.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 495.0, + 744.0, + 495.0, + 744.0, + 597.0, + 571.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 531.0, + 898.0, + 531.0, + 898.0, + 560.0, + 822.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 531.0, + 1035.0, + 531.0, + 1035.0, + 560.0, + 958.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 531.0, + 1171.0, + 531.0, + 1171.0, + 560.0, + 1091.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1227.0, + 535.0, + 1275.0, + 535.0, + 1275.0, + 555.0, + 1227.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 530.0, + 1306.0, + 530.0, + 1306.0, + 560.0, + 1277.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.75, + 351.5, + 874.75, + 351.5, + 874.75, + 375.5, + 814.75, + 375.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 351.5, + 1010.0, + 351.5, + 1010.0, + 375.5, + 950.0, + 375.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 351.5, + 1145.0, + 351.5, + 1145.0, + 375.5, + 1084.0, + 375.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.75, + 408.5, + 948.75, + 408.5, + 948.75, + 439.0, + 925.75, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 640.0, + 1406.0, + 640.0, + 1406.0, + 678.0, + 293.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 671.0, + 1404.0, + 671.0, + 1404.0, + 710.0, + 294.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 704.0, + 1404.0, + 704.0, + 1404.0, + 738.0, + 295.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 731.0, + 1407.0, + 731.0, + 1407.0, + 771.0, + 293.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 761.0, + 1133.0, + 761.0, + 1133.0, + 801.0, + 294.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1613.0, + 518.0, + 1613.0, + 518.0, + 1660.0, + 291.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 967.0, + 1404.0, + 967.0, + 1404.0, + 1004.0, + 295.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 998.0, + 1406.0, + 998.0, + 1406.0, + 1033.0, + 292.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1030.0, + 1406.0, + 1030.0, + 1406.0, + 1064.0, + 294.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1058.0, + 1408.0, + 1058.0, + 1408.0, + 1098.0, + 292.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1092.0, + 1404.0, + 1092.0, + 1404.0, + 1126.0, + 294.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1122.0, + 1405.0, + 1122.0, + 1405.0, + 1156.0, + 294.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1151.0, + 1406.0, + 1151.0, + 1406.0, + 1189.0, + 294.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1181.0, + 1405.0, + 1181.0, + 1405.0, + 1219.0, + 294.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1212.0, + 1405.0, + 1212.0, + 1405.0, + 1249.0, + 293.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1242.0, + 1405.0, + 1242.0, + 1405.0, + 1278.0, + 292.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1274.0, + 1405.0, + 1274.0, + 1405.0, + 1308.0, + 294.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1305.0, + 1366.0, + 1305.0, + 1366.0, + 1339.0, + 294.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1349.0, + 1405.0, + 1349.0, + 1405.0, + 1384.0, + 294.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1380.0, + 1403.0, + 1380.0, + 1403.0, + 1416.0, + 294.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1408.0, + 1405.0, + 1408.0, + 1405.0, + 1445.0, + 293.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1441.0, + 1405.0, + 1441.0, + 1405.0, + 1477.0, + 294.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1472.0, + 1407.0, + 1472.0, + 1407.0, + 1508.0, + 294.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1501.0, + 1156.0, + 1501.0, + 1156.0, + 1538.0, + 293.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1855.0, + 296.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1850.0, + 1406.0, + 1850.0, + 1406.0, + 1885.0, + 292.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 990.0, + 1882.0, + 990.0, + 1917.0, + 294.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 1882.0, + 1238.0, + 1882.0, + 1238.0, + 1917.0, + 1014.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1917.0, + 1266.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 998.0, + 1912.0, + 998.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1946.0, + 1114.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1940.0, + 297.0, + 1940.0, + 297.0, + 1977.0, + 292.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 1940.0, + 470.0, + 1940.0, + 470.0, + 1977.0, + 385.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1940.0, + 758.0, + 1940.0, + 758.0, + 1977.0, + 564.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 1940.0, + 1403.0, + 1940.0, + 1403.0, + 1977.0, + 777.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 789.0, + 1972.0, + 789.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2006.0, + 896.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2003.0, + 853.0, + 2003.0, + 853.0, + 2037.0, + 296.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1712.0, + 1406.0, + 1712.0, + 1406.0, + 1745.0, + 295.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1743.0, + 1404.0, + 1743.0, + 1404.0, + 1777.0, + 295.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1774.0, + 462.0, + 1774.0, + 462.0, + 1805.0, + 295.0, + 1805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 889.0, + 1402.0, + 889.0, + 1402.0, + 925.0, + 296.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 921.0, + 1041.0, + 921.0, + 1041.0, + 958.0, + 294.0, + 958.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 943, + 1404, + 943, + 1404, + 1430, + 297, + 1430 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 1667, + 1404, + 1667, + 1404, + 2034, + 297, + 2034 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1529, + 1403, + 1529, + 1403, + 1651, + 298, + 1651 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 306, + 244, + 1395, + 244, + 1395, + 627, + 306, + 627 + ], + "score": 0.971 + }, + { + "category_id": 4, + "poly": [ + 296, + 668, + 1405, + 668, + 1405, + 821, + 296, + 821 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 303, + 1470, + 765, + 1470, + 765, + 1502, + 303, + 1502 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 299, + 884, + 651, + 884, + 651, + 915, + 299, + 915 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 814, + 76, + 814, + 104, + 300, + 104 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.622 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.393 + }, + { + "category_id": 1, + "poly": [ + 296, + 668, + 1405, + 668, + 1405, + 821, + 296, + 821 + ], + "score": 0.099 + }, + { + "category_id": 13, + "poly": [ + 805, + 1668, + 940, + 1668, + 940, + 1701, + 805, + 1701 + ], + "score": 0.93, + "latex": "G = ( V , E )" + }, + { + "category_id": 13, + "poly": [ + 444, + 1759, + 540, + 1759, + 540, + 1793, + 444, + 1793 + ], + "score": 0.93, + "latex": "\\mathcal { N } _ { o u t } ( u )" + }, + { + "category_id": 13, + "poly": [ + 413, + 1820, + 628, + 1820, + 628, + 1853, + 413, + 1853 + ], + "score": 0.92, + "latex": "p _ { u } \\in \\{ 1 , 2 , \\ldots , P \\}" + }, + { + "category_id": 13, + "poly": [ + 813, + 1851, + 1069, + 1851, + 1069, + 1887, + 813, + 1887 + ], + "score": 0.91, + "latex": "c _ { ( u , v ) } \\in \\{ 1 , 2 , \\ldots , C \\}" + }, + { + "category_id": 13, + "poly": [ + 949, + 1791, + 1035, + 1791, + 1035, + 1823, + 949, + 1823 + ], + "score": 0.91, + "latex": "\\mathcal { N } _ { i n } ( u )" + }, + { + "category_id": 13, + "poly": [ + 951, + 1760, + 1016, + 1760, + 1016, + 1791, + 951, + 1791 + ], + "score": 0.9, + "latex": "( u , v )" + }, + { + "category_id": 13, + "poly": [ + 529, + 1854, + 594, + 1854, + 594, + 1884, + 529, + 1884 + ], + "score": 0.89, + "latex": "( u , v )" + }, + { + "category_id": 13, + "poly": [ + 1096, + 1670, + 1121, + 1670, + 1121, + 1696, + 1096, + 1696 + ], + "score": 0.82, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 1019, + 1669, + 1045, + 1669, + 1045, + 1697, + 1019, + 1697 + ], + "score": 0.76, + "latex": "V" + }, + { + "category_id": 13, + "poly": [ + 871, + 1765, + 889, + 1765, + 889, + 1787, + 871, + 1787 + ], + "score": 0.75, + "latex": "v" + }, + { + "category_id": 13, + "poly": [ + 895, + 1795, + 916, + 1795, + 916, + 1818, + 895, + 1818 + ], + "score": 0.75, + "latex": "u" + }, + { + "category_id": 13, + "poly": [ + 1183, + 1795, + 1202, + 1795, + 1202, + 1817, + 1183, + 1817 + ], + "score": 0.72, + "latex": "u" + }, + { + "category_id": 13, + "poly": [ + 389, + 1764, + 410, + 1764, + 410, + 1787, + 389, + 1787 + ], + "score": 0.7, + "latex": "u" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 248.0, + 427.0, + 248.0, + 427.0, + 268.0, + 349.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 245.0, + 564.0, + 245.0, + 564.0, + 275.0, + 518.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 252.0, + 909.0, + 252.0, + 909.0, + 272.0, + 831.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 246.0, + 1215.0, + 246.0, + 1215.0, + 274.0, + 1162.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1320.0, + 246.0, + 1396.0, + 246.0, + 1396.0, + 266.0, + 1320.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 263.0, + 412.0, + 263.0, + 412.0, + 285.0, + 364.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 263.0, + 563.0, + 263.0, + 563.0, + 285.0, + 517.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 266.0, + 893.0, + 266.0, + 893.0, + 287.0, + 847.0, + 287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 263.0, + 1211.0, + 263.0, + 1211.0, + 285.0, + 1166.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1334.0, + 262.0, + 1384.0, + 262.0, + 1384.0, + 283.0, + 1334.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 275.0, + 1383.0, + 275.0, + 1383.0, + 296.0, + 1335.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 320.0, + 360.0, + 320.0, + 360.0, + 338.0, + 312.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 317.0, + 764.0, + 317.0, + 764.0, + 327.0, + 754.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 317.0, + 879.0, + 317.0, + 879.0, + 327.0, + 868.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 317.0, + 1016.0, + 317.0, + 1016.0, + 327.0, + 1007.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 332.0, + 367.0, + 332.0, + 367.0, + 352.0, + 309.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 345.0, + 371.0, + 345.0, + 371.0, + 365.0, + 304.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 353.0, + 611.0, + 353.0, + 611.0, + 371.0, + 576.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 346.0, + 744.0, + 346.0, + 744.0, + 369.0, + 714.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 359.0, + 758.0, + 359.0, + 758.0, + 371.0, + 746.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 347.0, + 857.0, + 347.0, + 857.0, + 369.0, + 829.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 349.0, + 979.0, + 349.0, + 979.0, + 359.0, + 969.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1227.0, + 351.0, + 1264.0, + 351.0, + 1264.0, + 373.0, + 1227.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 373.0, + 739.0, + 373.0, + 739.0, + 383.0, + 729.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 371.0, + 770.0, + 371.0, + 770.0, + 385.0, + 754.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 373.0, + 853.0, + 373.0, + 853.0, + 383.0, + 843.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 373.0, + 1020.0, + 373.0, + 1020.0, + 384.0, + 1008.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 381.0, + 756.0, + 381.0, + 756.0, + 398.0, + 743.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 383.0, + 782.0, + 383.0, + 782.0, + 398.0, + 768.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 384.0, + 869.0, + 384.0, + 869.0, + 397.0, + 858.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 384.0, + 895.0, + 384.0, + 895.0, + 397.0, + 885.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 381.0, + 1008.0, + 381.0, + 1008.0, + 398.0, + 994.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 381.0, + 1035.0, + 381.0, + 1035.0, + 398.0, + 1020.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 408.0, + 363.0, + 408.0, + 363.0, + 442.0, + 314.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 428.0, + 472.0, + 428.0, + 472.0, + 447.0, + 444.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 412.0, + 608.0, + 412.0, + 608.0, + 433.0, + 572.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 418.0, + 640.0, + 418.0, + 640.0, + 432.0, + 619.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 407.0, + 766.0, + 407.0, + 766.0, + 442.0, + 750.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 407.0, + 881.0, + 407.0, + 881.0, + 442.0, + 864.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 419.0, + 951.0, + 419.0, + 951.0, + 445.0, + 911.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 405.0, + 1020.0, + 405.0, + 1020.0, + 444.0, + 999.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 428.0, + 1122.0, + 428.0, + 1122.0, + 447.0, + 1096.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 412.0, + 1260.0, + 412.0, + 1260.0, + 433.0, + 1225.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 441.0, + 564.0, + 441.0, + 564.0, + 458.0, + 532.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 442.0, + 724.0, + 442.0, + 724.0, + 452.0, + 714.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 442.0, + 839.0, + 442.0, + 839.0, + 453.0, + 828.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 436.0, + 958.0, + 436.0, + 958.0, + 461.0, + 904.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 442.0, + 978.0, + 442.0, + 978.0, + 453.0, + 966.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 441.0, + 1216.0, + 441.0, + 1216.0, + 458.0, + 1185.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 451.0, + 855.0, + 451.0, + 855.0, + 463.0, + 846.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 452.0, + 992.0, + 452.0, + 992.0, + 462.0, + 983.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 461.0, + 625.0, + 461.0, + 625.0, + 472.0, + 602.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 463.0, + 881.0, + 463.0, + 881.0, + 478.0, + 867.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 464.0, + 1016.0, + 464.0, + 1016.0, + 477.0, + 1007.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 471.0, + 501.0, + 471.0, + 501.0, + 493.0, + 464.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 473.0, + 580.0, + 473.0, + 580.0, + 491.0, + 553.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 478.0, + 753.0, + 478.0, + 753.0, + 488.0, + 743.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 478.0, + 779.0, + 478.0, + 779.0, + 488.0, + 769.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 478.0, + 868.0, + 478.0, + 868.0, + 488.0, + 857.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 477.0, + 895.0, + 477.0, + 895.0, + 488.0, + 883.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 478.0, + 1004.0, + 478.0, + 1004.0, + 488.0, + 994.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 478.0, + 1031.0, + 478.0, + 1031.0, + 488.0, + 1020.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 472.0, + 1144.0, + 472.0, + 1144.0, + 492.0, + 1118.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 472.0, + 1234.0, + 472.0, + 1234.0, + 491.0, + 1205.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 515.0, + 359.0, + 515.0, + 359.0, + 536.0, + 317.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 512.0, + 481.0, + 512.0, + 481.0, + 531.0, + 452.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 513.0, + 607.0, + 513.0, + 607.0, + 531.0, + 568.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 512.0, + 1134.0, + 512.0, + 1134.0, + 531.0, + 1103.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1228.0, + 515.0, + 1256.0, + 515.0, + 1256.0, + 529.0, + 1228.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 527.0, + 359.0, + 527.0, + 359.0, + 550.0, + 319.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 524.0, + 764.0, + 524.0, + 764.0, + 534.0, + 754.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 524.0, + 879.0, + 524.0, + 879.0, + 535.0, + 868.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 524.0, + 1016.0, + 524.0, + 1016.0, + 535.0, + 1004.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 540.0, + 358.0, + 540.0, + 358.0, + 561.0, + 319.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 549.0, + 530.0, + 549.0, + 530.0, + 568.0, + 502.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 549.0, + 604.0, + 549.0, + 604.0, + 568.0, + 578.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 555.0, + 724.0, + 555.0, + 724.0, + 564.0, + 714.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 555.0, + 977.0, + 555.0, + 977.0, + 564.0, + 967.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 551.0, + 1179.0, + 551.0, + 1179.0, + 565.0, + 1158.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 551.0, + 1253.0, + 551.0, + 1253.0, + 566.0, + 1232.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 562.0, + 608.0, + 562.0, + 608.0, + 602.0, + 472.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 564.0, + 855.0, + 564.0, + 855.0, + 574.0, + 846.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 581.0, + 607.0, + 581.0, + 607.0, + 613.0, + 473.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 591.0, + 1031.0, + 591.0, + 1031.0, + 602.0, + 1021.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.25, + 279.5, + 403.25, + 279.5, + 403.25, + 295.0, + 368.25, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 449.5, + 1119.0, + 449.5, + 1119.0, + 465.0, + 1105.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 495.5, + 628.0, + 495.5, + 628.0, + 504.5, + 619.0, + 504.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 665.0, + 1405.0, + 665.0, + 1405.0, + 704.0, + 293.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 698.0, + 1405.0, + 698.0, + 1405.0, + 731.0, + 295.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 727.0, + 1405.0, + 727.0, + 1405.0, + 764.0, + 295.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 759.0, + 1405.0, + 759.0, + 1405.0, + 796.0, + 294.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 789.0, + 1015.0, + 789.0, + 1015.0, + 825.0, + 294.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1470.0, + 770.0, + 1470.0, + 770.0, + 1506.0, + 296.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 881.0, + 654.0, + 881.0, + 654.0, + 920.0, + 293.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 941.0, + 1405.0, + 941.0, + 1405.0, + 977.0, + 294.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 969.0, + 1408.0, + 969.0, + 1408.0, + 1012.0, + 292.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1002.0, + 1405.0, + 1002.0, + 1405.0, + 1039.0, + 294.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1032.0, + 1407.0, + 1032.0, + 1407.0, + 1070.0, + 294.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1064.0, + 1406.0, + 1064.0, + 1406.0, + 1098.0, + 295.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1094.0, + 1406.0, + 1094.0, + 1406.0, + 1130.0, + 292.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1122.0, + 1405.0, + 1122.0, + 1405.0, + 1163.0, + 291.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1156.0, + 1406.0, + 1156.0, + 1406.0, + 1190.0, + 294.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1187.0, + 1405.0, + 1187.0, + 1405.0, + 1221.0, + 295.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1217.0, + 1404.0, + 1217.0, + 1404.0, + 1251.0, + 295.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1249.0, + 1404.0, + 1249.0, + 1404.0, + 1283.0, + 294.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1277.0, + 1405.0, + 1277.0, + 1405.0, + 1312.0, + 292.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1306.0, + 1407.0, + 1306.0, + 1407.0, + 1343.0, + 291.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1338.0, + 1404.0, + 1338.0, + 1404.0, + 1372.0, + 292.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1369.0, + 1405.0, + 1369.0, + 1405.0, + 1403.0, + 294.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1399.0, + 610.0, + 1399.0, + 610.0, + 1434.0, + 294.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1668.0, + 804.0, + 1668.0, + 804.0, + 1702.0, + 295.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1668.0, + 1018.0, + 1668.0, + 1018.0, + 1702.0, + 941.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 1668.0, + 1095.0, + 1668.0, + 1095.0, + 1702.0, + 1046.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 1668.0, + 1405.0, + 1668.0, + 1405.0, + 1702.0, + 1122.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1697.0, + 1406.0, + 1697.0, + 1406.0, + 1733.0, + 292.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1729.0, + 1405.0, + 1729.0, + 1405.0, + 1763.0, + 294.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1758.0, + 388.0, + 1758.0, + 388.0, + 1796.0, + 292.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 1758.0, + 443.0, + 1758.0, + 443.0, + 1796.0, + 411.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 1758.0, + 870.0, + 1758.0, + 870.0, + 1796.0, + 541.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1758.0, + 950.0, + 1758.0, + 950.0, + 1796.0, + 890.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 1758.0, + 1405.0, + 1758.0, + 1405.0, + 1796.0, + 1017.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1788.0, + 894.0, + 1788.0, + 894.0, + 1826.0, + 294.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1788.0, + 948.0, + 1788.0, + 948.0, + 1826.0, + 917.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 1788.0, + 1182.0, + 1788.0, + 1182.0, + 1826.0, + 1036.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 1788.0, + 1406.0, + 1788.0, + 1406.0, + 1826.0, + 1203.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1819.0, + 412.0, + 1819.0, + 412.0, + 1857.0, + 294.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 1819.0, + 1405.0, + 1819.0, + 1405.0, + 1857.0, + 629.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1849.0, + 528.0, + 1849.0, + 528.0, + 1891.0, + 291.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 1849.0, + 812.0, + 1849.0, + 812.0, + 1891.0, + 595.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 1849.0, + 1405.0, + 1849.0, + 1405.0, + 1891.0, + 1070.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1919.0, + 294.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1910.0, + 1406.0, + 1910.0, + 1406.0, + 1947.0, + 291.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1940.0, + 1406.0, + 1940.0, + 1406.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2001.0, + 1404.0, + 2001.0, + 1404.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1528.0, + 1405.0, + 1528.0, + 1405.0, + 1563.0, + 294.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1561.0, + 1403.0, + 1561.0, + 1403.0, + 1593.0, + 294.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1590.0, + 1404.0, + 1590.0, + 1404.0, + 1625.0, + 293.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1621.0, + 950.0, + 1621.0, + 950.0, + 1654.0, + 294.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 665.0, + 1405.0, + 665.0, + 1405.0, + 704.0, + 293.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 698.0, + 1405.0, + 698.0, + 1405.0, + 731.0, + 295.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 727.0, + 1405.0, + 727.0, + 1405.0, + 764.0, + 295.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 759.0, + 1405.0, + 759.0, + 1405.0, + 796.0, + 294.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 789.0, + 1015.0, + 789.0, + 1015.0, + 825.0, + 294.0, + 825.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 468, + 1403, + 468, + 1403, + 623, + 297, + 623 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1700, + 1404, + 1700, + 1404, + 1856, + 299, + 1856 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 862, + 1404, + 862, + 1404, + 987, + 297, + 987 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1941, + 1402, + 1941, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 1403, + 228, + 1403, + 385, + 298, + 385 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1144, + 1403, + 1144, + 1403, + 1269, + 299, + 1269 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 683, + 1403, + 683, + 1403, + 776, + 299, + 776 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1454, + 1403, + 1454, + 1403, + 1546, + 299, + 1546 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 300, + 1300, + 1404, + 1300, + 1404, + 1394, + 300, + 1394 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 293, + 1578, + 1399, + 1578, + 1399, + 1642, + 293, + 1642 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 298, + 1019, + 1404, + 1019, + 1404, + 1082, + 298, + 1082 + ], + "score": 0.952 + }, + { + "category_id": 8, + "poly": [ + 686, + 1403, + 1011, + 1403, + 1011, + 1445, + 686, + 1445 + ], + "score": 0.947 + }, + { + "category_id": 8, + "poly": [ + 765, + 633, + 931, + 633, + 931, + 674, + 765, + 674 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 719, + 1091, + 978, + 1091, + 978, + 1136, + 719, + 1136 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 722, + 1650, + 975, + 1650, + 975, + 1693, + 722, + 1693 + ], + "score": 0.941 + }, + { + "category_id": 0, + "poly": [ + 299, + 808, + 656, + 808, + 656, + 840, + 299, + 840 + ], + "score": 0.922 + }, + { + "category_id": 0, + "poly": [ + 299, + 1887, + 587, + 1887, + 587, + 1919, + 299, + 1919 + ], + "score": 0.921 + }, + { + "category_id": 0, + "poly": [ + 299, + 416, + 561, + 416, + 561, + 448, + 299, + 448 + ], + "score": 0.914 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.905 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1097, + 1399, + 1097, + 1399, + 1126, + 1366, + 1126 + ], + "score": 0.901 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1658, + 1399, + 1658, + 1399, + 1686, + 1366, + 1686 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1366, + 640, + 1399, + 640, + 1399, + 669, + 1366, + 669 + ], + "score": 0.887 + }, + { + "category_id": 9, + "poly": [ + 1367, + 1409, + 1399, + 1409, + 1399, + 1439, + 1367, + 1439 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.771 + }, + { + "category_id": 13, + "poly": [ + 649, + 1049, + 842, + 1049, + 842, + 1084, + 649, + 1084 + ], + "score": 0.94, + "latex": "( u , v ) \\in \\bar { \\mathcal { N } } _ { o u t } \\bar { ( u ) }" + }, + { + "category_id": 14, + "poly": [ + 723, + 1650, + 974, + 1650, + 974, + 1692, + 723, + 1692 + ], + "score": 0.93, + "latex": "h _ { u } ^ { t + 1 } = U _ { p _ { u } } ( h _ { u } ^ { t } , \\bar { m } _ { u } ^ { t } ) ," + }, + { + "category_id": 14, + "poly": [ + 765, + 631, + 932, + 631, + 932, + 673, + 765, + 673 + ], + "score": 0.93, + "latex": "h _ { u } ^ { 0 } = F _ { \\mathrm { i n } } ( x _ { u } ) ," + }, + { + "category_id": 14, + "poly": [ + 685, + 1402, + 1012, + 1402, + 1012, + 1444, + 685, + 1444 + ], + "score": 0.93, + "latex": "\\bar { m } _ { u } ^ { t } = A ( \\{ h _ { v } ^ { t } | v \\in \\mathcal { N } _ { i n } ( u ) \\} ) ," + }, + { + "category_id": 14, + "poly": [ + 719, + 1091, + 979, + 1091, + 979, + 1136, + 719, + 1136 + ], + "score": 0.92, + "latex": "m _ { ( u , v ) } ^ { t } = M _ { c _ { ( u , v ) } } ( h _ { u } ^ { t } ) ," + }, + { + "category_id": 13, + "poly": [ + 298, + 1821, + 441, + 1821, + 441, + 1858, + 298, + 1858 + ], + "score": 0.92, + "latex": "\\{ h _ { u } ^ { T } | u \\in \\bar { V } \\}" + }, + { + "category_id": 13, + "poly": [ + 553, + 713, + 586, + 713, + 586, + 747, + 553, + 747 + ], + "score": 0.91, + "latex": "h _ { u } ^ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 368, + 1484, + 410, + 1484, + 410, + 1518, + 368, + 1518 + ], + "score": 0.91, + "latex": "\\hat { m } _ { u } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 381, + 1147, + 465, + 1147, + 465, + 1183, + 381, + 1183 + ], + "score": 0.91, + "latex": "M _ { c _ { ( u , v ) } }" + }, + { + "category_id": 13, + "poly": [ + 1124, + 471, + 1206, + 471, + 1206, + 499, + 1124, + 499 + ], + "score": 0.9, + "latex": "s ^ { \\tau } \\in \\boldsymbol { S }" + }, + { + "category_id": 13, + "poly": [ + 431, + 1050, + 464, + 1050, + 464, + 1083, + 431, + 1083 + ], + "score": 0.9, + "latex": "h _ { u } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 406, + 1184, + 469, + 1184, + 469, + 1211, + 406, + 1211 + ], + "score": 0.89, + "latex": "c _ { ( u , v ) }" + }, + { + "category_id": 13, + "poly": [ + 298, + 716, + 335, + 716, + 335, + 745, + 298, + 745 + ], + "score": 0.88, + "latex": "F _ { \\mathrm { i n } }" + }, + { + "category_id": 13, + "poly": [ + 897, + 1737, + 930, + 1737, + 930, + 1764, + 897, + 1764 + ], + "score": 0.87, + "latex": "p _ { u }" + }, + { + "category_id": 13, + "poly": [ + 374, + 503, + 402, + 503, + 402, + 529, + 374, + 529 + ], + "score": 0.86, + "latex": "s ^ { \\tau }" + }, + { + "category_id": 13, + "poly": [ + 373, + 534, + 403, + 534, + 403, + 560, + 373, + 560 + ], + "score": 0.86, + "latex": "s ^ { \\tau }" + }, + { + "category_id": 13, + "poly": [ + 739, + 535, + 771, + 535, + 771, + 562, + 739, + 562 + ], + "score": 0.85, + "latex": "x _ { u }" + }, + { + "category_id": 13, + "poly": [ + 875, + 1795, + 899, + 1795, + 899, + 1821, + 875, + 1821 + ], + "score": 0.84, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 968, + 1734, + 992, + 1734, + 992, + 1761, + 968, + 1761 + ], + "score": 0.82, + "latex": "U" + }, + { + "category_id": 13, + "poly": [ + 379, + 1456, + 403, + 1456, + 403, + 1482, + 379, + 1482 + ], + "score": 0.8, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 386, + 1703, + 411, + 1703, + 411, + 1729, + 386, + 1729 + ], + "score": 0.79, + "latex": "U" + }, + { + "category_id": 13, + "poly": [ + 919, + 1618, + 938, + 1618, + 938, + 1638, + 919, + 1638 + ], + "score": 0.77, + "latex": "u" + }, + { + "category_id": 13, + "poly": [ + 507, + 477, + 526, + 477, + 526, + 498, + 507, + 498 + ], + "score": 0.77, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 660, + 537, + 679, + 537, + 679, + 559, + 660, + 559 + ], + "score": 0.75, + "latex": "u" + }, + { + "category_id": 13, + "poly": [ + 957, + 1024, + 970, + 1024, + 970, + 1047, + 957, + 1047 + ], + "score": 0.75, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 914, + 1055, + 934, + 1055, + 934, + 1077, + 914, + 1077 + ], + "score": 0.75, + "latex": "u" + }, + { + "category_id": 13, + "poly": [ + 877, + 720, + 897, + 720, + 897, + 742, + 877, + 742 + ], + "score": 0.75, + "latex": "u" + }, + { + "category_id": 13, + "poly": [ + 725, + 328, + 744, + 328, + 744, + 349, + 725, + 349 + ], + "score": 0.74, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1303, + 324, + 1318, + 324, + 1318, + 349, + 1303, + 349 + ], + "score": 0.74, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1189, + 1338, + 1208, + 1338, + 1208, + 1359, + 1189, + 1359 + ], + "score": 0.72, + "latex": "u" + }, + { + "category_id": 13, + "poly": [ + 1150, + 1027, + 1169, + 1027, + 1169, + 1047, + 1150, + 1047 + ], + "score": 0.71, + "latex": "u" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 805.0, + 659.0, + 805.0, + 659.0, + 844.0, + 294.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1884.0, + 592.0, + 1884.0, + 592.0, + 1924.0, + 293.0, + 1924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 415.0, + 566.0, + 415.0, + 566.0, + 452.0, + 295.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 472.0, + 506.0, + 472.0, + 506.0, + 502.0, + 297.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 472.0, + 1123.0, + 472.0, + 1123.0, + 502.0, + 527.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 472.0, + 1403.0, + 472.0, + 1403.0, + 502.0, + 1207.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 500.0, + 373.0, + 500.0, + 373.0, + 536.0, + 293.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 500.0, + 1405.0, + 500.0, + 1405.0, + 536.0, + 403.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 530.0, + 372.0, + 530.0, + 372.0, + 567.0, + 295.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 530.0, + 659.0, + 530.0, + 659.0, + 567.0, + 404.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 530.0, + 738.0, + 530.0, + 738.0, + 567.0, + 680.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 530.0, + 1405.0, + 530.0, + 1405.0, + 567.0, + 772.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 560.0, + 1407.0, + 560.0, + 1407.0, + 597.0, + 293.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 589.0, + 692.0, + 589.0, + 692.0, + 628.0, + 291.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1699.0, + 385.0, + 1699.0, + 385.0, + 1735.0, + 306.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 1699.0, + 1404.0, + 1699.0, + 1404.0, + 1735.0, + 412.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1733.0, + 896.0, + 1733.0, + 896.0, + 1766.0, + 297.0, + 1766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 1733.0, + 967.0, + 1733.0, + 967.0, + 1766.0, + 931.0, + 1766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 1733.0, + 1404.0, + 1733.0, + 1404.0, + 1766.0, + 993.0, + 1766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1761.0, + 1405.0, + 1761.0, + 1405.0, + 1798.0, + 295.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1791.0, + 874.0, + 1791.0, + 874.0, + 1830.0, + 293.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1791.0, + 1406.0, + 1791.0, + 1406.0, + 1830.0, + 900.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 1819.0, + 455.0, + 1819.0, + 455.0, + 1863.0, + 442.0, + 1863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 860.0, + 1405.0, + 860.0, + 1405.0, + 900.0, + 294.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 894.0, + 1405.0, + 894.0, + 1405.0, + 928.0, + 292.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 924.0, + 1405.0, + 924.0, + 1405.0, + 960.0, + 294.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 955.0, + 1267.0, + 955.0, + 1267.0, + 990.0, + 292.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1940.0, + 1406.0, + 1940.0, + 1406.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1973.0, + 1402.0, + 1973.0, + 1402.0, + 2003.0, + 297.0, + 2003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2001.0, + 1230.0, + 2001.0, + 1230.0, + 2039.0, + 292.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 228.0, + 1405.0, + 228.0, + 1405.0, + 266.0, + 292.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 262.0, + 1403.0, + 262.0, + 1403.0, + 296.0, + 296.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 290.0, + 1405.0, + 290.0, + 1405.0, + 328.0, + 294.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 319.0, + 724.0, + 319.0, + 724.0, + 357.0, + 292.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 319.0, + 1302.0, + 319.0, + 1302.0, + 357.0, + 745.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1319.0, + 319.0, + 1405.0, + 319.0, + 1405.0, + 357.0, + 1319.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 354.0, + 494.0, + 354.0, + 494.0, + 389.0, + 292.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 1144.0, + 1404.0, + 1144.0, + 1404.0, + 1178.0, + 466.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1173.0, + 405.0, + 1173.0, + 405.0, + 1216.0, + 292.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1173.0, + 1408.0, + 1173.0, + 1408.0, + 1216.0, + 470.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1205.0, + 1405.0, + 1205.0, + 1405.0, + 1240.0, + 294.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1236.0, + 739.0, + 1236.0, + 739.0, + 1272.0, + 294.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 1140.0, + 480.0, + 1140.0, + 480.0, + 1186.0, + 300.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 681.0, + 1403.0, + 681.0, + 1403.0, + 720.0, + 295.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 713.0, + 552.0, + 713.0, + 552.0, + 750.0, + 336.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 713.0, + 876.0, + 713.0, + 876.0, + 750.0, + 587.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 713.0, + 1407.0, + 713.0, + 1407.0, + 750.0, + 898.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 745.0, + 1330.0, + 745.0, + 1330.0, + 779.0, + 295.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 1451.0, + 378.0, + 1451.0, + 378.0, + 1490.0, + 302.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1451.0, + 1402.0, + 1451.0, + 1402.0, + 1490.0, + 404.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1480.0, + 367.0, + 1480.0, + 367.0, + 1522.0, + 293.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 1480.0, + 1407.0, + 1480.0, + 1407.0, + 1522.0, + 411.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1513.0, + 464.0, + 1513.0, + 464.0, + 1549.0, + 294.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1297.0, + 1405.0, + 1297.0, + 1405.0, + 1339.0, + 294.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1329.0, + 1188.0, + 1329.0, + 1188.0, + 1368.0, + 293.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 1329.0, + 1405.0, + 1329.0, + 1405.0, + 1368.0, + 1209.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1362.0, + 556.0, + 1362.0, + 556.0, + 1397.0, + 296.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1573.0, + 1405.0, + 1573.0, + 1405.0, + 1618.0, + 292.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1609.0, + 918.0, + 1609.0, + 918.0, + 1646.0, + 293.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 1609.0, + 1326.0, + 1609.0, + 1326.0, + 1646.0, + 939.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1019.0, + 956.0, + 1019.0, + 956.0, + 1054.0, + 296.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 1019.0, + 1149.0, + 1019.0, + 1149.0, + 1054.0, + 971.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 1019.0, + 1406.0, + 1019.0, + 1406.0, + 1054.0, + 1170.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1047.0, + 430.0, + 1047.0, + 430.0, + 1086.0, + 293.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1047.0, + 648.0, + 1047.0, + 648.0, + 1086.0, + 465.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1047.0, + 913.0, + 1047.0, + 913.0, + 1086.0, + 843.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 1047.0, + 1356.0, + 1047.0, + 1356.0, + 1086.0, + 935.0, + 1086.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1114, + 1405, + 1114, + 1405, + 1359, + 296, + 1359 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 227, + 1405, + 227, + 1405, + 475, + 297, + 475 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1817, + 1404, + 1817, + 1404, + 2035, + 297, + 2035 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1598, + 1405, + 1598, + 1405, + 1790, + 297, + 1790 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 521, + 1403, + 521, + 1403, + 643, + 299, + 643 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 884, + 1404, + 884, + 1404, + 1006, + 297, + 1006 + ], + "score": 0.978 + }, + { + "category_id": 8, + "poly": [ + 328, + 1363, + 1336, + 1363, + 1336, + 1592, + 328, + 1592 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 731, + 1401, + 731, + 1401, + 794, + 298, + 794 + ], + "score": 0.961 + }, + { + "category_id": 8, + "poly": [ + 671, + 1007, + 1025, + 1007, + 1025, + 1099, + 671, + 1099 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 493, + 650, + 1204, + 650, + 1204, + 726, + 493, + 726 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 741, + 480, + 956, + 480, + 956, + 520, + 741, + 520 + ], + "score": 0.938 + }, + { + "category_id": 0, + "poly": [ + 299, + 828, + 651, + 828, + 651, + 859, + 299, + 859 + ], + "score": 0.912 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 104, + 299, + 104 + ], + "score": 0.903 + }, + { + "category_id": 9, + "poly": [ + 1366, + 484, + 1400, + 484, + 1400, + 514, + 1366, + 514 + ], + "score": 0.879 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1036, + 1400, + 1036, + 1400, + 1066, + 1366, + 1066 + ], + "score": 0.876 + }, + { + "category_id": 9, + "poly": [ + 1366, + 665, + 1400, + 665, + 1400, + 695, + 1366, + 695 + ], + "score": 0.865 + }, + { + "category_id": 9, + "poly": [ + 1367, + 1530, + 1400, + 1530, + 1400, + 1560, + 1367, + 1560 + ], + "score": 0.834 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.74 + }, + { + "category_id": 14, + "poly": [ + 671, + 1003, + 1028, + 1003, + 1028, + 1099, + 671, + 1099 + ], + "score": 0.94, + "latex": "J ( \\theta ) = \\mathbb { E } _ { \\pi } \\left[ \\sum _ { \\tau = 0 } ^ { \\infty } \\gamma ^ { \\tau } r ( s ^ { \\tau } , a ^ { \\tau } ) \\right] ." + }, + { + "category_id": 14, + "poly": [ + 323, + 1357, + 1339, + 1357, + 1339, + 1596, + 323, + 1596 + ], + "score": 0.94, + "latex": "\\begin{array} { l } { \\displaystyle \\tilde { J } ( \\theta ) = J ( \\theta ) - \\beta L _ { K L } ( \\theta ) - \\alpha L _ { V } ( \\theta ) } \\\\ { \\displaystyle \\qquad = \\mathbb { E } _ { \\pi _ { \\theta } } \\left[ \\sum _ { \\tau = 0 } ^ { \\infty } \\operatorname* { m i n } \\left( \\hat { A } ^ { \\tau } r ^ { \\tau } ( \\theta ) , \\hat { A } ^ { \\tau } \\operatorname { c l i p } \\left( r ^ { \\tau } ( \\theta ) , 1 - \\epsilon , 1 + \\epsilon \\right) \\right) \\right] } \\\\ { \\displaystyle \\qquad - \\beta \\mathbb { E } _ { \\pi _ { \\theta } } \\left[ \\sum _ { \\tau = 0 } ^ { \\infty } \\mathrm { K L } \\left[ \\pi _ { \\theta } ( \\cdot | s ^ { \\tau } ) | \\pi _ { \\theta _ { o l d } } ( \\cdot | s ^ { \\tau } ) \\right] \\right] - \\alpha \\mathbb { E } _ { \\pi _ { \\theta } } \\left[ \\sum _ { \\tau = 0 } ^ { \\infty } \\left( V _ { \\theta } ( s ^ { \\tau } ) - V ( s ^ { \\tau } ) ^ { \\mathrm { t a r g e t } } \\right) ^ { 2 } \\right] , } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 495, + 646, + 1201, + 646, + 1201, + 728, + 495, + 728 + ], + "score": 0.94, + "latex": "\\pi _ { \\boldsymbol { \\theta } } ( a ^ { \\tau } \\vert s ^ { \\tau } ) = \\prod _ { \\boldsymbol { u } \\in \\mathcal { O } } \\pi _ { \\boldsymbol { \\theta } , \\boldsymbol { u } } ( a _ { \\boldsymbol { u } } ^ { \\tau } \\vert s ^ { \\tau } ) = \\prod _ { \\boldsymbol { u } \\in \\mathcal { O } } \\frac { 1 } { \\sqrt { 2 \\pi \\sigma _ { \\boldsymbol { u } } ^ { 2 } } } e ^ { ( a _ { \\boldsymbol { u } } ^ { \\tau } - \\mu _ { \\boldsymbol { u } } ) ^ { 2 } / ( 2 \\sigma _ { \\boldsymbol { u } } ^ { 2 } ) } ," + }, + { + "category_id": 13, + "poly": [ + 538, + 383, + 586, + 383, + 586, + 416, + 538, + 416 + ], + "score": 0.93, + "latex": "O _ { q _ { u } }" + }, + { + "category_id": 13, + "poly": [ + 807, + 1820, + 887, + 1820, + 887, + 1853, + 807, + 1853 + ], + "score": 0.93, + "latex": "V _ { \\theta } ( s ^ { \\tau } )" + }, + { + "category_id": 13, + "poly": [ + 927, + 945, + 1061, + 945, + 1061, + 980, + 927, + 980 + ], + "score": 0.93, + "latex": "\\mathsf { \\bar { P } } ( s ^ { \\tau + 1 } | s ^ { \\tau } )" + }, + { + "category_id": 14, + "poly": [ + 742, + 477, + 956, + 477, + 956, + 520, + 742, + 520 + ], + "score": 0.92, + "latex": "\\mu _ { u \\in \\mathcal { O } } = O _ { q _ { u } } ( h _ { u } ^ { T } ) ," + }, + { + "category_id": 13, + "poly": [ + 1188, + 289, + 1256, + 289, + 1256, + 327, + 1188, + 327 + ], + "score": 0.92, + "latex": "h _ { u \\in \\mathcal { O } } ^ { T }" + }, + { + "category_id": 13, + "poly": [ + 1138, + 918, + 1242, + 918, + 1242, + 949, + 1138, + 949 + ], + "score": 0.92, + "latex": "r ( s ^ { \\tau } , a ^ { \\tau } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1722, + 353, + 1722, + 353, + 1760, + 298, + 1760 + ], + "score": 0.91, + "latex": "\\tilde { J } ( \\theta )" + }, + { + "category_id": 13, + "poly": [ + 501, + 1267, + 928, + 1267, + 928, + 1301, + 501, + 1301 + ], + "score": 0.91, + "latex": "r ^ { \\tau } ( \\boldsymbol { \\theta } ; \\mathbf { \\dot { \\theta } } _ { o l d } ) = \\pi _ { \\boldsymbol { \\theta } } ( a ^ { \\tau } | s ^ { \\tau } ) / \\pi _ { \\theta _ { o l d } } ( a ^ { \\tau } | \\mathbf { \\bar { s } } ^ { \\tau } )" + }, + { + "category_id": 13, + "poly": [ + 419, + 1691, + 533, + 1691, + 533, + 1725, + 419, + 1725 + ], + "score": 0.91, + "latex": "V ( s _ { t } ) ^ { \\mathrm { t a r g e t } }" + }, + { + "category_id": 13, + "poly": [ + 379, + 732, + 468, + 732, + 468, + 759, + 379, + 759 + ], + "score": 0.91, + "latex": "a ^ { \\tau } \\in { \\mathcal { A } }" + }, + { + "category_id": 13, + "poly": [ + 1141, + 886, + 1255, + 886, + 1255, + 918, + 1141, + 918 + ], + "score": 0.9, + "latex": "\\pi _ { \\boldsymbol { \\theta } } ( \\boldsymbol { a } ^ { \\tau } | \\boldsymbol { s } ^ { \\tau } )" + }, + { + "category_id": 13, + "poly": [ + 371, + 1595, + 405, + 1595, + 405, + 1630, + 371, + 1630 + ], + "score": 0.9, + "latex": "\\hat { A } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 440, + 353, + 525, + 353, + 525, + 381, + 440, + 381 + ], + "score": 0.9, + "latex": "u \\in \\mathcal { O }" + }, + { + "category_id": 13, + "poly": [ + 381, + 527, + 449, + 527, + 449, + 554, + 381, + 554 + ], + "score": 0.88, + "latex": "\\mu _ { u \\in \\mathcal { O } }" + }, + { + "category_id": 13, + "poly": [ + 860, + 357, + 890, + 357, + 890, + 385, + 860, + 385 + ], + "score": 0.86, + "latex": "q _ { u }" + }, + { + "category_id": 13, + "poly": [ + 745, + 734, + 778, + 734, + 778, + 761, + 745, + 761 + ], + "score": 0.86, + "latex": "\\sigma _ { u }" + }, + { + "category_id": 13, + "poly": [ + 1139, + 1823, + 1168, + 1823, + 1168, + 1849, + 1139, + 1849 + ], + "score": 0.85, + "latex": "s ^ { \\tau }" + }, + { + "category_id": 13, + "poly": [ + 1090, + 1631, + 1111, + 1631, + 1111, + 1663, + 1090, + 1663 + ], + "score": 0.84, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 887, + 1665, + 910, + 1665, + 910, + 1688, + 887, + 1688 + ], + "score": 0.8, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 298, + 763, + 315, + 763, + 315, + 789, + 298, + 789 + ], + "score": 0.8, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 506, + 1636, + 522, + 1636, + 522, + 1658, + 506, + 1658 + ], + "score": 0.75, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 521, + 293, + 546, + 293, + 546, + 319, + 521, + 319 + ], + "score": 0.63, + "latex": "\\mathcal { O }" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 827.0, + 654.0, + 827.0, + 654.0, + 863.0, + 294.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1116.0, + 1403.0, + 1116.0, + 1403.0, + 1149.0, + 296.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1145.0, + 1405.0, + 1145.0, + 1405.0, + 1179.0, + 294.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1178.0, + 1405.0, + 1178.0, + 1405.0, + 1211.0, + 294.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1207.0, + 1405.0, + 1207.0, + 1405.0, + 1240.0, + 294.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1234.0, + 1406.0, + 1234.0, + 1406.0, + 1272.0, + 291.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1265.0, + 500.0, + 1265.0, + 500.0, + 1306.0, + 291.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1265.0, + 1406.0, + 1265.0, + 1406.0, + 1306.0, + 929.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1294.0, + 1403.0, + 1294.0, + 1403.0, + 1335.0, + 291.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1328.0, + 827.0, + 1328.0, + 827.0, + 1361.0, + 295.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 1403.0, + 230.0, + 1403.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 294.0, + 294.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 285.0, + 520.0, + 285.0, + 520.0, + 335.0, + 287.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 285.0, + 1187.0, + 285.0, + 1187.0, + 335.0, + 547.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 285.0, + 1410.0, + 285.0, + 1410.0, + 335.0, + 1257.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 322.0, + 1405.0, + 322.0, + 1405.0, + 356.0, + 294.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 353.0, + 439.0, + 353.0, + 439.0, + 387.0, + 294.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 353.0, + 859.0, + 353.0, + 859.0, + 387.0, + 526.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 353.0, + 1405.0, + 353.0, + 1405.0, + 387.0, + 891.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 381.0, + 537.0, + 381.0, + 537.0, + 419.0, + 292.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 381.0, + 1406.0, + 381.0, + 1406.0, + 419.0, + 587.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 410.0, + 1406.0, + 410.0, + 1406.0, + 450.0, + 292.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 442.0, + 835.0, + 442.0, + 835.0, + 477.0, + 294.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1818.0, + 806.0, + 1818.0, + 806.0, + 1856.0, + 296.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 1818.0, + 1138.0, + 1818.0, + 1138.0, + 1856.0, + 888.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1818.0, + 1405.0, + 1818.0, + 1405.0, + 1856.0, + 1169.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1849.0, + 1407.0, + 1849.0, + 1407.0, + 1887.0, + 292.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1880.0, + 1406.0, + 1880.0, + 1406.0, + 1919.0, + 295.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1909.0, + 1405.0, + 1909.0, + 1405.0, + 1950.0, + 294.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1976.0, + 292.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2005.0, + 294.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2003.0, + 920.0, + 2003.0, + 920.0, + 2038.0, + 296.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1595.0, + 370.0, + 1595.0, + 370.0, + 1634.0, + 295.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1595.0, + 1406.0, + 1595.0, + 1406.0, + 1634.0, + 406.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1629.0, + 505.0, + 1629.0, + 505.0, + 1664.0, + 294.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1629.0, + 1089.0, + 1629.0, + 1089.0, + 1664.0, + 523.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 1629.0, + 1405.0, + 1629.0, + 1405.0, + 1664.0, + 1112.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1660.0, + 886.0, + 1660.0, + 886.0, + 1693.0, + 295.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 1660.0, + 1403.0, + 1660.0, + 1403.0, + 1693.0, + 911.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1687.0, + 418.0, + 1687.0, + 418.0, + 1728.0, + 292.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 1687.0, + 1406.0, + 1687.0, + 1406.0, + 1728.0, + 534.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1720.0, + 297.0, + 1720.0, + 297.0, + 1766.0, + 294.0, + 1766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1720.0, + 1406.0, + 1720.0, + 1406.0, + 1766.0, + 354.0, + 1766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1757.0, + 452.0, + 1757.0, + 452.0, + 1792.0, + 293.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 521.0, + 380.0, + 521.0, + 380.0, + 557.0, + 303.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 521.0, + 1407.0, + 521.0, + 1407.0, + 557.0, + 450.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 552.0, + 1404.0, + 552.0, + 1404.0, + 584.0, + 294.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 582.0, + 1404.0, + 582.0, + 1404.0, + 620.0, + 292.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 614.0, + 734.0, + 614.0, + 734.0, + 647.0, + 295.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 884.0, + 1140.0, + 884.0, + 1140.0, + 920.0, + 294.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 884.0, + 1405.0, + 884.0, + 1405.0, + 920.0, + 1256.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 915.0, + 1137.0, + 915.0, + 1137.0, + 951.0, + 294.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 915.0, + 1405.0, + 915.0, + 1405.0, + 951.0, + 1243.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 943.0, + 926.0, + 943.0, + 926.0, + 982.0, + 292.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 943.0, + 1406.0, + 943.0, + 1406.0, + 982.0, + 1062.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 977.0, + 647.0, + 977.0, + 647.0, + 1009.0, + 294.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 727.0, + 378.0, + 727.0, + 378.0, + 765.0, + 302.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 727.0, + 744.0, + 727.0, + 744.0, + 765.0, + 469.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 727.0, + 1405.0, + 727.0, + 1405.0, + 765.0, + 779.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 761.0, + 297.0, + 761.0, + 297.0, + 797.0, + 294.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 761.0, + 858.0, + 761.0, + 858.0, + 797.0, + 316.0, + 797.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 519, + 1404, + 519, + 1404, + 887, + 298, + 887 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1094, + 1404, + 1094, + 1404, + 1551, + 298, + 1551 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 299, + 1589, + 1404, + 1589, + 1404, + 1896, + 299, + 1896 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 298, + 1404, + 298, + 1404, + 483, + 298, + 483 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 925, + 1403, + 925, + 1403, + 1079, + 299, + 1079 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 300, + 1911, + 1402, + 1911, + 1402, + 2034, + 300, + 2034 + ], + "score": 0.973 + }, + { + "category_id": 0, + "poly": [ + 301, + 225, + 587, + 225, + 587, + 262, + 301, + 262 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.766 + }, + { + "category_id": 15, + "poly": [ + 292.0, + 223.0, + 593.0, + 223.0, + 593.0, + 267.0, + 292.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 520.0, + 1401.0, + 520.0, + 1401.0, + 554.0, + 296.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 552.0, + 1405.0, + 552.0, + 1405.0, + 586.0, + 294.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 579.0, + 1405.0, + 579.0, + 1405.0, + 619.0, + 292.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 612.0, + 1406.0, + 612.0, + 1406.0, + 648.0, + 293.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 642.0, + 1402.0, + 642.0, + 1402.0, + 676.0, + 294.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 674.0, + 1404.0, + 674.0, + 1404.0, + 708.0, + 293.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 704.0, + 1404.0, + 704.0, + 1404.0, + 738.0, + 294.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 732.0, + 1404.0, + 732.0, + 1404.0, + 768.0, + 293.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 763.0, + 1406.0, + 763.0, + 1406.0, + 802.0, + 292.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 794.0, + 1405.0, + 794.0, + 1405.0, + 829.0, + 294.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 826.0, + 1406.0, + 826.0, + 1406.0, + 860.0, + 294.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 854.0, + 1303.0, + 854.0, + 1303.0, + 891.0, + 293.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1094.0, + 1405.0, + 1094.0, + 1405.0, + 1129.0, + 296.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1124.0, + 1403.0, + 1124.0, + 1403.0, + 1160.0, + 293.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1153.0, + 1404.0, + 1153.0, + 1404.0, + 1192.0, + 293.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1184.0, + 1404.0, + 1184.0, + 1404.0, + 1222.0, + 293.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1216.0, + 1405.0, + 1216.0, + 1405.0, + 1253.0, + 293.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1248.0, + 1405.0, + 1248.0, + 1405.0, + 1283.0, + 294.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1278.0, + 1405.0, + 1278.0, + 1405.0, + 1313.0, + 292.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1308.0, + 1406.0, + 1308.0, + 1406.0, + 1345.0, + 292.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1339.0, + 1405.0, + 1339.0, + 1405.0, + 1373.0, + 294.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1369.0, + 1406.0, + 1369.0, + 1406.0, + 1403.0, + 294.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1398.0, + 1405.0, + 1398.0, + 1405.0, + 1433.0, + 294.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1424.0, + 1405.0, + 1424.0, + 1405.0, + 1468.0, + 292.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1458.0, + 1404.0, + 1458.0, + 1404.0, + 1495.0, + 293.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1488.0, + 1406.0, + 1488.0, + 1406.0, + 1529.0, + 292.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1523.0, + 526.0, + 1523.0, + 526.0, + 1555.0, + 293.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1591.0, + 1404.0, + 1591.0, + 1404.0, + 1623.0, + 297.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1620.0, + 1403.0, + 1620.0, + 1403.0, + 1652.0, + 295.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1651.0, + 1405.0, + 1651.0, + 1405.0, + 1687.0, + 294.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1680.0, + 1404.0, + 1680.0, + 1404.0, + 1718.0, + 293.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1715.0, + 1404.0, + 1715.0, + 1404.0, + 1746.0, + 294.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1740.0, + 1405.0, + 1740.0, + 1405.0, + 1782.0, + 292.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 1405.0, + 1773.0, + 1405.0, + 1809.0, + 294.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1805.0, + 1404.0, + 1805.0, + 1404.0, + 1837.0, + 295.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1833.0, + 1404.0, + 1833.0, + 1404.0, + 1870.0, + 293.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1863.0, + 723.0, + 1863.0, + 723.0, + 1899.0, + 295.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 296.0, + 1408.0, + 296.0, + 1408.0, + 335.0, + 293.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 329.0, + 1405.0, + 329.0, + 1405.0, + 365.0, + 294.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 359.0, + 1406.0, + 359.0, + 1406.0, + 397.0, + 293.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 392.0, + 1402.0, + 392.0, + 1402.0, + 424.0, + 296.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 418.0, + 1405.0, + 418.0, + 1405.0, + 457.0, + 293.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 452.0, + 1389.0, + 452.0, + 1389.0, + 485.0, + 293.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 924.0, + 1402.0, + 924.0, + 1402.0, + 961.0, + 295.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 957.0, + 1404.0, + 957.0, + 1404.0, + 990.0, + 295.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 983.0, + 1405.0, + 983.0, + 1405.0, + 1022.0, + 293.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1013.0, + 1405.0, + 1013.0, + 1405.0, + 1052.0, + 293.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1047.0, + 904.0, + 1047.0, + 904.0, + 1084.0, + 294.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1912.0, + 1403.0, + 1912.0, + 1403.0, + 1945.0, + 295.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1974.0, + 1406.0, + 1974.0, + 1406.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2004.0, + 1406.0, + 2004.0, + 1406.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1559, + 1404, + 1559, + 1404, + 1774, + 298, + 1774 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1789, + 1403, + 1789, + 1403, + 2034, + 299, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 314, + 234, + 1375, + 234, + 1375, + 975, + 314, + 975 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1130, + 1403, + 1130, + 1403, + 1253, + 299, + 1253 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1370, + 1403, + 1370, + 1403, + 1462, + 299, + 1462 + ], + "score": 0.973 + }, + { + "category_id": 4, + "poly": [ + 298, + 1005, + 1400, + 1005, + 1400, + 1069, + 298, + 1069 + ], + "score": 0.939 + }, + { + "category_id": 0, + "poly": [ + 300, + 1300, + 557, + 1300, + 557, + 1334, + 300, + 1334 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.885 + }, + { + "category_id": 0, + "poly": [ + 297, + 1501, + 1041, + 1501, + 1041, + 1533, + 297, + 1533 + ], + "score": 0.832 + }, + { + "category_id": 2, + "poly": [ + 842, + 2087, + 858, + 2087, + 858, + 2111, + 842, + 2111 + ], + "score": 0.647 + }, + { + "category_id": 2, + "poly": [ + 842, + 2087, + 859, + 2087, + 859, + 2111, + 842, + 2111 + ], + "score": 0.272 + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 245.0, + 1240.0, + 245.0, + 1240.0, + 257.0, + 1206.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 259.0, + 345.0, + 259.0, + 345.0, + 272.0, + 327.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 256.0, + 397.0, + 256.0, + 397.0, + 275.0, + 346.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 260.0, + 1078.0, + 260.0, + 1078.0, + 269.0, + 1065.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 254.0, + 1133.0, + 254.0, + 1133.0, + 274.0, + 1086.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 268.0, + 386.0, + 268.0, + 386.0, + 281.0, + 364.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 268.0, + 754.0, + 268.0, + 754.0, + 281.0, + 734.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 269.0, + 1121.0, + 269.0, + 1121.0, + 281.0, + 1102.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 278.0, + 387.0, + 278.0, + 387.0, + 287.0, + 368.0, + 287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 281.0, + 344.0, + 281.0, + 344.0, + 294.0, + 326.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 305.0, + 1079.0, + 305.0, + 1079.0, + 313.0, + 1065.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 330.0, + 341.0, + 330.0, + 341.0, + 338.0, + 331.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 345.0, + 326.0, + 345.0, + 326.0, + 378.0, + 317.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 352.0, + 341.0, + 352.0, + 341.0, + 362.0, + 331.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 354.0, + 696.0, + 354.0, + 696.0, + 363.0, + 687.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 452.0, + 1001.0, + 452.0, + 1001.0, + 460.0, + 978.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 452.0, + 1369.0, + 452.0, + 1369.0, + 460.0, + 1355.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 460.0, + 499.0, + 460.0, + 499.0, + 468.0, + 476.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 457.0, + 887.0, + 457.0, + 887.0, + 471.0, + 820.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 456.0, + 1253.0, + 456.0, + 1253.0, + 472.0, + 1188.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 499.0, + 494.0, + 499.0, + 494.0, + 508.0, + 474.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 497.0, + 882.0, + 497.0, + 882.0, + 509.0, + 834.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 497.0, + 1263.0, + 497.0, + 1263.0, + 509.0, + 1199.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 513.0, + 393.0, + 513.0, + 393.0, + 524.0, + 355.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 515.0, + 622.0, + 515.0, + 622.0, + 612.0, + 395.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 510.0, + 1131.0, + 510.0, + 1131.0, + 526.0, + 1089.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 530.0, + 1290.0, + 530.0, + 1290.0, + 539.0, + 1277.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 594.0, + 327.0, + 594.0, + 327.0, + 606.0, + 318.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 606.0, + 328.0, + 606.0, + 328.0, + 631.0, + 320.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 653.0, + 710.0, + 653.0, + 710.0, + 663.0, + 697.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 654.0, + 991.0, + 654.0, + 991.0, + 674.0, + 944.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 670.0, + 978.0, + 670.0, + 978.0, + 679.0, + 962.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 705.0, + 415.0, + 705.0, + 415.0, + 713.0, + 399.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 712.0, + 497.0, + 712.0, + 497.0, + 720.0, + 474.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 708.0, + 888.0, + 708.0, + 888.0, + 724.0, + 819.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 705.0, + 944.0, + 705.0, + 944.0, + 713.0, + 929.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 702.0, + 1001.0, + 702.0, + 1001.0, + 714.0, + 976.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 708.0, + 1245.0, + 708.0, + 1245.0, + 724.0, + 1188.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 749.0, + 684.0, + 749.0, + 684.0, + 761.0, + 655.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 831.0, + 513.0, + 831.0, + 513.0, + 860.0, + 501.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 830.0, + 886.0, + 830.0, + 886.0, + 859.0, + 874.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 857.0, + 512.0, + 857.0, + 512.0, + 882.0, + 503.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 859.0, + 884.0, + 859.0, + 884.0, + 884.0, + 876.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 860.0, + 897.0, + 860.0, + 897.0, + 869.0, + 887.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 880.0, + 526.0, + 880.0, + 526.0, + 889.0, + 517.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 892.0, + 895.0, + 892.0, + 895.0, + 898.0, + 889.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 909.0, + 808.0, + 909.0, + 808.0, + 928.0, + 759.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 908.0, + 1175.0, + 908.0, + 1175.0, + 927.0, + 1129.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 921.0, + 795.0, + 921.0, + 795.0, + 933.0, + 776.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 923.0, + 1161.0, + 923.0, + 1161.0, + 932.0, + 1148.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 939.0, + 806.0, + 939.0, + 806.0, + 951.0, + 783.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 959.0, + 703.0, + 959.0, + 703.0, + 975.0, + 635.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 957.0, + 761.0, + 957.0, + 761.0, + 965.0, + 745.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 955.0, + 814.0, + 955.0, + 814.0, + 966.0, + 792.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 962.0, + 1073.0, + 962.0, + 1073.0, + 974.0, + 1005.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 955.0, + 1132.0, + 955.0, + 1132.0, + 966.0, + 1112.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 243.5, + 875.0, + 243.5, + 875.0, + 255.5, + 832.0, + 255.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 256.5, + 762.0, + 256.5, + 762.0, + 273.0, + 719.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 275.0, + 673.0, + 275.0, + 673.0, + 393.0, + 352.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 280.0, + 1010.0, + 280.0, + 1010.0, + 363.0, + 712.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 300.5, + 1399.0, + 300.5, + 1399.0, + 400.0, + 1074.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 502.5, + 1374.0, + 502.5, + 1374.0, + 612.5, + 1106.0, + 612.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.75, + 612.5, + 1359.75, + 612.5, + 1359.75, + 678.5, + 1184.75, + 678.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1004.0, + 1404.0, + 1004.0, + 1404.0, + 1040.0, + 294.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1032.0, + 373.0, + 1032.0, + 373.0, + 1077.0, + 291.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1297.0, + 561.0, + 1297.0, + 561.0, + 1340.0, + 293.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1501.0, + 1043.0, + 1501.0, + 1043.0, + 1537.0, + 295.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 861.0, + 2086.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1559.0, + 1404.0, + 1559.0, + 1404.0, + 1594.0, + 294.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1588.0, + 1406.0, + 1588.0, + 1406.0, + 1626.0, + 295.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1620.0, + 1406.0, + 1620.0, + 1406.0, + 1655.0, + 293.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1652.0, + 1405.0, + 1652.0, + 1405.0, + 1686.0, + 294.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1680.0, + 1402.0, + 1680.0, + 1402.0, + 1714.0, + 293.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1711.0, + 1408.0, + 1711.0, + 1408.0, + 1748.0, + 293.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1742.0, + 920.0, + 1742.0, + 920.0, + 1779.0, + 295.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1788.0, + 1403.0, + 1788.0, + 1403.0, + 1822.0, + 294.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1819.0, + 1404.0, + 1819.0, + 1404.0, + 1853.0, + 294.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1850.0, + 1403.0, + 1850.0, + 1403.0, + 1883.0, + 293.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1915.0, + 293.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1911.0, + 1403.0, + 1911.0, + 1403.0, + 1945.0, + 294.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1940.0, + 1404.0, + 1940.0, + 1404.0, + 1975.0, + 294.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2005.0, + 294.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1999.0, + 1358.0, + 1999.0, + 1358.0, + 2040.0, + 294.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1131.0, + 1403.0, + 1131.0, + 1403.0, + 1164.0, + 294.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1161.0, + 1403.0, + 1161.0, + 1403.0, + 1194.0, + 295.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1190.0, + 1403.0, + 1190.0, + 1403.0, + 1225.0, + 294.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1221.0, + 1244.0, + 1221.0, + 1244.0, + 1255.0, + 294.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1370.0, + 1405.0, + 1370.0, + 1405.0, + 1404.0, + 294.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1400.0, + 1400.0, + 1400.0, + 1400.0, + 1434.0, + 294.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1432.0, + 1393.0, + 1432.0, + 1393.0, + 1468.0, + 293.0, + 1468.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1180, + 1404, + 1180, + 1404, + 1576, + 297, + 1576 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 838, + 1404, + 838, + 1404, + 1144, + 298, + 1144 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 496, + 1404, + 496, + 1404, + 742, + 297, + 742 + ], + "score": 0.98 + }, + { + "category_id": 5, + "poly": [ + 314, + 282, + 1384, + 282, + 1384, + 440, + 314, + 440 + ], + "score": 0.976, + "html": "
C_FourReward AvgStdMaxC_SixReward AvgStdMax
NerveNet2799.91247.23833.9NerveNet2507.1738.42979.2
MLP2398.51390.43936.3MLP2793.01005.23529.5
TreeNet1429.6957.73021.7TreeNet2229.71109.43727.4
" + }, + { + "category_id": 1, + "poly": [ + 300, + 1612, + 1403, + 1612, + 1403, + 1704, + 300, + 1704 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 1792, + 1402, + 1792, + 1402, + 1854, + 299, + 1854 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 301, + 1890, + 1398, + 1890, + 1398, + 1952, + 301, + 1952 + ], + "score": 0.928 + }, + { + "category_id": 2, + "poly": [ + 296, + 1978, + 1402, + 1978, + 1402, + 2034, + 296, + 2034 + ], + "score": 0.927 + }, + { + "category_id": 6, + "poly": [ + 311, + 224, + 1371, + 224, + 1371, + 257, + 311, + 257 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 302, + 781, + 781, + 781, + 781, + 812, + 302, + 812 + ], + "score": 0.881 + }, + { + "category_id": 0, + "poly": [ + 300, + 1739, + 703, + 1739, + 703, + 1770, + 300, + 1770 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.802 + }, + { + "category_id": 13, + "poly": [ + 951, + 1214, + 969, + 1214, + 969, + 1241, + 951, + 1241 + ], + "score": 0.8, + "latex": "y" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 1976.0, + 1404.0, + 1976.0, + 1404.0, + 2010.0, + 333.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 780.0, + 2002.0, + 780.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 218.0, + 1376.0, + 218.0, + 1376.0, + 260.0, + 321.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 778.0, + 784.0, + 778.0, + 784.0, + 815.0, + 293.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1737.0, + 707.0, + 1737.0, + 707.0, + 1773.0, + 295.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2117.0, + 838.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1179.0, + 1401.0, + 1179.0, + 1401.0, + 1211.0, + 297.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1209.0, + 950.0, + 1209.0, + 950.0, + 1245.0, + 295.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 1209.0, + 1405.0, + 1209.0, + 1405.0, + 1245.0, + 970.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1242.0, + 1405.0, + 1242.0, + 1405.0, + 1275.0, + 296.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1272.0, + 1404.0, + 1272.0, + 1404.0, + 1303.0, + 294.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1300.0, + 1405.0, + 1300.0, + 1405.0, + 1338.0, + 294.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1332.0, + 1405.0, + 1332.0, + 1405.0, + 1368.0, + 294.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1362.0, + 1405.0, + 1362.0, + 1405.0, + 1399.0, + 295.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1392.0, + 1406.0, + 1392.0, + 1406.0, + 1428.0, + 294.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1423.0, + 1406.0, + 1423.0, + 1406.0, + 1459.0, + 295.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1452.0, + 1407.0, + 1452.0, + 1407.0, + 1492.0, + 294.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1483.0, + 1404.0, + 1483.0, + 1404.0, + 1519.0, + 295.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1516.0, + 1405.0, + 1516.0, + 1405.0, + 1548.0, + 294.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1545.0, + 799.0, + 1545.0, + 799.0, + 1581.0, + 295.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 837.0, + 1405.0, + 837.0, + 1405.0, + 872.0, + 296.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 868.0, + 1406.0, + 868.0, + 1406.0, + 905.0, + 292.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 901.0, + 1404.0, + 901.0, + 1404.0, + 933.0, + 296.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 929.0, + 1406.0, + 929.0, + 1406.0, + 966.0, + 292.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 957.0, + 1405.0, + 957.0, + 1405.0, + 999.0, + 292.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 990.0, + 1406.0, + 990.0, + 1406.0, + 1025.0, + 294.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1020.0, + 1406.0, + 1020.0, + 1406.0, + 1056.0, + 293.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1053.0, + 1402.0, + 1053.0, + 1402.0, + 1085.0, + 296.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1080.0, + 1405.0, + 1080.0, + 1405.0, + 1120.0, + 292.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1113.0, + 664.0, + 1113.0, + 664.0, + 1148.0, + 294.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 494.0, + 1402.0, + 494.0, + 1402.0, + 534.0, + 292.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 529.0, + 1405.0, + 529.0, + 1405.0, + 563.0, + 294.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 559.0, + 1404.0, + 559.0, + 1404.0, + 591.0, + 294.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 589.0, + 1405.0, + 589.0, + 1405.0, + 623.0, + 295.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 619.0, + 1404.0, + 619.0, + 1404.0, + 653.0, + 294.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 647.0, + 1405.0, + 647.0, + 1405.0, + 688.0, + 292.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 681.0, + 1404.0, + 681.0, + 1404.0, + 715.0, + 294.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 712.0, + 578.0, + 712.0, + 578.0, + 746.0, + 295.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1610.0, + 1404.0, + 1610.0, + 1404.0, + 1647.0, + 295.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1640.0, + 1405.0, + 1640.0, + 1405.0, + 1676.0, + 295.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1672.0, + 1343.0, + 1672.0, + 1343.0, + 1706.0, + 295.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1791.0, + 1406.0, + 1791.0, + 1406.0, + 1827.0, + 297.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1823.0, + 794.0, + 1823.0, + 794.0, + 1858.0, + 295.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1887.0, + 1402.0, + 1887.0, + 1402.0, + 1926.0, + 296.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1918.0, + 1403.0, + 1918.0, + 1403.0, + 1958.0, + 296.0, + 1958.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 318, + 913, + 1348, + 913, + 1348, + 1711, + 318, + 1711 + ], + "score": 0.976 + }, + { + "category_id": 4, + "poly": [ + 297, + 1732, + 1400, + 1732, + 1400, + 1794, + 297, + 1794 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 297, + 1863, + 1400, + 1863, + 1400, + 1927, + 297, + 1927 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 295, + 1971, + 1400, + 1971, + 1400, + 2035, + 295, + 2035 + ], + "score": 0.941 + }, + { + "category_id": 4, + "poly": [ + 294, + 785, + 1408, + 785, + 1408, + 850, + 294, + 850 + ], + "score": 0.908 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.802 + }, + { + "category_id": 3, + "poly": [ + 296, + 224, + 1421, + 224, + 1421, + 765, + 296, + 765 + ], + "score": 0.345 + }, + { + "category_id": 13, + "poly": [ + 1288, + 549, + 1327, + 549, + 1327, + 567, + 1288, + 567 + ], + "score": 0.46, + "latex": "( 9 9 \\% )" + }, + { + "category_id": 13, + "poly": [ + 1284, + 353, + 1324, + 353, + 1324, + 371, + 1284, + 371 + ], + "score": 0.44, + "latex": "( 8 7 \\% )" + }, + { + "category_id": 13, + "poly": [ + 1288, + 279, + 1329, + 279, + 1329, + 298, + 1288, + 298 + ], + "score": 0.41, + "latex": "( 9 2 \\% )" + }, + { + "category_id": 13, + "poly": [ + 1288, + 328, + 1328, + 328, + 1328, + 347, + 1288, + 347 + ], + "score": 0.31, + "latex": "( 9 0 \\% )" + }, + { + "category_id": 13, + "poly": [ + 406, + 233, + 440, + 233, + 440, + 251, + 406, + 251 + ], + "score": 0.29, + "latex": "( 3 2 \\% )" + }, + { + "category_id": 13, + "poly": [ + 777, + 307, + 817, + 307, + 817, + 325, + 777, + 325 + ], + "score": 0.29, + "latex": "( 8 8 \\% )" + }, + { + "category_id": 13, + "poly": [ + 1293, + 622, + 1330, + 622, + 1330, + 640, + 1293, + 640 + ], + "score": 0.29, + "latex": "( 9 9 \\% )" + }, + { + "category_id": 13, + "poly": [ + 1287, + 304, + 1328, + 304, + 1328, + 322, + 1287, + 322 + ], + "score": 0.26, + "latex": "( 9 1 \\% )" + }, + { + "category_id": 13, + "poly": [ + 494, + 306, + 528, + 306, + 528, + 326, + 494, + 326 + ], + "score": 0.26, + "latex": "( 7 2 \\% )" + }, + { + "category_id": 13, + "poly": [ + 1284, + 401, + 1325, + 401, + 1325, + 421, + 1284, + 421 + ], + "score": 0.25, + "latex": "( 9 8 \\% )" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 914.0, + 680.0, + 914.0, + 680.0, + 938.0, + 485.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 916.0, + 1231.0, + 916.0, + 1231.0, + 938.0, + 1028.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 933.0, + 375.0, + 933.0, + 375.0, + 954.0, + 340.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 937.0, + 460.0, + 937.0, + 460.0, + 960.0, + 408.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 937.0, + 1006.0, + 937.0, + 1006.0, + 960.0, + 954.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 957.0, + 445.0, + 957.0, + 445.0, + 973.0, + 412.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 962.0, + 920.0, + 962.0, + 920.0, + 983.0, + 886.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 957.0, + 990.0, + 957.0, + 990.0, + 973.0, + 957.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 971.0, + 375.0, + 971.0, + 375.0, + 992.0, + 340.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 972.0, + 473.0, + 972.0, + 473.0, + 987.0, + 412.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 971.0, + 1020.0, + 971.0, + 1020.0, + 989.0, + 957.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 987.0, + 449.0, + 987.0, + 449.0, + 1003.0, + 410.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 985.0, + 997.0, + 985.0, + 997.0, + 1004.0, + 955.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1007.0, + 375.0, + 1007.0, + 375.0, + 1028.0, + 340.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 1001.0, + 470.0, + 1001.0, + 470.0, + 1033.0, + 413.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 1000.0, + 920.0, + 1000.0, + 920.0, + 1021.0, + 886.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 1001.0, + 1028.0, + 1001.0, + 1028.0, + 1033.0, + 958.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1028.0, + 374.0, + 1028.0, + 374.0, + 1131.0, + 322.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 1033.0, + 440.0, + 1033.0, + 440.0, + 1058.0, + 414.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1027.0, + 920.0, + 1027.0, + 920.0, + 1133.0, + 867.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 1031.0, + 999.0, + 1031.0, + 999.0, + 1062.0, + 958.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 1062.0, + 467.0, + 1062.0, + 467.0, + 1077.0, + 433.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 1059.0, + 1033.0, + 1059.0, + 1033.0, + 1078.0, + 957.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1080.0, + 375.0, + 1080.0, + 375.0, + 1100.0, + 341.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 1078.0, + 920.0, + 1078.0, + 920.0, + 1098.0, + 886.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 1118.0, + 375.0, + 1118.0, + 375.0, + 1138.0, + 342.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 1116.0, + 921.0, + 1116.0, + 921.0, + 1135.0, + 888.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1154.0, + 375.0, + 1154.0, + 375.0, + 1176.0, + 346.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 1155.0, + 920.0, + 1155.0, + 920.0, + 1173.0, + 893.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1195.0, + 373.0, + 1195.0, + 373.0, + 1209.0, + 359.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1195.0, + 919.0, + 1195.0, + 919.0, + 1209.0, + 905.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1229.0, + 416.0, + 1229.0, + 416.0, + 1240.0, + 407.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 1223.0, + 523.0, + 1223.0, + 523.0, + 1243.0, + 480.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 1223.0, + 618.0, + 1223.0, + 618.0, + 1243.0, + 569.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1223.0, + 707.0, + 1223.0, + 707.0, + 1243.0, + 660.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 1223.0, + 799.0, + 1223.0, + 799.0, + 1243.0, + 749.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1225.0, + 964.0, + 1225.0, + 964.0, + 1241.0, + 951.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1026.0, + 1223.0, + 1070.0, + 1223.0, + 1070.0, + 1243.0, + 1026.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 1223.0, + 1163.0, + 1223.0, + 1163.0, + 1243.0, + 1114.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1223.0, + 1253.0, + 1223.0, + 1253.0, + 1243.0, + 1205.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 1223.0, + 1345.0, + 1223.0, + 1345.0, + 1242.0, + 1296.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 1235.0, + 646.0, + 1235.0, + 646.0, + 1260.0, + 520.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 1235.0, + 1192.0, + 1235.0, + 1192.0, + 1260.0, + 1066.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1265.0, + 602.0, + 1265.0, + 602.0, + 1307.0, + 553.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 1265.0, + 1148.0, + 1265.0, + 1148.0, + 1307.0, + 1098.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 1325.0, + 689.0, + 1325.0, + 689.0, + 1347.0, + 479.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 1325.0, + 1231.0, + 1325.0, + 1231.0, + 1347.0, + 1025.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 1338.0, + 921.0, + 1338.0, + 921.0, + 1359.0, + 886.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1347.0, + 460.0, + 1347.0, + 460.0, + 1370.0, + 408.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1349.0, + 1006.0, + 1349.0, + 1006.0, + 1369.0, + 954.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1360.0, + 375.0, + 1360.0, + 375.0, + 1380.0, + 340.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 1366.0, + 442.0, + 1366.0, + 442.0, + 1381.0, + 412.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 1365.0, + 990.0, + 1365.0, + 990.0, + 1381.0, + 958.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 1379.0, + 474.0, + 1379.0, + 474.0, + 1397.0, + 410.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 1376.0, + 920.0, + 1376.0, + 920.0, + 1397.0, + 886.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1377.0, + 1019.0, + 1377.0, + 1019.0, + 1397.0, + 955.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1397.0, + 376.0, + 1397.0, + 376.0, + 1422.0, + 339.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 1391.0, + 482.0, + 1391.0, + 482.0, + 1441.0, + 410.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1415.0, + 920.0, + 1415.0, + 920.0, + 1435.0, + 887.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1393.0, + 1030.0, + 1393.0, + 1030.0, + 1442.0, + 955.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1437.0, + 374.0, + 1437.0, + 374.0, + 1539.0, + 323.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 1441.0, + 447.0, + 1441.0, + 447.0, + 1456.0, + 412.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1436.0, + 894.0, + 1436.0, + 894.0, + 1538.0, + 868.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 1438.0, + 995.0, + 1438.0, + 995.0, + 1458.0, + 957.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 1456.0, + 452.0, + 1456.0, + 452.0, + 1471.0, + 413.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 1453.0, + 920.0, + 1453.0, + 920.0, + 1471.0, + 885.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 1455.0, + 999.0, + 1455.0, + 999.0, + 1471.0, + 957.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 1472.0, + 441.0, + 1472.0, + 441.0, + 1485.0, + 414.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 1469.0, + 1034.0, + 1469.0, + 1034.0, + 1487.0, + 957.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 1480.0, + 375.0, + 1480.0, + 375.0, + 1500.0, + 342.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1491.0, + 921.0, + 1491.0, + 921.0, + 1511.0, + 887.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 1522.0, + 373.0, + 1522.0, + 373.0, + 1540.0, + 343.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 1528.0, + 921.0, + 1528.0, + 921.0, + 1549.0, + 888.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1562.0, + 374.0, + 1562.0, + 374.0, + 1580.0, + 347.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 1567.0, + 921.0, + 1567.0, + 921.0, + 1588.0, + 892.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1605.0, + 373.0, + 1605.0, + 373.0, + 1620.0, + 359.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1609.0, + 918.0, + 1609.0, + 918.0, + 1620.0, + 906.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1637.0, + 416.0, + 1637.0, + 416.0, + 1648.0, + 406.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 1632.0, + 523.0, + 1632.0, + 523.0, + 1652.0, + 480.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 1632.0, + 618.0, + 1632.0, + 618.0, + 1651.0, + 569.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1632.0, + 708.0, + 1632.0, + 708.0, + 1652.0, + 660.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 1632.0, + 799.0, + 1632.0, + 799.0, + 1651.0, + 749.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1637.0, + 962.0, + 1637.0, + 962.0, + 1648.0, + 952.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1026.0, + 1632.0, + 1070.0, + 1632.0, + 1070.0, + 1652.0, + 1026.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 1632.0, + 1164.0, + 1632.0, + 1164.0, + 1652.0, + 1116.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 1632.0, + 1253.0, + 1632.0, + 1253.0, + 1652.0, + 1206.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 1632.0, + 1345.0, + 1632.0, + 1345.0, + 1651.0, + 1296.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 1643.0, + 646.0, + 1643.0, + 646.0, + 1669.0, + 520.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 1643.0, + 1192.0, + 1643.0, + 1192.0, + 1668.0, + 1066.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1674.0, + 601.0, + 1674.0, + 601.0, + 1715.0, + 553.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 1673.0, + 1148.0, + 1673.0, + 1148.0, + 1715.0, + 1098.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1728.0, + 1405.0, + 1728.0, + 1405.0, + 1771.0, + 293.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1764.0, + 689.0, + 1764.0, + 689.0, + 1796.0, + 296.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 784.0, + 1405.0, + 784.0, + 1405.0, + 820.0, + 296.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 815.0, + 1345.0, + 815.0, + 1345.0, + 851.0, + 296.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 233.0, + 363.0, + 233.0, + 363.0, + 255.0, + 324.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 229.0, + 405.0, + 229.0, + 405.0, + 254.0, + 369.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 229.0, + 445.0, + 229.0, + 445.0, + 254.0, + 441.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 225.0, + 542.0, + 225.0, + 542.0, + 257.0, + 459.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 229.0, + 636.0, + 229.0, + 636.0, + 254.0, + 554.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 225.0, + 727.0, + 225.0, + 727.0, + 257.0, + 646.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 228.0, + 830.0, + 228.0, + 830.0, + 257.0, + 731.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 227.0, + 927.0, + 227.0, + 927.0, + 255.0, + 851.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 228.0, + 1029.0, + 228.0, + 1029.0, + 253.0, + 947.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 227.0, + 1126.0, + 227.0, + 1126.0, + 255.0, + 1051.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 228.0, + 1226.0, + 228.0, + 1226.0, + 253.0, + 1149.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 228.0, + 1337.0, + 228.0, + 1337.0, + 253.0, + 1239.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 257.0, + 365.0, + 257.0, + 365.0, + 279.0, + 324.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 252.0, + 448.0, + 252.0, + 448.0, + 282.0, + 367.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 252.0, + 538.0, + 252.0, + 538.0, + 282.0, + 462.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 254.0, + 639.0, + 254.0, + 639.0, + 279.0, + 550.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 253.0, + 726.0, + 253.0, + 726.0, + 282.0, + 647.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 250.0, + 828.0, + 250.0, + 828.0, + 283.0, + 733.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 247.0, + 930.0, + 247.0, + 930.0, + 281.0, + 848.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 250.0, + 1027.0, + 250.0, + 1027.0, + 279.0, + 950.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 250.0, + 1128.0, + 250.0, + 1128.0, + 279.0, + 1050.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 250.0, + 1229.0, + 250.0, + 1229.0, + 279.0, + 1148.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 249.0, + 1340.0, + 249.0, + 1340.0, + 282.0, + 1236.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 279.0, + 366.0, + 279.0, + 366.0, + 305.0, + 323.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 275.0, + 446.0, + 275.0, + 446.0, + 305.0, + 367.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 274.0, + 539.0, + 274.0, + 539.0, + 308.0, + 460.0, + 308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 274.0, + 731.0, + 274.0, + 731.0, + 308.0, + 548.0, + 308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 274.0, + 828.0, + 274.0, + 828.0, + 307.0, + 733.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 272.0, + 931.0, + 272.0, + 931.0, + 305.0, + 848.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 274.0, + 1027.0, + 274.0, + 1027.0, + 303.0, + 950.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 274.0, + 1128.0, + 274.0, + 1128.0, + 304.0, + 1050.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 274.0, + 1227.0, + 274.0, + 1227.0, + 304.0, + 1148.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 273.0, + 1287.0, + 273.0, + 1287.0, + 302.0, + 1237.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1330.0, + 273.0, + 1338.0, + 273.0, + 1338.0, + 302.0, + 1330.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 300.0, + 448.0, + 300.0, + 448.0, + 354.0, + 323.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 299.0, + 493.0, + 299.0, + 493.0, + 358.0, + 463.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 299.0, + 537.0, + 299.0, + 537.0, + 358.0, + 529.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 299.0, + 642.0, + 299.0, + 642.0, + 358.0, + 547.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 302.0, + 728.0, + 302.0, + 728.0, + 354.0, + 647.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 299.0, + 776.0, + 299.0, + 776.0, + 358.0, + 732.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 299.0, + 828.0, + 299.0, + 828.0, + 358.0, + 818.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 295.0, + 932.0, + 295.0, + 932.0, + 356.0, + 848.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 298.0, + 1027.0, + 298.0, + 1027.0, + 353.0, + 950.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 298.0, + 1128.0, + 298.0, + 1128.0, + 353.0, + 1049.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 298.0, + 1229.0, + 298.0, + 1229.0, + 353.0, + 1148.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 297.0, + 1286.0, + 297.0, + 1286.0, + 354.0, + 1236.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 297.0, + 1340.0, + 297.0, + 1340.0, + 354.0, + 1329.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1391.0, + 310.0, + 1419.0, + 310.0, + 1419.0, + 334.0, + 1391.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 353.0, + 365.0, + 353.0, + 365.0, + 379.0, + 310.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 346.0, + 447.0, + 346.0, + 447.0, + 380.0, + 367.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 349.0, + 538.0, + 349.0, + 538.0, + 379.0, + 464.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 351.0, + 634.0, + 351.0, + 634.0, + 379.0, + 555.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 348.0, + 828.0, + 348.0, + 828.0, + 381.0, + 636.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 346.0, + 931.0, + 346.0, + 931.0, + 379.0, + 848.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 345.0, + 1029.0, + 345.0, + 1029.0, + 378.0, + 949.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 347.0, + 1128.0, + 347.0, + 1128.0, + 377.0, + 1050.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 345.0, + 1229.0, + 345.0, + 1229.0, + 378.0, + 1148.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 346.0, + 1283.0, + 346.0, + 1283.0, + 379.0, + 1240.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1325.0, + 346.0, + 1334.0, + 346.0, + 1334.0, + 379.0, + 1325.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 378.0, + 365.0, + 378.0, + 365.0, + 403.0, + 310.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 374.0, + 448.0, + 374.0, + 448.0, + 403.0, + 367.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 374.0, + 536.0, + 374.0, + 536.0, + 404.0, + 464.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 373.0, + 638.0, + 373.0, + 638.0, + 406.0, + 552.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 374.0, + 725.0, + 374.0, + 725.0, + 404.0, + 648.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 373.0, + 828.0, + 373.0, + 828.0, + 406.0, + 735.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 371.0, + 931.0, + 371.0, + 931.0, + 404.0, + 848.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 371.0, + 1029.0, + 371.0, + 1029.0, + 404.0, + 948.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 372.0, + 1128.0, + 372.0, + 1128.0, + 402.0, + 1049.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 370.0, + 1229.0, + 370.0, + 1229.0, + 403.0, + 1148.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 371.0, + 1334.0, + 371.0, + 1334.0, + 403.0, + 1240.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 404.0, + 365.0, + 404.0, + 365.0, + 426.0, + 308.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 398.0, + 446.0, + 398.0, + 446.0, + 428.0, + 367.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 398.0, + 541.0, + 398.0, + 541.0, + 428.0, + 462.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 397.0, + 638.0, + 397.0, + 638.0, + 429.0, + 552.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 398.0, + 725.0, + 398.0, + 725.0, + 428.0, + 648.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 397.0, + 828.0, + 397.0, + 828.0, + 429.0, + 735.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 394.0, + 930.0, + 394.0, + 930.0, + 428.0, + 848.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 397.0, + 1028.0, + 397.0, + 1028.0, + 427.0, + 949.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 396.0, + 1128.0, + 396.0, + 1128.0, + 426.0, + 1050.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 397.0, + 1229.0, + 397.0, + 1229.0, + 427.0, + 1149.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 396.0, + 1283.0, + 396.0, + 1283.0, + 428.0, + 1239.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 396.0, + 1334.0, + 396.0, + 1334.0, + 428.0, + 1326.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1391.0, + 403.0, + 1419.0, + 403.0, + 1419.0, + 427.0, + 1391.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 426.0, + 365.0, + 426.0, + 365.0, + 452.0, + 323.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 423.0, + 445.0, + 423.0, + 445.0, + 453.0, + 370.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 423.0, + 536.0, + 423.0, + 536.0, + 453.0, + 464.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 423.0, + 636.0, + 423.0, + 636.0, + 452.0, + 552.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 422.0, + 834.0, + 422.0, + 834.0, + 454.0, + 645.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 422.0, + 926.0, + 422.0, + 926.0, + 451.0, + 852.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 422.0, + 1024.0, + 422.0, + 1024.0, + 451.0, + 954.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 422.0, + 1125.0, + 422.0, + 1125.0, + 451.0, + 1052.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 422.0, + 1221.0, + 422.0, + 1221.0, + 451.0, + 1156.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 422.0, + 1343.0, + 422.0, + 1343.0, + 451.0, + 1232.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 443.0, + 319.0, + 443.0, + 319.0, + 484.0, + 296.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 451.0, + 366.0, + 451.0, + 366.0, + 477.0, + 323.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 445.0, + 444.0, + 445.0, + 444.0, + 479.0, + 369.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 447.0, + 538.0, + 447.0, + 538.0, + 477.0, + 464.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 448.0, + 636.0, + 448.0, + 636.0, + 477.0, + 552.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 447.0, + 726.0, + 447.0, + 726.0, + 477.0, + 648.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 446.0, + 832.0, + 446.0, + 832.0, + 478.0, + 728.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 440.0, + 928.0, + 440.0, + 928.0, + 477.0, + 851.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 446.0, + 1026.0, + 446.0, + 1026.0, + 476.0, + 952.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 446.0, + 1125.0, + 446.0, + 1125.0, + 476.0, + 1051.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 446.0, + 1226.0, + 446.0, + 1226.0, + 476.0, + 1152.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 442.0, + 1342.0, + 442.0, + 1342.0, + 477.0, + 1233.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 474.0, + 366.0, + 474.0, + 366.0, + 501.0, + 323.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 469.0, + 447.0, + 469.0, + 447.0, + 501.0, + 368.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 472.0, + 541.0, + 472.0, + 541.0, + 501.0, + 460.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 469.0, + 642.0, + 469.0, + 642.0, + 502.0, + 548.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 472.0, + 726.0, + 472.0, + 726.0, + 501.0, + 648.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 471.0, + 832.0, + 471.0, + 832.0, + 503.0, + 728.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 468.0, + 927.0, + 468.0, + 927.0, + 501.0, + 852.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 471.0, + 1023.0, + 471.0, + 1023.0, + 501.0, + 955.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 471.0, + 1125.0, + 471.0, + 1125.0, + 501.0, + 1052.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 471.0, + 1226.0, + 471.0, + 1226.0, + 500.0, + 1151.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 471.0, + 1341.0, + 471.0, + 1341.0, + 500.0, + 1235.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 498.0, + 366.0, + 498.0, + 366.0, + 525.0, + 321.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 496.0, + 446.0, + 496.0, + 446.0, + 526.0, + 367.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 496.0, + 541.0, + 496.0, + 541.0, + 526.0, + 460.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 494.0, + 640.0, + 494.0, + 640.0, + 527.0, + 548.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 494.0, + 832.0, + 494.0, + 832.0, + 528.0, + 645.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 492.0, + 926.0, + 492.0, + 926.0, + 522.0, + 852.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 494.0, + 1023.0, + 494.0, + 1023.0, + 525.0, + 957.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 494.0, + 1125.0, + 494.0, + 1125.0, + 525.0, + 1052.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 494.0, + 1226.0, + 494.0, + 1226.0, + 525.0, + 1152.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 494.0, + 1341.0, + 494.0, + 1341.0, + 523.0, + 1236.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1391.0, + 496.0, + 1419.0, + 496.0, + 1419.0, + 520.0, + 1391.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 523.0, + 366.0, + 523.0, + 366.0, + 550.0, + 321.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 518.0, + 447.0, + 518.0, + 447.0, + 551.0, + 367.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 521.0, + 539.0, + 521.0, + 539.0, + 551.0, + 462.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 522.0, + 638.0, + 522.0, + 638.0, + 547.0, + 550.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 518.0, + 832.0, + 518.0, + 832.0, + 551.0, + 640.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 518.0, + 926.0, + 518.0, + 926.0, + 548.0, + 852.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 518.0, + 1023.0, + 518.0, + 1023.0, + 548.0, + 957.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 515.0, + 1128.0, + 515.0, + 1128.0, + 549.0, + 1048.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 515.0, + 1229.0, + 515.0, + 1229.0, + 549.0, + 1148.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 518.0, + 1341.0, + 518.0, + 1341.0, + 547.0, + 1235.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 547.0, + 366.0, + 547.0, + 366.0, + 573.0, + 321.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 545.0, + 446.0, + 545.0, + 446.0, + 575.0, + 367.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 545.0, + 541.0, + 545.0, + 541.0, + 575.0, + 462.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 545.0, + 639.0, + 545.0, + 639.0, + 573.0, + 548.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 543.0, + 726.0, + 543.0, + 726.0, + 575.0, + 648.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 539.0, + 832.0, + 539.0, + 832.0, + 578.0, + 729.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 543.0, + 929.0, + 543.0, + 929.0, + 573.0, + 848.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 543.0, + 1024.0, + 543.0, + 1024.0, + 572.0, + 953.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 543.0, + 1128.0, + 543.0, + 1128.0, + 572.0, + 1049.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 543.0, + 1226.0, + 543.0, + 1226.0, + 573.0, + 1152.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 542.0, + 1287.0, + 542.0, + 1287.0, + 575.0, + 1236.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 542.0, + 1340.0, + 542.0, + 1340.0, + 575.0, + 1328.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 575.0, + 365.0, + 575.0, + 365.0, + 597.0, + 323.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 567.0, + 447.0, + 567.0, + 447.0, + 600.0, + 367.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 567.0, + 537.0, + 567.0, + 537.0, + 601.0, + 464.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 570.0, + 640.0, + 570.0, + 640.0, + 598.0, + 550.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 567.0, + 725.0, + 567.0, + 725.0, + 600.0, + 648.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 567.0, + 828.0, + 567.0, + 828.0, + 601.0, + 732.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 567.0, + 927.0, + 567.0, + 927.0, + 597.0, + 850.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 568.0, + 1029.0, + 568.0, + 1029.0, + 597.0, + 948.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 568.0, + 1126.0, + 568.0, + 1126.0, + 597.0, + 1050.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 568.0, + 1230.0, + 568.0, + 1230.0, + 597.0, + 1147.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 566.0, + 1340.0, + 566.0, + 1340.0, + 598.0, + 1235.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 597.0, + 365.0, + 597.0, + 365.0, + 623.0, + 309.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 593.0, + 448.0, + 593.0, + 448.0, + 622.0, + 367.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 593.0, + 539.0, + 593.0, + 539.0, + 622.0, + 462.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 593.0, + 638.0, + 593.0, + 638.0, + 622.0, + 551.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 592.0, + 832.0, + 592.0, + 832.0, + 625.0, + 644.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 592.0, + 926.0, + 592.0, + 926.0, + 622.0, + 852.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 593.0, + 1027.0, + 593.0, + 1027.0, + 622.0, + 952.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 592.0, + 1125.0, + 592.0, + 1125.0, + 622.0, + 1052.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 592.0, + 1226.0, + 592.0, + 1226.0, + 622.0, + 1152.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 591.0, + 1342.0, + 591.0, + 1342.0, + 623.0, + 1234.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1392.0, + 591.0, + 1417.0, + 591.0, + 1417.0, + 611.0, + 1392.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 621.0, + 365.0, + 621.0, + 365.0, + 647.0, + 309.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 618.0, + 448.0, + 618.0, + 448.0, + 647.0, + 367.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 616.0, + 541.0, + 616.0, + 541.0, + 650.0, + 462.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 617.0, + 639.0, + 617.0, + 639.0, + 650.0, + 551.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 616.0, + 832.0, + 616.0, + 832.0, + 650.0, + 643.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 616.0, + 926.0, + 616.0, + 926.0, + 646.0, + 852.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 616.0, + 1023.0, + 616.0, + 1023.0, + 646.0, + 954.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 616.0, + 1125.0, + 616.0, + 1125.0, + 646.0, + 1052.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 616.0, + 1225.0, + 616.0, + 1225.0, + 646.0, + 1152.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 617.0, + 1292.0, + 617.0, + 1292.0, + 646.0, + 1235.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 617.0, + 1341.0, + 617.0, + 1341.0, + 646.0, + 1331.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 646.0, + 366.0, + 646.0, + 366.0, + 672.0, + 309.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 642.0, + 446.0, + 642.0, + 446.0, + 671.0, + 367.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 642.0, + 539.0, + 642.0, + 539.0, + 671.0, + 463.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 642.0, + 639.0, + 642.0, + 639.0, + 671.0, + 550.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 642.0, + 728.0, + 642.0, + 728.0, + 671.0, + 647.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 641.0, + 830.0, + 641.0, + 830.0, + 674.0, + 730.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 638.0, + 927.0, + 638.0, + 927.0, + 671.0, + 853.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 643.0, + 1019.0, + 643.0, + 1019.0, + 670.0, + 958.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 639.0, + 1126.0, + 639.0, + 1126.0, + 671.0, + 1052.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 641.0, + 1226.0, + 641.0, + 1226.0, + 671.0, + 1152.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 640.0, + 1342.0, + 640.0, + 1342.0, + 672.0, + 1235.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 670.0, + 365.0, + 670.0, + 365.0, + 696.0, + 309.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 667.0, + 446.0, + 667.0, + 446.0, + 715.0, + 367.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 667.0, + 538.0, + 667.0, + 538.0, + 716.0, + 463.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 666.0, + 640.0, + 666.0, + 640.0, + 716.0, + 550.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 666.0, + 832.0, + 666.0, + 832.0, + 716.0, + 647.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 665.0, + 926.0, + 665.0, + 926.0, + 714.0, + 852.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 666.0, + 1022.0, + 666.0, + 1022.0, + 714.0, + 957.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 667.0, + 1123.0, + 667.0, + 1123.0, + 715.0, + 1052.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 665.0, + 1225.0, + 665.0, + 1225.0, + 714.0, + 1152.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 665.0, + 1342.0, + 665.0, + 1342.0, + 715.0, + 1234.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1392.0, + 684.0, + 1416.0, + 684.0, + 1416.0, + 702.0, + 1392.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 705.0, + 621.0, + 705.0, + 621.0, + 730.0, + 568.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 706.0, + 1114.0, + 706.0, + 1114.0, + 727.0, + 1063.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 729.0, + 735.0, + 729.0, + 735.0, + 762.0, + 395.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 727.0, + 1332.0, + 727.0, + 1332.0, + 767.0, + 902.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1860.0, + 1406.0, + 1860.0, + 1406.0, + 1903.0, + 291.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1895.0, + 516.0, + 1895.0, + 516.0, + 1928.0, + 292.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1970.0, + 1405.0, + 1970.0, + 1405.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2004.0, + 1401.0, + 2004.0, + 1401.0, + 2036.0, + 296.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 318, + 625, + 1372, + 625, + 1372, + 1438, + 318, + 1438 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1666, + 1404, + 1666, + 1404, + 1853, + 298, + 1853 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 300, + 1561, + 1400, + 1561, + 1400, + 1623, + 300, + 1623 + ], + "score": 0.95 + }, + { + "category_id": 4, + "poly": [ + 296, + 498, + 1403, + 498, + 1403, + 561, + 296, + 561 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 298, + 1896, + 1399, + 1896, + 1399, + 1958, + 298, + 1958 + ], + "score": 0.945 + }, + { + "category_id": 3, + "poly": [ + 404, + 228, + 1297, + 228, + 1297, + 478, + 404, + 478 + ], + "score": 0.933 + }, + { + "category_id": 4, + "poly": [ + 542, + 1460, + 1155, + 1460, + 1155, + 1492, + 542, + 1492 + ], + "score": 0.919 + }, + { + "category_id": 1, + "poly": [ + 293, + 2002, + 1380, + 2002, + 1380, + 2035, + 293, + 2035 + ], + "score": 0.908 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 865, + 2088, + 865, + 2112, + 837, + 2112 + ], + "score": 0.853 + }, + { + "category_id": 13, + "poly": [ + 1148, + 327, + 1191, + 327, + 1191, + 346, + 1148, + 346 + ], + "score": 0.28, + "latex": "( 9 5 \\% )" + }, + { + "category_id": 13, + "poly": [ + 1147, + 418, + 1192, + 418, + 1192, + 437, + 1147, + 437 + ], + "score": 0.27, + "latex": "( 9 4 \\% )" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 631.0, + 662.0, + 631.0, + 662.0, + 650.0, + 500.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 631.0, + 1216.0, + 631.0, + 1216.0, + 650.0, + 1057.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 659.0, + 368.0, + 659.0, + 368.0, + 677.0, + 340.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 680.0, + 922.0, + 680.0, + 922.0, + 697.0, + 897.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 718.0, + 368.0, + 718.0, + 368.0, + 735.0, + 340.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 731.0, + 923.0, + 731.0, + 923.0, + 749.0, + 895.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 745.0, + 349.0, + 745.0, + 349.0, + 847.0, + 324.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 744.0, + 904.0, + 744.0, + 904.0, + 850.0, + 877.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 770.0, + 1290.0, + 770.0, + 1290.0, + 778.0, + 1273.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 776.0, + 367.0, + 776.0, + 367.0, + 794.0, + 340.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 798.0, + 753.0, + 798.0, + 753.0, + 825.0, + 668.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 782.0, + 922.0, + 782.0, + 922.0, + 801.0, + 897.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 793.0, + 1083.0, + 793.0, + 1083.0, + 801.0, + 1069.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 796.0, + 1309.0, + 796.0, + 1309.0, + 826.0, + 1222.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 816.0, + 791.0, + 816.0, + 791.0, + 841.0, + 660.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 819.0, + 1346.0, + 819.0, + 1346.0, + 841.0, + 1216.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 835.0, + 368.0, + 835.0, + 368.0, + 853.0, + 341.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 834.0, + 748.0, + 834.0, + 748.0, + 857.0, + 660.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 832.0, + 1303.0, + 832.0, + 1303.0, + 858.0, + 1216.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 851.0, + 700.0, + 851.0, + 700.0, + 872.0, + 661.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 851.0, + 1254.0, + 851.0, + 1254.0, + 872.0, + 1218.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 866.0, + 762.0, + 866.0, + 762.0, + 889.0, + 661.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 866.0, + 1317.0, + 866.0, + 1317.0, + 889.0, + 1216.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 884.0, + 713.0, + 884.0, + 713.0, + 904.0, + 663.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 889.0, + 921.0, + 889.0, + 921.0, + 901.0, + 910.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 884.0, + 1268.0, + 884.0, + 1268.0, + 904.0, + 1219.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 896.0, + 367.0, + 896.0, + 367.0, + 911.0, + 352.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 900.0, + 719.0, + 900.0, + 719.0, + 920.0, + 662.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 900.0, + 1274.0, + 900.0, + 1274.0, + 920.0, + 1218.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 911.0, + 769.0, + 911.0, + 769.0, + 937.0, + 660.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 915.0, + 1323.0, + 915.0, + 1323.0, + 934.0, + 1218.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 946.0, + 382.0, + 946.0, + 382.0, + 961.0, + 368.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 945.0, + 480.0, + 945.0, + 480.0, + 961.0, + 437.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 945.0, + 563.0, + 945.0, + 563.0, + 960.0, + 522.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 945.0, + 645.0, + 945.0, + 645.0, + 961.0, + 603.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 942.0, + 731.0, + 942.0, + 731.0, + 963.0, + 685.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 944.0, + 816.0, + 944.0, + 816.0, + 963.0, + 767.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 946.0, + 938.0, + 946.0, + 938.0, + 960.0, + 923.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 945.0, + 1035.0, + 945.0, + 1035.0, + 961.0, + 993.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 945.0, + 1118.0, + 945.0, + 1118.0, + 961.0, + 1077.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 945.0, + 1201.0, + 945.0, + 1201.0, + 961.0, + 1159.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 942.0, + 1286.0, + 942.0, + 1286.0, + 963.0, + 1240.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1322.0, + 944.0, + 1372.0, + 944.0, + 1372.0, + 963.0, + 1322.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 953.0, + 647.0, + 953.0, + 647.0, + 978.0, + 517.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 952.0, + 1201.0, + 952.0, + 1201.0, + 979.0, + 1073.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 988.0, + 599.0, + 988.0, + 599.0, + 1026.0, + 552.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 985.0, + 1155.0, + 985.0, + 1155.0, + 1027.0, + 1106.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 1047.0, + 659.0, + 1047.0, + 659.0, + 1066.0, + 506.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 1047.0, + 1209.0, + 1047.0, + 1209.0, + 1066.0, + 1064.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1060.0, + 924.0, + 1060.0, + 924.0, + 1081.0, + 894.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1096.0, + 368.0, + 1096.0, + 368.0, + 1114.0, + 341.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1109.0, + 922.0, + 1109.0, + 922.0, + 1127.0, + 897.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1147.0, + 368.0, + 1147.0, + 368.0, + 1165.0, + 340.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1161.0, + 349.0, + 1161.0, + 349.0, + 1263.0, + 324.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 1165.0, + 740.0, + 1165.0, + 740.0, + 1173.0, + 722.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1160.0, + 892.0, + 1160.0, + 892.0, + 1262.0, + 869.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 1156.0, + 922.0, + 1156.0, + 922.0, + 1174.0, + 895.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1198.0, + 367.0, + 1198.0, + 367.0, + 1216.0, + 341.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 1206.0, + 531.0, + 1206.0, + 531.0, + 1218.0, + 511.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 1199.0, + 564.0, + 1199.0, + 564.0, + 1211.0, + 543.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1200.0, + 923.0, + 1200.0, + 923.0, + 1222.0, + 894.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 1214.0, + 752.0, + 1214.0, + 752.0, + 1241.0, + 668.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 1214.0, + 1307.0, + 1214.0, + 1307.0, + 1241.0, + 1224.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1232.0, + 790.0, + 1232.0, + 790.0, + 1258.0, + 663.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1232.0, + 1346.0, + 1232.0, + 1346.0, + 1258.0, + 1216.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1247.0, + 369.0, + 1247.0, + 369.0, + 1268.0, + 339.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1250.0, + 747.0, + 1250.0, + 747.0, + 1272.0, + 661.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1250.0, + 922.0, + 1250.0, + 922.0, + 1268.0, + 897.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1247.0, + 1303.0, + 1247.0, + 1303.0, + 1274.0, + 1216.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1267.0, + 698.0, + 1267.0, + 698.0, + 1288.0, + 661.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 1267.0, + 1254.0, + 1267.0, + 1254.0, + 1288.0, + 1218.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1282.0, + 762.0, + 1282.0, + 762.0, + 1305.0, + 661.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1282.0, + 1317.0, + 1282.0, + 1317.0, + 1305.0, + 1216.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1301.0, + 367.0, + 1301.0, + 367.0, + 1317.0, + 352.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1300.0, + 712.0, + 1300.0, + 712.0, + 1320.0, + 663.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 1298.0, + 923.0, + 1298.0, + 923.0, + 1313.0, + 908.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 1297.0, + 1267.0, + 1297.0, + 1267.0, + 1321.0, + 1217.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1316.0, + 719.0, + 1316.0, + 719.0, + 1336.0, + 662.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 1316.0, + 1274.0, + 1316.0, + 1274.0, + 1336.0, + 1218.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1332.0, + 768.0, + 1332.0, + 768.0, + 1351.0, + 662.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 1332.0, + 1323.0, + 1332.0, + 1323.0, + 1351.0, + 1218.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 1342.0, + 923.0, + 1342.0, + 923.0, + 1362.0, + 888.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1361.0, + 382.0, + 1361.0, + 382.0, + 1376.0, + 368.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 1358.0, + 481.0, + 1358.0, + 481.0, + 1379.0, + 436.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 1358.0, + 564.0, + 1358.0, + 564.0, + 1379.0, + 520.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1358.0, + 647.0, + 1358.0, + 647.0, + 1379.0, + 602.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1358.0, + 731.0, + 1358.0, + 731.0, + 1379.0, + 686.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1358.0, + 816.0, + 1358.0, + 816.0, + 1379.0, + 767.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 1361.0, + 938.0, + 1361.0, + 938.0, + 1377.0, + 923.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 1358.0, + 1036.0, + 1358.0, + 1036.0, + 1379.0, + 992.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 1358.0, + 1119.0, + 1358.0, + 1119.0, + 1379.0, + 1076.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1358.0, + 1202.0, + 1358.0, + 1202.0, + 1379.0, + 1158.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 1358.0, + 1286.0, + 1358.0, + 1286.0, + 1379.0, + 1240.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1322.0, + 1358.0, + 1372.0, + 1358.0, + 1372.0, + 1379.0, + 1322.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1370.0, + 647.0, + 1370.0, + 647.0, + 1395.0, + 517.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1370.0, + 1201.0, + 1370.0, + 1201.0, + 1395.0, + 1073.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1404.0, + 597.0, + 1404.0, + 597.0, + 1440.0, + 553.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 1402.0, + 1154.0, + 1402.0, + 1154.0, + 1440.0, + 1107.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.75, + 685.5, + 475.75, + 685.5, + 475.75, + 701.0, + 433.75, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.75, + 744.5, + 1332.75, + 744.5, + 1332.75, + 781.0, + 1107.75, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 779.5, + 1125.0, + 779.5, + 1125.0, + 796.0, + 1092.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 831.0, + 921.0, + 831.0, + 921.0, + 852.5, + 897.0, + 852.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 1170.0, + 691.0, + 1170.0, + 691.0, + 1183.0, + 657.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.25, + 1175.0, + 658.25, + 1175.0, + 658.25, + 1199.0, + 563.25, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 496.0, + 1405.0, + 496.0, + 1405.0, + 532.0, + 295.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 527.0, + 408.0, + 527.0, + 408.0, + 565.0, + 291.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 235.0, + 546.0, + 235.0, + 546.0, + 259.0, + 419.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 230.0, + 654.0, + 230.0, + 654.0, + 260.0, + 565.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 231.0, + 786.0, + 231.0, + 786.0, + 260.0, + 698.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 227.0, + 919.0, + 227.0, + 919.0, + 260.0, + 833.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 228.0, + 1051.0, + 228.0, + 1051.0, + 261.0, + 966.0, + 261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 231.0, + 1199.0, + 231.0, + 1199.0, + 260.0, + 1084.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 235.0, + 1261.0, + 235.0, + 1261.0, + 256.0, + 1250.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 266.0, + 545.0, + 266.0, + 545.0, + 288.0, + 413.0, + 288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 261.0, + 651.0, + 261.0, + 651.0, + 293.0, + 570.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 259.0, + 790.0, + 259.0, + 790.0, + 292.0, + 696.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 262.0, + 919.0, + 262.0, + 919.0, + 291.0, + 834.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 262.0, + 1052.0, + 262.0, + 1052.0, + 291.0, + 966.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 259.0, + 1198.0, + 259.0, + 1198.0, + 292.0, + 1084.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 262.0, + 1289.0, + 262.0, + 1289.0, + 284.0, + 1260.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 296.0, + 545.0, + 296.0, + 545.0, + 318.0, + 426.0, + 318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 291.0, + 644.0, + 291.0, + 644.0, + 322.0, + 575.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 290.0, + 790.0, + 290.0, + 790.0, + 322.0, + 696.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 290.0, + 919.0, + 290.0, + 919.0, + 322.0, + 833.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 292.0, + 1053.0, + 292.0, + 1053.0, + 322.0, + 964.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 291.0, + 1200.0, + 291.0, + 1200.0, + 323.0, + 1083.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 301.0, + 1291.0, + 301.0, + 1291.0, + 329.0, + 1258.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 327.0, + 545.0, + 327.0, + 545.0, + 349.0, + 413.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 320.0, + 654.0, + 320.0, + 654.0, + 353.0, + 565.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 319.0, + 788.0, + 319.0, + 788.0, + 352.0, + 697.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 320.0, + 918.0, + 320.0, + 918.0, + 352.0, + 833.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 322.0, + 1052.0, + 322.0, + 1052.0, + 352.0, + 965.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 321.0, + 1147.0, + 321.0, + 1147.0, + 353.0, + 1083.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 321.0, + 1200.0, + 321.0, + 1200.0, + 353.0, + 1192.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 356.0, + 546.0, + 356.0, + 546.0, + 380.0, + 411.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 352.0, + 647.0, + 352.0, + 647.0, + 382.0, + 571.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 352.0, + 785.0, + 352.0, + 785.0, + 382.0, + 699.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 352.0, + 919.0, + 352.0, + 919.0, + 382.0, + 833.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 352.0, + 1052.0, + 352.0, + 1052.0, + 382.0, + 966.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 351.0, + 1200.0, + 351.0, + 1200.0, + 383.0, + 1083.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 344.0, + 1291.0, + 344.0, + 1291.0, + 372.0, + 1259.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 386.0, + 545.0, + 386.0, + 545.0, + 410.0, + 406.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 377.0, + 651.0, + 377.0, + 651.0, + 415.0, + 570.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 383.0, + 787.0, + 383.0, + 787.0, + 413.0, + 698.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 381.0, + 919.0, + 381.0, + 919.0, + 414.0, + 833.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 383.0, + 1054.0, + 383.0, + 1054.0, + 413.0, + 963.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 382.0, + 1199.0, + 382.0, + 1199.0, + 413.0, + 1084.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 389.0, + 1289.0, + 389.0, + 1289.0, + 414.0, + 1260.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 415.0, + 544.0, + 415.0, + 544.0, + 440.0, + 418.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 412.0, + 648.0, + 412.0, + 648.0, + 442.0, + 572.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 410.0, + 785.0, + 410.0, + 785.0, + 443.0, + 701.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 412.0, + 919.0, + 412.0, + 919.0, + 441.0, + 833.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 413.0, + 1054.0, + 413.0, + 1054.0, + 442.0, + 964.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 410.0, + 1146.0, + 410.0, + 1146.0, + 443.0, + 1084.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 410.0, + 1199.0, + 410.0, + 1199.0, + 443.0, + 1193.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 440.0, + 648.0, + 440.0, + 648.0, + 466.0, + 573.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 440.0, + 779.0, + 440.0, + 779.0, + 464.0, + 704.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 440.0, + 908.0, + 440.0, + 908.0, + 464.0, + 844.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 441.0, + 1045.0, + 441.0, + 1045.0, + 464.0, + 974.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 441.0, + 1187.0, + 441.0, + 1187.0, + 464.0, + 1098.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 435.0, + 1287.0, + 435.0, + 1287.0, + 453.0, + 1263.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 455.0, + 906.0, + 455.0, + 906.0, + 482.0, + 846.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 1458.0, + 1156.0, + 1458.0, + 1156.0, + 1495.0, + 542.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1665.0, + 1402.0, + 1665.0, + 1402.0, + 1701.0, + 294.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1698.0, + 1405.0, + 1698.0, + 1405.0, + 1733.0, + 294.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1729.0, + 1405.0, + 1729.0, + 1405.0, + 1765.0, + 294.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1760.0, + 1405.0, + 1760.0, + 1405.0, + 1795.0, + 294.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1785.0, + 1406.0, + 1785.0, + 1406.0, + 1830.0, + 292.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1821.0, + 553.0, + 1821.0, + 553.0, + 1858.0, + 294.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1555.0, + 1404.0, + 1555.0, + 1404.0, + 1600.0, + 294.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1589.0, + 789.0, + 1589.0, + 789.0, + 1628.0, + 294.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1889.0, + 1404.0, + 1889.0, + 1404.0, + 1936.0, + 294.0, + 1936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1924.0, + 975.0, + 1924.0, + 975.0, + 1962.0, + 293.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1995.0, + 1386.0, + 1995.0, + 1386.0, + 2041.0, + 294.0, + 2041.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1589, + 1404, + 1589, + 1404, + 1866, + 297, + 1866 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 299, + 1023, + 1403, + 1023, + 1403, + 1238, + 299, + 1238 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 487, + 1404, + 487, + 1404, + 732, + 299, + 732 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 288, + 1403, + 288, + 1403, + 472, + 299, + 472 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 300, + 1881, + 1402, + 1881, + 1402, + 2034, + 300, + 2034 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1253, + 1402, + 1253, + 1402, + 1406, + 299, + 1406 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1421, + 1403, + 1421, + 1403, + 1546, + 298, + 1546 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 300, + 747, + 1403, + 747, + 1403, + 870, + 300, + 870 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 884, + 1403, + 884, + 1403, + 1008, + 299, + 1008 + ], + "score": 0.973 + }, + { + "category_id": 0, + "poly": [ + 299, + 229, + 498, + 229, + 498, + 261, + 299, + 261 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 862, + 2088, + 862, + 2112, + 836, + 2112 + ], + "score": 0.841 + }, + { + "category_id": 15, + "poly": [ + 294.0, + 225.0, + 502.0, + 225.0, + 502.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1587.0, + 1406.0, + 1587.0, + 1406.0, + 1625.0, + 294.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1619.0, + 1406.0, + 1619.0, + 1406.0, + 1657.0, + 292.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1653.0, + 1405.0, + 1653.0, + 1405.0, + 1686.0, + 295.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1680.0, + 1406.0, + 1680.0, + 1406.0, + 1717.0, + 295.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1710.0, + 1407.0, + 1710.0, + 1407.0, + 1750.0, + 291.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1742.0, + 1405.0, + 1742.0, + 1405.0, + 1779.0, + 295.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1775.0, + 1405.0, + 1775.0, + 1405.0, + 1807.0, + 296.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1805.0, + 1404.0, + 1805.0, + 1404.0, + 1837.0, + 295.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1836.0, + 722.0, + 1836.0, + 722.0, + 1869.0, + 296.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1022.0, + 1405.0, + 1022.0, + 1405.0, + 1060.0, + 295.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1055.0, + 1402.0, + 1055.0, + 1402.0, + 1090.0, + 294.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1084.0, + 1403.0, + 1084.0, + 1403.0, + 1119.0, + 294.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1116.0, + 1402.0, + 1116.0, + 1402.0, + 1150.0, + 294.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1144.0, + 1405.0, + 1144.0, + 1405.0, + 1182.0, + 294.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1177.0, + 1405.0, + 1177.0, + 1405.0, + 1212.0, + 294.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1206.0, + 1198.0, + 1206.0, + 1198.0, + 1244.0, + 293.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 487.0, + 1405.0, + 487.0, + 1405.0, + 521.0, + 297.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 517.0, + 1404.0, + 517.0, + 1404.0, + 551.0, + 294.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 546.0, + 1405.0, + 546.0, + 1405.0, + 584.0, + 293.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 581.0, + 1404.0, + 581.0, + 1404.0, + 611.0, + 297.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 612.0, + 1403.0, + 612.0, + 1403.0, + 642.0, + 294.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 637.0, + 1404.0, + 637.0, + 1404.0, + 674.0, + 294.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 669.0, + 1405.0, + 669.0, + 1405.0, + 705.0, + 293.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 699.0, + 1366.0, + 699.0, + 1366.0, + 735.0, + 294.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 287.0, + 1405.0, + 287.0, + 1405.0, + 323.0, + 295.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 317.0, + 1404.0, + 317.0, + 1404.0, + 352.0, + 295.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 346.0, + 1407.0, + 346.0, + 1407.0, + 386.0, + 292.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 375.0, + 1405.0, + 375.0, + 1405.0, + 417.0, + 293.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 409.0, + 1405.0, + 409.0, + 1405.0, + 445.0, + 294.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 442.0, + 1134.0, + 442.0, + 1134.0, + 474.0, + 295.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1878.0, + 1407.0, + 1878.0, + 1407.0, + 1917.0, + 295.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1909.0, + 1404.0, + 1909.0, + 1404.0, + 1948.0, + 295.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1976.0, + 298.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1968.0, + 1406.0, + 1968.0, + 1406.0, + 2012.0, + 293.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 1404.0, + 2002.0, + 1404.0, + 2039.0, + 295.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1252.0, + 1406.0, + 1252.0, + 1406.0, + 1287.0, + 293.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1285.0, + 1404.0, + 1285.0, + 1404.0, + 1318.0, + 295.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1315.0, + 1404.0, + 1315.0, + 1404.0, + 1348.0, + 295.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1347.0, + 1404.0, + 1347.0, + 1404.0, + 1380.0, + 295.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1374.0, + 433.0, + 1374.0, + 433.0, + 1405.0, + 294.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1420.0, + 1407.0, + 1420.0, + 1407.0, + 1458.0, + 293.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1453.0, + 1405.0, + 1453.0, + 1405.0, + 1489.0, + 294.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1481.0, + 1408.0, + 1481.0, + 1408.0, + 1520.0, + 292.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1514.0, + 1296.0, + 1514.0, + 1296.0, + 1549.0, + 293.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 745.0, + 1404.0, + 745.0, + 1404.0, + 783.0, + 294.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 778.0, + 1403.0, + 778.0, + 1403.0, + 811.0, + 296.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 808.0, + 1403.0, + 808.0, + 1403.0, + 843.0, + 293.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 836.0, + 1290.0, + 836.0, + 1290.0, + 875.0, + 294.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 885.0, + 1404.0, + 885.0, + 1404.0, + 918.0, + 294.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 917.0, + 1405.0, + 917.0, + 1405.0, + 949.0, + 295.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 947.0, + 1405.0, + 947.0, + 1405.0, + 980.0, + 294.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 978.0, + 1402.0, + 978.0, + 1402.0, + 1010.0, + 295.0, + 1010.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1758, + 1404, + 1758, + 1404, + 2035, + 297, + 2035 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1559, + 1403, + 1559, + 1403, + 1744, + 298, + 1744 + ], + "score": 0.98 + }, + { + "category_id": 5, + "poly": [ + 303, + 848, + 1395, + 848, + 1395, + 1125, + 303, + 1125 + ], + "score": 0.978, + "html": "
Model HalfHumanoidHopperOstrichWolfHorseAverage
MLPRewardRatio1775.7557.7%1369.5962.0%1198.8848.2%1249.2354.5%2084.0769.7%/58.6%
TreeNetRewardRatio237.8179.3%417.2798.0%224.0757.4%247.03141.2%223.3499.2%/94.8%
NerveNetRewardRatio2536.5296.3%2113.56101.8%1714.6398.8%2054.54105.9%2343.62106.4%1101.8%
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1346, + 1403, + 1346, + 1403, + 1440, + 299, + 1440 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 323, + 235, + 1357, + 235, + 1357, + 712, + 323, + 712 + ], + "score": 0.971 + }, + { + "category_id": 4, + "poly": [ + 295, + 737, + 1400, + 737, + 1400, + 802, + 295, + 802 + ], + "score": 0.936 + }, + { + "category_id": 0, + "poly": [ + 299, + 1495, + 654, + 1495, + 654, + 1526, + 299, + 1526 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.847 + }, + { + "category_id": 6, + "poly": [ + 297, + 1147, + 1404, + 1147, + 1404, + 1271, + 297, + 1271 + ], + "score": 0.654 + }, + { + "category_id": 1, + "poly": [ + 297, + 1147, + 1404, + 1147, + 1404, + 1271, + 297, + 1271 + ], + "score": 0.269 + }, + { + "category_id": 13, + "poly": [ + 915, + 1083, + 993, + 1083, + 993, + 1115, + 915, + 1115 + ], + "score": 0.71, + "latex": "9 8 . 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 1035, + 1084, + 1124, + 1084, + 1124, + 1115, + 1035, + 1115 + ], + "score": 0.68, + "latex": "1 0 5 . 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 1284, + 1083, + 1374, + 1083, + 1374, + 1115, + 1284, + 1115 + ], + "score": 0.65, + "latex": "1 0 1 . 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 916, + 1009, + 993, + 1009, + 993, + 1039, + 916, + 1039 + ], + "score": 0.6, + "latex": "5 7 . 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 1159, + 1084, + 1248, + 1084, + 1248, + 1114, + 1159, + 1114 + ], + "score": 0.59, + "latex": "1 0 6 . 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 786, + 1084, + 876, + 1084, + 876, + 1115, + 786, + 1115 + ], + "score": 0.57, + "latex": "1 0 1 . 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 792, + 1008, + 869, + 1008, + 869, + 1040, + 792, + 1040 + ], + "score": 0.5, + "latex": "9 8 . 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 1035, + 1009, + 1124, + 1009, + 1124, + 1039, + 1035, + 1039 + ], + "score": 0.44, + "latex": "1 4 1 . 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 1165, + 1009, + 1241, + 1009, + 1241, + 1039, + 1165, + 1039 + ], + "score": 0.39, + "latex": "9 9 . 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 1290, + 1008, + 1367, + 1008, + 1367, + 1040, + 1290, + 1040 + ], + "score": 0.3, + "latex": "9 4 . 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 628, + 1084, + 707, + 1084, + 707, + 1115, + 628, + 1115 + ], + "score": 0.29, + "latex": "9 6 . 3 \\%" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 241.0, + 519.0, + 241.0, + 519.0, + 257.0, + 464.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 241.0, + 877.0, + 241.0, + 877.0, + 257.0, + 809.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 241.0, + 1240.0, + 241.0, + 1240.0, + 257.0, + 1185.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 257.0, + 358.0, + 257.0, + 358.0, + 269.0, + 338.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 260.0, + 1074.0, + 260.0, + 1074.0, + 269.0, + 1059.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 266.0, + 712.0, + 266.0, + 712.0, + 274.0, + 701.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 286.0, + 358.0, + 286.0, + 358.0, + 298.0, + 338.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 291.0, + 717.0, + 291.0, + 717.0, + 304.0, + 698.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 287.0, + 1075.0, + 287.0, + 1075.0, + 299.0, + 1056.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 316.0, + 356.0, + 316.0, + 356.0, + 326.0, + 340.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 312.0, + 434.0, + 312.0, + 434.0, + 332.0, + 382.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 314.0, + 795.0, + 314.0, + 795.0, + 333.0, + 740.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 311.0, + 1152.0, + 311.0, + 1152.0, + 333.0, + 1099.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 320.0, + 716.0, + 320.0, + 716.0, + 332.0, + 696.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 328.0, + 340.0, + 328.0, + 340.0, + 350.0, + 329.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 328.0, + 450.0, + 328.0, + 450.0, + 339.0, + 397.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 328.0, + 801.0, + 328.0, + 801.0, + 340.0, + 743.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 327.0, + 1171.0, + 327.0, + 1171.0, + 341.0, + 1099.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 345.0, + 356.0, + 345.0, + 356.0, + 371.0, + 330.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 350.0, + 411.0, + 350.0, + 411.0, + 358.0, + 384.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 333.0, + 697.0, + 333.0, + 697.0, + 348.0, + 689.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 336.0, + 810.0, + 336.0, + 810.0, + 361.0, + 741.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 345.0, + 1076.0, + 345.0, + 1076.0, + 358.0, + 1055.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 336.0, + 1166.0, + 336.0, + 1166.0, + 361.0, + 1100.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 358.0, + 409.0, + 358.0, + 409.0, + 370.0, + 384.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 358.0, + 773.0, + 358.0, + 773.0, + 370.0, + 742.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1101.0, + 358.0, + 1131.0, + 358.0, + 1131.0, + 370.0, + 1101.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 374.0, + 356.0, + 374.0, + 356.0, + 382.0, + 342.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 376.0, + 717.0, + 376.0, + 717.0, + 389.0, + 698.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 371.0, + 757.0, + 371.0, + 757.0, + 380.0, + 743.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 374.0, + 1075.0, + 374.0, + 1075.0, + 387.0, + 1057.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 371.0, + 1132.0, + 371.0, + 1132.0, + 380.0, + 1102.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 404.0, + 717.0, + 404.0, + 717.0, + 417.0, + 701.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 405.0, + 1074.0, + 405.0, + 1074.0, + 413.0, + 1062.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 443.0, + 435.0, + 443.0, + 435.0, + 455.0, + 405.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 443.0, + 536.0, + 443.0, + 536.0, + 468.0, + 454.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 443.0, + 587.0, + 443.0, + 587.0, + 455.0, + 558.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 443.0, + 637.0, + 443.0, + 637.0, + 455.0, + 607.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 443.0, + 795.0, + 443.0, + 795.0, + 455.0, + 764.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 443.0, + 896.0, + 443.0, + 896.0, + 468.0, + 813.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 443.0, + 946.0, + 443.0, + 946.0, + 455.0, + 917.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 443.0, + 996.0, + 443.0, + 996.0, + 455.0, + 965.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 443.0, + 1153.0, + 443.0, + 1153.0, + 455.0, + 1123.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 445.0, + 1253.0, + 445.0, + 1253.0, + 465.0, + 1174.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 443.0, + 1305.0, + 443.0, + 1305.0, + 455.0, + 1276.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1324.0, + 443.0, + 1354.0, + 443.0, + 1354.0, + 455.0, + 1324.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 487.0, + 707.0, + 487.0, + 707.0, + 502.0, + 643.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 487.0, + 1060.0, + 487.0, + 1060.0, + 502.0, + 1008.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 502.0, + 538.0, + 502.0, + 538.0, + 514.0, + 516.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 512.0, + 894.0, + 512.0, + 894.0, + 520.0, + 879.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 537.0, + 538.0, + 537.0, + 538.0, + 549.0, + 518.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 544.0, + 894.0, + 544.0, + 894.0, + 554.0, + 879.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 556.0, + 615.0, + 556.0, + 615.0, + 578.0, + 560.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 560.0, + 970.0, + 560.0, + 970.0, + 576.0, + 920.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 570.0, + 539.0, + 570.0, + 539.0, + 586.0, + 512.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 571.0, + 631.0, + 571.0, + 631.0, + 585.0, + 560.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 576.0, + 897.0, + 576.0, + 897.0, + 589.0, + 875.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 574.0, + 966.0, + 574.0, + 966.0, + 583.0, + 923.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 582.0, + 629.0, + 582.0, + 629.0, + 597.0, + 561.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 582.0, + 963.0, + 582.0, + 963.0, + 597.0, + 920.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 591.0, + 617.0, + 591.0, + 617.0, + 607.0, + 561.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 596.0, + 878.0, + 596.0, + 878.0, + 606.0, + 870.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 593.0, + 976.0, + 593.0, + 976.0, + 608.0, + 920.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 607.0, + 536.0, + 607.0, + 536.0, + 620.0, + 510.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 603.0, + 593.0, + 603.0, + 593.0, + 615.0, + 562.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 611.0, + 894.0, + 611.0, + 894.0, + 620.0, + 879.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 605.0, + 950.0, + 605.0, + 950.0, + 613.0, + 924.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 615.0, + 597.0, + 615.0, + 597.0, + 624.0, + 565.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 615.0, + 953.0, + 615.0, + 953.0, + 624.0, + 927.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 644.0, + 536.0, + 644.0, + 536.0, + 654.0, + 522.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 644.0, + 896.0, + 644.0, + 896.0, + 654.0, + 881.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 689.0, + 616.0, + 689.0, + 616.0, + 701.0, + 585.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 689.0, + 665.0, + 689.0, + 665.0, + 701.0, + 637.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 689.0, + 767.0, + 689.0, + 767.0, + 701.0, + 737.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 689.0, + 816.0, + 689.0, + 816.0, + 701.0, + 787.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 686.0, + 976.0, + 686.0, + 976.0, + 702.0, + 943.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 689.0, + 1023.0, + 689.0, + 1023.0, + 701.0, + 996.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 690.0, + 1073.0, + 690.0, + 1073.0, + 698.0, + 1047.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 689.0, + 1126.0, + 689.0, + 1126.0, + 701.0, + 1096.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 686.0, + 1177.0, + 686.0, + 1177.0, + 702.0, + 1145.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 697.0, + 712.0, + 697.0, + 712.0, + 709.0, + 636.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 695.0, + 1074.0, + 695.0, + 1074.0, + 710.0, + 994.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 293.5, + 960.0, + 293.5, + 960.0, + 380.0, + 762.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 336.0, + 433.0, + 336.0, + 433.0, + 351.5, + 385.0, + 351.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.25, + 318.5, + 522.25, + 318.5, + 522.25, + 353.5, + 449.25, + 353.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.75, + 299.0, + 1357.75, + 299.0, + 1357.75, + 366.0, + 1184.75, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 733.0, + 1405.0, + 733.0, + 1405.0, + 775.0, + 293.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 767.0, + 534.0, + 767.0, + 534.0, + 803.0, + 294.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1493.0, + 656.0, + 1493.0, + 656.0, + 1529.0, + 295.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1148.0, + 1405.0, + 1148.0, + 1405.0, + 1181.0, + 295.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1177.0, + 1404.0, + 1177.0, + 1404.0, + 1211.0, + 294.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1208.0, + 1405.0, + 1208.0, + 1405.0, + 1244.0, + 294.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1239.0, + 565.0, + 1239.0, + 565.0, + 1269.0, + 296.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1760.0, + 1402.0, + 1760.0, + 1402.0, + 1793.0, + 295.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1785.0, + 1406.0, + 1785.0, + 1406.0, + 1828.0, + 291.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1822.0, + 1405.0, + 1822.0, + 1405.0, + 1855.0, + 295.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 1406.0, + 1851.0, + 1406.0, + 1887.0, + 295.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1916.0, + 294.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1906.0, + 1408.0, + 1906.0, + 1408.0, + 1952.0, + 291.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1944.0, + 1402.0, + 1944.0, + 1402.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1974.0, + 1405.0, + 1974.0, + 1405.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2001.0, + 850.0, + 2001.0, + 850.0, + 2037.0, + 292.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1556.0, + 1404.0, + 1556.0, + 1404.0, + 1597.0, + 292.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1589.0, + 1405.0, + 1589.0, + 1405.0, + 1626.0, + 292.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1617.0, + 1407.0, + 1617.0, + 1407.0, + 1659.0, + 291.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1653.0, + 1405.0, + 1653.0, + 1405.0, + 1685.0, + 296.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1683.0, + 1404.0, + 1683.0, + 1404.0, + 1715.0, + 296.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1713.0, + 1054.0, + 1713.0, + 1054.0, + 1748.0, + 294.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1344.0, + 1404.0, + 1344.0, + 1404.0, + 1385.0, + 293.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1377.0, + 1403.0, + 1377.0, + 1403.0, + 1411.0, + 295.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1409.0, + 698.0, + 1409.0, + 698.0, + 1443.0, + 295.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1148.0, + 1405.0, + 1148.0, + 1405.0, + 1181.0, + 295.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1177.0, + 1404.0, + 1177.0, + 1404.0, + 1211.0, + 294.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1208.0, + 1405.0, + 1208.0, + 1405.0, + 1244.0, + 294.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1239.0, + 565.0, + 1239.0, + 565.0, + 1269.0, + 296.0, + 1269.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 311, + 225, + 1387, + 225, + 1387, + 426, + 311, + 426 + ], + "score": 0.979, + "html": "
ModelHalfhumanoid HopperWolfOstrich HorseAverage
MassMLPNerveNet33.28%95.87%74.04%93.24%94.68%90.13%59.23% 40.61%80.2% 69.23%60.37%85.73%
StrengthMLPNerveNet25.96%31.11%21.77%42.20%27.32%42.84%30.08% 19.80%31.41% 36.54%24.99%36.82%
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1174, + 1403, + 1174, + 1403, + 1358, + 299, + 1358 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1619, + 1403, + 1619, + 1403, + 1866, + 298, + 1866 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 699, + 1404, + 699, + 1404, + 945, + 298, + 945 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1881, + 1402, + 1881, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 300, + 1405, + 1403, + 1405, + 1403, + 1497, + 300, + 1497 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 300, + 1513, + 1403, + 1513, + 1403, + 1604, + 300, + 1604 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 297, + 990, + 1399, + 990, + 1399, + 1052, + 297, + 1052 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 296, + 1097, + 1332, + 1097, + 1332, + 1129, + 296, + 1129 + ], + "score": 0.909 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 301, + 641, + 703, + 641, + 703, + 673, + 301, + 673 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 864, + 2088, + 864, + 2112, + 837, + 2112 + ], + "score": 0.84 + }, + { + "category_id": 1, + "poly": [ + 298, + 448, + 1403, + 448, + 1403, + 571, + 298, + 571 + ], + "score": 0.618 + }, + { + "category_id": 6, + "poly": [ + 298, + 448, + 1403, + 448, + 1403, + 571, + 298, + 571 + ], + "score": 0.387 + }, + { + "category_id": 13, + "poly": [ + 1197, + 1882, + 1253, + 1882, + 1253, + 1912, + 1197, + 1912 + ], + "score": 0.85, + "latex": "42 \\%" + }, + { + "category_id": 13, + "poly": [ + 1276, + 354, + 1364, + 354, + 1364, + 381, + 1276, + 381 + ], + "score": 0.7, + "latex": "2 4 . 9 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 1153, + 385, + 1242, + 385, + 1242, + 414, + 1153, + 414 + ], + "score": 0.67, + "latex": "3 6 . 5 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 1276, + 385, + 1365, + 385, + 1365, + 414, + 1276, + 414 + ], + "score": 0.67, + "latex": "3 6 . 8 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 1034, + 278, + 1122, + 278, + 1122, + 306, + 1034, + 306 + ], + "score": 0.65, + "latex": "5 9 . 2 3 \\%" + }, + { + "category_id": 13, + "poly": [ + 1034, + 353, + 1122, + 353, + 1122, + 381, + 1034, + 381 + ], + "score": 0.64, + "latex": "3 0 . 0 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 1154, + 353, + 1242, + 353, + 1242, + 382, + 1154, + 382 + ], + "score": 0.63, + "latex": "1 9 . 8 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 913, + 384, + 1003, + 384, + 1003, + 415, + 913, + 415 + ], + "score": 0.6, + "latex": "4 2 . 8 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 1034, + 385, + 1122, + 385, + 1122, + 414, + 1034, + 414 + ], + "score": 0.58, + "latex": "3 1 . 4 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 1276, + 278, + 1364, + 278, + 1364, + 306, + 1276, + 306 + ], + "score": 0.57, + "latex": "6 0 . 3 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 1040, + 309, + 1116, + 309, + 1116, + 339, + 1040, + 339 + ], + "score": 0.56, + "latex": "8 0 . 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 1153, + 309, + 1242, + 309, + 1242, + 339, + 1153, + 339 + ], + "score": 0.53, + "latex": "6 9 . 2 3 \\%" + }, + { + "category_id": 13, + "poly": [ + 913, + 353, + 1002, + 353, + 1002, + 381, + 913, + 381 + ], + "score": 0.52, + "latex": "2 7 . 3 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 1276, + 310, + 1365, + 310, + 1365, + 339, + 1276, + 339 + ], + "score": 0.5, + "latex": "8 5 . 7 3 \\%" + }, + { + "category_id": 13, + "poly": [ + 1153, + 277, + 1242, + 277, + 1242, + 306, + 1153, + 306 + ], + "score": 0.48, + "latex": "4 0 . 6 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 794, + 384, + 884, + 384, + 884, + 415, + 794, + 415 + ], + "score": 0.41, + "latex": "4 2 . 2 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 913, + 309, + 1003, + 309, + 1003, + 340, + 913, + 340 + ], + "score": 0.38, + "latex": "9 0 . 1 3 \\%" + }, + { + "category_id": 13, + "poly": [ + 913, + 277, + 1002, + 277, + 1002, + 306, + 913, + 306 + ], + "score": 0.34, + "latex": "9 4 . 6 8 \\%" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 638.0, + 707.0, + 638.0, + 707.0, + 677.0, + 294.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2086.0, + 868.0, + 2086.0, + 868.0, + 2121.0, + 830.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 444.0, + 1404.0, + 444.0, + 1404.0, + 484.0, + 292.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 478.0, + 1407.0, + 478.0, + 1407.0, + 513.0, + 292.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 511.0, + 1403.0, + 511.0, + 1403.0, + 543.0, + 294.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 541.0, + 480.0, + 541.0, + 480.0, + 570.0, + 294.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1175.0, + 1404.0, + 1175.0, + 1404.0, + 1210.0, + 295.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1204.0, + 1403.0, + 1204.0, + 1403.0, + 1242.0, + 294.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1234.0, + 1403.0, + 1234.0, + 1403.0, + 1272.0, + 294.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1267.0, + 1405.0, + 1267.0, + 1405.0, + 1303.0, + 295.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1298.0, + 1405.0, + 1298.0, + 1405.0, + 1331.0, + 294.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1329.0, + 976.0, + 1329.0, + 976.0, + 1361.0, + 297.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1622.0, + 1404.0, + 1622.0, + 1404.0, + 1652.0, + 297.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1651.0, + 1404.0, + 1651.0, + 1404.0, + 1685.0, + 294.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1681.0, + 1405.0, + 1681.0, + 1405.0, + 1717.0, + 294.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1711.0, + 1405.0, + 1711.0, + 1405.0, + 1746.0, + 293.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1743.0, + 1404.0, + 1743.0, + 1404.0, + 1776.0, + 294.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1771.0, + 1405.0, + 1771.0, + 1405.0, + 1808.0, + 293.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1804.0, + 1405.0, + 1804.0, + 1405.0, + 1838.0, + 294.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1835.0, + 1376.0, + 1835.0, + 1376.0, + 1869.0, + 294.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 701.0, + 1408.0, + 701.0, + 1408.0, + 734.0, + 296.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 731.0, + 1405.0, + 731.0, + 1405.0, + 763.0, + 294.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 762.0, + 1405.0, + 762.0, + 1405.0, + 796.0, + 294.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 791.0, + 1405.0, + 791.0, + 1405.0, + 827.0, + 292.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 822.0, + 1405.0, + 822.0, + 1405.0, + 858.0, + 293.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 852.0, + 1405.0, + 852.0, + 1405.0, + 886.0, + 294.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 884.0, + 1405.0, + 884.0, + 1405.0, + 918.0, + 294.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 915.0, + 634.0, + 915.0, + 634.0, + 948.0, + 294.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1878.0, + 1196.0, + 1878.0, + 1196.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 1878.0, + 1406.0, + 1878.0, + 1406.0, + 1917.0, + 1254.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1404.0, + 1910.0, + 1404.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1940.0, + 1404.0, + 1940.0, + 1404.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 963.0, + 2002.0, + 963.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1403.0, + 1402.0, + 1403.0, + 1402.0, + 1439.0, + 295.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1434.0, + 1404.0, + 1434.0, + 1404.0, + 1469.0, + 295.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1464.0, + 944.0, + 1464.0, + 944.0, + 1500.0, + 294.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1513.0, + 1405.0, + 1513.0, + 1405.0, + 1547.0, + 296.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1541.0, + 1405.0, + 1541.0, + 1405.0, + 1579.0, + 294.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1572.0, + 1295.0, + 1572.0, + 1295.0, + 1606.0, + 294.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 989.0, + 1401.0, + 989.0, + 1401.0, + 1025.0, + 296.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1020.0, + 1242.0, + 1020.0, + 1242.0, + 1059.0, + 296.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1094.0, + 1336.0, + 1094.0, + 1336.0, + 1135.0, + 295.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 444.0, + 1404.0, + 444.0, + 1404.0, + 484.0, + 292.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 478.0, + 1407.0, + 478.0, + 1407.0, + 513.0, + 292.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 511.0, + 1403.0, + 511.0, + 1403.0, + 543.0, + 294.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 541.0, + 480.0, + 541.0, + 480.0, + 570.0, + 294.0, + 570.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1850, + 1404, + 1850, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1711, + 1403, + 1711, + 1403, + 1835, + 298, + 1835 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 322, + 741, + 1367, + 741, + 1367, + 1419, + 322, + 1419 + ], + "score": 0.971 + }, + { + "category_id": 4, + "poly": [ + 297, + 1446, + 1404, + 1446, + 1404, + 1540, + 297, + 1540 + ], + "score": 0.958 + }, + { + "category_id": 3, + "poly": [ + 337, + 258, + 1359, + 258, + 1359, + 522, + 337, + 522 + ], + "score": 0.951 + }, + { + "category_id": 4, + "poly": [ + 296, + 561, + 1406, + 561, + 1406, + 656, + 296, + 656 + ], + "score": 0.947 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 302, + 1637, + 785, + 1637, + 785, + 1668, + 302, + 1668 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.846 + }, + { + "category_id": 15, + "poly": [ + 366.0, + 747.0, + 401.0, + 747.0, + 401.0, + 775.0, + 366.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 748.0, + 705.0, + 748.0, + 705.0, + 771.0, + 673.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 748.0, + 1014.0, + 748.0, + 1014.0, + 771.0, + 980.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 783.0, + 401.0, + 783.0, + 401.0, + 813.0, + 366.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 783.0, + 706.0, + 783.0, + 706.0, + 810.0, + 673.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 783.0, + 1016.0, + 783.0, + 1016.0, + 810.0, + 979.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 786.0, + 353.0, + 786.0, + 353.0, + 958.0, + 323.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 786.0, + 661.0, + 786.0, + 661.0, + 954.0, + 630.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 786.0, + 966.0, + 786.0, + 966.0, + 952.0, + 936.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 819.0, + 402.0, + 819.0, + 402.0, + 851.0, + 364.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 819.0, + 709.0, + 819.0, + 709.0, + 850.0, + 669.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 821.0, + 1016.0, + 821.0, + 1016.0, + 849.0, + 978.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 858.0, + 402.0, + 858.0, + 402.0, + 889.0, + 363.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 856.0, + 709.0, + 856.0, + 709.0, + 888.0, + 669.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 854.0, + 1018.0, + 854.0, + 1018.0, + 889.0, + 978.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 852.0, + 1345.0, + 852.0, + 1345.0, + 885.0, + 1246.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 896.0, + 402.0, + 896.0, + 402.0, + 928.0, + 355.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 895.0, + 707.0, + 895.0, + 707.0, + 922.0, + 666.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 895.0, + 1015.0, + 895.0, + 1015.0, + 922.0, + 973.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 875.0, + 1368.0, + 875.0, + 1368.0, + 937.0, + 1273.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 937.0, + 401.0, + 937.0, + 401.0, + 962.0, + 354.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 934.0, + 706.0, + 934.0, + 706.0, + 958.0, + 661.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 934.0, + 1015.0, + 934.0, + 1015.0, + 960.0, + 971.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 973.0, + 401.0, + 973.0, + 401.0, + 1000.0, + 360.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 968.0, + 709.0, + 968.0, + 709.0, + 997.0, + 660.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 968.0, + 1016.0, + 968.0, + 1016.0, + 999.0, + 969.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 992.0, + 447.0, + 992.0, + 447.0, + 1020.0, + 413.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 996.0, + 520.0, + 996.0, + 520.0, + 1018.0, + 501.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 995.0, + 604.0, + 995.0, + 604.0, + 1018.0, + 584.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 990.0, + 751.0, + 990.0, + 751.0, + 1016.0, + 724.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 992.0, + 825.0, + 992.0, + 825.0, + 1014.0, + 806.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 991.0, + 907.0, + 991.0, + 907.0, + 1016.0, + 887.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 990.0, + 1062.0, + 990.0, + 1062.0, + 1018.0, + 1028.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 994.0, + 1133.0, + 994.0, + 1133.0, + 1014.0, + 1114.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 992.0, + 1215.0, + 992.0, + 1215.0, + 1016.0, + 1196.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 1018.0, + 598.0, + 1018.0, + 598.0, + 1045.0, + 426.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 1012.0, + 901.0, + 1012.0, + 901.0, + 1042.0, + 730.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 1013.0, + 1209.0, + 1013.0, + 1209.0, + 1044.0, + 1038.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1078.0, + 835.0, + 1078.0, + 835.0, + 1099.0, + 702.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1073.0, + 904.0, + 1073.0, + 904.0, + 1096.0, + 876.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 1077.0, + 1204.0, + 1077.0, + 1204.0, + 1103.0, + 1123.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1114.0, + 533.0, + 1114.0, + 533.0, + 1136.0, + 505.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1125.0, + 903.0, + 1125.0, + 903.0, + 1148.0, + 877.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 1168.0, + 532.0, + 1168.0, + 532.0, + 1187.0, + 507.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 1170.0, + 502.0, + 1170.0, + 502.0, + 1285.0, + 480.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 1170.0, + 876.0, + 1170.0, + 876.0, + 1286.0, + 849.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1176.0, + 903.0, + 1176.0, + 903.0, + 1200.0, + 877.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1221.0, + 533.0, + 1221.0, + 533.0, + 1243.0, + 505.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1231.0, + 902.0, + 1231.0, + 902.0, + 1250.0, + 877.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 1273.0, + 533.0, + 1273.0, + 533.0, + 1295.0, + 500.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 1284.0, + 902.0, + 1284.0, + 902.0, + 1304.0, + 878.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 1329.0, + 532.0, + 1329.0, + 532.0, + 1347.0, + 502.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 1338.0, + 902.0, + 1338.0, + 902.0, + 1355.0, + 873.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1384.0, + 568.0, + 1384.0, + 568.0, + 1405.0, + 528.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1384.0, + 606.0, + 1384.0, + 606.0, + 1405.0, + 573.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 1384.0, + 650.0, + 1384.0, + 650.0, + 1405.0, + 617.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1384.0, + 694.0, + 1384.0, + 694.0, + 1405.0, + 661.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 1384.0, + 740.0, + 1384.0, + 740.0, + 1406.0, + 705.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 1384.0, + 784.0, + 1384.0, + 784.0, + 1406.0, + 750.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 1384.0, + 828.0, + 1384.0, + 828.0, + 1406.0, + 793.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 1384.0, + 945.0, + 1384.0, + 945.0, + 1406.0, + 910.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 1385.0, + 1017.0, + 1385.0, + 1017.0, + 1403.0, + 988.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 1384.0, + 1089.0, + 1384.0, + 1089.0, + 1406.0, + 1060.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 1385.0, + 1162.0, + 1385.0, + 1162.0, + 1405.0, + 1137.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 1398.0, + 743.0, + 1398.0, + 743.0, + 1422.0, + 631.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 1401.0, + 1110.0, + 1401.0, + 1110.0, + 1420.0, + 1003.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 795.5, + 823.0, + 795.5, + 823.0, + 837.5, + 762.0, + 837.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.75, + 885.5, + 870.75, + 885.5, + 870.75, + 930.5, + 753.75, + 930.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.25, + 1251.0, + 1070.25, + 1251.0, + 1070.25, + 1302.0, + 929.25, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1446.0, + 1402.0, + 1446.0, + 1402.0, + 1480.0, + 296.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1474.0, + 1405.0, + 1474.0, + 1405.0, + 1512.0, + 292.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1505.0, + 545.0, + 1505.0, + 545.0, + 1545.0, + 294.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 341.0, + 1342.0, + 341.0, + 1342.0, + 356.0, + 1297.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 355.0, + 1348.0, + 355.0, + 1348.0, + 366.0, + 1318.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 370.0, + 989.0, + 370.0, + 989.0, + 397.0, + 981.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 365.0, + 1010.0, + 365.0, + 1010.0, + 399.0, + 996.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 367.0, + 1347.0, + 367.0, + 1347.0, + 398.0, + 1318.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1323.0, + 399.0, + 1340.0, + 399.0, + 1340.0, + 407.0, + 1323.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1321.0, + 409.0, + 1345.0, + 409.0, + 1345.0, + 421.0, + 1321.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 416.0, + 1010.0, + 416.0, + 1010.0, + 425.0, + 996.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 492.0, + 1009.0, + 492.0, + 1009.0, + 500.0, + 994.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 498.0, + 1022.0, + 498.0, + 1022.0, + 511.0, + 1006.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 498.0, + 1061.0, + 498.0, + 1061.0, + 511.0, + 1043.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 498.0, + 1099.0, + 498.0, + 1099.0, + 509.0, + 1081.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 498.0, + 1139.0, + 498.0, + 1139.0, + 511.0, + 1120.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 498.0, + 1179.0, + 498.0, + 1179.0, + 509.0, + 1159.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 498.0, + 1217.0, + 498.0, + 1217.0, + 511.0, + 1196.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 498.0, + 1257.0, + 498.0, + 1257.0, + 511.0, + 1235.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1275.0, + 498.0, + 1296.0, + 498.0, + 1296.0, + 509.0, + 1275.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 498.0, + 1333.0, + 498.0, + 1333.0, + 509.0, + 1312.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 509.0, + 1204.0, + 509.0, + 1204.0, + 521.0, + 1162.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1320.25, + 375.5, + 1347.25, + 375.5, + 1347.25, + 387.5, + 1320.25, + 387.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 560.0, + 1406.0, + 560.0, + 1406.0, + 598.0, + 294.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 594.0, + 1404.0, + 594.0, + 1404.0, + 628.0, + 295.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 618.0, + 408.0, + 618.0, + 408.0, + 663.0, + 293.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1635.0, + 789.0, + 1635.0, + 789.0, + 1672.0, + 295.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1849.0, + 1406.0, + 1849.0, + 1406.0, + 1885.0, + 296.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1914.0, + 296.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1906.0, + 1405.0, + 1906.0, + 1405.0, + 1950.0, + 292.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1977.0, + 292.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1970.0, + 1405.0, + 1970.0, + 1405.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1999.0, + 614.0, + 1999.0, + 614.0, + 2040.0, + 294.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1710.0, + 1404.0, + 1710.0, + 1404.0, + 1749.0, + 292.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1744.0, + 1405.0, + 1744.0, + 1405.0, + 1776.0, + 294.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 1405.0, + 1773.0, + 1405.0, + 1809.0, + 294.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1805.0, + 1020.0, + 1805.0, + 1020.0, + 1837.0, + 294.0, + 1837.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1422, + 1404, + 1422, + 1404, + 1758, + 298, + 1758 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 644, + 1404, + 644, + 1404, + 918, + 298, + 918 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 300, + 1015, + 1402, + 1015, + 1402, + 1136, + 300, + 1136 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1153, + 1403, + 1153, + 1403, + 1306, + 299, + 1306 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 329, + 236, + 1359, + 236, + 1359, + 468, + 329, + 468 + ], + "score": 0.961 + }, + { + "category_id": 4, + "poly": [ + 356, + 493, + 1341, + 493, + 1341, + 526, + 356, + 526 + ], + "score": 0.914 + }, + { + "category_id": 0, + "poly": [ + 300, + 1352, + 543, + 1352, + 543, + 1387, + 300, + 1387 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 299, + 1805, + 488, + 1805, + 488, + 1838, + 299, + 1838 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.884 + }, + { + "category_id": 0, + "poly": [ + 303, + 957, + 791, + 957, + 791, + 987, + 303, + 987 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 864, + 2088, + 864, + 2112, + 837, + 2112 + ], + "score": 0.817 + }, + { + "category_id": 0, + "poly": [ + 298, + 585, + 954, + 585, + 954, + 617, + 298, + 617 + ], + "score": 0.602 + }, + { + "category_id": 1, + "poly": [ + 296, + 1971, + 1396, + 1971, + 1396, + 2034, + 296, + 2034 + ], + "score": 0.468 + }, + { + "category_id": 1, + "poly": [ + 301, + 1857, + 1400, + 1857, + 1400, + 1950, + 301, + 1950 + ], + "score": 0.429 + }, + { + "category_id": 1, + "poly": [ + 298, + 585, + 954, + 585, + 954, + 617, + 298, + 617 + ], + "score": 0.339 + }, + { + "category_id": 13, + "poly": [ + 1309, + 1219, + 1330, + 1219, + 1330, + 1242, + 1309, + 1242 + ], + "score": 0.79, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 950, + 1250, + 972, + 1250, + 972, + 1272, + 950, + 1272 + ], + "score": 0.76, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1247, + 675, + 1297, + 675, + 1297, + 703, + 1247, + 703 + ], + "score": 0.37, + "latex": "1 - D" + }, + { + "category_id": 13, + "poly": [ + 1375, + 675, + 1401, + 675, + 1401, + 702, + 1375, + 702 + ], + "score": 0.33, + "latex": "D" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 242.0, + 517.0, + 242.0, + 517.0, + 257.0, + 473.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 241.0, + 872.0, + 241.0, + 872.0, + 257.0, + 835.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 242.0, + 1240.0, + 242.0, + 1240.0, + 257.0, + 1187.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 255.0, + 358.0, + 255.0, + 358.0, + 268.0, + 341.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 259.0, + 426.0, + 259.0, + 426.0, + 274.0, + 373.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 252.0, + 1076.0, + 252.0, + 1076.0, + 266.0, + 1056.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 257.0, + 1145.0, + 257.0, + 1145.0, + 276.0, + 1090.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 285.0, + 358.0, + 285.0, + 358.0, + 299.0, + 341.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 270.0, + 437.0, + 270.0, + 437.0, + 305.0, + 383.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 276.0, + 717.0, + 276.0, + 717.0, + 289.0, + 700.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 272.0, + 1078.0, + 272.0, + 1078.0, + 309.0, + 1054.0, + 309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1101.0, + 269.0, + 1155.0, + 269.0, + 1155.0, + 302.0, + 1101.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 316.0, + 359.0, + 316.0, + 359.0, + 382.0, + 330.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 314.0, + 718.0, + 314.0, + 718.0, + 383.0, + 684.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 316.0, + 1076.0, + 316.0, + 1076.0, + 352.0, + 1041.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 335.0, + 1077.0, + 335.0, + 1077.0, + 351.0, + 1054.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 349.0, + 359.0, + 349.0, + 359.0, + 362.0, + 345.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 353.0, + 717.0, + 353.0, + 717.0, + 366.0, + 700.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 345.0, + 1058.0, + 345.0, + 1058.0, + 382.0, + 1043.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 359.0, + 1076.0, + 359.0, + 1076.0, + 373.0, + 1056.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 381.0, + 358.0, + 381.0, + 358.0, + 391.0, + 346.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 380.0, + 1076.0, + 380.0, + 1076.0, + 394.0, + 1058.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 391.0, + 717.0, + 391.0, + 717.0, + 404.0, + 700.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 390.0, + 979.0, + 390.0, + 979.0, + 411.0, + 921.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 404.0, + 987.0, + 404.0, + 987.0, + 417.0, + 932.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 403.0, + 1076.0, + 403.0, + 1076.0, + 412.0, + 1066.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 412.0, + 357.0, + 412.0, + 357.0, + 423.0, + 346.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 430.0, + 717.0, + 430.0, + 717.0, + 443.0, + 702.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 414.0, + 978.0, + 414.0, + 978.0, + 440.0, + 932.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 423.0, + 1076.0, + 423.0, + 1076.0, + 435.0, + 1056.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 445.0, + 370.0, + 445.0, + 370.0, + 455.0, + 361.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 442.0, + 433.0, + 442.0, + 433.0, + 457.0, + 403.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 443.0, + 484.0, + 443.0, + 484.0, + 456.0, + 457.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 443.0, + 536.0, + 443.0, + 536.0, + 455.0, + 507.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 442.0, + 590.0, + 442.0, + 590.0, + 457.0, + 559.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 442.0, + 644.0, + 442.0, + 644.0, + 457.0, + 610.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 445.0, + 747.0, + 445.0, + 747.0, + 455.0, + 738.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 443.0, + 802.0, + 443.0, + 802.0, + 456.0, + 775.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 443.0, + 846.0, + 443.0, + 846.0, + 455.0, + 819.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 443.0, + 891.0, + 443.0, + 891.0, + 456.0, + 864.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 442.0, + 937.0, + 442.0, + 937.0, + 457.0, + 906.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 442.0, + 984.0, + 442.0, + 984.0, + 457.0, + 950.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 442.0, + 1151.0, + 442.0, + 1151.0, + 457.0, + 1119.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 443.0, + 1203.0, + 443.0, + 1203.0, + 455.0, + 1175.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 443.0, + 1255.0, + 443.0, + 1255.0, + 455.0, + 1225.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 442.0, + 1309.0, + 442.0, + 1309.0, + 457.0, + 1277.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1327.0, + 442.0, + 1362.0, + 442.0, + 1362.0, + 457.0, + 1327.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 450.0, + 536.0, + 450.0, + 536.0, + 469.0, + 452.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 450.0, + 896.0, + 450.0, + 896.0, + 467.0, + 813.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 450.0, + 1255.0, + 450.0, + 1255.0, + 467.0, + 1172.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 304.0, + 1292.0, + 304.0, + 1292.0, + 342.0, + 1141.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 490.0, + 1345.0, + 490.0, + 1345.0, + 530.0, + 352.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1348.0, + 549.0, + 1348.0, + 549.0, + 1396.0, + 292.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1805.0, + 490.0, + 1805.0, + 490.0, + 1842.0, + 297.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 952.0, + 797.0, + 952.0, + 797.0, + 992.0, + 295.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 833.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 583.0, + 957.0, + 583.0, + 957.0, + 621.0, + 293.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1419.0, + 1405.0, + 1419.0, + 1405.0, + 1459.0, + 292.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1451.0, + 1405.0, + 1451.0, + 1405.0, + 1488.0, + 293.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1482.0, + 1407.0, + 1482.0, + 1407.0, + 1518.0, + 292.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1514.0, + 1404.0, + 1514.0, + 1404.0, + 1549.0, + 294.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1543.0, + 1406.0, + 1543.0, + 1406.0, + 1581.0, + 293.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1574.0, + 1408.0, + 1574.0, + 1408.0, + 1612.0, + 293.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1605.0, + 1404.0, + 1605.0, + 1404.0, + 1640.0, + 294.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1636.0, + 1402.0, + 1636.0, + 1402.0, + 1671.0, + 293.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1667.0, + 1404.0, + 1667.0, + 1404.0, + 1700.0, + 293.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1698.0, + 1405.0, + 1698.0, + 1405.0, + 1732.0, + 294.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1725.0, + 813.0, + 1725.0, + 813.0, + 1764.0, + 293.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 645.0, + 1404.0, + 645.0, + 1404.0, + 677.0, + 296.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 673.0, + 1246.0, + 673.0, + 1246.0, + 710.0, + 293.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1298.0, + 673.0, + 1374.0, + 673.0, + 1374.0, + 710.0, + 1298.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 673.0, + 1405.0, + 673.0, + 1405.0, + 710.0, + 1402.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 702.0, + 1406.0, + 702.0, + 1406.0, + 741.0, + 293.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 729.0, + 1405.0, + 729.0, + 1405.0, + 775.0, + 292.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 764.0, + 1404.0, + 764.0, + 1404.0, + 800.0, + 294.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 798.0, + 1404.0, + 798.0, + 1404.0, + 830.0, + 296.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 825.0, + 1402.0, + 825.0, + 1402.0, + 862.0, + 293.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 856.0, + 1404.0, + 856.0, + 1404.0, + 894.0, + 293.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 889.0, + 859.0, + 889.0, + 859.0, + 921.0, + 296.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1013.0, + 1404.0, + 1013.0, + 1404.0, + 1048.0, + 295.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1047.0, + 1404.0, + 1047.0, + 1404.0, + 1079.0, + 295.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1074.0, + 1407.0, + 1074.0, + 1407.0, + 1111.0, + 294.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1108.0, + 1100.0, + 1108.0, + 1100.0, + 1140.0, + 295.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1152.0, + 1405.0, + 1152.0, + 1405.0, + 1189.0, + 295.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1185.0, + 1404.0, + 1185.0, + 1404.0, + 1218.0, + 295.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1214.0, + 1308.0, + 1214.0, + 1308.0, + 1251.0, + 294.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 1214.0, + 1405.0, + 1214.0, + 1405.0, + 1251.0, + 1331.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1241.0, + 949.0, + 1241.0, + 949.0, + 1284.0, + 293.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 1241.0, + 1407.0, + 1241.0, + 1407.0, + 1284.0, + 973.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1274.0, + 673.0, + 1274.0, + 673.0, + 1313.0, + 293.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1401.0, + 1971.0, + 1401.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2003.0, + 961.0, + 2003.0, + 961.0, + 2036.0, + 321.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1854.0, + 1405.0, + 1854.0, + 1405.0, + 1893.0, + 293.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1889.0, + 1404.0, + 1889.0, + 1404.0, + 1922.0, + 323.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1919.0, + 753.0, + 1919.0, + 753.0, + 1950.0, + 322.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 583.0, + 957.0, + 583.0, + 957.0, + 621.0, + 293.0, + 621.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 105, + 300, + 105 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.827 + }, + { + "category_id": 1, + "poly": [ + 297, + 1680, + 1399, + 1680, + 1399, + 1743, + 297, + 1743 + ], + "score": 0.718 + }, + { + "category_id": 1, + "poly": [ + 299, + 1972, + 1398, + 1972, + 1398, + 2033, + 299, + 2033 + ], + "score": 0.715 + }, + { + "category_id": 1, + "poly": [ + 289, + 1767, + 1402, + 1767, + 1402, + 1831, + 289, + 1831 + ], + "score": 0.699 + }, + { + "category_id": 1, + "poly": [ + 299, + 1097, + 1402, + 1097, + 1402, + 1162, + 299, + 1162 + ], + "score": 0.673 + }, + { + "category_id": 1, + "poly": [ + 300, + 1185, + 1401, + 1185, + 1401, + 1279, + 300, + 1279 + ], + "score": 0.664 + }, + { + "category_id": 1, + "poly": [ + 297, + 1854, + 1402, + 1854, + 1402, + 1947, + 297, + 1947 + ], + "score": 0.658 + }, + { + "category_id": 1, + "poly": [ + 300, + 490, + 1404, + 490, + 1404, + 554, + 300, + 554 + ], + "score": 0.644 + }, + { + "category_id": 1, + "poly": [ + 295, + 1475, + 1399, + 1475, + 1399, + 1540, + 295, + 1540 + ], + "score": 0.644 + }, + { + "category_id": 1, + "poly": [ + 298, + 1011, + 1399, + 1011, + 1399, + 1075, + 298, + 1075 + ], + "score": 0.633 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1402, + 229, + 1402, + 294, + 294, + 294 + ], + "score": 0.629 + }, + { + "category_id": 1, + "poly": [ + 296, + 432, + 1117, + 432, + 1117, + 467, + 296, + 467 + ], + "score": 0.626 + }, + { + "category_id": 1, + "poly": [ + 300, + 316, + 1399, + 316, + 1399, + 411, + 300, + 411 + ], + "score": 0.621 + }, + { + "category_id": 1, + "poly": [ + 291, + 1302, + 1401, + 1302, + 1401, + 1366, + 291, + 1366 + ], + "score": 0.621 + }, + { + "category_id": 1, + "poly": [ + 295, + 1388, + 1400, + 1388, + 1400, + 1453, + 295, + 1453 + ], + "score": 0.619 + }, + { + "category_id": 1, + "poly": [ + 299, + 577, + 1401, + 577, + 1401, + 641, + 299, + 641 + ], + "score": 0.612 + }, + { + "category_id": 1, + "poly": [ + 297, + 749, + 1399, + 749, + 1399, + 814, + 297, + 814 + ], + "score": 0.608 + }, + { + "category_id": 1, + "poly": [ + 298, + 923, + 1399, + 923, + 1399, + 987, + 298, + 987 + ], + "score": 0.59 + }, + { + "category_id": 1, + "poly": [ + 296, + 662, + 1402, + 662, + 1402, + 729, + 296, + 729 + ], + "score": 0.583 + }, + { + "category_id": 1, + "poly": [ + 293, + 836, + 1400, + 836, + 1400, + 902, + 293, + 902 + ], + "score": 0.581 + }, + { + "category_id": 1, + "poly": [ + 298, + 1563, + 1402, + 1563, + 1402, + 1656, + 298, + 1656 + ], + "score": 0.304 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 871.0, + 2085.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1677.0, + 1403.0, + 1677.0, + 1403.0, + 1717.0, + 293.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1710.0, + 515.0, + 1710.0, + 515.0, + 1743.0, + 323.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1970.0, + 1404.0, + 1970.0, + 1404.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 2001.0, + 776.0, + 2001.0, + 776.0, + 2035.0, + 324.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1767.0, + 1404.0, + 1767.0, + 1404.0, + 1803.0, + 296.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1798.0, + 1173.0, + 1798.0, + 1173.0, + 1832.0, + 321.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1100.0, + 1404.0, + 1100.0, + 1404.0, + 1132.0, + 298.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1130.0, + 1404.0, + 1130.0, + 1404.0, + 1163.0, + 322.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1185.0, + 1405.0, + 1185.0, + 1405.0, + 1222.0, + 295.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1217.0, + 1403.0, + 1217.0, + 1403.0, + 1251.0, + 324.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1247.0, + 641.0, + 1247.0, + 641.0, + 1280.0, + 324.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1852.0, + 1404.0, + 1852.0, + 1404.0, + 1890.0, + 295.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1886.0, + 1404.0, + 1886.0, + 1404.0, + 1920.0, + 322.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1911.0, + 401.0, + 1911.0, + 401.0, + 1952.0, + 320.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 489.0, + 1403.0, + 489.0, + 1403.0, + 525.0, + 296.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 519.0, + 1157.0, + 519.0, + 1157.0, + 559.0, + 324.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1477.0, + 1404.0, + 1477.0, + 1404.0, + 1513.0, + 295.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1508.0, + 1236.0, + 1508.0, + 1236.0, + 1541.0, + 321.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1011.0, + 1404.0, + 1011.0, + 1404.0, + 1047.0, + 294.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1043.0, + 983.0, + 1043.0, + 983.0, + 1076.0, + 323.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 262.0, + 837.0, + 262.0, + 837.0, + 295.0, + 322.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 432.0, + 1118.0, + 432.0, + 1118.0, + 470.0, + 296.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 317.0, + 1404.0, + 317.0, + 1404.0, + 350.0, + 295.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 348.0, + 1404.0, + 348.0, + 1404.0, + 382.0, + 325.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 377.0, + 944.0, + 377.0, + 944.0, + 415.0, + 324.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1302.0, + 1403.0, + 1302.0, + 1403.0, + 1338.0, + 296.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1334.0, + 993.0, + 1334.0, + 993.0, + 1366.0, + 324.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1391.0, + 1402.0, + 1391.0, + 1402.0, + 1423.0, + 298.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1419.0, + 1147.0, + 1419.0, + 1147.0, + 1454.0, + 322.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 574.0, + 1403.0, + 574.0, + 1403.0, + 615.0, + 293.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 609.0, + 822.0, + 609.0, + 822.0, + 642.0, + 323.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 748.0, + 1403.0, + 748.0, + 1403.0, + 787.0, + 295.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 780.0, + 602.0, + 780.0, + 602.0, + 816.0, + 321.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 922.0, + 1404.0, + 922.0, + 1404.0, + 961.0, + 296.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 955.0, + 1187.0, + 955.0, + 1187.0, + 988.0, + 323.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 663.0, + 1406.0, + 663.0, + 1406.0, + 699.0, + 296.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 694.0, + 1279.0, + 694.0, + 1279.0, + 729.0, + 321.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 835.0, + 1403.0, + 835.0, + 1403.0, + 875.0, + 295.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 869.0, + 958.0, + 869.0, + 958.0, + 901.0, + 325.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1561.0, + 1403.0, + 1561.0, + 1403.0, + 1601.0, + 294.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1595.0, + 1404.0, + 1595.0, + 1404.0, + 1629.0, + 325.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1626.0, + 778.0, + 1626.0, + 778.0, + 1656.0, + 323.0, + 1656.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1072, + 1404, + 1072, + 1404, + 1167, + 298, + 1167 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.891 + }, + { + "category_id": 1, + "poly": [ + 298, + 1391, + 1404, + 1391, + 1404, + 1487, + 298, + 1487 + ], + "score": 0.884 + }, + { + "category_id": 1, + "poly": [ + 298, + 955, + 1404, + 955, + 1404, + 1049, + 298, + 1049 + ], + "score": 0.883 + }, + { + "category_id": 1, + "poly": [ + 300, + 229, + 1399, + 229, + 1399, + 324, + 300, + 324 + ], + "score": 0.877 + }, + { + "category_id": 1, + "poly": [ + 301, + 1274, + 1401, + 1274, + 1401, + 1371, + 301, + 1371 + ], + "score": 0.872 + }, + { + "category_id": 1, + "poly": [ + 299, + 431, + 1405, + 431, + 1405, + 525, + 299, + 525 + ], + "score": 0.871 + }, + { + "category_id": 1, + "poly": [ + 295, + 345, + 1398, + 345, + 1398, + 409, + 295, + 409 + ], + "score": 0.868 + }, + { + "category_id": 1, + "poly": [ + 305, + 666, + 1398, + 666, + 1398, + 761, + 305, + 761 + ], + "score": 0.867 + }, + { + "category_id": 1, + "poly": [ + 291, + 1189, + 1401, + 1189, + 1401, + 1254, + 291, + 1254 + ], + "score": 0.866 + }, + { + "category_id": 1, + "poly": [ + 298, + 868, + 1400, + 868, + 1400, + 933, + 298, + 933 + ], + "score": 0.866 + }, + { + "category_id": 1, + "poly": [ + 292, + 783, + 1401, + 783, + 1401, + 846, + 292, + 846 + ], + "score": 0.863 + }, + { + "category_id": 1, + "poly": [ + 295, + 1508, + 1399, + 1508, + 1399, + 1573, + 295, + 1573 + ], + "score": 0.852 + }, + { + "category_id": 1, + "poly": [ + 298, + 549, + 1406, + 549, + 1406, + 644, + 298, + 644 + ], + "score": 0.85 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.835 + }, + { + "category_id": 1, + "poly": [ + 300, + 1680, + 1404, + 1680, + 1404, + 1776, + 300, + 1776 + ], + "score": 0.835 + }, + { + "category_id": 1, + "poly": [ + 295, + 1594, + 1403, + 1594, + 1403, + 1660, + 295, + 1660 + ], + "score": 0.821 + }, + { + "category_id": 1, + "poly": [ + 293, + 1797, + 1402, + 1797, + 1402, + 1863, + 293, + 1863 + ], + "score": 0.794 + }, + { + "category_id": 1, + "poly": [ + 292, + 1884, + 1403, + 1884, + 1403, + 1949, + 292, + 1949 + ], + "score": 0.745 + }, + { + "category_id": 1, + "poly": [ + 302, + 1971, + 1398, + 1971, + 1398, + 2034, + 302, + 2034 + ], + "score": 0.714 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1070.0, + 1404.0, + 1070.0, + 1404.0, + 1110.0, + 293.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1102.0, + 1404.0, + 1102.0, + 1404.0, + 1139.0, + 320.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1134.0, + 699.0, + 1134.0, + 699.0, + 1166.0, + 322.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1391.0, + 1402.0, + 1391.0, + 1402.0, + 1429.0, + 296.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1424.0, + 1405.0, + 1424.0, + 1405.0, + 1456.0, + 323.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1454.0, + 1225.0, + 1454.0, + 1225.0, + 1490.0, + 322.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 955.0, + 1404.0, + 955.0, + 1404.0, + 992.0, + 294.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 989.0, + 1402.0, + 989.0, + 1402.0, + 1023.0, + 322.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1019.0, + 679.0, + 1019.0, + 679.0, + 1049.0, + 319.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 267.0, + 294.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 262.0, + 1404.0, + 262.0, + 1404.0, + 296.0, + 322.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 291.0, + 943.0, + 291.0, + 943.0, + 325.0, + 322.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1276.0, + 1404.0, + 1276.0, + 1404.0, + 1310.0, + 295.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1308.0, + 1405.0, + 1308.0, + 1405.0, + 1342.0, + 322.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1337.0, + 756.0, + 1337.0, + 756.0, + 1372.0, + 323.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 431.0, + 1402.0, + 431.0, + 1402.0, + 468.0, + 294.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 461.0, + 1406.0, + 461.0, + 1406.0, + 501.0, + 321.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 491.0, + 398.0, + 491.0, + 398.0, + 528.0, + 319.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 342.0, + 1403.0, + 342.0, + 1403.0, + 383.0, + 293.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 377.0, + 553.0, + 377.0, + 553.0, + 408.0, + 323.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 666.0, + 1403.0, + 666.0, + 1403.0, + 703.0, + 297.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 698.0, + 1403.0, + 698.0, + 1403.0, + 729.0, + 321.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 731.0, + 1153.0, + 731.0, + 1153.0, + 761.0, + 325.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1189.0, + 1406.0, + 1189.0, + 1406.0, + 1225.0, + 294.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1219.0, + 1123.0, + 1219.0, + 1123.0, + 1255.0, + 322.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 868.0, + 1402.0, + 868.0, + 1402.0, + 906.0, + 297.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 901.0, + 894.0, + 901.0, + 894.0, + 934.0, + 322.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 781.0, + 1406.0, + 781.0, + 1406.0, + 821.0, + 295.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 813.0, + 849.0, + 813.0, + 849.0, + 848.0, + 321.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1508.0, + 1403.0, + 1508.0, + 1403.0, + 1544.0, + 294.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1539.0, + 1223.0, + 1539.0, + 1223.0, + 1574.0, + 321.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 545.0, + 1407.0, + 545.0, + 1407.0, + 589.0, + 292.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 581.0, + 1406.0, + 581.0, + 1406.0, + 615.0, + 321.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 610.0, + 1027.0, + 610.0, + 1027.0, + 648.0, + 321.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1680.0, + 1404.0, + 1680.0, + 1404.0, + 1718.0, + 295.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1712.0, + 1405.0, + 1712.0, + 1405.0, + 1749.0, + 322.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1744.0, + 809.0, + 1744.0, + 809.0, + 1778.0, + 324.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1595.0, + 1405.0, + 1595.0, + 1405.0, + 1631.0, + 295.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1626.0, + 1137.0, + 1626.0, + 1137.0, + 1661.0, + 322.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1798.0, + 1404.0, + 1798.0, + 1404.0, + 1835.0, + 295.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1830.0, + 1203.0, + 1830.0, + 1203.0, + 1865.0, + 322.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1886.0, + 1404.0, + 1886.0, + 1404.0, + 1922.0, + 296.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1917.0, + 1204.0, + 1917.0, + 1204.0, + 1950.0, + 323.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1971.0, + 1402.0, + 1971.0, + 1402.0, + 2006.0, + 297.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 2001.0, + 907.0, + 2001.0, + 907.0, + 2037.0, + 324.0, + 2037.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.836 + }, + { + "category_id": 1, + "poly": [ + 299, + 1551, + 1403, + 1551, + 1403, + 1647, + 299, + 1647 + ], + "score": 0.708 + }, + { + "category_id": 1, + "poly": [ + 290, + 1242, + 1401, + 1242, + 1401, + 1306, + 290, + 1306 + ], + "score": 0.699 + }, + { + "category_id": 1, + "poly": [ + 298, + 1437, + 1403, + 1437, + 1403, + 1532, + 298, + 1532 + ], + "score": 0.697 + }, + { + "category_id": 1, + "poly": [ + 303, + 1325, + 1400, + 1325, + 1400, + 1421, + 303, + 1421 + ], + "score": 0.691 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1401, + 229, + 1401, + 324, + 299, + 324 + ], + "score": 0.686 + }, + { + "category_id": 1, + "poly": [ + 295, + 962, + 1401, + 962, + 1401, + 1027, + 295, + 1027 + ], + "score": 0.666 + }, + { + "category_id": 1, + "poly": [ + 293, + 1159, + 1402, + 1159, + 1402, + 1223, + 293, + 1223 + ], + "score": 0.665 + }, + { + "category_id": 1, + "poly": [ + 296, + 1665, + 1401, + 1665, + 1401, + 1760, + 296, + 1760 + ], + "score": 0.662 + }, + { + "category_id": 1, + "poly": [ + 304, + 736, + 1402, + 736, + 1402, + 830, + 304, + 830 + ], + "score": 0.658 + }, + { + "category_id": 1, + "poly": [ + 297, + 1779, + 1398, + 1779, + 1398, + 1843, + 297, + 1843 + ], + "score": 0.646 + }, + { + "category_id": 1, + "poly": [ + 295, + 539, + 1398, + 539, + 1398, + 603, + 295, + 603 + ], + "score": 0.643 + }, + { + "category_id": 1, + "poly": [ + 299, + 621, + 1403, + 621, + 1403, + 717, + 299, + 717 + ], + "score": 0.64 + }, + { + "category_id": 1, + "poly": [ + 298, + 343, + 1399, + 343, + 1399, + 436, + 298, + 436 + ], + "score": 0.638 + }, + { + "category_id": 1, + "poly": [ + 295, + 1044, + 1401, + 1044, + 1401, + 1140, + 295, + 1140 + ], + "score": 0.632 + }, + { + "category_id": 1, + "poly": [ + 294, + 455, + 1401, + 455, + 1401, + 519, + 294, + 519 + ], + "score": 0.605 + }, + { + "category_id": 1, + "poly": [ + 297, + 849, + 1401, + 849, + 1401, + 945, + 297, + 945 + ], + "score": 0.604 + }, + { + "category_id": 1, + "poly": [ + 298, + 1862, + 1405, + 1862, + 1405, + 1957, + 298, + 1957 + ], + "score": 0.519 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1552.0, + 1404.0, + 1552.0, + 1404.0, + 1587.0, + 295.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1583.0, + 1405.0, + 1583.0, + 1405.0, + 1617.0, + 324.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1616.0, + 708.0, + 1616.0, + 708.0, + 1645.0, + 320.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1238.0, + 1404.0, + 1238.0, + 1404.0, + 1282.0, + 292.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1273.0, + 1067.0, + 1273.0, + 1067.0, + 1307.0, + 322.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1436.0, + 1403.0, + 1436.0, + 1403.0, + 1475.0, + 294.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1471.0, + 1405.0, + 1471.0, + 1405.0, + 1505.0, + 322.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1500.0, + 537.0, + 1500.0, + 537.0, + 1532.0, + 323.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1324.0, + 1405.0, + 1324.0, + 1405.0, + 1362.0, + 297.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1353.0, + 1406.0, + 1353.0, + 1406.0, + 1394.0, + 322.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1388.0, + 1027.0, + 1388.0, + 1027.0, + 1422.0, + 323.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 227.0, + 1406.0, + 227.0, + 1406.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 261.0, + 1406.0, + 261.0, + 1406.0, + 298.0, + 323.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 291.0, + 722.0, + 291.0, + 722.0, + 325.0, + 323.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 958.0, + 1405.0, + 958.0, + 1405.0, + 1001.0, + 292.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 993.0, + 712.0, + 993.0, + 712.0, + 1028.0, + 323.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1156.0, + 1404.0, + 1156.0, + 1404.0, + 1196.0, + 295.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1190.0, + 1083.0, + 1190.0, + 1083.0, + 1224.0, + 322.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1663.0, + 1406.0, + 1663.0, + 1406.0, + 1704.0, + 294.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1698.0, + 1405.0, + 1698.0, + 1405.0, + 1732.0, + 323.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1727.0, + 757.0, + 1727.0, + 757.0, + 1761.0, + 323.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 732.0, + 1405.0, + 732.0, + 1405.0, + 775.0, + 295.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 762.0, + 1406.0, + 762.0, + 1406.0, + 806.0, + 323.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 798.0, + 1030.0, + 798.0, + 1030.0, + 832.0, + 325.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1779.0, + 1403.0, + 1779.0, + 1403.0, + 1815.0, + 295.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1811.0, + 707.0, + 1811.0, + 707.0, + 1843.0, + 321.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 538.0, + 1403.0, + 538.0, + 1403.0, + 576.0, + 295.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 570.0, + 1359.0, + 570.0, + 1359.0, + 604.0, + 323.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 620.0, + 1405.0, + 620.0, + 1405.0, + 659.0, + 294.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 651.0, + 1405.0, + 651.0, + 1405.0, + 691.0, + 321.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 685.0, + 1371.0, + 685.0, + 1371.0, + 719.0, + 324.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 344.0, + 1403.0, + 344.0, + 1403.0, + 377.0, + 294.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 375.0, + 1404.0, + 375.0, + 1404.0, + 409.0, + 324.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 405.0, + 623.0, + 405.0, + 623.0, + 434.0, + 323.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1043.0, + 1405.0, + 1043.0, + 1405.0, + 1081.0, + 295.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1076.0, + 1403.0, + 1076.0, + 1403.0, + 1112.0, + 322.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1106.0, + 848.0, + 1106.0, + 848.0, + 1143.0, + 323.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 451.0, + 1405.0, + 451.0, + 1405.0, + 496.0, + 293.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 487.0, + 1057.0, + 487.0, + 1057.0, + 520.0, + 323.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 846.0, + 1406.0, + 846.0, + 1406.0, + 886.0, + 292.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 881.0, + 1405.0, + 881.0, + 1405.0, + 915.0, + 324.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 913.0, + 882.0, + 913.0, + 882.0, + 944.0, + 323.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1860.0, + 1407.0, + 1860.0, + 1407.0, + 1901.0, + 293.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1896.0, + 1403.0, + 1896.0, + 1403.0, + 1927.0, + 325.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1923.0, + 1109.0, + 1923.0, + 1109.0, + 1959.0, + 323.0, + 1959.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 449, + 1825, + 1249, + 1825, + 1249, + 2014, + 449, + 2014 + ], + "score": 0.983, + "html": "
TreeNetValue Tried
Network Shape[64,64], [128,128], [256,256]
Number of Iteration Per Update10,20
Use KL PenaltyYes,No
Learning Rate SchedulerLinear Decay, Adaptive, Constant
" + }, + { + "category_id": 5, + "poly": [ + 407, + 577, + 1291, + 577, + 1291, + 737, + 407, + 737 + ], + "score": 0.981, + "html": "
ParametersValue SetParameters Value Set
Value Discount Factor y.99GAE入.95
PPO Clip Value0.2Starting Learning Rate3e-4
Gradient Clip Value5.0Target KL0.01
" + }, + { + "category_id": 1, + "poly": [ + 298, + 970, + 1404, + 970, + 1404, + 1126, + 298, + 1126 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1216, + 1405, + 1216, + 1405, + 1403, + 298, + 1403 + ], + "score": 0.977 + }, + { + "category_id": 5, + "poly": [ + 403, + 1534, + 1296, + 1534, + 1296, + 1725, + 403, + 1725 + ], + "score": 0.977, + "html": "
MLPValue Tried
Network Shape[64,64],[128,128], [256,256],[64,64,64]
Number of Iteration Per Update10,20
Use KL PenaltyYes, No
Learning Rate SchedulerLinear Decay, Adaptive, Constant
" + }, + { + "category_id": 1, + "poly": [ + 299, + 352, + 1404, + 352, + 1404, + 445, + 299, + 445 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 300, + 861, + 1403, + 861, + 1403, + 955, + 300, + 955 + ], + "score": 0.967 + }, + { + "category_id": 0, + "poly": [ + 299, + 805, + 583, + 805, + 583, + 836, + 299, + 836 + ], + "score": 0.923 + }, + { + "category_id": 6, + "poly": [ + 619, + 520, + 1079, + 520, + 1079, + 554, + 619, + 554 + ], + "score": 0.915 + }, + { + "category_id": 6, + "poly": [ + 527, + 1767, + 1170, + 1767, + 1170, + 1801, + 527, + 1801 + ], + "score": 0.915 + }, + { + "category_id": 6, + "poly": [ + 546, + 1477, + 1154, + 1477, + 1154, + 1511, + 546, + 1511 + ], + "score": 0.914 + }, + { + "category_id": 0, + "poly": [ + 299, + 294, + 651, + 294, + 651, + 326, + 299, + 326 + ], + "score": 0.913 + }, + { + "category_id": 0, + "poly": [ + 298, + 226, + 501, + 226, + 501, + 262, + 298, + 262 + ], + "score": 0.906 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 816, + 74, + 816, + 105, + 299, + 105 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 301, + 1161, + 701, + 1161, + 701, + 1192, + 301, + 1192 + ], + "score": 0.898 + }, + { + "category_id": 1, + "poly": [ + 291, + 459, + 1368, + 459, + 1368, + 494, + 291, + 494 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.858 + }, + { + "category_id": 1, + "poly": [ + 297, + 1417, + 981, + 1417, + 981, + 1450, + 297, + 1450 + ], + "score": 0.79 + }, + { + "category_id": 13, + "poly": [ + 512, + 1093, + 539, + 1093, + 539, + 1127, + 512, + 1127 + ], + "score": 0.88, + "latex": "\\theta _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1356, + 1067, + 1393, + 1067, + 1393, + 1097, + 1356, + 1097 + ], + "score": 0.86, + "latex": "\\varpi _ { j }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 801.0, + 587.0, + 801.0, + 587.0, + 840.0, + 295.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 514.0, + 1082.0, + 514.0, + 1082.0, + 559.0, + 616.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 1764.0, + 1173.0, + 1764.0, + 1173.0, + 1805.0, + 525.0, + 1805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 1476.0, + 1154.0, + 1476.0, + 1154.0, + 1513.0, + 544.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 655.0, + 291.0, + 655.0, + 331.0, + 294.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 220.0, + 507.0, + 220.0, + 507.0, + 269.0, + 292.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1159.0, + 705.0, + 1159.0, + 705.0, + 1195.0, + 295.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 971.0, + 1402.0, + 971.0, + 1402.0, + 1004.0, + 294.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1000.0, + 1406.0, + 1000.0, + 1406.0, + 1036.0, + 293.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1029.0, + 1405.0, + 1029.0, + 1405.0, + 1068.0, + 292.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1055.0, + 1355.0, + 1055.0, + 1355.0, + 1105.0, + 291.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1055.0, + 1409.0, + 1055.0, + 1409.0, + 1105.0, + 1394.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1089.0, + 511.0, + 1089.0, + 511.0, + 1131.0, + 293.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 1089.0, + 552.0, + 1089.0, + 552.0, + 1131.0, + 540.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1219.0, + 1403.0, + 1219.0, + 1403.0, + 1251.0, + 295.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1247.0, + 1405.0, + 1247.0, + 1405.0, + 1284.0, + 292.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1280.0, + 1403.0, + 1280.0, + 1403.0, + 1312.0, + 293.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1308.0, + 1405.0, + 1308.0, + 1405.0, + 1343.0, + 293.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1342.0, + 1406.0, + 1342.0, + 1406.0, + 1374.0, + 295.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1371.0, + 567.0, + 1371.0, + 567.0, + 1407.0, + 295.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 351.0, + 1403.0, + 351.0, + 1403.0, + 385.0, + 297.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 383.0, + 1405.0, + 383.0, + 1405.0, + 417.0, + 295.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 409.0, + 451.0, + 409.0, + 451.0, + 447.0, + 292.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 861.0, + 1403.0, + 861.0, + 1403.0, + 897.0, + 294.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 895.0, + 1404.0, + 895.0, + 1404.0, + 929.0, + 295.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 924.0, + 839.0, + 924.0, + 839.0, + 958.0, + 294.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 454.0, + 1371.0, + 454.0, + 1371.0, + 498.0, + 295.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1416.0, + 984.0, + 1416.0, + 984.0, + 1452.0, + 296.0, + 1452.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 18, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 375, + 282, + 1324, + 282, + 1324, + 684, + 375, + 684 + ], + "score": 0.984, + "html": "
NerveNetValue Tried
Network Shape[64,64],[128,128],[256, 256]
Number of Iteration Per Update10,20
Use KL PenaltyYes,No
Learning Rate SchedulerLinear Decay, Adaptive, Constant
Number of Propogation Steps3,4,5,6
NerveNetVariantsNerveNet-1,NerveNet-2,NerveNet-MLP
Size ofNodes'Hidden State32,64,128
Merge joint and body NodeYes,No
Output NetworkShared, Separate
Disable Edge TypeYes, No
Add Skip-connection from / to rootYes,No
" + }, + { + "category_id": 3, + "poly": [ + 361, + 919, + 1347, + 919, + 1347, + 1281, + 361, + 1281 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 333, + 1456, + 1364, + 1456, + 1364, + 1756, + 333, + 1756 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 299, + 797, + 1400, + 797, + 1400, + 860, + 299, + 860 + ], + "score": 0.954 + }, + { + "category_id": 4, + "poly": [ + 294, + 1310, + 1404, + 1310, + 1404, + 1372, + 294, + 1372 + ], + "score": 0.923 + }, + { + "category_id": 6, + "poly": [ + 518, + 224, + 1179, + 224, + 1179, + 258, + 518, + 258 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 297, + 739, + 921, + 739, + 921, + 772, + 297, + 772 + ], + "score": 0.911 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 866, + 2088, + 866, + 2113, + 834, + 2113 + ], + "score": 0.867 + }, + { + "category_id": 4, + "poly": [ + 293, + 1781, + 1400, + 1781, + 1400, + 1815, + 293, + 1815 + ], + "score": 0.745 + }, + { + "category_id": 1, + "poly": [ + 293, + 1781, + 1400, + 1781, + 1400, + 1815, + 293, + 1815 + ], + "score": 0.223 + }, + { + "category_id": 15, + "poly": [ + 502.0, + 925.0, + 658.0, + 925.0, + 658.0, + 949.0, + 502.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 925.0, + 1210.0, + 925.0, + 1210.0, + 949.0, + 1000.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 966.0, + 1282.0, + 966.0, + 1282.0, + 1002.0, + 1220.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 1027.0, + 797.0, + 1027.0, + 797.0, + 1037.0, + 788.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 1056.0, + 1282.0, + 1056.0, + 1282.0, + 1092.0, + 1220.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1100.0, + 671.0, + 1100.0, + 671.0, + 1130.0, + 610.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 1151.0, + 1191.0, + 1151.0, + 1191.0, + 1184.0, + 1130.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1191.0, + 756.0, + 1191.0, + 756.0, + 1222.0, + 706.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 1245.0, + 1276.0, + 1245.0, + 1276.0, + 1273.0, + 1226.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 1005.5, + 759.0, + 1005.5, + 759.0, + 1035.5, + 700.0, + 1035.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1145.0, + 1110.0, + 1145.0, + 1110.0, + 1161.0, + 1029.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 1467.0, + 600.0, + 1467.0, + 600.0, + 1486.0, + 532.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 1466.0, + 1131.0, + 1466.0, + 1131.0, + 1486.0, + 1072.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 1491.0, + 669.0, + 1491.0, + 669.0, + 1512.0, + 629.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1491.0, + 709.0, + 1491.0, + 709.0, + 1528.0, + 669.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 1549.0, + 1215.0, + 1549.0, + 1215.0, + 1572.0, + 1176.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 1549.0, + 1360.0, + 1549.0, + 1360.0, + 1572.0, + 1290.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 1570.0, + 701.0, + 1570.0, + 701.0, + 1597.0, + 652.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 1572.0, + 766.0, + 1572.0, + 766.0, + 1600.0, + 738.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 1589.0, + 1089.0, + 1589.0, + 1089.0, + 1605.0, + 1074.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 1580.0, + 1341.0, + 1580.0, + 1341.0, + 1613.0, + 1306.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1644.0, + 768.0, + 1644.0, + 768.0, + 1678.0, + 732.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 1645.0, + 687.0, + 1645.0, + 687.0, + 1673.0, + 653.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1721.0, + 691.0, + 1721.0, + 691.0, + 1740.0, + 654.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.75, + 1534.0, + 488.75, + 1534.0, + 488.75, + 1553.5, + 450.75, + 1553.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.75, + 1621.0, + 1163.75, + 1621.0, + 1163.75, + 1629.5, + 1112.75, + 1629.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1575.0, + 634.0, + 1575.0, + 634.0, + 1600.5, + 620.0, + 1600.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 1644.5, + 643.0, + 1644.5, + 643.0, + 1677.0, + 611.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 1675.5, + 1284.0, + 1675.5, + 1284.0, + 1699.5, + 1235.0, + 1699.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 1622.5, + 916.0, + 1622.5, + 916.0, + 1630.5, + 828.0, + 1630.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1309.0, + 1405.0, + 1309.0, + 1405.0, + 1345.0, + 296.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1342.0, + 675.0, + 1342.0, + 675.0, + 1374.0, + 297.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 222.0, + 1180.0, + 222.0, + 1180.0, + 262.0, + 519.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 739.0, + 921.0, + 739.0, + 921.0, + 774.0, + 295.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 2083.0, + 872.0, + 2083.0, + 872.0, + 2123.0, + 828.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1777.0, + 1403.0, + 1777.0, + 1403.0, + 1816.0, + 295.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 795.0, + 1401.0, + 795.0, + 1401.0, + 831.0, + 294.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 827.0, + 1122.0, + 827.0, + 1122.0, + 862.0, + 294.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1777.0, + 1403.0, + 1777.0, + 1403.0, + 1816.0, + 295.0, + 1816.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 19, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 376, + 707, + 1337, + 707, + 1337, + 1233, + 376, + 1233 + ], + "score": 0.968 + }, + { + "category_id": 3, + "poly": [ + 334, + 274, + 1355, + 274, + 1355, + 551, + 334, + 551 + ], + "score": 0.964 + }, + { + "category_id": 3, + "poly": [ + 342, + 1376, + 1359, + 1376, + 1359, + 1816, + 342, + 1816 + ], + "score": 0.963 + }, + { + "category_id": 4, + "poly": [ + 296, + 1873, + 1405, + 1873, + 1405, + 1936, + 296, + 1936 + ], + "score": 0.933 + }, + { + "category_id": 4, + "poly": [ + 420, + 1283, + 1276, + 1283, + 1276, + 1319, + 420, + 1319 + ], + "score": 0.927 + }, + { + "category_id": 4, + "poly": [ + 294, + 578, + 1400, + 578, + 1400, + 612, + 294, + 612 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 863, + 2088, + 863, + 2113, + 834, + 2113 + ], + "score": 0.661 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 863, + 2088, + 863, + 2113, + 834, + 2113 + ], + "score": 0.168 + }, + { + "category_id": 15, + "poly": [ + 823.0, + 710.0, + 865.0, + 710.0, + 865.0, + 737.0, + 823.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 791.0, + 1069.0, + 791.0, + 1069.0, + 824.0, + 1023.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 858.0, + 1334.0, + 858.0, + 1334.0, + 887.0, + 1270.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 897.0, + 1085.0, + 897.0, + 1085.0, + 928.0, + 1022.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 893.0, + 1313.0, + 893.0, + 1313.0, + 933.0, + 1271.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 950.0, + 1126.0, + 950.0, + 1126.0, + 984.0, + 1068.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 1037.0, + 547.0, + 1037.0, + 547.0, + 1115.0, + 466.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1072.0, + 898.0, + 1072.0, + 898.0, + 1109.0, + 861.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1078.0, + 1011.0, + 1078.0, + 1011.0, + 1102.0, + 990.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 1075.0, + 1261.0, + 1075.0, + 1261.0, + 1112.0, + 1213.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 1124.0, + 921.0, + 1124.0, + 921.0, + 1154.0, + 856.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1123.0, + 1035.0, + 1123.0, + 1035.0, + 1160.0, + 990.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1188.0, + 1279.0, + 1188.0, + 1279.0, + 1218.0, + 1216.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 857.5, + 1202.0, + 857.5, + 1202.0, + 889.5, + 1162.0, + 889.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 901.0, + 1004.0, + 901.0, + 1004.0, + 924.5, + 1000.0, + 924.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 985.0, + 1102.0, + 985.0, + 1102.0, + 1022.0, + 1077.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 279.0, + 656.0, + 279.0, + 656.0, + 303.0, + 581.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 279.0, + 1217.0, + 279.0, + 1217.0, + 303.0, + 1136.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 361.0, + 543.0, + 361.0, + 543.0, + 392.0, + 502.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 362.0, + 722.0, + 362.0, + 722.0, + 392.0, + 681.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 361.0, + 812.0, + 361.0, + 812.0, + 389.0, + 766.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 472.0, + 1166.0, + 472.0, + 1166.0, + 500.0, + 1126.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 472.0, + 1255.0, + 472.0, + 1255.0, + 501.0, + 1215.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 474.0, + 1347.0, + 474.0, + 1347.0, + 497.0, + 1302.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 508.0, + 1084.0, + 508.0, + 1084.0, + 539.0, + 1044.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1320.0, + 520.0, + 1331.0, + 520.0, + 1331.0, + 528.0, + 1320.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 337.0, + 1076.0, + 337.0, + 1076.0, + 357.0, + 1054.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 364.5, + 450.0, + 364.5, + 450.0, + 390.0, + 414.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 364.5, + 629.0, + 364.5, + 629.0, + 390.5, + 591.0, + 390.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 429.5, + 1275.0, + 429.5, + 1275.0, + 452.5, + 1215.0, + 452.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 473.0, + 983.0, + 473.0, + 983.0, + 498.0, + 951.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.25, + 472.5, + 1073.25, + 472.5, + 1073.25, + 498.5, + 1038.25, + 498.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 1382.0, + 644.0, + 1382.0, + 644.0, + 1413.0, + 577.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 1385.0, + 1116.0, + 1385.0, + 1116.0, + 1408.0, + 1018.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1421.0, + 741.0, + 1421.0, + 741.0, + 1459.0, + 699.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1422.0, + 679.0, + 1422.0, + 679.0, + 1448.0, + 635.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1494.0, + 1002.0, + 1494.0, + 1002.0, + 1512.0, + 981.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1533.0, + 688.0, + 1533.0, + 688.0, + 1566.0, + 634.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1538.0, + 740.0, + 1538.0, + 740.0, + 1573.0, + 703.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 1520.0, + 951.0, + 1520.0, + 951.0, + 1576.0, + 903.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 1615.0, + 1099.0, + 1615.0, + 1099.0, + 1639.0, + 1054.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1651.0, + 672.0, + 1651.0, + 672.0, + 1681.0, + 635.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 1648.0, + 741.0, + 1648.0, + 741.0, + 1691.0, + 697.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 1651.0, + 951.0, + 1651.0, + 951.0, + 1679.0, + 863.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 1649.0, + 1095.0, + 1649.0, + 1095.0, + 1684.0, + 1059.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1651.0, + 1303.0, + 1651.0, + 1303.0, + 1679.0, + 1211.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1704.0, + 940.0, + 1704.0, + 940.0, + 1732.0, + 864.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1704.0, + 1294.0, + 1704.0, + 1294.0, + 1732.0, + 1211.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1766.0, + 678.0, + 1766.0, + 678.0, + 1791.0, + 634.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 1762.0, + 867.0, + 1762.0, + 867.0, + 1785.0, + 791.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1760.0, + 916.0, + 1760.0, + 916.0, + 1791.0, + 887.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 1761.0, + 1358.0, + 1761.0, + 1358.0, + 1784.0, + 1276.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 1431.5, + 1262.0, + 1431.5, + 1262.0, + 1458.0, + 1155.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 1632.5, + 552.0, + 1632.5, + 552.0, + 1654.5, + 409.0, + 1654.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1869.0, + 1406.0, + 1869.0, + 1406.0, + 1911.0, + 293.0, + 1911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1905.0, + 494.0, + 1905.0, + 494.0, + 1936.0, + 295.0, + 1936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1282.0, + 1279.0, + 1282.0, + 1279.0, + 1322.0, + 421.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 575.0, + 1406.0, + 575.0, + 1406.0, + 615.0, + 296.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 830.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 830.0, + 2124.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 20, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.886 + }, + { + "category_id": 5, + "poly": [ + 938, + 389, + 1384, + 389, + 1384, + 1928, + 938, + 1928 + ], + "score": 0.871, + "html": "
971900933554ISL9816 48124111111
福848-R488845118101888636-
13-808-668-69L-498-486-29'88-8£6L-728-4155-58888-764156-
55446 8314448F4364学3999京08329
9'SL-0388756-6'1-6499L-28756-7I1-3678-896-8678-$'96-
8386'091590556699496007844481115446548814福683515554
3314541S2453 34251863
4568833853494
L094169 48248
518588408584548443244334
23225496970829449$'0998552455413331255551835454TOIIOI03134111441455514694445918041887551957138050675545
S66L339835369558635-8三61
14444431148B6953 国5058818
2533773828361949'101326.30059296078
993打99111
eeeereeneeeeeeeeet eereereeerreeeeegereeeeereraaergeeeedeereeeergereeeeeieeeereeneereergereeereeeeeeeeeeeeeceeeeeeeeeeeeeeeeeceeeereeeeeeeeeecreareeereseeeeeeeereeeetGeepeeeeeecegepeeeeetCeeereneerecegeeeeeeeeeerereeeeeneeeereeeeeeelereeeeeeeeeeneeeeneeeeeeeeeeeeeeereeeeeeeeeeereeeecreeeeererrereeneerectgereeetireeeeeeeseeetneeeeieeeeereretreresesaeeeeeeeeeeti
" + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2113, + 835, + 2113 + ], + "score": 0.856 + }, + { + "category_id": 5, + "poly": [ + 437, + 395, + 898, + 395, + 898, + 1933, + 437, + 1933 + ], + "score": 0.823, + "html": "
539396449484264448423252454福
91-101-101855 8004195593351 264 L'L9-
11.971-6'061-9'S61- 52586872111475113 461-14411
46018三99 5814 4444-1103435587-848 8'I
91552929-854福F44
2233344886335929'L6163289464
2806I3品188282
19453m0222222337494988881044
983841 9491 448888
4 4164216152
4199919189LI62312441111444449五T1LI1046101224494466
4004484854469413559 54944135583封9'$9937
1 667-中458334514
5334542283348
57611-46'0-059S01-61-409m6n6
eeeeeeeeeeeeeeeeeeeeeieeeegeeeeeeeeeeeeegeeeeeteeeeeeereeneeteeteereeeeeicreeeeeereeeeeeegeeer erneeeeeeeeeeeeereeeeeeeeeeeeeeeeeercreeeneeeeeeeeeerergereeeeereeteeereeieeieneeeeeeeeeeereeeeereeeeeeeeeeeeeeeeeeeeneeeeeneeeeeeeeeeeicrrereeerreteetgereeeneeeeeeeeeetereenetrectreeeetrerreeeeeerrcegeeeeeerereeeeeeecrgepeeeeelepeeeeeeseeneeeeeeeereeeeeeaeeeeeeeeeeeeeeeeteeea
" + }, + { + "category_id": 0, + "poly": [ + 301, + 231, + 900, + 231, + 900, + 261, + 301, + 261 + ], + "score": 0.303 + }, + { + "category_id": 6, + "poly": [ + 309, + 822, + 342, + 822, + 342, + 1481, + 309, + 1481 + ], + "score": 0.288 + }, + { + "category_id": 6, + "poly": [ + 380, + 371, + 435, + 371, + 435, + 1619, + 380, + 1619 + ], + "score": 0.175 + }, + { + "category_id": 7, + "poly": [ + 309, + 822, + 342, + 822, + 342, + 1481, + 309, + 1481 + ], + "score": 0.158 + }, + { + "category_id": 6, + "poly": [ + 904, + 802, + 941, + 802, + 941, + 1733, + 904, + 1733 + ], + "score": 0.128 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 906.0, + 229.0, + 906.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 825.0, + 347.0, + 825.0, + 347.0, + 1483.0, + 311.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 384.0, + 438.0, + 384.0, + 438.0, + 1617.0, + 403.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 520.0, + 406.0, + 520.0, + 406.0, + 667.0, + 374.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 946.0, + 404.0, + 946.0, + 404.0, + 1051.0, + 374.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 1346.0, + 404.0, + 1346.0, + 404.0, + 1463.0, + 374.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 825.0, + 347.0, + 825.0, + 347.0, + 1483.0, + 311.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 799.0, + 943.0, + 799.0, + 943.0, + 1614.0, + 910.0, + 1614.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 21, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 585, + 401, + 1198, + 401, + 1198, + 1483, + 585, + 1483 + ], + "score": 0.974, + "html": "
1.55290.00668744444435339LL35.55759'8116.44695511.441114£99L1143423360.6778'8311.15
555418.3554318.0044.5511915678IL'LS40.64488189011117118951511.351L0116410338130
66.4441442555511.09533.35154.31435.5514445198'0L0153.59£9.09.044/.6156.559.55S1855438.5053816.7916.1460.00434
111152585-56.95-6'L[-4441-9L'9S-1135911.51555331460'8L7'01-£6.33-11.04410'61-1571-89.97-L6'67-353335331531-88'L7-40.05-
15:114353555.815531966986.69355555.3511.333335.21448.1135.3845.55335.713233219'8466.55370.3380.5535.7128.85
10028B28281055131641621328388084660
ererreresdererertrreresdeeiereeerereeeeeeertrrereeeeer erraeraseeereerteeeeeer crrrrreereeretrrereeeeerrereedreeeeerrereeeeerCereresdeeeeerrtrreresdeeererrreeereereeeeetrereedeercrrereeeerreerrereeeetrCgrrreesdrerecrraedticerseerrretrrereeeirereeeereecrreaeeeeereereeseirreeeeeeeeeeareeeeereeeiceerreeeeteseseeaEreereedeereeretrrereseeereedeeeeeeeseeeeterreresdeeeeerereeseseeeeeeresdreeseeereseeererarrgeaeedeereeeereeseeeereeeeserrreerrareseeercarereeeerreetrrreeeeerreeeeteeeeseeeeercaaeesdeteeerteseeeear
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1723, + 1404, + 1723, + 1404, + 1819, + 298, + 1819 + ], + "score": 0.965 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2112, + 835, + 2112 + ], + "score": 0.847 + }, + { + "category_id": 6, + "poly": [ + 496, + 425, + 534, + 425, + 534, + 1458, + 496, + 1458 + ], + "score": 0.364 + }, + { + "category_id": 6, + "poly": [ + 490, + 432, + 535, + 432, + 535, + 1454, + 490, + 1454 + ], + "score": 0.313 + }, + { + "category_id": 6, + "poly": [ + 563, + 491, + 598, + 491, + 598, + 1394, + 563, + 1394 + ], + "score": 0.181 + }, + { + "category_id": 6, + "poly": [ + 568, + 497, + 594, + 497, + 594, + 1392, + 568, + 1392 + ], + "score": 0.125 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 869.0, + 2084.0, + 869.0, + 2122.0, + 830.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 475.0, + 535.0, + 475.0, + 535.0, + 1424.0, + 495.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 474.0, + 536.0, + 474.0, + 536.0, + 1427.0, + 494.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 484.0, + 599.0, + 484.0, + 599.0, + 1186.0, + 564.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1304.0, + 593.0, + 1304.0, + 593.0, + 1353.0, + 568.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 491.0, + 598.0, + 491.0, + 598.0, + 1185.0, + 564.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1304.0, + 592.0, + 1304.0, + 592.0, + 1354.0, + 568.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1722.0, + 1403.0, + 1722.0, + 1403.0, + 1760.0, + 293.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1755.0, + 1403.0, + 1755.0, + 1403.0, + 1791.0, + 295.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1784.0, + 1116.0, + 1784.0, + 1116.0, + 1821.0, + 293.0, + 1821.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 22, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 341, + 270, + 1355, + 270, + 1355, + 1557, + 341, + 1557 + ], + "score": 0.982, + "html": "
Reward Min
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v1-31.719-43.328-64.958-58.811308.994301.4
SnakeFive-v1-56.38229.262-55.562-39.346
SnakeSix-v1-44.888-5.306-42.744-76.42266.066
SnakeSeven-v1-52.764-25.511-49.765-59.52231.557329.456333.237253.828
SnakeFive-v1-56.382-30.466-49.953-47.339
SnakeSix-v1SnakeSeven-v1-44.888-35.655-52.18-57.334-90.837-52.015
-52.76489.358
Reward Std
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v19.5445.79210.76910.1465.5135.788
SnakeFive-v113.3186.19111.7157.883
SnakeSix-v111.51513.53510.30416.8697.3957.1795.856.60414.084
SnakeSeven-v111.2112.13110.9816.953
SnakeFive-v113.31852.7111.5448.5198.51923.10913.529
SnakeSix-v1SnakeSeven-v111.51511.2164.54910.68310.52811.615
Reward Max
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v13.863-19.40910.1180.918338.374326.974
SnakeFive-v120.58563.8663.136-2.329
SnakeSix-v17.658.7793.51-2.139305.011
SnakeSeven-v1-2.2851.2651.325-28.087275.536
SnakeFive-v120.585139.6175.3051.487356.366
SnakeSix-v1SnakeSeven-v17.6-2.28156.771143.8242.3162.596-4.1598.394367.768336.561
Reward Average
RandomMLPAAMLPPTreeNetNerveNet
SnakeFour-v1-13.531-30.185-12.363-21.824325.476314.919
SnakeFive-v1-17.82951.411-16.481-21.256
SnakeSix-v1-19.8932.2-19.674-42.461282.426
SnakeSeven-v1-22.09923.533-22.378-42.742256.293
SnakeFive-v1-17.82989.823-16.54-15.975342.881
SnakeSix-v1-19.89105.632-20.454-34.343351.85313.149
SnakeSeven-v1-22.099115.123-22.81-22.383
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1656, + 1405, + 1656, + 1405, + 1782, + 298, + 1782 + ], + "score": 0.973 + }, + { + "category_id": 5, + "poly": [ + 596, + 1900, + 1103, + 1900, + 1103, + 1995, + 596, + 1995 + ], + "score": 0.967, + "html": "
parametersVminIntLen
value-100,-100, -2030,30,30
" + }, + { + "category_id": 6, + "poly": [ + 296, + 1809, + 1402, + 1809, + 1402, + 1874, + 296, + 1874 + ], + "score": 0.923 + }, + { + "category_id": 6, + "poly": [ + 551, + 224, + 1149, + 224, + 1149, + 258, + 551, + 258 + ], + "score": 0.911 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2113, + 834, + 2113 + ], + "score": 0.872 + }, + { + "category_id": 6, + "poly": [ + 297, + 1602, + 1205, + 1602, + 1205, + 1635, + 297, + 1635 + ], + "score": 0.327 + }, + { + "category_id": 13, + "poly": [ + 703, + 1718, + 767, + 1718, + 767, + 1754, + 703, + 1754 + ], + "score": 0.9, + "latex": "V _ { \\mathrm { m i n } }" + }, + { + "category_id": 13, + "poly": [ + 1154, + 1752, + 1174, + 1752, + 1174, + 1776, + 1154, + 1776 + ], + "score": 0.61, + "latex": "^ *" + }, + { + "category_id": 13, + "poly": [ + 945, + 1747, + 1147, + 1747, + 1147, + 1782, + 945, + 1782 + ], + "score": 0.33, + "latex": "\\lfloor \\mathrm { m a x } ( V ) / \\mathrm { I n t L e n } \\rfloor" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1809.0, + 1402.0, + 1809.0, + 1402.0, + 1841.0, + 296.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1841.0, + 1115.0, + 1841.0, + 1115.0, + 1874.0, + 298.0, + 1874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 219.0, + 1149.0, + 219.0, + 1149.0, + 262.0, + 549.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 829.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1601.0, + 1214.0, + 1601.0, + 1214.0, + 1639.0, + 295.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1659.0, + 1403.0, + 1659.0, + 1403.0, + 1688.0, + 296.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1688.0, + 1406.0, + 1688.0, + 1406.0, + 1721.0, + 295.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1714.0, + 702.0, + 1714.0, + 702.0, + 1763.0, + 289.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1714.0, + 1410.0, + 1714.0, + 1410.0, + 1763.0, + 768.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1747.0, + 944.0, + 1747.0, + 944.0, + 1783.0, + 295.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 1747.0, + 1153.0, + 1747.0, + 1153.0, + 1783.0, + 1148.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 1747.0, + 1263.0, + 1747.0, + 1263.0, + 1783.0, + 1175.0, + 1783.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 23, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 433, + 299, + 1264, + 299, + 1264, + 1142, + 433, + 1142 + ], + "score": 0.984, + "html": "
ModelMass
Hopper Halfhumanoid HorseWolfOstrich
NerveNet1981.592047.921968.872062.361852.762519.62392.052282.842161.611804.522304.251900.431862.211955.16922.12932.48983.11910.29791.84
1703.471648.281576.211354.51
1239.72946.08
MLP2268.641932.981637.471591.511073.1421.82115.75278.921224.98353.561692.081083.621932.641754.94878.75945.94
690.22538.14443.921459.931039.9814.55879.66805.99802.41
ModelStrength
HopperHalfhumanoidHorseWolfOstrich
NerveNet1961.621117.91492.33468.4446.141528.89940.96594.53308.08249.62316.82960.341827.73862.19274.77189.49133.84794.56539.45371.19255.12197.84
488.51305.16
305.16201.42
MLP800.15431.97429.76426.27412.391439.891040.341236.881730.881108.32230.85483.36
7209.54
765.7306.08519.73163.36
117.0565.29228.26187.7288.28108.97232.8133.99
" + }, + { + "category_id": 5, + "poly": [ + 539, + 1385, + 1159, + 1385, + 1159, + 1846, + 539, + 1846 + ], + "score": 0.981, + "html": "
MLP, TreeNetand NerveNetValue Tried
Network Shape[64, 64]
Number of Iteration Per Update10
Use KL PenaltyNo
Learning Rate SchedulerAdaptive
NerveNetValue Tried
Number of Propogation Steps4
NerveNetVariants Size of Nodes'Hidden StateNerveNet-1
64
Merge joint and body NodeYes
Output Network Disable Edge Type Separate
Add Skip-connection from /to rootYes No
" + }, + { + "category_id": 6, + "poly": [ + 523, + 1326, + 1175, + 1326, + 1175, + 1361, + 523, + 1361 + ], + "score": 0.902 + }, + { + "category_id": 1, + "poly": [ + 297, + 1959, + 1229, + 1959, + 1229, + 1994, + 297, + 1994 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 296, + 1903, + 888, + 1903, + 888, + 1936, + 296, + 1936 + ], + "score": 0.865 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2113, + 834, + 2113 + ], + "score": 0.865 + }, + { + "category_id": 6, + "poly": [ + 593, + 1166, + 1106, + 1166, + 1106, + 1201, + 593, + 1201 + ], + "score": 0.829 + }, + { + "category_id": 6, + "poly": [ + 298, + 230, + 785, + 230, + 785, + 262, + 298, + 262 + ], + "score": 0.701 + }, + { + "category_id": 0, + "poly": [ + 296, + 1261, + 957, + 1261, + 957, + 1295, + 296, + 1295 + ], + "score": 0.416 + }, + { + "category_id": 1, + "poly": [ + 296, + 1261, + 957, + 1261, + 957, + 1295, + 296, + 1295 + ], + "score": 0.391 + }, + { + "category_id": 0, + "poly": [ + 298, + 230, + 785, + 230, + 785, + 262, + 298, + 262 + ], + "score": 0.203 + }, + { + "category_id": 15, + "poly": [ + 521.0, + 1320.0, + 1177.0, + 1320.0, + 1177.0, + 1368.0, + 521.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1903.0, + 889.0, + 1903.0, + 889.0, + 1938.0, + 296.0, + 1938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2083.0, + 871.0, + 2083.0, + 871.0, + 2124.0, + 829.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 1160.0, + 1109.0, + 1160.0, + 1109.0, + 1206.0, + 589.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 787.0, + 229.0, + 787.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1260.0, + 958.0, + 1260.0, + 958.0, + 1296.0, + 296.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 787.0, + 229.0, + 787.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1957.0, + 1233.0, + 1957.0, + 1233.0, + 1999.0, + 294.0, + 1999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1260.0, + 958.0, + 1260.0, + 958.0, + 1296.0, + 296.0, + 1296.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 24, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 428, + 891, + 1275, + 891, + 1275, + 1293, + 428, + 1293 + ], + "score": 0.964 + }, + { + "category_id": 3, + "poly": [ + 479, + 1566, + 1215, + 1566, + 1215, + 1882, + 479, + 1882 + ], + "score": 0.959 + }, + { + "category_id": 3, + "poly": [ + 315, + 314, + 1394, + 314, + 1394, + 625, + 315, + 625 + ], + "score": 0.953 + }, + { + "category_id": 4, + "poly": [ + 297, + 661, + 1405, + 661, + 1405, + 724, + 297, + 724 + ], + "score": 0.946 + }, + { + "category_id": 4, + "poly": [ + 365, + 1928, + 1330, + 1928, + 1330, + 1962, + 365, + 1962 + ], + "score": 0.917 + }, + { + "category_id": 4, + "poly": [ + 296, + 1327, + 1404, + 1327, + 1404, + 1392, + 296, + 1392 + ], + "score": 0.906 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2113, + 834, + 2113 + ], + "score": 0.863 + }, + { + "category_id": 2, + "poly": [ + 508, + 314, + 599, + 314, + 599, + 334, + 508, + 334 + ], + "score": 0.111 + }, + { + "category_id": 15, + "poly": [ + 557.0, + 896.0, + 688.0, + 896.0, + 688.0, + 928.0, + 557.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 897.0, + 1088.0, + 897.0, + 1088.0, + 924.0, + 901.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 965.0, + 1089.0, + 965.0, + 1089.0, + 989.0, + 1042.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 984.0, + 665.0, + 984.0, + 665.0, + 1009.0, + 619.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 1060.0, + 676.0, + 1060.0, + 676.0, + 1097.0, + 617.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1058.0, + 1128.0, + 1058.0, + 1128.0, + 1094.0, + 1068.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 1061.0, + 1273.0, + 1061.0, + 1273.0, + 1092.0, + 1215.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 1152.0, + 528.0, + 1152.0, + 528.0, + 1237.0, + 494.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 1230.0, + 666.0, + 1230.0, + 666.0, + 1260.0, + 618.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 1246.0, + 1118.0, + 1246.0, + 1118.0, + 1278.0, + 1071.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1248.0, + 1263.0, + 1248.0, + 1263.0, + 1275.0, + 1216.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 486.0, + 1030.5, + 560.0, + 1030.5, + 560.0, + 1043.5, + 486.0, + 1043.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 1148.5, + 657.0, + 1148.5, + 657.0, + 1176.0, + 618.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.75, + 1159.0, + 1108.75, + 1159.0, + 1108.75, + 1188.0, + 1070.75, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.75, + 1159.0, + 1253.75, + 1159.0, + 1253.75, + 1188.0, + 1216.75, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 1568.0, + 917.0, + 1568.0, + 917.0, + 1594.0, + 807.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 1631.0, + 1108.0, + 1631.0, + 1108.0, + 1655.0, + 1060.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1660.0, + 740.0, + 1660.0, + 740.0, + 1667.0, + 732.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1680.0, + 724.0, + 1680.0, + 724.0, + 1690.0, + 716.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 1674.0, + 1104.0, + 1674.0, + 1104.0, + 1699.0, + 1060.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1715.0, + 959.0, + 1715.0, + 959.0, + 1745.0, + 905.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 1721.0, + 987.0, + 1721.0, + 987.0, + 1740.0, + 971.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 1715.0, + 1168.0, + 1715.0, + 1168.0, + 1746.0, + 1085.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1770.0, + 944.0, + 1770.0, + 944.0, + 1794.0, + 919.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 1771.0, + 992.0, + 1771.0, + 992.0, + 1801.0, + 950.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1769.0, + 1199.0, + 1769.0, + 1199.0, + 1799.0, + 1133.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1844.0, + 1004.0, + 1844.0, + 1004.0, + 1869.0, + 956.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 1844.0, + 1212.0, + 1844.0, + 1212.0, + 1869.0, + 1166.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1680.5, + 741.0, + 1680.5, + 741.0, + 1707.5, + 661.0, + 1707.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 1735.0, + 802.0, + 1735.0, + 802.0, + 1766.0, + 756.0, + 1766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 314.0, + 602.0, + 314.0, + 602.0, + 335.0, + 508.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 336.0, + 1149.0, + 336.0, + 1149.0, + 353.0, + 1077.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 354.0, + 723.0, + 354.0, + 723.0, + 380.0, + 695.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 359.0, + 761.0, + 359.0, + 761.0, + 377.0, + 722.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 368.0, + 486.0, + 368.0, + 486.0, + 383.0, + 465.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 369.0, + 1062.0, + 369.0, + 1062.0, + 409.0, + 1025.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 371.0, + 1254.0, + 371.0, + 1254.0, + 389.0, + 1222.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 390.0, + 1245.0, + 390.0, + 1245.0, + 414.0, + 1222.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 390.0, + 1338.0, + 390.0, + 1338.0, + 413.0, + 1282.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 427.0, + 721.0, + 427.0, + 721.0, + 453.0, + 695.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 432.0, + 756.0, + 432.0, + 756.0, + 450.0, + 718.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 450.0, + 592.0, + 450.0, + 592.0, + 470.0, + 562.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 449.0, + 1305.0, + 449.0, + 1305.0, + 459.0, + 1292.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 480.0, + 581.0, + 480.0, + 581.0, + 486.0, + 574.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 468.0, + 668.0, + 468.0, + 668.0, + 489.0, + 648.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 464.0, + 1158.0, + 464.0, + 1158.0, + 488.0, + 1118.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 466.0, + 1212.0, + 466.0, + 1212.0, + 487.0, + 1173.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 466.0, + 1342.0, + 466.0, + 1342.0, + 487.0, + 1302.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1349.0, + 466.0, + 1389.0, + 466.0, + 1389.0, + 487.0, + 1349.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 503.0, + 503.0, + 503.0, + 503.0, + 609.0, + 346.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 507.0, + 641.0, + 507.0, + 641.0, + 526.0, + 620.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 502.0, + 696.0, + 502.0, + 696.0, + 527.0, + 670.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 508.0, + 1180.0, + 508.0, + 1180.0, + 527.0, + 1158.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 507.0, + 1358.0, + 507.0, + 1358.0, + 527.0, + 1335.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 552.0, + 606.0, + 552.0, + 606.0, + 574.0, + 571.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 554.0, + 755.0, + 554.0, + 755.0, + 572.0, + 722.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1336.0, + 568.0, + 1358.0, + 568.0, + 1358.0, + 589.0, + 1336.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 586.0, + 1158.0, + 586.0, + 1158.0, + 601.0, + 1130.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 584.0, + 1213.0, + 584.0, + 1213.0, + 602.0, + 1184.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 586.0, + 1342.0, + 586.0, + 1342.0, + 601.0, + 1314.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1360.0, + 584.0, + 1391.0, + 584.0, + 1391.0, + 602.0, + 1360.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 595.0, + 662.0, + 595.0, + 662.0, + 613.0, + 633.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 595.0, + 751.0, + 595.0, + 751.0, + 613.0, + 723.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 377.5, + 498.0, + 377.5, + 498.0, + 386.5, + 489.0, + 386.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 401.0, + 1046.0, + 401.0, + 1046.0, + 419.0, + 998.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 463.0, + 702.0, + 463.0, + 702.0, + 480.0, + 669.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 501.0, + 618.0, + 501.0, + 618.0, + 524.0, + 589.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 500.5, + 724.0, + 500.5, + 724.0, + 523.5, + 693.0, + 523.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 510.0, + 1306.0, + 510.0, + 1306.0, + 525.5, + 1290.0, + 525.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.75, + 526.0, + 1153.75, + 526.0, + 1153.75, + 544.0, + 1128.75, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 525.0, + 1209.0, + 525.0, + 1209.0, + 544.0, + 1183.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 526.0, + 1338.0, + 526.0, + 1338.0, + 544.0, + 1313.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1359.75, + 525.0, + 1385.75, + 525.0, + 1385.75, + 544.0, + 1359.75, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.25, + 570.5, + 1177.25, + 570.5, + 1177.25, + 585.0, + 1158.25, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 658.0, + 1406.0, + 658.0, + 1406.0, + 698.0, + 294.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 693.0, + 495.0, + 693.0, + 495.0, + 724.0, + 295.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 1926.0, + 1336.0, + 1926.0, + 1336.0, + 1965.0, + 365.0, + 1965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1327.0, + 1405.0, + 1327.0, + 1405.0, + 1363.0, + 295.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1359.0, + 626.0, + 1359.0, + 626.0, + 1392.0, + 295.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2121.0, + 830.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 316.0, + 600.0, + 316.0, + 600.0, + 333.0, + 509.0, + 333.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 25, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/SJ3rcZcxl/SJ3rcZcxl.md b/parse/train/SJ3rcZcxl/SJ3rcZcxl.md new file mode 100644 index 0000000000000000000000000000000000000000..447b97b60360c5d10b887682b06f195cd1afc4ce --- /dev/null +++ b/parse/train/SJ3rcZcxl/SJ3rcZcxl.md @@ -0,0 +1,360 @@ +# Q-PROP: SAMPLE-EFFICIENT POLICY GRADIENT WITH AN OFF-POLICY CRITIC + +Shixiang $\mathbf { G u } ^ { 1 2 3 }$ , Timothy Lillicrap4, Zoubin Ghahramani16, Richard E. Turner1, Sergey Levine35 + +sg717@cam.ac.uk,countzero@google.com,zoubin@eng.cam.ac.uk, +ret26@cam.ac.uk,svlevine@eecs.berkeley.edu +1University of Cambridge, UK +2Max Planck Institute for Intelligent Systems, Tubingen, Germany ¨ +3Google Brain, USA +4DeepMind, UK +5UC Berkeley, USA +6Uber AI Labs, USA + +# ABSTRACT + +Model-free deep reinforcement learning (RL) methods have been successful in a wide variety of simulated domains. However, a major obstacle facing deep RL in the real world is their high sample complexity. Batch policy gradient methods offer stable learning, but at the cost of high variance, which often requires large batches. TD-style methods, such as off-policy actor-critic and Q-learning, are more sample-efficient but biased, and often require costly hyperparameter sweeps to stabilize. In this work, we aim to develop methods that combine the stability of policy gradients with the efficiency of off-policy RL. We present Q-Prop, a policy gradient method that uses a Taylor expansion of the off-policy critic as a control variate. Q-Prop is both sample efficient and stable, and effectively combines the benefits of on-policy and off-policy methods. We analyze the connection between Q-Prop and existing model-free algorithms, and use control variate theory to derive two variants of Q-Prop with conservative and aggressive adaptation. We show that conservative Q-Prop provides substantial gains in sample efficiency over trust region policy optimization (TRPO) with generalized advantage estimation (GAE), and improves stability over deep deterministic policy gradient (DDPG), the stateof-the-art on-policy and off-policy methods, on OpenAI Gym’s MuJoCo continuous control environments. + +# 1 INTRODUCTION + +Model-free reinforcement learning is a promising approach for solving arbitrary goal-directed sequential decision-making problems with only high-level reward signals and no supervision. It has recently been extended to utilize large neural network policies and value functions, and has been shown to be successful in solving a range of difficult problems (Mnih et al., 2015; Schulman et al., 2015; Lillicrap et al., 2016; Silver et al., 2016; Gu et al., 2016b; Mnih et al., 2016). Deep neural network parametrization minimizes the need for manual feature and policy engineering, and allows learning end-to-end policies mapping from high-dimensional inputs, such as images, directly to actions. However, such expressive parametrization also introduces a number of practical problems. Deep reinforcement learning algorithms tend to be sensitive to hyperparameter settings, often requiring extensive hyperparameter sweeps to find good values. Poor hyperparameter settings tend to produce unstable or non-convergent learning. Deep RL algorithms also tend to exhibit high sample complexity, often to the point of being impractical to run on real physical systems. Although a number of recent techniques have sought to alleviate some of these issues (Hasselt, 2010; Mnih et al., 2015; Schulman et al., 2015; 2016), these recent advances still provide only a partial solution to the instability and sample complexity challenges. + +Model-free reinforcement learning consists of on- and off-policy methods. Monte Carlo policy gradient methods (Peters & Schaal, 2006; Schulman et al., 2015) are popular on-policy methods that directly maximize the cumulative future returns with respect to the policy. While these algorithms can offer unbiased (or nearly unbiased, as discussed in Section 2.1) estimates of the gradient, they rely on Monte Carlo estimation and often suffer from high variance. To cope with high variance gradient estimates and difficult optimization landscapes, a number of techniques have been proposed, including constraining the change in the policy at each gradient step (Kakade, 2001; Peters et al., 2010) and mixing value-based back-ups to trade off bias and variance in Monte Carlo return estimates (Schulman et al., 2015). However, these methods all tend to require very large numbers of samples to deal with the high variance when estimating gradients of high-dimensional neural network policies. The crux of the problem with policy gradient methods is that they can only effectively use on-policy samples, which means that they require collecting large amounts of on-policy experiences after each parameter update to the policy. This makes them very sample intensive. Offpolicy methods, such as Q-learning (Watkins & Dayan, 1992; Sutton et al., 1999; Mnih et al., 2015; Gu et al., 2016b) and off-policy actor-critic methods (Lever, 2014; Lillicrap et al., 2016), can instead use all samples, including off-policy samples, by adopting temporal difference learning with experience replay. Such methods are much more sample-efficient. However, convergence of these algorithms is in general not guaranteed with non-linear function approximators, and practical convergence and instability issues typically mean that extensive hyperparameter tuning is required to attain good results. + +In order to make deep reinforcement learning practical as a tool for tackling real-world tasks, we must develop methods that are both data efficient and stable. In this paper, we propose Q-Prop, a step in this direction that combines the advantages of on-policy policy gradient methods with the efficiency of off-policy learning. Unlike prior approaches for off-policy learning, which either introduce bias (Sutton et al., 1999; Silver et al., 2014) or increase variance (Precup, 2000; Levine & Koltun, 2013; Munos et al., 2016), Q-Prop can reduce the variance of gradient estimator without adding bias; unlike prior approaches for critic-based variance reduction (Schulman et al., 2016) which fit the value function on-policy, Q-Prop learns the action-value function off-policy. The core idea is to use the first-order Taylor expansion of the critic as a control variate, resulting in an analytical gradient term through the critic and a Monte Carlo policy gradient term consisting of the residuals in advantage approximations. The method helps unify policy gradient and actor-critic methods: it can be seen as using the off-policy critic to reduce variance in policy gradient or using on-policy Monte Carlo returns to correct for bias in the critic gradient. We further provide theoretical analysis of the control variate, and derive two additional variants of Q-Prop. The method can be easily incorporated into any policy gradient algorithm. We show that Q-Prop provides substantial gains in sample efficiency over trust region policy optimization (TRPO) with generalized advantage estimation (GAE) (Schulman et al., 2015; 2016), and improved stability over deep deterministic policy gradient (DDPG) (Lillicrap et al., 2016) across a repertoire of continuous control tasks. + +# 2 BACKGROUND + +Reinforcement learning (RL) aims to learn a policy for an agent such that it behaves optimally according to a reward function. At a time step $t$ and state $\scriptstyle { s _ { t } }$ , the agent chooses an action $\mathbf { } \mathbf { a } _ { t }$ according to its policy $\pi ( \boldsymbol { a } _ { t } | \boldsymbol { s } _ { t } )$ , the state of the agent and the environment changes to new state $\boldsymbol { s } _ { t + 1 }$ according to dynamics $p \big ( \pmb { s } _ { t + 1 } \big | \pmb { s } _ { t } , \pmb { a } _ { t } \big )$ , the agent receives a reward $r ( s _ { t } , a _ { t } )$ , and the process continues. Let $R _ { t }$ denote a $\gamma$ -discounted cumulative return from $t$ for an infinite horizon problem, i.e $\begin{array} { r } { R _ { t } = \sum _ { t ^ { \prime } = t } ^ { \infty } \gamma ^ { t ^ { \prime } - t } r ( s _ { t ^ { \prime } } , { \bf a } _ { t ^ { \prime } } ) } \end{array}$ . The goal of reinforcement learning is to maximize the expected return $J ( \theta ) = \mathbb { E } _ { \pi _ { \theta } } [ R _ { 0 } ]$ with respect to the policy parameters θ . In this section, we review several standard techniques for performing this optimization, and in the next section, we will discuss our proposed Q-Prop algorithm that combines the strengths of these approaches to achieve efficient, stable RL. Monte Carlo policy gradient refers to policy gradient methods that use full Monte Carlo returns, e.g. REINFORCE (Williams, 1992) and TRPO (Schulman et al., 2015), and policy gradient with function approximation refers to actor-critic methods (Sutton et al., 1999) which optimize the policy against a critic, e.g. deterministic policy gradient (Silver et al., 2014; Lillicrap et al., 2016). + +# 2.1 MONTE CARLO POLICY GRADIENT METHODS + +Monte Carlo policy gradient methods apply direct gradient-based optimization to the reinforcement learning objective. This involves directly differentiating the $J ( \theta )$ objective with respect to the policy + +parameters $\theta$ . The standard form, known as the REINFORCE algorithm (Williams, 1992), is shown below: + +$$ +\nabla _ { \theta } J ( \theta ) = \mathbb { E } _ { \pi } [ \sum _ { t = 0 } ^ { \infty } \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } | s _ { t } ) \gamma ^ { t } R _ { t } ] = \mathbb { E } _ { \pi } [ \sum _ { t = 0 } ^ { \infty } \gamma ^ { t } \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } | s _ { t } ) ( R _ { t } - b ( s _ { t } ) ) ] , +$$ + +where $b ( s _ { t } )$ is known as the baseline. For convenience of later derivations, Eq. 1 can also be written as below, where $\begin{array} { r } { \rho _ { \pi } ( s ) = \sum _ { t = 0 } ^ { \infty } \gamma ^ { t } p ( s _ { t } = s ) } \end{array}$ is the unnormalized discounted state visitation frequency, + +$$ +\nabla _ { \theta } J ( \theta ) = \mathbb { E } _ { s _ { t } \sim \rho _ { \pi } ( . ) , \mathbf { a } _ { t } \sim \pi ( \cdot | s _ { t } ) } [ \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } | s _ { t } ) ( R _ { t } - b ( s _ { t } ) ) ] . +$$ + +Eq. 2 is an unbiased gradient of the RL objective. However, in practice, most policy gradient methods effectively use undiscounted state visitation frequencies, i.e. $\gamma = 1$ in the equal for $\rho _ { \pi }$ , and are therefore biased; in fact, making them unbiased often hurts performance (Thomas, 2014). In this paper, we mainly discuss bias due to function approximation, off-policy learning, and value back-ups. + +The gradient is estimated using Monte Carlo samples in practice and has very high variance. A proper choice of baseline is necessary to reduce the variance sufficiently such that learning becomes feasible. A common choice is to estimate the value function of the state $V _ { \pi } ( \pmb { s } _ { t } )$ to use as the baseline, which provides an estimate of advantage function $A _ { \pi } ( s _ { t } , { \pmb a } _ { t } )$ , which is a centered action-value function $Q _ { \pi } ( s _ { t } , \pmb { a } _ { t } )$ , as defined below: + +$$ +\begin{array} { c } { { V _ { \pi } ( s _ { t } ) = \mathbb { E } _ { \pi } [ R _ { t } ] = \mathbb { E } _ { \pi _ { \theta } ( a _ { t } | s _ { t } ) } [ Q _ { \pi } ( s _ { t } , a _ { t } ) ] } } \\ { { Q _ { \pi } ( s _ { t } , a _ { t } ) = r ( s _ { t } , a _ { t } ) + \gamma \mathbb { E } _ { \pi } [ R _ { t + 1 } ] = r ( s _ { t } , a _ { t } ) + \gamma \mathbb { E } _ { p ( s _ { t + 1 } | s _ { t } , a _ { t } ) } [ V _ { \pi } ( s _ { t + 1 } ) ] } } \\ { { A _ { \pi } ( s _ { t } , a _ { t } ) = Q _ { \pi } ( s _ { t } , a _ { t } ) - V _ { \pi } ( s _ { t } ) . } } \end{array} +$$ + +$Q _ { \pi } ( s _ { t } , \pmb { a } _ { t } )$ summarizes the performance of each action from a given state, assuming it follows $\pi$ thereafter, and $A _ { \pi } ( s _ { t } , { \pmb a } _ { t } )$ provides a measure of how each action compares to the average performance at the state $\scriptstyle { s _ { t } }$ , which is given by $V _ { \pi } ( \pmb { s } _ { t } )$ . Using $A _ { \pi } ( s _ { t } , { \pmb a } _ { t } )$ centers the learning signal and reduces variance significantly. + +Besides high variance, another problem with the policy gradient is that it requires on-policy samples. This makes policy gradient optimization very sample intensive. To achieve similar sample efficiency as off-policy methods, we can attempt to include off-policy data. Prior attempts use importance sampling to include off-policy trajectories; however, these are known to be difficult scale to highdimensional action spaces because of rapidly degenerating importance weights (Precup, 2000). + +# 2.2 POLICY GRADIENT WITH FUNCTION APPROXIMATION + +Policy gradient methods with function approximation (Sutton et al., 1999), or actor-critic methods, include a policy evaluation step, which often uses temporal difference (TD) learning to fit a critic $Q _ { w }$ for the current policy $\pi ( \theta )$ , and a policy improvement step which greedily optimizes the policy $\pi$ against the critic estimate $Q _ { w }$ . Significant gains in sample efficiency may be achievable using offpolicy TD learning for the critic, as in Q-learning and deterministic policy gradient (Sutton, 1990; Silver et al., 2014), typically by means of experience replay for training deep Q networks (Mnih et al., 2015; Lillicrap et al., 2016; Gu et al., 2016b). + +One particularly relevant example of such a method is the deep deterministic policy gradient (DDPG) (Silver et al., 2014; Lillicrap et al., 2016). The updates for this method are given below, where $\pi _ { \boldsymbol { \theta } } ( \mathbf { a } _ { t } | \mathbf { s } _ { t } ) = \delta ( \mathbf { a } _ { t } = \mu _ { \boldsymbol { \theta } } ( \mathbf { s } _ { t } ) )$ is a deterministic policy, $\beta$ is arbitrary exploration distribution, and $\rho _ { \beta }$ corresponds to sampling from a replay buffer. $Q ( \cdot , \cdot )$ is the target network that slowly tracks $Q _ { w }$ (Lillicrap et al., 2016). + +$$ +\begin{array} { r l } & { w = \underset { w } { \arg \operatorname* { m i n } } \mathbb { E } _ { s _ { t } \sim \rho _ { \beta } ( \cdot ) , a _ { t } \sim \beta ( \cdot | s _ { t } ) } \big [ \big ( r \big ( s _ { t } , a _ { t } \big ) + \gamma Q ( s _ { t + 1 } , \mu _ { \theta } ( s _ { t + 1 } ) \big ) - Q _ { w } ( s _ { t } , a _ { t } ) \big ) ^ { 2 } \big ] } \\ & { \theta = \underset { \theta } { \arg \operatorname* { m a x } } \mathbb { E } _ { s _ { t } \sim \rho _ { \beta } ( \cdot ) } [ Q _ { w } \big ( s _ { t } , \mu _ { \theta } ( s _ { t } ) \big ) ] } \end{array} +$$ + +When the critic and policy are parametrized with neural networks, full optimization is expensive, and instead stochastic gradient optimization is used. The gradient in the policy improvement phase is given below, which is generally a biased gradient of $J ( \theta )$ . + +$$ +\nabla _ { \theta } J ( \theta ) \approx \mathbb { E } _ { s _ { t } \sim \rho _ { \beta } ( \cdot ) } [ \nabla _ { a } Q _ { w } \big ( s _ { t } , a \big ) | _ { a = \mu _ { \theta } ( s _ { t } ) } \nabla _ { \theta } \mu _ { \theta } \big ( s _ { t } \big ) ] +$$ + +The crucial benefits of DDPG are that it does not rely on high variance REINFORCE gradients and is trainable on off-policy data. These properties make DDPG and other analogous off-policy methods significantly more sample-efficient than policy gradient methods (Lillicrap et al., 2016; Gu et al., 2016b; Duan et al., 2016). However, the use of a biased policy gradient estimator makes analyzing its convergence and stability properties difficult. + +# 3 Q-PROP + +In this section, we derive the Q-Prop estimator for policy gradient. The key idea from this estimator comes from observing Equations 2 and 5 and noting that the former provides an almost unbiased (see Section 2.1), but high variance gradient, while the latter provides a deterministic, but biased gradient. By using the deterministic biased estimator as a particular form of control variate (Ross, 2006; Paisley et al., 2012) for the Monte Carlo policy gradient estimator, we can effectively use both types of gradient information to construct a new estimator that in practice exhibits improved sample efficiency through the inclusion of off-policy samples while preserving the stability of on-policy Monte Carlo policy gradient. + +# 3.1 Q-PROP ESTIMATOR + +To derive the Q-Prop gradient estimator, we start by using the first-order Taylor expansion of an arbitrary function $\hat { f ( s _ { t } , a _ { t } ) , ~ } \bar { f } ( s _ { t } , a _ { t } ) = f ( s _ { t } , \bar { a } _ { t } ) + \bar { \nabla } _ { a } f ( \bar { s _ { t } } , a ) | _ { a = \bar { a } _ { t } } ( a _ { t } - \bar { a } _ { t } ) \cdot$ as the control variate for the policy gradient estimator. We use $\begin{array} { r } { \hat { \boldsymbol { \mathcal Q } } ( \boldsymbol s _ { t } , \boldsymbol a _ { t } ) = \sum _ { t ^ { \prime } = t } ^ { \infty } \gamma ^ { t ^ { \prime } - t } \boldsymbol r ( \boldsymbol s _ { t ^ { \prime } } , \boldsymbol a _ { t ^ { \prime } } ) } \end{array}$ to denote Monte Carlo return from state $\scriptstyle { \pmb { s } } _ { t }$ and action $\mathbf { } \mathbf { } \mathbf { } \mathbf { a } _ { t }$ , i.e. $\mathbb { E } _ { \pi } [ \hat { \cal Q } ( s _ { t } , a _ { t } ) ] = r ( s _ { t } , a _ { t } ) + \gamma \mathbb { E } _ { p } [ V _ { \pi } ( s _ { t + 1 } ) ]$ , and $\mu _ { \theta } ( s _ { t } ) = \mathbb { E } _ { \pi _ { \theta } ( a _ { t } | s _ { t } ) } [ a _ { t } ]$ to denote the expected action of a stochastic policy $\pi _ { \theta }$ . Full derivation is in Appendix A. + +$$ +\begin{array} { r } { \nabla _ { \theta } J ( \theta ) = \mathbb { E } _ { \rho _ { \pi } , \pi } [ \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } | s _ { t } ) ( \hat { Q } ( s _ { t } , a _ { t } ) - \bar { f } ( s _ { t } , a _ { t } ) ] + \mathbb { E } _ { \rho _ { \pi } , \pi } [ \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } | s _ { t } ) \bar { f } ( s _ { t } , a _ { t } ) ] } \\ { = \mathbb { E } _ { \rho _ { \pi } , \pi } [ \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } | s _ { t } ) ( \hat { Q } ( s _ { t } , a _ { t } ) - \bar { f } ( s _ { t } , a _ { t } ) ] + \mathbb { E } _ { \rho _ { \pi } } [ \nabla _ { a } f ( s _ { t } , a ) | _ { a = \bar { a } _ { t } } \nabla _ { \theta } \mu _ { \theta } ( s _ { t } ) ] } \end{array} +$$ + +Eq. 6 is general for arbitrary function $f ( \mathbf { \boldsymbol { s } } _ { t } , \mathbf { \boldsymbol { a } } _ { t } )$ that is differentiable with respect to $\mathbf { } \mathbf { } \mathbf { } \mathbf { a } _ { t }$ at an arbitrary value of $\bar { \mathbf { a } } _ { t }$ ; however, a sensible choice is to use the critic $Q _ { w }$ for $f$ and $\pmb { \mu } _ { \pmb { \theta } } ( \pmb { s } _ { t } )$ for $\bar { \mathbf { a } } _ { t }$ to get, + +$$ +\nabla _ { \theta } J ( \theta ) = \mathbb { E } _ { \rho _ { \pi } , \pi } [ \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } | s _ { t } ) ( \hat { Q } ( s _ { t } , a _ { t } ) - \bar { Q } _ { w } ( s _ { t } , a _ { t } ) ] + \mathbb { E } _ { \rho _ { \pi } } [ \nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \mu _ { \theta } ( s _ { t } ) } \nabla _ { \theta } \mu _ { \theta } ( s _ { t } ) ] . +$$ + +Finally, since in practice we estimate advantages $\hat { A } ( s _ { t } , a _ { t } )$ , we write the Q-Prop estimator in terms of advantages to complete the basic derivation, + +$$ +\begin{array} { r l } & { \nabla _ { \theta } J ( \theta ) = \mathbb { E } _ { \rho _ { \pi } , \pi } [ \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } | s _ { t } ) ( \hat { A } ( s _ { t } , a _ { t } ) - \bar { A } _ { w } ( s _ { t } , a _ { t } ) ] + \mathbb { E } _ { \rho _ { \pi } } [ \nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \mu _ { \theta } ( s _ { t } ) } \nabla _ { \theta } \mu _ { \theta } ( s _ { t } ) ] } \\ & { \bar { A } ( s _ { t } , a _ { t } ) = \bar { Q } ( s _ { t } , a _ { t } ) - \mathbb { E } _ { \pi _ { \theta } } [ \bar { Q } ( s _ { t } , a _ { t } ) ] = \nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \mu _ { \theta } ( s _ { t } ) } ( a _ { t } - \mu _ { \theta } ( s _ { t } ) ) . } \end{array} +$$ + +Eq. 8 is composed of an analytic gradient through the critic as in Eq. 5 and a residual REINFORCE gradient in Eq. 2. From the above derivation, Q-Prop is simply a Monte Carlo policy gradient estimator with a special form of control variate. The important insight comes from the fact that $Q _ { w }$ can be trained using off-policy data as in Eq. 4. Under this setting, Q-Prop is no longer just a Monte Carlo policy gradient method, but more closely resembles an actor-critic method, where the critic can be updated off-policy but the actor is always updated on-policy with an additional REINFORCE correction term so that it remains a Monte Carlo policy gradient method regardless of the parametrization, training method, and performance of the critic. Therefore, Q-Prop can be directly combined with a number of prior techniques from both on-policy methods such as natural policy gradient (Kakade, 2001), trust-region policy optimization (TRPO) (Schulman et al., 2015) and generalized advantage estimation (GAE) (Schulman et al., 2016), and off-policy methods such as DDPG (Lillicrap et al., 2016) and Retrace $( \lambda )$ (Munos et al., 2016). + +Intuitively, if the critic $Q _ { w }$ approximates $Q _ { \pi }$ well, it provides a reliable gradient, reduces the estimator variance, and improves the convergence rate. Interestingly, control variate analysis in the next section shows that this is not the only circumstance where Q-Prop helps reduce variance. + +# 3.2 CONTROL VARIATE ANALYSIS AND ADAPTIVE Q-PROP + +For Q-Prop to be applied reliably, it is crucial to analyze how the variance of the estimator changes before and after the application of control variate. Following the prior work on control variates (Ross, 2006; Paisley et al., 2012), we first introduce $\eta ( s _ { t } )$ to Eq. 8, a weighing variable that modulates the strength of control variate. This additional variable $\eta ( s _ { t } )$ does not introduce bias to the estimator. + +$$ +\begin{array} { r l } & { \nabla _ { \theta } J ( \theta ) = \mathbb { E } _ { \rho _ { \pi } , \pi } [ \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } | s _ { t } ) ( \hat { A } ( s _ { t } , a _ { t } ) - \eta ( s _ { t } ) \bar { A } _ { w } ( s _ { t } , a _ { t } ) ] } \\ & { \qquad + \mathbb { E } _ { \rho _ { \pi } } [ \eta ( s _ { t } ) \nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \mu _ { \theta } ( s _ { t } ) } \nabla _ { \theta } \mu _ { \theta } ( s _ { t } ) ] } \end{array} +$$ + +The variance of this estimator is given below, where $m = 1 . . . M$ indexes the dimension of $\theta$ , + +$$ +\mathrm { V a r } ^ { * } = \mathbb { E } _ { \rho _ { \pi } } \left[ \sum _ { m } \mathrm { V a r } _ { a _ { t } } ( \nabla _ { \theta _ { m } } \log \pi _ { \theta } ( a _ { t } | s _ { t } ) ( \hat { A } ( s _ { t } , a _ { t } ) - \eta ( s _ { t } ) \bar { A } ( s _ { t } , a _ { t } ) ) ) \right] . +$$ + +If we choose $\eta ( s _ { t } )$ such that $\mathrm { V a r ^ { * } } < \mathrm { V a r }$ , where $\begin{array} { r } { \mathrm { V a r } = \mathbb E _ { \rho _ { \pi } } [ \sum _ { m } \mathrm { V a r } _ { \boldsymbol { a } _ { t } } ( \nabla _ { \boldsymbol { \theta } _ { m } } \log \pi _ { \boldsymbol { \theta } } ( \boldsymbol { a } _ { t } | \boldsymbol { s } _ { t } ) \hat { \boldsymbol { A } } ( \boldsymbol { s } _ { t } , \boldsymbol { a } _ { t } ) ) ] } \end{array}$ is the original estimator variance measure, then we have managed to reduce the variance. Directly analyzing the above variance measure is nontrivial, for the same reason that computing the optimal baseline is difficult (Weaver & Tao, 2001). In addition, it is often impractical to get multiple action samples from the same state, which prohibits using na¨ıve Monte Carlo to estimate the expectations. Instead, we propose a surrogate variance measure, $\operatorname { V a r } = \mathbb { E } _ { \rho _ { \pi } } [ \operatorname { V a r } _ { \mathbf { a } _ { t } } ( \hat { A } ( s _ { t } , \mathbf { a } _ { t } ) ) ]$ . A similar surrogate is also used by prior work on learning state-dependent baseline (Mnih & Gregor, 2014), and the benefit is that the measure becomes more tractable, + +$$ +\begin{array} { r l } & { \mathrm { V a r } ^ { * } = \mathbb { E } _ { \rho _ { \pi } } [ \mathrm { V a r } _ { a _ { t } } ( \hat { A } ( s _ { t } , a _ { t } ) - \eta ( s _ { t } ) \bar { A } ( s _ { t } , a _ { t } ) ) ] } \\ & { \qquad = \mathrm { V a r } + \mathbb { E } _ { \rho _ { \pi } } [ - 2 \eta ( s _ { t } ) \mathrm { C o v } _ { a _ { t } } ( \hat { A } ( s _ { t } , a _ { t } ) , \bar { A } ( s _ { t } , a _ { t } ) ) + \eta ( s _ { t } ) ^ { 2 } \mathrm { V a r } _ { a _ { t } } ( \bar { A } ( s _ { t } , a _ { t } ) ) ] . } \end{array} +$$ + +Since $\mathbb { E } _ { \pi } [ \hat { A } ( \pmb { s } _ { t } , \pmb { a } _ { t } ) ] = \mathbb { E } _ { \pi } [ \bar { A } ( \pmb { s } _ { t } , \pmb { a } _ { t } ) ] = 0$ , the terms can be simplified as below, + +$$ +\begin{array} { r l } & { \mathrm { C o v } _ { a _ { t } } ( \hat { A } , \bar { A } ) = \mathbb { E } _ { \pi } [ \hat { A } ( s _ { t } , a _ { t } ) \bar { A } ( s _ { t } , a _ { t } ) ] } \\ & { \quad \mathrm { V a r } _ { a _ { t } } ( \bar { A } ) = \mathbb { E } _ { \pi } [ \bar { A } ( s _ { t } , a _ { t } ) ^ { 2 } ] = \nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \mu _ { \theta } ( s _ { t } ) } ^ { T } \Sigma _ { \theta } ( s _ { t } ) \nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \mu _ { \theta } ( s _ { t } ) } , } \end{array} +$$ + +where $\Sigma _ { \theta } ( s _ { t } )$ is the covariance matrix of the stochastic policy $\pi _ { \theta }$ . The nice property of Eq. 11 is that $\mathrm { V a r } _ { \pmb { a } _ { t } } ( \bar { A } )$ is analytical and $\mathrm { C o v } _ { a _ { t } } ( \hat { A } , \bar { A } )$ can be estimated with single action sample. Using this estimate, we propose adaptive variants of Q-Prop that regulate the variance of the gradient estimate. + +Adaptive Q-Prop. The optimal state-dependent factor $\eta ( s _ { t } )$ can be computed per state, according to $\eta ^ { * } ( s _ { t } ) = \mathrm { C o v } _ { a _ { t } } ( \hat { A } , \bar { A } ) / \mathrm { V a r } _ { a _ { t } } ( \bar { A } )$ . This provides maximum reduction in variance according to Eq. 11. Substituting $\eta ^ { * } ( s _ { t } )$ into Eq. 11, we get $\mathrm { V a r } ^ { * } = \mathbb { E } _ { \rho _ { \pi } } [ ( 1 - \rho _ { c o r r } ( \hat { A } , \bar { A } ) ^ { 2 } ) \mathrm { V a r } _ { a _ { t } } ( \hat { A } ) ]$ , where $\rho _ { c o r r }$ is the correlation coefficient, which achieves guaranteed variance reduction if at any state $\bar { A }$ is correlated with $\hat { A }$ . We call this the fully adaptive Q-Prop method. An important conclusion from this analysis is that, in adaptive Q-Prop, the critic $Q _ { w }$ does not necessarily need to be approximating $Q _ { \pi }$ well to produce good results. Its Taylor expansion merely needs to be correlated with $\hat { A }$ , positively or even negatively. This is in contrast with actor-critic methods, where performance is greatly dependent on the absolute accuracy of the critic’s approximation. + +Conservative and Aggressive Q-Prop. In practice, the single-sample estimate of $\mathrm { C o v } _ { a _ { t } } ( \hat { A } , \bar { A } )$ has high variance itself, and we propose the following two practical implementations of adaptive Q-Prop: (1) $\eta ( s _ { t } ) = 1$ if $\hat { \mathrm { C o v } } _ { { \pmb a } _ { t } } ( \hat { A } , \bar { A } ) > \bar { 0 }$ and $\eta ( s _ { t } ) = 0$ if otherwise, and (2) $\eta ( \pmb { \mathscr { s } } _ { t } ) = \mathrm { s i g n } \big ( \mathsf { C } \hat { \operatorname { o v } } _ { \pmb { a } _ { t } } ( \hat { A } , \bar { A } ) \big )$ . The first implementation, which we call conservative Q-Prop, can be thought of as a more conservative version of Q-Prop, which effectively disables the control variate for some samples of the states. This is sensible as if $\hat { A }$ and $\bar { A }$ are negatively correlated, it is likely that the critic is very poor. The second variant can correspondingly be termed aggressive Q-Prop, since it makes more liberal use of the control variate. + +# 3.3 Q-PROP ALGORITHM + +Pseudo-code for the adaptive Q-Prop algorithm is provided in Algorithm 1. It is a mixture of policy gradient and actor-critic. At each iteration, it first rolls out the stochastic policy to collect on-policy + +# Algorithm 1 Adaptive Q-Prop + +1: Initialize $w$ for critic $Q _ { w }$ , $\theta$ for stochastic policy $\pi _ { \theta }$ , and replay buffer ${ \mathcal { R } } \gets \emptyset$ . +2 : repeat +3: for $e = 1 , \ldots , E$ do . Collect $E$ episodes of on-policy experience using $\pi _ { \theta }$ +4: $s _ { 0 , e } \sim p ( s _ { 0 } )$ +5: for t = 0, . . . , T − 1 do +6: $a _ { t , e } \sim \pi _ { \theta } ( \cdot | s _ { t , e } ) , s _ { t + 1 , e } \sim p ( \cdot | s _ { t , e } , a _ { t , e } ) , r _ { t , e } = r ( s _ { t , e } , a _ { t , e } )$ +7: Add batch data $\mathcal { B } = \left. s _ { 0 : T , 1 : E } , \pmb { a } _ { 0 : T - 1 , 1 : E } , r _ { 0 : T - 1 , 1 : E } \right.$ to replay buffer $\mathcal { R }$ +8: Take $E \cdot T$ gradient steps on $Q _ { w }$ using $\mathcal { R }$ and $\pi _ { \theta }$ +9: Fit $V _ { \phi } ( \pmb { s } _ { t } )$ using $\mathcal { B }$ +10: Compute $\hat { A } _ { t , e }$ using $\mathrm { G A E } ( \lambda )$ and $\bar { A } _ { t , e }$ using Eq. 7 +11: Set $\eta _ { t , e }$ based on Section 3.2 +12: Compute and center the learning signals $l _ { t , e } = \hat { A } _ { t , e } - \eta _ { t , e } \bar { A } _ { t , e }$ +13: Compute $\begin{array} { r } { \nabla _ { \theta } J ( \theta ) \approx \frac { 1 } { E T } \sum _ { e } \sum _ { t } \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t , e } | s _ { t , e } ) l _ { t , e } + \eta _ { t , e } \nabla _ { a } Q _ { w } ( s _ { t , e } , a ) | _ { a = \mu _ { \theta } ( s _ { t , e } ) } \nabla _ { \theta } \mu _ { \theta } ( s _ { t , e } ) } \end{array}$ +14: Take a gradient step on $\pi _ { \theta }$ using $\nabla _ { \boldsymbol { \theta } } J ( \boldsymbol { \theta } )$ , optionally with a trust-region constraint using $\mathcal { B }$ +15: until $\pi _ { \theta }$ converges. + +samples, adds the batch to a replay buffer, takes a few gradient steps on the critic, computes $\hat { A }$ and $\bar { A }$ , and finally applies a gradient step on the policy $\pi _ { \theta }$ . In our implementation, the critic $Q _ { w }$ is fitted with off-policy TD learning using the same techniques as in DDPG (Lillicrap et al., 2016): + +$$ +\begin{array} { r } { w = \underset { w } { \arg \operatorname* { m i n } } \mathbb { E } _ { s _ { t } \sim \rho _ { \beta } ( \cdot ) , a _ { t } \sim \beta ( \cdot | s _ { t } ) } [ \big ( r ( s _ { t } , a _ { t } ) + \gamma \mathbb { E } _ { \pi } [ Q ^ { \prime } ( s _ { t + 1 } , a _ { t + 1 } ) ] - Q _ { w } ( s _ { t } , a _ { t } ) \big ) ^ { 2 } ] . } \end{array} +$$ + +$V _ { \phi }$ is fitted with the same technique in (Schulman et al., 2016). Generalized advantage estimation (GAE) (Schulman et al., 2016) is used to estimate $\hat { A }$ . The policy update can be done by any method that utilizes the first-order gradient and possibly the on-policy batch data, which includes trust region policy optimization (TRPO) (Schulman et al., 2015). Importantly, this is just one possible implementation of Q-Prop, and in Appendix C we show a more general form that can interpolate between pure policy gradient and off-policy actor-critic. + +# 3.4 LIMITATIONS + +A limitation with Q-Prop is that if data collection is very fast, e.g. using fast simulators, the compute time per episode is bound by the critic training at each iteration, and similar to that of DDPG and usually much more than that of TRPO. However, in applications where data collection speed is the bottleneck, there is sufficient time between policy updates to fit $Q _ { w }$ well, which can be done asynchronously from the data collection, and the compute time of Q-Prop will be about the same as that of TRPO. + +Another limitation is the robustness to bad critics. We empirically show that our conservative Q-Prop is more robust than standard Q-Prop and much more robust than pure off-policy actor-critic methods such as DDPG; however, estimating when an off-policy critic is reliable or not is still a fundamental problem that shall be further investigated. We can also alleviate this limitation by adopting more stable off-policy critic learning techniques such as Retrace $( \lambda )$ (Munos et al., 2016). + +# 4 RELATED WORK + +Variance reduction in policy gradient methods is a long-standing problem with a large body of prior work (Weaver & Tao, 2001; Greensmith et al., 2004; Schulman et al., 2016). However, exploration of action-dependent control variates is relatively recent, with most work focusing instead on simpler baselining techniques (Ross, 2006). A subtle exception is compatible feature approximation (Sutton et al., 1999) which can be viewed as a control variate as explained in Appendix B. Another exception is doubly robust estimator in contextual bandits (Dud´ık et al., 2011), which uses a different control variate whose bias cannot be tractably corrected. Control variates were explored recently not in RL but for approximate inference in stochastic models (Paisley et al., 2012), and the closest related work in that domain is the MuProp algorithm (Gu et al., 2016a) which uses a mean-field network as a surrogate for backpropagating a deterministic gradient through stochastic discrete variables. MuProp is not directly applicable to model-free RL because the dynamics are unknown; however, it can be if the dynamics are learned as in model-based RL (Atkeson & Santamaria, 1997; Deisenroth & Rasmussen, 2011). This model-based Q-Prop is itself an interesting direction of research as it effectively corrects bias in model-based learning. + +Part of the benefit of Q-Prop is the ability to use off-policy data to improve on-policy policy gradient methods. Prior methods that combine off-policy data with policy gradients either introduce bias (Sutton et al., 1999; Silver et al., 2014) or use importance weighting, which is known to result in degenerate importance weights in high dimensions, resulting in very high variance (Precup, 2000; Levine & Koltun, 2013). Q-Prop provides a new approach for using off-policy data to reduce variance without introducing further bias. + +Lastly, since Q-Prop uses both on-policy policy updates and off-policy critic learning, it can take advantage of prior work along both lines of research. We chose to implement Q-Prop on top of TRPO-GAE primarily for the purpose of enabling a fair comparison in the experiments, but combining Q-Prop with other on-policy update schemes and off-policy critic training methods is an interesting direction for future work. For example, Q-Prop can also be used with other on-policy policy gradient methods such as A3C (Mnih et al., 2016) and off-policy advantage estimation methods such as Retrace $( \lambda )$ (Munos et al., 2016), GTD2 (Sutton et al., 2009), emphatic TD (Sutton et al., 2015), and WIS-LSTD (Mahmood et al., 2014). + +# 5 EXPERIMENTS + +![](images/8f3888758beaed39ef5a270faa0f5bf0d9ee7237dbc5a6eaa9d962ade37ee8e4.jpg) +Figure 1: Illustrations of OpenAI Gym MuJoCo domains (Brockman et al., 2016; Duan et al., 2016): (a) Ant, (b) HalfCheetah, (c) Hopper, (d) Humanoid, (e) Reacher, (f) Swimmer, (g) Walker. + +We evaluated Q-Prop and its variants on continuous control environments from the OpenAI Gym benchmark (Brockman et al., 2016) using the MuJoCo physics simulator (Todorov et al., 2012) as shown in Figure 1. Algorithms are identified by acronyms, followed by a number indicating batch size, except for DDPG, which is a prior online actor-critic algorithm (Lillicrap et al., 2016). “c-” and “v-” denote conservative and aggressive Q-Prop variants as described in Section 3.2. “TR-” denotes trust-region policy optimization (Schulman et al., 2015), while “V-” denotes vanilla policy gradient. For example, “TR-c-Q-Prop-5000” means convervative Q-Prop with the trust-region policy update, and a batch size of 5000. “VPG” and “TRPO” are vanilla policy gradient and trust-region policy optimization respectively (Schulman et al., 2016; Duan et al., 2016). Unless otherwise stated, all policy gradient methods are implemented with GAE $\lambda = 0 . 9 7$ ) (Schulman et al., 2016). Note that TRPOGAE is currently the state-of-the-art method on most of the OpenAI Gym benchmark tasks, though our experiments show that a well-tuned DDPG implementation sometimes achieves better results. Our algorithm implementations are built on top of the rllab TRPO and DDPG codes from Duan et al. (2016) and available at https://github.com/shaneshixiang/rllabplusplus. Policy and value function architectures and other training details including hyperparameter values are provided in Appendix D. + +# 5.1 ADAPTIVE Q-PROP + +First, it is useful to identify how reliable each variant of Q-Prop is. In this section, we analyze standard Q-Prop and two adaptive variants, c-Q-Prop and a-Q-Prop, and demonstrate the stability of the method across different batch sizes. Figure 2a shows a comparison of Q-Prop variants with trust-region updates on the HalfCheetah-v1 domain, along with the best performing TRPO hyperparameters. The results are consistent with theory: conservative Q-Prop achieves much more stable performance than the standard and aggressive variants, and all Q-Prop variants significantly outperform TRPO in terms of sample efficiency, e.g. conservative Q-Prop reaches average reward of 4000 using about 10 times less samples than TRPO. + +![](images/b8ef0cab31ceb57d2ca2f3840e44751e97cf886ce5ce81bc26d79da92a4d50c5.jpg) +(a) Standard Q-Prop vs adaptive variants. + +![](images/ac3e5492e7295b46b13f17d48d9490a7f429f569aab0d8bfb64532dc2013b0fc.jpg) +(b) Conservative Q-Prop vs TRPO across batch sizes. +Figure 2: Average return over episodes in HalfCheetah-v1 during learning, exploring adaptive QProp methods and different batch sizes. All variants of Q-Prop substantially outperform TRPO in terms of sample efficiency. TR-c-QP, conservative Q-Prop with trust-region update performs most stably across different batch sizes. + +Figure 2b shows the performance of conservative Q-Prop against TRPO across different batch sizes. Due to high variance in gradient estimates, TRPO typically requires very large batch sizes, e.g. 25000 steps or 25 episodes per update, to perform well. We show that our Q-Prop methods can learn even with just 1 episode per update, and achieves better sample efficiency with small batch sizes. This shows that Q-Prop significantly reduces the variance compared to the prior methods. + +As we discussed in Section 1, stability is a significant challenge with state-of-the-art deep RL methods, and is very important for being able to reliably use deep RL for real world tasks. In the rest of the experiments, we will use conservative Q-Prop as the main Q-Prop implementation. + +![](images/d75d5db85e34384abd1e96bb494b10087f6138234cb0f906d3e4c4dba89c33ff.jpg) + +(a) Comparing algorithms on HalfCheetah-v1. + +![](images/a609c411ab9a5449d74b3fe5b80daffd18f515d8845ee6448717ce07addc7bd5.jpg) +(b) Comparing algorithms on Humanoid-v1. +Figure 3: Average return over episodes in HalfCheetah-v1 and Humanoid-v1 during learning, comparing Q-Prop against other model-free algorithms. Q-Prop with vanilla policy gradient outperforms TRPO on HalfCheetah. Q-Prop significantly outperforms TRPO in convergence time on Humanoid. + +In this section, we evaluate two versions of conservative Q-Prop, v-c-Q-Prop using vanilla policy gradient and TR-c-Q-Prop using trust-region updates, against other model-free algorithms on the HalfCheetah-v1 domain. Figure 3a shows that c-Q-Prop methods significantly outperform the best TRPO and VPG methods. Even Q-Prop with vanilla policy gradient is comparable to TRPO, confirming the significant benefits from variance reduction. DDPG on the other hand exhibits inconsistent performances. With proper reward scaling, i.e. “DDPG-r0.1”, it outperforms other methods as well as the DDPG results reported in prior work (Duan et al., 2016; Amos et al., 2016). This illustrates the sensitivity of DDPG to hyperparameter settings, while Q-Prop exhibits more stable, monotonic learning behaviors when compared to DDPG. In the next section we show this improved stability allows Q-Prop to outperform DDPG in more complex domains. + +# 5.3 EVALUATION ACROSS DOMAINS + +Lastly, we evaluate Q-Prop against TRPO and DDPG across multiple domains. While the gym environments are biased toward locomotion, we expect we can achieve similar performance on manipulation tasks such as those in Lillicrap et al. (2016). Table 1 summarizes the results, including the best attained average rewards and the steps to convergence. Q-Prop consistently outperform TRPO in terms of sample complexity and sometimes achieves higher rewards than DDPG in more complex domains. A particularly notable case is shown in Figure 3b, where Q-Prop substantially improves sample efficiency over TRPO on Humanoid-v1 domain, while DDPG cannot find a good solution. + +The better performance on the more complex domains highlights the importance of stable deep RL algorithms: while costly hyperparameter sweeps may allow even less stable algorithms to perform well on simpler problems, more complex tasks might have such narrow regions of stable hyperparameters that discovering them becomes impractical. + +
TR-c-Q-PropTRPODDPG
DomainThresholdMaxReturn.EpisodesMaxReturnEpsisodesMaxReturnEpisodes
Ant350035344975423913825957N/A
HalfCheetah47004811207854734263707490600
Hopper200029575945248657152604965
Humanoid2500>349214750918>30000552N/A
Reacher-7-6.02060-6.72840-6.61800
Swimmer9010320451103025150500
Walker30004030368535671887536262125
+ +Table 1: Q-Prop, TRPO and DDPG results showing the max average rewards attained in the first $3 0 \mathrm { k }$ episodes and the episodes to cross specific reward thresholds. Q-Prop often learns more sample efficiently than TRPO and can solve difficult domains such as Humanoid better than DDPG. + +# 6 DISCUSSION AND CONCLUSION + +We presented Q-Prop, a policy gradient algorithm that combines reliable, consistent, and potentially unbiased on-policy gradient estimation with a sample-efficient off-policy critic that acts as a control variate. The method provides a large improvement in sample efficiency compared to stateof-the-art policy gradient methods such as TRPO, while outperforming state-of-the-art actor-critic methods on more challenging tasks such as humanoid locomotion. We hope that techniques like these, which combine on-policy Monte Carlo gradient estimation with sample-efficient variance reduction through off-policy critics, will eventually lead to deep reinforcement learning algorithms that are more stable and efficient, and therefore better suited for application to complex real-world learning tasks. + +# ACKNOWLEDGMENTS + +We thank Rocky Duan for sharing and answering questions about rllab code, and Yutian Chen and Laurent Dinh for discussion on control variates. SG and RT were funded by NSERC, Google, and EPSRC grants EP/L000776/1 and EP/M026957/1. ZG was funded by EPSRC grant EP/J012300/1 and the Alan Turing Institute (EP/N510129/1). + +# REFERENCES + +Brandon Amos, Lei Xu, and J Zico Kolter. Input convex neural networks. arXiv preprint arXiv:1609.07152, 2016. + +Christopher G Atkeson and Juan Carlos Santamaria. A comparison of direct and model-based reinforcement learning. In In International Conference on Robotics and Automation. Citeseer, 1997. + +Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016. + +Marc Deisenroth and Carl E Rasmussen. Pilco: A model-based and data-efficient approach to policy search. In Proceedings of the 28th International Conference on machine learning (ICML-11), pp. 465–472, 2011. + +Yan Duan, Xi Chen, Rein Houthooft, John Schulman, and Pieter Abbeel. Benchmarking deep reinforcement learning for continuous control. International Conference on Machine Learning (ICML), 2016. + +Miroslav Dud´ık, John Langford, and Lihong Li. Doubly robust policy evaluation and learning. arXiv preprint arXiv:1103.4601, 2011. + +Evan Greensmith, Peter L Bartlett, and Jonathan Baxter. Variance reduction techniques for gradient estimates in reinforcement learning. Journal of Machine Learning Research, 5(Nov):1471–1530, 2004. + +Shixiang Gu, Sergey Levine, Ilya Sutskever, and Andriy Mnih. Muprop: Unbiased backpropagation for stochastic neural networks. International Conference on Learning Representations (ICLR), 2016a. + +Shixiang Gu, Tim Lillicrap, Ilya Sutskever, and Sergey Levine. Continuous deep q-learning with model-based acceleration. In International Conference on Machine Learning (ICML), 2016b. + +Hado V Hasselt. Double q-learning. In Advances in Neural Information Processing Systems, pp. 2613–2621, 2010. + +Sham Kakade. A natural policy gradient. In NIPS, volume 14, pp. 1531–1538, 2001. + +Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. + +Guy Lever. Deterministic policy gradient algorithms. 2014. + +Sergey Levine and Vladlen Koltun. Guided policy search. In International Conference on Machine Learning (ICML), pp. 1–9, 2013. + +Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. International Conference on Learning Representations (ICLR), 2016. + +A Rupam Mahmood, Hado P van Hasselt, and Richard S Sutton. Weighted importance sampling for off-policy learning with linear function approximation. In Advances in Neural Information Processing Systems, pp. 3014–3022, 2014. + +Andriy Mnih and Karol Gregor. Neural variational inference and learning in belief networks. International Conference on Machine Learning (ICML), 2014. + +Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015. + +Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy P Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International Conference on Machine Learning (ICML), 2016. + +Remi Munos, Tom Stepleton, Anna Harutyunyan, and Marc G Bellemare. Safe and efficient off-´ policy reinforcement learning. arXiv preprint arXiv:1606.02647, 2016. + +John Paisley, David Blei, and Michael Jordan. Variational bayesian inference with stochastic search. International Conference on Machine Learning (ICML), 2012. + +Jan Peters and Stefan Schaal. Policy gradient methods for robotics. In International Conference on Intelligent Robots and Systems (IROS), pp. 2219–2225. IEEE, 2006. + +Jan Peters, Katharina Mulling, and Yasemin Altun. Relative entropy policy search. In ¨ AAAI. Atlanta, 2010. + +Doina Precup. Eligibility traces for off-policy policy evaluation. Computer Science Department Faculty Publication Series, pp. 80, 2000. + +Sheldon M Ross. Simulation. Burlington, MA: Elsevier, 2006. + +John Schulman, Sergey Levine, Pieter Abbeel, Michael I. Jordan, and Philipp Moritz. Trust region policy optimization. In International Conference on Machine Learning (ICML), pp. 1889–1897, 2015. + +John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. Highdimensional continuous control using generalized advantage estimation. International Conference on Learning Representations (ICLR), 2016. + +David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. Deterministic policy gradient algorithms. In International Conference on Machine Learning (ICML), 2014. + +David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016. + +Richard S Sutton. Integrated architectures for learning, planning, and reacting based on approximating dynamic programming. In International Conference on Machine Learning (ICML), pp. 216–224, 1990. + +Richard S Sutton, David A McAllester, Satinder P Singh, Yishay Mansour, et al. Policy gradient methods for reinforcement learning with function approximation. In Advances in Neural Information Processing Systems (NIPS), volume 99, pp. 1057–1063, 1999. + +Richard S Sutton, Hamid Reza Maei, Doina Precup, Shalabh Bhatnagar, David Silver, Csaba Szepesvari, and Eric Wiewiora. Fast gradient-descent methods for temporal-difference learning ´ with linear function approximation. In Proceedings of the 26th Annual International Conference on Machine Learning, pp. 993–1000. ACM, 2009. + +Richard S Sutton, A Rupam Mahmood, and Martha White. An emphatic approach to the problem of off-policy temporal-difference learning. The Journal of Machine Learning Research, 2015. + +Philip Thomas. Bias in natural actor-critic algorithms. In ICML, pp. 441–448, 2014. + +Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 5026–5033. IEEE, 2012. + +Christopher JCH Watkins and Peter Dayan. Q-learning. Machine learning, 8(3-4):279–292, 1992. + +Lex Weaver and Nigel Tao. The optimal reward baseline for gradient-based reinforcement learning. In Proceedings of the Seventeenth conference on Uncertainty in artificial intelligence, pp. 538– 545. Morgan Kaufmann Publishers Inc., 2001. + +Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3-4):229–256, 1992. + +# A Q-PROP ESTIMATOR DERIVATION + +The full derivation of the Q-Prop estimator is shown in Eq. 14. We make use of the following property that is commonly used in baseline derivations: + +$$ +\mathbb { E } _ { p _ { \theta } ( x ) } [ \nabla _ { \theta } \log p _ { \theta } ( x ) ] = \int _ { x } \nabla _ { \theta } p _ { \theta } ( x ) = \nabla _ { \theta } \int _ { x } p ( x ) = 0 +$$ + +This holds true when $f ( \mathbf { \boldsymbol { s } } _ { t } , \mathbf { \boldsymbol { a } } _ { t } )$ is an arbitrary function differentiable with respect to $\mathbf { } \mathbf { } \mathbf { } \mathbf { a } _ { t }$ and $\bar { f }$ is its first-order Taylor expansion around $\mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf \mathbf { } \mathbf { } \mathbf { } \mathbf \mathbf { } \mathbf { } \mathbf \Psi \mathbf { } \mathbf { } \mathbf \mathbf { } \mathbf { } \mathbf \Psi \mathbf { } \mathbf \mathbf { } \mathbf \Psi \mathbf { } \mathbf \mathbf { } \mathbf \Psi \mathbf { } \mathbf \mathbf { } \mathbf \mathbf \Psi \mathbf { } \mathbf \mathbf \Psi \mathbf { } \mathbf \mathbf \Psi \mathbf { } \mathbf \mathbf \Psi \mathbf \Psi \mathbf \Psi \mathbf \Psi \mathbf \Psi \mathbf \Psi \mathbf \Psi \Psi \mathbf \Psi \mathbf \Psi \mathbf \Psi \mathbf \Psi \mathbf \mathbf \Psi \mathbf \Psi \mathbf \mathbf \Psi \mathbf \Psi \mathbf \mathbf \Psi \mathbf \mathbf \mathbf \Psi \mathbf \mathbf \Psi \mathbf \mathbf \mathbf \mathbf \Psi \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \Psi \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf $ , i.e. $\bar { f } ( s _ { t } , a _ { t } ) = f ( s _ { t } , \bar { a } _ { t } ) + \nabla _ { a } \bar { f ( } s _ { t } , a ) | _ { a = \bar { a } _ { t } } ( a _ { t } - \bar { a } _ { t } ) .$ . Here, $\mu _ { \theta } ( \pmb { s } _ { t } ) \dot { = } \mathbb { E } _ { \pi } [ \bar { \mathbf { a } } _ { t } ]$ is the mean of stochastic policy $\pi _ { \theta }$ . The derivation appears below: + +$$ +\begin{array} { r l } { \nabla _ { \theta } J ( \theta ) = \mathrm { E } _ { \mathrm { g } , \pi } \big [ \nabla _ { \theta } \log \pi _ { \theta } \big ( \alpha _ { 1 } | s _ { 1 } \big ) \big ( \hat { Q } ( s _ { t } , a _ { t } ) - \hat { J } ( s _ { t } , a _ { t } ) \big ) \big ] + \mathrm { E } _ { \rho , \pi } \big [ \nabla _ { \theta } \log \pi _ { \theta } \big ( \alpha _ { t } | s _ { t } \big ) \hat { J } ( s _ { t } , a _ { t } ) \big ] } \\ { g ( \theta ) = \mathrm { E } _ { \mathrm { g } , \pi } \big [ \nabla _ { \theta } \log \pi _ { \theta } \big ( \alpha _ { 1 } | s _ { 1 } \big ) \hat { J } ( s _ { t } , a _ { t } ) \big ] } \\ & { = \mathrm { E } _ { \rho , \pi } \big [ \nabla _ { \theta } \log \pi _ { \theta } \big ( \alpha _ { 1 } | s _ { 1 } \big ) \big ( \hat { J } ( s _ { t } , a _ { t } ) + \nabla _ { a _ { t } } \big ) \big ( s _ { t } , a _ { t } \big ) \big ] } \\ & { = \mathrm { E } _ { \rho , \pi } \big [ \nabla _ { \theta } \log \pi _ { \theta } \big ( \alpha _ { 1 } | s _ { 1 } \big ) \mathrm { e } _ { n } f \big ( s _ { t } , a _ { t } \big ) \big ] = - \alpha _ { 1 } \alpha _ { 1 } \alpha _ { 1 } } \\ & { = \mathrm { E } _ { \rho , \pi } \bigg [ \int _ { a _ { t } } \mathrm { e } _ { g } \pi _ { \theta } \big ( a _ { t } | s _ { t } \big ) \mathrm { e } _ { n } f \big ( s _ { t } , a _ { t } \big ) \big ] \omega - \alpha _ { 2 } \alpha _ { 2 } \alpha _ { 2 } \bigg ] } \\ & { = \mathrm { E } _ { \rho , \pi } \bigg [ \sum _ { \theta } \mathrm { e } _ { g } \pi _ { \theta } \big ( a _ { t } \big ) \big ( \frac { \kappa } { \alpha _ { 1 } } \big ) \mathrm { e } _ { n } f \big ( \alpha _ { 1 } \big ) \omega _ { 1 } \omega _ { 2 } \bigg ] } \\ & = \mathrm { E } _ { \rho , \pi } \bigg [ \mathrm { v } _ { \alpha } f \big ( s _ { t } , a \big ) \big | \alpha - \alpha _ { 1 } \int _ a _ { t } \end{array} +$$ + +# B CONNECTION BETWEEN Q-PROP AND COMPATIBLE FEATURE APPROXIMATION + +In this section we show that actor-critic with compatible feature approximation is a form of control variate. A critic $Q _ { w }$ is compatible (Sutton et al., 1999) if it satisfies (1) $Q _ { w } ( \pmb { s } _ { t } , \pmb { a } _ { t } ) =$ $w ^ { T } \nabla _ { \theta } \log \pi _ { \theta } ( { \mathbf { \em a } } _ { t } | { \mathbf { \em s } } _ { t } )$ , i.e. $\nabla _ { w } \mathcal { Q } _ { w } ( s _ { t } , \bar { a } _ { t } ) = \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } | s _ { t } )$ , and (2) $w$ is fit with objective $w =$ a $\mathrm { r g } \operatorname* { m i n } _ { w } L ( w ) = \mathrm { a r g } \operatorname* { m i n } _ { w } \mathbb { E } _ { \rho _ { \pi } , \pi } [ ( \hat { Q } ( s _ { t } , a _ { t } ) - Q _ { w } ( { \pmb s } _ { t } , { \pmb a } _ { t } ) ) ^ { 2 } ]$ , that is fitting $Q _ { w }$ on on-policy Monte Carlo returns. Condition (2) implies the following identity, + +$$ +\begin{array} { r } { \nabla _ { w } L = 2 \mathbb { E } _ { \rho _ { \pi } , \pi } [ \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } | s _ { t } ) ( \hat { \mathcal { Q } } ( s _ { t } , a _ { t } ) - { \mathcal Q } _ { w } ( s _ { t } , a _ { t } ) ) ] = 0 . } \end{array} +$$ + +In compatible feature approximation, it directly uses $Q _ { w }$ as control variate, rather than its Taylor expansion $\bar { Q } _ { w }$ as in Q-Prop. Using Eq. 15, the Monte Carlo policy gradient is, + +$$ +\begin{array} { r l } & { \nabla _ { \theta } J ( \theta ) = \mathbb { E } _ { \rho _ { \pi } , \pi } [ \nabla _ { \theta } \log \pi _ { \theta } ( \boldsymbol { a } _ { t } | \boldsymbol { s } _ { t } ) Q _ { w } ( \boldsymbol { s } _ { t } , \boldsymbol { a } _ { t } ) ] } \\ & { \qquad = \mathbb { E } _ { \rho _ { \pi } , \pi } [ ( \nabla _ { \theta } \log \pi _ { \theta } ( \boldsymbol { a } _ { t } | \boldsymbol { s } _ { t } ) \nabla _ { \theta } \log \pi _ { \theta } ( \boldsymbol { a } _ { t } | \boldsymbol { s } _ { t } ) ^ { T } ) w ] } \\ & { \qquad = \mathbb { E } _ { \rho _ { \pi } } [ I ( \theta ; \boldsymbol { s } _ { t } ) w ] , } \end{array} +$$ + +where $I ( \theta ; s _ { t } ) = \mathbb { E } _ { \pi _ { \theta } } [ \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } \vert s _ { t } ) \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } \vert s _ { t } ) ^ { T } ]$ is Fisher’s information matrix. Thus, variance reduction depends on ability to compute or estimate $I ( \theta ; s _ { t } )$ and $w$ effectively. + +# C UNIFYING POLICY GRADIENT AND ACTOR-CRITIC + +Q-Prop closely ties together policy gradient and actor-critic algorithms. To analyze this point, we write a generalization of Eq. 9 below, introducing two additional variables $\alpha , \rho _ { C R }$ : + +$$ +\begin{array} { r l } & { \nabla _ { \theta } J ( \theta ) \propto \alpha \mathbb { E } _ { \rho _ { \pi } , \pi } [ \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } | s _ { t } ) ( \hat { \boldsymbol A } ( s _ { t } , a _ { t } ) - \eta \bar { \boldsymbol A } _ { w } ( s _ { t } , a _ { t } ) ] } \\ & { \qquad + \eta \mathbb { E } _ { \rho _ { C R } } [ \nabla _ { a } \mathcal { Q } _ { w } ( s _ { t } , a ) | _ { a = \mu _ { \theta } ( s _ { t } ) } \nabla _ { \theta } \mu _ { \theta } ( s _ { t } ) ] } \end{array} +$$ + +Eq. 17 enables more analysis where bias generally is introduced only when $\alpha \neq 1$ or $\rho _ { C R } \neq \rho _ { \pi }$ . Importantly, Eq. 17 covers both policy gradient and deterministic actor-critic algorithm as its special cases. Standard policy gradient is recovered by $\eta = 0$ , and deterministic actor-critic is recovered by $\alpha = 0$ and $\rho _ { C R } = \rho _ { \beta }$ . This allows heuristic or automatic methods for dynamically changing these variables through the learning process for optimizing different metrics, e.g. sample efficiency, convergence speed, stability. + +Table 2 summarizes the various edge cases of Eq. 17. For example, since we derive our method from a control variates standpoint, $Q _ { w }$ can be any function and the gradient remains almost unbiased (see + +Table 2: Implementation options and edge cases of the generalized Q-Prop estimator in Eq. 17. + +
ParameterImplementation optionsIntroduce bias?
Qwoff-policy TD; on-policy TD(λ); model-based; etc.No
Von-policy Monte Carlo fitting; Eπθ [Qw(St, at)}; etcNo
0≤λ≤1Yes,except=1
αα≥0Yes, except α=1
nany nNo
PCRp of any policyYes, except pcR=pπ
+ +Section 2.1). A natural choice is to use off-policy temporal difference learning to learn the critic $Q _ { w }$ corresponding to policy $\pi$ . This enables effectively utilizing off-policy samples without introducing further bias. An interesting alternative to this is to utilize model-based roll-outs to estimate the critic, which resembles MuProp in stochastic neural networks (Gu et al., 2016a). Unlike prior work on using fitted dynamics model to accelerate model-free learning (Gu et al., 2016b), this approach does not introduce bias to the gradient of the original objective. + +# D EXPERIMENT DETAILS + +Policy and value function architectures. The network architectures are largely based on the benchmark paper by Duan et al. (2016). For policy gradient methods, the stochastic policy $\pi _ { \boldsymbol { \theta } } ( \mathbf { a } _ { t } | \mathbf { s } _ { t } ) = \bar { \mathcal { N } } ( \mu _ { \boldsymbol { \theta } } ( \mathbf { \bar { s } } _ { t } ) , \Sigma _ { \boldsymbol { \theta } } )$ is a local Gaussian policy with a local state-dependent mean and a global covariance matrix. $\pmb { \mu } _ { \pmb { \theta } } ( \pmb { s } _ { t } )$ is a neural network with 3 hidden layers of sizes 100-50-25 and tanh nonlinearities at the first 2 layers, and $\Sigma _ { \theta }$ is diagonal. For DDPG, the policy is deterministic and has the same architecture as $\mu _ { \theta }$ except that it has an additional tanh layer at the output. $V _ { \phi } ( \pmb { s } _ { t } )$ for baselines and GAE is fit with the same technique by Schulman et al. (2016), a variant of linear regression on Monte Carlo returns with soft-update constraint. For Q-Prop and DDPG, $Q _ { w } ( s , \pmb { a } )$ is parametrized with a neural network with 2 hidden layers of size 100 and ReLU nonlinearity, where $^ { a }$ is included after the first hidden layer. + +Training details. This section describes parameters of the training algorithms and their hyperparameter search values in $\{ \}$ . The optimal performing hyperparameter results are reported. Policy gradient methods (VPG, TRPO, Q-Prop) used batch sizes of $\{ 1 0 0 0 , 5 0 0 0 , 2 5 0 0 0 \}$ time steps, step sizes of $\{ 0 . 1 , 0 . 0 1 , 0 . 0 0 1 \}$ for the trust-region method, and base learning rates of $\{ 0 . 0 0 1 , \bar { 0 } . 0 0 0 1 \}$ with Adam (Kingma & Ba, 2014) for vanilla policy gradient methods. For Q-Prop and DDPG, $Q _ { w }$ is learned with the same technique as in DDPG (Lillicrap et al., 2016), using soft target networks with $\tau = 0 . 9 9 9$ , a replay buffer of size $1 0 ^ { 6 }$ steps, a mini-batch size of 64, and a base learning rate of $\{ 0 . 0 0 1 , 0 . 0 0 0 1 \}$ with Adam (Kingma & Ba, 2014). For Q-Prop we also tuned the relative ratio of gradient steps on the critic $Q _ { w }$ against the number of steps on the policy, in the range $\{ 0 . 1 , 0 . 5$ , $1 . 0 \}$ , where 0.1 corresponds to 100 critic updates for every policy update if the batch size is 1000. For DDPG, we swept the reward scaling using $\{ 0 . 0 1 , 0 . 1 , \dot { 1 . } 0 \dot { \} }$ as it is sensitive to this parameter. \ No newline at end of file diff --git a/parse/train/SJ3rcZcxl/SJ3rcZcxl_content_list.json b/parse/train/SJ3rcZcxl/SJ3rcZcxl_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..01017c1873b5623b4fa936ba370703ec4d064c8b --- /dev/null +++ b/parse/train/SJ3rcZcxl/SJ3rcZcxl_content_list.json @@ -0,0 +1,1739 @@ +[ + { + "type": "text", + "text": "Q-PROP: SAMPLE-EFFICIENT POLICY GRADIENT WITH AN OFF-POLICY CRITIC ", + "text_level": 1, + "bbox": [ + 176, + 98, + 823, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shixiang $\\mathbf { G u } ^ { 1 2 3 }$ , Timothy Lillicrap4, Zoubin Ghahramani16, Richard E. Turner1, Sergey Levine35 ", + "bbox": [ + 184, + 169, + 852, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sg717@cam.ac.uk,countzero@google.com,zoubin@eng.cam.ac.uk, \nret26@cam.ac.uk,svlevine@eecs.berkeley.edu \n1University of Cambridge, UK \n2Max Planck Institute for Intelligent Systems, Tubingen, Germany ¨ \n3Google Brain, USA \n4DeepMind, UK \n5UC Berkeley, USA \n6Uber AI Labs, USA ", + "bbox": [ + 183, + 185, + 746, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 339, + 544, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Model-free deep reinforcement learning (RL) methods have been successful in a wide variety of simulated domains. However, a major obstacle facing deep RL in the real world is their high sample complexity. Batch policy gradient methods offer stable learning, but at the cost of high variance, which often requires large batches. TD-style methods, such as off-policy actor-critic and Q-learning, are more sample-efficient but biased, and often require costly hyperparameter sweeps to stabilize. In this work, we aim to develop methods that combine the stability of policy gradients with the efficiency of off-policy RL. We present Q-Prop, a policy gradient method that uses a Taylor expansion of the off-policy critic as a control variate. Q-Prop is both sample efficient and stable, and effectively combines the benefits of on-policy and off-policy methods. We analyze the connection between Q-Prop and existing model-free algorithms, and use control variate theory to derive two variants of Q-Prop with conservative and aggressive adaptation. We show that conservative Q-Prop provides substantial gains in sample efficiency over trust region policy optimization (TRPO) with generalized advantage estimation (GAE), and improves stability over deep deterministic policy gradient (DDPG), the stateof-the-art on-policy and off-policy methods, on OpenAI Gym’s MuJoCo continuous control environments. ", + "bbox": [ + 233, + 371, + 764, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 648, + 336, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Model-free reinforcement learning is a promising approach for solving arbitrary goal-directed sequential decision-making problems with only high-level reward signals and no supervision. It has recently been extended to utilize large neural network policies and value functions, and has been shown to be successful in solving a range of difficult problems (Mnih et al., 2015; Schulman et al., 2015; Lillicrap et al., 2016; Silver et al., 2016; Gu et al., 2016b; Mnih et al., 2016). Deep neural network parametrization minimizes the need for manual feature and policy engineering, and allows learning end-to-end policies mapping from high-dimensional inputs, such as images, directly to actions. However, such expressive parametrization also introduces a number of practical problems. Deep reinforcement learning algorithms tend to be sensitive to hyperparameter settings, often requiring extensive hyperparameter sweeps to find good values. Poor hyperparameter settings tend to produce unstable or non-convergent learning. Deep RL algorithms also tend to exhibit high sample complexity, often to the point of being impractical to run on real physical systems. Although a number of recent techniques have sought to alleviate some of these issues (Hasselt, 2010; Mnih et al., 2015; Schulman et al., 2015; 2016), these recent advances still provide only a partial solution to the instability and sample complexity challenges. ", + "bbox": [ + 174, + 681, + 825, + 888 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Model-free reinforcement learning consists of on- and off-policy methods. Monte Carlo policy gradient methods (Peters & Schaal, 2006; Schulman et al., 2015) are popular on-policy methods that directly maximize the cumulative future returns with respect to the policy. While these algorithms can offer unbiased (or nearly unbiased, as discussed in Section 2.1) estimates of the gradient, they rely on Monte Carlo estimation and often suffer from high variance. To cope with high variance gradient estimates and difficult optimization landscapes, a number of techniques have been proposed, including constraining the change in the policy at each gradient step (Kakade, 2001; Peters et al., 2010) and mixing value-based back-ups to trade off bias and variance in Monte Carlo return estimates (Schulman et al., 2015). However, these methods all tend to require very large numbers of samples to deal with the high variance when estimating gradients of high-dimensional neural network policies. The crux of the problem with policy gradient methods is that they can only effectively use on-policy samples, which means that they require collecting large amounts of on-policy experiences after each parameter update to the policy. This makes them very sample intensive. Offpolicy methods, such as Q-learning (Watkins & Dayan, 1992; Sutton et al., 1999; Mnih et al., 2015; Gu et al., 2016b) and off-policy actor-critic methods (Lever, 2014; Lillicrap et al., 2016), can instead use all samples, including off-policy samples, by adopting temporal difference learning with experience replay. Such methods are much more sample-efficient. However, convergence of these algorithms is in general not guaranteed with non-linear function approximators, and practical convergence and instability issues typically mean that extensive hyperparameter tuning is required to attain good results. ", + "bbox": [ + 174, + 895, + 820, + 924 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 104, + 825, + 353 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to make deep reinforcement learning practical as a tool for tackling real-world tasks, we must develop methods that are both data efficient and stable. In this paper, we propose Q-Prop, a step in this direction that combines the advantages of on-policy policy gradient methods with the efficiency of off-policy learning. Unlike prior approaches for off-policy learning, which either introduce bias (Sutton et al., 1999; Silver et al., 2014) or increase variance (Precup, 2000; Levine & Koltun, 2013; Munos et al., 2016), Q-Prop can reduce the variance of gradient estimator without adding bias; unlike prior approaches for critic-based variance reduction (Schulman et al., 2016) which fit the value function on-policy, Q-Prop learns the action-value function off-policy. The core idea is to use the first-order Taylor expansion of the critic as a control variate, resulting in an analytical gradient term through the critic and a Monte Carlo policy gradient term consisting of the residuals in advantage approximations. The method helps unify policy gradient and actor-critic methods: it can be seen as using the off-policy critic to reduce variance in policy gradient or using on-policy Monte Carlo returns to correct for bias in the critic gradient. We further provide theoretical analysis of the control variate, and derive two additional variants of Q-Prop. The method can be easily incorporated into any policy gradient algorithm. We show that Q-Prop provides substantial gains in sample efficiency over trust region policy optimization (TRPO) with generalized advantage estimation (GAE) (Schulman et al., 2015; 2016), and improved stability over deep deterministic policy gradient (DDPG) (Lillicrap et al., 2016) across a repertoire of continuous control tasks. ", + "bbox": [ + 174, + 359, + 825, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 BACKGROUND ", + "text_level": 1, + "bbox": [ + 176, + 632, + 326, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reinforcement learning (RL) aims to learn a policy for an agent such that it behaves optimally according to a reward function. At a time step $t$ and state $\\scriptstyle { s _ { t } }$ , the agent chooses an action $\\mathbf { } \\mathbf { a } _ { t }$ according to its policy $\\pi ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } )$ , the state of the agent and the environment changes to new state $\\boldsymbol { s } _ { t + 1 }$ according to dynamics $p \\big ( \\pmb { s } _ { t + 1 } \\big | \\pmb { s } _ { t } , \\pmb { a } _ { t } \\big )$ , the agent receives a reward $r ( s _ { t } , a _ { t } )$ , and the process continues. Let $R _ { t }$ denote a $\\gamma$ -discounted cumulative return from $t$ for an infinite horizon problem, i.e $\\begin{array} { r } { R _ { t } = \\sum _ { t ^ { \\prime } = t } ^ { \\infty } \\gamma ^ { t ^ { \\prime } - t } r ( s _ { t ^ { \\prime } } , { \\bf a } _ { t ^ { \\prime } } ) } \\end{array}$ . The goal of reinforcement learning is to maximize the expected return $J ( \\theta ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ R _ { 0 } ]$ with respect to the policy parameters θ . In this section, we review several standard techniques for performing this optimization, and in the next section, we will discuss our proposed Q-Prop algorithm that combines the strengths of these approaches to achieve efficient, stable RL. Monte Carlo policy gradient refers to policy gradient methods that use full Monte Carlo returns, e.g. REINFORCE (Williams, 1992) and TRPO (Schulman et al., 2015), and policy gradient with function approximation refers to actor-critic methods (Sutton et al., 1999) which optimize the policy against a critic, e.g. deterministic policy gradient (Silver et al., 2014; Lillicrap et al., 2016). ", + "bbox": [ + 173, + 665, + 825, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 MONTE CARLO POLICY GRADIENT METHODS ", + "text_level": 1, + "bbox": [ + 174, + 868, + 535, + 882 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Monte Carlo policy gradient methods apply direct gradient-based optimization to the reinforcement learning objective. This involves directly differentiating the $J ( \\theta )$ objective with respect to the policy ", + "bbox": [ + 176, + 895, + 823, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "parameters $\\theta$ . The standard form, known as the REINFORCE algorithm (Williams, 1992), is shown below: ", + "bbox": [ + 171, + 102, + 825, + 132 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/f6942cdf9ba69600fd352e5a9c3209b8ebb10c765cf4d7646f59830914b86a8c.jpg", + "text": "$$\n\\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\pi } [ \\sum _ { t = 0 } ^ { \\infty } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\gamma ^ { t } R _ { t } ] = \\mathbb { E } _ { \\pi } [ \\sum _ { t = 0 } ^ { \\infty } \\gamma ^ { t } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( R _ { t } - b ( s _ { t } ) ) ] ,\n$$", + "text_format": "latex", + "bbox": [ + 232, + 132, + 764, + 170 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $b ( s _ { t } )$ is known as the baseline. For convenience of later derivations, Eq. 1 can also be written as below, where $\\begin{array} { r } { \\rho _ { \\pi } ( s ) = \\sum _ { t = 0 } ^ { \\infty } \\gamma ^ { t } p ( s _ { t } = s ) } \\end{array}$ is the unnormalized discounted state visitation frequency, ", + "bbox": [ + 173, + 172, + 828, + 203 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/192f775e82ad0d09644c328424abf64f7014ba3648e4e893604deaff23956ce3.jpg", + "text": "$$\n\\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\pi } ( . ) , \\mathbf { a } _ { t } \\sim \\pi ( \\cdot | s _ { t } ) } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( R _ { t } - b ( s _ { t } ) ) ] .\n$$", + "text_format": "latex", + "bbox": [ + 302, + 219, + 694, + 238 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Eq. 2 is an unbiased gradient of the RL objective. However, in practice, most policy gradient methods effectively use undiscounted state visitation frequencies, i.e. $\\gamma = 1$ in the equal for $\\rho _ { \\pi }$ , and are therefore biased; in fact, making them unbiased often hurts performance (Thomas, 2014). In this paper, we mainly discuss bias due to function approximation, off-policy learning, and value back-ups. ", + "bbox": [ + 173, + 242, + 825, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The gradient is estimated using Monte Carlo samples in practice and has very high variance. A proper choice of baseline is necessary to reduce the variance sufficiently such that learning becomes feasible. A common choice is to estimate the value function of the state $V _ { \\pi } ( \\pmb { s } _ { t } )$ to use as the baseline, which provides an estimate of advantage function $A _ { \\pi } ( s _ { t } , { \\pmb a } _ { t } )$ , which is a centered action-value function $Q _ { \\pi } ( s _ { t } , \\pmb { a } _ { t } )$ , as defined below: ", + "bbox": [ + 176, + 319, + 825, + 390 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/9915c7f590b427849f27b3abc7fecc5232c681e28a5cfc2aab1cb8d4818fdf0b.jpg", + "text": "$$\n\\begin{array} { c } { { V _ { \\pi } ( s _ { t } ) = \\mathbb { E } _ { \\pi } [ R _ { t } ] = \\mathbb { E } _ { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } [ Q _ { \\pi } ( s _ { t } , a _ { t } ) ] } } \\\\ { { Q _ { \\pi } ( s _ { t } , a _ { t } ) = r ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { E } _ { \\pi } [ R _ { t + 1 } ] = r ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { E } _ { p ( s _ { t + 1 } | s _ { t } , a _ { t } ) } [ V _ { \\pi } ( s _ { t + 1 } ) ] } } \\\\ { { A _ { \\pi } ( s _ { t } , a _ { t } ) = Q _ { \\pi } ( s _ { t } , a _ { t } ) - V _ { \\pi } ( s _ { t } ) . } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 259, + 392, + 738, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$Q _ { \\pi } ( s _ { t } , \\pmb { a } _ { t } )$ summarizes the performance of each action from a given state, assuming it follows $\\pi$ thereafter, and $A _ { \\pi } ( s _ { t } , { \\pmb a } _ { t } )$ provides a measure of how each action compares to the average performance at the state $\\scriptstyle { s _ { t } }$ , which is given by $V _ { \\pi } ( \\pmb { s } _ { t } )$ . Using $A _ { \\pi } ( s _ { t } , { \\pmb a } _ { t } )$ centers the learning signal and reduces variance significantly. ", + "bbox": [ + 178, + 454, + 821, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Besides high variance, another problem with the policy gradient is that it requires on-policy samples. This makes policy gradient optimization very sample intensive. To achieve similar sample efficiency as off-policy methods, we can attempt to include off-policy data. Prior attempts use importance sampling to include off-policy trajectories; however, these are known to be difficult scale to highdimensional action spaces because of rapidly degenerating importance weights (Precup, 2000). ", + "bbox": [ + 173, + 516, + 825, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 POLICY GRADIENT WITH FUNCTION APPROXIMATION ", + "text_level": 1, + "bbox": [ + 174, + 603, + 589, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Policy gradient methods with function approximation (Sutton et al., 1999), or actor-critic methods, include a policy evaluation step, which often uses temporal difference (TD) learning to fit a critic $Q _ { w }$ for the current policy $\\pi ( \\theta )$ , and a policy improvement step which greedily optimizes the policy $\\pi$ against the critic estimate $Q _ { w }$ . Significant gains in sample efficiency may be achievable using offpolicy TD learning for the critic, as in Q-learning and deterministic policy gradient (Sutton, 1990; Silver et al., 2014), typically by means of experience replay for training deep Q networks (Mnih et al., 2015; Lillicrap et al., 2016; Gu et al., 2016b). ", + "bbox": [ + 173, + 628, + 825, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One particularly relevant example of such a method is the deep deterministic policy gradient (DDPG) (Silver et al., 2014; Lillicrap et al., 2016). The updates for this method are given below, where $\\pi _ { \\boldsymbol { \\theta } } ( \\mathbf { a } _ { t } | \\mathbf { s } _ { t } ) = \\delta ( \\mathbf { a } _ { t } = \\mu _ { \\boldsymbol { \\theta } } ( \\mathbf { s } _ { t } ) )$ is a deterministic policy, $\\beta$ is arbitrary exploration distribution, and $\\rho _ { \\beta }$ corresponds to sampling from a replay buffer. $Q ( \\cdot , \\cdot )$ is the target network that slowly tracks $Q _ { w }$ (Lillicrap et al., 2016). ", + "bbox": [ + 174, + 733, + 825, + 804 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/a249afcd4ae52b44b8cd04ef8cc47074157d56f3ad0e07af8bf36242d14bdcf1.jpg", + "text": "$$\n\\begin{array} { r l } & { w = \\underset { w } { \\arg \\operatorname* { m i n } } \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\beta } ( \\cdot ) , a _ { t } \\sim \\beta ( \\cdot | s _ { t } ) } \\big [ \\big ( r \\big ( s _ { t } , a _ { t } \\big ) + \\gamma Q ( s _ { t + 1 } , \\mu _ { \\theta } ( s _ { t + 1 } ) \\big ) - Q _ { w } ( s _ { t } , a _ { t } ) \\big ) ^ { 2 } \\big ] } \\\\ & { \\theta = \\underset { \\theta } { \\arg \\operatorname* { m a x } } \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\beta } ( \\cdot ) } [ Q _ { w } \\big ( s _ { t } , \\mu _ { \\theta } ( s _ { t } ) \\big ) ] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 240, + 806, + 758, + 857 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When the critic and policy are parametrized with neural networks, full optimization is expensive, and instead stochastic gradient optimization is used. The gradient in the policy improvement phase is given below, which is generally a biased gradient of $J ( \\theta )$ . ", + "bbox": [ + 176, + 858, + 821, + 901 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/3559dbdfcd748188a430af44d6b28d871d9806f9929c6b1b627837aaa3042cb4.jpg", + "text": "$$\n\\nabla _ { \\theta } J ( \\theta ) \\approx \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\beta } ( \\cdot ) } [ \\nabla _ { a } Q _ { w } \\big ( s _ { t } , a \\big ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } \\big ( s _ { t } \\big ) ]\n$$", + "text_format": "latex", + "bbox": [ + 323, + 905, + 674, + 925 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The crucial benefits of DDPG are that it does not rely on high variance REINFORCE gradients and is trainable on off-policy data. These properties make DDPG and other analogous off-policy methods significantly more sample-efficient than policy gradient methods (Lillicrap et al., 2016; Gu et al., 2016b; Duan et al., 2016). However, the use of a biased policy gradient estimator makes analyzing its convergence and stability properties difficult. ", + "bbox": [ + 174, + 103, + 825, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Q-PROP ", + "text_level": 1, + "bbox": [ + 176, + 194, + 272, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we derive the Q-Prop estimator for policy gradient. The key idea from this estimator comes from observing Equations 2 and 5 and noting that the former provides an almost unbiased (see Section 2.1), but high variance gradient, while the latter provides a deterministic, but biased gradient. By using the deterministic biased estimator as a particular form of control variate (Ross, 2006; Paisley et al., 2012) for the Monte Carlo policy gradient estimator, we can effectively use both types of gradient information to construct a new estimator that in practice exhibits improved sample efficiency through the inclusion of off-policy samples while preserving the stability of on-policy Monte Carlo policy gradient. ", + "bbox": [ + 173, + 227, + 825, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Q-PROP ESTIMATOR ", + "text_level": 1, + "bbox": [ + 176, + 356, + 356, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To derive the Q-Prop gradient estimator, we start by using the first-order Taylor expansion of an arbitrary function $\\hat { f ( s _ { t } , a _ { t } ) , ~ } \\bar { f } ( s _ { t } , a _ { t } ) = f ( s _ { t } , \\bar { a } _ { t } ) + \\bar { \\nabla } _ { a } f ( \\bar { s _ { t } } , a ) | _ { a = \\bar { a } _ { t } } ( a _ { t } - \\bar { a } _ { t } ) \\cdot$ as the control variate for the policy gradient estimator. We use $\\begin{array} { r } { \\hat { \\boldsymbol { \\mathcal Q } } ( \\boldsymbol s _ { t } , \\boldsymbol a _ { t } ) = \\sum _ { t ^ { \\prime } = t } ^ { \\infty } \\gamma ^ { t ^ { \\prime } - t } \\boldsymbol r ( \\boldsymbol s _ { t ^ { \\prime } } , \\boldsymbol a _ { t ^ { \\prime } } ) } \\end{array}$ to denote Monte Carlo return from state $\\scriptstyle { \\pmb { s } } _ { t }$ and action $\\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { a } _ { t }$ , i.e. $\\mathbb { E } _ { \\pi } [ \\hat { \\cal Q } ( s _ { t } , a _ { t } ) ] = r ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { E } _ { p } [ V _ { \\pi } ( s _ { t + 1 } ) ]$ , and $\\mu _ { \\theta } ( s _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } [ a _ { t } ]$ to denote the expected action of a stochastic policy $\\pi _ { \\theta }$ . Full derivation is in Appendix A. ", + "bbox": [ + 173, + 382, + 825, + 473 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/a9c36bb0c9ec213718f863f475c66da7c63792197b548ccb4fd90422835cad96.jpg", + "text": "$$\n\\begin{array} { r } { \\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { Q } ( s _ { t } , a _ { t } ) - \\bar { f } ( s _ { t } , a _ { t } ) ] + \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\bar { f } ( s _ { t } , a _ { t } ) ] } \\\\ { = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { Q } ( s _ { t } , a _ { t } ) - \\bar { f } ( s _ { t } , a _ { t } ) ] + \\mathbb { E } _ { \\rho _ { \\pi } } [ \\nabla _ { a } f ( s _ { t } , a ) | _ { a = \\bar { a } _ { t } } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 191, + 478, + 787, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eq. 6 is general for arbitrary function $f ( \\mathbf { \\boldsymbol { s } } _ { t } , \\mathbf { \\boldsymbol { a } } _ { t } )$ that is differentiable with respect to $\\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { a } _ { t }$ at an arbitrary value of $\\bar { \\mathbf { a } } _ { t }$ ; however, a sensible choice is to use the critic $Q _ { w }$ for $f$ and $\\pmb { \\mu } _ { \\pmb { \\theta } } ( \\pmb { s } _ { t } )$ for $\\bar { \\mathbf { a } } _ { t }$ to get, ", + "bbox": [ + 169, + 527, + 821, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/7ab4d9990df2732f4c0a7e5ab4367780adb6f5170139fdf3ae2d6112d988441c.jpg", + "text": "$$\n\\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { Q } ( s _ { t } , a _ { t } ) - \\bar { Q } _ { w } ( s _ { t } , a _ { t } ) ] + \\mathbb { E } _ { \\rho _ { \\pi } } [ \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] .\n$$", + "text_format": "latex", + "bbox": [ + 181, + 564, + 820, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, since in practice we estimate advantages $\\hat { A } ( s _ { t } , a _ { t } )$ , we write the Q-Prop estimator in terms of advantages to complete the basic derivation, ", + "bbox": [ + 174, + 608, + 821, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/7f30e2ad2a010faa58514730d9b6d8b8375efa01cd07bbebc391e6fd7d9a53fe.jpg", + "text": "$$\n\\begin{array} { r l } & { \\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { A } ( s _ { t } , a _ { t } ) - \\bar { A } _ { w } ( s _ { t } , a _ { t } ) ] + \\mathbb { E } _ { \\rho _ { \\pi } } [ \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] } \\\\ & { \\bar { A } ( s _ { t } , a _ { t } ) = \\bar { Q } ( s _ { t } , a _ { t } ) - \\mathbb { E } _ { \\pi _ { \\theta } } [ \\bar { Q } ( s _ { t } , a _ { t } ) ] = \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } ( a _ { t } - \\mu _ { \\theta } ( s _ { t } ) ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 178, + 643, + 815, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eq. 8 is composed of an analytic gradient through the critic as in Eq. 5 and a residual REINFORCE gradient in Eq. 2. From the above derivation, Q-Prop is simply a Monte Carlo policy gradient estimator with a special form of control variate. The important insight comes from the fact that $Q _ { w }$ can be trained using off-policy data as in Eq. 4. Under this setting, Q-Prop is no longer just a Monte Carlo policy gradient method, but more closely resembles an actor-critic method, where the critic can be updated off-policy but the actor is always updated on-policy with an additional REINFORCE correction term so that it remains a Monte Carlo policy gradient method regardless of the parametrization, training method, and performance of the critic. Therefore, Q-Prop can be directly combined with a number of prior techniques from both on-policy methods such as natural policy gradient (Kakade, 2001), trust-region policy optimization (TRPO) (Schulman et al., 2015) and generalized advantage estimation (GAE) (Schulman et al., 2016), and off-policy methods such as DDPG (Lillicrap et al., 2016) and Retrace $( \\lambda )$ (Munos et al., 2016). ", + "bbox": [ + 173, + 707, + 825, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, if the critic $Q _ { w }$ approximates $Q _ { \\pi }$ well, it provides a reliable gradient, reduces the estimator variance, and improves the convergence rate. Interestingly, control variate analysis in the next section shows that this is not the only circumstance where Q-Prop helps reduce variance. ", + "bbox": [ + 176, + 881, + 823, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 CONTROL VARIATE ANALYSIS AND ADAPTIVE Q-PROP ", + "text_level": 1, + "bbox": [ + 173, + 103, + 598, + 118 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For Q-Prop to be applied reliably, it is crucial to analyze how the variance of the estimator changes before and after the application of control variate. Following the prior work on control variates (Ross, 2006; Paisley et al., 2012), we first introduce $\\eta ( s _ { t } )$ to Eq. 8, a weighing variable that modulates the strength of control variate. This additional variable $\\eta ( s _ { t } )$ does not introduce bias to the estimator. ", + "bbox": [ + 173, + 128, + 825, + 199 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/8a29a313a26cb57497794ec002c4dd2d01c95f4e6b0f5774d17c0c9841dc44c4.jpg", + "text": "$$\n\\begin{array} { r l } & { \\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { A } ( s _ { t } , a _ { t } ) - \\eta ( s _ { t } ) \\bar { A } _ { w } ( s _ { t } , a _ { t } ) ] } \\\\ & { \\qquad + \\mathbb { E } _ { \\rho _ { \\pi } } [ \\eta ( s _ { t } ) \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 294, + 200, + 704, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The variance of this estimator is given below, where $m = 1 . . . M$ indexes the dimension of $\\theta$ , ", + "bbox": [ + 173, + 242, + 774, + 257 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/2c67b2666ae2f9272d5fd42a6de8a38b1c8d6b3514f7cf914c808530e04baac1.jpg", + "text": "$$\n\\mathrm { V a r } ^ { * } = \\mathbb { E } _ { \\rho _ { \\pi } } \\left[ \\sum _ { m } \\mathrm { V a r } _ { a _ { t } } ( \\nabla _ { \\theta _ { m } } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { A } ( s _ { t } , a _ { t } ) - \\eta ( s _ { t } ) \\bar { A } ( s _ { t } , a _ { t } ) ) ) \\right] .\n$$", + "text_format": "latex", + "bbox": [ + 264, + 260, + 735, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If we choose $\\eta ( s _ { t } )$ such that $\\mathrm { V a r ^ { * } } < \\mathrm { V a r }$ , where $\\begin{array} { r } { \\mathrm { V a r } = \\mathbb E _ { \\rho _ { \\pi } } [ \\sum _ { m } \\mathrm { V a r } _ { \\boldsymbol { a } _ { t } } ( \\nabla _ { \\boldsymbol { \\theta } _ { m } } \\log \\pi _ { \\boldsymbol { \\theta } } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ) \\hat { \\boldsymbol { A } } ( \\boldsymbol { s } _ { t } , \\boldsymbol { a } _ { t } ) ) ] } \\end{array}$ is the original estimator variance measure, then we have managed to reduce the variance. Directly analyzing the above variance measure is nontrivial, for the same reason that computing the optimal baseline is difficult (Weaver & Tao, 2001). In addition, it is often impractical to get multiple action samples from the same state, which prohibits using na¨ıve Monte Carlo to estimate the expectations. Instead, we propose a surrogate variance measure, $\\operatorname { V a r } = \\mathbb { E } _ { \\rho _ { \\pi } } [ \\operatorname { V a r } _ { \\mathbf { a } _ { t } } ( \\hat { A } ( s _ { t } , \\mathbf { a } _ { t } ) ) ]$ . A similar surrogate is also used by prior work on learning state-dependent baseline (Mnih & Gregor, 2014), and the benefit is that the measure becomes more tractable, ", + "bbox": [ + 173, + 301, + 825, + 415 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/be0f145346ff59e5f6dca67807e697ffebc0c2629f33f49e7b52beddc1e53d09.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathrm { V a r } ^ { * } = \\mathbb { E } _ { \\rho _ { \\pi } } [ \\mathrm { V a r } _ { a _ { t } } ( \\hat { A } ( s _ { t } , a _ { t } ) - \\eta ( s _ { t } ) \\bar { A } ( s _ { t } , a _ { t } ) ) ] } \\\\ & { \\qquad = \\mathrm { V a r } + \\mathbb { E } _ { \\rho _ { \\pi } } [ - 2 \\eta ( s _ { t } ) \\mathrm { C o v } _ { a _ { t } } ( \\hat { A } ( s _ { t } , a _ { t } ) , \\bar { A } ( s _ { t } , a _ { t } ) ) + \\eta ( s _ { t } ) ^ { 2 } \\mathrm { V a r } _ { a _ { t } } ( \\bar { A } ( s _ { t } , a _ { t } ) ) ] . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 232, + 417, + 764, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since $\\mathbb { E } _ { \\pi } [ \\hat { A } ( \\pmb { s } _ { t } , \\pmb { a } _ { t } ) ] = \\mathbb { E } _ { \\pi } [ \\bar { A } ( \\pmb { s } _ { t } , \\pmb { a } _ { t } ) ] = 0$ , the terms can be simplified as below, ", + "bbox": [ + 174, + 463, + 684, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/4221ff48c12b8d6e96304ffd21be86c42e62fe5f72fe9f6186c63a587bee192e.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathrm { C o v } _ { a _ { t } } ( \\hat { A } , \\bar { A } ) = \\mathbb { E } _ { \\pi } [ \\hat { A } ( s _ { t } , a _ { t } ) \\bar { A } ( s _ { t } , a _ { t } ) ] } \\\\ & { \\quad \\mathrm { V a r } _ { a _ { t } } ( \\bar { A } ) = \\mathbb { E } _ { \\pi } [ \\bar { A } ( s _ { t } , a _ { t } ) ^ { 2 } ] = \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } ^ { T } \\Sigma _ { \\theta } ( s _ { t } ) \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 209, + 483, + 759, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Sigma _ { \\theta } ( s _ { t } )$ is the covariance matrix of the stochastic policy $\\pi _ { \\theta }$ . The nice property of Eq. 11 is that $\\mathrm { V a r } _ { \\pmb { a } _ { t } } ( \\bar { A } )$ is analytical and $\\mathrm { C o v } _ { a _ { t } } ( \\hat { A } , \\bar { A } )$ can be estimated with single action sample. Using this estimate, we propose adaptive variants of Q-Prop that regulate the variance of the gradient estimate. ", + "bbox": [ + 176, + 530, + 821, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adaptive Q-Prop. The optimal state-dependent factor $\\eta ( s _ { t } )$ can be computed per state, according to $\\eta ^ { * } ( s _ { t } ) = \\mathrm { C o v } _ { a _ { t } } ( \\hat { A } , \\bar { A } ) / \\mathrm { V a r } _ { a _ { t } } ( \\bar { A } )$ . This provides maximum reduction in variance according to Eq. 11. Substituting $\\eta ^ { * } ( s _ { t } )$ into Eq. 11, we get $\\mathrm { V a r } ^ { * } = \\mathbb { E } _ { \\rho _ { \\pi } } [ ( 1 - \\rho _ { c o r r } ( \\hat { A } , \\bar { A } ) ^ { 2 } ) \\mathrm { V a r } _ { a _ { t } } ( \\hat { A } ) ]$ , where $\\rho _ { c o r r }$ is the correlation coefficient, which achieves guaranteed variance reduction if at any state $\\bar { A }$ is correlated with $\\hat { A }$ . We call this the fully adaptive Q-Prop method. An important conclusion from this analysis is that, in adaptive Q-Prop, the critic $Q _ { w }$ does not necessarily need to be approximating $Q _ { \\pi }$ well to produce good results. Its Taylor expansion merely needs to be correlated with $\\hat { A }$ , positively or even negatively. This is in contrast with actor-critic methods, where performance is greatly dependent on the absolute accuracy of the critic’s approximation. ", + "bbox": [ + 173, + 587, + 825, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conservative and Aggressive Q-Prop. In practice, the single-sample estimate of $\\mathrm { C o v } _ { a _ { t } } ( \\hat { A } , \\bar { A } )$ has high variance itself, and we propose the following two practical implementations of adaptive Q-Prop: (1) $\\eta ( s _ { t } ) = 1$ if $\\hat { \\mathrm { C o v } } _ { { \\pmb a } _ { t } } ( \\hat { A } , \\bar { A } ) > \\bar { 0 }$ and $\\eta ( s _ { t } ) = 0$ if otherwise, and (2) $\\eta ( \\pmb { \\mathscr { s } } _ { t } ) = \\mathrm { s i g n } \\big ( \\mathsf { C } \\hat { \\operatorname { o v } } _ { \\pmb { a } _ { t } } ( \\hat { A } , \\bar { A } ) \\big )$ . The first implementation, which we call conservative Q-Prop, can be thought of as a more conservative version of Q-Prop, which effectively disables the control variate for some samples of the states. This is sensible as if $\\hat { A }$ and $\\bar { A }$ are negatively correlated, it is likely that the critic is very poor. The second variant can correspondingly be termed aggressive Q-Prop, since it makes more liberal use of the control variate. ", + "bbox": [ + 173, + 737, + 825, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Q-PROP ALGORITHM ", + "text_level": 1, + "bbox": [ + 176, + 869, + 362, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pseudo-code for the adaptive Q-Prop algorithm is provided in Algorithm 1. It is a mixture of policy gradient and actor-critic. At each iteration, it first rolls out the stochastic policy to collect on-policy ", + "bbox": [ + 174, + 895, + 823, + 924 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Adaptive Q-Prop ", + "text_level": 1, + "bbox": [ + 174, + 103, + 377, + 118 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1: Initialize $w$ for critic $Q _ { w }$ , $\\theta$ for stochastic policy $\\pi _ { \\theta }$ , and replay buffer ${ \\mathcal { R } } \\gets \\emptyset$ . \n2 : repeat \n3: for $e = 1 , \\ldots , E$ do . Collect $E$ episodes of on-policy experience using $\\pi _ { \\theta }$ \n4: $s _ { 0 , e } \\sim p ( s _ { 0 } )$ \n5: for t = 0, . . . , T − 1 do \n6: $a _ { t , e } \\sim \\pi _ { \\theta } ( \\cdot | s _ { t , e } ) , s _ { t + 1 , e } \\sim p ( \\cdot | s _ { t , e } , a _ { t , e } ) , r _ { t , e } = r ( s _ { t , e } , a _ { t , e } )$ \n7: Add batch data $\\mathcal { B } = \\left. s _ { 0 : T , 1 : E } , \\pmb { a } _ { 0 : T - 1 , 1 : E } , r _ { 0 : T - 1 , 1 : E } \\right.$ to replay buffer $\\mathcal { R }$ \n8: Take $E \\cdot T$ gradient steps on $Q _ { w }$ using $\\mathcal { R }$ and $\\pi _ { \\theta }$ \n9: Fit $V _ { \\phi } ( \\pmb { s } _ { t } )$ using $\\mathcal { B }$ \n10: Compute $\\hat { A } _ { t , e }$ using $\\mathrm { G A E } ( \\lambda )$ and $\\bar { A } _ { t , e }$ using Eq. 7 \n11: Set $\\eta _ { t , e }$ based on Section 3.2 \n12: Compute and center the learning signals $l _ { t , e } = \\hat { A } _ { t , e } - \\eta _ { t , e } \\bar { A } _ { t , e }$ \n13: Compute $\\begin{array} { r } { \\nabla _ { \\theta } J ( \\theta ) \\approx \\frac { 1 } { E T } \\sum _ { e } \\sum _ { t } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t , e } | s _ { t , e } ) l _ { t , e } + \\eta _ { t , e } \\nabla _ { a } Q _ { w } ( s _ { t , e } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t , e } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t , e } ) } \\end{array}$ \n14: Take a gradient step on $\\pi _ { \\theta }$ using $\\nabla _ { \\boldsymbol { \\theta } } J ( \\boldsymbol { \\theta } )$ , optionally with a trust-region constraint using $\\mathcal { B }$ \n15: until $\\pi _ { \\theta }$ converges. ", + "bbox": [ + 174, + 122, + 821, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "samples, adds the batch to a replay buffer, takes a few gradient steps on the critic, computes $\\hat { A }$ and $\\bar { A }$ , and finally applies a gradient step on the policy $\\pi _ { \\theta }$ . In our implementation, the critic $Q _ { w }$ is fitted with off-policy TD learning using the same techniques as in DDPG (Lillicrap et al., 2016): ", + "bbox": [ + 174, + 354, + 823, + 397 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/80488ff0d73f8bbd7b698d7e98c99fb7aece1dddc06b9345ed8ffd6221da55b2.jpg", + "text": "$$\n\\begin{array} { r } { w = \\underset { w } { \\arg \\operatorname* { m i n } } \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\beta } ( \\cdot ) , a _ { t } \\sim \\beta ( \\cdot | s _ { t } ) } [ \\big ( r ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { E } _ { \\pi } [ Q ^ { \\prime } ( s _ { t + 1 } , a _ { t + 1 } ) ] - Q _ { w } ( s _ { t } , a _ { t } ) \\big ) ^ { 2 } ] . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 236, + 404, + 759, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$V _ { \\phi }$ is fitted with the same technique in (Schulman et al., 2016). Generalized advantage estimation (GAE) (Schulman et al., 2016) is used to estimate $\\hat { A }$ . The policy update can be done by any method that utilizes the first-order gradient and possibly the on-policy batch data, which includes trust region policy optimization (TRPO) (Schulman et al., 2015). Importantly, this is just one possible implementation of Q-Prop, and in Appendix C we show a more general form that can interpolate between pure policy gradient and off-policy actor-critic. ", + "bbox": [ + 174, + 429, + 825, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 LIMITATIONS ", + "text_level": 1, + "bbox": [ + 176, + 534, + 307, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A limitation with Q-Prop is that if data collection is very fast, e.g. using fast simulators, the compute time per episode is bound by the critic training at each iteration, and similar to that of DDPG and usually much more than that of TRPO. However, in applications where data collection speed is the bottleneck, there is sufficient time between policy updates to fit $Q _ { w }$ well, which can be done asynchronously from the data collection, and the compute time of Q-Prop will be about the same as that of TRPO. ", + "bbox": [ + 174, + 559, + 825, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Another limitation is the robustness to bad critics. We empirically show that our conservative Q-Prop is more robust than standard Q-Prop and much more robust than pure off-policy actor-critic methods such as DDPG; however, estimating when an off-policy critic is reliable or not is still a fundamental problem that shall be further investigated. We can also alleviate this limitation by adopting more stable off-policy critic learning techniques such as Retrace $( \\lambda )$ (Munos et al., 2016). ", + "bbox": [ + 174, + 650, + 825, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 739, + 344, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Variance reduction in policy gradient methods is a long-standing problem with a large body of prior work (Weaver & Tao, 2001; Greensmith et al., 2004; Schulman et al., 2016). However, exploration of action-dependent control variates is relatively recent, with most work focusing instead on simpler baselining techniques (Ross, 2006). A subtle exception is compatible feature approximation (Sutton et al., 1999) which can be viewed as a control variate as explained in Appendix B. Another exception is doubly robust estimator in contextual bandits (Dud´ık et al., 2011), which uses a different control variate whose bias cannot be tractably corrected. Control variates were explored recently not in RL but for approximate inference in stochastic models (Paisley et al., 2012), and the closest related work in that domain is the MuProp algorithm (Gu et al., 2016a) which uses a mean-field network as a surrogate for backpropagating a deterministic gradient through stochastic discrete variables. MuProp is not directly applicable to model-free RL because the dynamics are unknown; however, it can be if the dynamics are learned as in model-based RL (Atkeson & Santamaria, 1997; Deisenroth & Rasmussen, 2011). This model-based Q-Prop is itself an interesting direction of research as it effectively corrects bias in model-based learning. ", + "bbox": [ + 173, + 770, + 825, + 924 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 103, + 823, + 146 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Part of the benefit of Q-Prop is the ability to use off-policy data to improve on-policy policy gradient methods. Prior methods that combine off-policy data with policy gradients either introduce bias (Sutton et al., 1999; Silver et al., 2014) or use importance weighting, which is known to result in degenerate importance weights in high dimensions, resulting in very high variance (Precup, 2000; Levine & Koltun, 2013). Q-Prop provides a new approach for using off-policy data to reduce variance without introducing further bias. ", + "bbox": [ + 174, + 152, + 825, + 236 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lastly, since Q-Prop uses both on-policy policy updates and off-policy critic learning, it can take advantage of prior work along both lines of research. We chose to implement Q-Prop on top of TRPO-GAE primarily for the purpose of enabling a fair comparison in the experiments, but combining Q-Prop with other on-policy update schemes and off-policy critic training methods is an interesting direction for future work. For example, Q-Prop can also be used with other on-policy policy gradient methods such as A3C (Mnih et al., 2016) and off-policy advantage estimation methods such as Retrace $( \\lambda )$ (Munos et al., 2016), GTD2 (Sutton et al., 2009), emphatic TD (Sutton et al., 2015), and WIS-LSTD (Mahmood et al., 2014). ", + "bbox": [ + 174, + 243, + 825, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 375, + 326, + 390 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8f3888758beaed39ef5a270faa0f5bf0d9ee7237dbc5a6eaa9d962ade37ee8e4.jpg", + "image_caption": [ + "Figure 1: Illustrations of OpenAI Gym MuJoCo domains (Brockman et al., 2016; Duan et al., 2016): (a) Ant, (b) HalfCheetah, (c) Hopper, (d) Humanoid, (e) Reacher, (f) Swimmer, (g) Walker. " + ], + "image_footnote": [], + "bbox": [ + 189, + 405, + 810, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluated Q-Prop and its variants on continuous control environments from the OpenAI Gym benchmark (Brockman et al., 2016) using the MuJoCo physics simulator (Todorov et al., 2012) as shown in Figure 1. Algorithms are identified by acronyms, followed by a number indicating batch size, except for DDPG, which is a prior online actor-critic algorithm (Lillicrap et al., 2016). “c-” and “v-” denote conservative and aggressive Q-Prop variants as described in Section 3.2. “TR-” denotes trust-region policy optimization (Schulman et al., 2015), while “V-” denotes vanilla policy gradient. For example, “TR-c-Q-Prop-5000” means convervative Q-Prop with the trust-region policy update, and a batch size of 5000. “VPG” and “TRPO” are vanilla policy gradient and trust-region policy optimization respectively (Schulman et al., 2016; Duan et al., 2016). Unless otherwise stated, all policy gradient methods are implemented with GAE $\\lambda = 0 . 9 7$ ) (Schulman et al., 2016). Note that TRPOGAE is currently the state-of-the-art method on most of the OpenAI Gym benchmark tasks, though our experiments show that a well-tuned DDPG implementation sometimes achieves better results. Our algorithm implementations are built on top of the rllab TRPO and DDPG codes from Duan et al. (2016) and available at https://github.com/shaneshixiang/rllabplusplus. Policy and value function architectures and other training details including hyperparameter values are provided in Appendix D. ", + "bbox": [ + 174, + 547, + 825, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1 ADAPTIVE Q-PROP ", + "text_level": 1, + "bbox": [ + 176, + 786, + 348, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, it is useful to identify how reliable each variant of Q-Prop is. In this section, we analyze standard Q-Prop and two adaptive variants, c-Q-Prop and a-Q-Prop, and demonstrate the stability of the method across different batch sizes. Figure 2a shows a comparison of Q-Prop variants with trust-region updates on the HalfCheetah-v1 domain, along with the best performing TRPO hyperparameters. The results are consistent with theory: conservative Q-Prop achieves much more stable performance than the standard and aggressive variants, and all Q-Prop variants significantly outperform TRPO in terms of sample efficiency, e.g. conservative Q-Prop reaches average reward of 4000 using about 10 times less samples than TRPO. ", + "bbox": [ + 174, + 811, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8ef0cab31ceb57d2ca2f3840e44751e97cf886ce5ce81bc26d79da92a4d50c5.jpg", + "image_caption": [ + "(a) Standard Q-Prop vs adaptive variants. " + ], + "image_footnote": [], + "bbox": [ + 214, + 103, + 462, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac3e5492e7295b46b13f17d48d9490a7f429f569aab0d8bfb64532dc2013b0fc.jpg", + "image_caption": [ + "(b) Conservative Q-Prop vs TRPO across batch sizes. ", + "Figure 2: Average return over episodes in HalfCheetah-v1 during learning, exploring adaptive QProp methods and different batch sizes. All variants of Q-Prop substantially outperform TRPO in terms of sample efficiency. TR-c-QP, conservative Q-Prop with trust-region update performs most stably across different batch sizes. " + ], + "image_footnote": [], + "bbox": [ + 534, + 103, + 782, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 2b shows the performance of conservative Q-Prop against TRPO across different batch sizes. Due to high variance in gradient estimates, TRPO typically requires very large batch sizes, e.g. 25000 steps or 25 episodes per update, to perform well. We show that our Q-Prop methods can learn even with just 1 episode per update, and achieves better sample efficiency with small batch sizes. This shows that Q-Prop significantly reduces the variance compared to the prior methods. ", + "bbox": [ + 173, + 354, + 825, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As we discussed in Section 1, stability is a significant challenge with state-of-the-art deep RL methods, and is very important for being able to reliably use deep RL for real world tasks. In the rest of the experiments, we will use conservative Q-Prop as the main Q-Prop implementation. ", + "bbox": [ + 174, + 431, + 825, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d75d5db85e34384abd1e96bb494b10087f6138234cb0f906d3e4c4dba89c33ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 554, + 464, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(a) Comparing algorithms on HalfCheetah-v1. ", + "bbox": [ + 202, + 695, + 473, + 709 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a609c411ab9a5449d74b3fe5b80daffd18f515d8845ee6448717ce07addc7bd5.jpg", + "image_caption": [ + "(b) Comparing algorithms on Humanoid-v1. ", + "Figure 3: Average return over episodes in HalfCheetah-v1 and Humanoid-v1 during learning, comparing Q-Prop against other model-free algorithms. Q-Prop with vanilla policy gradient outperforms TRPO on HalfCheetah. Q-Prop significantly outperforms TRPO in convergence time on Humanoid. " + ], + "image_footnote": [], + "bbox": [ + 532, + 555, + 784, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we evaluate two versions of conservative Q-Prop, v-c-Q-Prop using vanilla policy gradient and TR-c-Q-Prop using trust-region updates, against other model-free algorithms on the HalfCheetah-v1 domain. Figure 3a shows that c-Q-Prop methods significantly outperform the best TRPO and VPG methods. Even Q-Prop with vanilla policy gradient is comparable to TRPO, confirming the significant benefits from variance reduction. DDPG on the other hand exhibits inconsistent performances. With proper reward scaling, i.e. “DDPG-r0.1”, it outperforms other methods as well as the DDPG results reported in prior work (Duan et al., 2016; Amos et al., 2016). This illustrates the sensitivity of DDPG to hyperparameter settings, while Q-Prop exhibits more stable, monotonic learning behaviors when compared to DDPG. In the next section we show this improved stability allows Q-Prop to outperform DDPG in more complex domains. ", + "bbox": [ + 173, + 785, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3 EVALUATION ACROSS DOMAINS ", + "text_level": 1, + "bbox": [ + 176, + 103, + 434, + 117 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Lastly, we evaluate Q-Prop against TRPO and DDPG across multiple domains. While the gym environments are biased toward locomotion, we expect we can achieve similar performance on manipulation tasks such as those in Lillicrap et al. (2016). Table 1 summarizes the results, including the best attained average rewards and the steps to convergence. Q-Prop consistently outperform TRPO in terms of sample complexity and sometimes achieves higher rewards than DDPG in more complex domains. A particularly notable case is shown in Figure 3b, where Q-Prop substantially improves sample efficiency over TRPO on Humanoid-v1 domain, while DDPG cannot find a good solution. ", + "bbox": [ + 173, + 131, + 825, + 228 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The better performance on the more complex domains highlights the importance of stable deep RL algorithms: while costly hyperparameter sweeps may allow even less stable algorithms to perform well on simpler problems, more complex tasks might have such narrow regions of stable hyperparameters that discovering them becomes impractical. ", + "bbox": [ + 174, + 236, + 825, + 291 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/727463be8b5bbf2c6633678d0ac605de2691434167ca6598b21e0e8b7985947c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TR-c-Q-PropTRPODDPG
DomainThresholdMaxReturn.EpisodesMaxReturnEpsisodesMaxReturnEpisodes
Ant350035344975423913825957N/A
HalfCheetah47004811207854734263707490600
Hopper200029575945248657152604965
Humanoid2500>349214750918>30000552N/A
Reacher-7-6.02060-6.72840-6.61800
Swimmer9010320451103025150500
Walker30004030368535671887536262125
", + "bbox": [ + 173, + 305, + 834, + 425 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 1: Q-Prop, TRPO and DDPG results showing the max average rewards attained in the first $3 0 \\mathrm { k }$ episodes and the episodes to cross specific reward thresholds. Q-Prop often learns more sample efficiently than TRPO and can solve difficult domains such as Humanoid better than DDPG. ", + "bbox": [ + 174, + 434, + 826, + 476 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 DISCUSSION AND CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 501, + 468, + 516 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We presented Q-Prop, a policy gradient algorithm that combines reliable, consistent, and potentially unbiased on-policy gradient estimation with a sample-efficient off-policy critic that acts as a control variate. The method provides a large improvement in sample efficiency compared to stateof-the-art policy gradient methods such as TRPO, while outperforming state-of-the-art actor-critic methods on more challenging tasks such as humanoid locomotion. We hope that techniques like these, which combine on-policy Monte Carlo gradient estimation with sample-efficient variance reduction through off-policy critics, will eventually lead to deep reinforcement learning algorithms that are more stable and efficient, and therefore better suited for application to complex real-world learning tasks. ", + "bbox": [ + 174, + 534, + 825, + 659 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ACKNOWLEDGMENTS ", + "text_level": 1, + "bbox": [ + 176, + 679, + 326, + 691 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Rocky Duan for sharing and answering questions about rllab code, and Yutian Chen and Laurent Dinh for discussion on control variates. SG and RT were funded by NSERC, Google, and EPSRC grants EP/L000776/1 and EP/M026957/1. ZG was funded by EPSRC grant EP/J012300/1 and the Alan Turing Institute (EP/N510129/1). ", + "bbox": [ + 174, + 703, + 825, + 758 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 784, + 285, + 799 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Brandon Amos, Lei Xu, and J Zico Kolter. Input convex neural networks. arXiv preprint arXiv:1609.07152, 2016. ", + "bbox": [ + 173, + 809, + 821, + 837 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Christopher G Atkeson and Juan Carlos Santamaria. A comparison of direct and model-based reinforcement learning. In In International Conference on Robotics and Automation. Citeseer, 1997. ", + "bbox": [ + 171, + 852, + 823, + 881 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016. ", + "bbox": [ + 176, + 895, + 823, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Marc Deisenroth and Carl E Rasmussen. Pilco: A model-based and data-efficient approach to policy search. In Proceedings of the 28th International Conference on machine learning (ICML-11), pp. 465–472, 2011. ", + "bbox": [ + 174, + 103, + 823, + 146 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yan Duan, Xi Chen, Rein Houthooft, John Schulman, and Pieter Abbeel. Benchmarking deep reinforcement learning for continuous control. International Conference on Machine Learning (ICML), 2016. ", + "bbox": [ + 174, + 155, + 821, + 196 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Miroslav Dud´ık, John Langford, and Lihong Li. Doubly robust policy evaluation and learning. arXiv preprint arXiv:1103.4601, 2011. ", + "bbox": [ + 171, + 205, + 823, + 234 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evan Greensmith, Peter L Bartlett, and Jonathan Baxter. Variance reduction techniques for gradient estimates in reinforcement learning. Journal of Machine Learning Research, 5(Nov):1471–1530, 2004. ", + "bbox": [ + 176, + 242, + 825, + 285 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Shixiang Gu, Sergey Levine, Ilya Sutskever, and Andriy Mnih. Muprop: Unbiased backpropagation for stochastic neural networks. International Conference on Learning Representations (ICLR), 2016a. ", + "bbox": [ + 173, + 294, + 825, + 337 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Shixiang Gu, Tim Lillicrap, Ilya Sutskever, and Sergey Levine. Continuous deep q-learning with model-based acceleration. In International Conference on Machine Learning (ICML), 2016b. ", + "bbox": [ + 173, + 344, + 821, + 375 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Hado V Hasselt. Double q-learning. In Advances in Neural Information Processing Systems, pp. 2613–2621, 2010. ", + "bbox": [ + 174, + 382, + 821, + 411 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sham Kakade. A natural policy gradient. In NIPS, volume 14, pp. 1531–1538, 2001. ", + "bbox": [ + 174, + 419, + 730, + 435 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. ", + "bbox": [ + 171, + 443, + 823, + 472 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Guy Lever. Deterministic policy gradient algorithms. 2014. ", + "bbox": [ + 173, + 481, + 566, + 496 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sergey Levine and Vladlen Koltun. Guided policy search. In International Conference on Machine Learning (ICML), pp. 1–9, 2013. ", + "bbox": [ + 171, + 503, + 823, + 534 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. International Conference on Learning Representations (ICLR), 2016. ", + "bbox": [ + 174, + 541, + 823, + 585 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A Rupam Mahmood, Hado P van Hasselt, and Richard S Sutton. Weighted importance sampling for off-policy learning with linear function approximation. In Advances in Neural Information Processing Systems, pp. 3014–3022, 2014. ", + "bbox": [ + 173, + 592, + 825, + 636 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Andriy Mnih and Karol Gregor. Neural variational inference and learning in belief networks. International Conference on Machine Learning (ICML), 2014. ", + "bbox": [ + 168, + 643, + 823, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015. ", + "bbox": [ + 174, + 680, + 825, + 724 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy P Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International Conference on Machine Learning (ICML), 2016. ", + "bbox": [ + 174, + 731, + 825, + 775 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Remi Munos, Tom Stepleton, Anna Harutyunyan, and Marc G Bellemare. Safe and efficient off-´ policy reinforcement learning. arXiv preprint arXiv:1606.02647, 2016. ", + "bbox": [ + 173, + 782, + 823, + 813 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "John Paisley, David Blei, and Michael Jordan. Variational bayesian inference with stochastic search. International Conference on Machine Learning (ICML), 2012. ", + "bbox": [ + 174, + 820, + 821, + 849 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jan Peters and Stefan Schaal. Policy gradient methods for robotics. In International Conference on Intelligent Robots and Systems (IROS), pp. 2219–2225. IEEE, 2006. ", + "bbox": [ + 174, + 858, + 821, + 887 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jan Peters, Katharina Mulling, and Yasemin Altun. Relative entropy policy search. In ¨ AAAI. Atlanta, 2010. ", + "bbox": [ + 174, + 895, + 823, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Doina Precup. Eligibility traces for off-policy policy evaluation. Computer Science Department Faculty Publication Series, pp. 80, 2000. ", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sheldon M Ross. Simulation. Burlington, MA: Elsevier, 2006. ", + "bbox": [ + 173, + 143, + 584, + 160 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "John Schulman, Sergey Levine, Pieter Abbeel, Michael I. Jordan, and Philipp Moritz. Trust region policy optimization. In International Conference on Machine Learning (ICML), pp. 1889–1897, 2015. ", + "bbox": [ + 174, + 170, + 826, + 213 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. Highdimensional continuous control using generalized advantage estimation. International Conference on Learning Representations (ICLR), 2016. ", + "bbox": [ + 173, + 226, + 825, + 268 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. Deterministic policy gradient algorithms. In International Conference on Machine Learning (ICML), 2014. ", + "bbox": [ + 174, + 280, + 825, + 321 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016. ", + "bbox": [ + 174, + 334, + 825, + 377 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Richard S Sutton. Integrated architectures for learning, planning, and reacting based on approximating dynamic programming. In International Conference on Machine Learning (ICML), pp. 216–224, 1990. ", + "bbox": [ + 173, + 390, + 821, + 431 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Richard S Sutton, David A McAllester, Satinder P Singh, Yishay Mansour, et al. Policy gradient methods for reinforcement learning with function approximation. In Advances in Neural Information Processing Systems (NIPS), volume 99, pp. 1057–1063, 1999. ", + "bbox": [ + 174, + 444, + 825, + 487 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Richard S Sutton, Hamid Reza Maei, Doina Precup, Shalabh Bhatnagar, David Silver, Csaba Szepesvari, and Eric Wiewiora. Fast gradient-descent methods for temporal-difference learning ´ with linear function approximation. In Proceedings of the 26th Annual International Conference on Machine Learning, pp. 993–1000. ACM, 2009. ", + "bbox": [ + 173, + 498, + 825, + 555 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Richard S Sutton, A Rupam Mahmood, and Martha White. An emphatic approach to the problem of off-policy temporal-difference learning. The Journal of Machine Learning Research, 2015. ", + "bbox": [ + 173, + 566, + 825, + 597 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Philip Thomas. Bias in natural actor-critic algorithms. In ICML, pp. 441–448, 2014. ", + "bbox": [ + 174, + 607, + 727, + 623 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 5026–5033. IEEE, 2012. ", + "bbox": [ + 174, + 633, + 825, + 678 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Christopher JCH Watkins and Peter Dayan. Q-learning. Machine learning, 8(3-4):279–292, 1992. ", + "bbox": [ + 173, + 689, + 815, + 705 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Lex Weaver and Nigel Tao. The optimal reward baseline for gradient-based reinforcement learning. In Proceedings of the Seventeenth conference on Uncertainty in artificial intelligence, pp. 538– 545. Morgan Kaufmann Publishers Inc., 2001. ", + "bbox": [ + 176, + 715, + 823, + 760 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3-4):229–256, 1992. ", + "bbox": [ + 174, + 771, + 823, + 799 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A Q-PROP ESTIMATOR DERIVATION ", + "text_level": 1, + "bbox": [ + 176, + 829, + 490, + 845 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The full derivation of the Q-Prop estimator is shown in Eq. 14. We make use of the following property that is commonly used in baseline derivations: ", + "bbox": [ + 173, + 861, + 823, + 890 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/d098e81861b78d720d0cc9ebc52a79d42144e032693153f767c944347369f851.jpg", + "text": "$$\n\\mathbb { E } _ { p _ { \\theta } ( x ) } [ \\nabla _ { \\theta } \\log p _ { \\theta } ( x ) ] = \\int _ { x } \\nabla _ { \\theta } p _ { \\theta } ( x ) = \\nabla _ { \\theta } \\int _ { x } p ( x ) = 0\n$$", + "text_format": "latex", + "bbox": [ + 325, + 898, + 671, + 929 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "This holds true when $f ( \\mathbf { \\boldsymbol { s } } _ { t } , \\mathbf { \\boldsymbol { a } } _ { t } )$ is an arbitrary function differentiable with respect to $\\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { a } _ { t }$ and $\\bar { f }$ is its first-order Taylor expansion around $\\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf $ , i.e. $\\bar { f } ( s _ { t } , a _ { t } ) = f ( s _ { t } , \\bar { a } _ { t } ) + \\nabla _ { a } \\bar { f ( } s _ { t } , a ) | _ { a = \\bar { a } _ { t } } ( a _ { t } - \\bar { a } _ { t } ) .$ . Here, $\\mu _ { \\theta } ( \\pmb { s } _ { t } ) \\dot { = } \\mathbb { E } _ { \\pi } [ \\bar { \\mathbf { a } } _ { t } ]$ is the mean of stochastic policy $\\pi _ { \\theta }$ . The derivation appears below: ", + "bbox": [ + 173, + 102, + 825, + 147 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/c9025713638b6bd3ac0c25065e2b04ac97b90ae6567399821a9170713a729d10.jpg", + "text": "$$\n\\begin{array} { r l } { \\nabla _ { \\theta } J ( \\theta ) = \\mathrm { E } _ { \\mathrm { g } , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { 1 } | s _ { 1 } \\big ) \\big ( \\hat { Q } ( s _ { t } , a _ { t } ) - \\hat { J } ( s _ { t } , a _ { t } ) \\big ) \\big ] + \\mathrm { E } _ { \\rho , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { t } | s _ { t } \\big ) \\hat { J } ( s _ { t } , a _ { t } ) \\big ] } \\\\ { g ( \\theta ) = \\mathrm { E } _ { \\mathrm { g } , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { 1 } | s _ { 1 } \\big ) \\hat { J } ( s _ { t } , a _ { t } ) \\big ] } \\\\ & { = \\mathrm { E } _ { \\rho , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { 1 } | s _ { 1 } \\big ) \\big ( \\hat { J } ( s _ { t } , a _ { t } ) + \\nabla _ { a _ { t } } \\big ) \\big ( s _ { t } , a _ { t } \\big ) \\big ] } \\\\ & { = \\mathrm { E } _ { \\rho , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { 1 } | s _ { 1 } \\big ) \\mathrm { e } _ { n } f \\big ( s _ { t } , a _ { t } \\big ) \\big ] = - \\alpha _ { 1 } \\alpha _ { 1 } \\alpha _ { 1 } } \\\\ & { = \\mathrm { E } _ { \\rho , \\pi } \\bigg [ \\int _ { a _ { t } } \\mathrm { e } _ { g } \\pi _ { \\theta } \\big ( a _ { t } | s _ { t } \\big ) \\mathrm { e } _ { n } f \\big ( s _ { t } , a _ { t } \\big ) \\big ] \\omega - \\alpha _ { 2 } \\alpha _ { 2 } \\alpha _ { 2 } \\bigg ] } \\\\ & { = \\mathrm { E } _ { \\rho , \\pi } \\bigg [ \\sum _ { \\theta } \\mathrm { e } _ { g } \\pi _ { \\theta } \\big ( a _ { t } \\big ) \\big ( \\frac { \\kappa } { \\alpha _ { 1 } } \\big ) \\mathrm { e } _ { n } f \\big ( \\alpha _ { 1 } \\big ) \\omega _ { 1 } \\omega _ { 2 } \\bigg ] } \\\\ & = \\mathrm { E } _ { \\rho , \\pi } \\bigg [ \\mathrm { v } _ { \\alpha } f \\big ( s _ { t } , a \\big ) \\big | \\alpha - \\alpha _ { 1 } \\int _ a _ { t } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 184, + 154, + 782, + 378 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "B CONNECTION BETWEEN Q-PROP AND COMPATIBLE FEATURE APPROXIMATION ", + "text_level": 1, + "bbox": [ + 174, + 390, + 722, + 425 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section we show that actor-critic with compatible feature approximation is a form of control variate. A critic $Q _ { w }$ is compatible (Sutton et al., 1999) if it satisfies (1) $Q _ { w } ( \\pmb { s } _ { t } , \\pmb { a } _ { t } ) =$ $w ^ { T } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( { \\mathbf { \\em a } } _ { t } | { \\mathbf { \\em s } } _ { t } )$ , i.e. $\\nabla _ { w } \\mathcal { Q } _ { w } ( s _ { t } , \\bar { a } _ { t } ) = \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } )$ , and (2) $w$ is fit with objective $w =$ a $\\mathrm { r g } \\operatorname* { m i n } _ { w } L ( w ) = \\mathrm { a r g } \\operatorname* { m i n } _ { w } \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ ( \\hat { Q } ( s _ { t } , a _ { t } ) - Q _ { w } ( { \\pmb s } _ { t } , { \\pmb a } _ { t } ) ) ^ { 2 } ]$ , that is fitting $Q _ { w }$ on on-policy Monte Carlo returns. Condition (2) implies the following identity, ", + "bbox": [ + 173, + 439, + 826, + 512 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/df93697e3d7e251afed3d6d2c3874531f249deeba1bf5f574eb2bd592d544520.jpg", + "text": "$$\n\\begin{array} { r } { \\nabla _ { w } L = 2 \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { \\mathcal { Q } } ( s _ { t } , a _ { t } ) - { \\mathcal Q } _ { w } ( s _ { t } , a _ { t } ) ) ] = 0 . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 297, + 518, + 700, + 539 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In compatible feature approximation, it directly uses $Q _ { w }$ as control variate, rather than its Taylor expansion $\\bar { Q } _ { w }$ as in Q-Prop. Using Eq. 15, the Monte Carlo policy gradient is, ", + "bbox": [ + 173, + 546, + 825, + 575 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/190a4f5c01b0e37dffa0037afb35be6b86f71f49759e6a8105d1953d9e424153.jpg", + "text": "$$\n\\begin{array} { r l } & { \\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ) Q _ { w } ( \\boldsymbol { s } _ { t } , \\boldsymbol { a } _ { t } ) ] } \\\\ & { \\qquad = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ ( \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ) \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ) ^ { T } ) w ] } \\\\ & { \\qquad = \\mathbb { E } _ { \\rho _ { \\pi } } [ I ( \\theta ; \\boldsymbol { s } _ { t } ) w ] , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 313, + 579, + 683, + 638 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "where $I ( \\theta ; s _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } \\vert s _ { t } ) \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } \\vert s _ { t } ) ^ { T } ]$ is Fisher’s information matrix. Thus, variance reduction depends on ability to compute or estimate $I ( \\theta ; s _ { t } )$ and $w$ effectively. ", + "bbox": [ + 176, + 643, + 820, + 674 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "C UNIFYING POLICY GRADIENT AND ACTOR-CRITIC ", + "text_level": 1, + "bbox": [ + 174, + 693, + 633, + 710 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Q-Prop closely ties together policy gradient and actor-critic algorithms. To analyze this point, we write a generalization of Eq. 9 below, introducing two additional variables $\\alpha , \\rho _ { C R }$ : ", + "bbox": [ + 174, + 724, + 823, + 755 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/81cb6159a91a178254bf4a790617f7e216698fa5af04494a2545828fb52a3a93.jpg", + "text": "$$\n\\begin{array} { r l } & { \\nabla _ { \\theta } J ( \\theta ) \\propto \\alpha \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { \\boldsymbol A } ( s _ { t } , a _ { t } ) - \\eta \\bar { \\boldsymbol A } _ { w } ( s _ { t } , a _ { t } ) ] } \\\\ & { \\qquad + \\eta \\mathbb { E } _ { \\rho _ { C R } } [ \\nabla _ { a } \\mathcal { Q } _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 300, + 760, + 696, + 801 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Eq. 17 enables more analysis where bias generally is introduced only when $\\alpha \\neq 1$ or $\\rho _ { C R } \\neq \\rho _ { \\pi }$ . Importantly, Eq. 17 covers both policy gradient and deterministic actor-critic algorithm as its special cases. Standard policy gradient is recovered by $\\eta = 0$ , and deterministic actor-critic is recovered by $\\alpha = 0$ and $\\rho _ { C R } = \\rho _ { \\beta }$ . This allows heuristic or automatic methods for dynamically changing these variables through the learning process for optimizing different metrics, e.g. sample efficiency, convergence speed, stability. ", + "bbox": [ + 173, + 804, + 825, + 890 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2 summarizes the various edge cases of Eq. 17. For example, since we derive our method from a control variates standpoint, $Q _ { w }$ can be any function and the gradient remains almost unbiased (see ", + "bbox": [ + 174, + 895, + 823, + 924 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/1aa12e84ce923bcad695b4e5e23bee23b45c30c0a31485d12c1fc031b384cf09.jpg", + "table_caption": [ + "Table 2: Implementation options and edge cases of the generalized Q-Prop estimator in Eq. 17. " + ], + "table_footnote": [], + "table_body": "
ParameterImplementation optionsIntroduce bias?
Qwoff-policy TD; on-policy TD(λ); model-based; etc.No
Von-policy Monte Carlo fitting; Eπθ [Qw(St, at)}; etcNo
0≤λ≤1Yes,except=1
αα≥0Yes, except α=1
nany nNo
PCRp of any policyYes, except pcR=pπ
", + "bbox": [ + 225, + 99, + 772, + 194 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Section 2.1). A natural choice is to use off-policy temporal difference learning to learn the critic $Q _ { w }$ corresponding to policy $\\pi$ . This enables effectively utilizing off-policy samples without introducing further bias. An interesting alternative to this is to utilize model-based roll-outs to estimate the critic, which resembles MuProp in stochastic neural networks (Gu et al., 2016a). Unlike prior work on using fitted dynamics model to accelerate model-free learning (Gu et al., 2016b), this approach does not introduce bias to the gradient of the original objective. ", + "bbox": [ + 173, + 243, + 825, + 328 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "D EXPERIMENT DETAILS ", + "text_level": 1, + "bbox": [ + 174, + 348, + 400, + 363 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Policy and value function architectures. The network architectures are largely based on the benchmark paper by Duan et al. (2016). For policy gradient methods, the stochastic policy $\\pi _ { \\boldsymbol { \\theta } } ( \\mathbf { a } _ { t } | \\mathbf { s } _ { t } ) = \\bar { \\mathcal { N } } ( \\mu _ { \\boldsymbol { \\theta } } ( \\mathbf { \\bar { s } } _ { t } ) , \\Sigma _ { \\boldsymbol { \\theta } } )$ is a local Gaussian policy with a local state-dependent mean and a global covariance matrix. $\\pmb { \\mu } _ { \\pmb { \\theta } } ( \\pmb { s } _ { t } )$ is a neural network with 3 hidden layers of sizes 100-50-25 and tanh nonlinearities at the first 2 layers, and $\\Sigma _ { \\theta }$ is diagonal. For DDPG, the policy is deterministic and has the same architecture as $\\mu _ { \\theta }$ except that it has an additional tanh layer at the output. $V _ { \\phi } ( \\pmb { s } _ { t } )$ for baselines and GAE is fit with the same technique by Schulman et al. (2016), a variant of linear regression on Monte Carlo returns with soft-update constraint. For Q-Prop and DDPG, $Q _ { w } ( s , \\pmb { a } )$ is parametrized with a neural network with 2 hidden layers of size 100 and ReLU nonlinearity, where $^ { a }$ is included after the first hidden layer. ", + "bbox": [ + 173, + 380, + 825, + 518 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Training details. This section describes parameters of the training algorithms and their hyperparameter search values in $\\{ \\}$ . The optimal performing hyperparameter results are reported. Policy gradient methods (VPG, TRPO, Q-Prop) used batch sizes of $\\{ 1 0 0 0 , 5 0 0 0 , 2 5 0 0 0 \\}$ time steps, step sizes of $\\{ 0 . 1 , 0 . 0 1 , 0 . 0 0 1 \\}$ for the trust-region method, and base learning rates of $\\{ 0 . 0 0 1 , \\bar { 0 } . 0 0 0 1 \\}$ with Adam (Kingma & Ba, 2014) for vanilla policy gradient methods. For Q-Prop and DDPG, $Q _ { w }$ is learned with the same technique as in DDPG (Lillicrap et al., 2016), using soft target networks with $\\tau = 0 . 9 9 9$ , a replay buffer of size $1 0 ^ { 6 }$ steps, a mini-batch size of 64, and a base learning rate of $\\{ 0 . 0 0 1 , 0 . 0 0 0 1 \\}$ with Adam (Kingma & Ba, 2014). For Q-Prop we also tuned the relative ratio of gradient steps on the critic $Q _ { w }$ against the number of steps on the policy, in the range $\\{ 0 . 1 , 0 . 5$ , $1 . 0 \\}$ , where 0.1 corresponds to 100 critic updates for every policy update if the batch size is 1000. For DDPG, we swept the reward scaling using $\\{ 0 . 0 1 , 0 . 1 , \\dot { 1 . } 0 \\dot { \\} }$ as it is sensitive to this parameter. ", + "bbox": [ + 173, + 534, + 825, + 688 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/parse/train/SJ3rcZcxl/SJ3rcZcxl_middle.json b/parse/train/SJ3rcZcxl/SJ3rcZcxl_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..c5f2ed16902dd14c8689f7a679b444ca9137a5a4 --- /dev/null +++ b/parse/train/SJ3rcZcxl/SJ3rcZcxl_middle.json @@ -0,0 +1,38243 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 76, + 506, + 99 + ], + "spans": [ + { + "bbox": [ + 105, + 76, + 506, + 99 + ], + "score": 1.0, + "content": "Q-PROP: SAMPLE-EFFICIENT POLICY GRADIENT", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 97, + 331, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 331, + 118 + ], + "score": 1.0, + "content": "WITH AN OFF-POLICY CRITIC", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 134, + 522, + 147 + ], + "lines": [ + { + "bbox": [ + 111, + 132, + 525, + 150 + ], + "spans": [ + { + "bbox": [ + 111, + 132, + 151, + 150 + ], + "score": 1.0, + "content": "Shixiang", + "type": "text" + }, + { + "bbox": [ + 152, + 134, + 177, + 146 + ], + "score": 0.83, + "content": "\\mathbf { G u } ^ { 1 2 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 132, + 525, + 150 + ], + "score": 1.0, + "content": ", Timothy Lillicrap4, Zoubin Ghahramani16, Richard E. Turner1, Sergey Levine35", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 112, + 147, + 457, + 239 + ], + "lines": [ + { + "bbox": [ + 112, + 146, + 460, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 460, + 159 + ], + "score": 1.0, + "content": "sg717@cam.ac.uk,countzero@google.com,zoubin@eng.cam.ac.uk,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 158, + 365, + 169 + ], + "spans": [ + { + "bbox": [ + 112, + 158, + 365, + 169 + ], + "score": 1.0, + "content": "ret26@cam.ac.uk,svlevine@eecs.berkeley.edu", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 167, + 239, + 181 + ], + "spans": [ + { + "bbox": [ + 111, + 167, + 239, + 181 + ], + "score": 1.0, + "content": "1University of Cambridge, UK", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 178, + 380, + 194 + ], + "spans": [ + { + "bbox": [ + 111, + 178, + 380, + 194 + ], + "score": 1.0, + "content": "2Max Planck Institute for Intelligent Systems, Tubingen, Germany ¨", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 190, + 199, + 206 + ], + "spans": [ + { + "bbox": [ + 111, + 190, + 199, + 206 + ], + "score": 1.0, + "content": "3Google Brain, USA", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 111, + 202, + 182, + 218 + ], + "spans": [ + { + "bbox": [ + 111, + 202, + 182, + 218 + ], + "score": 1.0, + "content": "4DeepMind, UK", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 111, + 214, + 196, + 230 + ], + "spans": [ + { + "bbox": [ + 111, + 214, + 196, + 230 + ], + "score": 1.0, + "content": "5UC Berkeley, USA", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 111, + 226, + 199, + 242 + ], + "spans": [ + { + "bbox": [ + 111, + 226, + 199, + 242 + ], + "score": 1.0, + "content": "6Uber AI Labs, USA", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 278, + 269, + 333, + 281 + ], + "lines": [ + { + "bbox": [ + 276, + 268, + 335, + 282 + ], + "spans": [ + { + "bbox": [ + 276, + 268, + 335, + 282 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 143, + 294, + 468, + 491 + ], + "lines": [ + { + "bbox": [ + 142, + 294, + 470, + 306 + ], + "spans": [ + { + "bbox": [ + 142, + 294, + 470, + 306 + ], + "score": 1.0, + "content": "Model-free deep reinforcement learning (RL) methods have been successful in a", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 305, + 470, + 317 + ], + "spans": [ + { + "bbox": [ + 142, + 305, + 470, + 317 + ], + "score": 1.0, + "content": "wide variety of simulated domains. However, a major obstacle facing deep RL", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 316, + 469, + 328 + ], + "spans": [ + { + "bbox": [ + 141, + 316, + 469, + 328 + ], + "score": 1.0, + "content": "in the real world is their high sample complexity. Batch policy gradient methods", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 325, + 470, + 341 + ], + "spans": [ + { + "bbox": [ + 141, + 325, + 470, + 341 + ], + "score": 1.0, + "content": "offer stable learning, but at the cost of high variance, which often requires large", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 338, + 470, + 351 + ], + "spans": [ + { + "bbox": [ + 141, + 338, + 470, + 351 + ], + "score": 1.0, + "content": "batches. TD-style methods, such as off-policy actor-critic and Q-learning, are", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 349, + 469, + 362 + ], + "spans": [ + { + "bbox": [ + 141, + 349, + 469, + 362 + ], + "score": 1.0, + "content": "more sample-efficient but biased, and often require costly hyperparameter sweeps", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 361, + 470, + 372 + ], + "spans": [ + { + "bbox": [ + 141, + 361, + 470, + 372 + ], + "score": 1.0, + "content": "to stabilize. In this work, we aim to develop methods that combine the stability of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 371, + 469, + 384 + ], + "spans": [ + { + "bbox": [ + 141, + 371, + 469, + 384 + ], + "score": 1.0, + "content": "policy gradients with the efficiency of off-policy RL. We present Q-Prop, a policy", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 382, + 470, + 394 + ], + "spans": [ + { + "bbox": [ + 141, + 382, + 470, + 394 + ], + "score": 1.0, + "content": "gradient method that uses a Taylor expansion of the off-policy critic as a control", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 393, + 469, + 405 + ], + "spans": [ + { + "bbox": [ + 142, + 393, + 469, + 405 + ], + "score": 1.0, + "content": "variate. Q-Prop is both sample efficient and stable, and effectively combines the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 403, + 470, + 417 + ], + "spans": [ + { + "bbox": [ + 141, + 403, + 470, + 417 + ], + "score": 1.0, + "content": "benefits of on-policy and off-policy methods. We analyze the connection between", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 414, + 469, + 427 + ], + "spans": [ + { + "bbox": [ + 141, + 414, + 469, + 427 + ], + "score": 1.0, + "content": "Q-Prop and existing model-free algorithms, and use control variate theory to de-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 140, + 425, + 470, + 439 + ], + "spans": [ + { + "bbox": [ + 140, + 425, + 470, + 439 + ], + "score": 1.0, + "content": "rive two variants of Q-Prop with conservative and aggressive adaptation. We show", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 437, + 470, + 450 + ], + "spans": [ + { + "bbox": [ + 141, + 437, + 470, + 450 + ], + "score": 1.0, + "content": "that conservative Q-Prop provides substantial gains in sample efficiency over trust", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 448, + 470, + 460 + ], + "spans": [ + { + "bbox": [ + 141, + 448, + 470, + 460 + ], + "score": 1.0, + "content": "region policy optimization (TRPO) with generalized advantage estimation (GAE),", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 459, + 469, + 471 + ], + "spans": [ + { + "bbox": [ + 141, + 459, + 469, + 471 + ], + "score": 1.0, + "content": "and improves stability over deep deterministic policy gradient (DDPG), the state-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 142, + 470, + 469, + 482 + ], + "spans": [ + { + "bbox": [ + 142, + 470, + 469, + 482 + ], + "score": 1.0, + "content": "of-the-art on-policy and off-policy methods, on OpenAI Gym’s MuJoCo continu-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 142, + 482, + 248, + 493 + ], + "spans": [ + { + "bbox": [ + 142, + 482, + 248, + 493 + ], + "score": 1.0, + "content": "ous control environments.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 20.5 + }, + { + "type": "title", + "bbox": [ + 108, + 514, + 206, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 208, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 208, + 529 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 540, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "Model-free reinforcement learning is a promising approach for solving arbitrary goal-directed se-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "quential decision-making problems with only high-level reward signals and no supervision. It has", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "recently been extended to utilize large neural network policies and value functions, and has been", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "shown to be successful in solving a range of difficult problems (Mnih et al., 2015; Schulman et al.,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "2015; Lillicrap et al., 2016; Silver et al., 2016; Gu et al., 2016b; Mnih et al., 2016). Deep neural", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 595, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 607 + ], + "score": 1.0, + "content": "network parametrization minimizes the need for manual feature and policy engineering, and allows", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 604, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 619 + ], + "score": 1.0, + "content": "learning end-to-end policies mapping from high-dimensional inputs, such as images, directly to ac-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "tions. However, such expressive parametrization also introduces a number of practical problems.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 626, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 641 + ], + "score": 1.0, + "content": "Deep reinforcement learning algorithms tend to be sensitive to hyperparameter settings, often re-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "score": 1.0, + "content": "quiring extensive hyperparameter sweeps to find good values. Poor hyperparameter settings tend to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "produce unstable or non-convergent learning. Deep RL algorithms also tend to exhibit high sample", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 660, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 506, + 673 + ], + "score": 1.0, + "content": "complexity, often to the point of being impractical to run on real physical systems. Although a num-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "ber of recent techniques have sought to alleviate some of these issues (Hasselt, 2010; Mnih et al.,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "2015; Schulman et al., 2015; 2016), these recent advances still provide only a partial solution to the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 693, + 290, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 290, + 705 + ], + "score": 1.0, + "content": "instability and sample complexity challenges.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 504, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 504, + 723 + ], + "score": 1.0, + "content": "Model-free reinforcement learning consists of on- and off-policy methods. Monte Carlo policy gra-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "dient methods (Peters & Schaal, 2006; Schulman et al., 2015) are popular on-policy methods that", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 76, + 506, + 99 + ], + "spans": [ + { + "bbox": [ + 105, + 76, + 506, + 99 + ], + "score": 1.0, + "content": "Q-PROP: SAMPLE-EFFICIENT POLICY GRADIENT", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 97, + 331, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 331, + 118 + ], + "score": 1.0, + "content": "WITH AN OFF-POLICY CRITIC", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 134, + 522, + 147 + ], + "lines": [ + { + "bbox": [ + 111, + 132, + 525, + 150 + ], + "spans": [ + { + "bbox": [ + 111, + 132, + 151, + 150 + ], + "score": 1.0, + "content": "Shixiang", + "type": "text" + }, + { + "bbox": [ + 152, + 134, + 177, + 146 + ], + "score": 0.83, + "content": "\\mathbf { G u } ^ { 1 2 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 132, + 525, + 150 + ], + "score": 1.0, + "content": ", Timothy Lillicrap4, Zoubin Ghahramani16, Richard E. Turner1, Sergey Levine35", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 111, + 132, + 525, + 150 + ] + }, + { + "type": "list", + "bbox": [ + 112, + 147, + 457, + 239 + ], + "lines": [ + { + "bbox": [ + 112, + 146, + 460, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 460, + 159 + ], + "score": 1.0, + "content": "sg717@cam.ac.uk,countzero@google.com,zoubin@eng.cam.ac.uk,", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 158, + 365, + 169 + ], + "spans": [ + { + "bbox": [ + 112, + 158, + 365, + 169 + ], + "score": 1.0, + "content": "ret26@cam.ac.uk,svlevine@eecs.berkeley.edu", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 167, + 239, + 181 + ], + "spans": [ + { + "bbox": [ + 111, + 167, + 239, + 181 + ], + "score": 1.0, + "content": "1University of Cambridge, UK", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 178, + 380, + 194 + ], + "spans": [ + { + "bbox": [ + 111, + 178, + 380, + 194 + ], + "score": 1.0, + "content": "2Max Planck Institute for Intelligent Systems, Tubingen, Germany ¨", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 190, + 199, + 206 + ], + "spans": [ + { + "bbox": [ + 111, + 190, + 199, + 206 + ], + "score": 1.0, + "content": "3Google Brain, USA", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 202, + 182, + 218 + ], + "spans": [ + { + "bbox": [ + 111, + 202, + 182, + 218 + ], + "score": 1.0, + "content": "4DeepMind, UK", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 214, + 196, + 230 + ], + "spans": [ + { + "bbox": [ + 111, + 214, + 196, + 230 + ], + "score": 1.0, + "content": "5UC Berkeley, USA", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 226, + 199, + 242 + ], + "spans": [ + { + "bbox": [ + 111, + 226, + 199, + 242 + ], + "score": 1.0, + "content": "6Uber AI Labs, USA", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + } + ], + "index": 6.5, + "bbox_fs": [ + 111, + 146, + 460, + 242 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 269, + 333, + 281 + ], + "lines": [ + { + "bbox": [ + 276, + 268, + 335, + 282 + ], + "spans": [ + { + "bbox": [ + 276, + 268, + 335, + 282 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 143, + 294, + 468, + 491 + ], + "lines": [ + { + "bbox": [ + 142, + 294, + 470, + 306 + ], + "spans": [ + { + "bbox": [ + 142, + 294, + 470, + 306 + ], + "score": 1.0, + "content": "Model-free deep reinforcement learning (RL) methods have been successful in a", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 305, + 470, + 317 + ], + "spans": [ + { + "bbox": [ + 142, + 305, + 470, + 317 + ], + "score": 1.0, + "content": "wide variety of simulated domains. However, a major obstacle facing deep RL", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 316, + 469, + 328 + ], + "spans": [ + { + "bbox": [ + 141, + 316, + 469, + 328 + ], + "score": 1.0, + "content": "in the real world is their high sample complexity. Batch policy gradient methods", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 325, + 470, + 341 + ], + "spans": [ + { + "bbox": [ + 141, + 325, + 470, + 341 + ], + "score": 1.0, + "content": "offer stable learning, but at the cost of high variance, which often requires large", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 338, + 470, + 351 + ], + "spans": [ + { + "bbox": [ + 141, + 338, + 470, + 351 + ], + "score": 1.0, + "content": "batches. TD-style methods, such as off-policy actor-critic and Q-learning, are", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 349, + 469, + 362 + ], + "spans": [ + { + "bbox": [ + 141, + 349, + 469, + 362 + ], + "score": 1.0, + "content": "more sample-efficient but biased, and often require costly hyperparameter sweeps", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 361, + 470, + 372 + ], + "spans": [ + { + "bbox": [ + 141, + 361, + 470, + 372 + ], + "score": 1.0, + "content": "to stabilize. In this work, we aim to develop methods that combine the stability of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 371, + 469, + 384 + ], + "spans": [ + { + "bbox": [ + 141, + 371, + 469, + 384 + ], + "score": 1.0, + "content": "policy gradients with the efficiency of off-policy RL. We present Q-Prop, a policy", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 382, + 470, + 394 + ], + "spans": [ + { + "bbox": [ + 141, + 382, + 470, + 394 + ], + "score": 1.0, + "content": "gradient method that uses a Taylor expansion of the off-policy critic as a control", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 393, + 469, + 405 + ], + "spans": [ + { + "bbox": [ + 142, + 393, + 469, + 405 + ], + "score": 1.0, + "content": "variate. Q-Prop is both sample efficient and stable, and effectively combines the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 403, + 470, + 417 + ], + "spans": [ + { + "bbox": [ + 141, + 403, + 470, + 417 + ], + "score": 1.0, + "content": "benefits of on-policy and off-policy methods. We analyze the connection between", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 414, + 469, + 427 + ], + "spans": [ + { + "bbox": [ + 141, + 414, + 469, + 427 + ], + "score": 1.0, + "content": "Q-Prop and existing model-free algorithms, and use control variate theory to de-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 140, + 425, + 470, + 439 + ], + "spans": [ + { + "bbox": [ + 140, + 425, + 470, + 439 + ], + "score": 1.0, + "content": "rive two variants of Q-Prop with conservative and aggressive adaptation. We show", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 437, + 470, + 450 + ], + "spans": [ + { + "bbox": [ + 141, + 437, + 470, + 450 + ], + "score": 1.0, + "content": "that conservative Q-Prop provides substantial gains in sample efficiency over trust", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 448, + 470, + 460 + ], + "spans": [ + { + "bbox": [ + 141, + 448, + 470, + 460 + ], + "score": 1.0, + "content": "region policy optimization (TRPO) with generalized advantage estimation (GAE),", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 459, + 469, + 471 + ], + "spans": [ + { + "bbox": [ + 141, + 459, + 469, + 471 + ], + "score": 1.0, + "content": "and improves stability over deep deterministic policy gradient (DDPG), the state-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 142, + 470, + 469, + 482 + ], + "spans": [ + { + "bbox": [ + 142, + 470, + 469, + 482 + ], + "score": 1.0, + "content": "of-the-art on-policy and off-policy methods, on OpenAI Gym’s MuJoCo continu-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 142, + 482, + 248, + 493 + ], + "spans": [ + { + "bbox": [ + 142, + 482, + 248, + 493 + ], + "score": 1.0, + "content": "ous control environments.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 20.5, + "bbox_fs": [ + 140, + 294, + 470, + 493 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 514, + 206, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 208, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 208, + 529 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 540, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "Model-free reinforcement learning is a promising approach for solving arbitrary goal-directed se-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "quential decision-making problems with only high-level reward signals and no supervision. It has", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "recently been extended to utilize large neural network policies and value functions, and has been", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "shown to be successful in solving a range of difficult problems (Mnih et al., 2015; Schulman et al.,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "2015; Lillicrap et al., 2016; Silver et al., 2016; Gu et al., 2016b; Mnih et al., 2016). Deep neural", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 595, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 607 + ], + "score": 1.0, + "content": "network parametrization minimizes the need for manual feature and policy engineering, and allows", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 604, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 619 + ], + "score": 1.0, + "content": "learning end-to-end policies mapping from high-dimensional inputs, such as images, directly to ac-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "tions. However, such expressive parametrization also introduces a number of practical problems.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 626, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 641 + ], + "score": 1.0, + "content": "Deep reinforcement learning algorithms tend to be sensitive to hyperparameter settings, often re-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "score": 1.0, + "content": "quiring extensive hyperparameter sweeps to find good values. Poor hyperparameter settings tend to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "produce unstable or non-convergent learning. Deep RL algorithms also tend to exhibit high sample", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 660, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 506, + 673 + ], + "score": 1.0, + "content": "complexity, often to the point of being impractical to run on real physical systems. Although a num-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "ber of recent techniques have sought to alleviate some of these issues (Hasselt, 2010; Mnih et al.,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "2015; Schulman et al., 2015; 2016), these recent advances still provide only a partial solution to the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 693, + 290, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 290, + 705 + ], + "score": 1.0, + "content": "instability and sample complexity challenges.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 539, + 506, + 705 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 504, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 504, + 723 + ], + "score": 1.0, + "content": "Model-free reinforcement learning consists of on- and off-policy methods. Monte Carlo policy gra-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "dient methods (Peters & Schaal, 2006; Schulman et al., 2015) are popular on-policy methods that", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "directly maximize the cumulative future returns with respect to the policy. While these algorithms", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 505, + 108 + ], + "score": 1.0, + "content": "can offer unbiased (or nearly unbiased, as discussed in Section 2.1) estimates of the gradient, they", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "rely on Monte Carlo estimation and often suffer from high variance. To cope with high variance", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "gradient estimates and difficult optimization landscapes, a number of techniques have been pro-", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "posed, including constraining the change in the policy at each gradient step (Kakade, 2001; Peters", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 149 + ], + "score": 1.0, + "content": "et al., 2010) and mixing value-based back-ups to trade off bias and variance in Monte Carlo return", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "estimates (Schulman et al., 2015). However, these methods all tend to require very large numbers", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "of samples to deal with the high variance when estimating gradients of high-dimensional neural", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "network policies. The crux of the problem with policy gradient methods is that they can only effec-", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "tively use on-policy samples, which means that they require collecting large amounts of on-policy", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 204 + ], + "score": 1.0, + "content": "experiences after each parameter update to the policy. This makes them very sample intensive. Off-", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "policy methods, such as Q-learning (Watkins & Dayan, 1992; Sutton et al., 1999; Mnih et al., 2015;", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "Gu et al., 2016b) and off-policy actor-critic methods (Lever, 2014; Lillicrap et al., 2016), can in-", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "stead use all samples, including off-policy samples, by adopting temporal difference learning with", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "score": 1.0, + "content": "experience replay. Such methods are much more sample-efficient. However, convergence of these", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 247, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 260 + ], + "score": 1.0, + "content": "algorithms is in general not guaranteed with non-linear function approximators, and practical con-", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 258, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 272 + ], + "score": 1.0, + "content": "vergence and instability issues typically mean that extensive hyperparameter tuning is required to", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 270, + 184, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 184, + 281 + ], + "score": 1.0, + "content": "attain good results.", + "type": "text", + "cross_page": true + } + ], + "index": 17 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 708, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 83, + 505, + 280 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "directly maximize the cumulative future returns with respect to the policy. While these algorithms", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 505, + 108 + ], + "score": 1.0, + "content": "can offer unbiased (or nearly unbiased, as discussed in Section 2.1) estimates of the gradient, they", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "rely on Monte Carlo estimation and often suffer from high variance. To cope with high variance", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "gradient estimates and difficult optimization landscapes, a number of techniques have been pro-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "posed, including constraining the change in the policy at each gradient step (Kakade, 2001; Peters", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 149 + ], + "score": 1.0, + "content": "et al., 2010) and mixing value-based back-ups to trade off bias and variance in Monte Carlo return", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "estimates (Schulman et al., 2015). However, these methods all tend to require very large numbers", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "of samples to deal with the high variance when estimating gradients of high-dimensional neural", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "network policies. The crux of the problem with policy gradient methods is that they can only effec-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "tively use on-policy samples, which means that they require collecting large amounts of on-policy", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 204 + ], + "score": 1.0, + "content": "experiences after each parameter update to the policy. This makes them very sample intensive. Off-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "policy methods, such as Q-learning (Watkins & Dayan, 1992; Sutton et al., 1999; Mnih et al., 2015;", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "Gu et al., 2016b) and off-policy actor-critic methods (Lever, 2014; Lillicrap et al., 2016), can in-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "stead use all samples, including off-policy samples, by adopting temporal difference learning with", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "score": 1.0, + "content": "experience replay. Such methods are much more sample-efficient. However, convergence of these", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 247, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 260 + ], + "score": 1.0, + "content": "algorithms is in general not guaranteed with non-linear function approximators, and practical con-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 258, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 272 + ], + "score": 1.0, + "content": "vergence and instability issues typically mean that extensive hyperparameter tuning is required to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 270, + 184, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 184, + 281 + ], + "score": 1.0, + "content": "attain good results.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 285, + 505, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "score": 1.0, + "content": "In order to make deep reinforcement learning practical as a tool for tackling real-world tasks, we", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 296, + 507, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 507, + 310 + ], + "score": 1.0, + "content": "must develop methods that are both data efficient and stable. In this paper, we propose Q-Prop, a", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "step in this direction that combines the advantages of on-policy policy gradient methods with the effi-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 319, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 505, + 332 + ], + "score": 1.0, + "content": "ciency of off-policy learning. Unlike prior approaches for off-policy learning, which either introduce", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 328, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 506, + 343 + ], + "score": 1.0, + "content": "bias (Sutton et al., 1999; Silver et al., 2014) or increase variance (Precup, 2000; Levine & Koltun,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 505, + 354 + ], + "score": 1.0, + "content": "2013; Munos et al., 2016), Q-Prop can reduce the variance of gradient estimator without adding", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "score": 1.0, + "content": "bias; unlike prior approaches for critic-based variance reduction (Schulman et al., 2016) which fit", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 363, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 505, + 375 + ], + "score": 1.0, + "content": "the value function on-policy, Q-Prop learns the action-value function off-policy. The core idea is", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 373, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 505, + 386 + ], + "score": 1.0, + "content": "to use the first-order Taylor expansion of the critic as a control variate, resulting in an analytical", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "score": 1.0, + "content": "gradient term through the critic and a Monte Carlo policy gradient term consisting of the residuals", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 396, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 408 + ], + "score": 1.0, + "content": "in advantage approximations. The method helps unify policy gradient and actor-critic methods: it", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "can be seen as using the off-policy critic to reduce variance in policy gradient or using on-policy", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "score": 1.0, + "content": "Monte Carlo returns to correct for bias in the critic gradient. We further provide theoretical analy-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "sis of the control variate, and derive two additional variants of Q-Prop. The method can be easily", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "incorporated into any policy gradient algorithm. We show that Q-Prop provides substantial gains", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "score": 1.0, + "content": "in sample efficiency over trust region policy optimization (TRPO) with generalized advantage esti-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 460, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 475 + ], + "score": 1.0, + "content": "mation (GAE) (Schulman et al., 2015; 2016), and improved stability over deep deterministic policy", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 472, + 456, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 456, + 484 + ], + "score": 1.0, + "content": "gradient (DDPG) (Lillicrap et al., 2016) across a repertoire of continuous control tasks.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 26.5 + }, + { + "type": "title", + "bbox": [ + 108, + 501, + 200, + 514 + ], + "lines": [ + { + "bbox": [ + 104, + 500, + 201, + 517 + ], + "spans": [ + { + "bbox": [ + 104, + 500, + 201, + 517 + ], + "score": 1.0, + "content": "2 BACKGROUND", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 527, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 504, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 504, + 541 + ], + "score": 1.0, + "content": "Reinforcement learning (RL) aims to learn a policy for an agent such that it behaves optimally", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 538, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 301, + 553 + ], + "score": 1.0, + "content": "according to a reward function. At a time step", + "type": "text" + }, + { + "bbox": [ + 301, + 541, + 307, + 549 + ], + "score": 0.55, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 538, + 347, + 553 + ], + "score": 1.0, + "content": "and state", + "type": "text" + }, + { + "bbox": [ + 348, + 541, + 357, + 550 + ], + "score": 0.83, + "content": "\\scriptstyle { s _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 538, + 477, + 553 + ], + "score": 1.0, + "content": ", the agent chooses an action", + "type": "text" + }, + { + "bbox": [ + 478, + 541, + 488, + 550 + ], + "score": 0.85, + "content": "\\mathbf { } \\mathbf { a } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 538, + 505, + 553 + ], + "score": 1.0, + "content": "ac-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 548, + 504, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 190, + 565 + ], + "score": 1.0, + "content": "cording to its policy", + "type": "text" + }, + { + "bbox": [ + 190, + 550, + 225, + 561 + ], + "score": 0.92, + "content": "\\pi ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 548, + 485, + 565 + ], + "score": 1.0, + "content": ", the state of the agent and the environment changes to new state", + "type": "text" + }, + { + "bbox": [ + 485, + 552, + 504, + 561 + ], + "score": 0.88, + "content": "\\boldsymbol { s } _ { t + 1 }", + "type": "inline_equation" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 559, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 104, + 559, + 202, + 575 + ], + "score": 1.0, + "content": "according to dynamics", + "type": "text" + }, + { + "bbox": [ + 202, + 561, + 258, + 573 + ], + "score": 0.93, + "content": "p \\big ( \\pmb { s } _ { t + 1 } \\big | \\pmb { s } _ { t } , \\pmb { a } _ { t } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 559, + 378, + 575 + ], + "score": 1.0, + "content": ", the agent receives a reward", + "type": "text" + }, + { + "bbox": [ + 379, + 560, + 413, + 573 + ], + "score": 0.93, + "content": "r ( s _ { t } , a _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 559, + 505, + 575 + ], + "score": 1.0, + "content": ", and the process con-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 571, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 154, + 585 + ], + "score": 1.0, + "content": "tinues. Let", + "type": "text" + }, + { + "bbox": [ + 154, + 572, + 164, + 583 + ], + "score": 0.83, + "content": "R _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 571, + 204, + 585 + ], + "score": 1.0, + "content": "denote a", + "type": "text" + }, + { + "bbox": [ + 204, + 574, + 211, + 583 + ], + "score": 0.61, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 571, + 354, + 585 + ], + "score": 1.0, + "content": "-discounted cumulative return from", + "type": "text" + }, + { + "bbox": [ + 354, + 573, + 360, + 582 + ], + "score": 0.59, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 571, + 505, + 585 + ], + "score": 1.0, + "content": "for an infinite horizon problem, i.e", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 582, + 507, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 206, + 596 + ], + "score": 0.91, + "content": "\\begin{array} { r } { R _ { t } = \\sum _ { t ^ { \\prime } = t } ^ { \\infty } \\gamma ^ { t ^ { \\prime } - t } r ( s _ { t ^ { \\prime } } , { \\bf a } _ { t ^ { \\prime } } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 582, + 507, + 599 + ], + "score": 1.0, + "content": ". The goal of reinforcement learning is to maximize the expected return", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 595, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 170, + 608 + ], + "score": 0.9, + "content": "J ( \\theta ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ R _ { 0 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 595, + 506, + 609 + ], + "score": 1.0, + "content": "with respect to the policy parameters θ . In this section, we review several standard", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 606, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 619 + ], + "score": 1.0, + "content": "techniques for performing this optimization, and in the next section, we will discuss our proposed", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "Q-Prop algorithm that combines the strengths of these approaches to achieve efficient, stable RL.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 627, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 104, + 627, + 506, + 642 + ], + "score": 1.0, + "content": "Monte Carlo policy gradient refers to policy gradient methods that use full Monte Carlo returns,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 103, + 639, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 103, + 639, + 506, + 653 + ], + "score": 1.0, + "content": "e.g. REINFORCE (Williams, 1992) and TRPO (Schulman et al., 2015), and policy gradient with", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 649, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 104, + 649, + 505, + 664 + ], + "score": 1.0, + "content": "function approximation refers to actor-critic methods (Sutton et al., 1999) which optimize the policy", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 662, + 472, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 472, + 674 + ], + "score": 1.0, + "content": "against a critic, e.g. deterministic policy gradient (Silver et al., 2014; Lillicrap et al., 2016).", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 43 + }, + { + "type": "title", + "bbox": [ + 107, + 688, + 328, + 699 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 329, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 329, + 701 + ], + "score": 1.0, + "content": "2.1 MONTE CARLO POLICY GRADIENT METHODS", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 50 + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Monte Carlo policy gradient methods apply direct gradient-based optimization to the reinforcement", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 343, + 734 + ], + "score": 1.0, + "content": "learning objective. This involves directly differentiating the", + "type": "text" + }, + { + "bbox": [ + 344, + 721, + 364, + 732 + ], + "score": 0.89, + "content": "J ( \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 719, + 505, + 734 + ], + "score": 1.0, + "content": "objective with respect to the policy", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 51.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 83, + 505, + 280 + ], + "lines": [], + "index": 8.5, + "bbox_fs": [ + 105, + 82, + 506, + 281 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 285, + 505, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "score": 1.0, + "content": "In order to make deep reinforcement learning practical as a tool for tackling real-world tasks, we", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 296, + 507, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 507, + 310 + ], + "score": 1.0, + "content": "must develop methods that are both data efficient and stable. In this paper, we propose Q-Prop, a", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "step in this direction that combines the advantages of on-policy policy gradient methods with the effi-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 319, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 505, + 332 + ], + "score": 1.0, + "content": "ciency of off-policy learning. Unlike prior approaches for off-policy learning, which either introduce", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 328, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 506, + 343 + ], + "score": 1.0, + "content": "bias (Sutton et al., 1999; Silver et al., 2014) or increase variance (Precup, 2000; Levine & Koltun,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 505, + 354 + ], + "score": 1.0, + "content": "2013; Munos et al., 2016), Q-Prop can reduce the variance of gradient estimator without adding", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "score": 1.0, + "content": "bias; unlike prior approaches for critic-based variance reduction (Schulman et al., 2016) which fit", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 363, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 505, + 375 + ], + "score": 1.0, + "content": "the value function on-policy, Q-Prop learns the action-value function off-policy. The core idea is", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 373, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 505, + 386 + ], + "score": 1.0, + "content": "to use the first-order Taylor expansion of the critic as a control variate, resulting in an analytical", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "score": 1.0, + "content": "gradient term through the critic and a Monte Carlo policy gradient term consisting of the residuals", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 396, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 408 + ], + "score": 1.0, + "content": "in advantage approximations. The method helps unify policy gradient and actor-critic methods: it", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "can be seen as using the off-policy critic to reduce variance in policy gradient or using on-policy", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "score": 1.0, + "content": "Monte Carlo returns to correct for bias in the critic gradient. We further provide theoretical analy-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "sis of the control variate, and derive two additional variants of Q-Prop. The method can be easily", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "incorporated into any policy gradient algorithm. We show that Q-Prop provides substantial gains", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "score": 1.0, + "content": "in sample efficiency over trust region policy optimization (TRPO) with generalized advantage esti-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 460, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 475 + ], + "score": 1.0, + "content": "mation (GAE) (Schulman et al., 2015; 2016), and improved stability over deep deterministic policy", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 472, + 456, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 456, + 484 + ], + "score": 1.0, + "content": "gradient (DDPG) (Lillicrap et al., 2016) across a repertoire of continuous control tasks.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 285, + 507, + 484 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 501, + 200, + 514 + ], + "lines": [ + { + "bbox": [ + 104, + 500, + 201, + 517 + ], + "spans": [ + { + "bbox": [ + 104, + 500, + 201, + 517 + ], + "score": 1.0, + "content": "2 BACKGROUND", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 527, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 504, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 504, + 541 + ], + "score": 1.0, + "content": "Reinforcement learning (RL) aims to learn a policy for an agent such that it behaves optimally", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 538, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 301, + 553 + ], + "score": 1.0, + "content": "according to a reward function. At a time step", + "type": "text" + }, + { + "bbox": [ + 301, + 541, + 307, + 549 + ], + "score": 0.55, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 538, + 347, + 553 + ], + "score": 1.0, + "content": "and state", + "type": "text" + }, + { + "bbox": [ + 348, + 541, + 357, + 550 + ], + "score": 0.83, + "content": "\\scriptstyle { s _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 538, + 477, + 553 + ], + "score": 1.0, + "content": ", the agent chooses an action", + "type": "text" + }, + { + "bbox": [ + 478, + 541, + 488, + 550 + ], + "score": 0.85, + "content": "\\mathbf { } \\mathbf { a } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 538, + 505, + 553 + ], + "score": 1.0, + "content": "ac-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 548, + 504, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 190, + 565 + ], + "score": 1.0, + "content": "cording to its policy", + "type": "text" + }, + { + "bbox": [ + 190, + 550, + 225, + 561 + ], + "score": 0.92, + "content": "\\pi ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 548, + 485, + 565 + ], + "score": 1.0, + "content": ", the state of the agent and the environment changes to new state", + "type": "text" + }, + { + "bbox": [ + 485, + 552, + 504, + 561 + ], + "score": 0.88, + "content": "\\boldsymbol { s } _ { t + 1 }", + "type": "inline_equation" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 559, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 104, + 559, + 202, + 575 + ], + "score": 1.0, + "content": "according to dynamics", + "type": "text" + }, + { + "bbox": [ + 202, + 561, + 258, + 573 + ], + "score": 0.93, + "content": "p \\big ( \\pmb { s } _ { t + 1 } \\big | \\pmb { s } _ { t } , \\pmb { a } _ { t } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 559, + 378, + 575 + ], + "score": 1.0, + "content": ", the agent receives a reward", + "type": "text" + }, + { + "bbox": [ + 379, + 560, + 413, + 573 + ], + "score": 0.93, + "content": "r ( s _ { t } , a _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 559, + 505, + 575 + ], + "score": 1.0, + "content": ", and the process con-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 571, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 154, + 585 + ], + "score": 1.0, + "content": "tinues. Let", + "type": "text" + }, + { + "bbox": [ + 154, + 572, + 164, + 583 + ], + "score": 0.83, + "content": "R _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 571, + 204, + 585 + ], + "score": 1.0, + "content": "denote a", + "type": "text" + }, + { + "bbox": [ + 204, + 574, + 211, + 583 + ], + "score": 0.61, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 571, + 354, + 585 + ], + "score": 1.0, + "content": "-discounted cumulative return from", + "type": "text" + }, + { + "bbox": [ + 354, + 573, + 360, + 582 + ], + "score": 0.59, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 571, + 505, + 585 + ], + "score": 1.0, + "content": "for an infinite horizon problem, i.e", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 582, + 507, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 206, + 596 + ], + "score": 0.91, + "content": "\\begin{array} { r } { R _ { t } = \\sum _ { t ^ { \\prime } = t } ^ { \\infty } \\gamma ^ { t ^ { \\prime } - t } r ( s _ { t ^ { \\prime } } , { \\bf a } _ { t ^ { \\prime } } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 582, + 507, + 599 + ], + "score": 1.0, + "content": ". The goal of reinforcement learning is to maximize the expected return", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 595, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 170, + 608 + ], + "score": 0.9, + "content": "J ( \\theta ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ R _ { 0 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 595, + 506, + 609 + ], + "score": 1.0, + "content": "with respect to the policy parameters θ . In this section, we review several standard", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 606, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 619 + ], + "score": 1.0, + "content": "techniques for performing this optimization, and in the next section, we will discuss our proposed", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "Q-Prop algorithm that combines the strengths of these approaches to achieve efficient, stable RL.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 627, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 104, + 627, + 506, + 642 + ], + "score": 1.0, + "content": "Monte Carlo policy gradient refers to policy gradient methods that use full Monte Carlo returns,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 103, + 639, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 103, + 639, + 506, + 653 + ], + "score": 1.0, + "content": "e.g. REINFORCE (Williams, 1992) and TRPO (Schulman et al., 2015), and policy gradient with", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 649, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 104, + 649, + 505, + 664 + ], + "score": 1.0, + "content": "function approximation refers to actor-critic methods (Sutton et al., 1999) which optimize the policy", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 662, + 472, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 472, + 674 + ], + "score": 1.0, + "content": "against a critic, e.g. deterministic policy gradient (Silver et al., 2014; Lillicrap et al., 2016).", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 43, + "bbox_fs": [ + 103, + 528, + 507, + 674 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 688, + 328, + 699 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 329, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 329, + 701 + ], + "score": 1.0, + "content": "2.1 MONTE CARLO POLICY GRADIENT METHODS", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 50 + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Monte Carlo policy gradient methods apply direct gradient-based optimization to the reinforcement", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 343, + 734 + ], + "score": 1.0, + "content": "learning objective. This involves directly differentiating the", + "type": "text" + }, + { + "bbox": [ + 344, + 721, + 364, + 732 + ], + "score": 0.89, + "content": "J ( \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 719, + 505, + 734 + ], + "score": 1.0, + "content": "objective with respect to the policy", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 51.5, + "bbox_fs": [ + 105, + 709, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 81, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 153, + 95 + ], + "score": 1.0, + "content": "parameters", + "type": "text" + }, + { + "bbox": [ + 153, + 83, + 160, + 92 + ], + "score": 0.46, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 82, + 506, + 95 + ], + "score": 1.0, + "content": ". The standard form, known as the REINFORCE algorithm (Williams, 1992), is shown", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 137, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 137, + 107 + ], + "score": 1.0, + "content": "below:", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 105, + 468, + 135 + ], + "lines": [ + { + "bbox": [ + 142, + 105, + 468, + 135 + ], + "spans": [ + { + "bbox": [ + 142, + 105, + 468, + 135 + ], + "score": 0.92, + "content": "\\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\pi } [ \\sum _ { t = 0 } ^ { \\infty } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\gamma ^ { t } R _ { t } ] = \\mathbb { E } _ { \\pi } [ \\sum _ { t = 0 } ^ { \\infty } \\gamma ^ { t } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( R _ { t } - b ( s _ { t } ) ) ] ,", + "type": "interline_equation", + "image_path": "f6942cdf9ba69600fd352e5a9c3209b8ebb10c765cf4d7646f59830914b86a8c.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 142, + 105, + 468, + 115.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 142, + 115.0, + 468, + 125.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 142, + 125.0, + 468, + 135.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 137, + 507, + 161 + ], + "lines": [ + { + "bbox": [ + 106, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 133, + 150 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 137, + 155, + 150 + ], + "score": 0.92, + "content": "b ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "is known as the baseline. For convenience of later derivations, Eq. 1 can also be written", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 146, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 104, + 146, + 172, + 163 + ], + "score": 1.0, + "content": "as below, where", + "type": "text" + }, + { + "bbox": [ + 172, + 148, + 275, + 161 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\rho _ { \\pi } ( s ) = \\sum _ { t = 0 } ^ { \\infty } \\gamma ^ { t } p ( s _ { t } = s ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 146, + 506, + 163 + ], + "score": 1.0, + "content": "is the unnormalized discounted state visitation frequency,", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 174, + 425, + 189 + ], + "lines": [ + { + "bbox": [ + 185, + 174, + 425, + 189 + ], + "spans": [ + { + "bbox": [ + 185, + 174, + 425, + 189 + ], + "score": 0.88, + "content": "\\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\pi } ( . ) , \\mathbf { a } _ { t } \\sim \\pi ( \\cdot | s _ { t } ) } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( R _ { t } - b ( s _ { t } ) ) ] .", + "type": "interline_equation", + "image_path": "192f775e82ad0d09644c328424abf64f7014ba3648e4e893604deaff23956ce3.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 185, + 174, + 425, + 189 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 192, + 505, + 248 + ], + "lines": [ + { + "bbox": [ + 106, + 193, + 504, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 504, + 205 + ], + "score": 1.0, + "content": "Eq. 2 is an unbiased gradient of the RL objective. However, in practice, most policy gradient meth-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 203, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 376, + 217 + ], + "score": 1.0, + "content": "ods effectively use undiscounted state visitation frequencies, i.e.", + "type": "text" + }, + { + "bbox": [ + 376, + 204, + 401, + 216 + ], + "score": 0.9, + "content": "\\gamma = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 203, + 471, + 217 + ], + "score": 1.0, + "content": "in the equal for", + "type": "text" + }, + { + "bbox": [ + 471, + 205, + 483, + 215 + ], + "score": 0.87, + "content": "\\rho _ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 203, + 505, + 217 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 214, + 504, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 504, + 226 + ], + "score": 1.0, + "content": "are therefore biased; in fact, making them unbiased often hurts performance (Thomas, 2014). In", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "this paper, we mainly discuss bias due to function approximation, off-policy learning, and value", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 235, + 149, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 149, + 251 + ], + "score": 1.0, + "content": "back-ups.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 108, + 253, + 505, + 309 + ], + "lines": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "The gradient is estimated using Monte Carlo samples in practice and has very high variance. A", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 264, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 279 + ], + "score": 1.0, + "content": "proper choice of baseline is necessary to reduce the variance sufficiently such that learning becomes", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 276, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 399, + 288 + ], + "score": 1.0, + "content": "feasible. A common choice is to estimate the value function of the state", + "type": "text" + }, + { + "bbox": [ + 399, + 276, + 427, + 288 + ], + "score": 0.93, + "content": "V _ { \\pi } ( \\pmb { s } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 276, + 505, + 288 + ], + "score": 1.0, + "content": "to use as the base-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 327, + 299 + ], + "score": 1.0, + "content": "line, which provides an estimate of advantage function", + "type": "text" + }, + { + "bbox": [ + 328, + 286, + 369, + 299 + ], + "score": 0.93, + "content": "A _ { \\pi } ( s _ { t } , { \\pmb a } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 286, + 505, + 299 + ], + "score": 1.0, + "content": ", which is a centered action-value", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 298, + 260, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 142, + 309 + ], + "score": 1.0, + "content": "function", + "type": "text" + }, + { + "bbox": [ + 142, + 298, + 185, + 309 + ], + "score": 0.94, + "content": "Q _ { \\pi } ( s _ { t } , \\pmb { a } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 298, + 260, + 309 + ], + "score": 1.0, + "content": ", as defined below:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 159, + 311, + 452, + 358 + ], + "lines": [ + { + "bbox": [ + 159, + 311, + 452, + 358 + ], + "spans": [ + { + "bbox": [ + 159, + 311, + 452, + 358 + ], + "score": 0.93, + "content": "\\begin{array} { c } { { V _ { \\pi } ( s _ { t } ) = \\mathbb { E } _ { \\pi } [ R _ { t } ] = \\mathbb { E } _ { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } [ Q _ { \\pi } ( s _ { t } , a _ { t } ) ] } } \\\\ { { Q _ { \\pi } ( s _ { t } , a _ { t } ) = r ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { E } _ { \\pi } [ R _ { t + 1 } ] = r ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { E } _ { p ( s _ { t + 1 } | s _ { t } , a _ { t } ) } [ V _ { \\pi } ( s _ { t + 1 } ) ] } } \\\\ { { A _ { \\pi } ( s _ { t } , a _ { t } ) = Q _ { \\pi } ( s _ { t } , a _ { t } ) - V _ { \\pi } ( s _ { t } ) . } } \\end{array}", + "type": "interline_equation", + "image_path": "9915c7f590b427849f27b3abc7fecc5232c681e28a5cfc2aab1cb8d4818fdf0b.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 159, + 311, + 452, + 326.6666666666667 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 159, + 326.6666666666667, + 452, + 342.33333333333337 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 159, + 342.33333333333337, + 452, + 358.00000000000006 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 109, + 360, + 503, + 405 + ], + "lines": [ + { + "bbox": [ + 106, + 360, + 504, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 149, + 372 + ], + "score": 0.92, + "content": "Q _ { \\pi } ( s _ { t } , \\pmb { a } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 360, + 496, + 373 + ], + "score": 1.0, + "content": "summarizes the performance of each action from a given state, assuming it follows", + "type": "text" + }, + { + "bbox": [ + 497, + 362, + 504, + 370 + ], + "score": 0.68, + "content": "\\pi", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 167, + 384 + ], + "score": 1.0, + "content": "thereafter, and", + "type": "text" + }, + { + "bbox": [ + 167, + 371, + 209, + 383 + ], + "score": 0.93, + "content": "A _ { \\pi } ( s _ { t } , { \\pmb a } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 370, + 505, + 384 + ], + "score": 1.0, + "content": "provides a measure of how each action compares to the average perfor-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 382, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 106, + 382, + 183, + 395 + ], + "score": 1.0, + "content": "mance at the state", + "type": "text" + }, + { + "bbox": [ + 183, + 384, + 192, + 393 + ], + "score": 0.85, + "content": "\\scriptstyle { s _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 382, + 273, + 395 + ], + "score": 1.0, + "content": ", which is given by", + "type": "text" + }, + { + "bbox": [ + 273, + 382, + 300, + 394 + ], + "score": 0.93, + "content": "V _ { \\pi } ( \\pmb { s } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 382, + 334, + 395 + ], + "score": 1.0, + "content": ". Using", + "type": "text" + }, + { + "bbox": [ + 334, + 382, + 376, + 394 + ], + "score": 0.93, + "content": "A _ { \\pi } ( s _ { t } , { \\pmb a } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 382, + 505, + 395 + ], + "score": 1.0, + "content": "centers the learning signal and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 394, + 229, + 406 + ], + "spans": [ + { + "bbox": [ + 107, + 394, + 229, + 406 + ], + "score": 1.0, + "content": "reduces variance significantly.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 409, + 505, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 423 + ], + "score": 1.0, + "content": "Besides high variance, another problem with the policy gradient is that it requires on-policy samples.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 420, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 505, + 434 + ], + "score": 1.0, + "content": "This makes policy gradient optimization very sample intensive. To achieve similar sample efficiency", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 431, + 506, + 445 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 506, + 445 + ], + "score": 1.0, + "content": "as off-policy methods, we can attempt to include off-policy data. Prior attempts use importance", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 442, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 456 + ], + "score": 1.0, + "content": "sampling to include off-policy trajectories; however, these are known to be difficult scale to high-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 453, + 487, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 487, + 466 + ], + "score": 1.0, + "content": "dimensional action spaces because of rapidly degenerating importance weights (Precup, 2000).", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 107, + 478, + 361, + 489 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 362, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 362, + 491 + ], + "score": 1.0, + "content": "2.2 POLICY GRADIENT WITH FUNCTION APPROXIMATION", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 498, + 505, + 576 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 504, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 504, + 511 + ], + "score": 1.0, + "content": "Policy gradient methods with function approximation (Sutton et al., 1999), or actor-critic methods,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "score": 1.0, + "content": "include a policy evaluation step, which often uses temporal difference (TD) learning to fit a critic", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 521, + 504, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 120, + 532 + ], + "score": 0.88, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 121, + 521, + 209, + 533 + ], + "score": 1.0, + "content": "for the current policy", + "type": "text" + }, + { + "bbox": [ + 210, + 521, + 230, + 532 + ], + "score": 0.91, + "content": "\\pi ( \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 521, + 504, + 533 + ], + "score": 1.0, + "content": ", and a policy improvement step which greedily optimizes the policy", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 107, + 531, + 504, + 545 + ], + "spans": [ + { + "bbox": [ + 107, + 533, + 114, + 542 + ], + "score": 0.7, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 531, + 218, + 545 + ], + "score": 1.0, + "content": "against the critic estimate", + "type": "text" + }, + { + "bbox": [ + 219, + 533, + 232, + 543 + ], + "score": 0.88, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 531, + 504, + 545 + ], + "score": 1.0, + "content": ". Significant gains in sample efficiency may be achievable using off-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 543, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 555 + ], + "score": 1.0, + "content": "policy TD learning for the critic, as in Q-learning and deterministic policy gradient (Sutton, 1990;", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 553, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 567 + ], + "score": 1.0, + "content": "Silver et al., 2014), typically by means of experience replay for training deep Q networks (Mnih", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 565, + 314, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 314, + 577 + ], + "score": 1.0, + "content": "et al., 2015; Lillicrap et al., 2016; Gu et al., 2016b).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "score": 1.0, + "content": "One particularly relevant example of such a method is the deep deterministic policy gradient", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 591, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 605 + ], + "score": 1.0, + "content": "(DDPG) (Silver et al., 2014; Lillicrap et al., 2016). The updates for this method are given below,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 603, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 133, + 617 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 603, + 249, + 615 + ], + "score": 0.91, + "content": "\\pi _ { \\boldsymbol { \\theta } } ( \\mathbf { a } _ { t } | \\mathbf { s } _ { t } ) = \\delta ( \\mathbf { a } _ { t } = \\mu _ { \\boldsymbol { \\theta } } ( \\mathbf { s } _ { t } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 603, + 350, + 617 + ], + "score": 1.0, + "content": "is a deterministic policy,", + "type": "text" + }, + { + "bbox": [ + 351, + 604, + 358, + 615 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 603, + 505, + 617 + ], + "score": 1.0, + "content": "is arbitrary exploration distribution,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 614, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 123, + 627 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 616, + 135, + 627 + ], + "score": 0.87, + "content": "\\rho _ { \\beta }", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 614, + 321, + 627 + ], + "score": 1.0, + "content": "corresponds to sampling from a replay buffer.", + "type": "text" + }, + { + "bbox": [ + 322, + 614, + 347, + 626 + ], + "score": 0.92, + "content": "Q ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 614, + 505, + 627 + ], + "score": 1.0, + "content": "is the target network that slowly tracks", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 107, + 625, + 214, + 639 + ], + "spans": [ + { + "bbox": [ + 107, + 626, + 120, + 637 + ], + "score": 0.86, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 625, + 214, + 639 + ], + "score": 1.0, + "content": "(Lillicrap et al., 2016).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + }, + { + "type": "interline_equation", + "bbox": [ + 147, + 639, + 464, + 679 + ], + "lines": [ + { + "bbox": [ + 147, + 639, + 464, + 679 + ], + "spans": [ + { + "bbox": [ + 147, + 639, + 464, + 679 + ], + "score": 0.9, + "content": "\\begin{array} { r l } & { w = \\underset { w } { \\arg \\operatorname* { m i n } } \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\beta } ( \\cdot ) , a _ { t } \\sim \\beta ( \\cdot | s _ { t } ) } \\big [ \\big ( r \\big ( s _ { t } , a _ { t } \\big ) + \\gamma Q ( s _ { t + 1 } , \\mu _ { \\theta } ( s _ { t + 1 } ) \\big ) - Q _ { w } ( s _ { t } , a _ { t } ) \\big ) ^ { 2 } \\big ] } \\\\ & { \\theta = \\underset { \\theta } { \\arg \\operatorname* { m a x } } \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\beta } ( \\cdot ) } [ Q _ { w } \\big ( s _ { t } , \\mu _ { \\theta } ( s _ { t } ) \\big ) ] } \\end{array}", + "type": "interline_equation", + "image_path": "a249afcd4ae52b44b8cd04ef8cc47074157d56f3ad0e07af8bf36242d14bdcf1.jpg" + } + ] + } + ], + "index": 44, + "virtual_lines": [ + { + "bbox": [ + 147, + 639, + 464, + 652.3333333333334 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 147, + 652.3333333333334, + 464, + 665.6666666666667 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 147, + 665.6666666666667, + 464, + 679.0000000000001 + ], + "spans": [], + "index": 45 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 680, + 503, + 714 + ], + "lines": [ + { + "bbox": [ + 105, + 679, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 505, + 694 + ], + "score": 1.0, + "content": "When the critic and policy are parametrized with neural networks, full optimization is expensive,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 691, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 505, + 704 + ], + "score": 1.0, + "content": "and instead stochastic gradient optimization is used. The gradient in the policy improvement phase", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 703, + 348, + 715 + ], + "spans": [ + { + "bbox": [ + 105, + 703, + 324, + 715 + ], + "score": 1.0, + "content": "is given below, which is generally a biased gradient of", + "type": "text" + }, + { + "bbox": [ + 325, + 703, + 344, + 714 + ], + "score": 0.92, + "content": "J ( \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 703, + 348, + 715 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47 + }, + { + "type": "interline_equation", + "bbox": [ + 198, + 717, + 413, + 733 + ], + "lines": [ + { + "bbox": [ + 198, + 717, + 413, + 733 + ], + "spans": [ + { + "bbox": [ + 198, + 717, + 413, + 733 + ], + "score": 0.9, + "content": "\\nabla _ { \\theta } J ( \\theta ) \\approx \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\beta } ( \\cdot ) } [ \\nabla _ { a } Q _ { w } \\big ( s _ { t } , a \\big ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } \\big ( s _ { t } \\big ) ]", + "type": "interline_equation", + "image_path": "3559dbdfcd748188a430af44d6b28d871d9806f9929c6b1b627837aaa3042cb4.jpg" + } + ] + } + ], + "index": 49, + "virtual_lines": [ + { + "bbox": [ + 198, + 717, + 413, + 733 + ], + "spans": [], + "index": 49 + } + ] + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 81, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 153, + 95 + ], + "score": 1.0, + "content": "parameters", + "type": "text" + }, + { + "bbox": [ + 153, + 83, + 160, + 92 + ], + "score": 0.46, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 82, + 506, + 95 + ], + "score": 1.0, + "content": ". The standard form, known as the REINFORCE algorithm (Williams, 1992), is shown", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 137, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 137, + 107 + ], + "score": 1.0, + "content": "below:", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 506, + 107 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 105, + 468, + 135 + ], + "lines": [ + { + "bbox": [ + 142, + 105, + 468, + 135 + ], + "spans": [ + { + "bbox": [ + 142, + 105, + 468, + 135 + ], + "score": 0.92, + "content": "\\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\pi } [ \\sum _ { t = 0 } ^ { \\infty } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\gamma ^ { t } R _ { t } ] = \\mathbb { E } _ { \\pi } [ \\sum _ { t = 0 } ^ { \\infty } \\gamma ^ { t } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( R _ { t } - b ( s _ { t } ) ) ] ,", + "type": "interline_equation", + "image_path": "f6942cdf9ba69600fd352e5a9c3209b8ebb10c765cf4d7646f59830914b86a8c.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 142, + 105, + 468, + 115.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 142, + 115.0, + 468, + 125.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 142, + 125.0, + 468, + 135.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 137, + 507, + 161 + ], + "lines": [ + { + "bbox": [ + 106, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 133, + 150 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 137, + 155, + 150 + ], + "score": 0.92, + "content": "b ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "is known as the baseline. For convenience of later derivations, Eq. 1 can also be written", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 146, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 104, + 146, + 172, + 163 + ], + "score": 1.0, + "content": "as below, where", + "type": "text" + }, + { + "bbox": [ + 172, + 148, + 275, + 161 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\rho _ { \\pi } ( s ) = \\sum _ { t = 0 } ^ { \\infty } \\gamma ^ { t } p ( s _ { t } = s ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 146, + 506, + 163 + ], + "score": 1.0, + "content": "is the unnormalized discounted state visitation frequency,", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 104, + 137, + 506, + 163 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 174, + 425, + 189 + ], + "lines": [ + { + "bbox": [ + 185, + 174, + 425, + 189 + ], + "spans": [ + { + "bbox": [ + 185, + 174, + 425, + 189 + ], + "score": 0.88, + "content": "\\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\pi } ( . ) , \\mathbf { a } _ { t } \\sim \\pi ( \\cdot | s _ { t } ) } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( R _ { t } - b ( s _ { t } ) ) ] .", + "type": "interline_equation", + "image_path": "192f775e82ad0d09644c328424abf64f7014ba3648e4e893604deaff23956ce3.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 185, + 174, + 425, + 189 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 192, + 505, + 248 + ], + "lines": [ + { + "bbox": [ + 106, + 193, + 504, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 504, + 205 + ], + "score": 1.0, + "content": "Eq. 2 is an unbiased gradient of the RL objective. However, in practice, most policy gradient meth-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 203, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 376, + 217 + ], + "score": 1.0, + "content": "ods effectively use undiscounted state visitation frequencies, i.e.", + "type": "text" + }, + { + "bbox": [ + 376, + 204, + 401, + 216 + ], + "score": 0.9, + "content": "\\gamma = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 203, + 471, + 217 + ], + "score": 1.0, + "content": "in the equal for", + "type": "text" + }, + { + "bbox": [ + 471, + 205, + 483, + 215 + ], + "score": 0.87, + "content": "\\rho _ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 203, + 505, + 217 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 214, + 504, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 504, + 226 + ], + "score": 1.0, + "content": "are therefore biased; in fact, making them unbiased often hurts performance (Thomas, 2014). In", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "this paper, we mainly discuss bias due to function approximation, off-policy learning, and value", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 235, + 149, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 149, + 251 + ], + "score": 1.0, + "content": "back-ups.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 193, + 505, + 251 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 253, + 505, + 309 + ], + "lines": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "The gradient is estimated using Monte Carlo samples in practice and has very high variance. A", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 264, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 279 + ], + "score": 1.0, + "content": "proper choice of baseline is necessary to reduce the variance sufficiently such that learning becomes", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 276, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 399, + 288 + ], + "score": 1.0, + "content": "feasible. A common choice is to estimate the value function of the state", + "type": "text" + }, + { + "bbox": [ + 399, + 276, + 427, + 288 + ], + "score": 0.93, + "content": "V _ { \\pi } ( \\pmb { s } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 276, + 505, + 288 + ], + "score": 1.0, + "content": "to use as the base-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 327, + 299 + ], + "score": 1.0, + "content": "line, which provides an estimate of advantage function", + "type": "text" + }, + { + "bbox": [ + 328, + 286, + 369, + 299 + ], + "score": 0.93, + "content": "A _ { \\pi } ( s _ { t } , { \\pmb a } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 286, + 505, + 299 + ], + "score": 1.0, + "content": ", which is a centered action-value", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 298, + 260, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 142, + 309 + ], + "score": 1.0, + "content": "function", + "type": "text" + }, + { + "bbox": [ + 142, + 298, + 185, + 309 + ], + "score": 0.94, + "content": "Q _ { \\pi } ( s _ { t } , \\pmb { a } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 298, + 260, + 309 + ], + "score": 1.0, + "content": ", as defined below:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 254, + 506, + 309 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 159, + 311, + 452, + 358 + ], + "lines": [ + { + "bbox": [ + 159, + 311, + 452, + 358 + ], + "spans": [ + { + "bbox": [ + 159, + 311, + 452, + 358 + ], + "score": 0.93, + "content": "\\begin{array} { c } { { V _ { \\pi } ( s _ { t } ) = \\mathbb { E } _ { \\pi } [ R _ { t } ] = \\mathbb { E } _ { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } [ Q _ { \\pi } ( s _ { t } , a _ { t } ) ] } } \\\\ { { Q _ { \\pi } ( s _ { t } , a _ { t } ) = r ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { E } _ { \\pi } [ R _ { t + 1 } ] = r ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { E } _ { p ( s _ { t + 1 } | s _ { t } , a _ { t } ) } [ V _ { \\pi } ( s _ { t + 1 } ) ] } } \\\\ { { A _ { \\pi } ( s _ { t } , a _ { t } ) = Q _ { \\pi } ( s _ { t } , a _ { t } ) - V _ { \\pi } ( s _ { t } ) . } } \\end{array}", + "type": "interline_equation", + "image_path": "9915c7f590b427849f27b3abc7fecc5232c681e28a5cfc2aab1cb8d4818fdf0b.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 159, + 311, + 452, + 326.6666666666667 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 159, + 326.6666666666667, + 452, + 342.33333333333337 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 159, + 342.33333333333337, + 452, + 358.00000000000006 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 109, + 360, + 503, + 405 + ], + "lines": [ + { + "bbox": [ + 106, + 360, + 504, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 149, + 372 + ], + "score": 0.92, + "content": "Q _ { \\pi } ( s _ { t } , \\pmb { a } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 360, + 496, + 373 + ], + "score": 1.0, + "content": "summarizes the performance of each action from a given state, assuming it follows", + "type": "text" + }, + { + "bbox": [ + 497, + 362, + 504, + 370 + ], + "score": 0.68, + "content": "\\pi", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 167, + 384 + ], + "score": 1.0, + "content": "thereafter, and", + "type": "text" + }, + { + "bbox": [ + 167, + 371, + 209, + 383 + ], + "score": 0.93, + "content": "A _ { \\pi } ( s _ { t } , { \\pmb a } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 370, + 505, + 384 + ], + "score": 1.0, + "content": "provides a measure of how each action compares to the average perfor-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 382, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 106, + 382, + 183, + 395 + ], + "score": 1.0, + "content": "mance at the state", + "type": "text" + }, + { + "bbox": [ + 183, + 384, + 192, + 393 + ], + "score": 0.85, + "content": "\\scriptstyle { s _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 382, + 273, + 395 + ], + "score": 1.0, + "content": ", which is given by", + "type": "text" + }, + { + "bbox": [ + 273, + 382, + 300, + 394 + ], + "score": 0.93, + "content": "V _ { \\pi } ( \\pmb { s } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 382, + 334, + 395 + ], + "score": 1.0, + "content": ". Using", + "type": "text" + }, + { + "bbox": [ + 334, + 382, + 376, + 394 + ], + "score": 0.93, + "content": "A _ { \\pi } ( s _ { t } , { \\pmb a } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 382, + 505, + 395 + ], + "score": 1.0, + "content": "centers the learning signal and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 394, + 229, + 406 + ], + "spans": [ + { + "bbox": [ + 107, + 394, + 229, + 406 + ], + "score": 1.0, + "content": "reduces variance significantly.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5, + "bbox_fs": [ + 106, + 360, + 505, + 406 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 409, + 505, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 423 + ], + "score": 1.0, + "content": "Besides high variance, another problem with the policy gradient is that it requires on-policy samples.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 420, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 505, + 434 + ], + "score": 1.0, + "content": "This makes policy gradient optimization very sample intensive. To achieve similar sample efficiency", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 431, + 506, + 445 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 506, + 445 + ], + "score": 1.0, + "content": "as off-policy methods, we can attempt to include off-policy data. Prior attempts use importance", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 442, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 456 + ], + "score": 1.0, + "content": "sampling to include off-policy trajectories; however, these are known to be difficult scale to high-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 453, + 487, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 487, + 466 + ], + "score": 1.0, + "content": "dimensional action spaces because of rapidly degenerating importance weights (Precup, 2000).", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 408, + 506, + 466 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 478, + 361, + 489 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 362, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 362, + 491 + ], + "score": 1.0, + "content": "2.2 POLICY GRADIENT WITH FUNCTION APPROXIMATION", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 498, + 505, + 576 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 504, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 504, + 511 + ], + "score": 1.0, + "content": "Policy gradient methods with function approximation (Sutton et al., 1999), or actor-critic methods,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "score": 1.0, + "content": "include a policy evaluation step, which often uses temporal difference (TD) learning to fit a critic", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 521, + 504, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 120, + 532 + ], + "score": 0.88, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 121, + 521, + 209, + 533 + ], + "score": 1.0, + "content": "for the current policy", + "type": "text" + }, + { + "bbox": [ + 210, + 521, + 230, + 532 + ], + "score": 0.91, + "content": "\\pi ( \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 521, + 504, + 533 + ], + "score": 1.0, + "content": ", and a policy improvement step which greedily optimizes the policy", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 107, + 531, + 504, + 545 + ], + "spans": [ + { + "bbox": [ + 107, + 533, + 114, + 542 + ], + "score": 0.7, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 531, + 218, + 545 + ], + "score": 1.0, + "content": "against the critic estimate", + "type": "text" + }, + { + "bbox": [ + 219, + 533, + 232, + 543 + ], + "score": 0.88, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 531, + 504, + 545 + ], + "score": 1.0, + "content": ". Significant gains in sample efficiency may be achievable using off-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 543, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 555 + ], + "score": 1.0, + "content": "policy TD learning for the critic, as in Q-learning and deterministic policy gradient (Sutton, 1990;", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 553, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 567 + ], + "score": 1.0, + "content": "Silver et al., 2014), typically by means of experience replay for training deep Q networks (Mnih", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 565, + 314, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 314, + 577 + ], + "score": 1.0, + "content": "et al., 2015; Lillicrap et al., 2016; Gu et al., 2016b).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 498, + 505, + 577 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "score": 1.0, + "content": "One particularly relevant example of such a method is the deep deterministic policy gradient", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 591, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 605 + ], + "score": 1.0, + "content": "(DDPG) (Silver et al., 2014; Lillicrap et al., 2016). The updates for this method are given below,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 603, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 133, + 617 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 603, + 249, + 615 + ], + "score": 0.91, + "content": "\\pi _ { \\boldsymbol { \\theta } } ( \\mathbf { a } _ { t } | \\mathbf { s } _ { t } ) = \\delta ( \\mathbf { a } _ { t } = \\mu _ { \\boldsymbol { \\theta } } ( \\mathbf { s } _ { t } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 603, + 350, + 617 + ], + "score": 1.0, + "content": "is a deterministic policy,", + "type": "text" + }, + { + "bbox": [ + 351, + 604, + 358, + 615 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 603, + 505, + 617 + ], + "score": 1.0, + "content": "is arbitrary exploration distribution,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 614, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 123, + 627 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 616, + 135, + 627 + ], + "score": 0.87, + "content": "\\rho _ { \\beta }", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 614, + 321, + 627 + ], + "score": 1.0, + "content": "corresponds to sampling from a replay buffer.", + "type": "text" + }, + { + "bbox": [ + 322, + 614, + 347, + 626 + ], + "score": 0.92, + "content": "Q ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 614, + 505, + 627 + ], + "score": 1.0, + "content": "is the target network that slowly tracks", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 107, + 625, + 214, + 639 + ], + "spans": [ + { + "bbox": [ + 107, + 626, + 120, + 637 + ], + "score": 0.86, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 625, + 214, + 639 + ], + "score": 1.0, + "content": "(Lillicrap et al., 2016).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 581, + 505, + 639 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 147, + 639, + 464, + 679 + ], + "lines": [ + { + "bbox": [ + 147, + 639, + 464, + 679 + ], + "spans": [ + { + "bbox": [ + 147, + 639, + 464, + 679 + ], + "score": 0.9, + "content": "\\begin{array} { r l } & { w = \\underset { w } { \\arg \\operatorname* { m i n } } \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\beta } ( \\cdot ) , a _ { t } \\sim \\beta ( \\cdot | s _ { t } ) } \\big [ \\big ( r \\big ( s _ { t } , a _ { t } \\big ) + \\gamma Q ( s _ { t + 1 } , \\mu _ { \\theta } ( s _ { t + 1 } ) \\big ) - Q _ { w } ( s _ { t } , a _ { t } ) \\big ) ^ { 2 } \\big ] } \\\\ & { \\theta = \\underset { \\theta } { \\arg \\operatorname* { m a x } } \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\beta } ( \\cdot ) } [ Q _ { w } \\big ( s _ { t } , \\mu _ { \\theta } ( s _ { t } ) \\big ) ] } \\end{array}", + "type": "interline_equation", + "image_path": "a249afcd4ae52b44b8cd04ef8cc47074157d56f3ad0e07af8bf36242d14bdcf1.jpg" + } + ] + } + ], + "index": 44, + "virtual_lines": [ + { + "bbox": [ + 147, + 639, + 464, + 652.3333333333334 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 147, + 652.3333333333334, + 464, + 665.6666666666667 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 147, + 665.6666666666667, + 464, + 679.0000000000001 + ], + "spans": [], + "index": 45 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 680, + 503, + 714 + ], + "lines": [ + { + "bbox": [ + 105, + 679, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 505, + 694 + ], + "score": 1.0, + "content": "When the critic and policy are parametrized with neural networks, full optimization is expensive,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 691, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 505, + 704 + ], + "score": 1.0, + "content": "and instead stochastic gradient optimization is used. The gradient in the policy improvement phase", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 703, + 348, + 715 + ], + "spans": [ + { + "bbox": [ + 105, + 703, + 324, + 715 + ], + "score": 1.0, + "content": "is given below, which is generally a biased gradient of", + "type": "text" + }, + { + "bbox": [ + 325, + 703, + 344, + 714 + ], + "score": 0.92, + "content": "J ( \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 703, + 348, + 715 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 679, + 505, + 715 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 198, + 717, + 413, + 733 + ], + "lines": [ + { + "bbox": [ + 198, + 717, + 413, + 733 + ], + "spans": [ + { + "bbox": [ + 198, + 717, + 413, + 733 + ], + "score": 0.9, + "content": "\\nabla _ { \\theta } J ( \\theta ) \\approx \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\beta } ( \\cdot ) } [ \\nabla _ { a } Q _ { w } \\big ( s _ { t } , a \\big ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } \\big ( s _ { t } \\big ) ]", + "type": "interline_equation", + "image_path": "3559dbdfcd748188a430af44d6b28d871d9806f9929c6b1b627837aaa3042cb4.jpg" + } + ] + } + ], + "index": 49, + "virtual_lines": [ + { + "bbox": [ + 198, + 717, + 413, + 733 + ], + "spans": [], + "index": 49 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "The crucial benefits of DDPG are that it does not rely on high variance REINFORCE gradients and is", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "trainable on off-policy data. These properties make DDPG and other analogous off-policy methods", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "significantly more sample-efficient than policy gradient methods (Lillicrap et al., 2016; Gu et al.,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 113, + 506, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 506, + 130 + ], + "score": 1.0, + "content": "2016b; Duan et al., 2016). However, the use of a biased policy gradient estimator makes analyzing", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 299, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 299, + 139 + ], + "score": 1.0, + "content": "its convergence and stability properties difficult.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 108, + 154, + 167, + 167 + ], + "lines": [ + { + "bbox": [ + 103, + 151, + 170, + 172 + ], + "spans": [ + { + "bbox": [ + 103, + 151, + 170, + 172 + ], + "score": 1.0, + "content": "3 Q-PROP", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 180, + 505, + 269 + ], + "lines": [ + { + "bbox": [ + 105, + 180, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 505, + 193 + ], + "score": 1.0, + "content": "In this section, we derive the Q-Prop estimator for policy gradient. The key idea from this estimator", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 192, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 505, + 203 + ], + "score": 1.0, + "content": "comes from observing Equations 2 and 5 and noting that the former provides an almost unbiased", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 203, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 505, + 214 + ], + "score": 1.0, + "content": "(see Section 2.1), but high variance gradient, while the latter provides a deterministic, but biased", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 213, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 506, + 226 + ], + "score": 1.0, + "content": "gradient. By using the deterministic biased estimator as a particular form of control variate (Ross,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 223, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 237 + ], + "score": 1.0, + "content": "2006; Paisley et al., 2012) for the Monte Carlo policy gradient estimator, we can effectively use both", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 235, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 506, + 248 + ], + "score": 1.0, + "content": "types of gradient information to construct a new estimator that in practice exhibits improved sample", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 246, + 504, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 504, + 259 + ], + "score": 1.0, + "content": "efficiency through the inclusion of off-policy samples while preserving the stability of on-policy", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 257, + 225, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 225, + 270 + ], + "score": 1.0, + "content": "Monte Carlo policy gradient.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 108, + 282, + 218, + 294 + ], + "lines": [ + { + "bbox": [ + 105, + 282, + 220, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 220, + 296 + ], + "score": 1.0, + "content": "3.1 Q-PROP ESTIMATOR", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 303, + 505, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 505, + 316 + ], + "score": 1.0, + "content": "To derive the Q-Prop gradient estimator, we start by using the first-order Taylor expansion of an", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 314, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 182, + 328 + ], + "score": 1.0, + "content": "arbitrary function", + "type": "text" + }, + { + "bbox": [ + 182, + 314, + 420, + 327 + ], + "score": 0.79, + "content": "\\hat { f ( s _ { t } , a _ { t } ) , ~ } \\bar { f } ( s _ { t } , a _ { t } ) = f ( s _ { t } , \\bar { a } _ { t } ) + \\bar { \\nabla } _ { a } f ( \\bar { s _ { t } } , a ) | _ { a = \\bar { a } _ { t } } ( a _ { t } - \\bar { a } _ { t } ) \\cdot", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 314, + 505, + 328 + ], + "score": 1.0, + "content": "as the control vari-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 326, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 303, + 342 + ], + "score": 1.0, + "content": "ate for the policy gradient estimator. We use", + "type": "text" + }, + { + "bbox": [ + 303, + 327, + 430, + 340 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\hat { \\boldsymbol { \\mathcal Q } } ( \\boldsymbol s _ { t } , \\boldsymbol a _ { t } ) = \\sum _ { t ^ { \\prime } = t } ^ { \\infty } \\gamma ^ { t ^ { \\prime } - t } \\boldsymbol r ( \\boldsymbol s _ { t ^ { \\prime } } , \\boldsymbol a _ { t ^ { \\prime } } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 326, + 506, + 342 + ], + "score": 1.0, + "content": "to denote Monte", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 209, + 354 + ], + "score": 1.0, + "content": "Carlo return from state", + "type": "text" + }, + { + "bbox": [ + 209, + 343, + 219, + 352 + ], + "score": 0.83, + "content": "\\scriptstyle { \\pmb { s } } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 340, + 270, + 354 + ], + "score": 1.0, + "content": "and action", + "type": "text" + }, + { + "bbox": [ + 270, + 343, + 280, + 352 + ], + "score": 0.82, + "content": "\\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { a } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 340, + 307, + 354 + ], + "score": 1.0, + "content": ", i.e.", + "type": "text" + }, + { + "bbox": [ + 308, + 341, + 482, + 353 + ], + "score": 0.88, + "content": "\\mathbb { E } _ { \\pi } [ \\hat { \\cal Q } ( s _ { t } , a _ { t } ) ] = r ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { E } _ { p } [ V _ { \\pi } ( s _ { t + 1 } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 340, + 505, + 354 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 351, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 107, + 352, + 201, + 365 + ], + "score": 0.93, + "content": "\\mu _ { \\theta } ( s _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } [ a _ { t } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 351, + 415, + 366 + ], + "score": 1.0, + "content": "to denote the expected action of a stochastic policy", + "type": "text" + }, + { + "bbox": [ + 415, + 354, + 426, + 363 + ], + "score": 0.84, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 351, + 506, + 366 + ], + "score": 1.0, + "content": ". Full derivation is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 363, + 171, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 171, + 375 + ], + "score": 1.0, + "content": "in Appendix A.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5 + }, + { + "type": "interline_equation", + "bbox": [ + 117, + 379, + 482, + 414 + ], + "lines": [ + { + "bbox": [ + 117, + 379, + 482, + 414 + ], + "spans": [ + { + "bbox": [ + 117, + 379, + 482, + 414 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { Q } ( s _ { t } , a _ { t } ) - \\bar { f } ( s _ { t } , a _ { t } ) ] + \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\bar { f } ( s _ { t } , a _ { t } ) ] } \\\\ { = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { Q } ( s _ { t } , a _ { t } ) - \\bar { f } ( s _ { t } , a _ { t } ) ] + \\mathbb { E } _ { \\rho _ { \\pi } } [ \\nabla _ { a } f ( s _ { t } , a ) | _ { a = \\bar { a } _ { t } } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] } \\end{array}", + "type": "interline_equation", + "image_path": "a9c36bb0c9ec213718f863f475c66da7c63792197b548ccb4fd90422835cad96.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 117, + 379, + 482, + 390.6666666666667 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 117, + 390.6666666666667, + 482, + 402.33333333333337 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 117, + 402.33333333333337, + 482, + 414.00000000000006 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 418, + 503, + 441 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 256, + 432 + ], + "score": 1.0, + "content": "Eq. 6 is general for arbitrary function", + "type": "text" + }, + { + "bbox": [ + 256, + 418, + 292, + 430 + ], + "score": 0.94, + "content": "f ( \\mathbf { \\boldsymbol { s } } _ { t } , \\mathbf { \\boldsymbol { a } } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 417, + 435, + 432 + ], + "score": 1.0, + "content": "that is differentiable with respect to", + "type": "text" + }, + { + "bbox": [ + 436, + 421, + 446, + 429 + ], + "score": 0.86, + "content": "\\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { a } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 417, + 505, + 432 + ], + "score": 1.0, + "content": "at an arbitrary", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 428, + 475, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 141, + 442 + ], + "score": 1.0, + "content": "value of", + "type": "text" + }, + { + "bbox": [ + 141, + 430, + 151, + 440 + ], + "score": 0.87, + "content": "\\bar { \\mathbf { a } } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 428, + 336, + 442 + ], + "score": 1.0, + "content": "; however, a sensible choice is to use the critic", + "type": "text" + }, + { + "bbox": [ + 337, + 430, + 351, + 441 + ], + "score": 0.89, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 428, + 365, + 442 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 366, + 430, + 373, + 441 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 428, + 391, + 442 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 391, + 429, + 420, + 441 + ], + "score": 0.92, + "content": "\\pmb { \\mu } _ { \\pmb { \\theta } } ( \\pmb { s } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 428, + 435, + 442 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 436, + 430, + 446, + 440 + ], + "score": 0.87, + "content": "\\bar { \\mathbf { a } } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 428, + 475, + 442 + ], + "score": 1.0, + "content": "to get,", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 447, + 502, + 463 + ], + "lines": [ + { + "bbox": [ + 111, + 447, + 502, + 463 + ], + "spans": [ + { + "bbox": [ + 111, + 447, + 502, + 463 + ], + "score": 0.86, + "content": "\\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { Q } ( s _ { t } , a _ { t } ) - \\bar { Q } _ { w } ( s _ { t } , a _ { t } ) ] + \\mathbb { E } _ { \\rho _ { \\pi } } [ \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] .", + "type": "interline_equation", + "image_path": "7ab4d9990df2732f4c0a7e5ab4367780adb6f5170139fdf3ae2d6112d988441c.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 111, + 447, + 502, + 463 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 482, + 503, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 480, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 304, + 496 + ], + "score": 1.0, + "content": "Finally, since in practice we estimate advantages", + "type": "text" + }, + { + "bbox": [ + 304, + 482, + 341, + 495 + ], + "score": 0.94, + "content": "\\hat { A } ( s _ { t } , a _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 480, + 505, + 496 + ], + "score": 1.0, + "content": ", we write the Q-Prop estimator in terms", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 493, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 296, + 507 + ], + "score": 1.0, + "content": "of advantages to complete the basic derivation,", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "interline_equation", + "bbox": [ + 109, + 510, + 499, + 545 + ], + "lines": [ + { + "bbox": [ + 109, + 510, + 499, + 545 + ], + "spans": [ + { + "bbox": [ + 109, + 510, + 499, + 545 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { \\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { A } ( s _ { t } , a _ { t } ) - \\bar { A } _ { w } ( s _ { t } , a _ { t } ) ] + \\mathbb { E } _ { \\rho _ { \\pi } } [ \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] } \\\\ & { \\bar { A } ( s _ { t } , a _ { t } ) = \\bar { Q } ( s _ { t } , a _ { t } ) - \\mathbb { E } _ { \\pi _ { \\theta } } [ \\bar { Q } ( s _ { t } , a _ { t } ) ] = \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } ( a _ { t } - \\mu _ { \\theta } ( s _ { t } ) ) . } \\end{array}", + "type": "interline_equation", + "image_path": "7f30e2ad2a010faa58514730d9b6d8b8375efa01cd07bbebc391e6fd7d9a53fe.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 109, + 510, + 499, + 521.6666666666666 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 109, + 521.6666666666666, + 499, + 533.3333333333333 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 109, + 533.3333333333333, + 499, + 544.9999999999999 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "Eq. 8 is composed of an analytic gradient through the critic as in Eq. 5 and a residual REINFORCE", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 572, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 586 + ], + "score": 1.0, + "content": "gradient in Eq. 2. From the above derivation, Q-Prop is simply a Monte Carlo policy gradient", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "estimator with a special form of control variate. The important insight comes from the fact that", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 594, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 120, + 605 + ], + "score": 0.87, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 121, + 594, + 506, + 608 + ], + "score": 1.0, + "content": "can be trained using off-policy data as in Eq. 4. Under this setting, Q-Prop is no longer just", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "a Monte Carlo policy gradient method, but more closely resembles an actor-critic method, where", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "the critic can be updated off-policy but the actor is always updated on-policy with an additional", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "REINFORCE correction term so that it remains a Monte Carlo policy gradient method regardless", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "of the parametrization, training method, and performance of the critic. Therefore, Q-Prop can be", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "directly combined with a number of prior techniques from both on-policy methods such as natural", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "policy gradient (Kakade, 2001), trust-region policy optimization (TRPO) (Schulman et al., 2015)", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "score": 1.0, + "content": "and generalized advantage estimation (GAE) (Schulman et al., 2016), and off-policy methods such", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 682, + 386, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 286, + 694 + ], + "score": 1.0, + "content": "as DDPG (Lillicrap et al., 2016) and Retrace", + "type": "text" + }, + { + "bbox": [ + 286, + 682, + 299, + 693 + ], + "score": 0.6, + "content": "( \\lambda )", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 682, + 386, + 694 + ], + "score": 1.0, + "content": "(Munos et al., 2016).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 197, + 712 + ], + "score": 1.0, + "content": "Intuitively, if the critic", + "type": "text" + }, + { + "bbox": [ + 198, + 699, + 212, + 710 + ], + "score": 0.89, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 698, + 268, + 712 + ], + "score": 1.0, + "content": "approximates", + "type": "text" + }, + { + "bbox": [ + 268, + 700, + 282, + 710 + ], + "score": 0.89, + "content": "Q _ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "well, it provides a reliable gradient, reduces the estima-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "tor variance, and improves the convergence rate. Interestingly, control variate analysis in the next", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 720, + 462, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 462, + 733 + ], + "score": 1.0, + "content": "section shows that this is not the only circumstance where Q-Prop helps reduce variance.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "The crucial benefits of DDPG are that it does not rely on high variance REINFORCE gradients and is", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "trainable on off-policy data. These properties make DDPG and other analogous off-policy methods", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "significantly more sample-efficient than policy gradient methods (Lillicrap et al., 2016; Gu et al.,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 113, + 506, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 506, + 130 + ], + "score": 1.0, + "content": "2016b; Duan et al., 2016). However, the use of a biased policy gradient estimator makes analyzing", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 299, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 299, + 139 + ], + "score": 1.0, + "content": "its convergence and stability properties difficult.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 82, + 506, + 139 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 154, + 167, + 167 + ], + "lines": [ + { + "bbox": [ + 103, + 151, + 170, + 172 + ], + "spans": [ + { + "bbox": [ + 103, + 151, + 170, + 172 + ], + "score": 1.0, + "content": "3 Q-PROP", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 180, + 505, + 269 + ], + "lines": [ + { + "bbox": [ + 105, + 180, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 505, + 193 + ], + "score": 1.0, + "content": "In this section, we derive the Q-Prop estimator for policy gradient. The key idea from this estimator", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 192, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 505, + 203 + ], + "score": 1.0, + "content": "comes from observing Equations 2 and 5 and noting that the former provides an almost unbiased", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 203, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 505, + 214 + ], + "score": 1.0, + "content": "(see Section 2.1), but high variance gradient, while the latter provides a deterministic, but biased", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 213, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 506, + 226 + ], + "score": 1.0, + "content": "gradient. By using the deterministic biased estimator as a particular form of control variate (Ross,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 223, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 237 + ], + "score": 1.0, + "content": "2006; Paisley et al., 2012) for the Monte Carlo policy gradient estimator, we can effectively use both", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 235, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 506, + 248 + ], + "score": 1.0, + "content": "types of gradient information to construct a new estimator that in practice exhibits improved sample", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 246, + 504, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 504, + 259 + ], + "score": 1.0, + "content": "efficiency through the inclusion of off-policy samples while preserving the stability of on-policy", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 257, + 225, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 225, + 270 + ], + "score": 1.0, + "content": "Monte Carlo policy gradient.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 180, + 506, + 270 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 282, + 218, + 294 + ], + "lines": [ + { + "bbox": [ + 105, + 282, + 220, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 220, + 296 + ], + "score": 1.0, + "content": "3.1 Q-PROP ESTIMATOR", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 303, + 505, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 505, + 316 + ], + "score": 1.0, + "content": "To derive the Q-Prop gradient estimator, we start by using the first-order Taylor expansion of an", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 314, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 182, + 328 + ], + "score": 1.0, + "content": "arbitrary function", + "type": "text" + }, + { + "bbox": [ + 182, + 314, + 420, + 327 + ], + "score": 0.79, + "content": "\\hat { f ( s _ { t } , a _ { t } ) , ~ } \\bar { f } ( s _ { t } , a _ { t } ) = f ( s _ { t } , \\bar { a } _ { t } ) + \\bar { \\nabla } _ { a } f ( \\bar { s _ { t } } , a ) | _ { a = \\bar { a } _ { t } } ( a _ { t } - \\bar { a } _ { t } ) \\cdot", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 314, + 505, + 328 + ], + "score": 1.0, + "content": "as the control vari-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 326, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 303, + 342 + ], + "score": 1.0, + "content": "ate for the policy gradient estimator. We use", + "type": "text" + }, + { + "bbox": [ + 303, + 327, + 430, + 340 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\hat { \\boldsymbol { \\mathcal Q } } ( \\boldsymbol s _ { t } , \\boldsymbol a _ { t } ) = \\sum _ { t ^ { \\prime } = t } ^ { \\infty } \\gamma ^ { t ^ { \\prime } - t } \\boldsymbol r ( \\boldsymbol s _ { t ^ { \\prime } } , \\boldsymbol a _ { t ^ { \\prime } } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 326, + 506, + 342 + ], + "score": 1.0, + "content": "to denote Monte", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 209, + 354 + ], + "score": 1.0, + "content": "Carlo return from state", + "type": "text" + }, + { + "bbox": [ + 209, + 343, + 219, + 352 + ], + "score": 0.83, + "content": "\\scriptstyle { \\pmb { s } } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 340, + 270, + 354 + ], + "score": 1.0, + "content": "and action", + "type": "text" + }, + { + "bbox": [ + 270, + 343, + 280, + 352 + ], + "score": 0.82, + "content": "\\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { a } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 340, + 307, + 354 + ], + "score": 1.0, + "content": ", i.e.", + "type": "text" + }, + { + "bbox": [ + 308, + 341, + 482, + 353 + ], + "score": 0.88, + "content": "\\mathbb { E } _ { \\pi } [ \\hat { \\cal Q } ( s _ { t } , a _ { t } ) ] = r ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { E } _ { p } [ V _ { \\pi } ( s _ { t + 1 } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 340, + 505, + 354 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 351, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 107, + 352, + 201, + 365 + ], + "score": 0.93, + "content": "\\mu _ { \\theta } ( s _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } [ a _ { t } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 351, + 415, + 366 + ], + "score": 1.0, + "content": "to denote the expected action of a stochastic policy", + "type": "text" + }, + { + "bbox": [ + 415, + 354, + 426, + 363 + ], + "score": 0.84, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 351, + 506, + 366 + ], + "score": 1.0, + "content": ". Full derivation is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 363, + 171, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 171, + 375 + ], + "score": 1.0, + "content": "in Appendix A.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 302, + 506, + 375 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 117, + 379, + 482, + 414 + ], + "lines": [ + { + "bbox": [ + 117, + 379, + 482, + 414 + ], + "spans": [ + { + "bbox": [ + 117, + 379, + 482, + 414 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { Q } ( s _ { t } , a _ { t } ) - \\bar { f } ( s _ { t } , a _ { t } ) ] + \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\bar { f } ( s _ { t } , a _ { t } ) ] } \\\\ { = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { Q } ( s _ { t } , a _ { t } ) - \\bar { f } ( s _ { t } , a _ { t } ) ] + \\mathbb { E } _ { \\rho _ { \\pi } } [ \\nabla _ { a } f ( s _ { t } , a ) | _ { a = \\bar { a } _ { t } } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] } \\end{array}", + "type": "interline_equation", + "image_path": "a9c36bb0c9ec213718f863f475c66da7c63792197b548ccb4fd90422835cad96.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 117, + 379, + 482, + 390.6666666666667 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 117, + 390.6666666666667, + 482, + 402.33333333333337 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 117, + 402.33333333333337, + 482, + 414.00000000000006 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 418, + 503, + 441 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 256, + 432 + ], + "score": 1.0, + "content": "Eq. 6 is general for arbitrary function", + "type": "text" + }, + { + "bbox": [ + 256, + 418, + 292, + 430 + ], + "score": 0.94, + "content": "f ( \\mathbf { \\boldsymbol { s } } _ { t } , \\mathbf { \\boldsymbol { a } } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 417, + 435, + 432 + ], + "score": 1.0, + "content": "that is differentiable with respect to", + "type": "text" + }, + { + "bbox": [ + 436, + 421, + 446, + 429 + ], + "score": 0.86, + "content": "\\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { a } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 417, + 505, + 432 + ], + "score": 1.0, + "content": "at an arbitrary", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 428, + 475, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 141, + 442 + ], + "score": 1.0, + "content": "value of", + "type": "text" + }, + { + "bbox": [ + 141, + 430, + 151, + 440 + ], + "score": 0.87, + "content": "\\bar { \\mathbf { a } } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 428, + 336, + 442 + ], + "score": 1.0, + "content": "; however, a sensible choice is to use the critic", + "type": "text" + }, + { + "bbox": [ + 337, + 430, + 351, + 441 + ], + "score": 0.89, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 428, + 365, + 442 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 366, + 430, + 373, + 441 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 428, + 391, + 442 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 391, + 429, + 420, + 441 + ], + "score": 0.92, + "content": "\\pmb { \\mu } _ { \\pmb { \\theta } } ( \\pmb { s } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 428, + 435, + 442 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 436, + 430, + 446, + 440 + ], + "score": 0.87, + "content": "\\bar { \\mathbf { a } } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 428, + 475, + 442 + ], + "score": 1.0, + "content": "to get,", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 417, + 505, + 442 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 447, + 502, + 463 + ], + "lines": [ + { + "bbox": [ + 111, + 447, + 502, + 463 + ], + "spans": [ + { + "bbox": [ + 111, + 447, + 502, + 463 + ], + "score": 0.86, + "content": "\\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { Q } ( s _ { t } , a _ { t } ) - \\bar { Q } _ { w } ( s _ { t } , a _ { t } ) ] + \\mathbb { E } _ { \\rho _ { \\pi } } [ \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] .", + "type": "interline_equation", + "image_path": "7ab4d9990df2732f4c0a7e5ab4367780adb6f5170139fdf3ae2d6112d988441c.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 111, + 447, + 502, + 463 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 482, + 503, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 480, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 304, + 496 + ], + "score": 1.0, + "content": "Finally, since in practice we estimate advantages", + "type": "text" + }, + { + "bbox": [ + 304, + 482, + 341, + 495 + ], + "score": 0.94, + "content": "\\hat { A } ( s _ { t } , a _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 480, + 505, + 496 + ], + "score": 1.0, + "content": ", we write the Q-Prop estimator in terms", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 493, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 296, + 507 + ], + "score": 1.0, + "content": "of advantages to complete the basic derivation,", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 480, + 505, + 507 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 109, + 510, + 499, + 545 + ], + "lines": [ + { + "bbox": [ + 109, + 510, + 499, + 545 + ], + "spans": [ + { + "bbox": [ + 109, + 510, + 499, + 545 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { \\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { A } ( s _ { t } , a _ { t } ) - \\bar { A } _ { w } ( s _ { t } , a _ { t } ) ] + \\mathbb { E } _ { \\rho _ { \\pi } } [ \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] } \\\\ & { \\bar { A } ( s _ { t } , a _ { t } ) = \\bar { Q } ( s _ { t } , a _ { t } ) - \\mathbb { E } _ { \\pi _ { \\theta } } [ \\bar { Q } ( s _ { t } , a _ { t } ) ] = \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } ( a _ { t } - \\mu _ { \\theta } ( s _ { t } ) ) . } \\end{array}", + "type": "interline_equation", + "image_path": "7f30e2ad2a010faa58514730d9b6d8b8375efa01cd07bbebc391e6fd7d9a53fe.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 109, + 510, + 499, + 521.6666666666666 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 109, + 521.6666666666666, + 499, + 533.3333333333333 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 109, + 533.3333333333333, + 499, + 544.9999999999999 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "Eq. 8 is composed of an analytic gradient through the critic as in Eq. 5 and a residual REINFORCE", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 572, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 586 + ], + "score": 1.0, + "content": "gradient in Eq. 2. From the above derivation, Q-Prop is simply a Monte Carlo policy gradient", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "estimator with a special form of control variate. The important insight comes from the fact that", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 594, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 120, + 605 + ], + "score": 0.87, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 121, + 594, + 506, + 608 + ], + "score": 1.0, + "content": "can be trained using off-policy data as in Eq. 4. Under this setting, Q-Prop is no longer just", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "a Monte Carlo policy gradient method, but more closely resembles an actor-critic method, where", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "the critic can be updated off-policy but the actor is always updated on-policy with an additional", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "REINFORCE correction term so that it remains a Monte Carlo policy gradient method regardless", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "of the parametrization, training method, and performance of the critic. Therefore, Q-Prop can be", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "directly combined with a number of prior techniques from both on-policy methods such as natural", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "policy gradient (Kakade, 2001), trust-region policy optimization (TRPO) (Schulman et al., 2015)", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "score": 1.0, + "content": "and generalized advantage estimation (GAE) (Schulman et al., 2016), and off-policy methods such", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 682, + 386, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 286, + 694 + ], + "score": 1.0, + "content": "as DDPG (Lillicrap et al., 2016) and Retrace", + "type": "text" + }, + { + "bbox": [ + 286, + 682, + 299, + 693 + ], + "score": 0.6, + "content": "( \\lambda )", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 682, + 386, + 694 + ], + "score": 1.0, + "content": "(Munos et al., 2016).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 561, + 506, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 197, + 712 + ], + "score": 1.0, + "content": "Intuitively, if the critic", + "type": "text" + }, + { + "bbox": [ + 198, + 699, + 212, + 710 + ], + "score": 0.89, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 698, + 268, + 712 + ], + "score": 1.0, + "content": "approximates", + "type": "text" + }, + { + "bbox": [ + 268, + 700, + 282, + 710 + ], + "score": 0.89, + "content": "Q _ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "well, it provides a reliable gradient, reduces the estima-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "tor variance, and improves the convergence rate. Interestingly, control variate analysis in the next", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 720, + 462, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 462, + 733 + ], + "score": 1.0, + "content": "section shows that this is not the only circumstance where Q-Prop helps reduce variance.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 698, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 82, + 366, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 367, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 367, + 95 + ], + "score": 1.0, + "content": "3.2 CONTROL VARIATE ANALYSIS AND ADAPTIVE Q-PROP", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 102, + 505, + 158 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 505, + 116 + ], + "score": 1.0, + "content": "For Q-Prop to be applied reliably, it is crucial to analyze how the variance of the estimator changes", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "before and after the application of control variate. Following the prior work on control vari-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 125, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 340, + 138 + ], + "score": 1.0, + "content": "ates (Ross, 2006; Paisley et al., 2012), we first introduce", + "type": "text" + }, + { + "bbox": [ + 340, + 125, + 364, + 137 + ], + "score": 0.92, + "content": "\\eta ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 125, + 506, + 138 + ], + "score": 1.0, + "content": "to Eq. 8, a weighing variable that", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 135, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 373, + 149 + ], + "score": 1.0, + "content": "modulates the strength of control variate. This additional variable", + "type": "text" + }, + { + "bbox": [ + 374, + 136, + 397, + 148 + ], + "score": 0.92, + "content": "\\eta ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 135, + 505, + 149 + ], + "score": 1.0, + "content": "does not introduce bias to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 146, + 163, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 163, + 160 + ], + "score": 1.0, + "content": "the estimator.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3 + }, + { + "type": "interline_equation", + "bbox": [ + 180, + 159, + 431, + 191 + ], + "lines": [ + { + "bbox": [ + 180, + 159, + 431, + 191 + ], + "spans": [ + { + "bbox": [ + 180, + 159, + 431, + 191 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { \\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { A } ( s _ { t } , a _ { t } ) - \\eta ( s _ { t } ) \\bar { A } _ { w } ( s _ { t } , a _ { t } ) ] } \\\\ & { \\qquad + \\mathbb { E } _ { \\rho _ { \\pi } } [ \\eta ( s _ { t } ) \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] } \\end{array}", + "type": "interline_equation", + "image_path": "8a29a313a26cb57497794ec002c4dd2d01c95f4e6b0f5774d17c0c9841dc44c4.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 180, + 159, + 431, + 169.66666666666666 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 180, + 169.66666666666666, + 431, + 180.33333333333331 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 180, + 180.33333333333331, + 431, + 190.99999999999997 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 192, + 474, + 204 + ], + "lines": [ + { + "bbox": [ + 106, + 192, + 474, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 316, + 204 + ], + "score": 1.0, + "content": "The variance of this estimator is given below, where", + "type": "text" + }, + { + "bbox": [ + 316, + 193, + 359, + 203 + ], + "score": 0.9, + "content": "m = 1 . . . M", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 192, + 463, + 204 + ], + "score": 1.0, + "content": "indexes the dimension of", + "type": "text" + }, + { + "bbox": [ + 463, + 193, + 469, + 202 + ], + "score": 0.79, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 192, + 474, + 204 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 162, + 206, + 450, + 235 + ], + "lines": [ + { + "bbox": [ + 162, + 206, + 450, + 235 + ], + "spans": [ + { + "bbox": [ + 162, + 206, + 450, + 235 + ], + "score": 0.93, + "content": "\\mathrm { V a r } ^ { * } = \\mathbb { E } _ { \\rho _ { \\pi } } \\left[ \\sum _ { m } \\mathrm { V a r } _ { a _ { t } } ( \\nabla _ { \\theta _ { m } } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { A } ( s _ { t } , a _ { t } ) - \\eta ( s _ { t } ) \\bar { A } ( s _ { t } , a _ { t } ) ) ) \\right] .", + "type": "interline_equation", + "image_path": "2c67b2666ae2f9272d5fd42a6de8a38b1c8d6b3514f7cf914c808530e04baac1.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 162, + 206, + 450, + 215.66666666666666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 162, + 215.66666666666666, + 450, + 225.33333333333331 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 162, + 225.33333333333331, + 450, + 234.99999999999997 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 104, + 239, + 504, + 254 + ], + "spans": [ + { + "bbox": [ + 104, + 239, + 162, + 254 + ], + "score": 1.0, + "content": "If we choose", + "type": "text" + }, + { + "bbox": [ + 162, + 240, + 186, + 252 + ], + "score": 0.94, + "content": "\\eta ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 239, + 227, + 254 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 228, + 240, + 274, + 251 + ], + "score": 0.88, + "content": "\\mathrm { V a r ^ { * } } < \\mathrm { V a r }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 239, + 306, + 254 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 306, + 239, + 504, + 253 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\mathrm { V a r } = \\mathbb E _ { \\rho _ { \\pi } } [ \\sum _ { m } \\mathrm { V a r } _ { \\boldsymbol { a } _ { t } } ( \\nabla _ { \\boldsymbol { \\theta } _ { m } } \\log \\pi _ { \\boldsymbol { \\theta } } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ) \\hat { \\boldsymbol { A } } ( \\boldsymbol { s } _ { t } , \\boldsymbol { a } _ { t } ) ) ] } \\end{array}", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 104, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "is the original estimator variance measure, then we have managed to reduce the variance. Directly", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 262, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 505, + 275 + ], + "score": 1.0, + "content": "analyzing the above variance measure is nontrivial, for the same reason that computing the optimal", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 272, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 505, + 285 + ], + "score": 1.0, + "content": "baseline is difficult (Weaver & Tao, 2001). In addition, it is often impractical to get multiple action", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 284, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 297 + ], + "score": 1.0, + "content": "samples from the same state, which prohibits using na¨ıve Monte Carlo to estimate the expectations.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 308, + 308 + ], + "score": 1.0, + "content": "Instead, we propose a surrogate variance measure,", + "type": "text" + }, + { + "bbox": [ + 309, + 295, + 421, + 308 + ], + "score": 0.91, + "content": "\\operatorname { V a r } = \\mathbb { E } _ { \\rho _ { \\pi } } [ \\operatorname { V a r } _ { \\mathbf { a } _ { t } } ( \\hat { A } ( s _ { t } , \\mathbf { a } _ { t } ) ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 296, + 505, + 308 + ], + "score": 1.0, + "content": ". A similar surrogate", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 308, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 505, + 319 + ], + "score": 1.0, + "content": "is also used by prior work on learning state-dependent baseline (Mnih & Gregor, 2014), and the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 318, + 312, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 312, + 329 + ], + "score": 1.0, + "content": "benefit is that the measure becomes more tractable,", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16.5 + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 331, + 468, + 364 + ], + "lines": [ + { + "bbox": [ + 142, + 331, + 468, + 364 + ], + "spans": [ + { + "bbox": [ + 142, + 331, + 468, + 364 + ], + "score": 0.88, + "content": "\\begin{array} { r l } & { \\mathrm { V a r } ^ { * } = \\mathbb { E } _ { \\rho _ { \\pi } } [ \\mathrm { V a r } _ { a _ { t } } ( \\hat { A } ( s _ { t } , a _ { t } ) - \\eta ( s _ { t } ) \\bar { A } ( s _ { t } , a _ { t } ) ) ] } \\\\ & { \\qquad = \\mathrm { V a r } + \\mathbb { E } _ { \\rho _ { \\pi } } [ - 2 \\eta ( s _ { t } ) \\mathrm { C o v } _ { a _ { t } } ( \\hat { A } ( s _ { t } , a _ { t } ) , \\bar { A } ( s _ { t } , a _ { t } ) ) + \\eta ( s _ { t } ) ^ { 2 } \\mathrm { V a r } _ { a _ { t } } ( \\bar { A } ( s _ { t } , a _ { t } ) ) ] . } \\end{array}", + "type": "interline_equation", + "image_path": "be0f145346ff59e5f6dca67807e697ffebc0c2629f33f49e7b52beddc1e53d09.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 142, + 331, + 468, + 342.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 142, + 342.0, + 468, + 353.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 142, + 353.0, + 468, + 364.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 367, + 419, + 380 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 420, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 131, + 382 + ], + "score": 1.0, + "content": "Since", + "type": "text" + }, + { + "bbox": [ + 131, + 367, + 267, + 380 + ], + "score": 0.92, + "content": "\\mathbb { E } _ { \\pi } [ \\hat { A } ( \\pmb { s } _ { t } , \\pmb { a } _ { t } ) ] = \\mathbb { E } _ { \\pi } [ \\bar { A } ( \\pmb { s } _ { t } , \\pmb { a } _ { t } ) ] = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 366, + 420, + 382 + ], + "score": 1.0, + "content": ", the terms can be simplified as below,", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "interline_equation", + "bbox": [ + 128, + 383, + 465, + 417 + ], + "lines": [ + { + "bbox": [ + 128, + 383, + 465, + 417 + ], + "spans": [ + { + "bbox": [ + 128, + 383, + 465, + 417 + ], + "score": 0.89, + "content": "\\begin{array} { r l } & { \\mathrm { C o v } _ { a _ { t } } ( \\hat { A } , \\bar { A } ) = \\mathbb { E } _ { \\pi } [ \\hat { A } ( s _ { t } , a _ { t } ) \\bar { A } ( s _ { t } , a _ { t } ) ] } \\\\ & { \\quad \\mathrm { V a r } _ { a _ { t } } ( \\bar { A } ) = \\mathbb { E } _ { \\pi } [ \\bar { A } ( s _ { t } , a _ { t } ) ^ { 2 } ] = \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } ^ { T } \\Sigma _ { \\theta } ( s _ { t } ) \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } , } \\end{array}", + "type": "interline_equation", + "image_path": "4221ff48c12b8d6e96304ffd21be86c42e62fe5f72fe9f6186c63a587bee192e.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 128, + 383, + 465, + 394.3333333333333 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 128, + 394.3333333333333, + 465, + 405.66666666666663 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 128, + 405.66666666666663, + 465, + 416.99999999999994 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 420, + 503, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 419, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 133, + 433 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 420, + 162, + 432 + ], + "score": 0.92, + "content": "\\Sigma _ { \\theta } ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 419, + 361, + 433 + ], + "score": 1.0, + "content": "is the covariance matrix of the stochastic policy", + "type": "text" + }, + { + "bbox": [ + 361, + 422, + 373, + 432 + ], + "score": 0.86, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 419, + 506, + 433 + ], + "score": 1.0, + "content": ". The nice property of Eq. 11 is", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 431, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 124, + 446 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 124, + 432, + 161, + 445 + ], + "score": 0.93, + "content": "\\mathrm { V a r } _ { \\pmb { a } _ { t } } ( \\bar { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 432, + 231, + 446 + ], + "score": 1.0, + "content": "is analytical and", + "type": "text" + }, + { + "bbox": [ + 231, + 431, + 280, + 444 + ], + "score": 0.94, + "content": "\\mathrm { C o v } _ { a _ { t } } ( \\hat { A } , \\bar { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "can be estimated with single action sample. Using this", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "score": 1.0, + "content": "estimate, we propose adaptive variants of Q-Prop that regulate the variance of the gradient estimate.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 465, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 466, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 336, + 480 + ], + "score": 1.0, + "content": "Adaptive Q-Prop. The optimal state-dependent factor", + "type": "text" + }, + { + "bbox": [ + 336, + 466, + 360, + 479 + ], + "score": 0.94, + "content": "\\eta ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 466, + 505, + 480 + ], + "score": 1.0, + "content": "can be computed per state, accord-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 478, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 134, + 492 + ], + "score": 1.0, + "content": "ing to", + "type": "text" + }, + { + "bbox": [ + 134, + 478, + 263, + 491 + ], + "score": 0.93, + "content": "\\eta ^ { * } ( s _ { t } ) = \\mathrm { C o v } _ { a _ { t } } ( \\hat { A } , \\bar { A } ) / \\mathrm { V a r } _ { a _ { t } } ( \\bar { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 478, + 506, + 492 + ], + "score": 1.0, + "content": ". This provides maximum reduction in variance according", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 490, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 104, + 491, + 203, + 505 + ], + "score": 1.0, + "content": "to Eq. 11. Substituting", + "type": "text" + }, + { + "bbox": [ + 204, + 492, + 232, + 504 + ], + "score": 0.92, + "content": "\\eta ^ { * } ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 491, + 315, + 505 + ], + "score": 1.0, + "content": "into Eq. 11, we get", + "type": "text" + }, + { + "bbox": [ + 315, + 490, + 473, + 504 + ], + "score": 0.91, + "content": "\\mathrm { V a r } ^ { * } = \\mathbb { E } _ { \\rho _ { \\pi } } [ ( 1 - \\rho _ { c o r r } ( \\hat { A } , \\bar { A } ) ^ { 2 } ) \\mathrm { V a r } _ { a _ { t } } ( \\hat { A } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 491, + 506, + 505 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 503, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 127, + 516 + ], + "score": 0.83, + "content": "\\rho _ { c o r r }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 504, + 486, + 516 + ], + "score": 1.0, + "content": "is the correlation coefficient, which achieves guaranteed variance reduction if at any state", + "type": "text" + }, + { + "bbox": [ + 487, + 503, + 495, + 514 + ], + "score": 0.84, + "content": "\\bar { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 504, + 506, + 516 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 169, + 529 + ], + "score": 1.0, + "content": "correlated with", + "type": "text" + }, + { + "bbox": [ + 170, + 515, + 178, + 526 + ], + "score": 0.82, + "content": "\\hat { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 516, + 505, + 529 + ], + "score": 1.0, + "content": ". We call this the fully adaptive Q-Prop method. An important conclusion from", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 526, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 304, + 541 + ], + "score": 1.0, + "content": "this analysis is that, in adaptive Q-Prop, the critic", + "type": "text" + }, + { + "bbox": [ + 304, + 528, + 318, + 538 + ], + "score": 0.89, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 526, + 505, + 541 + ], + "score": 1.0, + "content": "does not necessarily need to be approximating", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 538, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 120, + 551 + ], + "score": 0.87, + "content": "Q _ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 539, + 470, + 551 + ], + "score": 1.0, + "content": "well to produce good results. Its Taylor expansion merely needs to be correlated with", + "type": "text" + }, + { + "bbox": [ + 470, + 538, + 478, + 549 + ], + "score": 0.84, + "content": "\\hat { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 539, + 505, + 551 + ], + "score": 1.0, + "content": ", posi-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "tively or even negatively. This is in contrast with actor-critic methods, where performance is greatly", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 561, + 369, + 575 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 369, + 575 + ], + "score": 1.0, + "content": "dependent on the absolute accuracy of the critic’s approximation.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 584, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 105, + 584, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 439, + 599 + ], + "score": 1.0, + "content": "Conservative and Aggressive Q-Prop. In practice, the single-sample estimate of", + "type": "text" + }, + { + "bbox": [ + 439, + 584, + 488, + 597 + ], + "score": 0.94, + "content": "\\mathrm { C o v } _ { a _ { t } } ( \\hat { A } , \\bar { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 584, + 506, + 599 + ], + "score": 1.0, + "content": "has", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 595, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 104, + 595, + 505, + 610 + ], + "score": 1.0, + "content": "high variance itself, and we propose the following two practical implementations of adaptive Q-Prop:", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 607, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 120, + 622 + ], + "score": 1.0, + "content": "(1)", + "type": "text" + }, + { + "bbox": [ + 120, + 608, + 160, + 621 + ], + "score": 0.91, + "content": "\\eta ( s _ { t } ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 608, + 169, + 622 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 170, + 607, + 235, + 621 + ], + "score": 0.94, + "content": "\\hat { \\mathrm { C o v } } _ { { \\pmb a } _ { t } } ( \\hat { A } , \\bar { A } ) > \\bar { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 608, + 252, + 622 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 253, + 608, + 293, + 621 + ], + "score": 0.92, + "content": "\\eta ( s _ { t } ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 608, + 375, + 622 + ], + "score": 1.0, + "content": "if otherwise, and (2)", + "type": "text" + }, + { + "bbox": [ + 375, + 607, + 482, + 621 + ], + "score": 0.93, + "content": "\\eta ( \\pmb { \\mathscr { s } } _ { t } ) = \\mathrm { s i g n } \\big ( \\mathsf { C } \\hat { \\operatorname { o v } } _ { \\pmb { a } _ { t } } ( \\hat { A } , \\bar { A } ) \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 608, + 505, + 622 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 620, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 505, + 632 + ], + "score": 1.0, + "content": "first implementation, which we call conservative Q-Prop, can be thought of as a more conservative", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 631, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 505, + 643 + ], + "score": 1.0, + "content": "version of Q-Prop, which effectively disables the control variate for some samples of the states. This", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 641, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 169, + 655 + ], + "score": 1.0, + "content": "is sensible as if", + "type": "text" + }, + { + "bbox": [ + 169, + 641, + 177, + 653 + ], + "score": 0.84, + "content": "\\hat { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 642, + 194, + 655 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 195, + 642, + 203, + 653 + ], + "score": 0.82, + "content": "\\bar { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 642, + 505, + 655 + ], + "score": 1.0, + "content": "are negatively correlated, it is likely that the critic is very poor. The second", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 654, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 666 + ], + "score": 1.0, + "content": "variant can correspondingly be termed aggressive Q-Prop, since it makes more liberal use of the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 665, + 169, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 169, + 677 + ], + "score": 1.0, + "content": "control variate.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43.5 + }, + { + "type": "title", + "bbox": [ + 108, + 689, + 222, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 224, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 224, + 702 + ], + "score": 1.0, + "content": "3.3 Q-PROP ALGORITHM", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Pseudo-code for the adaptive Q-Prop algorithm is provided in Algorithm 1. It is a mixture of policy", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 505, + 734 + ], + "score": 1.0, + "content": "gradient and actor-critic. At each iteration, it first rolls out the stochastic policy to collect on-policy", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 82, + 366, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 367, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 367, + 95 + ], + "score": 1.0, + "content": "3.2 CONTROL VARIATE ANALYSIS AND ADAPTIVE Q-PROP", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 102, + 505, + 158 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 505, + 116 + ], + "score": 1.0, + "content": "For Q-Prop to be applied reliably, it is crucial to analyze how the variance of the estimator changes", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "before and after the application of control variate. Following the prior work on control vari-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 125, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 340, + 138 + ], + "score": 1.0, + "content": "ates (Ross, 2006; Paisley et al., 2012), we first introduce", + "type": "text" + }, + { + "bbox": [ + 340, + 125, + 364, + 137 + ], + "score": 0.92, + "content": "\\eta ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 125, + 506, + 138 + ], + "score": 1.0, + "content": "to Eq. 8, a weighing variable that", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 135, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 373, + 149 + ], + "score": 1.0, + "content": "modulates the strength of control variate. This additional variable", + "type": "text" + }, + { + "bbox": [ + 374, + 136, + 397, + 148 + ], + "score": 0.92, + "content": "\\eta ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 135, + 505, + 149 + ], + "score": 1.0, + "content": "does not introduce bias to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 146, + 163, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 163, + 160 + ], + "score": 1.0, + "content": "the estimator.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 102, + 506, + 160 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 180, + 159, + 431, + 191 + ], + "lines": [ + { + "bbox": [ + 180, + 159, + 431, + 191 + ], + "spans": [ + { + "bbox": [ + 180, + 159, + 431, + 191 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { \\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { A } ( s _ { t } , a _ { t } ) - \\eta ( s _ { t } ) \\bar { A } _ { w } ( s _ { t } , a _ { t } ) ] } \\\\ & { \\qquad + \\mathbb { E } _ { \\rho _ { \\pi } } [ \\eta ( s _ { t } ) \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] } \\end{array}", + "type": "interline_equation", + "image_path": "8a29a313a26cb57497794ec002c4dd2d01c95f4e6b0f5774d17c0c9841dc44c4.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 180, + 159, + 431, + 169.66666666666666 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 180, + 169.66666666666666, + 431, + 180.33333333333331 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 180, + 180.33333333333331, + 431, + 190.99999999999997 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 192, + 474, + 204 + ], + "lines": [ + { + "bbox": [ + 106, + 192, + 474, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 316, + 204 + ], + "score": 1.0, + "content": "The variance of this estimator is given below, where", + "type": "text" + }, + { + "bbox": [ + 316, + 193, + 359, + 203 + ], + "score": 0.9, + "content": "m = 1 . . . M", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 192, + 463, + 204 + ], + "score": 1.0, + "content": "indexes the dimension of", + "type": "text" + }, + { + "bbox": [ + 463, + 193, + 469, + 202 + ], + "score": 0.79, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 192, + 474, + 204 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 192, + 474, + 204 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 162, + 206, + 450, + 235 + ], + "lines": [ + { + "bbox": [ + 162, + 206, + 450, + 235 + ], + "spans": [ + { + "bbox": [ + 162, + 206, + 450, + 235 + ], + "score": 0.93, + "content": "\\mathrm { V a r } ^ { * } = \\mathbb { E } _ { \\rho _ { \\pi } } \\left[ \\sum _ { m } \\mathrm { V a r } _ { a _ { t } } ( \\nabla _ { \\theta _ { m } } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { A } ( s _ { t } , a _ { t } ) - \\eta ( s _ { t } ) \\bar { A } ( s _ { t } , a _ { t } ) ) ) \\right] .", + "type": "interline_equation", + "image_path": "2c67b2666ae2f9272d5fd42a6de8a38b1c8d6b3514f7cf914c808530e04baac1.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 162, + 206, + 450, + 215.66666666666666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 162, + 215.66666666666666, + 450, + 225.33333333333331 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 162, + 225.33333333333331, + 450, + 234.99999999999997 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 104, + 239, + 504, + 254 + ], + "spans": [ + { + "bbox": [ + 104, + 239, + 162, + 254 + ], + "score": 1.0, + "content": "If we choose", + "type": "text" + }, + { + "bbox": [ + 162, + 240, + 186, + 252 + ], + "score": 0.94, + "content": "\\eta ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 239, + 227, + 254 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 228, + 240, + 274, + 251 + ], + "score": 0.88, + "content": "\\mathrm { V a r ^ { * } } < \\mathrm { V a r }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 239, + 306, + 254 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 306, + 239, + 504, + 253 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\mathrm { V a r } = \\mathbb E _ { \\rho _ { \\pi } } [ \\sum _ { m } \\mathrm { V a r } _ { \\boldsymbol { a } _ { t } } ( \\nabla _ { \\boldsymbol { \\theta } _ { m } } \\log \\pi _ { \\boldsymbol { \\theta } } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ) \\hat { \\boldsymbol { A } } ( \\boldsymbol { s } _ { t } , \\boldsymbol { a } _ { t } ) ) ] } \\end{array}", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 104, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "is the original estimator variance measure, then we have managed to reduce the variance. Directly", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 262, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 505, + 275 + ], + "score": 1.0, + "content": "analyzing the above variance measure is nontrivial, for the same reason that computing the optimal", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 272, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 505, + 285 + ], + "score": 1.0, + "content": "baseline is difficult (Weaver & Tao, 2001). In addition, it is often impractical to get multiple action", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 284, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 297 + ], + "score": 1.0, + "content": "samples from the same state, which prohibits using na¨ıve Monte Carlo to estimate the expectations.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 308, + 308 + ], + "score": 1.0, + "content": "Instead, we propose a surrogate variance measure,", + "type": "text" + }, + { + "bbox": [ + 309, + 295, + 421, + 308 + ], + "score": 0.91, + "content": "\\operatorname { V a r } = \\mathbb { E } _ { \\rho _ { \\pi } } [ \\operatorname { V a r } _ { \\mathbf { a } _ { t } } ( \\hat { A } ( s _ { t } , \\mathbf { a } _ { t } ) ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 296, + 505, + 308 + ], + "score": 1.0, + "content": ". A similar surrogate", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 308, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 505, + 319 + ], + "score": 1.0, + "content": "is also used by prior work on learning state-dependent baseline (Mnih & Gregor, 2014), and the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 318, + 312, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 312, + 329 + ], + "score": 1.0, + "content": "benefit is that the measure becomes more tractable,", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16.5, + "bbox_fs": [ + 104, + 239, + 505, + 329 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 331, + 468, + 364 + ], + "lines": [ + { + "bbox": [ + 142, + 331, + 468, + 364 + ], + "spans": [ + { + "bbox": [ + 142, + 331, + 468, + 364 + ], + "score": 0.88, + "content": "\\begin{array} { r l } & { \\mathrm { V a r } ^ { * } = \\mathbb { E } _ { \\rho _ { \\pi } } [ \\mathrm { V a r } _ { a _ { t } } ( \\hat { A } ( s _ { t } , a _ { t } ) - \\eta ( s _ { t } ) \\bar { A } ( s _ { t } , a _ { t } ) ) ] } \\\\ & { \\qquad = \\mathrm { V a r } + \\mathbb { E } _ { \\rho _ { \\pi } } [ - 2 \\eta ( s _ { t } ) \\mathrm { C o v } _ { a _ { t } } ( \\hat { A } ( s _ { t } , a _ { t } ) , \\bar { A } ( s _ { t } , a _ { t } ) ) + \\eta ( s _ { t } ) ^ { 2 } \\mathrm { V a r } _ { a _ { t } } ( \\bar { A } ( s _ { t } , a _ { t } ) ) ] . } \\end{array}", + "type": "interline_equation", + "image_path": "be0f145346ff59e5f6dca67807e697ffebc0c2629f33f49e7b52beddc1e53d09.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 142, + 331, + 468, + 342.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 142, + 342.0, + 468, + 353.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 142, + 353.0, + 468, + 364.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 367, + 419, + 380 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 420, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 131, + 382 + ], + "score": 1.0, + "content": "Since", + "type": "text" + }, + { + "bbox": [ + 131, + 367, + 267, + 380 + ], + "score": 0.92, + "content": "\\mathbb { E } _ { \\pi } [ \\hat { A } ( \\pmb { s } _ { t } , \\pmb { a } _ { t } ) ] = \\mathbb { E } _ { \\pi } [ \\bar { A } ( \\pmb { s } _ { t } , \\pmb { a } _ { t } ) ] = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 366, + 420, + 382 + ], + "score": 1.0, + "content": ", the terms can be simplified as below,", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 106, + 366, + 420, + 382 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 128, + 383, + 465, + 417 + ], + "lines": [ + { + "bbox": [ + 128, + 383, + 465, + 417 + ], + "spans": [ + { + "bbox": [ + 128, + 383, + 465, + 417 + ], + "score": 0.89, + "content": "\\begin{array} { r l } & { \\mathrm { C o v } _ { a _ { t } } ( \\hat { A } , \\bar { A } ) = \\mathbb { E } _ { \\pi } [ \\hat { A } ( s _ { t } , a _ { t } ) \\bar { A } ( s _ { t } , a _ { t } ) ] } \\\\ & { \\quad \\mathrm { V a r } _ { a _ { t } } ( \\bar { A } ) = \\mathbb { E } _ { \\pi } [ \\bar { A } ( s _ { t } , a _ { t } ) ^ { 2 } ] = \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } ^ { T } \\Sigma _ { \\theta } ( s _ { t } ) \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } , } \\end{array}", + "type": "interline_equation", + "image_path": "4221ff48c12b8d6e96304ffd21be86c42e62fe5f72fe9f6186c63a587bee192e.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 128, + 383, + 465, + 394.3333333333333 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 128, + 394.3333333333333, + 465, + 405.66666666666663 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 128, + 405.66666666666663, + 465, + 416.99999999999994 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 420, + 503, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 419, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 133, + 433 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 420, + 162, + 432 + ], + "score": 0.92, + "content": "\\Sigma _ { \\theta } ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 419, + 361, + 433 + ], + "score": 1.0, + "content": "is the covariance matrix of the stochastic policy", + "type": "text" + }, + { + "bbox": [ + 361, + 422, + 373, + 432 + ], + "score": 0.86, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 419, + 506, + 433 + ], + "score": 1.0, + "content": ". The nice property of Eq. 11 is", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 431, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 124, + 446 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 124, + 432, + 161, + 445 + ], + "score": 0.93, + "content": "\\mathrm { V a r } _ { \\pmb { a } _ { t } } ( \\bar { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 432, + 231, + 446 + ], + "score": 1.0, + "content": "is analytical and", + "type": "text" + }, + { + "bbox": [ + 231, + 431, + 280, + 444 + ], + "score": 0.94, + "content": "\\mathrm { C o v } _ { a _ { t } } ( \\hat { A } , \\bar { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "can be estimated with single action sample. Using this", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "score": 1.0, + "content": "estimate, we propose adaptive variants of Q-Prop that regulate the variance of the gradient estimate.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 419, + 506, + 456 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 465, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 466, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 336, + 480 + ], + "score": 1.0, + "content": "Adaptive Q-Prop. The optimal state-dependent factor", + "type": "text" + }, + { + "bbox": [ + 336, + 466, + 360, + 479 + ], + "score": 0.94, + "content": "\\eta ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 466, + 505, + 480 + ], + "score": 1.0, + "content": "can be computed per state, accord-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 478, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 134, + 492 + ], + "score": 1.0, + "content": "ing to", + "type": "text" + }, + { + "bbox": [ + 134, + 478, + 263, + 491 + ], + "score": 0.93, + "content": "\\eta ^ { * } ( s _ { t } ) = \\mathrm { C o v } _ { a _ { t } } ( \\hat { A } , \\bar { A } ) / \\mathrm { V a r } _ { a _ { t } } ( \\bar { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 478, + 506, + 492 + ], + "score": 1.0, + "content": ". This provides maximum reduction in variance according", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 490, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 104, + 491, + 203, + 505 + ], + "score": 1.0, + "content": "to Eq. 11. Substituting", + "type": "text" + }, + { + "bbox": [ + 204, + 492, + 232, + 504 + ], + "score": 0.92, + "content": "\\eta ^ { * } ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 491, + 315, + 505 + ], + "score": 1.0, + "content": "into Eq. 11, we get", + "type": "text" + }, + { + "bbox": [ + 315, + 490, + 473, + 504 + ], + "score": 0.91, + "content": "\\mathrm { V a r } ^ { * } = \\mathbb { E } _ { \\rho _ { \\pi } } [ ( 1 - \\rho _ { c o r r } ( \\hat { A } , \\bar { A } ) ^ { 2 } ) \\mathrm { V a r } _ { a _ { t } } ( \\hat { A } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 491, + 506, + 505 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 503, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 127, + 516 + ], + "score": 0.83, + "content": "\\rho _ { c o r r }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 504, + 486, + 516 + ], + "score": 1.0, + "content": "is the correlation coefficient, which achieves guaranteed variance reduction if at any state", + "type": "text" + }, + { + "bbox": [ + 487, + 503, + 495, + 514 + ], + "score": 0.84, + "content": "\\bar { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 504, + 506, + 516 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 169, + 529 + ], + "score": 1.0, + "content": "correlated with", + "type": "text" + }, + { + "bbox": [ + 170, + 515, + 178, + 526 + ], + "score": 0.82, + "content": "\\hat { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 516, + 505, + 529 + ], + "score": 1.0, + "content": ". We call this the fully adaptive Q-Prop method. An important conclusion from", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 526, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 304, + 541 + ], + "score": 1.0, + "content": "this analysis is that, in adaptive Q-Prop, the critic", + "type": "text" + }, + { + "bbox": [ + 304, + 528, + 318, + 538 + ], + "score": 0.89, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 526, + 505, + 541 + ], + "score": 1.0, + "content": "does not necessarily need to be approximating", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 538, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 120, + 551 + ], + "score": 0.87, + "content": "Q _ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 539, + 470, + 551 + ], + "score": 1.0, + "content": "well to produce good results. Its Taylor expansion merely needs to be correlated with", + "type": "text" + }, + { + "bbox": [ + 470, + 538, + 478, + 549 + ], + "score": 0.84, + "content": "\\hat { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 539, + 505, + 551 + ], + "score": 1.0, + "content": ", posi-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "tively or even negatively. This is in contrast with actor-critic methods, where performance is greatly", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 561, + 369, + 575 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 369, + 575 + ], + "score": 1.0, + "content": "dependent on the absolute accuracy of the critic’s approximation.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35, + "bbox_fs": [ + 104, + 466, + 506, + 575 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 584, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 105, + 584, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 439, + 599 + ], + "score": 1.0, + "content": "Conservative and Aggressive Q-Prop. In practice, the single-sample estimate of", + "type": "text" + }, + { + "bbox": [ + 439, + 584, + 488, + 597 + ], + "score": 0.94, + "content": "\\mathrm { C o v } _ { a _ { t } } ( \\hat { A } , \\bar { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 584, + 506, + 599 + ], + "score": 1.0, + "content": "has", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 595, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 104, + 595, + 505, + 610 + ], + "score": 1.0, + "content": "high variance itself, and we propose the following two practical implementations of adaptive Q-Prop:", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 607, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 120, + 622 + ], + "score": 1.0, + "content": "(1)", + "type": "text" + }, + { + "bbox": [ + 120, + 608, + 160, + 621 + ], + "score": 0.91, + "content": "\\eta ( s _ { t } ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 608, + 169, + 622 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 170, + 607, + 235, + 621 + ], + "score": 0.94, + "content": "\\hat { \\mathrm { C o v } } _ { { \\pmb a } _ { t } } ( \\hat { A } , \\bar { A } ) > \\bar { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 608, + 252, + 622 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 253, + 608, + 293, + 621 + ], + "score": 0.92, + "content": "\\eta ( s _ { t } ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 608, + 375, + 622 + ], + "score": 1.0, + "content": "if otherwise, and (2)", + "type": "text" + }, + { + "bbox": [ + 375, + 607, + 482, + 621 + ], + "score": 0.93, + "content": "\\eta ( \\pmb { \\mathscr { s } } _ { t } ) = \\mathrm { s i g n } \\big ( \\mathsf { C } \\hat { \\operatorname { o v } } _ { \\pmb { a } _ { t } } ( \\hat { A } , \\bar { A } ) \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 608, + 505, + 622 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 620, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 505, + 632 + ], + "score": 1.0, + "content": "first implementation, which we call conservative Q-Prop, can be thought of as a more conservative", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 631, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 505, + 643 + ], + "score": 1.0, + "content": "version of Q-Prop, which effectively disables the control variate for some samples of the states. This", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 641, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 169, + 655 + ], + "score": 1.0, + "content": "is sensible as if", + "type": "text" + }, + { + "bbox": [ + 169, + 641, + 177, + 653 + ], + "score": 0.84, + "content": "\\hat { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 642, + 194, + 655 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 195, + 642, + 203, + 653 + ], + "score": 0.82, + "content": "\\bar { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 642, + 505, + 655 + ], + "score": 1.0, + "content": "are negatively correlated, it is likely that the critic is very poor. The second", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 654, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 666 + ], + "score": 1.0, + "content": "variant can correspondingly be termed aggressive Q-Prop, since it makes more liberal use of the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 665, + 169, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 169, + 677 + ], + "score": 1.0, + "content": "control variate.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43.5, + "bbox_fs": [ + 104, + 584, + 506, + 677 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 689, + 222, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 224, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 224, + 702 + ], + "score": 1.0, + "content": "3.3 Q-PROP ALGORITHM", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Pseudo-code for the adaptive Q-Prop algorithm is provided in Algorithm 1. It is a mixture of policy", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 505, + 734 + ], + "score": 1.0, + "content": "gradient and actor-critic. At each iteration, it first rolls out the stochastic policy to collect on-policy", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49.5, + "bbox_fs": [ + 105, + 709, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 231, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 232, + 99 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 232, + 99 + ], + "score": 1.0, + "content": "Algorithm 1 Adaptive Q-Prop", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 97, + 503, + 259 + ], + "lines": [ + { + "bbox": [ + 109, + 97, + 405, + 109 + ], + "spans": [ + { + "bbox": [ + 109, + 97, + 155, + 109 + ], + "score": 1.0, + "content": "1: Initialize", + "type": "text" + }, + { + "bbox": [ + 156, + 100, + 163, + 107 + ], + "score": 0.72, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 97, + 197, + 109 + ], + "score": 1.0, + "content": "for critic", + "type": "text" + }, + { + "bbox": [ + 198, + 98, + 210, + 108 + ], + "score": 0.84, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 97, + 214, + 109 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 214, + 98, + 220, + 107 + ], + "score": 0.69, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 97, + 296, + 109 + ], + "score": 1.0, + "content": "for stochastic policy", + "type": "text" + }, + { + "bbox": [ + 297, + 100, + 307, + 108 + ], + "score": 0.84, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 97, + 374, + 109 + ], + "score": 1.0, + "content": ", and replay buffer", + "type": "text" + }, + { + "bbox": [ + 374, + 98, + 401, + 107 + ], + "score": 0.89, + "content": "{ \\mathcal { R } } \\gets \\emptyset", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 97, + 405, + 109 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 109, + 107, + 149, + 118 + ], + "spans": [ + { + "bbox": [ + 109, + 107, + 121, + 118 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 116, + 107, + 149, + 118 + ], + "score": 1.0, + "content": ": repeat", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 116, + 504, + 130 + ], + "spans": [ + { + "bbox": [ + 109, + 116, + 120, + 128 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 134, + 117, + 148, + 128 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 149, + 118, + 193, + 127 + ], + "score": 0.73, + "content": "e = 1 , \\ldots , E", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 117, + 207, + 128 + ], + "score": 1.0, + "content": "do", + "type": "text" + }, + { + "bbox": [ + 306, + 116, + 342, + 130 + ], + "score": 1.0, + "content": ". Collect", + "type": "text" + }, + { + "bbox": [ + 342, + 118, + 350, + 127 + ], + "score": 0.77, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 116, + 492, + 130 + ], + "score": 1.0, + "content": "episodes of on-policy experience using", + "type": "text" + }, + { + "bbox": [ + 493, + 119, + 504, + 128 + ], + "score": 0.77, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 109, + 127, + 196, + 139 + ], + "spans": [ + { + "bbox": [ + 109, + 127, + 120, + 139 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 149, + 128, + 196, + 138 + ], + "score": 0.3, + "content": "s _ { 0 , e } \\sim p ( s _ { 0 } )", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 136, + 232, + 148 + ], + "spans": [ + { + "bbox": [ + 109, + 136, + 120, + 148 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 147, + 137, + 232, + 148 + ], + "score": 1.0, + "content": "for t = 0, . . . , T − 1 do", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 109, + 147, + 373, + 159 + ], + "spans": [ + { + "bbox": [ + 109, + 147, + 120, + 159 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 161, + 147, + 373, + 159 + ], + "score": 0.28, + "content": "a _ { t , e } \\sim \\pi _ { \\theta } ( \\cdot | s _ { t , e } ) , s _ { t + 1 , e } \\sim p ( \\cdot | s _ { t , e } , a _ { t , e } ) , r _ { t , e } = r ( s _ { t , e } , a _ { t , e } )", + "type": "inline_equation", + "image_path": "4bfd6b522607a374f16a61ce64090e3ac2c267f80cca6c905140338a81b3cb6a.jpg" + } + ], + "index": 6 + }, + { + "bbox": [ + 109, + 159, + 397, + 174 + ], + "spans": [ + { + "bbox": [ + 109, + 159, + 120, + 171 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 134, + 159, + 191, + 174 + ], + "score": 1.0, + "content": "Add batch data", + "type": "text" + }, + { + "bbox": [ + 191, + 161, + 328, + 171 + ], + "score": 0.9, + "content": "\\mathcal { B } = \\left. s _ { 0 : T , 1 : E } , \\pmb { a } _ { 0 : T - 1 , 1 : E } , r _ { 0 : T - 1 , 1 : E } \\right.", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 159, + 386, + 174 + ], + "score": 1.0, + "content": "to replay buffer", + "type": "text" + }, + { + "bbox": [ + 387, + 160, + 397, + 169 + ], + "score": 0.79, + "content": "\\mathcal { R }", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 109, + 169, + 311, + 182 + ], + "spans": [ + { + "bbox": [ + 109, + 169, + 120, + 181 + ], + "score": 1.0, + "content": "8:", + "type": "text" + }, + { + "bbox": [ + 134, + 169, + 154, + 182 + ], + "score": 1.0, + "content": "Take", + "type": "text" + }, + { + "bbox": [ + 154, + 171, + 173, + 180 + ], + "score": 0.87, + "content": "E \\cdot T", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 169, + 237, + 182 + ], + "score": 1.0, + "content": "gradient steps on", + "type": "text" + }, + { + "bbox": [ + 237, + 172, + 250, + 181 + ], + "score": 0.86, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 169, + 272, + 182 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 272, + 171, + 282, + 180 + ], + "score": 0.8, + "content": "\\mathcal { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 169, + 298, + 182 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 299, + 172, + 309, + 181 + ], + "score": 0.83, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 169, + 311, + 182 + ], + "score": 0.0, + "content": "", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 109, + 179, + 205, + 192 + ], + "spans": [ + { + "bbox": [ + 109, + 180, + 120, + 192 + ], + "score": 1.0, + "content": "9:", + "type": "text" + }, + { + "bbox": [ + 135, + 179, + 147, + 192 + ], + "score": 1.0, + "content": "Fit", + "type": "text" + }, + { + "bbox": [ + 147, + 180, + 172, + 192 + ], + "score": 0.91, + "content": "V _ { \\phi } ( \\pmb { s } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 179, + 195, + 192 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 195, + 181, + 205, + 190 + ], + "score": 0.79, + "content": "\\mathcal { B }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 191, + 315, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 120, + 203 + ], + "score": 1.0, + "content": "10:", + "type": "text" + }, + { + "bbox": [ + 135, + 191, + 170, + 204 + ], + "score": 1.0, + "content": "Compute", + "type": "text" + }, + { + "bbox": [ + 171, + 191, + 185, + 204 + ], + "score": 0.91, + "content": "\\hat { A } _ { t , e }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 191, + 208, + 204 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 208, + 192, + 240, + 203 + ], + "score": 0.66, + "content": "\\mathrm { G A E } ( \\lambda )", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 191, + 255, + 204 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 255, + 192, + 270, + 204 + ], + "score": 0.91, + "content": "\\bar { A } _ { t , e }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 191, + 315, + 204 + ], + "score": 1.0, + "content": "using Eq. 7", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 202, + 242, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 121, + 213 + ], + "score": 1.0, + "content": "11:", + "type": "text" + }, + { + "bbox": [ + 135, + 203, + 150, + 213 + ], + "score": 1.0, + "content": "Set", + "type": "text" + }, + { + "bbox": [ + 150, + 204, + 164, + 213 + ], + "score": 0.88, + "content": "\\eta _ { t , e }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 203, + 242, + 213 + ], + "score": 1.0, + "content": "based on Section 3.2", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 212, + 354, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 121, + 225 + ], + "score": 1.0, + "content": "12:", + "type": "text" + }, + { + "bbox": [ + 135, + 212, + 282, + 227 + ], + "score": 1.0, + "content": "Compute and center the learning signals", + "type": "text" + }, + { + "bbox": [ + 282, + 213, + 354, + 226 + ], + "score": 0.92, + "content": "l _ { t , e } = \\hat { A } _ { t , e } - \\eta _ { t , e } \\bar { A } _ { t , e }", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 223, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 121, + 237 + ], + "score": 1.0, + "content": "13:", + "type": "text" + }, + { + "bbox": [ + 133, + 223, + 171, + 240 + ], + "score": 1.0, + "content": "Compute", + "type": "text" + }, + { + "bbox": [ + 171, + 225, + 482, + 237 + ], + "score": 0.84, + "content": "\\begin{array} { r } { \\nabla _ { \\theta } J ( \\theta ) \\approx \\frac { 1 } { E T } \\sum _ { e } \\sum _ { t } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t , e } | s _ { t , e } ) l _ { t , e } + \\eta _ { t , e } \\nabla _ { a } Q _ { w } ( s _ { t , e } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t , e } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t , e } ) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 237, + 464, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 122, + 249 + ], + "score": 1.0, + "content": "14:", + "type": "text" + }, + { + "bbox": [ + 135, + 237, + 221, + 250 + ], + "score": 1.0, + "content": "Take a gradient step on", + "type": "text" + }, + { + "bbox": [ + 221, + 239, + 232, + 248 + ], + "score": 0.84, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 237, + 254, + 250 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 255, + 238, + 284, + 249 + ], + "score": 0.89, + "content": "\\nabla _ { \\boldsymbol { \\theta } } J ( \\boldsymbol { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 237, + 453, + 250 + ], + "score": 1.0, + "content": ", optionally with a trust-region constraint using", + "type": "text" + }, + { + "bbox": [ + 454, + 239, + 464, + 247 + ], + "score": 0.65, + "content": "\\mathcal { B }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 246, + 195, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 142, + 261 + ], + "score": 1.0, + "content": "15: until", + "type": "text" + }, + { + "bbox": [ + 143, + 249, + 153, + 258 + ], + "score": 0.84, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 246, + 195, + 261 + ], + "score": 1.0, + "content": "converges.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 281, + 504, + 315 + ], + "lines": [ + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 478, + 294 + ], + "score": 1.0, + "content": "samples, adds the batch to a replay buffer, takes a few gradient steps on the critic, computes", + "type": "text" + }, + { + "bbox": [ + 479, + 281, + 487, + 292 + ], + "score": 0.83, + "content": "\\hat { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 281, + 506, + 294 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 291, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 114, + 303 + ], + "score": 0.8, + "content": "\\bar { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 291, + 309, + 306 + ], + "score": 1.0, + "content": ", and finally applies a gradient step on the policy", + "type": "text" + }, + { + "bbox": [ + 309, + 294, + 320, + 304 + ], + "score": 0.86, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 291, + 457, + 306 + ], + "score": 1.0, + "content": ". In our implementation, the critic", + "type": "text" + }, + { + "bbox": [ + 457, + 293, + 471, + 304 + ], + "score": 0.9, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 291, + 506, + 306 + ], + "score": 1.0, + "content": "is fitted", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 304, + 470, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 470, + 316 + ], + "score": 1.0, + "content": "with off-policy TD learning using the same techniques as in DDPG (Lillicrap et al., 2016):", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 145, + 320, + 465, + 338 + ], + "lines": [ + { + "bbox": [ + 145, + 320, + 465, + 338 + ], + "spans": [ + { + "bbox": [ + 145, + 320, + 465, + 338 + ], + "score": 0.88, + "content": "\\begin{array} { r } { w = \\underset { w } { \\arg \\operatorname* { m i n } } \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\beta } ( \\cdot ) , a _ { t } \\sim \\beta ( \\cdot | s _ { t } ) } [ \\big ( r ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { E } _ { \\pi } [ Q ^ { \\prime } ( s _ { t + 1 } , a _ { t + 1 } ) ] - Q _ { w } ( s _ { t } , a _ { t } ) \\big ) ^ { 2 } ] . } \\end{array}", + "type": "interline_equation", + "image_path": "80488ff0d73f8bbd7b698d7e98c99fb7aece1dddc06b9345ed8ffd6221da55b2.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 145, + 320, + 465, + 338 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 340, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 106, + 341, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 118, + 354 + ], + "score": 0.85, + "content": "V _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 341, + 505, + 354 + ], + "score": 1.0, + "content": "is fitted with the same technique in (Schulman et al., 2016). Generalized advantage estimation", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 307, + 367 + ], + "score": 1.0, + "content": "(GAE) (Schulman et al., 2016) is used to estimate", + "type": "text" + }, + { + "bbox": [ + 307, + 353, + 315, + 364 + ], + "score": 0.81, + "content": "\\hat { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 353, + 505, + 367 + ], + "score": 1.0, + "content": ". The policy update can be done by any method", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "score": 1.0, + "content": "that utilizes the first-order gradient and possibly the on-policy batch data, which includes trust region", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 376, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 505, + 389 + ], + "score": 1.0, + "content": "policy optimization (TRPO) (Schulman et al., 2015). Importantly, this is just one possible imple-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 387, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 505, + 400 + ], + "score": 1.0, + "content": "mentation of Q-Prop, and in Appendix C we show a more general form that can interpolate between", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 398, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 296, + 411 + ], + "score": 1.0, + "content": "pure policy gradient and off-policy actor-critic.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5 + }, + { + "type": "title", + "bbox": [ + 108, + 423, + 188, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 422, + 189, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 189, + 435 + ], + "score": 1.0, + "content": "3.4 LIMITATIONS", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 456 + ], + "score": 1.0, + "content": "A limitation with Q-Prop is that if data collection is very fast, e.g. using fast simulators, the compute", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 454, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 505, + 466 + ], + "score": 1.0, + "content": "time per episode is bound by the critic training at each iteration, and similar to that of DDPG and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "score": 1.0, + "content": "usually much more than that of TRPO. However, in applications where data collection speed is", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 385, + 488 + ], + "score": 1.0, + "content": "the bottleneck, there is sufficient time between policy updates to fit", + "type": "text" + }, + { + "bbox": [ + 386, + 477, + 400, + 488 + ], + "score": 0.87, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 477, + 505, + 488 + ], + "score": 1.0, + "content": "well, which can be done", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 487, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 500 + ], + "score": 1.0, + "content": "asynchronously from the data collection, and the compute time of Q-Prop will be about the same as", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 498, + 165, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 165, + 510 + ], + "score": 1.0, + "content": "that of TRPO.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 515, + 505, + 570 + ], + "lines": [ + { + "bbox": [ + 105, + 514, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 529 + ], + "score": 1.0, + "content": "Another limitation is the robustness to bad critics. We empirically show that our conservative Q-Prop", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "is more robust than standard Q-Prop and much more robust than pure off-policy actor-critic methods", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "such as DDPG; however, estimating when an off-policy critic is reliable or not is still a fundamental", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 548, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 505, + 561 + ], + "score": 1.0, + "content": "problem that shall be further investigated. We can also alleviate this limitation by adopting more", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 559, + 441, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 340, + 571 + ], + "score": 1.0, + "content": "stable off-policy critic learning techniques such as Retrace", + "type": "text" + }, + { + "bbox": [ + 341, + 559, + 354, + 570 + ], + "score": 0.4, + "content": "( \\lambda )", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 559, + 441, + 571 + ], + "score": 1.0, + "content": "(Munos et al., 2016).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35 + }, + { + "type": "title", + "bbox": [ + 108, + 586, + 211, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 213, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 213, + 601 + ], + "score": 1.0, + "content": "4 RELATED WORK", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 610, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 609, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 505, + 624 + ], + "score": 1.0, + "content": "Variance reduction in policy gradient methods is a long-standing problem with a large body of prior", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "score": 1.0, + "content": "work (Weaver & Tao, 2001; Greensmith et al., 2004; Schulman et al., 2016). However, exploration", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "of action-dependent control variates is relatively recent, with most work focusing instead on simpler", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "baselining techniques (Ross, 2006). A subtle exception is compatible feature approximation (Sutton", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "et al., 1999) which can be viewed as a control variate as explained in Appendix B. Another exception", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "is doubly robust estimator in contextual bandits (Dud´ık et al., 2011), which uses a different control", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "variate whose bias cannot be tractably corrected. Control variates were explored recently not in", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "RL but for approximate inference in stochastic models (Paisley et al., 2012), and the closest related", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "work in that domain is the MuProp algorithm (Gu et al., 2016a) which uses a mean-field network", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "as a surrogate for backpropagating a deterministic gradient through stochastic discrete variables.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "MuProp is not directly applicable to model-free RL because the dynamics are unknown; however, it", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 44 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 231, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 232, + 99 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 232, + 99 + ], + "score": 1.0, + "content": "Algorithm 1 Adaptive Q-Prop", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "index", + "bbox": [ + 107, + 97, + 503, + 259 + ], + "lines": [ + { + "bbox": [ + 109, + 97, + 405, + 109 + ], + "spans": [ + { + "bbox": [ + 109, + 97, + 155, + 109 + ], + "score": 1.0, + "content": "1: Initialize", + "type": "text" + }, + { + "bbox": [ + 156, + 100, + 163, + 107 + ], + "score": 0.72, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 97, + 197, + 109 + ], + "score": 1.0, + "content": "for critic", + "type": "text" + }, + { + "bbox": [ + 198, + 98, + 210, + 108 + ], + "score": 0.84, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 97, + 214, + 109 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 214, + 98, + 220, + 107 + ], + "score": 0.69, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 97, + 296, + 109 + ], + "score": 1.0, + "content": "for stochastic policy", + "type": "text" + }, + { + "bbox": [ + 297, + 100, + 307, + 108 + ], + "score": 0.84, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 97, + 374, + 109 + ], + "score": 1.0, + "content": ", and replay buffer", + "type": "text" + }, + { + "bbox": [ + 374, + 98, + 401, + 107 + ], + "score": 0.89, + "content": "{ \\mathcal { R } } \\gets \\emptyset", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 97, + 405, + 109 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 107, + 149, + 118 + ], + "spans": [ + { + "bbox": [ + 109, + 107, + 121, + 118 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 116, + 107, + 149, + 118 + ], + "score": 1.0, + "content": ": repeat", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 116, + 504, + 130 + ], + "spans": [ + { + "bbox": [ + 109, + 116, + 120, + 128 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 134, + 117, + 148, + 128 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 149, + 118, + 193, + 127 + ], + "score": 0.73, + "content": "e = 1 , \\ldots , E", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 117, + 207, + 128 + ], + "score": 1.0, + "content": "do", + "type": "text" + }, + { + "bbox": [ + 306, + 116, + 342, + 130 + ], + "score": 1.0, + "content": ". Collect", + "type": "text" + }, + { + "bbox": [ + 342, + 118, + 350, + 127 + ], + "score": 0.77, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 116, + 492, + 130 + ], + "score": 1.0, + "content": "episodes of on-policy experience using", + "type": "text" + }, + { + "bbox": [ + 493, + 119, + 504, + 128 + ], + "score": 0.77, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 127, + 196, + 139 + ], + "spans": [ + { + "bbox": [ + 109, + 127, + 120, + 139 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 149, + 128, + 196, + 138 + ], + "score": 0.3, + "content": "s _ { 0 , e } \\sim p ( s _ { 0 } )", + "type": "inline_equation" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 136, + 232, + 148 + ], + "spans": [ + { + "bbox": [ + 109, + 136, + 120, + 148 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 147, + 137, + 232, + 148 + ], + "score": 1.0, + "content": "for t = 0, . . . , T − 1 do", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 147, + 373, + 159 + ], + "spans": [ + { + "bbox": [ + 109, + 147, + 120, + 159 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 161, + 147, + 373, + 159 + ], + "score": 0.28, + "content": "a _ { t , e } \\sim \\pi _ { \\theta } ( \\cdot | s _ { t , e } ) , s _ { t + 1 , e } \\sim p ( \\cdot | s _ { t , e } , a _ { t , e } ) , r _ { t , e } = r ( s _ { t , e } , a _ { t , e } )", + "type": "inline_equation", + "image_path": "4bfd6b522607a374f16a61ce64090e3ac2c267f80cca6c905140338a81b3cb6a.jpg" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 159, + 397, + 174 + ], + "spans": [ + { + "bbox": [ + 109, + 159, + 120, + 171 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 134, + 159, + 191, + 174 + ], + "score": 1.0, + "content": "Add batch data", + "type": "text" + }, + { + "bbox": [ + 191, + 161, + 328, + 171 + ], + "score": 0.9, + "content": "\\mathcal { B } = \\left. s _ { 0 : T , 1 : E } , \\pmb { a } _ { 0 : T - 1 , 1 : E } , r _ { 0 : T - 1 , 1 : E } \\right.", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 159, + 386, + 174 + ], + "score": 1.0, + "content": "to replay buffer", + "type": "text" + }, + { + "bbox": [ + 387, + 160, + 397, + 169 + ], + "score": 0.79, + "content": "\\mathcal { R }", + "type": "inline_equation" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 169, + 311, + 182 + ], + "spans": [ + { + "bbox": [ + 109, + 169, + 120, + 181 + ], + "score": 1.0, + "content": "8:", + "type": "text" + }, + { + "bbox": [ + 134, + 169, + 154, + 182 + ], + "score": 1.0, + "content": "Take", + "type": "text" + }, + { + "bbox": [ + 154, + 171, + 173, + 180 + ], + "score": 0.87, + "content": "E \\cdot T", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 169, + 237, + 182 + ], + "score": 1.0, + "content": "gradient steps on", + "type": "text" + }, + { + "bbox": [ + 237, + 172, + 250, + 181 + ], + "score": 0.86, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 169, + 272, + 182 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 272, + 171, + 282, + 180 + ], + "score": 0.8, + "content": "\\mathcal { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 169, + 298, + 182 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 299, + 172, + 309, + 181 + ], + "score": 0.83, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 169, + 311, + 182 + ], + "score": 0.0, + "content": "", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 179, + 205, + 192 + ], + "spans": [ + { + "bbox": [ + 109, + 180, + 120, + 192 + ], + "score": 1.0, + "content": "9:", + "type": "text" + }, + { + "bbox": [ + 135, + 179, + 147, + 192 + ], + "score": 1.0, + "content": "Fit", + "type": "text" + }, + { + "bbox": [ + 147, + 180, + 172, + 192 + ], + "score": 0.91, + "content": "V _ { \\phi } ( \\pmb { s } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 179, + 195, + 192 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 195, + 181, + 205, + 190 + ], + "score": 0.79, + "content": "\\mathcal { B }", + "type": "inline_equation" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 191, + 315, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 120, + 203 + ], + "score": 1.0, + "content": "10:", + "type": "text" + }, + { + "bbox": [ + 135, + 191, + 170, + 204 + ], + "score": 1.0, + "content": "Compute", + "type": "text" + }, + { + "bbox": [ + 171, + 191, + 185, + 204 + ], + "score": 0.91, + "content": "\\hat { A } _ { t , e }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 191, + 208, + 204 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 208, + 192, + 240, + 203 + ], + "score": 0.66, + "content": "\\mathrm { G A E } ( \\lambda )", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 191, + 255, + 204 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 255, + 192, + 270, + 204 + ], + "score": 0.91, + "content": "\\bar { A } _ { t , e }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 191, + 315, + 204 + ], + "score": 1.0, + "content": "using Eq. 7", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 202, + 242, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 121, + 213 + ], + "score": 1.0, + "content": "11:", + "type": "text" + }, + { + "bbox": [ + 135, + 203, + 150, + 213 + ], + "score": 1.0, + "content": "Set", + "type": "text" + }, + { + "bbox": [ + 150, + 204, + 164, + 213 + ], + "score": 0.88, + "content": "\\eta _ { t , e }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 203, + 242, + 213 + ], + "score": 1.0, + "content": "based on Section 3.2", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 212, + 354, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 121, + 225 + ], + "score": 1.0, + "content": "12:", + "type": "text" + }, + { + "bbox": [ + 135, + 212, + 282, + 227 + ], + "score": 1.0, + "content": "Compute and center the learning signals", + "type": "text" + }, + { + "bbox": [ + 282, + 213, + 354, + 226 + ], + "score": 0.92, + "content": "l _ { t , e } = \\hat { A } _ { t , e } - \\eta _ { t , e } \\bar { A } _ { t , e }", + "type": "inline_equation" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 223, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 121, + 237 + ], + "score": 1.0, + "content": "13:", + "type": "text" + }, + { + "bbox": [ + 133, + 223, + 171, + 240 + ], + "score": 1.0, + "content": "Compute", + "type": "text" + }, + { + "bbox": [ + 171, + 225, + 482, + 237 + ], + "score": 0.84, + "content": "\\begin{array} { r } { \\nabla _ { \\theta } J ( \\theta ) \\approx \\frac { 1 } { E T } \\sum _ { e } \\sum _ { t } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t , e } | s _ { t , e } ) l _ { t , e } + \\eta _ { t , e } \\nabla _ { a } Q _ { w } ( s _ { t , e } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t , e } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t , e } ) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 237, + 464, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 122, + 249 + ], + "score": 1.0, + "content": "14:", + "type": "text" + }, + { + "bbox": [ + 135, + 237, + 221, + 250 + ], + "score": 1.0, + "content": "Take a gradient step on", + "type": "text" + }, + { + "bbox": [ + 221, + 239, + 232, + 248 + ], + "score": 0.84, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 237, + 254, + 250 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 255, + 238, + 284, + 249 + ], + "score": 0.89, + "content": "\\nabla _ { \\boldsymbol { \\theta } } J ( \\boldsymbol { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 237, + 453, + 250 + ], + "score": 1.0, + "content": ", optionally with a trust-region constraint using", + "type": "text" + }, + { + "bbox": [ + 454, + 239, + 464, + 247 + ], + "score": 0.65, + "content": "\\mathcal { B }", + "type": "inline_equation" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 246, + 195, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 142, + 261 + ], + "score": 1.0, + "content": "15: until", + "type": "text" + }, + { + "bbox": [ + 143, + 249, + 153, + 258 + ], + "score": 0.84, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 246, + 195, + 261 + ], + "score": 1.0, + "content": "converges.", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + } + ], + "index": 8, + "bbox_fs": [ + 105, + 97, + 504, + 261 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 281, + 504, + 315 + ], + "lines": [ + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 478, + 294 + ], + "score": 1.0, + "content": "samples, adds the batch to a replay buffer, takes a few gradient steps on the critic, computes", + "type": "text" + }, + { + "bbox": [ + 479, + 281, + 487, + 292 + ], + "score": 0.83, + "content": "\\hat { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 281, + 506, + 294 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 291, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 114, + 303 + ], + "score": 0.8, + "content": "\\bar { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 291, + 309, + 306 + ], + "score": 1.0, + "content": ", and finally applies a gradient step on the policy", + "type": "text" + }, + { + "bbox": [ + 309, + 294, + 320, + 304 + ], + "score": 0.86, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 291, + 457, + 306 + ], + "score": 1.0, + "content": ". In our implementation, the critic", + "type": "text" + }, + { + "bbox": [ + 457, + 293, + 471, + 304 + ], + "score": 0.9, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 291, + 506, + 306 + ], + "score": 1.0, + "content": "is fitted", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 304, + 470, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 470, + 316 + ], + "score": 1.0, + "content": "with off-policy TD learning using the same techniques as in DDPG (Lillicrap et al., 2016):", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 281, + 506, + 316 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 145, + 320, + 465, + 338 + ], + "lines": [ + { + "bbox": [ + 145, + 320, + 465, + 338 + ], + "spans": [ + { + "bbox": [ + 145, + 320, + 465, + 338 + ], + "score": 0.88, + "content": "\\begin{array} { r } { w = \\underset { w } { \\arg \\operatorname* { m i n } } \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\beta } ( \\cdot ) , a _ { t } \\sim \\beta ( \\cdot | s _ { t } ) } [ \\big ( r ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { E } _ { \\pi } [ Q ^ { \\prime } ( s _ { t + 1 } , a _ { t + 1 } ) ] - Q _ { w } ( s _ { t } , a _ { t } ) \\big ) ^ { 2 } ] . } \\end{array}", + "type": "interline_equation", + "image_path": "80488ff0d73f8bbd7b698d7e98c99fb7aece1dddc06b9345ed8ffd6221da55b2.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 145, + 320, + 465, + 338 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 340, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 106, + 341, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 118, + 354 + ], + "score": 0.85, + "content": "V _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 341, + 505, + 354 + ], + "score": 1.0, + "content": "is fitted with the same technique in (Schulman et al., 2016). Generalized advantage estimation", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 307, + 367 + ], + "score": 1.0, + "content": "(GAE) (Schulman et al., 2016) is used to estimate", + "type": "text" + }, + { + "bbox": [ + 307, + 353, + 315, + 364 + ], + "score": 0.81, + "content": "\\hat { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 353, + 505, + 367 + ], + "score": 1.0, + "content": ". The policy update can be done by any method", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "score": 1.0, + "content": "that utilizes the first-order gradient and possibly the on-policy batch data, which includes trust region", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 376, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 505, + 389 + ], + "score": 1.0, + "content": "policy optimization (TRPO) (Schulman et al., 2015). Importantly, this is just one possible imple-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 387, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 505, + 400 + ], + "score": 1.0, + "content": "mentation of Q-Prop, and in Appendix C we show a more general form that can interpolate between", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 398, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 296, + 411 + ], + "score": 1.0, + "content": "pure policy gradient and off-policy actor-critic.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 341, + 505, + 411 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 423, + 188, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 422, + 189, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 189, + 435 + ], + "score": 1.0, + "content": "3.4 LIMITATIONS", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 456 + ], + "score": 1.0, + "content": "A limitation with Q-Prop is that if data collection is very fast, e.g. using fast simulators, the compute", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 454, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 505, + 466 + ], + "score": 1.0, + "content": "time per episode is bound by the critic training at each iteration, and similar to that of DDPG and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "score": 1.0, + "content": "usually much more than that of TRPO. However, in applications where data collection speed is", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 385, + 488 + ], + "score": 1.0, + "content": "the bottleneck, there is sufficient time between policy updates to fit", + "type": "text" + }, + { + "bbox": [ + 386, + 477, + 400, + 488 + ], + "score": 0.87, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 477, + 505, + 488 + ], + "score": 1.0, + "content": "well, which can be done", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 487, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 500 + ], + "score": 1.0, + "content": "asynchronously from the data collection, and the compute time of Q-Prop will be about the same as", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 498, + 165, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 165, + 510 + ], + "score": 1.0, + "content": "that of TRPO.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 442, + 506, + 510 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 515, + 505, + 570 + ], + "lines": [ + { + "bbox": [ + 105, + 514, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 529 + ], + "score": 1.0, + "content": "Another limitation is the robustness to bad critics. We empirically show that our conservative Q-Prop", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "is more robust than standard Q-Prop and much more robust than pure off-policy actor-critic methods", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "such as DDPG; however, estimating when an off-policy critic is reliable or not is still a fundamental", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 548, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 505, + 561 + ], + "score": 1.0, + "content": "problem that shall be further investigated. We can also alleviate this limitation by adopting more", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 559, + 441, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 340, + 571 + ], + "score": 1.0, + "content": "stable off-policy critic learning techniques such as Retrace", + "type": "text" + }, + { + "bbox": [ + 341, + 559, + 354, + 570 + ], + "score": 0.4, + "content": "( \\lambda )", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 559, + 441, + 571 + ], + "score": 1.0, + "content": "(Munos et al., 2016).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 514, + 505, + 571 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 586, + 211, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 213, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 213, + 601 + ], + "score": 1.0, + "content": "4 RELATED WORK", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 610, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 609, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 505, + 624 + ], + "score": 1.0, + "content": "Variance reduction in policy gradient methods is a long-standing problem with a large body of prior", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "score": 1.0, + "content": "work (Weaver & Tao, 2001; Greensmith et al., 2004; Schulman et al., 2016). However, exploration", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "of action-dependent control variates is relatively recent, with most work focusing instead on simpler", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "baselining techniques (Ross, 2006). A subtle exception is compatible feature approximation (Sutton", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "et al., 1999) which can be viewed as a control variate as explained in Appendix B. Another exception", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "is doubly robust estimator in contextual bandits (Dud´ık et al., 2011), which uses a different control", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "variate whose bias cannot be tractably corrected. Control variates were explored recently not in", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "RL but for approximate inference in stochastic models (Paisley et al., 2012), and the closest related", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "work in that domain is the MuProp algorithm (Gu et al., 2016a) which uses a mean-field network", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "as a surrogate for backpropagating a deterministic gradient through stochastic discrete variables.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "MuProp is not directly applicable to model-free RL because the dynamics are unknown; however, it", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "can be if the dynamics are learned as in model-based RL (Atkeson & Santamaria, 1997; Deisenroth", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "& Rasmussen, 2011). This model-based Q-Prop is itself an interesting direction of research as it", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 304, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 304, + 118 + ], + "score": 1.0, + "content": "effectively corrects bias in model-based learning.", + "type": "text", + "cross_page": true + } + ], + "index": 2 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 609, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "can be if the dynamics are learned as in model-based RL (Atkeson & Santamaria, 1997; Deisenroth", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "& Rasmussen, 2011). This model-based Q-Prop is itself an interesting direction of research as it", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 304, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 304, + 118 + ], + "score": 1.0, + "content": "effectively corrects bias in model-based learning.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 119, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 505, + 135 + ], + "score": 1.0, + "content": "Part of the benefit of Q-Prop is the ability to use off-policy data to improve on-policy policy gra-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "dient methods. Prior methods that combine off-policy data with policy gradients either introduce", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "bias (Sutton et al., 1999; Silver et al., 2014) or use importance weighting, which is known to re-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 506, + 168 + ], + "score": 1.0, + "content": "sult in degenerate importance weights in high dimensions, resulting in very high variance (Precup,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "score": 1.0, + "content": "2000; Levine & Koltun, 2013). Q-Prop provides a new approach for using off-policy data to reduce", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 177, + 274, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 274, + 188 + ], + "score": 1.0, + "content": "variance without introducing further bias.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 193, + 505, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "score": 1.0, + "content": "Lastly, since Q-Prop uses both on-policy policy updates and off-policy critic learning, it can take", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 505, + 217 + ], + "score": 1.0, + "content": "advantage of prior work along both lines of research. We chose to implement Q-Prop on top of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "score": 1.0, + "content": "TRPO-GAE primarily for the purpose of enabling a fair comparison in the experiments, but com-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 225, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 240 + ], + "score": 1.0, + "content": "bining Q-Prop with other on-policy update schemes and off-policy critic training methods is an", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "score": 1.0, + "content": "interesting direction for future work. For example, Q-Prop can also be used with other on-policy", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 247, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 261 + ], + "score": 1.0, + "content": "policy gradient methods such as A3C (Mnih et al., 2016) and off-policy advantage estimation meth-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 258, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 185, + 272 + ], + "score": 1.0, + "content": "ods such as Retrace", + "type": "text" + }, + { + "bbox": [ + 185, + 259, + 198, + 270 + ], + "score": 0.46, + "content": "( \\lambda )", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 258, + 505, + 272 + ], + "score": 1.0, + "content": "(Munos et al., 2016), GTD2 (Sutton et al., 2009), emphatic TD (Sutton et al.,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 269, + 300, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 300, + 282 + ], + "score": 1.0, + "content": "2015), and WIS-LSTD (Mahmood et al., 2014).", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 108, + 297, + 200, + 309 + ], + "lines": [ + { + "bbox": [ + 104, + 295, + 202, + 312 + ], + "spans": [ + { + "bbox": [ + 104, + 295, + 202, + 312 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 116, + 321, + 496, + 390 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 321, + 496, + 390 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 321, + 496, + 390 + ], + "spans": [ + { + "bbox": [ + 116, + 321, + 496, + 390 + ], + "score": 0.968, + "type": "image", + "image_path": "8f3888758beaed39ef5a270faa0f5bf0d9ee7237dbc5a6eaa9d962ade37ee8e4.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 116, + 321, + 496, + 344.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 116, + 344.0, + 496, + 367.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 116, + 367.0, + 496, + 390.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 399, + 503, + 422 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "score": 1.0, + "content": "Figure 1: Illustrations of OpenAI Gym MuJoCo domains (Brockman et al., 2016; Duan et al., 2016):", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 410, + 472, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 472, + 423 + ], + "score": 1.0, + "content": "(a) Ant, (b) HalfCheetah, (c) Hopper, (d) Humanoid, (e) Reacher, (f) Swimmer, (g) Walker.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + } + ], + "index": 20.25 + }, + { + "type": "text", + "bbox": [ + 107, + 434, + 505, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "We evaluated Q-Prop and its variants on continuous control environments from the OpenAI Gym", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 446, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 458 + ], + "score": 1.0, + "content": "benchmark (Brockman et al., 2016) using the MuJoCo physics simulator (Todorov et al., 2012) as", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "shown in Figure 1. Algorithms are identified by acronyms, followed by a number indicating batch", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "size, except for DDPG, which is a prior online actor-critic algorithm (Lillicrap et al., 2016). “c-” and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "“v-” denote conservative and aggressive Q-Prop variants as described in Section 3.2. “TR-” denotes", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "trust-region policy optimization (Schulman et al., 2015), while “V-” denotes vanilla policy gradient.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 500, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 514 + ], + "score": 1.0, + "content": "For example, “TR-c-Q-Prop-5000” means convervative Q-Prop with the trust-region policy update,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 510, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 525 + ], + "score": 1.0, + "content": "and a batch size of 5000. “VPG” and “TRPO” are vanilla policy gradient and trust-region policy op-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "timization respectively (Schulman et al., 2016; Duan et al., 2016). Unless otherwise stated, all policy", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 291, + 546 + ], + "score": 1.0, + "content": "gradient methods are implemented with GAE", + "type": "text" + }, + { + "bbox": [ + 292, + 533, + 331, + 544 + ], + "score": 0.76, + "content": "\\lambda = 0 . 9 7", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 533, + 506, + 546 + ], + "score": 1.0, + "content": ") (Schulman et al., 2016). Note that TRPO-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "GAE is currently the state-of-the-art method on most of the OpenAI Gym benchmark tasks, though", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "our experiments show that a well-tuned DDPG implementation sometimes achieves better results.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "Our algorithm implementations are built on top of the rllab TRPO and DDPG codes from Duan", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "et al. (2016) and available at https://github.com/shaneshixiang/rllabplusplus.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 587, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 601 + ], + "score": 1.0, + "content": "Policy and value function architectures and other training details including hyperparameter values", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 599, + 223, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 223, + 612 + ], + "score": 1.0, + "content": "are provided in Appendix D.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 30.5 + }, + { + "type": "title", + "bbox": [ + 108, + 623, + 213, + 634 + ], + "lines": [ + { + "bbox": [ + 105, + 622, + 214, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 214, + 636 + ], + "score": 1.0, + "content": "5.1 ADAPTIVE Q-PROP", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "First, it is useful to identify how reliable each variant of Q-Prop is. In this section, we analyze", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "standard Q-Prop and two adaptive variants, c-Q-Prop and a-Q-Prop, and demonstrate the stability", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "of the method across different batch sizes. Figure 2a shows a comparison of Q-Prop variants with", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "trust-region updates on the HalfCheetah-v1 domain, along with the best performing TRPO hyper-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "parameters. The results are consistent with theory: conservative Q-Prop achieves much more stable", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "performance than the standard and aggressive variants, and all Q-Prop variants significantly outper-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "form TRPO in terms of sample efficiency, e.g. conservative Q-Prop reaches average reward of 4000", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 721, + 294, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 294, + 733 + ], + "score": 1.0, + "content": "using about 10 times less samples than TRPO.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 7 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [], + "index": 1, + "bbox_fs": [ + 105, + 82, + 505, + 118 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 119, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 505, + 135 + ], + "score": 1.0, + "content": "Part of the benefit of Q-Prop is the ability to use off-policy data to improve on-policy policy gra-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "dient methods. Prior methods that combine off-policy data with policy gradients either introduce", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "bias (Sutton et al., 1999; Silver et al., 2014) or use importance weighting, which is known to re-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 506, + 168 + ], + "score": 1.0, + "content": "sult in degenerate importance weights in high dimensions, resulting in very high variance (Precup,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "score": 1.0, + "content": "2000; Levine & Koltun, 2013). Q-Prop provides a new approach for using off-policy data to reduce", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 177, + 274, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 274, + 188 + ], + "score": 1.0, + "content": "variance without introducing further bias.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 119, + 506, + 188 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 193, + 505, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "score": 1.0, + "content": "Lastly, since Q-Prop uses both on-policy policy updates and off-policy critic learning, it can take", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 505, + 217 + ], + "score": 1.0, + "content": "advantage of prior work along both lines of research. We chose to implement Q-Prop on top of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "score": 1.0, + "content": "TRPO-GAE primarily for the purpose of enabling a fair comparison in the experiments, but com-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 225, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 240 + ], + "score": 1.0, + "content": "bining Q-Prop with other on-policy update schemes and off-policy critic training methods is an", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "score": 1.0, + "content": "interesting direction for future work. For example, Q-Prop can also be used with other on-policy", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 247, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 261 + ], + "score": 1.0, + "content": "policy gradient methods such as A3C (Mnih et al., 2016) and off-policy advantage estimation meth-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 258, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 185, + 272 + ], + "score": 1.0, + "content": "ods such as Retrace", + "type": "text" + }, + { + "bbox": [ + 185, + 259, + 198, + 270 + ], + "score": 0.46, + "content": "( \\lambda )", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 258, + 505, + 272 + ], + "score": 1.0, + "content": "(Munos et al., 2016), GTD2 (Sutton et al., 2009), emphatic TD (Sutton et al.,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 269, + 300, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 300, + 282 + ], + "score": 1.0, + "content": "2015), and WIS-LSTD (Mahmood et al., 2014).", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 192, + 506, + 282 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 297, + 200, + 309 + ], + "lines": [ + { + "bbox": [ + 104, + 295, + 202, + 312 + ], + "spans": [ + { + "bbox": [ + 104, + 295, + 202, + 312 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 116, + 321, + 496, + 390 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 321, + 496, + 390 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 321, + 496, + 390 + ], + "spans": [ + { + "bbox": [ + 116, + 321, + 496, + 390 + ], + "score": 0.968, + "type": "image", + "image_path": "8f3888758beaed39ef5a270faa0f5bf0d9ee7237dbc5a6eaa9d962ade37ee8e4.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 116, + 321, + 496, + 344.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 116, + 344.0, + 496, + 367.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 116, + 367.0, + 496, + 390.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 399, + 503, + 422 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "score": 1.0, + "content": "Figure 1: Illustrations of OpenAI Gym MuJoCo domains (Brockman et al., 2016; Duan et al., 2016):", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 410, + 472, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 472, + 423 + ], + "score": 1.0, + "content": "(a) Ant, (b) HalfCheetah, (c) Hopper, (d) Humanoid, (e) Reacher, (f) Swimmer, (g) Walker.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + } + ], + "index": 20.25 + }, + { + "type": "text", + "bbox": [ + 107, + 434, + 505, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "We evaluated Q-Prop and its variants on continuous control environments from the OpenAI Gym", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 446, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 458 + ], + "score": 1.0, + "content": "benchmark (Brockman et al., 2016) using the MuJoCo physics simulator (Todorov et al., 2012) as", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "shown in Figure 1. Algorithms are identified by acronyms, followed by a number indicating batch", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "size, except for DDPG, which is a prior online actor-critic algorithm (Lillicrap et al., 2016). “c-” and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "“v-” denote conservative and aggressive Q-Prop variants as described in Section 3.2. “TR-” denotes", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "trust-region policy optimization (Schulman et al., 2015), while “V-” denotes vanilla policy gradient.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 500, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 514 + ], + "score": 1.0, + "content": "For example, “TR-c-Q-Prop-5000” means convervative Q-Prop with the trust-region policy update,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 510, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 525 + ], + "score": 1.0, + "content": "and a batch size of 5000. “VPG” and “TRPO” are vanilla policy gradient and trust-region policy op-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "timization respectively (Schulman et al., 2016; Duan et al., 2016). Unless otherwise stated, all policy", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 291, + 546 + ], + "score": 1.0, + "content": "gradient methods are implemented with GAE", + "type": "text" + }, + { + "bbox": [ + 292, + 533, + 331, + 544 + ], + "score": 0.76, + "content": "\\lambda = 0 . 9 7", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 533, + 506, + 546 + ], + "score": 1.0, + "content": ") (Schulman et al., 2016). Note that TRPO-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "GAE is currently the state-of-the-art method on most of the OpenAI Gym benchmark tasks, though", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "our experiments show that a well-tuned DDPG implementation sometimes achieves better results.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "Our algorithm implementations are built on top of the rllab TRPO and DDPG codes from Duan", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "et al. (2016) and available at https://github.com/shaneshixiang/rllabplusplus.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 587, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 601 + ], + "score": 1.0, + "content": "Policy and value function architectures and other training details including hyperparameter values", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 599, + 223, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 223, + 612 + ], + "score": 1.0, + "content": "are provided in Appendix D.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 434, + 506, + 612 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 623, + 213, + 634 + ], + "lines": [ + { + "bbox": [ + 105, + 622, + 214, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 214, + 636 + ], + "score": 1.0, + "content": "5.1 ADAPTIVE Q-PROP", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "First, it is useful to identify how reliable each variant of Q-Prop is. In this section, we analyze", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "standard Q-Prop and two adaptive variants, c-Q-Prop and a-Q-Prop, and demonstrate the stability", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "of the method across different batch sizes. Figure 2a shows a comparison of Q-Prop variants with", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "trust-region updates on the HalfCheetah-v1 domain, along with the best performing TRPO hyper-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "parameters. The results are consistent with theory: conservative Q-Prop achieves much more stable", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "performance than the standard and aggressive variants, and all Q-Prop variants significantly outper-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "form TRPO in terms of sample efficiency, e.g. conservative Q-Prop reaches average reward of 4000", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 721, + 294, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 294, + 733 + ], + "score": 1.0, + "content": "using about 10 times less samples than TRPO.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 644, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 131, + 82, + 283, + 188 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 131, + 82, + 283, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 131, + 82, + 283, + 188 + ], + "spans": [ + { + "bbox": [ + 131, + 82, + 283, + 188 + ], + "score": 0.651, + "type": "image", + "image_path": "b8ef0cab31ceb57d2ca2f3840e44751e97cf886ce5ce81bc26d79da92a4d50c5.jpg" + } + ] + } + ], + "index": 1.0, + "virtual_lines": [ + { + "bbox": [ + 131, + 82, + 283, + 135.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 131, + 135.0, + 283, + 188.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 133, + 193, + 282, + 204 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 282, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 282, + 205 + ], + "score": 1.0, + "content": "(a) Standard Q-Prop vs adaptive variants.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 327, + 82, + 479, + 186 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 327, + 82, + 479, + 186 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 327, + 82, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 327, + 82, + 479, + 186 + ], + "score": 0.851, + "type": "image", + "image_path": "ac3e5492e7295b46b13f17d48d9490a7f429f569aab0d8bfb64532dc2013b0fc.jpg" + } + ] + } + ], + "index": 2.0, + "virtual_lines": [ + { + "bbox": [ + 327, + 82, + 479, + 134.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 327, + 134.0, + 479, + 186.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 307, + 193, + 500, + 204 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 306, + 192, + 501, + 205 + ], + "spans": [ + { + "bbox": [ + 306, + 192, + 501, + 205 + ], + "score": 1.0, + "content": "(b) Conservative Q-Prop vs TRPO across batch sizes.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "image_caption", + "bbox": [ + 106, + 213, + 505, + 257 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 212, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 226 + ], + "score": 1.0, + "content": "Figure 2: Average return over episodes in HalfCheetah-v1 during learning, exploring adaptive Q-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 223, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 236 + ], + "score": 1.0, + "content": "Prop methods and different batch sizes. All variants of Q-Prop substantially outperform TRPO in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 235, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 505, + 247 + ], + "score": 1.0, + "content": "terms of sample efficiency. TR-c-QP, conservative Q-Prop with trust-region update performs most", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 246, + 245, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 245, + 258 + ], + "score": 1.0, + "content": "stably across different batch sizes.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 281, + 505, + 337 + ], + "lines": [ + { + "bbox": [ + 106, + 282, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 294 + ], + "score": 1.0, + "content": "Figure 2b shows the performance of conservative Q-Prop against TRPO across different batch", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 292, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 505, + 306 + ], + "score": 1.0, + "content": "sizes. Due to high variance in gradient estimates, TRPO typically requires very large batch sizes,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 303, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 104, + 303, + 506, + 317 + ], + "score": 1.0, + "content": "e.g. 25000 steps or 25 episodes per update, to perform well. We show that our Q-Prop methods can", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "learn even with just 1 episode per update, and achieves better sample efficiency with small batch", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 325, + 490, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 490, + 339 + ], + "score": 1.0, + "content": "sizes. This shows that Q-Prop significantly reduces the variance compared to the prior methods.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 342, + 505, + 376 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "As we discussed in Section 1, stability is a significant challenge with state-of-the-art deep RL meth-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 353, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 505, + 366 + ], + "score": 1.0, + "content": "ods, and is very important for being able to reliably use deep RL for real world tasks. In the rest of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 364, + 453, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 453, + 378 + ], + "score": 1.0, + "content": "the experiments, we will use conservative Q-Prop as the main Q-Prop implementation.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 130, + 439, + 284, + 544 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 130, + 439, + 284, + 544 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 130, + 439, + 284, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 284, + 544 + ], + "score": 0.731, + "type": "image", + "image_path": "d75d5db85e34384abd1e96bb494b10087f6138234cb0f906d3e4c4dba89c33ff.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 130, + 439, + 284, + 452.125 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 130, + 452.125, + 284, + 465.25 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 130, + 465.25, + 284, + 478.375 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 130, + 478.375, + 284, + 491.5 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 130, + 491.5, + 284, + 504.625 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 130, + 504.625, + 284, + 517.75 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 130, + 517.75, + 284, + 530.875 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 130, + 530.875, + 284, + 544.0 + ], + "spans": [], + "index": 25 + } + ] + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 124, + 551, + 290, + 562 + ], + "lines": [ + { + "bbox": [ + 122, + 550, + 292, + 563 + ], + "spans": [ + { + "bbox": [ + 122, + 550, + 292, + 563 + ], + "score": 1.0, + "content": "(a) Comparing algorithms on HalfCheetah-v1.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 326, + 440, + 480, + 545 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 326, + 440, + 480, + 545 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 326, + 440, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 326, + 440, + 480, + 545 + ], + "score": 0.747, + "type": "image", + "image_path": "a609c411ab9a5449d74b3fe5b80daffd18f515d8845ee6448717ce07addc7bd5.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 326, + 440, + 480, + 453.125 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 326, + 453.125, + 480, + 466.25 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 326, + 466.25, + 480, + 479.375 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 326, + 479.375, + 480, + 492.5 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 326, + 492.5, + 480, + 505.625 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 326, + 505.625, + 480, + 518.75 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 326, + 518.75, + 480, + 531.875 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 326, + 531.875, + 480, + 545.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 325, + 551, + 483, + 562 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 322, + 550, + 485, + 563 + ], + "spans": [ + { + "bbox": [ + 322, + 550, + 485, + 563 + ], + "score": 1.0, + "content": "(b) Comparing algorithms on Humanoid-v1.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "image_caption", + "bbox": [ + 106, + 571, + 505, + 605 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 106, + 571, + 504, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 504, + 584 + ], + "score": 1.0, + "content": "Figure 3: Average return over episodes in HalfCheetah-v1 and Humanoid-v1 during learning, com-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "paring Q-Prop against other model-free algorithms. Q-Prop with vanilla policy gradient outperforms", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 592, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 606 + ], + "score": 1.0, + "content": "TRPO on HalfCheetah. Q-Prop significantly outperforms TRPO in convergence time on Humanoid.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 622, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "In this section, we evaluate two versions of conservative Q-Prop, v-c-Q-Prop using vanilla pol-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "icy gradient and TR-c-Q-Prop using trust-region updates, against other model-free algorithms on", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "the HalfCheetah-v1 domain. Figure 3a shows that c-Q-Prop methods significantly outperform the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "best TRPO and VPG methods. Even Q-Prop with vanilla policy gradient is comparable to TRPO,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "confirming the significant benefits from variance reduction. DDPG on the other hand exhibits incon-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "sistent performances. With proper reward scaling, i.e. “DDPG-r0.1”, it outperforms other methods", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "as well as the DDPG results reported in prior work (Duan et al., 2016; Amos et al., 2016). This", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "illustrates the sensitivity of DDPG to hyperparameter settings, while Q-Prop exhibits more stable,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "monotonic learning behaviors when compared to DDPG. In the next section we show this improved", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 721, + 397, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 397, + 733 + ], + "score": 1.0, + "content": "stability allows Q-Prop to outperform DDPG in more complex domains.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 43.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 401, + 285, + 412 + ], + "lines": [ + { + "bbox": [ + 106, + 400, + 286, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 286, + 413 + ], + "score": 1.0, + "content": "5.2 EVALUATION ACROSS ALGORITHMS", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 131, + 82, + 283, + 188 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 131, + 82, + 283, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 131, + 82, + 283, + 188 + ], + "spans": [ + { + "bbox": [ + 131, + 82, + 283, + 188 + ], + "score": 0.651, + "type": "image", + "image_path": "b8ef0cab31ceb57d2ca2f3840e44751e97cf886ce5ce81bc26d79da92a4d50c5.jpg" + } + ] + } + ], + "index": 1.0, + "virtual_lines": [ + { + "bbox": [ + 131, + 82, + 283, + 135.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 131, + 135.0, + 283, + 188.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 133, + 193, + 282, + 204 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 282, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 282, + 205 + ], + "score": 1.0, + "content": "(a) Standard Q-Prop vs adaptive variants.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 327, + 82, + 479, + 186 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 327, + 82, + 479, + 186 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 327, + 82, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 327, + 82, + 479, + 186 + ], + "score": 0.851, + "type": "image", + "image_path": "ac3e5492e7295b46b13f17d48d9490a7f429f569aab0d8bfb64532dc2013b0fc.jpg" + } + ] + } + ], + "index": 2.0, + "virtual_lines": [ + { + "bbox": [ + 327, + 82, + 479, + 134.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 327, + 134.0, + 479, + 186.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 307, + 193, + 500, + 204 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 306, + 192, + 501, + 205 + ], + "spans": [ + { + "bbox": [ + 306, + 192, + 501, + 205 + ], + "score": 1.0, + "content": "(b) Conservative Q-Prop vs TRPO across batch sizes.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "image_caption", + "bbox": [ + 106, + 213, + 505, + 257 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 212, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 226 + ], + "score": 1.0, + "content": "Figure 2: Average return over episodes in HalfCheetah-v1 during learning, exploring adaptive Q-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 223, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 236 + ], + "score": 1.0, + "content": "Prop methods and different batch sizes. All variants of Q-Prop substantially outperform TRPO in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 235, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 505, + 247 + ], + "score": 1.0, + "content": "terms of sample efficiency. TR-c-QP, conservative Q-Prop with trust-region update performs most", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 246, + 245, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 245, + 258 + ], + "score": 1.0, + "content": "stably across different batch sizes.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 281, + 505, + 337 + ], + "lines": [ + { + "bbox": [ + 106, + 282, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 294 + ], + "score": 1.0, + "content": "Figure 2b shows the performance of conservative Q-Prop against TRPO across different batch", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 292, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 505, + 306 + ], + "score": 1.0, + "content": "sizes. Due to high variance in gradient estimates, TRPO typically requires very large batch sizes,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 303, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 104, + 303, + 506, + 317 + ], + "score": 1.0, + "content": "e.g. 25000 steps or 25 episodes per update, to perform well. We show that our Q-Prop methods can", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "learn even with just 1 episode per update, and achieves better sample efficiency with small batch", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 325, + 490, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 490, + 339 + ], + "score": 1.0, + "content": "sizes. This shows that Q-Prop significantly reduces the variance compared to the prior methods.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12, + "bbox_fs": [ + 104, + 282, + 506, + 339 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 342, + 505, + 376 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "As we discussed in Section 1, stability is a significant challenge with state-of-the-art deep RL meth-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 353, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 505, + 366 + ], + "score": 1.0, + "content": "ods, and is very important for being able to reliably use deep RL for real world tasks. In the rest of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 364, + 453, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 453, + 378 + ], + "score": 1.0, + "content": "the experiments, we will use conservative Q-Prop as the main Q-Prop implementation.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 342, + 505, + 378 + ] + }, + { + "type": "image", + "bbox": [ + 130, + 439, + 284, + 544 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 130, + 439, + 284, + 544 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 130, + 439, + 284, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 284, + 544 + ], + "score": 0.731, + "type": "image", + "image_path": "d75d5db85e34384abd1e96bb494b10087f6138234cb0f906d3e4c4dba89c33ff.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 130, + 439, + 284, + 452.125 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 130, + 452.125, + 284, + 465.25 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 130, + 465.25, + 284, + 478.375 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 130, + 478.375, + 284, + 491.5 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 130, + 491.5, + 284, + 504.625 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 130, + 504.625, + 284, + 517.75 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 130, + 517.75, + 284, + 530.875 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 130, + 530.875, + 284, + 544.0 + ], + "spans": [], + "index": 25 + } + ] + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 124, + 551, + 290, + 562 + ], + "lines": [ + { + "bbox": [ + 122, + 550, + 292, + 563 + ], + "spans": [ + { + "bbox": [ + 122, + 550, + 292, + 563 + ], + "score": 1.0, + "content": "(a) Comparing algorithms on HalfCheetah-v1.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34, + "bbox_fs": [ + 122, + 550, + 292, + 563 + ] + }, + { + "type": "image", + "bbox": [ + 326, + 440, + 480, + 545 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 326, + 440, + 480, + 545 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 326, + 440, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 326, + 440, + 480, + 545 + ], + "score": 0.747, + "type": "image", + "image_path": "a609c411ab9a5449d74b3fe5b80daffd18f515d8845ee6448717ce07addc7bd5.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 326, + 440, + 480, + 453.125 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 326, + 453.125, + 480, + 466.25 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 326, + 466.25, + 480, + 479.375 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 326, + 479.375, + 480, + 492.5 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 326, + 492.5, + 480, + 505.625 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 326, + 505.625, + 480, + 518.75 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 326, + 518.75, + 480, + 531.875 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 326, + 531.875, + 480, + 545.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 325, + 551, + 483, + 562 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 322, + 550, + 485, + 563 + ], + "spans": [ + { + "bbox": [ + 322, + 550, + 485, + 563 + ], + "score": 1.0, + "content": "(b) Comparing algorithms on Humanoid-v1.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "image_caption", + "bbox": [ + 106, + 571, + 505, + 605 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 106, + 571, + 504, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 504, + 584 + ], + "score": 1.0, + "content": "Figure 3: Average return over episodes in HalfCheetah-v1 and Humanoid-v1 during learning, com-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "paring Q-Prop against other model-free algorithms. Q-Prop with vanilla policy gradient outperforms", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 592, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 606 + ], + "score": 1.0, + "content": "TRPO on HalfCheetah. Q-Prop significantly outperforms TRPO in convergence time on Humanoid.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 622, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "In this section, we evaluate two versions of conservative Q-Prop, v-c-Q-Prop using vanilla pol-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "icy gradient and TR-c-Q-Prop using trust-region updates, against other model-free algorithms on", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "the HalfCheetah-v1 domain. Figure 3a shows that c-Q-Prop methods significantly outperform the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "best TRPO and VPG methods. Even Q-Prop with vanilla policy gradient is comparable to TRPO,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "confirming the significant benefits from variance reduction. DDPG on the other hand exhibits incon-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "sistent performances. With proper reward scaling, i.e. “DDPG-r0.1”, it outperforms other methods", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "as well as the DDPG results reported in prior work (Duan et al., 2016; Amos et al., 2016). This", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "illustrates the sensitivity of DDPG to hyperparameter settings, while Q-Prop exhibits more stable,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "monotonic learning behaviors when compared to DDPG. In the next section we show this improved", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 721, + 397, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 397, + 733 + ], + "score": 1.0, + "content": "stability allows Q-Prop to outperform DDPG in more complex domains.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 621, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 266, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 267, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 267, + 95 + ], + "score": 1.0, + "content": "5.3 EVALUATION ACROSS DOMAINS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 104, + 505, + 181 + ], + "lines": [ + { + "bbox": [ + 104, + 103, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 505, + 118 + ], + "score": 1.0, + "content": "Lastly, we evaluate Q-Prop against TRPO and DDPG across multiple domains. While the gym", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "score": 1.0, + "content": "environments are biased toward locomotion, we expect we can achieve similar performance on ma-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "nipulation tasks such as those in Lillicrap et al. (2016). Table 1 summarizes the results, including the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "best attained average rewards and the steps to convergence. Q-Prop consistently outperform TRPO", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "in terms of sample complexity and sometimes achieves higher rewards than DDPG in more complex", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "score": 1.0, + "content": "domains. A particularly notable case is shown in Figure 3b, where Q-Prop substantially improves", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 169, + 499, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 499, + 183 + ], + "score": 1.0, + "content": "sample efficiency over TRPO on Humanoid-v1 domain, while DDPG cannot find a good solution.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 187, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 186, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 199 + ], + "score": 1.0, + "content": "The better performance on the more complex domains highlights the importance of stable deep RL", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 197, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 505, + 210 + ], + "score": 1.0, + "content": "algorithms: while costly hyperparameter sweeps may allow even less stable algorithms to perform", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "well on simpler problems, more complex tasks might have such narrow regions of stable hyperpa-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 220, + 319, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 319, + 232 + ], + "score": 1.0, + "content": "rameters that discovering them becomes impractical.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "table", + "bbox": [ + 106, + 242, + 511, + 337 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 242, + 511, + 337 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 242, + 511, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 511, + 337 + ], + "score": 0.98, + "html": "
TR-c-Q-PropTRPODDPG
DomainThresholdMaxReturn.EpisodesMaxReturnEpsisodesMaxReturnEpisodes
Ant350035344975423913825957N/A
HalfCheetah47004811207854734263707490600
Hopper200029575945248657152604965
Humanoid2500>349214750918>30000552N/A
Reacher-7-6.02060-6.72840-6.61800
Swimmer9010320451103025150500
Walker30004030368535671887536262125
", + "type": "table", + "image_path": "727463be8b5bbf2c6633678d0ac605de2691434167ca6598b21e0e8b7985947c.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 106, + 242, + 511, + 273.6666666666667 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 106, + 273.6666666666667, + 511, + 305.33333333333337 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 106, + 305.33333333333337, + 511, + 337.00000000000006 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 344, + 506, + 377 + ], + "lines": [ + { + "bbox": [ + 105, + 344, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 506, + 357 + ], + "score": 1.0, + "content": "Table 1: Q-Prop, TRPO and DDPG results showing the max average rewards attained in the first", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 355, + 123, + 366 + ], + "score": 0.34, + "content": "3 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "episodes and the episodes to cross specific reward thresholds. Q-Prop often learns more sample", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 366, + 476, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 476, + 378 + ], + "score": 1.0, + "content": "efficiently than TRPO and can solve difficult domains such as Humanoid better than DDPG.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 108, + 397, + 287, + 409 + ], + "lines": [ + { + "bbox": [ + 104, + 394, + 290, + 412 + ], + "spans": [ + { + "bbox": [ + 104, + 394, + 290, + 412 + ], + "score": 1.0, + "content": "6 DISCUSSION AND CONCLUSION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 505, + 522 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 436 + ], + "score": 1.0, + "content": "We presented Q-Prop, a policy gradient algorithm that combines reliable, consistent, and poten-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "score": 1.0, + "content": "tially unbiased on-policy gradient estimation with a sample-efficient off-policy critic that acts as a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "score": 1.0, + "content": "control variate. The method provides a large improvement in sample efficiency compared to state-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 457, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 468 + ], + "score": 1.0, + "content": "of-the-art policy gradient methods such as TRPO, while outperforming state-of-the-art actor-critic", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "methods on more challenging tasks such as humanoid locomotion. We hope that techniques like", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 477, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 491 + ], + "score": 1.0, + "content": "these, which combine on-policy Monte Carlo gradient estimation with sample-efficient variance re-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 489, + 504, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 504, + 501 + ], + "score": 1.0, + "content": "duction through off-policy critics, will eventually lead to deep reinforcement learning algorithms", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "that are more stable and efficient, and therefore better suited for application to complex real-world", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 511, + 166, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 166, + 524 + ], + "score": 1.0, + "content": "learning tasks.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 108, + 538, + 200, + 548 + ], + "lines": [ + { + "bbox": [ + 107, + 538, + 200, + 549 + ], + "spans": [ + { + "bbox": [ + 107, + 538, + 200, + 549 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 557, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 557, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 506, + 570 + ], + "score": 1.0, + "content": "We thank Rocky Duan for sharing and answering questions about rllab code, and Yutian Chen and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "Laurent Dinh for discussion on control variates. SG and RT were funded by NSERC, Google, and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "EPSRC grants EP/L000776/1 and EP/M026957/1. ZG was funded by EPSRC grant EP/J012300/1", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 590, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 295, + 603 + ], + "score": 1.0, + "content": "and the Alan Turing Institute (EP/N510129/1).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "title", + "bbox": [ + 108, + 621, + 175, + 633 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 176, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 176, + 634 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 641, + 503, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 640, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 505, + 654 + ], + "score": 1.0, + "content": "Brandon Amos, Lei Xu, and J Zico Kolter. Input convex neural networks. arXiv preprint", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 651, + 219, + 663 + ], + "spans": [ + { + "bbox": [ + 115, + 651, + 219, + 663 + ], + "score": 1.0, + "content": "arXiv:1609.07152, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 105, + 675, + 504, + 698 + ], + "lines": [ + { + "bbox": [ + 107, + 676, + 504, + 687 + ], + "spans": [ + { + "bbox": [ + 107, + 676, + 504, + 687 + ], + "score": 1.0, + "content": "Christopher G Atkeson and Juan Carlos Santamaria. A comparison of direct and model-based rein-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 686, + 503, + 698 + ], + "spans": [ + { + "bbox": [ + 116, + 686, + 503, + 698 + ], + "score": 1.0, + "content": "forcement learning. In In International Conference on Robotics and Automation. Citeseer, 1997.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 720, + 415, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 415, + 733 + ], + "score": 1.0, + "content": "Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 266, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 267, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 267, + 95 + ], + "score": 1.0, + "content": "5.3 EVALUATION ACROSS DOMAINS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 104, + 505, + 181 + ], + "lines": [ + { + "bbox": [ + 104, + 103, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 505, + 118 + ], + "score": 1.0, + "content": "Lastly, we evaluate Q-Prop against TRPO and DDPG across multiple domains. While the gym", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "score": 1.0, + "content": "environments are biased toward locomotion, we expect we can achieve similar performance on ma-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "nipulation tasks such as those in Lillicrap et al. (2016). Table 1 summarizes the results, including the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "best attained average rewards and the steps to convergence. Q-Prop consistently outperform TRPO", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "in terms of sample complexity and sometimes achieves higher rewards than DDPG in more complex", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "score": 1.0, + "content": "domains. A particularly notable case is shown in Figure 3b, where Q-Prop substantially improves", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 169, + 499, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 499, + 183 + ], + "score": 1.0, + "content": "sample efficiency over TRPO on Humanoid-v1 domain, while DDPG cannot find a good solution.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4, + "bbox_fs": [ + 104, + 103, + 505, + 183 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 187, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 186, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 199 + ], + "score": 1.0, + "content": "The better performance on the more complex domains highlights the importance of stable deep RL", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 197, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 505, + 210 + ], + "score": 1.0, + "content": "algorithms: while costly hyperparameter sweeps may allow even less stable algorithms to perform", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "well on simpler problems, more complex tasks might have such narrow regions of stable hyperpa-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 220, + 319, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 319, + 232 + ], + "score": 1.0, + "content": "rameters that discovering them becomes impractical.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 186, + 505, + 232 + ] + }, + { + "type": "table", + "bbox": [ + 106, + 242, + 511, + 337 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 242, + 511, + 337 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 242, + 511, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 511, + 337 + ], + "score": 0.98, + "html": "
TR-c-Q-PropTRPODDPG
DomainThresholdMaxReturn.EpisodesMaxReturnEpsisodesMaxReturnEpisodes
Ant350035344975423913825957N/A
HalfCheetah47004811207854734263707490600
Hopper200029575945248657152604965
Humanoid2500>349214750918>30000552N/A
Reacher-7-6.02060-6.72840-6.61800
Swimmer9010320451103025150500
Walker30004030368535671887536262125
", + "type": "table", + "image_path": "727463be8b5bbf2c6633678d0ac605de2691434167ca6598b21e0e8b7985947c.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 106, + 242, + 511, + 273.6666666666667 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 106, + 273.6666666666667, + 511, + 305.33333333333337 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 106, + 305.33333333333337, + 511, + 337.00000000000006 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 344, + 506, + 377 + ], + "lines": [ + { + "bbox": [ + 105, + 344, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 506, + 357 + ], + "score": 1.0, + "content": "Table 1: Q-Prop, TRPO and DDPG results showing the max average rewards attained in the first", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 355, + 123, + 366 + ], + "score": 0.34, + "content": "3 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "episodes and the episodes to cross specific reward thresholds. Q-Prop often learns more sample", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 366, + 476, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 476, + 378 + ], + "score": 1.0, + "content": "efficiently than TRPO and can solve difficult domains such as Humanoid better than DDPG.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 344, + 506, + 378 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 397, + 287, + 409 + ], + "lines": [ + { + "bbox": [ + 104, + 394, + 290, + 412 + ], + "spans": [ + { + "bbox": [ + 104, + 394, + 290, + 412 + ], + "score": 1.0, + "content": "6 DISCUSSION AND CONCLUSION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 505, + 522 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 436 + ], + "score": 1.0, + "content": "We presented Q-Prop, a policy gradient algorithm that combines reliable, consistent, and poten-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "score": 1.0, + "content": "tially unbiased on-policy gradient estimation with a sample-efficient off-policy critic that acts as a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "score": 1.0, + "content": "control variate. The method provides a large improvement in sample efficiency compared to state-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 457, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 468 + ], + "score": 1.0, + "content": "of-the-art policy gradient methods such as TRPO, while outperforming state-of-the-art actor-critic", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "methods on more challenging tasks such as humanoid locomotion. We hope that techniques like", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 477, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 491 + ], + "score": 1.0, + "content": "these, which combine on-policy Monte Carlo gradient estimation with sample-efficient variance re-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 489, + 504, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 504, + 501 + ], + "score": 1.0, + "content": "duction through off-policy critics, will eventually lead to deep reinforcement learning algorithms", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "that are more stable and efficient, and therefore better suited for application to complex real-world", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 511, + 166, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 166, + 524 + ], + "score": 1.0, + "content": "learning tasks.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 423, + 506, + 524 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 538, + 200, + 548 + ], + "lines": [ + { + "bbox": [ + 107, + 538, + 200, + 549 + ], + "spans": [ + { + "bbox": [ + 107, + 538, + 200, + 549 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 557, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 557, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 506, + 570 + ], + "score": 1.0, + "content": "We thank Rocky Duan for sharing and answering questions about rllab code, and Yutian Chen and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "Laurent Dinh for discussion on control variates. SG and RT were funded by NSERC, Google, and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "EPSRC grants EP/L000776/1 and EP/M026957/1. ZG was funded by EPSRC grant EP/J012300/1", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 590, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 295, + 603 + ], + "score": 1.0, + "content": "and the Alan Turing Institute (EP/N510129/1).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 557, + 506, + 603 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 621, + 175, + 633 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 176, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 176, + 634 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 641, + 503, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 640, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 505, + 654 + ], + "score": 1.0, + "content": "Brandon Amos, Lei Xu, and J Zico Kolter. Input convex neural networks. arXiv preprint", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 651, + 219, + 663 + ], + "spans": [ + { + "bbox": [ + 115, + 651, + 219, + 663 + ], + "score": 1.0, + "content": "arXiv:1609.07152, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 640, + 505, + 663 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 675, + 504, + 698 + ], + "lines": [ + { + "bbox": [ + 107, + 676, + 504, + 687 + ], + "spans": [ + { + "bbox": [ + 107, + 676, + 504, + 687 + ], + "score": 1.0, + "content": "Christopher G Atkeson and Juan Carlos Santamaria. A comparison of direct and model-based rein-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 686, + 503, + 698 + ], + "spans": [ + { + "bbox": [ + 116, + 686, + 503, + 698 + ], + "score": 1.0, + "content": "forcement learning. In In International Conference on Robotics and Automation. Citeseer, 1997.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 107, + 676, + 504, + 698 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 720, + 415, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 415, + 733 + ], + "score": 1.0, + "content": "Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 106, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 505, + 96 + ], + "score": 1.0, + "content": "Marc Deisenroth and Carl E Rasmussen. Pilco: A model-based and data-efficient approach to policy", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 505, + 108 + ], + "score": 1.0, + "content": "search. In Proceedings of the 28th International Conference on machine learning (ICML-11), pp.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 181, + 116 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 181, + 116 + ], + "score": 1.0, + "content": "465–472, 2011.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 503, + 156 + ], + "lines": [ + { + "bbox": [ + 105, + 121, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 136 + ], + "score": 1.0, + "content": "Yan Duan, Xi Chen, Rein Houthooft, John Schulman, and Pieter Abbeel. Benchmarking deep", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 133, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 133, + 505, + 147 + ], + "score": 1.0, + "content": "reinforcement learning for continuous control. International Conference on Machine Learning", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 145, + 176, + 156 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 176, + 156 + ], + "score": 1.0, + "content": "(ICML), 2016.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 105, + 163, + 504, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 505, + 176 + ], + "score": 1.0, + "content": "Miroslav Dud´ık, John Langford, and Lihong Li. Doubly robust policy evaluation and learning. arXiv", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 175, + 249, + 186 + ], + "spans": [ + { + "bbox": [ + 114, + 175, + 249, + 186 + ], + "score": 1.0, + "content": "preprint arXiv:1103.4601, 2011.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 108, + 192, + 505, + 226 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 506, + 206 + ], + "score": 1.0, + "content": "Evan Greensmith, Peter L Bartlett, and Jonathan Baxter. Variance reduction techniques for gradient", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 204, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 115, + 204, + 505, + 216 + ], + "score": 1.0, + "content": "estimates in reinforcement learning. Journal of Machine Learning Research, 5(Nov):1471–1530,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 114, + 214, + 143, + 227 + ], + "spans": [ + { + "bbox": [ + 114, + 214, + 143, + 227 + ], + "score": 1.0, + "content": "2004.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 233, + 505, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 233, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 505, + 247 + ], + "score": 1.0, + "content": "Shixiang Gu, Sergey Levine, Ilya Sutskever, and Andriy Mnih. Muprop: Unbiased backpropagation", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 115, + 244, + 505, + 257 + ], + "score": 1.0, + "content": "for stochastic neural networks. International Conference on Learning Representations (ICLR),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 253, + 147, + 269 + ], + "spans": [ + { + "bbox": [ + 115, + 253, + 147, + 269 + ], + "score": 1.0, + "content": "2016a.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 273, + 503, + 297 + ], + "lines": [ + { + "bbox": [ + 106, + 273, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 505, + 286 + ], + "score": 1.0, + "content": "Shixiang Gu, Tim Lillicrap, Ilya Sutskever, and Sergey Levine. Continuous deep q-learning with", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 285, + 491, + 297 + ], + "spans": [ + { + "bbox": [ + 116, + 285, + 491, + 297 + ], + "score": 1.0, + "content": "model-based acceleration. In International Conference on Machine Learning (ICML), 2016b.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 303, + 503, + 326 + ], + "lines": [ + { + "bbox": [ + 104, + 300, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 104, + 300, + 505, + 318 + ], + "score": 1.0, + "content": "Hado V Hasselt. Double q-learning. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 314, + 191, + 326 + ], + "spans": [ + { + "bbox": [ + 116, + 314, + 191, + 326 + ], + "score": 1.0, + "content": "2613–2621, 2010.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 332, + 447, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 332, + 448, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 448, + 346 + ], + "score": 1.0, + "content": "Sham Kakade. A natural policy gradient. In NIPS, volume 14, pp. 1531–1538, 2001.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 105, + 351, + 504, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 365 + ], + "score": 1.0, + "content": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 362, + 216, + 374 + ], + "spans": [ + { + "bbox": [ + 115, + 362, + 216, + 374 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 381, + 347, + 393 + ], + "lines": [ + { + "bbox": [ + 106, + 380, + 347, + 395 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 347, + 395 + ], + "score": 1.0, + "content": "Guy Lever. Deterministic policy gradient algorithms. 2014.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 105, + 399, + 504, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 398, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 506, + 413 + ], + "score": 1.0, + "content": "Sergey Levine and Vladlen Koltun. Guided policy search. In International Conference on Machine", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 410, + 251, + 423 + ], + "spans": [ + { + "bbox": [ + 115, + 410, + 251, + 423 + ], + "score": 1.0, + "content": "Learning (ICML), pp. 1–9, 2013.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 429, + 504, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 443 + ], + "score": 1.0, + "content": "Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 116, + 440, + 505, + 453 + ], + "score": 1.0, + "content": "David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. Interna-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 117, + 451, + 366, + 464 + ], + "spans": [ + { + "bbox": [ + 117, + 451, + 366, + 464 + ], + "score": 1.0, + "content": "tional Conference on Learning Representations (ICLR), 2016.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 469, + 505, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 483 + ], + "score": 1.0, + "content": "A Rupam Mahmood, Hado P van Hasselt, and Richard S Sutton. Weighted importance sampling", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 481, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 116, + 481, + 505, + 493 + ], + "score": 1.0, + "content": "for off-policy learning with linear function approximation. In Advances in Neural Information", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 492, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 116, + 492, + 289, + 504 + ], + "score": 1.0, + "content": "Processing Systems, pp. 3014–3022, 2014.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 103, + 510, + 504, + 533 + ], + "lines": [ + { + "bbox": [ + 106, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "Andriy Mnih and Karol Gregor. Neural variational inference and learning in belief networks. Inter-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 522, + 349, + 533 + ], + "spans": [ + { + "bbox": [ + 115, + 522, + 349, + 533 + ], + "score": 1.0, + "content": "national Conference on Machine Learning (ICML), 2014.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 539, + 505, + 574 + ], + "lines": [ + { + "bbox": [ + 106, + 539, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 505, + 553 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Belle-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 551, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 116, + 551, + 505, + 563 + ], + "score": 1.0, + "content": "mare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 562, + 441, + 574 + ], + "spans": [ + { + "bbox": [ + 116, + 562, + 441, + 574 + ], + "score": 1.0, + "content": "control through deep reinforcement learning. Nature, 518(7540):529–533, 2015.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 579, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 505, + 594 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy P Lillicrap, Tim", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 115, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 602, + 419, + 615 + ], + "spans": [ + { + "bbox": [ + 115, + 602, + 419, + 615 + ], + "score": 1.0, + "content": "learning. In International Conference on Machine Learning (ICML), 2016.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 620, + 504, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 620, + 504, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 504, + 633 + ], + "score": 1.0, + "content": "Remi Munos, Tom Stepleton, Anna Harutyunyan, and Marc G Bellemare. Safe and efficient off-´", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 632, + 404, + 644 + ], + "spans": [ + { + "bbox": [ + 115, + 632, + 404, + 644 + ], + "score": 1.0, + "content": "policy reinforcement learning. arXiv preprint arXiv:1606.02647, 2016.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 503, + 673 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 504, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 504, + 663 + ], + "score": 1.0, + "content": "John Paisley, David Blei, and Michael Jordan. Variational bayesian inference with stochastic search.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 661, + 369, + 673 + ], + "spans": [ + { + "bbox": [ + 116, + 661, + 369, + 673 + ], + "score": 1.0, + "content": "International Conference on Machine Learning (ICML), 2012.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 680, + 503, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 679, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 505, + 693 + ], + "score": 1.0, + "content": "Jan Peters and Stefan Schaal. Policy gradient methods for robotics. In International Conference on", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 691, + 390, + 703 + ], + "spans": [ + { + "bbox": [ + 116, + 691, + 390, + 703 + ], + "score": 1.0, + "content": "Intelligent Robots and Systems (IROS), pp. 2219–2225. IEEE, 2006.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "Jan Peters, Katharina Mulling, and Yasemin Altun. Relative entropy policy search. In ¨ AAAI. Atlanta,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 720, + 142, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 142, + 732 + ], + "score": 1.0, + "content": "2010.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 505, + 96 + ], + "score": 1.0, + "content": "Marc Deisenroth and Carl E Rasmussen. Pilco: A model-based and data-efficient approach to policy", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 505, + 108 + ], + "score": 1.0, + "content": "search. In Proceedings of the 28th International Conference on machine learning (ICML-11), pp.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 181, + 116 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 181, + 116 + ], + "score": 1.0, + "content": "465–472, 2011.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 80, + 505, + 116 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 503, + 156 + ], + "lines": [ + { + "bbox": [ + 105, + 121, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 136 + ], + "score": 1.0, + "content": "Yan Duan, Xi Chen, Rein Houthooft, John Schulman, and Pieter Abbeel. Benchmarking deep", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 133, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 133, + 505, + 147 + ], + "score": 1.0, + "content": "reinforcement learning for continuous control. International Conference on Machine Learning", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 145, + 176, + 156 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 176, + 156 + ], + "score": 1.0, + "content": "(ICML), 2016.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 121, + 505, + 156 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 163, + 504, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 505, + 176 + ], + "score": 1.0, + "content": "Miroslav Dud´ık, John Langford, and Lihong Li. Doubly robust policy evaluation and learning. arXiv", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 175, + 249, + 186 + ], + "spans": [ + { + "bbox": [ + 114, + 175, + 249, + 186 + ], + "score": 1.0, + "content": "preprint arXiv:1103.4601, 2011.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 162, + 505, + 186 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 192, + 505, + 226 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 506, + 206 + ], + "score": 1.0, + "content": "Evan Greensmith, Peter L Bartlett, and Jonathan Baxter. Variance reduction techniques for gradient", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 204, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 115, + 204, + 505, + 216 + ], + "score": 1.0, + "content": "estimates in reinforcement learning. Journal of Machine Learning Research, 5(Nov):1471–1530,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 114, + 214, + 143, + 227 + ], + "spans": [ + { + "bbox": [ + 114, + 214, + 143, + 227 + ], + "score": 1.0, + "content": "2004.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 191, + 506, + 227 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 233, + 505, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 233, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 505, + 247 + ], + "score": 1.0, + "content": "Shixiang Gu, Sergey Levine, Ilya Sutskever, and Andriy Mnih. Muprop: Unbiased backpropagation", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 115, + 244, + 505, + 257 + ], + "score": 1.0, + "content": "for stochastic neural networks. International Conference on Learning Representations (ICLR),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 253, + 147, + 269 + ], + "spans": [ + { + "bbox": [ + 115, + 253, + 147, + 269 + ], + "score": 1.0, + "content": "2016a.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 233, + 505, + 269 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 273, + 503, + 297 + ], + "lines": [ + { + "bbox": [ + 106, + 273, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 505, + 286 + ], + "score": 1.0, + "content": "Shixiang Gu, Tim Lillicrap, Ilya Sutskever, and Sergey Levine. Continuous deep q-learning with", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 285, + 491, + 297 + ], + "spans": [ + { + "bbox": [ + 116, + 285, + 491, + 297 + ], + "score": 1.0, + "content": "model-based acceleration. In International Conference on Machine Learning (ICML), 2016b.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 273, + 505, + 297 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 303, + 503, + 326 + ], + "lines": [ + { + "bbox": [ + 104, + 300, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 104, + 300, + 505, + 318 + ], + "score": 1.0, + "content": "Hado V Hasselt. Double q-learning. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 314, + 191, + 326 + ], + "spans": [ + { + "bbox": [ + 116, + 314, + 191, + 326 + ], + "score": 1.0, + "content": "2613–2621, 2010.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 104, + 300, + 505, + 326 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 332, + 447, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 332, + 448, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 448, + 346 + ], + "score": 1.0, + "content": "Sham Kakade. A natural policy gradient. In NIPS, volume 14, pp. 1531–1538, 2001.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 332, + 448, + 346 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 351, + 504, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 365 + ], + "score": 1.0, + "content": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 362, + 216, + 374 + ], + "spans": [ + { + "bbox": [ + 115, + 362, + 216, + 374 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 349, + 506, + 374 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 381, + 347, + 393 + ], + "lines": [ + { + "bbox": [ + 106, + 380, + 347, + 395 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 347, + 395 + ], + "score": 1.0, + "content": "Guy Lever. Deterministic policy gradient algorithms. 2014.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 380, + 347, + 395 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 399, + 504, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 398, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 506, + 413 + ], + "score": 1.0, + "content": "Sergey Levine and Vladlen Koltun. Guided policy search. In International Conference on Machine", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 410, + 251, + 423 + ], + "spans": [ + { + "bbox": [ + 115, + 410, + 251, + 423 + ], + "score": 1.0, + "content": "Learning (ICML), pp. 1–9, 2013.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 398, + 506, + 423 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 429, + 504, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 443 + ], + "score": 1.0, + "content": "Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 116, + 440, + 505, + 453 + ], + "score": 1.0, + "content": "David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. Interna-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 117, + 451, + 366, + 464 + ], + "spans": [ + { + "bbox": [ + 117, + 451, + 366, + 464 + ], + "score": 1.0, + "content": "tional Conference on Learning Representations (ICLR), 2016.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 428, + 506, + 464 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 469, + 505, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 483 + ], + "score": 1.0, + "content": "A Rupam Mahmood, Hado P van Hasselt, and Richard S Sutton. Weighted importance sampling", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 481, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 116, + 481, + 505, + 493 + ], + "score": 1.0, + "content": "for off-policy learning with linear function approximation. In Advances in Neural Information", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 492, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 116, + 492, + 289, + 504 + ], + "score": 1.0, + "content": "Processing Systems, pp. 3014–3022, 2014.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 468, + 505, + 504 + ] + }, + { + "type": "text", + "bbox": [ + 103, + 510, + 504, + 533 + ], + "lines": [ + { + "bbox": [ + 106, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "Andriy Mnih and Karol Gregor. Neural variational inference and learning in belief networks. Inter-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 522, + 349, + 533 + ], + "spans": [ + { + "bbox": [ + 115, + 522, + 349, + 533 + ], + "score": 1.0, + "content": "national Conference on Machine Learning (ICML), 2014.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 510, + 505, + 533 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 539, + 505, + 574 + ], + "lines": [ + { + "bbox": [ + 106, + 539, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 505, + 553 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Belle-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 551, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 116, + 551, + 505, + 563 + ], + "score": 1.0, + "content": "mare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 562, + 441, + 574 + ], + "spans": [ + { + "bbox": [ + 116, + 562, + 441, + 574 + ], + "score": 1.0, + "content": "control through deep reinforcement learning. Nature, 518(7540):529–533, 2015.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 106, + 539, + 505, + 574 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 579, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 505, + 594 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy P Lillicrap, Tim", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 115, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 602, + 419, + 615 + ], + "spans": [ + { + "bbox": [ + 115, + 602, + 419, + 615 + ], + "score": 1.0, + "content": "learning. In International Conference on Machine Learning (ICML), 2016.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 106, + 579, + 505, + 615 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 620, + 504, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 620, + 504, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 504, + 633 + ], + "score": 1.0, + "content": "Remi Munos, Tom Stepleton, Anna Harutyunyan, and Marc G Bellemare. Safe and efficient off-´", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 632, + 404, + 644 + ], + "spans": [ + { + "bbox": [ + 115, + 632, + 404, + 644 + ], + "score": 1.0, + "content": "policy reinforcement learning. arXiv preprint arXiv:1606.02647, 2016.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 106, + 620, + 504, + 644 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 503, + 673 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 504, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 504, + 663 + ], + "score": 1.0, + "content": "John Paisley, David Blei, and Michael Jordan. Variational bayesian inference with stochastic search.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 661, + 369, + 673 + ], + "spans": [ + { + "bbox": [ + 116, + 661, + 369, + 673 + ], + "score": 1.0, + "content": "International Conference on Machine Learning (ICML), 2012.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 650, + 504, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 680, + 503, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 679, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 505, + 693 + ], + "score": 1.0, + "content": "Jan Peters and Stefan Schaal. Policy gradient methods for robotics. In International Conference on", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 691, + 390, + 703 + ], + "spans": [ + { + "bbox": [ + 116, + 691, + 390, + 703 + ], + "score": 1.0, + "content": "Intelligent Robots and Systems (IROS), pp. 2219–2225. IEEE, 2006.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 679, + 505, + 703 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "Jan Peters, Katharina Mulling, and Yasemin Altun. Relative entropy policy search. In ¨ AAAI. Atlanta,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 720, + 142, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 142, + 732 + ], + "score": 1.0, + "content": "2010.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 708, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Doina Precup. Eligibility traces for off-policy policy evaluation. Computer Science Department", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 284, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 284, + 106 + ], + "score": 1.0, + "content": "Faculty Publication Series, pp. 80, 2000.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 114, + 358, + 127 + ], + "lines": [ + { + "bbox": [ + 106, + 114, + 358, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 358, + 128 + ], + "score": 1.0, + "content": "Sheldon M Ross. Simulation. Burlington, MA: Elsevier, 2006.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 135, + 506, + 169 + ], + "lines": [ + { + "bbox": [ + 105, + 135, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 505, + 149 + ], + "score": 1.0, + "content": "John Schulman, Sergey Levine, Pieter Abbeel, Michael I. Jordan, and Philipp Moritz. Trust region", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 147, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 115, + 147, + 505, + 159 + ], + "score": 1.0, + "content": "policy optimization. In International Conference on Machine Learning (ICML), pp. 1889–1897,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 156, + 144, + 171 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 144, + 171 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 179, + 505, + 213 + ], + "lines": [ + { + "bbox": [ + 105, + 179, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 505, + 192 + ], + "score": 1.0, + "content": "John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 190, + 504, + 203 + ], + "spans": [ + { + "bbox": [ + 115, + 190, + 504, + 203 + ], + "score": 1.0, + "content": "dimensional continuous control using generalized advantage estimation. International Confer-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 202, + 313, + 213 + ], + "spans": [ + { + "bbox": [ + 115, + 202, + 313, + 213 + ], + "score": 1.0, + "content": "ence on Learning Representations (ICLR), 2016.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 222, + 505, + 255 + ], + "lines": [ + { + "bbox": [ + 106, + 222, + 504, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 504, + 235 + ], + "score": 1.0, + "content": "David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. De-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 233, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 115, + 233, + 506, + 247 + ], + "score": 1.0, + "content": "terministic policy gradient algorithms. In International Conference on Machine Learning (ICML),", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 114, + 242, + 143, + 257 + ], + "spans": [ + { + "bbox": [ + 114, + 242, + 143, + 257 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 265, + 505, + 299 + ], + "lines": [ + { + "bbox": [ + 105, + 264, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 279 + ], + "score": 1.0, + "content": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 276, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 115, + 276, + 505, + 290 + ], + "score": 1.0, + "content": "Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 117, + 288, + 493, + 300 + ], + "spans": [ + { + "bbox": [ + 117, + 288, + 493, + 300 + ], + "score": 1.0, + "content": "the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 309, + 503, + 342 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "score": 1.0, + "content": "Richard S Sutton. Integrated architectures for learning, planning, and reacting based on approxi-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 319, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 115, + 319, + 505, + 334 + ], + "score": 1.0, + "content": "mating dynamic programming. In International Conference on Machine Learning (ICML), pp.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 331, + 181, + 342 + ], + "spans": [ + { + "bbox": [ + 116, + 331, + 181, + 342 + ], + "score": 1.0, + "content": "216–224, 1990.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 505, + 386 + ], + "lines": [ + { + "bbox": [ + 105, + 351, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 365 + ], + "score": 1.0, + "content": "Richard S Sutton, David A McAllester, Satinder P Singh, Yishay Mansour, et al. Policy gradient", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 117, + 363, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 117, + 363, + 505, + 375 + ], + "score": 1.0, + "content": "methods for reinforcement learning with function approximation. In Advances in Neural Infor-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 375, + 396, + 387 + ], + "spans": [ + { + "bbox": [ + 116, + 375, + 396, + 387 + ], + "score": 1.0, + "content": "mation Processing Systems (NIPS), volume 99, pp. 1057–1063, 1999.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 395, + 505, + 440 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "score": 1.0, + "content": "Richard S Sutton, Hamid Reza Maei, Doina Precup, Shalabh Bhatnagar, David Silver, Csaba", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 406, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 115, + 406, + 505, + 420 + ], + "score": 1.0, + "content": "Szepesvari, and Eric Wiewiora. Fast gradient-descent methods for temporal-difference learning ´", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 115, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "with linear function approximation. In Proceedings of the 26th Annual International Conference", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 429, + 320, + 441 + ], + "spans": [ + { + "bbox": [ + 115, + 429, + 320, + 441 + ], + "score": 1.0, + "content": "on Machine Learning, pp. 993–1000. ACM, 2009.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 449, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 106, + 449, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 505, + 462 + ], + "score": 1.0, + "content": "Richard S Sutton, A Rupam Mahmood, and Martha White. An emphatic approach to the problem", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 460, + 492, + 473 + ], + "spans": [ + { + "bbox": [ + 115, + 460, + 492, + 473 + ], + "score": 1.0, + "content": "of off-policy temporal-difference learning. The Journal of Machine Learning Research, 2015.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 481, + 445, + 494 + ], + "lines": [ + { + "bbox": [ + 106, + 481, + 446, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 481, + 446, + 495 + ], + "score": 1.0, + "content": "Philip Thomas. Bias in natural actor-critic algorithms. In ICML, pp. 441–448, 2014.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 505, + 537 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "score": 1.0, + "content": "Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 514, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 115, + 514, + 505, + 527 + ], + "score": 1.0, + "content": "In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 5026–5033.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 525, + 168, + 537 + ], + "spans": [ + { + "bbox": [ + 116, + 525, + 168, + 537 + ], + "score": 1.0, + "content": "IEEE, 2012.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 546, + 499, + 559 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 500, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 500, + 560 + ], + "score": 1.0, + "content": "Christopher JCH Watkins and Peter Dayan. Q-learning. Machine learning, 8(3-4):279–292, 1992.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 108, + 567, + 504, + 602 + ], + "lines": [ + { + "bbox": [ + 104, + 566, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 104, + 566, + 506, + 582 + ], + "score": 1.0, + "content": "Lex Weaver and Nigel Tao. The optimal reward baseline for gradient-based reinforcement learning.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 115, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "In Proceedings of the Seventeenth conference on Uncertainty in artificial intelligence, pp. 538–", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 117, + 591, + 303, + 602 + ], + "spans": [ + { + "bbox": [ + 117, + 591, + 303, + 602 + ], + "score": 1.0, + "content": "545. Morgan Kaufmann Publishers Inc., 2001.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 611, + 504, + 633 + ], + "lines": [ + { + "bbox": [ + 106, + 611, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 624 + ], + "score": 1.0, + "content": "Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 622, + 323, + 635 + ], + "spans": [ + { + "bbox": [ + 116, + 622, + 323, + 635 + ], + "score": 1.0, + "content": "learning. Machine learning, 8(3-4):229–256, 1992.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "title", + "bbox": [ + 108, + 657, + 300, + 670 + ], + "lines": [ + { + "bbox": [ + 106, + 657, + 302, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 302, + 672 + ], + "score": 1.0, + "content": "A Q-PROP ESTIMATOR DERIVATION", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 682, + 504, + 705 + ], + "lines": [ + { + "bbox": [ + 105, + 681, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 696 + ], + "score": 1.0, + "content": "The full derivation of the Q-Prop estimator is shown in Eq. 14. We make use of the following", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 694, + 330, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 330, + 706 + ], + "score": 1.0, + "content": "property that is commonly used in baseline derivations:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 712, + 411, + 736 + ], + "lines": [ + { + "bbox": [ + 199, + 712, + 411, + 736 + ], + "spans": [ + { + "bbox": [ + 199, + 712, + 411, + 736 + ], + "score": 0.91, + "content": "\\mathbb { E } _ { p _ { \\theta } ( x ) } [ \\nabla _ { \\theta } \\log p _ { \\theta } ( x ) ] = \\int _ { x } \\nabla _ { \\theta } p _ { \\theta } ( x ) = \\nabla _ { \\theta } \\int _ { x } p ( x ) = 0", + "type": "interline_equation", + "image_path": "d098e81861b78d720d0cc9ebc52a79d42144e032693153f767c944347369f851.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 199, + 712, + 411, + 736 + ], + "spans": [], + "index": 40 + } + ] + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Doina Precup. Eligibility traces for off-policy policy evaluation. Computer Science Department", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 284, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 284, + 106 + ], + "score": 1.0, + "content": "Faculty Publication Series, pp. 80, 2000.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 114, + 358, + 127 + ], + "lines": [ + { + "bbox": [ + 106, + 114, + 358, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 358, + 128 + ], + "score": 1.0, + "content": "Sheldon M Ross. Simulation. Burlington, MA: Elsevier, 2006.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 106, + 114, + 358, + 128 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 135, + 506, + 169 + ], + "lines": [ + { + "bbox": [ + 105, + 135, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 505, + 149 + ], + "score": 1.0, + "content": "John Schulman, Sergey Levine, Pieter Abbeel, Michael I. Jordan, and Philipp Moritz. Trust region", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 147, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 115, + 147, + 505, + 159 + ], + "score": 1.0, + "content": "policy optimization. In International Conference on Machine Learning (ICML), pp. 1889–1897,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 156, + 144, + 171 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 144, + 171 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 135, + 505, + 171 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 179, + 505, + 213 + ], + "lines": [ + { + "bbox": [ + 105, + 179, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 505, + 192 + ], + "score": 1.0, + "content": "John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 190, + 504, + 203 + ], + "spans": [ + { + "bbox": [ + 115, + 190, + 504, + 203 + ], + "score": 1.0, + "content": "dimensional continuous control using generalized advantage estimation. International Confer-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 202, + 313, + 213 + ], + "spans": [ + { + "bbox": [ + 115, + 202, + 313, + 213 + ], + "score": 1.0, + "content": "ence on Learning Representations (ICLR), 2016.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 179, + 505, + 213 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 222, + 505, + 255 + ], + "lines": [ + { + "bbox": [ + 106, + 222, + 504, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 504, + 235 + ], + "score": 1.0, + "content": "David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. De-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 233, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 115, + 233, + 506, + 247 + ], + "score": 1.0, + "content": "terministic policy gradient algorithms. In International Conference on Machine Learning (ICML),", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 114, + 242, + 143, + 257 + ], + "spans": [ + { + "bbox": [ + 114, + 242, + 143, + 257 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 106, + 222, + 506, + 257 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 265, + 505, + 299 + ], + "lines": [ + { + "bbox": [ + 105, + 264, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 279 + ], + "score": 1.0, + "content": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 276, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 115, + 276, + 505, + 290 + ], + "score": 1.0, + "content": "Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 117, + 288, + 493, + 300 + ], + "spans": [ + { + "bbox": [ + 117, + 288, + 493, + 300 + ], + "score": 1.0, + "content": "the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 264, + 505, + 300 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 309, + 503, + 342 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "score": 1.0, + "content": "Richard S Sutton. Integrated architectures for learning, planning, and reacting based on approxi-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 319, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 115, + 319, + 505, + 334 + ], + "score": 1.0, + "content": "mating dynamic programming. In International Conference on Machine Learning (ICML), pp.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 331, + 181, + 342 + ], + "spans": [ + { + "bbox": [ + 116, + 331, + 181, + 342 + ], + "score": 1.0, + "content": "216–224, 1990.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 308, + 505, + 342 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 505, + 386 + ], + "lines": [ + { + "bbox": [ + 105, + 351, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 365 + ], + "score": 1.0, + "content": "Richard S Sutton, David A McAllester, Satinder P Singh, Yishay Mansour, et al. Policy gradient", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 117, + 363, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 117, + 363, + 505, + 375 + ], + "score": 1.0, + "content": "methods for reinforcement learning with function approximation. In Advances in Neural Infor-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 375, + 396, + 387 + ], + "spans": [ + { + "bbox": [ + 116, + 375, + 396, + 387 + ], + "score": 1.0, + "content": "mation Processing Systems (NIPS), volume 99, pp. 1057–1063, 1999.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 351, + 505, + 387 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 395, + 505, + 440 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "score": 1.0, + "content": "Richard S Sutton, Hamid Reza Maei, Doina Precup, Shalabh Bhatnagar, David Silver, Csaba", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 406, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 115, + 406, + 505, + 420 + ], + "score": 1.0, + "content": "Szepesvari, and Eric Wiewiora. Fast gradient-descent methods for temporal-difference learning ´", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 115, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "with linear function approximation. In Proceedings of the 26th Annual International Conference", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 429, + 320, + 441 + ], + "spans": [ + { + "bbox": [ + 115, + 429, + 320, + 441 + ], + "score": 1.0, + "content": "on Machine Learning, pp. 993–1000. ACM, 2009.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 394, + 505, + 441 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 449, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 106, + 449, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 505, + 462 + ], + "score": 1.0, + "content": "Richard S Sutton, A Rupam Mahmood, and Martha White. An emphatic approach to the problem", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 460, + 492, + 473 + ], + "spans": [ + { + "bbox": [ + 115, + 460, + 492, + 473 + ], + "score": 1.0, + "content": "of off-policy temporal-difference learning. The Journal of Machine Learning Research, 2015.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 449, + 505, + 473 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 481, + 445, + 494 + ], + "lines": [ + { + "bbox": [ + 106, + 481, + 446, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 481, + 446, + 495 + ], + "score": 1.0, + "content": "Philip Thomas. Bias in natural actor-critic algorithms. In ICML, pp. 441–448, 2014.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 481, + 446, + 495 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 505, + 537 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "score": 1.0, + "content": "Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 514, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 115, + 514, + 505, + 527 + ], + "score": 1.0, + "content": "In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 5026–5033.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 525, + 168, + 537 + ], + "spans": [ + { + "bbox": [ + 116, + 525, + 168, + 537 + ], + "score": 1.0, + "content": "IEEE, 2012.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 502, + 505, + 537 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 546, + 499, + 559 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 500, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 500, + 560 + ], + "score": 1.0, + "content": "Christopher JCH Watkins and Peter Dayan. Q-learning. Machine learning, 8(3-4):279–292, 1992.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31, + "bbox_fs": [ + 106, + 546, + 500, + 560 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 567, + 504, + 602 + ], + "lines": [ + { + "bbox": [ + 104, + 566, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 104, + 566, + 506, + 582 + ], + "score": 1.0, + "content": "Lex Weaver and Nigel Tao. The optimal reward baseline for gradient-based reinforcement learning.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 115, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "In Proceedings of the Seventeenth conference on Uncertainty in artificial intelligence, pp. 538–", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 117, + 591, + 303, + 602 + ], + "spans": [ + { + "bbox": [ + 117, + 591, + 303, + 602 + ], + "score": 1.0, + "content": "545. Morgan Kaufmann Publishers Inc., 2001.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 104, + 566, + 506, + 602 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 611, + 504, + 633 + ], + "lines": [ + { + "bbox": [ + 106, + 611, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 624 + ], + "score": 1.0, + "content": "Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 622, + 323, + 635 + ], + "spans": [ + { + "bbox": [ + 116, + 622, + 323, + 635 + ], + "score": 1.0, + "content": "learning. Machine learning, 8(3-4):229–256, 1992.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 106, + 611, + 505, + 635 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 657, + 300, + 670 + ], + "lines": [ + { + "bbox": [ + 106, + 657, + 302, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 302, + 672 + ], + "score": 1.0, + "content": "A Q-PROP ESTIMATOR DERIVATION", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 682, + 504, + 705 + ], + "lines": [ + { + "bbox": [ + 105, + 681, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 696 + ], + "score": 1.0, + "content": "The full derivation of the Q-Prop estimator is shown in Eq. 14. We make use of the following", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 694, + 330, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 330, + 706 + ], + "score": 1.0, + "content": "property that is commonly used in baseline derivations:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 681, + 506, + 706 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 712, + 411, + 736 + ], + "lines": [ + { + "bbox": [ + 199, + 712, + 411, + 736 + ], + "spans": [ + { + "bbox": [ + 199, + 712, + 411, + 736 + ], + "score": 0.91, + "content": "\\mathbb { E } _ { p _ { \\theta } ( x ) } [ \\nabla _ { \\theta } \\log p _ { \\theta } ( x ) ] = \\int _ { x } \\nabla _ { \\theta } p _ { \\theta } ( x ) = \\nabla _ { \\theta } \\int _ { x } p ( x ) = 0", + "type": "interline_equation", + "image_path": "d098e81861b78d720d0cc9ebc52a79d42144e032693153f767c944347369f851.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 199, + 712, + 411, + 736 + ], + "spans": [], + "index": 40 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 194, + 95 + ], + "score": 1.0, + "content": "This holds true when", + "type": "text" + }, + { + "bbox": [ + 194, + 82, + 230, + 95 + ], + "score": 0.93, + "content": "f ( \\mathbf { \\boldsymbol { s } } _ { t } , \\mathbf { \\boldsymbol { a } } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 82, + 446, + 95 + ], + "score": 1.0, + "content": "is an arbitrary function differentiable with respect to", + "type": "text" + }, + { + "bbox": [ + 446, + 84, + 456, + 93 + ], + "score": 0.82, + "content": "\\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { a } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 82, + 474, + 95 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 475, + 82, + 482, + 94 + ], + "score": 0.85, + "content": "\\bar { f }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "is its", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 252, + 107 + ], + "score": 1.0, + "content": "first-order Taylor expansion around", + "type": "text" + }, + { + "bbox": [ + 252, + 95, + 284, + 105 + ], + "score": 0.91, + "content": "\\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf ", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 93, + 305, + 107 + ], + "score": 1.0, + "content": ", i.e.", + "type": "text" + }, + { + "bbox": [ + 305, + 93, + 502, + 106 + ], + "score": 0.89, + "content": "\\bar { f } ( s _ { t } , a _ { t } ) = f ( s _ { t } , \\bar { a } _ { t } ) + \\nabla _ { a } \\bar { f ( } s _ { t } , a ) | _ { a = \\bar { a } _ { t } } ( a _ { t } - \\bar { a } _ { t } ) .", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 93, + 505, + 107 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 468, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 131, + 118 + ], + "score": 1.0, + "content": "Here,", + "type": "text" + }, + { + "bbox": [ + 131, + 104, + 199, + 117 + ], + "score": 0.94, + "content": "\\mu _ { \\theta } ( \\pmb { s } _ { t } ) \\dot { = } \\mathbb { E } _ { \\pi } [ \\bar { \\mathbf { a } } _ { t } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 103, + 328, + 118 + ], + "score": 1.0, + "content": "is the mean of stochastic policy", + "type": "text" + }, + { + "bbox": [ + 328, + 106, + 339, + 116 + ], + "score": 0.85, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 103, + 468, + 118 + ], + "score": 1.0, + "content": ". The derivation appears below:", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "interline_equation", + "bbox": [ + 113, + 122, + 479, + 300 + ], + "lines": [ + { + "bbox": [ + 113, + 122, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 113, + 122, + 479, + 300 + ], + "score": 0.95, + "content": "\\begin{array} { r l } { \\nabla _ { \\theta } J ( \\theta ) = \\mathrm { E } _ { \\mathrm { g } , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { 1 } | s _ { 1 } \\big ) \\big ( \\hat { Q } ( s _ { t } , a _ { t } ) - \\hat { J } ( s _ { t } , a _ { t } ) \\big ) \\big ] + \\mathrm { E } _ { \\rho , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { t } | s _ { t } \\big ) \\hat { J } ( s _ { t } , a _ { t } ) \\big ] } \\\\ { g ( \\theta ) = \\mathrm { E } _ { \\mathrm { g } , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { 1 } | s _ { 1 } \\big ) \\hat { J } ( s _ { t } , a _ { t } ) \\big ] } \\\\ & { = \\mathrm { E } _ { \\rho , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { 1 } | s _ { 1 } \\big ) \\big ( \\hat { J } ( s _ { t } , a _ { t } ) + \\nabla _ { a _ { t } } \\big ) \\big ( s _ { t } , a _ { t } \\big ) \\big ] } \\\\ & { = \\mathrm { E } _ { \\rho , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { 1 } | s _ { 1 } \\big ) \\mathrm { e } _ { n } f \\big ( s _ { t } , a _ { t } \\big ) \\big ] = - \\alpha _ { 1 } \\alpha _ { 1 } \\alpha _ { 1 } } \\\\ & { = \\mathrm { E } _ { \\rho , \\pi } \\bigg [ \\int _ { a _ { t } } \\mathrm { e } _ { g } \\pi _ { \\theta } \\big ( a _ { t } | s _ { t } \\big ) \\mathrm { e } _ { n } f \\big ( s _ { t } , a _ { t } \\big ) \\big ] \\omega - \\alpha _ { 2 } \\alpha _ { 2 } \\alpha _ { 2 } \\bigg ] } \\\\ & { = \\mathrm { E } _ { \\rho , \\pi } \\bigg [ \\sum _ { \\theta } \\mathrm { e } _ { g } \\pi _ { \\theta } \\big ( a _ { t } \\big ) \\big ( \\frac { \\kappa } { \\alpha _ { 1 } } \\big ) \\mathrm { e } _ { n } f \\big ( \\alpha _ { 1 } \\big ) \\omega _ { 1 } \\omega _ { 2 } \\bigg ] } \\\\ & = \\mathrm { E } _ { \\rho , \\pi } \\bigg [ \\mathrm { v } _ { \\alpha } f \\big ( s _ { t } , a \\big ) \\big | \\alpha - \\alpha _ { 1 } \\int _ a _ { t } \\end{array}", + "type": "interline_equation", + "image_path": "c9025713638b6bd3ac0c25065e2b04ac97b90ae6567399821a9170713a729d10.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 113, + 122, + 479, + 181.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 113, + 181.33333333333334, + 479, + 240.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 113, + 240.66666666666669, + 479, + 300.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 309, + 442, + 337 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 444, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 444, + 324 + ], + "score": 1.0, + "content": "B CONNECTION BETWEEN Q-PROP AND COMPATIBLE FEATURE", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 325, + 216, + 337 + ], + "spans": [ + { + "bbox": [ + 127, + 325, + 216, + 337 + ], + "score": 1.0, + "content": "APPROXIMATION", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 348, + 506, + 406 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "In this section we show that actor-critic with compatible feature approximation is a form of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 359, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 212, + 373 + ], + "score": 1.0, + "content": "control variate. A critic", + "type": "text" + }, + { + "bbox": [ + 213, + 361, + 227, + 371 + ], + "score": 0.87, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 359, + 450, + 373 + ], + "score": 1.0, + "content": "is compatible (Sutton et al., 1999) if it satisfies (1)", + "type": "text" + }, + { + "bbox": [ + 450, + 360, + 505, + 372 + ], + "score": 0.89, + "content": "Q _ { w } ( \\pmb { s } _ { t } , \\pmb { a } _ { t } ) =", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 368, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 185, + 383 + ], + "score": 0.91, + "content": "w ^ { T } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( { \\mathbf { \\em a } } _ { t } | { \\mathbf { \\em s } } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 368, + 212, + 385 + ], + "score": 1.0, + "content": ", i.e.", + "type": "text" + }, + { + "bbox": [ + 212, + 371, + 348, + 383 + ], + "score": 0.88, + "content": "\\nabla _ { w } \\mathcal { Q } _ { w } ( s _ { t } , \\bar { a } _ { t } ) = \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 368, + 387, + 385 + ], + "score": 1.0, + "content": ", and (2)", + "type": "text" + }, + { + "bbox": [ + 388, + 372, + 397, + 381 + ], + "score": 0.63, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 368, + 484, + 385 + ], + "score": 1.0, + "content": "is fit with objective", + "type": "text" + }, + { + "bbox": [ + 485, + 373, + 505, + 382 + ], + "score": 0.83, + "content": "w =", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 382, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 104, + 382, + 113, + 397 + ], + "score": 1.0, + "content": "a", + "type": "text" + }, + { + "bbox": [ + 113, + 383, + 342, + 396 + ], + "score": 0.73, + "content": "\\mathrm { r g } \\operatorname* { m i n } _ { w } L ( w ) = \\mathrm { a r g } \\operatorname* { m i n } _ { w } \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ ( \\hat { Q } ( s _ { t } , a _ { t } ) - Q _ { w } ( { \\pmb s } _ { t } , { \\pmb a } _ { t } ) ) ^ { 2 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 382, + 404, + 397 + ], + "score": 1.0, + "content": ", that is fitting", + "type": "text" + }, + { + "bbox": [ + 404, + 384, + 418, + 395 + ], + "score": 0.88, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 382, + 506, + 397 + ], + "score": 1.0, + "content": "on on-policy Monte", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 393, + 344, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 344, + 408 + ], + "score": 1.0, + "content": "Carlo returns. Condition (2) implies the following identity,", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 182, + 411, + 429, + 427 + ], + "lines": [ + { + "bbox": [ + 182, + 411, + 429, + 427 + ], + "spans": [ + { + "bbox": [ + 182, + 411, + 429, + 427 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\nabla _ { w } L = 2 \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { \\mathcal { Q } } ( s _ { t } , a _ { t } ) - { \\mathcal Q } _ { w } ( s _ { t } , a _ { t } ) ) ] = 0 . } \\end{array}", + "type": "interline_equation", + "image_path": "df93697e3d7e251afed3d6d2c3874531f249deeba1bf5f574eb2bd592d544520.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 182, + 411, + 429, + 427 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 433, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 106, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 324, + 446 + ], + "score": 1.0, + "content": "In compatible feature approximation, it directly uses", + "type": "text" + }, + { + "bbox": [ + 325, + 434, + 338, + 444 + ], + "score": 0.89, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "as control variate, rather than its Taylor", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 443, + 419, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 149, + 457 + ], + "score": 1.0, + "content": "expansion", + "type": "text" + }, + { + "bbox": [ + 150, + 443, + 163, + 455 + ], + "score": 0.9, + "content": "\\bar { Q } _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 444, + 419, + 457 + ], + "score": 1.0, + "content": "as in Q-Prop. Using Eq. 15, the Monte Carlo policy gradient is,", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "interline_equation", + "bbox": [ + 192, + 459, + 418, + 506 + ], + "lines": [ + { + "bbox": [ + 192, + 459, + 418, + 506 + ], + "spans": [ + { + "bbox": [ + 192, + 459, + 418, + 506 + ], + "score": 0.92, + "content": "\\begin{array} { r l } & { \\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ) Q _ { w } ( \\boldsymbol { s } _ { t } , \\boldsymbol { a } _ { t } ) ] } \\\\ & { \\qquad = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ ( \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ) \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ) ^ { T } ) w ] } \\\\ & { \\qquad = \\mathbb { E } _ { \\rho _ { \\pi } } [ I ( \\theta ; \\boldsymbol { s } _ { t } ) w ] , } \\end{array}", + "type": "interline_equation", + "image_path": "190a4f5c01b0e37dffa0037afb35be6b86f71f49759e6a8105d1953d9e424153.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 192, + 459, + 418, + 474.6666666666667 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 192, + 474.6666666666667, + 418, + 490.33333333333337 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 192, + 490.33333333333337, + 418, + 506.00000000000006 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 510, + 502, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 134, + 524 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 511, + 334, + 524 + ], + "score": 0.85, + "content": "I ( \\theta ; s _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } \\vert s _ { t } ) \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } \\vert s _ { t } ) ^ { T } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "is Fisher’s information matrix. Thus, vari-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 522, + 441, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 336, + 535 + ], + "score": 1.0, + "content": "ance reduction depends on ability to compute or estimate", + "type": "text" + }, + { + "bbox": [ + 336, + 523, + 367, + 534 + ], + "score": 0.93, + "content": "I ( \\theta ; s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 522, + 385, + 535 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 386, + 524, + 394, + 532 + ], + "score": 0.78, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 522, + 441, + 535 + ], + "score": 1.0, + "content": "effectively.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "title", + "bbox": [ + 107, + 549, + 388, + 563 + ], + "lines": [ + { + "bbox": [ + 106, + 549, + 389, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 389, + 564 + ], + "score": 1.0, + "content": "C UNIFYING POLICY GRADIENT AND ACTOR-CRITIC", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 574, + 504, + 598 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 588 + ], + "score": 1.0, + "content": "Q-Prop closely ties together policy gradient and actor-critic algorithms. To analyze this point, we", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 585, + 437, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 405, + 600 + ], + "score": 1.0, + "content": "write a generalization of Eq. 9 below, introducing two additional variables", + "type": "text" + }, + { + "bbox": [ + 405, + 587, + 432, + 598 + ], + "score": 0.9, + "content": "\\alpha , \\rho _ { C R }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 585, + 437, + 600 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "interline_equation", + "bbox": [ + 184, + 602, + 426, + 635 + ], + "lines": [ + { + "bbox": [ + 184, + 602, + 426, + 635 + ], + "spans": [ + { + "bbox": [ + 184, + 602, + 426, + 635 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { \\nabla _ { \\theta } J ( \\theta ) \\propto \\alpha \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { \\boldsymbol A } ( s _ { t } , a _ { t } ) - \\eta \\bar { \\boldsymbol A } _ { w } ( s _ { t } , a _ { t } ) ] } \\\\ & { \\qquad + \\eta \\mathbb { E } _ { \\rho _ { C R } } [ \\nabla _ { a } \\mathcal { Q } _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] } \\end{array}", + "type": "interline_equation", + "image_path": "81cb6159a91a178254bf4a790617f7e216698fa5af04494a2545828fb52a3a93.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 184, + 602, + 426, + 618.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 184, + 618.5, + 426, + 635.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 637, + 505, + 705 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 420, + 651 + ], + "score": 1.0, + "content": "Eq. 17 enables more analysis where bias generally is introduced only when", + "type": "text" + }, + { + "bbox": [ + 421, + 638, + 447, + 649 + ], + "score": 0.91, + "content": "\\alpha \\neq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 637, + 460, + 651 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 461, + 639, + 501, + 650 + ], + "score": 0.92, + "content": "\\rho _ { C R } \\neq \\rho _ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 637, + 505, + 651 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 649, + 504, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 504, + 660 + ], + "score": 1.0, + "content": "Importantly, Eq. 17 covers both policy gradient and deterministic actor-critic algorithm as its special", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 660, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 303, + 673 + ], + "score": 1.0, + "content": "cases. Standard policy gradient is recovered by", + "type": "text" + }, + { + "bbox": [ + 303, + 660, + 329, + 672 + ], + "score": 0.9, + "content": "\\eta = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 660, + 506, + 673 + ], + "score": 1.0, + "content": ", and deterministic actor-critic is recovered", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 669, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 120, + 685 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 120, + 671, + 147, + 681 + ], + "score": 0.89, + "content": "\\alpha = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 669, + 167, + 685 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 167, + 672, + 208, + 684 + ], + "score": 0.9, + "content": "\\rho _ { C R } = \\rho _ { \\beta }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 669, + 506, + 685 + ], + "score": 1.0, + "content": ". This allows heuristic or automatic methods for dynamically changing", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 680, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 506, + 695 + ], + "score": 1.0, + "content": "these variables through the learning process for optimizing different metrics, e.g. sample efficiency,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 693, + 223, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 223, + 705 + ], + "score": 1.0, + "content": "convergence speed, stability.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "Table 2 summarizes the various edge cases of Eq. 17. For example, since we derive our method from", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 223, + 733 + ], + "score": 1.0, + "content": "a control variates standpoint,", + "type": "text" + }, + { + "bbox": [ + 224, + 721, + 237, + 732 + ], + "score": 0.89, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "can be any function and the gradient remains almost unbiased (see", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 194, + 95 + ], + "score": 1.0, + "content": "This holds true when", + "type": "text" + }, + { + "bbox": [ + 194, + 82, + 230, + 95 + ], + "score": 0.93, + "content": "f ( \\mathbf { \\boldsymbol { s } } _ { t } , \\mathbf { \\boldsymbol { a } } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 82, + 446, + 95 + ], + "score": 1.0, + "content": "is an arbitrary function differentiable with respect to", + "type": "text" + }, + { + "bbox": [ + 446, + 84, + 456, + 93 + ], + "score": 0.82, + "content": "\\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { a } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 82, + 474, + 95 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 475, + 82, + 482, + 94 + ], + "score": 0.85, + "content": "\\bar { f }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "is its", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 252, + 107 + ], + "score": 1.0, + "content": "first-order Taylor expansion around", + "type": "text" + }, + { + "bbox": [ + 252, + 95, + 284, + 105 + ], + "score": 0.91, + "content": "\\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf ", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 93, + 305, + 107 + ], + "score": 1.0, + "content": ", i.e.", + "type": "text" + }, + { + "bbox": [ + 305, + 93, + 502, + 106 + ], + "score": 0.89, + "content": "\\bar { f } ( s _ { t } , a _ { t } ) = f ( s _ { t } , \\bar { a } _ { t } ) + \\nabla _ { a } \\bar { f ( } s _ { t } , a ) | _ { a = \\bar { a } _ { t } } ( a _ { t } - \\bar { a } _ { t } ) .", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 93, + 505, + 107 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 468, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 131, + 118 + ], + "score": 1.0, + "content": "Here,", + "type": "text" + }, + { + "bbox": [ + 131, + 104, + 199, + 117 + ], + "score": 0.94, + "content": "\\mu _ { \\theta } ( \\pmb { s } _ { t } ) \\dot { = } \\mathbb { E } _ { \\pi } [ \\bar { \\mathbf { a } } _ { t } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 103, + 328, + 118 + ], + "score": 1.0, + "content": "is the mean of stochastic policy", + "type": "text" + }, + { + "bbox": [ + 328, + 106, + 339, + 116 + ], + "score": 0.85, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 103, + 468, + 118 + ], + "score": 1.0, + "content": ". The derivation appears below:", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 82, + 505, + 118 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 113, + 122, + 479, + 300 + ], + "lines": [ + { + "bbox": [ + 113, + 122, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 113, + 122, + 479, + 300 + ], + "score": 0.95, + "content": "\\begin{array} { r l } { \\nabla _ { \\theta } J ( \\theta ) = \\mathrm { E } _ { \\mathrm { g } , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { 1 } | s _ { 1 } \\big ) \\big ( \\hat { Q } ( s _ { t } , a _ { t } ) - \\hat { J } ( s _ { t } , a _ { t } ) \\big ) \\big ] + \\mathrm { E } _ { \\rho , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { t } | s _ { t } \\big ) \\hat { J } ( s _ { t } , a _ { t } ) \\big ] } \\\\ { g ( \\theta ) = \\mathrm { E } _ { \\mathrm { g } , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { 1 } | s _ { 1 } \\big ) \\hat { J } ( s _ { t } , a _ { t } ) \\big ] } \\\\ & { = \\mathrm { E } _ { \\rho , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { 1 } | s _ { 1 } \\big ) \\big ( \\hat { J } ( s _ { t } , a _ { t } ) + \\nabla _ { a _ { t } } \\big ) \\big ( s _ { t } , a _ { t } \\big ) \\big ] } \\\\ & { = \\mathrm { E } _ { \\rho , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { 1 } | s _ { 1 } \\big ) \\mathrm { e } _ { n } f \\big ( s _ { t } , a _ { t } \\big ) \\big ] = - \\alpha _ { 1 } \\alpha _ { 1 } \\alpha _ { 1 } } \\\\ & { = \\mathrm { E } _ { \\rho , \\pi } \\bigg [ \\int _ { a _ { t } } \\mathrm { e } _ { g } \\pi _ { \\theta } \\big ( a _ { t } | s _ { t } \\big ) \\mathrm { e } _ { n } f \\big ( s _ { t } , a _ { t } \\big ) \\big ] \\omega - \\alpha _ { 2 } \\alpha _ { 2 } \\alpha _ { 2 } \\bigg ] } \\\\ & { = \\mathrm { E } _ { \\rho , \\pi } \\bigg [ \\sum _ { \\theta } \\mathrm { e } _ { g } \\pi _ { \\theta } \\big ( a _ { t } \\big ) \\big ( \\frac { \\kappa } { \\alpha _ { 1 } } \\big ) \\mathrm { e } _ { n } f \\big ( \\alpha _ { 1 } \\big ) \\omega _ { 1 } \\omega _ { 2 } \\bigg ] } \\\\ & = \\mathrm { E } _ { \\rho , \\pi } \\bigg [ \\mathrm { v } _ { \\alpha } f \\big ( s _ { t } , a \\big ) \\big | \\alpha - \\alpha _ { 1 } \\int _ a _ { t } \\end{array}", + "type": "interline_equation", + "image_path": "c9025713638b6bd3ac0c25065e2b04ac97b90ae6567399821a9170713a729d10.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 113, + 122, + 479, + 181.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 113, + 181.33333333333334, + 479, + 240.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 113, + 240.66666666666669, + 479, + 300.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 309, + 442, + 337 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 444, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 444, + 324 + ], + "score": 1.0, + "content": "B CONNECTION BETWEEN Q-PROP AND COMPATIBLE FEATURE", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 325, + 216, + 337 + ], + "spans": [ + { + "bbox": [ + 127, + 325, + 216, + 337 + ], + "score": 1.0, + "content": "APPROXIMATION", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 348, + 506, + 406 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "In this section we show that actor-critic with compatible feature approximation is a form of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 359, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 212, + 373 + ], + "score": 1.0, + "content": "control variate. A critic", + "type": "text" + }, + { + "bbox": [ + 213, + 361, + 227, + 371 + ], + "score": 0.87, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 359, + 450, + 373 + ], + "score": 1.0, + "content": "is compatible (Sutton et al., 1999) if it satisfies (1)", + "type": "text" + }, + { + "bbox": [ + 450, + 360, + 505, + 372 + ], + "score": 0.89, + "content": "Q _ { w } ( \\pmb { s } _ { t } , \\pmb { a } _ { t } ) =", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 368, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 185, + 383 + ], + "score": 0.91, + "content": "w ^ { T } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( { \\mathbf { \\em a } } _ { t } | { \\mathbf { \\em s } } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 368, + 212, + 385 + ], + "score": 1.0, + "content": ", i.e.", + "type": "text" + }, + { + "bbox": [ + 212, + 371, + 348, + 383 + ], + "score": 0.88, + "content": "\\nabla _ { w } \\mathcal { Q } _ { w } ( s _ { t } , \\bar { a } _ { t } ) = \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 368, + 387, + 385 + ], + "score": 1.0, + "content": ", and (2)", + "type": "text" + }, + { + "bbox": [ + 388, + 372, + 397, + 381 + ], + "score": 0.63, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 368, + 484, + 385 + ], + "score": 1.0, + "content": "is fit with objective", + "type": "text" + }, + { + "bbox": [ + 485, + 373, + 505, + 382 + ], + "score": 0.83, + "content": "w =", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 382, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 104, + 382, + 113, + 397 + ], + "score": 1.0, + "content": "a", + "type": "text" + }, + { + "bbox": [ + 113, + 383, + 342, + 396 + ], + "score": 0.73, + "content": "\\mathrm { r g } \\operatorname* { m i n } _ { w } L ( w ) = \\mathrm { a r g } \\operatorname* { m i n } _ { w } \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ ( \\hat { Q } ( s _ { t } , a _ { t } ) - Q _ { w } ( { \\pmb s } _ { t } , { \\pmb a } _ { t } ) ) ^ { 2 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 382, + 404, + 397 + ], + "score": 1.0, + "content": ", that is fitting", + "type": "text" + }, + { + "bbox": [ + 404, + 384, + 418, + 395 + ], + "score": 0.88, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 382, + 506, + 397 + ], + "score": 1.0, + "content": "on on-policy Monte", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 393, + 344, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 344, + 408 + ], + "score": 1.0, + "content": "Carlo returns. Condition (2) implies the following identity,", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10, + "bbox_fs": [ + 104, + 349, + 506, + 408 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 182, + 411, + 429, + 427 + ], + "lines": [ + { + "bbox": [ + 182, + 411, + 429, + 427 + ], + "spans": [ + { + "bbox": [ + 182, + 411, + 429, + 427 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\nabla _ { w } L = 2 \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { \\mathcal { Q } } ( s _ { t } , a _ { t } ) - { \\mathcal Q } _ { w } ( s _ { t } , a _ { t } ) ) ] = 0 . } \\end{array}", + "type": "interline_equation", + "image_path": "df93697e3d7e251afed3d6d2c3874531f249deeba1bf5f574eb2bd592d544520.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 182, + 411, + 429, + 427 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 433, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 106, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 324, + 446 + ], + "score": 1.0, + "content": "In compatible feature approximation, it directly uses", + "type": "text" + }, + { + "bbox": [ + 325, + 434, + 338, + 444 + ], + "score": 0.89, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "as control variate, rather than its Taylor", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 443, + 419, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 149, + 457 + ], + "score": 1.0, + "content": "expansion", + "type": "text" + }, + { + "bbox": [ + 150, + 443, + 163, + 455 + ], + "score": 0.9, + "content": "\\bar { Q } _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 444, + 419, + 457 + ], + "score": 1.0, + "content": "as in Q-Prop. Using Eq. 15, the Monte Carlo policy gradient is,", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 433, + 505, + 457 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 192, + 459, + 418, + 506 + ], + "lines": [ + { + "bbox": [ + 192, + 459, + 418, + 506 + ], + "spans": [ + { + "bbox": [ + 192, + 459, + 418, + 506 + ], + "score": 0.92, + "content": "\\begin{array} { r l } & { \\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ) Q _ { w } ( \\boldsymbol { s } _ { t } , \\boldsymbol { a } _ { t } ) ] } \\\\ & { \\qquad = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ ( \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ) \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ) ^ { T } ) w ] } \\\\ & { \\qquad = \\mathbb { E } _ { \\rho _ { \\pi } } [ I ( \\theta ; \\boldsymbol { s } _ { t } ) w ] , } \\end{array}", + "type": "interline_equation", + "image_path": "190a4f5c01b0e37dffa0037afb35be6b86f71f49759e6a8105d1953d9e424153.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 192, + 459, + 418, + 474.6666666666667 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 192, + 474.6666666666667, + 418, + 490.33333333333337 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 192, + 490.33333333333337, + 418, + 506.00000000000006 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 510, + 502, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 134, + 524 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 511, + 334, + 524 + ], + "score": 0.85, + "content": "I ( \\theta ; s _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } \\vert s _ { t } ) \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } \\vert s _ { t } ) ^ { T } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "is Fisher’s information matrix. Thus, vari-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 522, + 441, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 336, + 535 + ], + "score": 1.0, + "content": "ance reduction depends on ability to compute or estimate", + "type": "text" + }, + { + "bbox": [ + 336, + 523, + 367, + 534 + ], + "score": 0.93, + "content": "I ( \\theta ; s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 522, + 385, + 535 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 386, + 524, + 394, + 532 + ], + "score": 0.78, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 522, + 441, + 535 + ], + "score": 1.0, + "content": "effectively.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 510, + 505, + 535 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 549, + 388, + 563 + ], + "lines": [ + { + "bbox": [ + 106, + 549, + 389, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 389, + 564 + ], + "score": 1.0, + "content": "C UNIFYING POLICY GRADIENT AND ACTOR-CRITIC", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 574, + 504, + 598 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 588 + ], + "score": 1.0, + "content": "Q-Prop closely ties together policy gradient and actor-critic algorithms. To analyze this point, we", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 585, + 437, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 405, + 600 + ], + "score": 1.0, + "content": "write a generalization of Eq. 9 below, introducing two additional variables", + "type": "text" + }, + { + "bbox": [ + 405, + 587, + 432, + 598 + ], + "score": 0.9, + "content": "\\alpha , \\rho _ { C R }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 585, + 437, + 600 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 574, + 505, + 600 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 184, + 602, + 426, + 635 + ], + "lines": [ + { + "bbox": [ + 184, + 602, + 426, + 635 + ], + "spans": [ + { + "bbox": [ + 184, + 602, + 426, + 635 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { \\nabla _ { \\theta } J ( \\theta ) \\propto \\alpha \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { \\boldsymbol A } ( s _ { t } , a _ { t } ) - \\eta \\bar { \\boldsymbol A } _ { w } ( s _ { t } , a _ { t } ) ] } \\\\ & { \\qquad + \\eta \\mathbb { E } _ { \\rho _ { C R } } [ \\nabla _ { a } \\mathcal { Q } _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] } \\end{array}", + "type": "interline_equation", + "image_path": "81cb6159a91a178254bf4a790617f7e216698fa5af04494a2545828fb52a3a93.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 184, + 602, + 426, + 618.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 184, + 618.5, + 426, + 635.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 637, + 505, + 705 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 420, + 651 + ], + "score": 1.0, + "content": "Eq. 17 enables more analysis where bias generally is introduced only when", + "type": "text" + }, + { + "bbox": [ + 421, + 638, + 447, + 649 + ], + "score": 0.91, + "content": "\\alpha \\neq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 637, + 460, + 651 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 461, + 639, + 501, + 650 + ], + "score": 0.92, + "content": "\\rho _ { C R } \\neq \\rho _ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 637, + 505, + 651 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 649, + 504, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 504, + 660 + ], + "score": 1.0, + "content": "Importantly, Eq. 17 covers both policy gradient and deterministic actor-critic algorithm as its special", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 660, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 303, + 673 + ], + "score": 1.0, + "content": "cases. Standard policy gradient is recovered by", + "type": "text" + }, + { + "bbox": [ + 303, + 660, + 329, + 672 + ], + "score": 0.9, + "content": "\\eta = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 660, + 506, + 673 + ], + "score": 1.0, + "content": ", and deterministic actor-critic is recovered", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 669, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 120, + 685 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 120, + 671, + 147, + 681 + ], + "score": 0.89, + "content": "\\alpha = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 669, + 167, + 685 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 167, + 672, + 208, + 684 + ], + "score": 0.9, + "content": "\\rho _ { C R } = \\rho _ { \\beta }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 669, + 506, + 685 + ], + "score": 1.0, + "content": ". This allows heuristic or automatic methods for dynamically changing", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 680, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 506, + 695 + ], + "score": 1.0, + "content": "these variables through the learning process for optimizing different metrics, e.g. sample efficiency,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 693, + 223, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 223, + 705 + ], + "score": 1.0, + "content": "convergence speed, stability.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28.5, + "bbox_fs": [ + 104, + 637, + 506, + 705 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "Table 2 summarizes the various edge cases of Eq. 17. For example, since we derive our method from", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 223, + 733 + ], + "score": 1.0, + "content": "a control variates standpoint,", + "type": "text" + }, + { + "bbox": [ + 224, + 721, + 237, + 732 + ], + "score": 0.89, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "can be any function and the gradient remains almost unbiased (see", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 710, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 79, + 473, + 154 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 138, + 79, + 473, + 154 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 138, + 79, + 473, + 154 + ], + "spans": [ + { + "bbox": [ + 138, + 79, + 473, + 154 + ], + "score": 0.974, + "html": "
ParameterImplementation optionsIntroduce bias?
Qwoff-policy TD; on-policy TD(λ); model-based; etc.No
Von-policy Monte Carlo fitting; Eπθ [Qw(St, at)}; etcNo
0≤λ≤1Yes,except=1
αα≥0Yes, except α=1
nany nNo
PCRp of any policyYes, except pcR=pπ
", + "type": "table", + "image_path": "1aa12e84ce923bcad695b4e5e23bee23b45c30c0a31485d12c1fc031b384cf09.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 138, + 79, + 473, + 104.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 138, + 104.0, + 473, + 129.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 138, + 129.0, + 473, + 154.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 114, + 160, + 495, + 173 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 159, + 496, + 174 + ], + "spans": [ + { + "bbox": [ + 115, + 159, + 496, + 174 + ], + "score": 1.0, + "content": "Table 2: Implementation options and edge cases of the generalized Q-Prop estimator in Eq. 17.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 106, + 193, + 505, + 260 + ], + "lines": [ + { + "bbox": [ + 105, + 192, + 504, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 489, + 207 + ], + "score": 1.0, + "content": "Section 2.1). A natural choice is to use off-policy temporal difference learning to learn the critic", + "type": "text" + }, + { + "bbox": [ + 490, + 194, + 504, + 205 + ], + "score": 0.87, + "content": "Q _ { w }", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 204, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 203, + 218 + ], + "score": 1.0, + "content": "corresponding to policy", + "type": "text" + }, + { + "bbox": [ + 204, + 206, + 211, + 215 + ], + "score": 0.7, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 204, + 506, + 218 + ], + "score": 1.0, + "content": ". This enables effectively utilizing off-policy samples without introducing", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "score": 1.0, + "content": "further bias. An interesting alternative to this is to utilize model-based roll-outs to estimate the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 227, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 505, + 239 + ], + "score": 1.0, + "content": "critic, which resembles MuProp in stochastic neural networks (Gu et al., 2016a). Unlike prior work", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "on using fitted dynamics model to accelerate model-free learning (Gu et al., 2016b), this approach", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 249, + 361, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 361, + 261 + ], + "score": 1.0, + "content": "does not introduce bias to the gradient of the original objective.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 107, + 276, + 245, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 246, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 246, + 291 + ], + "score": 1.0, + "content": "D EXPERIMENT DETAILS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 301, + 505, + 411 + ], + "lines": [ + { + "bbox": [ + 106, + 300, + 504, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 504, + 312 + ], + "score": 1.0, + "content": "Policy and value function architectures. The network architectures are largely based on the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 310, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 104, + 310, + 505, + 326 + ], + "score": 1.0, + "content": "benchmark paper by Duan et al. (2016). For policy gradient methods, the stochastic policy", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 323, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 223, + 335 + ], + "score": 0.9, + "content": "\\pi _ { \\boldsymbol { \\theta } } ( \\mathbf { a } _ { t } | \\mathbf { s } _ { t } ) = \\bar { \\mathcal { N } } ( \\mu _ { \\boldsymbol { \\theta } } ( \\mathbf { \\bar { s } } _ { t } ) , \\Sigma _ { \\boldsymbol { \\theta } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 323, + 506, + 336 + ], + "score": 1.0, + "content": "is a local Gaussian policy with a local state-dependent mean and a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 213, + 347 + ], + "score": 1.0, + "content": "global covariance matrix.", + "type": "text" + }, + { + "bbox": [ + 213, + 334, + 242, + 345 + ], + "score": 0.92, + "content": "\\pmb { \\mu } _ { \\pmb { \\theta } } ( \\pmb { s } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 333, + 506, + 347 + ], + "score": 1.0, + "content": "is a neural network with 3 hidden layers of sizes 100-50-25 and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 344, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 282, + 357 + ], + "score": 1.0, + "content": "tanh nonlinearities at the first 2 layers, and", + "type": "text" + }, + { + "bbox": [ + 282, + 345, + 295, + 356 + ], + "score": 0.87, + "content": "\\Sigma _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 344, + 506, + 357 + ], + "score": 1.0, + "content": "is diagonal. For DDPG, the policy is deterministic", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 355, + 504, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 238, + 369 + ], + "score": 1.0, + "content": "and has the same architecture as", + "type": "text" + }, + { + "bbox": [ + 238, + 357, + 252, + 367 + ], + "score": 0.85, + "content": "\\mu _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 355, + 476, + 369 + ], + "score": 1.0, + "content": "except that it has an additional tanh layer at the output.", + "type": "text" + }, + { + "bbox": [ + 476, + 356, + 504, + 368 + ], + "score": 0.91, + "content": "V _ { \\phi } ( \\pmb { s } _ { t } )", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 366, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 506, + 379 + ], + "score": 1.0, + "content": "for baselines and GAE is fit with the same technique by Schulman et al. (2016), a variant of linear", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 457, + 390 + ], + "score": 1.0, + "content": "regression on Monte Carlo returns with soft-update constraint. For Q-Prop and DDPG,", + "type": "text" + }, + { + "bbox": [ + 457, + 378, + 494, + 390 + ], + "score": 0.93, + "content": "Q _ { w } ( s , \\pmb { a } )", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 378, + 505, + 390 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "parametrized with a neural network with 2 hidden layers of size 100 and ReLU nonlinearity, where", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 399, + 268, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 114, + 410 + ], + "score": 0.71, + "content": "^ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 399, + 268, + 412 + ], + "score": 1.0, + "content": "is included after the first hidden layer.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 423, + 505, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 505, + 437 + ], + "score": 1.0, + "content": "Training details. This section describes parameters of the training algorithms and their hyperpa-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 433, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 104, + 433, + 208, + 448 + ], + "score": 1.0, + "content": "rameter search values in", + "type": "text" + }, + { + "bbox": [ + 208, + 434, + 219, + 446 + ], + "score": 0.88, + "content": "\\{ \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 433, + 505, + 448 + ], + "score": 1.0, + "content": ". The optimal performing hyperparameter results are reported. Policy", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 444, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 104, + 444, + 352, + 460 + ], + "score": 1.0, + "content": "gradient methods (VPG, TRPO, Q-Prop) used batch sizes of", + "type": "text" + }, + { + "bbox": [ + 353, + 445, + 439, + 457 + ], + "score": 0.79, + "content": "\\{ 1 0 0 0 , 5 0 0 0 , 2 5 0 0 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 444, + 506, + 460 + ], + "score": 1.0, + "content": "time steps, step", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 456, + 504, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 140, + 469 + ], + "score": 1.0, + "content": "sizes of", + "type": "text" + }, + { + "bbox": [ + 141, + 456, + 215, + 468 + ], + "score": 0.85, + "content": "\\{ 0 . 1 , 0 . 0 1 , 0 . 0 0 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 456, + 438, + 469 + ], + "score": 1.0, + "content": "for the trust-region method, and base learning rates of", + "type": "text" + }, + { + "bbox": [ + 438, + 456, + 504, + 468 + ], + "score": 0.87, + "content": "\\{ 0 . 0 0 1 , \\bar { 0 } . 0 0 0 1 \\}", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 465, + 504, + 482 + ], + "spans": [ + { + "bbox": [ + 104, + 465, + 489, + 482 + ], + "score": 1.0, + "content": "with Adam (Kingma & Ba, 2014) for vanilla policy gradient methods. For Q-Prop and DDPG,", + "type": "text" + }, + { + "bbox": [ + 490, + 468, + 504, + 479 + ], + "score": 0.85, + "content": "Q _ { w }", + "type": "inline_equation" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 477, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 104, + 477, + 506, + 492 + ], + "score": 1.0, + "content": "is learned with the same technique as in DDPG (Lillicrap et al., 2016), using soft target networks", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 488, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 127, + 502 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 128, + 489, + 169, + 499 + ], + "score": 0.89, + "content": "\\tau = 0 . 9 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 488, + 265, + 502 + ], + "score": 1.0, + "content": ", a replay buffer of size", + "type": "text" + }, + { + "bbox": [ + 266, + 488, + 281, + 499 + ], + "score": 0.88, + "content": "1 0 ^ { 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 488, + 506, + 502 + ], + "score": 1.0, + "content": "steps, a mini-batch size of 64, and a base learning rate", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 118, + 513 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 500, + 184, + 512 + ], + "score": 0.88, + "content": "\\{ 0 . 0 0 1 , 0 . 0 0 0 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 500, + 506, + 513 + ], + "score": 1.0, + "content": "with Adam (Kingma & Ba, 2014). For Q-Prop we also tuned the relative ratio", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 227, + 524 + ], + "score": 1.0, + "content": "of gradient steps on the critic", + "type": "text" + }, + { + "bbox": [ + 227, + 512, + 241, + 522 + ], + "score": 0.89, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 511, + 465, + 524 + ], + "score": 1.0, + "content": "against the number of steps on the policy, in the range", + "type": "text" + }, + { + "bbox": [ + 465, + 511, + 501, + 523 + ], + "score": 0.6, + "content": "\\{ 0 . 1 , 0 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 511, + 505, + 524 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 125, + 534 + ], + "score": 0.82, + "content": "1 . 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 522, + 505, + 534 + ], + "score": 1.0, + "content": ", where 0.1 corresponds to 100 critic updates for every policy update if the batch size is 1000.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 533, + 489, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 294, + 546 + ], + "score": 1.0, + "content": "For DDPG, we swept the reward scaling using", + "type": "text" + }, + { + "bbox": [ + 294, + 533, + 352, + 545 + ], + "score": 0.92, + "content": "\\{ 0 . 0 1 , 0 . 1 , \\dot { 1 . } 0 \\dot { \\} }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 533, + 489, + 546 + ], + "score": 1.0, + "content": "as it is sensitive to this parameter.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 26 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 79, + 473, + 154 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 138, + 79, + 473, + 154 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 138, + 79, + 473, + 154 + ], + "spans": [ + { + "bbox": [ + 138, + 79, + 473, + 154 + ], + "score": 0.974, + "html": "
ParameterImplementation optionsIntroduce bias?
Qwoff-policy TD; on-policy TD(λ); model-based; etc.No
Von-policy Monte Carlo fitting; Eπθ [Qw(St, at)}; etcNo
0≤λ≤1Yes,except=1
αα≥0Yes, except α=1
nany nNo
PCRp of any policyYes, except pcR=pπ
", + "type": "table", + "image_path": "1aa12e84ce923bcad695b4e5e23bee23b45c30c0a31485d12c1fc031b384cf09.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 138, + 79, + 473, + 104.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 138, + 104.0, + 473, + 129.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 138, + 129.0, + 473, + 154.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 114, + 160, + 495, + 173 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 159, + 496, + 174 + ], + "spans": [ + { + "bbox": [ + 115, + 159, + 496, + 174 + ], + "score": 1.0, + "content": "Table 2: Implementation options and edge cases of the generalized Q-Prop estimator in Eq. 17.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 106, + 193, + 505, + 260 + ], + "lines": [ + { + "bbox": [ + 105, + 192, + 504, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 489, + 207 + ], + "score": 1.0, + "content": "Section 2.1). A natural choice is to use off-policy temporal difference learning to learn the critic", + "type": "text" + }, + { + "bbox": [ + 490, + 194, + 504, + 205 + ], + "score": 0.87, + "content": "Q _ { w }", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 204, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 203, + 218 + ], + "score": 1.0, + "content": "corresponding to policy", + "type": "text" + }, + { + "bbox": [ + 204, + 206, + 211, + 215 + ], + "score": 0.7, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 204, + 506, + 218 + ], + "score": 1.0, + "content": ". This enables effectively utilizing off-policy samples without introducing", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "score": 1.0, + "content": "further bias. An interesting alternative to this is to utilize model-based roll-outs to estimate the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 227, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 505, + 239 + ], + "score": 1.0, + "content": "critic, which resembles MuProp in stochastic neural networks (Gu et al., 2016a). Unlike prior work", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "on using fitted dynamics model to accelerate model-free learning (Gu et al., 2016b), this approach", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 249, + 361, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 361, + 261 + ], + "score": 1.0, + "content": "does not introduce bias to the gradient of the original objective.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 192, + 506, + 261 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 276, + 245, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 246, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 246, + 291 + ], + "score": 1.0, + "content": "D EXPERIMENT DETAILS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 301, + 505, + 411 + ], + "lines": [ + { + "bbox": [ + 106, + 300, + 504, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 504, + 312 + ], + "score": 1.0, + "content": "Policy and value function architectures. The network architectures are largely based on the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 310, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 104, + 310, + 505, + 326 + ], + "score": 1.0, + "content": "benchmark paper by Duan et al. (2016). For policy gradient methods, the stochastic policy", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 323, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 223, + 335 + ], + "score": 0.9, + "content": "\\pi _ { \\boldsymbol { \\theta } } ( \\mathbf { a } _ { t } | \\mathbf { s } _ { t } ) = \\bar { \\mathcal { N } } ( \\mu _ { \\boldsymbol { \\theta } } ( \\mathbf { \\bar { s } } _ { t } ) , \\Sigma _ { \\boldsymbol { \\theta } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 323, + 506, + 336 + ], + "score": 1.0, + "content": "is a local Gaussian policy with a local state-dependent mean and a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 213, + 347 + ], + "score": 1.0, + "content": "global covariance matrix.", + "type": "text" + }, + { + "bbox": [ + 213, + 334, + 242, + 345 + ], + "score": 0.92, + "content": "\\pmb { \\mu } _ { \\pmb { \\theta } } ( \\pmb { s } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 333, + 506, + 347 + ], + "score": 1.0, + "content": "is a neural network with 3 hidden layers of sizes 100-50-25 and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 344, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 282, + 357 + ], + "score": 1.0, + "content": "tanh nonlinearities at the first 2 layers, and", + "type": "text" + }, + { + "bbox": [ + 282, + 345, + 295, + 356 + ], + "score": 0.87, + "content": "\\Sigma _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 344, + 506, + 357 + ], + "score": 1.0, + "content": "is diagonal. For DDPG, the policy is deterministic", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 355, + 504, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 238, + 369 + ], + "score": 1.0, + "content": "and has the same architecture as", + "type": "text" + }, + { + "bbox": [ + 238, + 357, + 252, + 367 + ], + "score": 0.85, + "content": "\\mu _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 355, + 476, + 369 + ], + "score": 1.0, + "content": "except that it has an additional tanh layer at the output.", + "type": "text" + }, + { + "bbox": [ + 476, + 356, + 504, + 368 + ], + "score": 0.91, + "content": "V _ { \\phi } ( \\pmb { s } _ { t } )", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 366, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 506, + 379 + ], + "score": 1.0, + "content": "for baselines and GAE is fit with the same technique by Schulman et al. (2016), a variant of linear", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 457, + 390 + ], + "score": 1.0, + "content": "regression on Monte Carlo returns with soft-update constraint. For Q-Prop and DDPG,", + "type": "text" + }, + { + "bbox": [ + 457, + 378, + 494, + 390 + ], + "score": 0.93, + "content": "Q _ { w } ( s , \\pmb { a } )", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 378, + 505, + 390 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "parametrized with a neural network with 2 hidden layers of size 100 and ReLU nonlinearity, where", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 399, + 268, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 114, + 410 + ], + "score": 0.71, + "content": "^ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 399, + 268, + 412 + ], + "score": 1.0, + "content": "is included after the first hidden layer.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15.5, + "bbox_fs": [ + 104, + 300, + 506, + 412 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 423, + 505, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 505, + 437 + ], + "score": 1.0, + "content": "Training details. This section describes parameters of the training algorithms and their hyperpa-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 433, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 104, + 433, + 208, + 448 + ], + "score": 1.0, + "content": "rameter search values in", + "type": "text" + }, + { + "bbox": [ + 208, + 434, + 219, + 446 + ], + "score": 0.88, + "content": "\\{ \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 433, + 505, + 448 + ], + "score": 1.0, + "content": ". The optimal performing hyperparameter results are reported. Policy", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 444, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 104, + 444, + 352, + 460 + ], + "score": 1.0, + "content": "gradient methods (VPG, TRPO, Q-Prop) used batch sizes of", + "type": "text" + }, + { + "bbox": [ + 353, + 445, + 439, + 457 + ], + "score": 0.79, + "content": "\\{ 1 0 0 0 , 5 0 0 0 , 2 5 0 0 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 444, + 506, + 460 + ], + "score": 1.0, + "content": "time steps, step", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 456, + 504, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 140, + 469 + ], + "score": 1.0, + "content": "sizes of", + "type": "text" + }, + { + "bbox": [ + 141, + 456, + 215, + 468 + ], + "score": 0.85, + "content": "\\{ 0 . 1 , 0 . 0 1 , 0 . 0 0 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 456, + 438, + 469 + ], + "score": 1.0, + "content": "for the trust-region method, and base learning rates of", + "type": "text" + }, + { + "bbox": [ + 438, + 456, + 504, + 468 + ], + "score": 0.87, + "content": "\\{ 0 . 0 0 1 , \\bar { 0 } . 0 0 0 1 \\}", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 465, + 504, + 482 + ], + "spans": [ + { + "bbox": [ + 104, + 465, + 489, + 482 + ], + "score": 1.0, + "content": "with Adam (Kingma & Ba, 2014) for vanilla policy gradient methods. For Q-Prop and DDPG,", + "type": "text" + }, + { + "bbox": [ + 490, + 468, + 504, + 479 + ], + "score": 0.85, + "content": "Q _ { w }", + "type": "inline_equation" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 477, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 104, + 477, + 506, + 492 + ], + "score": 1.0, + "content": "is learned with the same technique as in DDPG (Lillicrap et al., 2016), using soft target networks", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 488, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 127, + 502 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 128, + 489, + 169, + 499 + ], + "score": 0.89, + "content": "\\tau = 0 . 9 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 488, + 265, + 502 + ], + "score": 1.0, + "content": ", a replay buffer of size", + "type": "text" + }, + { + "bbox": [ + 266, + 488, + 281, + 499 + ], + "score": 0.88, + "content": "1 0 ^ { 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 488, + 506, + 502 + ], + "score": 1.0, + "content": "steps, a mini-batch size of 64, and a base learning rate", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 118, + 513 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 500, + 184, + 512 + ], + "score": 0.88, + "content": "\\{ 0 . 0 0 1 , 0 . 0 0 0 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 500, + 506, + 513 + ], + "score": 1.0, + "content": "with Adam (Kingma & Ba, 2014). For Q-Prop we also tuned the relative ratio", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 227, + 524 + ], + "score": 1.0, + "content": "of gradient steps on the critic", + "type": "text" + }, + { + "bbox": [ + 227, + 512, + 241, + 522 + ], + "score": 0.89, + "content": "Q _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 511, + 465, + 524 + ], + "score": 1.0, + "content": "against the number of steps on the policy, in the range", + "type": "text" + }, + { + "bbox": [ + 465, + 511, + 501, + 523 + ], + "score": 0.6, + "content": "\\{ 0 . 1 , 0 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 511, + 505, + 524 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 125, + 534 + ], + "score": 0.82, + "content": "1 . 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 522, + 505, + 534 + ], + "score": 1.0, + "content": ", where 0.1 corresponds to 100 critic updates for every policy update if the batch size is 1000.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 533, + 489, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 294, + 546 + ], + "score": 1.0, + "content": "For DDPG, we swept the reward scaling using", + "type": "text" + }, + { + "bbox": [ + 294, + 533, + 352, + 545 + ], + "score": 0.92, + "content": "\\{ 0 . 0 1 , 0 . 1 , \\dot { 1 . } 0 \\dot { \\} }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 533, + 489, + 546 + ], + "score": 1.0, + "content": "as it is sensitive to this parameter.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 26, + "bbox_fs": [ + 104, + 422, + 506, + 546 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/SJ3rcZcxl/SJ3rcZcxl_model.json b/parse/train/SJ3rcZcxl/SJ3rcZcxl_model.json new file mode 100644 index 0000000000000000000000000000000000000000..05ee1b14b1912662b7bc489b549d3ba1861f00b8 --- /dev/null +++ b/parse/train/SJ3rcZcxl/SJ3rcZcxl_model.json @@ -0,0 +1,18570 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 817, + 1302, + 817, + 1302, + 1366, + 398, + 1366 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1501, + 1403, + 1501, + 1403, + 1956, + 298, + 1956 + ], + "score": 0.983 + }, + { + "category_id": 0, + "poly": [ + 300, + 218, + 1402, + 218, + 1402, + 324, + 300, + 324 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 299, + 1972, + 1397, + 1972, + 1397, + 2034, + 299, + 2034 + ], + "score": 0.936 + }, + { + "category_id": 0, + "poly": [ + 302, + 1428, + 573, + 1428, + 573, + 1463, + 302, + 1463 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 773, + 748, + 926, + 748, + 926, + 781, + 773, + 781 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.876 + }, + { + "category_id": 1, + "poly": [ + 315, + 374, + 1452, + 374, + 1452, + 409, + 315, + 409 + ], + "score": 0.837 + }, + { + "category_id": 1, + "poly": [ + 313, + 411, + 1272, + 411, + 1272, + 666, + 313, + 666 + ], + "score": 0.745 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 857, + 2088, + 857, + 2112, + 842, + 2112 + ], + "score": 0.655 + }, + { + "category_id": 13, + "poly": [ + 423, + 374, + 494, + 374, + 494, + 408, + 423, + 408 + ], + "score": 0.83, + "latex": "\\mathbf { G u } ^ { 1 2 3 }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 213.0, + 1406.0, + 213.0, + 1406.0, + 275.0, + 292.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 272.0, + 921.0, + 272.0, + 921.0, + 328.0, + 295.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1425.0, + 579.0, + 1425.0, + 579.0, + 1472.0, + 294.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 747.0, + 932.0, + 747.0, + 932.0, + 784.0, + 769.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 858.0, + 2088.0, + 858.0, + 2116.0, + 841.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 819.0, + 1307.0, + 819.0, + 1307.0, + 852.0, + 395.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 849.0, + 1306.0, + 849.0, + 1306.0, + 882.0, + 395.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 880.0, + 1304.0, + 880.0, + 1304.0, + 913.0, + 394.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 905.0, + 1307.0, + 905.0, + 1307.0, + 949.0, + 392.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 939.0, + 1307.0, + 939.0, + 1307.0, + 977.0, + 393.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 971.0, + 1305.0, + 971.0, + 1305.0, + 1006.0, + 393.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1003.0, + 1306.0, + 1003.0, + 1306.0, + 1036.0, + 394.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1031.0, + 1304.0, + 1031.0, + 1304.0, + 1069.0, + 393.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1063.0, + 1307.0, + 1063.0, + 1307.0, + 1097.0, + 393.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1094.0, + 1305.0, + 1094.0, + 1305.0, + 1127.0, + 395.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1120.0, + 1306.0, + 1120.0, + 1306.0, + 1159.0, + 392.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1151.0, + 1305.0, + 1151.0, + 1305.0, + 1188.0, + 394.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1183.0, + 1306.0, + 1183.0, + 1306.0, + 1220.0, + 391.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1214.0, + 1307.0, + 1214.0, + 1307.0, + 1251.0, + 394.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1245.0, + 1306.0, + 1245.0, + 1306.0, + 1280.0, + 393.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1276.0, + 1303.0, + 1276.0, + 1303.0, + 1310.0, + 394.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1307.0, + 1305.0, + 1307.0, + 1305.0, + 1340.0, + 395.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1340.0, + 690.0, + 1340.0, + 690.0, + 1370.0, + 396.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1498.0, + 1405.0, + 1498.0, + 1405.0, + 1536.0, + 293.0, + 1536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1529.0, + 1407.0, + 1529.0, + 1407.0, + 1565.0, + 293.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1560.0, + 1405.0, + 1560.0, + 1405.0, + 1595.0, + 294.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1591.0, + 1404.0, + 1591.0, + 1404.0, + 1626.0, + 294.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1621.0, + 1405.0, + 1621.0, + 1405.0, + 1655.0, + 294.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1653.0, + 1405.0, + 1653.0, + 1405.0, + 1688.0, + 293.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1679.0, + 1405.0, + 1679.0, + 1405.0, + 1720.0, + 292.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1711.0, + 1403.0, + 1711.0, + 1403.0, + 1748.0, + 293.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1740.0, + 1407.0, + 1740.0, + 1407.0, + 1781.0, + 292.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1774.0, + 1406.0, + 1774.0, + 1406.0, + 1809.0, + 293.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1803.0, + 1404.0, + 1803.0, + 1404.0, + 1838.0, + 294.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1834.0, + 1407.0, + 1834.0, + 1407.0, + 1871.0, + 292.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1864.0, + 1405.0, + 1864.0, + 1405.0, + 1899.0, + 293.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1893.0, + 1405.0, + 1893.0, + 1405.0, + 1930.0, + 294.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1927.0, + 808.0, + 1927.0, + 808.0, + 1961.0, + 294.0, + 1961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1968.0, + 1402.0, + 1968.0, + 1402.0, + 2011.0, + 293.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2000.0, + 1403.0, + 2000.0, + 1403.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 369.0, + 422.0, + 369.0, + 422.0, + 417.0, + 310.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 369.0, + 1459.0, + 369.0, + 1459.0, + 417.0, + 495.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 406.0, + 1278.0, + 406.0, + 1278.0, + 443.0, + 312.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 439.0, + 1016.0, + 439.0, + 1016.0, + 470.0, + 313.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 466.0, + 664.0, + 466.0, + 664.0, + 505.0, + 311.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 497.0, + 1057.0, + 497.0, + 1057.0, + 541.0, + 309.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 530.0, + 553.0, + 530.0, + 553.0, + 574.0, + 310.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 563.0, + 507.0, + 563.0, + 507.0, + 606.0, + 309.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 595.0, + 546.0, + 595.0, + 546.0, + 640.0, + 309.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 628.0, + 555.0, + 628.0, + 555.0, + 673.0, + 309.0, + 673.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 232, + 1403, + 232, + 1403, + 779, + 298, + 779 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 1466, + 1404, + 1466, + 1404, + 1870, + 297, + 1870 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 794, + 1404, + 794, + 1404, + 1343, + 298, + 1343 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 300, + 1972, + 1400, + 1972, + 1400, + 2034, + 300, + 2034 + ], + "score": 0.94 + }, + { + "category_id": 0, + "poly": [ + 298, + 1912, + 912, + 1912, + 912, + 1944, + 298, + 1944 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 300, + 1394, + 557, + 1394, + 557, + 1429, + 300, + 1429 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 814, + 76, + 814, + 104, + 300, + 104 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2112, + 841, + 2112 + ], + "score": 0.698 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.144 + }, + { + "category_id": 13, + "poly": [ + 1053, + 1558, + 1148, + 1558, + 1148, + 1592, + 1053, + 1592 + ], + "score": 0.93, + "latex": "r ( s _ { t } , a _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 563, + 1559, + 718, + 1559, + 718, + 1592, + 563, + 1592 + ], + "score": 0.93, + "latex": "p \\big ( \\pmb { s } _ { t + 1 } \\big | \\pmb { s } _ { t } , \\pmb { a } _ { t } \\big )" + }, + { + "category_id": 13, + "poly": [ + 530, + 1528, + 626, + 1528, + 626, + 1561, + 530, + 1561 + ], + "score": 0.92, + "latex": "\\pi ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 296, + 1621, + 574, + 1621, + 574, + 1658, + 296, + 1658 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { R _ { t } = \\sum _ { t ^ { \\prime } = t } ^ { \\infty } \\gamma ^ { t ^ { \\prime } - t } r ( s _ { t ^ { \\prime } } , { \\bf a } _ { t ^ { \\prime } } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 296, + 1658, + 474, + 1658, + 474, + 1690, + 296, + 1690 + ], + "score": 0.9, + "latex": "J ( \\theta ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ R _ { 0 } ]" + }, + { + "category_id": 13, + "poly": [ + 956, + 2003, + 1012, + 2003, + 1012, + 2036, + 956, + 2036 + ], + "score": 0.89, + "latex": "J ( \\theta )" + }, + { + "category_id": 13, + "poly": [ + 1349, + 1534, + 1400, + 1534, + 1400, + 1561, + 1349, + 1561 + ], + "score": 0.88, + "latex": "\\boldsymbol { s } _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 1328, + 1504, + 1358, + 1504, + 1358, + 1529, + 1328, + 1529 + ], + "score": 0.85, + "latex": "\\mathbf { } \\mathbf { a } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 429, + 1591, + 458, + 1591, + 458, + 1620, + 429, + 1620 + ], + "score": 0.83, + "latex": "R _ { t }" + }, + { + "category_id": 13, + "poly": [ + 967, + 1503, + 992, + 1503, + 992, + 1529, + 967, + 1529 + ], + "score": 0.83, + "latex": "\\scriptstyle { s _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 568, + 1595, + 587, + 1595, + 587, + 1622, + 568, + 1622 + ], + "score": 0.61, + "latex": "\\gamma" + }, + { + "category_id": 13, + "poly": [ + 986, + 1594, + 1001, + 1594, + 1001, + 1617, + 986, + 1617 + ], + "score": 0.59, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 838, + 1503, + 853, + 1503, + 853, + 1527, + 838, + 1527 + ], + "score": 0.55, + "latex": "t" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1912.0, + 914.0, + 1912.0, + 914.0, + 1948.0, + 295.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1390.0, + 561.0, + 1390.0, + 561.0, + 1437.0, + 291.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2121.0, + 839.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 1405.0, + 230.0, + 1405.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 258.0, + 1403.0, + 258.0, + 1403.0, + 300.0, + 293.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 291.0, + 1405.0, + 291.0, + 1405.0, + 327.0, + 293.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 322.0, + 1403.0, + 322.0, + 1403.0, + 358.0, + 294.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 351.0, + 1406.0, + 351.0, + 1406.0, + 388.0, + 292.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 383.0, + 1403.0, + 383.0, + 1403.0, + 415.0, + 294.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 412.0, + 1405.0, + 412.0, + 1405.0, + 449.0, + 293.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 444.0, + 1405.0, + 444.0, + 1405.0, + 479.0, + 294.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 475.0, + 1405.0, + 475.0, + 1405.0, + 511.0, + 293.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 504.0, + 1403.0, + 504.0, + 1403.0, + 541.0, + 293.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 536.0, + 1403.0, + 536.0, + 1403.0, + 569.0, + 293.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 565.0, + 1405.0, + 565.0, + 1405.0, + 600.0, + 292.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 595.0, + 1405.0, + 595.0, + 1405.0, + 631.0, + 294.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 627.0, + 1405.0, + 627.0, + 1405.0, + 662.0, + 294.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 657.0, + 1404.0, + 657.0, + 1404.0, + 693.0, + 294.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 687.0, + 1405.0, + 687.0, + 1405.0, + 723.0, + 294.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 717.0, + 1407.0, + 717.0, + 1407.0, + 756.0, + 293.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 751.0, + 512.0, + 751.0, + 512.0, + 782.0, + 296.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1467.0, + 1402.0, + 1467.0, + 1402.0, + 1504.0, + 295.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1496.0, + 837.0, + 1496.0, + 837.0, + 1537.0, + 294.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 1496.0, + 966.0, + 1496.0, + 966.0, + 1537.0, + 854.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 1496.0, + 1327.0, + 1496.0, + 1327.0, + 1537.0, + 993.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1359.0, + 1496.0, + 1405.0, + 1496.0, + 1405.0, + 1537.0, + 1359.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1524.0, + 529.0, + 1524.0, + 529.0, + 1570.0, + 292.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 1524.0, + 1348.0, + 1524.0, + 1348.0, + 1570.0, + 627.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1524.0, + 1405.0, + 1524.0, + 1405.0, + 1570.0, + 1401.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1555.0, + 562.0, + 1555.0, + 562.0, + 1598.0, + 291.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 1555.0, + 1052.0, + 1555.0, + 1052.0, + 1598.0, + 719.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 1555.0, + 1405.0, + 1555.0, + 1405.0, + 1598.0, + 1149.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1588.0, + 428.0, + 1588.0, + 428.0, + 1625.0, + 295.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 1588.0, + 567.0, + 1588.0, + 567.0, + 1625.0, + 459.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 1588.0, + 985.0, + 1588.0, + 985.0, + 1625.0, + 588.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 1588.0, + 1404.0, + 1588.0, + 1404.0, + 1625.0, + 1002.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1619.0, + 295.0, + 1619.0, + 295.0, + 1664.0, + 291.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1619.0, + 1409.0, + 1619.0, + 1409.0, + 1664.0, + 575.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1654.0, + 295.0, + 1654.0, + 295.0, + 1694.0, + 292.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 1654.0, + 1407.0, + 1654.0, + 1407.0, + 1694.0, + 475.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1685.0, + 1406.0, + 1685.0, + 1406.0, + 1722.0, + 294.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1715.0, + 1405.0, + 1715.0, + 1405.0, + 1752.0, + 294.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1743.0, + 1406.0, + 1743.0, + 1406.0, + 1785.0, + 291.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1776.0, + 1406.0, + 1776.0, + 1406.0, + 1814.0, + 288.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1803.0, + 1405.0, + 1803.0, + 1405.0, + 1847.0, + 290.0, + 1847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1839.0, + 1312.0, + 1839.0, + 1312.0, + 1874.0, + 294.0, + 1874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 793.0, + 1406.0, + 793.0, + 1406.0, + 831.0, + 293.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 824.0, + 1409.0, + 824.0, + 1409.0, + 863.0, + 292.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 856.0, + 1404.0, + 856.0, + 1404.0, + 891.0, + 294.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 887.0, + 1404.0, + 887.0, + 1404.0, + 923.0, + 294.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 913.0, + 1408.0, + 913.0, + 1408.0, + 954.0, + 292.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 946.0, + 1405.0, + 946.0, + 1405.0, + 984.0, + 293.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 977.0, + 1406.0, + 977.0, + 1406.0, + 1013.0, + 294.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1010.0, + 1405.0, + 1010.0, + 1405.0, + 1042.0, + 296.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1038.0, + 1405.0, + 1038.0, + 1405.0, + 1074.0, + 292.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1069.0, + 1407.0, + 1069.0, + 1407.0, + 1105.0, + 292.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1100.0, + 1406.0, + 1100.0, + 1406.0, + 1136.0, + 294.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1131.0, + 1404.0, + 1131.0, + 1404.0, + 1166.0, + 294.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1158.0, + 1404.0, + 1158.0, + 1404.0, + 1197.0, + 292.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1189.0, + 1404.0, + 1189.0, + 1404.0, + 1226.0, + 293.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1222.0, + 1405.0, + 1222.0, + 1405.0, + 1257.0, + 294.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1252.0, + 1404.0, + 1252.0, + 1404.0, + 1288.0, + 294.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1279.0, + 1405.0, + 1279.0, + 1405.0, + 1321.0, + 292.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1312.0, + 1269.0, + 1312.0, + 1269.0, + 1347.0, + 292.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1999.0, + 955.0, + 1999.0, + 955.0, + 2041.0, + 293.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 1999.0, + 1404.0, + 1999.0, + 1404.0, + 2041.0, + 1013.0, + 2041.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1384, + 1404, + 1384, + 1404, + 1602, + 296, + 1602 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1138, + 1404, + 1138, + 1404, + 1295, + 297, + 1295 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 301, + 704, + 1403, + 704, + 1403, + 859, + 301, + 859 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1615, + 1403, + 1615, + 1403, + 1770, + 298, + 1770 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 296, + 535, + 1404, + 535, + 1404, + 690, + 296, + 690 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 303, + 1001, + 1399, + 1001, + 1399, + 1125, + 303, + 1125 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 300, + 1891, + 1399, + 1891, + 1399, + 1985, + 300, + 1985 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 408, + 1780, + 1288, + 1780, + 1288, + 1880, + 408, + 1880 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 440, + 868, + 1256, + 868, + 1256, + 993, + 440, + 993 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 295, + 382, + 1410, + 382, + 1410, + 448, + 295, + 448 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 294, + 227, + 1403, + 227, + 1403, + 293, + 294, + 293 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 395, + 297, + 1303, + 297, + 1303, + 374, + 395, + 374 + ], + "score": 0.937 + }, + { + "category_id": 8, + "poly": [ + 552, + 1993, + 1150, + 1993, + 1150, + 2035, + 552, + 2035 + ], + "score": 0.933 + }, + { + "category_id": 8, + "poly": [ + 512, + 485, + 1182, + 485, + 1182, + 527, + 512, + 527 + ], + "score": 0.933 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 815, + 74, + 815, + 105, + 299, + 105 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 298, + 1328, + 1005, + 1328, + 1005, + 1361, + 298, + 1361 + ], + "score": 0.902 + }, + { + "category_id": 9, + "poly": [ + 1366, + 491, + 1400, + 491, + 1400, + 522, + 1366, + 522 + ], + "score": 0.895 + }, + { + "category_id": 9, + "poly": [ + 1366, + 915, + 1400, + 915, + 1400, + 945, + 1366, + 945 + ], + "score": 0.886 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1817, + 1400, + 1817, + 1400, + 1846, + 1366, + 1846 + ], + "score": 0.882 + }, + { + "category_id": 9, + "poly": [ + 1365, + 2000, + 1400, + 2000, + 1400, + 2029, + 1365, + 2029 + ], + "score": 0.863 + }, + { + "category_id": 9, + "poly": [ + 1365, + 319, + 1400, + 319, + 1400, + 350, + 1365, + 350 + ], + "score": 0.861 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.676 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.456 + }, + { + "category_id": 13, + "poly": [ + 397, + 828, + 514, + 828, + 514, + 861, + 397, + 861 + ], + "score": 0.94, + "latex": "Q _ { \\pi } ( s _ { t } , \\pmb { a } _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 466, + 1032, + 582, + 1032, + 582, + 1065, + 466, + 1065 + ], + "score": 0.93, + "latex": "A _ { \\pi } ( s _ { t } , { \\pmb a } _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 929, + 1063, + 1046, + 1063, + 1046, + 1096, + 929, + 1096 + ], + "score": 0.93, + "latex": "A _ { \\pi } ( s _ { t } , { \\pmb a } _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 912, + 797, + 1027, + 797, + 1027, + 831, + 912, + 831 + ], + "score": 0.93, + "latex": "A _ { \\pi } ( s _ { t } , { \\pmb a } _ { t } )" + }, + { + "category_id": 14, + "poly": [ + 444, + 865, + 1256, + 865, + 1256, + 996, + 444, + 996 + ], + "score": 0.93, + "latex": "\\begin{array} { c } { { V _ { \\pi } ( s _ { t } ) = \\mathbb { E } _ { \\pi } [ R _ { t } ] = \\mathbb { E } _ { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } [ Q _ { \\pi } ( s _ { t } , a _ { t } ) ] } } \\\\ { { Q _ { \\pi } ( s _ { t } , a _ { t } ) = r ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { E } _ { \\pi } [ R _ { t + 1 } ] = r ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { E } _ { p ( s _ { t + 1 } | s _ { t } , a _ { t } ) } [ V _ { \\pi } ( s _ { t + 1 } ) ] } } \\\\ { { A _ { \\pi } ( s _ { t } , a _ { t } ) = Q _ { \\pi } ( s _ { t } , a _ { t } ) - V _ { \\pi } ( s _ { t } ) . } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 760, + 1063, + 836, + 1063, + 836, + 1096, + 760, + 1096 + ], + "score": 0.93, + "latex": "V _ { \\pi } ( \\pmb { s } _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 1111, + 767, + 1187, + 767, + 1187, + 800, + 1111, + 800 + ], + "score": 0.93, + "latex": "V _ { \\pi } ( \\pmb { s } _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 895, + 1708, + 966, + 1708, + 966, + 1741, + 895, + 1741 + ], + "score": 0.92, + "latex": "Q ( \\cdot , \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 903, + 1953, + 958, + 1953, + 958, + 1986, + 903, + 1986 + ], + "score": 0.92, + "latex": "J ( \\theta )" + }, + { + "category_id": 13, + "poly": [ + 372, + 383, + 432, + 383, + 432, + 417, + 372, + 417 + ], + "score": 0.92, + "latex": "b ( s _ { t } )" + }, + { + "category_id": 14, + "poly": [ + 397, + 293, + 1300, + 293, + 1300, + 376, + 397, + 376 + ], + "score": 0.92, + "latex": "\\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\pi } [ \\sum _ { t = 0 } ^ { \\infty } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\gamma ^ { t } R _ { t } ] = \\mathbb { E } _ { \\pi } [ \\sum _ { t = 0 } ^ { \\infty } \\gamma ^ { t } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( R _ { t } - b ( s _ { t } ) ) ] ," + }, + { + "category_id": 13, + "poly": [ + 297, + 1001, + 414, + 1001, + 414, + 1035, + 297, + 1035 + ], + "score": 0.92, + "latex": "Q _ { \\pi } ( s _ { t } , \\pmb { a } _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 372, + 1677, + 692, + 1677, + 692, + 1711, + 372, + 1711 + ], + "score": 0.91, + "latex": "\\pi _ { \\boldsymbol { \\theta } } ( \\mathbf { a } _ { t } | \\mathbf { s } _ { t } ) = \\delta ( \\mathbf { a } _ { t } = \\mu _ { \\boldsymbol { \\theta } } ( \\mathbf { s } _ { t } ) )" + }, + { + "category_id": 13, + "poly": [ + 479, + 413, + 764, + 413, + 764, + 449, + 479, + 449 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\rho _ { \\pi } ( s ) = \\sum _ { t = 0 } ^ { \\infty } \\gamma ^ { t } p ( s _ { t } = s ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 584, + 1448, + 641, + 1448, + 641, + 1480, + 584, + 1480 + ], + "score": 0.91, + "latex": "\\pi ( \\theta )" + }, + { + "category_id": 13, + "poly": [ + 1047, + 569, + 1115, + 569, + 1115, + 600, + 1047, + 600 + ], + "score": 0.9, + "latex": "\\gamma = 1" + }, + { + "category_id": 14, + "poly": [ + 409, + 1776, + 1291, + 1776, + 1291, + 1887, + 409, + 1887 + ], + "score": 0.9, + "latex": "\\begin{array} { r l } & { w = \\underset { w } { \\arg \\operatorname* { m i n } } \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\beta } ( \\cdot ) , a _ { t } \\sim \\beta ( \\cdot | s _ { t } ) } \\big [ \\big ( r \\big ( s _ { t } , a _ { t } \\big ) + \\gamma Q ( s _ { t + 1 } , \\mu _ { \\theta } ( s _ { t + 1 } ) \\big ) - Q _ { w } ( s _ { t } , a _ { t } ) \\big ) ^ { 2 } \\big ] } \\\\ & { \\theta = \\underset { \\theta } { \\arg \\operatorname* { m a x } } \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\beta } ( \\cdot ) } [ Q _ { w } \\big ( s _ { t } , \\mu _ { \\theta } ( s _ { t } ) \\big ) ] } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 551, + 1992, + 1149, + 1992, + 1149, + 2038, + 551, + 2038 + ], + "score": 0.9, + "latex": "\\nabla _ { \\theta } J ( \\theta ) \\approx \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\beta } ( \\cdot ) } [ \\nabla _ { a } Q _ { w } \\big ( s _ { t } , a \\big ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } \\big ( s _ { t } \\big ) ]" + }, + { + "category_id": 14, + "poly": [ + 514, + 485, + 1183, + 485, + 1183, + 527, + 514, + 527 + ], + "score": 0.88, + "latex": "\\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\pi } ( . ) , \\mathbf { a } _ { t } \\sim \\pi ( \\cdot | s _ { t } ) } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( R _ { t } - b ( s _ { t } ) ) ] ." + }, + { + "category_id": 13, + "poly": [ + 609, + 1481, + 647, + 1481, + 647, + 1510, + 609, + 1510 + ], + "score": 0.88, + "latex": "Q _ { w }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1449, + 336, + 1449, + 336, + 1479, + 297, + 1479 + ], + "score": 0.88, + "latex": "Q _ { w }" + }, + { + "category_id": 13, + "poly": [ + 344, + 1712, + 377, + 1712, + 377, + 1744, + 344, + 1744 + ], + "score": 0.87, + "latex": "\\rho _ { \\beta }" + }, + { + "category_id": 13, + "poly": [ + 1310, + 572, + 1342, + 572, + 1342, + 599, + 1310, + 599 + ], + "score": 0.87, + "latex": "\\rho _ { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1741, + 335, + 1741, + 335, + 1770, + 298, + 1770 + ], + "score": 0.86, + "latex": "Q _ { w }" + }, + { + "category_id": 13, + "poly": [ + 511, + 1069, + 536, + 1069, + 536, + 1093, + 511, + 1093 + ], + "score": 0.85, + "latex": "\\scriptstyle { s _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 975, + 1678, + 997, + 1678, + 997, + 1710, + 975, + 1710 + ], + "score": 0.85, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 298, + 1483, + 319, + 1483, + 319, + 1506, + 298, + 1506 + ], + "score": 0.7, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1381, + 1007, + 1401, + 1007, + 1401, + 1029, + 1381, + 1029 + ], + "score": 0.68, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 427, + 233, + 445, + 233, + 445, + 258, + 427, + 258 + ], + "score": 0.46, + "latex": "\\theta" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1328.0, + 1008.0, + 1328.0, + 1008.0, + 1364.0, + 294.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1386.0, + 1402.0, + 1386.0, + 1402.0, + 1421.0, + 294.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1415.0, + 1405.0, + 1415.0, + 1405.0, + 1455.0, + 293.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 1448.0, + 583.0, + 1448.0, + 583.0, + 1483.0, + 337.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1448.0, + 1402.0, + 1448.0, + 1402.0, + 1483.0, + 642.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1477.0, + 297.0, + 1477.0, + 297.0, + 1514.0, + 293.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1477.0, + 608.0, + 1477.0, + 608.0, + 1514.0, + 320.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1477.0, + 1402.0, + 1477.0, + 1402.0, + 1514.0, + 648.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1509.0, + 1404.0, + 1509.0, + 1404.0, + 1544.0, + 294.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1538.0, + 1405.0, + 1538.0, + 1405.0, + 1576.0, + 293.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1570.0, + 874.0, + 1570.0, + 874.0, + 1605.0, + 294.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1136.0, + 1404.0, + 1136.0, + 1404.0, + 1175.0, + 294.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1167.0, + 1404.0, + 1167.0, + 1404.0, + 1208.0, + 295.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1199.0, + 1406.0, + 1199.0, + 1406.0, + 1237.0, + 295.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1230.0, + 1404.0, + 1230.0, + 1404.0, + 1267.0, + 294.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1261.0, + 1353.0, + 1261.0, + 1353.0, + 1297.0, + 294.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 706.0, + 1404.0, + 706.0, + 1404.0, + 739.0, + 297.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 734.0, + 1407.0, + 734.0, + 1407.0, + 775.0, + 294.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 767.0, + 1110.0, + 767.0, + 1110.0, + 800.0, + 296.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 767.0, + 1404.0, + 767.0, + 1404.0, + 800.0, + 1188.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 796.0, + 911.0, + 796.0, + 911.0, + 833.0, + 295.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 796.0, + 1404.0, + 796.0, + 1404.0, + 833.0, + 1028.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 828.0, + 396.0, + 828.0, + 396.0, + 861.0, + 296.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 828.0, + 723.0, + 828.0, + 723.0, + 861.0, + 515.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1615.0, + 1403.0, + 1615.0, + 1403.0, + 1649.0, + 297.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1642.0, + 1405.0, + 1642.0, + 1405.0, + 1683.0, + 293.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1675.0, + 371.0, + 1675.0, + 371.0, + 1714.0, + 293.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 1675.0, + 974.0, + 1675.0, + 974.0, + 1714.0, + 693.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 1675.0, + 1405.0, + 1675.0, + 1405.0, + 1714.0, + 998.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1706.0, + 343.0, + 1706.0, + 343.0, + 1743.0, + 293.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 1706.0, + 894.0, + 1706.0, + 894.0, + 1743.0, + 378.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1706.0, + 1405.0, + 1706.0, + 1405.0, + 1743.0, + 967.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1738.0, + 297.0, + 1738.0, + 297.0, + 1775.0, + 294.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 1738.0, + 596.0, + 1738.0, + 596.0, + 1775.0, + 336.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 537.0, + 1401.0, + 537.0, + 1401.0, + 570.0, + 296.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 566.0, + 1046.0, + 566.0, + 1046.0, + 603.0, + 295.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 566.0, + 1309.0, + 566.0, + 1309.0, + 603.0, + 1116.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1343.0, + 566.0, + 1405.0, + 566.0, + 1405.0, + 603.0, + 1343.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 597.0, + 1402.0, + 597.0, + 1402.0, + 630.0, + 295.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 627.0, + 1405.0, + 627.0, + 1405.0, + 664.0, + 294.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 655.0, + 414.0, + 655.0, + 414.0, + 698.0, + 294.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 1001.0, + 1380.0, + 1001.0, + 1380.0, + 1037.0, + 415.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1001.0, + 1405.0, + 1001.0, + 1405.0, + 1037.0, + 1402.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1030.0, + 465.0, + 1030.0, + 465.0, + 1068.0, + 295.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 1030.0, + 1404.0, + 1030.0, + 1404.0, + 1068.0, + 583.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1062.0, + 510.0, + 1062.0, + 510.0, + 1098.0, + 295.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 1062.0, + 759.0, + 1062.0, + 759.0, + 1098.0, + 537.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1062.0, + 928.0, + 1062.0, + 928.0, + 1098.0, + 837.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 1062.0, + 1405.0, + 1062.0, + 1405.0, + 1098.0, + 1047.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1095.0, + 638.0, + 1095.0, + 638.0, + 1128.0, + 298.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1887.0, + 1404.0, + 1887.0, + 1404.0, + 1928.0, + 294.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1920.0, + 1403.0, + 1920.0, + 1403.0, + 1957.0, + 295.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1953.0, + 902.0, + 1953.0, + 902.0, + 1988.0, + 292.0, + 1988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 1953.0, + 969.0, + 1953.0, + 969.0, + 1988.0, + 959.0, + 1988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 381.0, + 371.0, + 381.0, + 371.0, + 418.0, + 296.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 381.0, + 1406.0, + 381.0, + 1406.0, + 418.0, + 433.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 407.0, + 478.0, + 407.0, + 478.0, + 454.0, + 291.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 407.0, + 1407.0, + 407.0, + 1407.0, + 454.0, + 765.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 228.0, + 426.0, + 228.0, + 426.0, + 266.0, + 292.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 228.0, + 1406.0, + 228.0, + 1406.0, + 266.0, + 446.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 256.0, + 383.0, + 256.0, + 383.0, + 298.0, + 293.0, + 298.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1558, + 1404, + 1558, + 1404, + 1927, + 297, + 1927 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 500, + 1405, + 500, + 1405, + 748, + 297, + 748 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 843, + 1405, + 843, + 1405, + 1042, + 297, + 1042 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 1404, + 228, + 1404, + 384, + 298, + 384 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 301, + 1941, + 1401, + 1941, + 1401, + 2034, + 301, + 2034 + ], + "score": 0.971 + }, + { + "category_id": 8, + "poly": [ + 325, + 1058, + 1334, + 1058, + 1334, + 1148, + 325, + 1148 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 291, + 1162, + 1399, + 1162, + 1399, + 1227, + 291, + 1227 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 299, + 1339, + 1399, + 1339, + 1399, + 1406, + 299, + 1406 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 307, + 1421, + 1393, + 1421, + 1393, + 1511, + 307, + 1511 + ], + "score": 0.945 + }, + { + "category_id": 0, + "poly": [ + 300, + 786, + 608, + 786, + 608, + 818, + 300, + 818 + ], + "score": 0.91 + }, + { + "category_id": 0, + "poly": [ + 300, + 428, + 466, + 428, + 466, + 466, + 300, + 466 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 104, + 299, + 104 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1509, + 1400, + 1509, + 1400, + 1538, + 1366, + 1538 + ], + "score": 0.88 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1088, + 1400, + 1088, + 1400, + 1116, + 1366, + 1116 + ], + "score": 0.854 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1287, + 1400, + 1287, + 1400, + 1315, + 1366, + 1315 + ], + "score": 0.822 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.76 + }, + { + "category_id": 8, + "poly": [ + 298, + 1244, + 1394, + 1244, + 1394, + 1286, + 298, + 1286 + ], + "score": 0.657 + }, + { + "category_id": 8, + "poly": [ + 300, + 1242, + 1387, + 1242, + 1387, + 1318, + 300, + 1318 + ], + "score": 0.143 + }, + { + "category_id": 13, + "poly": [ + 847, + 1339, + 948, + 1339, + 948, + 1375, + 847, + 1375 + ], + "score": 0.94, + "latex": "\\hat { A } ( s _ { t } , a _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 713, + 1163, + 812, + 1163, + 812, + 1197, + 713, + 1197 + ], + "score": 0.94, + "latex": "f ( \\mathbf { \\boldsymbol { s } } _ { t } , \\mathbf { \\boldsymbol { a } } _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 299, + 978, + 559, + 978, + 559, + 1015, + 299, + 1015 + ], + "score": 0.93, + "latex": "\\mu _ { \\theta } ( s _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } [ a _ { t } ]" + }, + { + "category_id": 13, + "poly": [ + 1088, + 1194, + 1168, + 1194, + 1168, + 1227, + 1088, + 1227 + ], + "score": 0.92, + "latex": "\\pmb { \\mu } _ { \\pmb { \\theta } } ( \\pmb { s } _ { t } )" + }, + { + "category_id": 14, + "poly": [ + 305, + 1418, + 1388, + 1418, + 1388, + 1515, + 305, + 1515 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } & { \\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { A } ( s _ { t } , a _ { t } ) - \\bar { A } _ { w } ( s _ { t } , a _ { t } ) ] + \\mathbb { E } _ { \\rho _ { \\pi } } [ \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] } \\\\ & { \\bar { A } ( s _ { t } , a _ { t } ) = \\bar { Q } ( s _ { t } , a _ { t } ) - \\mathbb { E } _ { \\pi _ { \\theta } } [ \\bar { Q } ( s _ { t } , a _ { t } ) ] = \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } ( a _ { t } - \\mu _ { \\theta } ( s _ { t } ) ) . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 327, + 1055, + 1340, + 1055, + 1340, + 1151, + 327, + 1151 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { Q } ( s _ { t } , a _ { t } ) - \\bar { f } ( s _ { t } , a _ { t } ) ] + \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\bar { f } ( s _ { t } , a _ { t } ) ] } \\\\ { = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { Q } ( s _ { t } , a _ { t } ) - \\bar { f } ( s _ { t } , a _ { t } ) ] + \\mathbb { E } _ { \\rho _ { \\pi } } [ \\nabla _ { a } f ( s _ { t } , a ) | _ { a = \\bar { a } _ { t } } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 937, + 1196, + 975, + 1196, + 975, + 1225, + 937, + 1225 + ], + "score": 0.89, + "latex": "Q _ { w }" + }, + { + "category_id": 13, + "poly": [ + 550, + 1944, + 589, + 1944, + 589, + 1974, + 550, + 1974 + ], + "score": 0.89, + "latex": "Q _ { w }" + }, + { + "category_id": 13, + "poly": [ + 747, + 1945, + 785, + 1945, + 785, + 1974, + 747, + 1974 + ], + "score": 0.89, + "latex": "Q _ { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 856, + 948, + 1339, + 948, + 1339, + 982, + 856, + 982 + ], + "score": 0.88, + "latex": "\\mathbb { E } _ { \\pi } [ \\hat { \\cal Q } ( s _ { t } , a _ { t } ) ] = r ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { E } _ { p } [ V _ { \\pi } ( s _ { t + 1 } ) ]" + }, + { + "category_id": 13, + "poly": [ + 843, + 910, + 1197, + 910, + 1197, + 946, + 843, + 946 + ], + "score": 0.88, + "latex": "\\begin{array} { r } { \\hat { \\boldsymbol { \\mathcal Q } } ( \\boldsymbol s _ { t } , \\boldsymbol a _ { t } ) = \\sum _ { t ^ { \\prime } = t } ^ { \\infty } \\gamma ^ { t ^ { \\prime } - t } \\boldsymbol r ( \\boldsymbol s _ { t ^ { \\prime } } , \\boldsymbol a _ { t ^ { \\prime } } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 394, + 1197, + 421, + 1197, + 421, + 1224, + 394, + 1224 + ], + "score": 0.87, + "latex": "\\bar { \\mathbf { a } } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1652, + 336, + 1652, + 336, + 1683, + 297, + 1683 + ], + "score": 0.87, + "latex": "Q _ { w }" + }, + { + "category_id": 13, + "poly": [ + 1212, + 1197, + 1239, + 1197, + 1239, + 1224, + 1212, + 1224 + ], + "score": 0.87, + "latex": "\\bar { \\mathbf { a } } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1212, + 1170, + 1240, + 1170, + 1240, + 1194, + 1212, + 1194 + ], + "score": 0.86, + "latex": "\\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { a } _ { t }" + }, + { + "category_id": 14, + "poly": [ + 310, + 1242, + 1395, + 1242, + 1395, + 1287, + 310, + 1287 + ], + "score": 0.86, + "latex": "\\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { Q } ( s _ { t } , a _ { t } ) - \\bar { Q } _ { w } ( s _ { t } , a _ { t } ) ] + \\mathbb { E } _ { \\rho _ { \\pi } } [ \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] ." + }, + { + "category_id": 13, + "poly": [ + 1017, + 1196, + 1037, + 1196, + 1037, + 1226, + 1017, + 1226 + ], + "score": 0.86, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1155, + 984, + 1185, + 984, + 1185, + 1010, + 1155, + 1010 + ], + "score": 0.84, + "latex": "\\pi _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 582, + 954, + 609, + 954, + 609, + 979, + 582, + 979 + ], + "score": 0.83, + "latex": "\\scriptstyle { \\pmb { s } } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 752, + 953, + 780, + 953, + 780, + 979, + 752, + 979 + ], + "score": 0.82, + "latex": "\\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { a } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 507, + 874, + 1169, + 874, + 1169, + 909, + 507, + 909 + ], + "score": 0.79, + "latex": "\\hat { f ( s _ { t } , a _ { t } ) , ~ } \\bar { f } ( s _ { t } , a _ { t } ) = f ( s _ { t } , \\bar { a } _ { t } ) + \\bar { \\nabla } _ { a } f ( \\bar { s _ { t } } , a ) | _ { a = \\bar { a } _ { t } } ( a _ { t } - \\bar { a } _ { t } ) \\cdot" + }, + { + "category_id": 13, + "poly": [ + 796, + 1895, + 832, + 1895, + 832, + 1926, + 796, + 1926 + ], + "score": 0.6, + "latex": "( \\lambda )" + }, + { + "category_id": 13, + "poly": [ + 507, + 875, + 607, + 875, + 607, + 909, + 507, + 909 + ], + "score": 0.58, + "latex": "f ( \\mathbf { \\sigma } _ \\mathbf { \\sigma } _ { \\mathbf { \\sigma } } , \\mathbf { \\sigma } _ { \\mathbf { \\sigma } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } } )" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 784.0, + 613.0, + 784.0, + 613.0, + 824.0, + 294.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 422.0, + 474.0, + 422.0, + 474.0, + 480.0, + 288.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1559.0, + 1405.0, + 1559.0, + 1405.0, + 1597.0, + 295.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1590.0, + 1406.0, + 1590.0, + 1406.0, + 1628.0, + 294.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1622.0, + 1406.0, + 1622.0, + 1406.0, + 1656.0, + 295.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 1651.0, + 1406.0, + 1651.0, + 1406.0, + 1689.0, + 337.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1683.0, + 1406.0, + 1683.0, + 1406.0, + 1717.0, + 294.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1713.0, + 1405.0, + 1713.0, + 1405.0, + 1747.0, + 295.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1741.0, + 1404.0, + 1741.0, + 1404.0, + 1778.0, + 292.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 1405.0, + 1773.0, + 1405.0, + 1811.0, + 294.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1802.0, + 1407.0, + 1802.0, + 1407.0, + 1840.0, + 294.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1834.0, + 1405.0, + 1834.0, + 1405.0, + 1870.0, + 292.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1863.0, + 1406.0, + 1863.0, + 1406.0, + 1901.0, + 294.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1895.0, + 795.0, + 1895.0, + 795.0, + 1929.0, + 295.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 1895.0, + 1074.0, + 1895.0, + 1074.0, + 1929.0, + 833.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 501.0, + 1405.0, + 501.0, + 1405.0, + 537.0, + 294.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 535.0, + 1403.0, + 535.0, + 1403.0, + 565.0, + 296.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 564.0, + 1405.0, + 564.0, + 1405.0, + 597.0, + 294.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 594.0, + 1406.0, + 594.0, + 1406.0, + 629.0, + 292.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 620.0, + 1406.0, + 620.0, + 1406.0, + 659.0, + 292.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 654.0, + 1406.0, + 654.0, + 1406.0, + 690.0, + 292.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 684.0, + 1402.0, + 684.0, + 1402.0, + 722.0, + 294.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 715.0, + 625.0, + 715.0, + 625.0, + 752.0, + 294.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 841.0, + 1405.0, + 841.0, + 1405.0, + 880.0, + 295.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 874.0, + 506.0, + 874.0, + 506.0, + 912.0, + 295.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 874.0, + 1403.0, + 874.0, + 1403.0, + 912.0, + 1170.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 906.0, + 842.0, + 906.0, + 842.0, + 952.0, + 292.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 906.0, + 1407.0, + 906.0, + 1407.0, + 952.0, + 1198.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 947.0, + 581.0, + 947.0, + 581.0, + 984.0, + 296.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 947.0, + 751.0, + 947.0, + 751.0, + 984.0, + 610.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 947.0, + 855.0, + 947.0, + 855.0, + 984.0, + 781.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1340.0, + 947.0, + 1405.0, + 947.0, + 1405.0, + 984.0, + 1340.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 976.0, + 298.0, + 976.0, + 298.0, + 1017.0, + 292.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 976.0, + 1154.0, + 976.0, + 1154.0, + 1017.0, + 560.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 976.0, + 1407.0, + 976.0, + 1407.0, + 1017.0, + 1186.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1009.0, + 476.0, + 1009.0, + 476.0, + 1044.0, + 295.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 266.0, + 296.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 262.0, + 1404.0, + 262.0, + 1404.0, + 295.0, + 296.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 289.0, + 1404.0, + 289.0, + 1404.0, + 326.0, + 293.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 314.0, + 1406.0, + 314.0, + 1406.0, + 362.0, + 292.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 354.0, + 833.0, + 354.0, + 833.0, + 387.0, + 294.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1939.0, + 549.0, + 1939.0, + 549.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1939.0, + 746.0, + 1939.0, + 746.0, + 1978.0, + 590.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1939.0, + 1405.0, + 1939.0, + 1405.0, + 1978.0, + 786.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1286.0, + 2002.0, + 1286.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1160.0, + 712.0, + 1160.0, + 712.0, + 1201.0, + 293.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1160.0, + 1211.0, + 1160.0, + 1211.0, + 1201.0, + 813.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 1160.0, + 1404.0, + 1160.0, + 1404.0, + 1201.0, + 1241.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1191.0, + 393.0, + 1191.0, + 393.0, + 1229.0, + 294.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1191.0, + 936.0, + 1191.0, + 936.0, + 1229.0, + 422.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 1191.0, + 1016.0, + 1191.0, + 1016.0, + 1229.0, + 976.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 1191.0, + 1087.0, + 1191.0, + 1087.0, + 1229.0, + 1038.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1191.0, + 1211.0, + 1191.0, + 1211.0, + 1229.0, + 1169.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 1191.0, + 1321.0, + 1191.0, + 1321.0, + 1229.0, + 1240.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1336.0, + 846.0, + 1336.0, + 846.0, + 1380.0, + 293.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1336.0, + 1404.0, + 1336.0, + 1404.0, + 1380.0, + 949.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1371.0, + 823.0, + 1371.0, + 823.0, + 1409.0, + 294.0, + 1409.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1294, + 1405, + 1294, + 1405, + 1593, + 296, + 1593 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 296, + 665, + 1405, + 665, + 1405, + 916, + 296, + 916 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1624, + 1405, + 1624, + 1405, + 1879, + 297, + 1879 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 285, + 1405, + 285, + 1405, + 440, + 297, + 440 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 300, + 1168, + 1398, + 1168, + 1398, + 1267, + 300, + 1267 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 298, + 1971, + 1401, + 1971, + 1401, + 2035, + 298, + 2035 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 378, + 1069, + 1292, + 1069, + 1292, + 1159, + 378, + 1159 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 502, + 446, + 1196, + 446, + 1196, + 527, + 502, + 527 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 456, + 576, + 1246, + 576, + 1246, + 652, + 456, + 652 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 298, + 1022, + 1166, + 1022, + 1166, + 1058, + 298, + 1058 + ], + "score": 0.926 + }, + { + "category_id": 8, + "poly": [ + 400, + 925, + 1299, + 925, + 1299, + 1011, + 400, + 1011 + ], + "score": 0.923 + }, + { + "category_id": 0, + "poly": [ + 300, + 1914, + 618, + 1914, + 618, + 1947, + 300, + 1947 + ], + "score": 0.91 + }, + { + "category_id": 1, + "poly": [ + 295, + 535, + 1317, + 535, + 1317, + 568, + 295, + 568 + ], + "score": 0.907 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1097, + 1400, + 1097, + 1400, + 1127, + 1352, + 1127 + ], + "score": 0.905 + }, + { + "category_id": 9, + "poly": [ + 1352, + 599, + 1400, + 599, + 1400, + 630, + 1352, + 630 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.895 + }, + { + "category_id": 9, + "poly": [ + 1351, + 952, + 1400, + 952, + 1400, + 982, + 1351, + 982 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1365, + 472, + 1400, + 472, + 1400, + 501, + 1365, + 501 + ], + "score": 0.87 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.736 + }, + { + "category_id": 0, + "poly": [ + 297, + 229, + 1018, + 229, + 1018, + 262, + 297, + 262 + ], + "score": 0.664 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1018, + 229, + 1018, + 262, + 297, + 262 + ], + "score": 0.196 + }, + { + "category_id": 13, + "poly": [ + 935, + 1297, + 1001, + 1297, + 1001, + 1331, + 935, + 1331 + ], + "score": 0.94, + "latex": "\\eta ( s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 473, + 1688, + 655, + 1688, + 655, + 1725, + 473, + 1725 + ], + "score": 0.94, + "latex": "\\hat { \\mathrm { C o v } } _ { { \\pmb a } _ { t } } ( \\hat { A } , \\bar { A } ) > \\bar { 0 }" + }, + { + "category_id": 13, + "poly": [ + 452, + 668, + 518, + 668, + 518, + 702, + 452, + 702 + ], + "score": 0.94, + "latex": "\\eta ( s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 1221, + 1624, + 1357, + 1624, + 1357, + 1661, + 1221, + 1661 + ], + "score": 0.94, + "latex": "\\mathrm { C o v } _ { a _ { t } } ( \\hat { A } , \\bar { A } )" + }, + { + "category_id": 13, + "poly": [ + 643, + 1199, + 780, + 1199, + 780, + 1236, + 643, + 1236 + ], + "score": 0.94, + "latex": "\\mathrm { C o v } _ { a _ { t } } ( \\hat { A } , \\bar { A } )" + }, + { + "category_id": 14, + "poly": [ + 452, + 574, + 1250, + 574, + 1250, + 653, + 452, + 653 + ], + "score": 0.93, + "latex": "\\mathrm { V a r } ^ { * } = \\mathbb { E } _ { \\rho _ { \\pi } } \\left[ \\sum _ { m } \\mathrm { V a r } _ { a _ { t } } ( \\nabla _ { \\theta _ { m } } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { A } ( s _ { t } , a _ { t } ) - \\eta ( s _ { t } ) \\bar { A } ( s _ { t } , a _ { t } ) ) ) \\right] ." + }, + { + "category_id": 13, + "poly": [ + 347, + 1202, + 448, + 1202, + 448, + 1237, + 347, + 1237 + ], + "score": 0.93, + "latex": "\\mathrm { V a r } _ { \\pmb { a } _ { t } } ( \\bar { A } )" + }, + { + "category_id": 13, + "poly": [ + 1043, + 1688, + 1341, + 1688, + 1341, + 1726, + 1043, + 1726 + ], + "score": 0.93, + "latex": "\\eta ( \\pmb { \\mathscr { s } } _ { t } ) = \\mathrm { s i g n } \\big ( \\mathsf { C } \\hat { \\operatorname { o v } } _ { \\pmb { a } _ { t } } ( \\hat { A } , \\bar { A } ) \\big )" + }, + { + "category_id": 13, + "poly": [ + 374, + 1329, + 733, + 1329, + 733, + 1366, + 374, + 1366 + ], + "score": 0.93, + "latex": "\\eta ^ { * } ( s _ { t } ) = \\mathrm { C o v } _ { a _ { t } } ( \\hat { A } , \\bar { A } ) / \\mathrm { V a r } _ { a _ { t } } ( \\bar { A } )" + }, + { + "category_id": 13, + "poly": [ + 567, + 1368, + 645, + 1368, + 645, + 1401, + 567, + 1401 + ], + "score": 0.92, + "latex": "\\eta ^ { * } ( s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 703, + 1691, + 815, + 1691, + 815, + 1725, + 703, + 1725 + ], + "score": 0.92, + "latex": "\\eta ( s _ { t } ) = 0" + }, + { + "category_id": 13, + "poly": [ + 947, + 349, + 1012, + 349, + 1012, + 382, + 947, + 382 + ], + "score": 0.92, + "latex": "\\eta ( s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 1039, + 379, + 1104, + 379, + 1104, + 412, + 1039, + 412 + ], + "score": 0.92, + "latex": "\\eta ( s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 372, + 1168, + 451, + 1168, + 451, + 1201, + 372, + 1201 + ], + "score": 0.92, + "latex": "\\Sigma _ { \\theta } ( s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 366, + 1020, + 742, + 1020, + 742, + 1058, + 366, + 1058 + ], + "score": 0.92, + "latex": "\\mathbb { E } _ { \\pi } [ \\hat { A } ( \\pmb { s } _ { t } , \\pmb { a } _ { t } ) ] = \\mathbb { E } _ { \\pi } [ \\bar { A } ( \\pmb { s } _ { t } , \\pmb { a } _ { t } ) ] = 0" + }, + { + "category_id": 13, + "poly": [ + 859, + 820, + 1171, + 820, + 1171, + 858, + 859, + 858 + ], + "score": 0.91, + "latex": "\\operatorname { V a r } = \\mathbb { E } _ { \\rho _ { \\pi } } [ \\operatorname { V a r } _ { \\mathbf { a } _ { t } } ( \\hat { A } ( s _ { t } , \\mathbf { a } _ { t } ) ) ]" + }, + { + "category_id": 13, + "poly": [ + 877, + 1363, + 1316, + 1363, + 1316, + 1402, + 877, + 1402 + ], + "score": 0.91, + "latex": "\\mathrm { V a r } ^ { * } = \\mathbb { E } _ { \\rho _ { \\pi } } [ ( 1 - \\rho _ { c o r r } ( \\hat { A } , \\bar { A } ) ^ { 2 } ) \\mathrm { V a r } _ { a _ { t } } ( \\hat { A } ) ]" + }, + { + "category_id": 14, + "poly": [ + 501, + 442, + 1198, + 442, + 1198, + 531, + 501, + 531 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } & { \\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { A } ( s _ { t } , a _ { t } ) - \\eta ( s _ { t } ) \\bar { A } _ { w } ( s _ { t } , a _ { t } ) ] } \\\\ & { \\qquad + \\mathbb { E } _ { \\rho _ { \\pi } } [ \\eta ( s _ { t } ) \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 336, + 1690, + 447, + 1690, + 447, + 1725, + 336, + 1725 + ], + "score": 0.91, + "latex": "\\eta ( s _ { t } ) = 1" + }, + { + "category_id": 13, + "poly": [ + 880, + 537, + 999, + 537, + 999, + 564, + 880, + 564 + ], + "score": 0.9, + "latex": "m = 1 . . . M" + }, + { + "category_id": 13, + "poly": [ + 846, + 1468, + 884, + 1468, + 884, + 1497, + 846, + 1497 + ], + "score": 0.89, + "latex": "Q _ { w }" + }, + { + "category_id": 14, + "poly": [ + 358, + 1066, + 1294, + 1066, + 1294, + 1160, + 358, + 1160 + ], + "score": 0.89, + "latex": "\\begin{array} { r l } & { \\mathrm { C o v } _ { a _ { t } } ( \\hat { A } , \\bar { A } ) = \\mathbb { E } _ { \\pi } [ \\hat { A } ( s _ { t } , a _ { t } ) \\bar { A } ( s _ { t } , a _ { t } ) ] } \\\\ & { \\quad \\mathrm { V a r } _ { a _ { t } } ( \\bar { A } ) = \\mathbb { E } _ { \\pi } [ \\bar { A } ( s _ { t } , a _ { t } ) ^ { 2 } ] = \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } ^ { T } \\Sigma _ { \\theta } ( s _ { t } ) \\nabla _ { a } Q _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } , } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 395, + 921, + 1302, + 921, + 1302, + 1013, + 395, + 1013 + ], + "score": 0.88, + "latex": "\\begin{array} { r l } & { \\mathrm { V a r } ^ { * } = \\mathbb { E } _ { \\rho _ { \\pi } } [ \\mathrm { V a r } _ { a _ { t } } ( \\hat { A } ( s _ { t } , a _ { t } ) - \\eta ( s _ { t } ) \\bar { A } ( s _ { t } , a _ { t } ) ) ] } \\\\ & { \\qquad = \\mathrm { V a r } + \\mathbb { E } _ { \\rho _ { \\pi } } [ - 2 \\eta ( s _ { t } ) \\mathrm { C o v } _ { a _ { t } } ( \\hat { A } ( s _ { t } , a _ { t } ) , \\bar { A } ( s _ { t } , a _ { t } ) ) + \\eta ( s _ { t } ) ^ { 2 } \\mathrm { V a r } _ { a _ { t } } ( \\bar { A } ( s _ { t } , a _ { t } ) ) ] . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 634, + 668, + 763, + 668, + 763, + 699, + 634, + 699 + ], + "score": 0.88, + "latex": "\\mathrm { V a r ^ { * } } < \\mathrm { V a r }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1500, + 335, + 1500, + 335, + 1531, + 297, + 1531 + ], + "score": 0.87, + "latex": "Q _ { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 852, + 664, + 1402, + 664, + 1402, + 704, + 852, + 704 + ], + "score": 0.87, + "latex": "\\begin{array} { r } { \\mathrm { V a r } = \\mathbb E _ { \\rho _ { \\pi } } [ \\sum _ { m } \\mathrm { V a r } _ { \\boldsymbol { a } _ { t } } ( \\nabla _ { \\boldsymbol { \\theta } _ { m } } \\log \\pi _ { \\boldsymbol { \\theta } } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ) \\hat { \\boldsymbol { A } } ( \\boldsymbol { s } _ { t } , \\boldsymbol { a } _ { t } ) ) ] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1005, + 1173, + 1037, + 1173, + 1037, + 1200, + 1005, + 1200 + ], + "score": 0.86, + "latex": "\\pi _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1353, + 1399, + 1375, + 1399, + 1375, + 1429, + 1353, + 1429 + ], + "score": 0.84, + "latex": "\\bar { A }" + }, + { + "category_id": 13, + "poly": [ + 1308, + 1496, + 1330, + 1496, + 1330, + 1527, + 1308, + 1527 + ], + "score": 0.84, + "latex": "\\hat { A }" + }, + { + "category_id": 13, + "poly": [ + 472, + 1783, + 494, + 1783, + 494, + 1814, + 472, + 1814 + ], + "score": 0.84, + "latex": "\\hat { A }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1404, + 354, + 1404, + 354, + 1434, + 297, + 1434 + ], + "score": 0.83, + "latex": "\\rho _ { c o r r }" + }, + { + "category_id": 13, + "poly": [ + 473, + 1432, + 495, + 1432, + 495, + 1463, + 473, + 1463 + ], + "score": 0.82, + "latex": "\\hat { A }" + }, + { + "category_id": 13, + "poly": [ + 542, + 1784, + 564, + 1784, + 564, + 1814, + 542, + 1814 + ], + "score": 0.82, + "latex": "\\bar { A }" + }, + { + "category_id": 13, + "poly": [ + 1288, + 538, + 1305, + 538, + 1305, + 563, + 1288, + 563 + ], + "score": 0.79, + "latex": "\\theta" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1912.0, + 623.0, + 1912.0, + 623.0, + 1951.0, + 293.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 226.0, + 1021.0, + 226.0, + 1021.0, + 266.0, + 292.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1295.0, + 934.0, + 1295.0, + 934.0, + 1334.0, + 295.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 1295.0, + 1403.0, + 1295.0, + 1403.0, + 1334.0, + 1002.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1330.0, + 373.0, + 1330.0, + 373.0, + 1369.0, + 294.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1330.0, + 1406.0, + 1330.0, + 1406.0, + 1369.0, + 734.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1364.0, + 566.0, + 1364.0, + 566.0, + 1405.0, + 291.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 1364.0, + 876.0, + 1364.0, + 876.0, + 1405.0, + 646.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 1364.0, + 1406.0, + 1364.0, + 1406.0, + 1405.0, + 1317.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1401.0, + 296.0, + 1401.0, + 296.0, + 1436.0, + 291.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1401.0, + 1352.0, + 1401.0, + 1352.0, + 1436.0, + 355.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 1401.0, + 1406.0, + 1401.0, + 1406.0, + 1436.0, + 1376.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1435.0, + 472.0, + 1435.0, + 472.0, + 1470.0, + 294.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 1435.0, + 1405.0, + 1435.0, + 1405.0, + 1470.0, + 496.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1462.0, + 845.0, + 1462.0, + 845.0, + 1503.0, + 293.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 1462.0, + 1405.0, + 1462.0, + 1405.0, + 1503.0, + 885.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 1498.0, + 1307.0, + 1498.0, + 1307.0, + 1533.0, + 336.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 1498.0, + 1403.0, + 1498.0, + 1403.0, + 1533.0, + 1331.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1528.0, + 1403.0, + 1528.0, + 1403.0, + 1566.0, + 294.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1559.0, + 1027.0, + 1559.0, + 1027.0, + 1598.0, + 295.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 665.0, + 451.0, + 665.0, + 451.0, + 706.0, + 291.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 665.0, + 633.0, + 665.0, + 633.0, + 706.0, + 519.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 665.0, + 851.0, + 665.0, + 851.0, + 706.0, + 764.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 665.0, + 1406.0, + 665.0, + 1406.0, + 706.0, + 1403.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 698.0, + 1403.0, + 698.0, + 1403.0, + 733.0, + 291.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 728.0, + 1405.0, + 728.0, + 1405.0, + 764.0, + 293.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 758.0, + 1405.0, + 758.0, + 1405.0, + 794.0, + 294.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 790.0, + 1403.0, + 790.0, + 1403.0, + 825.0, + 295.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 823.0, + 858.0, + 823.0, + 858.0, + 858.0, + 294.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 823.0, + 1405.0, + 823.0, + 1405.0, + 858.0, + 1172.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 856.0, + 1403.0, + 856.0, + 1403.0, + 887.0, + 295.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 886.0, + 869.0, + 886.0, + 869.0, + 916.0, + 296.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1623.0, + 1220.0, + 1623.0, + 1220.0, + 1665.0, + 294.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1358.0, + 1623.0, + 1406.0, + 1623.0, + 1406.0, + 1665.0, + 1358.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1655.0, + 1405.0, + 1655.0, + 1405.0, + 1696.0, + 291.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1690.0, + 335.0, + 1690.0, + 335.0, + 1728.0, + 295.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1690.0, + 472.0, + 1690.0, + 472.0, + 1728.0, + 448.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 1690.0, + 702.0, + 1690.0, + 702.0, + 1728.0, + 656.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 1690.0, + 1042.0, + 1690.0, + 1042.0, + 1728.0, + 816.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1342.0, + 1690.0, + 1403.0, + 1690.0, + 1403.0, + 1728.0, + 1342.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1723.0, + 1405.0, + 1723.0, + 1405.0, + 1758.0, + 295.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1753.0, + 1405.0, + 1753.0, + 1405.0, + 1787.0, + 295.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1786.0, + 471.0, + 1786.0, + 471.0, + 1821.0, + 295.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1786.0, + 541.0, + 1786.0, + 541.0, + 1821.0, + 495.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 1786.0, + 1405.0, + 1786.0, + 1405.0, + 1821.0, + 565.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1817.0, + 1406.0, + 1817.0, + 1406.0, + 1850.0, + 294.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1849.0, + 471.0, + 1849.0, + 471.0, + 1881.0, + 295.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 285.0, + 1405.0, + 285.0, + 1405.0, + 323.0, + 294.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 317.0, + 1405.0, + 317.0, + 1405.0, + 355.0, + 294.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 348.0, + 946.0, + 348.0, + 946.0, + 384.0, + 292.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 348.0, + 1406.0, + 348.0, + 1406.0, + 384.0, + 1013.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 376.0, + 1038.0, + 376.0, + 1038.0, + 414.0, + 292.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 376.0, + 1405.0, + 376.0, + 1405.0, + 414.0, + 1105.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 407.0, + 454.0, + 407.0, + 454.0, + 445.0, + 294.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1164.0, + 371.0, + 1164.0, + 371.0, + 1205.0, + 294.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 1164.0, + 1004.0, + 1164.0, + 1004.0, + 1205.0, + 452.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 1164.0, + 1407.0, + 1164.0, + 1407.0, + 1205.0, + 1038.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1200.0, + 346.0, + 1200.0, + 346.0, + 1239.0, + 295.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1200.0, + 642.0, + 1200.0, + 642.0, + 1239.0, + 449.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 1200.0, + 1404.0, + 1200.0, + 1404.0, + 1239.0, + 781.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1234.0, + 1403.0, + 1234.0, + 1403.0, + 1268.0, + 296.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1970.0, + 1403.0, + 1970.0, + 1403.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1999.0, + 1405.0, + 1999.0, + 1405.0, + 2041.0, + 292.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1018.0, + 365.0, + 1018.0, + 365.0, + 1063.0, + 295.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 1018.0, + 1167.0, + 1018.0, + 1167.0, + 1063.0, + 743.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 534.0, + 879.0, + 534.0, + 879.0, + 569.0, + 295.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 534.0, + 1287.0, + 534.0, + 1287.0, + 569.0, + 1000.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 534.0, + 1319.0, + 534.0, + 1319.0, + 569.0, + 1306.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 226.0, + 1021.0, + 226.0, + 1021.0, + 266.0, + 292.0, + 266.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1697, + 1404, + 1697, + 1404, + 2034, + 297, + 2034 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1231, + 1403, + 1231, + 1403, + 1416, + 298, + 1416 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 947, + 1403, + 947, + 1403, + 1140, + 298, + 1140 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1431, + 1403, + 1431, + 1403, + 1585, + 298, + 1585 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 783, + 1402, + 783, + 1402, + 877, + 298, + 877 + ], + "score": 0.976 + }, + { + "category_id": 8, + "poly": [ + 404, + 891, + 1291, + 891, + 1291, + 935, + 404, + 935 + ], + "score": 0.916 + }, + { + "category_id": 0, + "poly": [ + 301, + 1629, + 587, + 1629, + 587, + 1664, + 301, + 1664 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 814, + 75, + 814, + 105, + 300, + 105 + ], + "score": 0.899 + }, + { + "category_id": 9, + "poly": [ + 1353, + 898, + 1399, + 898, + 1399, + 929, + 1353, + 929 + ], + "score": 0.88 + }, + { + "category_id": 0, + "poly": [ + 300, + 1175, + 524, + 1175, + 524, + 1206, + 300, + 1206 + ], + "score": 0.869 + }, + { + "category_id": 0, + "poly": [ + 299, + 229, + 644, + 229, + 644, + 262, + 299, + 262 + ], + "score": 0.824 + }, + { + "category_id": 1, + "poly": [ + 299, + 272, + 1398, + 272, + 1398, + 722, + 299, + 722 + ], + "score": 0.801 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 859, + 2089, + 859, + 2112, + 840, + 2112 + ], + "score": 0.784 + }, + { + "category_id": 13, + "poly": [ + 786, + 593, + 984, + 593, + 984, + 628, + 786, + 628 + ], + "score": 0.92, + "latex": "l _ { t , e } = \\hat { A } _ { t , e } - \\eta _ { t , e } \\bar { A } _ { t , e }" + }, + { + "category_id": 13, + "poly": [ + 411, + 502, + 480, + 502, + 480, + 535, + 411, + 535 + ], + "score": 0.91, + "latex": "V _ { \\phi } ( \\pmb { s } _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 475, + 533, + 516, + 533, + 516, + 567, + 475, + 567 + ], + "score": 0.91, + "latex": "\\hat { A } _ { t , e }" + }, + { + "category_id": 13, + "poly": [ + 711, + 534, + 752, + 534, + 752, + 567, + 711, + 567 + ], + "score": 0.91, + "latex": "\\bar { A } _ { t , e }" + }, + { + "category_id": 13, + "poly": [ + 1272, + 816, + 1311, + 816, + 1311, + 846, + 1272, + 846 + ], + "score": 0.9, + "latex": "Q _ { w }" + }, + { + "category_id": 13, + "poly": [ + 533, + 448, + 913, + 448, + 913, + 477, + 533, + 477 + ], + "score": 0.9, + "latex": "\\mathcal { B } = \\left. s _ { 0 : T , 1 : E } , \\pmb { a } _ { 0 : T - 1 , 1 : E } , r _ { 0 : T - 1 , 1 : E } \\right." + }, + { + "category_id": 13, + "poly": [ + 709, + 662, + 790, + 662, + 790, + 692, + 709, + 692 + ], + "score": 0.89, + "latex": "\\nabla _ { \\boldsymbol { \\theta } } J ( \\boldsymbol { \\theta } )" + }, + { + "category_id": 13, + "poly": [ + 1041, + 273, + 1115, + 273, + 1115, + 298, + 1041, + 298 + ], + "score": 0.89, + "latex": "{ \\mathcal { R } } \\gets \\emptyset" + }, + { + "category_id": 14, + "poly": [ + 404, + 889, + 1293, + 889, + 1293, + 940, + 404, + 940 + ], + "score": 0.88, + "latex": "\\begin{array} { r } { w = \\underset { w } { \\arg \\operatorname* { m i n } } \\mathbb { E } _ { s _ { t } \\sim \\rho _ { \\beta } ( \\cdot ) , a _ { t } \\sim \\beta ( \\cdot | s _ { t } ) } [ \\big ( r ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { E } _ { \\pi } [ Q ^ { \\prime } ( s _ { t + 1 } , a _ { t + 1 } ) ] - Q _ { w } ( s _ { t } , a _ { t } ) \\big ) ^ { 2 } ] . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 418, + 567, + 456, + 567, + 456, + 594, + 418, + 594 + ], + "score": 0.88, + "latex": "\\eta _ { t , e }" + }, + { + "category_id": 13, + "poly": [ + 430, + 476, + 481, + 476, + 481, + 500, + 430, + 500 + ], + "score": 0.87, + "latex": "E \\cdot T" + }, + { + "category_id": 13, + "poly": [ + 1073, + 1325, + 1113, + 1325, + 1113, + 1356, + 1073, + 1356 + ], + "score": 0.87, + "latex": "Q _ { w }" + }, + { + "category_id": 13, + "poly": [ + 660, + 478, + 695, + 478, + 695, + 503, + 660, + 503 + ], + "score": 0.86, + "latex": "Q _ { w }" + }, + { + "category_id": 13, + "poly": [ + 860, + 819, + 891, + 819, + 891, + 846, + 860, + 846 + ], + "score": 0.86, + "latex": "\\pi _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 297, + 951, + 328, + 951, + 328, + 985, + 297, + 985 + ], + "score": 0.85, + "latex": "V _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 825, + 278, + 854, + 278, + 854, + 301, + 825, + 301 + ], + "score": 0.84, + "latex": "\\pi _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 398, + 693, + 427, + 693, + 427, + 719, + 398, + 719 + ], + "score": 0.84, + "latex": "\\pi _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 616, + 666, + 645, + 666, + 645, + 691, + 616, + 691 + ], + "score": 0.84, + "latex": "\\pi _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 477, + 627, + 1341, + 627, + 1341, + 661, + 477, + 661 + ], + "score": 0.84, + "latex": "\\begin{array} { r } { \\nabla _ { \\theta } J ( \\theta ) \\approx \\frac { 1 } { E T } \\sum _ { e } \\sum _ { t } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t , e } | s _ { t , e } ) l _ { t , e } + \\eta _ { t , e } \\nabla _ { a } Q _ { w } ( s _ { t , e } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t , e } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t , e } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 550, + 274, + 584, + 274, + 584, + 301, + 550, + 301 + ], + "score": 0.84, + "latex": "Q _ { w }" + }, + { + "category_id": 13, + "poly": [ + 1331, + 781, + 1354, + 781, + 1354, + 812, + 1331, + 812 + ], + "score": 0.83, + "latex": "\\hat { A }" + }, + { + "category_id": 13, + "poly": [ + 831, + 480, + 859, + 480, + 859, + 503, + 831, + 503 + ], + "score": 0.83, + "latex": "\\pi _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 854, + 982, + 876, + 982, + 876, + 1013, + 854, + 1013 + ], + "score": 0.81, + "latex": "\\hat { A }" + }, + { + "category_id": 13, + "poly": [ + 758, + 477, + 785, + 477, + 785, + 500, + 758, + 500 + ], + "score": 0.8, + "latex": "\\mathcal { R }" + }, + { + "category_id": 13, + "poly": [ + 295, + 812, + 317, + 812, + 317, + 842, + 295, + 842 + ], + "score": 0.8, + "latex": "\\bar { A }" + }, + { + "category_id": 13, + "poly": [ + 544, + 503, + 572, + 503, + 572, + 528, + 544, + 528 + ], + "score": 0.79, + "latex": "\\mathcal { B }" + }, + { + "category_id": 13, + "poly": [ + 1075, + 447, + 1103, + 447, + 1103, + 472, + 1075, + 472 + ], + "score": 0.79, + "latex": "\\mathcal { R }" + }, + { + "category_id": 13, + "poly": [ + 1370, + 332, + 1400, + 332, + 1400, + 357, + 1370, + 357 + ], + "score": 0.77, + "latex": "\\pi _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 952, + 330, + 973, + 330, + 973, + 353, + 952, + 353 + ], + "score": 0.77, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 414, + 329, + 537, + 329, + 537, + 355, + 414, + 355 + ], + "score": 0.73, + "latex": "e = 1 , \\ldots , E" + }, + { + "category_id": 13, + "poly": [ + 434, + 278, + 455, + 278, + 455, + 298, + 434, + 298 + ], + "score": 0.72, + "latex": "w" + }, + { + "category_id": 13, + "poly": [ + 596, + 274, + 613, + 274, + 613, + 298, + 596, + 298 + ], + "score": 0.69, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 580, + 535, + 667, + 535, + 667, + 564, + 580, + 564 + ], + "score": 0.66, + "latex": "\\mathrm { G A E } ( \\lambda )" + }, + { + "category_id": 13, + "poly": [ + 1262, + 664, + 1290, + 664, + 1290, + 687, + 1262, + 687 + ], + "score": 0.65, + "latex": "\\mathcal { B }" + }, + { + "category_id": 13, + "poly": [ + 628, + 410, + 851, + 410, + 851, + 443, + 628, + 443 + ], + "score": 0.55, + "latex": "\\pmb { s } _ { t + 1 , e } \\sim p ( \\cdot | \\pmb { s } _ { t , e } , \\pmb { a } _ { t , e } )" + }, + { + "category_id": 13, + "poly": [ + 948, + 1554, + 985, + 1554, + 985, + 1585, + 948, + 1585 + ], + "score": 0.4, + "latex": "( \\lambda )" + }, + { + "category_id": 13, + "poly": [ + 860, + 410, + 1038, + 410, + 1038, + 443, + 860, + 443 + ], + "score": 0.33, + "latex": "r _ { t , e } = r ( \\pmb { s } _ { t , e } , \\pmb { a } _ { t , e } )" + }, + { + "category_id": 13, + "poly": [ + 414, + 356, + 545, + 356, + 545, + 386, + 414, + 386 + ], + "score": 0.3, + "latex": "s _ { 0 , e } \\sim p ( s _ { 0 } )" + }, + { + "category_id": 14, + "poly": [ + 449, + 411, + 1037, + 411, + 1037, + 442, + 449, + 442 + ], + "score": 0.28, + "latex": "a _ { t , e } \\sim \\pi _ { \\theta } ( \\cdot | s _ { t , e } ) , s _ { t + 1 , e } \\sim p ( \\cdot | s _ { t , e } , a _ { t , e } ) , r _ { t , e } = r ( s _ { t , e } , a _ { t , e } )" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1626.0, + 593.0, + 1626.0, + 593.0, + 1670.0, + 292.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1174.0, + 527.0, + 1174.0, + 527.0, + 1210.0, + 295.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 219.0, + 646.0, + 219.0, + 646.0, + 275.0, + 293.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 862.0, + 2087.0, + 862.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1694.0, + 1405.0, + 1694.0, + 1405.0, + 1735.0, + 294.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1726.0, + 1404.0, + 1726.0, + 1404.0, + 1764.0, + 294.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1759.0, + 1404.0, + 1759.0, + 1404.0, + 1795.0, + 292.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1789.0, + 1405.0, + 1789.0, + 1405.0, + 1826.0, + 292.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1817.0, + 1404.0, + 1817.0, + 1404.0, + 1856.0, + 292.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1851.0, + 1406.0, + 1851.0, + 1406.0, + 1885.0, + 292.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1883.0, + 1404.0, + 1883.0, + 1404.0, + 1914.0, + 296.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1911.0, + 1406.0, + 1911.0, + 1406.0, + 1946.0, + 295.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1977.0, + 295.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1974.0, + 1405.0, + 1974.0, + 1405.0, + 2006.0, + 292.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2005.0, + 1405.0, + 2005.0, + 1405.0, + 2036.0, + 296.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1229.0, + 1405.0, + 1229.0, + 1405.0, + 1268.0, + 293.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1262.0, + 1405.0, + 1262.0, + 1405.0, + 1297.0, + 293.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1292.0, + 1405.0, + 1292.0, + 1405.0, + 1329.0, + 293.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1326.0, + 1072.0, + 1326.0, + 1072.0, + 1358.0, + 296.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 1326.0, + 1403.0, + 1326.0, + 1403.0, + 1358.0, + 1114.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1354.0, + 1407.0, + 1354.0, + 1407.0, + 1390.0, + 293.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1386.0, + 461.0, + 1386.0, + 461.0, + 1418.0, + 294.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 948.0, + 1405.0, + 948.0, + 1405.0, + 985.0, + 329.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 983.0, + 853.0, + 983.0, + 853.0, + 1020.0, + 294.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 983.0, + 1405.0, + 983.0, + 1405.0, + 1020.0, + 877.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1012.0, + 1405.0, + 1012.0, + 1405.0, + 1051.0, + 293.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1045.0, + 1404.0, + 1045.0, + 1404.0, + 1083.0, + 292.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1076.0, + 1405.0, + 1076.0, + 1405.0, + 1113.0, + 294.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1108.0, + 824.0, + 1108.0, + 824.0, + 1143.0, + 293.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1429.0, + 1405.0, + 1429.0, + 1405.0, + 1470.0, + 293.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1462.0, + 1403.0, + 1462.0, + 1403.0, + 1495.0, + 294.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1493.0, + 1404.0, + 1493.0, + 1404.0, + 1527.0, + 294.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1523.0, + 1405.0, + 1523.0, + 1405.0, + 1560.0, + 294.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1554.0, + 947.0, + 1554.0, + 947.0, + 1587.0, + 296.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 1554.0, + 1225.0, + 1554.0, + 1225.0, + 1587.0, + 986.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 782.0, + 1330.0, + 782.0, + 1330.0, + 819.0, + 293.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 782.0, + 1406.0, + 782.0, + 1406.0, + 819.0, + 1355.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 811.0, + 859.0, + 811.0, + 859.0, + 851.0, + 318.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 811.0, + 1271.0, + 811.0, + 1271.0, + 851.0, + 892.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 811.0, + 1407.0, + 811.0, + 1407.0, + 851.0, + 1312.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 846.0, + 1307.0, + 846.0, + 1307.0, + 880.0, + 296.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 270.0, + 433.0, + 270.0, + 433.0, + 304.0, + 304.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 270.0, + 549.0, + 270.0, + 549.0, + 304.0, + 456.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 270.0, + 595.0, + 270.0, + 595.0, + 304.0, + 585.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 270.0, + 824.0, + 270.0, + 824.0, + 304.0, + 614.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 270.0, + 1040.0, + 270.0, + 1040.0, + 304.0, + 855.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 270.0, + 1126.0, + 270.0, + 1126.0, + 304.0, + 1116.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 298.0, + 415.0, + 298.0, + 415.0, + 330.0, + 323.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 299.0, + 338.0, + 299.0, + 338.0, + 330.0, + 305.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 324.0, + 334.0, + 324.0, + 334.0, + 358.0, + 305.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 326.0, + 413.0, + 326.0, + 413.0, + 358.0, + 374.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 326.0, + 575.0, + 326.0, + 575.0, + 358.0, + 538.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 323.0, + 951.0, + 323.0, + 951.0, + 363.0, + 851.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 323.0, + 1369.0, + 323.0, + 1369.0, + 363.0, + 974.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 353.0, + 335.0, + 353.0, + 335.0, + 387.0, + 304.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 356.0, + 549.0, + 356.0, + 549.0, + 388.0, + 546.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 379.0, + 334.0, + 379.0, + 334.0, + 412.0, + 304.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 382.0, + 647.0, + 382.0, + 647.0, + 412.0, + 411.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 410.0, + 334.0, + 410.0, + 334.0, + 442.0, + 304.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 444.0, + 334.0, + 444.0, + 334.0, + 477.0, + 305.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 442.0, + 532.0, + 442.0, + 532.0, + 485.0, + 373.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 442.0, + 1074.0, + 442.0, + 1074.0, + 485.0, + 914.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 442.0, + 1110.0, + 442.0, + 1110.0, + 485.0, + 1104.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 472.0, + 335.0, + 472.0, + 335.0, + 504.0, + 304.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 471.0, + 429.0, + 471.0, + 429.0, + 508.0, + 374.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 471.0, + 659.0, + 471.0, + 659.0, + 508.0, + 482.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 471.0, + 757.0, + 471.0, + 757.0, + 508.0, + 696.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 471.0, + 830.0, + 471.0, + 830.0, + 508.0, + 786.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 471.0, + 865.0, + 471.0, + 865.0, + 508.0, + 860.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 501.0, + 334.0, + 501.0, + 334.0, + 534.0, + 303.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 499.0, + 410.0, + 499.0, + 410.0, + 535.0, + 375.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 499.0, + 543.0, + 499.0, + 543.0, + 535.0, + 481.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 533.0, + 336.0, + 533.0, + 336.0, + 566.0, + 295.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 531.0, + 474.0, + 531.0, + 474.0, + 568.0, + 375.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 531.0, + 579.0, + 531.0, + 579.0, + 568.0, + 517.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 531.0, + 710.0, + 531.0, + 710.0, + 568.0, + 668.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 531.0, + 875.0, + 531.0, + 875.0, + 568.0, + 753.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 562.0, + 338.0, + 562.0, + 338.0, + 594.0, + 294.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 565.0, + 417.0, + 565.0, + 417.0, + 593.0, + 375.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 565.0, + 673.0, + 565.0, + 673.0, + 593.0, + 457.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 593.0, + 338.0, + 593.0, + 338.0, + 626.0, + 294.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 590.0, + 785.0, + 590.0, + 785.0, + 631.0, + 375.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 626.0, + 338.0, + 626.0, + 338.0, + 659.0, + 294.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 621.0, + 476.0, + 621.0, + 476.0, + 669.0, + 371.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 659.0, + 339.0, + 659.0, + 339.0, + 692.0, + 294.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 660.0, + 615.0, + 660.0, + 615.0, + 695.0, + 375.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 660.0, + 708.0, + 660.0, + 708.0, + 695.0, + 646.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 660.0, + 1261.0, + 660.0, + 1261.0, + 695.0, + 791.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 686.0, + 397.0, + 686.0, + 397.0, + 726.0, + 293.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 686.0, + 543.0, + 686.0, + 543.0, + 726.0, + 428.0, + 726.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1208, + 1404, + 1208, + 1404, + 1696, + 298, + 1696 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 537, + 1404, + 537, + 1404, + 782, + 298, + 782 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1788, + 1403, + 1788, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 338, + 1403, + 338, + 1403, + 522, + 299, + 522 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 324, + 892, + 1378, + 892, + 1378, + 1085, + 324, + 1085 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 300, + 229, + 1401, + 229, + 1401, + 323, + 300, + 323 + ], + "score": 0.963 + }, + { + "category_id": 4, + "poly": [ + 289, + 1109, + 1398, + 1109, + 1398, + 1173, + 289, + 1173 + ], + "score": 0.932 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 814, + 76, + 814, + 104, + 300, + 104 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 300, + 826, + 557, + 826, + 557, + 859, + 300, + 859 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 300, + 1731, + 594, + 1731, + 594, + 1763, + 300, + 1763 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2111, + 842, + 2111 + ], + "score": 0.653 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2111, + 842, + 2111 + ], + "score": 0.196 + }, + { + "category_id": 13, + "poly": [ + 812, + 1483, + 920, + 1483, + 920, + 1513, + 812, + 1513 + ], + "score": 0.76, + "latex": "\\lambda = 0 . 9 7" + }, + { + "category_id": 13, + "poly": [ + 515, + 720, + 552, + 720, + 552, + 751, + 515, + 751 + ], + "score": 0.46, + "latex": "( \\lambda )" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 917.0, + 872.0, + 917.0, + 872.0, + 982.0, + 843.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 956.0, + 1161.0, + 956.0, + 1161.0, + 986.0, + 1140.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 978.0, + 863.0, + 978.0, + 863.0, + 1018.0, + 843.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 1055.0, + 422.0, + 1055.0, + 422.0, + 1087.0, + 381.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1054.0, + 576.0, + 1054.0, + 576.0, + 1087.0, + 528.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 1059.0, + 648.0, + 1059.0, + 648.0, + 1082.0, + 599.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 1056.0, + 722.0, + 1056.0, + 722.0, + 1087.0, + 679.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 1054.0, + 873.0, + 1054.0, + 873.0, + 1088.0, + 827.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1055.0, + 1020.0, + 1055.0, + 1020.0, + 1087.0, + 981.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 1055.0, + 1168.0, + 1055.0, + 1168.0, + 1087.0, + 1131.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 1056.0, + 1319.0, + 1056.0, + 1319.0, + 1088.0, + 1280.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1109.0, + 1404.0, + 1109.0, + 1404.0, + 1145.0, + 294.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1140.0, + 1312.0, + 1140.0, + 1312.0, + 1176.0, + 294.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 822.0, + 562.0, + 822.0, + 562.0, + 867.0, + 291.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1729.0, + 597.0, + 1729.0, + 597.0, + 1769.0, + 294.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1207.0, + 1404.0, + 1207.0, + 1404.0, + 1244.0, + 295.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1240.0, + 1406.0, + 1240.0, + 1406.0, + 1274.0, + 294.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1269.0, + 1404.0, + 1269.0, + 1404.0, + 1304.0, + 294.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1300.0, + 1405.0, + 1300.0, + 1405.0, + 1334.0, + 293.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1328.0, + 1405.0, + 1328.0, + 1405.0, + 1366.0, + 293.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1359.0, + 1405.0, + 1359.0, + 1405.0, + 1397.0, + 293.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1391.0, + 1404.0, + 1391.0, + 1404.0, + 1428.0, + 293.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1418.0, + 1404.0, + 1418.0, + 1404.0, + 1461.0, + 292.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1453.0, + 1404.0, + 1453.0, + 1404.0, + 1488.0, + 294.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1481.0, + 811.0, + 1481.0, + 811.0, + 1517.0, + 292.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1481.0, + 1406.0, + 1481.0, + 1406.0, + 1517.0, + 921.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1512.0, + 1404.0, + 1512.0, + 1404.0, + 1548.0, + 293.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1544.0, + 1404.0, + 1544.0, + 1404.0, + 1577.0, + 292.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1572.0, + 1404.0, + 1572.0, + 1404.0, + 1606.0, + 294.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1603.0, + 1404.0, + 1603.0, + 1404.0, + 1639.0, + 292.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1633.0, + 1405.0, + 1633.0, + 1405.0, + 1670.0, + 292.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1666.0, + 620.0, + 1666.0, + 620.0, + 1701.0, + 293.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 535.0, + 1406.0, + 535.0, + 1406.0, + 574.0, + 292.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 570.0, + 1405.0, + 570.0, + 1405.0, + 603.0, + 294.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 598.0, + 1404.0, + 598.0, + 1404.0, + 636.0, + 293.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 627.0, + 1405.0, + 627.0, + 1405.0, + 667.0, + 292.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 657.0, + 1405.0, + 657.0, + 1405.0, + 697.0, + 292.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 688.0, + 1405.0, + 688.0, + 1405.0, + 725.0, + 292.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 717.0, + 514.0, + 717.0, + 514.0, + 758.0, + 292.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 717.0, + 1405.0, + 717.0, + 1405.0, + 758.0, + 553.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 748.0, + 835.0, + 748.0, + 835.0, + 784.0, + 293.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1790.0, + 1404.0, + 1790.0, + 1404.0, + 1823.0, + 294.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1821.0, + 1403.0, + 1821.0, + 1403.0, + 1855.0, + 294.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1851.0, + 1404.0, + 1851.0, + 1404.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1880.0, + 1407.0, + 1880.0, + 1407.0, + 1918.0, + 292.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1912.0, + 1406.0, + 1912.0, + 1406.0, + 1946.0, + 292.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1979.0, + 292.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 1407.0, + 1972.0, + 1407.0, + 2010.0, + 293.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2004.0, + 818.0, + 2004.0, + 818.0, + 2037.0, + 296.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 332.0, + 1404.0, + 332.0, + 1404.0, + 377.0, + 292.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 367.0, + 1404.0, + 367.0, + 1404.0, + 403.0, + 293.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 397.0, + 1405.0, + 397.0, + 1405.0, + 435.0, + 293.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 426.0, + 1407.0, + 426.0, + 1407.0, + 468.0, + 292.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 459.0, + 1407.0, + 459.0, + 1407.0, + 498.0, + 293.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 492.0, + 763.0, + 492.0, + 763.0, + 524.0, + 297.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 230.0, + 1403.0, + 230.0, + 1403.0, + 263.0, + 294.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 260.0, + 1405.0, + 260.0, + 1405.0, + 294.0, + 295.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 289.0, + 847.0, + 289.0, + 847.0, + 328.0, + 295.0, + 328.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1728, + 1404, + 1728, + 1404, + 2034, + 297, + 2034 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 782, + 1404, + 782, + 1404, + 938, + 297, + 938 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 951, + 1403, + 951, + 1403, + 1046, + 299, + 1046 + ], + "score": 0.97 + }, + { + "category_id": 4, + "poly": [ + 297, + 1587, + 1403, + 1587, + 1403, + 1681, + 297, + 1681 + ], + "score": 0.918 + }, + { + "category_id": 4, + "poly": [ + 296, + 592, + 1403, + 592, + 1403, + 715, + 296, + 715 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 814, + 76, + 814, + 104, + 298, + 104 + ], + "score": 0.889 + }, + { + "category_id": 3, + "poly": [ + 909, + 229, + 1333, + 229, + 1333, + 518, + 909, + 518 + ], + "score": 0.851 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.795 + }, + { + "category_id": 4, + "poly": [ + 853, + 537, + 1391, + 537, + 1391, + 568, + 853, + 568 + ], + "score": 0.753 + }, + { + "category_id": 3, + "poly": [ + 908, + 1223, + 1335, + 1223, + 1335, + 1514, + 908, + 1514 + ], + "score": 0.747 + }, + { + "category_id": 3, + "poly": [ + 362, + 1221, + 789, + 1221, + 789, + 1513, + 362, + 1513 + ], + "score": 0.731 + }, + { + "category_id": 3, + "poly": [ + 364, + 228, + 787, + 228, + 787, + 523, + 364, + 523 + ], + "score": 0.651 + }, + { + "category_id": 4, + "poly": [ + 903, + 1533, + 1343, + 1533, + 1343, + 1563, + 903, + 1563 + ], + "score": 0.537 + }, + { + "category_id": 1, + "poly": [ + 347, + 1533, + 808, + 1533, + 808, + 1562, + 347, + 1562 + ], + "score": 0.52 + }, + { + "category_id": 0, + "poly": [ + 300, + 1115, + 793, + 1115, + 793, + 1145, + 300, + 1145 + ], + "score": 0.496 + }, + { + "category_id": 2, + "poly": [ + 300, + 1115, + 793, + 1115, + 793, + 1145, + 300, + 1145 + ], + "score": 0.274 + }, + { + "category_id": 4, + "poly": [ + 371, + 538, + 784, + 538, + 784, + 567, + 371, + 567 + ], + "score": 0.163 + }, + { + "category_id": 1, + "poly": [ + 371, + 538, + 784, + 538, + 784, + 567, + 371, + 567 + ], + "score": 0.118 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1587.0, + 1401.0, + 1587.0, + 1401.0, + 1624.0, + 295.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1618.0, + 1404.0, + 1618.0, + 1404.0, + 1655.0, + 292.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1645.0, + 1403.0, + 1645.0, + 1403.0, + 1685.0, + 294.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 591.0, + 1404.0, + 591.0, + 1404.0, + 628.0, + 293.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 622.0, + 1404.0, + 622.0, + 1404.0, + 657.0, + 293.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 653.0, + 1405.0, + 653.0, + 1405.0, + 688.0, + 294.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 685.0, + 682.0, + 685.0, + 682.0, + 717.0, + 296.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 229.0, + 982.0, + 229.0, + 982.0, + 254.0, + 938.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 263.0, + 982.0, + 263.0, + 982.0, + 287.0, + 938.0, + 287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 295.0, + 982.0, + 295.0, + 982.0, + 321.0, + 938.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 302.0, + 935.0, + 302.0, + 935.0, + 420.0, + 911.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 312.0, + 1312.0, + 312.0, + 1312.0, + 325.0, + 1277.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 316.0, + 1275.0, + 316.0, + 1275.0, + 326.0, + 1241.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 330.0, + 982.0, + 330.0, + 982.0, + 354.0, + 938.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 363.0, + 982.0, + 363.0, + 982.0, + 388.0, + 938.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 372.0, + 1164.0, + 372.0, + 1164.0, + 383.0, + 1133.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 368.0, + 1313.0, + 368.0, + 1313.0, + 379.0, + 1265.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 385.0, + 1268.0, + 385.0, + 1268.0, + 404.0, + 1152.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 400.0, + 982.0, + 400.0, + 982.0, + 418.0, + 964.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 400.0, + 1188.0, + 400.0, + 1188.0, + 414.0, + 1155.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 398.0, + 1268.0, + 398.0, + 1268.0, + 418.0, + 1189.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 415.0, + 1177.0, + 415.0, + 1177.0, + 425.0, + 1157.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 411.0, + 1268.0, + 411.0, + 1268.0, + 431.0, + 1189.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 429.0, + 984.0, + 429.0, + 984.0, + 457.0, + 929.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 422.0, + 1307.0, + 422.0, + 1307.0, + 460.0, + 1188.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 449.0, + 1307.0, + 449.0, + 1307.0, + 472.0, + 1188.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 462.0, + 984.0, + 462.0, + 984.0, + 489.0, + 929.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 480.0, + 990.0, + 480.0, + 990.0, + 496.0, + 974.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 477.0, + 1334.0, + 477.0, + 1334.0, + 499.0, + 998.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 495.0, + 1185.0, + 495.0, + 1185.0, + 518.0, + 1110.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.75, + 269.5, + 1172.75, + 269.5, + 1172.75, + 280.5, + 1146.75, + 280.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.25, + 283.5, + 1059.25, + 283.5, + 1059.25, + 299.5, + 1034.25, + 299.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.75, + 288.5, + 1241.75, + 288.5, + 1241.75, + 298.5, + 1199.75, + 298.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 278.5, + 1316.0, + 278.5, + 1316.0, + 295.0, + 1237.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 296.5, + 1206.0, + 296.5, + 1206.0, + 310.0, + 1171.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 389.5, + 1075.0, + 389.5, + 1075.0, + 417.5, + 973.0, + 417.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2117.0, + 838.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 534.0, + 1394.0, + 534.0, + 1394.0, + 572.0, + 852.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1224.0, + 973.0, + 1224.0, + 973.0, + 1249.0, + 928.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1258.0, + 973.0, + 1258.0, + 973.0, + 1283.0, + 929.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 1293.0, + 971.0, + 1293.0, + 971.0, + 1315.0, + 926.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 1294.0, + 938.0, + 1294.0, + 938.0, + 1416.0, + 908.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1323.0, + 974.0, + 1323.0, + 974.0, + 1351.0, + 928.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 1357.0, + 973.0, + 1357.0, + 973.0, + 1383.0, + 930.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1390.0, + 974.0, + 1390.0, + 974.0, + 1418.0, + 929.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 1411.0, + 1311.0, + 1411.0, + 1311.0, + 1425.0, + 1232.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 1425.0, + 974.0, + 1425.0, + 974.0, + 1452.0, + 936.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 1434.0, + 1265.0, + 1434.0, + 1265.0, + 1453.0, + 1150.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1446.0, + 1304.0, + 1446.0, + 1304.0, + 1466.0, + 1182.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1461.0, + 976.0, + 1461.0, + 976.0, + 1483.0, + 955.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 1474.0, + 980.0, + 1474.0, + 980.0, + 1494.0, + 965.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 1471.0, + 1062.0, + 1471.0, + 1062.0, + 1497.0, + 1018.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 1472.0, + 1201.0, + 1472.0, + 1201.0, + 1496.0, + 1081.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 1472.0, + 1269.0, + 1472.0, + 1269.0, + 1496.0, + 1213.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 1472.0, + 1335.0, + 1472.0, + 1335.0, + 1496.0, + 1281.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 1491.0, + 1180.0, + 1491.0, + 1180.0, + 1514.0, + 1103.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 1262.5, + 1334.0, + 1262.5, + 1334.0, + 1339.5, + 1095.0, + 1339.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 1419.5, + 1232.0, + 1419.5, + 1232.0, + 1449.0, + 979.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1225.0, + 436.0, + 1225.0, + 436.0, + 1250.0, + 391.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1263.0, + 618.0, + 1263.0, + 618.0, + 1271.0, + 606.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1267.0, + 437.0, + 1267.0, + 437.0, + 1292.0, + 391.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 1272.0, + 509.0, + 1272.0, + 509.0, + 1284.0, + 491.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1295.0, + 392.0, + 1295.0, + 392.0, + 1417.0, + 362.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1310.0, + 437.0, + 1310.0, + 437.0, + 1334.0, + 391.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1353.0, + 436.0, + 1353.0, + 436.0, + 1377.0, + 391.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 1355.0, + 619.0, + 1355.0, + 619.0, + 1369.0, + 555.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 1364.0, + 463.0, + 1364.0, + 463.0, + 1374.0, + 455.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 1376.0, + 461.0, + 1376.0, + 461.0, + 1383.0, + 452.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 1374.0, + 554.0, + 1374.0, + 554.0, + 1382.0, + 545.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 1386.0, + 638.0, + 1386.0, + 638.0, + 1394.0, + 614.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1380.0, + 711.0, + 1380.0, + 711.0, + 1400.0, + 642.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1401.0, + 433.0, + 1401.0, + 433.0, + 1415.0, + 420.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1394.0, + 712.0, + 1394.0, + 712.0, + 1414.0, + 642.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 1407.0, + 715.0, + 1407.0, + 715.0, + 1426.0, + 643.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 1423.0, + 640.0, + 1423.0, + 640.0, + 1436.0, + 608.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 1419.0, + 722.0, + 1419.0, + 722.0, + 1439.0, + 643.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 1436.0, + 438.0, + 1436.0, + 438.0, + 1464.0, + 382.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 1432.0, + 761.0, + 1432.0, + 761.0, + 1466.0, + 640.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 1474.0, + 444.0, + 1474.0, + 444.0, + 1494.0, + 427.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 1472.0, + 788.0, + 1472.0, + 788.0, + 1496.0, + 472.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 1488.0, + 642.0, + 1488.0, + 642.0, + 1517.0, + 562.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1318.5, + 766.0, + 1318.5, + 766.0, + 1338.0, + 597.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 1335.0, + 606.0, + 1335.0, + 606.0, + 1349.0, + 572.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.25, + 1347.5, + 653.25, + 1347.5, + 653.25, + 1360.5, + 625.25, + 1360.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 1358.5, + 554.0, + 1358.5, + 554.0, + 1377.5, + 516.0, + 1377.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.25, + 1377.5, + 529.25, + 1377.5, + 529.25, + 1405.5, + 450.25, + 1405.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.25, + 1411.5, + 628.25, + 1411.5, + 628.25, + 1422.0, + 610.25, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 229.0, + 436.0, + 229.0, + 436.0, + 254.0, + 392.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 262.0, + 436.0, + 262.0, + 436.0, + 288.0, + 392.0, + 288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 296.0, + 436.0, + 296.0, + 436.0, + 321.0, + 392.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 300.0, + 391.0, + 300.0, + 391.0, + 420.0, + 364.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 331.0, + 435.0, + 331.0, + 435.0, + 353.0, + 393.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 363.0, + 436.0, + 363.0, + 436.0, + 388.0, + 393.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 388.0, + 483.0, + 388.0, + 483.0, + 397.0, + 475.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 396.0, + 436.0, + 396.0, + 436.0, + 421.0, + 393.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 419.0, + 466.0, + 419.0, + 466.0, + 431.0, + 458.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 411.0, + 718.0, + 411.0, + 718.0, + 429.0, + 606.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 424.0, + 745.0, + 424.0, + 745.0, + 444.0, + 638.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 433.0, + 435.0, + 433.0, + 435.0, + 453.0, + 418.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 436.0, + 757.0, + 436.0, + 757.0, + 459.0, + 638.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 449.0, + 757.0, + 449.0, + 757.0, + 471.0, + 638.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 464.0, + 437.0, + 464.0, + 437.0, + 488.0, + 384.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 480.0, + 442.0, + 480.0, + 442.0, + 496.0, + 429.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 475.0, + 624.0, + 475.0, + 624.0, + 501.0, + 467.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 475.0, + 788.0, + 475.0, + 788.0, + 501.0, + 630.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 495.0, + 638.0, + 495.0, + 638.0, + 518.0, + 562.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1530.0, + 1348.0, + 1530.0, + 1348.0, + 1565.0, + 897.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1113.0, + 797.0, + 1113.0, + 797.0, + 1148.0, + 295.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1113.0, + 797.0, + 1113.0, + 797.0, + 1148.0, + 295.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 536.0, + 786.0, + 536.0, + 786.0, + 572.0, + 368.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1726.0, + 1404.0, + 1726.0, + 1404.0, + 1763.0, + 292.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1758.0, + 1406.0, + 1758.0, + 1406.0, + 1797.0, + 292.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1787.0, + 1405.0, + 1787.0, + 1405.0, + 1825.0, + 292.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1819.0, + 1406.0, + 1819.0, + 1406.0, + 1858.0, + 294.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1852.0, + 1405.0, + 1852.0, + 1405.0, + 1884.0, + 296.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1882.0, + 1404.0, + 1882.0, + 1404.0, + 1918.0, + 295.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1404.0, + 1910.0, + 1404.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1980.0, + 294.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1974.0, + 1404.0, + 1974.0, + 1404.0, + 2005.0, + 295.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 1103.0, + 2003.0, + 1103.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 784.0, + 1404.0, + 784.0, + 1404.0, + 817.0, + 296.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 812.0, + 1405.0, + 812.0, + 1405.0, + 851.0, + 294.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 843.0, + 1406.0, + 843.0, + 1406.0, + 883.0, + 291.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 876.0, + 1404.0, + 876.0, + 1404.0, + 909.0, + 295.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 905.0, + 1363.0, + 905.0, + 1363.0, + 943.0, + 295.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 951.0, + 1404.0, + 951.0, + 1404.0, + 989.0, + 294.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 983.0, + 1405.0, + 983.0, + 1405.0, + 1017.0, + 295.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1013.0, + 1261.0, + 1013.0, + 1261.0, + 1051.0, + 294.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1530.0, + 812.0, + 1530.0, + 812.0, + 1565.0, + 341.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 536.0, + 786.0, + 536.0, + 786.0, + 572.0, + 368.0, + 572.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 297, + 674, + 1420, + 674, + 1420, + 938, + 297, + 938 + ], + "score": 0.98, + "html": "
TR-c-Q-PropTRPODDPG
DomainThresholdMaxReturn.EpisodesMaxReturnEpsisodesMaxReturnEpisodes
Ant350035344975423913825957N/A
HalfCheetah47004811207854734263707490600
Hopper200029575945248657152604965
Humanoid2500>349214750918>30000552N/A
Reacher-7-6.02060-6.72840-6.61800
Swimmer9010320451103025150500
Walker30004030368535671887536262125
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1176, + 1404, + 1176, + 1404, + 1452, + 298, + 1452 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 289, + 1404, + 289, + 1404, + 505, + 297, + 505 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 520, + 1404, + 520, + 1404, + 643, + 299, + 643 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1549, + 1403, + 1549, + 1403, + 1672, + 299, + 1672 + ], + "score": 0.969 + }, + { + "category_id": 0, + "poly": [ + 301, + 1103, + 799, + 1103, + 799, + 1138, + 301, + 1138 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.88 + }, + { + "category_id": 0, + "poly": [ + 301, + 1495, + 557, + 1495, + 557, + 1523, + 301, + 1523 + ], + "score": 0.863 + }, + { + "category_id": 0, + "poly": [ + 301, + 230, + 740, + 230, + 740, + 260, + 301, + 260 + ], + "score": 0.827 + }, + { + "category_id": 1, + "poly": [ + 295, + 1781, + 1399, + 1781, + 1399, + 1843, + 295, + 1843 + ], + "score": 0.823 + }, + { + "category_id": 0, + "poly": [ + 300, + 1726, + 488, + 1726, + 488, + 1759, + 300, + 1759 + ], + "score": 0.804 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.763 + }, + { + "category_id": 1, + "poly": [ + 298, + 957, + 1406, + 957, + 1406, + 1049, + 298, + 1049 + ], + "score": 0.752 + }, + { + "category_id": 1, + "poly": [ + 294, + 1876, + 1401, + 1876, + 1401, + 1939, + 294, + 1939 + ], + "score": 0.706 + }, + { + "category_id": 1, + "poly": [ + 300, + 1971, + 1400, + 1971, + 1400, + 2034, + 300, + 2034 + ], + "score": 0.676 + }, + { + "category_id": 6, + "poly": [ + 298, + 957, + 1406, + 957, + 1406, + 1049, + 298, + 1049 + ], + "score": 0.154 + }, + { + "category_id": 13, + "poly": [ + 297, + 988, + 344, + 988, + 344, + 1017, + 297, + 1017 + ], + "score": 0.34, + "latex": "3 0 \\mathrm { k }" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1097.0, + 806.0, + 1097.0, + 806.0, + 1147.0, + 291.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1497.0, + 558.0, + 1497.0, + 558.0, + 1526.0, + 299.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 744.0, + 230.0, + 744.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1725.0, + 491.0, + 1725.0, + 491.0, + 1763.0, + 295.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 956.0, + 1407.0, + 956.0, + 1407.0, + 993.0, + 293.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 986.0, + 1404.0, + 986.0, + 1404.0, + 1022.0, + 345.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1017.0, + 1323.0, + 1017.0, + 1323.0, + 1051.0, + 295.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1176.0, + 1404.0, + 1176.0, + 1404.0, + 1213.0, + 295.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1208.0, + 1406.0, + 1208.0, + 1406.0, + 1245.0, + 294.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1238.0, + 1404.0, + 1238.0, + 1404.0, + 1275.0, + 293.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1270.0, + 1404.0, + 1270.0, + 1404.0, + 1302.0, + 296.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1301.0, + 1404.0, + 1301.0, + 1404.0, + 1334.0, + 294.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1327.0, + 1404.0, + 1327.0, + 1404.0, + 1366.0, + 293.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1361.0, + 1402.0, + 1361.0, + 1402.0, + 1394.0, + 296.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1390.0, + 1405.0, + 1390.0, + 1405.0, + 1426.0, + 294.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1422.0, + 462.0, + 1422.0, + 462.0, + 1456.0, + 294.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 287.0, + 1405.0, + 287.0, + 1405.0, + 328.0, + 291.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 323.0, + 1404.0, + 323.0, + 1404.0, + 354.0, + 295.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 351.0, + 1404.0, + 351.0, + 1404.0, + 386.0, + 294.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 382.0, + 1404.0, + 382.0, + 1404.0, + 417.0, + 294.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 413.0, + 1404.0, + 413.0, + 1404.0, + 448.0, + 294.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 441.0, + 1404.0, + 441.0, + 1404.0, + 478.0, + 294.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 472.0, + 1387.0, + 472.0, + 1387.0, + 509.0, + 294.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 518.0, + 1403.0, + 518.0, + 1403.0, + 555.0, + 294.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 549.0, + 1405.0, + 549.0, + 1405.0, + 586.0, + 293.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 582.0, + 1404.0, + 582.0, + 1404.0, + 618.0, + 294.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 613.0, + 887.0, + 613.0, + 887.0, + 645.0, + 294.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1549.0, + 1407.0, + 1549.0, + 1407.0, + 1585.0, + 294.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1581.0, + 1404.0, + 1581.0, + 1404.0, + 1614.0, + 294.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1611.0, + 1403.0, + 1611.0, + 1403.0, + 1647.0, + 294.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1641.0, + 821.0, + 1641.0, + 821.0, + 1677.0, + 294.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1778.0, + 1405.0, + 1778.0, + 1405.0, + 1817.0, + 294.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1811.0, + 611.0, + 1811.0, + 611.0, + 1844.0, + 321.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 956.0, + 1407.0, + 956.0, + 1407.0, + 993.0, + 293.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 986.0, + 1404.0, + 986.0, + 1404.0, + 1022.0, + 345.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1017.0, + 1323.0, + 1017.0, + 1323.0, + 1051.0, + 295.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1878.0, + 1402.0, + 1878.0, + 1402.0, + 1910.0, + 298.0, + 1910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1906.0, + 1399.0, + 1906.0, + 1399.0, + 1941.0, + 323.0, + 1941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2008.0, + 296.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 2001.0, + 1154.0, + 2001.0, + 1154.0, + 2037.0, + 324.0, + 2037.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 294, + 977, + 1401, + 977, + 1401, + 1041, + 294, + 1041 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.883 + }, + { + "category_id": 1, + "poly": [ + 299, + 228, + 1400, + 228, + 1400, + 323, + 299, + 323 + ], + "score": 0.878 + }, + { + "category_id": 1, + "poly": [ + 296, + 1059, + 964, + 1059, + 964, + 1094, + 296, + 1094 + ], + "score": 0.866 + }, + { + "category_id": 1, + "poly": [ + 301, + 536, + 1405, + 536, + 1405, + 630, + 301, + 630 + ], + "score": 0.863 + }, + { + "category_id": 1, + "poly": [ + 294, + 453, + 1402, + 453, + 1402, + 519, + 294, + 519 + ], + "score": 0.859 + }, + { + "category_id": 1, + "poly": [ + 293, + 1111, + 1402, + 1111, + 1402, + 1175, + 293, + 1175 + ], + "score": 0.858 + }, + { + "category_id": 1, + "poly": [ + 299, + 924, + 1244, + 924, + 1244, + 960, + 299, + 960 + ], + "score": 0.846 + }, + { + "category_id": 1, + "poly": [ + 298, + 342, + 1399, + 342, + 1399, + 435, + 298, + 435 + ], + "score": 0.839 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 866, + 2088, + 866, + 2113, + 836, + 2113 + ], + "score": 0.836 + }, + { + "category_id": 1, + "poly": [ + 299, + 842, + 1399, + 842, + 1399, + 907, + 299, + 907 + ], + "score": 0.831 + }, + { + "category_id": 1, + "poly": [ + 288, + 1417, + 1400, + 1417, + 1400, + 1483, + 288, + 1483 + ], + "score": 0.831 + }, + { + "category_id": 1, + "poly": [ + 297, + 1304, + 1403, + 1304, + 1403, + 1400, + 297, + 1400 + ], + "score": 0.822 + }, + { + "category_id": 1, + "poly": [ + 299, + 1971, + 1400, + 1971, + 1400, + 2034, + 299, + 2034 + ], + "score": 0.819 + }, + { + "category_id": 1, + "poly": [ + 296, + 760, + 1399, + 760, + 1399, + 826, + 296, + 826 + ], + "score": 0.817 + }, + { + "category_id": 1, + "poly": [ + 298, + 1611, + 1404, + 1611, + 1404, + 1708, + 298, + 1708 + ], + "score": 0.813 + }, + { + "category_id": 1, + "poly": [ + 297, + 648, + 1405, + 648, + 1405, + 742, + 297, + 742 + ], + "score": 0.802 + }, + { + "category_id": 1, + "poly": [ + 298, + 1499, + 1403, + 1499, + 1403, + 1596, + 298, + 1596 + ], + "score": 0.802 + }, + { + "category_id": 1, + "poly": [ + 297, + 1724, + 1400, + 1724, + 1400, + 1790, + 297, + 1790 + ], + "score": 0.795 + }, + { + "category_id": 1, + "poly": [ + 298, + 1889, + 1399, + 1889, + 1399, + 1954, + 298, + 1954 + ], + "score": 0.79 + }, + { + "category_id": 1, + "poly": [ + 298, + 1192, + 1402, + 1192, + 1402, + 1289, + 298, + 1289 + ], + "score": 0.785 + }, + { + "category_id": 1, + "poly": [ + 299, + 1806, + 1398, + 1806, + 1398, + 1872, + 299, + 1872 + ], + "score": 0.784 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 972.0, + 1406.0, + 972.0, + 1406.0, + 1015.0, + 293.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1007.0, + 600.0, + 1007.0, + 600.0, + 1041.0, + 321.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 224.0, + 1404.0, + 224.0, + 1404.0, + 268.0, + 293.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 256.0, + 1405.0, + 256.0, + 1405.0, + 301.0, + 320.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 291.0, + 505.0, + 291.0, + 505.0, + 323.0, + 323.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1058.0, + 965.0, + 1058.0, + 965.0, + 1098.0, + 296.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 533.0, + 1406.0, + 533.0, + 1406.0, + 573.0, + 294.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 568.0, + 1404.0, + 568.0, + 1404.0, + 602.0, + 322.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 595.0, + 398.0, + 595.0, + 398.0, + 632.0, + 319.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 452.0, + 1404.0, + 452.0, + 1404.0, + 490.0, + 294.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 487.0, + 694.0, + 487.0, + 694.0, + 518.0, + 319.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1108.0, + 1407.0, + 1108.0, + 1407.0, + 1149.0, + 293.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1140.0, + 699.0, + 1140.0, + 699.0, + 1176.0, + 322.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 923.0, + 1246.0, + 923.0, + 1246.0, + 962.0, + 294.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 338.0, + 1404.0, + 338.0, + 1404.0, + 380.0, + 294.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 370.0, + 1404.0, + 370.0, + 1404.0, + 410.0, + 321.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 404.0, + 491.0, + 404.0, + 491.0, + 436.0, + 323.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 834.0, + 1405.0, + 834.0, + 1405.0, + 886.0, + 290.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 874.0, + 531.0, + 874.0, + 531.0, + 908.0, + 324.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1418.0, + 1403.0, + 1418.0, + 1403.0, + 1454.0, + 295.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1450.0, + 971.0, + 1450.0, + 971.0, + 1482.0, + 321.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1302.0, + 1404.0, + 1302.0, + 1404.0, + 1343.0, + 292.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1337.0, + 1404.0, + 1337.0, + 1404.0, + 1372.0, + 324.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1367.0, + 805.0, + 1367.0, + 805.0, + 1401.0, + 324.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1968.0, + 1404.0, + 1968.0, + 1404.0, + 2008.0, + 292.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 2000.0, + 395.0, + 2000.0, + 395.0, + 2036.0, + 320.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 761.0, + 1403.0, + 761.0, + 1403.0, + 797.0, + 296.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 794.0, + 1365.0, + 794.0, + 1365.0, + 827.0, + 324.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1610.0, + 1403.0, + 1610.0, + 1403.0, + 1650.0, + 296.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1642.0, + 1404.0, + 1642.0, + 1404.0, + 1680.0, + 322.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1673.0, + 1164.0, + 1673.0, + 1164.0, + 1711.0, + 322.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 648.0, + 1405.0, + 648.0, + 1405.0, + 687.0, + 294.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 678.0, + 1403.0, + 678.0, + 1403.0, + 716.0, + 321.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 705.0, + 410.0, + 705.0, + 410.0, + 748.0, + 320.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1499.0, + 1404.0, + 1499.0, + 1404.0, + 1538.0, + 296.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1531.0, + 1405.0, + 1531.0, + 1405.0, + 1566.0, + 323.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1562.0, + 1225.0, + 1562.0, + 1225.0, + 1597.0, + 323.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1723.0, + 1401.0, + 1723.0, + 1401.0, + 1760.0, + 295.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1757.0, + 1123.0, + 1757.0, + 1123.0, + 1791.0, + 321.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1887.0, + 1403.0, + 1887.0, + 1403.0, + 1927.0, + 293.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1920.0, + 1085.0, + 1920.0, + 1085.0, + 1954.0, + 323.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1189.0, + 1406.0, + 1189.0, + 1406.0, + 1232.0, + 293.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1224.0, + 1404.0, + 1224.0, + 1404.0, + 1259.0, + 323.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1255.0, + 1019.0, + 1255.0, + 1019.0, + 1290.0, + 325.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1807.0, + 1402.0, + 1807.0, + 1402.0, + 1843.0, + 294.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1838.0, + 1025.0, + 1838.0, + 1025.0, + 1871.0, + 323.0, + 1871.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 978, + 1404, + 978, + 1404, + 1074, + 298, + 1074 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 297, + 1098, + 1404, + 1098, + 1404, + 1224, + 297, + 1224 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 298, + 737, + 1404, + 737, + 1404, + 833, + 298, + 833 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 553, + 1979, + 1145, + 1979, + 1145, + 2045, + 553, + 2045 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 298, + 617, + 1403, + 617, + 1403, + 711, + 298, + 711 + ], + "score": 0.942 + }, + { + "category_id": 1, + "poly": [ + 297, + 1897, + 1402, + 1897, + 1402, + 1960, + 297, + 1960 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 297, + 498, + 1404, + 498, + 1404, + 592, + 297, + 592 + ], + "score": 0.932 + }, + { + "category_id": 1, + "poly": [ + 298, + 1397, + 1404, + 1397, + 1404, + 1492, + 298, + 1492 + ], + "score": 0.929 + }, + { + "category_id": 1, + "poly": [ + 299, + 1699, + 1401, + 1699, + 1401, + 1761, + 299, + 1761 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 298, + 377, + 1406, + 377, + 1406, + 470, + 298, + 470 + ], + "score": 0.927 + }, + { + "category_id": 1, + "poly": [ + 296, + 859, + 1398, + 859, + 1398, + 951, + 296, + 951 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 301, + 1577, + 1401, + 1577, + 1401, + 1673, + 301, + 1673 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 293, + 229, + 1402, + 229, + 1402, + 294, + 293, + 294 + ], + "score": 0.917 + }, + { + "category_id": 1, + "poly": [ + 295, + 1249, + 1403, + 1249, + 1403, + 1314, + 295, + 1314 + ], + "score": 0.899 + }, + { + "category_id": 1, + "poly": [ + 297, + 318, + 995, + 318, + 995, + 353, + 297, + 353 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 300, + 1825, + 835, + 1825, + 835, + 1863, + 300, + 1863 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.886 + }, + { + "category_id": 1, + "poly": [ + 298, + 1338, + 1237, + 1338, + 1237, + 1373, + 298, + 1373 + ], + "score": 0.885 + }, + { + "category_id": 1, + "poly": [ + 295, + 1518, + 1388, + 1518, + 1388, + 1554, + 295, + 1554 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.859 + }, + { + "category_id": 14, + "poly": [ + 555, + 1978, + 1144, + 1978, + 1144, + 2045, + 555, + 2045 + ], + "score": 0.91, + "latex": "\\mathbb { E } _ { p _ { \\theta } ( x ) } [ \\nabla _ { \\theta } \\log p _ { \\theta } ( x ) ] = \\int _ { x } \\nabla _ { \\theta } p _ { \\theta } ( x ) = \\nabla _ { \\theta } \\int _ { x } p ( x ) = 0" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1826.0, + 839.0, + 1826.0, + 839.0, + 1867.0, + 295.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 976.0, + 1404.0, + 976.0, + 1404.0, + 1016.0, + 293.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1010.0, + 1404.0, + 1010.0, + 1404.0, + 1044.0, + 325.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1042.0, + 1101.0, + 1042.0, + 1101.0, + 1075.0, + 323.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1097.0, + 1404.0, + 1097.0, + 1404.0, + 1135.0, + 294.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1128.0, + 1405.0, + 1128.0, + 1405.0, + 1168.0, + 321.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1161.0, + 1404.0, + 1161.0, + 1404.0, + 1196.0, + 322.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1192.0, + 889.0, + 1192.0, + 889.0, + 1225.0, + 322.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 735.0, + 1405.0, + 735.0, + 1405.0, + 776.0, + 293.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 767.0, + 1405.0, + 767.0, + 1405.0, + 808.0, + 320.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 801.0, + 1371.0, + 801.0, + 1371.0, + 835.0, + 325.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 619.0, + 1402.0, + 619.0, + 1402.0, + 653.0, + 296.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 648.0, + 1407.0, + 648.0, + 1407.0, + 687.0, + 320.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 674.0, + 399.0, + 674.0, + 399.0, + 715.0, + 319.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1892.0, + 1406.0, + 1892.0, + 1406.0, + 1936.0, + 294.0, + 1936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1929.0, + 917.0, + 1929.0, + 917.0, + 1963.0, + 293.0, + 1963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 498.0, + 1404.0, + 498.0, + 1404.0, + 535.0, + 294.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 530.0, + 1402.0, + 530.0, + 1402.0, + 564.0, + 322.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 562.0, + 872.0, + 562.0, + 872.0, + 593.0, + 322.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1396.0, + 1404.0, + 1396.0, + 1404.0, + 1435.0, + 293.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1429.0, + 1404.0, + 1429.0, + 1404.0, + 1465.0, + 321.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1460.0, + 469.0, + 1460.0, + 469.0, + 1492.0, + 323.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1698.0, + 1403.0, + 1698.0, + 1403.0, + 1734.0, + 295.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1729.0, + 899.0, + 1729.0, + 899.0, + 1764.0, + 324.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 376.0, + 1404.0, + 376.0, + 1404.0, + 415.0, + 292.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 410.0, + 1404.0, + 410.0, + 1404.0, + 444.0, + 322.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 435.0, + 400.0, + 435.0, + 400.0, + 476.0, + 320.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 857.0, + 1403.0, + 857.0, + 1403.0, + 895.0, + 294.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 888.0, + 1403.0, + 888.0, + 1403.0, + 928.0, + 321.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 921.0, + 504.0, + 921.0, + 504.0, + 952.0, + 324.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1573.0, + 1406.0, + 1573.0, + 1406.0, + 1619.0, + 291.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1610.0, + 1404.0, + 1610.0, + 1404.0, + 1645.0, + 322.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1642.0, + 844.0, + 1642.0, + 844.0, + 1673.0, + 325.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 228.0, + 1405.0, + 228.0, + 1405.0, + 268.0, + 293.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 260.0, + 790.0, + 260.0, + 790.0, + 296.0, + 324.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1249.0, + 1404.0, + 1249.0, + 1404.0, + 1285.0, + 295.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1280.0, + 1369.0, + 1280.0, + 1369.0, + 1315.0, + 322.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 318.0, + 997.0, + 318.0, + 997.0, + 356.0, + 296.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1338.0, + 1240.0, + 1338.0, + 1240.0, + 1377.0, + 295.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1518.0, + 1391.0, + 1518.0, + 1391.0, + 1557.0, + 296.0, + 1557.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 968, + 1406, + 968, + 1406, + 1130, + 296, + 1130 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 296, + 1771, + 1405, + 1771, + 1405, + 1959, + 296, + 1959 + ], + "score": 0.98 + }, + { + "category_id": 8, + "poly": [ + 536, + 1280, + 1165, + 1280, + 1165, + 1404, + 536, + 1404 + ], + "score": 0.966 + }, + { + "category_id": 8, + "poly": [ + 514, + 1677, + 1184, + 1677, + 1184, + 1760, + 514, + 1760 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 297, + 227, + 1405, + 227, + 1405, + 325, + 297, + 325 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 298, + 1597, + 1401, + 1597, + 1401, + 1662, + 298, + 1662 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 295, + 1203, + 1405, + 1203, + 1405, + 1267, + 295, + 1267 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 300, + 1419, + 1397, + 1419, + 1397, + 1486, + 300, + 1486 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 298, + 1970, + 1400, + 1970, + 1400, + 2036, + 298, + 2036 + ], + "score": 0.943 + }, + { + "category_id": 8, + "poly": [ + 503, + 1147, + 1193, + 1147, + 1193, + 1188, + 503, + 1188 + ], + "score": 0.921 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 815, + 74, + 815, + 105, + 298, + 105 + ], + "score": 0.916 + }, + { + "category_id": 0, + "poly": [ + 298, + 1527, + 1080, + 1527, + 1080, + 1566, + 298, + 1566 + ], + "score": 0.91 + }, + { + "category_id": 0, + "poly": [ + 299, + 860, + 1229, + 860, + 1229, + 937, + 299, + 937 + ], + "score": 0.904 + }, + { + "category_id": 8, + "poly": [ + 316, + 340, + 1338, + 340, + 1338, + 736, + 316, + 736 + ], + "score": 0.903 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1151, + 1400, + 1151, + 1400, + 1181, + 1352, + 1181 + ], + "score": 0.901 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1701, + 1400, + 1701, + 1400, + 1733, + 1352, + 1733 + ], + "score": 0.898 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1327, + 1400, + 1327, + 1400, + 1358, + 1351, + 1358 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1351, + 567, + 1402, + 567, + 1402, + 598, + 1351, + 598 + ], + "score": 0.874 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.866 + }, + { + "category_id": 8, + "poly": [ + 313, + 742, + 1337, + 742, + 1337, + 829, + 313, + 829 + ], + "score": 0.856 + }, + { + "category_id": 14, + "poly": [ + 315, + 340, + 1332, + 340, + 1332, + 835, + 315, + 835 + ], + "score": 0.95, + "latex": "\\begin{array} { r l } { \\nabla _ { \\theta } J ( \\theta ) = \\mathrm { E } _ { \\mathrm { g } , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { 1 } | s _ { 1 } \\big ) \\big ( \\hat { Q } ( s _ { t } , a _ { t } ) - \\hat { J } ( s _ { t } , a _ { t } ) \\big ) \\big ] + \\mathrm { E } _ { \\rho , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { t } | s _ { t } \\big ) \\hat { J } ( s _ { t } , a _ { t } ) \\big ] } \\\\ { g ( \\theta ) = \\mathrm { E } _ { \\mathrm { g } , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { 1 } | s _ { 1 } \\big ) \\hat { J } ( s _ { t } , a _ { t } ) \\big ] } \\\\ & { = \\mathrm { E } _ { \\rho , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { 1 } | s _ { 1 } \\big ) \\big ( \\hat { J } ( s _ { t } , a _ { t } ) + \\nabla _ { a _ { t } } \\big ) \\big ( s _ { t } , a _ { t } \\big ) \\big ] } \\\\ & { = \\mathrm { E } _ { \\rho , \\pi } \\big [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } \\big ( \\alpha _ { 1 } | s _ { 1 } \\big ) \\mathrm { e } _ { n } f \\big ( s _ { t } , a _ { t } \\big ) \\big ] = - \\alpha _ { 1 } \\alpha _ { 1 } \\alpha _ { 1 } } \\\\ & { = \\mathrm { E } _ { \\rho , \\pi } \\bigg [ \\int _ { a _ { t } } \\mathrm { e } _ { g } \\pi _ { \\theta } \\big ( a _ { t } | s _ { t } \\big ) \\mathrm { e } _ { n } f \\big ( s _ { t } , a _ { t } \\big ) \\big ] \\omega - \\alpha _ { 2 } \\alpha _ { 2 } \\alpha _ { 2 } \\bigg ] } \\\\ & { = \\mathrm { E } _ { \\rho , \\pi } \\bigg [ \\sum _ { \\theta } \\mathrm { e } _ { g } \\pi _ { \\theta } \\big ( a _ { t } \\big ) \\big ( \\frac { \\kappa } { \\alpha _ { 1 } } \\big ) \\mathrm { e } _ { n } f \\big ( \\alpha _ { 1 } \\big ) \\omega _ { 1 } \\omega _ { 2 } \\bigg ] } \\\\ & = \\mathrm { E } _ { \\rho , \\pi } \\bigg [ \\mathrm { v } _ { \\alpha } f \\big ( s _ { t } , a \\big ) \\big | \\alpha - \\alpha _ { 1 } \\int _ a _ { t } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 366, + 291, + 553, + 291, + 553, + 325, + 366, + 325 + ], + "score": 0.94, + "latex": "\\mu _ { \\theta } ( \\pmb { s } _ { t } ) \\dot { = } \\mathbb { E } _ { \\pi } [ \\bar { \\mathbf { a } } _ { t } ]" + }, + { + "category_id": 13, + "poly": [ + 541, + 230, + 640, + 230, + 640, + 264, + 541, + 264 + ], + "score": 0.93, + "latex": "f ( \\mathbf { \\boldsymbol { s } } _ { t } , \\mathbf { \\boldsymbol { a } } _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 936, + 1453, + 1022, + 1453, + 1022, + 1486, + 936, + 1486 + ], + "score": 0.93, + "latex": "I ( \\theta ; s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 1281, + 1775, + 1393, + 1775, + 1393, + 1806, + 1281, + 1806 + ], + "score": 0.92, + "latex": "\\rho _ { C R } \\neq \\rho _ { \\pi }" + }, + { + "category_id": 14, + "poly": [ + 535, + 1277, + 1162, + 1277, + 1162, + 1408, + 535, + 1408 + ], + "score": 0.92, + "latex": "\\begin{array} { r l } & { \\nabla _ { \\theta } J ( \\theta ) = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ) Q _ { w } ( \\boldsymbol { s } _ { t } , \\boldsymbol { a } _ { t } ) ] } \\\\ & { \\qquad = \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ ( \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ) \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ) ^ { T } ) w ] } \\\\ & { \\qquad = \\mathbb { E } _ { \\rho _ { \\pi } } [ I ( \\theta ; \\boldsymbol { s } _ { t } ) w ] , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 701, + 264, + 790, + 264, + 790, + 292, + 701, + 292 + ], + "score": 0.91, + "latex": "\\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf " + }, + { + "category_id": 13, + "poly": [ + 1170, + 1774, + 1243, + 1774, + 1243, + 1805, + 1170, + 1805 + ], + "score": 0.91, + "latex": "\\alpha \\neq 1" + }, + { + "category_id": 14, + "poly": [ + 513, + 1674, + 1184, + 1674, + 1184, + 1764, + 513, + 1764 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } & { \\nabla _ { \\theta } J ( \\theta ) \\propto \\alpha \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { \\boldsymbol A } ( s _ { t } , a _ { t } ) - \\eta \\bar { \\boldsymbol A } _ { w } ( s _ { t } , a _ { t } ) ] } \\\\ & { \\qquad + \\eta \\mathbb { E } _ { \\rho _ { C R } } [ \\nabla _ { a } \\mathcal { Q } _ { w } ( s _ { t } , a ) | _ { a = \\mu _ { \\theta } ( s _ { t } ) } \\nabla _ { \\theta } \\mu _ { \\theta } ( s _ { t } ) ] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1029, + 516, + 1029, + 516, + 1065, + 297, + 1065 + ], + "score": 0.91, + "latex": "w ^ { T } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( { \\mathbf { \\em a } } _ { t } | { \\mathbf { \\em s } } _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 844, + 1835, + 916, + 1835, + 916, + 1867, + 844, + 1867 + ], + "score": 0.9, + "latex": "\\eta = 0" + }, + { + "category_id": 13, + "poly": [ + 417, + 1233, + 454, + 1233, + 454, + 1266, + 417, + 1266 + ], + "score": 0.9, + "latex": "\\bar { Q } _ { w }" + }, + { + "category_id": 13, + "poly": [ + 466, + 1868, + 579, + 1868, + 579, + 1901, + 466, + 1901 + ], + "score": 0.9, + "latex": "\\rho _ { C R } = \\rho _ { \\beta }" + }, + { + "category_id": 13, + "poly": [ + 1127, + 1633, + 1200, + 1633, + 1200, + 1662, + 1127, + 1662 + ], + "score": 0.9, + "latex": "\\alpha , \\rho _ { C R }" + }, + { + "category_id": 13, + "poly": [ + 903, + 1207, + 941, + 1207, + 941, + 1236, + 903, + 1236 + ], + "score": 0.89, + "latex": "Q _ { w }" + }, + { + "category_id": 13, + "poly": [ + 849, + 260, + 1396, + 260, + 1396, + 295, + 849, + 295 + ], + "score": 0.89, + "latex": "\\bar { f } ( s _ { t } , a _ { t } ) = f ( s _ { t } , \\bar { a } _ { t } ) + \\nabla _ { a } \\bar { f ( } s _ { t } , a ) | _ { a = \\bar { a } _ { t } } ( a _ { t } - \\bar { a } _ { t } ) ." + }, + { + "category_id": 13, + "poly": [ + 336, + 1866, + 410, + 1866, + 410, + 1893, + 336, + 1893 + ], + "score": 0.89, + "latex": "\\alpha = 0" + }, + { + "category_id": 13, + "poly": [ + 1252, + 1001, + 1403, + 1001, + 1403, + 1035, + 1252, + 1035 + ], + "score": 0.89, + "latex": "Q _ { w } ( \\pmb { s } _ { t } , \\pmb { a } _ { t } ) =" + }, + { + "category_id": 13, + "poly": [ + 623, + 2004, + 661, + 2004, + 661, + 2034, + 623, + 2034 + ], + "score": 0.89, + "latex": "Q _ { w }" + }, + { + "category_id": 13, + "poly": [ + 1124, + 1067, + 1163, + 1067, + 1163, + 1098, + 1124, + 1098 + ], + "score": 0.88, + "latex": "Q _ { w }" + }, + { + "category_id": 13, + "poly": [ + 590, + 1031, + 967, + 1031, + 967, + 1064, + 590, + 1064 + ], + "score": 0.88, + "latex": "\\nabla _ { w } \\mathcal { Q } _ { w } ( s _ { t } , \\bar { a } _ { t } ) = \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } )" + }, + { + "category_id": 14, + "poly": [ + 506, + 1144, + 1193, + 1144, + 1193, + 1188, + 506, + 1188 + ], + "score": 0.87, + "latex": "\\begin{array} { r } { \\nabla _ { w } L = 2 \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) ( \\hat { \\mathcal { Q } } ( s _ { t } , a _ { t } ) - { \\mathcal Q } _ { w } ( s _ { t } , a _ { t } ) ) ] = 0 . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 592, + 1003, + 632, + 1003, + 632, + 1032, + 592, + 1032 + ], + "score": 0.87, + "latex": "Q _ { w }" + }, + { + "category_id": 13, + "poly": [ + 1320, + 229, + 1341, + 229, + 1341, + 262, + 1320, + 262 + ], + "score": 0.85, + "latex": "\\bar { f }" + }, + { + "category_id": 13, + "poly": [ + 375, + 1421, + 930, + 1421, + 930, + 1457, + 375, + 1457 + ], + "score": 0.85, + "latex": "I ( \\theta ; s _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } \\vert s _ { t } ) \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } \\vert s _ { t } ) ^ { T } ]" + }, + { + "category_id": 13, + "poly": [ + 913, + 297, + 944, + 297, + 944, + 323, + 913, + 323 + ], + "score": 0.85, + "latex": "\\pi _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1348, + 1037, + 1404, + 1037, + 1404, + 1062, + 1348, + 1062 + ], + "score": 0.83, + "latex": "w =" + }, + { + "category_id": 13, + "poly": [ + 1240, + 236, + 1268, + 236, + 1268, + 261, + 1240, + 261 + ], + "score": 0.82, + "latex": "\\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { a } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1073, + 1456, + 1096, + 1456, + 1096, + 1480, + 1073, + 1480 + ], + "score": 0.78, + "latex": "w" + }, + { + "category_id": 13, + "poly": [ + 316, + 1065, + 950, + 1065, + 950, + 1101, + 316, + 1101 + ], + "score": 0.73, + "latex": "\\mathrm { r g } \\operatorname* { m i n } _ { w } L ( w ) = \\mathrm { a r g } \\operatorname* { m i n } _ { w } \\mathbb { E } _ { \\rho _ { \\pi } , \\pi } [ ( \\hat { Q } ( s _ { t } , a _ { t } ) - Q _ { w } ( { \\pmb s } _ { t } , { \\pmb a } _ { t } ) ) ^ { 2 } ]" + }, + { + "category_id": 13, + "poly": [ + 1078, + 1035, + 1103, + 1035, + 1103, + 1060, + 1078, + 1060 + ], + "score": 0.63, + "latex": "w" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1525.0, + 1082.0, + 1525.0, + 1082.0, + 1569.0, + 295.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 860.0, + 1234.0, + 860.0, + 1234.0, + 901.0, + 294.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 903.0, + 602.0, + 903.0, + 602.0, + 938.0, + 354.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 970.0, + 1404.0, + 970.0, + 1404.0, + 1004.0, + 295.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 998.0, + 591.0, + 998.0, + 591.0, + 1038.0, + 293.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 998.0, + 1251.0, + 998.0, + 1251.0, + 1038.0, + 633.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1024.0, + 589.0, + 1024.0, + 589.0, + 1070.0, + 517.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 1024.0, + 1077.0, + 1024.0, + 1077.0, + 1070.0, + 968.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 1024.0, + 1347.0, + 1024.0, + 1347.0, + 1070.0, + 1104.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1063.0, + 315.0, + 1063.0, + 315.0, + 1104.0, + 291.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1063.0, + 1123.0, + 1063.0, + 1123.0, + 1104.0, + 951.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 1063.0, + 1406.0, + 1063.0, + 1406.0, + 1104.0, + 1164.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1093.0, + 957.0, + 1093.0, + 957.0, + 1134.0, + 294.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1770.0, + 1169.0, + 1770.0, + 1169.0, + 1810.0, + 293.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 1770.0, + 1280.0, + 1770.0, + 1280.0, + 1810.0, + 1244.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1770.0, + 1403.0, + 1770.0, + 1403.0, + 1810.0, + 1394.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1803.0, + 1402.0, + 1803.0, + 1402.0, + 1835.0, + 295.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1834.0, + 843.0, + 1834.0, + 843.0, + 1870.0, + 294.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1834.0, + 1406.0, + 1834.0, + 1406.0, + 1870.0, + 917.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1861.0, + 335.0, + 1861.0, + 335.0, + 1904.0, + 291.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 1861.0, + 465.0, + 1861.0, + 465.0, + 1904.0, + 411.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 1861.0, + 1406.0, + 1861.0, + 1406.0, + 1904.0, + 580.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1891.0, + 1406.0, + 1891.0, + 1406.0, + 1933.0, + 293.0, + 1933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1926.0, + 620.0, + 1926.0, + 620.0, + 1960.0, + 294.0, + 1960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 540.0, + 229.0, + 540.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 229.0, + 1239.0, + 229.0, + 1239.0, + 264.0, + 641.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 229.0, + 1319.0, + 229.0, + 1319.0, + 264.0, + 1269.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1342.0, + 229.0, + 1403.0, + 229.0, + 1403.0, + 264.0, + 1342.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 260.0, + 700.0, + 260.0, + 700.0, + 298.0, + 294.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 260.0, + 848.0, + 260.0, + 848.0, + 298.0, + 791.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 260.0, + 1405.0, + 260.0, + 1405.0, + 298.0, + 1397.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 288.0, + 365.0, + 288.0, + 365.0, + 328.0, + 292.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 288.0, + 912.0, + 288.0, + 912.0, + 328.0, + 554.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 288.0, + 1300.0, + 288.0, + 1300.0, + 328.0, + 945.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1596.0, + 1405.0, + 1596.0, + 1405.0, + 1636.0, + 294.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1626.0, + 1126.0, + 1626.0, + 1126.0, + 1668.0, + 293.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 1626.0, + 1214.0, + 1626.0, + 1214.0, + 1668.0, + 1201.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1203.0, + 902.0, + 1203.0, + 902.0, + 1239.0, + 295.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 1203.0, + 1405.0, + 1203.0, + 1405.0, + 1239.0, + 942.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1234.0, + 416.0, + 1234.0, + 416.0, + 1270.0, + 295.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 1234.0, + 1166.0, + 1234.0, + 1166.0, + 1270.0, + 455.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1419.0, + 374.0, + 1419.0, + 374.0, + 1458.0, + 295.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 1419.0, + 1403.0, + 1419.0, + 1403.0, + 1458.0, + 931.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1451.0, + 935.0, + 1451.0, + 935.0, + 1488.0, + 295.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 1451.0, + 1072.0, + 1451.0, + 1072.0, + 1488.0, + 1023.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 1451.0, + 1226.0, + 1451.0, + 1226.0, + 1488.0, + 1097.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2006.0, + 297.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 622.0, + 2002.0, + 622.0, + 2038.0, + 293.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 2002.0, + 1404.0, + 2002.0, + 1404.0, + 2038.0, + 662.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 837, + 1404, + 837, + 1404, + 1143, + 297, + 1143 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1177, + 1404, + 1177, + 1404, + 1514, + 297, + 1514 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 538, + 1404, + 538, + 1404, + 723, + 297, + 723 + ], + "score": 0.98 + }, + { + "category_id": 5, + "poly": [ + 385, + 221, + 1314, + 221, + 1314, + 428, + 385, + 428 + ], + "score": 0.974, + "html": "
ParameterImplementation optionsIntroduce bias?
Qwoff-policy TD; on-policy TD(λ); model-based; etc.No
Von-policy Monte Carlo fitting; Eπθ [Qw(St, at)}; etcNo
0≤λ≤1Yes,except=1
αα≥0Yes, except α=1
nany nNo
PCRp of any policyYes, except pcR=pπ
" + }, + { + "category_id": 0, + "poly": [ + 299, + 767, + 681, + 767, + 681, + 802, + 299, + 802 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 814, + 75, + 814, + 105, + 299, + 105 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.84 + }, + { + "category_id": 6, + "poly": [ + 319, + 447, + 1376, + 447, + 1376, + 481, + 319, + 481 + ], + "score": 0.621 + }, + { + "category_id": 1, + "poly": [ + 319, + 447, + 1376, + 447, + 1376, + 481, + 319, + 481 + ], + "score": 0.112 + }, + { + "category_id": 13, + "poly": [ + 1272, + 1050, + 1374, + 1050, + 1374, + 1084, + 1272, + 1084 + ], + "score": 0.93, + "latex": "Q _ { w } ( s , \\pmb { a } )" + }, + { + "category_id": 13, + "poly": [ + 843, + 281, + 1004, + 281, + 1004, + 311, + 843, + 311 + ], + "score": 0.93, + "latex": "\\mathbb { E } _ { \\pi _ { \\theta } } [ Q _ { w } ( \\pmb { s } _ { t } , \\pmb { a } _ { t } ) ]" + }, + { + "category_id": 13, + "poly": [ + 819, + 1481, + 980, + 1481, + 980, + 1515, + 819, + 1515 + ], + "score": 0.92, + "latex": "\\{ 0 . 0 1 , 0 . 1 , \\dot { 1 . } 0 \\dot { \\} }" + }, + { + "category_id": 13, + "poly": [ + 594, + 930, + 674, + 930, + 674, + 961, + 594, + 961 + ], + "score": 0.92, + "latex": "\\pmb { \\mu } _ { \\pmb { \\theta } } ( \\pmb { s } _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 1324, + 989, + 1400, + 989, + 1400, + 1024, + 1324, + 1024 + ], + "score": 0.91, + "latex": "V _ { \\phi } ( \\pmb { s } _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 297, + 898, + 621, + 898, + 621, + 932, + 297, + 932 + ], + "score": 0.9, + "latex": "\\pi _ { \\boldsymbol { \\theta } } ( \\mathbf { a } _ { t } | \\mathbf { s } _ { t } ) = \\bar { \\mathcal { N } } ( \\mu _ { \\boldsymbol { \\theta } } ( \\mathbf { \\bar { s } } _ { t } ) , \\Sigma _ { \\boldsymbol { \\theta } } )" + }, + { + "category_id": 13, + "poly": [ + 356, + 1360, + 470, + 1360, + 470, + 1388, + 356, + 1388 + ], + "score": 0.89, + "latex": "\\tau = 0 . 9 9 9" + }, + { + "category_id": 13, + "poly": [ + 633, + 1423, + 671, + 1423, + 671, + 1452, + 633, + 1452 + ], + "score": 0.89, + "latex": "Q _ { w }" + }, + { + "category_id": 13, + "poly": [ + 580, + 1207, + 610, + 1207, + 610, + 1241, + 580, + 1241 + ], + "score": 0.88, + "latex": "\\{ \\}" + }, + { + "category_id": 13, + "poly": [ + 739, + 1357, + 781, + 1357, + 781, + 1388, + 739, + 1388 + ], + "score": 0.88, + "latex": "1 0 ^ { 6 }" + }, + { + "category_id": 13, + "poly": [ + 330, + 1391, + 513, + 1391, + 513, + 1424, + 330, + 1424 + ], + "score": 0.88, + "latex": "\\{ 0 . 0 0 1 , 0 . 0 0 0 1 \\}" + }, + { + "category_id": 13, + "poly": [ + 1218, + 1268, + 1401, + 1268, + 1401, + 1302, + 1218, + 1302 + ], + "score": 0.87, + "latex": "\\{ 0 . 0 0 1 , \\bar { 0 } . 0 0 0 1 \\}" + }, + { + "category_id": 13, + "poly": [ + 786, + 960, + 821, + 960, + 821, + 990, + 786, + 990 + ], + "score": 0.87, + "latex": "\\Sigma _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1362, + 541, + 1401, + 541, + 1401, + 571, + 1362, + 571 + ], + "score": 0.87, + "latex": "Q _ { w }" + }, + { + "category_id": 13, + "poly": [ + 392, + 1268, + 598, + 1268, + 598, + 1301, + 392, + 1301 + ], + "score": 0.85, + "latex": "\\{ 0 . 1 , 0 . 0 1 , 0 . 0 0 1 \\}" + }, + { + "category_id": 13, + "poly": [ + 663, + 994, + 700, + 994, + 700, + 1022, + 663, + 1022 + ], + "score": 0.85, + "latex": "\\mu _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1362, + 1302, + 1401, + 1302, + 1401, + 1331, + 1362, + 1331 + ], + "score": 0.85, + "latex": "Q _ { w }" + }, + { + "category_id": 13, + "poly": [ + 1196, + 396, + 1296, + 396, + 1296, + 422, + 1196, + 422 + ], + "score": 0.84, + "latex": "\\rho _ { C R } = \\rho _ { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1451, + 348, + 1451, + 348, + 1485, + 297, + 1485 + ], + "score": 0.82, + "latex": "1 . 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 981, + 1238, + 1221, + 1238, + 1221, + 1271, + 981, + 1271 + ], + "score": 0.79, + "latex": "\\{ 1 0 0 0 , 5 0 0 0 , 2 5 0 0 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 437, + 254, + 471, + 254, + 471, + 282, + 437, + 282 + ], + "score": 0.79, + "latex": "\\overline { { Q _ { w } } }" + }, + { + "category_id": 13, + "poly": [ + 788, + 253, + 857, + 253, + 857, + 281, + 788, + 281 + ], + "score": 0.78, + "latex": "\\overline { { \\mathrm { T D } ( \\lambda ) } }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1117, + 318, + 1117, + 318, + 1139, + 297, + 1139 + ], + "score": 0.71, + "latex": "^ { a }" + }, + { + "category_id": 13, + "poly": [ + 567, + 574, + 587, + 574, + 587, + 598, + 567, + 598 + ], + "score": 0.7, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 444, + 342, + 464, + 342, + 464, + 362, + 444, + 362 + ], + "score": 0.66, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1215, + 308, + 1278, + 308, + 1278, + 335, + 1215, + 335 + ], + "score": 0.66, + "latex": "\\lambda = 1" + }, + { + "category_id": 13, + "poly": [ + 1215, + 338, + 1278, + 338, + 1278, + 363, + 1215, + 363 + ], + "score": 0.66, + "latex": "\\alpha = 1" + }, + { + "category_id": 13, + "poly": [ + 1293, + 1420, + 1394, + 1420, + 1394, + 1453, + 1293, + 1453 + ], + "score": 0.6, + "latex": "\\{ 0 . 1 , 0 . 5" + }, + { + "category_id": 13, + "poly": [ + 444, + 370, + 463, + 370, + 463, + 395, + 444, + 395 + ], + "score": 0.56, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 803, + 370, + 822, + 370, + 822, + 394, + 803, + 394 + ], + "score": 0.48, + "latex": "\\eta" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 763.0, + 684.0, + 763.0, + 684.0, + 809.0, + 293.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 444.0, + 1378.0, + 444.0, + 1378.0, + 484.0, + 320.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 836.0, + 1402.0, + 836.0, + 1402.0, + 868.0, + 295.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 863.0, + 1405.0, + 863.0, + 1405.0, + 906.0, + 291.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 899.0, + 1406.0, + 899.0, + 1406.0, + 934.0, + 622.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 927.0, + 593.0, + 927.0, + 593.0, + 966.0, + 292.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 927.0, + 1407.0, + 927.0, + 1407.0, + 966.0, + 675.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 957.0, + 785.0, + 957.0, + 785.0, + 994.0, + 292.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 957.0, + 1406.0, + 957.0, + 1406.0, + 994.0, + 822.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 987.0, + 662.0, + 987.0, + 662.0, + 1027.0, + 292.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 987.0, + 1323.0, + 987.0, + 1323.0, + 1027.0, + 701.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 987.0, + 1405.0, + 987.0, + 1405.0, + 1027.0, + 1401.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1019.0, + 1406.0, + 1019.0, + 1406.0, + 1055.0, + 294.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1050.0, + 1271.0, + 1050.0, + 1271.0, + 1085.0, + 292.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1375.0, + 1050.0, + 1405.0, + 1050.0, + 1405.0, + 1085.0, + 1375.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1081.0, + 1405.0, + 1081.0, + 1405.0, + 1116.0, + 292.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1110.0, + 746.0, + 1110.0, + 746.0, + 1145.0, + 319.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1173.0, + 1405.0, + 1173.0, + 1405.0, + 1215.0, + 294.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1205.0, + 579.0, + 1205.0, + 579.0, + 1245.0, + 291.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 1205.0, + 1404.0, + 1205.0, + 1404.0, + 1245.0, + 611.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1234.0, + 980.0, + 1234.0, + 980.0, + 1278.0, + 291.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 1234.0, + 1406.0, + 1234.0, + 1406.0, + 1278.0, + 1222.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1269.0, + 391.0, + 1269.0, + 391.0, + 1303.0, + 294.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 1269.0, + 1217.0, + 1269.0, + 1217.0, + 1303.0, + 599.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1293.0, + 1361.0, + 1293.0, + 1361.0, + 1339.0, + 291.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1293.0, + 1406.0, + 1293.0, + 1406.0, + 1339.0, + 1402.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1326.0, + 1406.0, + 1326.0, + 1406.0, + 1368.0, + 291.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1357.0, + 355.0, + 1357.0, + 355.0, + 1395.0, + 294.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 1357.0, + 738.0, + 1357.0, + 738.0, + 1395.0, + 471.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 1357.0, + 1407.0, + 1357.0, + 1407.0, + 1395.0, + 782.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1390.0, + 329.0, + 1390.0, + 329.0, + 1425.0, + 294.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 1390.0, + 1406.0, + 1390.0, + 1406.0, + 1425.0, + 514.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1421.0, + 632.0, + 1421.0, + 632.0, + 1456.0, + 294.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 1421.0, + 1292.0, + 1421.0, + 1292.0, + 1456.0, + 672.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 1421.0, + 1405.0, + 1421.0, + 1405.0, + 1456.0, + 1395.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1451.0, + 1404.0, + 1451.0, + 1404.0, + 1486.0, + 349.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1482.0, + 818.0, + 1482.0, + 818.0, + 1517.0, + 294.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1482.0, + 1361.0, + 1482.0, + 1361.0, + 1517.0, + 981.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 536.0, + 1361.0, + 536.0, + 1361.0, + 576.0, + 294.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 567.0, + 566.0, + 567.0, + 566.0, + 608.0, + 292.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 567.0, + 1406.0, + 567.0, + 1406.0, + 608.0, + 588.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 599.0, + 1405.0, + 599.0, + 1405.0, + 636.0, + 292.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 632.0, + 1405.0, + 632.0, + 1405.0, + 664.0, + 296.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 660.0, + 1404.0, + 660.0, + 1404.0, + 695.0, + 294.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 692.0, + 1005.0, + 692.0, + 1005.0, + 727.0, + 295.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 444.0, + 1378.0, + 444.0, + 1378.0, + 484.0, + 320.0, + 484.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/SJyVzQ-C-/SJyVzQ-C-.md b/parse/train/SJyVzQ-C-/SJyVzQ-C-.md new file mode 100644 index 0000000000000000000000000000000000000000..ad50c2c3043f233ecd1270d13380e97a97ee568c --- /dev/null +++ b/parse/train/SJyVzQ-C-/SJyVzQ-C-.md @@ -0,0 +1,335 @@ +# FRATERNAL DROPOUT + +Konrad $\dot { \mathbf Z } \mathbf { o } \mathbf { f n } \mathbf { a } ^ { 1 , 2 , * } ;$ Devansh Arpit2, Dendi Suhubdy2 & Yoshua Bengio2,3 + +1Jagiellonian University +2MILA, Université de Montréal +3CIFAR Senior Fellow + +# ABSTRACT + +Recurrent neural networks (RNNs) form an important class of architectures among neural networks useful for language modeling and sequential prediction. However, optimizing RNNs is known to be harder compared to feed-forward neural networks. A number of techniques have been proposed in literature to address this problem. In this paper we propose a simple technique called fraternal dropout that takes advantage of dropout to achieve this goal. Specifically, we propose to train two identical copies of an RNN (that share parameters) with different dropout masks while minimizing the difference between their (pre-softmax) predictions. In this way our regularization encourages the representations of RNNs to be invariant to dropout mask, thus being robust. We show that our regularization term is upper bounded by the expectation-linear dropout objective which has been shown to address the gap due to the difference between the train and inference phases of dropout. We evaluate our model and achieve state-of-the-art results in sequence modeling tasks on two benchmark datasets – Penn Treebank and Wikitext-2. We also show that our approach leads to performance improvement by a significant margin in image captioning (Microsoft COCO) and semi-supervised (CIFAR-10) tasks. + +# 1 INTRODUCTION + +Recurrent neural networks (RNNs) like long short-term memory (LSTM; Hochreiter & Schmidhuber (1997)) networks and gated recurrent unit (GRU; Chung et al. (2014)) are popular architectures for sequence modeling tasks like language generation, translation, speech synthesis, and machine comprehension. However, they are harder to optimize compared to feed-forward networks due to challenges like variable length input sequences, repeated application of the same transition operator at each time step, and largely-dense embedding matrix that depends on the vocabulary size. Due to these optimization challenges in RNNs, the application of batch normalization and its variants (layer normalization, recurrent batch normalization, recurrent normalization propagation) have not been as successful as their counterparts in feed-forward networks (Laurent et al., 2016), although they do considerably provide performance gains. Similarly, naive application of dropout (Srivastava et al., 2014) has been shown to be ineffective in RNNs (Zaremba et al., 2014). Therefore, regularization techniques for RNNs is an active area of research. + +To address these challenges, Zaremba et al. (2014) proposed to apply dropout only to the nonrecurrent connections in multi-layer RNNs. Variational dropout (Gal & Ghahramani (2016)) uses the same dropout mask throughout a sequence during training. DropConnect (Wan et al., 2013) applies the dropout operation on the weight matrices. Zoneout (Krueger et al. (2016)), in a similar spirit with dropout, randomly chooses to use the previous time step hidden state instead of using the current one. Similarly as a substitute for batch normalization, layer normalization normalizes the hidden units within each sample to have zero mean and unit standard deviation. Recurrent batch normalization applies batch normalization but with unshared mini-batch statistics for each time step (Cooijmans et al., 2016). + +Merity et al. (2017a) and Merity et al. (2017b) on the other hand show that activity regularization (AR) and temporal activation regularization (TAR)1 are also effective methods for regularizing LSTMs. Another more recent way of regularizing RNNs, that is similar in spirit to the approach we take, involves minimizing the difference between the hidden states of the original and the auxiliary network Serdyuk et al. (2017). + +In this paper we propose a simple regularization based on dropout that we call fraternal dropout, where we minimize an equally weighted sum of prediction losses from two identical copies of the same LSTM with different dropout masks, and add as a regularization the $\ell ^ { 2 }$ difference between the predictions (pre-softmax) of the two networks. We analytically show that our regularization objective is equivalent to minimizing the variance in predictions from different i.i.d. dropout masks; thus encouraging the predictions to be invariant to dropout masks. We also discuss how our regularization is related to expectation linear dropout Ma et al. (2016), Π-model Laine & Aila (2016) and activity regularization Merity et al. (2017b), and empirically show that our method provides non-trivial gains over these related methods which we explain furthermore in our ablation study (Section 5). + +# 2 FRATERNAL DROPOUT + +Dropout is a powerful regularization for neural networks. It is usually more effective on densely connected layers because they suffer more from overfitting compared with convolution layers where the parameters are shared. For this reason dropout is an important regularization for RNNs. However, dropout has a gap between its training and inference phase since the latter phase assumes linear activations to correct for the factor by which the expected value of each activation would be different Ma et al. (2016). In addition, the prediction of models with dropout generally vary with different dropout mask. However, the desirable property in such cases would be to have final predictions be invariant to dropout masks. + +As such, the idea behind fraternal dropout is to train a neural network model in a way that encourages the variance in predictions under different dropout masks to be as small as possible. Specifically, consider we have an RNN model denoted by $\mathcal { M } ( \boldsymbol { \theta } )$ that takes as input $\mathbf { X }$ , where $\theta$ denotes the model parameters. Let $\mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \boldsymbol { \theta } ) \in \mathbb { R } ^ { m }$ be the prediction of the model for input sample $\mathbf { X }$ at time $t$ , for dropout mask $s _ { i } ^ { t }$ and current input $\mathbf { z } ^ { t }$ , where $\mathbf { z } ^ { t }$ is a function of $\mathbf { X }$ and the hidden states corresponding to the previous time steps. Similarly, let $\ell ^ { t } ( \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \theta ) , \mathbf { Y } )$ be the corresponding $t ^ { t h }$ time step loss value for the overall input-target sample pair $( \mathbf { X } , \mathbf { Y } )$ . + +Then in fraternal dropout, we simultaneously feed-forward the input sample $\mathbf { X }$ through two identical copies of the RNN that share the same parameters $\theta$ but with different dropout masks $s _ { i } ^ { t }$ and $\boldsymbol { s } _ { j } ^ { t }$ at each time step $t$ . This yields two loss values at each time step $t$ given by $\ell ^ { t } ( \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \theta ) , \mathbf { Y } )$ , and $\ell ^ { t } ( \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { j } ^ { t } ; \bar { \theta } ) , \mathbf { Y } )$ . Then the overall loss function of fraternal dropout is given by, + +$$ +\ell _ { F D } ( \mathbf { X } , \mathbf { Y } ) = \sum _ { t = 1 } ^ { T } \frac { 1 } { 2 } \left( \ell ^ { t } ( \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \theta ) , \mathbf { Y } ) + \ell ^ { t } ( \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { j } ^ { t } ; \theta ) , \mathbf { Y } ) \right) + \frac { \kappa } { m T } \sum _ { t = 1 } ^ { T } \mathcal { R } _ { F D } ( \mathbf { z } ^ { t } ; \theta ) +$$ + +where $\kappa$ is the regularization coefficient, $m$ is the dimensions of $\mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \boldsymbol { \theta } )$ and $\mathcal { R } _ { F D } ( { \bf z } ^ { t } ; \boldsymbol { \theta } )$ is the fraternal dropout regularization given by, + +$$ +\mathcal { R } _ { F D } ( \mathbf { z } ^ { t } ; \boldsymbol { \theta } ) : = \mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \left[ \| \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \boldsymbol { \theta } ) - \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { j } ^ { t } ; \boldsymbol { \theta } ) \| _ { 2 } ^ { 2 } \right] . +$$ + +We use Monte Carlo sampling to approximate $\mathcal { R } _ { F D } ( { \bf z } ^ { t } ; \boldsymbol { \theta } )$ where $\mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \boldsymbol { \theta } )$ and $\mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { j } ^ { t } ; \theta )$ are the same as the one used to calculate $\ell ^ { t }$ values. Hence, the additional computation is negligible. + +We note that the regularization term of our objective is equivalent to minimizing the variance in the prediction function with different dropout masks as shown below (proof in the appendix). + +Remark 1. Let $s _ { i } ^ { t }$ and $\boldsymbol { s } _ { j } ^ { t }$ be i.i.d. dropout masks and $\mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \boldsymbol { \theta } ) \in \mathbb { R } ^ { m }$ be the prediction function as described above. Then, + +$$ +\mathcal { R } _ { F D } ( \mathbf { z } ^ { t } ; \theta ) = \mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \left[ \| \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \theta ) - \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { j } ^ { t } ; \theta ) \| _ { 2 } ^ { 2 } \right] = 2 \sum _ { q = 1 } ^ { m } \mathrm { v a r } _ { s _ { i } ^ { t } } ( p _ { q } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \theta ) ) . +$$ + +Note that a generalization of our approach would be to minimize the difference between the predictions of the two networks with different data/model augmentations. However, in this paper we focus on using different dropout masks and experiment mainly with $\mathrm { R N N s } ^ { 2 }$ . + +# 3 RELATED WORK + +# 3.1 RELATION TO EXPECTATION LINEAR DROPOUT (ELD) + +Ma et al. (2016) analytically showed that the expected error (over samples) between a model’s expected prediction over all dropout masks, and the prediction using the average mask, is upper bounded. Based on this result, they propose to explicitly minimize the difference (we have adapted their regularization to our notations), + +$$ +\mathcal { R } _ { E L D } ( \mathbf { z } ^ { t } ; \boldsymbol { \theta } ) = \| \mathbb { E } _ { s } \left[ \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s ; \boldsymbol { \theta } ) \right] - \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , \mathbb { E } _ { s } [ s ] ; \boldsymbol { \theta } ) \| _ { 2 } +$$ + +where $s$ is the dropout mask. However, due to feasibility consideration, they instead propose to use the following regularization in practice, + +$$ +\begin{array} { r } { \tilde { \mathcal { R } } _ { E L D } ( \mathbf { z } ^ { t } ; \theta ) = \mathbb { E } _ { s _ { i } } \left[ \Vert \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ; \theta ) - \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , \mathbb { E } _ { s } [ s ] ; \theta ) \Vert _ { 2 } ^ { 2 } \right] . } \end{array} +$$ + +Specifically, this is achieved by feed-forwarding the input twice through the network, with and without dropout mask, and minimizing the main network loss (with dropout) along with the regularization term specified above (but without back-propagating the gradients through the network without dropout). The goal of Ma et al. (2016) is to minimize the network loss along with the expected difference between the prediction from individual dropout mask and the prediction from the expected dropout mask. We note that our regularization objective is upper bounded by the expectation-linear dropout regularization as shown below (proof in the appendix). + +Proposition 1. $\mathcal { R } _ { F D } ( \mathbf { z } ^ { t } ; \boldsymbol { \theta } ) \leq 4 \tilde { \mathcal { R } } _ { E L D } ( \mathbf { z } ^ { t } ; \boldsymbol { \theta } ) .$ . + +This result shows that minimizing the ELD objective indirectly minimizes our regularization term. Finally as indicated above, they apply the target loss only on the network with dropout. In fact, in our own ablation studies (see Section 5) we find that back-propagating target loss through the network (without dropout) makes optimizing the model harder. However, in our setting, simultaneously backpropagating target loss through both networks yields both performance gain as well as convergence gain. We believe convergence is faster for our regularization because network weights are more likely to get target based updates from back-propagation in our case. This is especially true for weight dropout (Wan et al., 2013) since in this case dropped weights do not get updated in the training iteration. + +# 3.2 RELATION TO Π-MODEL + +Laine & Aila (2016) propose $\Pi$ -model with the goal of improving performance on classification tasks in the semi-supervised setting. They propose a model similar to ours (considering the equivalent deep feed-forward version of our model) except they apply target loss only on one of the networks and use time-dependent weighting function $\omega ( t )$ (while we use constant $\frac { \kappa } { m T }$ ). The intuition in their case is to leverage unlabeled data by using them to minimize the difference in prediction between the two copies of the network with different dropout masks. Further, they also test their model in the supervised setting but fail to explain the improvements they obtain by using this regularization. + +We note that in our case we analytically show that minimizing our regularizer (also used in $\Pi$ -model) is equivalent to minimizing the variance in the model predictions (Remark 1). Furthermore, we also show the relation of our regularizer to expectation linear dropout (Proposition 1). In Section 5, we study the effects of target based loss on both networks, which is not used in the $\Pi$ -model. We find that applying target loss on both the networks leads to significantly faster convergence. Finally, we bring to attention that temporal embedding (another model proposed by Laine & Aila (2016), claimed to be a better version of $\Pi$ -model for semi-supervised, learning) is intractable in natural language processing applications because storing averaged predictions over all of the time steps would be memory exhaustive (since predictions are usually huge - tens of thousands values). On a final note, we argue that in the supervised case, using a time-dependent weighting function $\omega ( t )$ instead of a constant value $\frac { \kappa } { m T }$ is not needed. Since the ground truth labels are known, we have not observed the problem mentioned by Laine & Aila (2016), that the network gets stuck in a degenerate solution when $\omega ( t )$ is too large in earlier epochs of training. We note that it is much easier to search for an optimal constant value, which is true in our case, as opposed to tuning the time-dependent function. + +Similarity to $\Pi$ -model makes our method related to other semi-supervised works, mainly Rasmus et al. (2015) and Sajjadi et al. (2016). Since semi-supervised learning is not a primary focus of this paper, we refer to Laine & Aila (2016) for more details. + +Another way to address the gap between the train and evaluation mode of dropout is to perform Monte Carlo sampling of masks and average the predictions during evaluation, and this has been used for feed-forward networks. We find that this technique does not work well for RNNs. The details of these experiments can be found in the appendix. + +# 4 EXPERIMENTS + +# 4.1 LANGUAGE MODELS + +In the case of language modeling we test our model3 on two benchmark datasets – Penn Tree-bank (PTB) dataset (Marcus et al., 1993) and WikiText-2 (WT2) dataset (Merity et al., 2016). We use preprocessing as specified by Mikolov et al. (2010) (for PTB corpus) and Moses tokenizer Koehn et al. (2007) (for the WT2 dataset). + +For both datasets we use the AWD-LSTM 3-layer architecture described in Merity et al. (2017a)4 which we call the baseline model. The number of parameters in the model used for PTB is 24 million as compared to 34 million in the case of WT2 because WT2 has a larger vocabulary size for which we use a larger embedding matrix. Apart from those differences, the architectures are identical. When we use fraternal dropout, we simply add our regularization on top of this baseline model. + +# Word level Penn Treebank (PTB) + +Influenced by Melis et al. (2017), our goal here is to make sure that fraternal dropout outperforms existing methods not simply because of extensive hyper-parameter grid search but rather due to its regularization effects. Hence, in our experiments we leave a vast majority of hyper-parameters used in the baseline model (Melis et al., 2017) unchanged i.e. embedding and hidden states sizes, gradient clipping value, weight decay and the values used for all dropout layers (dropout on the word vectors, the output between LSTM layers, the output of the final LSTM, and embedding dropout). However, a few changes are necessary: + +• the coefficients for AR and TAR needed to be altered because fraternal dropout also affects RNNs activation (as explained in Subsection 5.3) – we did not run grid search to obtain the best values but simply deactivated AR and TAR regularizers; • since fraternal dropout needs twice as much memory, batch size is halved so the model needs approximately the same amount of memory and hence fits on the same GPU. + +The final change in hyper-parameters is to alter the non-monotone interval $n$ used in nonmonotonically triggered averaged SGD (NT-ASGD) optimizer Polyak & Juditsky (1992); Mandt et al. (2017); Melis et al. (2017). We run a grid search on $n \in \{ 5 , 2 5 , 4 0 , 5 0 , 6 0 \}$ and obtain very similar results for the largest values (40, 50 and 60) in the candidate set. Hence, our model is trained longer using ordinary SGD optimizer as compared to the baseline model (Melis et al., 2017). + +Table 1: Perplexity on Penn Treebank word level language modeling task. + +
ModelParametersValidationTest
Zaremba et al. (2014) - LSTM (medium)10M86.282.7
Zaremba et al. (2014) - LSTM (large)24M82.278.4
Gal & Ghahramani (2016) - Variational LSTM (medium)20M81.979.7
Gal & Ghahramani (2016) - Variational LSTM (large)66M77.975.2
Inan et al. (2016) - Variational LSTM51M71.168.5
Inan et al. (2016) - Variational RHN24M68.166.0
Zilly et al. (2016) - Variational RHN23M67.965.4
Melis et al. (2017) - 5-layer RHN24M64.862.2
Melis et al. (2017) - 4-layer skip connection LSTM24M60.958.3
Merity etal. (2017a) - AWD-LSTM3-layer (baseline)24M60.057.3
Fraternal dropout + AWD-LSTM3-layer24M58.956.8
+ +Table 2: Perplexity on WikiText-2 word level language modeling task. + +
ModelParametersValidationTest
Merity et al. (2016) - Variational LSTM + Zoneout20M108.7100.9
Merity et al. (2016) - Variational LSTM20M101.796.3
Inan et al. (2016) - Variational LSTM28M91.587.0
Melis et al. (2017) - 5-layer RHN24M78.175.6
Melis et al. (2017) - 1-layer LSTM24M69.365.9
Melis et al. (2017) - 2-layer skip connection LSTM24M69.165.9
Merity et al. (2017a) - AWD-LSTM3-layer (baseline)34M68.665.8
Fraternal dropout + AWD-LSTM 3-layer34M66.864.1
+ +We evaluate our model using the perplexity metric and compare the results that we obtain against the existing state-of-the-art results. The results are reported in Table 1. Our approach achieves the state-of-the-art performance compared with existing benchmarks. + +To confirm that the gains are robust to initialization, we run ten experiments for the baseline model with different seeds (without fine-tuning) for PTB dataset to compute confidence intervals. The average best validation perplexity is $6 0 . 6 4 \pm 0 . 1 5$ with the minimum value equals 60.33. The same for test perplexity is $5 8 . 3 2 \pm 0 . 1 4$ and 58.05, respectively. Our score (59.8 validation and 58.0 test perplexity) beats ordinal dropout minimum values. + +We also perform experiments using fraternal dropout with a grid search on all the hyper-parameters and find that it leads to further improvements in performance. The details of this experiment can be found in section 5.5. + +# Word level WikiText-2 (WT2) + +In the case of WikiText-2 language modeling task, we outperform the current state-of-the-art using the perplexity metric by a significant margin. Due to the lack of computational power, we run a single training procedure for fraternal dropout on WT2 dataset because it is larger than PTB. In this experiment, we use the best hyper-parameters found for PTB dataset $\kappa = 0 . 1$ , non-monotone interval $n = 6 0$ and halved batch size; the rest of the hyper-parameters are the same as described in Melis et al. (2017) for WT2). The final results are presented in Table 2. + +# 4.2 IMAGE CAPTIONING + +We also apply fraternal dropout on an image captioning task. We use the well-known show and tell model as a baseline5 (Vinyals et al., 2014). We emphasize that in the image captioning task, the image encoder and sentence decoder architectures are usually learned together. Since we want to focus on the benefits of using fraternal dropout in RNNs we use frozen pretrained ResNet-101 (He et al., 2015) model as our image encoder. It means that our results are not directly comparable with other state-of-the-art methods, however we report results for the original methods so readers can see that our baseline performs well. The final results are presented in Table 3. + +Table 3: BLEU scores for the Microsoft COCO image captioning task. Using fraternal dropout is the only difference between models. The rest of hyper-parameters are the same. + +
ModelBLEU-1BLEU-2BLEU-3BLEU-4
Show and Tell Xu et al. (2015)66.646.132.924.6
Baseline68.850.836.125.6
Fraternal dropout, K = 0.01569.351.436.626.1
Fraternal dropout, K = 0.00569.351.536.926.3
+ +We argue that in this task smaller $\kappa$ values are optimal because the image captioning encoder is given all information in the beginning and hence the variance of consecutive predictions is smaller that in unconditioned natural language processing tasks. Fraternal dropout may benefits here mainly due to averaging gradients for different mask and hence updating weights more frequently. + +# 5 ABLATION STUDIES + +In this section, the goal is to study existing methods closely related to ours – expectation linear dropout Ma et al. (2016), Π-model Laine & Aila (2016) and activity regularization Merity et al. (2017b). All of our experiments for ablation studies, which apply a single layer LSTM, use the same hyper-parameters and model architecture6 as Melis et al. (2017). + +# 5.1 EXPECTATION-LINEAR DROPOUT (ELD) + +The relation with expectation-linear dropout Ma et al. (2016) has been discussed in Section 2. Here we perform experiments to study the difference in performance when using the ELD regularization versus our regularization (FD). In addition to ELD, we also study a modification (ELDM) of ELD which applies target loss to both copies of LSTMs in ELD similar to FD (notice in their case they only have dropout on one LSTM). Finally we also evaluate a baseline model without any of these regularizations. The learning dynamics curves are shown in Figure 1. Our regularization performs better in terms of convergence compared with other methods. In terms of generalization, we find that FD is similar to ELD, but baseline and ELDM are much worse. Interestingly, looking at the train and validation curves together, ELDM seems to be suffering from optimization problems. + +# 5.2 Π-MODEL + +Since Π-model Laine & Aila (2016) is similar to our algorithm (even though it is designed for semi-supervised learning in feed-forward networks), we study the difference in performance with Π-model7 both qualitatively and quantitatively to establish the advantage of our approach. First, we run both single layer LSTM and 3-layer AWD-LSTM on PTB task to check how their model compares with ours in the case of language modeling. The results are shown in Figure 1 and 2. We find that our model converges significantly faster than Π-model. We believe this happens because we back-propagate the target loss through both networks (in contrast to Π-model) that leads to weights getting updated using target-based gradients more often. + +![](images/4b8d4833a24b6748ef3cd0f76c9fe0025e06c99555d7e766f2a29c044f96997c.jpg) +Figure 1: Ablation study: Train (left) and validation (right) perplexity on PTB word level modeling with single layer LSTM (10M parameters). These curves study the learning dynamics of the baseline model, Π-model, Expectation-linear dropout (ELD), Expectation-linear dropout with modification (ELDM) and fraternal dropout (FD, our algorithm). We find that FD converges faster than the regularizers in comparison, and generalizes at par. + +![](images/d64e450f08346c6bc468f12061b79ab092e7cc2cf8930eb3af1ac794100c063b.jpg) + +![](images/87020ced434d3499701675350c507f939c1046c58cb17698501cdf12b3ae97a7.jpg) +Figure 2: Ablation study: Validation perplexity on PTB word level modeling for $\Pi$ -model and fraternal dropout. We find that FD converges faster and generalizes at par. + +![](images/463bd2f8083be91f8df7ca8a7e5292781fe18c685b3e04efd116713fc227bc09.jpg) +Figure 3: Ablation study: Average hidden state activation is reduced when any of the regularizer described is used. The y-axis is the value of $\frac { 1 } { d } \| \mathbf { m } \cdot \mathbf { h } ^ { t } \| _ { 2 } ^ { 2 }$ . + +Even though we designed our algorithm specifically to address problems in RNNs, to have a fair comparison, we compare with $\Pi$ -model on a semi-supervised task which is their goal. Specifically, we use the CIFAR-10 dataset that consists of $3 2 \times 3 2$ images from 10 classes. Following the usual splits used in semi-supervised learning literature, we use 4 thousand labeled and 41 thousand unlabeled samples for training, 5 thousand labeled samples for validation and 10 thousand labeled samples for test set. We use the original ResNet-56 (He et al., 2015) architecture. We run grid search on $\kappa \in \{ 0 . 0 5 , 0 . 1 , 0 . 1 5 , 0 . 2 \}$ , dropout rates in $\{ 0 . 0 5 , 0 . 1 , 0 . 1 5 , 0 . 2 \}$ and leave the rest of the hyperparameters unchanged. We additionally check importance of using unlabeled data. The results are reported in Table 4. We find that our algorithm performs at par with $\Pi$ -model. When unlabeled data is not used, fraternal dropout provides slightly better results as compared to traditional dropout. + +# 5.3 ACTIVITY REGULARIZATION AND TEMPORAL ACTIVITY REGULARIZATION ANALYSIS + +The authors of Merity et al. (2017b) study the importance of activity regularization $( \mathsf { A R } ) ^ { 8 }$ and temporal activity regularization (TAR) in LSTMs given as, + +$$ +\begin{array} { c } { \mathcal { R } _ { A R } ( { \bf z } ^ { t } ; \theta ) = \displaystyle \frac { \alpha } { d } \| { \bf h } ^ { t } \| _ { 2 } ^ { 2 } } \\ { \mathcal { R } _ { T A R } ( { \bf z } ^ { t } ; \theta ) = \displaystyle \frac { \beta } { d } \| { \bf h } ^ { t } - { \bf h } ^ { t - 1 } \| _ { 2 } ^ { 2 } } \end{array} +$$ + +where $\mathbf { h } ^ { t } \in \mathbb { R } ^ { d }$ is the LSTM’s output activation at time step $t$ (hence depends on both current input $\mathbf { z } ^ { t }$ and the model parameters $\theta$ ). Notice that AR and TAR regularizations are applied on the output of + +
ModelDropout rateUnlabeled dataValidationTest
Traditional dropout0.1No78.4 (± 0.25)76.9 (± 0.31)
No dropout0.0No78.8 (± 0.59)77.1 (± 0.3)
Fraternal dropout (k = 0.05)0.05No79.3 (± 0.38)77.6 (± 0.35)
Traditional dropout + I-model0.1Yes80.2 (± 0.33)78.5 (± 0.46)
Fraternal dropout (k = 0.15)0.1Yes80.5 (± 0.18)79.1 (± 0.37)
+ +Table 4: Ablation study: Accuracy on altered (semi-supervised) CIFAR-10 dataset for ResNet-56 based models. We find that our algorithm performs at par with $\Pi$ -model. When unlabeled data is not used traditional dropout hurts performance while fraternal dropout provides slightly better results. It means that our methods may be beneficial when we lack data and have to use additional regularizing methods. + +![](images/45583947707d0455d79d12a5db380878e6174220941656e29500ab8024936b0e.jpg) +Figure 4: Ablation study: Train (left) and validation (right) perplexity on PTB word level modeling with single layer LSTM (10M parameters). These curves study the learning dynamics of the baseline model, temporal activity regularization (TAR), prediction regularization (PR), activity regularization (AR) and fraternal dropout (FD, our algorithm). We find that FD both converges faster and generalizes better than the regularizers in comparison. + +the LSTM, while our regularization is applied on the pre-softmax output $\mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \boldsymbol { \theta } )$ of the LSTM. However, since our regularization can be decomposed as + +$$ +\begin{array} { r l } & { \mathcal { R } _ { F D } ( \mathbf { z } ^ { t } ; \boldsymbol { \theta } ) = \mathbb { E } _ { s _ { i } , s _ { j } } \left[ | | \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \boldsymbol { \theta } ) - \mathbf { p } ^ { t } ( \mathbf { h } ^ { t } , s _ { j } ^ { t } ; \boldsymbol { \theta } ) | | _ { 2 } ^ { 2 } \right] } \\ & { \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \boldsymbol { \theta } ) ^ { T } \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { j } ^ { t } ; \boldsymbol { \theta } ) | | _ { 2 } ^ { 2 } + | | \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { j } ^ { t } ; \boldsymbol { \theta } ) | | _ { 2 } ^ { 2 } - 2 \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \boldsymbol { \theta } ) ^ { T } \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { j } ^ { t } ; \boldsymbol { \theta } ) \Big ] } \\ & { \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad } \\ & { \quad \quad \quad \quad \quad \quad \quad \quad \quad = \mathbb { E } _ { s _ { i } , s _ { j } } \left[ | | \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \boldsymbol { \theta } ) | | _ { 2 } ^ { 2 } + | | \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { j } ^ { t } ; \boldsymbol { \theta } ) | | _ { 2 } ^ { 2 } - 2 \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \boldsymbol { \theta } ) ^ { T } \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { j } ^ { t } ; \boldsymbol { \theta } ) \right] } \end{array} +$$ + +and encapsulates an $\ell ^ { 2 }$ term along with the dot product term, we perform experiments to confirm that the gains in our approach is not due to the $\bar { \ell } ^ { 2 }$ regularization alone. A similar argument goes for the TAR objective. We run a grid search on $\alpha \in \{ 1 , 2 , \ldots , 1 2 \}$ , $\beta \in \{ 1 , 2 , \ldots , 1 2 \}$ , which include the hyper-parameters mentioned in Merity et al. (2017a). For our regularization, we use $\kappa \in \{ 0 . 0 5 , 0 . 1 , \ldots , 0 . 4 \}$ . Furthermore, we also compare with a regularization (PR) that regularizes $\| \mathbf { p } ^ { t } ( \bar { \mathbf { z } } ^ { t } , s _ { i } ^ { t } ; \theta ) \| _ { 2 } ^ { 2 }$ to further rule-out any gains only from $\ell ^ { 2 }$ regularization. Based on this grid search, we pick the best model on the validation set for all the regularizations, and additionally report a baseline model without any of these four mentioned regularizations. The learning dynamics is shown in Figure 4. Our regularization performs better both in terms of convergence and generalization compared with other methods. Average hidden state activation is reduced when any of the regularizer described is applied (see Figure 3). + +# 5.4 IMPROVEMENTS USING FINE-TUNING + +We confirm that models trained with fraternal dropout benefit from the NT-ASGD fine-tuning step (as also used in Merity et al. (2017a)). However, this is a very time-consuming practice and since different hyper-parameters may be used in this additional part of the learning procedure, the probability of obtaining better results due to the extensive grid search is higher. Hence, in our experiments we use the same fine-tuning procedure as implemented in the official repository (even fraternal dropout was not used). We present the importance of fine-tuning in Table 5. + +Table 5: Ablation study: Importance of fine-tuning for AWD-LSTM 3-layer model. Perplexity for the Penn Treebank and WikiText-2 language modeling tasks. + +
PTBWT2
DropoutFine-tuningValidationTestValidationTest
TraditionalNone60.758.869.166.0
TraditionalOne60.057.368.665.8
FraternalNone59.858.068.365.3
FraternalOne58.956.866.864.1
FraternalTwo58.556.211
+ +Table 6: Ablation study: Candidate hyper-parameters possible used in the grid search for comparing fraternal dropout and expectation linear dropout. $U ( a , b )$ is the uniform distribution on the interval $[ a , b ]$ . For finite sets, each value is drawn with equal probability. + +
Hyper-parameterPossible values
batch size non-monotone interval[10,20,30, 40] [5,10,20,40,60,100]
K -FD or ELD strength weight decayU(0, 0.3) U(0.6 × 10-6,2.4 × 10-6)
+ +Table 7: Ablation study: Fraternal dropout and expectation linear dropout comparison. Perplexity on the Penn Treebank validation dateset. Fraternal dropout is more robust to different hyperparameters choice as twice as much runs finished performing better than the baseline model (60.7). + +
ModelBestTop5 avgTop10 avgBeating baseline runs (out of)
Expectation lineardropout59.460.160.56(208)
Fraternal dropout59.459.659.914 (203)
+ +# 5.5 FRATERNAL DROPOUT AND EXPECTATION LINEAR DROPOUT COMPARISON + +We perform extensive grid search for the baseline model from Subsection 4.1 (an AWD-LSTM 3-layer architecture) trained with either fraternal dropout or expectation linear dropout regularizations, to further contrast the performance of these two methods. The experiments are run without fine-tuning on the PTB dataset. + +In each run, all five dropout rates are randomly altered (they are set to their original value, as in Merity et al. (2017a), multiplied by a value drawn from the uniform distribution on the interval [0.5, 1.5]) and the rest of the hyper-parameters are drawn as shown in Table 6. As in Subsection 4.1, AR and TAR regularizers are deactivated. + +Together we run more than 400 experiments. The results are presented in Table 7. Both FD and ELD perform better than the baseline model that instead uses AR and TAR regularizers. Hence, we confirm our previous finding (see Subsection 5.3) that both FD and ELD are better. However, as found previously for smaller model in Subsection 5.1, the convergence of FD is faster than that of ELD. Additionally, fraternal dropout is more robust to different hyper-parameters choice (more runs performing better than the baseline and better average for top performing runs). + +# 6 CONCLUSION + +In this paper we propose a simple regularization method for RNNs called fraternal dropout that acts as a regularization by reducing the variance in model predictions across different dropout masks. We show that our model achieves state-of-the-art results on benchmark language modeling tasks along with faster convergence. We also analytically study the relationship between our regularization and expectation linear dropout Ma et al. (2016). We perform a number of ablation studies to evaluate our model from different aspects and carefully compare it with related methods both qualitatively and quantitatively. + +# ACKNOWLEDGEMENTS + +The authors would like to acknowledge the support of the following agencies for research funding and computing support: NSERC, CIFAR, and IVADO. We would like to thank Rosemary Nan Ke and Philippe Lacaille for their thoughts and comments throughout the project. We would also like to thank Stanisław Jastrz˛ebski† and Evan Racah† for useful discussions. + +# REFERENCES + +Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. + +Tim Cooijmans, Nicolas Ballas, César Laurent, and Aaron C. Courville. Recurrent batch normalization. CoRR, abs/1603.09025, 2016. URL http://arxiv.org/abs/1603.09025. + +Yarin Gal and Zoubin Ghahramani. A theoretically grounded application of dropout in recurrent neural networks. In Advances in neural information processing systems, pp. 1019–1027, 2016. + +Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CoRR, abs/1512.03385, 2015. URL http://arxiv.org/abs/1512.03385. + +Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8): 1735–1780, 1997. + +Hakan Inan, Khashayar Khosravi, and Richard Socher. Tying word vectors and word classifiers: A loss framework for language modeling. arXiv preprint arXiv:1611.01462, 2016. + +Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, et al. Moses: Open source toolkit for statistical machine translation. In Proceedings of the 45th annual meeting of the ACL on interactive poster and demonstration sessions, pp. 177–180. Association for Computational Linguistics, 2007. + +David Krueger, Tegan Maharaj, János Kramár, Mohammad Pezeshki, Nicolas Ballas, Nan Rosemary Ke, Anirudh Goyal, Yoshua Bengio, Hugo Larochelle, Aaron Courville, et al. Zoneout: Regularizing rnns by randomly preserving hidden activations. arXiv preprint arXiv:1606.01305, 2016. + +Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. + +César Laurent, Gabriel Pereyra, Philémon Brakel, Ying Zhang, and Yoshua Bengio. Batch normalized recurrent neural networks. In Acoustics, Speech and Signal Processing (ICASSP), 2016 IEEE International Conference on, pp. 2657–2661. IEEE, 2016. + +Xuezhe Ma, Yingkai Gao, Zhiting Hu, Yaoliang Yu, Yuntian Deng, and Eduard Hovy. Dropout with expectation-linear regularization. arXiv preprint arXiv:1609.08017, 2016. + +Stephan Mandt, Matthew D Hoffman, and David M Blei. Stochastic gradient descent as approximate bayesian inference. arXiv preprint arXiv:1704.04289, 2017. + +Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated corpus of english: The penn treebank. Computational linguistics, 19(2):313–330, 1993. + +Gábor Melis, Chris Dyer, and Phil Blunsom. On the state of the art of evaluation in neural language models. arXiv preprint arXiv:1707.05589, 2017. + +Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. CoRR, abs/1609.07843, 2016. URL http://arxiv.org/abs/1609.07843. + +Stephen Merity, Nitish Shirish Keskar, and Richard Socher. Regularizing and optimizing lstm language models. arXiv preprint arXiv:1708.02182, 2017a. + +Stephen Merity, Bryan McCann, and Richard Socher. Revisiting activation regularization for language rnns. arXiv preprint arXiv:1708.01009, 2017b. + +Tomas Mikolov, Martin Karafiát, Lukas Burget, Jan Cernocky, and Sanjeev Khudanpur. Recurrent \` neural network based language model. In Interspeech, volume 2, pp. 3, 2010. + +Boris T Polyak and Anatoli B Juditsky. Acceleration of stochastic approximation by averaging. SIAM Journal on Control and Optimization, 30(4):838–855, 1992. + +Ofir Press and Lior Wolf. Using the output embedding to improve language models. arXiv preprint arXiv:1608.05859, 2016. + +Antti Rasmus, Harri Valpola, Mikko Honkala, Mathias Berglund, and Tapani Raiko. Semisupervised learning with ladder network. CoRR, abs/1507.02672, 2015. URL http://arxiv. org/abs/1507.02672. + +Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In Advances in Neural Information Processing Systems, pp. 1163–1171, 2016. + +Dmitriy Serdyuk, Nan Rosemary Ke, Alessandro Sordoni, Adam Trischler, Chris Pal, and Yoshua Bengio. Twin networks: Matching the future for sequence generation. 2017. + +Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. Journal of machine learning research, 15(1):1929–1958, 2014. + +Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. CoRR, abs/1411.4555, 2014. URL http://arxiv.org/abs/1411. 4555. + +Li Wan, Matthew Zeiler, Sixin Zhang, Yann L Cun, and Rob Fergus. Regularization of neural networks using dropconnect. In Proceedings of the 30th international conference on machine learning (ICML-13), pp. 1058–1066, 2013. + +Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron C. Courville, Ruslan Salakhutdinov, Richard S. Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. CoRR, abs/1502.03044, 2015. URL http://arxiv.org/abs/1502. 03044. + +Wojciech Zaremba, Ilya Sutskever, and Oriol Vinyals. Recurrent neural network regularization. arXiv preprint arXiv:1409.2329, 2014. + +Julian Georg Zilly, Rupesh Kumar Srivastava, Jan Koutník, and Jürgen Schmidhuber. Recurrent highway networks. arXiv preprint arXiv:1607.03474, 2016. + +# APPENDIX + +# MONTE CARLO EVALUATION + +A well known way to address the gap between the train and evaluation mode of dropout is to perform Monte Carlo sampling of masks and average the predictions during evaluation (MC-eval), and this has been used for feed-forward networks. Since fraternal dropout addresses the same problem, we would like to clarify that it is not straight-forward and feasible to apply MC-eval for RNNs. In feed-forward networks, we average the output prediction scores from different masks. However, in the case RNNs (for next step predictions), there is more than one way to perform such evaluation, but each one is problematic. They are as follows: + +# 1. Online averaging + +Consider that we first make the prediction at time step 1 using different masks by averaging the prediction score. Then we use this output to feed as input to the time step 2, then use different masks at time step 2 to generate the output at time step 2, and so on. But in order to do so, because of the way RNNs work, we also need to feed the previous time hidden state to time step 2. One way would be to average the hidden states over different masks at time step 1. But the hidden space can in general be highly nonlinear, and it is not clear if averaging in this space is a good strategy. This approach is not justified. + +Besides, this strategy as a whole is extremely time consuming because we would need to sequentially make predictions with multiple masks at each time step. + +# 2. Sequence averaging + +Let’s consider that we use a different mask each time we want to generate a sequence, and then we average the prediction scores, and compute the argmax (at each time step) to get the actual generated sequence. + +In this case, notice it is not guaranteed that the predicted word at time step $t$ due to averaging the predictions would lead to the next word (generated by the same process) if we were to feed the time step $t$ output as input to the time step $t + 1$ . For example, with different dropout masks, if the probability of $\bar { 1 } ^ { s t }$ time step outputs are: $1 4 0 \%$ ), he $( 3 0 \% )$ , she $( 3 0 \% )$ , and the probability of the 2nd time step outputs are: am $( 3 0 \% )$ , is $( 6 0 \% )$ , was $( 1 0 \% )$ . Then the averaged prediction score followed by argmax will result in the prediction “I is”, but this would be incorrect. A similar concern applies for output predictions varying in temporal length. + +Hence, this approach can not be used to generate a sequence (it has to be done by by sampling a mask and generating a single sequence). However, this approach may be used to estimate the probability assigned by the model to a given sequence. + +Nonetheless, we run experiments on the PTB dataset using MC-eval (the results are summarized in Table 8). We start with a simple comparison that compares fraternal dropout with the averaged mask and the AWD-LSTM 3-layer baseline with a single fixed mask that we call MC1. The MC1 model performs much worse than fraternal dropout. Hence, it would be hard to use MC1 model in practice because a single sample is inaccurate. We also check MC-eval for a larger number of models (MC50) (50 models were used since we were not able to fit more models simultaneously on a single GPU). The final results for MC50 are worse than the baseline which uses the averaged mask. For comparison, we also evaluate MC10. Note that no fine-tuning is used for the above experiments. + +Table 8: Appendix: Monte Carlo evaluation. Perplexity on Penn Treebank word level language modeling task using Monte Carlo sampling, fraternal dropout or average mask. + +
ModelValidationTest
MC1 MC10 MC50 Baseline (average mask) Fraternal dropout92.2 (± 0.5) 66.2 (± 0.2) 64.4 (± 0.1) 60.7 59.889.2 (± 0.5) 63.7 (± 0.2) 62.1 (± 0.1) 58.8 58.0
+ +# REASONS FOR FOCUSING ON RNNS + +The fraternal dropout method is general and may be applied in feed-forward architectures (as shown in Subsection 5.2 for CIFAR-10 semisupervised example). However, we believe that it is more powerful in the case of RNNs because: + +1. Variance in prediction accumulates among time steps in RNNs and since we share parameters for all time steps, one may use the same $\kappa$ value at each step. In feed-forward networks the layers usually do not share parameters and hence one may want to use different $\kappa$ values for different layers (which may be hard to tune). The simple way to alleviate this problem is to apply the regularization term on the pre-softmax predictions only (as shown in the paper) or use the same $\kappa$ value for all layers. However, we believe that it may limit possible gains. +2. The best performing RNN architectures (state-of-the-art) usually use some kind of dropout (embedding dropout, word dropout, weight dropout etc.), very often with high dropout rates (even larger than $50 \%$ for input word embedding in NLP tasks). However, this is not true for feed-forward networks. For instance, ResNet architectures very often do not use dropout at all (probably because batch normalization is often better to use). It can be seen in the paper (Subsection 5.2, semisupervised CIFAR-10 task) that when unlabeled data is not used the regular dropout hurts performance and using fraternal dropout seems to improve just a little. +3. On a final note, the Monte Carlo sampling (a well known method that adresses the gap betweem the train and evaluation mode of dropout) can not be easily applied for RNNs and fraternal dropout may be seen as an alternative. + +To conclude, we believe that when the use of dropout benefits in a given architecture, applying fraternal dropout should improve performance even more. + +As mentioned before, in image recognition tasks, one may experiment with something what we would temporarily dub fraternal augmentation (even though dropout is not used, one can use random data augmentation such as random crop or random flip). Hence, one may force a given neural network to have the same predictions for different augmentations. + +PROOFS + +Remark 1. Let $s _ { i } ^ { t }$ and $\boldsymbol { s } _ { j } ^ { t }$ be i.i.d. dropout masks and $\mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \boldsymbol { \theta } ) \in \mathbb { R } ^ { m }$ be the prediction function as described above. Then, + +$$ +\mathcal { R } _ { F D } ( \mathbf { z } ^ { t } ; \theta ) = \mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \left[ \| \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \theta ) - \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { j } ^ { t } ; \theta ) \| _ { 2 } ^ { 2 } \right] = 2 \sum _ { q = 1 } ^ { m } \mathrm { v a r } _ { s _ { i } ^ { t } } ( p _ { q } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \theta ) ) . +$$ + +Proof. For simplicity of notation, we omit the time index $t$ + +$$ +\begin{array} { r l } { \mathcal { R } _ { F S } [ \mathbf { z } , \boldsymbol { \theta } ] } & { = \mathrm { E } _ { \boldsymbol { \phi } , \boldsymbol { \xi } } [ \| \mathbf { P } ( \mathbf { z } , \boldsymbol { s } , \boldsymbol { \xi } , \boldsymbol { \theta } ) - \mathbf { p } ( \mathbf { z } , \boldsymbol { s } , \boldsymbol { \xi } ) \boldsymbol { \theta } \| ^ { 2 } ] } \\ & { = \mathrm { E } _ { \boldsymbol { \phi } , \boldsymbol { \xi } } \Bigg [ \| \mathbf { p } ( \mathbf { z } , \boldsymbol { s } , \boldsymbol { \xi } ) \| _ { \mathbf { z } } ^ { 2 } + \mathrm { E } _ { \boldsymbol { \xi } , \boldsymbol { \xi } } \Bigg [ \| \mathbf { p } ( \mathbf { z } , \boldsymbol { s } , \boldsymbol { \xi } ) \| _ { \mathbf { z } } ^ { 2 } \Bigg ] } \\ & { - 2 \mathrm { E } _ { \boldsymbol { \xi } , \boldsymbol { s } , \boldsymbol { \xi } } [ \| \mathbf { p } ( \mathbf { z } , \boldsymbol { s } , \boldsymbol { \xi } ) \boldsymbol { \theta } \| ^ { 2 } \mathbf { p } ( \mathbf { z } , \boldsymbol { s } , \boldsymbol { \xi } ) \boldsymbol { \theta } ] } \\ & { = 2 \sum _ { \boldsymbol { \phi } = 1 } ^ { \infty } ( \mathrm { E } _ { \boldsymbol { \phi } } [ \partial _ { \boldsymbol { \phi } } ( \mathbf { z } , \boldsymbol { s } , \boldsymbol { \xi } ) \boldsymbol { \theta } ] ^ { 2 } ) - \mathrm { E } _ { \boldsymbol { \phi } , \boldsymbol { \xi } } \Bigg [ \mathcal { P } _ { \boldsymbol { \phi } } ( \mathbf { z } , \boldsymbol { s } , \boldsymbol { \xi } ) \boldsymbol { \theta } ] _ { \boldsymbol { \phi } } ( \mathbf { z } , \boldsymbol { s } , \boldsymbol { \xi } ; \boldsymbol { \theta } ) \Bigg ] } \\ & = 2 \sum _ { \boldsymbol { \phi } = 1 } ^ { \infty } ( \mathrm { E } _ { \boldsymbol { \phi } } [ \partial _ { \boldsymbol { \phi } } ( \mathbf { z } , \boldsymbol { s } , \boldsymbol { \xi } , \boldsymbol { \theta } ) ] - \mathrm { E } _ { \boldsymbol { \phi } , \boldsymbol { \xi } } [ \mathcal { P } _ { \boldsymbol { \phi } } ( \mathbf { z } , \boldsymbol { s } , \boldsymbol { \xi } ) ] \mathrm { E } _ { \boldsymbol { \phi } } [ \mathcal { P } _ { \boldsymbol { \phi } } ( \mathbf { z } , \boldsymbol { s } , \boldsymbol \ \end{array} +$$ + +Proposition 1. $\mathcal { R } _ { F D } ( \mathbf { z } ^ { t } ; \boldsymbol { \theta } ) \leq 4 \tilde { \mathcal { R } } _ { E L D } ( \mathbf { z } ^ { t } ; \boldsymbol { \theta } ) .$ . + +Proof. Let $\bar { s } : = \mathbb { E } _ { s } [ s ]$ , then + +$$ +\begin{array} { r l } & { \mathcal { R } _ { t } ( \mathbf { z } ^ { t } ) : = \mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \left[ \| \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \theta ) - \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { j } ^ { t } ; \theta ) \| _ { 2 } ^ { 2 } \right] } \\ & { \quad \quad \quad \quad = \mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \left[ \| \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \theta ) - \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , \bar { s } ; \theta ) + \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , \bar { s } ; \theta ) - \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { j } ^ { t } ; \theta ) \| _ { 2 } ^ { 2 } \right] } \\ & { \quad \quad \quad \quad = 4 \mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \left[ \| \frac { \mathbf { p } ^ { t } \left( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \theta \right) - \mathbf { p } ^ { t } \left( \mathbf { z } ^ { t } , \bar { s } ; \theta \right) } { 2 } + \frac { \mathbf { p } ^ { t } \left( \mathbf { z } ^ { t } , \bar { s } ; \theta \right) - \mathbf { p } ^ { t } \left( \mathbf { z } ^ { t } , s _ { j } ^ { t } ; \theta \right) } { 2 } \| _ { 2 } ^ { 2 } \right] . } \end{array} +$$ + +Then using Jensen’s inequality, + +$$ +\begin{array} { r l } & { \mathcal { R } _ { t } ( \mathbf { z } ^ { t } ) \leq 4 \mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \left[ \frac { 1 } { 2 } \| \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \theta ) - \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , \bar { s } ; \theta ) \| _ { 2 } ^ { 2 } + \frac { 1 } { 2 } \| \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , \bar { s } ; \theta ) - \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { j } ^ { t } ; \theta ) \| _ { 2 } ^ { 2 } \right] } \\ & { \qquad ( 2 \pi ) } \\ & { \qquad = 4 \mathbb { E } _ { s _ { i } ^ { t } } \left[ \| \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , s _ { i } ^ { t } ; \theta ) - \mathbf { p } ^ { t } ( \mathbf { z } ^ { t } , \bar { s } ; \theta ) \| _ { 2 } ^ { 2 } \right] = 4 \tilde { \mathcal { R } } _ { E L D } ( \mathbf { z } ^ { t } ; \theta ) . } \end{array} +$$ \ No newline at end of file diff --git a/parse/train/SJyVzQ-C-/SJyVzQ-C-_content_list.json b/parse/train/SJyVzQ-C-/SJyVzQ-C-_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e768a419021ae1e27d64cda821f050d740c51e46 --- /dev/null +++ b/parse/train/SJyVzQ-C-/SJyVzQ-C-_content_list.json @@ -0,0 +1,1745 @@ +[ + { + "type": "text", + "text": "FRATERNAL DROPOUT ", + "text_level": 1, + "bbox": [ + 176, + 99, + 444, + 121 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Konrad $\\dot { \\mathbf Z } \\mathbf { o } \\mathbf { f n } \\mathbf { a } ^ { 1 , 2 , * } ;$ Devansh Arpit2, Dendi Suhubdy2 & Yoshua Bengio2,3 ", + "bbox": [ + 186, + 143, + 694, + 160 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Jagiellonian University \n2MILA, Université de Montréal \n3CIFAR Senior Fellow ", + "bbox": [ + 183, + 172, + 393, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 253, + 544, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recurrent neural networks (RNNs) form an important class of architectures among neural networks useful for language modeling and sequential prediction. However, optimizing RNNs is known to be harder compared to feed-forward neural networks. A number of techniques have been proposed in literature to address this problem. In this paper we propose a simple technique called fraternal dropout that takes advantage of dropout to achieve this goal. Specifically, we propose to train two identical copies of an RNN (that share parameters) with different dropout masks while minimizing the difference between their (pre-softmax) predictions. In this way our regularization encourages the representations of RNNs to be invariant to dropout mask, thus being robust. We show that our regularization term is upper bounded by the expectation-linear dropout objective which has been shown to address the gap due to the difference between the train and inference phases of dropout. We evaluate our model and achieve state-of-the-art results in sequence modeling tasks on two benchmark datasets – Penn Treebank and Wikitext-2. We also show that our approach leads to performance improvement by a significant margin in image captioning (Microsoft COCO) and semi-supervised (CIFAR-10) tasks. ", + "bbox": [ + 233, + 287, + 764, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 559, + 336, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recurrent neural networks (RNNs) like long short-term memory (LSTM; Hochreiter & Schmidhuber (1997)) networks and gated recurrent unit (GRU; Chung et al. (2014)) are popular architectures for sequence modeling tasks like language generation, translation, speech synthesis, and machine comprehension. However, they are harder to optimize compared to feed-forward networks due to challenges like variable length input sequences, repeated application of the same transition operator at each time step, and largely-dense embedding matrix that depends on the vocabulary size. Due to these optimization challenges in RNNs, the application of batch normalization and its variants (layer normalization, recurrent batch normalization, recurrent normalization propagation) have not been as successful as their counterparts in feed-forward networks (Laurent et al., 2016), although they do considerably provide performance gains. Similarly, naive application of dropout (Srivastava et al., 2014) has been shown to be ineffective in RNNs (Zaremba et al., 2014). Therefore, regularization techniques for RNNs is an active area of research. ", + "bbox": [ + 174, + 593, + 825, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, Zaremba et al. (2014) proposed to apply dropout only to the nonrecurrent connections in multi-layer RNNs. Variational dropout (Gal & Ghahramani (2016)) uses the same dropout mask throughout a sequence during training. DropConnect (Wan et al., 2013) applies the dropout operation on the weight matrices. Zoneout (Krueger et al. (2016)), in a similar spirit with dropout, randomly chooses to use the previous time step hidden state instead of using the current one. Similarly as a substitute for batch normalization, layer normalization normalizes the hidden units within each sample to have zero mean and unit standard deviation. Recurrent batch normalization applies batch normalization but with unshared mini-batch statistics for each time step (Cooijmans et al., 2016). ", + "bbox": [ + 174, + 767, + 825, + 891 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Merity et al. (2017a) and Merity et al. (2017b) on the other hand show that activity regularization (AR) and temporal activation regularization (TAR)1 are also effective methods for regularizing LSTMs. Another more recent way of regularizing RNNs, that is similar in spirit to the approach we take, involves minimizing the difference between the hidden states of the original and the auxiliary network Serdyuk et al. (2017). ", + "bbox": [ + 174, + 103, + 823, + 172 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper we propose a simple regularization based on dropout that we call fraternal dropout, where we minimize an equally weighted sum of prediction losses from two identical copies of the same LSTM with different dropout masks, and add as a regularization the $\\ell ^ { 2 }$ difference between the predictions (pre-softmax) of the two networks. We analytically show that our regularization objective is equivalent to minimizing the variance in predictions from different i.i.d. dropout masks; thus encouraging the predictions to be invariant to dropout masks. We also discuss how our regularization is related to expectation linear dropout Ma et al. (2016), Π-model Laine & Aila (2016) and activity regularization Merity et al. (2017b), and empirically show that our method provides non-trivial gains over these related methods which we explain furthermore in our ablation study (Section 5). ", + "bbox": [ + 174, + 180, + 825, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 FRATERNAL DROPOUT ", + "text_level": 1, + "bbox": [ + 176, + 332, + 390, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dropout is a powerful regularization for neural networks. It is usually more effective on densely connected layers because they suffer more from overfitting compared with convolution layers where the parameters are shared. For this reason dropout is an important regularization for RNNs. However, dropout has a gap between its training and inference phase since the latter phase assumes linear activations to correct for the factor by which the expected value of each activation would be different Ma et al. (2016). In addition, the prediction of models with dropout generally vary with different dropout mask. However, the desirable property in such cases would be to have final predictions be invariant to dropout masks. ", + "bbox": [ + 173, + 367, + 825, + 478 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As such, the idea behind fraternal dropout is to train a neural network model in a way that encourages the variance in predictions under different dropout masks to be as small as possible. Specifically, consider we have an RNN model denoted by $\\mathcal { M } ( \\boldsymbol { \\theta } )$ that takes as input $\\mathbf { X }$ , where $\\theta$ denotes the model parameters. Let $\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) \\in \\mathbb { R } ^ { m }$ be the prediction of the model for input sample $\\mathbf { X }$ at time $t$ , for dropout mask $s _ { i } ^ { t }$ and current input $\\mathbf { z } ^ { t }$ , where $\\mathbf { z } ^ { t }$ is a function of $\\mathbf { X }$ and the hidden states corresponding to the previous time steps. Similarly, let $\\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) , \\mathbf { Y } )$ be the corresponding $t ^ { t h }$ time step loss value for the overall input-target sample pair $( \\mathbf { X } , \\mathbf { Y } )$ . ", + "bbox": [ + 174, + 486, + 825, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Then in fraternal dropout, we simultaneously feed-forward the input sample $\\mathbf { X }$ through two identical copies of the RNN that share the same parameters $\\theta$ but with different dropout masks $s _ { i } ^ { t }$ and $\\boldsymbol { s } _ { j } ^ { t }$ at each time step $t$ . This yields two loss values at each time step $t$ given by $\\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) , \\mathbf { Y } )$ , and $\\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\bar { \\theta } ) , \\mathbf { Y } )$ . Then the overall loss function of fraternal dropout is given by, ", + "bbox": [ + 174, + 590, + 825, + 651 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/93bdccc72eea7a43b2752ee0bbada1f1fb8f1cc819ea811ccae19b60a4d61061.jpg", + "text": "$$\n\\ell _ { F D } ( \\mathbf { X } , \\mathbf { Y } ) = \\sum _ { t = 1 } ^ { T } \\frac { 1 } { 2 } \\left( \\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) , \\mathbf { Y } ) + \\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) , \\mathbf { Y } ) \\right) + \\frac { \\kappa } { m T } \\sum _ { t = 1 } ^ { T } \\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\theta )\n$$", + "text_format": "latex", + "bbox": [ + 232, + 662, + 820, + 708 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\kappa$ is the regularization coefficient, $m$ is the dimensions of $\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } )$ and $\\mathcal { R } _ { F D } ( { \\bf z } ^ { t } ; \\boldsymbol { \\theta } )$ is the fraternal dropout regularization given by, ", + "bbox": [ + 171, + 734, + 826, + 765 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/19a7aa964ddb00f5896584ab9a2ee80ad80921cffd3df6a4d3833ec209159062.jpg", + "text": "$$\n\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) : = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) \\| _ { 2 } ^ { 2 } \\right] .\n$$", + "text_format": "latex", + "bbox": [ + 232, + 775, + 606, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We use Monte Carlo sampling to approximate $\\mathcal { R } _ { F D } ( { \\bf z } ^ { t } ; \\boldsymbol { \\theta } )$ where $\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } )$ and $\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta )$ are the same as the one used to calculate $\\ell ^ { t }$ values. Hence, the additional computation is negligible. ", + "bbox": [ + 174, + 809, + 825, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We note that the regularization term of our objective is equivalent to minimizing the variance in the prediction function with different dropout masks as shown below (proof in the appendix). ", + "bbox": [ + 176, + 847, + 825, + 876 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Remark 1. Let $s _ { i } ^ { t }$ and $\\boldsymbol { s } _ { j } ^ { t }$ be i.i.d. dropout masks and $\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) \\in \\mathbb { R } ^ { m }$ be the prediction function as described above. Then, ", + "bbox": [ + 169, + 102, + 825, + 132 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/0a168a922a4cdc23ffbcba643900e9fa7e60d2fa738443aed02c1cf592b69992.jpg", + "text": "$$\n\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\theta ) = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] = 2 \\sum _ { q = 1 } ^ { m } \\mathrm { v a r } _ { s _ { i } ^ { t } } ( p _ { q } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) ) .\n$$", + "text_format": "latex", + "bbox": [ + 232, + 136, + 781, + 179 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that a generalization of our approach would be to minimize the difference between the predictions of the two networks with different data/model augmentations. However, in this paper we focus on using different dropout masks and experiment mainly with $\\mathrm { R N N s } ^ { 2 }$ . ", + "bbox": [ + 174, + 191, + 825, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 253, + 346, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 RELATION TO EXPECTATION LINEAR DROPOUT (ELD) ", + "text_level": 1, + "bbox": [ + 174, + 285, + 594, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ma et al. (2016) analytically showed that the expected error (over samples) between a model’s expected prediction over all dropout masks, and the prediction using the average mask, is upper bounded. Based on this result, they propose to explicitly minimize the difference (we have adapted their regularization to our notations), ", + "bbox": [ + 174, + 311, + 825, + 367 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/d5b9503e2f3580f8faa094454aa7daa1f3502bb502f6109681ecc8d523451ea2.jpg", + "text": "$$\n\\mathcal { R } _ { E L D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) = \\| \\mathbb { E } _ { s } \\left[ \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s ; \\boldsymbol { \\theta } ) \\right] - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\mathbb { E } _ { s } [ s ] ; \\boldsymbol { \\theta } ) \\| _ { 2 }\n$$", + "text_format": "latex", + "bbox": [ + 232, + 372, + 594, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $s$ is the dropout mask. However, due to feasibility consideration, they instead propose to use the following regularization in practice, ", + "bbox": [ + 173, + 396, + 825, + 425 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/1e947ec086b9b0b95ed800567914b15a7c2e4666bcc4d4fde3424b3f5195e5ea.jpg", + "text": "$$\n\\begin{array} { r } { \\tilde { \\mathcal { R } } _ { E L D } ( \\mathbf { z } ^ { t } ; \\theta ) = \\mathbb { E } _ { s _ { i } } \\left[ \\Vert \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\mathbb { E } _ { s } [ s ] ; \\theta ) \\Vert _ { 2 } ^ { 2 } \\right] . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 230, + 431, + 612, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, this is achieved by feed-forwarding the input twice through the network, with and without dropout mask, and minimizing the main network loss (with dropout) along with the regularization term specified above (but without back-propagating the gradients through the network without dropout). The goal of Ma et al. (2016) is to minimize the network loss along with the expected difference between the prediction from individual dropout mask and the prediction from the expected dropout mask. We note that our regularization objective is upper bounded by the expectation-linear dropout regularization as shown below (proof in the appendix). ", + "bbox": [ + 173, + 455, + 825, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 1. $\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) \\leq 4 \\tilde { \\mathcal { R } } _ { E L D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) .$ . ", + "bbox": [ + 174, + 558, + 478, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This result shows that minimizing the ELD objective indirectly minimizes our regularization term. Finally as indicated above, they apply the target loss only on the network with dropout. In fact, in our own ablation studies (see Section 5) we find that back-propagating target loss through the network (without dropout) makes optimizing the model harder. However, in our setting, simultaneously backpropagating target loss through both networks yields both performance gain as well as convergence gain. We believe convergence is faster for our regularization because network weights are more likely to get target based updates from back-propagation in our case. This is especially true for weight dropout (Wan et al., 2013) since in this case dropped weights do not get updated in the training iteration. ", + "bbox": [ + 173, + 584, + 825, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 RELATION TO Π-MODEL ", + "text_level": 1, + "bbox": [ + 176, + 727, + 382, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Laine & Aila (2016) propose $\\Pi$ -model with the goal of improving performance on classification tasks in the semi-supervised setting. They propose a model similar to ours (considering the equivalent deep feed-forward version of our model) except they apply target loss only on one of the networks and use time-dependent weighting function $\\omega ( t )$ (while we use constant $\\frac { \\kappa } { m T }$ ). The intuition in their case is to leverage unlabeled data by using them to minimize the difference in prediction between the two copies of the network with different dropout masks. Further, they also test their model in the supervised setting but fail to explain the improvements they obtain by using this regularization. ", + "bbox": [ + 173, + 753, + 825, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We note that in our case we analytically show that minimizing our regularizer (also used in $\\Pi$ -model) is equivalent to minimizing the variance in the model predictions (Remark 1). Furthermore, we also show the relation of our regularizer to expectation linear dropout (Proposition 1). In Section 5, we study the effects of target based loss on both networks, which is not used in the $\\Pi$ -model. We find that applying target loss on both the networks leads to significantly faster convergence. Finally, we bring to attention that temporal embedding (another model proposed by Laine & Aila (2016), claimed to be a better version of $\\Pi$ -model for semi-supervised, learning) is intractable in natural language processing applications because storing averaged predictions over all of the time steps would be memory exhaustive (since predictions are usually huge - tens of thousands values). On a final note, we argue that in the supervised case, using a time-dependent weighting function $\\omega ( t )$ instead of a constant value $\\frac { \\kappa } { m T }$ is not needed. Since the ground truth labels are known, we have not observed the problem mentioned by Laine & Aila (2016), that the network gets stuck in a degenerate solution when $\\omega ( t )$ is too large in earlier epochs of training. We note that it is much easier to search for an optimal constant value, which is true in our case, as opposed to tuning the time-dependent function. ", + "bbox": [ + 176, + 858, + 825, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarity to $\\Pi$ -model makes our method related to other semi-supervised works, mainly Rasmus et al. (2015) and Sajjadi et al. (2016). Since semi-supervised learning is not a primary focus of this paper, we refer to Laine & Aila (2016) for more details. ", + "bbox": [ + 176, + 277, + 821, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another way to address the gap between the train and evaluation mode of dropout is to perform Monte Carlo sampling of masks and average the predictions during evaluation, and this has been used for feed-forward networks. We find that this technique does not work well for RNNs. The details of these experiments can be found in the appendix. ", + "bbox": [ + 174, + 327, + 825, + 382 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 402, + 326, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1 LANGUAGE MODELS ", + "text_level": 1, + "bbox": [ + 176, + 434, + 359, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the case of language modeling we test our model3 on two benchmark datasets – Penn Tree-bank (PTB) dataset (Marcus et al., 1993) and WikiText-2 (WT2) dataset (Merity et al., 2016). We use preprocessing as specified by Mikolov et al. (2010) (for PTB corpus) and Moses tokenizer Koehn et al. (2007) (for the WT2 dataset). ", + "bbox": [ + 174, + 459, + 823, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For both datasets we use the AWD-LSTM 3-layer architecture described in Merity et al. (2017a)4 which we call the baseline model. The number of parameters in the model used for PTB is 24 million as compared to 34 million in the case of WT2 because WT2 has a larger vocabulary size for which we use a larger embedding matrix. Apart from those differences, the architectures are identical. When we use fraternal dropout, we simply add our regularization on top of this baseline model. ", + "bbox": [ + 174, + 522, + 825, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Word level Penn Treebank (PTB) ", + "text_level": 1, + "bbox": [ + 176, + 613, + 406, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Influenced by Melis et al. (2017), our goal here is to make sure that fraternal dropout outperforms existing methods not simply because of extensive hyper-parameter grid search but rather due to its regularization effects. Hence, in our experiments we leave a vast majority of hyper-parameters used in the baseline model (Melis et al., 2017) unchanged i.e. embedding and hidden states sizes, gradient clipping value, weight decay and the values used for all dropout layers (dropout on the word vectors, the output between LSTM layers, the output of the final LSTM, and embedding dropout). However, a few changes are necessary: ", + "bbox": [ + 174, + 635, + 825, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "• the coefficients for AR and TAR needed to be altered because fraternal dropout also affects RNNs activation (as explained in Subsection 5.3) – we did not run grid search to obtain the best values but simply deactivated AR and TAR regularizers; • since fraternal dropout needs twice as much memory, batch size is halved so the model needs approximately the same amount of memory and hence fits on the same GPU. ", + "bbox": [ + 215, + 743, + 825, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final change in hyper-parameters is to alter the non-monotone interval $n$ used in nonmonotonically triggered averaged SGD (NT-ASGD) optimizer Polyak & Juditsky (1992); Mandt et al. (2017); Melis et al. (2017). We run a grid search on $n \\in \\{ 5 , 2 5 , 4 0 , 5 0 , 6 0 \\}$ and obtain very similar results for the largest values (40, 50 and 60) in the candidate set. Hence, our model is trained longer using ordinary SGD optimizer as compared to the baseline model (Melis et al., 2017). ", + "bbox": [ + 176, + 832, + 825, + 873 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/58fbc95757bda8d8b5297de81e2a587e22f8bcf476f9fcb6913b30e2a0ced2ed.jpg", + "table_caption": [ + "Table 1: Perplexity on Penn Treebank word level language modeling task. " + ], + "table_footnote": [], + "table_body": "
ModelParametersValidationTest
Zaremba et al. (2014) - LSTM (medium)10M86.282.7
Zaremba et al. (2014) - LSTM (large)24M82.278.4
Gal & Ghahramani (2016) - Variational LSTM (medium)20M81.979.7
Gal & Ghahramani (2016) - Variational LSTM (large)66M77.975.2
Inan et al. (2016) - Variational LSTM51M71.168.5
Inan et al. (2016) - Variational RHN24M68.166.0
Zilly et al. (2016) - Variational RHN23M67.965.4
Melis et al. (2017) - 5-layer RHN24M64.862.2
Melis et al. (2017) - 4-layer skip connection LSTM24M60.958.3
Merity etal. (2017a) - AWD-LSTM3-layer (baseline)24M60.057.3
Fraternal dropout + AWD-LSTM3-layer24M58.956.8
", + "bbox": [ + 181, + 98, + 818, + 268 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9f4bf8304f3c42f131fdef6f693f45615c4ab57d1fc38de924e0cfd7fb9a8e70.jpg", + "table_caption": [ + "Table 2: Perplexity on WikiText-2 word level language modeling task. " + ], + "table_footnote": [], + "table_body": "
ModelParametersValidationTest
Merity et al. (2016) - Variational LSTM + Zoneout20M108.7100.9
Merity et al. (2016) - Variational LSTM20M101.796.3
Inan et al. (2016) - Variational LSTM28M91.587.0
Melis et al. (2017) - 5-layer RHN24M78.175.6
Melis et al. (2017) - 1-layer LSTM24M69.365.9
Melis et al. (2017) - 2-layer skip connection LSTM24M69.165.9
Merity et al. (2017a) - AWD-LSTM3-layer (baseline)34M68.665.8
Fraternal dropout + AWD-LSTM 3-layer34M66.864.1
", + "bbox": [ + 186, + 297, + 812, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 473, + 823, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our model using the perplexity metric and compare the results that we obtain against the existing state-of-the-art results. The results are reported in Table 1. Our approach achieves the state-of-the-art performance compared with existing benchmarks. ", + "bbox": [ + 174, + 508, + 825, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To confirm that the gains are robust to initialization, we run ten experiments for the baseline model with different seeds (without fine-tuning) for PTB dataset to compute confidence intervals. The average best validation perplexity is $6 0 . 6 4 \\pm 0 . 1 5$ with the minimum value equals 60.33. The same for test perplexity is $5 8 . 3 2 \\pm 0 . 1 4$ and 58.05, respectively. Our score (59.8 validation and 58.0 test perplexity) beats ordinal dropout minimum values. ", + "bbox": [ + 174, + 558, + 825, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also perform experiments using fraternal dropout with a grid search on all the hyper-parameters and find that it leads to further improvements in performance. The details of this experiment can be found in section 5.5. ", + "bbox": [ + 176, + 633, + 825, + 675 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Word level WikiText-2 (WT2) ", + "text_level": 1, + "bbox": [ + 176, + 683, + 383, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the case of WikiText-2 language modeling task, we outperform the current state-of-the-art using the perplexity metric by a significant margin. Due to the lack of computational power, we run a single training procedure for fraternal dropout on WT2 dataset because it is larger than PTB. In this experiment, we use the best hyper-parameters found for PTB dataset $\\kappa = 0 . 1$ , non-monotone interval $n = 6 0$ and halved batch size; the rest of the hyper-parameters are the same as described in Melis et al. (2017) for WT2). The final results are presented in Table 2. ", + "bbox": [ + 174, + 704, + 825, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2 IMAGE CAPTIONING ", + "text_level": 1, + "bbox": [ + 176, + 805, + 354, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also apply fraternal dropout on an image captioning task. We use the well-known show and tell model as a baseline5 (Vinyals et al., 2014). We emphasize that in the image captioning task, the image encoder and sentence decoder architectures are usually learned together. Since we want to focus on the benefits of using fraternal dropout in RNNs we use frozen pretrained ResNet-101 (He et al., 2015) model as our image encoder. It means that our results are not directly comparable with other state-of-the-art methods, however we report results for the original methods so readers can see that our baseline performs well. The final results are presented in Table 3. ", + "bbox": [ + 174, + 832, + 825, + 887 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e8caa9152ecfc21710dfc1a2d30cc6455c9f4a6ee49165aa732df86c66312e73.jpg", + "table_caption": [ + "Table 3: BLEU scores for the Microsoft COCO image captioning task. Using fraternal dropout is the only difference between models. The rest of hyper-parameters are the same. " + ], + "table_footnote": [], + "table_body": "
ModelBLEU-1BLEU-2BLEU-3BLEU-4
Show and Tell Xu et al. (2015)66.646.132.924.6
Baseline68.850.836.125.6
Fraternal dropout, K = 0.01569.351.436.626.1
Fraternal dropout, K = 0.00569.351.536.926.3
", + "bbox": [ + 230, + 97, + 764, + 171 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 231, + 825, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We argue that in this task smaller $\\kappa$ values are optimal because the image captioning encoder is given all information in the beginning and hence the variance of consecutive predictions is smaller that in unconditioned natural language processing tasks. Fraternal dropout may benefits here mainly due to averaging gradients for different mask and hence updating weights more frequently. ", + "bbox": [ + 174, + 280, + 825, + 337 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 ABLATION STUDIES ", + "text_level": 1, + "bbox": [ + 176, + 358, + 372, + 373 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, the goal is to study existing methods closely related to ours – expectation linear dropout Ma et al. (2016), Π-model Laine & Aila (2016) and activity regularization Merity et al. (2017b). All of our experiments for ablation studies, which apply a single layer LSTM, use the same hyper-parameters and model architecture6 as Melis et al. (2017). ", + "bbox": [ + 173, + 390, + 825, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1 EXPECTATION-LINEAR DROPOUT (ELD) ", + "text_level": 1, + "bbox": [ + 176, + 464, + 493, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The relation with expectation-linear dropout Ma et al. (2016) has been discussed in Section 2. Here we perform experiments to study the difference in performance when using the ELD regularization versus our regularization (FD). In addition to ELD, we also study a modification (ELDM) of ELD which applies target loss to both copies of LSTMs in ELD similar to FD (notice in their case they only have dropout on one LSTM). Finally we also evaluate a baseline model without any of these regularizations. The learning dynamics curves are shown in Figure 1. Our regularization performs better in terms of convergence compared with other methods. In terms of generalization, we find that FD is similar to ELD, but baseline and ELDM are much worse. Interestingly, looking at the train and validation curves together, ELDM seems to be suffering from optimization problems. ", + "bbox": [ + 174, + 489, + 825, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2 Π-MODEL ", + "text_level": 1, + "bbox": [ + 176, + 633, + 282, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since Π-model Laine & Aila (2016) is similar to our algorithm (even though it is designed for semi-supervised learning in feed-forward networks), we study the difference in performance with Π-model7 both qualitatively and quantitatively to establish the advantage of our approach. First, we run both single layer LSTM and 3-layer AWD-LSTM on PTB task to check how their model compares with ours in the case of language modeling. The results are shown in Figure 1 and 2. We find that our model converges significantly faster than Π-model. We believe this happens because we back-propagate the target loss through both networks (in contrast to Π-model) that leads to weights getting updated using target-based gradients more often. ", + "bbox": [ + 173, + 660, + 825, + 771 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b8d4833a24b6748ef3cd0f76c9fe0025e06c99555d7e766f2a29c044f96997c.jpg", + "image_caption": [ + "Figure 1: Ablation study: Train (left) and validation (right) perplexity on PTB word level modeling with single layer LSTM (10M parameters). These curves study the learning dynamics of the baseline model, Π-model, Expectation-linear dropout (ELD), Expectation-linear dropout with modification (ELDM) and fraternal dropout (FD, our algorithm). We find that FD converges faster than the regularizers in comparison, and generalizes at par. " + ], + "image_footnote": [], + "bbox": [ + 186, + 109, + 457, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d64e450f08346c6bc468f12061b79ab092e7cc2cf8930eb3af1ac794100c063b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 109, + 795, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87020ced434d3499701675350c507f939c1046c58cb17698501cdf12b3ae97a7.jpg", + "image_caption": [ + "Figure 2: Ablation study: Validation perplexity on PTB word level modeling for $\\Pi$ -model and fraternal dropout. We find that FD converges faster and generalizes at par. " + ], + "image_footnote": [], + "bbox": [ + 186, + 347, + 457, + 486 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/463bd2f8083be91f8df7ca8a7e5292781fe18c685b3e04efd116713fc227bc09.jpg", + "image_caption": [ + "Figure 3: Ablation study: Average hidden state activation is reduced when any of the regularizer described is used. The y-axis is the value of $\\frac { 1 } { d } \\| \\mathbf { m } \\cdot \\mathbf { h } ^ { t } \\| _ { 2 } ^ { 2 }$ . " + ], + "image_footnote": [], + "bbox": [ + 514, + 347, + 789, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Even though we designed our algorithm specifically to address problems in RNNs, to have a fair comparison, we compare with $\\Pi$ -model on a semi-supervised task which is their goal. Specifically, we use the CIFAR-10 dataset that consists of $3 2 \\times 3 2$ images from 10 classes. Following the usual splits used in semi-supervised learning literature, we use 4 thousand labeled and 41 thousand unlabeled samples for training, 5 thousand labeled samples for validation and 10 thousand labeled samples for test set. We use the original ResNet-56 (He et al., 2015) architecture. We run grid search on $\\kappa \\in \\{ 0 . 0 5 , 0 . 1 , 0 . 1 5 , 0 . 2 \\}$ , dropout rates in $\\{ 0 . 0 5 , 0 . 1 , 0 . 1 5 , 0 . 2 \\}$ and leave the rest of the hyperparameters unchanged. We additionally check importance of using unlabeled data. The results are reported in Table 4. We find that our algorithm performs at par with $\\Pi$ -model. When unlabeled data is not used, fraternal dropout provides slightly better results as compared to traditional dropout. ", + "bbox": [ + 173, + 575, + 825, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3 ACTIVITY REGULARIZATION AND TEMPORAL ACTIVITY REGULARIZATION ANALYSIS ", + "text_level": 1, + "bbox": [ + 173, + 733, + 800, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The authors of Merity et al. (2017b) study the importance of activity regularization $( \\mathsf { A R } ) ^ { 8 }$ and temporal activity regularization (TAR) in LSTMs given as, ", + "bbox": [ + 174, + 757, + 820, + 786 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/b38dd65a7f67fddfa81341093884dcd3c38ba12153ff03f3da3f09bec8e564e5.jpg", + "text": "$$\n\\begin{array} { c } { \\mathcal { R } _ { A R } ( { \\bf z } ^ { t } ; \\theta ) = \\displaystyle \\frac { \\alpha } { d } \\| { \\bf h } ^ { t } \\| _ { 2 } ^ { 2 } } \\\\ { \\mathcal { R } _ { T A R } ( { \\bf z } ^ { t } ; \\theta ) = \\displaystyle \\frac { \\beta } { d } \\| { \\bf h } ^ { t } - { \\bf h } ^ { t - 1 } \\| _ { 2 } ^ { 2 } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 232, + 792, + 449, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf { h } ^ { t } \\in \\mathbb { R } ^ { d }$ is the LSTM’s output activation at time step $t$ (hence depends on both current input $\\mathbf { z } ^ { t }$ and the model parameters $\\theta$ ). Notice that AR and TAR regularizations are applied on the output of ", + "bbox": [ + 174, + 859, + 825, + 887 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2aeea17ebcb0188853ad5bb84b6d7253418d9774807341160e5fb918f328b85d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelDropout rateUnlabeled dataValidationTest
Traditional dropout0.1No78.4 (± 0.25)76.9 (± 0.31)
No dropout0.0No78.8 (± 0.59)77.1 (± 0.3)
Fraternal dropout (k = 0.05)0.05No79.3 (± 0.38)77.6 (± 0.35)
Traditional dropout + I-model0.1Yes80.2 (± 0.33)78.5 (± 0.46)
Fraternal dropout (k = 0.15)0.1Yes80.5 (± 0.18)79.1 (± 0.37)
", + "bbox": [ + 173, + 97, + 849, + 198 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4: Ablation study: Accuracy on altered (semi-supervised) CIFAR-10 dataset for ResNet-56 based models. We find that our algorithm performs at par with $\\Pi$ -model. When unlabeled data is not used traditional dropout hurts performance while fraternal dropout provides slightly better results. It means that our methods may be beneficial when we lack data and have to use additional regularizing methods. ", + "bbox": [ + 173, + 202, + 825, + 272 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/45583947707d0455d79d12a5db380878e6174220941656e29500ab8024936b0e.jpg", + "image_caption": [ + "Figure 4: Ablation study: Train (left) and validation (right) perplexity on PTB word level modeling with single layer LSTM (10M parameters). These curves study the learning dynamics of the baseline model, temporal activity regularization (TAR), prediction regularization (PR), activity regularization (AR) and fraternal dropout (FD, our algorithm). We find that FD both converges faster and generalizes better than the regularizers in comparison. " + ], + "image_footnote": [], + "bbox": [ + 186, + 295, + 795, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the LSTM, while our regularization is applied on the pre-softmax output $\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } )$ of the LSTM. However, since our regularization can be decomposed as ", + "bbox": [ + 169, + 540, + 821, + 569 + ], + "page_idx": 7 + }, + { + "type": "equation", + "img_path": "images/a9990d1706822c4d925d6c9a10839ee14f0260c58e403a40356f53bb93dc5647.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) = \\mathbb { E } _ { s _ { i } , s _ { j } } \\left[ | | \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) - \\mathbf { p } ^ { t } ( \\mathbf { h } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } \\right] } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) ^ { T } \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } + | | \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } - 2 \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) ^ { T } \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) \\Big ] } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad = \\mathbb { E } _ { s _ { i } , s _ { j } } \\left[ | | \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } + | | \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } - 2 \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) ^ { T } \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) \\right] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 230, + 575, + 812, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and encapsulates an $\\ell ^ { 2 }$ term along with the dot product term, we perform experiments to confirm that the gains in our approach is not due to the $\\bar { \\ell } ^ { 2 }$ regularization alone. A similar argument goes for the TAR objective. We run a grid search on $\\alpha \\in \\{ 1 , 2 , \\ldots , 1 2 \\}$ , $\\beta \\in \\{ 1 , 2 , \\ldots , 1 2 \\}$ , which include the hyper-parameters mentioned in Merity et al. (2017a). For our regularization, we use $\\kappa \\in \\{ 0 . 0 5 , 0 . 1 , \\ldots , 0 . 4 \\}$ . Furthermore, we also compare with a regularization (PR) that regularizes $\\| \\mathbf { p } ^ { t } ( \\bar { \\mathbf { z } } ^ { t } , s _ { i } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 }$ to further rule-out any gains only from $\\ell ^ { 2 }$ regularization. Based on this grid search, we pick the best model on the validation set for all the regularizations, and additionally report a baseline model without any of these four mentioned regularizations. The learning dynamics is shown in Figure 4. Our regularization performs better both in terms of convergence and generalization compared with other methods. Average hidden state activation is reduced when any of the regularizer described is applied (see Figure 3). ", + "bbox": [ + 173, + 641, + 825, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4 IMPROVEMENTS USING FINE-TUNING ", + "text_level": 1, + "bbox": [ + 178, + 814, + 472, + 827 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We confirm that models trained with fraternal dropout benefit from the NT-ASGD fine-tuning step (as also used in Merity et al. (2017a)). However, this is a very time-consuming practice and since different hyper-parameters may be used in this additional part of the learning procedure, the probability of obtaining better results due to the extensive grid search is higher. Hence, in our experiments we use the same fine-tuning procedure as implemented in the official repository (even fraternal dropout was not used). We present the importance of fine-tuning in Table 5. ", + "bbox": [ + 174, + 839, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cb441eff58ff23d6d3734a8c592089c38ee86cc0af61b90822ae152ec84ef3dc.jpg", + "table_caption": [ + "Table 5: Ablation study: Importance of fine-tuning for AWD-LSTM 3-layer model. Perplexity for the Penn Treebank and WikiText-2 language modeling tasks. " + ], + "table_footnote": [], + "table_body": "
PTBWT2
DropoutFine-tuningValidationTestValidationTest
TraditionalNone60.758.869.166.0
TraditionalOne60.057.368.665.8
FraternalNone59.858.068.365.3
FraternalOne58.956.866.864.1
FraternalTwo58.556.211
", + "bbox": [ + 259, + 98, + 736, + 200 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/81288747a210a7748b2bb2edb3d9e19287db61a5f4d0244a78b709167c48133c.jpg", + "table_caption": [ + "Table 6: Ablation study: Candidate hyper-parameters possible used in the grid search for comparing fraternal dropout and expectation linear dropout. $U ( a , b )$ is the uniform distribution on the interval $[ a , b ]$ . For finite sets, each value is drawn with equal probability. " + ], + "table_footnote": [], + "table_body": "
Hyper-parameterPossible values
batch size non-monotone interval[10,20,30, 40] [5,10,20,40,60,100]
K -FD or ELD strength weight decayU(0, 0.3) U(0.6 × 10-6,2.4 × 10-6)
", + "bbox": [ + 307, + 251, + 687, + 325 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/080b0198eaea0fe2ad695f2300f5f44e370afb3f693385f868df02e311bdb4d7.jpg", + "table_caption": [ + "Table 7: Ablation study: Fraternal dropout and expectation linear dropout comparison. Perplexity on the Penn Treebank validation dateset. Fraternal dropout is more robust to different hyperparameters choice as twice as much runs finished performing better than the baseline model (60.7). " + ], + "table_footnote": [], + "table_body": "
ModelBestTop5 avgTop10 avgBeating baseline runs (out of)
Expectation lineardropout59.460.160.56(208)
Fraternal dropout59.459.659.914 (203)
", + "bbox": [ + 176, + 393, + 820, + 439 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.5 FRATERNAL DROPOUT AND EXPECTATION LINEAR DROPOUT COMPARISON ", + "text_level": 1, + "bbox": [ + 174, + 521, + 725, + 534 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We perform extensive grid search for the baseline model from Subsection 4.1 (an AWD-LSTM 3-layer architecture) trained with either fraternal dropout or expectation linear dropout regularizations, to further contrast the performance of these two methods. The experiments are run without fine-tuning on the PTB dataset. ", + "bbox": [ + 174, + 549, + 825, + 606 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In each run, all five dropout rates are randomly altered (they are set to their original value, as in Merity et al. (2017a), multiplied by a value drawn from the uniform distribution on the interval [0.5, 1.5]) and the rest of the hyper-parameters are drawn as shown in Table 6. As in Subsection 4.1, AR and TAR regularizers are deactivated. ", + "bbox": [ + 174, + 613, + 825, + 667 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Together we run more than 400 experiments. The results are presented in Table 7. Both FD and ELD perform better than the baseline model that instead uses AR and TAR regularizers. Hence, we confirm our previous finding (see Subsection 5.3) that both FD and ELD are better. However, as found previously for smaller model in Subsection 5.1, the convergence of FD is faster than that of ELD. Additionally, fraternal dropout is more robust to different hyper-parameters choice (more runs performing better than the baseline and better average for top performing runs). ", + "bbox": [ + 174, + 675, + 825, + 760 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 CONCLUSION ", + "text_level": 1, + "bbox": [ + 174, + 789, + 318, + 805 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper we propose a simple regularization method for RNNs called fraternal dropout that acts as a regularization by reducing the variance in model predictions across different dropout masks. We show that our model achieves state-of-the-art results on benchmark language modeling tasks along with faster convergence. We also analytically study the relationship between our regularization and expectation linear dropout Ma et al. (2016). We perform a number of ablation studies to evaluate our model from different aspects and carefully compare it with related methods both qualitatively and quantitatively. ", + "bbox": [ + 173, + 825, + 825, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ACKNOWLEDGEMENTS ", + "text_level": 1, + "bbox": [ + 176, + 103, + 367, + 117 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The authors would like to acknowledge the support of the following agencies for research funding and computing support: NSERC, CIFAR, and IVADO. We would like to thank Rosemary Nan Ke and Philippe Lacaille for their thoughts and comments throughout the project. We would also like to thank Stanisław Jastrz˛ebski† and Evan Racah† for useful discussions. ", + "bbox": [ + 174, + 133, + 825, + 189 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 210, + 287, + 224 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. ", + "bbox": [ + 174, + 233, + 823, + 262 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tim Cooijmans, Nicolas Ballas, César Laurent, and Aaron C. Courville. Recurrent batch normalization. CoRR, abs/1603.09025, 2016. URL http://arxiv.org/abs/1603.09025. ", + "bbox": [ + 176, + 268, + 821, + 299 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yarin Gal and Zoubin Ghahramani. A theoretically grounded application of dropout in recurrent neural networks. In Advances in neural information processing systems, pp. 1019–1027, 2016. ", + "bbox": [ + 176, + 306, + 820, + 335 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CoRR, abs/1512.03385, 2015. URL http://arxiv.org/abs/1512.03385. ", + "bbox": [ + 176, + 343, + 820, + 372 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8): 1735–1780, 1997. ", + "bbox": [ + 174, + 381, + 823, + 409 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Hakan Inan, Khashayar Khosravi, and Richard Socher. Tying word vectors and word classifiers: A loss framework for language modeling. arXiv preprint arXiv:1611.01462, 2016. ", + "bbox": [ + 173, + 417, + 823, + 446 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, et al. Moses: Open source toolkit for statistical machine translation. In Proceedings of the 45th annual meeting of the ACL on interactive poster and demonstration sessions, pp. 177–180. Association for Computational Linguistics, 2007. ", + "bbox": [ + 174, + 454, + 825, + 525 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "David Krueger, Tegan Maharaj, János Kramár, Mohammad Pezeshki, Nicolas Ballas, Nan Rosemary Ke, Anirudh Goyal, Yoshua Bengio, Hugo Larochelle, Aaron Courville, et al. Zoneout: Regularizing rnns by randomly preserving hidden activations. arXiv preprint arXiv:1606.01305, 2016. ", + "bbox": [ + 174, + 534, + 825, + 589 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. ", + "bbox": [ + 173, + 598, + 823, + 627 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "César Laurent, Gabriel Pereyra, Philémon Brakel, Ying Zhang, and Yoshua Bengio. Batch normalized recurrent neural networks. In Acoustics, Speech and Signal Processing (ICASSP), 2016 IEEE International Conference on, pp. 2657–2661. IEEE, 2016. ", + "bbox": [ + 173, + 635, + 826, + 679 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Xuezhe Ma, Yingkai Gao, Zhiting Hu, Yaoliang Yu, Yuntian Deng, and Eduard Hovy. Dropout with expectation-linear regularization. arXiv preprint arXiv:1609.08017, 2016. ", + "bbox": [ + 174, + 686, + 823, + 715 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Stephan Mandt, Matthew D Hoffman, and David M Blei. Stochastic gradient descent as approximate bayesian inference. arXiv preprint arXiv:1704.04289, 2017. ", + "bbox": [ + 174, + 723, + 823, + 752 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated corpus of english: The penn treebank. Computational linguistics, 19(2):313–330, 1993. ", + "bbox": [ + 174, + 760, + 825, + 790 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Gábor Melis, Chris Dyer, and Phil Blunsom. On the state of the art of evaluation in neural language models. arXiv preprint arXiv:1707.05589, 2017. ", + "bbox": [ + 174, + 797, + 823, + 827 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. CoRR, abs/1609.07843, 2016. URL http://arxiv.org/abs/1609.07843. ", + "bbox": [ + 176, + 834, + 823, + 863 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Stephen Merity, Nitish Shirish Keskar, and Richard Socher. Regularizing and optimizing lstm language models. arXiv preprint arXiv:1708.02182, 2017a. ", + "bbox": [ + 174, + 872, + 823, + 901 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Stephen Merity, Bryan McCann, and Richard Socher. Revisiting activation regularization for language rnns. arXiv preprint arXiv:1708.01009, 2017b. ", + "bbox": [ + 173, + 103, + 823, + 132 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tomas Mikolov, Martin Karafiát, Lukas Burget, Jan Cernocky, and Sanjeev Khudanpur. Recurrent \\` neural network based language model. In Interspeech, volume 2, pp. 3, 2010. ", + "bbox": [ + 174, + 140, + 823, + 171 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Boris T Polyak and Anatoli B Juditsky. Acceleration of stochastic approximation by averaging. SIAM Journal on Control and Optimization, 30(4):838–855, 1992. ", + "bbox": [ + 171, + 179, + 821, + 208 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ofir Press and Lior Wolf. Using the output embedding to improve language models. arXiv preprint arXiv:1608.05859, 2016. ", + "bbox": [ + 173, + 215, + 823, + 246 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Antti Rasmus, Harri Valpola, Mikko Honkala, Mathias Berglund, and Tapani Raiko. Semisupervised learning with ladder network. CoRR, abs/1507.02672, 2015. URL http://arxiv. org/abs/1507.02672. ", + "bbox": [ + 173, + 253, + 823, + 296 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In Advances in Neural Information Processing Systems, pp. 1163–1171, 2016. ", + "bbox": [ + 174, + 306, + 821, + 349 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Dmitriy Serdyuk, Nan Rosemary Ke, Alessandro Sordoni, Adam Trischler, Chris Pal, and Yoshua Bengio. Twin networks: Matching the future for sequence generation. 2017. ", + "bbox": [ + 171, + 357, + 820, + 387 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. Journal of machine learning research, 15(1):1929–1958, 2014. ", + "bbox": [ + 173, + 395, + 823, + 438 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. CoRR, abs/1411.4555, 2014. URL http://arxiv.org/abs/1411. 4555. ", + "bbox": [ + 173, + 446, + 823, + 489 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Li Wan, Matthew Zeiler, Sixin Zhang, Yann L Cun, and Rob Fergus. Regularization of neural networks using dropconnect. In Proceedings of the 30th international conference on machine learning (ICML-13), pp. 1058–1066, 2013. ", + "bbox": [ + 173, + 497, + 823, + 541 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron C. Courville, Ruslan Salakhutdinov, Richard S. Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. CoRR, abs/1502.03044, 2015. URL http://arxiv.org/abs/1502. 03044. ", + "bbox": [ + 173, + 549, + 825, + 606 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Wojciech Zaremba, Ilya Sutskever, and Oriol Vinyals. Recurrent neural network regularization. arXiv preprint arXiv:1409.2329, 2014. ", + "bbox": [ + 169, + 614, + 823, + 645 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Julian Georg Zilly, Rupesh Kumar Srivastava, Jan Koutník, and Jürgen Schmidhuber. Recurrent highway networks. arXiv preprint arXiv:1607.03474, 2016. ", + "bbox": [ + 169, + 652, + 823, + 683 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "APPENDIX ", + "text_level": 1, + "bbox": [ + 176, + 103, + 263, + 117 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "MONTE CARLO EVALUATION ", + "text_level": 1, + "bbox": [ + 176, + 133, + 377, + 147 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A well known way to address the gap between the train and evaluation mode of dropout is to perform Monte Carlo sampling of masks and average the predictions during evaluation (MC-eval), and this has been used for feed-forward networks. Since fraternal dropout addresses the same problem, we would like to clarify that it is not straight-forward and feasible to apply MC-eval for RNNs. In feed-forward networks, we average the output prediction scores from different masks. However, in the case RNNs (for next step predictions), there is more than one way to perform such evaluation, but each one is problematic. They are as follows: ", + "bbox": [ + 174, + 160, + 825, + 257 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "1. Online averaging ", + "text_level": 1, + "bbox": [ + 176, + 263, + 313, + 279 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Consider that we first make the prediction at time step 1 using different masks by averaging the prediction score. Then we use this output to feed as input to the time step 2, then use different masks at time step 2 to generate the output at time step 2, and so on. But in order to do so, because of the way RNNs work, we also need to feed the previous time hidden state to time step 2. One way would be to average the hidden states over different masks at time step 1. But the hidden space can in general be highly nonlinear, and it is not clear if averaging in this space is a good strategy. This approach is not justified. ", + "bbox": [ + 174, + 285, + 825, + 382 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Besides, this strategy as a whole is extremely time consuming because we would need to sequentially make predictions with multiple masks at each time step. ", + "bbox": [ + 174, + 388, + 821, + 417 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "2. Sequence averaging ", + "text_level": 1, + "bbox": [ + 174, + 425, + 331, + 439 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Let’s consider that we use a different mask each time we want to generate a sequence, and then we average the prediction scores, and compute the argmax (at each time step) to get the actual generated sequence. ", + "bbox": [ + 174, + 446, + 825, + 488 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this case, notice it is not guaranteed that the predicted word at time step $t$ due to averaging the predictions would lead to the next word (generated by the same process) if we were to feed the time step $t$ output as input to the time step $t + 1$ . For example, with different dropout masks, if the probability of $\\bar { 1 } ^ { s t }$ time step outputs are: $1 4 0 \\%$ ), he $( 3 0 \\% )$ , she $( 3 0 \\% )$ , and the probability of the 2nd time step outputs are: am $( 3 0 \\% )$ , is $( 6 0 \\% )$ , was $( 1 0 \\% )$ . Then the averaged prediction score followed by argmax will result in the prediction “I is”, but this would be incorrect. A similar concern applies for output predictions varying in temporal length. ", + "bbox": [ + 174, + 494, + 825, + 593 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Hence, this approach can not be used to generate a sequence (it has to be done by by sampling a mask and generating a single sequence). However, this approach may be used to estimate the probability assigned by the model to a given sequence. ", + "bbox": [ + 174, + 599, + 825, + 642 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Nonetheless, we run experiments on the PTB dataset using MC-eval (the results are summarized in Table 8). We start with a simple comparison that compares fraternal dropout with the averaged mask and the AWD-LSTM 3-layer baseline with a single fixed mask that we call MC1. The MC1 model performs much worse than fraternal dropout. Hence, it would be hard to use MC1 model in practice because a single sample is inaccurate. We also check MC-eval for a larger number of models (MC50) (50 models were used since we were not able to fit more models simultaneously on a single GPU). The final results for MC50 are worse than the baseline which uses the averaged mask. For comparison, we also evaluate MC10. Note that no fine-tuning is used for the above experiments. ", + "bbox": [ + 173, + 648, + 825, + 761 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/910137973af0702bba6415fa366c5234d0c42e20a24f609d5bd82422af5dbf96.jpg", + "table_caption": [ + "Table 8: Appendix: Monte Carlo evaluation. Perplexity on Penn Treebank word level language modeling task using Monte Carlo sampling, fraternal dropout or average mask. " + ], + "table_footnote": [], + "table_body": "
ModelValidationTest
MC1 MC10 MC50 Baseline (average mask) Fraternal dropout92.2 (± 0.5) 66.2 (± 0.2) 64.4 (± 0.1) 60.7 59.889.2 (± 0.5) 63.7 (± 0.2) 62.1 (± 0.1) 58.8 58.0
", + "bbox": [ + 308, + 770, + 691, + 871 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "REASONS FOR FOCUSING ON RNNS ", + "text_level": 1, + "bbox": [ + 176, + 103, + 423, + 117 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The fraternal dropout method is general and may be applied in feed-forward architectures (as shown in Subsection 5.2 for CIFAR-10 semisupervised example). However, we believe that it is more powerful in the case of RNNs because: ", + "bbox": [ + 176, + 128, + 823, + 171 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "1. Variance in prediction accumulates among time steps in RNNs and since we share parameters for all time steps, one may use the same $\\kappa$ value at each step. In feed-forward networks the layers usually do not share parameters and hence one may want to use different $\\kappa$ values for different layers (which may be hard to tune). The simple way to alleviate this problem is to apply the regularization term on the pre-softmax predictions only (as shown in the paper) or use the same $\\kappa$ value for all layers. However, we believe that it may limit possible gains. \n2. The best performing RNN architectures (state-of-the-art) usually use some kind of dropout (embedding dropout, word dropout, weight dropout etc.), very often with high dropout rates (even larger than $50 \\%$ for input word embedding in NLP tasks). However, this is not true for feed-forward networks. For instance, ResNet architectures very often do not use dropout at all (probably because batch normalization is often better to use). It can be seen in the paper (Subsection 5.2, semisupervised CIFAR-10 task) that when unlabeled data is not used the regular dropout hurts performance and using fraternal dropout seems to improve just a little. \n3. On a final note, the Monte Carlo sampling (a well known method that adresses the gap betweem the train and evaluation mode of dropout) can not be easily applied for RNNs and fraternal dropout may be seen as an alternative. ", + "bbox": [ + 210, + 184, + 825, + 443 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To conclude, we believe that when the use of dropout benefits in a given architecture, applying fraternal dropout should improve performance even more. ", + "bbox": [ + 173, + 455, + 823, + 483 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As mentioned before, in image recognition tasks, one may experiment with something what we would temporarily dub fraternal augmentation (even though dropout is not used, one can use random data augmentation such as random crop or random flip). Hence, one may force a given neural network to have the same predictions for different augmentations. ", + "bbox": [ + 174, + 491, + 825, + 546 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "PROOFS ", + "bbox": [ + 174, + 103, + 232, + 117 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Remark 1. Let $s _ { i } ^ { t }$ and $\\boldsymbol { s } _ { j } ^ { t }$ be i.i.d. dropout masks and $\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) \\in \\mathbb { R } ^ { m }$ be the prediction function as described above. Then, ", + "bbox": [ + 169, + 128, + 825, + 159 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/19b899377283a2897e4e2f965b3f5c99c16188c1df7a49acc53da71265300747.jpg", + "text": "$$\n\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\theta ) = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] = 2 \\sum _ { q = 1 } ^ { m } \\mathrm { v a r } _ { s _ { i } ^ { t } } ( p _ { q } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) ) .\n$$", + "text_format": "latex", + "bbox": [ + 232, + 162, + 781, + 205 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Proof. For simplicity of notation, we omit the time index $t$ ", + "bbox": [ + 173, + 219, + 563, + 234 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/ee9b4078e855855dd4a7060bf6140cdc2c1570eb2de4814747098ead92f77b38.jpg", + "text": "$$\n\\begin{array} { r l } { \\mathcal { R } _ { F S } [ \\mathbf { z } , \\boldsymbol { \\theta } ] } & { = \\mathrm { E } _ { \\boldsymbol { \\phi } , \\boldsymbol { \\xi } } [ \\| \\mathbf { P } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } , \\boldsymbol { \\theta } ) - \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } \\| ^ { 2 } ] } \\\\ & { = \\mathrm { E } _ { \\boldsymbol { \\phi } , \\boldsymbol { \\xi } } \\Bigg [ \\| \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\| _ { \\mathbf { z } } ^ { 2 } + \\mathrm { E } _ { \\boldsymbol { \\xi } , \\boldsymbol { \\xi } } \\Bigg [ \\| \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\| _ { \\mathbf { z } } ^ { 2 } \\Bigg ] } \\\\ & { - 2 \\mathrm { E } _ { \\boldsymbol { \\xi } , \\boldsymbol { s } , \\boldsymbol { \\xi } } [ \\| \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } \\| ^ { 2 } \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } ] } \\\\ & { = 2 \\sum _ { \\boldsymbol { \\phi } = 1 } ^ { \\infty } ( \\mathrm { E } _ { \\boldsymbol { \\phi } } [ \\partial _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } ] ^ { 2 } ) - \\mathrm { E } _ { \\boldsymbol { \\phi } , \\boldsymbol { \\xi } } \\Bigg [ \\mathcal { P } _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } ] _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ; \\boldsymbol { \\theta } ) \\Bigg ] } \\\\ & = 2 \\sum _ { \\boldsymbol { \\phi } = 1 } ^ { \\infty } ( \\mathrm { E } _ { \\boldsymbol { \\phi } } [ \\partial _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } , \\boldsymbol { \\theta } ) ] - \\mathrm { E } _ { \\boldsymbol { \\phi } , \\boldsymbol { \\xi } } [ \\mathcal { P } _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) ] \\mathrm { E } _ { \\boldsymbol { \\phi } } [ \\mathcal { P } _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol \\ \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 232, + 239, + 733, + 476 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Proposition 1. $\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) \\leq 4 \\tilde { \\mathcal { R } } _ { E L D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) .$ . ", + "bbox": [ + 174, + 501, + 478, + 518 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Proof. Let $\\bar { s } : = \\mathbb { E } _ { s } [ s ]$ , then ", + "bbox": [ + 173, + 531, + 356, + 547 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/c56e7b117430a0f08c020211f1d54d6de45a191dab907f9d4cb05bab8c3cfa93.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathcal { R } _ { t } ( \\mathbf { z } ^ { t } ) : = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] } \\\\ & { \\quad \\quad \\quad \\quad = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) + \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] } \\\\ & { \\quad \\quad \\quad \\quad = 4 \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\frac { \\mathbf { p } ^ { t } \\left( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta \\right) - \\mathbf { p } ^ { t } \\left( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta \\right) } { 2 } + \\frac { \\mathbf { p } ^ { t } \\left( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta \\right) - \\mathbf { p } ^ { t } \\left( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta \\right) } { 2 } \\| _ { 2 } ^ { 2 } \\right] . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 230, + 553, + 776, + 637 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Then using Jensen’s inequality, ", + "bbox": [ + 174, + 641, + 380, + 656 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/24b9c68592299ce9300b633f9611f2195b71342c78020834a7e086496852e274.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathcal { R } _ { t } ( \\mathbf { z } ^ { t } ) \\leq 4 \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\frac { 1 } { 2 } \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) \\| _ { 2 } ^ { 2 } + \\frac { 1 } { 2 } \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] } \\\\ & { \\qquad ( 2 \\pi ) } \\\\ & { \\qquad = 4 \\mathbb { E } _ { s _ { i } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] = 4 \\tilde { \\mathcal { R } } _ { E L D } ( \\mathbf { z } ^ { t } ; \\theta ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 230, + 662, + 807, + 736 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/parse/train/SJyVzQ-C-/SJyVzQ-C-_middle.json b/parse/train/SJyVzQ-C-/SJyVzQ-C-_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..cd0e43b1dac70ab8ba56e9d2b24c4bbeb0f5ffcd --- /dev/null +++ b/parse/train/SJyVzQ-C-/SJyVzQ-C-_middle.json @@ -0,0 +1,34596 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 79, + 272, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 275, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 275, + 98 + ], + "score": 1.0, + "content": "FRATERNAL DROPOUT", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 114, + 114, + 425, + 127 + ], + "lines": [ + { + "bbox": [ + 114, + 112, + 428, + 129 + ], + "spans": [ + { + "bbox": [ + 114, + 112, + 150, + 129 + ], + "score": 1.0, + "content": "Konrad", + "type": "text" + }, + { + "bbox": [ + 150, + 113, + 193, + 126 + ], + "score": 0.33, + "content": "\\dot { \\mathbf Z } \\mathbf { o } \\mathbf { f n } \\mathbf { a } ^ { 1 , 2 , * } ;", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 112, + 428, + 129 + ], + "score": 1.0, + "content": "Devansh Arpit2, Dendi Suhubdy2 & Yoshua Bengio2,3", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 112, + 137, + 241, + 171 + ], + "lines": [ + { + "bbox": [ + 112, + 136, + 212, + 150 + ], + "spans": [ + { + "bbox": [ + 112, + 136, + 212, + 150 + ], + "score": 1.0, + "content": "1Jagiellonian University", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 147, + 243, + 162 + ], + "spans": [ + { + "bbox": [ + 111, + 147, + 243, + 162 + ], + "score": 1.0, + "content": "2MILA, Université de Montréal", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 158, + 207, + 173 + ], + "spans": [ + { + "bbox": [ + 111, + 158, + 207, + 173 + ], + "score": 1.0, + "content": "3CIFAR Senior Fellow", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 278, + 201, + 333, + 213 + ], + "lines": [ + { + "bbox": [ + 276, + 200, + 335, + 214 + ], + "spans": [ + { + "bbox": [ + 276, + 200, + 335, + 214 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 228, + 468, + 415 + ], + "lines": [ + { + "bbox": [ + 142, + 228, + 469, + 241 + ], + "spans": [ + { + "bbox": [ + 142, + 228, + 469, + 241 + ], + "score": 1.0, + "content": "Recurrent neural networks (RNNs) form an important class of architectures", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 240, + 469, + 252 + ], + "spans": [ + { + "bbox": [ + 141, + 240, + 469, + 252 + ], + "score": 1.0, + "content": "among neural networks useful for language modeling and sequential prediction.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 251, + 469, + 263 + ], + "spans": [ + { + "bbox": [ + 142, + 251, + 469, + 263 + ], + "score": 1.0, + "content": "However, optimizing RNNs is known to be harder compared to feed-forward neu-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 262, + 469, + 274 + ], + "spans": [ + { + "bbox": [ + 141, + 262, + 469, + 274 + ], + "score": 1.0, + "content": "ral networks. A number of techniques have been proposed in literature to address", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 272, + 470, + 285 + ], + "spans": [ + { + "bbox": [ + 141, + 272, + 470, + 285 + ], + "score": 1.0, + "content": "this problem. In this paper we propose a simple technique called fraternal dropout", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 283, + 470, + 297 + ], + "spans": [ + { + "bbox": [ + 141, + 283, + 470, + 297 + ], + "score": 1.0, + "content": "that takes advantage of dropout to achieve this goal. Specifically, we propose to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 294, + 469, + 307 + ], + "spans": [ + { + "bbox": [ + 141, + 294, + 469, + 307 + ], + "score": 1.0, + "content": "train two identical copies of an RNN (that share parameters) with different dropout", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 305, + 469, + 317 + ], + "spans": [ + { + "bbox": [ + 141, + 305, + 469, + 317 + ], + "score": 1.0, + "content": "masks while minimizing the difference between their (pre-softmax) predictions.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 316, + 469, + 329 + ], + "spans": [ + { + "bbox": [ + 141, + 316, + 469, + 329 + ], + "score": 1.0, + "content": "In this way our regularization encourages the representations of RNNs to be in-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 327, + 469, + 339 + ], + "spans": [ + { + "bbox": [ + 141, + 327, + 469, + 339 + ], + "score": 1.0, + "content": "variant to dropout mask, thus being robust. We show that our regularization term is", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 338, + 470, + 350 + ], + "spans": [ + { + "bbox": [ + 141, + 338, + 470, + 350 + ], + "score": 1.0, + "content": "upper bounded by the expectation-linear dropout objective which has been shown", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 349, + 470, + 362 + ], + "spans": [ + { + "bbox": [ + 141, + 349, + 470, + 362 + ], + "score": 1.0, + "content": "to address the gap due to the difference between the train and inference phases of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 142, + 360, + 470, + 372 + ], + "spans": [ + { + "bbox": [ + 142, + 360, + 470, + 372 + ], + "score": 1.0, + "content": "dropout. We evaluate our model and achieve state-of-the-art results in sequence", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 371, + 469, + 383 + ], + "spans": [ + { + "bbox": [ + 142, + 371, + 469, + 383 + ], + "score": 1.0, + "content": "modeling tasks on two benchmark datasets – Penn Treebank and Wikitext-2. We", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 142, + 383, + 470, + 394 + ], + "spans": [ + { + "bbox": [ + 142, + 383, + 470, + 394 + ], + "score": 1.0, + "content": "also show that our approach leads to performance improvement by a significant", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 393, + 469, + 405 + ], + "spans": [ + { + "bbox": [ + 141, + 393, + 469, + 405 + ], + "score": 1.0, + "content": "margin in image captioning (Microsoft COCO) and semi-supervised (CIFAR-10)", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 405, + 167, + 415 + ], + "spans": [ + { + "bbox": [ + 142, + 405, + 167, + 415 + ], + "score": 1.0, + "content": "tasks.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 14 + }, + { + "type": "title", + "bbox": [ + 108, + 443, + 206, + 455 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 208, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 208, + 459 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 470, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 483 + ], + "score": 1.0, + "content": "Recurrent neural networks (RNNs) like long short-term memory (LSTM; Hochreiter & Schmidhu-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 482, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 494 + ], + "score": 1.0, + "content": "ber (1997)) networks and gated recurrent unit (GRU; Chung et al. (2014)) are popular architectures", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 493, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 505 + ], + "score": 1.0, + "content": "for sequence modeling tasks like language generation, translation, speech synthesis, and machine", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "score": 1.0, + "content": "comprehension. However, they are harder to optimize compared to feed-forward networks due to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 514, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 528 + ], + "score": 1.0, + "content": "challenges like variable length input sequences, repeated application of the same transition operator", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 526, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 506, + 538 + ], + "score": 1.0, + "content": "at each time step, and largely-dense embedding matrix that depends on the vocabulary size. Due to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "these optimization challenges in RNNs, the application of batch normalization and its variants (layer", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 548, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 506, + 560 + ], + "score": 1.0, + "content": "normalization, recurrent batch normalization, recurrent normalization propagation) have not been as", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 558, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 571 + ], + "score": 1.0, + "content": "successful as their counterparts in feed-forward networks (Laurent et al., 2016), although they do", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "considerably provide performance gains. Similarly, naive application of dropout (Srivastava et al.,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "score": 1.0, + "content": "2014) has been shown to be ineffective in RNNs (Zaremba et al., 2014). Therefore, regularization", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 591, + 308, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 308, + 604 + ], + "score": 1.0, + "content": "techniques for RNNs is an active area of research.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 608, + 505, + 706 + ], + "lines": [ + { + "bbox": [ + 106, + 609, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 620 + ], + "score": 1.0, + "content": "To address these challenges, Zaremba et al. (2014) proposed to apply dropout only to the non-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "recurrent connections in multi-layer RNNs. Variational dropout (Gal & Ghahramani (2016)) uses", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "the same dropout mask throughout a sequence during training. DropConnect (Wan et al., 2013)", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 640, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 506, + 654 + ], + "score": 1.0, + "content": "applies the dropout operation on the weight matrices. Zoneout (Krueger et al. (2016)), in a similar", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 650, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 506, + 666 + ], + "score": 1.0, + "content": "spirit with dropout, randomly chooses to use the previous time step hidden state instead of using", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 663, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 505, + 675 + ], + "score": 1.0, + "content": "the current one. Similarly as a substitute for batch normalization, layer normalization normalizes", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "score": 1.0, + "content": "the hidden units within each sample to have zero mean and unit standard deviation. Recurrent batch", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 684, + 506, + 699 + ], + "spans": [ + { + "bbox": [ + 104, + 684, + 506, + 699 + ], + "score": 1.0, + "content": "normalization applies batch normalization but with unshared mini-batch statistics for each time step", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 695, + 208, + 708 + ], + "spans": [ + { + "bbox": [ + 105, + 695, + 208, + 708 + ], + "score": 1.0, + "content": "(Cooijmans et al., 2016).", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 40 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 722, + 243, + 731 + ], + "lines": [ + { + "bbox": [ + 119, + 720, + 244, + 733 + ], + "spans": [ + { + "bbox": [ + 119, + 720, + 244, + 733 + ], + "score": 1.0, + "content": "∗konrad.zolna@gmail.com", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 79, + 272, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 275, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 275, + 98 + ], + "score": 1.0, + "content": "FRATERNAL DROPOUT", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 114, + 114, + 425, + 127 + ], + "lines": [ + { + "bbox": [ + 114, + 112, + 428, + 129 + ], + "spans": [ + { + "bbox": [ + 114, + 112, + 150, + 129 + ], + "score": 1.0, + "content": "Konrad", + "type": "text" + }, + { + "bbox": [ + 150, + 113, + 193, + 126 + ], + "score": 0.33, + "content": "\\dot { \\mathbf Z } \\mathbf { o } \\mathbf { f n } \\mathbf { a } ^ { 1 , 2 , * } ;", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 112, + 428, + 129 + ], + "score": 1.0, + "content": "Devansh Arpit2, Dendi Suhubdy2 & Yoshua Bengio2,3", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 114, + 112, + 428, + 129 + ] + }, + { + "type": "index", + "bbox": [ + 112, + 137, + 241, + 171 + ], + "lines": [ + { + "bbox": [ + 112, + 136, + 212, + 150 + ], + "spans": [ + { + "bbox": [ + 112, + 136, + 212, + 150 + ], + "score": 1.0, + "content": "1Jagiellonian University", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 147, + 243, + 162 + ], + "spans": [ + { + "bbox": [ + 111, + 147, + 243, + 162 + ], + "score": 1.0, + "content": "2MILA, Université de Montréal", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 158, + 207, + 173 + ], + "spans": [ + { + "bbox": [ + 111, + 158, + 207, + 173 + ], + "score": 1.0, + "content": "3CIFAR Senior Fellow", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + } + ], + "index": 3, + "bbox_fs": [ + 111, + 136, + 243, + 173 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 201, + 333, + 213 + ], + "lines": [ + { + "bbox": [ + 276, + 200, + 335, + 214 + ], + "spans": [ + { + "bbox": [ + 276, + 200, + 335, + 214 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 228, + 468, + 415 + ], + "lines": [ + { + "bbox": [ + 142, + 228, + 469, + 241 + ], + "spans": [ + { + "bbox": [ + 142, + 228, + 469, + 241 + ], + "score": 1.0, + "content": "Recurrent neural networks (RNNs) form an important class of architectures", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 240, + 469, + 252 + ], + "spans": [ + { + "bbox": [ + 141, + 240, + 469, + 252 + ], + "score": 1.0, + "content": "among neural networks useful for language modeling and sequential prediction.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 251, + 469, + 263 + ], + "spans": [ + { + "bbox": [ + 142, + 251, + 469, + 263 + ], + "score": 1.0, + "content": "However, optimizing RNNs is known to be harder compared to feed-forward neu-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 262, + 469, + 274 + ], + "spans": [ + { + "bbox": [ + 141, + 262, + 469, + 274 + ], + "score": 1.0, + "content": "ral networks. A number of techniques have been proposed in literature to address", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 272, + 470, + 285 + ], + "spans": [ + { + "bbox": [ + 141, + 272, + 470, + 285 + ], + "score": 1.0, + "content": "this problem. In this paper we propose a simple technique called fraternal dropout", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 283, + 470, + 297 + ], + "spans": [ + { + "bbox": [ + 141, + 283, + 470, + 297 + ], + "score": 1.0, + "content": "that takes advantage of dropout to achieve this goal. Specifically, we propose to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 294, + 469, + 307 + ], + "spans": [ + { + "bbox": [ + 141, + 294, + 469, + 307 + ], + "score": 1.0, + "content": "train two identical copies of an RNN (that share parameters) with different dropout", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 305, + 469, + 317 + ], + "spans": [ + { + "bbox": [ + 141, + 305, + 469, + 317 + ], + "score": 1.0, + "content": "masks while minimizing the difference between their (pre-softmax) predictions.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 316, + 469, + 329 + ], + "spans": [ + { + "bbox": [ + 141, + 316, + 469, + 329 + ], + "score": 1.0, + "content": "In this way our regularization encourages the representations of RNNs to be in-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 327, + 469, + 339 + ], + "spans": [ + { + "bbox": [ + 141, + 327, + 469, + 339 + ], + "score": 1.0, + "content": "variant to dropout mask, thus being robust. We show that our regularization term is", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 338, + 470, + 350 + ], + "spans": [ + { + "bbox": [ + 141, + 338, + 470, + 350 + ], + "score": 1.0, + "content": "upper bounded by the expectation-linear dropout objective which has been shown", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 349, + 470, + 362 + ], + "spans": [ + { + "bbox": [ + 141, + 349, + 470, + 362 + ], + "score": 1.0, + "content": "to address the gap due to the difference between the train and inference phases of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 142, + 360, + 470, + 372 + ], + "spans": [ + { + "bbox": [ + 142, + 360, + 470, + 372 + ], + "score": 1.0, + "content": "dropout. We evaluate our model and achieve state-of-the-art results in sequence", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 371, + 469, + 383 + ], + "spans": [ + { + "bbox": [ + 142, + 371, + 469, + 383 + ], + "score": 1.0, + "content": "modeling tasks on two benchmark datasets – Penn Treebank and Wikitext-2. We", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 142, + 383, + 470, + 394 + ], + "spans": [ + { + "bbox": [ + 142, + 383, + 470, + 394 + ], + "score": 1.0, + "content": "also show that our approach leads to performance improvement by a significant", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 393, + 469, + 405 + ], + "spans": [ + { + "bbox": [ + 141, + 393, + 469, + 405 + ], + "score": 1.0, + "content": "margin in image captioning (Microsoft COCO) and semi-supervised (CIFAR-10)", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 405, + 167, + 415 + ], + "spans": [ + { + "bbox": [ + 142, + 405, + 167, + 415 + ], + "score": 1.0, + "content": "tasks.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 14, + "bbox_fs": [ + 141, + 228, + 470, + 415 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 443, + 206, + 455 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 208, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 208, + 459 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 470, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 483 + ], + "score": 1.0, + "content": "Recurrent neural networks (RNNs) like long short-term memory (LSTM; Hochreiter & Schmidhu-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 482, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 494 + ], + "score": 1.0, + "content": "ber (1997)) networks and gated recurrent unit (GRU; Chung et al. (2014)) are popular architectures", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 493, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 505 + ], + "score": 1.0, + "content": "for sequence modeling tasks like language generation, translation, speech synthesis, and machine", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "score": 1.0, + "content": "comprehension. However, they are harder to optimize compared to feed-forward networks due to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 514, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 528 + ], + "score": 1.0, + "content": "challenges like variable length input sequences, repeated application of the same transition operator", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 526, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 506, + 538 + ], + "score": 1.0, + "content": "at each time step, and largely-dense embedding matrix that depends on the vocabulary size. Due to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "these optimization challenges in RNNs, the application of batch normalization and its variants (layer", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 548, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 506, + 560 + ], + "score": 1.0, + "content": "normalization, recurrent batch normalization, recurrent normalization propagation) have not been as", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 558, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 571 + ], + "score": 1.0, + "content": "successful as their counterparts in feed-forward networks (Laurent et al., 2016), although they do", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "considerably provide performance gains. Similarly, naive application of dropout (Srivastava et al.,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "score": 1.0, + "content": "2014) has been shown to be ineffective in RNNs (Zaremba et al., 2014). Therefore, regularization", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 591, + 308, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 308, + 604 + ], + "score": 1.0, + "content": "techniques for RNNs is an active area of research.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 471, + 506, + 604 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 608, + 505, + 706 + ], + "lines": [ + { + "bbox": [ + 106, + 609, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 620 + ], + "score": 1.0, + "content": "To address these challenges, Zaremba et al. (2014) proposed to apply dropout only to the non-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "recurrent connections in multi-layer RNNs. Variational dropout (Gal & Ghahramani (2016)) uses", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "the same dropout mask throughout a sequence during training. DropConnect (Wan et al., 2013)", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 640, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 506, + 654 + ], + "score": 1.0, + "content": "applies the dropout operation on the weight matrices. Zoneout (Krueger et al. (2016)), in a similar", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 650, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 506, + 666 + ], + "score": 1.0, + "content": "spirit with dropout, randomly chooses to use the previous time step hidden state instead of using", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 663, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 505, + 675 + ], + "score": 1.0, + "content": "the current one. Similarly as a substitute for batch normalization, layer normalization normalizes", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "score": 1.0, + "content": "the hidden units within each sample to have zero mean and unit standard deviation. Recurrent batch", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 684, + 506, + 699 + ], + "spans": [ + { + "bbox": [ + 104, + 684, + 506, + 699 + ], + "score": 1.0, + "content": "normalization applies batch normalization but with unshared mini-batch statistics for each time step", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 695, + 208, + 708 + ], + "spans": [ + { + "bbox": [ + 105, + 695, + 208, + 708 + ], + "score": 1.0, + "content": "(Cooijmans et al., 2016).", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 40, + "bbox_fs": [ + 104, + 609, + 506, + 708 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 137 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "Merity et al. (2017a) and Merity et al. (2017b) on the other hand show that activity regulariza-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 506, + 107 + ], + "score": 1.0, + "content": "tion (AR) and temporal activation regularization (TAR)1 are also effective methods for regularizing", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "LSTMs. Another more recent way of regularizing RNNs, that is similar in spirit to the approach we", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 129 + ], + "score": 1.0, + "content": "take, involves minimizing the difference between the hidden states of the original and the auxiliary", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 230, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 230, + 139 + ], + "score": 1.0, + "content": "network Serdyuk et al. (2017).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 242 + ], + "lines": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "In this paper we propose a simple regularization based on dropout that we call fraternal dropout,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 155, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 505, + 167 + ], + "score": 1.0, + "content": "where we minimize an equally weighted sum of prediction losses from two identical copies of the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 401, + 178 + ], + "score": 1.0, + "content": "same LSTM with different dropout masks, and add as a regularization the", + "type": "text" + }, + { + "bbox": [ + 401, + 165, + 411, + 175 + ], + "score": 0.84, + "content": "\\ell ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 164, + 506, + 178 + ], + "score": 1.0, + "content": "difference between the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "score": 1.0, + "content": "predictions (pre-softmax) of the two networks. We analytically show that our regularization objec-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "tive is equivalent to minimizing the variance in predictions from different i.i.d. dropout masks; thus", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "encouraging the predictions to be invariant to dropout masks. We also discuss how our regularization", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 208, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 104, + 208, + 505, + 223 + ], + "score": 1.0, + "content": "is related to expectation linear dropout Ma et al. (2016), Π-model Laine & Aila (2016) and activity", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 234 + ], + "score": 1.0, + "content": "regularization Merity et al. (2017b), and empirically show that our method provides non-trivial gains", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 470, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 470, + 244 + ], + "score": 1.0, + "content": "over these related methods which we explain furthermore in our ablation study (Section 5).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9 + }, + { + "type": "title", + "bbox": [ + 108, + 263, + 239, + 276 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 241, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 241, + 278 + ], + "score": 1.0, + "content": "2 FRATERNAL DROPOUT", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 291, + 505, + 379 + ], + "lines": [ + { + "bbox": [ + 106, + 291, + 504, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 504, + 303 + ], + "score": 1.0, + "content": "Dropout is a powerful regularization for neural networks. It is usually more effective on densely", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 303, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 314 + ], + "score": 1.0, + "content": "connected layers because they suffer more from overfitting compared with convolution layers where", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "the parameters are shared. For this reason dropout is an important regularization for RNNs. How-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "ever, dropout has a gap between its training and inference phase since the latter phase assumes linear", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 335, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 505, + 348 + ], + "score": 1.0, + "content": "activations to correct for the factor by which the expected value of each activation would be different", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 346, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 505, + 359 + ], + "score": 1.0, + "content": "Ma et al. (2016). In addition, the prediction of models with dropout generally vary with different", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "dropout mask. However, the desirable property in such cases would be to have final predictions be", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 369, + 217, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 217, + 380 + ], + "score": 1.0, + "content": "invariant to dropout masks.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "As such, the idea behind fraternal dropout is to train a neural network model in a way that encour-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "score": 1.0, + "content": "ages the variance in predictions under different dropout masks to be as small as possible. Specif-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 318, + 419 + ], + "score": 1.0, + "content": "ically, consider we have an RNN model denoted by", + "type": "text" + }, + { + "bbox": [ + 319, + 407, + 344, + 419 + ], + "score": 0.92, + "content": "\\mathcal { M } ( \\boldsymbol { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 407, + 422, + 419 + ], + "score": 1.0, + "content": "that takes as input", + "type": "text" + }, + { + "bbox": [ + 423, + 407, + 432, + 417 + ], + "score": 0.55, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 407, + 464, + 419 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 464, + 407, + 470, + 417 + ], + "score": 0.78, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 417, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 214, + 431 + ], + "score": 1.0, + "content": "the model parameters. Let", + "type": "text" + }, + { + "bbox": [ + 214, + 417, + 291, + 430 + ], + "score": 0.93, + "content": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) \\in \\mathbb { R } ^ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 417, + 484, + 431 + ], + "score": 1.0, + "content": "be the prediction of the model for input sample", + "type": "text" + }, + { + "bbox": [ + 484, + 418, + 494, + 428 + ], + "score": 0.53, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 417, + 506, + 431 + ], + "score": 1.0, + "content": "at", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 126, + 442 + ], + "score": 1.0, + "content": "time", + "type": "text" + }, + { + "bbox": [ + 127, + 430, + 132, + 439 + ], + "score": 0.69, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 428, + 208, + 442 + ], + "score": 1.0, + "content": ", for dropout mask", + "type": "text" + }, + { + "bbox": [ + 208, + 429, + 218, + 441 + ], + "score": 0.88, + "content": "s _ { i } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 428, + 290, + 442 + ], + "score": 1.0, + "content": "and current input", + "type": "text" + }, + { + "bbox": [ + 290, + 429, + 300, + 439 + ], + "score": 0.84, + "content": "\\mathbf { z } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 428, + 331, + 442 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 331, + 429, + 342, + 439 + ], + "score": 0.87, + "content": "\\mathbf { z } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 428, + 407, + 442 + ], + "score": 1.0, + "content": "is a function of", + "type": "text" + }, + { + "bbox": [ + 407, + 429, + 417, + 439 + ], + "score": 0.33, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 428, + 505, + 442 + ], + "score": 1.0, + "content": "and the hidden states", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 440, + 503, + 455 + ], + "spans": [ + { + "bbox": [ + 104, + 440, + 326, + 455 + ], + "score": 1.0, + "content": "corresponding to the previous time steps. Similarly, let", + "type": "text" + }, + { + "bbox": [ + 326, + 440, + 404, + 453 + ], + "score": 0.91, + "content": "\\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) , \\mathbf { Y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 440, + 491, + 455 + ], + "score": 1.0, + "content": "be the corresponding", + "type": "text" + }, + { + "bbox": [ + 491, + 440, + 503, + 451 + ], + "score": 0.86, + "content": "t ^ { t h }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 451, + 379, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 344, + 465 + ], + "score": 1.0, + "content": "time step loss value for the overall input-target sample pair", + "type": "text" + }, + { + "bbox": [ + 344, + 453, + 374, + 464 + ], + "score": 0.89, + "content": "( \\mathbf { X } , \\mathbf { Y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 451, + 379, + 465 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 468, + 505, + 516 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 408, + 482 + ], + "score": 1.0, + "content": "Then in fraternal dropout, we simultaneously feed-forward the input sample", + "type": "text" + }, + { + "bbox": [ + 408, + 469, + 418, + 479 + ], + "score": 0.37, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 468, + 505, + 482 + ], + "score": 1.0, + "content": "through two identical", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 479, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 312, + 493 + ], + "score": 1.0, + "content": "copies of the RNN that share the same parameters", + "type": "text" + }, + { + "bbox": [ + 312, + 480, + 318, + 490 + ], + "score": 0.77, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 479, + 455, + 493 + ], + "score": 1.0, + "content": "but with different dropout masks", + "type": "text" + }, + { + "bbox": [ + 455, + 480, + 464, + 492 + ], + "score": 0.87, + "content": "s _ { i } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 479, + 483, + 493 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 484, + 480, + 493, + 493 + ], + "score": 0.88, + "content": "\\boldsymbol { s } _ { j } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 479, + 506, + 493 + ], + "score": 1.0, + "content": "at", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 167, + 505 + ], + "score": 1.0, + "content": "each time step", + "type": "text" + }, + { + "bbox": [ + 167, + 493, + 173, + 502 + ], + "score": 0.65, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 492, + 361, + 505 + ], + "score": 1.0, + "content": ". This yields two loss values at each time step", + "type": "text" + }, + { + "bbox": [ + 361, + 493, + 367, + 503 + ], + "score": 0.69, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 492, + 406, + 505 + ], + "score": 1.0, + "content": "given by", + "type": "text" + }, + { + "bbox": [ + 406, + 492, + 483, + 505 + ], + "score": 0.92, + "content": "\\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) , \\mathbf { Y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 492, + 505, + 505 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 502, + 441, + 517 + ], + "spans": [ + { + "bbox": [ + 107, + 503, + 185, + 517 + ], + "score": 0.93, + "content": "\\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\bar { \\theta } ) , \\mathbf { Y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 502, + 441, + 517 + ], + "score": 1.0, + "content": ". Then the overall loss function of fraternal dropout is given by,", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 525, + 502, + 561 + ], + "lines": [ + { + "bbox": [ + 142, + 525, + 502, + 561 + ], + "spans": [ + { + "bbox": [ + 142, + 525, + 502, + 561 + ], + "score": 0.91, + "content": "\\ell _ { F D } ( \\mathbf { X } , \\mathbf { Y } ) = \\sum _ { t = 1 } ^ { T } \\frac { 1 } { 2 } \\left( \\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) , \\mathbf { Y } ) + \\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) , \\mathbf { Y } ) \\right) + \\frac { \\kappa } { m T } \\sum _ { t = 1 } ^ { T } \\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\theta )", + "type": "interline_equation", + "image_path": "93bdccc72eea7a43b2752ee0bbada1f1fb8f1cc819ea811ccae19b60a4d61061.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 142, + 525, + 502, + 537.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 142, + 537.0, + 502, + 549.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 142, + 549.0, + 502, + 561.0 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 582, + 506, + 606 + ], + "lines": [ + { + "bbox": [ + 105, + 581, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 133, + 596 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 585, + 140, + 593 + ], + "score": 0.79, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 581, + 270, + 596 + ], + "score": 1.0, + "content": "is the regularization coefficient,", + "type": "text" + }, + { + "bbox": [ + 270, + 585, + 280, + 593 + ], + "score": 0.78, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 581, + 363, + 596 + ], + "score": 1.0, + "content": "is the dimensions of", + "type": "text" + }, + { + "bbox": [ + 364, + 582, + 413, + 595 + ], + "score": 0.93, + "content": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 581, + 431, + 596 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 432, + 582, + 480, + 595 + ], + "score": 0.93, + "content": "\\mathcal { R } _ { F D } ( { \\bf z } ^ { t } ; \\boldsymbol { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 581, + 505, + 596 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 594, + 274, + 608 + ], + "spans": [ + { + "bbox": [ + 104, + 594, + 274, + 608 + ], + "score": 1.0, + "content": "fraternal dropout regularization given by,", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 614, + 371, + 632 + ], + "lines": [ + { + "bbox": [ + 142, + 614, + 371, + 632 + ], + "spans": [ + { + "bbox": [ + 142, + 614, + 371, + 632 + ], + "score": 0.89, + "content": "\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) : = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) \\| _ { 2 } ^ { 2 } \\right] .", + "type": "interline_equation", + "image_path": "19a7aa964ddb00f5896584ab9a2ee80ad80921cffd3df6a4d3833ec209159062.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 142, + 614, + 371, + 632 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 641, + 505, + 666 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 294, + 655 + ], + "score": 1.0, + "content": "We use Monte Carlo sampling to approximate", + "type": "text" + }, + { + "bbox": [ + 294, + 641, + 342, + 654 + ], + "score": 0.93, + "content": "\\mathcal { R } _ { F D } ( { \\bf z } ^ { t } ; \\boldsymbol { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 640, + 370, + 655 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 371, + 641, + 420, + 654 + ], + "score": 0.95, + "content": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 640, + 439, + 655 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 439, + 641, + 489, + 655 + ], + "score": 0.93, + "content": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 640, + 506, + 655 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 653, + 489, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 255, + 668 + ], + "score": 1.0, + "content": "the same as the one used to calculate", + "type": "text" + }, + { + "bbox": [ + 255, + 654, + 264, + 664 + ], + "score": 0.85, + "content": "\\ell ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 653, + 489, + 668 + ], + "score": 1.0, + "content": "values. Hence, the additional computation is negligible.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 108, + 671, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 106, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 505, + 684 + ], + "score": 1.0, + "content": "We note that the regularization term of our objective is equivalent to minimizing the variance in the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 682, + 465, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 465, + 696 + ], + "score": 1.0, + "content": "prediction function with different dropout masks as shown below (proof in the appendix).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 105, + 711, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 119, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "1TAR and Zoneout are similar in their motivations because both leads to adjacent time step hidden states to", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 720, + 182, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 182, + 733 + ], + "score": 1.0, + "content": "be close on average.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 137 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "Merity et al. (2017a) and Merity et al. (2017b) on the other hand show that activity regulariza-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 506, + 107 + ], + "score": 1.0, + "content": "tion (AR) and temporal activation regularization (TAR)1 are also effective methods for regularizing", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "LSTMs. Another more recent way of regularizing RNNs, that is similar in spirit to the approach we", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 129 + ], + "score": 1.0, + "content": "take, involves minimizing the difference between the hidden states of the original and the auxiliary", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 230, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 230, + 139 + ], + "score": 1.0, + "content": "network Serdyuk et al. (2017).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 82, + 506, + 139 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 242 + ], + "lines": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "In this paper we propose a simple regularization based on dropout that we call fraternal dropout,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 155, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 505, + 167 + ], + "score": 1.0, + "content": "where we minimize an equally weighted sum of prediction losses from two identical copies of the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 401, + 178 + ], + "score": 1.0, + "content": "same LSTM with different dropout masks, and add as a regularization the", + "type": "text" + }, + { + "bbox": [ + 401, + 165, + 411, + 175 + ], + "score": 0.84, + "content": "\\ell ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 164, + 506, + 178 + ], + "score": 1.0, + "content": "difference between the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "score": 1.0, + "content": "predictions (pre-softmax) of the two networks. We analytically show that our regularization objec-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "tive is equivalent to minimizing the variance in predictions from different i.i.d. dropout masks; thus", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "encouraging the predictions to be invariant to dropout masks. We also discuss how our regularization", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 208, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 104, + 208, + 505, + 223 + ], + "score": 1.0, + "content": "is related to expectation linear dropout Ma et al. (2016), Π-model Laine & Aila (2016) and activity", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 234 + ], + "score": 1.0, + "content": "regularization Merity et al. (2017b), and empirically show that our method provides non-trivial gains", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 470, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 470, + 244 + ], + "score": 1.0, + "content": "over these related methods which we explain furthermore in our ablation study (Section 5).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9, + "bbox_fs": [ + 104, + 142, + 506, + 244 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 263, + 239, + 276 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 241, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 241, + 278 + ], + "score": 1.0, + "content": "2 FRATERNAL DROPOUT", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 291, + 505, + 379 + ], + "lines": [ + { + "bbox": [ + 106, + 291, + 504, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 504, + 303 + ], + "score": 1.0, + "content": "Dropout is a powerful regularization for neural networks. It is usually more effective on densely", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 303, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 314 + ], + "score": 1.0, + "content": "connected layers because they suffer more from overfitting compared with convolution layers where", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "the parameters are shared. For this reason dropout is an important regularization for RNNs. How-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "ever, dropout has a gap between its training and inference phase since the latter phase assumes linear", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 335, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 505, + 348 + ], + "score": 1.0, + "content": "activations to correct for the factor by which the expected value of each activation would be different", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 346, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 505, + 359 + ], + "score": 1.0, + "content": "Ma et al. (2016). In addition, the prediction of models with dropout generally vary with different", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "dropout mask. However, the desirable property in such cases would be to have final predictions be", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 369, + 217, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 217, + 380 + ], + "score": 1.0, + "content": "invariant to dropout masks.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 291, + 505, + 380 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "As such, the idea behind fraternal dropout is to train a neural network model in a way that encour-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "score": 1.0, + "content": "ages the variance in predictions under different dropout masks to be as small as possible. Specif-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 318, + 419 + ], + "score": 1.0, + "content": "ically, consider we have an RNN model denoted by", + "type": "text" + }, + { + "bbox": [ + 319, + 407, + 344, + 419 + ], + "score": 0.92, + "content": "\\mathcal { M } ( \\boldsymbol { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 407, + 422, + 419 + ], + "score": 1.0, + "content": "that takes as input", + "type": "text" + }, + { + "bbox": [ + 423, + 407, + 432, + 417 + ], + "score": 0.55, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 407, + 464, + 419 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 464, + 407, + 470, + 417 + ], + "score": 0.78, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 417, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 214, + 431 + ], + "score": 1.0, + "content": "the model parameters. Let", + "type": "text" + }, + { + "bbox": [ + 214, + 417, + 291, + 430 + ], + "score": 0.93, + "content": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) \\in \\mathbb { R } ^ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 417, + 484, + 431 + ], + "score": 1.0, + "content": "be the prediction of the model for input sample", + "type": "text" + }, + { + "bbox": [ + 484, + 418, + 494, + 428 + ], + "score": 0.53, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 417, + 506, + 431 + ], + "score": 1.0, + "content": "at", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 126, + 442 + ], + "score": 1.0, + "content": "time", + "type": "text" + }, + { + "bbox": [ + 127, + 430, + 132, + 439 + ], + "score": 0.69, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 428, + 208, + 442 + ], + "score": 1.0, + "content": ", for dropout mask", + "type": "text" + }, + { + "bbox": [ + 208, + 429, + 218, + 441 + ], + "score": 0.88, + "content": "s _ { i } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 428, + 290, + 442 + ], + "score": 1.0, + "content": "and current input", + "type": "text" + }, + { + "bbox": [ + 290, + 429, + 300, + 439 + ], + "score": 0.84, + "content": "\\mathbf { z } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 428, + 331, + 442 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 331, + 429, + 342, + 439 + ], + "score": 0.87, + "content": "\\mathbf { z } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 428, + 407, + 442 + ], + "score": 1.0, + "content": "is a function of", + "type": "text" + }, + { + "bbox": [ + 407, + 429, + 417, + 439 + ], + "score": 0.33, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 428, + 505, + 442 + ], + "score": 1.0, + "content": "and the hidden states", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 440, + 503, + 455 + ], + "spans": [ + { + "bbox": [ + 104, + 440, + 326, + 455 + ], + "score": 1.0, + "content": "corresponding to the previous time steps. Similarly, let", + "type": "text" + }, + { + "bbox": [ + 326, + 440, + 404, + 453 + ], + "score": 0.91, + "content": "\\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) , \\mathbf { Y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 440, + 491, + 455 + ], + "score": 1.0, + "content": "be the corresponding", + "type": "text" + }, + { + "bbox": [ + 491, + 440, + 503, + 451 + ], + "score": 0.86, + "content": "t ^ { t h }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 451, + 379, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 344, + 465 + ], + "score": 1.0, + "content": "time step loss value for the overall input-target sample pair", + "type": "text" + }, + { + "bbox": [ + 344, + 453, + 374, + 464 + ], + "score": 0.89, + "content": "( \\mathbf { X } , \\mathbf { Y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 451, + 379, + 465 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26, + "bbox_fs": [ + 104, + 385, + 506, + 465 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 468, + 505, + 516 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 408, + 482 + ], + "score": 1.0, + "content": "Then in fraternal dropout, we simultaneously feed-forward the input sample", + "type": "text" + }, + { + "bbox": [ + 408, + 469, + 418, + 479 + ], + "score": 0.37, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 468, + 505, + 482 + ], + "score": 1.0, + "content": "through two identical", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 479, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 312, + 493 + ], + "score": 1.0, + "content": "copies of the RNN that share the same parameters", + "type": "text" + }, + { + "bbox": [ + 312, + 480, + 318, + 490 + ], + "score": 0.77, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 479, + 455, + 493 + ], + "score": 1.0, + "content": "but with different dropout masks", + "type": "text" + }, + { + "bbox": [ + 455, + 480, + 464, + 492 + ], + "score": 0.87, + "content": "s _ { i } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 479, + 483, + 493 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 484, + 480, + 493, + 493 + ], + "score": 0.88, + "content": "\\boldsymbol { s } _ { j } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 479, + 506, + 493 + ], + "score": 1.0, + "content": "at", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 167, + 505 + ], + "score": 1.0, + "content": "each time step", + "type": "text" + }, + { + "bbox": [ + 167, + 493, + 173, + 502 + ], + "score": 0.65, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 492, + 361, + 505 + ], + "score": 1.0, + "content": ". This yields two loss values at each time step", + "type": "text" + }, + { + "bbox": [ + 361, + 493, + 367, + 503 + ], + "score": 0.69, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 492, + 406, + 505 + ], + "score": 1.0, + "content": "given by", + "type": "text" + }, + { + "bbox": [ + 406, + 492, + 483, + 505 + ], + "score": 0.92, + "content": "\\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) , \\mathbf { Y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 492, + 505, + 505 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 502, + 441, + 517 + ], + "spans": [ + { + "bbox": [ + 107, + 503, + 185, + 517 + ], + "score": 0.93, + "content": "\\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\bar { \\theta } ) , \\mathbf { Y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 502, + 441, + 517 + ], + "score": 1.0, + "content": ". Then the overall loss function of fraternal dropout is given by,", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 468, + 506, + 517 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 525, + 502, + 561 + ], + "lines": [ + { + "bbox": [ + 142, + 525, + 502, + 561 + ], + "spans": [ + { + "bbox": [ + 142, + 525, + 502, + 561 + ], + "score": 0.91, + "content": "\\ell _ { F D } ( \\mathbf { X } , \\mathbf { Y } ) = \\sum _ { t = 1 } ^ { T } \\frac { 1 } { 2 } \\left( \\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) , \\mathbf { Y } ) + \\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) , \\mathbf { Y } ) \\right) + \\frac { \\kappa } { m T } \\sum _ { t = 1 } ^ { T } \\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\theta )", + "type": "interline_equation", + "image_path": "93bdccc72eea7a43b2752ee0bbada1f1fb8f1cc819ea811ccae19b60a4d61061.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 142, + 525, + 502, + 537.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 142, + 537.0, + 502, + 549.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 142, + 549.0, + 502, + 561.0 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 582, + 506, + 606 + ], + "lines": [ + { + "bbox": [ + 105, + 581, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 133, + 596 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 585, + 140, + 593 + ], + "score": 0.79, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 581, + 270, + 596 + ], + "score": 1.0, + "content": "is the regularization coefficient,", + "type": "text" + }, + { + "bbox": [ + 270, + 585, + 280, + 593 + ], + "score": 0.78, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 581, + 363, + 596 + ], + "score": 1.0, + "content": "is the dimensions of", + "type": "text" + }, + { + "bbox": [ + 364, + 582, + 413, + 595 + ], + "score": 0.93, + "content": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 581, + 431, + 596 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 432, + 582, + 480, + 595 + ], + "score": 0.93, + "content": "\\mathcal { R } _ { F D } ( { \\bf z } ^ { t } ; \\boldsymbol { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 581, + 505, + 596 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 594, + 274, + 608 + ], + "spans": [ + { + "bbox": [ + 104, + 594, + 274, + 608 + ], + "score": 1.0, + "content": "fraternal dropout regularization given by,", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 104, + 581, + 505, + 608 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 614, + 371, + 632 + ], + "lines": [ + { + "bbox": [ + 142, + 614, + 371, + 632 + ], + "spans": [ + { + "bbox": [ + 142, + 614, + 371, + 632 + ], + "score": 0.89, + "content": "\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) : = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) \\| _ { 2 } ^ { 2 } \\right] .", + "type": "interline_equation", + "image_path": "19a7aa964ddb00f5896584ab9a2ee80ad80921cffd3df6a4d3833ec209159062.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 142, + 614, + 371, + 632 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 641, + 505, + 666 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 294, + 655 + ], + "score": 1.0, + "content": "We use Monte Carlo sampling to approximate", + "type": "text" + }, + { + "bbox": [ + 294, + 641, + 342, + 654 + ], + "score": 0.93, + "content": "\\mathcal { R } _ { F D } ( { \\bf z } ^ { t } ; \\boldsymbol { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 640, + 370, + 655 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 371, + 641, + 420, + 654 + ], + "score": 0.95, + "content": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 640, + 439, + 655 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 439, + 641, + 489, + 655 + ], + "score": 0.93, + "content": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 640, + 506, + 655 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 653, + 489, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 255, + 668 + ], + "score": 1.0, + "content": "the same as the one used to calculate", + "type": "text" + }, + { + "bbox": [ + 255, + 654, + 264, + 664 + ], + "score": 0.85, + "content": "\\ell ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 653, + 489, + 668 + ], + "score": 1.0, + "content": "values. Hence, the additional computation is negligible.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 640, + 506, + 668 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 671, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 106, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 505, + 684 + ], + "score": 1.0, + "content": "We note that the regularization term of our objective is equivalent to minimizing the variance in the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 682, + 465, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 465, + 696 + ], + "score": 1.0, + "content": "prediction function with different dropout masks as shown below (proof in the appendix).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 671, + 505, + 696 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 81, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 171, + 97 + ], + "score": 1.0, + "content": "Remark 1. Let", + "type": "text" + }, + { + "bbox": [ + 172, + 82, + 181, + 95 + ], + "score": 0.88, + "content": "s _ { i } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 81, + 200, + 97 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 200, + 82, + 209, + 96 + ], + "score": 0.87, + "content": "\\boldsymbol { s } _ { j } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 81, + 322, + 97 + ], + "score": 1.0, + "content": "be i.i.d. dropout masks and", + "type": "text" + }, + { + "bbox": [ + 322, + 82, + 399, + 95 + ], + "score": 0.93, + "content": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) \\in \\mathbb { R } ^ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 81, + 506, + 97 + ], + "score": 1.0, + "content": "be the prediction function", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 213, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 213, + 106 + ], + "score": 1.0, + "content": "as described above. Then,", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 108, + 478, + 142 + ], + "lines": [ + { + "bbox": [ + 142, + 108, + 478, + 142 + ], + "spans": [ + { + "bbox": [ + 142, + 108, + 478, + 142 + ], + "score": 0.93, + "content": "\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\theta ) = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] = 2 \\sum _ { q = 1 } ^ { m } \\mathrm { v a r } _ { s _ { i } ^ { t } } ( p _ { q } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) ) .", + "type": "interline_equation", + "image_path": "0a168a922a4cdc23ffbcba643900e9fa7e60d2fa738443aed02c1cf592b69992.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 142, + 108, + 478, + 119.33333333333333 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 142, + 119.33333333333333, + 478, + 130.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 142, + 130.66666666666666, + 478, + 142.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 152, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 152, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 505, + 166 + ], + "score": 1.0, + "content": "Note that a generalization of our approach would be to minimize the difference between the predic-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 163, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 505, + 176 + ], + "score": 1.0, + "content": "tions of the two networks with different data/model augmentations. However, in this paper we focus", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 174, + 389, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 354, + 187 + ], + "score": 1.0, + "content": "on using different dropout masks and experiment mainly with", + "type": "text" + }, + { + "bbox": [ + 355, + 174, + 385, + 186 + ], + "score": 0.47, + "content": "\\mathrm { R N N s } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 174, + 389, + 187 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "title", + "bbox": [ + 108, + 201, + 212, + 215 + ], + "lines": [ + { + "bbox": [ + 104, + 200, + 213, + 217 + ], + "spans": [ + { + "bbox": [ + 104, + 200, + 213, + 217 + ], + "score": 1.0, + "content": "3 RELATED WORK", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "title", + "bbox": [ + 107, + 226, + 364, + 239 + ], + "lines": [ + { + "bbox": [ + 104, + 225, + 365, + 241 + ], + "spans": [ + { + "bbox": [ + 104, + 225, + 365, + 241 + ], + "score": 1.0, + "content": "3.1 RELATION TO EXPECTATION LINEAR DROPOUT (ELD)", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 247, + 505, + 291 + ], + "lines": [ + { + "bbox": [ + 105, + 247, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 506, + 260 + ], + "score": 1.0, + "content": "Ma et al. (2016) analytically showed that the expected error (over samples) between a model’s", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "score": 1.0, + "content": "expected prediction over all dropout masks, and the prediction using the average mask, is upper", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "score": 1.0, + "content": "bounded. Based on this result, they propose to explicitly minimize the difference (we have adapted", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 280, + 256, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 256, + 293 + ], + "score": 1.0, + "content": "their regularization to our notations),", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 295, + 364, + 311 + ], + "lines": [ + { + "bbox": [ + 142, + 295, + 364, + 311 + ], + "spans": [ + { + "bbox": [ + 142, + 295, + 364, + 311 + ], + "score": 0.89, + "content": "\\mathcal { R } _ { E L D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) = \\| \\mathbb { E } _ { s } \\left[ \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s ; \\boldsymbol { \\theta } ) \\right] - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\mathbb { E } _ { s } [ s ] ; \\boldsymbol { \\theta } ) \\| _ { 2 }", + "type": "interline_equation", + "image_path": "d5b9503e2f3580f8faa094454aa7daa1f3502bb502f6109681ecc8d523451ea2.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 142, + 295, + 364, + 311 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 314, + 505, + 337 + ], + "lines": [ + { + "bbox": [ + 105, + 313, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 133, + 329 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 317, + 139, + 325 + ], + "score": 0.74, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 313, + 505, + 329 + ], + "score": 1.0, + "content": "is the dropout mask. However, due to feasibility consideration, they instead propose to use", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 326, + 266, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 266, + 338 + ], + "score": 1.0, + "content": "the following regularization in practice,", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "interline_equation", + "bbox": [ + 141, + 342, + 375, + 358 + ], + "lines": [ + { + "bbox": [ + 141, + 342, + 375, + 358 + ], + "spans": [ + { + "bbox": [ + 141, + 342, + 375, + 358 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\tilde { \\mathcal { R } } _ { E L D } ( \\mathbf { z } ^ { t } ; \\theta ) = \\mathbb { E } _ { s _ { i } } \\left[ \\Vert \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\mathbb { E } _ { s } [ s ] ; \\theta ) \\Vert _ { 2 } ^ { 2 } \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "1e947ec086b9b0b95ed800567914b15a7c2e4666bcc4d4fde3424b3f5195e5ea.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 141, + 342, + 375, + 358 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 361, + 505, + 439 + ], + "lines": [ + { + "bbox": [ + 106, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "Specifically, this is achieved by feed-forwarding the input twice through the network, with and with-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "out dropout mask, and minimizing the main network loss (with dropout) along with the regulariza-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "tion term specified above (but without back-propagating the gradients through the network without", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 395, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 505, + 407 + ], + "score": 1.0, + "content": "dropout). The goal of Ma et al. (2016) is to minimize the network loss along with the expected dif-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "score": 1.0, + "content": "ference between the prediction from individual dropout mask and the prediction from the expected", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "score": 1.0, + "content": "dropout mask. We note that our regularization objective is upper bounded by the expectation-linear", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 428, + 361, + 440 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 361, + 440 + ], + "score": 1.0, + "content": "dropout regularization as shown below (proof in the appendix).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 442, + 293, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 171, + 458 + ], + "score": 1.0, + "content": "Proposition 1.", + "type": "text" + }, + { + "bbox": [ + 171, + 441, + 291, + 456 + ], + "score": 0.91, + "content": "\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) \\leq 4 \\tilde { \\mathcal { R } } _ { E L D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) .", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 441, + 294, + 458 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 463, + 505, + 563 + ], + "lines": [ + { + "bbox": [ + 106, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 477 + ], + "score": 1.0, + "content": "This result shows that minimizing the ELD objective indirectly minimizes our regularization term.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 475, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 506, + 488 + ], + "score": 1.0, + "content": "Finally as indicated above, they apply the target loss only on the network with dropout. In fact, in our", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 486, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 499 + ], + "score": 1.0, + "content": "own ablation studies (see Section 5) we find that back-propagating target loss through the network", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "(without dropout) makes optimizing the model harder. However, in our setting, simultaneously back-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 507, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 104, + 507, + 506, + 522 + ], + "score": 1.0, + "content": "propagating target loss through both networks yields both performance gain as well as convergence", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 518, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 104, + 518, + 506, + 532 + ], + "score": 1.0, + "content": "gain. We believe convergence is faster for our regularization because network weights are more", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "likely to get target based updates from back-propagation in our case. This is especially true for", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 540, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 506, + 554 + ], + "score": 1.0, + "content": "weight dropout (Wan et al., 2013) since in this case dropped weights do not get updated in the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 552, + 177, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 177, + 565 + ], + "score": 1.0, + "content": "training iteration.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30 + }, + { + "type": "title", + "bbox": [ + 108, + 576, + 234, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 576, + 236, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 236, + 589 + ], + "score": 1.0, + "content": "3.2 RELATION TO Π-MODEL", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 597, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 229, + 610 + ], + "score": 1.0, + "content": "Laine & Aila (2016) propose", + "type": "text" + }, + { + "bbox": [ + 230, + 597, + 238, + 608 + ], + "score": 0.28, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "-model with the goal of improving performance on classification", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "tasks in the semi-supervised setting. They propose a model similar to ours (considering the equiv-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "score": 1.0, + "content": "alent deep feed-forward version of our model) except they apply target loss only on one of the net-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 629, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 104, + 629, + 310, + 644 + ], + "score": 1.0, + "content": "works and use time-dependent weighting function", + "type": "text" + }, + { + "bbox": [ + 310, + 630, + 329, + 642 + ], + "score": 0.91, + "content": "\\omega ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 629, + 426, + 644 + ], + "score": 1.0, + "content": "(while we use constant", + "type": "text" + }, + { + "bbox": [ + 426, + 631, + 442, + 643 + ], + "score": 0.89, + "content": "\\frac { \\kappa } { m T }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 629, + 506, + 644 + ], + "score": 1.0, + "content": "). The intuition", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 642, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 505, + 653 + ], + "score": 1.0, + "content": "in their case is to leverage unlabeled data by using them to minimize the difference in prediction be-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 652, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 505, + 664 + ], + "score": 1.0, + "content": "tween the two copies of the network with different dropout masks. Further, they also test their model", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 663, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 505, + 675 + ], + "score": 1.0, + "content": "in the supervised setting but fail to explain the improvements they obtain by using this regularization.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 108, + 680, + 505, + 713 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 463, + 693 + ], + "score": 1.0, + "content": "We note that in our case we analytically show that minimizing our regularizer (also used in", + "type": "text" + }, + { + "bbox": [ + 464, + 680, + 472, + 690 + ], + "score": 0.26, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 678, + 505, + 693 + ], + "score": 1.0, + "content": "-model)", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 691, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 505, + 703 + ], + "score": 1.0, + "content": "is equivalent to minimizing the variance in the model predictions (Remark 1). Furthermore, we also", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 701, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 701, + 505, + 714 + ], + "score": 1.0, + "content": "show the relation of our regularizer to expectation linear dropout (Proposition 1). In Section 5, we", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 721, + 358, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 720, + 358, + 733 + ], + "spans": [ + { + "bbox": [ + 118, + 720, + 358, + 733 + ], + "score": 1.0, + "content": "2The reasons of our focus on RNNs are described in the appendix.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 81, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 171, + 97 + ], + "score": 1.0, + "content": "Remark 1. Let", + "type": "text" + }, + { + "bbox": [ + 172, + 82, + 181, + 95 + ], + "score": 0.88, + "content": "s _ { i } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 81, + 200, + 97 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 200, + 82, + 209, + 96 + ], + "score": 0.87, + "content": "\\boldsymbol { s } _ { j } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 81, + 322, + 97 + ], + "score": 1.0, + "content": "be i.i.d. dropout masks and", + "type": "text" + }, + { + "bbox": [ + 322, + 82, + 399, + 95 + ], + "score": 0.93, + "content": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) \\in \\mathbb { R } ^ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 81, + 506, + 97 + ], + "score": 1.0, + "content": "be the prediction function", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 213, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 213, + 106 + ], + "score": 1.0, + "content": "as described above. Then,", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 506, + 106 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 108, + 478, + 142 + ], + "lines": [ + { + "bbox": [ + 142, + 108, + 478, + 142 + ], + "spans": [ + { + "bbox": [ + 142, + 108, + 478, + 142 + ], + "score": 0.93, + "content": "\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\theta ) = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] = 2 \\sum _ { q = 1 } ^ { m } \\mathrm { v a r } _ { s _ { i } ^ { t } } ( p _ { q } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) ) .", + "type": "interline_equation", + "image_path": "0a168a922a4cdc23ffbcba643900e9fa7e60d2fa738443aed02c1cf592b69992.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 142, + 108, + 478, + 119.33333333333333 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 142, + 119.33333333333333, + 478, + 130.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 142, + 130.66666666666666, + 478, + 142.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 152, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 152, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 505, + 166 + ], + "score": 1.0, + "content": "Note that a generalization of our approach would be to minimize the difference between the predic-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 163, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 505, + 176 + ], + "score": 1.0, + "content": "tions of the two networks with different data/model augmentations. However, in this paper we focus", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 174, + 389, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 354, + 187 + ], + "score": 1.0, + "content": "on using different dropout masks and experiment mainly with", + "type": "text" + }, + { + "bbox": [ + 355, + 174, + 385, + 186 + ], + "score": 0.47, + "content": "\\mathrm { R N N s } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 174, + 389, + 187 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 152, + 505, + 187 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 201, + 212, + 215 + ], + "lines": [ + { + "bbox": [ + 104, + 200, + 213, + 217 + ], + "spans": [ + { + "bbox": [ + 104, + 200, + 213, + 217 + ], + "score": 1.0, + "content": "3 RELATED WORK", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "title", + "bbox": [ + 107, + 226, + 364, + 239 + ], + "lines": [ + { + "bbox": [ + 104, + 225, + 365, + 241 + ], + "spans": [ + { + "bbox": [ + 104, + 225, + 365, + 241 + ], + "score": 1.0, + "content": "3.1 RELATION TO EXPECTATION LINEAR DROPOUT (ELD)", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 247, + 505, + 291 + ], + "lines": [ + { + "bbox": [ + 105, + 247, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 506, + 260 + ], + "score": 1.0, + "content": "Ma et al. (2016) analytically showed that the expected error (over samples) between a model’s", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "score": 1.0, + "content": "expected prediction over all dropout masks, and the prediction using the average mask, is upper", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "score": 1.0, + "content": "bounded. Based on this result, they propose to explicitly minimize the difference (we have adapted", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 280, + 256, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 256, + 293 + ], + "score": 1.0, + "content": "their regularization to our notations),", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 247, + 506, + 293 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 295, + 364, + 311 + ], + "lines": [ + { + "bbox": [ + 142, + 295, + 364, + 311 + ], + "spans": [ + { + "bbox": [ + 142, + 295, + 364, + 311 + ], + "score": 0.89, + "content": "\\mathcal { R } _ { E L D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) = \\| \\mathbb { E } _ { s } \\left[ \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s ; \\boldsymbol { \\theta } ) \\right] - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\mathbb { E } _ { s } [ s ] ; \\boldsymbol { \\theta } ) \\| _ { 2 }", + "type": "interline_equation", + "image_path": "d5b9503e2f3580f8faa094454aa7daa1f3502bb502f6109681ecc8d523451ea2.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 142, + 295, + 364, + 311 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 314, + 505, + 337 + ], + "lines": [ + { + "bbox": [ + 105, + 313, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 133, + 329 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 317, + 139, + 325 + ], + "score": 0.74, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 313, + 505, + 329 + ], + "score": 1.0, + "content": "is the dropout mask. However, due to feasibility consideration, they instead propose to use", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 326, + 266, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 266, + 338 + ], + "score": 1.0, + "content": "the following regularization in practice,", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 313, + 505, + 338 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 141, + 342, + 375, + 358 + ], + "lines": [ + { + "bbox": [ + 141, + 342, + 375, + 358 + ], + "spans": [ + { + "bbox": [ + 141, + 342, + 375, + 358 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\tilde { \\mathcal { R } } _ { E L D } ( \\mathbf { z } ^ { t } ; \\theta ) = \\mathbb { E } _ { s _ { i } } \\left[ \\Vert \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\mathbb { E } _ { s } [ s ] ; \\theta ) \\Vert _ { 2 } ^ { 2 } \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "1e947ec086b9b0b95ed800567914b15a7c2e4666bcc4d4fde3424b3f5195e5ea.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 141, + 342, + 375, + 358 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 361, + 505, + 439 + ], + "lines": [ + { + "bbox": [ + 106, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "Specifically, this is achieved by feed-forwarding the input twice through the network, with and with-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "out dropout mask, and minimizing the main network loss (with dropout) along with the regulariza-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "tion term specified above (but without back-propagating the gradients through the network without", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 395, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 505, + 407 + ], + "score": 1.0, + "content": "dropout). The goal of Ma et al. (2016) is to minimize the network loss along with the expected dif-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "score": 1.0, + "content": "ference between the prediction from individual dropout mask and the prediction from the expected", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "score": 1.0, + "content": "dropout mask. We note that our regularization objective is upper bounded by the expectation-linear", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 428, + 361, + 440 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 361, + 440 + ], + "score": 1.0, + "content": "dropout regularization as shown below (proof in the appendix).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 361, + 505, + 440 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 442, + 293, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 171, + 458 + ], + "score": 1.0, + "content": "Proposition 1.", + "type": "text" + }, + { + "bbox": [ + 171, + 441, + 291, + 456 + ], + "score": 0.91, + "content": "\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) \\leq 4 \\tilde { \\mathcal { R } } _ { E L D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) .", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 441, + 294, + 458 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 441, + 294, + 458 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 463, + 505, + 563 + ], + "lines": [ + { + "bbox": [ + 106, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 477 + ], + "score": 1.0, + "content": "This result shows that minimizing the ELD objective indirectly minimizes our regularization term.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 475, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 506, + 488 + ], + "score": 1.0, + "content": "Finally as indicated above, they apply the target loss only on the network with dropout. In fact, in our", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 486, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 499 + ], + "score": 1.0, + "content": "own ablation studies (see Section 5) we find that back-propagating target loss through the network", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "(without dropout) makes optimizing the model harder. However, in our setting, simultaneously back-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 507, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 104, + 507, + 506, + 522 + ], + "score": 1.0, + "content": "propagating target loss through both networks yields both performance gain as well as convergence", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 518, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 104, + 518, + 506, + 532 + ], + "score": 1.0, + "content": "gain. We believe convergence is faster for our regularization because network weights are more", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "likely to get target based updates from back-propagation in our case. This is especially true for", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 540, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 506, + 554 + ], + "score": 1.0, + "content": "weight dropout (Wan et al., 2013) since in this case dropped weights do not get updated in the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 552, + 177, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 177, + 565 + ], + "score": 1.0, + "content": "training iteration.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30, + "bbox_fs": [ + 104, + 464, + 506, + 565 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 576, + 234, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 576, + 236, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 236, + 589 + ], + "score": 1.0, + "content": "3.2 RELATION TO Π-MODEL", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 597, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 229, + 610 + ], + "score": 1.0, + "content": "Laine & Aila (2016) propose", + "type": "text" + }, + { + "bbox": [ + 230, + 597, + 238, + 608 + ], + "score": 0.28, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "-model with the goal of improving performance on classification", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "tasks in the semi-supervised setting. They propose a model similar to ours (considering the equiv-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "score": 1.0, + "content": "alent deep feed-forward version of our model) except they apply target loss only on one of the net-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 629, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 104, + 629, + 310, + 644 + ], + "score": 1.0, + "content": "works and use time-dependent weighting function", + "type": "text" + }, + { + "bbox": [ + 310, + 630, + 329, + 642 + ], + "score": 0.91, + "content": "\\omega ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 629, + 426, + 644 + ], + "score": 1.0, + "content": "(while we use constant", + "type": "text" + }, + { + "bbox": [ + 426, + 631, + 442, + 643 + ], + "score": 0.89, + "content": "\\frac { \\kappa } { m T }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 629, + 506, + 644 + ], + "score": 1.0, + "content": "). The intuition", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 642, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 505, + 653 + ], + "score": 1.0, + "content": "in their case is to leverage unlabeled data by using them to minimize the difference in prediction be-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 652, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 505, + 664 + ], + "score": 1.0, + "content": "tween the two copies of the network with different dropout masks. Further, they also test their model", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 663, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 505, + 675 + ], + "score": 1.0, + "content": "in the supervised setting but fail to explain the improvements they obtain by using this regularization.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39, + "bbox_fs": [ + 104, + 597, + 506, + 675 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 680, + 505, + 713 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 463, + 693 + ], + "score": 1.0, + "content": "We note that in our case we analytically show that minimizing our regularizer (also used in", + "type": "text" + }, + { + "bbox": [ + 464, + 680, + 472, + 690 + ], + "score": 0.26, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 678, + 505, + 693 + ], + "score": 1.0, + "content": "-model)", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 691, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 505, + 703 + ], + "score": 1.0, + "content": "is equivalent to minimizing the variance in the model predictions (Remark 1). Furthermore, we also", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 701, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 701, + 505, + 714 + ], + "score": 1.0, + "content": "show the relation of our regularizer to expectation linear dropout (Proposition 1). In Section 5, we", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 429, + 94 + ], + "score": 1.0, + "content": "study the effects of target based loss on both networks, which is not used in the", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 429, + 83, + 438, + 93 + ], + "score": 0.27, + "content": "\\Pi", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 438, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "-model. We find", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "that applying target loss on both the networks leads to significantly faster convergence. Finally,", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "we bring to attention that temporal embedding (another model proposed by Laine & Aila (2016),", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 244, + 127 + ], + "score": 1.0, + "content": "claimed to be a better version of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 244, + 115, + 253, + 126 + ], + "score": 0.31, + "content": "\\Pi", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 253, + 115, + 505, + 127 + ], + "score": 1.0, + "content": "-model for semi-supervised, learning) is intractable in natural", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "language processing applications because storing averaged predictions over all of the time steps", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "would be memory exhaustive (since predictions are usually huge - tens of thousands values). On", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 149, + 504, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 484, + 161 + ], + "score": 1.0, + "content": "a final note, we argue that in the supervised case, using a time-dependent weighting function", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 485, + 149, + 504, + 160 + ], + "score": 0.92, + "content": "\\omega ( t )", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 506, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 215, + 171 + ], + "score": 1.0, + "content": "instead of a constant value", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 215, + 159, + 231, + 172 + ], + "score": 0.9, + "content": "\\frac { \\kappa } { m T }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 231, + 159, + 506, + 172 + ], + "score": 1.0, + "content": "is not needed. Since the ground truth labels are known, we have not", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "observed the problem mentioned by Laine & Aila (2016), that the network gets stuck in a degenerate", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 165, + 194 + ], + "score": 1.0, + "content": "solution when", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 165, + 181, + 184, + 193 + ], + "score": 0.92, + "content": "\\omega ( t )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 185, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "is too large in earlier epochs of training. We note that it is much easier to search", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "for an optimal constant value, which is true in our case, as opposed to tuning the time-dependent", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 145, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 145, + 215 + ], + "score": 1.0, + "content": "function.", + "type": "text", + "cross_page": true + } + ], + "index": 11 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 678, + 505, + 714 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 213 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 429, + 94 + ], + "score": 1.0, + "content": "study the effects of target based loss on both networks, which is not used in the", + "type": "text" + }, + { + "bbox": [ + 429, + 83, + 438, + 93 + ], + "score": 0.27, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "-model. We find", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "that applying target loss on both the networks leads to significantly faster convergence. Finally,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "we bring to attention that temporal embedding (another model proposed by Laine & Aila (2016),", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 244, + 127 + ], + "score": 1.0, + "content": "claimed to be a better version of", + "type": "text" + }, + { + "bbox": [ + 244, + 115, + 253, + 126 + ], + "score": 0.31, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 115, + 505, + 127 + ], + "score": 1.0, + "content": "-model for semi-supervised, learning) is intractable in natural", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "language processing applications because storing averaged predictions over all of the time steps", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "would be memory exhaustive (since predictions are usually huge - tens of thousands values). On", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 149, + 504, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 484, + 161 + ], + "score": 1.0, + "content": "a final note, we argue that in the supervised case, using a time-dependent weighting function", + "type": "text" + }, + { + "bbox": [ + 485, + 149, + 504, + 160 + ], + "score": 0.92, + "content": "\\omega ( t )", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 506, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 215, + 171 + ], + "score": 1.0, + "content": "instead of a constant value", + "type": "text" + }, + { + "bbox": [ + 215, + 159, + 231, + 172 + ], + "score": 0.9, + "content": "\\frac { \\kappa } { m T }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 159, + 506, + 172 + ], + "score": 1.0, + "content": "is not needed. Since the ground truth labels are known, we have not", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "observed the problem mentioned by Laine & Aila (2016), that the network gets stuck in a degenerate", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 165, + 194 + ], + "score": 1.0, + "content": "solution when", + "type": "text" + }, + { + "bbox": [ + 165, + 181, + 184, + 193 + ], + "score": 0.92, + "content": "\\omega ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "is too large in earlier epochs of training. We note that it is much easier to search", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "for an optimal constant value, which is true in our case, as opposed to tuning the time-dependent", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 145, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 145, + 215 + ], + "score": 1.0, + "content": "function.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 108, + 220, + 503, + 253 + ], + "lines": [ + { + "bbox": [ + 106, + 219, + 504, + 233 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 160, + 233 + ], + "score": 1.0, + "content": "Similarity to", + "type": "text" + }, + { + "bbox": [ + 160, + 220, + 169, + 230 + ], + "score": 0.38, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 219, + 504, + 233 + ], + "score": 1.0, + "content": "-model makes our method related to other semi-supervised works, mainly Rasmus", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "et al. (2015) and Sajjadi et al. (2016). Since semi-supervised learning is not a primary focus of this", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 242, + 331, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 331, + 254 + ], + "score": 1.0, + "content": "paper, we refer to Laine & Aila (2016) for more details.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 259, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "Another way to address the gap between the train and evaluation mode of dropout is to perform", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 270, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 506, + 282 + ], + "score": 1.0, + "content": "Monte Carlo sampling of masks and average the predictions during evaluation, and this has been", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "used for feed-forward networks. We find that this technique does not work well for RNNs. The", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 292, + 339, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 339, + 303 + ], + "score": 1.0, + "content": "details of these experiments can be found in the appendix.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 108, + 319, + 200, + 331 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 201, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 201, + 333 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 108, + 344, + 220, + 355 + ], + "lines": [ + { + "bbox": [ + 105, + 343, + 221, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 221, + 357 + ], + "score": 1.0, + "content": "4.1 LANGUAGE MODELS", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 364, + 504, + 408 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 506, + 377 + ], + "score": 1.0, + "content": "In the case of language modeling we test our model3 on two benchmark datasets – Penn Tree-bank", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "(PTB) dataset (Marcus et al., 1993) and WikiText-2 (WT2) dataset (Merity et al., 2016). We use", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 387, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 399 + ], + "score": 1.0, + "content": "preprocessing as specified by Mikolov et al. (2010) (for PTB corpus) and Moses tokenizer Koehn", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 397, + 248, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 248, + 410 + ], + "score": 1.0, + "content": "et al. (2007) (for the WT2 dataset).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "For both datasets we use the AWD-LSTM 3-layer architecture described in Merity et al. (2017a)4", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 505, + 437 + ], + "score": 1.0, + "content": "which we call the baseline model. The number of parameters in the model used for PTB is 24", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 435, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 506, + 449 + ], + "score": 1.0, + "content": "million as compared to 34 million in the case of WT2 because WT2 has a larger vocabulary size", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "for which we use a larger embedding matrix. Apart from those differences, the architectures are", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "score": 1.0, + "content": "identical. When we use fraternal dropout, we simply add our regularization on top of this baseline", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 469, + 137, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 137, + 481 + ], + "score": 1.0, + "content": "model.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5 + }, + { + "type": "title", + "bbox": [ + 108, + 486, + 249, + 498 + ], + "lines": [ + { + "bbox": [ + 106, + 485, + 250, + 499 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 250, + 499 + ], + "score": 1.0, + "content": "Word level Penn Treebank (PTB)", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 503, + 505, + 580 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 504, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 504, + 516 + ], + "score": 1.0, + "content": "Influenced by Melis et al. (2017), our goal here is to make sure that fraternal dropout outperforms", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "score": 1.0, + "content": "existing methods not simply because of extensive hyper-parameter grid search but rather due to its", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 526, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 537 + ], + "score": 1.0, + "content": "regularization effects. Hence, in our experiments we leave a vast majority of hyper-parameters used", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "in the baseline model (Melis et al., 2017) unchanged i.e. embedding and hidden states sizes, gradient", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 546, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 505, + 561 + ], + "score": 1.0, + "content": "clipping value, weight decay and the values used for all dropout layers (dropout on the word vectors,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 557, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 505, + 572 + ], + "score": 1.0, + "content": "the output between LSTM layers, the output of the final LSTM, and embedding dropout). However,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 569, + 224, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 224, + 582 + ], + "score": 1.0, + "content": "a few changes are necessary:", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 132, + 589, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 132, + 590, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 505, + 602 + ], + "score": 1.0, + "content": "• the coefficients for AR and TAR needed to be altered because fraternal dropout also affects", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 141, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 141, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "RNNs activation (as explained in Subsection 5.3) – we did not run grid search to obtain the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 142, + 612, + 388, + 624 + ], + "spans": [ + { + "bbox": [ + 142, + 612, + 388, + 624 + ], + "score": 1.0, + "content": "best values but simply deactivated AR and TAR regularizers;", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 133, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 133, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "• since fraternal dropout needs twice as much memory, batch size is halved so the model", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 141, + 638, + 476, + 649 + ], + "spans": [ + { + "bbox": [ + 141, + 638, + 476, + 649 + ], + "score": 1.0, + "content": "needs approximately the same amount of memory and hence fits on the same GPU.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 108, + 659, + 505, + 692 + ], + "lines": [ + { + "bbox": [ + 106, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 436, + 671 + ], + "score": 1.0, + "content": "The final change in hyper-parameters is to alter the non-monotone interval", + "type": "text" + }, + { + "bbox": [ + 436, + 660, + 444, + 668 + ], + "score": 0.64, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "used in non-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 670, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 505, + 681 + ], + "score": 1.0, + "content": "monotonically triggered averaged SGD (NT-ASGD) optimizer Polyak & Juditsky (1992); Mandt", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 680, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 344, + 694 + ], + "score": 1.0, + "content": "et al. (2017); Melis et al. (2017). We run a grid search on", + "type": "text" + }, + { + "bbox": [ + 344, + 680, + 438, + 693 + ], + "score": 0.91, + "content": "n \\in \\{ 5 , 2 5 , 4 0 , 5 0 , 6 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 680, + 505, + 694 + ], + "score": 1.0, + "content": "and obtain very", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 701, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 699, + 405, + 712 + ], + "spans": [ + { + "bbox": [ + 118, + 699, + 405, + 712 + ], + "score": 1.0, + "content": "3Our code is available at github.com/kondiz/fraternal-dropout .", + "type": "text" + } + ] + }, + { + "bbox": [ + 117, + 709, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 117, + 709, + 505, + 724 + ], + "score": 1.0, + "content": "4We used the official GitHub repository code for this paper github.com/salesforce/", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 721, + 174, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 174, + 732 + ], + "score": 1.0, + "content": "awd-lstm-lm .", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 213 + ], + "lines": [], + "index": 5.5, + "bbox_fs": [ + 105, + 82, + 506, + 215 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 108, + 220, + 503, + 253 + ], + "lines": [ + { + "bbox": [ + 106, + 219, + 504, + 233 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 160, + 233 + ], + "score": 1.0, + "content": "Similarity to", + "type": "text" + }, + { + "bbox": [ + 160, + 220, + 169, + 230 + ], + "score": 0.38, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 219, + 504, + 233 + ], + "score": 1.0, + "content": "-model makes our method related to other semi-supervised works, mainly Rasmus", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "et al. (2015) and Sajjadi et al. (2016). Since semi-supervised learning is not a primary focus of this", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 242, + 331, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 331, + 254 + ], + "score": 1.0, + "content": "paper, we refer to Laine & Aila (2016) for more details.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 219, + 505, + 254 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 259, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "Another way to address the gap between the train and evaluation mode of dropout is to perform", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 270, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 506, + 282 + ], + "score": 1.0, + "content": "Monte Carlo sampling of masks and average the predictions during evaluation, and this has been", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "used for feed-forward networks. We find that this technique does not work well for RNNs. The", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 292, + 339, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 339, + 303 + ], + "score": 1.0, + "content": "details of these experiments can be found in the appendix.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 259, + 506, + 303 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 319, + 200, + 331 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 201, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 201, + 333 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 108, + 344, + 220, + 355 + ], + "lines": [ + { + "bbox": [ + 105, + 343, + 221, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 221, + 357 + ], + "score": 1.0, + "content": "4.1 LANGUAGE MODELS", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 364, + 504, + 408 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 506, + 377 + ], + "score": 1.0, + "content": "In the case of language modeling we test our model3 on two benchmark datasets – Penn Tree-bank", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "(PTB) dataset (Marcus et al., 1993) and WikiText-2 (WT2) dataset (Merity et al., 2016). We use", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 387, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 399 + ], + "score": 1.0, + "content": "preprocessing as specified by Mikolov et al. (2010) (for PTB corpus) and Moses tokenizer Koehn", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 397, + 248, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 248, + 410 + ], + "score": 1.0, + "content": "et al. (2007) (for the WT2 dataset).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 364, + 506, + 410 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "For both datasets we use the AWD-LSTM 3-layer architecture described in Merity et al. (2017a)4", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 505, + 437 + ], + "score": 1.0, + "content": "which we call the baseline model. The number of parameters in the model used for PTB is 24", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 435, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 506, + 449 + ], + "score": 1.0, + "content": "million as compared to 34 million in the case of WT2 because WT2 has a larger vocabulary size", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "for which we use a larger embedding matrix. Apart from those differences, the architectures are", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "score": 1.0, + "content": "identical. When we use fraternal dropout, we simply add our regularization on top of this baseline", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 469, + 137, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 137, + 481 + ], + "score": 1.0, + "content": "model.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 414, + 506, + 481 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 486, + 249, + 498 + ], + "lines": [ + { + "bbox": [ + 106, + 485, + 250, + 499 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 250, + 499 + ], + "score": 1.0, + "content": "Word level Penn Treebank (PTB)", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 503, + 505, + 580 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 504, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 504, + 516 + ], + "score": 1.0, + "content": "Influenced by Melis et al. (2017), our goal here is to make sure that fraternal dropout outperforms", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "score": 1.0, + "content": "existing methods not simply because of extensive hyper-parameter grid search but rather due to its", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 526, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 537 + ], + "score": 1.0, + "content": "regularization effects. Hence, in our experiments we leave a vast majority of hyper-parameters used", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "in the baseline model (Melis et al., 2017) unchanged i.e. embedding and hidden states sizes, gradient", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 546, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 505, + 561 + ], + "score": 1.0, + "content": "clipping value, weight decay and the values used for all dropout layers (dropout on the word vectors,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 557, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 505, + 572 + ], + "score": 1.0, + "content": "the output between LSTM layers, the output of the final LSTM, and embedding dropout). However,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 569, + 224, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 224, + 582 + ], + "score": 1.0, + "content": "a few changes are necessary:", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 503, + 506, + 582 + ] + }, + { + "type": "text", + "bbox": [ + 132, + 589, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 132, + 590, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 505, + 602 + ], + "score": 1.0, + "content": "• the coefficients for AR and TAR needed to be altered because fraternal dropout also affects", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 141, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 141, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "RNNs activation (as explained in Subsection 5.3) – we did not run grid search to obtain the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 142, + 612, + 388, + 624 + ], + "spans": [ + { + "bbox": [ + 142, + 612, + 388, + 624 + ], + "score": 1.0, + "content": "best values but simply deactivated AR and TAR regularizers;", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 133, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 133, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "• since fraternal dropout needs twice as much memory, batch size is halved so the model", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 141, + 638, + 476, + 649 + ], + "spans": [ + { + "bbox": [ + 141, + 638, + 476, + 649 + ], + "score": 1.0, + "content": "needs approximately the same amount of memory and hence fits on the same GPU.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41, + "bbox_fs": [ + 132, + 590, + 506, + 649 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 659, + 505, + 692 + ], + "lines": [ + { + "bbox": [ + 106, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 436, + 671 + ], + "score": 1.0, + "content": "The final change in hyper-parameters is to alter the non-monotone interval", + "type": "text" + }, + { + "bbox": [ + 436, + 660, + 444, + 668 + ], + "score": 0.64, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "used in non-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 670, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 505, + 681 + ], + "score": 1.0, + "content": "monotonically triggered averaged SGD (NT-ASGD) optimizer Polyak & Juditsky (1992); Mandt", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 680, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 344, + 694 + ], + "score": 1.0, + "content": "et al. (2017); Melis et al. (2017). We run a grid search on", + "type": "text" + }, + { + "bbox": [ + 344, + 680, + 438, + 693 + ], + "score": 0.91, + "content": "n \\in \\{ 5 , 2 5 , 4 0 , 5 0 , 6 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 680, + 505, + 694 + ], + "score": 1.0, + "content": "and obtain very", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 375, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 505, + 387 + ], + "score": 1.0, + "content": "similar results for the largest values (40, 50 and 60) in the candidate set. Hence, our model is trained", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 385, + 479, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 479, + 398 + ], + "score": 1.0, + "content": "longer using ordinary SGD optimizer as compared to the baseline model (Melis et al., 2017).", + "type": "text", + "cross_page": true + } + ], + "index": 9 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 658, + 505, + 694 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 78, + 501, + 213 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 111, + 78, + 501, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 78, + 501, + 213 + ], + "spans": [ + { + "bbox": [ + 111, + 78, + 501, + 213 + ], + "score": 0.975, + "html": "
ModelParametersValidationTest
Zaremba et al. (2014) - LSTM (medium)10M86.282.7
Zaremba et al. (2014) - LSTM (large)24M82.278.4
Gal & Ghahramani (2016) - Variational LSTM (medium)20M81.979.7
Gal & Ghahramani (2016) - Variational LSTM (large)66M77.975.2
Inan et al. (2016) - Variational LSTM51M71.168.5
Inan et al. (2016) - Variational RHN24M68.166.0
Zilly et al. (2016) - Variational RHN23M67.965.4
Melis et al. (2017) - 5-layer RHN24M64.862.2
Melis et al. (2017) - 4-layer skip connection LSTM24M60.958.3
Merity etal. (2017a) - AWD-LSTM3-layer (baseline)24M60.057.3
Fraternal dropout + AWD-LSTM3-layer24M58.956.8
", + "type": "table", + "image_path": "58fbc95757bda8d8b5297de81e2a587e22f8bcf476f9fcb6913b30e2a0ced2ed.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 78, + 501, + 123.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 123.0, + 501, + 168.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 168.0, + 501, + 213.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 156, + 217, + 453, + 229 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 156, + 217, + 453, + 231 + ], + "spans": [ + { + "bbox": [ + 156, + 217, + 453, + 231 + ], + "score": 1.0, + "content": "Table 1: Perplexity on Penn Treebank word level language modeling task.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "table", + "bbox": [ + 114, + 236, + 497, + 339 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 114, + 236, + 497, + 339 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 114, + 236, + 497, + 339 + ], + "spans": [ + { + "bbox": [ + 114, + 236, + 497, + 339 + ], + "score": 0.976, + "html": "
ModelParametersValidationTest
Merity et al. (2016) - Variational LSTM + Zoneout20M108.7100.9
Merity et al. (2016) - Variational LSTM20M101.796.3
Inan et al. (2016) - Variational LSTM28M91.587.0
Melis et al. (2017) - 5-layer RHN24M78.175.6
Melis et al. (2017) - 1-layer LSTM24M69.365.9
Melis et al. (2017) - 2-layer skip connection LSTM24M69.165.9
Merity et al. (2017a) - AWD-LSTM3-layer (baseline)34M68.665.8
Fraternal dropout + AWD-LSTM 3-layer34M66.864.1
", + "type": "table", + "image_path": "9f4bf8304f3c42f131fdef6f693f45615c4ab57d1fc38de924e0cfd7fb9a8e70.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 114, + 236, + 497, + 270.3333333333333 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 114, + 270.3333333333333, + 497, + 304.66666666666663 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 114, + 304.66666666666663, + 497, + 338.99999999999994 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 164, + 344, + 446, + 355 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 163, + 342, + 447, + 357 + ], + "spans": [ + { + "bbox": [ + 163, + 342, + 447, + 357 + ], + "score": 1.0, + "content": "Table 2: Perplexity on WikiText-2 word level language modeling task.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 108, + 375, + 504, + 397 + ], + "lines": [ + { + "bbox": [ + 106, + 375, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 505, + 387 + ], + "score": 1.0, + "content": "similar results for the largest values (40, 50 and 60) in the candidate set. Hence, our model is trained", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 385, + 479, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 479, + 398 + ], + "score": 1.0, + "content": "longer using ordinary SGD optimizer as compared to the baseline model (Melis et al., 2017).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 403, + 505, + 436 + ], + "lines": [ + { + "bbox": [ + 106, + 402, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 506, + 416 + ], + "score": 1.0, + "content": "We evaluate our model using the perplexity metric and compare the results that we obtain against", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "the existing state-of-the-art results. The results are reported in Table 1. Our approach achieves the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 425, + 370, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 370, + 437 + ], + "score": 1.0, + "content": "state-of-the-art performance compared with existing benchmarks.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 442, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 106, + 441, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 455 + ], + "score": 1.0, + "content": "To confirm that the gains are robust to initialization, we run ten experiments for the baseline model", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 453, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 505, + 465 + ], + "score": 1.0, + "content": "with different seeds (without fine-tuning) for PTB dataset to compute confidence intervals. The", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 463, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 252, + 477 + ], + "score": 1.0, + "content": "average best validation perplexity is", + "type": "text" + }, + { + "bbox": [ + 253, + 464, + 306, + 475 + ], + "score": 0.89, + "content": "6 0 . 6 4 \\pm 0 . 1 5", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 463, + 505, + 477 + ], + "score": 1.0, + "content": "with the minimum value equals 60.33. The same", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 189, + 488 + ], + "score": 1.0, + "content": "for test perplexity is", + "type": "text" + }, + { + "bbox": [ + 190, + 475, + 244, + 486 + ], + "score": 0.88, + "content": "5 8 . 3 2 \\pm 0 . 1 4", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "and 58.05, respectively. Our score (59.8 validation and 58.0 test", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 486, + 311, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 311, + 499 + ], + "score": 1.0, + "content": "perplexity) beats ordinal dropout minimum values.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 108, + 502, + 505, + 535 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 515 + ], + "score": 1.0, + "content": "We also perform experiments using fraternal dropout with a grid search on all the hyper-parameters", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 513, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 526 + ], + "score": 1.0, + "content": "and find that it leads to further improvements in performance. The details of this experiment can be", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 525, + 191, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 191, + 536 + ], + "score": 1.0, + "content": "found in section 5.5.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 108, + 541, + 235, + 552 + ], + "lines": [ + { + "bbox": [ + 106, + 540, + 235, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 235, + 555 + ], + "score": 1.0, + "content": "Word level WikiText-2 (WT2)", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 558, + 505, + 624 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 572 + ], + "score": 1.0, + "content": "In the case of WikiText-2 language modeling task, we outperform the current state-of-the-art using", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "the perplexity metric by a significant margin. Due to the lack of computational power, we run a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 580, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 506, + 593 + ], + "score": 1.0, + "content": "single training procedure for fraternal dropout on WT2 dataset because it is larger than PTB. In", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 405, + 604 + ], + "score": 1.0, + "content": "this experiment, we use the best hyper-parameters found for PTB dataset", + "type": "text" + }, + { + "bbox": [ + 406, + 592, + 440, + 602 + ], + "score": 0.86, + "content": "\\kappa = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 591, + 505, + 604 + ], + "score": 1.0, + "content": ", non-monotone", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 602, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 139, + 615 + ], + "score": 1.0, + "content": "interval", + "type": "text" + }, + { + "bbox": [ + 139, + 603, + 170, + 613 + ], + "score": 0.89, + "content": "n = 6 0", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 602, + 505, + 615 + ], + "score": 1.0, + "content": "and halved batch size; the rest of the hyper-parameters are the same as described in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 613, + 393, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 393, + 625 + ], + "score": 1.0, + "content": "Melis et al. (2017) for WT2). The final results are presented in Table 2.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 108, + 638, + 217, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 637, + 218, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 218, + 651 + ], + "score": 1.0, + "content": "4.2 IMAGE CAPTIONING", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 659, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "We also apply fraternal dropout on an image captioning task. We use the well-known show and tell", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "score": 1.0, + "content": "model as a baseline5 (Vinyals et al., 2014). We emphasize that in the image captioning task, the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 681, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 693 + ], + "score": 1.0, + "content": "image encoder and sentence decoder architectures are usually learned together. Since we want to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 691, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 505, + 704 + ], + "score": 1.0, + "content": "focus on the benefits of using fraternal dropout in RNNs we use frozen pretrained ResNet-101 (He", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 712, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 117, + 708, + 505, + 725 + ], + "spans": [ + { + "bbox": [ + 117, + 708, + 505, + 725 + ], + "score": 1.0, + "content": "5We used PyTorch implementation with default hyper-parameters from github.com/ruotianluo/", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 721, + 216, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 216, + 732 + ], + "score": 1.0, + "content": "neuraltalk2.pytorch .", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 78, + 501, + 213 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 111, + 78, + 501, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 78, + 501, + 213 + ], + "spans": [ + { + "bbox": [ + 111, + 78, + 501, + 213 + ], + "score": 0.975, + "html": "
ModelParametersValidationTest
Zaremba et al. (2014) - LSTM (medium)10M86.282.7
Zaremba et al. (2014) - LSTM (large)24M82.278.4
Gal & Ghahramani (2016) - Variational LSTM (medium)20M81.979.7
Gal & Ghahramani (2016) - Variational LSTM (large)66M77.975.2
Inan et al. (2016) - Variational LSTM51M71.168.5
Inan et al. (2016) - Variational RHN24M68.166.0
Zilly et al. (2016) - Variational RHN23M67.965.4
Melis et al. (2017) - 5-layer RHN24M64.862.2
Melis et al. (2017) - 4-layer skip connection LSTM24M60.958.3
Merity etal. (2017a) - AWD-LSTM3-layer (baseline)24M60.057.3
Fraternal dropout + AWD-LSTM3-layer24M58.956.8
", + "type": "table", + "image_path": "58fbc95757bda8d8b5297de81e2a587e22f8bcf476f9fcb6913b30e2a0ced2ed.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 78, + 501, + 123.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 123.0, + 501, + 168.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 168.0, + 501, + 213.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 156, + 217, + 453, + 229 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 156, + 217, + 453, + 231 + ], + "spans": [ + { + "bbox": [ + 156, + 217, + 453, + 231 + ], + "score": 1.0, + "content": "Table 1: Perplexity on Penn Treebank word level language modeling task.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "table", + "bbox": [ + 114, + 236, + 497, + 339 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 114, + 236, + 497, + 339 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 114, + 236, + 497, + 339 + ], + "spans": [ + { + "bbox": [ + 114, + 236, + 497, + 339 + ], + "score": 0.976, + "html": "
ModelParametersValidationTest
Merity et al. (2016) - Variational LSTM + Zoneout20M108.7100.9
Merity et al. (2016) - Variational LSTM20M101.796.3
Inan et al. (2016) - Variational LSTM28M91.587.0
Melis et al. (2017) - 5-layer RHN24M78.175.6
Melis et al. (2017) - 1-layer LSTM24M69.365.9
Melis et al. (2017) - 2-layer skip connection LSTM24M69.165.9
Merity et al. (2017a) - AWD-LSTM3-layer (baseline)34M68.665.8
Fraternal dropout + AWD-LSTM 3-layer34M66.864.1
", + "type": "table", + "image_path": "9f4bf8304f3c42f131fdef6f693f45615c4ab57d1fc38de924e0cfd7fb9a8e70.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 114, + 236, + 497, + 270.3333333333333 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 114, + 270.3333333333333, + 497, + 304.66666666666663 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 114, + 304.66666666666663, + 497, + 338.99999999999994 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 164, + 344, + 446, + 355 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 163, + 342, + 447, + 357 + ], + "spans": [ + { + "bbox": [ + 163, + 342, + 447, + 357 + ], + "score": 1.0, + "content": "Table 2: Perplexity on WikiText-2 word level language modeling task.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 108, + 375, + 504, + 397 + ], + "lines": [], + "index": 8.5, + "bbox_fs": [ + 105, + 375, + 505, + 398 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 403, + 505, + 436 + ], + "lines": [ + { + "bbox": [ + 106, + 402, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 506, + 416 + ], + "score": 1.0, + "content": "We evaluate our model using the perplexity metric and compare the results that we obtain against", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "the existing state-of-the-art results. The results are reported in Table 1. Our approach achieves the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 425, + 370, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 370, + 437 + ], + "score": 1.0, + "content": "state-of-the-art performance compared with existing benchmarks.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 402, + 506, + 437 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 442, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 106, + 441, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 455 + ], + "score": 1.0, + "content": "To confirm that the gains are robust to initialization, we run ten experiments for the baseline model", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 453, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 505, + 465 + ], + "score": 1.0, + "content": "with different seeds (without fine-tuning) for PTB dataset to compute confidence intervals. The", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 463, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 252, + 477 + ], + "score": 1.0, + "content": "average best validation perplexity is", + "type": "text" + }, + { + "bbox": [ + 253, + 464, + 306, + 475 + ], + "score": 0.89, + "content": "6 0 . 6 4 \\pm 0 . 1 5", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 463, + 505, + 477 + ], + "score": 1.0, + "content": "with the minimum value equals 60.33. The same", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 189, + 488 + ], + "score": 1.0, + "content": "for test perplexity is", + "type": "text" + }, + { + "bbox": [ + 190, + 475, + 244, + 486 + ], + "score": 0.88, + "content": "5 8 . 3 2 \\pm 0 . 1 4", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "and 58.05, respectively. Our score (59.8 validation and 58.0 test", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 486, + 311, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 311, + 499 + ], + "score": 1.0, + "content": "perplexity) beats ordinal dropout minimum values.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 441, + 505, + 499 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 502, + 505, + 535 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 515 + ], + "score": 1.0, + "content": "We also perform experiments using fraternal dropout with a grid search on all the hyper-parameters", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 513, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 526 + ], + "score": 1.0, + "content": "and find that it leads to further improvements in performance. The details of this experiment can be", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 525, + 191, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 191, + 536 + ], + "score": 1.0, + "content": "found in section 5.5.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 502, + 505, + 536 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 541, + 235, + 552 + ], + "lines": [ + { + "bbox": [ + 106, + 540, + 235, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 235, + 555 + ], + "score": 1.0, + "content": "Word level WikiText-2 (WT2)", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 558, + 505, + 624 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 572 + ], + "score": 1.0, + "content": "In the case of WikiText-2 language modeling task, we outperform the current state-of-the-art using", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "the perplexity metric by a significant margin. Due to the lack of computational power, we run a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 580, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 506, + 593 + ], + "score": 1.0, + "content": "single training procedure for fraternal dropout on WT2 dataset because it is larger than PTB. In", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 405, + 604 + ], + "score": 1.0, + "content": "this experiment, we use the best hyper-parameters found for PTB dataset", + "type": "text" + }, + { + "bbox": [ + 406, + 592, + 440, + 602 + ], + "score": 0.86, + "content": "\\kappa = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 591, + 505, + 604 + ], + "score": 1.0, + "content": ", non-monotone", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 602, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 139, + 615 + ], + "score": 1.0, + "content": "interval", + "type": "text" + }, + { + "bbox": [ + 139, + 603, + 170, + 613 + ], + "score": 0.89, + "content": "n = 6 0", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 602, + 505, + 615 + ], + "score": 1.0, + "content": "and halved batch size; the rest of the hyper-parameters are the same as described in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 613, + 393, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 393, + 625 + ], + "score": 1.0, + "content": "Melis et al. (2017) for WT2). The final results are presented in Table 2.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 556, + 506, + 625 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 638, + 217, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 637, + 218, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 218, + 651 + ], + "score": 1.0, + "content": "4.2 IMAGE CAPTIONING", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 659, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "We also apply fraternal dropout on an image captioning task. We use the well-known show and tell", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "score": 1.0, + "content": "model as a baseline5 (Vinyals et al., 2014). We emphasize that in the image captioning task, the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 681, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 693 + ], + "score": 1.0, + "content": "image encoder and sentence decoder architectures are usually learned together. Since we want to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 691, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 505, + 704 + ], + "score": 1.0, + "content": "focus on the benefits of using fraternal dropout in RNNs we use frozen pretrained ResNet-101 (He", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 184, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 184, + 505, + 196 + ], + "score": 1.0, + "content": "et al., 2015) model as our image encoder. It means that our results are not directly comparable with", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 195, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 505, + 207 + ], + "score": 1.0, + "content": "other state-of-the-art methods, however we report results for the original methods so readers can see", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 206, + 402, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 206, + 402, + 218 + ], + "score": 1.0, + "content": "that our baseline performs well. The final results are presented in Table 3.", + "type": "text", + "cross_page": true + } + ], + "index": 7 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 658, + 506, + 704 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 77, + 468, + 136 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 141, + 77, + 468, + 136 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 141, + 77, + 468, + 136 + ], + "spans": [ + { + "bbox": [ + 141, + 77, + 468, + 136 + ], + "score": 0.977, + "html": "
ModelBLEU-1BLEU-2BLEU-3BLEU-4
Show and Tell Xu et al. (2015)66.646.132.924.6
Baseline68.850.836.125.6
Fraternal dropout, K = 0.01569.351.436.626.1
Fraternal dropout, K = 0.00569.351.536.926.3
", + "type": "table", + "image_path": "e8caa9152ecfc21710dfc1a2d30cc6455c9f4a6ee49165aa732df86c66312e73.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 141, + 77, + 468, + 96.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 141, + 96.66666666666667, + 468, + 116.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 141, + 116.33333333333334, + 468, + 136.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 104, + 140, + 505, + 163 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 140, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 505, + 153 + ], + "score": 1.0, + "content": "Table 3: BLEU scores for the Microsoft COCO image captioning task. Using fraternal dropout is", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 151, + 426, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 426, + 164 + ], + "score": 1.0, + "content": "the only difference between models. The rest of hyper-parameters are the same.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 183, + 505, + 217 + ], + "lines": [ + { + "bbox": [ + 105, + 184, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 184, + 505, + 196 + ], + "score": 1.0, + "content": "et al., 2015) model as our image encoder. It means that our results are not directly comparable with", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 195, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 505, + 207 + ], + "score": 1.0, + "content": "other state-of-the-art methods, however we report results for the original methods so readers can see", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 206, + 402, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 206, + 402, + 218 + ], + "score": 1.0, + "content": "that our baseline performs well. The final results are presented in Table 3.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 222, + 505, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 238, + 235 + ], + "score": 1.0, + "content": "We argue that in this task smaller", + "type": "text" + }, + { + "bbox": [ + 238, + 225, + 245, + 232 + ], + "score": 0.74, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "values are optimal because the image captioning encoder is given", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 233, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 505, + 246 + ], + "score": 1.0, + "content": "all information in the beginning and hence the variance of consecutive predictions is smaller that in", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "unconditioned natural language processing tasks. Fraternal dropout may benefits here mainly due", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 255, + 453, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 453, + 268 + ], + "score": 1.0, + "content": "to averaging gradients for different mask and hence updating weights more frequently.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 108, + 284, + 228, + 296 + ], + "lines": [ + { + "bbox": [ + 105, + 282, + 230, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 230, + 299 + ], + "score": 1.0, + "content": "5 ABLATION STUDIES", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 309, + 505, + 353 + ], + "lines": [ + { + "bbox": [ + 106, + 309, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 505, + 321 + ], + "score": 1.0, + "content": "In this section, the goal is to study existing methods closely related to ours – expectation linear", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 321, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 505, + 332 + ], + "score": 1.0, + "content": "dropout Ma et al. (2016), Π-model Laine & Aila (2016) and activity regularization Merity et al.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "(2017b). All of our experiments for ablation studies, which apply a single layer LSTM, use the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 342, + 389, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 389, + 354 + ], + "score": 1.0, + "content": "same hyper-parameters and model architecture6 as Melis et al. (2017).", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 108, + 368, + 302, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 303, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 303, + 382 + ], + "score": 1.0, + "content": "5.1 EXPECTATION-LINEAR DROPOUT (ELD)", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 388, + 505, + 488 + ], + "lines": [ + { + "bbox": [ + 106, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "The relation with expectation-linear dropout Ma et al. (2016) has been discussed in Section 2. Here", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 399, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 506, + 412 + ], + "score": 1.0, + "content": "we perform experiments to study the difference in performance when using the ELD regularization", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "versus our regularization (FD). In addition to ELD, we also study a modification (ELDM) of ELD", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 420, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 505, + 436 + ], + "score": 1.0, + "content": "which applies target loss to both copies of LSTMs in ELD similar to FD (notice in their case they", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 433, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 506, + 446 + ], + "score": 1.0, + "content": "only have dropout on one LSTM). Finally we also evaluate a baseline model without any of these", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "regularizations. The learning dynamics curves are shown in Figure 1. Our regularization performs", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "better in terms of convergence compared with other methods. In terms of generalization, we find", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "score": 1.0, + "content": "that FD is similar to ELD, but baseline and ELDM are much worse. Interestingly, looking at the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 477, + 486, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 486, + 489 + ], + "score": 1.0, + "content": "train and validation curves together, ELDM seems to be suffering from optimization problems.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 108, + 502, + 173, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 175, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 175, + 515 + ], + "score": 1.0, + "content": "5.2 Π-MODEL", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 523, + 505, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "Since Π-model Laine & Aila (2016) is similar to our algorithm (even though it is designed for", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 535, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 505, + 546 + ], + "score": 1.0, + "content": "semi-supervised learning in feed-forward networks), we study the difference in performance with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 544, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 558 + ], + "score": 1.0, + "content": "Π-model7 both qualitatively and quantitatively to establish the advantage of our approach. First,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "we run both single layer LSTM and 3-layer AWD-LSTM on PTB task to check how their model", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 567, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 505, + 579 + ], + "score": 1.0, + "content": "compares with ours in the case of language modeling. The results are shown in Figure 1 and 2. We", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 104, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "find that our model converges significantly faster than Π-model. We believe this happens because we", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 589, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 505, + 601 + ], + "score": 1.0, + "content": "back-propagate the target loss through both networks (in contrast to Π-model) that leads to weights", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 601, + 333, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 333, + 612 + ], + "score": 1.0, + "content": "getting updated using target-based gradients more often.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 621, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 117, + 618, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 117, + 618, + 506, + 635 + ], + "score": 1.0, + "content": "6We use a batch size of 64, truncated back-propagation with 35 time steps, a constant zero state is provided", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "as the initial state with probability 0.01 (similar to Melis et al. (2017)), SGD with learning rate 30 (no mo-", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 641, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 505, + 653 + ], + "score": 1.0, + "content": "mentum) which is multiplied by 0.1 whenever validation performance does not improve ever during 20 epochs,", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "weight dropout on the hidden to hidden matrix 0.5, dropout every word in a mini-batch with probability 0.1,", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "embedding dropout 0.65, output dropout 0.4 (final value of LSTM), gradient clipping of 0.25, weight decay", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 669, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 150, + 681 + ], + "score": 0.92, + "content": "1 . 2 \\times 1 0 ^ { - 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 669, + 506, + 683 + ], + "score": 1.0, + "content": ", input embedding size of 655, the input/output size of LSTM is the same as embedding size (655)", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 680, + 380, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 380, + 693 + ], + "score": 1.0, + "content": "and the embedding weights are tied (Inan et al., 2016; Press & Wolf, 2016).", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 692, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 689, + 506, + 704 + ], + "spans": [ + { + "bbox": [ + 118, + 689, + 223, + 704 + ], + "score": 1.0, + "content": "7We use a constant function", + "type": "text" + }, + { + "bbox": [ + 223, + 691, + 268, + 703 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\omega ( t ) = \\frac { \\kappa } { m T } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 689, + 506, + 704 + ], + "score": 1.0, + "content": "as a coefficient for Π-model (similar to our regularization term).", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 702, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 505, + 712 + ], + "score": 1.0, + "content": "Hence, the focus of our experiment is to evaluate the difference in performance when target loss is back-", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "propagated through one of the networks (Π-model) vs. both (ours). Additionally, we find that tuning a function", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 722, + 291, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 722, + 291, + 732 + ], + "score": 1.0, + "content": "instead of using a constant coefficient is infeasible.", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 77, + 468, + 136 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 141, + 77, + 468, + 136 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 141, + 77, + 468, + 136 + ], + "spans": [ + { + "bbox": [ + 141, + 77, + 468, + 136 + ], + "score": 0.977, + "html": "
ModelBLEU-1BLEU-2BLEU-3BLEU-4
Show and Tell Xu et al. (2015)66.646.132.924.6
Baseline68.850.836.125.6
Fraternal dropout, K = 0.01569.351.436.626.1
Fraternal dropout, K = 0.00569.351.536.926.3
", + "type": "table", + "image_path": "e8caa9152ecfc21710dfc1a2d30cc6455c9f4a6ee49165aa732df86c66312e73.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 141, + 77, + 468, + 96.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 141, + 96.66666666666667, + 468, + 116.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 141, + 116.33333333333334, + 468, + 136.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 104, + 140, + 505, + 163 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 140, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 505, + 153 + ], + "score": 1.0, + "content": "Table 3: BLEU scores for the Microsoft COCO image captioning task. Using fraternal dropout is", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 151, + 426, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 426, + 164 + ], + "score": 1.0, + "content": "the only difference between models. The rest of hyper-parameters are the same.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 183, + 505, + 217 + ], + "lines": [], + "index": 6, + "bbox_fs": [ + 105, + 184, + 505, + 218 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 222, + 505, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 238, + 235 + ], + "score": 1.0, + "content": "We argue that in this task smaller", + "type": "text" + }, + { + "bbox": [ + 238, + 225, + 245, + 232 + ], + "score": 0.74, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "values are optimal because the image captioning encoder is given", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 233, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 505, + 246 + ], + "score": 1.0, + "content": "all information in the beginning and hence the variance of consecutive predictions is smaller that in", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "unconditioned natural language processing tasks. Fraternal dropout may benefits here mainly due", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 255, + 453, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 453, + 268 + ], + "score": 1.0, + "content": "to averaging gradients for different mask and hence updating weights more frequently.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 222, + 505, + 268 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 284, + 228, + 296 + ], + "lines": [ + { + "bbox": [ + 105, + 282, + 230, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 230, + 299 + ], + "score": 1.0, + "content": "5 ABLATION STUDIES", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 309, + 505, + 353 + ], + "lines": [ + { + "bbox": [ + 106, + 309, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 505, + 321 + ], + "score": 1.0, + "content": "In this section, the goal is to study existing methods closely related to ours – expectation linear", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 321, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 505, + 332 + ], + "score": 1.0, + "content": "dropout Ma et al. (2016), Π-model Laine & Aila (2016) and activity regularization Merity et al.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "(2017b). All of our experiments for ablation studies, which apply a single layer LSTM, use the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 342, + 389, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 389, + 354 + ], + "score": 1.0, + "content": "same hyper-parameters and model architecture6 as Melis et al. (2017).", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 309, + 505, + 354 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 368, + 302, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 303, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 303, + 382 + ], + "score": 1.0, + "content": "5.1 EXPECTATION-LINEAR DROPOUT (ELD)", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 388, + 505, + 488 + ], + "lines": [ + { + "bbox": [ + 106, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "The relation with expectation-linear dropout Ma et al. (2016) has been discussed in Section 2. Here", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 399, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 506, + 412 + ], + "score": 1.0, + "content": "we perform experiments to study the difference in performance when using the ELD regularization", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "versus our regularization (FD). In addition to ELD, we also study a modification (ELDM) of ELD", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 420, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 505, + 436 + ], + "score": 1.0, + "content": "which applies target loss to both copies of LSTMs in ELD similar to FD (notice in their case they", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 433, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 506, + 446 + ], + "score": 1.0, + "content": "only have dropout on one LSTM). Finally we also evaluate a baseline model without any of these", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "regularizations. The learning dynamics curves are shown in Figure 1. Our regularization performs", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "better in terms of convergence compared with other methods. In terms of generalization, we find", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "score": 1.0, + "content": "that FD is similar to ELD, but baseline and ELDM are much worse. Interestingly, looking at the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 477, + 486, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 486, + 489 + ], + "score": 1.0, + "content": "train and validation curves together, ELDM seems to be suffering from optimization problems.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 389, + 506, + 489 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 502, + 173, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 175, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 175, + 515 + ], + "score": 1.0, + "content": "5.2 Π-MODEL", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 523, + 505, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "Since Π-model Laine & Aila (2016) is similar to our algorithm (even though it is designed for", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 535, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 505, + 546 + ], + "score": 1.0, + "content": "semi-supervised learning in feed-forward networks), we study the difference in performance with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 544, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 558 + ], + "score": 1.0, + "content": "Π-model7 both qualitatively and quantitatively to establish the advantage of our approach. First,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "we run both single layer LSTM and 3-layer AWD-LSTM on PTB task to check how their model", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 567, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 505, + 579 + ], + "score": 1.0, + "content": "compares with ours in the case of language modeling. The results are shown in Figure 1 and 2. We", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 104, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "find that our model converges significantly faster than Π-model. We believe this happens because we", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 589, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 505, + 601 + ], + "score": 1.0, + "content": "back-propagate the target loss through both networks (in contrast to Π-model) that leads to weights", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 601, + 333, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 333, + 612 + ], + "score": 1.0, + "content": "getting updated using target-based gradients more often.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5, + "bbox_fs": [ + 104, + 523, + 506, + 612 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 114, + 87, + 280, + 196 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 87, + 280, + 196 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 87, + 280, + 196 + ], + "spans": [ + { + "bbox": [ + 114, + 87, + 280, + 196 + ], + "score": 0.964, + "type": "image", + "image_path": "4b8d4833a24b6748ef3cd0f76c9fe0025e06c99555d7e766f2a29c044f96997c.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 114, + 87, + 280, + 100.625 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 114, + 100.625, + 280, + 114.25 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 114, + 114.25, + 280, + 127.875 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 114, + 127.875, + 280, + 141.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 114, + 141.5, + 280, + 155.125 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 114, + 155.125, + 280, + 168.75 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 114, + 168.75, + 280, + 182.375 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 114, + 182.375, + 280, + 196.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 203, + 506, + 259 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 203, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 505, + 217 + ], + "score": 1.0, + "content": "Figure 1: Ablation study: Train (left) and validation (right) perplexity on PTB word level modeling", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "score": 1.0, + "content": "with single layer LSTM (10M parameters). These curves study the learning dynamics of the baseline", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "model, Π-model, Expectation-linear dropout (ELD), Expectation-linear dropout with modification", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 235, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 505, + 250 + ], + "score": 1.0, + "content": "(ELDM) and fraternal dropout (FD, our algorithm). We find that FD converges faster than the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 247, + 308, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 308, + 261 + ], + "score": 1.0, + "content": "regularizers in comparison, and generalizes at par.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18 + } + ], + "index": 10.75 + }, + { + "type": "image", + "bbox": [ + 320, + 87, + 487, + 196 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 320, + 87, + 487, + 196 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 320, + 87, + 487, + 196 + ], + "spans": [ + { + "bbox": [ + 320, + 87, + 487, + 196 + ], + "score": 0.961, + "type": "image", + "image_path": "d64e450f08346c6bc468f12061b79ab092e7cc2cf8930eb3af1ac794100c063b.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 320, + 87, + 487, + 100.625 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 320, + 100.625, + 487, + 114.25 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 320, + 114.25, + 487, + 127.875 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 320, + 127.875, + 487, + 141.5 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 320, + 141.5, + 487, + 155.125 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 320, + 155.125, + 487, + 168.75 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 320, + 168.75, + 487, + 182.375 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 320, + 182.375, + 487, + 196.0 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 11.5 + }, + { + "type": "image", + "bbox": [ + 114, + 275, + 280, + 385 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 275, + 280, + 385 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 114, + 275, + 280, + 385 + ], + "spans": [ + { + "bbox": [ + 114, + 275, + 280, + 385 + ], + "score": 0.968, + "type": "image", + "image_path": "87020ced434d3499701675350c507f939c1046c58cb17698501cdf12b3ae97a7.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 114, + 275, + 280, + 288.75 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 114, + 288.75, + 280, + 302.5 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 114, + 302.5, + 280, + 316.25 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 114, + 316.25, + 280, + 330.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 114, + 330.0, + 280, + 343.75 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 114, + 343.75, + 280, + 357.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 114, + 357.5, + 280, + 371.25 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 114, + 371.25, + 280, + 385.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 392, + 299, + 437 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 392, + 299, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 299, + 405 + ], + "score": 1.0, + "content": "Figure 2: Ablation study: Validation perplexity", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 403, + 299, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 243, + 415 + ], + "score": 1.0, + "content": "on PTB word level modeling for", + "type": "text" + }, + { + "bbox": [ + 243, + 403, + 252, + 414 + ], + "score": 0.43, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 403, + 299, + 415 + ], + "score": 1.0, + "content": "-model and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 413, + 300, + 427 + ], + "spans": [ + { + "bbox": [ + 104, + 413, + 300, + 427 + ], + "score": 1.0, + "content": "fraternal dropout. We find that FD converges", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 425, + 222, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 222, + 437 + ], + "score": 1.0, + "content": "faster and generalizes at par.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5 + } + ], + "index": 31.5 + }, + { + "type": "image", + "bbox": [ + 315, + 275, + 483, + 385 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 315, + 275, + 483, + 385 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 315, + 275, + 483, + 385 + ], + "spans": [ + { + "bbox": [ + 315, + 275, + 483, + 385 + ], + "score": 0.966, + "type": "image", + "image_path": "463bd2f8083be91f8df7ca8a7e5292781fe18c685b3e04efd116713fc227bc09.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 315, + 275, + 483, + 288.75 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 315, + 288.75, + 483, + 302.5 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 315, + 302.5, + 483, + 316.25 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 315, + 316.25, + 483, + 330.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 315, + 330.0, + 483, + 343.75 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 315, + 343.75, + 483, + 357.5 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 315, + 357.5, + 483, + 371.25 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 315, + 371.25, + 483, + 385.0 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 309, + 392, + 502, + 438 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 309, + 391, + 502, + 404 + ], + "spans": [ + { + "bbox": [ + 309, + 391, + 502, + 404 + ], + "score": 1.0, + "content": "Figure 3: Ablation study: Average hidden state", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 309, + 403, + 502, + 415 + ], + "spans": [ + { + "bbox": [ + 309, + 403, + 502, + 415 + ], + "score": 1.0, + "content": "activation is reduced when any of the regular-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 309, + 413, + 503, + 426 + ], + "spans": [ + { + "bbox": [ + 309, + 413, + 503, + 426 + ], + "score": 1.0, + "content": "izer described is used. The y-axis is the value of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 311, + 423, + 363, + 439 + ], + "spans": [ + { + "bbox": [ + 311, + 424, + 358, + 438 + ], + "score": 0.92, + "content": "\\frac { 1 } { d } \\| \\mathbf { m } \\cdot \\mathbf { h } ^ { t } \\| _ { 2 } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 423, + 363, + 439 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 106, + 456, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "score": 1.0, + "content": "Even though we designed our algorithm specifically to address problems in RNNs, to have a fair", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 234, + 480 + ], + "score": 1.0, + "content": "comparison, we compare with", + "type": "text" + }, + { + "bbox": [ + 235, + 468, + 243, + 478 + ], + "score": 0.46, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "-model on a semi-supervised task which is their goal. Specifi-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 313, + 491 + ], + "score": 1.0, + "content": "cally, we use the CIFAR-10 dataset that consists of", + "type": "text" + }, + { + "bbox": [ + 313, + 479, + 347, + 489 + ], + "score": 0.9, + "content": "3 2 \\times 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "images from 10 classes. Following the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "score": 1.0, + "content": "usual splits used in semi-supervised learning literature, we use 4 thousand labeled and 41 thousand", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "score": 1.0, + "content": "unlabeled samples for training, 5 thousand labeled samples for validation and 10 thousand labeled", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "samples for test set. We use the original ResNet-56 (He et al., 2015) architecture. We run grid search", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 119, + 536 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 119, + 523, + 223, + 535 + ], + "score": 0.91, + "content": "\\kappa \\in \\{ 0 . 0 5 , 0 . 1 , 0 . 1 5 , 0 . 2 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 523, + 293, + 536 + ], + "score": 1.0, + "content": ", dropout rates in", + "type": "text" + }, + { + "bbox": [ + 293, + 523, + 378, + 535 + ], + "score": 0.89, + "content": "\\{ 0 . 0 5 , 0 . 1 , 0 . 1 5 , 0 . 2 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "and leave the rest of the hyper-", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 104, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 104, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "parameters unchanged. We additionally check importance of using unlabeled data. The results are", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 104, + 544, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 104, + 544, + 377, + 558 + ], + "score": 1.0, + "content": "reported in Table 4. We find that our algorithm performs at par with", + "type": "text" + }, + { + "bbox": [ + 378, + 545, + 386, + 555 + ], + "score": 0.34, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 544, + 506, + 558 + ], + "score": 1.0, + "content": "-model. When unlabeled data", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 556, + 490, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 490, + 568 + ], + "score": 1.0, + "content": "is not used, fraternal dropout provides slightly better results as compared to traditional dropout.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 49.5 + }, + { + "type": "title", + "bbox": [ + 106, + 581, + 490, + 591 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 491, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 491, + 592 + ], + "score": 1.0, + "content": "5.3 ACTIVITY REGULARIZATION AND TEMPORAL ACTIVITY REGULARIZATION ANALYSIS", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 55 + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 502, + 623 + ], + "lines": [ + { + "bbox": [ + 106, + 600, + 504, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 441, + 613 + ], + "score": 1.0, + "content": "The authors of Merity et al. (2017b) study the importance of activity regularization", + "type": "text" + }, + { + "bbox": [ + 441, + 600, + 466, + 612 + ], + "score": 0.8, + "content": "( \\mathsf { A R } ) ^ { 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 600, + 504, + 613 + ], + "score": 1.0, + "content": "and tem-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 611, + 328, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 328, + 625 + ], + "score": 1.0, + "content": "poral activity regularization (TAR) in LSTMs given as,", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 56.5 + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 628, + 275, + 676 + ], + "lines": [ + { + "bbox": [ + 142, + 628, + 275, + 676 + ], + "spans": [ + { + "bbox": [ + 142, + 628, + 275, + 676 + ], + "score": 0.92, + "content": "\\begin{array} { c } { \\mathcal { R } _ { A R } ( { \\bf z } ^ { t } ; \\theta ) = \\displaystyle \\frac { \\alpha } { d } \\| { \\bf h } ^ { t } \\| _ { 2 } ^ { 2 } } \\\\ { \\mathcal { R } _ { T A R } ( { \\bf z } ^ { t } ; \\theta ) = \\displaystyle \\frac { \\beta } { d } \\| { \\bf h } ^ { t } - { \\bf h } ^ { t - 1 } \\| _ { 2 } ^ { 2 } } \\end{array}", + "type": "interline_equation", + "image_path": "b38dd65a7f67fddfa81341093884dcd3c38ba12153ff03f3da3f09bec8e564e5.jpg" + } + ] + } + ], + "index": 58.5, + "virtual_lines": [ + { + "bbox": [ + 142, + 628, + 275, + 652.0 + ], + "spans": [], + "index": 58 + }, + { + "bbox": [ + 142, + 652.0, + 275, + 676.0 + ], + "spans": [], + "index": 59 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 681, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 679, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 133, + 694 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 680, + 169, + 691 + ], + "score": 0.92, + "content": "\\mathbf { h } ^ { t } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 679, + 348, + 694 + ], + "score": 1.0, + "content": "is the LSTM’s output activation at time step", + "type": "text" + }, + { + "bbox": [ + 348, + 682, + 353, + 691 + ], + "score": 0.73, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 679, + 506, + 694 + ], + "score": 1.0, + "content": "(hence depends on both current input", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 107, + 691, + 506, + 704 + ], + "spans": [ + { + "bbox": [ + 107, + 692, + 116, + 702 + ], + "score": 0.83, + "content": "\\mathbf { z } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 691, + 221, + 704 + ], + "score": 1.0, + "content": "and the model parameters", + "type": "text" + }, + { + "bbox": [ + 221, + 693, + 226, + 702 + ], + "score": 0.76, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 691, + 506, + 704 + ], + "score": 1.0, + "content": "). Notice that AR and TAR regularizations are applied on the output of", + "type": "text" + } + ], + "index": 61 + } + ], + "index": 60.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 105, + 711, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 708, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 118, + 708, + 158, + 725 + ], + "score": 1.0, + "content": "8We used", + "type": "text" + }, + { + "bbox": [ + 158, + 711, + 196, + 722 + ], + "score": 0.9, + "content": "\\| \\mathbf { m } \\cdot \\mathbf { h } ^ { t } \\| _ { 2 } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 708, + 226, + 725 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 227, + 716, + 235, + 721 + ], + "score": 0.78, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 708, + 506, + 725 + ], + "score": 1.0, + "content": "is the dropout mask, in our actual experiments with AR because it was", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 721, + 410, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 410, + 732 + ], + "score": 1.0, + "content": "implemented as such in the original paper’s Github repository Merity et al. (2017a).", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 114, + 87, + 280, + 196 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 87, + 280, + 196 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 87, + 280, + 196 + ], + "spans": [ + { + "bbox": [ + 114, + 87, + 280, + 196 + ], + "score": 0.964, + "type": "image", + "image_path": "4b8d4833a24b6748ef3cd0f76c9fe0025e06c99555d7e766f2a29c044f96997c.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 114, + 87, + 280, + 100.625 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 114, + 100.625, + 280, + 114.25 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 114, + 114.25, + 280, + 127.875 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 114, + 127.875, + 280, + 141.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 114, + 141.5, + 280, + 155.125 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 114, + 155.125, + 280, + 168.75 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 114, + 168.75, + 280, + 182.375 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 114, + 182.375, + 280, + 196.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 203, + 506, + 259 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 203, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 505, + 217 + ], + "score": 1.0, + "content": "Figure 1: Ablation study: Train (left) and validation (right) perplexity on PTB word level modeling", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "score": 1.0, + "content": "with single layer LSTM (10M parameters). These curves study the learning dynamics of the baseline", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "model, Π-model, Expectation-linear dropout (ELD), Expectation-linear dropout with modification", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 235, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 505, + 250 + ], + "score": 1.0, + "content": "(ELDM) and fraternal dropout (FD, our algorithm). We find that FD converges faster than the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 247, + 308, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 308, + 261 + ], + "score": 1.0, + "content": "regularizers in comparison, and generalizes at par.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18 + } + ], + "index": 10.75 + }, + { + "type": "image", + "bbox": [ + 320, + 87, + 487, + 196 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 320, + 87, + 487, + 196 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 320, + 87, + 487, + 196 + ], + "spans": [ + { + "bbox": [ + 320, + 87, + 487, + 196 + ], + "score": 0.961, + "type": "image", + "image_path": "d64e450f08346c6bc468f12061b79ab092e7cc2cf8930eb3af1ac794100c063b.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 320, + 87, + 487, + 100.625 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 320, + 100.625, + 487, + 114.25 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 320, + 114.25, + 487, + 127.875 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 320, + 127.875, + 487, + 141.5 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 320, + 141.5, + 487, + 155.125 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 320, + 155.125, + 487, + 168.75 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 320, + 168.75, + 487, + 182.375 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 320, + 182.375, + 487, + 196.0 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 11.5 + }, + { + "type": "image", + "bbox": [ + 114, + 275, + 280, + 385 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 275, + 280, + 385 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 114, + 275, + 280, + 385 + ], + "spans": [ + { + "bbox": [ + 114, + 275, + 280, + 385 + ], + "score": 0.968, + "type": "image", + "image_path": "87020ced434d3499701675350c507f939c1046c58cb17698501cdf12b3ae97a7.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 114, + 275, + 280, + 288.75 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 114, + 288.75, + 280, + 302.5 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 114, + 302.5, + 280, + 316.25 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 114, + 316.25, + 280, + 330.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 114, + 330.0, + 280, + 343.75 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 114, + 343.75, + 280, + 357.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 114, + 357.5, + 280, + 371.25 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 114, + 371.25, + 280, + 385.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 392, + 299, + 437 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 392, + 299, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 299, + 405 + ], + "score": 1.0, + "content": "Figure 2: Ablation study: Validation perplexity", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 403, + 299, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 243, + 415 + ], + "score": 1.0, + "content": "on PTB word level modeling for", + "type": "text" + }, + { + "bbox": [ + 243, + 403, + 252, + 414 + ], + "score": 0.43, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 403, + 299, + 415 + ], + "score": 1.0, + "content": "-model and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 413, + 300, + 427 + ], + "spans": [ + { + "bbox": [ + 104, + 413, + 300, + 427 + ], + "score": 1.0, + "content": "fraternal dropout. We find that FD converges", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 425, + 222, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 222, + 437 + ], + "score": 1.0, + "content": "faster and generalizes at par.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5 + } + ], + "index": 31.5 + }, + { + "type": "image", + "bbox": [ + 315, + 275, + 483, + 385 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 315, + 275, + 483, + 385 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 315, + 275, + 483, + 385 + ], + "spans": [ + { + "bbox": [ + 315, + 275, + 483, + 385 + ], + "score": 0.966, + "type": "image", + "image_path": "463bd2f8083be91f8df7ca8a7e5292781fe18c685b3e04efd116713fc227bc09.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 315, + 275, + 483, + 288.75 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 315, + 288.75, + 483, + 302.5 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 315, + 302.5, + 483, + 316.25 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 315, + 316.25, + 483, + 330.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 315, + 330.0, + 483, + 343.75 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 315, + 343.75, + 483, + 357.5 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 315, + 357.5, + 483, + 371.25 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 315, + 371.25, + 483, + 385.0 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 309, + 392, + 502, + 438 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 309, + 391, + 502, + 404 + ], + "spans": [ + { + "bbox": [ + 309, + 391, + 502, + 404 + ], + "score": 1.0, + "content": "Figure 3: Ablation study: Average hidden state", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 309, + 403, + 502, + 415 + ], + "spans": [ + { + "bbox": [ + 309, + 403, + 502, + 415 + ], + "score": 1.0, + "content": "activation is reduced when any of the regular-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 309, + 413, + 503, + 426 + ], + "spans": [ + { + "bbox": [ + 309, + 413, + 503, + 426 + ], + "score": 1.0, + "content": "izer described is used. The y-axis is the value of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 311, + 423, + 363, + 439 + ], + "spans": [ + { + "bbox": [ + 311, + 424, + 358, + 438 + ], + "score": 0.92, + "content": "\\frac { 1 } { d } \\| \\mathbf { m } \\cdot \\mathbf { h } ^ { t } \\| _ { 2 } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 423, + 363, + 439 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 106, + 456, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "score": 1.0, + "content": "Even though we designed our algorithm specifically to address problems in RNNs, to have a fair", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 234, + 480 + ], + "score": 1.0, + "content": "comparison, we compare with", + "type": "text" + }, + { + "bbox": [ + 235, + 468, + 243, + 478 + ], + "score": 0.46, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "-model on a semi-supervised task which is their goal. Specifi-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 313, + 491 + ], + "score": 1.0, + "content": "cally, we use the CIFAR-10 dataset that consists of", + "type": "text" + }, + { + "bbox": [ + 313, + 479, + 347, + 489 + ], + "score": 0.9, + "content": "3 2 \\times 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "images from 10 classes. Following the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "score": 1.0, + "content": "usual splits used in semi-supervised learning literature, we use 4 thousand labeled and 41 thousand", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "score": 1.0, + "content": "unlabeled samples for training, 5 thousand labeled samples for validation and 10 thousand labeled", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "samples for test set. We use the original ResNet-56 (He et al., 2015) architecture. We run grid search", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 119, + 536 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 119, + 523, + 223, + 535 + ], + "score": 0.91, + "content": "\\kappa \\in \\{ 0 . 0 5 , 0 . 1 , 0 . 1 5 , 0 . 2 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 523, + 293, + 536 + ], + "score": 1.0, + "content": ", dropout rates in", + "type": "text" + }, + { + "bbox": [ + 293, + 523, + 378, + 535 + ], + "score": 0.89, + "content": "\\{ 0 . 0 5 , 0 . 1 , 0 . 1 5 , 0 . 2 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "and leave the rest of the hyper-", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 104, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 104, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "parameters unchanged. We additionally check importance of using unlabeled data. The results are", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 104, + 544, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 104, + 544, + 377, + 558 + ], + "score": 1.0, + "content": "reported in Table 4. We find that our algorithm performs at par with", + "type": "text" + }, + { + "bbox": [ + 378, + 545, + 386, + 555 + ], + "score": 0.34, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 544, + 506, + 558 + ], + "score": 1.0, + "content": "-model. When unlabeled data", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 556, + 490, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 490, + 568 + ], + "score": 1.0, + "content": "is not used, fraternal dropout provides slightly better results as compared to traditional dropout.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 49.5, + "bbox_fs": [ + 104, + 457, + 506, + 568 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 581, + 490, + 591 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 491, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 491, + 592 + ], + "score": 1.0, + "content": "5.3 ACTIVITY REGULARIZATION AND TEMPORAL ACTIVITY REGULARIZATION ANALYSIS", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 55 + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 502, + 623 + ], + "lines": [ + { + "bbox": [ + 106, + 600, + 504, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 441, + 613 + ], + "score": 1.0, + "content": "The authors of Merity et al. (2017b) study the importance of activity regularization", + "type": "text" + }, + { + "bbox": [ + 441, + 600, + 466, + 612 + ], + "score": 0.8, + "content": "( \\mathsf { A R } ) ^ { 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 600, + 504, + 613 + ], + "score": 1.0, + "content": "and tem-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 611, + 328, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 328, + 625 + ], + "score": 1.0, + "content": "poral activity regularization (TAR) in LSTMs given as,", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 56.5, + "bbox_fs": [ + 105, + 600, + 504, + 625 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 628, + 275, + 676 + ], + "lines": [ + { + "bbox": [ + 142, + 628, + 275, + 676 + ], + "spans": [ + { + "bbox": [ + 142, + 628, + 275, + 676 + ], + "score": 0.92, + "content": "\\begin{array} { c } { \\mathcal { R } _ { A R } ( { \\bf z } ^ { t } ; \\theta ) = \\displaystyle \\frac { \\alpha } { d } \\| { \\bf h } ^ { t } \\| _ { 2 } ^ { 2 } } \\\\ { \\mathcal { R } _ { T A R } ( { \\bf z } ^ { t } ; \\theta ) = \\displaystyle \\frac { \\beta } { d } \\| { \\bf h } ^ { t } - { \\bf h } ^ { t - 1 } \\| _ { 2 } ^ { 2 } } \\end{array}", + "type": "interline_equation", + "image_path": "b38dd65a7f67fddfa81341093884dcd3c38ba12153ff03f3da3f09bec8e564e5.jpg" + } + ] + } + ], + "index": 58.5, + "virtual_lines": [ + { + "bbox": [ + 142, + 628, + 275, + 652.0 + ], + "spans": [], + "index": 58 + }, + { + "bbox": [ + 142, + 652.0, + 275, + 676.0 + ], + "spans": [], + "index": 59 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 681, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 679, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 133, + 694 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 680, + 169, + 691 + ], + "score": 0.92, + "content": "\\mathbf { h } ^ { t } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 679, + 348, + 694 + ], + "score": 1.0, + "content": "is the LSTM’s output activation at time step", + "type": "text" + }, + { + "bbox": [ + 348, + 682, + 353, + 691 + ], + "score": 0.73, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 679, + 506, + 694 + ], + "score": 1.0, + "content": "(hence depends on both current input", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 107, + 691, + 506, + 704 + ], + "spans": [ + { + "bbox": [ + 107, + 692, + 116, + 702 + ], + "score": 0.83, + "content": "\\mathbf { z } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 691, + 221, + 704 + ], + "score": 1.0, + "content": "and the model parameters", + "type": "text" + }, + { + "bbox": [ + 221, + 693, + 226, + 702 + ], + "score": 0.76, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 691, + 506, + 704 + ], + "score": 1.0, + "content": "). Notice that AR and TAR regularizations are applied on the output of", + "type": "text" + } + ], + "index": 61 + } + ], + "index": 60.5, + "bbox_fs": [ + 105, + 679, + 506, + 704 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 77, + 520, + 157 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 77, + 520, + 157 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 77, + 520, + 157 + ], + "spans": [ + { + "bbox": [ + 106, + 77, + 520, + 157 + ], + "score": 0.972, + "html": "
ModelDropout rateUnlabeled dataValidationTest
Traditional dropout0.1No78.4 (± 0.25)76.9 (± 0.31)
No dropout0.0No78.8 (± 0.59)77.1 (± 0.3)
Fraternal dropout (k = 0.05)0.05No79.3 (± 0.38)77.6 (± 0.35)
Traditional dropout + I-model0.1Yes80.2 (± 0.33)78.5 (± 0.46)
Fraternal dropout (k = 0.15)0.1Yes80.5 (± 0.18)79.1 (± 0.37)
", + "type": "table", + "image_path": "2aeea17ebcb0188853ad5bb84b6d7253418d9774807341160e5fb918f328b85d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 77, + 520, + 103.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 103.66666666666667, + 520, + 130.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 130.33333333333334, + 520, + 157.0 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 160, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 160, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 506, + 173 + ], + "score": 1.0, + "content": "Table 4: Ablation study: Accuracy on altered (semi-supervised) CIFAR-10 dataset for ResNet-56", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 171, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 371, + 185 + ], + "score": 1.0, + "content": "based models. We find that our algorithm performs at par with", + "type": "text" + }, + { + "bbox": [ + 372, + 172, + 380, + 182 + ], + "score": 0.32, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 171, + 506, + 185 + ], + "score": 1.0, + "content": "-model. When unlabeled data", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 183, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 506, + 195 + ], + "score": 1.0, + "content": "is not used traditional dropout hurts performance while fraternal dropout provides slightly better", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 194, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 205 + ], + "score": 1.0, + "content": "results. It means that our methods may be beneficial when we lack data and have to use additional", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 205, + 195, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 195, + 217 + ], + "score": 1.0, + "content": "regularizing methods.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 114, + 234, + 487, + 345 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 234, + 487, + 345 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 234, + 487, + 345 + ], + "spans": [ + { + "bbox": [ + 114, + 234, + 487, + 345 + ], + "score": 0.825, + "type": "image", + "image_path": "45583947707d0455d79d12a5db380878e6174220941656e29500ab8024936b0e.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 114, + 234, + 487, + 271.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 114, + 271.0, + 487, + 308.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 114, + 308.0, + 487, + 345.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 352, + 505, + 407 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 352, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 505, + 365 + ], + "score": 1.0, + "content": "Figure 4: Ablation study: Train (left) and validation (right) perplexity on PTB word level model-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "ing with single layer LSTM (10M parameters). These curves study the learning dynamics of the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 372, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 389 + ], + "score": 1.0, + "content": "baseline model, temporal activity regularization (TAR), prediction regularization (PR), activity reg-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 384, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 398 + ], + "score": 1.0, + "content": "ularization (AR) and fraternal dropout (FD, our algorithm). We find that FD both converges faster", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 396, + 340, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 340, + 410 + ], + "score": 1.0, + "content": "and generalizes better than the regularizers in comparison.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + } + ], + "index": 11.0 + }, + { + "type": "text", + "bbox": [ + 104, + 428, + 503, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 427, + 504, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 397, + 442 + ], + "score": 1.0, + "content": "the LSTM, while our regularization is applied on the pre-softmax output", + "type": "text" + }, + { + "bbox": [ + 397, + 428, + 447, + 441 + ], + "score": 0.93, + "content": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 427, + 504, + 442 + ], + "score": 1.0, + "content": "of the LSTM.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 440, + 335, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 335, + 452 + ], + "score": 1.0, + "content": "However, since our regularization can be decomposed as", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "interline_equation", + "bbox": [ + 141, + 456, + 497, + 491 + ], + "lines": [ + { + "bbox": [ + 141, + 456, + 497, + 491 + ], + "spans": [ + { + "bbox": [ + 141, + 456, + 497, + 491 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) = \\mathbb { E } _ { s _ { i } , s _ { j } } \\left[ | | \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) - \\mathbf { p } ^ { t } ( \\mathbf { h } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } \\right] } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) ^ { T } \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } + | | \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } - 2 \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) ^ { T } \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) \\Big ] } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad = \\mathbb { E } _ { s _ { i } , s _ { j } } \\left[ | | \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } + | | \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } - 2 \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) ^ { T } \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) \\right] } \\end{array}", + "type": "interline_equation", + "image_path": "a9990d1706822c4d925d6c9a10839ee14f0260c58e403a40356f53bb93dc5647.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 141, + 456, + 497, + 467.6666666666667 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 141, + 467.6666666666667, + 497, + 479.33333333333337 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 141, + 479.33333333333337, + 497, + 491.00000000000006 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 508, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 508, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 190, + 522 + ], + "score": 1.0, + "content": "and encapsulates an", + "type": "text" + }, + { + "bbox": [ + 190, + 508, + 200, + 519 + ], + "score": 0.85, + "content": "\\ell ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 508, + 505, + 522 + ], + "score": 1.0, + "content": "term along with the dot product term, we perform experiments to confirm", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 519, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 302, + 533 + ], + "score": 1.0, + "content": "that the gains in our approach is not due to the", + "type": "text" + }, + { + "bbox": [ + 302, + 519, + 312, + 530 + ], + "score": 0.85, + "content": "\\bar { \\ell } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 519, + 506, + 533 + ], + "score": 1.0, + "content": "regularization alone. A similar argument goes", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 530, + 506, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 308, + 544 + ], + "score": 1.0, + "content": "for the TAR objective. We run a grid search on", + "type": "text" + }, + { + "bbox": [ + 308, + 531, + 388, + 542 + ], + "score": 0.88, + "content": "\\alpha \\in \\{ 1 , 2 , \\ldots , 1 2 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 530, + 393, + 544 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 393, + 531, + 473, + 543 + ], + "score": 0.88, + "content": "\\beta \\in \\{ 1 , 2 , \\ldots , 1 2 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 530, + 506, + 544 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 542, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 554 + ], + "score": 1.0, + "content": "include the hyper-parameters mentioned in Merity et al. (2017a). For our regularization, we use", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 552, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 206, + 564 + ], + "score": 0.89, + "content": "\\kappa \\in \\{ 0 . 0 5 , 0 . 1 , \\ldots , 0 . 4 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 552, + 505, + 565 + ], + "score": 1.0, + "content": ". Furthermore, we also compare with a regularization (PR) that regularizes", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 561, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 107, + 564, + 170, + 576 + ], + "score": 0.91, + "content": "\\| \\mathbf { p } ^ { t } ( \\bar { \\mathbf { z } } ^ { t } , s _ { i } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 561, + 327, + 578 + ], + "score": 1.0, + "content": "to further rule-out any gains only from", + "type": "text" + }, + { + "bbox": [ + 328, + 563, + 338, + 573 + ], + "score": 0.84, + "content": "\\ell ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 561, + 506, + 578 + ], + "score": 1.0, + "content": "regularization. Based on this grid search,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 573, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 588 + ], + "score": 1.0, + "content": "we pick the best model on the validation set for all the regularizations, and additionally report a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 585, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 598 + ], + "score": 1.0, + "content": "baseline model without any of these four mentioned regularizations. The learning dynamics is shown", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 596, + 504, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 504, + 609 + ], + "score": 1.0, + "content": "in Figure 4. Our regularization performs better both in terms of convergence and generalization", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "compared with other methods. Average hidden state activation is reduced when any of the regularizer", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 619, + 248, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 248, + 631 + ], + "score": 1.0, + "content": "described is applied (see Figure 3).", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 26 + }, + { + "type": "title", + "bbox": [ + 109, + 645, + 289, + 655 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 290, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 290, + 657 + ], + "score": 1.0, + "content": "5.4 IMPROVEMENTS USING FINE-TUNING", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 664, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 505, + 679 + ], + "score": 1.0, + "content": "We confirm that models trained with fraternal dropout benefit from the NT-ASGD fine-tuning step", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "(as also used in Merity et al. (2017a)). However, this is a very time-consuming practice and since dif-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "ferent hyper-parameters may be used in this additional part of the learning procedure, the probability", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "of obtaining better results due to the extensive grid search is higher. Hence, in our experiments we", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "use the same fine-tuning procedure as implemented in the official repository (even fraternal dropout", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 721, + 378, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 378, + 733 + ], + "score": 1.0, + "content": "was not used). We present the importance of fine-tuning in Table 5.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 77, + 520, + 157 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 77, + 520, + 157 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 77, + 520, + 157 + ], + "spans": [ + { + "bbox": [ + 106, + 77, + 520, + 157 + ], + "score": 0.972, + "html": "
ModelDropout rateUnlabeled dataValidationTest
Traditional dropout0.1No78.4 (± 0.25)76.9 (± 0.31)
No dropout0.0No78.8 (± 0.59)77.1 (± 0.3)
Fraternal dropout (k = 0.05)0.05No79.3 (± 0.38)77.6 (± 0.35)
Traditional dropout + I-model0.1Yes80.2 (± 0.33)78.5 (± 0.46)
Fraternal dropout (k = 0.15)0.1Yes80.5 (± 0.18)79.1 (± 0.37)
", + "type": "table", + "image_path": "2aeea17ebcb0188853ad5bb84b6d7253418d9774807341160e5fb918f328b85d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 77, + 520, + 103.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 103.66666666666667, + 520, + 130.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 130.33333333333334, + 520, + 157.0 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 160, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 160, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 506, + 173 + ], + "score": 1.0, + "content": "Table 4: Ablation study: Accuracy on altered (semi-supervised) CIFAR-10 dataset for ResNet-56", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 171, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 371, + 185 + ], + "score": 1.0, + "content": "based models. We find that our algorithm performs at par with", + "type": "text" + }, + { + "bbox": [ + 372, + 172, + 380, + 182 + ], + "score": 0.32, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 171, + 506, + 185 + ], + "score": 1.0, + "content": "-model. When unlabeled data", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 183, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 506, + 195 + ], + "score": 1.0, + "content": "is not used traditional dropout hurts performance while fraternal dropout provides slightly better", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 194, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 205 + ], + "score": 1.0, + "content": "results. It means that our methods may be beneficial when we lack data and have to use additional", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 205, + 195, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 195, + 217 + ], + "score": 1.0, + "content": "regularizing methods.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 160, + 506, + 217 + ] + }, + { + "type": "image", + "bbox": [ + 114, + 234, + 487, + 345 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 234, + 487, + 345 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 234, + 487, + 345 + ], + "spans": [ + { + "bbox": [ + 114, + 234, + 487, + 345 + ], + "score": 0.825, + "type": "image", + "image_path": "45583947707d0455d79d12a5db380878e6174220941656e29500ab8024936b0e.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 114, + 234, + 487, + 271.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 114, + 271.0, + 487, + 308.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 114, + 308.0, + 487, + 345.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 352, + 505, + 407 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 352, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 505, + 365 + ], + "score": 1.0, + "content": "Figure 4: Ablation study: Train (left) and validation (right) perplexity on PTB word level model-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "ing with single layer LSTM (10M parameters). These curves study the learning dynamics of the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 372, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 389 + ], + "score": 1.0, + "content": "baseline model, temporal activity regularization (TAR), prediction regularization (PR), activity reg-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 384, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 398 + ], + "score": 1.0, + "content": "ularization (AR) and fraternal dropout (FD, our algorithm). We find that FD both converges faster", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 396, + 340, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 340, + 410 + ], + "score": 1.0, + "content": "and generalizes better than the regularizers in comparison.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + } + ], + "index": 11.0 + }, + { + "type": "text", + "bbox": [ + 104, + 428, + 503, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 427, + 504, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 397, + 442 + ], + "score": 1.0, + "content": "the LSTM, while our regularization is applied on the pre-softmax output", + "type": "text" + }, + { + "bbox": [ + 397, + 428, + 447, + 441 + ], + "score": 0.93, + "content": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 427, + 504, + 442 + ], + "score": 1.0, + "content": "of the LSTM.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 440, + 335, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 335, + 452 + ], + "score": 1.0, + "content": "However, since our regularization can be decomposed as", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 427, + 504, + 452 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 141, + 456, + 497, + 491 + ], + "lines": [ + { + "bbox": [ + 141, + 456, + 497, + 491 + ], + "spans": [ + { + "bbox": [ + 141, + 456, + 497, + 491 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) = \\mathbb { E } _ { s _ { i } , s _ { j } } \\left[ | | \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) - \\mathbf { p } ^ { t } ( \\mathbf { h } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } \\right] } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) ^ { T } \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } + | | \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } - 2 \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) ^ { T } \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) \\Big ] } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad = \\mathbb { E } _ { s _ { i } , s _ { j } } \\left[ | | \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } + | | \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } - 2 \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) ^ { T } \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) \\right] } \\end{array}", + "type": "interline_equation", + "image_path": "a9990d1706822c4d925d6c9a10839ee14f0260c58e403a40356f53bb93dc5647.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 141, + 456, + 497, + 467.6666666666667 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 141, + 467.6666666666667, + 497, + 479.33333333333337 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 141, + 479.33333333333337, + 497, + 491.00000000000006 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 508, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 508, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 190, + 522 + ], + "score": 1.0, + "content": "and encapsulates an", + "type": "text" + }, + { + "bbox": [ + 190, + 508, + 200, + 519 + ], + "score": 0.85, + "content": "\\ell ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 508, + 505, + 522 + ], + "score": 1.0, + "content": "term along with the dot product term, we perform experiments to confirm", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 519, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 302, + 533 + ], + "score": 1.0, + "content": "that the gains in our approach is not due to the", + "type": "text" + }, + { + "bbox": [ + 302, + 519, + 312, + 530 + ], + "score": 0.85, + "content": "\\bar { \\ell } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 519, + 506, + 533 + ], + "score": 1.0, + "content": "regularization alone. A similar argument goes", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 530, + 506, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 308, + 544 + ], + "score": 1.0, + "content": "for the TAR objective. We run a grid search on", + "type": "text" + }, + { + "bbox": [ + 308, + 531, + 388, + 542 + ], + "score": 0.88, + "content": "\\alpha \\in \\{ 1 , 2 , \\ldots , 1 2 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 530, + 393, + 544 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 393, + 531, + 473, + 543 + ], + "score": 0.88, + "content": "\\beta \\in \\{ 1 , 2 , \\ldots , 1 2 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 530, + 506, + 544 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 542, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 554 + ], + "score": 1.0, + "content": "include the hyper-parameters mentioned in Merity et al. (2017a). For our regularization, we use", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 552, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 206, + 564 + ], + "score": 0.89, + "content": "\\kappa \\in \\{ 0 . 0 5 , 0 . 1 , \\ldots , 0 . 4 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 552, + 505, + 565 + ], + "score": 1.0, + "content": ". Furthermore, we also compare with a regularization (PR) that regularizes", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 561, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 107, + 564, + 170, + 576 + ], + "score": 0.91, + "content": "\\| \\mathbf { p } ^ { t } ( \\bar { \\mathbf { z } } ^ { t } , s _ { i } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 561, + 327, + 578 + ], + "score": 1.0, + "content": "to further rule-out any gains only from", + "type": "text" + }, + { + "bbox": [ + 328, + 563, + 338, + 573 + ], + "score": 0.84, + "content": "\\ell ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 561, + 506, + 578 + ], + "score": 1.0, + "content": "regularization. Based on this grid search,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 573, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 588 + ], + "score": 1.0, + "content": "we pick the best model on the validation set for all the regularizations, and additionally report a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 585, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 598 + ], + "score": 1.0, + "content": "baseline model without any of these four mentioned regularizations. The learning dynamics is shown", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 596, + 504, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 504, + 609 + ], + "score": 1.0, + "content": "in Figure 4. Our regularization performs better both in terms of convergence and generalization", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "compared with other methods. Average hidden state activation is reduced when any of the regularizer", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 619, + 248, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 248, + 631 + ], + "score": 1.0, + "content": "described is applied (see Figure 3).", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 508, + 506, + 631 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 645, + 289, + 655 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 290, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 290, + 657 + ], + "score": 1.0, + "content": "5.4 IMPROVEMENTS USING FINE-TUNING", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 664, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 505, + 679 + ], + "score": 1.0, + "content": "We confirm that models trained with fraternal dropout benefit from the NT-ASGD fine-tuning step", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "(as also used in Merity et al. (2017a)). However, this is a very time-consuming practice and since dif-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "ferent hyper-parameters may be used in this additional part of the learning procedure, the probability", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "of obtaining better results due to the extensive grid search is higher. Hence, in our experiments we", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "use the same fine-tuning procedure as implemented in the official repository (even fraternal dropout", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 721, + 378, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 378, + 733 + ], + "score": 1.0, + "content": "was not used). We present the importance of fine-tuning in Table 5.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 664, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 78, + 451, + 159 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 159, + 78, + 451, + 159 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 159, + 78, + 451, + 159 + ], + "spans": [ + { + "bbox": [ + 159, + 78, + 451, + 159 + ], + "score": 0.98, + "html": "
PTBWT2
DropoutFine-tuningValidationTestValidationTest
TraditionalNone60.758.869.166.0
TraditionalOne60.057.368.665.8
FraternalNone59.858.068.365.3
FraternalOne58.956.866.864.1
FraternalTwo58.556.211
", + "type": "table", + "image_path": "cb441eff58ff23d6d3734a8c592089c38ee86cc0af61b90822ae152ec84ef3dc.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 159, + 78, + 451, + 105.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 159, + 105.0, + 451, + 132.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 159, + 132.0, + 451, + 159.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 103, + 163, + 504, + 186 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 162, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 505, + 176 + ], + "score": 1.0, + "content": "Table 5: Ablation study: Importance of fine-tuning for AWD-LSTM 3-layer model. Perplexity for", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 173, + 351, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 351, + 186 + ], + "score": 1.0, + "content": "the Penn Treebank and WikiText-2 language modeling tasks.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 188, + 199, + 421, + 258 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 188, + 199, + 421, + 258 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 188, + 199, + 421, + 258 + ], + "spans": [ + { + "bbox": [ + 188, + 199, + 421, + 258 + ], + "score": 0.974, + "html": "
Hyper-parameterPossible values
batch size non-monotone interval[10,20,30, 40] [5,10,20,40,60,100]
K -FD or ELD strength weight decayU(0, 0.3) U(0.6 × 10-6,2.4 × 10-6)
", + "type": "table", + "image_path": "81288747a210a7748b2bb2edb3d9e19287db61a5f4d0244a78b709167c48133c.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 188, + 199, + 421, + 213.75 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 188, + 213.75, + 421, + 228.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 188, + 228.5, + 421, + 243.25 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 188, + 243.25, + 421, + 258.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 261, + 504, + 294 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 259, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 506, + 275 + ], + "score": 1.0, + "content": "Table 6: Ablation study: Candidate hyper-parameters possible used in the grid search for comparing", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 306, + 284 + ], + "score": 1.0, + "content": "fraternal dropout and expectation linear dropout.", + "type": "text" + }, + { + "bbox": [ + 307, + 272, + 337, + 283 + ], + "score": 0.93, + "content": "U ( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "is the uniform distribution on the interval", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 282, + 364, + 296 + ], + "spans": [ + { + "bbox": [ + 107, + 282, + 127, + 294 + ], + "score": 0.9, + "content": "[ a , b ]", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 282, + 364, + 296 + ], + "score": 1.0, + "content": ". For finite sets, each value is drawn with equal probability.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.25 + }, + { + "type": "table", + "bbox": [ + 108, + 312, + 502, + 348 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 312, + 502, + 348 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 108, + 312, + 502, + 348 + ], + "spans": [ + { + "bbox": [ + 108, + 312, + 502, + 348 + ], + "score": 0.968, + "html": "
ModelBestTop5 avgTop10 avgBeating baseline runs (out of)
Expectation lineardropout59.460.160.56(208)
Fraternal dropout59.459.659.914 (203)
", + "type": "table", + "image_path": "080b0198eaea0fe2ad695f2300f5f44e370afb3f693385f868df02e311bdb4d7.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 108, + 312, + 502, + 324.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 108, + 324.0, + 502, + 336.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 108, + 336.0, + 502, + 348.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 352, + 504, + 386 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 351, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 506, + 365 + ], + "score": 1.0, + "content": "Table 7: Ablation study: Fraternal dropout and expectation linear dropout comparison. Perplex-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "ity on the Penn Treebank validation dateset. Fraternal dropout is more robust to different hyper-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 374, + 504, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 504, + 387 + ], + "score": 1.0, + "content": "parameters choice as twice as much runs finished performing better than the baseline model (60.7).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 107, + 413, + 444, + 423 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 447, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 447, + 424 + ], + "score": 1.0, + "content": "5.5 FRATERNAL DROPOUT AND EXPECTATION LINEAR DROPOUT COMPARISON", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 435, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 436, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 505, + 448 + ], + "score": 1.0, + "content": "We perform extensive grid search for the baseline model from Subsection 4.1 (an AWD-LSTM", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "score": 1.0, + "content": "3-layer architecture) trained with either fraternal dropout or expectation linear dropout regulariza-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "tions, to further contrast the performance of these two methods. The experiments are run without", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 469, + 233, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 233, + 480 + ], + "score": 1.0, + "content": "fine-tuning on the PTB dataset.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 505, + 529 + ], + "lines": [ + { + "bbox": [ + 105, + 486, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 497 + ], + "score": 1.0, + "content": "In each run, all five dropout rates are randomly altered (they are set to their original value, as in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "Merity et al. (2017a), multiplied by a value drawn from the uniform distribution on the interval", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "[0.5, 1.5]) and the rest of the hyper-parameters are drawn as shown in Table 6. As in Subsection 4.1,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 519, + 275, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 275, + 531 + ], + "score": 1.0, + "content": "AR and TAR regularizers are deactivated.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 535, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 506, + 548 + ], + "score": 1.0, + "content": "Together we run more than 400 experiments. The results are presented in Table 7. Both FD and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 545, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 506, + 560 + ], + "score": 1.0, + "content": "ELD perform better than the baseline model that instead uses AR and TAR regularizers. Hence,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "score": 1.0, + "content": "we confirm our previous finding (see Subsection 5.3) that both FD and ELD are better. However,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "as found previously for smaller model in Subsection 5.1, the convergence of FD is faster than that", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "of ELD. Additionally, fraternal dropout is more robust to different hyper-parameters choice (more", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 591, + 445, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 445, + 603 + ], + "score": 1.0, + "content": "runs performing better than the baseline and better average for top performing runs).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + }, + { + "type": "title", + "bbox": [ + 107, + 625, + 195, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 624, + 197, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 197, + 641 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "In this paper we propose a simple regularization method for RNNs called fraternal dropout that acts", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "as a regularization by reducing the variance in model predictions across different dropout masks. We", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 676, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 691 + ], + "score": 1.0, + "content": "show that our model achieves state-of-the-art results on benchmark language modeling tasks along", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "with faster convergence. We also analytically study the relationship between our regularization and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "expectation linear dropout Ma et al. (2016). We perform a number of ablation studies to evaluate", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "our model from different aspects and carefully compare it with related methods both qualitatively", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 721, + 182, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 182, + 734 + ], + "score": 1.0, + "content": "and quantitatively.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 78, + 451, + 159 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 159, + 78, + 451, + 159 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 159, + 78, + 451, + 159 + ], + "spans": [ + { + "bbox": [ + 159, + 78, + 451, + 159 + ], + "score": 0.98, + "html": "
PTBWT2
DropoutFine-tuningValidationTestValidationTest
TraditionalNone60.758.869.166.0
TraditionalOne60.057.368.665.8
FraternalNone59.858.068.365.3
FraternalOne58.956.866.864.1
FraternalTwo58.556.211
", + "type": "table", + "image_path": "cb441eff58ff23d6d3734a8c592089c38ee86cc0af61b90822ae152ec84ef3dc.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 159, + 78, + 451, + 105.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 159, + 105.0, + 451, + 132.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 159, + 132.0, + 451, + 159.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 103, + 163, + 504, + 186 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 162, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 505, + 176 + ], + "score": 1.0, + "content": "Table 5: Ablation study: Importance of fine-tuning for AWD-LSTM 3-layer model. Perplexity for", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 173, + 351, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 351, + 186 + ], + "score": 1.0, + "content": "the Penn Treebank and WikiText-2 language modeling tasks.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 188, + 199, + 421, + 258 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 188, + 199, + 421, + 258 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 188, + 199, + 421, + 258 + ], + "spans": [ + { + "bbox": [ + 188, + 199, + 421, + 258 + ], + "score": 0.974, + "html": "
Hyper-parameterPossible values
batch size non-monotone interval[10,20,30, 40] [5,10,20,40,60,100]
K -FD or ELD strength weight decayU(0, 0.3) U(0.6 × 10-6,2.4 × 10-6)
", + "type": "table", + "image_path": "81288747a210a7748b2bb2edb3d9e19287db61a5f4d0244a78b709167c48133c.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 188, + 199, + 421, + 213.75 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 188, + 213.75, + 421, + 228.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 188, + 228.5, + 421, + 243.25 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 188, + 243.25, + 421, + 258.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 261, + 504, + 294 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 259, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 506, + 275 + ], + "score": 1.0, + "content": "Table 6: Ablation study: Candidate hyper-parameters possible used in the grid search for comparing", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 306, + 284 + ], + "score": 1.0, + "content": "fraternal dropout and expectation linear dropout.", + "type": "text" + }, + { + "bbox": [ + 307, + 272, + 337, + 283 + ], + "score": 0.93, + "content": "U ( a , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "is the uniform distribution on the interval", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 282, + 364, + 296 + ], + "spans": [ + { + "bbox": [ + 107, + 282, + 127, + 294 + ], + "score": 0.9, + "content": "[ a , b ]", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 282, + 364, + 296 + ], + "score": 1.0, + "content": ". For finite sets, each value is drawn with equal probability.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.25 + }, + { + "type": "table", + "bbox": [ + 108, + 312, + 502, + 348 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 312, + 502, + 348 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 108, + 312, + 502, + 348 + ], + "spans": [ + { + "bbox": [ + 108, + 312, + 502, + 348 + ], + "score": 0.968, + "html": "
ModelBestTop5 avgTop10 avgBeating baseline runs (out of)
Expectation lineardropout59.460.160.56(208)
Fraternal dropout59.459.659.914 (203)
", + "type": "table", + "image_path": "080b0198eaea0fe2ad695f2300f5f44e370afb3f693385f868df02e311bdb4d7.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 108, + 312, + 502, + 324.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 108, + 324.0, + 502, + 336.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 108, + 336.0, + 502, + 348.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 352, + 504, + 386 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 351, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 506, + 365 + ], + "score": 1.0, + "content": "Table 7: Ablation study: Fraternal dropout and expectation linear dropout comparison. Perplex-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "ity on the Penn Treebank validation dateset. Fraternal dropout is more robust to different hyper-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 374, + 504, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 504, + 387 + ], + "score": 1.0, + "content": "parameters choice as twice as much runs finished performing better than the baseline model (60.7).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 107, + 413, + 444, + 423 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 447, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 447, + 424 + ], + "score": 1.0, + "content": "5.5 FRATERNAL DROPOUT AND EXPECTATION LINEAR DROPOUT COMPARISON", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 435, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 436, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 505, + 448 + ], + "score": 1.0, + "content": "We perform extensive grid search for the baseline model from Subsection 4.1 (an AWD-LSTM", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "score": 1.0, + "content": "3-layer architecture) trained with either fraternal dropout or expectation linear dropout regulariza-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "tions, to further contrast the performance of these two methods. The experiments are run without", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 469, + 233, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 233, + 480 + ], + "score": 1.0, + "content": "fine-tuning on the PTB dataset.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 436, + 505, + 480 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 505, + 529 + ], + "lines": [ + { + "bbox": [ + 105, + 486, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 497 + ], + "score": 1.0, + "content": "In each run, all five dropout rates are randomly altered (they are set to their original value, as in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "Merity et al. (2017a), multiplied by a value drawn from the uniform distribution on the interval", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "[0.5, 1.5]) and the rest of the hyper-parameters are drawn as shown in Table 6. As in Subsection 4.1,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 519, + 275, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 275, + 531 + ], + "score": 1.0, + "content": "AR and TAR regularizers are deactivated.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 486, + 506, + 531 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 535, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 506, + 548 + ], + "score": 1.0, + "content": "Together we run more than 400 experiments. The results are presented in Table 7. Both FD and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 545, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 506, + 560 + ], + "score": 1.0, + "content": "ELD perform better than the baseline model that instead uses AR and TAR regularizers. Hence,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "score": 1.0, + "content": "we confirm our previous finding (see Subsection 5.3) that both FD and ELD are better. However,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "as found previously for smaller model in Subsection 5.1, the convergence of FD is faster than that", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "of ELD. Additionally, fraternal dropout is more robust to different hyper-parameters choice (more", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 591, + 445, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 445, + 603 + ], + "score": 1.0, + "content": "runs performing better than the baseline and better average for top performing runs).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 534, + 506, + 603 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 625, + 195, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 624, + 197, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 197, + 641 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "In this paper we propose a simple regularization method for RNNs called fraternal dropout that acts", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "as a regularization by reducing the variance in model predictions across different dropout masks. We", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 676, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 691 + ], + "score": 1.0, + "content": "show that our model achieves state-of-the-art results on benchmark language modeling tasks along", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "with faster convergence. We also analytically study the relationship between our regularization and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "expectation linear dropout Ma et al. (2016). We perform a number of ablation studies to evaluate", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "our model from different aspects and carefully compare it with related methods both qualitatively", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 721, + 182, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 182, + 734 + ], + "score": 1.0, + "content": "and quantitatively.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 655, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 225, + 93 + ], + "lines": [ + { + "bbox": [ + 107, + 80, + 226, + 96 + ], + "spans": [ + { + "bbox": [ + 107, + 80, + 226, + 96 + ], + "score": 1.0, + "content": "ACKNOWLEDGEMENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "The authors would like to acknowledge the support of the following agencies for research funding", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "score": 1.0, + "content": "and computing support: NSERC, CIFAR, and IVADO. We would like to thank Rosemary Nan Ke", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 128, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 505, + 140 + ], + "score": 1.0, + "content": "and Philippe Lacaille for their thoughts and comments throughout the project. We would also like", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 138, + 396, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 396, + 151 + ], + "score": 1.0, + "content": "to thank Stanisław Jastrz˛ebski† and Evan Racah† for useful discussions.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 107, + 167, + 176, + 178 + ], + "lines": [ + { + "bbox": [ + 106, + 167, + 176, + 180 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 176, + 180 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 185, + 504, + 208 + ], + "lines": [ + { + "bbox": [ + 106, + 185, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 506, + 198 + ], + "score": 1.0, + "content": "Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 196, + 497, + 208 + ], + "spans": [ + { + "bbox": [ + 115, + 196, + 497, + 208 + ], + "score": 1.0, + "content": "gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 108, + 213, + 503, + 237 + ], + "lines": [ + { + "bbox": [ + 107, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 107, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "Tim Cooijmans, Nicolas Ballas, César Laurent, and Aaron C. Courville. Recurrent batch normal-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 224, + 486, + 237 + ], + "spans": [ + { + "bbox": [ + 116, + 224, + 486, + 237 + ], + "score": 1.0, + "content": "ization. CoRR, abs/1603.09025, 2016. URL http://arxiv.org/abs/1603.09025.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 108, + 243, + 502, + 266 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 504, + 256 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 504, + 256 + ], + "score": 1.0, + "content": "Yarin Gal and Zoubin Ghahramani. A theoretically grounded application of dropout in recurrent", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 254, + 496, + 268 + ], + "spans": [ + { + "bbox": [ + 115, + 254, + 496, + 268 + ], + "score": 1.0, + "content": "neural networks. In Advances in neural information processing systems, pp. 1019–1027, 2016.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 108, + 272, + 502, + 295 + ], + "lines": [ + { + "bbox": [ + 105, + 271, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 288 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 284, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 115, + 284, + 482, + 297 + ], + "score": 1.0, + "content": "nition. CoRR, abs/1512.03385, 2015. URL http://arxiv.org/abs/1512.03385.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 504, + 324 + ], + "lines": [ + { + "bbox": [ + 105, + 301, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 505, + 316 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 313, + 191, + 325 + ], + "spans": [ + { + "bbox": [ + 116, + 313, + 191, + 325 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 331, + 504, + 354 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "Hakan Inan, Khashayar Khosravi, and Richard Socher. Tying word vectors and word classifiers: A", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 342, + 438, + 355 + ], + "spans": [ + { + "bbox": [ + 115, + 342, + 438, + 355 + ], + "score": 1.0, + "content": "loss framework for language modeling. arXiv preprint arXiv:1611.01462, 2016.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 360, + 505, + 416 + ], + "lines": [ + { + "bbox": [ + 106, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 371, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 115, + 371, + 506, + 385 + ], + "score": 1.0, + "content": "Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, et al. Moses: Open source", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 382, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 115, + 382, + 505, + 396 + ], + "score": 1.0, + "content": "toolkit for statistical machine translation. In Proceedings of the 45th annual meeting of the ACL", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 115, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "on interactive poster and demonstration sessions, pp. 177–180. Association for Computational", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 405, + 190, + 417 + ], + "spans": [ + { + "bbox": [ + 115, + 405, + 190, + 417 + ], + "score": 1.0, + "content": "Linguistics, 2007.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 505, + 467 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 504, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 504, + 435 + ], + "score": 1.0, + "content": "David Krueger, Tegan Maharaj, János Kramár, Mohammad Pezeshki, Nicolas Ballas, Nan Rose-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 434, + 504, + 446 + ], + "spans": [ + { + "bbox": [ + 116, + 434, + 504, + 446 + ], + "score": 1.0, + "content": "mary Ke, Anirudh Goyal, Yoshua Bengio, Hugo Larochelle, Aaron Courville, et al. Zoneout:", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 114, + 444, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 114, + 444, + 506, + 459 + ], + "score": 1.0, + "content": "Regularizing rnns by randomly preserving hidden activations. arXiv preprint arXiv:1606.01305,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 455, + 142, + 468 + ], + "spans": [ + { + "bbox": [ + 115, + 455, + 142, + 468 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 474, + 504, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 506, + 488 + ], + "score": 1.0, + "content": "Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 486, + 219, + 497 + ], + "spans": [ + { + "bbox": [ + 115, + 486, + 219, + 497 + ], + "score": 1.0, + "content": "arXiv:1610.02242, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 503, + 506, + 538 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 517 + ], + "score": 1.0, + "content": "César Laurent, Gabriel Pereyra, Philémon Brakel, Ying Zhang, and Yoshua Bengio. Batch normal-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 116, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "ized recurrent neural networks. In Acoustics, Speech and Signal Processing (ICASSP), 2016 IEEE", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 117, + 526, + 351, + 538 + ], + "spans": [ + { + "bbox": [ + 117, + 526, + 351, + 538 + ], + "score": 1.0, + "content": "International Conference on, pp. 2657–2661. IEEE, 2016.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 544, + 504, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 557 + ], + "score": 1.0, + "content": "Xuezhe Ma, Yingkai Gao, Zhiting Hu, Yaoliang Yu, Yuntian Deng, and Eduard Hovy. Dropout with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 555, + 414, + 567 + ], + "spans": [ + { + "bbox": [ + 116, + 555, + 414, + 567 + ], + "score": 1.0, + "content": "expectation-linear regularization. arXiv preprint arXiv:1609.08017, 2016.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 573, + 504, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 572, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 587 + ], + "score": 1.0, + "content": "Stephan Mandt, Matthew D Hoffman, and David M Blei. Stochastic gradient descent as approximate", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 585, + 359, + 597 + ], + "spans": [ + { + "bbox": [ + 115, + 585, + 359, + 597 + ], + "score": 1.0, + "content": "bayesian inference. arXiv preprint arXiv:1704.04289, 2017.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 602, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "score": 1.0, + "content": "Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 613, + 470, + 626 + ], + "spans": [ + { + "bbox": [ + 115, + 613, + 470, + 626 + ], + "score": 1.0, + "content": "corpus of english: The penn treebank. Computational linguistics, 19(2):313–330, 1993.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 632, + 504, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 630, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 506, + 646 + ], + "score": 1.0, + "content": "Gábor Melis, Chris Dyer, and Phil Blunsom. On the state of the art of evaluation in neural language", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 644, + 313, + 655 + ], + "spans": [ + { + "bbox": [ + 116, + 644, + 313, + 655 + ], + "score": 1.0, + "content": "models. arXiv preprint arXiv:1707.05589, 2017.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 108, + 661, + 504, + 684 + ], + "lines": [ + { + "bbox": [ + 106, + 661, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 661, + 505, + 674 + ], + "score": 1.0, + "content": "Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 672, + 487, + 685 + ], + "spans": [ + { + "bbox": [ + 115, + 672, + 487, + 685 + ], + "score": 1.0, + "content": "models. CoRR, abs/1609.07843, 2016. URL http://arxiv.org/abs/1609.07843.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 691, + 504, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 691, + 504, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 504, + 704 + ], + "score": 1.0, + "content": "Stephen Merity, Nitish Shirish Keskar, and Richard Socher. Regularizing and optimizing lstm lan-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 702, + 344, + 714 + ], + "spans": [ + { + "bbox": [ + 115, + 702, + 344, + 714 + ], + "score": 1.0, + "content": "guage models. arXiv preprint arXiv:1708.02182, 2017a.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 120, + 721, + 200, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 719, + 201, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 201, + 734 + ], + "score": 1.0, + "content": "†equal contribution ^¨", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 225, + 93 + ], + "lines": [ + { + "bbox": [ + 107, + 80, + 226, + 96 + ], + "spans": [ + { + "bbox": [ + 107, + 80, + 226, + 96 + ], + "score": 1.0, + "content": "ACKNOWLEDGEMENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "The authors would like to acknowledge the support of the following agencies for research funding", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "score": 1.0, + "content": "and computing support: NSERC, CIFAR, and IVADO. We would like to thank Rosemary Nan Ke", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 128, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 505, + 140 + ], + "score": 1.0, + "content": "and Philippe Lacaille for their thoughts and comments throughout the project. We would also like", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 138, + 396, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 396, + 151 + ], + "score": 1.0, + "content": "to thank Stanisław Jastrz˛ebski† and Evan Racah† for useful discussions.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 106, + 505, + 151 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 167, + 176, + 178 + ], + "lines": [ + { + "bbox": [ + 106, + 167, + 176, + 180 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 176, + 180 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 185, + 504, + 208 + ], + "lines": [ + { + "bbox": [ + 106, + 185, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 506, + 198 + ], + "score": 1.0, + "content": "Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 196, + 497, + 208 + ], + "spans": [ + { + "bbox": [ + 115, + 196, + 497, + 208 + ], + "score": 1.0, + "content": "gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 106, + 185, + 506, + 208 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 213, + 503, + 237 + ], + "lines": [ + { + "bbox": [ + 107, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 107, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "Tim Cooijmans, Nicolas Ballas, César Laurent, and Aaron C. Courville. Recurrent batch normal-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 224, + 486, + 237 + ], + "spans": [ + { + "bbox": [ + 116, + 224, + 486, + 237 + ], + "score": 1.0, + "content": "ization. CoRR, abs/1603.09025, 2016. URL http://arxiv.org/abs/1603.09025.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 107, + 214, + 505, + 237 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 243, + 502, + 266 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 504, + 256 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 504, + 256 + ], + "score": 1.0, + "content": "Yarin Gal and Zoubin Ghahramani. A theoretically grounded application of dropout in recurrent", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 254, + 496, + 268 + ], + "spans": [ + { + "bbox": [ + 115, + 254, + 496, + 268 + ], + "score": 1.0, + "content": "neural networks. In Advances in neural information processing systems, pp. 1019–1027, 2016.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 106, + 243, + 504, + 268 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 272, + 502, + 295 + ], + "lines": [ + { + "bbox": [ + 105, + 271, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 288 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 284, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 115, + 284, + 482, + 297 + ], + "score": 1.0, + "content": "nition. CoRR, abs/1512.03385, 2015. URL http://arxiv.org/abs/1512.03385.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 271, + 505, + 297 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 504, + 324 + ], + "lines": [ + { + "bbox": [ + 105, + 301, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 505, + 316 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 313, + 191, + 325 + ], + "spans": [ + { + "bbox": [ + 116, + 313, + 191, + 325 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 301, + 505, + 325 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 331, + 504, + 354 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "Hakan Inan, Khashayar Khosravi, and Richard Socher. Tying word vectors and word classifiers: A", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 342, + 438, + 355 + ], + "spans": [ + { + "bbox": [ + 115, + 342, + 438, + 355 + ], + "score": 1.0, + "content": "loss framework for language modeling. arXiv preprint arXiv:1611.01462, 2016.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 331, + 505, + 355 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 360, + 505, + 416 + ], + "lines": [ + { + "bbox": [ + 106, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 371, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 115, + 371, + 506, + 385 + ], + "score": 1.0, + "content": "Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, et al. Moses: Open source", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 382, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 115, + 382, + 505, + 396 + ], + "score": 1.0, + "content": "toolkit for statistical machine translation. In Proceedings of the 45th annual meeting of the ACL", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 115, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "on interactive poster and demonstration sessions, pp. 177–180. Association for Computational", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 405, + 190, + 417 + ], + "spans": [ + { + "bbox": [ + 115, + 405, + 190, + 417 + ], + "score": 1.0, + "content": "Linguistics, 2007.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20, + "bbox_fs": [ + 106, + 361, + 506, + 417 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 505, + 467 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 504, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 504, + 435 + ], + "score": 1.0, + "content": "David Krueger, Tegan Maharaj, János Kramár, Mohammad Pezeshki, Nicolas Ballas, Nan Rose-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 434, + 504, + 446 + ], + "spans": [ + { + "bbox": [ + 116, + 434, + 504, + 446 + ], + "score": 1.0, + "content": "mary Ke, Anirudh Goyal, Yoshua Bengio, Hugo Larochelle, Aaron Courville, et al. Zoneout:", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 114, + 444, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 114, + 444, + 506, + 459 + ], + "score": 1.0, + "content": "Regularizing rnns by randomly preserving hidden activations. arXiv preprint arXiv:1606.01305,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 455, + 142, + 468 + ], + "spans": [ + { + "bbox": [ + 115, + 455, + 142, + 468 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 106, + 423, + 506, + 468 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 474, + 504, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 506, + 488 + ], + "score": 1.0, + "content": "Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 486, + 219, + 497 + ], + "spans": [ + { + "bbox": [ + 115, + 486, + 219, + 497 + ], + "score": 1.0, + "content": "arXiv:1610.02242, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 473, + 506, + 497 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 503, + 506, + 538 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 517 + ], + "score": 1.0, + "content": "César Laurent, Gabriel Pereyra, Philémon Brakel, Ying Zhang, and Yoshua Bengio. Batch normal-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 116, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "ized recurrent neural networks. In Acoustics, Speech and Signal Processing (ICASSP), 2016 IEEE", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 117, + 526, + 351, + 538 + ], + "spans": [ + { + "bbox": [ + 117, + 526, + 351, + 538 + ], + "score": 1.0, + "content": "International Conference on, pp. 2657–2661. IEEE, 2016.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 106, + 503, + 505, + 538 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 544, + 504, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 557 + ], + "score": 1.0, + "content": "Xuezhe Ma, Yingkai Gao, Zhiting Hu, Yaoliang Yu, Yuntian Deng, and Eduard Hovy. Dropout with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 555, + 414, + 567 + ], + "spans": [ + { + "bbox": [ + 116, + 555, + 414, + 567 + ], + "score": 1.0, + "content": "expectation-linear regularization. arXiv preprint arXiv:1609.08017, 2016.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 543, + 505, + 567 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 573, + 504, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 572, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 587 + ], + "score": 1.0, + "content": "Stephan Mandt, Matthew D Hoffman, and David M Blei. Stochastic gradient descent as approximate", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 585, + 359, + 597 + ], + "spans": [ + { + "bbox": [ + 115, + 585, + 359, + 597 + ], + "score": 1.0, + "content": "bayesian inference. arXiv preprint arXiv:1704.04289, 2017.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 572, + 505, + 597 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 602, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "score": 1.0, + "content": "Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 613, + 470, + 626 + ], + "spans": [ + { + "bbox": [ + 115, + 613, + 470, + 626 + ], + "score": 1.0, + "content": "corpus of english: The penn treebank. Computational linguistics, 19(2):313–330, 1993.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 106, + 603, + 505, + 626 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 632, + 504, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 630, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 506, + 646 + ], + "score": 1.0, + "content": "Gábor Melis, Chris Dyer, and Phil Blunsom. On the state of the art of evaluation in neural language", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 644, + 313, + 655 + ], + "spans": [ + { + "bbox": [ + 116, + 644, + 313, + 655 + ], + "score": 1.0, + "content": "models. arXiv preprint arXiv:1707.05589, 2017.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 630, + 506, + 655 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 661, + 504, + 684 + ], + "lines": [ + { + "bbox": [ + 106, + 661, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 661, + 505, + 674 + ], + "score": 1.0, + "content": "Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 672, + 487, + 685 + ], + "spans": [ + { + "bbox": [ + 115, + 672, + 487, + 685 + ], + "score": 1.0, + "content": "models. CoRR, abs/1609.07843, 2016. URL http://arxiv.org/abs/1609.07843.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 661, + 505, + 685 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 691, + 504, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 691, + 504, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 504, + 704 + ], + "score": 1.0, + "content": "Stephen Merity, Nitish Shirish Keskar, and Richard Socher. Regularizing and optimizing lstm lan-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 702, + 344, + 714 + ], + "spans": [ + { + "bbox": [ + 115, + 702, + 344, + 714 + ], + "score": 1.0, + "content": "guage models. arXiv preprint arXiv:1708.02182, 2017a.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 106, + 691, + 504, + 714 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Stephen Merity, Bryan McCann, and Richard Socher. Revisiting activation regularization for lan-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 94, + 334, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 334, + 106 + ], + "score": 1.0, + "content": "guage rnns. arXiv preprint arXiv:1708.01009, 2017b.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 111, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "score": 1.0, + "content": "Tomas Mikolov, Martin Karafiát, Lukas Burget, Jan Cernocky, and Sanjeev Khudanpur. Recurrent `", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 123, + 430, + 136 + ], + "spans": [ + { + "bbox": [ + 116, + 123, + 430, + 136 + ], + "score": 1.0, + "content": "neural network based language model. In Interspeech, volume 2, pp. 3, 2010.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 105, + 142, + 503, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 140, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 505, + 157 + ], + "score": 1.0, + "content": "Boris T Polyak and Anatoli B Juditsky. Acceleration of stochastic approximation by averaging.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 153, + 384, + 166 + ], + "spans": [ + { + "bbox": [ + 116, + 153, + 384, + 166 + ], + "score": 1.0, + "content": "SIAM Journal on Control and Optimization, 30(4):838–855, 1992.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 106, + 171, + 504, + 195 + ], + "lines": [ + { + "bbox": [ + 106, + 171, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 505, + 185 + ], + "score": 1.0, + "content": "Ofir Press and Lior Wolf. Using the output embedding to improve language models. arXiv preprint", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 183, + 219, + 194 + ], + "spans": [ + { + "bbox": [ + 116, + 183, + 219, + 194 + ], + "score": 1.0, + "content": "arXiv:1608.05859, 2016.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 201, + 504, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 201, + 504, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 504, + 214 + ], + "score": 1.0, + "content": "Antti Rasmus, Harri Valpola, Mikko Honkala, Mathias Berglund, and Tapani Raiko. Semi-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 213, + 504, + 226 + ], + "spans": [ + { + "bbox": [ + 116, + 213, + 504, + 226 + ], + "score": 1.0, + "content": "supervised learning with ladder network. CoRR, abs/1507.02672, 2015. URL http://arxiv.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 225, + 229, + 236 + ], + "spans": [ + { + "bbox": [ + 115, + 225, + 229, + 236 + ], + "score": 1.0, + "content": "org/abs/1507.02672.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 243, + 503, + 277 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 504, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 504, + 255 + ], + "score": 1.0, + "content": "Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transfor-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 255, + 504, + 266 + ], + "spans": [ + { + "bbox": [ + 116, + 255, + 504, + 266 + ], + "score": 1.0, + "content": "mations and perturbations for deep semi-supervised learning. In Advances in Neural Information", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 265, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 116, + 265, + 289, + 277 + ], + "score": 1.0, + "content": "Processing Systems, pp. 1163–1171, 2016.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 105, + 283, + 502, + 307 + ], + "lines": [ + { + "bbox": [ + 105, + 283, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 505, + 296 + ], + "score": 1.0, + "content": "Dmitriy Serdyuk, Nan Rosemary Ke, Alessandro Sordoni, Adam Trischler, Chris Pal, and Yoshua", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 294, + 424, + 307 + ], + "spans": [ + { + "bbox": [ + 115, + 294, + 424, + 307 + ], + "score": 1.0, + "content": "Bengio. Twin networks: Matching the future for sequence generation. 2017.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 313, + 504, + 347 + ], + "lines": [ + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 324, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 115, + 324, + 505, + 338 + ], + "score": 1.0, + "content": "Dropout: a simple way to prevent neural networks from overfitting. Journal of machine learning", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 336, + 255, + 348 + ], + "spans": [ + { + "bbox": [ + 116, + 336, + 255, + 348 + ], + "score": 1.0, + "content": "research, 15(1):1929–1958, 2014.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 354, + 504, + 388 + ], + "lines": [ + { + "bbox": [ + 105, + 352, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 506, + 368 + ], + "score": 1.0, + "content": "Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 365, + 504, + 378 + ], + "spans": [ + { + "bbox": [ + 115, + 365, + 504, + 378 + ], + "score": 1.0, + "content": "caption generator. CoRR, abs/1411.4555, 2014. URL http://arxiv.org/abs/1411.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 375, + 147, + 389 + ], + "spans": [ + { + "bbox": [ + 115, + 375, + 147, + 389 + ], + "score": 1.0, + "content": "4555.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 394, + 504, + 429 + ], + "lines": [ + { + "bbox": [ + 104, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 104, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "Li Wan, Matthew Zeiler, Sixin Zhang, Yann L Cun, and Rob Fergus. Regularization of neural", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 116, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "networks using dropconnect. In Proceedings of the 30th international conference on machine", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 417, + 291, + 430 + ], + "spans": [ + { + "bbox": [ + 116, + 417, + 291, + 430 + ], + "score": 1.0, + "content": "learning (ICML-13), pp. 1058–1066, 2013.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 435, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 506, + 448 + ], + "score": 1.0, + "content": "Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron C. Courville, Ruslan Salakhutdinov,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 446, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 115, + 446, + 505, + 460 + ], + "score": 1.0, + "content": "Richard S. Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 458, + 504, + 470 + ], + "spans": [ + { + "bbox": [ + 116, + 458, + 504, + 470 + ], + "score": 1.0, + "content": "with visual attention. CoRR, abs/1502.03044, 2015. URL http://arxiv.org/abs/1502.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 469, + 152, + 480 + ], + "spans": [ + { + "bbox": [ + 115, + 469, + 152, + 480 + ], + "score": 1.0, + "content": "03044.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 104, + 487, + 504, + 511 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 504, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 504, + 500 + ], + "score": 1.0, + "content": "Wojciech Zaremba, Ilya Sutskever, and Oriol Vinyals. Recurrent neural network regularization.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 499, + 273, + 510 + ], + "spans": [ + { + "bbox": [ + 116, + 499, + 273, + 510 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1409.2329, 2014.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 104, + 517, + 504, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "score": 1.0, + "content": "Julian Georg Zilly, Rupesh Kumar Srivastava, Jan Koutník, and Jürgen Schmidhuber. Recurrent", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 529, + 358, + 542 + ], + "spans": [ + { + "bbox": [ + 115, + 529, + 358, + 542 + ], + "score": 1.0, + "content": "highway networks. arXiv preprint arXiv:1607.03474, 2016.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Stephen Merity, Bryan McCann, and Richard Socher. Revisiting activation regularization for lan-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 94, + 334, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 334, + 106 + ], + "score": 1.0, + "content": "guage rnns. arXiv preprint arXiv:1708.01009, 2017b.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 81, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 111, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "score": 1.0, + "content": "Tomas Mikolov, Martin Karafiát, Lukas Burget, Jan Cernocky, and Sanjeev Khudanpur. Recurrent `", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 123, + 430, + 136 + ], + "spans": [ + { + "bbox": [ + 116, + 123, + 430, + 136 + ], + "score": 1.0, + "content": "neural network based language model. In Interspeech, volume 2, pp. 3, 2010.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 111, + 505, + 136 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 142, + 503, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 140, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 505, + 157 + ], + "score": 1.0, + "content": "Boris T Polyak and Anatoli B Juditsky. Acceleration of stochastic approximation by averaging.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 153, + 384, + 166 + ], + "spans": [ + { + "bbox": [ + 116, + 153, + 384, + 166 + ], + "score": 1.0, + "content": "SIAM Journal on Control and Optimization, 30(4):838–855, 1992.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 140, + 505, + 166 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 171, + 504, + 195 + ], + "lines": [ + { + "bbox": [ + 106, + 171, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 505, + 185 + ], + "score": 1.0, + "content": "Ofir Press and Lior Wolf. Using the output embedding to improve language models. arXiv preprint", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 183, + 219, + 194 + ], + "spans": [ + { + "bbox": [ + 116, + 183, + 219, + 194 + ], + "score": 1.0, + "content": "arXiv:1608.05859, 2016.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 106, + 171, + 505, + 194 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 201, + 504, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 201, + 504, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 504, + 214 + ], + "score": 1.0, + "content": "Antti Rasmus, Harri Valpola, Mikko Honkala, Mathias Berglund, and Tapani Raiko. Semi-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 213, + 504, + 226 + ], + "spans": [ + { + "bbox": [ + 116, + 213, + 504, + 226 + ], + "score": 1.0, + "content": "supervised learning with ladder network. CoRR, abs/1507.02672, 2015. URL http://arxiv.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 225, + 229, + 236 + ], + "spans": [ + { + "bbox": [ + 115, + 225, + 229, + 236 + ], + "score": 1.0, + "content": "org/abs/1507.02672.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 201, + 504, + 236 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 243, + 503, + 277 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 504, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 504, + 255 + ], + "score": 1.0, + "content": "Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transfor-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 255, + 504, + 266 + ], + "spans": [ + { + "bbox": [ + 116, + 255, + 504, + 266 + ], + "score": 1.0, + "content": "mations and perturbations for deep semi-supervised learning. In Advances in Neural Information", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 265, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 116, + 265, + 289, + 277 + ], + "score": 1.0, + "content": "Processing Systems, pp. 1163–1171, 2016.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 106, + 243, + 504, + 277 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 283, + 502, + 307 + ], + "lines": [ + { + "bbox": [ + 105, + 283, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 505, + 296 + ], + "score": 1.0, + "content": "Dmitriy Serdyuk, Nan Rosemary Ke, Alessandro Sordoni, Adam Trischler, Chris Pal, and Yoshua", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 294, + 424, + 307 + ], + "spans": [ + { + "bbox": [ + 115, + 294, + 424, + 307 + ], + "score": 1.0, + "content": "Bengio. Twin networks: Matching the future for sequence generation. 2017.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 283, + 505, + 307 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 313, + 504, + 347 + ], + "lines": [ + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 324, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 115, + 324, + 505, + 338 + ], + "score": 1.0, + "content": "Dropout: a simple way to prevent neural networks from overfitting. Journal of machine learning", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 336, + 255, + 348 + ], + "spans": [ + { + "bbox": [ + 116, + 336, + 255, + 348 + ], + "score": 1.0, + "content": "research, 15(1):1929–1958, 2014.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 313, + 505, + 348 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 354, + 504, + 388 + ], + "lines": [ + { + "bbox": [ + 105, + 352, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 506, + 368 + ], + "score": 1.0, + "content": "Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 365, + 504, + 378 + ], + "spans": [ + { + "bbox": [ + 115, + 365, + 504, + 378 + ], + "score": 1.0, + "content": "caption generator. CoRR, abs/1411.4555, 2014. URL http://arxiv.org/abs/1411.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 375, + 147, + 389 + ], + "spans": [ + { + "bbox": [ + 115, + 375, + 147, + 389 + ], + "score": 1.0, + "content": "4555.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 352, + 506, + 389 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 394, + 504, + 429 + ], + "lines": [ + { + "bbox": [ + 104, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 104, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "Li Wan, Matthew Zeiler, Sixin Zhang, Yann L Cun, and Rob Fergus. Regularization of neural", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 116, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "networks using dropconnect. In Proceedings of the 30th international conference on machine", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 417, + 291, + 430 + ], + "spans": [ + { + "bbox": [ + 116, + 417, + 291, + 430 + ], + "score": 1.0, + "content": "learning (ICML-13), pp. 1058–1066, 2013.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 104, + 394, + 506, + 430 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 435, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 506, + 448 + ], + "score": 1.0, + "content": "Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron C. Courville, Ruslan Salakhutdinov,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 446, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 115, + 446, + 505, + 460 + ], + "score": 1.0, + "content": "Richard S. Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 458, + 504, + 470 + ], + "spans": [ + { + "bbox": [ + 116, + 458, + 504, + 470 + ], + "score": 1.0, + "content": "with visual attention. CoRR, abs/1502.03044, 2015. URL http://arxiv.org/abs/1502.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 469, + 152, + 480 + ], + "spans": [ + { + "bbox": [ + 115, + 469, + 152, + 480 + ], + "score": 1.0, + "content": "03044.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5, + "bbox_fs": [ + 106, + 435, + 506, + 480 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 487, + 504, + 511 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 504, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 504, + 500 + ], + "score": 1.0, + "content": "Wojciech Zaremba, Ilya Sutskever, and Oriol Vinyals. Recurrent neural network regularization.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 499, + 273, + 510 + ], + "spans": [ + { + "bbox": [ + 116, + 499, + 273, + 510 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1409.2329, 2014.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 487, + 504, + 510 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 517, + 504, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "score": 1.0, + "content": "Julian Georg Zilly, Rupesh Kumar Srivastava, Jan Koutník, and Jürgen Schmidhuber. Recurrent", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 529, + 358, + 542 + ], + "spans": [ + { + "bbox": [ + 115, + 529, + 358, + 542 + ], + "score": 1.0, + "content": "highway networks. arXiv preprint arXiv:1607.03474, 2016.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 518, + 505, + 542 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 161, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 163, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 163, + 96 + ], + "score": 1.0, + "content": "APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 106, + 231, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 232, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 232, + 119 + ], + "score": 1.0, + "content": "MONTE CARLO EVALUATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 127, + 505, + 204 + ], + "lines": [ + { + "bbox": [ + 106, + 127, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 139 + ], + "score": 1.0, + "content": "A well known way to address the gap between the train and evaluation mode of dropout is to perform", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "Monte Carlo sampling of masks and average the predictions during evaluation (MC-eval), and this", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "score": 1.0, + "content": "has been used for feed-forward networks. Since fraternal dropout addresses the same problem, we", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 160, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 505, + 171 + ], + "score": 1.0, + "content": "would like to clarify that it is not straight-forward and feasible to apply MC-eval for RNNs. In", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 170, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 506, + 184 + ], + "score": 1.0, + "content": "feed-forward networks, we average the output prediction scores from different masks. However, in", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "the case RNNs (for next step predictions), there is more than one way to perform such evaluation,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 192, + 305, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 305, + 205 + ], + "score": 1.0, + "content": "but each one is problematic. They are as follows:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 108, + 209, + 192, + 221 + ], + "lines": [ + { + "bbox": [ + 105, + 207, + 194, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 194, + 224 + ], + "score": 1.0, + "content": "1. Online averaging", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 226, + 504, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 504, + 239 + ], + "score": 1.0, + "content": "Consider that we first make the prediction at time step 1 using different masks by averaging the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 238, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 506, + 249 + ], + "score": 1.0, + "content": "prediction score. Then we use this output to feed as input to the time step 2, then use different masks", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 248, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 506, + 260 + ], + "score": 1.0, + "content": "at time step 2 to generate the output at time step 2, and so on. But in order to do so, because of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "the way RNNs work, we also need to feed the previous time hidden state to time step 2. One way", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "score": 1.0, + "content": "would be to average the hidden states over different masks at time step 1. But the hidden space can", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "in general be highly nonlinear, and it is not clear if averaging in this space is a good strategy. This", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 293, + 207, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 207, + 304 + ], + "score": 1.0, + "content": "approach is not justified.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 308, + 503, + 331 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 504, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 504, + 322 + ], + "score": 1.0, + "content": "Besides, this strategy as a whole is extremely time consuming because we would need to sequentially", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 319, + 332, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 332, + 333 + ], + "score": 1.0, + "content": "make predictions with multiple masks at each time step.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 107, + 337, + 203, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 334, + 205, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 205, + 353 + ], + "score": 1.0, + "content": "2. Sequence averaging", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 505, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "Let’s consider that we use a different mask each time we want to generate a sequence, and then we", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 364, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 506, + 377 + ], + "score": 1.0, + "content": "average the prediction scores, and compute the argmax (at each time step) to get the actual generated", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 376, + 149, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 149, + 389 + ], + "score": 1.0, + "content": "sequence.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 505, + 470 + ], + "lines": [ + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 413, + 405 + ], + "score": 1.0, + "content": "In this case, notice it is not guaranteed that the predicted word at time step", + "type": "text" + }, + { + "bbox": [ + 413, + 394, + 419, + 403 + ], + "score": 0.57, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "due to averaging the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "predictions would lead to the next word (generated by the same process) if we were to feed the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 415, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 145, + 427 + ], + "score": 1.0, + "content": "time step", + "type": "text" + }, + { + "bbox": [ + 145, + 416, + 151, + 425 + ], + "score": 0.68, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 415, + 278, + 427 + ], + "score": 1.0, + "content": "output as input to the time step", + "type": "text" + }, + { + "bbox": [ + 279, + 415, + 300, + 425 + ], + "score": 0.88, + "content": "t + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 415, + 505, + 427 + ], + "score": 1.0, + "content": ". For example, with different dropout masks, if the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 424, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 104, + 424, + 163, + 439 + ], + "score": 1.0, + "content": "probability of", + "type": "text" + }, + { + "bbox": [ + 163, + 425, + 177, + 436 + ], + "score": 0.87, + "content": "\\bar { 1 } ^ { s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 424, + 265, + 439 + ], + "score": 1.0, + "content": "time step outputs are:", + "type": "text" + }, + { + "bbox": [ + 266, + 426, + 292, + 436 + ], + "score": 0.85, + "content": "1 4 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 424, + 310, + 439 + ], + "score": 1.0, + "content": "), he", + "type": "text" + }, + { + "bbox": [ + 310, + 426, + 336, + 437 + ], + "score": 0.87, + "content": "( 3 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 424, + 355, + 439 + ], + "score": 1.0, + "content": ", she", + "type": "text" + }, + { + "bbox": [ + 356, + 426, + 381, + 437 + ], + "score": 0.89, + "content": "( 3 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 424, + 506, + 439 + ], + "score": 1.0, + "content": ", and the probability of the 2nd", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 437, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 209, + 449 + ], + "score": 1.0, + "content": "time step outputs are: am", + "type": "text" + }, + { + "bbox": [ + 210, + 437, + 235, + 448 + ], + "score": 0.86, + "content": "( 3 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 437, + 248, + 449 + ], + "score": 1.0, + "content": ", is", + "type": "text" + }, + { + "bbox": [ + 248, + 437, + 273, + 448 + ], + "score": 0.87, + "content": "( 6 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 437, + 296, + 449 + ], + "score": 1.0, + "content": ", was", + "type": "text" + }, + { + "bbox": [ + 296, + 437, + 321, + 448 + ], + "score": 0.86, + "content": "( 1 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 437, + 505, + 449 + ], + "score": 1.0, + "content": ". Then the averaged prediction score followed", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "score": 1.0, + "content": "by argmax will result in the prediction “I is”, but this would be incorrect. A similar concern applies", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 458, + 305, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 305, + 471 + ], + "score": 1.0, + "content": "for output predictions varying in temporal length.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 475, + 505, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 489 + ], + "score": 1.0, + "content": "Hence, this approach can not be used to generate a sequence (it has to be done by by sampling a mask", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 486, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 500 + ], + "score": 1.0, + "content": "and generating a single sequence). However, this approach may be used to estimate the probability", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 498, + 281, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 281, + 510 + ], + "score": 1.0, + "content": "assigned by the model to a given sequence.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 514, + 505, + 603 + ], + "lines": [ + { + "bbox": [ + 106, + 514, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 506, + 527 + ], + "score": 1.0, + "content": "Nonetheless, we run experiments on the PTB dataset using MC-eval (the results are summarized", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "in Table 8). We start with a simple comparison that compares fraternal dropout with the averaged", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 536, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 548 + ], + "score": 1.0, + "content": "mask and the AWD-LSTM 3-layer baseline with a single fixed mask that we call MC1. The MC1", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 547, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 505, + 559 + ], + "score": 1.0, + "content": "model performs much worse than fraternal dropout. Hence, it would be hard to use MC1 model", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 558, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 570 + ], + "score": 1.0, + "content": "in practice because a single sample is inaccurate. We also check MC-eval for a larger number of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "models (MC50) (50 models were used since we were not able to fit more models simultaneously on", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 580, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 104, + 580, + 505, + 592 + ], + "score": 1.0, + "content": "a single GPU). The final results for MC50 are worse than the baseline which uses the averaged mask.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "For comparison, we also evaluate MC10. Note that no fine-tuning is used for the above experiments.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5 + }, + { + "type": "table", + "bbox": [ + 189, + 610, + 423, + 690 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 189, + 610, + 423, + 690 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 189, + 610, + 423, + 690 + ], + "spans": [ + { + "bbox": [ + 189, + 610, + 423, + 690 + ], + "score": 0.967, + "html": "
ModelValidationTest
MC1 MC10 MC50 Baseline (average mask) Fraternal dropout92.2 (± 0.5) 66.2 (± 0.2) 64.4 (± 0.1) 60.7 59.889.2 (± 0.5) 63.7 (± 0.2) 62.1 (± 0.1) 58.8 58.0
", + "type": "table", + "image_path": "910137973af0702bba6415fa366c5234d0c42e20a24f609d5bd82422af5dbf96.jpg" + } + ] + } + ], + "index": 43.5, + "virtual_lines": [ + { + "bbox": [ + 189, + 610, + 423, + 623.3333333333334 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 189, + 623.3333333333334, + 423, + 636.6666666666667 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 189, + 636.6666666666667, + 423, + 650.0000000000001 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 189, + 650.0000000000001, + 423, + 663.3333333333335 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 189, + 663.3333333333335, + 423, + 676.6666666666669 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 189, + 676.6666666666669, + 423, + 690.0000000000002 + ], + "spans": [], + "index": 46 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 693, + 504, + 716 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 690, + 505, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 505, + 707 + ], + "score": 1.0, + "content": "Table 8: Appendix: Monte Carlo evaluation. Perplexity on Penn Treebank word level language", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 704, + 425, + 717 + ], + "spans": [ + { + "bbox": [ + 106, + 704, + 425, + 717 + ], + "score": 1.0, + "content": "modeling task using Monte Carlo sampling, fraternal dropout or average mask.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5 + } + ], + "index": 45.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 161, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 163, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 163, + 96 + ], + "score": 1.0, + "content": "APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 106, + 231, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 232, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 232, + 119 + ], + "score": 1.0, + "content": "MONTE CARLO EVALUATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 127, + 505, + 204 + ], + "lines": [ + { + "bbox": [ + 106, + 127, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 139 + ], + "score": 1.0, + "content": "A well known way to address the gap between the train and evaluation mode of dropout is to perform", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "Monte Carlo sampling of masks and average the predictions during evaluation (MC-eval), and this", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "score": 1.0, + "content": "has been used for feed-forward networks. Since fraternal dropout addresses the same problem, we", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 160, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 505, + 171 + ], + "score": 1.0, + "content": "would like to clarify that it is not straight-forward and feasible to apply MC-eval for RNNs. In", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 170, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 506, + 184 + ], + "score": 1.0, + "content": "feed-forward networks, we average the output prediction scores from different masks. However, in", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "the case RNNs (for next step predictions), there is more than one way to perform such evaluation,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 192, + 305, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 305, + 205 + ], + "score": 1.0, + "content": "but each one is problematic. They are as follows:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 127, + 506, + 205 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 209, + 192, + 221 + ], + "lines": [ + { + "bbox": [ + 105, + 207, + 194, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 194, + 224 + ], + "score": 1.0, + "content": "1. Online averaging", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 226, + 504, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 504, + 239 + ], + "score": 1.0, + "content": "Consider that we first make the prediction at time step 1 using different masks by averaging the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 238, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 506, + 249 + ], + "score": 1.0, + "content": "prediction score. Then we use this output to feed as input to the time step 2, then use different masks", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 248, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 506, + 260 + ], + "score": 1.0, + "content": "at time step 2 to generate the output at time step 2, and so on. But in order to do so, because of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "the way RNNs work, we also need to feed the previous time hidden state to time step 2. One way", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "score": 1.0, + "content": "would be to average the hidden states over different masks at time step 1. But the hidden space can", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "in general be highly nonlinear, and it is not clear if averaging in this space is a good strategy. This", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 293, + 207, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 207, + 304 + ], + "score": 1.0, + "content": "approach is not justified.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 226, + 506, + 304 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 308, + 503, + 331 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 504, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 504, + 322 + ], + "score": 1.0, + "content": "Besides, this strategy as a whole is extremely time consuming because we would need to sequentially", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 319, + 332, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 332, + 333 + ], + "score": 1.0, + "content": "make predictions with multiple masks at each time step.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 308, + 504, + 333 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 337, + 203, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 334, + 205, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 205, + 353 + ], + "score": 1.0, + "content": "2. Sequence averaging", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 505, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "Let’s consider that we use a different mask each time we want to generate a sequence, and then we", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 364, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 506, + 377 + ], + "score": 1.0, + "content": "average the prediction scores, and compute the argmax (at each time step) to get the actual generated", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 376, + 149, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 149, + 389 + ], + "score": 1.0, + "content": "sequence.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 353, + 506, + 389 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 505, + 470 + ], + "lines": [ + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 413, + 405 + ], + "score": 1.0, + "content": "In this case, notice it is not guaranteed that the predicted word at time step", + "type": "text" + }, + { + "bbox": [ + 413, + 394, + 419, + 403 + ], + "score": 0.57, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "due to averaging the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "predictions would lead to the next word (generated by the same process) if we were to feed the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 415, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 145, + 427 + ], + "score": 1.0, + "content": "time step", + "type": "text" + }, + { + "bbox": [ + 145, + 416, + 151, + 425 + ], + "score": 0.68, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 415, + 278, + 427 + ], + "score": 1.0, + "content": "output as input to the time step", + "type": "text" + }, + { + "bbox": [ + 279, + 415, + 300, + 425 + ], + "score": 0.88, + "content": "t + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 415, + 505, + 427 + ], + "score": 1.0, + "content": ". For example, with different dropout masks, if the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 424, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 104, + 424, + 163, + 439 + ], + "score": 1.0, + "content": "probability of", + "type": "text" + }, + { + "bbox": [ + 163, + 425, + 177, + 436 + ], + "score": 0.87, + "content": "\\bar { 1 } ^ { s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 424, + 265, + 439 + ], + "score": 1.0, + "content": "time step outputs are:", + "type": "text" + }, + { + "bbox": [ + 266, + 426, + 292, + 436 + ], + "score": 0.85, + "content": "1 4 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 424, + 310, + 439 + ], + "score": 1.0, + "content": "), he", + "type": "text" + }, + { + "bbox": [ + 310, + 426, + 336, + 437 + ], + "score": 0.87, + "content": "( 3 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 424, + 355, + 439 + ], + "score": 1.0, + "content": ", she", + "type": "text" + }, + { + "bbox": [ + 356, + 426, + 381, + 437 + ], + "score": 0.89, + "content": "( 3 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 424, + 506, + 439 + ], + "score": 1.0, + "content": ", and the probability of the 2nd", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 437, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 209, + 449 + ], + "score": 1.0, + "content": "time step outputs are: am", + "type": "text" + }, + { + "bbox": [ + 210, + 437, + 235, + 448 + ], + "score": 0.86, + "content": "( 3 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 437, + 248, + 449 + ], + "score": 1.0, + "content": ", is", + "type": "text" + }, + { + "bbox": [ + 248, + 437, + 273, + 448 + ], + "score": 0.87, + "content": "( 6 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 437, + 296, + 449 + ], + "score": 1.0, + "content": ", was", + "type": "text" + }, + { + "bbox": [ + 296, + 437, + 321, + 448 + ], + "score": 0.86, + "content": "( 1 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 437, + 505, + 449 + ], + "score": 1.0, + "content": ". Then the averaged prediction score followed", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "score": 1.0, + "content": "by argmax will result in the prediction “I is”, but this would be incorrect. A similar concern applies", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 458, + 305, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 305, + 471 + ], + "score": 1.0, + "content": "for output predictions varying in temporal length.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26, + "bbox_fs": [ + 104, + 393, + 506, + 471 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 475, + 505, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 489 + ], + "score": 1.0, + "content": "Hence, this approach can not be used to generate a sequence (it has to be done by by sampling a mask", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 486, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 500 + ], + "score": 1.0, + "content": "and generating a single sequence). However, this approach may be used to estimate the probability", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 498, + 281, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 281, + 510 + ], + "score": 1.0, + "content": "assigned by the model to a given sequence.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 474, + 506, + 510 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 514, + 505, + 603 + ], + "lines": [ + { + "bbox": [ + 106, + 514, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 506, + 527 + ], + "score": 1.0, + "content": "Nonetheless, we run experiments on the PTB dataset using MC-eval (the results are summarized", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "in Table 8). We start with a simple comparison that compares fraternal dropout with the averaged", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 536, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 548 + ], + "score": 1.0, + "content": "mask and the AWD-LSTM 3-layer baseline with a single fixed mask that we call MC1. The MC1", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 547, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 505, + 559 + ], + "score": 1.0, + "content": "model performs much worse than fraternal dropout. Hence, it would be hard to use MC1 model", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 558, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 570 + ], + "score": 1.0, + "content": "in practice because a single sample is inaccurate. We also check MC-eval for a larger number of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "models (MC50) (50 models were used since we were not able to fit more models simultaneously on", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 580, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 104, + 580, + 505, + 592 + ], + "score": 1.0, + "content": "a single GPU). The final results for MC50 are worse than the baseline which uses the averaged mask.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "For comparison, we also evaluate MC10. Note that no fine-tuning is used for the above experiments.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5, + "bbox_fs": [ + 104, + 514, + 506, + 603 + ] + }, + { + "type": "table", + "bbox": [ + 189, + 610, + 423, + 690 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 189, + 610, + 423, + 690 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 189, + 610, + 423, + 690 + ], + "spans": [ + { + "bbox": [ + 189, + 610, + 423, + 690 + ], + "score": 0.967, + "html": "
ModelValidationTest
MC1 MC10 MC50 Baseline (average mask) Fraternal dropout92.2 (± 0.5) 66.2 (± 0.2) 64.4 (± 0.1) 60.7 59.889.2 (± 0.5) 63.7 (± 0.2) 62.1 (± 0.1) 58.8 58.0
", + "type": "table", + "image_path": "910137973af0702bba6415fa366c5234d0c42e20a24f609d5bd82422af5dbf96.jpg" + } + ] + } + ], + "index": 43.5, + "virtual_lines": [ + { + "bbox": [ + 189, + 610, + 423, + 623.3333333333334 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 189, + 623.3333333333334, + 423, + 636.6666666666667 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 189, + 636.6666666666667, + 423, + 650.0000000000001 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 189, + 650.0000000000001, + 423, + 663.3333333333335 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 189, + 663.3333333333335, + 423, + 676.6666666666669 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 189, + 676.6666666666669, + 423, + 690.0000000000002 + ], + "spans": [], + "index": 46 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 693, + 504, + 716 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 690, + 505, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 505, + 707 + ], + "score": 1.0, + "content": "Table 8: Appendix: Monte Carlo evaluation. Perplexity on Penn Treebank word level language", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 704, + 425, + 717 + ], + "spans": [ + { + "bbox": [ + 106, + 704, + 425, + 717 + ], + "score": 1.0, + "content": "modeling task using Monte Carlo sampling, fraternal dropout or average mask.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5 + } + ], + "index": 45.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 259, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 260, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 260, + 95 + ], + "score": 1.0, + "content": "REASONS FOR FOCUSING ON RNNS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 102, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 505, + 116 + ], + "score": 1.0, + "content": "The fraternal dropout method is general and may be applied in feed-forward architectures (as shown", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "in Subsection 5.2 for CIFAR-10 semisupervised example). However, we believe that it is more", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 125, + 264, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 264, + 137 + ], + "score": 1.0, + "content": "powerful in the case of RNNs because:", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 129, + 146, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 129, + 146, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 129, + 146, + 505, + 158 + ], + "score": 1.0, + "content": "1. Variance in prediction accumulates among time steps in RNNs and since we share parame-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 157, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 141, + 157, + 318, + 169 + ], + "score": 1.0, + "content": "ters for all time steps, one may use the same", + "type": "text" + }, + { + "bbox": [ + 318, + 159, + 326, + 167 + ], + "score": 0.69, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 157, + 505, + 169 + ], + "score": 1.0, + "content": "value at each step. In feed-forward networks", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 168, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 141, + 168, + 469, + 181 + ], + "score": 1.0, + "content": "the layers usually do not share parameters and hence one may want to use different", + "type": "text" + }, + { + "bbox": [ + 470, + 170, + 477, + 178 + ], + "score": 0.75, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 168, + 506, + 181 + ], + "score": 1.0, + "content": "values", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 178, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 141, + 178, + 505, + 191 + ], + "score": 1.0, + "content": "for different layers (which may be hard to tune). The simple way to alleviate this problem", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 190, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 141, + 190, + 505, + 202 + ], + "score": 1.0, + "content": "is to apply the regularization term on the pre-softmax predictions only (as shown in the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 140, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 140, + 201, + 233, + 213 + ], + "score": 1.0, + "content": "paper) or use the same", + "type": "text" + }, + { + "bbox": [ + 233, + 203, + 240, + 210 + ], + "score": 0.63, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "value for all layers. However, we believe that it may limit possible", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 140, + 213, + 168, + 224 + ], + "spans": [ + { + "bbox": [ + 140, + 213, + 168, + 224 + ], + "score": 1.0, + "content": "gains.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 226, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 506, + 239 + ], + "score": 1.0, + "content": "2. The best performing RNN architectures (state-of-the-art) usually use some kind of dropout", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 238, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 142, + 238, + 505, + 250 + ], + "score": 1.0, + "content": "(embedding dropout, word dropout, weight dropout etc.), very often with high dropout", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 249, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 141, + 249, + 233, + 261 + ], + "score": 1.0, + "content": "rates (even larger than", + "type": "text" + }, + { + "bbox": [ + 234, + 249, + 253, + 259 + ], + "score": 0.85, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 249, + 505, + 261 + ], + "score": 1.0, + "content": "for input word embedding in NLP tasks). However, this is not", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 259, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 141, + 259, + 506, + 272 + ], + "score": 1.0, + "content": "true for feed-forward networks. For instance, ResNet architectures very often do not use", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 271, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 141, + 271, + 505, + 282 + ], + "score": 1.0, + "content": "dropout at all (probably because batch normalization is often better to use). It can be seen in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 141, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "the paper (Subsection 5.2, semisupervised CIFAR-10 task) that when unlabeled data is not", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 293, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 141, + 293, + 505, + 305 + ], + "score": 1.0, + "content": "used the regular dropout hurts performance and using fraternal dropout seems to improve", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 304, + 190, + 315 + ], + "spans": [ + { + "bbox": [ + 141, + 304, + 190, + 315 + ], + "score": 1.0, + "content": "just a little.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 128, + 317, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 128, + 317, + 505, + 331 + ], + "score": 1.0, + "content": "3. On a final note, the Monte Carlo sampling (a well known method that adresses the gap", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 329, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 141, + 329, + 505, + 341 + ], + "score": 1.0, + "content": "betweem the train and evaluation mode of dropout) can not be easily applied for RNNs and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 341, + 335, + 352 + ], + "spans": [ + { + "bbox": [ + 141, + 341, + 335, + 352 + ], + "score": 1.0, + "content": "fraternal dropout may be seen as an alternative.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 361, + 504, + 383 + ], + "lines": [ + { + "bbox": [ + 105, + 359, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 505, + 375 + ], + "score": 1.0, + "content": "To conclude, we believe that when the use of dropout benefits in a given architecture, applying", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 372, + 342, + 385 + ], + "spans": [ + { + "bbox": [ + 104, + 372, + 342, + 385 + ], + "score": 1.0, + "content": "fraternal dropout should improve performance even more.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 389, + 505, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "As mentioned before, in image recognition tasks, one may experiment with something what we", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "score": 1.0, + "content": "would temporarily dub fraternal augmentation (even though dropout is not used, one can use random", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "data augmentation such as random crop or random flip). Hence, one may force a given neural", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 422, + 370, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 370, + 434 + ], + "score": 1.0, + "content": "network to have the same predictions for different augmentations.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 259, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 260, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 260, + 95 + ], + "score": 1.0, + "content": "REASONS FOR FOCUSING ON RNNS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 102, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 505, + 116 + ], + "score": 1.0, + "content": "The fraternal dropout method is general and may be applied in feed-forward architectures (as shown", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "in Subsection 5.2 for CIFAR-10 semisupervised example). However, we believe that it is more", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 125, + 264, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 264, + 137 + ], + "score": 1.0, + "content": "powerful in the case of RNNs because:", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 102, + 505, + 137 + ] + }, + { + "type": "list", + "bbox": [ + 129, + 146, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 129, + 146, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 129, + 146, + 505, + 158 + ], + "score": 1.0, + "content": "1. Variance in prediction accumulates among time steps in RNNs and since we share parame-", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 157, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 141, + 157, + 318, + 169 + ], + "score": 1.0, + "content": "ters for all time steps, one may use the same", + "type": "text" + }, + { + "bbox": [ + 318, + 159, + 326, + 167 + ], + "score": 0.69, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 157, + 505, + 169 + ], + "score": 1.0, + "content": "value at each step. In feed-forward networks", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 168, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 141, + 168, + 469, + 181 + ], + "score": 1.0, + "content": "the layers usually do not share parameters and hence one may want to use different", + "type": "text" + }, + { + "bbox": [ + 470, + 170, + 477, + 178 + ], + "score": 0.75, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 168, + 506, + 181 + ], + "score": 1.0, + "content": "values", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 178, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 141, + 178, + 505, + 191 + ], + "score": 1.0, + "content": "for different layers (which may be hard to tune). The simple way to alleviate this problem", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 190, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 141, + 190, + 505, + 202 + ], + "score": 1.0, + "content": "is to apply the regularization term on the pre-softmax predictions only (as shown in the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 140, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 140, + 201, + 233, + 213 + ], + "score": 1.0, + "content": "paper) or use the same", + "type": "text" + }, + { + "bbox": [ + 233, + 203, + 240, + 210 + ], + "score": 0.63, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "value for all layers. However, we believe that it may limit possible", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 140, + 213, + 168, + 224 + ], + "spans": [ + { + "bbox": [ + 140, + 213, + 168, + 224 + ], + "score": 1.0, + "content": "gains.", + "type": "text" + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 130, + 226, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 506, + 239 + ], + "score": 1.0, + "content": "2. The best performing RNN architectures (state-of-the-art) usually use some kind of dropout", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 238, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 142, + 238, + 505, + 250 + ], + "score": 1.0, + "content": "(embedding dropout, word dropout, weight dropout etc.), very often with high dropout", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 249, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 141, + 249, + 233, + 261 + ], + "score": 1.0, + "content": "rates (even larger than", + "type": "text" + }, + { + "bbox": [ + 234, + 249, + 253, + 259 + ], + "score": 0.85, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 249, + 505, + 261 + ], + "score": 1.0, + "content": "for input word embedding in NLP tasks). However, this is not", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 259, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 141, + 259, + 506, + 272 + ], + "score": 1.0, + "content": "true for feed-forward networks. For instance, ResNet architectures very often do not use", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 271, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 141, + 271, + 505, + 282 + ], + "score": 1.0, + "content": "dropout at all (probably because batch normalization is often better to use). It can be seen in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 141, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "the paper (Subsection 5.2, semisupervised CIFAR-10 task) that when unlabeled data is not", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 293, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 141, + 293, + 505, + 305 + ], + "score": 1.0, + "content": "used the regular dropout hurts performance and using fraternal dropout seems to improve", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 304, + 190, + 315 + ], + "spans": [ + { + "bbox": [ + 141, + 304, + 190, + 315 + ], + "score": 1.0, + "content": "just a little.", + "type": "text" + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 128, + 317, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 128, + 317, + 505, + 331 + ], + "score": 1.0, + "content": "3. On a final note, the Monte Carlo sampling (a well known method that adresses the gap", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 329, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 141, + 329, + 505, + 341 + ], + "score": 1.0, + "content": "betweem the train and evaluation mode of dropout) can not be easily applied for RNNs and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 341, + 335, + 352 + ], + "spans": [ + { + "bbox": [ + 141, + 341, + 335, + 352 + ], + "score": 1.0, + "content": "fraternal dropout may be seen as an alternative.", + "type": "text" + } + ], + "index": 21, + "is_list_end_line": true + } + ], + "index": 12.5, + "bbox_fs": [ + 128, + 146, + 506, + 352 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 361, + 504, + 383 + ], + "lines": [ + { + "bbox": [ + 105, + 359, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 505, + 375 + ], + "score": 1.0, + "content": "To conclude, we believe that when the use of dropout benefits in a given architecture, applying", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 372, + 342, + 385 + ], + "spans": [ + { + "bbox": [ + 104, + 372, + 342, + 385 + ], + "score": 1.0, + "content": "fraternal dropout should improve performance even more.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 104, + 359, + 505, + 385 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 389, + 505, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "As mentioned before, in image recognition tasks, one may experiment with something what we", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "score": 1.0, + "content": "would temporarily dub fraternal augmentation (even though dropout is not used, one can use random", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "data augmentation such as random crop or random flip). Hence, one may force a given neural", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 422, + 370, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 370, + 434 + ], + "score": 1.0, + "content": "network to have the same predictions for different augmentations.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 388, + 506, + 434 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 142, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 143, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 143, + 95 + ], + "score": 1.0, + "content": "PROOFS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 104, + 102, + 505, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 101, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 171, + 117 + ], + "score": 1.0, + "content": "Remark 1. Let", + "type": "text" + }, + { + "bbox": [ + 172, + 103, + 181, + 115 + ], + "score": 0.88, + "content": "s _ { i } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 101, + 200, + 117 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 200, + 103, + 209, + 116 + ], + "score": 0.87, + "content": "\\boldsymbol { s } _ { j } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 101, + 322, + 117 + ], + "score": 1.0, + "content": "be i.i.d. dropout masks and", + "type": "text" + }, + { + "bbox": [ + 322, + 102, + 399, + 115 + ], + "score": 0.93, + "content": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) \\in \\mathbb { R } ^ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 101, + 506, + 117 + ], + "score": 1.0, + "content": "be the prediction function", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 113, + 213, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 213, + 127 + ], + "score": 1.0, + "content": "as described above. Then,", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 129, + 478, + 163 + ], + "lines": [ + { + "bbox": [ + 142, + 129, + 478, + 163 + ], + "spans": [ + { + "bbox": [ + 142, + 129, + 478, + 163 + ], + "score": 0.92, + "content": "\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\theta ) = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] = 2 \\sum _ { q = 1 } ^ { m } \\mathrm { v a r } _ { s _ { i } ^ { t } } ( p _ { q } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) ) .", + "type": "interline_equation", + "image_path": "19b899377283a2897e4e2f965b3f5c99c16188c1df7a49acc53da71265300747.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 142, + 129, + 478, + 140.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 142, + 140.33333333333334, + 478, + 151.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 142, + 151.66666666666669, + 478, + 163.00000000000003 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 174, + 345, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 173, + 342, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 337, + 187 + ], + "score": 1.0, + "content": "Proof. For simplicity of notation, we omit the time index", + "type": "text" + }, + { + "bbox": [ + 338, + 176, + 342, + 184 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 190, + 449, + 377 + ], + "lines": [ + { + "bbox": [ + 142, + 190, + 449, + 377 + ], + "spans": [ + { + "bbox": [ + 142, + 190, + 449, + 377 + ], + "score": 0.95, + "content": "\\begin{array} { r l } { \\mathcal { R } _ { F S } [ \\mathbf { z } , \\boldsymbol { \\theta } ] } & { = \\mathrm { E } _ { \\boldsymbol { \\phi } , \\boldsymbol { \\xi } } [ \\| \\mathbf { P } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } , \\boldsymbol { \\theta } ) - \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } \\| ^ { 2 } ] } \\\\ & { = \\mathrm { E } _ { \\boldsymbol { \\phi } , \\boldsymbol { \\xi } } \\Bigg [ \\| \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\| _ { \\mathbf { z } } ^ { 2 } + \\mathrm { E } _ { \\boldsymbol { \\xi } , \\boldsymbol { \\xi } } \\Bigg [ \\| \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\| _ { \\mathbf { z } } ^ { 2 } \\Bigg ] } \\\\ & { - 2 \\mathrm { E } _ { \\boldsymbol { \\xi } , \\boldsymbol { s } , \\boldsymbol { \\xi } } [ \\| \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } \\| ^ { 2 } \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } ] } \\\\ & { = 2 \\sum _ { \\boldsymbol { \\phi } = 1 } ^ { \\infty } ( \\mathrm { E } _ { \\boldsymbol { \\phi } } [ \\partial _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } ] ^ { 2 } ) - \\mathrm { E } _ { \\boldsymbol { \\phi } , \\boldsymbol { \\xi } } \\Bigg [ \\mathcal { P } _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } ] _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ; \\boldsymbol { \\theta } ) \\Bigg ] } \\\\ & = 2 \\sum _ { \\boldsymbol { \\phi } = 1 } ^ { \\infty } ( \\mathrm { E } _ { \\boldsymbol { \\phi } } [ \\partial _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } , \\boldsymbol { \\theta } ) ] - \\mathrm { E } _ { \\boldsymbol { \\phi } , \\boldsymbol { \\xi } } [ \\mathcal { P } _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) ] \\mathrm { E } _ { \\boldsymbol { \\phi } } [ \\mathcal { P } _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol \\ \\end{array}", + "type": "interline_equation", + "image_path": "ee9b4078e855855dd4a7060bf6140cdc2c1570eb2de4814747098ead92f77b38.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 142, + 190, + 449, + 252.33333333333334 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 142, + 252.33333333333334, + 449, + 314.6666666666667 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 142, + 314.6666666666667, + 449, + 377.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 397, + 293, + 411 + ], + "lines": [ + { + "bbox": [ + 106, + 397, + 294, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 170, + 412 + ], + "score": 1.0, + "content": "Proposition 1.", + "type": "text" + }, + { + "bbox": [ + 171, + 397, + 291, + 411 + ], + "score": 0.91, + "content": "\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) \\leq 4 \\tilde { \\mathcal { R } } _ { E L D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) .", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 397, + 294, + 412 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 421, + 218, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 422, + 218, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 152, + 435 + ], + "score": 1.0, + "content": "Proof. Let", + "type": "text" + }, + { + "bbox": [ + 152, + 422, + 195, + 434 + ], + "score": 0.92, + "content": "\\bar { s } : = \\mathbb { E } _ { s } [ s ]", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 422, + 218, + 435 + ], + "score": 1.0, + "content": ", then", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "interline_equation", + "bbox": [ + 141, + 438, + 475, + 505 + ], + "lines": [ + { + "bbox": [ + 141, + 438, + 475, + 505 + ], + "spans": [ + { + "bbox": [ + 141, + 438, + 475, + 505 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\mathcal { R } _ { t } ( \\mathbf { z } ^ { t } ) : = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] } \\\\ & { \\quad \\quad \\quad \\quad = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) + \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] } \\\\ & { \\quad \\quad \\quad \\quad = 4 \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\frac { \\mathbf { p } ^ { t } \\left( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta \\right) - \\mathbf { p } ^ { t } \\left( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta \\right) } { 2 } + \\frac { \\mathbf { p } ^ { t } \\left( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta \\right) - \\mathbf { p } ^ { t } \\left( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta \\right) } { 2 } \\| _ { 2 } ^ { 2 } \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "c56e7b117430a0f08c020211f1d54d6de45a191dab907f9d4cb05bab8c3cfa93.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 141, + 438, + 475, + 460.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 141, + 460.3333333333333, + 475, + 482.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 141, + 482.66666666666663, + 475, + 504.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 508, + 233, + 520 + ], + "lines": [ + { + "bbox": [ + 105, + 506, + 234, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 234, + 523 + ], + "score": 1.0, + "content": "Then using Jensen’s inequality,", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 141, + 525, + 494, + 583 + ], + "lines": [ + { + "bbox": [ + 141, + 525, + 494, + 583 + ], + "spans": [ + { + "bbox": [ + 141, + 525, + 494, + 583 + ], + "score": 0.73, + "content": "\\begin{array} { r l } & { \\mathcal { R } _ { t } ( \\mathbf { z } ^ { t } ) \\leq 4 \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\frac { 1 } { 2 } \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) \\| _ { 2 } ^ { 2 } + \\frac { 1 } { 2 } \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] } \\\\ & { \\qquad ( 2 \\pi ) } \\\\ & { \\qquad = 4 \\mathbb { E } _ { s _ { i } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] = 4 \\tilde { \\mathcal { R } } _ { E L D } ( \\mathbf { z } ^ { t } ; \\theta ) . } \\end{array}", + "type": "interline_equation", + "image_path": "24b9c68592299ce9300b633f9611f2195b71342c78020834a7e086496852e274.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 141, + 525, + 494, + 544.3333333333334 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 141, + 544.3333333333334, + 494, + 563.6666666666667 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 141, + 563.6666666666667, + 494, + 583.0000000000001 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 587, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 495, + 588, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 495, + 588, + 505, + 599 + ], + "score": 0.998, + "content": "□", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 380, + 505, + 391 + ], + "lines": [ + { + "bbox": [ + 495, + 382, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 495, + 382, + 505, + 391 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 142, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 143, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 143, + 95 + ], + "score": 1.0, + "content": "PROOFS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 106, + 82, + 143, + 95 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 102, + 505, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 101, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 171, + 117 + ], + "score": 1.0, + "content": "Remark 1. Let", + "type": "text" + }, + { + "bbox": [ + 172, + 103, + 181, + 115 + ], + "score": 0.88, + "content": "s _ { i } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 101, + 200, + 117 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 200, + 103, + 209, + 116 + ], + "score": 0.87, + "content": "\\boldsymbol { s } _ { j } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 101, + 322, + 117 + ], + "score": 1.0, + "content": "be i.i.d. dropout masks and", + "type": "text" + }, + { + "bbox": [ + 322, + 102, + 399, + 115 + ], + "score": 0.93, + "content": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) \\in \\mathbb { R } ^ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 101, + 506, + 117 + ], + "score": 1.0, + "content": "be the prediction function", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 113, + 213, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 213, + 127 + ], + "score": 1.0, + "content": "as described above. Then,", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 101, + 506, + 127 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 129, + 478, + 163 + ], + "lines": [ + { + "bbox": [ + 142, + 129, + 478, + 163 + ], + "spans": [ + { + "bbox": [ + 142, + 129, + 478, + 163 + ], + "score": 0.92, + "content": "\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\theta ) = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] = 2 \\sum _ { q = 1 } ^ { m } \\mathrm { v a r } _ { s _ { i } ^ { t } } ( p _ { q } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) ) .", + "type": "interline_equation", + "image_path": "19b899377283a2897e4e2f965b3f5c99c16188c1df7a49acc53da71265300747.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 142, + 129, + 478, + 140.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 142, + 140.33333333333334, + 478, + 151.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 142, + 151.66666666666669, + 478, + 163.00000000000003 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 174, + 345, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 173, + 342, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 337, + 187 + ], + "score": 1.0, + "content": "Proof. For simplicity of notation, we omit the time index", + "type": "text" + }, + { + "bbox": [ + 338, + 176, + 342, + 184 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + } + ], + "index": 6 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 173, + 342, + 187 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 190, + 449, + 377 + ], + "lines": [ + { + "bbox": [ + 142, + 190, + 449, + 377 + ], + "spans": [ + { + "bbox": [ + 142, + 190, + 449, + 377 + ], + "score": 0.95, + "content": "\\begin{array} { r l } { \\mathcal { R } _ { F S } [ \\mathbf { z } , \\boldsymbol { \\theta } ] } & { = \\mathrm { E } _ { \\boldsymbol { \\phi } , \\boldsymbol { \\xi } } [ \\| \\mathbf { P } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } , \\boldsymbol { \\theta } ) - \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } \\| ^ { 2 } ] } \\\\ & { = \\mathrm { E } _ { \\boldsymbol { \\phi } , \\boldsymbol { \\xi } } \\Bigg [ \\| \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\| _ { \\mathbf { z } } ^ { 2 } + \\mathrm { E } _ { \\boldsymbol { \\xi } , \\boldsymbol { \\xi } } \\Bigg [ \\| \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\| _ { \\mathbf { z } } ^ { 2 } \\Bigg ] } \\\\ & { - 2 \\mathrm { E } _ { \\boldsymbol { \\xi } , \\boldsymbol { s } , \\boldsymbol { \\xi } } [ \\| \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } \\| ^ { 2 } \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } ] } \\\\ & { = 2 \\sum _ { \\boldsymbol { \\phi } = 1 } ^ { \\infty } ( \\mathrm { E } _ { \\boldsymbol { \\phi } } [ \\partial _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } ] ^ { 2 } ) - \\mathrm { E } _ { \\boldsymbol { \\phi } , \\boldsymbol { \\xi } } \\Bigg [ \\mathcal { P } _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } ] _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ; \\boldsymbol { \\theta } ) \\Bigg ] } \\\\ & = 2 \\sum _ { \\boldsymbol { \\phi } = 1 } ^ { \\infty } ( \\mathrm { E } _ { \\boldsymbol { \\phi } } [ \\partial _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } , \\boldsymbol { \\theta } ) ] - \\mathrm { E } _ { \\boldsymbol { \\phi } , \\boldsymbol { \\xi } } [ \\mathcal { P } _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) ] \\mathrm { E } _ { \\boldsymbol { \\phi } } [ \\mathcal { P } _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol \\ \\end{array}", + "type": "interline_equation", + "image_path": "ee9b4078e855855dd4a7060bf6140cdc2c1570eb2de4814747098ead92f77b38.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 142, + 190, + 449, + 252.33333333333334 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 142, + 252.33333333333334, + 449, + 314.6666666666667 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 142, + 314.6666666666667, + 449, + 377.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 397, + 293, + 411 + ], + "lines": [ + { + "bbox": [ + 106, + 397, + 294, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 170, + 412 + ], + "score": 1.0, + "content": "Proposition 1.", + "type": "text" + }, + { + "bbox": [ + 171, + 397, + 291, + 411 + ], + "score": 0.91, + "content": "\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) \\leq 4 \\tilde { \\mathcal { R } } _ { E L D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) .", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 397, + 294, + 412 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 106, + 397, + 294, + 412 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 421, + 218, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 422, + 218, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 152, + 435 + ], + "score": 1.0, + "content": "Proof. Let", + "type": "text" + }, + { + "bbox": [ + 152, + 422, + 195, + 434 + ], + "score": 0.92, + "content": "\\bar { s } : = \\mathbb { E } _ { s } [ s ]", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 422, + 218, + 435 + ], + "score": 1.0, + "content": ", then", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 422, + 218, + 435 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 141, + 438, + 475, + 505 + ], + "lines": [ + { + "bbox": [ + 141, + 438, + 475, + 505 + ], + "spans": [ + { + "bbox": [ + 141, + 438, + 475, + 505 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\mathcal { R } _ { t } ( \\mathbf { z } ^ { t } ) : = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] } \\\\ & { \\quad \\quad \\quad \\quad = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) + \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] } \\\\ & { \\quad \\quad \\quad \\quad = 4 \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\frac { \\mathbf { p } ^ { t } \\left( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta \\right) - \\mathbf { p } ^ { t } \\left( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta \\right) } { 2 } + \\frac { \\mathbf { p } ^ { t } \\left( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta \\right) - \\mathbf { p } ^ { t } \\left( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta \\right) } { 2 } \\| _ { 2 } ^ { 2 } \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "c56e7b117430a0f08c020211f1d54d6de45a191dab907f9d4cb05bab8c3cfa93.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 141, + 438, + 475, + 460.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 141, + 460.3333333333333, + 475, + 482.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 141, + 482.66666666666663, + 475, + 504.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 508, + 233, + 520 + ], + "lines": [ + { + "bbox": [ + 105, + 506, + 234, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 234, + 523 + ], + "score": 1.0, + "content": "Then using Jensen’s inequality,", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 506, + 234, + 523 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 141, + 525, + 494, + 583 + ], + "lines": [ + { + "bbox": [ + 141, + 525, + 494, + 583 + ], + "spans": [ + { + "bbox": [ + 141, + 525, + 494, + 583 + ], + "score": 0.73, + "content": "\\begin{array} { r l } & { \\mathcal { R } _ { t } ( \\mathbf { z } ^ { t } ) \\leq 4 \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\frac { 1 } { 2 } \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) \\| _ { 2 } ^ { 2 } + \\frac { 1 } { 2 } \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] } \\\\ & { \\qquad ( 2 \\pi ) } \\\\ & { \\qquad = 4 \\mathbb { E } _ { s _ { i } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] = 4 \\tilde { \\mathcal { R } } _ { E L D } ( \\mathbf { z } ^ { t } ; \\theta ) . } \\end{array}", + "type": "interline_equation", + "image_path": "24b9c68592299ce9300b633f9611f2195b71342c78020834a7e086496852e274.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 141, + 525, + 494, + 544.3333333333334 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 141, + 544.3333333333334, + 494, + 563.6666666666667 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 141, + 563.6666666666667, + 494, + 583.0000000000001 + ], + "spans": [], + "index": 18 + } + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/SJyVzQ-C-/SJyVzQ-C-_model.json b/parse/train/SJyVzQ-C-/SJyVzQ-C-_model.json new file mode 100644 index 0000000000000000000000000000000000000000..25b4bc87575f3c72ba32593f1af00c88f8462b31 --- /dev/null +++ b/parse/train/SJyVzQ-C-/SJyVzQ-C-_model.json @@ -0,0 +1,17579 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 635, + 1302, + 635, + 1302, + 1153, + 398, + 1153 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1308, + 1404, + 1308, + 1404, + 1674, + 298, + 1674 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1690, + 1403, + 1690, + 1403, + 1963, + 298, + 1963 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 313, + 382, + 671, + 382, + 671, + 476, + 313, + 476 + ], + "score": 0.955 + }, + { + "category_id": 0, + "poly": [ + 302, + 220, + 758, + 220, + 758, + 269, + 302, + 269 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 333, + 2008, + 677, + 2008, + 677, + 2033, + 333, + 2033 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 302, + 1231, + 573, + 1231, + 573, + 1266, + 302, + 1266 + ], + "score": 0.894 + }, + { + "category_id": 1, + "poly": [ + 319, + 318, + 1181, + 318, + 1181, + 353, + 319, + 353 + ], + "score": 0.871 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 104, + 299, + 104 + ], + "score": 0.869 + }, + { + "category_id": 0, + "poly": [ + 773, + 559, + 926, + 559, + 926, + 592, + 773, + 592 + ], + "score": 0.858 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 857, + 2088, + 857, + 2112, + 842, + 2112 + ], + "score": 0.703 + }, + { + "category_id": 13, + "poly": [ + 419, + 316, + 538, + 316, + 538, + 352, + 419, + 352 + ], + "score": 0.33, + "latex": "\\dot { \\mathbf Z } \\mathbf { o } \\mathbf { f n } \\mathbf { a } ^ { 1 , 2 , * } ;" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 218.0, + 765.0, + 218.0, + 765.0, + 273.0, + 292.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 2001.0, + 679.0, + 2001.0, + 679.0, + 2037.0, + 331.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1228.0, + 579.0, + 1228.0, + 579.0, + 1275.0, + 294.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 558.0, + 932.0, + 558.0, + 932.0, + 595.0, + 769.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 858.0, + 2088.0, + 858.0, + 2116.0, + 841.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 636.0, + 1305.0, + 636.0, + 1305.0, + 670.0, + 395.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 668.0, + 1305.0, + 668.0, + 1305.0, + 701.0, + 394.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 698.0, + 1305.0, + 698.0, + 1305.0, + 731.0, + 395.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 729.0, + 1305.0, + 729.0, + 1305.0, + 762.0, + 394.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 758.0, + 1307.0, + 758.0, + 1307.0, + 794.0, + 393.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 787.0, + 1307.0, + 787.0, + 1307.0, + 825.0, + 392.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 818.0, + 1305.0, + 818.0, + 1305.0, + 853.0, + 394.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 848.0, + 1305.0, + 848.0, + 1305.0, + 883.0, + 394.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 879.0, + 1304.0, + 879.0, + 1304.0, + 915.0, + 394.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 910.0, + 1305.0, + 910.0, + 1305.0, + 943.0, + 394.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 940.0, + 1307.0, + 940.0, + 1307.0, + 974.0, + 393.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 970.0, + 1307.0, + 970.0, + 1307.0, + 1007.0, + 393.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1002.0, + 1306.0, + 1002.0, + 1306.0, + 1036.0, + 395.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1033.0, + 1305.0, + 1033.0, + 1305.0, + 1064.0, + 396.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1064.0, + 1306.0, + 1064.0, + 1306.0, + 1097.0, + 395.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1094.0, + 1304.0, + 1094.0, + 1304.0, + 1127.0, + 393.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1125.0, + 465.0, + 1125.0, + 465.0, + 1154.0, + 395.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1309.0, + 1404.0, + 1309.0, + 1404.0, + 1343.0, + 294.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1339.0, + 1405.0, + 1339.0, + 1405.0, + 1373.0, + 294.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1371.0, + 1404.0, + 1371.0, + 1404.0, + 1405.0, + 294.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1400.0, + 1405.0, + 1400.0, + 1405.0, + 1434.0, + 294.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1429.0, + 1408.0, + 1429.0, + 1408.0, + 1467.0, + 293.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1462.0, + 1406.0, + 1462.0, + 1406.0, + 1496.0, + 294.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1490.0, + 1405.0, + 1490.0, + 1405.0, + 1526.0, + 293.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1523.0, + 1406.0, + 1523.0, + 1406.0, + 1557.0, + 294.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1550.0, + 1406.0, + 1550.0, + 1406.0, + 1588.0, + 293.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1583.0, + 1404.0, + 1583.0, + 1404.0, + 1617.0, + 294.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1611.0, + 1404.0, + 1611.0, + 1404.0, + 1648.0, + 293.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1644.0, + 856.0, + 1644.0, + 856.0, + 1678.0, + 294.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1692.0, + 1404.0, + 1692.0, + 1404.0, + 1724.0, + 297.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1719.0, + 1405.0, + 1719.0, + 1405.0, + 1755.0, + 293.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1750.0, + 1405.0, + 1750.0, + 1405.0, + 1785.0, + 293.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1780.0, + 1408.0, + 1780.0, + 1408.0, + 1817.0, + 292.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1808.0, + 1407.0, + 1808.0, + 1407.0, + 1850.0, + 292.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1843.0, + 1404.0, + 1843.0, + 1404.0, + 1876.0, + 296.0, + 1876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1873.0, + 1405.0, + 1873.0, + 1405.0, + 1906.0, + 296.0, + 1906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1900.0, + 1407.0, + 1900.0, + 1407.0, + 1942.0, + 291.0, + 1942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1933.0, + 578.0, + 1933.0, + 578.0, + 1969.0, + 294.0, + 1969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 379.0, + 589.0, + 379.0, + 589.0, + 418.0, + 312.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 409.0, + 675.0, + 409.0, + 675.0, + 451.0, + 310.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 441.0, + 576.0, + 441.0, + 576.0, + 481.0, + 310.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 313.0, + 418.0, + 313.0, + 418.0, + 360.0, + 318.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 313.0, + 1190.0, + 313.0, + 1190.0, + 360.0, + 539.0, + 360.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1070, + 1403, + 1070, + 1403, + 1290, + 298, + 1290 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 399, + 1404, + 399, + 1404, + 674, + 298, + 674 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 809, + 1404, + 809, + 1404, + 1055, + 297, + 1055 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 1302, + 1403, + 1302, + 1403, + 1434, + 299, + 1434 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1402, + 229, + 1402, + 383, + 298, + 383 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1782, + 1403, + 1782, + 1403, + 1852, + 299, + 1852 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 393, + 1466, + 1395, + 1466, + 1395, + 1557, + 393, + 1557 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 300, + 1866, + 1404, + 1866, + 1404, + 1930, + 300, + 1930 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 395, + 1706, + 1029, + 1706, + 1029, + 1755, + 395, + 1755 + ], + "score": 0.94 + }, + { + "category_id": 2, + "poly": [ + 294, + 1977, + 1406, + 1977, + 1406, + 2035, + 294, + 2035 + ], + "score": 0.94 + }, + { + "category_id": 0, + "poly": [ + 301, + 732, + 666, + 732, + 666, + 767, + 301, + 767 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1715, + 1400, + 1715, + 1400, + 1744, + 1366, + 1744 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 104, + 300, + 104 + ], + "score": 0.889 + }, + { + "category_id": 1, + "poly": [ + 292, + 1618, + 1406, + 1618, + 1406, + 1684, + 292, + 1684 + ], + "score": 0.87 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1557, + 1400, + 1557, + 1400, + 1587, + 1366, + 1587 + ], + "score": 0.856 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.757 + }, + { + "category_id": 13, + "poly": [ + 1031, + 1783, + 1169, + 1783, + 1169, + 1818, + 1031, + 1818 + ], + "score": 0.95, + "latex": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } )" + }, + { + "category_id": 13, + "poly": [ + 1200, + 1619, + 1334, + 1619, + 1334, + 1654, + 1200, + 1654 + ], + "score": 0.93, + "latex": "\\mathcal { R } _ { F D } ( { \\bf z } ^ { t } ; \\boldsymbol { \\theta } )" + }, + { + "category_id": 13, + "poly": [ + 818, + 1783, + 952, + 1783, + 952, + 1818, + 818, + 1818 + ], + "score": 0.93, + "latex": "\\mathcal { R } _ { F D } ( { \\bf z } ^ { t } ; \\boldsymbol { \\theta } )" + }, + { + "category_id": 13, + "poly": [ + 1012, + 1619, + 1148, + 1619, + 1148, + 1655, + 1012, + 1655 + ], + "score": 0.93, + "latex": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } )" + }, + { + "category_id": 13, + "poly": [ + 1221, + 1783, + 1359, + 1783, + 1359, + 1822, + 1221, + 1822 + ], + "score": 0.93, + "latex": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1399, + 515, + 1399, + 515, + 1438, + 298, + 1438 + ], + "score": 0.93, + "latex": "\\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\bar { \\theta } ) , \\mathbf { Y } )" + }, + { + "category_id": 13, + "poly": [ + 597, + 1161, + 809, + 1161, + 809, + 1196, + 597, + 1196 + ], + "score": 0.93, + "latex": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) \\in \\mathbb { R } ^ { m }" + }, + { + "category_id": 13, + "poly": [ + 1129, + 1368, + 1344, + 1368, + 1344, + 1404, + 1129, + 1404 + ], + "score": 0.92, + "latex": "\\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) , \\mathbf { Y } )" + }, + { + "category_id": 13, + "poly": [ + 887, + 1131, + 958, + 1131, + 958, + 1164, + 887, + 1164 + ], + "score": 0.92, + "latex": "\\mathcal { M } ( \\boldsymbol { \\theta } )" + }, + { + "category_id": 13, + "poly": [ + 907, + 1224, + 1123, + 1224, + 1123, + 1259, + 907, + 1259 + ], + "score": 0.91, + "latex": "\\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) , \\mathbf { Y } )" + }, + { + "category_id": 14, + "poly": [ + 395, + 1461, + 1397, + 1461, + 1397, + 1561, + 395, + 1561 + ], + "score": 0.91, + "latex": "\\ell _ { F D } ( \\mathbf { X } , \\mathbf { Y } ) = \\sum _ { t = 1 } ^ { T } \\frac { 1 } { 2 } \\left( \\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) , \\mathbf { Y } ) + \\ell ^ { t } ( \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) , \\mathbf { Y } ) \\right) + \\frac { \\kappa } { m T } \\sum _ { t = 1 } ^ { T } \\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\theta )" + }, + { + "category_id": 13, + "poly": [ + 957, + 1259, + 1040, + 1259, + 1040, + 1290, + 957, + 1290 + ], + "score": 0.89, + "latex": "( \\mathbf { X } , \\mathbf { Y } )" + }, + { + "category_id": 14, + "poly": [ + 395, + 1706, + 1031, + 1706, + 1031, + 1756, + 395, + 1756 + ], + "score": 0.89, + "latex": "\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) : = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) \\| _ { 2 } ^ { 2 } \\right] ." + }, + { + "category_id": 13, + "poly": [ + 579, + 1194, + 606, + 1194, + 606, + 1226, + 579, + 1226 + ], + "score": 0.88, + "latex": "s _ { i } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 1345, + 1334, + 1372, + 1334, + 1372, + 1372, + 1345, + 1372 + ], + "score": 0.88, + "latex": "\\boldsymbol { s } _ { j } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 1266, + 1334, + 1291, + 1334, + 1291, + 1368, + 1266, + 1368 + ], + "score": 0.87, + "latex": "s _ { i } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 922, + 1192, + 950, + 1192, + 950, + 1220, + 922, + 1220 + ], + "score": 0.87, + "latex": "\\mathbf { z } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 1365, + 1224, + 1399, + 1224, + 1399, + 1254, + 1365, + 1254 + ], + "score": 0.86, + "latex": "t ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 711, + 1819, + 735, + 1819, + 735, + 1847, + 711, + 1847 + ], + "score": 0.85, + "latex": "\\ell ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 808, + 1192, + 835, + 1192, + 835, + 1221, + 808, + 1221 + ], + "score": 0.84, + "latex": "\\mathbf { z } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 1115, + 459, + 1143, + 459, + 1143, + 488, + 1115, + 488 + ], + "score": 0.84, + "latex": "\\ell ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 372, + 1627, + 391, + 1627, + 391, + 1648, + 372, + 1648 + ], + "score": 0.79, + "latex": "\\kappa" + }, + { + "category_id": 13, + "poly": [ + 1290, + 1133, + 1308, + 1133, + 1308, + 1159, + 1290, + 1159 + ], + "score": 0.78, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 751, + 1626, + 779, + 1626, + 779, + 1648, + 751, + 1648 + ], + "score": 0.78, + "latex": "m" + }, + { + "category_id": 13, + "poly": [ + 869, + 1336, + 886, + 1336, + 886, + 1362, + 869, + 1362 + ], + "score": 0.77, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 1004, + 1372, + 1020, + 1372, + 1020, + 1398, + 1004, + 1398 + ], + "score": 0.69, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 353, + 1195, + 368, + 1195, + 368, + 1221, + 353, + 1221 + ], + "score": 0.69, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 466, + 1372, + 481, + 1372, + 481, + 1397, + 466, + 1397 + ], + "score": 0.65, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1175, + 1133, + 1202, + 1133, + 1202, + 1160, + 1175, + 1160 + ], + "score": 0.55, + "latex": "\\mathbf { X }" + }, + { + "category_id": 13, + "poly": [ + 1346, + 1163, + 1374, + 1163, + 1374, + 1191, + 1346, + 1191 + ], + "score": 0.53, + "latex": "\\mathbf { X }" + }, + { + "category_id": 13, + "poly": [ + 1135, + 1304, + 1162, + 1304, + 1162, + 1332, + 1135, + 1332 + ], + "score": 0.37, + "latex": "\\mathbf { X }" + }, + { + "category_id": 13, + "poly": [ + 1132, + 1193, + 1159, + 1193, + 1159, + 1221, + 1132, + 1221 + ], + "score": 0.33, + "latex": "\\mathbf { X }" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2011.0, + 331.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2001.0, + 506.0, + 2001.0, + 506.0, + 2038.0, + 295.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 729.0, + 671.0, + 729.0, + 671.0, + 773.0, + 293.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1071.0, + 1403.0, + 1071.0, + 1403.0, + 1106.0, + 294.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1100.0, + 1404.0, + 1100.0, + 1404.0, + 1137.0, + 292.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1131.0, + 886.0, + 1131.0, + 886.0, + 1166.0, + 294.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 1131.0, + 1174.0, + 1131.0, + 1174.0, + 1166.0, + 959.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 1131.0, + 1289.0, + 1131.0, + 1289.0, + 1166.0, + 1203.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 1131.0, + 1404.0, + 1131.0, + 1404.0, + 1166.0, + 1309.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1160.0, + 596.0, + 1160.0, + 596.0, + 1199.0, + 293.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 1160.0, + 1345.0, + 1160.0, + 1345.0, + 1199.0, + 810.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1375.0, + 1160.0, + 1407.0, + 1160.0, + 1407.0, + 1199.0, + 1375.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1189.0, + 352.0, + 1189.0, + 352.0, + 1228.0, + 293.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1189.0, + 578.0, + 1189.0, + 578.0, + 1228.0, + 369.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 1189.0, + 807.0, + 1189.0, + 807.0, + 1228.0, + 607.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 1189.0, + 921.0, + 1189.0, + 921.0, + 1228.0, + 836.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1189.0, + 1131.0, + 1189.0, + 1131.0, + 1228.0, + 951.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 1189.0, + 1405.0, + 1189.0, + 1405.0, + 1228.0, + 1160.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1224.0, + 906.0, + 1224.0, + 906.0, + 1265.0, + 289.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 1224.0, + 1364.0, + 1224.0, + 1364.0, + 1265.0, + 1124.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1400.0, + 1224.0, + 1406.0, + 1224.0, + 1406.0, + 1265.0, + 1400.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1255.0, + 956.0, + 1255.0, + 956.0, + 1294.0, + 293.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1255.0, + 1053.0, + 1255.0, + 1053.0, + 1294.0, + 1041.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 397.0, + 1405.0, + 397.0, + 1405.0, + 435.0, + 293.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 431.0, + 1404.0, + 431.0, + 1404.0, + 464.0, + 296.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 457.0, + 1114.0, + 457.0, + 1114.0, + 496.0, + 292.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 457.0, + 1406.0, + 457.0, + 1406.0, + 496.0, + 1144.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 488.0, + 1404.0, + 488.0, + 1404.0, + 529.0, + 293.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 521.0, + 1405.0, + 521.0, + 1405.0, + 557.0, + 292.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 551.0, + 1405.0, + 551.0, + 1405.0, + 587.0, + 292.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 578.0, + 1404.0, + 578.0, + 1404.0, + 621.0, + 291.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 611.0, + 1406.0, + 611.0, + 1406.0, + 650.0, + 292.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 642.0, + 1308.0, + 642.0, + 1308.0, + 678.0, + 294.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 811.0, + 1402.0, + 811.0, + 1402.0, + 844.0, + 296.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 842.0, + 1404.0, + 842.0, + 1404.0, + 874.0, + 294.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 872.0, + 1404.0, + 872.0, + 1404.0, + 904.0, + 294.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 903.0, + 1404.0, + 903.0, + 1404.0, + 937.0, + 295.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 933.0, + 1404.0, + 933.0, + 1404.0, + 967.0, + 295.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 962.0, + 1404.0, + 962.0, + 1404.0, + 998.0, + 292.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 994.0, + 1405.0, + 994.0, + 1405.0, + 1028.0, + 295.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1025.0, + 605.0, + 1025.0, + 605.0, + 1058.0, + 294.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1302.0, + 1134.0, + 1302.0, + 1134.0, + 1339.0, + 294.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 1302.0, + 1405.0, + 1302.0, + 1405.0, + 1339.0, + 1163.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1333.0, + 868.0, + 1333.0, + 868.0, + 1370.0, + 294.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1333.0, + 1265.0, + 1333.0, + 1265.0, + 1370.0, + 887.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 1333.0, + 1344.0, + 1333.0, + 1344.0, + 1370.0, + 1292.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 1333.0, + 1407.0, + 1333.0, + 1407.0, + 1370.0, + 1373.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1368.0, + 465.0, + 1368.0, + 465.0, + 1405.0, + 294.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1368.0, + 1003.0, + 1368.0, + 1003.0, + 1405.0, + 482.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 1368.0, + 1128.0, + 1368.0, + 1128.0, + 1405.0, + 1021.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 1368.0, + 1405.0, + 1368.0, + 1405.0, + 1405.0, + 1345.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1397.0, + 297.0, + 1397.0, + 297.0, + 1438.0, + 293.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 1397.0, + 1227.0, + 1397.0, + 1227.0, + 1438.0, + 516.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 1406.0, + 229.0, + 1406.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 258.0, + 1406.0, + 258.0, + 1406.0, + 298.0, + 293.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 289.0, + 1404.0, + 289.0, + 1404.0, + 326.0, + 293.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 317.0, + 1403.0, + 317.0, + 1403.0, + 360.0, + 292.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 351.0, + 641.0, + 351.0, + 641.0, + 388.0, + 293.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1778.0, + 817.0, + 1778.0, + 817.0, + 1822.0, + 295.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1778.0, + 1030.0, + 1778.0, + 1030.0, + 1822.0, + 953.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 1778.0, + 1220.0, + 1778.0, + 1220.0, + 1822.0, + 1170.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1360.0, + 1778.0, + 1407.0, + 1778.0, + 1407.0, + 1822.0, + 1360.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1816.0, + 710.0, + 1816.0, + 710.0, + 1856.0, + 294.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1816.0, + 1360.0, + 1816.0, + 1360.0, + 1856.0, + 736.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1865.0, + 1404.0, + 1865.0, + 1404.0, + 1902.0, + 295.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1895.0, + 1293.0, + 1895.0, + 1293.0, + 1934.0, + 294.0, + 1934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1616.0, + 371.0, + 1616.0, + 371.0, + 1656.0, + 294.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1616.0, + 750.0, + 1616.0, + 750.0, + 1656.0, + 392.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 1616.0, + 1011.0, + 1616.0, + 1011.0, + 1656.0, + 780.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 1616.0, + 1199.0, + 1616.0, + 1199.0, + 1656.0, + 1149.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 1616.0, + 1405.0, + 1616.0, + 1405.0, + 1656.0, + 1335.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1650.0, + 763.0, + 1650.0, + 763.0, + 1689.0, + 290.0, + 1689.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1005, + 1404, + 1005, + 1404, + 1220, + 297, + 1220 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1288, + 1405, + 1288, + 1405, + 1566, + 297, + 1566 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 687, + 1404, + 687, + 1404, + 810, + 298, + 810 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 424, + 1403, + 424, + 1403, + 519, + 299, + 519 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1659, + 1403, + 1659, + 1403, + 1875, + 297, + 1875 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 300, + 1889, + 1403, + 1889, + 1403, + 1983, + 300, + 1983 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 295, + 874, + 1404, + 874, + 1404, + 938, + 295, + 938 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 289, + 227, + 1404, + 227, + 1404, + 294, + 289, + 294 + ], + "score": 0.943 + }, + { + "category_id": 8, + "poly": [ + 395, + 307, + 1328, + 307, + 1328, + 393, + 395, + 393 + ], + "score": 0.941 + }, + { + "category_id": 8, + "poly": [ + 395, + 822, + 1014, + 822, + 1014, + 865, + 395, + 865 + ], + "score": 0.933 + }, + { + "category_id": 8, + "poly": [ + 394, + 952, + 1042, + 952, + 1042, + 995, + 394, + 995 + ], + "score": 0.933 + }, + { + "category_id": 0, + "poly": [ + 300, + 561, + 589, + 561, + 589, + 598, + 300, + 598 + ], + "score": 0.926 + }, + { + "category_id": 0, + "poly": [ + 298, + 629, + 1013, + 629, + 1013, + 664, + 298, + 664 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 330, + 2005, + 996, + 2005, + 996, + 2033, + 330, + 2033 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 816, + 74, + 816, + 105, + 299, + 105 + ], + "score": 0.914 + }, + { + "category_id": 0, + "poly": [ + 300, + 1602, + 652, + 1602, + 652, + 1634, + 300, + 1634 + ], + "score": 0.91 + }, + { + "category_id": 9, + "poly": [ + 1366, + 959, + 1399, + 959, + 1399, + 988, + 1366, + 988 + ], + "score": 0.9 + }, + { + "category_id": 9, + "poly": [ + 1366, + 829, + 1399, + 829, + 1399, + 858, + 1366, + 858 + ], + "score": 0.877 + }, + { + "category_id": 1, + "poly": [ + 299, + 1230, + 815, + 1230, + 815, + 1267, + 299, + 1267 + ], + "score": 0.842 + }, + { + "category_id": 9, + "poly": [ + 1365, + 331, + 1401, + 331, + 1401, + 362, + 1365, + 362 + ], + "score": 0.83 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.651 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.437 + }, + { + "category_id": 13, + "poly": [ + 897, + 229, + 1109, + 229, + 1109, + 264, + 897, + 264 + ], + "score": 0.93, + "latex": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) \\in \\mathbb { R } ^ { m }" + }, + { + "category_id": 14, + "poly": [ + 395, + 302, + 1330, + 302, + 1330, + 396, + 395, + 396 + ], + "score": 0.93, + "latex": "\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\theta ) = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] = 2 \\sum _ { q = 1 } ^ { m } \\mathrm { v a r } _ { s _ { i } ^ { t } } ( p _ { q } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) ) ." + }, + { + "category_id": 14, + "poly": [ + 394, + 951, + 1044, + 951, + 1044, + 996, + 394, + 996 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\tilde { \\mathcal { R } } _ { E L D } ( \\mathbf { z } ^ { t } ; \\theta ) = \\mathbb { E } _ { s _ { i } } \\left[ \\Vert \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\mathbb { E } _ { s } [ s ] ; \\theta ) \\Vert _ { 2 } ^ { 2 } \\right] . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 863, + 1752, + 916, + 1752, + 916, + 1784, + 863, + 1784 + ], + "score": 0.91, + "latex": "\\omega ( t )" + }, + { + "category_id": 13, + "poly": [ + 476, + 1227, + 809, + 1227, + 809, + 1268, + 476, + 1268 + ], + "score": 0.91, + "latex": "\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) \\leq 4 \\tilde { \\mathcal { R } } _ { E L D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) ." + }, + { + "category_id": 14, + "poly": [ + 396, + 822, + 1013, + 822, + 1013, + 865, + 396, + 865 + ], + "score": 0.89, + "latex": "\\mathcal { R } _ { E L D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) = \\| \\mathbb { E } _ { s } \\left[ \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s ; \\boldsymbol { \\theta } ) \\right] - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\mathbb { E } _ { s } [ s ] ; \\boldsymbol { \\theta } ) \\| _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1185, + 1753, + 1229, + 1753, + 1229, + 1788, + 1185, + 1788 + ], + "score": 0.89, + "latex": "\\frac { \\kappa } { m T }" + }, + { + "category_id": 13, + "poly": [ + 478, + 230, + 504, + 230, + 504, + 264, + 478, + 264 + ], + "score": 0.88, + "latex": "s _ { i } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 557, + 230, + 583, + 230, + 583, + 267, + 557, + 267 + ], + "score": 0.87, + "latex": "\\boldsymbol { s } _ { j } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 372, + 882, + 388, + 882, + 388, + 903, + 372, + 903 + ], + "score": 0.74, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 987, + 485, + 1070, + 485, + 1070, + 517, + 987, + 517 + ], + "score": 0.47, + "latex": "\\mathrm { R N N s } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 639, + 1661, + 663, + 1661, + 663, + 1689, + 639, + 1689 + ], + "score": 0.28, + "latex": "\\Pi" + }, + { + "category_id": 13, + "poly": [ + 1289, + 1890, + 1313, + 1890, + 1313, + 1918, + 1289, + 1918 + ], + "score": 0.26, + "latex": "\\Pi" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 557.0, + 593.0, + 557.0, + 593.0, + 604.0, + 291.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 625.0, + 1015.0, + 625.0, + 1015.0, + 671.0, + 291.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 2000.0, + 997.0, + 2000.0, + 997.0, + 2037.0, + 330.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1602.0, + 656.0, + 1602.0, + 656.0, + 1638.0, + 294.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1004.0, + 1404.0, + 1004.0, + 1404.0, + 1041.0, + 296.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1037.0, + 1404.0, + 1037.0, + 1404.0, + 1072.0, + 295.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1067.0, + 1405.0, + 1067.0, + 1405.0, + 1101.0, + 295.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1098.0, + 1404.0, + 1098.0, + 1404.0, + 1132.0, + 295.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1127.0, + 1405.0, + 1127.0, + 1405.0, + 1162.0, + 294.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1159.0, + 1405.0, + 1159.0, + 1405.0, + 1194.0, + 295.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1190.0, + 1003.0, + 1190.0, + 1003.0, + 1224.0, + 296.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1289.0, + 1403.0, + 1289.0, + 1403.0, + 1325.0, + 295.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1321.0, + 1406.0, + 1321.0, + 1406.0, + 1357.0, + 295.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1351.0, + 1405.0, + 1351.0, + 1405.0, + 1388.0, + 294.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1382.0, + 1403.0, + 1382.0, + 1403.0, + 1415.0, + 296.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1411.0, + 1407.0, + 1411.0, + 1407.0, + 1451.0, + 291.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1441.0, + 1406.0, + 1441.0, + 1406.0, + 1479.0, + 291.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1472.0, + 1403.0, + 1472.0, + 1403.0, + 1508.0, + 295.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1502.0, + 1406.0, + 1502.0, + 1406.0, + 1540.0, + 292.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1536.0, + 492.0, + 1536.0, + 492.0, + 1570.0, + 296.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 687.0, + 1406.0, + 687.0, + 1406.0, + 723.0, + 293.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 718.0, + 1406.0, + 718.0, + 1406.0, + 753.0, + 293.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 747.0, + 1406.0, + 747.0, + 1406.0, + 786.0, + 292.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 779.0, + 713.0, + 779.0, + 713.0, + 814.0, + 295.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 424.0, + 1404.0, + 424.0, + 1404.0, + 462.0, + 294.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 454.0, + 1404.0, + 454.0, + 1404.0, + 490.0, + 294.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 486.0, + 986.0, + 486.0, + 986.0, + 520.0, + 293.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 486.0, + 1082.0, + 486.0, + 1082.0, + 520.0, + 1071.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1661.0, + 638.0, + 1661.0, + 638.0, + 1695.0, + 295.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 1661.0, + 1403.0, + 1661.0, + 1403.0, + 1695.0, + 664.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1692.0, + 1403.0, + 1692.0, + 1403.0, + 1726.0, + 295.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1720.0, + 1405.0, + 1720.0, + 1405.0, + 1758.0, + 293.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1749.0, + 862.0, + 1749.0, + 862.0, + 1789.0, + 291.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1749.0, + 1184.0, + 1749.0, + 1184.0, + 1789.0, + 917.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 1749.0, + 1408.0, + 1749.0, + 1408.0, + 1789.0, + 1230.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1784.0, + 1403.0, + 1784.0, + 1403.0, + 1815.0, + 295.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1813.0, + 1404.0, + 1813.0, + 1404.0, + 1847.0, + 293.0, + 1847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1842.0, + 1403.0, + 1842.0, + 1403.0, + 1877.0, + 295.0, + 1877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1886.0, + 1288.0, + 1886.0, + 1288.0, + 1926.0, + 294.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 1886.0, + 1405.0, + 1886.0, + 1405.0, + 1926.0, + 1314.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1921.0, + 1405.0, + 1921.0, + 1405.0, + 1955.0, + 294.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1949.0, + 1405.0, + 1949.0, + 1405.0, + 1986.0, + 295.0, + 1986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 870.0, + 371.0, + 870.0, + 371.0, + 914.0, + 293.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 870.0, + 1405.0, + 870.0, + 1405.0, + 914.0, + 389.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 906.0, + 741.0, + 906.0, + 741.0, + 941.0, + 295.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 225.0, + 477.0, + 225.0, + 477.0, + 270.0, + 293.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 225.0, + 556.0, + 225.0, + 556.0, + 270.0, + 505.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 225.0, + 896.0, + 225.0, + 896.0, + 270.0, + 584.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 225.0, + 1406.0, + 225.0, + 1406.0, + 270.0, + 1110.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 258.0, + 593.0, + 258.0, + 593.0, + 297.0, + 294.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1227.0, + 475.0, + 1227.0, + 475.0, + 1274.0, + 294.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 1227.0, + 818.0, + 1227.0, + 818.0, + 1274.0, + 810.0, + 1274.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 230, + 1404, + 230, + 1404, + 594, + 298, + 594 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1399, + 1404, + 1399, + 1404, + 1613, + 298, + 1613 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1151, + 1403, + 1151, + 1403, + 1335, + 298, + 1335 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1013, + 1402, + 1013, + 1402, + 1136, + 299, + 1136 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 720, + 1403, + 720, + 1403, + 843, + 299, + 843 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 300, + 1831, + 1404, + 1831, + 1404, + 1924, + 300, + 1924 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 301, + 612, + 1398, + 612, + 1398, + 704, + 301, + 704 + ], + "score": 0.959 + }, + { + "category_id": 2, + "poly": [ + 298, + 1948, + 1402, + 1948, + 1402, + 2034, + 298, + 2034 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 368, + 1638, + 1403, + 1638, + 1403, + 1805, + 368, + 1805 + ], + "score": 0.945 + }, + { + "category_id": 0, + "poly": [ + 300, + 1351, + 693, + 1351, + 693, + 1384, + 300, + 1384 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 300, + 956, + 613, + 956, + 613, + 988, + 300, + 988 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.885 + }, + { + "category_id": 0, + "poly": [ + 301, + 887, + 558, + 887, + 558, + 922, + 301, + 922 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.771 + }, + { + "category_id": 13, + "poly": [ + 1348, + 414, + 1400, + 414, + 1400, + 447, + 1348, + 447 + ], + "score": 0.92, + "latex": "\\omega ( t )" + }, + { + "category_id": 13, + "poly": [ + 460, + 505, + 513, + 505, + 513, + 537, + 460, + 537 + ], + "score": 0.92, + "latex": "\\omega ( t )" + }, + { + "category_id": 13, + "poly": [ + 958, + 1891, + 1218, + 1891, + 1218, + 1925, + 958, + 1925 + ], + "score": 0.91, + "latex": "n \\in \\{ 5 , 2 5 , 4 0 , 5 0 , 6 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 599, + 444, + 643, + 444, + 643, + 480, + 599, + 480 + ], + "score": 0.9, + "latex": "\\frac { \\kappa } { m T }" + }, + { + "category_id": 13, + "poly": [ + 1213, + 1836, + 1234, + 1836, + 1234, + 1858, + 1213, + 1858 + ], + "score": 0.64, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 447, + 613, + 471, + 613, + 471, + 641, + 447, + 641 + ], + "score": 0.38, + "latex": "\\Pi" + }, + { + "category_id": 13, + "poly": [ + 679, + 322, + 703, + 322, + 703, + 350, + 679, + 350 + ], + "score": 0.31, + "latex": "\\Pi" + }, + { + "category_id": 13, + "poly": [ + 1193, + 231, + 1217, + 231, + 1217, + 259, + 1193, + 259 + ], + "score": 0.27, + "latex": "\\Pi" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1942.0, + 1126.0, + 1942.0, + 1126.0, + 1979.0, + 330.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1971.0, + 1403.0, + 1971.0, + 1403.0, + 2012.0, + 327.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2005.0, + 486.0, + 2005.0, + 486.0, + 2035.0, + 296.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1349.0, + 696.0, + 1349.0, + 696.0, + 1387.0, + 297.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 953.0, + 616.0, + 953.0, + 616.0, + 992.0, + 294.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 884.0, + 561.0, + 884.0, + 561.0, + 927.0, + 292.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 1192.0, + 230.0, + 1192.0, + 263.0, + 296.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 230.0, + 1404.0, + 230.0, + 1404.0, + 263.0, + 1218.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 259.0, + 1405.0, + 259.0, + 1405.0, + 298.0, + 292.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 292.0, + 1404.0, + 292.0, + 1404.0, + 326.0, + 294.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 320.0, + 678.0, + 320.0, + 678.0, + 355.0, + 293.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 320.0, + 1405.0, + 320.0, + 1405.0, + 355.0, + 704.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 352.0, + 1405.0, + 352.0, + 1405.0, + 389.0, + 293.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 381.0, + 1407.0, + 381.0, + 1407.0, + 417.0, + 292.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 414.0, + 1347.0, + 414.0, + 1347.0, + 448.0, + 292.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 414.0, + 1404.0, + 414.0, + 1404.0, + 448.0, + 1401.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 443.0, + 598.0, + 443.0, + 598.0, + 476.0, + 293.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 442.0, + 1407.0, + 442.0, + 1407.0, + 478.0, + 644.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 475.0, + 1406.0, + 475.0, + 1406.0, + 509.0, + 296.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 503.0, + 459.0, + 503.0, + 459.0, + 539.0, + 293.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 503.0, + 1405.0, + 503.0, + 1405.0, + 539.0, + 514.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 534.0, + 1406.0, + 534.0, + 1406.0, + 571.0, + 292.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 564.0, + 404.0, + 564.0, + 404.0, + 598.0, + 294.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1399.0, + 1402.0, + 1399.0, + 1402.0, + 1434.0, + 296.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1429.0, + 1406.0, + 1429.0, + 1406.0, + 1466.0, + 294.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1462.0, + 1405.0, + 1462.0, + 1405.0, + 1494.0, + 293.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1490.0, + 1405.0, + 1490.0, + 1405.0, + 1526.0, + 293.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1518.0, + 1405.0, + 1518.0, + 1405.0, + 1559.0, + 292.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1549.0, + 1404.0, + 1549.0, + 1404.0, + 1589.0, + 293.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1582.0, + 624.0, + 1582.0, + 624.0, + 1619.0, + 292.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1150.0, + 1403.0, + 1150.0, + 1403.0, + 1186.0, + 294.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1182.0, + 1405.0, + 1182.0, + 1405.0, + 1216.0, + 294.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1211.0, + 1407.0, + 1211.0, + 1407.0, + 1248.0, + 293.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1242.0, + 1405.0, + 1242.0, + 1405.0, + 1280.0, + 293.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1272.0, + 1405.0, + 1272.0, + 1405.0, + 1310.0, + 292.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1303.0, + 381.0, + 1303.0, + 381.0, + 1337.0, + 292.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1013.0, + 1406.0, + 1013.0, + 1406.0, + 1049.0, + 293.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1044.0, + 1407.0, + 1044.0, + 1407.0, + 1080.0, + 294.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1076.0, + 1406.0, + 1076.0, + 1406.0, + 1110.0, + 292.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1105.0, + 691.0, + 1105.0, + 691.0, + 1140.0, + 294.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 720.0, + 1404.0, + 720.0, + 1404.0, + 756.0, + 294.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 750.0, + 1407.0, + 750.0, + 1407.0, + 786.0, + 294.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 783.0, + 1405.0, + 783.0, + 1405.0, + 815.0, + 295.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 812.0, + 944.0, + 812.0, + 944.0, + 844.0, + 295.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1830.0, + 1212.0, + 1830.0, + 1212.0, + 1864.0, + 296.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 1830.0, + 1404.0, + 1830.0, + 1404.0, + 1864.0, + 1235.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1862.0, + 1405.0, + 1862.0, + 1405.0, + 1894.0, + 293.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1890.0, + 957.0, + 1890.0, + 957.0, + 1929.0, + 293.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 1890.0, + 1404.0, + 1890.0, + 1404.0, + 1929.0, + 1219.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 610.0, + 446.0, + 610.0, + 446.0, + 648.0, + 295.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 610.0, + 1402.0, + 610.0, + 1402.0, + 648.0, + 472.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 644.0, + 1403.0, + 644.0, + 1403.0, + 677.0, + 295.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 674.0, + 922.0, + 674.0, + 922.0, + 708.0, + 293.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1639.0, + 1403.0, + 1639.0, + 1403.0, + 1673.0, + 369.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1669.0, + 1405.0, + 1669.0, + 1405.0, + 1705.0, + 394.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1702.0, + 1078.0, + 1702.0, + 1078.0, + 1736.0, + 396.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 1742.0, + 1406.0, + 1742.0, + 1406.0, + 1779.0, + 372.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1774.0, + 1324.0, + 1774.0, + 1324.0, + 1805.0, + 394.0, + 1805.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1551, + 1404, + 1551, + 1404, + 1736, + 299, + 1736 + ], + "score": 0.978 + }, + { + "category_id": 5, + "poly": [ + 318, + 658, + 1381, + 658, + 1381, + 943, + 318, + 943 + ], + "score": 0.976, + "html": "
ModelParametersValidationTest
Merity et al. (2016) - Variational LSTM + Zoneout20M108.7100.9
Merity et al. (2016) - Variational LSTM20M101.796.3
Inan et al. (2016) - Variational LSTM28M91.587.0
Melis et al. (2017) - 5-layer RHN24M78.175.6
Melis et al. (2017) - 1-layer LSTM24M69.365.9
Melis et al. (2017) - 2-layer skip connection LSTM24M69.165.9
Merity et al. (2017a) - AWD-LSTM3-layer (baseline)34M68.665.8
Fraternal dropout + AWD-LSTM 3-layer34M66.864.1
" + }, + { + "category_id": 5, + "poly": [ + 309, + 218, + 1392, + 218, + 1392, + 592, + 309, + 592 + ], + "score": 0.975, + "html": "
ModelParametersValidationTest
Zaremba et al. (2014) - LSTM (medium)10M86.282.7
Zaremba et al. (2014) - LSTM (large)24M82.278.4
Gal & Ghahramani (2016) - Variational LSTM (medium)20M81.979.7
Gal & Ghahramani (2016) - Variational LSTM (large)66M77.975.2
Inan et al. (2016) - Variational LSTM51M71.168.5
Inan et al. (2016) - Variational RHN24M68.166.0
Zilly et al. (2016) - Variational RHN23M67.965.4
Melis et al. (2017) - 5-layer RHN24M64.862.2
Melis et al. (2017) - 4-layer skip connection LSTM24M60.958.3
Merity etal. (2017a) - AWD-LSTM3-layer (baseline)24M60.057.3
Fraternal dropout + AWD-LSTM3-layer24M58.956.8
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1228, + 1403, + 1228, + 1403, + 1382, + 299, + 1382 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 1831, + 1403, + 1831, + 1403, + 1954, + 299, + 1954 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 300, + 1397, + 1403, + 1397, + 1403, + 1488, + 300, + 1488 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 299, + 1120, + 1403, + 1120, + 1403, + 1212, + 299, + 1212 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 302, + 1043, + 1402, + 1043, + 1402, + 1104, + 302, + 1104 + ], + "score": 0.949 + }, + { + "category_id": 2, + "poly": [ + 299, + 1978, + 1403, + 1978, + 1403, + 2034, + 299, + 2034 + ], + "score": 0.931 + }, + { + "category_id": 0, + "poly": [ + 300, + 1504, + 653, + 1504, + 653, + 1536, + 300, + 1536 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 300, + 1774, + 603, + 1774, + 603, + 1805, + 300, + 1805 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.888 + }, + { + "category_id": 6, + "poly": [ + 435, + 605, + 1261, + 605, + 1261, + 638, + 435, + 638 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.76 + }, + { + "category_id": 6, + "poly": [ + 456, + 956, + 1239, + 956, + 1239, + 987, + 456, + 987 + ], + "score": 0.732 + }, + { + "category_id": 7, + "poly": [ + 456, + 956, + 1239, + 956, + 1239, + 987, + 456, + 987 + ], + "score": 0.108 + }, + { + "category_id": 13, + "poly": [ + 388, + 1675, + 474, + 1675, + 474, + 1703, + 388, + 1703 + ], + "score": 0.89, + "latex": "n = 6 0" + }, + { + "category_id": 13, + "poly": [ + 703, + 1290, + 852, + 1290, + 852, + 1320, + 703, + 1320 + ], + "score": 0.89, + "latex": "6 0 . 6 4 \\pm 0 . 1 5" + }, + { + "category_id": 13, + "poly": [ + 528, + 1321, + 679, + 1321, + 679, + 1350, + 528, + 1350 + ], + "score": 0.88, + "latex": "5 8 . 3 2 \\pm 0 . 1 4" + }, + { + "category_id": 13, + "poly": [ + 1128, + 1645, + 1223, + 1645, + 1223, + 1673, + 1128, + 1673 + ], + "score": 0.86, + "latex": "\\kappa = 0 . 1" + }, + { + "category_id": 13, + "poly": [ + 619, + 564, + 641, + 564, + 641, + 585, + 619, + 585 + ], + "score": 0.6, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 611, + 913, + 635, + 913, + 635, + 936, + 611, + 936 + ], + "score": 0.57, + "latex": "^ +" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1968.0, + 1404.0, + 1968.0, + 1404.0, + 2015.0, + 327.0, + 2015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2005.0, + 600.0, + 2005.0, + 600.0, + 2035.0, + 295.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1501.0, + 655.0, + 1501.0, + 655.0, + 1544.0, + 297.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1771.0, + 607.0, + 1771.0, + 607.0, + 1809.0, + 295.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 603.0, + 1261.0, + 603.0, + 1261.0, + 642.0, + 436.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2119.0, + 838.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 952.0, + 1242.0, + 952.0, + 1242.0, + 994.0, + 455.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 952.0, + 1242.0, + 952.0, + 1242.0, + 994.0, + 455.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1547.0, + 1406.0, + 1547.0, + 1406.0, + 1591.0, + 292.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1581.0, + 1408.0, + 1581.0, + 1408.0, + 1619.0, + 294.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1613.0, + 1406.0, + 1613.0, + 1406.0, + 1649.0, + 294.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1643.0, + 1127.0, + 1643.0, + 1127.0, + 1679.0, + 295.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 1643.0, + 1405.0, + 1643.0, + 1405.0, + 1679.0, + 1224.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1674.0, + 387.0, + 1674.0, + 387.0, + 1710.0, + 294.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 1674.0, + 1405.0, + 1674.0, + 1405.0, + 1710.0, + 475.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1703.0, + 1092.0, + 1703.0, + 1092.0, + 1738.0, + 293.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1227.0, + 1405.0, + 1227.0, + 1405.0, + 1264.0, + 297.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1260.0, + 1404.0, + 1260.0, + 1404.0, + 1293.0, + 297.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1288.0, + 702.0, + 1288.0, + 702.0, + 1326.0, + 292.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1288.0, + 1405.0, + 1288.0, + 1405.0, + 1326.0, + 853.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1320.0, + 527.0, + 1320.0, + 527.0, + 1357.0, + 294.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 1320.0, + 1405.0, + 1320.0, + 1405.0, + 1357.0, + 680.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1350.0, + 865.0, + 1350.0, + 865.0, + 1387.0, + 294.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1829.0, + 1405.0, + 1829.0, + 1405.0, + 1866.0, + 294.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1859.0, + 1407.0, + 1859.0, + 1407.0, + 1897.0, + 293.0, + 1897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1892.0, + 1405.0, + 1892.0, + 1405.0, + 1927.0, + 294.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1920.0, + 1405.0, + 1920.0, + 1405.0, + 1958.0, + 293.0, + 1958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1396.0, + 1405.0, + 1396.0, + 1405.0, + 1433.0, + 295.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1426.0, + 1404.0, + 1426.0, + 1404.0, + 1462.0, + 294.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1459.0, + 531.0, + 1459.0, + 531.0, + 1489.0, + 296.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1119.0, + 1407.0, + 1119.0, + 1407.0, + 1156.0, + 295.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1152.0, + 1404.0, + 1152.0, + 1404.0, + 1185.0, + 295.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1183.0, + 1029.0, + 1183.0, + 1029.0, + 1215.0, + 294.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1042.0, + 1405.0, + 1042.0, + 1405.0, + 1077.0, + 297.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1072.0, + 1331.0, + 1072.0, + 1331.0, + 1107.0, + 294.0, + 1107.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1080, + 1404, + 1080, + 1404, + 1356, + 298, + 1356 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 860, + 1404, + 860, + 1404, + 982, + 297, + 982 + ], + "score": 0.977 + }, + { + "category_id": 5, + "poly": [ + 394, + 216, + 1302, + 216, + 1302, + 380, + 394, + 380 + ], + "score": 0.977, + "html": "
ModelBLEU-1BLEU-2BLEU-3BLEU-4
Show and Tell Xu et al. (2015)66.646.132.924.6
Baseline68.850.836.125.6
Fraternal dropout, K = 0.01569.351.436.626.1
Fraternal dropout, K = 0.00569.351.536.926.3
" + }, + { + "category_id": 1, + "poly": [ + 299, + 511, + 1404, + 511, + 1404, + 604, + 299, + 604 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 619, + 1404, + 619, + 1404, + 743, + 299, + 743 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1453, + 1404, + 1453, + 1404, + 1699, + 297, + 1699 + ], + "score": 0.953 + }, + { + "category_id": 0, + "poly": [ + 301, + 789, + 635, + 789, + 635, + 824, + 301, + 824 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 302, + 1023, + 840, + 1023, + 840, + 1054, + 302, + 1054 + ], + "score": 0.873 + }, + { + "category_id": 0, + "poly": [ + 300, + 1395, + 482, + 1395, + 482, + 1427, + 300, + 1427 + ], + "score": 0.872 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.773 + }, + { + "category_id": 2, + "poly": [ + 297, + 1726, + 1404, + 1726, + 1404, + 1920, + 297, + 1920 + ], + "score": 0.585 + }, + { + "category_id": 2, + "poly": [ + 298, + 1923, + 1402, + 1923, + 1402, + 2032, + 298, + 2032 + ], + "score": 0.547 + }, + { + "category_id": 6, + "poly": [ + 289, + 391, + 1403, + 391, + 1403, + 454, + 289, + 454 + ], + "score": 0.341 + }, + { + "category_id": 13, + "poly": [ + 622, + 1922, + 745, + 1922, + 745, + 1954, + 622, + 1954 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\omega ( t ) = \\frac { \\kappa } { m T } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1862, + 418, + 1862, + 418, + 1892, + 297, + 1892 + ], + "score": 0.92, + "latex": "1 . 2 \\times 1 0 ^ { - 6 }" + }, + { + "category_id": 13, + "poly": [ + 663, + 625, + 683, + 625, + 683, + 647, + 663, + 647 + ], + "score": 0.74, + "latex": "\\kappa" + }, + { + "category_id": 13, + "poly": [ + 632, + 315, + 753, + 315, + 753, + 343, + 632, + 343 + ], + "score": 0.61, + "latex": "\\kappa = 0 . 0 1 5" + }, + { + "category_id": 13, + "poly": [ + 632, + 346, + 753, + 346, + 753, + 376, + 632, + 376 + ], + "score": 0.58, + "latex": "\\kappa = 0 . 0 0 5" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 784.0, + 639.0, + 784.0, + 639.0, + 831.0, + 292.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1021.0, + 842.0, + 1021.0, + 842.0, + 1062.0, + 293.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1390.0, + 488.0, + 1390.0, + 488.0, + 1433.0, + 293.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1718.0, + 1408.0, + 1718.0, + 1408.0, + 1764.0, + 326.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1752.0, + 1405.0, + 1752.0, + 1405.0, + 1786.0, + 294.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1781.0, + 1405.0, + 1781.0, + 1405.0, + 1814.0, + 295.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1807.0, + 1405.0, + 1807.0, + 1405.0, + 1844.0, + 295.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1832.0, + 1404.0, + 1832.0, + 1404.0, + 1872.0, + 292.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 1860.0, + 1406.0, + 1860.0, + 1406.0, + 1899.0, + 419.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1890.0, + 1057.0, + 1890.0, + 1057.0, + 1927.0, + 294.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1916.0, + 621.0, + 1916.0, + 621.0, + 1958.0, + 330.0, + 1958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 1916.0, + 1408.0, + 1916.0, + 1408.0, + 1958.0, + 746.0, + 1958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1951.0, + 1403.0, + 1951.0, + 1403.0, + 1980.0, + 296.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1975.0, + 1407.0, + 1975.0, + 1407.0, + 2011.0, + 292.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2006.0, + 810.0, + 2006.0, + 810.0, + 2035.0, + 296.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1922.5, + 738.0, + 1922.5, + 738.0, + 1941.5, + 702.0, + 1941.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 391.0, + 1405.0, + 391.0, + 1405.0, + 426.0, + 296.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 420.0, + 1186.0, + 420.0, + 1186.0, + 457.0, + 293.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1082.0, + 1405.0, + 1082.0, + 1405.0, + 1115.0, + 297.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1111.0, + 1406.0, + 1111.0, + 1406.0, + 1147.0, + 293.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1142.0, + 1406.0, + 1142.0, + 1406.0, + 1178.0, + 293.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1168.0, + 1405.0, + 1168.0, + 1405.0, + 1212.0, + 292.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1204.0, + 1406.0, + 1204.0, + 1406.0, + 1240.0, + 294.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1234.0, + 1405.0, + 1234.0, + 1405.0, + 1270.0, + 294.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1264.0, + 1405.0, + 1264.0, + 1405.0, + 1300.0, + 293.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1291.0, + 1404.0, + 1291.0, + 1404.0, + 1330.0, + 292.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1326.0, + 1351.0, + 1326.0, + 1351.0, + 1359.0, + 296.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 861.0, + 1404.0, + 861.0, + 1404.0, + 893.0, + 295.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 892.0, + 1404.0, + 892.0, + 1404.0, + 924.0, + 295.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 920.0, + 1404.0, + 920.0, + 1404.0, + 957.0, + 295.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 950.0, + 1082.0, + 950.0, + 1082.0, + 986.0, + 292.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 512.0, + 1404.0, + 512.0, + 1404.0, + 545.0, + 294.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 543.0, + 1405.0, + 543.0, + 1405.0, + 577.0, + 294.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 573.0, + 1119.0, + 573.0, + 1119.0, + 607.0, + 295.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 618.0, + 662.0, + 618.0, + 662.0, + 654.0, + 294.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 618.0, + 1404.0, + 618.0, + 1404.0, + 654.0, + 684.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 649.0, + 1405.0, + 649.0, + 1405.0, + 685.0, + 294.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 681.0, + 1404.0, + 681.0, + 1404.0, + 717.0, + 294.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 711.0, + 1261.0, + 711.0, + 1261.0, + 747.0, + 294.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1455.0, + 1404.0, + 1455.0, + 1404.0, + 1488.0, + 296.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1487.0, + 1404.0, + 1487.0, + 1404.0, + 1517.0, + 296.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1512.0, + 1404.0, + 1512.0, + 1404.0, + 1552.0, + 294.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1546.0, + 1404.0, + 1546.0, + 1404.0, + 1578.0, + 294.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1577.0, + 1404.0, + 1577.0, + 1404.0, + 1611.0, + 295.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1604.0, + 1406.0, + 1604.0, + 1406.0, + 1643.0, + 291.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1638.0, + 1404.0, + 1638.0, + 1404.0, + 1672.0, + 295.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1670.0, + 926.0, + 1670.0, + 926.0, + 1701.0, + 293.0, + 1701.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1269, + 1405, + 1269, + 1405, + 1576, + 296, + 1576 + ], + "score": 0.981 + }, + { + "category_id": 4, + "poly": [ + 295, + 1091, + 831, + 1091, + 831, + 1214, + 295, + 1214 + ], + "score": 0.968 + }, + { + "category_id": 3, + "poly": [ + 319, + 765, + 780, + 765, + 780, + 1070, + 319, + 1070 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 861, + 1091, + 1396, + 1091, + 1396, + 1217, + 861, + 1217 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 876, + 766, + 1344, + 766, + 1344, + 1070, + 876, + 1070 + ], + "score": 0.966 + }, + { + "category_id": 3, + "poly": [ + 318, + 242, + 779, + 242, + 779, + 545, + 318, + 545 + ], + "score": 0.964 + }, + { + "category_id": 3, + "poly": [ + 890, + 242, + 1355, + 242, + 1355, + 546, + 890, + 546 + ], + "score": 0.961 + }, + { + "category_id": 4, + "poly": [ + 296, + 566, + 1406, + 566, + 1406, + 721, + 296, + 721 + ], + "score": 0.958 + }, + { + "category_id": 2, + "poly": [ + 293, + 1977, + 1406, + 1977, + 1406, + 2035, + 293, + 2035 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 397, + 1812, + 766, + 1812, + 766, + 1873, + 397, + 1873 + ], + "score": 0.938 + }, + { + "category_id": 1, + "poly": [ + 298, + 1892, + 1403, + 1892, + 1403, + 1955, + 298, + 1955 + ], + "score": 0.936 + }, + { + "category_id": 1, + "poly": [ + 298, + 1669, + 1396, + 1669, + 1396, + 1733, + 298, + 1733 + ], + "score": 0.931 + }, + { + "category_id": 8, + "poly": [ + 413, + 1749, + 676, + 1749, + 676, + 1806, + 413, + 1806 + ], + "score": 0.911 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.886 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1829, + 1400, + 1829, + 1400, + 1859, + 1366, + 1859 + ], + "score": 0.867 + }, + { + "category_id": 9, + "poly": [ + 1367, + 1762, + 1400, + 1762, + 1400, + 1790, + 1367, + 1790 + ], + "score": 0.867 + }, + { + "category_id": 0, + "poly": [ + 296, + 1615, + 1362, + 1615, + 1362, + 1644, + 296, + 1644 + ], + "score": 0.815 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 858, + 2087, + 858, + 2111, + 841, + 2111 + ], + "score": 0.712 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.156 + }, + { + "category_id": 1, + "poly": [ + 296, + 1615, + 1362, + 1615, + 1362, + 1644, + 296, + 1644 + ], + "score": 0.098 + }, + { + "category_id": 13, + "poly": [ + 864, + 1179, + 997, + 1179, + 997, + 1218, + 864, + 1218 + ], + "score": 0.92, + "latex": "\\frac { 1 } { d } \\| \\mathbf { m } \\cdot \\mathbf { h } ^ { t } \\| _ { 2 } ^ { 2 }" + }, + { + "category_id": 14, + "poly": [ + 395, + 1747, + 766, + 1747, + 766, + 1880, + 395, + 1880 + ], + "score": 0.92, + "latex": "\\begin{array} { c } { \\mathcal { R } _ { A R } ( { \\bf z } ^ { t } ; \\theta ) = \\displaystyle \\frac { \\alpha } { d } \\| { \\bf h } ^ { t } \\| _ { 2 } ^ { 2 } } \\\\ { \\mathcal { R } _ { T A R } ( { \\bf z } ^ { t } ; \\theta ) = \\displaystyle \\frac { \\beta } { d } \\| { \\bf h } ^ { t } - { \\bf h } ^ { t - 1 } \\| _ { 2 } ^ { 2 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 372, + 1890, + 470, + 1890, + 470, + 1921, + 372, + 1921 + ], + "score": 0.92, + "latex": "\\mathbf { h } ^ { t } \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 332, + 1453, + 620, + 1453, + 620, + 1487, + 332, + 1487 + ], + "score": 0.91, + "latex": "\\kappa \\in \\{ 0 . 0 5 , 0 . 1 , 0 . 1 5 , 0 . 2 \\}" + }, + { + "category_id": 13, + "poly": [ + 871, + 1332, + 964, + 1332, + 964, + 1361, + 871, + 1361 + ], + "score": 0.9, + "latex": "3 2 \\times 3 2" + }, + { + "category_id": 13, + "poly": [ + 440, + 1976, + 546, + 1976, + 546, + 2008, + 440, + 2008 + ], + "score": 0.9, + "latex": "\\| \\mathbf { m } \\cdot \\mathbf { h } ^ { t } \\| _ { 2 } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 815, + 1453, + 1050, + 1453, + 1050, + 1487, + 815, + 1487 + ], + "score": 0.89, + "latex": "\\{ 0 . 0 5 , 0 . 1 , 0 . 1 5 , 0 . 2 \\}" + }, + { + "category_id": 13, + "poly": [ + 1101, + 313, + 1154, + 313, + 1154, + 332, + 1101, + 332 + ], + "score": 0.85, + "latex": "\\kappa { = } 0 . 1 5" + }, + { + "category_id": 13, + "poly": [ + 543, + 314, + 596, + 314, + 596, + 333, + 543, + 333 + ], + "score": 0.84, + "latex": "\\kappa { = } 0 . 1 5" + }, + { + "category_id": 13, + "poly": [ + 512, + 336, + 566, + 336, + 566, + 355, + 512, + 355 + ], + "score": 0.83, + "latex": "\\kappa { = } 0 . 2 5" + }, + { + "category_id": 13, + "poly": [ + 298, + 1923, + 324, + 1923, + 324, + 1950, + 298, + 1950 + ], + "score": 0.83, + "latex": "\\mathbf { z } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 1070, + 334, + 1122, + 334, + 1122, + 353, + 1070, + 353 + ], + "score": 0.82, + "latex": "\\kappa { = } 0 . 2 5" + }, + { + "category_id": 13, + "poly": [ + 1082, + 272, + 1135, + 272, + 1135, + 290, + 1082, + 290 + ], + "score": 0.82, + "latex": "\\kappa { = } 0 . 2 5" + }, + { + "category_id": 13, + "poly": [ + 504, + 358, + 556, + 358, + 556, + 377, + 504, + 377 + ], + "score": 0.8, + "latex": "\\kappa { = } 0 . 1 5" + }, + { + "category_id": 13, + "poly": [ + 1226, + 1669, + 1295, + 1669, + 1295, + 1702, + 1226, + 1702 + ], + "score": 0.8, + "latex": "( \\mathsf { A R } ) ^ { 8 }" + }, + { + "category_id": 13, + "poly": [ + 524, + 271, + 577, + 271, + 577, + 290, + 524, + 290 + ], + "score": 0.79, + "latex": "\\kappa { = } 0 . 2 5" + }, + { + "category_id": 13, + "poly": [ + 1062, + 356, + 1115, + 356, + 1115, + 374, + 1062, + 374 + ], + "score": 0.79, + "latex": "\\kappa { = } 0 . 1 5" + }, + { + "category_id": 13, + "poly": [ + 631, + 1990, + 655, + 1990, + 655, + 2003, + 631, + 2003 + ], + "score": 0.78, + "latex": "m" + }, + { + "category_id": 13, + "poly": [ + 616, + 1927, + 630, + 1927, + 630, + 1950, + 616, + 1950 + ], + "score": 0.76, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 968, + 1896, + 981, + 1896, + 981, + 1920, + 968, + 1920 + ], + "score": 0.73, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 653, + 1302, + 677, + 1302, + 677, + 1330, + 653, + 1330 + ], + "score": 0.46, + "latex": "\\Pi" + }, + { + "category_id": 13, + "poly": [ + 677, + 1122, + 702, + 1122, + 702, + 1150, + 677, + 1150 + ], + "score": 0.43, + "latex": "\\Pi" + }, + { + "category_id": 13, + "poly": [ + 1050, + 1515, + 1074, + 1515, + 1074, + 1543, + 1050, + 1543 + ], + "score": 0.34, + "latex": "\\Pi" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1090.0, + 831.0, + 1090.0, + 831.0, + 1126.0, + 295.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1122.0, + 676.0, + 1122.0, + 676.0, + 1153.0, + 296.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1122.0, + 832.0, + 1122.0, + 832.0, + 1153.0, + 703.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1149.0, + 834.0, + 1149.0, + 834.0, + 1188.0, + 291.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1182.0, + 618.0, + 1182.0, + 618.0, + 1216.0, + 295.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 764.0, + 785.0, + 764.0, + 785.0, + 791.0, + 356.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 783.0, + 370.0, + 783.0, + 370.0, + 812.0, + 328.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 791.0, + 770.0, + 791.0, + 770.0, + 818.0, + 516.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 810.0, + 370.0, + 810.0, + 370.0, + 842.0, + 337.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 811.0, + 730.0, + 811.0, + 730.0, + 839.0, + 517.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 840.0, + 371.0, + 840.0, + 371.0, + 871.0, + 337.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 867.0, + 350.0, + 867.0, + 350.0, + 959.0, + 314.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 870.0, + 368.0, + 870.0, + 368.0, + 897.0, + 338.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 898.0, + 369.0, + 898.0, + 369.0, + 925.0, + 339.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 927.0, + 368.0, + 927.0, + 368.0, + 954.0, + 339.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 953.0, + 370.0, + 953.0, + 370.0, + 985.0, + 337.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 984.0, + 369.0, + 984.0, + 369.0, + 1012.0, + 338.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1013.0, + 369.0, + 1013.0, + 369.0, + 1040.0, + 339.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 1035.0, + 392.0, + 1035.0, + 392.0, + 1054.0, + 375.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 1031.0, + 461.0, + 1031.0, + 461.0, + 1058.0, + 430.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 1030.0, + 530.0, + 1030.0, + 530.0, + 1059.0, + 487.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 1030.0, + 592.0, + 1030.0, + 592.0, + 1058.0, + 549.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 1032.0, + 655.0, + 1032.0, + 655.0, + 1057.0, + 612.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 1032.0, + 717.0, + 1032.0, + 717.0, + 1057.0, + 673.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1033.0, + 776.0, + 1033.0, + 776.0, + 1057.0, + 736.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 1046.0, + 604.0, + 1046.0, + 604.0, + 1075.0, + 539.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1088.0, + 1397.0, + 1088.0, + 1397.0, + 1124.0, + 860.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1122.0, + 1395.0, + 1122.0, + 1395.0, + 1153.0, + 861.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 1148.0, + 1399.0, + 1148.0, + 1399.0, + 1185.0, + 859.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 1175.0, + 1011.0, + 1175.0, + 1011.0, + 1222.0, + 998.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 765.0, + 1311.0, + 765.0, + 1311.0, + 789.0, + 962.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 790.0, + 1336.0, + 790.0, + 1336.0, + 815.0, + 1263.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 804.0, + 935.0, + 804.0, + 935.0, + 833.0, + 889.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 808.0, + 1308.0, + 808.0, + 1308.0, + 839.0, + 1260.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 829.0, + 1297.0, + 829.0, + 1297.0, + 860.0, + 1260.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 849.0, + 935.0, + 849.0, + 935.0, + 877.0, + 889.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 850.0, + 1297.0, + 850.0, + 1297.0, + 879.0, + 1259.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 876.0, + 936.0, + 876.0, + 936.0, + 946.0, + 870.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 869.0, + 1297.0, + 869.0, + 1297.0, + 900.0, + 1260.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 938.0, + 935.0, + 938.0, + 935.0, + 967.0, + 889.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 983.0, + 935.0, + 983.0, + 935.0, + 1011.0, + 889.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 1035.0, + 961.0, + 1035.0, + 961.0, + 1055.0, + 939.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1032.0, + 1075.0, + 1032.0, + 1075.0, + 1057.0, + 981.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 1032.0, + 1177.0, + 1032.0, + 1177.0, + 1057.0, + 1077.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 1032.0, + 1328.0, + 1032.0, + 1328.0, + 1057.0, + 1180.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 1048.0, + 1169.0, + 1048.0, + 1169.0, + 1074.0, + 1104.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 241.0, + 776.0, + 241.0, + 776.0, + 266.0, + 365.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 261.0, + 369.0, + 261.0, + 369.0, + 286.0, + 331.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 269.0, + 523.0, + 269.0, + 523.0, + 292.0, + 474.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 269.0, + 749.0, + 269.0, + 749.0, + 292.0, + 578.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 290.0, + 711.0, + 290.0, + 711.0, + 314.0, + 474.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 301.0, + 369.0, + 301.0, + 369.0, + 324.0, + 341.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 311.0, + 542.0, + 311.0, + 542.0, + 335.0, + 474.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 311.0, + 770.0, + 311.0, + 770.0, + 335.0, + 597.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 339.0, + 368.0, + 339.0, + 368.0, + 362.0, + 340.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 333.0, + 511.0, + 333.0, + 511.0, + 355.0, + 474.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 333.0, + 737.0, + 333.0, + 737.0, + 355.0, + 567.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 349.0, + 340.0, + 349.0, + 340.0, + 430.0, + 318.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 355.0, + 503.0, + 355.0, + 503.0, + 379.0, + 474.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 355.0, + 730.0, + 355.0, + 730.0, + 379.0, + 557.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 378.0, + 367.0, + 378.0, + 367.0, + 400.0, + 341.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 416.0, + 367.0, + 416.0, + 367.0, + 439.0, + 341.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 454.0, + 369.0, + 454.0, + 369.0, + 480.0, + 339.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 488.0, + 564.0, + 488.0, + 564.0, + 495.0, + 553.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 491.0, + 371.0, + 491.0, + 371.0, + 517.0, + 339.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 510.0, + 394.0, + 510.0, + 394.0, + 531.0, + 375.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 510.0, + 457.0, + 510.0, + 457.0, + 531.0, + 421.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 509.0, + 512.0, + 509.0, + 512.0, + 533.0, + 473.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 510.0, + 567.0, + 510.0, + 567.0, + 532.0, + 529.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 510.0, + 622.0, + 510.0, + 622.0, + 532.0, + 584.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 510.0, + 674.0, + 510.0, + 674.0, + 531.0, + 638.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 692.0, + 510.0, + 730.0, + 510.0, + 730.0, + 532.0, + 692.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 510.0, + 782.0, + 510.0, + 782.0, + 531.0, + 746.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 525.0, + 603.0, + 525.0, + 603.0, + 549.0, + 540.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 401.5, + 508.0, + 401.5, + 508.0, + 414.0, + 479.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 242.0, + 1356.0, + 242.0, + 1356.0, + 266.0, + 929.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 262.0, + 941.0, + 262.0, + 941.0, + 286.0, + 903.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 270.0, + 1081.0, + 270.0, + 1081.0, + 291.0, + 1032.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 270.0, + 1322.0, + 270.0, + 1322.0, + 291.0, + 1136.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 297.0, + 940.0, + 297.0, + 940.0, + 321.0, + 913.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 290.0, + 1283.0, + 290.0, + 1283.0, + 311.0, + 1032.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 311.0, + 1100.0, + 311.0, + 1100.0, + 332.0, + 1030.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 311.0, + 1340.0, + 311.0, + 1340.0, + 332.0, + 1155.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 331.0, + 940.0, + 331.0, + 940.0, + 354.0, + 913.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 332.0, + 1069.0, + 332.0, + 1069.0, + 353.0, + 1032.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 332.0, + 1310.0, + 332.0, + 1310.0, + 353.0, + 1123.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 347.0, + 915.0, + 347.0, + 915.0, + 431.0, + 889.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 352.0, + 1061.0, + 352.0, + 1061.0, + 376.0, + 1031.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 352.0, + 1301.0, + 352.0, + 1301.0, + 376.0, + 1116.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 367.0, + 939.0, + 367.0, + 939.0, + 390.0, + 913.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 401.0, + 939.0, + 401.0, + 939.0, + 424.0, + 913.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 435.0, + 940.0, + 435.0, + 940.0, + 462.0, + 912.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 471.0, + 940.0, + 471.0, + 940.0, + 495.0, + 913.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 512.0, + 965.0, + 512.0, + 965.0, + 530.0, + 948.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 510.0, + 1032.0, + 510.0, + 1032.0, + 532.0, + 994.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 509.0, + 1084.0, + 509.0, + 1084.0, + 532.0, + 1044.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 510.0, + 1139.0, + 510.0, + 1139.0, + 531.0, + 1103.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 509.0, + 1195.0, + 509.0, + 1195.0, + 532.0, + 1155.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 509.0, + 1250.0, + 509.0, + 1250.0, + 532.0, + 1209.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 511.0, + 1303.0, + 511.0, + 1303.0, + 531.0, + 1257.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 510.0, + 1354.0, + 510.0, + 1354.0, + 531.0, + 1318.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 525.0, + 1174.0, + 525.0, + 1174.0, + 548.0, + 1113.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 564.0, + 1404.0, + 564.0, + 1404.0, + 604.0, + 293.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 596.0, + 1404.0, + 596.0, + 1404.0, + 634.0, + 294.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 625.0, + 1404.0, + 625.0, + 1404.0, + 665.0, + 293.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 655.0, + 1404.0, + 655.0, + 1404.0, + 695.0, + 293.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 687.0, + 856.0, + 687.0, + 856.0, + 725.0, + 295.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1968.0, + 439.0, + 1968.0, + 439.0, + 2016.0, + 328.0, + 2016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 1968.0, + 630.0, + 1968.0, + 630.0, + 2016.0, + 547.0, + 2016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 1968.0, + 1408.0, + 1968.0, + 1408.0, + 2016.0, + 656.0, + 2016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 2005.0, + 1139.0, + 2005.0, + 1139.0, + 2036.0, + 297.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1609.0, + 1366.0, + 1609.0, + 1366.0, + 1647.0, + 295.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1270.0, + 1406.0, + 1270.0, + 1406.0, + 1305.0, + 294.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1302.0, + 652.0, + 1302.0, + 652.0, + 1335.0, + 293.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1302.0, + 1403.0, + 1302.0, + 1403.0, + 1335.0, + 678.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1331.0, + 870.0, + 1331.0, + 870.0, + 1365.0, + 294.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 1331.0, + 1405.0, + 1331.0, + 1405.0, + 1365.0, + 965.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1360.0, + 1406.0, + 1360.0, + 1406.0, + 1395.0, + 293.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1391.0, + 1407.0, + 1391.0, + 1407.0, + 1426.0, + 293.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1423.0, + 1405.0, + 1423.0, + 1405.0, + 1456.0, + 293.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1453.0, + 331.0, + 1453.0, + 331.0, + 1489.0, + 294.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1453.0, + 814.0, + 1453.0, + 814.0, + 1489.0, + 621.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 1453.0, + 1405.0, + 1453.0, + 1405.0, + 1489.0, + 1051.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1482.0, + 1405.0, + 1482.0, + 1405.0, + 1519.0, + 291.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1513.0, + 1049.0, + 1513.0, + 1049.0, + 1550.0, + 291.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 1513.0, + 1406.0, + 1513.0, + 1406.0, + 1550.0, + 1075.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1545.0, + 1363.0, + 1545.0, + 1363.0, + 1580.0, + 293.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1887.0, + 371.0, + 1887.0, + 371.0, + 1930.0, + 293.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 1887.0, + 967.0, + 1887.0, + 967.0, + 1930.0, + 471.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 1887.0, + 1407.0, + 1887.0, + 1407.0, + 1930.0, + 982.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1921.0, + 297.0, + 1921.0, + 297.0, + 1957.0, + 294.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1921.0, + 615.0, + 1921.0, + 615.0, + 1957.0, + 325.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 1921.0, + 1407.0, + 1921.0, + 1407.0, + 1957.0, + 631.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1669.0, + 1225.0, + 1669.0, + 1225.0, + 1705.0, + 295.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 1669.0, + 1400.0, + 1669.0, + 1400.0, + 1705.0, + 1296.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1698.0, + 913.0, + 1698.0, + 913.0, + 1738.0, + 293.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1609.0, + 1366.0, + 1609.0, + 1366.0, + 1647.0, + 295.0, + 1647.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1412, + 1404, + 1412, + 1404, + 1752, + 297, + 1752 + ], + "score": 0.986 + }, + { + "category_id": 1, + "poly": [ + 298, + 1849, + 1403, + 1849, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.981 + }, + { + "category_id": 5, + "poly": [ + 296, + 216, + 1446, + 216, + 1446, + 438, + 296, + 438 + ], + "score": 0.972, + "html": "
ModelDropout rateUnlabeled dataValidationTest
Traditional dropout0.1No78.4 (± 0.25)76.9 (± 0.31)
No dropout0.0No78.8 (± 0.59)77.1 (± 0.3)
Fraternal dropout (k = 0.05)0.05No79.3 (± 0.38)77.6 (± 0.35)
Traditional dropout + I-model0.1Yes80.2 (± 0.33)78.5 (± 0.46)
Fraternal dropout (k = 0.15)0.1Yes80.5 (± 0.18)79.1 (± 0.37)
" + }, + { + "category_id": 4, + "poly": [ + 297, + 978, + 1405, + 978, + 1405, + 1133, + 297, + 1133 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 291, + 1190, + 1398, + 1190, + 1398, + 1255, + 291, + 1255 + ], + "score": 0.939 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 105, + 299, + 105 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 303, + 1793, + 803, + 1793, + 803, + 1822, + 303, + 1822 + ], + "score": 0.889 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1361, + 1399, + 1361, + 1399, + 1388, + 1366, + 1388 + ], + "score": 0.88 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1278, + 1399, + 1278, + 1399, + 1307, + 1366, + 1307 + ], + "score": 0.871 + }, + { + "category_id": 3, + "poly": [ + 318, + 652, + 1354, + 652, + 1354, + 959, + 318, + 959 + ], + "score": 0.825 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.798 + }, + { + "category_id": 1, + "poly": [ + 296, + 447, + 1405, + 447, + 1405, + 602, + 296, + 602 + ], + "score": 0.748 + }, + { + "category_id": 8, + "poly": [ + 404, + 1318, + 1369, + 1318, + 1369, + 1362, + 404, + 1362 + ], + "score": 0.567 + }, + { + "category_id": 8, + "poly": [ + 396, + 1269, + 1374, + 1269, + 1374, + 1362, + 396, + 1362 + ], + "score": 0.522 + }, + { + "category_id": 6, + "poly": [ + 296, + 447, + 1405, + 447, + 1405, + 602, + 296, + 602 + ], + "score": 0.107 + }, + { + "category_id": 13, + "poly": [ + 1105, + 1191, + 1242, + 1191, + 1242, + 1227, + 1105, + 1227 + ], + "score": 0.93, + "latex": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } )" + }, + { + "category_id": 14, + "poly": [ + 394, + 1268, + 1382, + 1268, + 1382, + 1365, + 394, + 1365 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) = \\mathbb { E } _ { s _ { i } , s _ { j } } \\left[ | | \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) - \\mathbf { p } ^ { t } ( \\mathbf { h } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } \\right] } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) ^ { T } \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } + | | \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } - 2 \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) ^ { T } \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) \\Big ] } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad = \\mathbb { E } _ { s _ { i } , s _ { j } } \\left[ | | \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } + | | \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) | | _ { 2 } ^ { 2 } - 2 \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) ^ { T } \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) \\right] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 298, + 1567, + 474, + 1567, + 474, + 1601, + 298, + 1601 + ], + "score": 0.91, + "latex": "\\| \\mathbf { p } ^ { t } ( \\bar { \\mathbf { z } } ^ { t } , s _ { i } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1536, + 573, + 1536, + 573, + 1568, + 297, + 1568 + ], + "score": 0.89, + "latex": "\\kappa \\in \\{ 0 . 0 5 , 0 . 1 , \\ldots , 0 . 4 \\}" + }, + { + "category_id": 13, + "poly": [ + 1094, + 1475, + 1315, + 1475, + 1315, + 1509, + 1094, + 1509 + ], + "score": 0.88, + "latex": "\\beta \\in \\{ 1 , 2 , \\ldots , 1 2 \\}" + }, + { + "category_id": 13, + "poly": [ + 858, + 1476, + 1079, + 1476, + 1079, + 1508, + 858, + 1508 + ], + "score": 0.88, + "latex": "\\alpha \\in \\{ 1 , 2 , \\ldots , 1 2 \\}" + }, + { + "category_id": 13, + "poly": [ + 840, + 1444, + 868, + 1444, + 868, + 1473, + 840, + 1473 + ], + "score": 0.85, + "latex": "\\bar { \\ell } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 530, + 1413, + 557, + 1413, + 557, + 1442, + 530, + 1442 + ], + "score": 0.85, + "latex": "\\ell ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 912, + 1565, + 939, + 1565, + 939, + 1594, + 912, + 1594 + ], + "score": 0.84, + "latex": "\\ell ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1109, + 705, + 1155, + 705, + 1155, + 724, + 1109, + 724 + ], + "score": 0.8, + "latex": "\\beta { = } 4 . 0" + }, + { + "category_id": 13, + "poly": [ + 551, + 705, + 596, + 705, + 596, + 724, + 551, + 724 + ], + "score": 0.79, + "latex": "\\beta { = } 4 . 0" + }, + { + "category_id": 13, + "poly": [ + 1100, + 769, + 1153, + 769, + 1153, + 788, + 1100, + 788 + ], + "score": 0.77, + "latex": "\\kappa { = } 0 . 1 5" + }, + { + "category_id": 13, + "poly": [ + 543, + 749, + 588, + 749, + 588, + 767, + 543, + 767 + ], + "score": 0.77, + "latex": "\\alpha { = } 7 . 0" + }, + { + "category_id": 13, + "poly": [ + 1100, + 727, + 1145, + 727, + 1145, + 745, + 1100, + 745 + ], + "score": 0.77, + "latex": "\\kappa { = } 0 . 1 ," + }, + { + "category_id": 13, + "poly": [ + 1101, + 749, + 1146, + 749, + 1146, + 766, + 1101, + 766 + ], + "score": 0.76, + "latex": "\\alpha { = } 7 . 0" + }, + { + "category_id": 13, + "poly": [ + 542, + 771, + 595, + 771, + 595, + 790, + 542, + 790 + ], + "score": 0.76, + "latex": "\\kappa { = } 0 . 1 5" + }, + { + "category_id": 13, + "poly": [ + 535, + 376, + 647, + 376, + 647, + 405, + 535, + 405 + ], + "score": 0.74, + "latex": "( \\kappa = 0 . 1 5 )" + }, + { + "category_id": 13, + "poly": [ + 542, + 728, + 586, + 728, + 586, + 745, + 542, + 745 + ], + "score": 0.72, + "latex": "\\kappa { = } 0 . 1" + }, + { + "category_id": 13, + "poly": [ + 535, + 350, + 557, + 350, + 557, + 372, + 535, + 372 + ], + "score": 0.71, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 535, + 313, + 648, + 313, + 648, + 343, + 535, + 343 + ], + "score": 0.7, + "latex": "( \\kappa = 0 . 0 5 )" + }, + { + "category_id": 13, + "poly": [ + 1034, + 479, + 1058, + 479, + 1058, + 506, + 1034, + 506 + ], + "score": 0.32, + "latex": "\\Pi" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 978.0, + 1405.0, + 978.0, + 1405.0, + 1016.0, + 295.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1009.0, + 1405.0, + 1009.0, + 1405.0, + 1046.0, + 295.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1036.0, + 1405.0, + 1036.0, + 1405.0, + 1081.0, + 292.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1067.0, + 1406.0, + 1067.0, + 1406.0, + 1107.0, + 294.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1102.0, + 947.0, + 1102.0, + 947.0, + 1139.0, + 295.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 71.0, + 817.0, + 71.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1792.0, + 806.0, + 1792.0, + 806.0, + 1827.0, + 296.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 655.0, + 774.0, + 655.0, + 774.0, + 677.0, + 365.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 655.0, + 1355.0, + 655.0, + 1355.0, + 677.0, + 931.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 674.0, + 370.0, + 674.0, + 370.0, + 699.0, + 330.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 682.0, + 750.0, + 682.0, + 750.0, + 704.0, + 512.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 674.0, + 941.0, + 674.0, + 941.0, + 699.0, + 902.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 682.0, + 1322.0, + 682.0, + 1322.0, + 704.0, + 1070.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 710.0, + 370.0, + 710.0, + 370.0, + 737.0, + 339.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 701.0, + 550.0, + 701.0, + 550.0, + 727.0, + 513.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 701.0, + 770.0, + 701.0, + 770.0, + 727.0, + 597.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 708.0, + 942.0, + 708.0, + 942.0, + 735.0, + 911.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 703.0, + 1108.0, + 703.0, + 1108.0, + 725.0, + 1071.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 703.0, + 1341.0, + 703.0, + 1341.0, + 725.0, + 1156.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 723.0, + 541.0, + 723.0, + 541.0, + 749.0, + 512.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 723.0, + 759.0, + 723.0, + 759.0, + 749.0, + 587.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 725.0, + 1099.0, + 725.0, + 1099.0, + 747.0, + 1070.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 725.0, + 1331.0, + 725.0, + 1331.0, + 747.0, + 1146.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 749.0, + 369.0, + 749.0, + 369.0, + 772.0, + 340.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 745.0, + 542.0, + 745.0, + 542.0, + 770.0, + 509.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 745.0, + 763.0, + 745.0, + 763.0, + 770.0, + 589.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 743.0, + 942.0, + 743.0, + 942.0, + 770.0, + 912.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 747.0, + 1100.0, + 747.0, + 1100.0, + 768.0, + 1070.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 747.0, + 1334.0, + 747.0, + 1334.0, + 768.0, + 1147.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 760.0, + 344.0, + 760.0, + 344.0, + 844.0, + 317.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 766.0, + 541.0, + 766.0, + 541.0, + 792.0, + 512.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 766.0, + 769.0, + 766.0, + 769.0, + 792.0, + 596.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 759.0, + 916.0, + 759.0, + 916.0, + 844.0, + 888.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 767.0, + 1099.0, + 767.0, + 1099.0, + 789.0, + 1070.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 767.0, + 1340.0, + 767.0, + 1340.0, + 789.0, + 1154.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 786.0, + 367.0, + 786.0, + 367.0, + 809.0, + 341.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 780.0, + 940.0, + 780.0, + 940.0, + 803.0, + 912.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 823.0, + 366.0, + 823.0, + 366.0, + 846.0, + 340.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 815.0, + 940.0, + 815.0, + 940.0, + 838.0, + 912.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 858.0, + 370.0, + 858.0, + 370.0, + 885.0, + 339.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 847.0, + 941.0, + 847.0, + 941.0, + 875.0, + 912.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 844.0, + 1049.0, + 844.0, + 1049.0, + 860.0, + 1025.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 867.0, + 1105.0, + 867.0, + 1105.0, + 880.0, + 1083.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 883.0, + 942.0, + 883.0, + 942.0, + 910.0, + 911.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 896.0, + 370.0, + 896.0, + 370.0, + 922.0, + 339.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 923.0, + 395.0, + 923.0, + 395.0, + 944.0, + 374.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 921.0, + 508.0, + 921.0, + 508.0, + 948.0, + 415.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 923.0, + 558.0, + 923.0, + 558.0, + 944.0, + 517.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 922.0, + 610.0, + 922.0, + 610.0, + 947.0, + 570.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 922.0, + 664.0, + 922.0, + 664.0, + 947.0, + 622.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 922.0, + 716.0, + 922.0, + 716.0, + 945.0, + 672.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 923.0, + 763.0, + 923.0, + 763.0, + 944.0, + 724.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 923.0, + 969.0, + 923.0, + 969.0, + 945.0, + 946.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 923.0, + 1029.0, + 923.0, + 1029.0, + 944.0, + 989.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 922.0, + 1080.0, + 922.0, + 1080.0, + 947.0, + 1036.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 921.0, + 1132.0, + 921.0, + 1132.0, + 947.0, + 1090.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 919.0, + 1183.0, + 919.0, + 1183.0, + 947.0, + 1143.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 922.0, + 1237.0, + 922.0, + 1237.0, + 947.0, + 1195.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 922.0, + 1337.0, + 922.0, + 1337.0, + 945.0, + 1244.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 936.0, + 606.0, + 936.0, + 606.0, + 963.0, + 538.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 936.0, + 1179.0, + 936.0, + 1179.0, + 963.0, + 1111.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.25, + 821.0, + 1110.25, + 821.0, + 1110.25, + 853.0, + 1027.25, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 445.0, + 1406.0, + 445.0, + 1406.0, + 483.0, + 294.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 477.0, + 1033.0, + 477.0, + 1033.0, + 514.0, + 295.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 477.0, + 1406.0, + 477.0, + 1406.0, + 514.0, + 1059.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 510.0, + 1406.0, + 510.0, + 1406.0, + 543.0, + 294.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 539.0, + 1403.0, + 539.0, + 1403.0, + 572.0, + 294.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 570.0, + 543.0, + 570.0, + 543.0, + 605.0, + 292.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1412.0, + 529.0, + 1412.0, + 529.0, + 1451.0, + 294.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 1412.0, + 1405.0, + 1412.0, + 1405.0, + 1451.0, + 558.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1443.0, + 839.0, + 1443.0, + 839.0, + 1482.0, + 292.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1443.0, + 1407.0, + 1443.0, + 1407.0, + 1482.0, + 869.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1473.0, + 857.0, + 1473.0, + 857.0, + 1512.0, + 292.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 1473.0, + 1093.0, + 1473.0, + 1093.0, + 1512.0, + 1080.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 1473.0, + 1406.0, + 1473.0, + 1406.0, + 1512.0, + 1316.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1506.0, + 1405.0, + 1506.0, + 1405.0, + 1541.0, + 294.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1534.0, + 296.0, + 1534.0, + 296.0, + 1572.0, + 291.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 1534.0, + 1405.0, + 1534.0, + 1405.0, + 1572.0, + 574.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1561.0, + 297.0, + 1561.0, + 297.0, + 1606.0, + 292.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 1561.0, + 911.0, + 1561.0, + 911.0, + 1606.0, + 475.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 1561.0, + 1408.0, + 1561.0, + 1408.0, + 1606.0, + 940.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1594.0, + 1407.0, + 1594.0, + 1407.0, + 1635.0, + 292.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1626.0, + 1405.0, + 1626.0, + 1405.0, + 1663.0, + 294.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1658.0, + 1402.0, + 1658.0, + 1402.0, + 1693.0, + 294.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1690.0, + 1404.0, + 1690.0, + 1404.0, + 1724.0, + 295.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1720.0, + 691.0, + 1720.0, + 691.0, + 1754.0, + 295.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1846.0, + 1405.0, + 1846.0, + 1405.0, + 1887.0, + 295.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1883.0, + 1404.0, + 1883.0, + 1404.0, + 1915.0, + 296.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1910.0, + 1403.0, + 1910.0, + 1403.0, + 1948.0, + 293.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1944.0, + 1403.0, + 1944.0, + 1403.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1408.0, + 1971.0, + 1408.0, + 2010.0, + 293.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2003.0, + 1050.0, + 2003.0, + 1050.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1188.0, + 1104.0, + 1188.0, + 1104.0, + 1228.0, + 294.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 1188.0, + 1401.0, + 1188.0, + 1401.0, + 1228.0, + 1243.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1223.0, + 932.0, + 1223.0, + 932.0, + 1257.0, + 293.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 445.0, + 1406.0, + 445.0, + 1406.0, + 483.0, + 294.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 477.0, + 1033.0, + 477.0, + 1033.0, + 514.0, + 295.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 477.0, + 1406.0, + 477.0, + 1406.0, + 514.0, + 1059.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 510.0, + 1406.0, + 510.0, + 1406.0, + 543.0, + 294.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 539.0, + 1403.0, + 539.0, + 1403.0, + 572.0, + 294.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 570.0, + 543.0, + 570.0, + 543.0, + 605.0, + 292.0, + 605.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 444, + 218, + 1254, + 218, + 1254, + 443, + 444, + 443 + ], + "score": 0.98, + "html": "
PTBWT2
DropoutFine-tuningValidationTestValidationTest
TraditionalNone60.758.869.166.0
TraditionalOne60.057.368.665.8
FraternalNone59.858.068.365.3
FraternalOne58.956.866.864.1
FraternalTwo58.556.211
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1819, + 1403, + 1819, + 1403, + 2035, + 297, + 2035 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1487, + 1404, + 1487, + 1404, + 1674, + 298, + 1674 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1350, + 1403, + 1350, + 1403, + 1472, + 298, + 1472 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1211, + 1404, + 1211, + 1404, + 1334, + 298, + 1334 + ], + "score": 0.975 + }, + { + "category_id": 5, + "poly": [ + 523, + 555, + 1172, + 555, + 1172, + 718, + 523, + 718 + ], + "score": 0.974, + "html": "
Hyper-parameterPossible values
batch size non-monotone interval[10,20,30, 40] [5,10,20,40,60,100]
K -FD or ELD strength weight decayU(0, 0.3) U(0.6 × 10-6,2.4 × 10-6)
" + }, + { + "category_id": 5, + "poly": [ + 302, + 868, + 1395, + 868, + 1395, + 967, + 302, + 967 + ], + "score": 0.968, + "html": "
ModelBestTop5 avgTop10 avgBeating baseline runs (out of)
Expectation lineardropout59.460.160.56(208)
Fraternal dropout59.459.659.914 (203)
" + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 815, + 75, + 815, + 105, + 298, + 105 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 299, + 1738, + 543, + 1738, + 543, + 1773, + 299, + 1773 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.793 + }, + { + "category_id": 6, + "poly": [ + 297, + 980, + 1402, + 980, + 1402, + 1074, + 297, + 1074 + ], + "score": 0.763 + }, + { + "category_id": 6, + "poly": [ + 288, + 453, + 1402, + 453, + 1402, + 517, + 288, + 517 + ], + "score": 0.667 + }, + { + "category_id": 6, + "poly": [ + 296, + 725, + 1402, + 725, + 1402, + 819, + 296, + 819 + ], + "score": 0.641 + }, + { + "category_id": 0, + "poly": [ + 298, + 1148, + 1236, + 1148, + 1236, + 1177, + 298, + 1177 + ], + "score": 0.599 + }, + { + "category_id": 1, + "poly": [ + 298, + 1148, + 1236, + 1148, + 1236, + 1177, + 298, + 1177 + ], + "score": 0.229 + }, + { + "category_id": 1, + "poly": [ + 297, + 980, + 1402, + 980, + 1402, + 1074, + 297, + 1074 + ], + "score": 0.129 + }, + { + "category_id": 13, + "poly": [ + 853, + 756, + 937, + 756, + 937, + 788, + 853, + 788 + ], + "score": 0.93, + "latex": "U ( a , b )" + }, + { + "category_id": 13, + "poly": [ + 298, + 786, + 354, + 786, + 354, + 818, + 298, + 818 + ], + "score": 0.9, + "latex": "[ a , b ]" + }, + { + "category_id": 13, + "poly": [ + 943, + 651, + 1052, + 651, + 1052, + 682, + 943, + 682 + ], + "score": 0.56, + "latex": "U ( 0 , 0 . 3 )" + }, + { + "category_id": 13, + "poly": [ + 836, + 681, + 1152, + 681, + 1152, + 716, + 836, + 716 + ], + "score": 0.51, + "latex": "U ( 0 . 6 \\times 1 0 ^ { - 6 } , 2 . 4 \\times 1 0 ^ { - 6 } )" + }, + { + "category_id": 13, + "poly": [ + 540, + 651, + 623, + 651, + 623, + 680, + 540, + 680 + ], + "score": 0.48, + "latex": "\\kappa - \\mathrm { F D }" + }, + { + "category_id": 14, + "poly": [ + 835, + 681, + 1152, + 681, + 1152, + 716, + 835, + 716 + ], + "score": 0.43, + "latex": "U ( 0 . 6 \\times 1 0 ^ { - 6 } , 2 . 4 \\times 1 0 ^ { - 6 } )" + }, + { + "category_id": 13, + "poly": [ + 861, + 620, + 1129, + 620, + 1129, + 650, + 861, + 650 + ], + "score": 0.29, + "latex": "[ 5 , 1 0 , 2 0 , 4 0 , 6 0 , 1 0 0 ]" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1734.0, + 548.0, + 1734.0, + 548.0, + 1782.0, + 292.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 976.0, + 1406.0, + 976.0, + 1406.0, + 1016.0, + 292.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1009.0, + 1404.0, + 1009.0, + 1404.0, + 1045.0, + 293.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1040.0, + 1402.0, + 1040.0, + 1402.0, + 1075.0, + 292.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 450.0, + 1404.0, + 450.0, + 1404.0, + 489.0, + 295.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 482.0, + 977.0, + 482.0, + 977.0, + 518.0, + 295.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 722.0, + 1406.0, + 722.0, + 1406.0, + 764.0, + 293.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 756.0, + 852.0, + 756.0, + 852.0, + 790.0, + 292.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 756.0, + 1404.0, + 756.0, + 1404.0, + 790.0, + 938.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 784.0, + 297.0, + 784.0, + 297.0, + 823.0, + 294.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 784.0, + 1012.0, + 784.0, + 1012.0, + 823.0, + 355.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1146.0, + 1242.0, + 1146.0, + 1242.0, + 1180.0, + 295.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1820.0, + 1407.0, + 1820.0, + 1407.0, + 1858.0, + 293.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1850.0, + 1405.0, + 1850.0, + 1405.0, + 1885.0, + 292.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1878.0, + 1405.0, + 1878.0, + 1405.0, + 1920.0, + 292.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1947.0, + 295.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1978.0, + 292.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1970.0, + 1403.0, + 1970.0, + 1403.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2004.0, + 507.0, + 2004.0, + 507.0, + 2039.0, + 295.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1486.0, + 1406.0, + 1486.0, + 1406.0, + 1523.0, + 293.0, + 1523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1516.0, + 1406.0, + 1516.0, + 1406.0, + 1556.0, + 293.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1546.0, + 1406.0, + 1546.0, + 1406.0, + 1586.0, + 293.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1582.0, + 1405.0, + 1582.0, + 1405.0, + 1614.0, + 296.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1609.0, + 1405.0, + 1609.0, + 1405.0, + 1646.0, + 293.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1642.0, + 1238.0, + 1642.0, + 1238.0, + 1677.0, + 293.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1351.0, + 1405.0, + 1351.0, + 1405.0, + 1383.0, + 294.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1382.0, + 1405.0, + 1382.0, + 1405.0, + 1414.0, + 294.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1411.0, + 1407.0, + 1411.0, + 1407.0, + 1447.0, + 296.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1442.0, + 765.0, + 1442.0, + 765.0, + 1475.0, + 296.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1212.0, + 1405.0, + 1212.0, + 1405.0, + 1245.0, + 296.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1240.0, + 1404.0, + 1240.0, + 1404.0, + 1278.0, + 293.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1272.0, + 1405.0, + 1272.0, + 1405.0, + 1307.0, + 293.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1304.0, + 649.0, + 1304.0, + 649.0, + 1336.0, + 294.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1146.0, + 1242.0, + 1146.0, + 1242.0, + 1180.0, + 295.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 976.0, + 1406.0, + 976.0, + 1406.0, + 1016.0, + 292.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1009.0, + 1404.0, + 1009.0, + 1404.0, + 1045.0, + 293.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1040.0, + 1402.0, + 1040.0, + 1402.0, + 1075.0, + 292.0, + 1075.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 295, + 1404, + 295, + 1404, + 417, + 299, + 417 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 298, + 1175, + 1404, + 1175, + 1404, + 1299, + 298, + 1299 + ], + "score": 0.917 + }, + { + "category_id": 1, + "poly": [ + 298, + 1002, + 1405, + 1002, + 1405, + 1158, + 298, + 1158 + ], + "score": 0.903 + }, + { + "category_id": 1, + "poly": [ + 298, + 840, + 1401, + 840, + 1401, + 902, + 298, + 902 + ], + "score": 0.9 + }, + { + "category_id": 1, + "poly": [ + 299, + 514, + 1401, + 514, + 1401, + 578, + 299, + 578 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 335, + 2005, + 557, + 2005, + 557, + 2034, + 335, + 2034 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.883 + }, + { + "category_id": 1, + "poly": [ + 302, + 594, + 1398, + 594, + 1398, + 660, + 302, + 660 + ], + "score": 0.879 + }, + { + "category_id": 1, + "poly": [ + 300, + 758, + 1397, + 758, + 1397, + 822, + 300, + 822 + ], + "score": 0.873 + }, + { + "category_id": 1, + "poly": [ + 296, + 921, + 1401, + 921, + 1401, + 986, + 296, + 986 + ], + "score": 0.871 + }, + { + "category_id": 1, + "poly": [ + 296, + 1318, + 1401, + 1318, + 1401, + 1382, + 296, + 1382 + ], + "score": 0.858 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 866, + 2088, + 866, + 2113, + 836, + 2113 + ], + "score": 0.847 + }, + { + "category_id": 1, + "poly": [ + 300, + 677, + 1395, + 677, + 1395, + 741, + 300, + 741 + ], + "score": 0.845 + }, + { + "category_id": 0, + "poly": [ + 299, + 464, + 489, + 464, + 489, + 497, + 299, + 497 + ], + "score": 0.841 + }, + { + "category_id": 0, + "poly": [ + 302, + 229, + 625, + 229, + 625, + 261, + 302, + 261 + ], + "score": 0.834 + }, + { + "category_id": 1, + "poly": [ + 297, + 1399, + 1406, + 1399, + 1406, + 1495, + 297, + 1495 + ], + "score": 0.809 + }, + { + "category_id": 1, + "poly": [ + 298, + 1674, + 1403, + 1674, + 1403, + 1739, + 298, + 1739 + ], + "score": 0.809 + }, + { + "category_id": 1, + "poly": [ + 299, + 1593, + 1401, + 1593, + 1401, + 1658, + 299, + 1658 + ], + "score": 0.805 + }, + { + "category_id": 1, + "poly": [ + 299, + 1512, + 1402, + 1512, + 1402, + 1576, + 299, + 1576 + ], + "score": 0.8 + }, + { + "category_id": 1, + "poly": [ + 298, + 1756, + 1402, + 1756, + 1402, + 1821, + 298, + 1821 + ], + "score": 0.771 + }, + { + "category_id": 1, + "poly": [ + 300, + 1838, + 1400, + 1838, + 1400, + 1902, + 300, + 1902 + ], + "score": 0.77 + }, + { + "category_id": 1, + "poly": [ + 299, + 1920, + 1400, + 1920, + 1400, + 1984, + 299, + 1984 + ], + "score": 0.72 + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1998.0, + 561.0, + 1998.0, + 561.0, + 2039.0, + 329.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 464.0, + 490.0, + 464.0, + 490.0, + 501.0, + 297.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 224.0, + 630.0, + 224.0, + 630.0, + 269.0, + 298.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 295.0, + 1405.0, + 295.0, + 1405.0, + 331.0, + 294.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 326.0, + 1405.0, + 326.0, + 1405.0, + 361.0, + 293.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 357.0, + 1404.0, + 357.0, + 1404.0, + 390.0, + 294.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 385.0, + 1100.0, + 385.0, + 1100.0, + 420.0, + 294.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1177.0, + 1402.0, + 1177.0, + 1402.0, + 1210.0, + 296.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1208.0, + 1402.0, + 1208.0, + 1402.0, + 1241.0, + 323.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1235.0, + 1406.0, + 1235.0, + 1406.0, + 1276.0, + 318.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1266.0, + 396.0, + 1266.0, + 396.0, + 1302.0, + 320.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1003.0, + 1406.0, + 1003.0, + 1406.0, + 1040.0, + 295.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1031.0, + 1406.0, + 1031.0, + 1406.0, + 1071.0, + 321.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1063.0, + 1405.0, + 1063.0, + 1405.0, + 1101.0, + 321.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1095.0, + 1405.0, + 1095.0, + 1405.0, + 1132.0, + 321.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1125.0, + 530.0, + 1125.0, + 530.0, + 1160.0, + 321.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 837.0, + 1405.0, + 837.0, + 1405.0, + 879.0, + 293.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 871.0, + 532.0, + 871.0, + 532.0, + 903.0, + 324.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 514.0, + 1406.0, + 514.0, + 1406.0, + 550.0, + 295.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 546.0, + 1383.0, + 546.0, + 1383.0, + 579.0, + 322.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 595.0, + 1403.0, + 595.0, + 1403.0, + 631.0, + 298.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 624.0, + 1350.0, + 624.0, + 1350.0, + 661.0, + 323.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 753.0, + 1403.0, + 753.0, + 1403.0, + 800.0, + 292.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 789.0, + 1340.0, + 789.0, + 1340.0, + 825.0, + 322.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 921.0, + 1405.0, + 921.0, + 1405.0, + 957.0, + 295.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 952.0, + 1219.0, + 952.0, + 1219.0, + 987.0, + 321.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1314.0, + 1406.0, + 1314.0, + 1406.0, + 1356.0, + 294.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1350.0, + 610.0, + 1350.0, + 610.0, + 1381.0, + 322.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 677.0, + 1401.0, + 677.0, + 1401.0, + 713.0, + 297.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 706.0, + 1380.0, + 706.0, + 1380.0, + 746.0, + 321.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1399.0, + 1404.0, + 1399.0, + 1404.0, + 1437.0, + 296.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1432.0, + 1404.0, + 1432.0, + 1404.0, + 1467.0, + 324.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1463.0, + 976.0, + 1463.0, + 976.0, + 1497.0, + 325.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1677.0, + 1404.0, + 1677.0, + 1404.0, + 1709.0, + 296.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1705.0, + 1306.0, + 1705.0, + 1306.0, + 1741.0, + 322.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1591.0, + 1403.0, + 1591.0, + 1403.0, + 1631.0, + 294.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1626.0, + 999.0, + 1626.0, + 999.0, + 1659.0, + 322.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1511.0, + 1403.0, + 1511.0, + 1403.0, + 1549.0, + 294.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1544.0, + 1151.0, + 1544.0, + 1151.0, + 1577.0, + 323.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1750.0, + 1406.0, + 1750.0, + 1406.0, + 1796.0, + 293.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1789.0, + 872.0, + 1789.0, + 872.0, + 1821.0, + 324.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1838.0, + 1405.0, + 1838.0, + 1405.0, + 1874.0, + 296.0, + 1874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1869.0, + 1355.0, + 1869.0, + 1355.0, + 1903.0, + 321.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1920.0, + 1402.0, + 1920.0, + 1402.0, + 1956.0, + 297.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1952.0, + 958.0, + 1952.0, + 958.0, + 1985.0, + 321.0, + 1985.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.833 + }, + { + "category_id": 1, + "poly": [ + 289, + 1438, + 1401, + 1438, + 1401, + 1504, + 289, + 1504 + ], + "score": 0.55 + }, + { + "category_id": 1, + "poly": [ + 291, + 1354, + 1400, + 1354, + 1400, + 1420, + 291, + 1420 + ], + "score": 0.54 + }, + { + "category_id": 1, + "poly": [ + 296, + 1097, + 1402, + 1097, + 1402, + 1193, + 296, + 1193 + ], + "score": 0.517 + }, + { + "category_id": 1, + "poly": [ + 296, + 985, + 1401, + 985, + 1401, + 1078, + 296, + 1078 + ], + "score": 0.478 + }, + { + "category_id": 1, + "poly": [ + 296, + 870, + 1400, + 870, + 1400, + 966, + 296, + 966 + ], + "score": 0.437 + }, + { + "category_id": 1, + "poly": [ + 295, + 477, + 1401, + 477, + 1401, + 543, + 295, + 543 + ], + "score": 0.435 + }, + { + "category_id": 1, + "poly": [ + 296, + 228, + 1402, + 228, + 1402, + 294, + 296, + 294 + ], + "score": 0.42 + }, + { + "category_id": 1, + "poly": [ + 298, + 311, + 1400, + 311, + 1400, + 378, + 298, + 378 + ], + "score": 0.413 + }, + { + "category_id": 1, + "poly": [ + 296, + 560, + 1400, + 560, + 1400, + 655, + 296, + 655 + ], + "score": 0.41 + }, + { + "category_id": 1, + "poly": [ + 294, + 395, + 1399, + 395, + 1399, + 460, + 294, + 460 + ], + "score": 0.4 + }, + { + "category_id": 1, + "poly": [ + 299, + 675, + 1398, + 675, + 1398, + 770, + 299, + 770 + ], + "score": 0.393 + }, + { + "category_id": 1, + "poly": [ + 293, + 787, + 1397, + 787, + 1397, + 854, + 293, + 854 + ], + "score": 0.387 + }, + { + "category_id": 1, + "poly": [ + 297, + 1210, + 1403, + 1210, + 1403, + 1336, + 297, + 1336 + ], + "score": 0.315 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1439.0, + 1404.0, + 1439.0, + 1404.0, + 1475.0, + 293.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1470.0, + 995.0, + 1470.0, + 995.0, + 1506.0, + 322.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1355.0, + 1402.0, + 1355.0, + 1402.0, + 1391.0, + 296.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1388.0, + 761.0, + 1388.0, + 761.0, + 1418.0, + 323.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1096.0, + 1406.0, + 1096.0, + 1406.0, + 1135.0, + 291.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1130.0, + 1404.0, + 1130.0, + 1404.0, + 1165.0, + 323.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1161.0, + 811.0, + 1161.0, + 811.0, + 1195.0, + 324.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 980.0, + 1406.0, + 980.0, + 1406.0, + 1024.0, + 294.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1015.0, + 1400.0, + 1015.0, + 1400.0, + 1051.0, + 321.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1042.0, + 411.0, + 1042.0, + 411.0, + 1081.0, + 322.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 870.0, + 1404.0, + 870.0, + 1404.0, + 907.0, + 294.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 901.0, + 1405.0, + 901.0, + 1405.0, + 941.0, + 320.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 934.0, + 709.0, + 934.0, + 709.0, + 968.0, + 323.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 476.0, + 1405.0, + 476.0, + 1405.0, + 515.0, + 295.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 510.0, + 609.0, + 510.0, + 609.0, + 541.0, + 323.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 226.0, + 1404.0, + 226.0, + 1404.0, + 267.0, + 295.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 262.0, + 928.0, + 262.0, + 928.0, + 296.0, + 320.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 309.0, + 1404.0, + 309.0, + 1404.0, + 350.0, + 294.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 343.0, + 1195.0, + 343.0, + 1195.0, + 380.0, + 323.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 561.0, + 1402.0, + 561.0, + 1402.0, + 595.0, + 295.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 592.0, + 1401.0, + 592.0, + 1401.0, + 629.0, + 323.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 625.0, + 638.0, + 625.0, + 638.0, + 656.0, + 322.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 390.0, + 1403.0, + 390.0, + 1403.0, + 437.0, + 293.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 426.0, + 1067.0, + 426.0, + 1067.0, + 462.0, + 323.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 676.0, + 1402.0, + 676.0, + 1402.0, + 710.0, + 295.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 709.0, + 1402.0, + 709.0, + 1402.0, + 740.0, + 324.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 737.0, + 805.0, + 737.0, + 805.0, + 771.0, + 324.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 788.0, + 1403.0, + 788.0, + 1403.0, + 824.0, + 294.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 819.0, + 1180.0, + 819.0, + 1180.0, + 854.0, + 322.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1211.0, + 1407.0, + 1211.0, + 1407.0, + 1247.0, + 295.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1240.0, + 1404.0, + 1240.0, + 1404.0, + 1279.0, + 322.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1274.0, + 1401.0, + 1274.0, + 1401.0, + 1307.0, + 324.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1304.0, + 423.0, + 1304.0, + 423.0, + 1335.0, + 322.0, + 1335.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 630, + 1404, + 630, + 1404, + 844, + 298, + 844 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 353, + 1405, + 353, + 1405, + 567, + 298, + 567 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1091, + 1403, + 1091, + 1403, + 1307, + 298, + 1307 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1428, + 1405, + 1428, + 1405, + 1675, + 297, + 1675 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1322, + 1403, + 1322, + 1403, + 1414, + 299, + 1414 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 984, + 1403, + 984, + 1403, + 1076, + 299, + 1076 + ], + "score": 0.974 + }, + { + "category_id": 5, + "poly": [ + 525, + 1697, + 1175, + 1697, + 1175, + 1917, + 525, + 1917 + ], + "score": 0.967, + "html": "
ModelValidationTest
MC1 MC10 MC50 Baseline (average mask) Fraternal dropout92.2 (± 0.5) 66.2 (± 0.2) 64.4 (± 0.1) 60.7 59.889.2 (± 0.5) 63.7 (± 0.2) 62.1 (± 0.1) 58.8 58.0
" + }, + { + "category_id": 1, + "poly": [ + 299, + 858, + 1399, + 858, + 1399, + 922, + 299, + 922 + ], + "score": 0.951 + }, + { + "category_id": 0, + "poly": [ + 299, + 937, + 565, + 937, + 565, + 969, + 299, + 969 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 300, + 582, + 534, + 582, + 534, + 614, + 300, + 614 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.861 + }, + { + "category_id": 0, + "poly": [ + 301, + 228, + 449, + 228, + 449, + 261, + 301, + 261 + ], + "score": 0.838 + }, + { + "category_id": 6, + "poly": [ + 295, + 1926, + 1400, + 1926, + 1400, + 1990, + 295, + 1990 + ], + "score": 0.731 + }, + { + "category_id": 0, + "poly": [ + 301, + 296, + 642, + 296, + 642, + 327, + 301, + 327 + ], + "score": 0.666 + }, + { + "category_id": 1, + "poly": [ + 301, + 296, + 642, + 296, + 642, + 327, + 301, + 327 + ], + "score": 0.247 + }, + { + "category_id": 13, + "poly": [ + 989, + 1184, + 1060, + 1184, + 1060, + 1214, + 989, + 1214 + ], + "score": 0.89, + "latex": "( 3 0 \\% )" + }, + { + "category_id": 13, + "poly": [ + 775, + 1155, + 836, + 1155, + 836, + 1182, + 775, + 1182 + ], + "score": 0.88, + "latex": "t + 1" + }, + { + "category_id": 13, + "poly": [ + 691, + 1215, + 761, + 1215, + 761, + 1245, + 691, + 1245 + ], + "score": 0.87, + "latex": "( 6 0 \\% )" + }, + { + "category_id": 13, + "poly": [ + 455, + 1183, + 492, + 1183, + 492, + 1212, + 455, + 1212 + ], + "score": 0.87, + "latex": "\\bar { 1 } ^ { s t }" + }, + { + "category_id": 13, + "poly": [ + 863, + 1184, + 934, + 1184, + 934, + 1214, + 863, + 1214 + ], + "score": 0.87, + "latex": "( 3 0 \\% )" + }, + { + "category_id": 13, + "poly": [ + 824, + 1215, + 893, + 1215, + 893, + 1245, + 824, + 1245 + ], + "score": 0.86, + "latex": "( 1 0 \\% )" + }, + { + "category_id": 13, + "poly": [ + 584, + 1215, + 654, + 1215, + 654, + 1245, + 584, + 1245 + ], + "score": 0.86, + "latex": "( 3 0 \\% )" + }, + { + "category_id": 13, + "poly": [ + 739, + 1185, + 812, + 1185, + 812, + 1213, + 739, + 1213 + ], + "score": 0.85, + "latex": "1 4 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 405, + 1157, + 420, + 1157, + 420, + 1181, + 405, + 1181 + ], + "score": 0.68, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1149, + 1096, + 1164, + 1096, + 1164, + 1120, + 1149, + 1120 + ], + "score": 0.57, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 903, + 1734, + 985, + 1734, + 985, + 1762, + 903, + 1762 + ], + "score": 0.26, + "latex": "( \\pm 0 . 5 )" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 930.0, + 571.0, + 930.0, + 571.0, + 981.0, + 292.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 576.0, + 539.0, + 576.0, + 539.0, + 623.0, + 294.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 225.0, + 454.0, + 225.0, + 454.0, + 268.0, + 296.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1919.0, + 1404.0, + 1919.0, + 1404.0, + 1966.0, + 293.0, + 1966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1957.0, + 1183.0, + 1957.0, + 1183.0, + 1993.0, + 295.0, + 1993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 293.0, + 646.0, + 293.0, + 646.0, + 331.0, + 295.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 630.0, + 1402.0, + 630.0, + 1402.0, + 665.0, + 296.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 662.0, + 1406.0, + 662.0, + 1406.0, + 693.0, + 293.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 690.0, + 1406.0, + 690.0, + 1406.0, + 723.0, + 293.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 720.0, + 1404.0, + 720.0, + 1404.0, + 757.0, + 295.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 750.0, + 1405.0, + 750.0, + 1405.0, + 786.0, + 293.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 781.0, + 1404.0, + 781.0, + 1404.0, + 818.0, + 293.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 814.0, + 576.0, + 814.0, + 576.0, + 846.0, + 294.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 353.0, + 1403.0, + 353.0, + 1403.0, + 388.0, + 295.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 384.0, + 1405.0, + 384.0, + 1405.0, + 418.0, + 293.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 413.0, + 1405.0, + 413.0, + 1405.0, + 450.0, + 293.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 447.0, + 1403.0, + 447.0, + 1403.0, + 477.0, + 296.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 473.0, + 1406.0, + 473.0, + 1406.0, + 512.0, + 293.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 504.0, + 1405.0, + 504.0, + 1405.0, + 541.0, + 293.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 536.0, + 848.0, + 536.0, + 848.0, + 570.0, + 295.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1093.0, + 1148.0, + 1093.0, + 1148.0, + 1127.0, + 294.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 1093.0, + 1405.0, + 1093.0, + 1405.0, + 1127.0, + 1165.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1122.0, + 1405.0, + 1122.0, + 1405.0, + 1156.0, + 293.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1153.0, + 404.0, + 1153.0, + 404.0, + 1188.0, + 294.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1153.0, + 774.0, + 1153.0, + 774.0, + 1188.0, + 421.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1153.0, + 1405.0, + 1153.0, + 1405.0, + 1188.0, + 837.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1179.0, + 454.0, + 1179.0, + 454.0, + 1220.0, + 291.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 1179.0, + 738.0, + 1179.0, + 738.0, + 1220.0, + 493.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1179.0, + 862.0, + 1179.0, + 862.0, + 1220.0, + 813.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 1179.0, + 988.0, + 1179.0, + 988.0, + 1220.0, + 935.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 1179.0, + 1408.0, + 1179.0, + 1408.0, + 1220.0, + 1061.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1214.0, + 583.0, + 1214.0, + 583.0, + 1248.0, + 294.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 1214.0, + 690.0, + 1214.0, + 690.0, + 1248.0, + 655.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 1214.0, + 823.0, + 1214.0, + 823.0, + 1248.0, + 762.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1214.0, + 1405.0, + 1214.0, + 1405.0, + 1248.0, + 894.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1242.0, + 1407.0, + 1242.0, + 1407.0, + 1280.0, + 293.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1273.0, + 849.0, + 1273.0, + 849.0, + 1310.0, + 293.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1430.0, + 1406.0, + 1430.0, + 1406.0, + 1464.0, + 295.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1461.0, + 1405.0, + 1461.0, + 1405.0, + 1495.0, + 294.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1490.0, + 1403.0, + 1490.0, + 1403.0, + 1524.0, + 292.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1521.0, + 1405.0, + 1521.0, + 1405.0, + 1554.0, + 294.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1552.0, + 1406.0, + 1552.0, + 1406.0, + 1585.0, + 294.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1583.0, + 1405.0, + 1583.0, + 1405.0, + 1617.0, + 295.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1612.0, + 1405.0, + 1612.0, + 1405.0, + 1647.0, + 291.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1643.0, + 1403.0, + 1643.0, + 1403.0, + 1677.0, + 294.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1319.0, + 1407.0, + 1319.0, + 1407.0, + 1360.0, + 293.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1352.0, + 1403.0, + 1352.0, + 1403.0, + 1390.0, + 294.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1385.0, + 781.0, + 1385.0, + 781.0, + 1418.0, + 295.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 981.0, + 1403.0, + 981.0, + 1403.0, + 1020.0, + 293.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1013.0, + 1407.0, + 1013.0, + 1407.0, + 1049.0, + 292.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1047.0, + 415.0, + 1047.0, + 415.0, + 1083.0, + 294.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 857.0, + 1402.0, + 857.0, + 1402.0, + 895.0, + 294.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 887.0, + 923.0, + 887.0, + 923.0, + 927.0, + 295.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 293.0, + 646.0, + 293.0, + 646.0, + 331.0, + 295.0, + 331.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1081, + 1403, + 1081, + 1403, + 1204, + 299, + 1204 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 300, + 286, + 1400, + 286, + 1400, + 379, + 300, + 379 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 295, + 1004, + 1400, + 1004, + 1400, + 1065, + 295, + 1065 + ], + "score": 0.92 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.836 + }, + { + "category_id": 1, + "poly": [ + 361, + 407, + 1405, + 407, + 1405, + 977, + 361, + 977 + ], + "score": 0.794 + }, + { + "category_id": 0, + "poly": [ + 301, + 230, + 721, + 230, + 721, + 260, + 301, + 260 + ], + "score": 0.544 + }, + { + "category_id": 1, + "poly": [ + 301, + 230, + 721, + 230, + 721, + 260, + 301, + 260 + ], + "score": 0.35 + }, + { + "category_id": 13, + "poly": [ + 650, + 692, + 705, + 692, + 705, + 721, + 650, + 721 + ], + "score": 0.85, + "latex": "50 \\%" + }, + { + "category_id": 13, + "poly": [ + 1306, + 473, + 1325, + 473, + 1325, + 495, + 1306, + 495 + ], + "score": 0.75, + "latex": "\\kappa" + }, + { + "category_id": 13, + "poly": [ + 886, + 443, + 906, + 443, + 906, + 465, + 886, + 465 + ], + "score": 0.69, + "latex": "\\kappa" + }, + { + "category_id": 13, + "poly": [ + 649, + 564, + 669, + 564, + 669, + 586, + 649, + 586 + ], + "score": 0.63, + "latex": "\\kappa" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 232.0, + 724.0, + 232.0, + 724.0, + 264.0, + 297.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1080.0, + 1407.0, + 1080.0, + 1407.0, + 1117.0, + 293.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1112.0, + 1403.0, + 1112.0, + 1403.0, + 1145.0, + 295.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1142.0, + 1404.0, + 1142.0, + 1404.0, + 1177.0, + 293.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1174.0, + 1030.0, + 1174.0, + 1030.0, + 1208.0, + 294.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 285.0, + 1405.0, + 285.0, + 1405.0, + 323.0, + 294.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 317.0, + 1405.0, + 317.0, + 1405.0, + 353.0, + 293.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 348.0, + 735.0, + 348.0, + 735.0, + 383.0, + 294.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 998.0, + 1405.0, + 998.0, + 1405.0, + 1042.0, + 293.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1034.0, + 952.0, + 1034.0, + 952.0, + 1070.0, + 291.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 406.0, + 1405.0, + 406.0, + 1405.0, + 440.0, + 361.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 438.0, + 885.0, + 438.0, + 885.0, + 471.0, + 394.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 438.0, + 1405.0, + 438.0, + 1405.0, + 471.0, + 907.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 468.0, + 1305.0, + 468.0, + 1305.0, + 504.0, + 394.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 468.0, + 1406.0, + 468.0, + 1406.0, + 504.0, + 1326.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 495.0, + 1404.0, + 495.0, + 1404.0, + 533.0, + 393.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 530.0, + 1404.0, + 530.0, + 1404.0, + 563.0, + 393.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 559.0, + 648.0, + 559.0, + 648.0, + 593.0, + 391.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 559.0, + 1405.0, + 559.0, + 1405.0, + 593.0, + 670.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 594.0, + 469.0, + 594.0, + 469.0, + 624.0, + 391.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 628.0, + 1406.0, + 628.0, + 1406.0, + 665.0, + 362.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 663.0, + 1405.0, + 663.0, + 1405.0, + 695.0, + 396.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 692.0, + 649.0, + 692.0, + 649.0, + 725.0, + 394.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 692.0, + 1404.0, + 692.0, + 1404.0, + 725.0, + 706.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 721.0, + 1407.0, + 721.0, + 1407.0, + 758.0, + 392.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 754.0, + 1405.0, + 754.0, + 1405.0, + 786.0, + 394.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 783.0, + 1404.0, + 783.0, + 1404.0, + 815.0, + 394.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 815.0, + 1405.0, + 815.0, + 1405.0, + 848.0, + 394.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 845.0, + 529.0, + 845.0, + 529.0, + 876.0, + 392.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 882.0, + 1404.0, + 882.0, + 1404.0, + 922.0, + 356.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 916.0, + 1405.0, + 916.0, + 1405.0, + 949.0, + 394.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 948.0, + 932.0, + 948.0, + 932.0, + 980.0, + 393.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 232.0, + 724.0, + 232.0, + 724.0, + 264.0, + 297.0, + 264.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 291, + 284, + 1404, + 284, + 1404, + 351, + 291, + 351 + ], + "score": 0.94 + }, + { + "category_id": 8, + "poly": [ + 396, + 364, + 1325, + 364, + 1325, + 451, + 396, + 451 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 298, + 1413, + 649, + 1413, + 649, + 1447, + 298, + 1447 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 297, + 1172, + 607, + 1172, + 607, + 1207, + 297, + 1207 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 816, + 73, + 816, + 106, + 298, + 106 + ], + "score": 0.921 + }, + { + "category_id": 8, + "poly": [ + 524, + 860, + 1087, + 860, + 1087, + 945, + 524, + 945 + ], + "score": 0.914 + }, + { + "category_id": 1, + "poly": [ + 298, + 1104, + 815, + 1104, + 815, + 1143, + 298, + 1143 + ], + "score": 0.91 + }, + { + "category_id": 1, + "poly": [ + 296, + 484, + 961, + 484, + 961, + 519, + 296, + 519 + ], + "score": 0.908 + }, + { + "category_id": 8, + "poly": [ + 387, + 1462, + 1371, + 1462, + 1371, + 1534, + 387, + 1534 + ], + "score": 0.897 + }, + { + "category_id": 8, + "poly": [ + 527, + 956, + 832, + 956, + 832, + 1042, + 527, + 1042 + ], + "score": 0.884 + }, + { + "category_id": 9, + "poly": [ + 1351, + 884, + 1401, + 884, + 1401, + 915, + 1351, + 915 + ], + "score": 0.884 + }, + { + "category_id": 9, + "poly": [ + 1351, + 980, + 1401, + 980, + 1401, + 1011, + 1351, + 1011 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1351, + 692, + 1401, + 692, + 1401, + 725, + 1351, + 725 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1579, + 1401, + 1579, + 1401, + 1611, + 1351, + 1611 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1351, + 536, + 1401, + 536, + 1401, + 567, + 1351, + 567 + ], + "score": 0.878 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1227, + 1401, + 1227, + 1401, + 1259, + 1351, + 1259 + ], + "score": 0.878 + }, + { + "category_id": 8, + "poly": [ + 484, + 1572, + 1135, + 1572, + 1135, + 1619, + 484, + 1619 + ], + "score": 0.878 + }, + { + "category_id": 9, + "poly": [ + 1351, + 788, + 1401, + 788, + 1401, + 819, + 1351, + 819 + ], + "score": 0.877 + }, + { + "category_id": 8, + "poly": [ + 521, + 576, + 1039, + 576, + 1039, + 618, + 521, + 618 + ], + "score": 0.876 + }, + { + "category_id": 9, + "poly": [ + 1351, + 581, + 1401, + 581, + 1401, + 612, + 1351, + 612 + ], + "score": 0.87 + }, + { + "category_id": 8, + "poly": [ + 474, + 1323, + 1313, + 1323, + 1313, + 1396, + 474, + 1396 + ], + "score": 0.869 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1278, + 1401, + 1278, + 1401, + 1309, + 1352, + 1309 + ], + "score": 0.868 + }, + { + "category_id": 9, + "poly": [ + 1353, + 1534, + 1400, + 1534, + 1400, + 1563, + 1353, + 1563 + ], + "score": 0.866 + }, + { + "category_id": 8, + "poly": [ + 470, + 1273, + 1268, + 1273, + 1268, + 1317, + 470, + 1317 + ], + "score": 0.863 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1347, + 1401, + 1347, + 1401, + 1378, + 1352, + 1378 + ], + "score": 0.86 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.853 + }, + { + "category_id": 9, + "poly": [ + 1352, + 390, + 1401, + 390, + 1401, + 422, + 1352, + 422 + ], + "score": 0.835 + }, + { + "category_id": 8, + "poly": [ + 524, + 763, + 1255, + 763, + 1255, + 850, + 524, + 850 + ], + "score": 0.825 + }, + { + "category_id": 2, + "poly": [ + 1374, + 1631, + 1403, + 1631, + 1403, + 1661, + 1374, + 1661 + ], + "score": 0.811 + }, + { + "category_id": 2, + "poly": [ + 1374, + 1058, + 1403, + 1058, + 1403, + 1088, + 1374, + 1088 + ], + "score": 0.797 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 397, + 229, + 397, + 261, + 298, + 261 + ], + "score": 0.734 + }, + { + "category_id": 8, + "poly": [ + 525, + 668, + 1220, + 668, + 1220, + 753, + 525, + 753 + ], + "score": 0.734 + }, + { + "category_id": 8, + "poly": [ + 395, + 1222, + 965, + 1222, + 965, + 1267, + 395, + 1267 + ], + "score": 0.56 + }, + { + "category_id": 8, + "poly": [ + 524, + 622, + 911, + 622, + 911, + 662, + 524, + 662 + ], + "score": 0.519 + }, + { + "category_id": 8, + "poly": [ + 406, + 531, + 955, + 531, + 955, + 573, + 406, + 573 + ], + "score": 0.457 + }, + { + "category_id": 0, + "poly": [ + 298, + 229, + 397, + 229, + 397, + 261, + 298, + 261 + ], + "score": 0.2 + }, + { + "category_id": 8, + "poly": [ + 397, + 1225, + 1312, + 1225, + 1312, + 1400, + 397, + 1400 + ], + "score": 0.168 + }, + { + "category_id": 14, + "poly": [ + 396, + 530, + 1248, + 530, + 1248, + 1048, + 396, + 1048 + ], + "score": 0.95, + "latex": "\\begin{array} { r l } { \\mathcal { R } _ { F S } [ \\mathbf { z } , \\boldsymbol { \\theta } ] } & { = \\mathrm { E } _ { \\boldsymbol { \\phi } , \\boldsymbol { \\xi } } [ \\| \\mathbf { P } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } , \\boldsymbol { \\theta } ) - \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } \\| ^ { 2 } ] } \\\\ & { = \\mathrm { E } _ { \\boldsymbol { \\phi } , \\boldsymbol { \\xi } } \\Bigg [ \\| \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\| _ { \\mathbf { z } } ^ { 2 } + \\mathrm { E } _ { \\boldsymbol { \\xi } , \\boldsymbol { \\xi } } \\Bigg [ \\| \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\| _ { \\mathbf { z } } ^ { 2 } \\Bigg ] } \\\\ & { - 2 \\mathrm { E } _ { \\boldsymbol { \\xi } , \\boldsymbol { s } , \\boldsymbol { \\xi } } [ \\| \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } \\| ^ { 2 } \\mathbf { p } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } ] } \\\\ & { = 2 \\sum _ { \\boldsymbol { \\phi } = 1 } ^ { \\infty } ( \\mathrm { E } _ { \\boldsymbol { \\phi } } [ \\partial _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } ] ^ { 2 } ) - \\mathrm { E } _ { \\boldsymbol { \\phi } , \\boldsymbol { \\xi } } \\Bigg [ \\mathcal { P } _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) \\boldsymbol { \\theta } ] _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ; \\boldsymbol { \\theta } ) \\Bigg ] } \\\\ & = 2 \\sum _ { \\boldsymbol { \\phi } = 1 } ^ { \\infty } ( \\mathrm { E } _ { \\boldsymbol { \\phi } } [ \\partial _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } , \\boldsymbol { \\theta } ) ] - \\mathrm { E } _ { \\boldsymbol { \\phi } , \\boldsymbol { \\xi } } [ \\mathcal { P } _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol { \\xi } ) ] \\mathrm { E } _ { \\boldsymbol { \\phi } } [ \\mathcal { P } _ { \\boldsymbol { \\phi } } ( \\mathbf { z } , \\boldsymbol { s } , \\boldsymbol \\ \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 897, + 286, + 1109, + 286, + 1109, + 321, + 897, + 321 + ], + "score": 0.93, + "latex": "\\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) \\in \\mathbb { R } ^ { m }" + }, + { + "category_id": 14, + "poly": [ + 393, + 1219, + 1321, + 1219, + 1321, + 1405, + 393, + 1405 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\mathcal { R } _ { t } ( \\mathbf { z } ^ { t } ) : = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] } \\\\ & { \\quad \\quad \\quad \\quad = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) + \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] } \\\\ & { \\quad \\quad \\quad \\quad = 4 \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\frac { \\mathbf { p } ^ { t } \\left( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta \\right) - \\mathbf { p } ^ { t } \\left( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta \\right) } { 2 } + \\frac { \\mathbf { p } ^ { t } \\left( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta \\right) - \\mathbf { p } ^ { t } \\left( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta \\right) } { 2 } \\| _ { 2 } ^ { 2 } \\right] . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 395, + 360, + 1329, + 360, + 1329, + 455, + 395, + 455 + ], + "score": 0.92, + "latex": "\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\theta ) = \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] = 2 \\sum _ { q = 1 } ^ { m } \\mathrm { v a r } _ { s _ { i } ^ { t } } ( p _ { q } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) ) ." + }, + { + "category_id": 13, + "poly": [ + 424, + 1174, + 542, + 1174, + 542, + 1207, + 424, + 1207 + ], + "score": 0.92, + "latex": "\\bar { s } : = \\mathbb { E } _ { s } [ s ]" + }, + { + "category_id": 13, + "poly": [ + 475, + 1103, + 809, + 1103, + 809, + 1142, + 475, + 1142 + ], + "score": 0.91, + "latex": "\\mathcal { R } _ { F D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) \\leq 4 \\tilde { \\mathcal { R } } _ { E L D } ( \\mathbf { z } ^ { t } ; \\boldsymbol { \\theta } ) ." + }, + { + "category_id": 13, + "poly": [ + 478, + 287, + 504, + 287, + 504, + 321, + 478, + 321 + ], + "score": 0.88, + "latex": "s _ { i } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 557, + 287, + 583, + 287, + 583, + 324, + 557, + 324 + ], + "score": 0.87, + "latex": "\\boldsymbol { s } _ { j } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 939, + 490, + 952, + 490, + 952, + 512, + 939, + 512 + ], + "score": 0.78, + "latex": "t" + }, + { + "category_id": 14, + "poly": [ + 394, + 1459, + 1373, + 1459, + 1373, + 1620, + 394, + 1620 + ], + "score": 0.73, + "latex": "\\begin{array} { r l } & { \\mathcal { R } _ { t } ( \\mathbf { z } ^ { t } ) \\leq 4 \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\frac { 1 } { 2 } \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) \\| _ { 2 } ^ { 2 } + \\frac { 1 } { 2 } \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] } \\\\ & { \\qquad ( 2 \\pi ) } \\\\ & { \\qquad = 4 \\mathbb { E } _ { s _ { i } ^ { t } } \\left[ \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\theta ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\theta ) \\| _ { 2 } ^ { 2 } \\right] = 4 \\tilde { \\mathcal { R } } _ { E L D } ( \\mathbf { z } ^ { t } ; \\theta ) . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 393, + 1462, + 1371, + 1462, + 1371, + 1534, + 393, + 1534 + ], + "score": 0.49, + "latex": "\\mathcal { R } _ { t } ( \\mathbf { z } ^ { t } ) \\leq 4 \\mathbb { E } _ { s _ { i } ^ { t } , s _ { j } ^ { t } } \\left[ \\frac { 1 } { 2 } \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { i } ^ { t } ; \\boldsymbol { \\theta } ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\boldsymbol { \\theta } ) \\| _ { 2 } ^ { 2 } + \\frac { 1 } { 2 } \\| \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , \\bar { s } ; \\boldsymbol { \\theta } ) - \\mathbf { p } ^ { t } ( \\mathbf { z } ^ { t } , s _ { j } ^ { t } ; \\boldsymbol { \\theta } ) \\| _ { 2 } ^ { 2 } \\right]" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 818.0, + 72.0, + 818.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 871.0, + 2085.0, + 871.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 1634.0, + 1404.0, + 1634.0, + 1404.0, + 1665.0, + 1376.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1377.0, + 1063.0, + 1403.0, + 1063.0, + 1403.0, + 1087.0, + 1377.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 398.0, + 230.0, + 398.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 282.0, + 477.0, + 282.0, + 477.0, + 327.0, + 293.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 282.0, + 556.0, + 282.0, + 556.0, + 327.0, + 505.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 282.0, + 896.0, + 282.0, + 896.0, + 327.0, + 584.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 282.0, + 1406.0, + 282.0, + 1406.0, + 327.0, + 1110.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 315.0, + 594.0, + 315.0, + 594.0, + 354.0, + 294.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1407.0, + 650.0, + 1407.0, + 650.0, + 1454.0, + 294.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1173.0, + 423.0, + 1173.0, + 423.0, + 1210.0, + 296.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 1173.0, + 608.0, + 1173.0, + 608.0, + 1210.0, + 543.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1103.0, + 474.0, + 1103.0, + 474.0, + 1147.0, + 295.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 1103.0, + 817.0, + 1103.0, + 817.0, + 1147.0, + 810.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 483.0, + 938.0, + 483.0, + 938.0, + 522.0, + 294.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 483.0, + 964.0, + 483.0, + 964.0, + 522.0, + 953.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 398.0, + 230.0, + 398.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/r1-4BLaQz/r1-4BLaQz.md b/parse/train/r1-4BLaQz/r1-4BLaQz.md new file mode 100644 index 0000000000000000000000000000000000000000..cbfefa4cb521ef33641fc189ae21dc96182b16a6 --- /dev/null +++ b/parse/train/r1-4BLaQz/r1-4BLaQz.md @@ -0,0 +1,202 @@ +# CLUSTER-BASED WARM-START NETS + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +Theories in cognitive psychology postulate that humans use similarity as a basis for object categorization. However, work in image classification generally assumes disjoint and equally dissimilar classes to achieve super-human levels of performance on certain datasets. In our work, we adapt notions of similarity using weak labels over multiple hierarchical levels to boost classification performance. Instead of pitting clustering directly against classification, we use a warm-start based evaluation to explicitly provide value to a clustering representation by its ability to aid classification. We evaluate on CIFAR10 and a fine-grained classification dataset to show improvements in performance with the procedural addition of intermediate losses and weak labels based on multiple hierarchy levels. Furthermore, we show that pretraining AlexNet on hierarchical weak labels in conjunction with intermediate losses outperforms a classification baseline by over $17 \%$ on a subset of Birdsnap dataset. Finally, we show improvement over AlexNet trained using ImageNet pre-trained weights as initializations which further supports our claim of the importance of similarity. + +# 1 INTRODUCTION + +Similarity is one of the bases of object categorization in humans (Rosch, 1999). Theories of perceptual categorization in cognitive psychology postulate that humans construct categories by grouping similar stimuli to construct one or several prototypes. New instances are then labeled as the category that they are most similar to. For instance, when one categorizes a specific animal as a dog, they are saying that it is more similar to previously observed dogs than it is to all other objects. This view of categorization better explains the often fuzzy boundaries between real-life classes, where a new object may be equally similar to multiple prototypes. For example, while a dolphin is technically a mammal, its visual appearance is similar to fish, resulting in a lot of people misclassifying it. + +While research in cognitive psychology on object categorization might indicate that similarity should play a central role in object classification in humans, image classification in computer vision seems to do pretty well without using it. The task is commonly interpreted as one of classifying an image into one of multiple classes that are assumed to be disjoint and equally dissimilar. The use of softmax-based loss functions assumes that the classes are disjoint, while the use of one-hot labels assumes that classes are equally dissimilar. Despite those strong assumptions, which seem to violate human notions of categories, image classification has shown remarkable progress, achieving superhuman performance on multiple datasets (Ioffe & Szegedy, 2015; He et al., 2015). Far from being an anomaly, state-of-the-art models across image classification benchmark datasets use losses, such as cross-entropy loss, that make the explicit assumption of disjoint classes. However, Hinton et al. (2015) notes that the predictions of ensembles of image classification models produces soft labels that “define a rich similarity structure over the data” despite the predictions of individual models not capturing this structure. + +Can similarity-based metrics improve classification performance in convolutional neural networks? Previous work has tried to answer this question in different problems such as metric learning (Bellet et al., 2013), clustering (Seldin et al., 2003), and hierarchical classification (Lee & Crawford, 2005). We focus on applications of similarity in the context of convolutional neural networks such as Hsu & Kira (2015) who use a contrastive loss to perform end-to-end clustering using weak labels. While they show impressive performance on MNIST (LeCun et al., 1998) and CIFAR-10 (Krizhevsky & Hinton, 2009), their method does not scale well to more complex datasets. Rippel et al. (2016) propose a new loss function, magnet loss, to learn a representational space where distance corresponds to similarity. They show that clustering in this space allows them to achieve state-of-the-art performance on multiple fine-grained classification datasets.1 However, they initialize their model with pre-trained weights on ImageNet(Russakovsky et al., 2015), so it is not clear whether their model is learning a good representation, or if it is finding a good mapping between already learned representations and the labels. Wu et al. (2017) pose the problem as one of hierarchical classification and propose a loss based on an ultra-metric tree representation of a hierarchy over the labels. While their loss has interesting properties, it only outperforms classification losses for datasets with a small number of instances per class. In this work, we use a contrastive loss to improve the classification performance of randomly initialized convolutional neural networks on a fine-grained classification task. + +What measures of similarity should we teach our model? We represent the relations between our classes in a hierarchy where the labels are all leaf nodes. Therefore, there are two kinds of similarities that we want our model to capture. The first is intra-class similarity—all cats are similar to each other and they are dissimilar to other animals. The second is inter-class similarity—dogs and cats are more similar to each other than they are to non-living objects. For a more complex hierarchy, our model should learn similarity at different levels of the class hierarchy. Wu et al. (2017) use their hierarchical loss function to learn those similarities, however, they observe that reducing the similarity between two classes across all levels of the hierarchy to a single value biases the model towards correct classification at the higher levels of the hierarchy resulting in poor classification performance. Rippel et al. (2016) also observe that applying classification losses to the final layer reduces the entire representation of each class to a single scalar value which destroys intra- and inter-class variation. However, these are the same variations that we want our model to capture. We overcome those limitations by explicitly training the model to capture different grains of similarity at different levels of the network through applying an intermediate pairwise contrastive loss at those levels. In this manner, we can use hierarchical information, such as species of birds having the same coarse-grained category of bird while having different fine-grained labels. Despite being able to encode different levels of similarity, a contrastive loss does not require an explicit hierarchy; we only need weak labels for pairs of instances. + +How can we evaluate the representations learned by a clustering algorithm? Previous work has always tried to compete against classification using metrics biased towards the latter task. While some researchers have used hierarchical metrics (Wu et al., 2017) or qualitative measures (Rippel et al., 2016) to evaluate the quality of their representations, their primary evaluation criteria has consistently been the classification accuracy of their model. We propose another way to evaluate the clustering representations by using the clustering model as a “warm start” from which they can train on classification. The intuition is that if the clustering model learned a representations that captures similarity in a space, it would be be at an advantage when it is fine-tuned for classification. + +We propose the use of a contrastive loss at different layers of a convolutional neural network to learn a notion of similarity across a set of labels. We also propose the use of the accuracy of a model pre-trained for clustering and fine-tuned for classification as an evaluation metric for clustering algorithms. + +# 2 INTERMEDIATE PAIRWISE CONTRASTIVE LOSS + +Pairwise contrastive losses were first proposed by Hadsell et al. (2006) to learn a function that maps high dimensional inputs to lower dimensional outputs such that distances in the lower dimensional space approximate relationships in the input space. Relationships in the input space are not restricted to simple distance measures. In the context of image classification, the resulting function would have to learn the complex invariances that make two images of the same class similar, such that they are both mapped to points that are close to each other in the output space. This loss can be calculated using the representation of two points in the same embedding space and a binary label for whether or not they are similar. Hence, the loss only requires weak labels in the form of pairwise relationships. + +![](images/a72c159113b92f8c536176837b61a6ca7237947ff8e534d0cbf2286e02239b7b.jpg) +Figure 1: Pairwise losses are applied to intermediate layers of the network as they are calculated through pairwise comparisons of the activations at those layers. + +# 2.1 LOSS FORMULATION + +Let $( X _ { 1 } , X _ { 2 } )$ be vectors that are calculated as mappings of two input instances in the same embedding space. $\mathbb { 1 } _ { X _ { 1 } , X _ { 2 } }$ is an indicator variable that takes a value of one if $( X _ { 1 } , X _ { 2 } )$ are similar and zero otherwise. $D ( X , Y )$ can be any distance function between two vectors in the same space. $L _ { S } ( \cdot )$ is a partial loss function for a similar pair, and $L _ { D } ( \cdot )$ is a partial loss function for dissimilar pairs. Then the general formulation of the pairwise contrastive loss function is as shown in Eq. (1): + +$$ +\boldsymbol { \mathcal { L } } ( \mathbf { X _ { 1 } } , \mathbf { X _ { 2 } } ) = \mathbb { 1 } _ { X _ { 1 } , X _ { 2 } } L _ { S } ( D ( \mathbf { X _ { 1 } } , \mathbf { X _ { 2 } } ) ) + ( 1 - \mathbb { 1 } _ { X _ { 1 } , X _ { 2 } } ) L _ { D } ( D ( \mathbf { X _ { 1 } } , \mathbf { X _ { 2 } } ) ) . +$$ + +The first half of Eq. (1) penalizes any divergence between the two vectors for similar instances while the second half of the loss rewards divergence up to a margin. Following the formulation proposed by (Hsu & Kira, 2015), we use Kullback-Leibler (KL) divergence as our distance metric and a hinge loss for dissimilar cases. The final formulation of the loss we use is shown in Eq. (2): + +$$ +\begin{array} { r } { \mathcal { L } ( \mathbf { P } , \mathbf { Q } ) = l o s s ( \mathbf { P } , \mathbf { Q } ) + l o s s ( \mathbf { Q } , \mathbf { P } ) , } \end{array} +$$ + +$$ +l o s s ( \mathbf { P } , \mathbf { Q } ) = \mathbb { 1 } _ { P , Q } K L ( \mathbf { P } | | \mathbf { Q } ) + ( 1 - \mathbb { 1 } _ { P , Q } ) \operatorname* { m a x } \{ m a r g i n - K L ( \mathbf { P } | | \mathbf { Q } ) , 0 \} , +$$ + +$$ +K L ( \mathbf { P } \parallel \mathbf { Q } ) = \sum _ { i = 1 } ^ { k } P _ { i } \log \Big ( \frac { P _ { i } } { Q _ { i } } \Big ) . +$$ + +We have found that a value of two for the margin works well. Since we apply the loss to activations at intermediate layers of the network, we first apply a softmax operation on the activations to obtain the two vectors, $P$ and $Q$ , used in Eq. (2). + +# 2.2 HIERARCHICAL SIMILARITY + +Given that one of our objectives is to encode different levels of similarity into the representations learned by the network, we explore different ways of using the intermediate loss to achieve that goal as shown in Fig. 2. We establish a baseline network where the losses are all applied to the output of the last layer. We then use three different variants by applying intermediate losses at different levels of the network. In variant B, we apply the intermediate loss at the penultimate layer. This allows us to apply the contrastive-loss to a higher dimensional space where it could potentially learn a richer representation. In variants C and D, we provide the network with two different kinds of weak labels based on fine-grained and coarse-grained. Since we expect that a weaker form of similarity would be more beneficial earlier on in the network pipeline, we apply the coarse-grained loss to the secondto-last fully-connected layer in Network C. It should be noted that many instance pairs will have contradicting losses since they may match at the coarse-grained level but conflict at the fine-grained level. To tackle this, we augment the network with a skip layer around the second-to-last layer and apply the coarse-grained loss to the skip layer only. Hence, the coarse-grained loss will only provide additional information to the network without directly detracting from the representations passed to subsequent layers. While we show the network variants on AlexNet, the same ideas are applied to a variant of LeNet that has three fully-connected layers. + +![](images/c751ec74231b40986e76cf6183577dc664e8807abf7061a09418db3bd1e4d8cc.jpg) +Figure 2: Different network architecture and losses. Network A is the baseline network. Network B applies the fine-grained intermediate loss to the penultimate layer as well. Network C extends Network B by applying the coarse-grained intermediate loss at the second-to-last layer. Network D extends Network B by adding a branch around the second-to-last layer, and applying the coarsegrained intermediate loss there. + +We follow the efficient implementation regime proposed by (Hsu & Kira, 2015) in order to train a single network based on pairwise constraints. The original network formulation to handle pairwise constraints is a Siamese network. For every pairwise constraint, at least two data passes need to be performed which results in redundancies. Instead, we feed the softmax outcomes from a single network and the pairwise similarity constraints to the contrastive loss function which externally handles all possible pairwise constraints that are available within a mini-batch. We further extend the application of contrastive loss functions to multiple intermediate layers with a minimal overhead of extra processing and pairwise constraints, if alternate hierarchical labels are used. Details regarding the exact parameters and setup used to perform experiments are furnished in the Appendices. PyTorch2 is used for all implementations discussed within this paper. + +# 3 EXPERIMENTS + +We evaluate our approach on two different datasets using two different convolutional neural networks: LeNet on CIFAR10 and AlexNet on Birdsnap34–a 34-class subset of the Birdsnap dataset (Berg et al., 2014). We generate coarse-grained labels for CIFAR10 by separating the ten classes into six animal classes and four vehicle classes. Below is a discussion of how we constructed Birdsnap34. + +Birdsnap34 Birdsnap is a popular fine-grained classification dataset that has 500 different classes and over 40,000 images. We annotated Birdsnap using the scientific classification of each bird to create a six-level hierarchy over all the labels. In this work, we use the birds’ subfamily and family as our fine-grained and coarse-grained labels, respectively. Using a pairwise loss with a large number of classes poses a significant implementation challenge. For a uniform dataset with $n$ classes, the ratio of similar pairs to dissimilar pairs is roughly $1 : ( n - 1 )$ . Hence, for a large number of classes, most of the pairs will be dissimilar resulting in a very weak signal to the model for clustering instances together. This challenge is analogous to that of imbalanced datasets (Van Horn & Perona, 2017), which is a major problem that lies beyond the scope of our paper. Instead of abandoning Birdsnap, we curate a 34-class subset, which we will refer to as Birdsnap34, and apply our approach to it. The classes are chosen such that each class comes from a unique biological subfamily of birds. Through sampling form unique subfamilies, we increase the interclass variance between our classes. The resulting dataset has 2982 training images and 158 test images that are divided into 34 fine-grained and 18 coarse-grained classes. We plan on extending our model to the entire Birdsnap dataset in future work. + +All the network architectures used have randomly initialized weights unless specified otherwise. The models are evaluated using accuracy, purity, and Normalized Mutual Information (NMI) (Strehl & Ghosh, 2002). We train each model for a fixed number of epochs, and we report the top performing results out of 5 runs. Details regarding the exact parameters and setups used to perform the experiments are available in Appendix B. + +# 3.1 QUALITY OF WARM START AS AN EVALUATION METRIC + +In our first experiment, we show the value of evaluating the representations learned by a clustering loss through using it as a warm start from training on classification. We replicate the setup used by Hsu & Kira (2015) by using LeNet on CIFAR10, and extend their framework to AlexNet on Birdsnap34. We use the Hungarian algorithm (Kuhn, 1955) to evaluate clustering accuracy. This method of evaluation assumes that the clusters will be centered across the different dimensions of the softmax output of the final layer. The Hungarian algorithm is then used to find the assignment of dimensions to classes that would achieve the highest accuracy. We use the Hungarian algorithm to evaluate clustering performance in all of our experiments. + +Table 1: Baseline classification performance of LeNet and AlexNet + +
CIFAR-10Birdsnap-34
Classification75.4853.50
Clustering74.2117.20
Warm Start Classification75.7550.32
+ +From Table 1 we observe that for CIFAR10, the application of the loss to the final layer results in clustering performance almost matching classification performance. However, in more complex dataset such as Birdsnap34, this pattern disappears. This supports the pattern shown in (Hsu & Kira, 2015) where the gap between clustering and classification performance increases when they move from MNIST to CIFAR10. The Cross-Entropy loss formulation, as mentioned in (Hsu & Kira, 2015), is extremely harsh and seeks to segment each and every sample into one absolute class. The tSNE (van der Maaten & Hinton, 2008) outcomes from the last fully connected layer of LeNet in Fig. 3 match our observations. + +![](images/a8990135f6d1aafc86247b26a23967600823155298c1769f10985048200a7f11.jpg) +Figure 3: t-SNE Visualization of the outputs of LeNet on CIFAR10 + +# 3.2 INTERMEDIATE LOSS + +The previous experiment indicates that the representations learned from end-to-end clustering do not extend beyond simple datasets. We posit that this is a result of how the clustering loss is applied. Most machine learning models only apply losses to their last layer. This is generally a reasonable thing to do since we only have labels for the output of the last layer. On the other hand, pairwise losses can be applied at any level of the network since they only require a pair of vectors in the same embedding space. In this experiment we explore the potential of applying losses at intermediate layers on the learned representations. + +Table 2: Using intermediate losses + +
Clustering Acc.NMIPurityClassification Acc.
CIFAR10Cross-Entropy LeNet A LeNet B1 74.21 72.661 0.551 0.7475.48 75.75
Birdsnap34Cross-Entropy AlexNet A AlexNetB1 17.20 20.380.54 1 0.39 0.450.73 1 0.17 0.2075.42 53.50 50.32 45.86
+ +We train two variants of LeNet and AlexNet using the pairwise contrastive loss. Variant A follows the setup used by Hsu & Kira (2015), while Variant B applies the same loss at both the final and penultimate layer of the convolutional neural network. We also train the same network using CrossEntropy loss to establish baseline performance. As shown in Table 2, performance on CIFAR10 is almost the same across all three settings. Meanwhile, we see clear differences on Birdsnap34. There is a significant increase across all metrics for the clustering models for the intermediate loss version, however, this improvement does not carry over to the classification accuracy. + +# 3.3 HIERARCHICAL CLUSTERING + +Previous work on embedding spaces has shown that adding structure to the embedding space or learning it using a hierarchical structure often validates its integrity as well as improves performance (Mikolov et al., 2013). In an effort to understand the impact of adding a hierarchical-labelbased contrastive loss to our baseline networks, we apply it on variants C and D of LeNet and AlexNet. Since the restriction on dimensionality is removed by using a contrastive loss with KL divergence, we can directly apply the same loss to an alternate set of weak labels and representations. Our ultimate expectation is that adding hierarchical structure should aid in improving classification, given their potency to improve embedding spaces as shown in previous works. + +Table 3: Impact of adding heirarchy based weak labels + +
Clustering Acc.NMIPurityClassification Acc.
CIFAR10Cross-Entropy75.48
LeNet A74.210.550.7475.75
LeNet B72.660.540.7375.42
LeNet C72.840.550.7375.11
LeNet D74.170.560.7475.53
Birdsnap34Cross-Entropy11153.50
AlexNet A17.200.390.1750.32
AlexNet B20.380.450.2045.86
AlexNet C24.840.510.2570.70
AlexNet D22.290.490.2265.61
+ +Table 3 shows that the experimental results align with our expectations in terms of overall improvement in classification. As seen in other experiments, the performance on CIFAR10 is almost the same across all metrics. Meanwhile, there is a spike in performance for Birdsnap34 for both AlexNet C and D with the metrics far exceeding classification performance. Given that Birdsnap34 has a more well-defined hierarchy, it would be expected that using the hierarchical labels to learn similarity would result in learning good representations. Another thing worth noting is that the dataset is curated by having subfamilies as fine-grained labels and biological families as coarsegrained labels. This structure results in similarity along this tree corresponding to distinct visual features. This might be indicating that choice of hierarchy can have an impact on performance. A surprising finding is that AlexNet C outperforms AlexNet D. One would expect AlexNet D to be able to extract features for both the coarse and fine levels of the hierarchy. We plan on investigating this in future work. + +# 3.4 SENSITIVITY TO NETWORK SIZE AND WEIGHT INITIALIZATION + +Given the ability of standard end-to-end clustering networks to perform fine-tuned classification reasonably well, we test their sensitivity towards network size in order to understand if they would be able to achieve or outperform their simpler baseline classification counterparts. In order to do so, we double the number of convolutional filters and the sizes of fully-connected layers within AlexNet. We use their doubled networks to perform clustering and fine-tune them for classification. We denote these modified networks with a suffix of “Double.” + +Table 4: Sensitivity to network capacity + +
Clustering Acc.NMIPurityClassification Acc.
Cross-Entropy一1153.50
AlexNet A (Double)22.930.490.2359.24
AlexNet C24.840.510.2570.70
+ +Table 5: Comparison against ImageNet pre-trained weights + +
Clustering Acc.NMIPurityClassification Acc.
Cross-Entropy11181.53
AlexNet C24.200.510.2482.80
+ +From Table 4, we clearly observe that clustering is highly sensitive to network capacity. Compared to AlexNet A, its doubled counterpart allows clustering to learn better representations which leads to classificaiton performance being higher than the baseline classification model. However, our AlexNet C variant outperforms the double capacity network. This points to the efficacy of the combination of weak labels over multiple levels of heirarchy and the use of intermediate losses. Furthermore, Table 5 shows that even we use ImageNet pre-trained weights, variant C outperforms the ImageNet pre-trained classification baseline. This points to a tractable and viable approach to extending performance over more complex datasets without the need to increase the capacity of neural networks. + +# 4 DISCUSSION AND FUTURE WORK + +In this work, we argue that similarity-based losses can be applied as a warm start to boost the performance of image classification models. We show that applying a pairwise contrastive loss to intermediate layers of the network results in better clustering performance, as well as better finetuned classification performance. Furthermore, we demonstrate how the pairwise loss can be used to train a model using weak hierarchical labels. + +We find that training with hierarchical labels results in the highest performance beating models with more parameters and approaches the performance of models pre-trained with ImageNet weights. This is a very significant finding since ImageNet contains multiple bird classes, so a model pretrained with ImageNet has seen many more birds than a model pre-trained with pairwise constraints on Birdsnap34. Nevertheless, it only outperforms a cluster-based warm-start model by just $10 \%$ . Regardless, we also show that applying our approach to ImageNet weights still results in a boost of $1 . 2 7 \%$ . This supports our claim that similarity-based metrics can improve classification performance, but it suggests that the gains we expect decrease if we start of with a good representational space. + +We hope to expand this in future work in multiple ways. We plan on extending our approach to the entire Birdsnap dataset, as well as to other fine-grained classification datasets. We also hope to perform more extensive analysis of the quality of the embedding spaces learned as it is obvious that the use of the Hungarian algorithm does not accurately reflect the performance of the clustering model. + +# REFERENCES + +Aurelien Bellet, Amaury Habrard, and Marc Sebban. A survey on metric learning for feature vectors ´ and structured data. arXiv preprint arXiv:1306.6709, 2013. + +Thomas Berg, Jiongxin Liu, Seung Woo Lee, Michelle L Alexander, David W Jacobs, and Peter N Belhumeur. Birdsnap: Large-scale fine-grained visual categorization of birds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2011–2018, 2014. + +Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In Computer vision and pattern recognition, 2006 IEEE computer society conference on, volume 2, pp. 1735–1742. IEEE, 2006. + +Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pp. 1026–1034, 2015. + +Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. + +Yen-Chang Hsu and Zsolt Kira. Neural network-based clustering using pairwise constraints. arXiv preprint arXiv:1511.06321, 2015. + +Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456, 2015. + +Jonathan Krause, Benjamin Sapp, Andrew Howard, Howard Zhou, Alexander Toshev, Tom Duerig, James Philbin, and Li Fei-Fei. The unreasonable effectiveness of noisy data for fine-grained recognition. In European Conference on Computer Vision, pp. 301–320. Springer, 2016. + +Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009. + +Harold W Kuhn. The hungarian method for the assignment problem. Naval Research Logistics (NRL), 2(1-2):83–97, 1955. + +Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´ document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. + +Sanghoon Lee and Melba M Crawford. Unsupervised multistage image classification using hierarchical clustering with a bayesian similarity measure. IEEE Transactions on Image Processing, 14 (3):312–320, 2005. + +Tomas Mikolov, Wen-tau Yih, and Geoffrey Zweig. Linguistic regularities in continuous space word representations. In hlt-Naacl, volume 13, pp. 746–751, 2013. + +Oren Rippel, Manohar Paluri, Piotr Dollar, and Lubomir Bourdev. Metric learning with adaptive density discrimination. International Conference on Learning Representations, 2016. + +Eleanor Rosch. Principles of categorization. Concepts: core readings, 189, 1999. + +Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015. + +Yevgeny Seldin, Sonia Starik, and Michael Werman. Unsupervised clustering of images using their joint segmentation. In Proceedings of the 3rd International Workshop on Statistical and Computational Theories of Vision (SCTV 2003), 2003. + +Alexander Strehl and Joydeep Ghosh. Cluster ensembles—a knowledge reuse framework for combining multiple partitions. Journal of machine learning research, 3(Dec):583–617, 2002. + +Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(Nov):2579–2605, 2008. + +Grant Van Horn and Pietro Perona. The devil is in the tails: Fine-grained classification in the wild. arXiv preprint arXiv:1709.01450, 2017. + +Cinna Wu, Mark Tygert, and Yann LeCun. Hierarchical loss for classification. arXiv preprint arXiv:1709.01062, 2017. + +# A PREPROCESSING STEPS + +We use two primary datasets, CIFAR10 and Birdsnap34, in experiments that substantiate our claims of improving on the classification task by means of using clustering representations as a warm-start. For the CIFAR10 dataset, we follow the preprocessing steps used by (Hsu & Kira, 2015). They are listed below, + +• Convert the images to YUV format. +• Calculate the mean and standard deviation of U and V channels over the entire training set. +• For each image, normalize by its Y channel mean and standard deviation while normalizing over the remaining channels using the aggregate mean and standard deviation over the entire training set. + +NOTE: The mean and standard deviation for U and V channel as per our calculations are (0.001, 0.003) and (0.227, 0.105). + +For the Birdsnap34 dataset, we perform the standard normalization using mean and standard deviations for R, G and B channels as (0.485, 0.229), (0.456, 0.224) and (0.406, 0.225). + +# B EXPERIMENTAL SETUP + +Table 6: Parameters used for Classification in Network A + +
ParametersValues
CIFAR10Total Epochs Batch Size Learning Rate Learning Rate Scale Factor95 16 0.01
Learning Rate Schedule Momentum Optimizer margin0.5 25,50,75,100 0.9 SGD
Weight Decay Weights of Losses Total Epochs Batch Size Learning Rate Learning1 gRate Scale Factor Birdsnap34 Learning Rate Schedule2.0 0.0005 N/A 195 16 0.01 0.5
+ +Table 7: Parameters used for Clustering in Network A + +
ParametersValues
CIFAR10Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay Weights of Losses0.5 25,50,75,100 0.9 SGD 2.0 0.0005
Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of LossesN/A 195 1024 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.001 N/A
+ +Table 8: Parameters used for Clustering in Network B + +
Parameters Total EpochsValues
CIFAR10Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay0.5 25,50,75,100 0.9 SGD 2.0
Weights of Losses Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of Losses0.0005 0.2, 0.8 195 128 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.0005 0.2, 0.8
+ +Table 9: Parameters used for Clustering in Network C + +
Parameters Total EpochsValues
CIFAR10Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay Weights of Losses0.5 25,50, 75,100 0.9 SGD 2.0 0.0005
Total Epochs Batch Size Learning Rate Learning gRate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of Losses0.4, 0.2, 0.4 195 128 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.0005 0.4, 0.2, 0.4
+ +Table 10: Parameters used for Clustering in Network D + +
Parameters Total EpochsValues
CIFAR10Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay0.5 25,50,75,100 0.9 SGD 2.0
Weights of Losses Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of Losses0.0005 0.4, 0.2, 0.4 195 128 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.0005 0.4, 0.2, 0.4
\ No newline at end of file diff --git a/parse/train/r1-4BLaQz/r1-4BLaQz_content_list.json b/parse/train/r1-4BLaQz/r1-4BLaQz_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..44a95dabc69d326b973d42ce25ec44d1f3be368b --- /dev/null +++ b/parse/train/r1-4BLaQz/r1-4BLaQz_content_list.json @@ -0,0 +1,1077 @@ +[ + { + "type": "text", + "text": "CLUSTER-BASED WARM-START NETS ", + "text_level": 1, + "bbox": [ + 176, + 98, + 629, + 121 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 145, + 398, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 210, + 544, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Theories in cognitive psychology postulate that humans use similarity as a basis for object categorization. However, work in image classification generally assumes disjoint and equally dissimilar classes to achieve super-human levels of performance on certain datasets. In our work, we adapt notions of similarity using weak labels over multiple hierarchical levels to boost classification performance. Instead of pitting clustering directly against classification, we use a warm-start based evaluation to explicitly provide value to a clustering representation by its ability to aid classification. We evaluate on CIFAR10 and a fine-grained classification dataset to show improvements in performance with the procedural addition of intermediate losses and weak labels based on multiple hierarchy levels. Furthermore, we show that pretraining AlexNet on hierarchical weak labels in conjunction with intermediate losses outperforms a classification baseline by over $17 \\%$ on a subset of Birdsnap dataset. Finally, we show improvement over AlexNet trained using ImageNet pre-trained weights as initializations which further supports our claim of the importance of similarity. ", + "bbox": [ + 233, + 241, + 764, + 449 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 474, + 336, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Similarity is one of the bases of object categorization in humans (Rosch, 1999). Theories of perceptual categorization in cognitive psychology postulate that humans construct categories by grouping similar stimuli to construct one or several prototypes. New instances are then labeled as the category that they are most similar to. For instance, when one categorizes a specific animal as a dog, they are saying that it is more similar to previously observed dogs than it is to all other objects. This view of categorization better explains the often fuzzy boundaries between real-life classes, where a new object may be equally similar to multiple prototypes. For example, while a dolphin is technically a mammal, its visual appearance is similar to fish, resulting in a lot of people misclassifying it. ", + "bbox": [ + 174, + 506, + 823, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While research in cognitive psychology on object categorization might indicate that similarity should play a central role in object classification in humans, image classification in computer vision seems to do pretty well without using it. The task is commonly interpreted as one of classifying an image into one of multiple classes that are assumed to be disjoint and equally dissimilar. The use of softmax-based loss functions assumes that the classes are disjoint, while the use of one-hot labels assumes that classes are equally dissimilar. Despite those strong assumptions, which seem to violate human notions of categories, image classification has shown remarkable progress, achieving superhuman performance on multiple datasets (Ioffe & Szegedy, 2015; He et al., 2015). Far from being an anomaly, state-of-the-art models across image classification benchmark datasets use losses, such as cross-entropy loss, that make the explicit assumption of disjoint classes. However, Hinton et al. (2015) notes that the predictions of ensembles of image classification models produces soft labels that “define a rich similarity structure over the data” despite the predictions of individual models not capturing this structure. ", + "bbox": [ + 174, + 625, + 825, + 804 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Can similarity-based metrics improve classification performance in convolutional neural networks? Previous work has tried to answer this question in different problems such as metric learning (Bellet et al., 2013), clustering (Seldin et al., 2003), and hierarchical classification (Lee & Crawford, 2005). We focus on applications of similarity in the context of convolutional neural networks such as Hsu & Kira (2015) who use a contrastive loss to perform end-to-end clustering using weak labels. While they show impressive performance on MNIST (LeCun et al., 1998) and CIFAR-10 (Krizhevsky & Hinton, 2009), their method does not scale well to more complex datasets. Rippel et al. (2016) propose a new loss function, magnet loss, to learn a representational space where distance corresponds to similarity. They show that clustering in this space allows them to achieve state-of-the-art performance on multiple fine-grained classification datasets.1 However, they initialize their model with pre-trained weights on ImageNet(Russakovsky et al., 2015), so it is not clear whether their model is learning a good representation, or if it is finding a good mapping between already learned representations and the labels. Wu et al. (2017) pose the problem as one of hierarchical classification and propose a loss based on an ultra-metric tree representation of a hierarchy over the labels. While their loss has interesting properties, it only outperforms classification losses for datasets with a small number of instances per class. In this work, we use a contrastive loss to improve the classification performance of randomly initialized convolutional neural networks on a fine-grained classification task. ", + "bbox": [ + 174, + 813, + 825, + 924 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 825, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "What measures of similarity should we teach our model? We represent the relations between our classes in a hierarchy where the labels are all leaf nodes. Therefore, there are two kinds of similarities that we want our model to capture. The first is intra-class similarity—all cats are similar to each other and they are dissimilar to other animals. The second is inter-class similarity—dogs and cats are more similar to each other than they are to non-living objects. For a more complex hierarchy, our model should learn similarity at different levels of the class hierarchy. Wu et al. (2017) use their hierarchical loss function to learn those similarities, however, they observe that reducing the similarity between two classes across all levels of the hierarchy to a single value biases the model towards correct classification at the higher levels of the hierarchy resulting in poor classification performance. Rippel et al. (2016) also observe that applying classification losses to the final layer reduces the entire representation of each class to a single scalar value which destroys intra- and inter-class variation. However, these are the same variations that we want our model to capture. We overcome those limitations by explicitly training the model to capture different grains of similarity at different levels of the network through applying an intermediate pairwise contrastive loss at those levels. In this manner, we can use hierarchical information, such as species of birds having the same coarse-grained category of bird while having different fine-grained labels. Despite being able to encode different levels of similarity, a contrastive loss does not require an explicit hierarchy; we only need weak labels for pairs of instances. ", + "bbox": [ + 173, + 250, + 825, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "How can we evaluate the representations learned by a clustering algorithm? Previous work has always tried to compete against classification using metrics biased towards the latter task. While some researchers have used hierarchical metrics (Wu et al., 2017) or qualitative measures (Rippel et al., 2016) to evaluate the quality of their representations, their primary evaluation criteria has consistently been the classification accuracy of their model. We propose another way to evaluate the clustering representations by using the clustering model as a “warm start” from which they can train on classification. The intuition is that if the clustering model learned a representations that captures similarity in a space, it would be be at an advantage when it is fine-tuned for classification. ", + "bbox": [ + 174, + 506, + 823, + 617 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose the use of a contrastive loss at different layers of a convolutional neural network to learn a notion of similarity across a set of labels. We also propose the use of the accuracy of a model pre-trained for clustering and fine-tuned for classification as an evaluation metric for clustering algorithms. ", + "bbox": [ + 174, + 625, + 823, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 INTERMEDIATE PAIRWISE CONTRASTIVE LOSS ", + "text_level": 1, + "bbox": [ + 174, + 713, + 596, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pairwise contrastive losses were first proposed by Hadsell et al. (2006) to learn a function that maps high dimensional inputs to lower dimensional outputs such that distances in the lower dimensional space approximate relationships in the input space. Relationships in the input space are not restricted to simple distance measures. In the context of image classification, the resulting function would have to learn the complex invariances that make two images of the same class similar, such that they are both mapped to points that are close to each other in the output space. This loss can be calculated using the representation of two points in the same embedding space and a binary label for whether or not they are similar. Hence, the loss only requires weak labels in the form of pairwise relationships. ", + "bbox": [ + 174, + 752, + 825, + 864 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a72c159113b92f8c536176837b61a6ca7237947ff8e534d0cbf2286e02239b7b.jpg", + "image_caption": [ + "Figure 1: Pairwise losses are applied to intermediate layers of the network as they are calculated through pairwise comparisons of the activations at those layers. " + ], + "image_footnote": [], + "bbox": [ + 250, + 108, + 758, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 LOSS FORMULATION ", + "text_level": 1, + "bbox": [ + 174, + 349, + 359, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $( X _ { 1 } , X _ { 2 } )$ be vectors that are calculated as mappings of two input instances in the same embedding space. $\\mathbb { 1 } _ { X _ { 1 } , X _ { 2 } }$ is an indicator variable that takes a value of one if $( X _ { 1 } , X _ { 2 } )$ are similar and zero otherwise. $D ( X , Y )$ can be any distance function between two vectors in the same space. $L _ { S } ( \\cdot )$ is a partial loss function for a similar pair, and $L _ { D } ( \\cdot )$ is a partial loss function for dissimilar pairs. Then the general formulation of the pairwise contrastive loss function is as shown in Eq. (1): ", + "bbox": [ + 173, + 376, + 825, + 446 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/ec1b27e93ebb2ef0da41150796dbb3489c15ac11fc98dafb1229eaa3f236febf.jpg", + "text": "$$\n\\boldsymbol { \\mathcal { L } } ( \\mathbf { X _ { 1 } } , \\mathbf { X _ { 2 } } ) = \\mathbb { 1 } _ { X _ { 1 } , X _ { 2 } } L _ { S } ( D ( \\mathbf { X _ { 1 } } , \\mathbf { X _ { 2 } } ) ) + ( 1 - \\mathbb { 1 } _ { X _ { 1 } , X _ { 2 } } ) L _ { D } ( D ( \\mathbf { X _ { 1 } } , \\mathbf { X _ { 2 } } ) ) .\n$$", + "text_format": "latex", + "bbox": [ + 251, + 465, + 746, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first half of Eq. (1) penalizes any divergence between the two vectors for similar instances while the second half of the loss rewards divergence up to a margin. Following the formulation proposed by (Hsu & Kira, 2015), we use Kullback-Leibler (KL) divergence as our distance metric and a hinge loss for dissimilar cases. The final formulation of the loss we use is shown in Eq. (2): ", + "bbox": [ + 174, + 493, + 825, + 550 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/0418768456cd74f938b93dc2cbe35706c6e52515ae1a5d795838820068268b5b.jpg", + "text": "$$\n\\begin{array} { r } { \\mathcal { L } ( \\mathbf { P } , \\mathbf { Q } ) = l o s s ( \\mathbf { P } , \\mathbf { Q } ) + l o s s ( \\mathbf { Q } , \\mathbf { P } ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 369, + 569, + 627, + 587 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/2f72a2f29051112b9ff5efb86e026de477cc48c737ced8c19365ed766c97a788.jpg", + "text": "$$\nl o s s ( \\mathbf { P } , \\mathbf { Q } ) = \\mathbb { 1 } _ { P , Q } K L ( \\mathbf { P } | | \\mathbf { Q } ) + ( 1 - \\mathbb { 1 } _ { P , Q } ) \\operatorname* { m a x } \\{ m a r g i n - K L ( \\mathbf { P } | | \\mathbf { Q } ) , 0 \\} ,\n$$", + "text_format": "latex", + "bbox": [ + 235, + 602, + 759, + 621 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/7f8124c73dced2aa0f78106eab2a851360b420115872be0ee912903754e96da6.jpg", + "text": "$$\nK L ( \\mathbf { P } \\parallel \\mathbf { Q } ) = \\sum _ { i = 1 } ^ { k } P _ { i } \\log \\Big ( \\frac { P _ { i } } { Q _ { i } } \\Big ) .\n$$", + "text_format": "latex", + "bbox": [ + 387, + 632, + 611, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We have found that a value of two for the margin works well. Since we apply the loss to activations at intermediate layers of the network, we first apply a softmax operation on the activations to obtain the two vectors, $P$ and $Q$ , used in Eq. (2). ", + "bbox": [ + 174, + 685, + 825, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 HIERARCHICAL SIMILARITY ", + "text_level": 1, + "bbox": [ + 176, + 744, + 413, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given that one of our objectives is to encode different levels of similarity into the representations learned by the network, we explore different ways of using the intermediate loss to achieve that goal as shown in Fig. 2. We establish a baseline network where the losses are all applied to the output of the last layer. We then use three different variants by applying intermediate losses at different levels of the network. In variant B, we apply the intermediate loss at the penultimate layer. This allows us to apply the contrastive-loss to a higher dimensional space where it could potentially learn a richer representation. In variants C and D, we provide the network with two different kinds of weak labels based on fine-grained and coarse-grained. Since we expect that a weaker form of similarity would be more beneficial earlier on in the network pipeline, we apply the coarse-grained loss to the secondto-last fully-connected layer in Network C. It should be noted that many instance pairs will have contradicting losses since they may match at the coarse-grained level but conflict at the fine-grained level. To tackle this, we augment the network with a skip layer around the second-to-last layer and apply the coarse-grained loss to the skip layer only. Hence, the coarse-grained loss will only provide additional information to the network without directly detracting from the representations passed to subsequent layers. While we show the network variants on AlexNet, the same ideas are applied to a variant of LeNet that has three fully-connected layers. ", + "bbox": [ + 173, + 770, + 825, + 924 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c751ec74231b40986e76cf6183577dc664e8807abf7061a09418db3bd1e4d8cc.jpg", + "image_caption": [ + "Figure 2: Different network architecture and losses. Network A is the baseline network. Network B applies the fine-grained intermediate loss to the penultimate layer as well. Network C extends Network B by applying the coarse-grained intermediate loss at the second-to-last layer. Network D extends Network B by adding a branch around the second-to-last layer, and applying the coarsegrained intermediate loss there. " + ], + "image_footnote": [], + "bbox": [ + 173, + 103, + 823, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 391, + 825, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We follow the efficient implementation regime proposed by (Hsu & Kira, 2015) in order to train a single network based on pairwise constraints. The original network formulation to handle pairwise constraints is a Siamese network. For every pairwise constraint, at least two data passes need to be performed which results in redundancies. Instead, we feed the softmax outcomes from a single network and the pairwise similarity constraints to the contrastive loss function which externally handles all possible pairwise constraints that are available within a mini-batch. We further extend the application of contrastive loss functions to multiple intermediate layers with a minimal overhead of extra processing and pairwise constraints, if alternate hierarchical labels are used. Details regarding the exact parameters and setup used to perform experiments are furnished in the Appendices. PyTorch2 is used for all implementations discussed within this paper. ", + "bbox": [ + 174, + 468, + 825, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 628, + 326, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate our approach on two different datasets using two different convolutional neural networks: LeNet on CIFAR10 and AlexNet on Birdsnap34–a 34-class subset of the Birdsnap dataset (Berg et al., 2014). We generate coarse-grained labels for CIFAR10 by separating the ten classes into six animal classes and four vehicle classes. Below is a discussion of how we constructed Birdsnap34. ", + "bbox": [ + 174, + 660, + 825, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Birdsnap34 Birdsnap is a popular fine-grained classification dataset that has 500 different classes and over 40,000 images. We annotated Birdsnap using the scientific classification of each bird to create a six-level hierarchy over all the labels. In this work, we use the birds’ subfamily and family as our fine-grained and coarse-grained labels, respectively. Using a pairwise loss with a large number of classes poses a significant implementation challenge. For a uniform dataset with $n$ classes, the ratio of similar pairs to dissimilar pairs is roughly $1 : ( n - 1 )$ . Hence, for a large number of classes, most of the pairs will be dissimilar resulting in a very weak signal to the model for clustering instances together. This challenge is analogous to that of imbalanced datasets (Van Horn & Perona, 2017), which is a major problem that lies beyond the scope of our paper. Instead of abandoning Birdsnap, we curate a 34-class subset, which we will refer to as Birdsnap34, and apply our approach to it. The classes are chosen such that each class comes from a unique biological subfamily of birds. Through sampling form unique subfamilies, we increase the interclass variance between our classes. The resulting dataset has 2982 training images and 158 test images that are divided into 34 fine-grained and 18 coarse-grained classes. We plan on extending our model to the entire Birdsnap dataset in future work. ", + "bbox": [ + 174, + 746, + 825, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 159 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All the network architectures used have randomly initialized weights unless specified otherwise. The models are evaluated using accuracy, purity, and Normalized Mutual Information (NMI) (Strehl & Ghosh, 2002). We train each model for a fixed number of epochs, and we report the top performing results out of 5 runs. Details regarding the exact parameters and setups used to perform the experiments are available in Appendix B. ", + "bbox": [ + 174, + 166, + 825, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 QUALITY OF WARM START AS AN EVALUATION METRIC ", + "text_level": 1, + "bbox": [ + 174, + 257, + 604, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our first experiment, we show the value of evaluating the representations learned by a clustering loss through using it as a warm start from training on classification. We replicate the setup used by Hsu & Kira (2015) by using LeNet on CIFAR10, and extend their framework to AlexNet on Birdsnap34. We use the Hungarian algorithm (Kuhn, 1955) to evaluate clustering accuracy. This method of evaluation assumes that the clusters will be centered across the different dimensions of the softmax output of the final layer. The Hungarian algorithm is then used to find the assignment of dimensions to classes that would achieve the highest accuracy. We use the Hungarian algorithm to evaluate clustering performance in all of our experiments. ", + "bbox": [ + 173, + 286, + 825, + 397 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/dda884d66cab9d7d7b697f66bc83d1770b76b7c4337ef8e82246a1eb8133716c.jpg", + "table_caption": [ + "Table 1: Baseline classification performance of LeNet and AlexNet " + ], + "table_footnote": [], + "table_body": "
CIFAR-10Birdsnap-34
Classification75.4853.50
Clustering74.2117.20
Warm Start Classification75.7550.32
", + "bbox": [ + 305, + 446, + 689, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From Table 1 we observe that for CIFAR10, the application of the loss to the final layer results in clustering performance almost matching classification performance. However, in more complex dataset such as Birdsnap34, this pattern disappears. This supports the pattern shown in (Hsu & Kira, 2015) where the gap between clustering and classification performance increases when they move from MNIST to CIFAR10. The Cross-Entropy loss formulation, as mentioned in (Hsu & Kira, 2015), is extremely harsh and seeks to segment each and every sample into one absolute class. The tSNE (van der Maaten & Hinton, 2008) outcomes from the last fully connected layer of LeNet in Fig. 3 match our observations. ", + "bbox": [ + 173, + 531, + 825, + 643 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a8990135f6d1aafc86247b26a23967600823155298c1769f10985048200a7f11.jpg", + "image_caption": [ + "Figure 3: t-SNE Visualization of the outputs of LeNet on CIFAR10 " + ], + "image_footnote": [], + "bbox": [ + 210, + 664, + 792, + 882 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 INTERMEDIATE LOSS ", + "text_level": 1, + "bbox": [ + 176, + 103, + 362, + 117 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The previous experiment indicates that the representations learned from end-to-end clustering do not extend beyond simple datasets. We posit that this is a result of how the clustering loss is applied. Most machine learning models only apply losses to their last layer. This is generally a reasonable thing to do since we only have labels for the output of the last layer. On the other hand, pairwise losses can be applied at any level of the network since they only require a pair of vectors in the same embedding space. In this experiment we explore the potential of applying losses at intermediate layers on the learned representations. ", + "bbox": [ + 173, + 128, + 825, + 227 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/560b62ba4f40d2eea4ed79f5986a5c606e20e79a7a30c96a948e7e38acfe160e.jpg", + "table_caption": [ + "Table 2: Using intermediate losses " + ], + "table_footnote": [], + "table_body": "
Clustering Acc.NMIPurityClassification Acc.
CIFAR10Cross-Entropy LeNet A LeNet B1 74.21 72.661 0.551 0.7475.48 75.75
Birdsnap34Cross-Entropy AlexNet A AlexNetB1 17.20 20.380.54 1 0.39 0.450.73 1 0.17 0.2075.42 53.50 50.32 45.86
", + "bbox": [ + 202, + 272, + 795, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train two variants of LeNet and AlexNet using the pairwise contrastive loss. Variant A follows the setup used by Hsu & Kira (2015), while Variant B applies the same loss at both the final and penultimate layer of the convolutional neural network. We also train the same network using CrossEntropy loss to establish baseline performance. As shown in Table 2, performance on CIFAR10 is almost the same across all three settings. Meanwhile, we see clear differences on Birdsnap34. There is a significant increase across all metrics for the clustering models for the intermediate loss version, however, this improvement does not carry over to the classification accuracy. ", + "bbox": [ + 173, + 393, + 825, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 HIERARCHICAL CLUSTERING ", + "text_level": 1, + "bbox": [ + 176, + 507, + 419, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Previous work on embedding spaces has shown that adding structure to the embedding space or learning it using a hierarchical structure often validates its integrity as well as improves performance (Mikolov et al., 2013). In an effort to understand the impact of adding a hierarchical-labelbased contrastive loss to our baseline networks, we apply it on variants C and D of LeNet and AlexNet. Since the restriction on dimensionality is removed by using a contrastive loss with KL divergence, we can directly apply the same loss to an alternate set of weak labels and representations. Our ultimate expectation is that adding hierarchical structure should aid in improving classification, given their potency to improve embedding spaces as shown in previous works. ", + "bbox": [ + 173, + 534, + 825, + 646 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/267382a8c11ea2d13a8405e789dc3ec7b077d5b50a014517a25d438aa1015309.jpg", + "table_caption": [ + "Table 3: Impact of adding heirarchy based weak labels " + ], + "table_footnote": [], + "table_body": "
Clustering Acc.NMIPurityClassification Acc.
CIFAR10Cross-Entropy75.48
LeNet A74.210.550.7475.75
LeNet B72.660.540.7375.42
LeNet C72.840.550.7375.11
LeNet D74.170.560.7475.53
Birdsnap34Cross-Entropy11153.50
AlexNet A17.200.390.1750.32
AlexNet B20.380.450.2045.86
AlexNet C24.840.510.2570.70
AlexNet D22.290.490.2265.61
", + "bbox": [ + 202, + 690, + 795, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 shows that the experimental results align with our expectations in terms of overall improvement in classification. As seen in other experiments, the performance on CIFAR10 is almost the same across all metrics. Meanwhile, there is a spike in performance for Birdsnap34 for both AlexNet C and D with the metrics far exceeding classification performance. Given that Birdsnap34 has a more well-defined hierarchy, it would be expected that using the hierarchical labels to learn similarity would result in learning good representations. Another thing worth noting is that the dataset is curated by having subfamilies as fine-grained labels and biological families as coarsegrained labels. This structure results in similarity along this tree corresponding to distinct visual features. This might be indicating that choice of hierarchy can have an impact on performance. A surprising finding is that AlexNet C outperforms AlexNet D. One would expect AlexNet D to be able to extract features for both the coarse and fine levels of the hierarchy. We plan on investigating this in future work. ", + "bbox": [ + 173, + 867, + 825, + 924 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 214 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 SENSITIVITY TO NETWORK SIZE AND WEIGHT INITIALIZATION", + "text_level": 1, + "bbox": [ + 176, + 234, + 655, + 248 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the ability of standard end-to-end clustering networks to perform fine-tuned classification reasonably well, we test their sensitivity towards network size in order to understand if they would be able to achieve or outperform their simpler baseline classification counterparts. In order to do so, we double the number of convolutional filters and the sizes of fully-connected layers within AlexNet. We use their doubled networks to perform clustering and fine-tune them for classification. We denote these modified networks with a suffix of “Double.” ", + "bbox": [ + 174, + 261, + 825, + 344 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fe2b3905cbc46bc7cf256753eb284946a22b77a879476085c03d23a987fe7014.jpg", + "table_caption": [ + "Table 4: Sensitivity to network capacity " + ], + "table_footnote": [], + "table_body": "
Clustering Acc.NMIPurityClassification Acc.
Cross-Entropy一1153.50
AlexNet A (Double)22.930.490.2359.24
AlexNet C24.840.510.2570.70
", + "bbox": [ + 232, + 390, + 764, + 450 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f8cfe17876cbbaf94ee482764253089d623f6643bb4551bbd2477838da5a2c46.jpg", + "table_caption": [ + "Table 5: Comparison against ImageNet pre-trained weights " + ], + "table_footnote": [], + "table_body": "
Clustering Acc.NMIPurityClassification Acc.
Cross-Entropy11181.53
AlexNet C24.200.510.2482.80
", + "bbox": [ + 250, + 512, + 743, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From Table 4, we clearly observe that clustering is highly sensitive to network capacity. Compared to AlexNet A, its doubled counterpart allows clustering to learn better representations which leads to classificaiton performance being higher than the baseline classification model. However, our AlexNet C variant outperforms the double capacity network. This points to the efficacy of the combination of weak labels over multiple levels of heirarchy and the use of intermediate losses. Furthermore, Table 5 shows that even we use ImageNet pre-trained weights, variant C outperforms the ImageNet pre-trained classification baseline. This points to a tractable and viable approach to extending performance over more complex datasets without the need to increase the capacity of neural networks. ", + "bbox": [ + 173, + 582, + 825, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 DISCUSSION AND FUTURE WORK ", + "text_level": 1, + "bbox": [ + 176, + 731, + 485, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this work, we argue that similarity-based losses can be applied as a warm start to boost the performance of image classification models. We show that applying a pairwise contrastive loss to intermediate layers of the network results in better clustering performance, as well as better finetuned classification performance. Furthermore, we demonstrate how the pairwise loss can be used to train a model using weak hierarchical labels. ", + "bbox": [ + 174, + 763, + 823, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We find that training with hierarchical labels results in the highest performance beating models with more parameters and approaches the performance of models pre-trained with ImageNet weights. This is a very significant finding since ImageNet contains multiple bird classes, so a model pretrained with ImageNet has seen many more birds than a model pre-trained with pairwise constraints on Birdsnap34. Nevertheless, it only outperforms a cluster-based warm-start model by just $10 \\%$ . Regardless, we also show that applying our approach to ImageNet weights still results in a boost of $1 . 2 7 \\%$ . This supports our claim that similarity-based metrics can improve classification performance, but it suggests that the gains we expect decrease if we start of with a good representational space. ", + "bbox": [ + 174, + 840, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 821, + 145 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We hope to expand this in future work in multiple ways. We plan on extending our approach to the entire Birdsnap dataset, as well as to other fine-grained classification datasets. We also hope to perform more extensive analysis of the quality of the embedding spaces learned as it is obvious that the use of the Hungarian algorithm does not accurately reflect the performance of the clustering model. ", + "bbox": [ + 174, + 152, + 825, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 243, + 285, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Aurelien Bellet, Amaury Habrard, and Marc Sebban. A survey on metric learning for feature vectors ´ and structured data. arXiv preprint arXiv:1306.6709, 2013. ", + "bbox": [ + 171, + 266, + 825, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Thomas Berg, Jiongxin Liu, Seung Woo Lee, Michelle L Alexander, David W Jacobs, and Peter N Belhumeur. Birdsnap: Large-scale fine-grained visual categorization of birds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2011–2018, 2014. ", + "bbox": [ + 178, + 303, + 823, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In Computer vision and pattern recognition, 2006 IEEE computer society conference on, volume 2, pp. 1735–1742. IEEE, 2006. ", + "bbox": [ + 178, + 354, + 821, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pp. 1026–1034, 2015. ", + "bbox": [ + 174, + 405, + 823, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. ", + "bbox": [ + 174, + 457, + 821, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Yen-Chang Hsu and Zsolt Kira. Neural network-based clustering using pairwise constraints. arXiv preprint arXiv:1511.06321, 2015. ", + "bbox": [ + 176, + 494, + 821, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456, 2015. ", + "bbox": [ + 174, + 531, + 825, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Jonathan Krause, Benjamin Sapp, Andrew Howard, Howard Zhou, Alexander Toshev, Tom Duerig, James Philbin, and Li Fei-Fei. The unreasonable effectiveness of noisy data for fine-grained recognition. In European Conference on Computer Vision, pp. 301–320. Springer, 2016. ", + "bbox": [ + 174, + 582, + 825, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009. ", + "bbox": [ + 171, + 633, + 820, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Harold W Kuhn. The hungarian method for the assignment problem. Naval Research Logistics (NRL), 2(1-2):83–97, 1955. ", + "bbox": [ + 174, + 657, + 823, + 685 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´ document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. ", + "bbox": [ + 174, + 694, + 821, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sanghoon Lee and Melba M Crawford. Unsupervised multistage image classification using hierarchical clustering with a bayesian similarity measure. IEEE Transactions on Image Processing, 14 (3):312–320, 2005. ", + "bbox": [ + 173, + 732, + 825, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tomas Mikolov, Wen-tau Yih, and Geoffrey Zweig. Linguistic regularities in continuous space word representations. In hlt-Naacl, volume 13, pp. 746–751, 2013. ", + "bbox": [ + 174, + 784, + 821, + 813 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Oren Rippel, Manohar Paluri, Piotr Dollar, and Lubomir Bourdev. Metric learning with adaptive density discrimination. International Conference on Learning Representations, 2016. ", + "bbox": [ + 176, + 820, + 825, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Eleanor Rosch. Principles of categorization. Concepts: core readings, 189, 1999. ", + "bbox": [ + 174, + 858, + 710, + 873 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015. ", + "bbox": [ + 176, + 882, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Yevgeny Seldin, Sonia Starik, and Michael Werman. Unsupervised clustering of images using their joint segmentation. In Proceedings of the 3rd International Workshop on Statistical and Computational Theories of Vision (SCTV 2003), 2003. ", + "bbox": [ + 176, + 103, + 823, + 146 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Alexander Strehl and Joydeep Ghosh. Cluster ensembles—a knowledge reuse framework for combining multiple partitions. Journal of machine learning research, 3(Dec):583–617, 2002. ", + "bbox": [ + 173, + 155, + 821, + 184 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(Nov):2579–2605, 2008. ", + "bbox": [ + 171, + 193, + 823, + 222 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Grant Van Horn and Pietro Perona. The devil is in the tails: Fine-grained classification in the wild. arXiv preprint arXiv:1709.01450, 2017. ", + "bbox": [ + 171, + 229, + 823, + 258 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Cinna Wu, Mark Tygert, and Yann LeCun. Hierarchical loss for classification. arXiv preprint arXiv:1709.01062, 2017. ", + "bbox": [ + 169, + 268, + 823, + 296 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A PREPROCESSING STEPS ", + "text_level": 1, + "bbox": [ + 176, + 102, + 406, + 118 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We use two primary datasets, CIFAR10 and Birdsnap34, in experiments that substantiate our claims of improving on the classification task by means of using clustering representations as a warm-start. For the CIFAR10 dataset, we follow the preprocessing steps used by (Hsu & Kira, 2015). They are listed below, ", + "bbox": [ + 174, + 133, + 825, + 189 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "• Convert the images to YUV format. \n• Calculate the mean and standard deviation of U and V channels over the entire training set. \n• For each image, normalize by its Y channel mean and standard deviation while normalizing over the remaining channels using the aggregate mean and standard deviation over the entire training set. ", + "bbox": [ + 215, + 200, + 825, + 281 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "NOTE: The mean and standard deviation for U and V channel as per our calculations are (0.001, 0.003) and (0.227, 0.105). ", + "bbox": [ + 171, + 292, + 823, + 321 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For the Birdsnap34 dataset, we perform the standard normalization using mean and standard deviations for R, G and B channels as (0.485, 0.229), (0.456, 0.224) and (0.406, 0.225). ", + "bbox": [ + 173, + 328, + 823, + 357 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "B EXPERIMENTAL SETUP ", + "text_level": 1, + "bbox": [ + 176, + 377, + 401, + 393 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/d109f64e1e19bf4dd834b8e96d8eaa5fdfa21c10b6af5980ce49aef3aade5439.jpg", + "table_caption": [ + "Table 6: Parameters used for Classification in Network A " + ], + "table_footnote": [], + "table_body": "
ParametersValues
CIFAR10Total Epochs Batch Size Learning Rate Learning Rate Scale Factor95 16 0.01
Learning Rate Schedule Momentum Optimizer margin0.5 25,50,75,100 0.9 SGD
Weight Decay Weights of Losses Total Epochs Batch Size Learning Rate Learning1 gRate Scale Factor Birdsnap34 Learning Rate Schedule2.0 0.0005 N/A 195 16 0.01 0.5
", + "bbox": [ + 284, + 438, + 714, + 734 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/2b2c8f03bfe04e64f0b4715f232253e163b02cfb498b4c6bf2c253df1ccd9cbc.jpg", + "table_caption": [ + "Table 7: Parameters used for Clustering in Network A " + ], + "table_footnote": [], + "table_body": "
ParametersValues
CIFAR10Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay Weights of Losses0.5 25,50,75,100 0.9 SGD 2.0 0.0005
Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of LossesN/A 195 1024 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.001 N/A
", + "bbox": [ + 284, + 169, + 714, + 464 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d965d87dc528603f4b223475197e0c8e5e3b8a84147858383ec90a6638cf1c8a.jpg", + "table_caption": [ + "Table 8: Parameters used for Clustering in Network B " + ], + "table_footnote": [], + "table_body": "
Parameters Total EpochsValues
CIFAR10Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay0.5 25,50,75,100 0.9 SGD 2.0
Weights of Losses Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of Losses0.0005 0.2, 0.8 195 128 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.0005 0.2, 0.8
", + "bbox": [ + 284, + 583, + 714, + 878 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4dc15f0a3b71bc97215371bdfe733a0c14e1a8afac9bcc2012e4a382a9a5eae5.jpg", + "table_caption": [ + "Table 9: Parameters used for Clustering in Network C " + ], + "table_footnote": [], + "table_body": "
Parameters Total EpochsValues
CIFAR10Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay Weights of Losses0.5 25,50, 75,100 0.9 SGD 2.0 0.0005
Total Epochs Batch Size Learning Rate Learning gRate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of Losses0.4, 0.2, 0.4 195 128 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.0005 0.4, 0.2, 0.4
", + "bbox": [ + 284, + 169, + 714, + 464 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3b325157bb9a4216b8f5fa8baff35b1fb724c5eb554b439ebc802f6db4a8f58f.jpg", + "table_caption": [ + "Table 10: Parameters used for Clustering in Network D " + ], + "table_footnote": [], + "table_body": "
Parameters Total EpochsValues
CIFAR10Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay0.5 25,50,75,100 0.9 SGD 2.0
Weights of Losses Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of Losses0.0005 0.4, 0.2, 0.4 195 128 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.0005 0.4, 0.2, 0.4
", + "bbox": [ + 284, + 583, + 714, + 878 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/parse/train/r1-4BLaQz/r1-4BLaQz_middle.json b/parse/train/r1-4BLaQz/r1-4BLaQz_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..2cfac5efbc9c13ba0f4ff6dbbb3c62c3549c3de9 --- /dev/null +++ b/parse/train/r1-4BLaQz/r1-4BLaQz_middle.json @@ -0,0 +1,22718 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 385, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 387, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 387, + 97 + ], + "score": 1.0, + "content": "CLUSTER-BASED WARM-START NETS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 112, + 115, + 244, + 137 + ], + "lines": [ + { + "bbox": [ + 113, + 115, + 201, + 127 + ], + "spans": [ + { + "bbox": [ + 113, + 115, + 201, + 127 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 111, + 126, + 245, + 138 + ], + "spans": [ + { + "bbox": [ + 111, + 126, + 245, + 138 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 278, + 167, + 333, + 178 + ], + "lines": [ + { + "bbox": [ + 276, + 165, + 336, + 181 + ], + "spans": [ + { + "bbox": [ + 276, + 165, + 336, + 181 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 143, + 191, + 468, + 356 + ], + "lines": [ + { + "bbox": [ + 142, + 191, + 470, + 204 + ], + "spans": [ + { + "bbox": [ + 142, + 191, + 470, + 204 + ], + "score": 1.0, + "content": "Theories in cognitive psychology postulate that humans use similarity as a basis", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 203, + 470, + 215 + ], + "spans": [ + { + "bbox": [ + 141, + 203, + 470, + 215 + ], + "score": 1.0, + "content": "for object categorization. However, work in image classification generally as-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 214, + 470, + 225 + ], + "spans": [ + { + "bbox": [ + 141, + 214, + 470, + 225 + ], + "score": 1.0, + "content": "sumes disjoint and equally dissimilar classes to achieve super-human levels of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 140, + 224, + 470, + 238 + ], + "spans": [ + { + "bbox": [ + 140, + 224, + 470, + 238 + ], + "score": 1.0, + "content": "performance on certain datasets. In our work, we adapt notions of similarity using", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 235, + 469, + 248 + ], + "spans": [ + { + "bbox": [ + 141, + 235, + 469, + 248 + ], + "score": 1.0, + "content": "weak labels over multiple hierarchical levels to boost classification performance.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 246, + 470, + 259 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 470, + 259 + ], + "score": 1.0, + "content": "Instead of pitting clustering directly against classification, we use a warm-start", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 257, + 469, + 270 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 469, + 270 + ], + "score": 1.0, + "content": "based evaluation to explicitly provide value to a clustering representation by its", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 268, + 470, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 470, + 281 + ], + "score": 1.0, + "content": "ability to aid classification. We evaluate on CIFAR10 and a fine-grained classifi-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 280, + 469, + 291 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 469, + 291 + ], + "score": 1.0, + "content": "cation dataset to show improvements in performance with the procedural addition", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 290, + 469, + 302 + ], + "spans": [ + { + "bbox": [ + 142, + 290, + 469, + 302 + ], + "score": 1.0, + "content": "of intermediate losses and weak labels based on multiple hierarchy levels. Further-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 301, + 469, + 313 + ], + "spans": [ + { + "bbox": [ + 141, + 301, + 469, + 313 + ], + "score": 1.0, + "content": "more, we show that pretraining AlexNet on hierarchical weak labels in conjunc-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 312, + 469, + 324 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 436, + 324 + ], + "score": 1.0, + "content": "tion with intermediate losses outperforms a classification baseline by over", + "type": "text" + }, + { + "bbox": [ + 437, + 312, + 456, + 322 + ], + "score": 0.88, + "content": "17 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 312, + 469, + 324 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 323, + 470, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 470, + 335 + ], + "score": 1.0, + "content": "a subset of Birdsnap dataset. Finally, we show improvement over AlexNet trained", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 333, + 470, + 347 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 470, + 347 + ], + "score": 1.0, + "content": "using ImageNet pre-trained weights as initializations which further supports our", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 345, + 292, + 358 + ], + "spans": [ + { + "bbox": [ + 141, + 345, + 292, + 358 + ], + "score": 1.0, + "content": "claim of the importance of similarity.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 108, + 376, + 206, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 375, + 208, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 208, + 392 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 401, + 504, + 490 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 505, + 415 + ], + "score": 1.0, + "content": "Similarity is one of the bases of object categorization in humans (Rosch, 1999). Theories of percep-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 412, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 426 + ], + "score": 1.0, + "content": "tual categorization in cognitive psychology postulate that humans construct categories by grouping", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "score": 1.0, + "content": "similar stimuli to construct one or several prototypes. New instances are then labeled as the category", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "score": 1.0, + "content": "that they are most similar to. For instance, when one categorizes a specific animal as a dog, they are", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "score": 1.0, + "content": "saying that it is more similar to previously observed dogs than it is to all other objects. This view", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "of categorization better explains the often fuzzy boundaries between real-life classes, where a new", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 467, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 482 + ], + "score": 1.0, + "content": "object may be equally similar to multiple prototypes. For example, while a dolphin is technically a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 479, + 478, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 478, + 491 + ], + "score": 1.0, + "content": "mammal, its visual appearance is similar to fish, resulting in a lot of people misclassifying it.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 495, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "While research in cognitive psychology on object categorization might indicate that similarity should", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "score": 1.0, + "content": "play a central role in object classification in humans, image classification in computer vision seems", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 517, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 506, + 531 + ], + "score": 1.0, + "content": "to do pretty well without using it. The task is commonly interpreted as one of classifying an image", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 529, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 506, + 540 + ], + "score": 1.0, + "content": "into one of multiple classes that are assumed to be disjoint and equally dissimilar. The use of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 540, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 506, + 552 + ], + "score": 1.0, + "content": "softmax-based loss functions assumes that the classes are disjoint, while the use of one-hot labels", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "assumes that classes are equally dissimilar. Despite those strong assumptions, which seem to violate", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "human notions of categories, image classification has shown remarkable progress, achieving super-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 570, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 506, + 586 + ], + "score": 1.0, + "content": "human performance on multiple datasets (Ioffe & Szegedy, 2015; He et al., 2015). Far from being", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "an anomaly, state-of-the-art models across image classification benchmark datasets use losses, such", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "as cross-entropy loss, that make the explicit assumption of disjoint classes. However, Hinton et al.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "(2015) notes that the predictions of ensembles of image classification models produces soft labels", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "that “define a rich similarity structure over the data” despite the predictions of individual models not", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 628, + 203, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 203, + 639 + ], + "score": 1.0, + "content": "capturing this structure.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "Can similarity-based metrics improve classification performance in convolutional neural networks?", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "Previous work has tried to answer this question in different problems such as metric learning (Bellet", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "et al., 2013), clustering (Seldin et al., 2003), and hierarchical classification (Lee & Crawford, 2005).", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "We focus on applications of similarity in the context of convolutional neural networks such as Hsu", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "& Kira (2015) who use a contrastive loss to perform end-to-end clustering using weak labels. While", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "they show impressive performance on MNIST (LeCun et al., 1998) and CIFAR-10 (Krizhevsky &", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Hinton, 2009), their method does not scale well to more complex datasets. Rippel et al. (2016) pro-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "pose a new loss function, magnet loss, to learn a representational space where distance corresponds", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 44.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 385, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 387, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 387, + 97 + ], + "score": 1.0, + "content": "CLUSTER-BASED WARM-START NETS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 112, + 115, + 244, + 137 + ], + "lines": [ + { + "bbox": [ + 113, + 115, + 201, + 127 + ], + "spans": [ + { + "bbox": [ + 113, + 115, + 201, + 127 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 111, + 126, + 245, + 138 + ], + "spans": [ + { + "bbox": [ + 111, + 126, + 245, + 138 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 111, + 115, + 245, + 138 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 167, + 333, + 178 + ], + "lines": [ + { + "bbox": [ + 276, + 165, + 336, + 181 + ], + "spans": [ + { + "bbox": [ + 276, + 165, + 336, + 181 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 143, + 191, + 468, + 356 + ], + "lines": [ + { + "bbox": [ + 142, + 191, + 470, + 204 + ], + "spans": [ + { + "bbox": [ + 142, + 191, + 470, + 204 + ], + "score": 1.0, + "content": "Theories in cognitive psychology postulate that humans use similarity as a basis", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 203, + 470, + 215 + ], + "spans": [ + { + "bbox": [ + 141, + 203, + 470, + 215 + ], + "score": 1.0, + "content": "for object categorization. However, work in image classification generally as-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 214, + 470, + 225 + ], + "spans": [ + { + "bbox": [ + 141, + 214, + 470, + 225 + ], + "score": 1.0, + "content": "sumes disjoint and equally dissimilar classes to achieve super-human levels of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 140, + 224, + 470, + 238 + ], + "spans": [ + { + "bbox": [ + 140, + 224, + 470, + 238 + ], + "score": 1.0, + "content": "performance on certain datasets. In our work, we adapt notions of similarity using", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 235, + 469, + 248 + ], + "spans": [ + { + "bbox": [ + 141, + 235, + 469, + 248 + ], + "score": 1.0, + "content": "weak labels over multiple hierarchical levels to boost classification performance.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 246, + 470, + 259 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 470, + 259 + ], + "score": 1.0, + "content": "Instead of pitting clustering directly against classification, we use a warm-start", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 257, + 469, + 270 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 469, + 270 + ], + "score": 1.0, + "content": "based evaluation to explicitly provide value to a clustering representation by its", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 268, + 470, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 470, + 281 + ], + "score": 1.0, + "content": "ability to aid classification. We evaluate on CIFAR10 and a fine-grained classifi-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 280, + 469, + 291 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 469, + 291 + ], + "score": 1.0, + "content": "cation dataset to show improvements in performance with the procedural addition", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 290, + 469, + 302 + ], + "spans": [ + { + "bbox": [ + 142, + 290, + 469, + 302 + ], + "score": 1.0, + "content": "of intermediate losses and weak labels based on multiple hierarchy levels. Further-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 301, + 469, + 313 + ], + "spans": [ + { + "bbox": [ + 141, + 301, + 469, + 313 + ], + "score": 1.0, + "content": "more, we show that pretraining AlexNet on hierarchical weak labels in conjunc-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 312, + 469, + 324 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 436, + 324 + ], + "score": 1.0, + "content": "tion with intermediate losses outperforms a classification baseline by over", + "type": "text" + }, + { + "bbox": [ + 437, + 312, + 456, + 322 + ], + "score": 0.88, + "content": "17 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 312, + 469, + 324 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 323, + 470, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 470, + 335 + ], + "score": 1.0, + "content": "a subset of Birdsnap dataset. Finally, we show improvement over AlexNet trained", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 333, + 470, + 347 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 470, + 347 + ], + "score": 1.0, + "content": "using ImageNet pre-trained weights as initializations which further supports our", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 345, + 292, + 358 + ], + "spans": [ + { + "bbox": [ + 141, + 345, + 292, + 358 + ], + "score": 1.0, + "content": "claim of the importance of similarity.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 11, + "bbox_fs": [ + 140, + 191, + 470, + 358 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 376, + 206, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 375, + 208, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 208, + 392 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 401, + 504, + 490 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 505, + 415 + ], + "score": 1.0, + "content": "Similarity is one of the bases of object categorization in humans (Rosch, 1999). Theories of percep-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 412, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 426 + ], + "score": 1.0, + "content": "tual categorization in cognitive psychology postulate that humans construct categories by grouping", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "score": 1.0, + "content": "similar stimuli to construct one or several prototypes. New instances are then labeled as the category", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "score": 1.0, + "content": "that they are most similar to. For instance, when one categorizes a specific animal as a dog, they are", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "score": 1.0, + "content": "saying that it is more similar to previously observed dogs than it is to all other objects. This view", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "of categorization better explains the often fuzzy boundaries between real-life classes, where a new", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 467, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 482 + ], + "score": 1.0, + "content": "object may be equally similar to multiple prototypes. For example, while a dolphin is technically a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 479, + 478, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 478, + 491 + ], + "score": 1.0, + "content": "mammal, its visual appearance is similar to fish, resulting in a lot of people misclassifying it.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 400, + 506, + 491 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 495, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "While research in cognitive psychology on object categorization might indicate that similarity should", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "score": 1.0, + "content": "play a central role in object classification in humans, image classification in computer vision seems", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 517, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 506, + 531 + ], + "score": 1.0, + "content": "to do pretty well without using it. The task is commonly interpreted as one of classifying an image", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 529, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 506, + 540 + ], + "score": 1.0, + "content": "into one of multiple classes that are assumed to be disjoint and equally dissimilar. The use of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 540, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 506, + 552 + ], + "score": 1.0, + "content": "softmax-based loss functions assumes that the classes are disjoint, while the use of one-hot labels", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "assumes that classes are equally dissimilar. Despite those strong assumptions, which seem to violate", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "human notions of categories, image classification has shown remarkable progress, achieving super-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 570, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 506, + 586 + ], + "score": 1.0, + "content": "human performance on multiple datasets (Ioffe & Szegedy, 2015; He et al., 2015). Far from being", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "an anomaly, state-of-the-art models across image classification benchmark datasets use losses, such", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "as cross-entropy loss, that make the explicit assumption of disjoint classes. However, Hinton et al.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "(2015) notes that the predictions of ensembles of image classification models produces soft labels", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "that “define a rich similarity structure over the data” despite the predictions of individual models not", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 628, + 203, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 203, + 639 + ], + "score": 1.0, + "content": "capturing this structure.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 495, + 506, + 639 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "Can similarity-based metrics improve classification performance in convolutional neural networks?", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "Previous work has tried to answer this question in different problems such as metric learning (Bellet", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "et al., 2013), clustering (Seldin et al., 2003), and hierarchical classification (Lee & Crawford, 2005).", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "We focus on applications of similarity in the context of convolutional neural networks such as Hsu", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "& Kira (2015) who use a contrastive loss to perform end-to-end clustering using weak labels. While", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "they show impressive performance on MNIST (LeCun et al., 1998) and CIFAR-10 (Krizhevsky &", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Hinton, 2009), their method does not scale well to more complex datasets. Rippel et al. (2016) pro-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "pose a new loss function, magnet loss, to learn a representational space where distance corresponds", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "to similarity. They show that clustering in this space allows them to achieve state-of-the-art perfor-", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "mance on multiple fine-grained classification datasets.1 However, they initialize their model with", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "score": 1.0, + "content": "pre-trained weights on ImageNet(Russakovsky et al., 2015), so it is not clear whether their model", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 114, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 104, + 114, + 505, + 129 + ], + "score": 1.0, + "content": "is learning a good representation, or if it is finding a good mapping between already learned rep-", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "resentations and the labels. Wu et al. (2017) pose the problem as one of hierarchical classification", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 138, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 149 + ], + "score": 1.0, + "content": "and propose a loss based on an ultra-metric tree representation of a hierarchy over the labels. While", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "their loss has interesting properties, it only outperforms classification losses for datasets with a small", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "number of instances per class. In this work, we use a contrastive loss to improve the classification", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "performance of randomly initialized convolutional neural networks on a fine-grained classification", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 128, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 128, + 193 + ], + "score": 1.0, + "content": "task.", + "type": "text", + "cross_page": true + } + ], + "index": 9 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 643, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "to similarity. They show that clustering in this space allows them to achieve state-of-the-art perfor-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "mance on multiple fine-grained classification datasets.1 However, they initialize their model with", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "score": 1.0, + "content": "pre-trained weights on ImageNet(Russakovsky et al., 2015), so it is not clear whether their model", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 114, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 104, + 114, + 505, + 129 + ], + "score": 1.0, + "content": "is learning a good representation, or if it is finding a good mapping between already learned rep-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "resentations and the labels. Wu et al. (2017) pose the problem as one of hierarchical classification", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 138, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 149 + ], + "score": 1.0, + "content": "and propose a loss based on an ultra-metric tree representation of a hierarchy over the labels. While", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "their loss has interesting properties, it only outperforms classification losses for datasets with a small", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "number of instances per class. In this work, we use a contrastive loss to improve the classification", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "performance of randomly initialized convolutional neural networks on a fine-grained classification", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 128, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 128, + 193 + ], + "score": 1.0, + "content": "task.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 106, + 199, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 210 + ], + "score": 1.0, + "content": "What measures of similarity should we teach our model? We represent the relations between our", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 209, + 504, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 504, + 221 + ], + "score": 1.0, + "content": "classes in a hierarchy where the labels are all leaf nodes. Therefore, there are two kinds of similari-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 221, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 505, + 232 + ], + "score": 1.0, + "content": "ties that we want our model to capture. The first is intra-class similarity—all cats are similar to each", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "other and they are dissimilar to other animals. The second is inter-class similarity—dogs and cats", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 241, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 104, + 241, + 506, + 255 + ], + "score": 1.0, + "content": "are more similar to each other than they are to non-living objects. For a more complex hierarchy,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "our model should learn similarity at different levels of the class hierarchy. Wu et al. (2017) use", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "their hierarchical loss function to learn those similarities, however, they observe that reducing the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 275, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 286 + ], + "score": 1.0, + "content": "similarity between two classes across all levels of the hierarchy to a single value biases the model", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 505, + 298 + ], + "score": 1.0, + "content": "towards correct classification at the higher levels of the hierarchy resulting in poor classification", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 298, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 505, + 309 + ], + "score": 1.0, + "content": "performance. Rippel et al. (2016) also observe that applying classification losses to the final layer", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "reduces the entire representation of each class to a single scalar value which destroys intra- and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 319, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 505, + 330 + ], + "score": 1.0, + "content": "inter-class variation. However, these are the same variations that we want our model to capture. We", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "overcome those limitations by explicitly training the model to capture different grains of similarity", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 340, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 353 + ], + "score": 1.0, + "content": "at different levels of the network through applying an intermediate pairwise contrastive loss at those", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 349, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 104, + 349, + 506, + 365 + ], + "score": 1.0, + "content": "levels. In this manner, we can use hierarchical information, such as species of birds having the same", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 363, + 504, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 504, + 375 + ], + "score": 1.0, + "content": "coarse-grained category of bird while having different fine-grained labels. Despite being able to en-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 373, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 505, + 387 + ], + "score": 1.0, + "content": "code different levels of similarity, a contrastive loss does not require an explicit hierarchy; we only", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 385, + 265, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 265, + 397 + ], + "score": 1.0, + "content": "need weak labels for pairs of instances.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 401, + 504, + 489 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "score": 1.0, + "content": "How can we evaluate the representations learned by a clustering algorithm? Previous work has", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 412, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 506, + 424 + ], + "score": 1.0, + "content": "always tried to compete against classification using metrics biased towards the latter task. While", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "some researchers have used hierarchical metrics (Wu et al., 2017) or qualitative measures (Rippel", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "et al., 2016) to evaluate the quality of their representations, their primary evaluation criteria has", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 445, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 457 + ], + "score": 1.0, + "content": "consistently been the classification accuracy of their model. We propose another way to evaluate the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "clustering representations by using the clustering model as a “warm start” from which they can train", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 468, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 479 + ], + "score": 1.0, + "content": "on classification. The intuition is that if the clustering model learned a representations that captures", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 478, + 469, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 469, + 490 + ], + "score": 1.0, + "content": "similarity in a space, it would be be at an advantage when it is fine-tuned for classification.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 495, + 504, + 539 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "score": 1.0, + "content": "We propose the use of a contrastive loss at different layers of a convolutional neural network to learn", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "a notion of similarity across a set of labels. We also propose the use of the accuracy of a model", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 517, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 506, + 529 + ], + "score": 1.0, + "content": "pre-trained for clustering and fine-tuned for classification as an evaluation metric for clustering al-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 528, + 148, + 540 + ], + "spans": [ + { + "bbox": [ + 104, + 528, + 148, + 540 + ], + "score": 1.0, + "content": "gorithms.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + }, + { + "type": "title", + "bbox": [ + 107, + 565, + 365, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 366, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 366, + 579 + ], + "score": 1.0, + "content": "2 INTERMEDIATE PAIRWISE CONTRASTIVE LOSS", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 505, + 685 + ], + "lines": [ + { + "bbox": [ + 105, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "Pairwise contrastive losses were first proposed by Hadsell et al. (2006) to learn a function that maps", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 608, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 619 + ], + "score": 1.0, + "content": "high dimensional inputs to lower dimensional outputs such that distances in the lower dimensional", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 619, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 631 + ], + "score": 1.0, + "content": "space approximate relationships in the input space. Relationships in the input space are not restricted", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 631, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 505, + 641 + ], + "score": 1.0, + "content": "to simple distance measures. In the context of image classification, the resulting function would have", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 640, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 506, + 653 + ], + "score": 1.0, + "content": "to learn the complex invariances that make two images of the same class similar, such that they are", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 652, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 506, + 664 + ], + "score": 1.0, + "content": "both mapped to points that are close to each other in the output space. This loss can be calculated", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 662, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 506, + 675 + ], + "score": 1.0, + "content": "using the representation of two points in the same embedding space and a binary label for whether or", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 673, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 505, + 687 + ], + "score": 1.0, + "content": "not they are similar. Hence, the loss only requires weak labels in the form of pairwise relationships.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 44.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 712, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 709, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 506, + 724 + ], + "score": 1.0, + "content": "1 It should be noted that in work following that paper, Krause et al. (2016) achieve a better performance", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 721, + 369, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 369, + 732 + ], + "score": 1.0, + "content": "using softmax, which can be seen as their baseline classification models.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 192 + ], + "lines": [], + "index": 4.5, + "bbox_fs": [ + 104, + 83, + 506, + 193 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 106, + 199, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 210 + ], + "score": 1.0, + "content": "What measures of similarity should we teach our model? We represent the relations between our", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 209, + 504, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 504, + 221 + ], + "score": 1.0, + "content": "classes in a hierarchy where the labels are all leaf nodes. Therefore, there are two kinds of similari-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 221, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 505, + 232 + ], + "score": 1.0, + "content": "ties that we want our model to capture. The first is intra-class similarity—all cats are similar to each", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "other and they are dissimilar to other animals. The second is inter-class similarity—dogs and cats", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 241, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 104, + 241, + 506, + 255 + ], + "score": 1.0, + "content": "are more similar to each other than they are to non-living objects. For a more complex hierarchy,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "our model should learn similarity at different levels of the class hierarchy. Wu et al. (2017) use", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "their hierarchical loss function to learn those similarities, however, they observe that reducing the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 275, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 286 + ], + "score": 1.0, + "content": "similarity between two classes across all levels of the hierarchy to a single value biases the model", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 505, + 298 + ], + "score": 1.0, + "content": "towards correct classification at the higher levels of the hierarchy resulting in poor classification", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 298, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 505, + 309 + ], + "score": 1.0, + "content": "performance. Rippel et al. (2016) also observe that applying classification losses to the final layer", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "reduces the entire representation of each class to a single scalar value which destroys intra- and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 319, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 505, + 330 + ], + "score": 1.0, + "content": "inter-class variation. However, these are the same variations that we want our model to capture. We", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "overcome those limitations by explicitly training the model to capture different grains of similarity", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 340, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 353 + ], + "score": 1.0, + "content": "at different levels of the network through applying an intermediate pairwise contrastive loss at those", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 349, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 104, + 349, + 506, + 365 + ], + "score": 1.0, + "content": "levels. In this manner, we can use hierarchical information, such as species of birds having the same", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 363, + 504, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 504, + 375 + ], + "score": 1.0, + "content": "coarse-grained category of bird while having different fine-grained labels. Despite being able to en-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 373, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 505, + 387 + ], + "score": 1.0, + "content": "code different levels of similarity, a contrastive loss does not require an explicit hierarchy; we only", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 385, + 265, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 265, + 397 + ], + "score": 1.0, + "content": "need weak labels for pairs of instances.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 18.5, + "bbox_fs": [ + 104, + 199, + 506, + 397 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 401, + 504, + 489 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "score": 1.0, + "content": "How can we evaluate the representations learned by a clustering algorithm? Previous work has", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 412, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 506, + 424 + ], + "score": 1.0, + "content": "always tried to compete against classification using metrics biased towards the latter task. While", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "some researchers have used hierarchical metrics (Wu et al., 2017) or qualitative measures (Rippel", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "et al., 2016) to evaluate the quality of their representations, their primary evaluation criteria has", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 445, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 457 + ], + "score": 1.0, + "content": "consistently been the classification accuracy of their model. We propose another way to evaluate the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "clustering representations by using the clustering model as a “warm start” from which they can train", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 468, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 479 + ], + "score": 1.0, + "content": "on classification. The intuition is that if the clustering model learned a representations that captures", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 478, + 469, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 469, + 490 + ], + "score": 1.0, + "content": "similarity in a space, it would be be at an advantage when it is fine-tuned for classification.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 401, + 506, + 490 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 495, + 504, + 539 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "score": 1.0, + "content": "We propose the use of a contrastive loss at different layers of a convolutional neural network to learn", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "a notion of similarity across a set of labels. We also propose the use of the accuracy of a model", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 517, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 506, + 529 + ], + "score": 1.0, + "content": "pre-trained for clustering and fine-tuned for classification as an evaluation metric for clustering al-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 528, + 148, + 540 + ], + "spans": [ + { + "bbox": [ + 104, + 528, + 148, + 540 + ], + "score": 1.0, + "content": "gorithms.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5, + "bbox_fs": [ + 104, + 495, + 506, + 540 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 565, + 365, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 366, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 366, + 579 + ], + "score": 1.0, + "content": "2 INTERMEDIATE PAIRWISE CONTRASTIVE LOSS", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 505, + 685 + ], + "lines": [ + { + "bbox": [ + 105, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "Pairwise contrastive losses were first proposed by Hadsell et al. (2006) to learn a function that maps", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 608, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 619 + ], + "score": 1.0, + "content": "high dimensional inputs to lower dimensional outputs such that distances in the lower dimensional", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 619, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 631 + ], + "score": 1.0, + "content": "space approximate relationships in the input space. Relationships in the input space are not restricted", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 631, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 505, + 641 + ], + "score": 1.0, + "content": "to simple distance measures. In the context of image classification, the resulting function would have", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 640, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 506, + 653 + ], + "score": 1.0, + "content": "to learn the complex invariances that make two images of the same class similar, such that they are", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 652, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 506, + 664 + ], + "score": 1.0, + "content": "both mapped to points that are close to each other in the output space. This loss can be calculated", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 662, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 506, + 675 + ], + "score": 1.0, + "content": "using the representation of two points in the same embedding space and a binary label for whether or", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 673, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 505, + 687 + ], + "score": 1.0, + "content": "not they are similar. Hence, the loss only requires weak labels in the form of pairwise relationships.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 597, + 506, + 687 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 86, + 464, + 219 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 153, + 86, + 464, + 219 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 153, + 86, + 464, + 219 + ], + "spans": [ + { + "bbox": [ + 153, + 86, + 464, + 219 + ], + "score": 0.972, + "type": "image", + "image_path": "a72c159113b92f8c536176837b61a6ca7237947ff8e534d0cbf2286e02239b7b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 153, + 86, + 464, + 130.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 153, + 130.33333333333334, + 464, + 174.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 153, + 174.66666666666669, + 464, + 219.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 233, + 504, + 256 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 233, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 505, + 246 + ], + "score": 1.0, + "content": "Figure 1: Pairwise losses are applied to intermediate layers of the network as they are calculated", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 244, + 362, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 362, + 257 + ], + "score": 1.0, + "content": "through pairwise comparisons of the activations at those layers.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "title", + "bbox": [ + 107, + 277, + 220, + 289 + ], + "lines": [ + { + "bbox": [ + 106, + 277, + 222, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 222, + 290 + ], + "score": 1.0, + "content": "2.1 LOSS FORMULATION", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 298, + 505, + 354 + ], + "lines": [ + { + "bbox": [ + 105, + 297, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 122, + 311 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 122, + 299, + 157, + 310 + ], + "score": 0.89, + "content": "( X _ { 1 } , X _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 297, + 505, + 311 + ], + "score": 1.0, + "content": "be vectors that are calculated as mappings of two input instances in the same embedding", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 104, + 309, + 137, + 322 + ], + "score": 1.0, + "content": "space.", + "type": "text" + }, + { + "bbox": [ + 137, + 310, + 167, + 321 + ], + "score": 0.9, + "content": "\\mathbb { 1 } _ { X _ { 1 } , X _ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 309, + 380, + 322 + ], + "score": 1.0, + "content": "is an indicator variable that takes a value of one if", + "type": "text" + }, + { + "bbox": [ + 381, + 309, + 419, + 321 + ], + "score": 0.93, + "content": "( X _ { 1 } , X _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "are similar and zero", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 319, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 151, + 333 + ], + "score": 1.0, + "content": "otherwise.", + "type": "text" + }, + { + "bbox": [ + 151, + 321, + 189, + 332 + ], + "score": 0.92, + "content": "D ( X , Y )", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 319, + 464, + 333 + ], + "score": 1.0, + "content": "can be any distance function between two vectors in the same space.", + "type": "text" + }, + { + "bbox": [ + 464, + 320, + 488, + 332 + ], + "score": 0.91, + "content": "L _ { S } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 319, + 506, + 333 + ], + "score": 1.0, + "content": "is a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 330, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 104, + 330, + 276, + 344 + ], + "score": 1.0, + "content": "partial loss function for a similar pair, and", + "type": "text" + }, + { + "bbox": [ + 276, + 331, + 302, + 343 + ], + "score": 0.92, + "content": "L _ { D } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 330, + 506, + 344 + ], + "score": 1.0, + "content": "is a partial loss function for dissimilar pairs. Then", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 343, + 455, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 455, + 354 + ], + "score": 1.0, + "content": "the general formulation of the pairwise contrastive loss function is as shown in Eq. (1):", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8 + }, + { + "type": "interline_equation", + "bbox": [ + 154, + 369, + 457, + 383 + ], + "lines": [ + { + "bbox": [ + 154, + 369, + 457, + 383 + ], + "spans": [ + { + "bbox": [ + 154, + 369, + 457, + 383 + ], + "score": 0.9, + "content": "\\boldsymbol { \\mathcal { L } } ( \\mathbf { X _ { 1 } } , \\mathbf { X _ { 2 } } ) = \\mathbb { 1 } _ { X _ { 1 } , X _ { 2 } } L _ { S } ( D ( \\mathbf { X _ { 1 } } , \\mathbf { X _ { 2 } } ) ) + ( 1 - \\mathbb { 1 } _ { X _ { 1 } , X _ { 2 } } ) L _ { D } ( D ( \\mathbf { X _ { 1 } } , \\mathbf { X _ { 2 } } ) ) .", + "type": "interline_equation", + "image_path": "ec1b27e93ebb2ef0da41150796dbb3489c15ac11fc98dafb1229eaa3f236febf.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 154, + 369, + 457, + 383 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 391, + 505, + 436 + ], + "lines": [ + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "score": 1.0, + "content": "The first half of Eq. (1) penalizes any divergence between the two vectors for similar instances while", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "the second half of the loss rewards divergence up to a margin. Following the formulation proposed", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "by (Hsu & Kira, 2015), we use Kullback-Leibler (KL) divergence as our distance metric and a hinge", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 424, + 448, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 448, + 437 + ], + "score": 1.0, + "content": "loss for dissimilar cases. The final formulation of the loss we use is shown in Eq. (2):", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "interline_equation", + "bbox": [ + 226, + 451, + 384, + 465 + ], + "lines": [ + { + "bbox": [ + 226, + 451, + 384, + 465 + ], + "spans": [ + { + "bbox": [ + 226, + 451, + 384, + 465 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\mathcal { L } ( \\mathbf { P } , \\mathbf { Q } ) = l o s s ( \\mathbf { P } , \\mathbf { Q } ) + l o s s ( \\mathbf { Q } , \\mathbf { P } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "0418768456cd74f938b93dc2cbe35706c6e52515ae1a5d795838820068268b5b.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 226, + 451, + 384, + 465 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 144, + 477, + 465, + 492 + ], + "lines": [ + { + "bbox": [ + 144, + 477, + 465, + 492 + ], + "spans": [ + { + "bbox": [ + 144, + 477, + 465, + 492 + ], + "score": 0.79, + "content": "l o s s ( \\mathbf { P } , \\mathbf { Q } ) = \\mathbb { 1 } _ { P , Q } K L ( \\mathbf { P } | | \\mathbf { Q } ) + ( 1 - \\mathbb { 1 } _ { P , Q } ) \\operatorname* { m a x } \\{ m a r g i n - K L ( \\mathbf { P } | | \\mathbf { Q } ) , 0 \\} ,", + "type": "interline_equation", + "image_path": "2f72a2f29051112b9ff5efb86e026de477cc48c737ced8c19365ed766c97a788.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 144, + 477, + 465, + 492 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 237, + 501, + 374, + 536 + ], + "lines": [ + { + "bbox": [ + 237, + 501, + 374, + 536 + ], + "spans": [ + { + "bbox": [ + 237, + 501, + 374, + 536 + ], + "score": 0.93, + "content": "K L ( \\mathbf { P } \\parallel \\mathbf { Q } ) = \\sum _ { i = 1 } ^ { k } P _ { i } \\log \\Big ( \\frac { P _ { i } } { Q _ { i } } \\Big ) .", + "type": "interline_equation", + "image_path": "7f8124c73dced2aa0f78106eab2a851360b420115872be0ee912903754e96da6.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 237, + 501, + 374, + 518.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 237, + 518.5, + 374, + 536.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 543, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "score": 1.0, + "content": "We have found that a value of two for the margin works well. Since we apply the loss to activations", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 555, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 566 + ], + "score": 1.0, + "content": "at intermediate layers of the network, we first apply a softmax operation on the activations to obtain", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 565, + 275, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 172, + 578 + ], + "score": 1.0, + "content": "the two vectors,", + "type": "text" + }, + { + "bbox": [ + 172, + 566, + 181, + 575 + ], + "score": 0.84, + "content": "P", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 565, + 199, + 578 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 199, + 566, + 208, + 577 + ], + "score": 0.86, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 565, + 275, + 578 + ], + "score": 1.0, + "content": ", used in Eq. (2).", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 108, + 590, + 253, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 590, + 254, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 254, + 603 + ], + "score": 1.0, + "content": "2.2 HIERARCHICAL SIMILARITY", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 610, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "Given that one of our objectives is to encode different levels of similarity into the representations", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 623, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 505, + 634 + ], + "score": 1.0, + "content": "learned by the network, we explore different ways of using the intermediate loss to achieve that goal", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "score": 1.0, + "content": "as shown in Fig. 2. We establish a baseline network where the losses are all applied to the output of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "score": 1.0, + "content": "the last layer. We then use three different variants by applying intermediate losses at different levels", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "of the network. In variant B, we apply the intermediate loss at the penultimate layer. This allows us", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "to apply the contrastive-loss to a higher dimensional space where it could potentially learn a richer", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "representation. In variants C and D, we provide the network with two different kinds of weak labels", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "based on fine-grained and coarse-grained. Since we expect that a weaker form of similarity would be", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 699, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 504, + 712 + ], + "score": 1.0, + "content": "more beneficial earlier on in the network pipeline, we apply the coarse-grained loss to the second-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "to-last fully-connected layer in Network C. It should be noted that many instance pairs will have", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "contradicting losses since they may match at the coarse-grained level but conflict at the fine-grained", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 29 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 86, + 464, + 219 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 153, + 86, + 464, + 219 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 153, + 86, + 464, + 219 + ], + "spans": [ + { + "bbox": [ + 153, + 86, + 464, + 219 + ], + "score": 0.972, + "type": "image", + "image_path": "a72c159113b92f8c536176837b61a6ca7237947ff8e534d0cbf2286e02239b7b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 153, + 86, + 464, + 130.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 153, + 130.33333333333334, + 464, + 174.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 153, + 174.66666666666669, + 464, + 219.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 233, + 504, + 256 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 233, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 505, + 246 + ], + "score": 1.0, + "content": "Figure 1: Pairwise losses are applied to intermediate layers of the network as they are calculated", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 244, + 362, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 362, + 257 + ], + "score": 1.0, + "content": "through pairwise comparisons of the activations at those layers.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "title", + "bbox": [ + 107, + 277, + 220, + 289 + ], + "lines": [ + { + "bbox": [ + 106, + 277, + 222, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 222, + 290 + ], + "score": 1.0, + "content": "2.1 LOSS FORMULATION", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 298, + 505, + 354 + ], + "lines": [ + { + "bbox": [ + 105, + 297, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 122, + 311 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 122, + 299, + 157, + 310 + ], + "score": 0.89, + "content": "( X _ { 1 } , X _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 297, + 505, + 311 + ], + "score": 1.0, + "content": "be vectors that are calculated as mappings of two input instances in the same embedding", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 104, + 309, + 137, + 322 + ], + "score": 1.0, + "content": "space.", + "type": "text" + }, + { + "bbox": [ + 137, + 310, + 167, + 321 + ], + "score": 0.9, + "content": "\\mathbb { 1 } _ { X _ { 1 } , X _ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 309, + 380, + 322 + ], + "score": 1.0, + "content": "is an indicator variable that takes a value of one if", + "type": "text" + }, + { + "bbox": [ + 381, + 309, + 419, + 321 + ], + "score": 0.93, + "content": "( X _ { 1 } , X _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "are similar and zero", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 319, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 151, + 333 + ], + "score": 1.0, + "content": "otherwise.", + "type": "text" + }, + { + "bbox": [ + 151, + 321, + 189, + 332 + ], + "score": 0.92, + "content": "D ( X , Y )", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 319, + 464, + 333 + ], + "score": 1.0, + "content": "can be any distance function between two vectors in the same space.", + "type": "text" + }, + { + "bbox": [ + 464, + 320, + 488, + 332 + ], + "score": 0.91, + "content": "L _ { S } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 319, + 506, + 333 + ], + "score": 1.0, + "content": "is a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 330, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 104, + 330, + 276, + 344 + ], + "score": 1.0, + "content": "partial loss function for a similar pair, and", + "type": "text" + }, + { + "bbox": [ + 276, + 331, + 302, + 343 + ], + "score": 0.92, + "content": "L _ { D } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 330, + 506, + 344 + ], + "score": 1.0, + "content": "is a partial loss function for dissimilar pairs. Then", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 343, + 455, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 455, + 354 + ], + "score": 1.0, + "content": "the general formulation of the pairwise contrastive loss function is as shown in Eq. (1):", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8, + "bbox_fs": [ + 104, + 297, + 506, + 354 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 154, + 369, + 457, + 383 + ], + "lines": [ + { + "bbox": [ + 154, + 369, + 457, + 383 + ], + "spans": [ + { + "bbox": [ + 154, + 369, + 457, + 383 + ], + "score": 0.9, + "content": "\\boldsymbol { \\mathcal { L } } ( \\mathbf { X _ { 1 } } , \\mathbf { X _ { 2 } } ) = \\mathbb { 1 } _ { X _ { 1 } , X _ { 2 } } L _ { S } ( D ( \\mathbf { X _ { 1 } } , \\mathbf { X _ { 2 } } ) ) + ( 1 - \\mathbb { 1 } _ { X _ { 1 } , X _ { 2 } } ) L _ { D } ( D ( \\mathbf { X _ { 1 } } , \\mathbf { X _ { 2 } } ) ) .", + "type": "interline_equation", + "image_path": "ec1b27e93ebb2ef0da41150796dbb3489c15ac11fc98dafb1229eaa3f236febf.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 154, + 369, + 457, + 383 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 391, + 505, + 436 + ], + "lines": [ + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "score": 1.0, + "content": "The first half of Eq. (1) penalizes any divergence between the two vectors for similar instances while", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "the second half of the loss rewards divergence up to a margin. Following the formulation proposed", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "by (Hsu & Kira, 2015), we use Kullback-Leibler (KL) divergence as our distance metric and a hinge", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 424, + 448, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 448, + 437 + ], + "score": 1.0, + "content": "loss for dissimilar cases. The final formulation of the loss we use is shown in Eq. (2):", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 390, + 506, + 437 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 226, + 451, + 384, + 465 + ], + "lines": [ + { + "bbox": [ + 226, + 451, + 384, + 465 + ], + "spans": [ + { + "bbox": [ + 226, + 451, + 384, + 465 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\mathcal { L } ( \\mathbf { P } , \\mathbf { Q } ) = l o s s ( \\mathbf { P } , \\mathbf { Q } ) + l o s s ( \\mathbf { Q } , \\mathbf { P } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "0418768456cd74f938b93dc2cbe35706c6e52515ae1a5d795838820068268b5b.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 226, + 451, + 384, + 465 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 144, + 477, + 465, + 492 + ], + "lines": [ + { + "bbox": [ + 144, + 477, + 465, + 492 + ], + "spans": [ + { + "bbox": [ + 144, + 477, + 465, + 492 + ], + "score": 0.79, + "content": "l o s s ( \\mathbf { P } , \\mathbf { Q } ) = \\mathbb { 1 } _ { P , Q } K L ( \\mathbf { P } | | \\mathbf { Q } ) + ( 1 - \\mathbb { 1 } _ { P , Q } ) \\operatorname* { m a x } \\{ m a r g i n - K L ( \\mathbf { P } | | \\mathbf { Q } ) , 0 \\} ,", + "type": "interline_equation", + "image_path": "2f72a2f29051112b9ff5efb86e026de477cc48c737ced8c19365ed766c97a788.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 144, + 477, + 465, + 492 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 237, + 501, + 374, + 536 + ], + "lines": [ + { + "bbox": [ + 237, + 501, + 374, + 536 + ], + "spans": [ + { + "bbox": [ + 237, + 501, + 374, + 536 + ], + "score": 0.93, + "content": "K L ( \\mathbf { P } \\parallel \\mathbf { Q } ) = \\sum _ { i = 1 } ^ { k } P _ { i } \\log \\Big ( \\frac { P _ { i } } { Q _ { i } } \\Big ) .", + "type": "interline_equation", + "image_path": "7f8124c73dced2aa0f78106eab2a851360b420115872be0ee912903754e96da6.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 237, + 501, + 374, + 518.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 237, + 518.5, + 374, + 536.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 543, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "score": 1.0, + "content": "We have found that a value of two for the margin works well. Since we apply the loss to activations", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 555, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 566 + ], + "score": 1.0, + "content": "at intermediate layers of the network, we first apply a softmax operation on the activations to obtain", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 565, + 275, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 172, + 578 + ], + "score": 1.0, + "content": "the two vectors,", + "type": "text" + }, + { + "bbox": [ + 172, + 566, + 181, + 575 + ], + "score": 0.84, + "content": "P", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 565, + 199, + 578 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 199, + 566, + 208, + 577 + ], + "score": 0.86, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 565, + 275, + 578 + ], + "score": 1.0, + "content": ", used in Eq. (2).", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 542, + 505, + 578 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 590, + 253, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 590, + 254, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 254, + 603 + ], + "score": 1.0, + "content": "2.2 HIERARCHICAL SIMILARITY", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 610, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "Given that one of our objectives is to encode different levels of similarity into the representations", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 623, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 505, + 634 + ], + "score": 1.0, + "content": "learned by the network, we explore different ways of using the intermediate loss to achieve that goal", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "score": 1.0, + "content": "as shown in Fig. 2. We establish a baseline network where the losses are all applied to the output of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "score": 1.0, + "content": "the last layer. We then use three different variants by applying intermediate losses at different levels", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "of the network. In variant B, we apply the intermediate loss at the penultimate layer. This allows us", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "to apply the contrastive-loss to a higher dimensional space where it could potentially learn a richer", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "representation. In variants C and D, we provide the network with two different kinds of weak labels", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "based on fine-grained and coarse-grained. Since we expect that a weaker form of similarity would be", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 699, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 504, + 712 + ], + "score": 1.0, + "content": "more beneficial earlier on in the network pipeline, we apply the coarse-grained loss to the second-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "to-last fully-connected layer in Network C. It should be noted that many instance pairs will have", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "contradicting losses since they may match at the coarse-grained level but conflict at the fine-grained", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 311, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 323 + ], + "score": 1.0, + "content": "level. To tackle this, we augment the network with a skip layer around the second-to-last layer and", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "apply the coarse-grained loss to the skip layer only. Hence, the coarse-grained loss will only provide", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 332, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 505, + 345 + ], + "score": 1.0, + "content": "additional information to the network without directly detracting from the representations passed to", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 344, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 506, + 357 + ], + "score": 1.0, + "content": "subsequent layers. While we show the network variants on AlexNet, the same ideas are applied to a", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 354, + 324, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 324, + 368 + ], + "score": 1.0, + "content": "variant of LeNet that has three fully-connected layers.", + "type": "text", + "cross_page": true + } + ], + "index": 12 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 611, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 82, + 504, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 82, + 504, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 82, + 504, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 504, + 225 + ], + "score": 0.968, + "type": "image", + "image_path": "c751ec74231b40986e76cf6183577dc664e8807abf7061a09418db3bd1e4d8cc.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 82, + 504, + 129.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 129.66666666666666, + 504, + 177.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 177.33333333333331, + 504, + 224.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 233, + 505, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 234, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 505, + 245 + ], + "score": 1.0, + "content": "Figure 2: Different network architecture and losses. Network A is the baseline network. Network", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 243, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 257 + ], + "score": 1.0, + "content": "B applies the fine-grained intermediate loss to the penultimate layer as well. Network C extends", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 254, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 505, + 268 + ], + "score": 1.0, + "content": "Network B by applying the coarse-grained intermediate loss at the second-to-last layer. Network", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 265, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 506, + 280 + ], + "score": 1.0, + "content": "D extends Network B by adding a branch around the second-to-last layer, and applying the coarse-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 277, + 234, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 234, + 290 + ], + "score": 1.0, + "content": "grained intermediate loss there.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 107, + 310, + 505, + 366 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 323 + ], + "score": 1.0, + "content": "level. To tackle this, we augment the network with a skip layer around the second-to-last layer and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "apply the coarse-grained loss to the skip layer only. Hence, the coarse-grained loss will only provide", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 332, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 505, + 345 + ], + "score": 1.0, + "content": "additional information to the network without directly detracting from the representations passed to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 344, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 506, + 357 + ], + "score": 1.0, + "content": "subsequent layers. While we show the network variants on AlexNet, the same ideas are applied to a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 354, + 324, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 324, + 368 + ], + "score": 1.0, + "content": "variant of LeNet that has three fully-connected layers.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 371, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 371, + 506, + 384 + ], + "score": 1.0, + "content": "We follow the efficient implementation regime proposed by (Hsu & Kira, 2015) in order to train a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 383, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 506, + 396 + ], + "score": 1.0, + "content": "single network based on pairwise constraints. The original network formulation to handle pairwise", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 394, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 406 + ], + "score": 1.0, + "content": "constraints is a Siamese network. For every pairwise constraint, at least two data passes need to be", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 404, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 505, + 417 + ], + "score": 1.0, + "content": "performed which results in redundancies. Instead, we feed the softmax outcomes from a single net-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 415, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 505, + 428 + ], + "score": 1.0, + "content": "work and the pairwise similarity constraints to the contrastive loss function which externally handles", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 426, + 504, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 504, + 438 + ], + "score": 1.0, + "content": "all possible pairwise constraints that are available within a mini-batch. We further extend the appli-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 438, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 449 + ], + "score": 1.0, + "content": "cation of contrastive loss functions to multiple intermediate layers with a minimal overhead of extra", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "score": 1.0, + "content": "processing and pairwise constraints, if alternate hierarchical labels are used. Details regarding the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 104, + 459, + 505, + 472 + ], + "score": 1.0, + "content": "exact parameters and setup used to perform experiments are furnished in the Appendices. PyTorch2", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 470, + 344, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 344, + 483 + ], + "score": 1.0, + "content": "is used for all implementations discussed within this paper.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 108, + 498, + 200, + 510 + ], + "lines": [ + { + "bbox": [ + 104, + 496, + 202, + 513 + ], + "spans": [ + { + "bbox": [ + 104, + 496, + 202, + 513 + ], + "score": 1.0, + "content": "3 EXPERIMENTS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "We evaluate our approach on two different datasets using two different convolutional neural", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "networks: LeNet on CIFAR10 and AlexNet on Birdsnap34–a 34-class subset of the Birdsnap", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "score": 1.0, + "content": "dataset (Berg et al., 2014). We generate coarse-grained labels for CIFAR10 by separating the ten", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "classes into six animal classes and four vehicle classes. Below is a discussion of how we constructed", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 565, + 158, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 158, + 581 + ], + "score": 1.0, + "content": "Birdsnap34.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 591, + 505, + 712 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "Birdsnap34 Birdsnap is a popular fine-grained classification dataset that has 500 different classes", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 602, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 602, + 506, + 615 + ], + "score": 1.0, + "content": "and over 40,000 images. We annotated Birdsnap using the scientific classification of each bird to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "create a six-level hierarchy over all the labels. In this work, we use the birds’ subfamily and family as", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "our fine-grained and coarse-grained labels, respectively. Using a pairwise loss with a large number of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 635, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 428, + 647 + ], + "score": 1.0, + "content": "classes poses a significant implementation challenge. For a uniform dataset with", + "type": "text" + }, + { + "bbox": [ + 429, + 637, + 436, + 645 + ], + "score": 0.73, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 635, + 505, + 647 + ], + "score": 1.0, + "content": "classes, the ratio", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 646, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 285, + 659 + ], + "score": 1.0, + "content": "of similar pairs to dissimilar pairs is roughly", + "type": "text" + }, + { + "bbox": [ + 285, + 646, + 329, + 658 + ], + "score": 0.92, + "content": "1 : ( n - 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 646, + 506, + 659 + ], + "score": 1.0, + "content": ". Hence, for a large number of classes, most", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 657, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 670 + ], + "score": 1.0, + "content": "of the pairs will be dissimilar resulting in a very weak signal to the model for clustering instances", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "together. This challenge is analogous to that of imbalanced datasets (Van Horn & Perona, 2017),", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 677, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 693 + ], + "score": 1.0, + "content": "which is a major problem that lies beyond the scope of our paper. Instead of abandoning Birdsnap,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 690, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 505, + 702 + ], + "score": 1.0, + "content": "we curate a 34-class subset, which we will refer to as Birdsnap34, and apply our approach to it. The", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 701, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 701, + 505, + 714 + ], + "score": 1.0, + "content": "classes are chosen such that each class comes from a unique biological subfamily of birds. Through", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 120, + 722, + 189, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 718, + 191, + 735 + ], + "spans": [ + { + "bbox": [ + 118, + 718, + 191, + 735 + ], + "score": 1.0, + "content": "2http://pytorch.org/", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 82, + 504, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 82, + 504, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 82, + 504, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 504, + 225 + ], + "score": 0.968, + "type": "image", + "image_path": "c751ec74231b40986e76cf6183577dc664e8807abf7061a09418db3bd1e4d8cc.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 82, + 504, + 129.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 129.66666666666666, + 504, + 177.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 177.33333333333331, + 504, + 224.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 233, + 505, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 234, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 505, + 245 + ], + "score": 1.0, + "content": "Figure 2: Different network architecture and losses. Network A is the baseline network. Network", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 243, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 257 + ], + "score": 1.0, + "content": "B applies the fine-grained intermediate loss to the penultimate layer as well. Network C extends", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 254, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 505, + 268 + ], + "score": 1.0, + "content": "Network B by applying the coarse-grained intermediate loss at the second-to-last layer. Network", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 265, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 506, + 280 + ], + "score": 1.0, + "content": "D extends Network B by adding a branch around the second-to-last layer, and applying the coarse-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 277, + 234, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 234, + 290 + ], + "score": 1.0, + "content": "grained intermediate loss there.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 107, + 310, + 505, + 366 + ], + "lines": [], + "index": 10, + "bbox_fs": [ + 105, + 311, + 506, + 368 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 371, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 371, + 506, + 384 + ], + "score": 1.0, + "content": "We follow the efficient implementation regime proposed by (Hsu & Kira, 2015) in order to train a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 383, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 506, + 396 + ], + "score": 1.0, + "content": "single network based on pairwise constraints. The original network formulation to handle pairwise", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 394, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 406 + ], + "score": 1.0, + "content": "constraints is a Siamese network. For every pairwise constraint, at least two data passes need to be", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 404, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 505, + 417 + ], + "score": 1.0, + "content": "performed which results in redundancies. Instead, we feed the softmax outcomes from a single net-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 415, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 505, + 428 + ], + "score": 1.0, + "content": "work and the pairwise similarity constraints to the contrastive loss function which externally handles", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 426, + 504, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 504, + 438 + ], + "score": 1.0, + "content": "all possible pairwise constraints that are available within a mini-batch. We further extend the appli-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 438, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 449 + ], + "score": 1.0, + "content": "cation of contrastive loss functions to multiple intermediate layers with a minimal overhead of extra", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "score": 1.0, + "content": "processing and pairwise constraints, if alternate hierarchical labels are used. Details regarding the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 104, + 459, + 505, + 472 + ], + "score": 1.0, + "content": "exact parameters and setup used to perform experiments are furnished in the Appendices. PyTorch2", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 470, + 344, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 344, + 483 + ], + "score": 1.0, + "content": "is used for all implementations discussed within this paper.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 17.5, + "bbox_fs": [ + 104, + 371, + 506, + 483 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 498, + 200, + 510 + ], + "lines": [ + { + "bbox": [ + 104, + 496, + 202, + 513 + ], + "spans": [ + { + "bbox": [ + 104, + 496, + 202, + 513 + ], + "score": 1.0, + "content": "3 EXPERIMENTS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "We evaluate our approach on two different datasets using two different convolutional neural", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "networks: LeNet on CIFAR10 and AlexNet on Birdsnap34–a 34-class subset of the Birdsnap", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "score": 1.0, + "content": "dataset (Berg et al., 2014). We generate coarse-grained labels for CIFAR10 by separating the ten", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "classes into six animal classes and four vehicle classes. Below is a discussion of how we constructed", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 565, + 158, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 158, + 581 + ], + "score": 1.0, + "content": "Birdsnap34.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 523, + 505, + 581 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 591, + 505, + 712 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "Birdsnap34 Birdsnap is a popular fine-grained classification dataset that has 500 different classes", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 602, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 602, + 506, + 615 + ], + "score": 1.0, + "content": "and over 40,000 images. We annotated Birdsnap using the scientific classification of each bird to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "create a six-level hierarchy over all the labels. In this work, we use the birds’ subfamily and family as", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "our fine-grained and coarse-grained labels, respectively. Using a pairwise loss with a large number of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 635, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 428, + 647 + ], + "score": 1.0, + "content": "classes poses a significant implementation challenge. For a uniform dataset with", + "type": "text" + }, + { + "bbox": [ + 429, + 637, + 436, + 645 + ], + "score": 0.73, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 635, + 505, + 647 + ], + "score": 1.0, + "content": "classes, the ratio", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 646, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 285, + 659 + ], + "score": 1.0, + "content": "of similar pairs to dissimilar pairs is roughly", + "type": "text" + }, + { + "bbox": [ + 285, + 646, + 329, + 658 + ], + "score": 0.92, + "content": "1 : ( n - 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 646, + 506, + 659 + ], + "score": 1.0, + "content": ". Hence, for a large number of classes, most", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 657, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 670 + ], + "score": 1.0, + "content": "of the pairs will be dissimilar resulting in a very weak signal to the model for clustering instances", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "together. This challenge is analogous to that of imbalanced datasets (Van Horn & Perona, 2017),", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 677, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 693 + ], + "score": 1.0, + "content": "which is a major problem that lies beyond the scope of our paper. Instead of abandoning Birdsnap,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 690, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 505, + 702 + ], + "score": 1.0, + "content": "we curate a 34-class subset, which we will refer to as Birdsnap34, and apply our approach to it. The", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 701, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 701, + 505, + 714 + ], + "score": 1.0, + "content": "classes are chosen such that each class comes from a unique biological subfamily of birds. Through", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "sampling form unique subfamilies, we increase the interclass variance between our classes. The", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "resulting dataset has 2982 training images and 158 test images that are divided into 34 fine-grained", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "and 18 coarse-grained classes. We plan on extending our model to the entire Birdsnap dataset in", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 157, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 157, + 126 + ], + "score": 1.0, + "content": "future work.", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 591, + 506, + 714 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 126 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "sampling form unique subfamilies, we increase the interclass variance between our classes. The", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "resulting dataset has 2982 training images and 158 test images that are divided into 34 fine-grained", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "and 18 coarse-grained classes. We plan on extending our model to the entire Birdsnap dataset in", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 157, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 157, + 126 + ], + "score": 1.0, + "content": "future work.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "All the network architectures used have randomly initialized weights unless specified otherwise.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "The models are evaluated using accuracy, purity, and Normalized Mutual Information (NMI) (Strehl", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 168 + ], + "score": 1.0, + "content": "& Ghosh, 2002). We train each model for a fixed number of epochs, and we report the top per-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "forming results out of 5 runs. Details regarding the exact parameters and setups used to perform the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 274, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 274, + 189 + ], + "score": 1.0, + "content": "experiments are available in Appendix B.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6 + }, + { + "type": "title", + "bbox": [ + 107, + 204, + 370, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 204, + 372, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 372, + 218 + ], + "score": 1.0, + "content": "3.1 QUALITY OF WARM START AS AN EVALUATION METRIC", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 227, + 505, + 315 + ], + "lines": [ + { + "bbox": [ + 104, + 226, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 104, + 226, + 505, + 241 + ], + "score": 1.0, + "content": "In our first experiment, we show the value of evaluating the representations learned by a clustering", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 237, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 506, + 250 + ], + "score": 1.0, + "content": "loss through using it as a warm start from training on classification. We replicate the setup used", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 249, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 505, + 262 + ], + "score": 1.0, + "content": "by Hsu & Kira (2015) by using LeNet on CIFAR10, and extend their framework to AlexNet on", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 260, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 272 + ], + "score": 1.0, + "content": "Birdsnap34. We use the Hungarian algorithm (Kuhn, 1955) to evaluate clustering accuracy. This", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 272, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 506, + 282 + ], + "score": 1.0, + "content": "method of evaluation assumes that the clusters will be centered across the different dimensions of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 282, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 294 + ], + "score": 1.0, + "content": "the softmax output of the final layer. The Hungarian algorithm is then used to find the assignment", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 292, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 505, + 306 + ], + "score": 1.0, + "content": "of dimensions to classes that would achieve the highest accuracy. We use the Hungarian algorithm", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 304, + 349, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 349, + 316 + ], + "score": 1.0, + "content": "to evaluate clustering performance in all of our experiments.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13.5 + }, + { + "type": "table", + "bbox": [ + 187, + 354, + 422, + 402 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 170, + 329, + 441, + 341 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 170, + 329, + 442, + 342 + ], + "spans": [ + { + "bbox": [ + 170, + 329, + 442, + 342 + ], + "score": 1.0, + "content": "Table 1: Baseline classification performance of LeNet and AlexNet", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "table_body", + "bbox": [ + 187, + 354, + 422, + 402 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 187, + 354, + 422, + 402 + ], + "spans": [ + { + "bbox": [ + 187, + 354, + 422, + 402 + ], + "score": 0.974, + "html": "
CIFAR-10Birdsnap-34
Classification75.4853.50
Clustering74.2117.20
Warm Start Classification75.7550.32
", + "type": "table", + "image_path": "dda884d66cab9d7d7b697f66bc83d1770b76b7c4337ef8e82246a1eb8133716c.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 187, + 354, + 422, + 370.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 187, + 370.0, + 422, + 386.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 187, + 386.0, + 422, + 402.0 + ], + "spans": [], + "index": 21 + } + ] + } + ], + "index": 19.0 + }, + { + "type": "text", + "bbox": [ + 106, + 421, + 505, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 422, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 505, + 434 + ], + "score": 1.0, + "content": "From Table 1 we observe that for CIFAR10, the application of the loss to the final layer results", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 432, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 104, + 432, + 506, + 446 + ], + "score": 1.0, + "content": "in clustering performance almost matching classification performance. However, in more complex", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "dataset such as Birdsnap34, this pattern disappears. This supports the pattern shown in (Hsu &", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 454, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 505, + 467 + ], + "score": 1.0, + "content": "Kira, 2015) where the gap between clustering and classification performance increases when they", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 465, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 506, + 478 + ], + "score": 1.0, + "content": "move from MNIST to CIFAR10. The Cross-Entropy loss formulation, as mentioned in (Hsu & Kira,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "2015), is extremely harsh and seeks to segment each and every sample into one absolute class. The", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "tSNE (van der Maaten & Hinton, 2008) outcomes from the last fully connected layer of LeNet in", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 499, + 230, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 230, + 511 + ], + "score": 1.0, + "content": "Fig. 3 match our observations.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5 + }, + { + "type": "image", + "bbox": [ + 129, + 526, + 485, + 699 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 129, + 526, + 485, + 699 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 526, + 485, + 699 + ], + "spans": [ + { + "bbox": [ + 129, + 526, + 485, + 699 + ], + "score": 0.971, + "type": "image", + "image_path": "a8990135f6d1aafc86247b26a23967600823155298c1769f10985048200a7f11.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 129, + 526, + 485, + 583.6666666666666 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 129, + 583.6666666666666, + 485, + 641.3333333333333 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 129, + 641.3333333333333, + 485, + 698.9999999999999 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 170, + 708, + 442, + 720 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 169, + 707, + 441, + 721 + ], + "spans": [ + { + "bbox": [ + 169, + 707, + 441, + 721 + ], + "score": 1.0, + "content": "Figure 3: t-SNE Visualization of the outputs of LeNet on CIFAR10", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + } + ], + "index": 32.0 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 126 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 105, + 83, + 505, + 126 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "All the network architectures used have randomly initialized weights unless specified otherwise.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "The models are evaluated using accuracy, purity, and Normalized Mutual Information (NMI) (Strehl", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 168 + ], + "score": 1.0, + "content": "& Ghosh, 2002). We train each model for a fixed number of epochs, and we report the top per-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "forming results out of 5 runs. Details regarding the exact parameters and setups used to perform the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 274, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 274, + 189 + ], + "score": 1.0, + "content": "experiments are available in Appendix B.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 132, + 505, + 189 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 204, + 370, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 204, + 372, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 372, + 218 + ], + "score": 1.0, + "content": "3.1 QUALITY OF WARM START AS AN EVALUATION METRIC", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 227, + 505, + 315 + ], + "lines": [ + { + "bbox": [ + 104, + 226, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 104, + 226, + 505, + 241 + ], + "score": 1.0, + "content": "In our first experiment, we show the value of evaluating the representations learned by a clustering", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 237, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 506, + 250 + ], + "score": 1.0, + "content": "loss through using it as a warm start from training on classification. We replicate the setup used", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 249, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 505, + 262 + ], + "score": 1.0, + "content": "by Hsu & Kira (2015) by using LeNet on CIFAR10, and extend their framework to AlexNet on", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 260, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 272 + ], + "score": 1.0, + "content": "Birdsnap34. We use the Hungarian algorithm (Kuhn, 1955) to evaluate clustering accuracy. This", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 272, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 506, + 282 + ], + "score": 1.0, + "content": "method of evaluation assumes that the clusters will be centered across the different dimensions of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 282, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 294 + ], + "score": 1.0, + "content": "the softmax output of the final layer. The Hungarian algorithm is then used to find the assignment", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 292, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 505, + 306 + ], + "score": 1.0, + "content": "of dimensions to classes that would achieve the highest accuracy. We use the Hungarian algorithm", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 304, + 349, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 349, + 316 + ], + "score": 1.0, + "content": "to evaluate clustering performance in all of our experiments.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13.5, + "bbox_fs": [ + 104, + 226, + 506, + 316 + ] + }, + { + "type": "table", + "bbox": [ + 187, + 354, + 422, + 402 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 170, + 329, + 441, + 341 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 170, + 329, + 442, + 342 + ], + "spans": [ + { + "bbox": [ + 170, + 329, + 442, + 342 + ], + "score": 1.0, + "content": "Table 1: Baseline classification performance of LeNet and AlexNet", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "table_body", + "bbox": [ + 187, + 354, + 422, + 402 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 187, + 354, + 422, + 402 + ], + "spans": [ + { + "bbox": [ + 187, + 354, + 422, + 402 + ], + "score": 0.974, + "html": "
CIFAR-10Birdsnap-34
Classification75.4853.50
Clustering74.2117.20
Warm Start Classification75.7550.32
", + "type": "table", + "image_path": "dda884d66cab9d7d7b697f66bc83d1770b76b7c4337ef8e82246a1eb8133716c.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 187, + 354, + 422, + 370.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 187, + 370.0, + 422, + 386.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 187, + 386.0, + 422, + 402.0 + ], + "spans": [], + "index": 21 + } + ] + } + ], + "index": 19.0 + }, + { + "type": "text", + "bbox": [ + 106, + 421, + 505, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 422, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 505, + 434 + ], + "score": 1.0, + "content": "From Table 1 we observe that for CIFAR10, the application of the loss to the final layer results", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 432, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 104, + 432, + 506, + 446 + ], + "score": 1.0, + "content": "in clustering performance almost matching classification performance. However, in more complex", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "dataset such as Birdsnap34, this pattern disappears. This supports the pattern shown in (Hsu &", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 454, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 505, + 467 + ], + "score": 1.0, + "content": "Kira, 2015) where the gap between clustering and classification performance increases when they", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 465, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 506, + 478 + ], + "score": 1.0, + "content": "move from MNIST to CIFAR10. The Cross-Entropy loss formulation, as mentioned in (Hsu & Kira,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "2015), is extremely harsh and seeks to segment each and every sample into one absolute class. The", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "tSNE (van der Maaten & Hinton, 2008) outcomes from the last fully connected layer of LeNet in", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 499, + 230, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 230, + 511 + ], + "score": 1.0, + "content": "Fig. 3 match our observations.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5, + "bbox_fs": [ + 104, + 422, + 506, + 511 + ] + }, + { + "type": "image", + "bbox": [ + 129, + 526, + 485, + 699 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 129, + 526, + 485, + 699 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 526, + 485, + 699 + ], + "spans": [ + { + "bbox": [ + 129, + 526, + 485, + 699 + ], + "score": 0.971, + "type": "image", + "image_path": "a8990135f6d1aafc86247b26a23967600823155298c1769f10985048200a7f11.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 129, + 526, + 485, + 583.6666666666666 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 129, + 583.6666666666666, + 485, + 641.3333333333333 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 129, + 641.3333333333333, + 485, + 698.9999999999999 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 170, + 708, + 442, + 720 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 169, + 707, + 441, + 721 + ], + "spans": [ + { + "bbox": [ + 169, + 707, + 441, + 721 + ], + "score": 1.0, + "content": "Figure 3: t-SNE Visualization of the outputs of LeNet on CIFAR10", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + } + ], + "index": 32.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 222, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 223, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 223, + 95 + ], + "score": 1.0, + "content": "3.2 INTERMEDIATE LOSS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 102, + 505, + 180 + ], + "lines": [ + { + "bbox": [ + 106, + 103, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 505, + 115 + ], + "score": 1.0, + "content": "The previous experiment indicates that the representations learned from end-to-end clustering do", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 504, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 504, + 127 + ], + "score": 1.0, + "content": "not extend beyond simple datasets. We posit that this is a result of how the clustering loss is applied.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 125, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 505, + 137 + ], + "score": 1.0, + "content": "Most machine learning models only apply losses to their last layer. This is generally a reasonable", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 136, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 136, + 505, + 149 + ], + "score": 1.0, + "content": "thing to do since we only have labels for the output of the last layer. On the other hand, pairwise", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 147, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 505, + 160 + ], + "score": 1.0, + "content": "losses can be applied at any level of the network since they only require a pair of vectors in the same", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 157, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 506, + 171 + ], + "score": 1.0, + "content": "embedding space. In this experiment we explore the potential of applying losses at intermediate", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 169, + 257, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 169, + 257, + 182 + ], + "score": 1.0, + "content": "layers on the learned representations.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 124, + 216, + 487, + 297 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 236, + 190, + 375, + 202 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 235, + 189, + 376, + 203 + ], + "spans": [ + { + "bbox": [ + 235, + 189, + 376, + 203 + ], + "score": 1.0, + "content": "Table 2: Using intermediate losses", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "table_body", + "bbox": [ + 124, + 216, + 487, + 297 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 216, + 487, + 297 + ], + "spans": [ + { + "bbox": [ + 124, + 216, + 487, + 297 + ], + "score": 0.984, + "html": "
Clustering Acc.NMIPurityClassification Acc.
CIFAR10Cross-Entropy LeNet A LeNet B1 74.21 72.661 0.551 0.7475.48 75.75
Birdsnap34Cross-Entropy AlexNet A AlexNetB1 17.20 20.380.54 1 0.39 0.450.73 1 0.17 0.2075.42 53.50 50.32 45.86
", + "type": "table", + "image_path": "560b62ba4f40d2eea4ed79f5986a5c606e20e79a7a30c96a948e7e38acfe160e.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 124, + 216, + 487, + 243.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 124, + 243.0, + 487, + 270.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 124, + 270.0, + 487, + 297.0 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 9.0 + }, + { + "type": "text", + "bbox": [ + 106, + 312, + 505, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "We train two variants of LeNet and AlexNet using the pairwise contrastive loss. Variant A follows", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "the setup used by Hsu & Kira (2015), while Variant B applies the same loss at both the final and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "penultimate layer of the convolutional neural network. We also train the same network using Cross-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "Entropy loss to establish baseline performance. As shown in Table 2, performance on CIFAR10 is", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "almost the same across all three settings. Meanwhile, we see clear differences on Birdsnap34. There", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 367, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 380 + ], + "score": 1.0, + "content": "is a significant increase across all metrics for the clustering models for the intermediate loss version,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 378, + 415, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 415, + 391 + ], + "score": 1.0, + "content": "however, this improvement does not carry over to the classification accuracy.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 108, + 402, + 257, + 414 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 258, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 258, + 415 + ], + "score": 1.0, + "content": "3.3 HIERARCHICAL CLUSTERING", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 423, + 505, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 506, + 437 + ], + "score": 1.0, + "content": "Previous work on embedding spaces has shown that adding structure to the embedding space or", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 434, + 504, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 504, + 447 + ], + "score": 1.0, + "content": "learning it using a hierarchical structure often validates its integrity as well as improves perfor-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 445, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 457 + ], + "score": 1.0, + "content": "mance (Mikolov et al., 2013). In an effort to understand the impact of adding a hierarchical-label-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 456, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 470 + ], + "score": 1.0, + "content": "based contrastive loss to our baseline networks, we apply it on variants C and D of LeNet and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "score": 1.0, + "content": "AlexNet. Since the restriction on dimensionality is removed by using a contrastive loss with KL di-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 478, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 492 + ], + "score": 1.0, + "content": "vergence, we can directly apply the same loss to an alternate set of weak labels and representations.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "score": 1.0, + "content": "Our ultimate expectation is that adding hierarchical structure should aid in improving classification,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 501, + 421, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 421, + 513 + ], + "score": 1.0, + "content": "given their potency to improve embedding spaces as shown in previous works.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + }, + { + "type": "table", + "bbox": [ + 124, + 547, + 487, + 672 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 196, + 522, + 415, + 534 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 195, + 521, + 416, + 534 + ], + "spans": [ + { + "bbox": [ + 195, + 521, + 416, + 534 + ], + "score": 1.0, + "content": "Table 3: Impact of adding heirarchy based weak labels", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "table_body", + "bbox": [ + 124, + 547, + 487, + 672 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 547, + 487, + 672 + ], + "spans": [ + { + "bbox": [ + 124, + 547, + 487, + 672 + ], + "score": 0.986, + "html": "
Clustering Acc.NMIPurityClassification Acc.
CIFAR10Cross-Entropy75.48
LeNet A74.210.550.7475.75
LeNet B72.660.540.7375.42
LeNet C72.840.550.7375.11
LeNet D74.170.560.7475.53
Birdsnap34Cross-Entropy11153.50
AlexNet A17.200.390.1750.32
AlexNet B20.380.450.2045.86
AlexNet C24.840.510.2570.70
AlexNet D22.290.490.2265.61
", + "type": "table", + "image_path": "267382a8c11ea2d13a8405e789dc3ec7b077d5b50a014517a25d438aa1015309.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 124, + 547, + 487, + 588.6666666666666 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 124, + 588.6666666666666, + 487, + 630.3333333333333 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 124, + 630.3333333333333, + 487, + 671.9999999999999 + ], + "spans": [], + "index": 31 + } + ] + } + ], + "index": 29.0 + }, + { + "type": "text", + "bbox": [ + 106, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "Table 3 shows that the experimental results align with our expectations in terms of overall im-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "provement in classification. As seen in other experiments, the performance on CIFAR10 is almost", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "the same across all metrics. Meanwhile, there is a spike in performance for Birdsnap34 for both", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "AlexNet C and D with the metrics far exceeding classification performance. Given that Birdsnap34", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 222, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 223, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 223, + 95 + ], + "score": 1.0, + "content": "3.2 INTERMEDIATE LOSS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 102, + 505, + 180 + ], + "lines": [ + { + "bbox": [ + 106, + 103, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 505, + 115 + ], + "score": 1.0, + "content": "The previous experiment indicates that the representations learned from end-to-end clustering do", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 504, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 504, + 127 + ], + "score": 1.0, + "content": "not extend beyond simple datasets. We posit that this is a result of how the clustering loss is applied.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 125, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 505, + 137 + ], + "score": 1.0, + "content": "Most machine learning models only apply losses to their last layer. This is generally a reasonable", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 136, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 136, + 505, + 149 + ], + "score": 1.0, + "content": "thing to do since we only have labels for the output of the last layer. On the other hand, pairwise", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 147, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 505, + 160 + ], + "score": 1.0, + "content": "losses can be applied at any level of the network since they only require a pair of vectors in the same", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 157, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 506, + 171 + ], + "score": 1.0, + "content": "embedding space. In this experiment we explore the potential of applying losses at intermediate", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 169, + 257, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 169, + 257, + 182 + ], + "score": 1.0, + "content": "layers on the learned representations.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 103, + 506, + 182 + ] + }, + { + "type": "table", + "bbox": [ + 124, + 216, + 487, + 297 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 236, + 190, + 375, + 202 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 235, + 189, + 376, + 203 + ], + "spans": [ + { + "bbox": [ + 235, + 189, + 376, + 203 + ], + "score": 1.0, + "content": "Table 2: Using intermediate losses", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "table_body", + "bbox": [ + 124, + 216, + 487, + 297 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 216, + 487, + 297 + ], + "spans": [ + { + "bbox": [ + 124, + 216, + 487, + 297 + ], + "score": 0.984, + "html": "
Clustering Acc.NMIPurityClassification Acc.
CIFAR10Cross-Entropy LeNet A LeNet B1 74.21 72.661 0.551 0.7475.48 75.75
Birdsnap34Cross-Entropy AlexNet A AlexNetB1 17.20 20.380.54 1 0.39 0.450.73 1 0.17 0.2075.42 53.50 50.32 45.86
", + "type": "table", + "image_path": "560b62ba4f40d2eea4ed79f5986a5c606e20e79a7a30c96a948e7e38acfe160e.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 124, + 216, + 487, + 243.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 124, + 243.0, + 487, + 270.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 124, + 270.0, + 487, + 297.0 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 9.0 + }, + { + "type": "text", + "bbox": [ + 106, + 312, + 505, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "We train two variants of LeNet and AlexNet using the pairwise contrastive loss. Variant A follows", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "the setup used by Hsu & Kira (2015), while Variant B applies the same loss at both the final and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "penultimate layer of the convolutional neural network. We also train the same network using Cross-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "Entropy loss to establish baseline performance. As shown in Table 2, performance on CIFAR10 is", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "almost the same across all three settings. Meanwhile, we see clear differences on Birdsnap34. There", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 367, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 380 + ], + "score": 1.0, + "content": "is a significant increase across all metrics for the clustering models for the intermediate loss version,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 378, + 415, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 415, + 391 + ], + "score": 1.0, + "content": "however, this improvement does not carry over to the classification accuracy.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 312, + 506, + 391 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 402, + 257, + 414 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 258, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 258, + 415 + ], + "score": 1.0, + "content": "3.3 HIERARCHICAL CLUSTERING", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 423, + 505, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 506, + 437 + ], + "score": 1.0, + "content": "Previous work on embedding spaces has shown that adding structure to the embedding space or", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 434, + 504, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 504, + 447 + ], + "score": 1.0, + "content": "learning it using a hierarchical structure often validates its integrity as well as improves perfor-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 445, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 457 + ], + "score": 1.0, + "content": "mance (Mikolov et al., 2013). In an effort to understand the impact of adding a hierarchical-label-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 456, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 470 + ], + "score": 1.0, + "content": "based contrastive loss to our baseline networks, we apply it on variants C and D of LeNet and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "score": 1.0, + "content": "AlexNet. Since the restriction on dimensionality is removed by using a contrastive loss with KL di-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 478, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 492 + ], + "score": 1.0, + "content": "vergence, we can directly apply the same loss to an alternate set of weak labels and representations.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "score": 1.0, + "content": "Our ultimate expectation is that adding hierarchical structure should aid in improving classification,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 501, + 421, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 421, + 513 + ], + "score": 1.0, + "content": "given their potency to improve embedding spaces as shown in previous works.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 422, + 506, + 513 + ] + }, + { + "type": "table", + "bbox": [ + 124, + 547, + 487, + 672 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 196, + 522, + 415, + 534 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 195, + 521, + 416, + 534 + ], + "spans": [ + { + "bbox": [ + 195, + 521, + 416, + 534 + ], + "score": 1.0, + "content": "Table 3: Impact of adding heirarchy based weak labels", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "table_body", + "bbox": [ + 124, + 547, + 487, + 672 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 547, + 487, + 672 + ], + "spans": [ + { + "bbox": [ + 124, + 547, + 487, + 672 + ], + "score": 0.986, + "html": "
Clustering Acc.NMIPurityClassification Acc.
CIFAR10Cross-Entropy75.48
LeNet A74.210.550.7475.75
LeNet B72.660.540.7375.42
LeNet C72.840.550.7375.11
LeNet D74.170.560.7475.53
Birdsnap34Cross-Entropy11153.50
AlexNet A17.200.390.1750.32
AlexNet B20.380.450.2045.86
AlexNet C24.840.510.2570.70
AlexNet D22.290.490.2265.61
", + "type": "table", + "image_path": "267382a8c11ea2d13a8405e789dc3ec7b077d5b50a014517a25d438aa1015309.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 124, + 547, + 487, + 588.6666666666666 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 124, + 588.6666666666666, + 487, + 630.3333333333333 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 124, + 630.3333333333333, + 487, + 671.9999999999999 + ], + "spans": [], + "index": 31 + } + ] + } + ], + "index": 29.0 + }, + { + "type": "text", + "bbox": [ + 106, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "Table 3 shows that the experimental results align with our expectations in terms of overall im-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "provement in classification. As seen in other experiments, the performance on CIFAR10 is almost", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "the same across all metrics. Meanwhile, there is a spike in performance for Birdsnap34 for both", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "AlexNet C and D with the metrics far exceeding classification performance. Given that Birdsnap34", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "has a more well-defined hierarchy, it would be expected that using the hierarchical labels to learn", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "similarity would result in learning good representations. Another thing worth noting is that the", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "dataset is curated by having subfamilies as fine-grained labels and biological families as coarse-", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "grained labels. This structure results in similarity along this tree corresponding to distinct visual", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "features. This might be indicating that choice of hierarchy can have an impact on performance. A", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "surprising finding is that AlexNet C outperforms AlexNet D. One would expect AlexNet D to be", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 147, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 505, + 162 + ], + "score": 1.0, + "content": "able to extract features for both the coarse and fine levels of the hierarchy. We plan on investigating", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 185, + 170 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 185, + 170 + ], + "score": 1.0, + "content": "this in future work.", + "type": "text", + "cross_page": true + } + ], + "index": 7 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 687, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 170 + ], + "lines": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "has a more well-defined hierarchy, it would be expected that using the hierarchical labels to learn", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "similarity would result in learning good representations. Another thing worth noting is that the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "dataset is curated by having subfamilies as fine-grained labels and biological families as coarse-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "grained labels. This structure results in similarity along this tree corresponding to distinct visual", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "features. This might be indicating that choice of hierarchy can have an impact on performance. A", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "surprising finding is that AlexNet C outperforms AlexNet D. One would expect AlexNet D to be", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 147, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 505, + 162 + ], + "score": 1.0, + "content": "able to extract features for both the coarse and fine levels of the hierarchy. We plan on investigating", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 185, + 170 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 185, + 170 + ], + "score": 1.0, + "content": "this in future work.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 108, + 186, + 401, + 197 + ], + "lines": [ + { + "bbox": [ + 105, + 185, + 403, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 403, + 199 + ], + "score": 1.0, + "content": "3.4 SENSITIVITY TO NETWORK SIZE AND WEIGHT INITIALIZATION", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 207, + 505, + 273 + ], + "lines": [ + { + "bbox": [ + 106, + 207, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 505, + 219 + ], + "score": 1.0, + "content": "Given the ability of standard end-to-end clustering networks to perform fine-tuned classification", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 218, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 230 + ], + "score": 1.0, + "content": "reasonably well, we test their sensitivity towards network size in order to understand if they would", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 229, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 505, + 241 + ], + "score": 1.0, + "content": "be able to achieve or outperform their simpler baseline classification counterparts. In order to do", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "so, we double the number of convolutional filters and the sizes of fully-connected layers within", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "AlexNet. We use their doubled networks to perform clustering and fine-tune them for classification.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 261, + 356, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 356, + 273 + ], + "score": 1.0, + "content": "We denote these modified networks with a suffix of “Double.”", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11.5 + }, + { + "type": "table", + "bbox": [ + 142, + 309, + 468, + 357 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 226, + 283, + 385, + 295 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 224, + 281, + 386, + 298 + ], + "spans": [ + { + "bbox": [ + 224, + 281, + 386, + 298 + ], + "score": 1.0, + "content": "Table 4: Sensitivity to network capacity", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "table_body", + "bbox": [ + 142, + 309, + 468, + 357 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 142, + 309, + 468, + 357 + ], + "spans": [ + { + "bbox": [ + 142, + 309, + 468, + 357 + ], + "score": 0.977, + "html": "
Clustering Acc.NMIPurityClassification Acc.
Cross-Entropy一1153.50
AlexNet A (Double)22.930.490.2359.24
AlexNet C24.840.510.2570.70
", + "type": "table", + "image_path": "fe2b3905cbc46bc7cf256753eb284946a22b77a879476085c03d23a987fe7014.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 142, + 309, + 468, + 325.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 142, + 325.0, + 468, + 341.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 142, + 341.0, + 468, + 357.0 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 16.0 + }, + { + "type": "table", + "bbox": [ + 153, + 406, + 455, + 442 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 186, + 380, + 425, + 392 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 186, + 379, + 425, + 394 + ], + "spans": [ + { + "bbox": [ + 186, + 379, + 425, + 394 + ], + "score": 1.0, + "content": "Table 5: Comparison against ImageNet pre-trained weights", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "table_body", + "bbox": [ + 153, + 406, + 455, + 442 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 153, + 406, + 455, + 442 + ], + "spans": [ + { + "bbox": [ + 153, + 406, + 455, + 442 + ], + "score": 0.97, + "html": "
Clustering Acc.NMIPurityClassification Acc.
Cross-Entropy11181.53
AlexNet C24.200.510.2482.80
", + "type": "table", + "image_path": "f8cfe17876cbbaf94ee482764253089d623f6643bb4551bbd2477838da5a2c46.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 153, + 406, + 455, + 418.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 153, + 418.0, + 455, + 430.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 153, + 430.0, + 455, + 442.0 + ], + "spans": [], + "index": 22 + } + ] + } + ], + "index": 20.0 + }, + { + "type": "text", + "bbox": [ + 106, + 461, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "From Table 4, we clearly observe that clustering is highly sensitive to network capacity. Compared", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 473, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 505, + 485 + ], + "score": 1.0, + "content": "to AlexNet A, its doubled counterpart allows clustering to learn better representations which leads", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "score": 1.0, + "content": "to classificaiton performance being higher than the baseline classification model. However, our", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "AlexNet C variant outperforms the double capacity network. This points to the efficacy of the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 506, + 504, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 504, + 518 + ], + "score": 1.0, + "content": "combination of weak labels over multiple levels of heirarchy and the use of intermediate losses.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 516, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 530 + ], + "score": 1.0, + "content": "Furthermore, Table 5 shows that even we use ImageNet pre-trained weights, variant C outperforms", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "the ImageNet pre-trained classification baseline. This points to a tractable and viable approach to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 538, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 104, + 538, + 506, + 552 + ], + "score": 1.0, + "content": "extending performance over more complex datasets without the need to increase the capacity of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 550, + 175, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 175, + 562 + ], + "score": 1.0, + "content": "neural networks.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 108, + 579, + 297, + 591 + ], + "lines": [ + { + "bbox": [ + 105, + 577, + 299, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 299, + 594 + ], + "score": 1.0, + "content": "4 DISCUSSION AND FUTURE WORK", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 605, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "In this work, we argue that similarity-based losses can be applied as a warm start to boost the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "performance of image classification models. We show that applying a pairwise contrastive loss to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "intermediate layers of the network results in better clustering performance, as well as better fine-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "tuned classification performance. Furthermore, we demonstrate how the pairwise loss can be used", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 649, + 297, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 297, + 661 + ], + "score": 1.0, + "content": "to train a model using weak hierarchical labels.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "We find that training with hierarchical labels results in the highest performance beating models with", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "more parameters and approaches the performance of models pre-trained with ImageNet weights.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "This is a very significant finding since ImageNet contains multiple bird classes, so a model pre-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "trained with ImageNet has seen many more birds than a model pre-trained with pairwise constraints", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 482, + 721 + ], + "score": 1.0, + "content": "on Birdsnap34. Nevertheless, it only outperforms a cluster-based warm-start model by just", + "type": "text" + }, + { + "bbox": [ + 482, + 710, + 501, + 720 + ], + "score": 0.86, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 710, + 505, + 721 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "score": 1.0, + "content": "Regardless, we also show that applying our approach to ImageNet weights still results in a boost", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 170 + ], + "lines": [], + "index": 3.5, + "bbox_fs": [ + 105, + 83, + 505, + 170 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 186, + 401, + 197 + ], + "lines": [ + { + "bbox": [ + 105, + 185, + 403, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 403, + 199 + ], + "score": 1.0, + "content": "3.4 SENSITIVITY TO NETWORK SIZE AND WEIGHT INITIALIZATION", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 207, + 505, + 273 + ], + "lines": [ + { + "bbox": [ + 106, + 207, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 505, + 219 + ], + "score": 1.0, + "content": "Given the ability of standard end-to-end clustering networks to perform fine-tuned classification", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 218, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 230 + ], + "score": 1.0, + "content": "reasonably well, we test their sensitivity towards network size in order to understand if they would", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 229, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 505, + 241 + ], + "score": 1.0, + "content": "be able to achieve or outperform their simpler baseline classification counterparts. In order to do", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "so, we double the number of convolutional filters and the sizes of fully-connected layers within", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "AlexNet. We use their doubled networks to perform clustering and fine-tune them for classification.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 261, + 356, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 356, + 273 + ], + "score": 1.0, + "content": "We denote these modified networks with a suffix of “Double.”", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 207, + 505, + 273 + ] + }, + { + "type": "table", + "bbox": [ + 142, + 309, + 468, + 357 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 226, + 283, + 385, + 295 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 224, + 281, + 386, + 298 + ], + "spans": [ + { + "bbox": [ + 224, + 281, + 386, + 298 + ], + "score": 1.0, + "content": "Table 4: Sensitivity to network capacity", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "table_body", + "bbox": [ + 142, + 309, + 468, + 357 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 142, + 309, + 468, + 357 + ], + "spans": [ + { + "bbox": [ + 142, + 309, + 468, + 357 + ], + "score": 0.977, + "html": "
Clustering Acc.NMIPurityClassification Acc.
Cross-Entropy一1153.50
AlexNet A (Double)22.930.490.2359.24
AlexNet C24.840.510.2570.70
", + "type": "table", + "image_path": "fe2b3905cbc46bc7cf256753eb284946a22b77a879476085c03d23a987fe7014.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 142, + 309, + 468, + 325.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 142, + 325.0, + 468, + 341.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 142, + 341.0, + 468, + 357.0 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 16.0 + }, + { + "type": "table", + "bbox": [ + 153, + 406, + 455, + 442 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 186, + 380, + 425, + 392 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 186, + 379, + 425, + 394 + ], + "spans": [ + { + "bbox": [ + 186, + 379, + 425, + 394 + ], + "score": 1.0, + "content": "Table 5: Comparison against ImageNet pre-trained weights", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "table_body", + "bbox": [ + 153, + 406, + 455, + 442 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 153, + 406, + 455, + 442 + ], + "spans": [ + { + "bbox": [ + 153, + 406, + 455, + 442 + ], + "score": 0.97, + "html": "
Clustering Acc.NMIPurityClassification Acc.
Cross-Entropy11181.53
AlexNet C24.200.510.2482.80
", + "type": "table", + "image_path": "f8cfe17876cbbaf94ee482764253089d623f6643bb4551bbd2477838da5a2c46.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 153, + 406, + 455, + 418.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 153, + 418.0, + 455, + 430.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 153, + 430.0, + 455, + 442.0 + ], + "spans": [], + "index": 22 + } + ] + } + ], + "index": 20.0 + }, + { + "type": "text", + "bbox": [ + 106, + 461, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "From Table 4, we clearly observe that clustering is highly sensitive to network capacity. Compared", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 473, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 505, + 485 + ], + "score": 1.0, + "content": "to AlexNet A, its doubled counterpart allows clustering to learn better representations which leads", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "score": 1.0, + "content": "to classificaiton performance being higher than the baseline classification model. However, our", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "AlexNet C variant outperforms the double capacity network. This points to the efficacy of the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 506, + 504, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 504, + 518 + ], + "score": 1.0, + "content": "combination of weak labels over multiple levels of heirarchy and the use of intermediate losses.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 516, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 530 + ], + "score": 1.0, + "content": "Furthermore, Table 5 shows that even we use ImageNet pre-trained weights, variant C outperforms", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "the ImageNet pre-trained classification baseline. This points to a tractable and viable approach to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 538, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 104, + 538, + 506, + 552 + ], + "score": 1.0, + "content": "extending performance over more complex datasets without the need to increase the capacity of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 550, + 175, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 175, + 562 + ], + "score": 1.0, + "content": "neural networks.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27, + "bbox_fs": [ + 104, + 462, + 506, + 562 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 579, + 297, + 591 + ], + "lines": [ + { + "bbox": [ + 105, + 577, + 299, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 299, + 594 + ], + "score": 1.0, + "content": "4 DISCUSSION AND FUTURE WORK", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 605, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "In this work, we argue that similarity-based losses can be applied as a warm start to boost the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "performance of image classification models. We show that applying a pairwise contrastive loss to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "intermediate layers of the network results in better clustering performance, as well as better fine-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "tuned classification performance. Furthermore, we demonstrate how the pairwise loss can be used", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 649, + 297, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 297, + 661 + ], + "score": 1.0, + "content": "to train a model using weak hierarchical labels.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 605, + 506, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "We find that training with hierarchical labels results in the highest performance beating models with", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "more parameters and approaches the performance of models pre-trained with ImageNet weights.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "This is a very significant finding since ImageNet contains multiple bird classes, so a model pre-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "trained with ImageNet has seen many more birds than a model pre-trained with pairwise constraints", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 482, + 721 + ], + "score": 1.0, + "content": "on Birdsnap34. Nevertheless, it only outperforms a cluster-based warm-start model by just", + "type": "text" + }, + { + "bbox": [ + 482, + 710, + 501, + 720 + ], + "score": 0.86, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 710, + 505, + 721 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "score": 1.0, + "content": "Regardless, we also show that applying our approach to ImageNet weights still results in a boost", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 118, + 95 + ], + "score": 1.0, + "content": "of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 118, + 82, + 145, + 93 + ], + "score": 0.87, + "content": "1 . 2 7 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 145, + 81, + 505, + 95 + ], + "score": 1.0, + "content": ". This supports our claim that similarity-based metrics can improve classification perfor-", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 504, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 504, + 106 + ], + "score": 1.0, + "content": "mance, but it suggests that the gains we expect decrease if we start of with a good representational", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 106, + 134, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 106, + 134, + 118 + ], + "score": 1.0, + "content": "space.", + "type": "text", + "cross_page": true + } + ], + "index": 2 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 665, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 503, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 118, + 95 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 82, + 145, + 93 + ], + "score": 0.87, + "content": "1 . 2 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 81, + 505, + 95 + ], + "score": 1.0, + "content": ". This supports our claim that similarity-based metrics can improve classification perfor-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 504, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 504, + 106 + ], + "score": 1.0, + "content": "mance, but it suggests that the gains we expect decrease if we start of with a good representational", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 106, + 134, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 106, + 134, + 118 + ], + "score": 1.0, + "content": "space.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 505, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 506, + 134 + ], + "score": 1.0, + "content": "We hope to expand this in future work in multiple ways. We plan on extending our approach to", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "the entire Birdsnap dataset, as well as to other fine-grained classification datasets. We also hope", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 505, + 155 + ], + "score": 1.0, + "content": "to perform more extensive analysis of the quality of the embedding spaces learned as it is obvious", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 152, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 104, + 152, + 505, + 168 + ], + "score": 1.0, + "content": "that the use of the Hungarian algorithm does not accurately reflect the performance of the clustering", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 165, + 137, + 177 + ], + "spans": [ + { + "bbox": [ + 104, + 165, + 137, + 177 + ], + "score": 1.0, + "content": "model.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 107, + 193, + 175, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 193, + 176, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 176, + 207 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 105, + 211, + 505, + 234 + ], + "lines": [ + { + "bbox": [ + 106, + 211, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 211, + 505, + 224 + ], + "score": 1.0, + "content": "Aurelien Bellet, Amaury Habrard, and Marc Sebban. A survey on metric learning for feature vectors ´", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 222, + 356, + 234 + ], + "spans": [ + { + "bbox": [ + 115, + 222, + 356, + 234 + ], + "score": 1.0, + "content": "and structured data. arXiv preprint arXiv:1306.6709, 2013.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 109, + 240, + 504, + 275 + ], + "lines": [ + { + "bbox": [ + 107, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 107, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "Thomas Berg, Jiongxin Liu, Seung Woo Lee, Michelle L Alexander, David W Jacobs, and Peter N", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 252, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 115, + 252, + 506, + 265 + ], + "score": 1.0, + "content": "Belhumeur. Birdsnap: Large-scale fine-grained visual categorization of birds. In Proceedings of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 263, + 477, + 276 + ], + "spans": [ + { + "bbox": [ + 115, + 263, + 477, + 276 + ], + "score": 1.0, + "content": "the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2011–2018, 2014.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 109, + 281, + 503, + 315 + ], + "lines": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 292, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 116, + 292, + 505, + 306 + ], + "score": 1.0, + "content": "mapping. In Computer vision and pattern recognition, 2006 IEEE computer society conference", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 304, + 289, + 316 + ], + "spans": [ + { + "bbox": [ + 116, + 304, + 289, + 316 + ], + "score": 1.0, + "content": "on, volume 2, pp. 1735–1742. IEEE, 2006.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 321, + 504, + 356 + ], + "lines": [ + { + "bbox": [ + 105, + 321, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 335 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 334, + 504, + 345 + ], + "spans": [ + { + "bbox": [ + 116, + 334, + 504, + 345 + ], + "score": 1.0, + "content": "human-level performance on imagenet classification. In Proceedings of the IEEE international", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 344, + 334, + 356 + ], + "spans": [ + { + "bbox": [ + 116, + 344, + 334, + 356 + ], + "score": 1.0, + "content": "conference on computer vision, pp. 1026–1034, 2015.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 362, + 503, + 385 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 114, + 374, + 254, + 385 + ], + "spans": [ + { + "bbox": [ + 114, + 374, + 254, + 385 + ], + "score": 1.0, + "content": "preprint arXiv:1503.02531, 2015.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 108, + 392, + 503, + 415 + ], + "lines": [ + { + "bbox": [ + 106, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 505, + 405 + ], + "score": 1.0, + "content": "Yen-Chang Hsu and Zsolt Kira. Neural network-based clustering using pairwise constraints. arXiv", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 403, + 254, + 415 + ], + "spans": [ + { + "bbox": [ + 115, + 403, + 254, + 415 + ], + "score": 1.0, + "content": "preprint arXiv:1511.06321, 2015.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 421, + 505, + 455 + ], + "lines": [ + { + "bbox": [ + 105, + 420, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 505, + 435 + ], + "score": 1.0, + "content": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 432, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 115, + 432, + 505, + 445 + ], + "score": 1.0, + "content": "reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 442, + 143, + 456 + ], + "spans": [ + { + "bbox": [ + 115, + 442, + 143, + 456 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 461, + 505, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "score": 1.0, + "content": "Jonathan Krause, Benjamin Sapp, Andrew Howard, Howard Zhou, Alexander Toshev, Tom Duerig,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 115, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "James Philbin, and Li Fei-Fei. The unreasonable effectiveness of noisy data for fine-grained", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 484, + 472, + 497 + ], + "spans": [ + { + "bbox": [ + 115, + 484, + 472, + 497 + ], + "score": 1.0, + "content": "recognition. In European Conference on Computer Vision, pp. 301–320. Springer, 2016.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 105, + 502, + 502, + 514 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 504, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 504, + 516 + ], + "score": 1.0, + "content": "Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 521, + 504, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "score": 1.0, + "content": "Harold W Kuhn. The hungarian method for the assignment problem. Naval Research Logistics", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 532, + 229, + 543 + ], + "spans": [ + { + "bbox": [ + 116, + 532, + 229, + 543 + ], + "score": 1.0, + "content": "(NRL), 2(1-2):83–97, 1955.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 550, + 503, + 573 + ], + "lines": [ + { + "bbox": [ + 107, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 107, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 561, + 417, + 574 + ], + "spans": [ + { + "bbox": [ + 116, + 561, + 417, + 574 + ], + "score": 1.0, + "content": "document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 580, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "score": 1.0, + "content": "Sanghoon Lee and Melba M Crawford. Unsupervised multistage image classification using hierar-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 116, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "chical clustering with a bayesian similarity measure. IEEE Transactions on Image Processing, 14", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 602, + 196, + 614 + ], + "spans": [ + { + "bbox": [ + 115, + 602, + 196, + 614 + ], + "score": 1.0, + "content": "(3):312–320, 2005.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 503, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 620, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 505, + 634 + ], + "score": 1.0, + "content": "Tomas Mikolov, Wen-tau Yih, and Geoffrey Zweig. Linguistic regularities in continuous space word", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 632, + 364, + 644 + ], + "spans": [ + { + "bbox": [ + 115, + 632, + 364, + 644 + ], + "score": 1.0, + "content": "representations. In hlt-Naacl, volume 13, pp. 746–751, 2013.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 108, + 650, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "Oren Rippel, Manohar Paluri, Piotr Dollar, and Lubomir Bourdev. Metric learning with adaptive", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 661, + 460, + 673 + ], + "spans": [ + { + "bbox": [ + 115, + 661, + 460, + 673 + ], + "score": 1.0, + "content": "density discrimination. International Conference on Learning Representations, 2016.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 680, + 435, + 692 + ], + "lines": [ + { + "bbox": [ + 105, + 679, + 433, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 433, + 693 + ], + "score": 1.0, + "content": "Eleanor Rosch. Principles of categorization. Concepts: core readings, 189, 1999.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 117, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 117, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 720, + 474, + 734 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 474, + 734 + ], + "score": 1.0, + "content": "recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 503, + 115 + ], + "lines": [], + "index": 1, + "bbox_fs": [ + 104, + 81, + 505, + 118 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 505, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 506, + 134 + ], + "score": 1.0, + "content": "We hope to expand this in future work in multiple ways. We plan on extending our approach to", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "the entire Birdsnap dataset, as well as to other fine-grained classification datasets. We also hope", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 505, + 155 + ], + "score": 1.0, + "content": "to perform more extensive analysis of the quality of the embedding spaces learned as it is obvious", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 152, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 104, + 152, + 505, + 168 + ], + "score": 1.0, + "content": "that the use of the Hungarian algorithm does not accurately reflect the performance of the clustering", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 165, + 137, + 177 + ], + "spans": [ + { + "bbox": [ + 104, + 165, + 137, + 177 + ], + "score": 1.0, + "content": "model.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5, + "bbox_fs": [ + 104, + 120, + 506, + 177 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 193, + 175, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 193, + 176, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 176, + 207 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 105, + 211, + 505, + 234 + ], + "lines": [ + { + "bbox": [ + 106, + 211, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 211, + 505, + 224 + ], + "score": 1.0, + "content": "Aurelien Bellet, Amaury Habrard, and Marc Sebban. A survey on metric learning for feature vectors ´", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 222, + 356, + 234 + ], + "spans": [ + { + "bbox": [ + 115, + 222, + 356, + 234 + ], + "score": 1.0, + "content": "and structured data. arXiv preprint arXiv:1306.6709, 2013.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 211, + 505, + 234 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 240, + 504, + 275 + ], + "lines": [ + { + "bbox": [ + 107, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 107, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "Thomas Berg, Jiongxin Liu, Seung Woo Lee, Michelle L Alexander, David W Jacobs, and Peter N", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 252, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 115, + 252, + 506, + 265 + ], + "score": 1.0, + "content": "Belhumeur. Birdsnap: Large-scale fine-grained visual categorization of birds. In Proceedings of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 263, + 477, + 276 + ], + "spans": [ + { + "bbox": [ + 115, + 263, + 477, + 276 + ], + "score": 1.0, + "content": "the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2011–2018, 2014.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 107, + 240, + 506, + 276 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 281, + 503, + 315 + ], + "lines": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 292, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 116, + 292, + 505, + 306 + ], + "score": 1.0, + "content": "mapping. In Computer vision and pattern recognition, 2006 IEEE computer society conference", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 304, + 289, + 316 + ], + "spans": [ + { + "bbox": [ + 116, + 304, + 289, + 316 + ], + "score": 1.0, + "content": "on, volume 2, pp. 1735–1742. IEEE, 2006.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 106, + 281, + 505, + 316 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 321, + 504, + 356 + ], + "lines": [ + { + "bbox": [ + 105, + 321, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 335 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 334, + 504, + 345 + ], + "spans": [ + { + "bbox": [ + 116, + 334, + 504, + 345 + ], + "score": 1.0, + "content": "human-level performance on imagenet classification. In Proceedings of the IEEE international", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 344, + 334, + 356 + ], + "spans": [ + { + "bbox": [ + 116, + 344, + 334, + 356 + ], + "score": 1.0, + "content": "conference on computer vision, pp. 1026–1034, 2015.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 321, + 505, + 356 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 362, + 503, + 385 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 114, + 374, + 254, + 385 + ], + "spans": [ + { + "bbox": [ + 114, + 374, + 254, + 385 + ], + "score": 1.0, + "content": "preprint arXiv:1503.02531, 2015.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 362, + 505, + 385 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 392, + 503, + 415 + ], + "lines": [ + { + "bbox": [ + 106, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 505, + 405 + ], + "score": 1.0, + "content": "Yen-Chang Hsu and Zsolt Kira. Neural network-based clustering using pairwise constraints. arXiv", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 403, + 254, + 415 + ], + "spans": [ + { + "bbox": [ + 115, + 403, + 254, + 415 + ], + "score": 1.0, + "content": "preprint arXiv:1511.06321, 2015.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 106, + 392, + 505, + 415 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 421, + 505, + 455 + ], + "lines": [ + { + "bbox": [ + 105, + 420, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 505, + 435 + ], + "score": 1.0, + "content": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 432, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 115, + 432, + 505, + 445 + ], + "score": 1.0, + "content": "reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 442, + 143, + 456 + ], + "spans": [ + { + "bbox": [ + 115, + 442, + 143, + 456 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 420, + 505, + 456 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 461, + 505, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "score": 1.0, + "content": "Jonathan Krause, Benjamin Sapp, Andrew Howard, Howard Zhou, Alexander Toshev, Tom Duerig,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 115, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "James Philbin, and Li Fei-Fei. The unreasonable effectiveness of noisy data for fine-grained", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 484, + 472, + 497 + ], + "spans": [ + { + "bbox": [ + 115, + 484, + 472, + 497 + ], + "score": 1.0, + "content": "recognition. In European Conference on Computer Vision, pp. 301–320. Springer, 2016.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 461, + 505, + 497 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 502, + 502, + 514 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 504, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 504, + 516 + ], + "score": 1.0, + "content": "Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30, + "bbox_fs": [ + 106, + 502, + 504, + 516 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 521, + 504, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "score": 1.0, + "content": "Harold W Kuhn. The hungarian method for the assignment problem. Naval Research Logistics", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 532, + 229, + 543 + ], + "spans": [ + { + "bbox": [ + 116, + 532, + 229, + 543 + ], + "score": 1.0, + "content": "(NRL), 2(1-2):83–97, 1955.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 520, + 505, + 543 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 550, + 503, + 573 + ], + "lines": [ + { + "bbox": [ + 107, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 107, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 561, + 417, + 574 + ], + "spans": [ + { + "bbox": [ + 116, + 561, + 417, + 574 + ], + "score": 1.0, + "content": "document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 107, + 550, + 505, + 574 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 580, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "score": 1.0, + "content": "Sanghoon Lee and Melba M Crawford. Unsupervised multistage image classification using hierar-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 116, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "chical clustering with a bayesian similarity measure. IEEE Transactions on Image Processing, 14", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 602, + 196, + 614 + ], + "spans": [ + { + "bbox": [ + 115, + 602, + 196, + 614 + ], + "score": 1.0, + "content": "(3):312–320, 2005.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 579, + 505, + 614 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 503, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 620, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 505, + 634 + ], + "score": 1.0, + "content": "Tomas Mikolov, Wen-tau Yih, and Geoffrey Zweig. Linguistic regularities in continuous space word", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 632, + 364, + 644 + ], + "spans": [ + { + "bbox": [ + 115, + 632, + 364, + 644 + ], + "score": 1.0, + "content": "representations. In hlt-Naacl, volume 13, pp. 746–751, 2013.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 106, + 620, + 505, + 644 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 650, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "Oren Rippel, Manohar Paluri, Piotr Dollar, and Lubomir Bourdev. Metric learning with adaptive", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 661, + 460, + 673 + ], + "spans": [ + { + "bbox": [ + 115, + 661, + 460, + 673 + ], + "score": 1.0, + "content": "density discrimination. International Conference on Learning Representations, 2016.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 650, + 505, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 680, + 435, + 692 + ], + "lines": [ + { + "bbox": [ + 105, + 679, + 433, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 433, + 693 + ], + "score": 1.0, + "content": "Eleanor Rosch. Principles of categorization. Concepts: core readings, 189, 1999.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 679, + 433, + 693 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 117, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 117, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 720, + 474, + 734 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 474, + 734 + ], + "score": 1.0, + "content": "recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 698, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Yevgeny Seldin, Sonia Starik, and Michael Werman. Unsupervised clustering of images using their", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "joint segmentation. In Proceedings of the 3rd International Workshop on Statistical and Compu-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 105, + 308, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 308, + 117 + ], + "score": 1.0, + "content": "tational Theories of Vision (SCTV 2003), 2003.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 503, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 505, + 136 + ], + "score": 1.0, + "content": "Alexander Strehl and Joydeep Ghosh. Cluster ensembles—a knowledge reuse framework for com-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 135, + 473, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 135, + 473, + 146 + ], + "score": 1.0, + "content": "bining multiple partitions. Journal of machine learning research, 3(Dec):583–617, 2002.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 105, + 153, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 164, + 303, + 176 + ], + "spans": [ + { + "bbox": [ + 116, + 164, + 303, + 176 + ], + "score": 1.0, + "content": "Learning Research, 9(Nov):2579–2605, 2008.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 105, + 182, + 504, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 182, + 504, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 504, + 195 + ], + "score": 1.0, + "content": "Grant Van Horn and Pietro Perona. The devil is in the tails: Fine-grained classification in the wild.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 194, + 278, + 205 + ], + "spans": [ + { + "bbox": [ + 116, + 194, + 278, + 205 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1709.01450, 2017.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 104, + 213, + 504, + 235 + ], + "lines": [ + { + "bbox": [ + 105, + 211, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 506, + 226 + ], + "score": 1.0, + "content": "Cinna Wu, Mark Tygert, and Yann LeCun. Hierarchical loss for classification. arXiv preprint", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 223, + 220, + 235 + ], + "spans": [ + { + "bbox": [ + 115, + 223, + 220, + 235 + ], + "score": 1.0, + "content": "arXiv:1709.01062, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 26, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Yevgeny Seldin, Sonia Starik, and Michael Werman. Unsupervised clustering of images using their", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "joint segmentation. In Proceedings of the 3rd International Workshop on Statistical and Compu-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 105, + 308, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 308, + 117 + ], + "score": 1.0, + "content": "tational Theories of Vision (SCTV 2003), 2003.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 82, + 505, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 503, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 505, + 136 + ], + "score": 1.0, + "content": "Alexander Strehl and Joydeep Ghosh. Cluster ensembles—a knowledge reuse framework for com-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 135, + 473, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 135, + 473, + 146 + ], + "score": 1.0, + "content": "bining multiple partitions. Journal of machine learning research, 3(Dec):583–617, 2002.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 106, + 122, + 505, + 146 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 153, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 164, + 303, + 176 + ], + "spans": [ + { + "bbox": [ + 116, + 164, + 303, + 176 + ], + "score": 1.0, + "content": "Learning Research, 9(Nov):2579–2605, 2008.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 106, + 153, + 505, + 176 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 182, + 504, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 182, + 504, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 504, + 195 + ], + "score": 1.0, + "content": "Grant Van Horn and Pietro Perona. The devil is in the tails: Fine-grained classification in the wild.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 194, + 278, + 205 + ], + "spans": [ + { + "bbox": [ + 116, + 194, + 278, + 205 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1709.01450, 2017.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 182, + 504, + 205 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 213, + 504, + 235 + ], + "lines": [ + { + "bbox": [ + 105, + 211, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 506, + 226 + ], + "score": 1.0, + "content": "Cinna Wu, Mark Tygert, and Yann LeCun. Hierarchical loss for classification. arXiv preprint", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 223, + 220, + 235 + ], + "spans": [ + { + "bbox": [ + 115, + 223, + 220, + 235 + ], + "score": 1.0, + "content": "arXiv:1709.01062, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 211, + 506, + 235 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 249, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 250, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 250, + 96 + ], + "score": 1.0, + "content": "A PREPROCESSING STEPS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 505, + 118 + ], + "score": 1.0, + "content": "We use two primary datasets, CIFAR10 and Birdsnap34, in experiments that substantiate our claims", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "score": 1.0, + "content": "of improving on the classification task by means of using clustering representations as a warm-start.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 127, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 505, + 141 + ], + "score": 1.0, + "content": "For the CIFAR10 dataset, we follow the preprocessing steps used by (Hsu & Kira, 2015). They are", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 138, + 160, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 160, + 152 + ], + "score": 1.0, + "content": "listed below,", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 132, + 159, + 505, + 223 + ], + "lines": [ + { + "bbox": [ + 133, + 160, + 286, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 286, + 172 + ], + "score": 1.0, + "content": "• Convert the images to YUV format.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 174, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 131, + 174, + 505, + 188 + ], + "score": 1.0, + "content": "• Calculate the mean and standard deviation of U and V channels over the entire training set.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 190, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 505, + 202 + ], + "score": 1.0, + "content": "• For each image, normalize by its Y channel mean and standard deviation while normalizing", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 141, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "over the remaining channels using the aggregate mean and standard deviation over the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 212, + 217, + 224 + ], + "spans": [ + { + "bbox": [ + 142, + 212, + 217, + 224 + ], + "score": 1.0, + "content": "entire training set.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 105, + 232, + 504, + 255 + ], + "lines": [ + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "score": 1.0, + "content": "NOTE: The mean and standard deviation for U and V channel as per our calculations are (0.001,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 243, + 213, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 213, + 256 + ], + "score": 1.0, + "content": "0.003) and (0.227, 0.105).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 260, + 504, + 283 + ], + "lines": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "For the Birdsnap34 dataset, we perform the standard normalization using mean and standard devia-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 272, + 436, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 436, + 283 + ], + "score": 1.0, + "content": "tions for R, G and B channels as (0.485, 0.229), (0.456, 0.224) and (0.406, 0.225).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 108, + 299, + 246, + 312 + ], + "lines": [ + { + "bbox": [ + 104, + 297, + 248, + 313 + ], + "spans": [ + { + "bbox": [ + 104, + 297, + 248, + 313 + ], + "score": 1.0, + "content": "B EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 174, + 347, + 437, + 582 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 190, + 327, + 421, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 190, + 326, + 421, + 339 + ], + "spans": [ + { + "bbox": [ + 190, + 326, + 421, + 339 + ], + "score": 1.0, + "content": "Table 6: Parameters used for Classification in Network A", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "table_body", + "bbox": [ + 174, + 347, + 437, + 582 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 174, + 347, + 437, + 582 + ], + "spans": [ + { + "bbox": [ + 174, + 347, + 437, + 582 + ], + "score": 0.985, + "html": "
ParametersValues
CIFAR10Total Epochs Batch Size Learning Rate Learning Rate Scale Factor95 16 0.01
Learning Rate Schedule Momentum Optimizer margin0.5 25,50,75,100 0.9 SGD
Weight Decay Weights of Losses Total Epochs Batch Size Learning Rate Learning1 gRate Scale Factor Birdsnap34 Learning Rate Schedule2.0 0.0005 N/A 195 16 0.01 0.5
", + "type": "table", + "image_path": "d109f64e1e19bf4dd834b8e96d8eaa5fdfa21c10b6af5980ce49aef3aade5439.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 174, + 347, + 437, + 360.05555555555554 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 174, + 360.05555555555554, + 437, + 373.1111111111111 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 174, + 373.1111111111111, + 437, + 386.16666666666663 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 174, + 386.16666666666663, + 437, + 399.2222222222222 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 174, + 399.2222222222222, + 437, + 412.2777777777777 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 174, + 412.2777777777777, + 437, + 425.33333333333326 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 174, + 425.33333333333326, + 437, + 438.3888888888888 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 174, + 438.3888888888888, + 437, + 451.44444444444434 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 174, + 451.44444444444434, + 437, + 464.4999999999999 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 174, + 464.4999999999999, + 437, + 477.55555555555543 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 174, + 477.55555555555543, + 437, + 490.611111111111 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 174, + 490.611111111111, + 437, + 503.6666666666665 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 174, + 503.6666666666665, + 437, + 516.7222222222221 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 174, + 516.7222222222221, + 437, + 529.7777777777776 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 174, + 529.7777777777776, + 437, + 542.8333333333331 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 174, + 542.8333333333331, + 437, + 555.8888888888887 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 174, + 555.8888888888887, + 437, + 568.9444444444442 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 174, + 568.9444444444442, + 437, + 581.9999999999998 + ], + "spans": [], + "index": 33 + } + ] + } + ], + "index": 19.75 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 249, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 250, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 250, + 96 + ], + "score": 1.0, + "content": "A PREPROCESSING STEPS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 505, + 118 + ], + "score": 1.0, + "content": "We use two primary datasets, CIFAR10 and Birdsnap34, in experiments that substantiate our claims", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "score": 1.0, + "content": "of improving on the classification task by means of using clustering representations as a warm-start.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 127, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 505, + 141 + ], + "score": 1.0, + "content": "For the CIFAR10 dataset, we follow the preprocessing steps used by (Hsu & Kira, 2015). They are", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 138, + 160, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 160, + 152 + ], + "score": 1.0, + "content": "listed below,", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 106, + 505, + 152 + ] + }, + { + "type": "list", + "bbox": [ + 132, + 159, + 505, + 223 + ], + "lines": [ + { + "bbox": [ + 133, + 160, + 286, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 286, + 172 + ], + "score": 1.0, + "content": "• Convert the images to YUV format.", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 131, + 174, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 131, + 174, + 505, + 188 + ], + "score": 1.0, + "content": "• Calculate the mean and standard deviation of U and V channels over the entire training set.", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 190, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 505, + 202 + ], + "score": 1.0, + "content": "• For each image, normalize by its Y channel mean and standard deviation while normalizing", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 141, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "over the remaining channels using the aggregate mean and standard deviation over the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 212, + 217, + 224 + ], + "spans": [ + { + "bbox": [ + 142, + 212, + 217, + 224 + ], + "score": 1.0, + "content": "entire training set.", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + } + ], + "index": 7, + "bbox_fs": [ + 131, + 160, + 505, + 224 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 232, + 504, + 255 + ], + "lines": [ + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "score": 1.0, + "content": "NOTE: The mean and standard deviation for U and V channel as per our calculations are (0.001,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 243, + 213, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 213, + 256 + ], + "score": 1.0, + "content": "0.003) and (0.227, 0.105).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 231, + 505, + 256 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 260, + 504, + 283 + ], + "lines": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "For the Birdsnap34 dataset, we perform the standard normalization using mean and standard devia-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 272, + 436, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 436, + 283 + ], + "score": 1.0, + "content": "tions for R, G and B channels as (0.485, 0.229), (0.456, 0.224) and (0.406, 0.225).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 260, + 505, + 283 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 299, + 246, + 312 + ], + "lines": [ + { + "bbox": [ + 104, + 297, + 248, + 313 + ], + "spans": [ + { + "bbox": [ + 104, + 297, + 248, + 313 + ], + "score": 1.0, + "content": "B EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 174, + 347, + 437, + 582 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 190, + 327, + 421, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 190, + 326, + 421, + 339 + ], + "spans": [ + { + "bbox": [ + 190, + 326, + 421, + 339 + ], + "score": 1.0, + "content": "Table 6: Parameters used for Classification in Network A", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "table_body", + "bbox": [ + 174, + 347, + 437, + 582 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 174, + 347, + 437, + 582 + ], + "spans": [ + { + "bbox": [ + 174, + 347, + 437, + 582 + ], + "score": 0.985, + "html": "
ParametersValues
CIFAR10Total Epochs Batch Size Learning Rate Learning Rate Scale Factor95 16 0.01
Learning Rate Schedule Momentum Optimizer margin0.5 25,50,75,100 0.9 SGD
Weight Decay Weights of Losses Total Epochs Batch Size Learning Rate Learning1 gRate Scale Factor Birdsnap34 Learning Rate Schedule2.0 0.0005 N/A 195 16 0.01 0.5
", + "type": "table", + "image_path": "d109f64e1e19bf4dd834b8e96d8eaa5fdfa21c10b6af5980ce49aef3aade5439.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 174, + 347, + 437, + 360.05555555555554 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 174, + 360.05555555555554, + 437, + 373.1111111111111 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 174, + 373.1111111111111, + 437, + 386.16666666666663 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 174, + 386.16666666666663, + 437, + 399.2222222222222 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 174, + 399.2222222222222, + 437, + 412.2777777777777 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 174, + 412.2777777777777, + 437, + 425.33333333333326 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 174, + 425.33333333333326, + 437, + 438.3888888888888 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 174, + 438.3888888888888, + 437, + 451.44444444444434 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 174, + 451.44444444444434, + 437, + 464.4999999999999 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 174, + 464.4999999999999, + 437, + 477.55555555555543 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 174, + 477.55555555555543, + 437, + 490.611111111111 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 174, + 490.611111111111, + 437, + 503.6666666666665 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 174, + 503.6666666666665, + 437, + 516.7222222222221 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 174, + 516.7222222222221, + 437, + 529.7777777777776 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 174, + 529.7777777777776, + 437, + 542.8333333333331 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 174, + 542.8333333333331, + 437, + 555.8888888888887 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 174, + 555.8888888888887, + 437, + 568.9444444444442 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 174, + 568.9444444444442, + 437, + 581.9999999999998 + ], + "spans": [], + "index": 33 + } + ] + } + ], + "index": 19.75 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 174, + 134, + 437, + 368 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 197, + 114, + 414, + 126 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 196, + 113, + 414, + 126 + ], + "spans": [ + { + "bbox": [ + 196, + 113, + 414, + 126 + ], + "score": 1.0, + "content": "Table 7: Parameters used for Clustering in Network A", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 174, + 134, + 437, + 368 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 174, + 134, + 437, + 368 + ], + "spans": [ + { + "bbox": [ + 174, + 134, + 437, + 368 + ], + "score": 0.983, + "html": "
ParametersValues
CIFAR10Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay Weights of Losses0.5 25,50,75,100 0.9 SGD 2.0 0.0005
Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of LossesN/A 195 1024 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.001 N/A
", + "type": "table", + "image_path": "2b2c8f03bfe04e64f0b4715f232253e163b02cfb498b4c6bf2c253df1ccd9cbc.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 174, + 134, + 437, + 147.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 174, + 147.0, + 437, + 160.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 174, + 160.0, + 437, + 173.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 174, + 173.0, + 437, + 186.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 174, + 186.0, + 437, + 199.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 174, + 199.0, + 437, + 212.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 174, + 212.0, + 437, + 225.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 174, + 225.0, + 437, + 238.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 174, + 238.0, + 437, + 251.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 174, + 251.0, + 437, + 264.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 174, + 264.0, + 437, + 277.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 174, + 277.0, + 437, + 290.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 174, + 290.0, + 437, + 303.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 174, + 303.0, + 437, + 316.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 174, + 316.0, + 437, + 329.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 174, + 329.0, + 437, + 342.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 174, + 342.0, + 437, + 355.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 174, + 355.0, + 437, + 368.0 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 4.75 + }, + { + "type": "table", + "bbox": [ + 174, + 462, + 437, + 696 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 197, + 442, + 414, + 453 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 196, + 441, + 414, + 454 + ], + "spans": [ + { + "bbox": [ + 196, + 441, + 414, + 454 + ], + "score": 1.0, + "content": "Table 8: Parameters used for Clustering in Network B", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "table_body", + "bbox": [ + 174, + 462, + 437, + 696 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 174, + 462, + 437, + 696 + ], + "spans": [ + { + "bbox": [ + 174, + 462, + 437, + 696 + ], + "score": 0.984, + "html": "
Parameters Total EpochsValues
CIFAR10Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay0.5 25,50,75,100 0.9 SGD 2.0
Weights of Losses Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of Losses0.0005 0.2, 0.8 195 128 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.0005 0.2, 0.8
", + "type": "table", + "image_path": "d965d87dc528603f4b223475197e0c8e5e3b8a84147858383ec90a6638cf1c8a.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 174, + 462, + 437, + 475.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 174, + 475.0, + 437, + 488.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 174, + 488.0, + 437, + 501.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 174, + 501.0, + 437, + 514.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 174, + 514.0, + 437, + 527.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 174, + 527.0, + 437, + 540.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 174, + 540.0, + 437, + 553.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 174, + 553.0, + 437, + 566.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 174, + 566.0, + 437, + 579.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 174, + 579.0, + 437, + 592.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 174, + 592.0, + 437, + 605.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 174, + 605.0, + 437, + 618.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 174, + 618.0, + 437, + 631.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 174, + 631.0, + 437, + 644.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 174, + 644.0, + 437, + 657.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 174, + 657.0, + 437, + 670.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 174, + 670.0, + 437, + 683.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 174, + 683.0, + 437, + 696.0 + ], + "spans": [], + "index": 37 + } + ] + } + ], + "index": 23.75 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 174, + 134, + 437, + 368 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 197, + 114, + 414, + 126 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 196, + 113, + 414, + 126 + ], + "spans": [ + { + "bbox": [ + 196, + 113, + 414, + 126 + ], + "score": 1.0, + "content": "Table 7: Parameters used for Clustering in Network A", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 174, + 134, + 437, + 368 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 174, + 134, + 437, + 368 + ], + "spans": [ + { + "bbox": [ + 174, + 134, + 437, + 368 + ], + "score": 0.983, + "html": "
ParametersValues
CIFAR10Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay Weights of Losses0.5 25,50,75,100 0.9 SGD 2.0 0.0005
Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of LossesN/A 195 1024 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.001 N/A
", + "type": "table", + "image_path": "2b2c8f03bfe04e64f0b4715f232253e163b02cfb498b4c6bf2c253df1ccd9cbc.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 174, + 134, + 437, + 147.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 174, + 147.0, + 437, + 160.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 174, + 160.0, + 437, + 173.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 174, + 173.0, + 437, + 186.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 174, + 186.0, + 437, + 199.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 174, + 199.0, + 437, + 212.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 174, + 212.0, + 437, + 225.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 174, + 225.0, + 437, + 238.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 174, + 238.0, + 437, + 251.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 174, + 251.0, + 437, + 264.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 174, + 264.0, + 437, + 277.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 174, + 277.0, + 437, + 290.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 174, + 290.0, + 437, + 303.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 174, + 303.0, + 437, + 316.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 174, + 316.0, + 437, + 329.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 174, + 329.0, + 437, + 342.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 174, + 342.0, + 437, + 355.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 174, + 355.0, + 437, + 368.0 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 4.75 + }, + { + "type": "table", + "bbox": [ + 174, + 462, + 437, + 696 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 197, + 442, + 414, + 453 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 196, + 441, + 414, + 454 + ], + "spans": [ + { + "bbox": [ + 196, + 441, + 414, + 454 + ], + "score": 1.0, + "content": "Table 8: Parameters used for Clustering in Network B", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "table_body", + "bbox": [ + 174, + 462, + 437, + 696 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 174, + 462, + 437, + 696 + ], + "spans": [ + { + "bbox": [ + 174, + 462, + 437, + 696 + ], + "score": 0.984, + "html": "
Parameters Total EpochsValues
CIFAR10Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay0.5 25,50,75,100 0.9 SGD 2.0
Weights of Losses Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of Losses0.0005 0.2, 0.8 195 128 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.0005 0.2, 0.8
", + "type": "table", + "image_path": "d965d87dc528603f4b223475197e0c8e5e3b8a84147858383ec90a6638cf1c8a.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 174, + 462, + 437, + 475.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 174, + 475.0, + 437, + 488.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 174, + 488.0, + 437, + 501.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 174, + 501.0, + 437, + 514.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 174, + 514.0, + 437, + 527.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 174, + 527.0, + 437, + 540.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 174, + 540.0, + 437, + 553.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 174, + 553.0, + 437, + 566.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 174, + 566.0, + 437, + 579.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 174, + 579.0, + 437, + 592.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 174, + 592.0, + 437, + 605.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 174, + 605.0, + 437, + 618.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 174, + 618.0, + 437, + 631.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 174, + 631.0, + 437, + 644.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 174, + 644.0, + 437, + 657.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 174, + 657.0, + 437, + 670.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 174, + 670.0, + 437, + 683.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 174, + 683.0, + 437, + 696.0 + ], + "spans": [], + "index": 37 + } + ] + } + ], + "index": 23.75 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 174, + 134, + 437, + 368 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 197, + 114, + 414, + 126 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 196, + 113, + 414, + 126 + ], + "spans": [ + { + "bbox": [ + 196, + 113, + 414, + 126 + ], + "score": 1.0, + "content": "Table 9: Parameters used for Clustering in Network C", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 174, + 134, + 437, + 368 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 174, + 134, + 437, + 368 + ], + "spans": [ + { + "bbox": [ + 174, + 134, + 437, + 368 + ], + "score": 0.983, + "html": "
Parameters Total EpochsValues
CIFAR10Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay Weights of Losses0.5 25,50, 75,100 0.9 SGD 2.0 0.0005
Total Epochs Batch Size Learning Rate Learning gRate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of Losses0.4, 0.2, 0.4 195 128 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.0005 0.4, 0.2, 0.4
", + "type": "table", + "image_path": "4dc15f0a3b71bc97215371bdfe733a0c14e1a8afac9bcc2012e4a382a9a5eae5.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 174, + 134, + 437, + 147.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 174, + 147.0, + 437, + 160.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 174, + 160.0, + 437, + 173.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 174, + 173.0, + 437, + 186.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 174, + 186.0, + 437, + 199.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 174, + 199.0, + 437, + 212.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 174, + 212.0, + 437, + 225.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 174, + 225.0, + 437, + 238.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 174, + 238.0, + 437, + 251.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 174, + 251.0, + 437, + 264.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 174, + 264.0, + 437, + 277.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 174, + 277.0, + 437, + 290.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 174, + 290.0, + 437, + 303.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 174, + 303.0, + 437, + 316.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 174, + 316.0, + 437, + 329.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 174, + 329.0, + 437, + 342.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 174, + 342.0, + 437, + 355.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 174, + 355.0, + 437, + 368.0 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 4.75 + }, + { + "type": "table", + "bbox": [ + 174, + 462, + 437, + 696 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 194, + 442, + 417, + 453 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 194, + 441, + 417, + 454 + ], + "spans": [ + { + "bbox": [ + 194, + 441, + 417, + 454 + ], + "score": 1.0, + "content": "Table 10: Parameters used for Clustering in Network D", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "table_body", + "bbox": [ + 174, + 462, + 437, + 696 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 174, + 462, + 437, + 696 + ], + "spans": [ + { + "bbox": [ + 174, + 462, + 437, + 696 + ], + "score": 0.984, + "html": "
Parameters Total EpochsValues
CIFAR10Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay0.5 25,50,75,100 0.9 SGD 2.0
Weights of Losses Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of Losses0.0005 0.4, 0.2, 0.4 195 128 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.0005 0.4, 0.2, 0.4
", + "type": "table", + "image_path": "3b325157bb9a4216b8f5fa8baff35b1fb724c5eb554b439ebc802f6db4a8f58f.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 174, + 462, + 437, + 475.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 174, + 475.0, + 437, + 488.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 174, + 488.0, + 437, + 501.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 174, + 501.0, + 437, + 514.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 174, + 514.0, + 437, + 527.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 174, + 527.0, + 437, + 540.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 174, + 540.0, + 437, + 553.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 174, + 553.0, + 437, + 566.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 174, + 566.0, + 437, + 579.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 174, + 579.0, + 437, + 592.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 174, + 592.0, + 437, + 605.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 174, + 605.0, + 437, + 618.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 174, + 618.0, + 437, + 631.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 174, + 631.0, + 437, + 644.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 174, + 644.0, + 437, + 657.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 174, + 657.0, + 437, + 670.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 174, + 670.0, + 437, + 683.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 174, + 683.0, + 437, + 696.0 + ], + "spans": [], + "index": 37 + } + ] + } + ], + "index": 23.75 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 174, + 134, + 437, + 368 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 197, + 114, + 414, + 126 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 196, + 113, + 414, + 126 + ], + "spans": [ + { + "bbox": [ + 196, + 113, + 414, + 126 + ], + "score": 1.0, + "content": "Table 9: Parameters used for Clustering in Network C", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 174, + 134, + 437, + 368 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 174, + 134, + 437, + 368 + ], + "spans": [ + { + "bbox": [ + 174, + 134, + 437, + 368 + ], + "score": 0.983, + "html": "
Parameters Total EpochsValues
CIFAR10Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay Weights of Losses0.5 25,50, 75,100 0.9 SGD 2.0 0.0005
Total Epochs Batch Size Learning Rate Learning gRate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of Losses0.4, 0.2, 0.4 195 128 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.0005 0.4, 0.2, 0.4
", + "type": "table", + "image_path": "4dc15f0a3b71bc97215371bdfe733a0c14e1a8afac9bcc2012e4a382a9a5eae5.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 174, + 134, + 437, + 147.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 174, + 147.0, + 437, + 160.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 174, + 160.0, + 437, + 173.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 174, + 173.0, + 437, + 186.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 174, + 186.0, + 437, + 199.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 174, + 199.0, + 437, + 212.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 174, + 212.0, + 437, + 225.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 174, + 225.0, + 437, + 238.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 174, + 238.0, + 437, + 251.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 174, + 251.0, + 437, + 264.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 174, + 264.0, + 437, + 277.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 174, + 277.0, + 437, + 290.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 174, + 290.0, + 437, + 303.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 174, + 303.0, + 437, + 316.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 174, + 316.0, + 437, + 329.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 174, + 329.0, + 437, + 342.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 174, + 342.0, + 437, + 355.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 174, + 355.0, + 437, + 368.0 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 4.75 + }, + { + "type": "table", + "bbox": [ + 174, + 462, + 437, + 696 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 194, + 442, + 417, + 453 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 194, + 441, + 417, + 454 + ], + "spans": [ + { + "bbox": [ + 194, + 441, + 417, + 454 + ], + "score": 1.0, + "content": "Table 10: Parameters used for Clustering in Network D", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "table_body", + "bbox": [ + 174, + 462, + 437, + 696 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 174, + 462, + 437, + 696 + ], + "spans": [ + { + "bbox": [ + 174, + 462, + 437, + 696 + ], + "score": 0.984, + "html": "
Parameters Total EpochsValues
CIFAR10Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay0.5 25,50,75,100 0.9 SGD 2.0
Weights of Losses Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of Losses0.0005 0.4, 0.2, 0.4 195 128 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.0005 0.4, 0.2, 0.4
", + "type": "table", + "image_path": "3b325157bb9a4216b8f5fa8baff35b1fb724c5eb554b439ebc802f6db4a8f58f.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 174, + 462, + 437, + 475.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 174, + 475.0, + 437, + 488.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 174, + 488.0, + 437, + 501.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 174, + 501.0, + 437, + 514.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 174, + 514.0, + 437, + 527.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 174, + 527.0, + 437, + 540.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 174, + 540.0, + 437, + 553.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 174, + 553.0, + 437, + 566.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 174, + 566.0, + 437, + 579.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 174, + 579.0, + 437, + 592.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 174, + 592.0, + 437, + 605.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 174, + 605.0, + 437, + 618.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 174, + 618.0, + 437, + 631.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 174, + 631.0, + 437, + 644.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 174, + 644.0, + 437, + 657.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 174, + 657.0, + 437, + 670.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 174, + 670.0, + 437, + 683.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 174, + 683.0, + 437, + 696.0 + ], + "spans": [], + "index": 37 + } + ] + } + ], + "index": 23.75 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/r1-4BLaQz/r1-4BLaQz_model.json b/parse/train/r1-4BLaQz/r1-4BLaQz_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7c1d442e25ec46ede38bada182840335944ff6f3 --- /dev/null +++ b/parse/train/r1-4BLaQz/r1-4BLaQz_model.json @@ -0,0 +1,9084 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1377, + 1404, + 1377, + 1404, + 1772, + 298, + 1772 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 398, + 532, + 1302, + 532, + 1302, + 990, + 398, + 990 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 299, + 1116, + 1402, + 1116, + 1402, + 1362, + 299, + 1362 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 299, + 1789, + 1403, + 1789, + 1403, + 2034, + 299, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 0, + "poly": [ + 301, + 219, + 1070, + 219, + 1070, + 268, + 301, + 268 + ], + "score": 0.941 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 104, + 298, + 104 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 302, + 1047, + 573, + 1047, + 573, + 1082, + 302, + 1082 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 773, + 464, + 927, + 464, + 927, + 497, + 773, + 497 + ], + "score": 0.882 + }, + { + "category_id": 1, + "poly": [ + 313, + 322, + 680, + 322, + 680, + 383, + 313, + 383 + ], + "score": 0.824 + }, + { + "category_id": 2, + "poly": [ + 842, + 2089, + 856, + 2089, + 856, + 2112, + 842, + 2112 + ], + "score": 0.592 + }, + { + "category_id": 13, + "poly": [ + 1214, + 868, + 1268, + 868, + 1268, + 897, + 1214, + 897 + ], + "score": 0.88, + "latex": "17 \\%" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 216.0, + 1077.0, + 216.0, + 1077.0, + 272.0, + 293.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1044.0, + 579.0, + 1044.0, + 579.0, + 1091.0, + 294.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 460.0, + 934.0, + 460.0, + 934.0, + 503.0, + 768.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2088.0, + 861.0, + 2088.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1377.0, + 1405.0, + 1377.0, + 1405.0, + 1413.0, + 296.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1407.0, + 1406.0, + 1407.0, + 1406.0, + 1444.0, + 293.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1438.0, + 1406.0, + 1438.0, + 1406.0, + 1475.0, + 293.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1470.0, + 1406.0, + 1470.0, + 1406.0, + 1502.0, + 296.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1500.0, + 1406.0, + 1500.0, + 1406.0, + 1536.0, + 294.0, + 1536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1529.0, + 1405.0, + 1529.0, + 1405.0, + 1565.0, + 294.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1559.0, + 1405.0, + 1559.0, + 1405.0, + 1596.0, + 293.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1586.0, + 1406.0, + 1586.0, + 1406.0, + 1629.0, + 292.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1621.0, + 1405.0, + 1621.0, + 1405.0, + 1657.0, + 293.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1652.0, + 1404.0, + 1652.0, + 1404.0, + 1685.0, + 292.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1683.0, + 1405.0, + 1683.0, + 1405.0, + 1715.0, + 296.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1710.0, + 1406.0, + 1710.0, + 1406.0, + 1749.0, + 293.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1745.0, + 564.0, + 1745.0, + 564.0, + 1775.0, + 294.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 531.0, + 1306.0, + 531.0, + 1306.0, + 569.0, + 395.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 564.0, + 1306.0, + 564.0, + 1306.0, + 599.0, + 394.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 595.0, + 1306.0, + 595.0, + 1306.0, + 627.0, + 393.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 623.0, + 1306.0, + 623.0, + 1306.0, + 663.0, + 391.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 653.0, + 1305.0, + 653.0, + 1305.0, + 689.0, + 394.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 684.0, + 1307.0, + 684.0, + 1307.0, + 722.0, + 392.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 716.0, + 1305.0, + 716.0, + 1305.0, + 751.0, + 393.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 747.0, + 1306.0, + 747.0, + 1306.0, + 782.0, + 394.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 778.0, + 1305.0, + 778.0, + 1305.0, + 810.0, + 394.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 806.0, + 1304.0, + 806.0, + 1304.0, + 839.0, + 395.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 838.0, + 1303.0, + 838.0, + 1303.0, + 872.0, + 393.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 869.0, + 1213.0, + 869.0, + 1213.0, + 900.0, + 394.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 869.0, + 1305.0, + 869.0, + 1305.0, + 900.0, + 1269.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 898.0, + 1306.0, + 898.0, + 1306.0, + 931.0, + 392.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 927.0, + 1307.0, + 927.0, + 1307.0, + 965.0, + 393.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 959.0, + 813.0, + 959.0, + 813.0, + 996.0, + 394.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1113.0, + 1403.0, + 1113.0, + 1403.0, + 1155.0, + 293.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1146.0, + 1404.0, + 1146.0, + 1404.0, + 1184.0, + 294.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1175.0, + 1404.0, + 1175.0, + 1404.0, + 1216.0, + 293.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1208.0, + 1406.0, + 1208.0, + 1406.0, + 1244.0, + 294.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1239.0, + 1406.0, + 1239.0, + 1406.0, + 1275.0, + 293.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1269.0, + 1404.0, + 1269.0, + 1404.0, + 1305.0, + 293.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1298.0, + 1408.0, + 1298.0, + 1408.0, + 1339.0, + 293.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1332.0, + 1329.0, + 1332.0, + 1329.0, + 1365.0, + 295.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1788.0, + 1407.0, + 1788.0, + 1407.0, + 1825.0, + 295.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1819.0, + 1407.0, + 1819.0, + 1407.0, + 1856.0, + 292.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 1404.0, + 1850.0, + 1404.0, + 1884.0, + 294.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1915.0, + 295.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1909.0, + 1407.0, + 1909.0, + 1407.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1939.0, + 1408.0, + 1939.0, + 1408.0, + 1980.0, + 292.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2007.0, + 293.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2003.0, + 1404.0, + 2003.0, + 1404.0, + 2038.0, + 292.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 322.0, + 560.0, + 322.0, + 560.0, + 355.0, + 315.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 352.0, + 682.0, + 352.0, + 682.0, + 385.0, + 311.0, + 385.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 552, + 1404, + 552, + 1404, + 1101, + 297, + 1101 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1116, + 1402, + 1116, + 1402, + 1361, + 298, + 1361 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 230, + 1403, + 230, + 1403, + 535, + 297, + 535 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1657, + 1404, + 1657, + 1404, + 1904, + 298, + 1904 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1376, + 1402, + 1376, + 1402, + 1499, + 299, + 1499 + ], + "score": 0.973 + }, + { + "category_id": 2, + "poly": [ + 296, + 1978, + 1402, + 1978, + 1402, + 2033, + 296, + 2033 + ], + "score": 0.942 + }, + { + "category_id": 0, + "poly": [ + 298, + 1572, + 1015, + 1572, + 1015, + 1608, + 298, + 1608 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.748 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 856, + 76, + 856, + 104, + 299, + 104 + ], + "score": 0.741 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 855, + 76, + 855, + 104, + 300, + 104 + ], + "score": 0.406 + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1970.0, + 1406.0, + 1970.0, + 1406.0, + 2013.0, + 329.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1027.0, + 2003.0, + 1027.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1571.0, + 1017.0, + 1571.0, + 1017.0, + 1611.0, + 294.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 553.0, + 1405.0, + 553.0, + 1405.0, + 584.0, + 297.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 583.0, + 1402.0, + 583.0, + 1402.0, + 615.0, + 295.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 614.0, + 1404.0, + 614.0, + 1404.0, + 645.0, + 295.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 644.0, + 1405.0, + 644.0, + 1405.0, + 676.0, + 296.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 671.0, + 1406.0, + 671.0, + 1406.0, + 711.0, + 291.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 702.0, + 1406.0, + 702.0, + 1406.0, + 739.0, + 292.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 731.0, + 1406.0, + 731.0, + 1406.0, + 769.0, + 292.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 766.0, + 1405.0, + 766.0, + 1405.0, + 797.0, + 295.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 797.0, + 1405.0, + 797.0, + 1405.0, + 829.0, + 295.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 828.0, + 1405.0, + 828.0, + 1405.0, + 859.0, + 295.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 857.0, + 1405.0, + 857.0, + 1405.0, + 889.0, + 295.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 887.0, + 1405.0, + 887.0, + 1405.0, + 919.0, + 295.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 916.0, + 1404.0, + 916.0, + 1404.0, + 952.0, + 294.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 946.0, + 1406.0, + 946.0, + 1406.0, + 982.0, + 292.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 972.0, + 1408.0, + 972.0, + 1408.0, + 1015.0, + 291.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1010.0, + 1402.0, + 1010.0, + 1402.0, + 1042.0, + 295.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1037.0, + 1404.0, + 1037.0, + 1404.0, + 1075.0, + 294.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1070.0, + 737.0, + 1070.0, + 737.0, + 1104.0, + 294.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1115.0, + 1407.0, + 1115.0, + 1407.0, + 1152.0, + 293.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1146.0, + 1406.0, + 1146.0, + 1406.0, + 1180.0, + 293.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1179.0, + 1404.0, + 1179.0, + 1404.0, + 1212.0, + 294.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1209.0, + 1404.0, + 1209.0, + 1404.0, + 1242.0, + 294.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1238.0, + 1406.0, + 1238.0, + 1406.0, + 1272.0, + 294.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1268.0, + 1403.0, + 1268.0, + 1403.0, + 1302.0, + 294.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1300.0, + 1406.0, + 1300.0, + 1406.0, + 1333.0, + 294.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1330.0, + 1305.0, + 1330.0, + 1305.0, + 1363.0, + 294.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 232.0, + 1404.0, + 232.0, + 1404.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 295.0, + 292.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 291.0, + 1405.0, + 291.0, + 1405.0, + 324.0, + 292.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 318.0, + 1404.0, + 318.0, + 1404.0, + 360.0, + 291.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 352.0, + 1407.0, + 352.0, + 1407.0, + 388.0, + 292.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 384.0, + 1404.0, + 384.0, + 1404.0, + 415.0, + 296.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 413.0, + 1405.0, + 413.0, + 1405.0, + 448.0, + 295.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 443.0, + 1405.0, + 443.0, + 1405.0, + 479.0, + 293.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 475.0, + 1405.0, + 475.0, + 1405.0, + 509.0, + 292.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 504.0, + 358.0, + 504.0, + 358.0, + 538.0, + 292.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1659.0, + 1404.0, + 1659.0, + 1404.0, + 1693.0, + 294.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1691.0, + 1404.0, + 1691.0, + 1404.0, + 1721.0, + 296.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1721.0, + 1405.0, + 1721.0, + 1405.0, + 1753.0, + 293.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1753.0, + 1404.0, + 1753.0, + 1404.0, + 1783.0, + 296.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1780.0, + 1406.0, + 1780.0, + 1406.0, + 1815.0, + 293.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1812.0, + 1406.0, + 1812.0, + 1406.0, + 1846.0, + 294.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1841.0, + 1407.0, + 1841.0, + 1407.0, + 1877.0, + 293.0, + 1877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1871.0, + 1404.0, + 1871.0, + 1404.0, + 1909.0, + 292.0, + 1909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1376.0, + 1404.0, + 1376.0, + 1404.0, + 1408.0, + 295.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1407.0, + 1404.0, + 1407.0, + 1404.0, + 1440.0, + 294.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1438.0, + 1406.0, + 1438.0, + 1406.0, + 1472.0, + 292.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1467.0, + 412.0, + 1467.0, + 412.0, + 1501.0, + 291.0, + 1501.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1696, + 1405, + 1696, + 1405, + 2035, + 297, + 2035 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 296, + 828, + 1404, + 828, + 1404, + 984, + 296, + 984 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1087, + 1403, + 1087, + 1403, + 1212, + 298, + 1212 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1509, + 1403, + 1509, + 1403, + 1605, + 299, + 1605 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 425, + 240, + 1290, + 240, + 1290, + 610, + 425, + 610 + ], + "score": 0.972 + }, + { + "category_id": 8, + "poly": [ + 659, + 1399, + 1036, + 1399, + 1036, + 1488, + 659, + 1488 + ], + "score": 0.955 + }, + { + "category_id": 4, + "poly": [ + 297, + 649, + 1400, + 649, + 1400, + 713, + 297, + 713 + ], + "score": 0.935 + }, + { + "category_id": 8, + "poly": [ + 429, + 1025, + 1267, + 1025, + 1267, + 1066, + 429, + 1066 + ], + "score": 0.93 + }, + { + "category_id": 8, + "poly": [ + 632, + 1254, + 1067, + 1254, + 1067, + 1293, + 632, + 1293 + ], + "score": 0.928 + }, + { + "category_id": 0, + "poly": [ + 299, + 771, + 613, + 771, + 613, + 803, + 299, + 803 + ], + "score": 0.919 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 857, + 75, + 857, + 105, + 297, + 105 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 300, + 1640, + 704, + 1640, + 704, + 1672, + 300, + 1672 + ], + "score": 0.904 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1030, + 1400, + 1030, + 1400, + 1061, + 1366, + 1061 + ], + "score": 0.887 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1427, + 1400, + 1427, + 1400, + 1457, + 1366, + 1457 + ], + "score": 0.887 + }, + { + "category_id": 8, + "poly": [ + 402, + 1326, + 1295, + 1326, + 1295, + 1366, + 402, + 1366 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1259, + 1400, + 1259, + 1400, + 1288, + 1366, + 1288 + ], + "score": 0.875 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1331, + 1400, + 1331, + 1400, + 1361, + 1366, + 1361 + ], + "score": 0.863 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.638 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.447 + }, + { + "category_id": 14, + "poly": [ + 660, + 1394, + 1039, + 1394, + 1039, + 1489, + 660, + 1489 + ], + "score": 0.93, + "latex": "K L ( \\mathbf { P } \\parallel \\mathbf { Q } ) = \\sum _ { i = 1 } ^ { k } P _ { i } \\log \\Big ( \\frac { P _ { i } } { Q _ { i } } \\Big ) ." + }, + { + "category_id": 13, + "poly": [ + 1059, + 861, + 1164, + 861, + 1164, + 894, + 1059, + 894 + ], + "score": 0.93, + "latex": "( X _ { 1 } , X _ { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 769, + 921, + 840, + 921, + 840, + 955, + 769, + 955 + ], + "score": 0.92, + "latex": "L _ { D } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 421, + 892, + 526, + 892, + 526, + 924, + 421, + 924 + ], + "score": 0.92, + "latex": "D ( X , Y )" + }, + { + "category_id": 13, + "poly": [ + 1290, + 891, + 1356, + 891, + 1356, + 924, + 1290, + 924 + ], + "score": 0.91, + "latex": "L _ { S } ( \\cdot )" + }, + { + "category_id": 14, + "poly": [ + 429, + 1025, + 1270, + 1025, + 1270, + 1066, + 429, + 1066 + ], + "score": 0.9, + "latex": "\\boldsymbol { \\mathcal { L } } ( \\mathbf { X _ { 1 } } , \\mathbf { X _ { 2 } } ) = \\mathbb { 1 } _ { X _ { 1 } , X _ { 2 } } L _ { S } ( D ( \\mathbf { X _ { 1 } } , \\mathbf { X _ { 2 } } ) ) + ( 1 - \\mathbb { 1 } _ { X _ { 1 } , X _ { 2 } } ) L _ { D } ( D ( \\mathbf { X _ { 1 } } , \\mathbf { X _ { 2 } } ) ) ." + }, + { + "category_id": 13, + "poly": [ + 383, + 863, + 466, + 863, + 466, + 893, + 383, + 893 + ], + "score": 0.9, + "latex": "\\mathbb { 1 } _ { X _ { 1 } , X _ { 2 } }" + }, + { + "category_id": 14, + "poly": [ + 629, + 1254, + 1068, + 1254, + 1068, + 1293, + 629, + 1293 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { \\mathcal { L } ( \\mathbf { P } , \\mathbf { Q } ) = l o s s ( \\mathbf { P } , \\mathbf { Q } ) + l o s s ( \\mathbf { Q } , \\mathbf { P } ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 340, + 831, + 438, + 831, + 438, + 862, + 340, + 862 + ], + "score": 0.89, + "latex": "( X _ { 1 } , X _ { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 555, + 1573, + 580, + 1573, + 580, + 1604, + 555, + 1604 + ], + "score": 0.86, + "latex": "Q" + }, + { + "category_id": 13, + "poly": [ + 480, + 1573, + 504, + 1573, + 504, + 1599, + 480, + 1599 + ], + "score": 0.84, + "latex": "P" + }, + { + "category_id": 14, + "poly": [ + 402, + 1325, + 1294, + 1325, + 1294, + 1368, + 402, + 1368 + ], + "score": 0.79, + "latex": "l o s s ( \\mathbf { P } , \\mathbf { Q } ) = \\mathbb { 1 } _ { P , Q } K L ( \\mathbf { P } | | \\mathbf { Q } ) + ( 1 - \\mathbb { 1 } _ { P , Q } ) \\operatorname* { m a x } \\{ m a r g i n - K L ( \\mathbf { P } | | \\mathbf { Q } ) , 0 \\} ," + }, + { + "category_id": 15, + "poly": [ + 750.0, + 285.0, + 895.0, + 285.0, + 895.0, + 313.0, + 750.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 285.0, + 1145.0, + 285.0, + 1145.0, + 313.0, + 1001.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 305.0, + 629.0, + 305.0, + 629.0, + 329.0, + 459.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 309.0, + 862.0, + 309.0, + 862.0, + 336.0, + 786.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 307.0, + 1112.0, + 307.0, + 1112.0, + 336.0, + 1036.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 481.0, + 629.0, + 481.0, + 629.0, + 506.0, + 459.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 461.0, + 895.0, + 461.0, + 895.0, + 512.0, + 750.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 461.0, + 1145.0, + 461.0, + 1145.0, + 513.0, + 1000.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 465.0, + 1292.0, + 465.0, + 1292.0, + 506.0, + 1252.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.75, + 300.5, + 1269.75, + 300.5, + 1269.75, + 320.5, + 1265.75, + 320.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.25, + 577.0, + 850.25, + 577.0, + 850.25, + 608.0, + 820.25, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 649.0, + 1405.0, + 649.0, + 1405.0, + 685.0, + 296.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 680.0, + 1006.0, + 680.0, + 1006.0, + 716.0, + 296.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 771.0, + 617.0, + 771.0, + 617.0, + 807.0, + 295.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1640.0, + 706.0, + 1640.0, + 706.0, + 1676.0, + 294.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1698.0, + 1405.0, + 1698.0, + 1405.0, + 1733.0, + 296.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1731.0, + 1403.0, + 1731.0, + 1403.0, + 1762.0, + 296.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1759.0, + 1406.0, + 1759.0, + 1406.0, + 1794.0, + 294.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1789.0, + 1406.0, + 1789.0, + 1406.0, + 1824.0, + 294.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1856.0, + 294.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1851.0, + 1407.0, + 1851.0, + 1407.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1880.0, + 1405.0, + 1880.0, + 1405.0, + 1915.0, + 292.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1947.0, + 295.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1943.0, + 1402.0, + 1943.0, + 1402.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1405.0, + 2003.0, + 1405.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 827.0, + 339.0, + 827.0, + 339.0, + 866.0, + 293.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 827.0, + 1405.0, + 827.0, + 1405.0, + 866.0, + 439.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 859.0, + 382.0, + 859.0, + 382.0, + 895.0, + 291.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 859.0, + 1058.0, + 859.0, + 1058.0, + 895.0, + 467.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 859.0, + 1406.0, + 859.0, + 1406.0, + 895.0, + 1165.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 887.0, + 420.0, + 887.0, + 420.0, + 927.0, + 293.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 887.0, + 1289.0, + 887.0, + 1289.0, + 927.0, + 527.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1357.0, + 887.0, + 1406.0, + 887.0, + 1406.0, + 927.0, + 1357.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 918.0, + 768.0, + 918.0, + 768.0, + 958.0, + 291.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 918.0, + 1407.0, + 918.0, + 1407.0, + 958.0, + 841.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 953.0, + 1264.0, + 953.0, + 1264.0, + 986.0, + 296.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1085.0, + 1407.0, + 1085.0, + 1407.0, + 1124.0, + 292.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1117.0, + 1407.0, + 1117.0, + 1407.0, + 1154.0, + 292.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1148.0, + 1407.0, + 1148.0, + 1407.0, + 1185.0, + 293.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1178.0, + 1246.0, + 1178.0, + 1246.0, + 1214.0, + 293.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1508.0, + 1405.0, + 1508.0, + 1405.0, + 1545.0, + 294.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1544.0, + 1404.0, + 1544.0, + 1404.0, + 1574.0, + 295.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1572.0, + 479.0, + 1572.0, + 479.0, + 1606.0, + 295.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1572.0, + 554.0, + 1572.0, + 554.0, + 1606.0, + 505.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 1572.0, + 766.0, + 1572.0, + 766.0, + 1606.0, + 581.0, + 1606.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1033, + 1404, + 1033, + 1404, + 1339, + 298, + 1339 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1642, + 1405, + 1642, + 1405, + 1980, + 298, + 1980 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1455, + 1403, + 1455, + 1403, + 1608, + 299, + 1608 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 863, + 1403, + 863, + 1403, + 1017, + 299, + 1017 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 295, + 229, + 1400, + 229, + 1400, + 627, + 295, + 627 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 296, + 648, + 1404, + 648, + 1404, + 802, + 296, + 802 + ], + "score": 0.965 + }, + { + "category_id": 0, + "poly": [ + 300, + 1384, + 557, + 1384, + 557, + 1419, + 300, + 1419 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 334, + 2006, + 527, + 2006, + 527, + 2035, + 334, + 2035 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.862 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.761 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 855, + 76, + 855, + 104, + 299, + 104 + ], + "score": 0.12 + }, + { + "category_id": 13, + "poly": [ + 793, + 1796, + 916, + 1796, + 916, + 1829, + 793, + 1829 + ], + "score": 0.92, + "latex": "1 : ( n - 1 )" + }, + { + "category_id": 13, + "poly": [ + 1192, + 1770, + 1213, + 1770, + 1213, + 1793, + 1192, + 1793 + ], + "score": 0.73, + "latex": "n" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 224.0, + 387.0, + 224.0, + 387.0, + 250.0, + 296.0, + 250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 257.0, + 429.0, + 257.0, + 429.0, + 288.0, + 313.0, + 288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 259.0, + 569.0, + 259.0, + 569.0, + 284.0, + 473.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 261.0, + 757.0, + 261.0, + 757.0, + 285.0, + 662.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 258.0, + 947.0, + 258.0, + 947.0, + 287.0, + 773.0, + 287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 249.0, + 1198.0, + 249.0, + 1198.0, + 274.0, + 1016.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 311.0, + 387.0, + 311.0, + 387.0, + 336.0, + 298.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 341.0, + 795.0, + 341.0, + 795.0, + 350.0, + 785.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 345.0, + 428.0, + 345.0, + 428.0, + 374.0, + 314.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 345.0, + 571.0, + 345.0, + 571.0, + 371.0, + 476.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 351.0, + 653.0, + 351.0, + 653.0, + 366.0, + 637.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 346.0, + 760.0, + 346.0, + 760.0, + 371.0, + 663.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 346.0, + 939.0, + 346.0, + 939.0, + 371.0, + 790.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 366.0, + 1401.0, + 366.0, + 1401.0, + 397.0, + 1209.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 388.0, + 1374.0, + 388.0, + 1374.0, + 415.0, + 1221.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 402.0, + 387.0, + 402.0, + 387.0, + 428.0, + 298.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 434.0, + 430.0, + 434.0, + 430.0, + 469.0, + 315.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 446.0, + 462.0, + 446.0, + 462.0, + 457.0, + 448.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 436.0, + 572.0, + 436.0, + 572.0, + 465.0, + 475.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 444.0, + 654.0, + 444.0, + 654.0, + 459.0, + 635.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 438.0, + 760.0, + 438.0, + 760.0, + 463.0, + 663.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 438.0, + 938.0, + 438.0, + 938.0, + 463.0, + 767.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 462.0, + 1399.0, + 462.0, + 1399.0, + 486.0, + 1232.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 482.0, + 1391.0, + 482.0, + 1391.0, + 507.0, + 1240.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 495.0, + 388.0, + 495.0, + 388.0, + 521.0, + 298.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 527.0, + 430.0, + 527.0, + 430.0, + 562.0, + 315.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 541.0, + 457.0, + 541.0, + 457.0, + 551.0, + 448.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 530.0, + 570.0, + 530.0, + 570.0, + 555.0, + 475.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 537.0, + 627.0, + 537.0, + 627.0, + 558.0, + 607.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 525.0, + 653.0, + 525.0, + 653.0, + 550.0, + 639.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 532.0, + 758.0, + 532.0, + 758.0, + 558.0, + 662.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 528.0, + 944.0, + 528.0, + 944.0, + 559.0, + 768.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 554.0, + 777.0, + 554.0, + 777.0, + 564.0, + 767.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 586.0, + 585.0, + 586.0, + 585.0, + 618.0, + 448.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 350.0, + 798.0, + 350.0, + 798.0, + 367.0, + 765.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 650.0, + 1405.0, + 650.0, + 1405.0, + 683.0, + 295.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 677.0, + 1405.0, + 677.0, + 1405.0, + 716.0, + 293.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 708.0, + 1405.0, + 708.0, + 1405.0, + 746.0, + 293.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 738.0, + 1406.0, + 738.0, + 1406.0, + 778.0, + 293.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 771.0, + 652.0, + 771.0, + 652.0, + 807.0, + 295.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1380.0, + 562.0, + 1380.0, + 562.0, + 1426.0, + 291.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1996.0, + 533.0, + 1996.0, + 533.0, + 2043.0, + 329.0, + 2043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1032.0, + 1408.0, + 1032.0, + 1408.0, + 1068.0, + 296.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1064.0, + 1406.0, + 1064.0, + 1406.0, + 1100.0, + 296.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1095.0, + 1404.0, + 1095.0, + 1404.0, + 1130.0, + 294.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1123.0, + 1405.0, + 1123.0, + 1405.0, + 1161.0, + 292.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1154.0, + 1404.0, + 1154.0, + 1404.0, + 1189.0, + 293.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1186.0, + 1401.0, + 1186.0, + 1401.0, + 1218.0, + 296.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1218.0, + 1404.0, + 1218.0, + 1404.0, + 1249.0, + 296.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1247.0, + 1404.0, + 1247.0, + 1404.0, + 1281.0, + 293.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1276.0, + 1404.0, + 1276.0, + 1404.0, + 1313.0, + 291.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1307.0, + 957.0, + 1307.0, + 957.0, + 1343.0, + 292.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1643.0, + 1405.0, + 1643.0, + 1405.0, + 1678.0, + 296.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1674.0, + 1406.0, + 1674.0, + 1406.0, + 1709.0, + 295.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1703.0, + 1406.0, + 1703.0, + 1406.0, + 1741.0, + 293.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1736.0, + 1406.0, + 1736.0, + 1406.0, + 1770.0, + 292.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1766.0, + 1191.0, + 1766.0, + 1191.0, + 1799.0, + 293.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1214.0, + 1766.0, + 1405.0, + 1766.0, + 1405.0, + 1799.0, + 1214.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1796.0, + 792.0, + 1796.0, + 792.0, + 1831.0, + 293.0, + 1831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1796.0, + 1406.0, + 1796.0, + 1406.0, + 1831.0, + 917.0, + 1831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1825.0, + 1405.0, + 1825.0, + 1405.0, + 1862.0, + 293.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1857.0, + 1403.0, + 1857.0, + 1403.0, + 1892.0, + 295.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1883.0, + 1406.0, + 1883.0, + 1406.0, + 1926.0, + 292.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1917.0, + 1403.0, + 1917.0, + 1403.0, + 1952.0, + 295.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1949.0, + 1403.0, + 1949.0, + 1403.0, + 1984.0, + 295.0, + 1984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1454.0, + 1405.0, + 1454.0, + 1405.0, + 1491.0, + 297.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1481.0, + 1405.0, + 1481.0, + 1405.0, + 1522.0, + 293.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1517.0, + 1405.0, + 1517.0, + 1405.0, + 1550.0, + 297.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1546.0, + 1404.0, + 1546.0, + 1404.0, + 1579.0, + 295.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1572.0, + 439.0, + 1572.0, + 439.0, + 1615.0, + 293.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 865.0, + 1404.0, + 865.0, + 1404.0, + 898.0, + 295.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 895.0, + 1405.0, + 895.0, + 1405.0, + 931.0, + 293.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 923.0, + 1404.0, + 923.0, + 1404.0, + 961.0, + 294.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 956.0, + 1407.0, + 956.0, + 1407.0, + 993.0, + 295.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 984.0, + 902.0, + 984.0, + 902.0, + 1023.0, + 293.0, + 1023.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1171, + 1404, + 1171, + 1404, + 1417, + 297, + 1417 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 631, + 1405, + 631, + 1405, + 877, + 297, + 877 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 367, + 1403, + 367, + 1403, + 522, + 298, + 522 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1403, + 229, + 1403, + 352, + 299, + 352 + ], + "score": 0.975 + }, + { + "category_id": 5, + "poly": [ + 521, + 985, + 1174, + 985, + 1174, + 1119, + 521, + 1119 + ], + "score": 0.974, + "html": "
CIFAR-10Birdsnap-34
Classification75.4853.50
Clustering74.2117.20
Warm Start Classification75.7550.32
" + }, + { + "category_id": 3, + "poly": [ + 359, + 1462, + 1348, + 1462, + 1348, + 1942, + 359, + 1942 + ], + "score": 0.971 + }, + { + "category_id": 4, + "poly": [ + 474, + 1967, + 1229, + 1967, + 1229, + 2000, + 474, + 2000 + ], + "score": 0.927 + }, + { + "category_id": 6, + "poly": [ + 474, + 915, + 1225, + 915, + 1225, + 949, + 474, + 949 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.887 + }, + { + "category_id": 0, + "poly": [ + 299, + 569, + 1030, + 569, + 1030, + 601, + 299, + 601 + ], + "score": 0.854 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.752 + }, + { + "category_id": 15, + "poly": [ + 1227.0, + 1520.0, + 1238.0, + 1520.0, + 1238.0, + 1539.0, + 1227.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 1606.0, + 519.0, + 1606.0, + 519.0, + 1619.0, + 503.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 1606.0, + 1248.0, + 1606.0, + 1248.0, + 1618.0, + 1231.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 1623.0, + 519.0, + 1623.0, + 519.0, + 1645.0, + 502.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 1718.0, + 947.0, + 1718.0, + 947.0, + 1727.0, + 939.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1739.0, + 504.0, + 1739.0, + 504.0, + 1795.0, + 470.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 1737.0, + 1331.0, + 1737.0, + 1331.0, + 1784.0, + 1306.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 1813.0, + 638.0, + 1813.0, + 638.0, + 1822.0, + 624.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 486.0, + 1912.0, + 664.0, + 1912.0, + 664.0, + 1944.0, + 486.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 1908.0, + 1200.0, + 1908.0, + 1200.0, + 1949.0, + 1049.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.25, + 1754.0, + 661.25, + 1754.0, + 661.25, + 1816.0, + 599.25, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1727.5, + 966.0, + 1727.5, + 966.0, + 1746.5, + 941.0, + 1746.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.25, + 1730.0, + 1126.25, + 1730.0, + 1126.25, + 1742.0, + 1100.25, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1818.5, + 503.0, + 1818.5, + 503.0, + 1837.0, + 495.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 1785.5, + 560.0, + 1785.5, + 560.0, + 1802.5, + 541.0, + 1802.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 1821.0, + 588.0, + 1821.0, + 588.0, + 1845.0, + 557.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.25, + 1863.0, + 609.25, + 1863.0, + 609.25, + 1878.5, + 587.25, + 1878.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 1965.0, + 1227.0, + 1965.0, + 1227.0, + 2003.0, + 471.0, + 2003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 914.0, + 1228.0, + 914.0, + 1228.0, + 950.0, + 474.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 567.0, + 1035.0, + 567.0, + 1035.0, + 607.0, + 294.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1173.0, + 1404.0, + 1173.0, + 1404.0, + 1206.0, + 295.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1200.0, + 1406.0, + 1200.0, + 1406.0, + 1239.0, + 291.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1233.0, + 1405.0, + 1233.0, + 1405.0, + 1270.0, + 294.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1263.0, + 1404.0, + 1263.0, + 1404.0, + 1299.0, + 294.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1293.0, + 1406.0, + 1293.0, + 1406.0, + 1329.0, + 294.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1323.0, + 1405.0, + 1323.0, + 1405.0, + 1359.0, + 294.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1354.0, + 1405.0, + 1354.0, + 1405.0, + 1390.0, + 294.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1387.0, + 641.0, + 1387.0, + 641.0, + 1420.0, + 295.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 628.0, + 1405.0, + 628.0, + 1405.0, + 670.0, + 291.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 661.0, + 1406.0, + 661.0, + 1406.0, + 696.0, + 292.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 694.0, + 1405.0, + 694.0, + 1405.0, + 728.0, + 295.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 723.0, + 1405.0, + 723.0, + 1405.0, + 758.0, + 292.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 756.0, + 1406.0, + 756.0, + 1406.0, + 786.0, + 295.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 785.0, + 1405.0, + 785.0, + 1405.0, + 819.0, + 294.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 813.0, + 1405.0, + 813.0, + 1405.0, + 850.0, + 292.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 847.0, + 972.0, + 847.0, + 972.0, + 880.0, + 294.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 367.0, + 1405.0, + 367.0, + 1405.0, + 405.0, + 294.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 397.0, + 1405.0, + 397.0, + 1405.0, + 435.0, + 293.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 426.0, + 1405.0, + 426.0, + 1405.0, + 467.0, + 293.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 459.0, + 1405.0, + 459.0, + 1405.0, + 496.0, + 294.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 491.0, + 763.0, + 491.0, + 763.0, + 525.0, + 293.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 231.0, + 1404.0, + 231.0, + 1404.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 294.0, + 294.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 289.0, + 1405.0, + 289.0, + 1405.0, + 328.0, + 292.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 322.0, + 438.0, + 322.0, + 438.0, + 352.0, + 295.0, + 352.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 345, + 1521, + 1353, + 1521, + 1353, + 1868, + 345, + 1868 + ], + "score": 0.986, + "html": "
Clustering Acc.NMIPurityClassification Acc.
CIFAR10Cross-Entropy75.48
LeNet A74.210.550.7475.75
LeNet B72.660.540.7375.42
LeNet C72.840.550.7375.11
LeNet D74.170.560.7475.53
Birdsnap34Cross-Entropy11153.50
AlexNet A17.200.390.1750.32
AlexNet B20.380.450.2045.86
AlexNet C24.840.510.2570.70
AlexNet D22.290.490.2265.61
" + }, + { + "category_id": 5, + "poly": [ + 345, + 600, + 1353, + 600, + 1353, + 825, + 345, + 825 + ], + "score": 0.984, + "html": "
Clustering Acc.NMIPurityClassification Acc.
CIFAR10Cross-Entropy LeNet A LeNet B1 74.21 72.661 0.551 0.7475.48 75.75
Birdsnap34Cross-Entropy AlexNet A AlexNetB1 17.20 20.380.54 1 0.39 0.450.73 1 0.17 0.2075.42 53.50 50.32 45.86
" + }, + { + "category_id": 1, + "poly": [ + 297, + 286, + 1404, + 286, + 1404, + 502, + 297, + 502 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1176, + 1404, + 1176, + 1404, + 1423, + 297, + 1423 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 868, + 1405, + 868, + 1405, + 1084, + 297, + 1084 + ], + "score": 0.976 + }, + { + "category_id": 6, + "poly": [ + 546, + 1450, + 1155, + 1450, + 1155, + 1484, + 546, + 1484 + ], + "score": 0.915 + }, + { + "category_id": 6, + "poly": [ + 657, + 530, + 1044, + 530, + 1044, + 563, + 657, + 563 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.904 + }, + { + "category_id": 0, + "poly": [ + 300, + 229, + 617, + 229, + 617, + 261, + 300, + 261 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 301, + 1119, + 714, + 1119, + 714, + 1151, + 301, + 1151 + ], + "score": 0.88 + }, + { + "category_id": 1, + "poly": [ + 297, + 1911, + 1404, + 1911, + 1404, + 2035, + 297, + 2035 + ], + "score": 0.844 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 859, + 2089, + 859, + 2112, + 840, + 2112 + ], + "score": 0.793 + }, + { + "category_id": 7, + "poly": [ + 297, + 1911, + 1404, + 1911, + 1404, + 2035, + 297, + 2035 + ], + "score": 0.095 + }, + { + "category_id": 15, + "poly": [ + 543.0, + 1449.0, + 1157.0, + 1449.0, + 1157.0, + 1485.0, + 543.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 527.0, + 1045.0, + 527.0, + 1045.0, + 565.0, + 654.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 620.0, + 229.0, + 620.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1116.0, + 718.0, + 1116.0, + 718.0, + 1155.0, + 293.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 862.0, + 2087.0, + 862.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1404.0, + 1910.0, + 1404.0, + 1945.0, + 294.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1943.0, + 1406.0, + 1943.0, + 1406.0, + 1975.0, + 294.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2003.0, + 1404.0, + 2003.0, + 1404.0, + 2036.0, + 296.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 288.0, + 1404.0, + 288.0, + 1404.0, + 322.0, + 295.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 319.0, + 1402.0, + 319.0, + 1402.0, + 353.0, + 294.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 348.0, + 1404.0, + 348.0, + 1404.0, + 383.0, + 294.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 379.0, + 1405.0, + 379.0, + 1405.0, + 414.0, + 295.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 410.0, + 1404.0, + 410.0, + 1404.0, + 445.0, + 294.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 437.0, + 1406.0, + 437.0, + 1406.0, + 477.0, + 292.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 472.0, + 716.0, + 472.0, + 716.0, + 506.0, + 295.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1174.0, + 1406.0, + 1174.0, + 1406.0, + 1214.0, + 292.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1208.0, + 1402.0, + 1208.0, + 1402.0, + 1242.0, + 295.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1238.0, + 1404.0, + 1238.0, + 1404.0, + 1271.0, + 292.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1268.0, + 1406.0, + 1268.0, + 1406.0, + 1306.0, + 294.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1300.0, + 1404.0, + 1300.0, + 1404.0, + 1333.0, + 295.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1330.0, + 1405.0, + 1330.0, + 1405.0, + 1367.0, + 294.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1359.0, + 1406.0, + 1359.0, + 1406.0, + 1397.0, + 294.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1392.0, + 1171.0, + 1392.0, + 1171.0, + 1426.0, + 296.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 867.0, + 1405.0, + 867.0, + 1405.0, + 904.0, + 295.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 899.0, + 1405.0, + 899.0, + 1405.0, + 934.0, + 295.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 929.0, + 1406.0, + 929.0, + 1406.0, + 966.0, + 292.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 960.0, + 1405.0, + 960.0, + 1405.0, + 995.0, + 294.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 990.0, + 1405.0, + 990.0, + 1405.0, + 1025.0, + 294.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1022.0, + 1405.0, + 1022.0, + 1405.0, + 1056.0, + 294.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1050.0, + 1153.0, + 1050.0, + 1153.0, + 1087.0, + 294.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1404.0, + 1910.0, + 1404.0, + 1945.0, + 294.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1943.0, + 1406.0, + 1943.0, + 1406.0, + 1975.0, + 294.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2003.0, + 1404.0, + 2003.0, + 1404.0, + 2036.0, + 296.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 229, + 1404, + 229, + 1404, + 474, + 298, + 474 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 575, + 1403, + 575, + 1403, + 760, + 298, + 760 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1850, + 1403, + 1850, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1283, + 1405, + 1283, + 1405, + 1560, + 297, + 1560 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1681, + 1402, + 1681, + 1402, + 1834, + 298, + 1834 + ], + "score": 0.979 + }, + { + "category_id": 5, + "poly": [ + 395, + 859, + 1302, + 859, + 1302, + 992, + 395, + 992 + ], + "score": 0.977, + "html": "
Clustering Acc.NMIPurityClassification Acc.
Cross-Entropy一1153.50
AlexNet A (Double)22.930.490.2359.24
AlexNet C24.840.510.2570.70
" + }, + { + "category_id": 5, + "poly": [ + 425, + 1128, + 1265, + 1128, + 1265, + 1230, + 425, + 1230 + ], + "score": 0.97, + "html": "
Clustering Acc.NMIPurityClassification Acc.
Cross-Entropy11181.53
AlexNet C24.200.510.2482.80
" + }, + { + "category_id": 6, + "poly": [ + 518, + 1057, + 1182, + 1057, + 1182, + 1091, + 518, + 1091 + ], + "score": 0.92 + }, + { + "category_id": 6, + "poly": [ + 628, + 788, + 1071, + 788, + 1071, + 821, + 628, + 821 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 302, + 1609, + 825, + 1609, + 825, + 1644, + 302, + 1644 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.868 + }, + { + "category_id": 0, + "poly": [ + 301, + 517, + 1115, + 517, + 1115, + 549, + 301, + 549 + ], + "score": 0.801 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2111, + 842, + 2111 + ], + "score": 0.634 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2111, + 842, + 2111 + ], + "score": 0.263 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 855, + 76, + 855, + 104, + 300, + 104 + ], + "score": 0.124 + }, + { + "category_id": 1, + "poly": [ + 301, + 517, + 1115, + 517, + 1115, + 549, + 301, + 549 + ], + "score": 0.107 + }, + { + "category_id": 13, + "poly": [ + 1340, + 1973, + 1393, + 1973, + 1393, + 2002, + 1340, + 2002 + ], + "score": 0.86, + "latex": "10 \\%" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 1055.0, + 1181.0, + 1055.0, + 1181.0, + 1095.0, + 518.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 782.0, + 1073.0, + 782.0, + 1073.0, + 829.0, + 623.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1605.0, + 832.0, + 1605.0, + 832.0, + 1651.0, + 292.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 514.0, + 1121.0, + 514.0, + 1121.0, + 554.0, + 293.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 231.0, + 1404.0, + 231.0, + 1404.0, + 264.0, + 294.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 260.0, + 1405.0, + 260.0, + 1405.0, + 298.0, + 293.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 293.0, + 1404.0, + 293.0, + 1404.0, + 326.0, + 294.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 322.0, + 1405.0, + 322.0, + 1405.0, + 356.0, + 292.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 351.0, + 1404.0, + 351.0, + 1404.0, + 387.0, + 293.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 381.0, + 1405.0, + 381.0, + 1405.0, + 417.0, + 292.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 410.0, + 1404.0, + 410.0, + 1404.0, + 451.0, + 293.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 444.0, + 516.0, + 444.0, + 516.0, + 474.0, + 296.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 577.0, + 1404.0, + 577.0, + 1404.0, + 609.0, + 296.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 608.0, + 1404.0, + 608.0, + 1404.0, + 639.0, + 293.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 638.0, + 1404.0, + 638.0, + 1404.0, + 670.0, + 296.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 668.0, + 1405.0, + 668.0, + 1405.0, + 702.0, + 292.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 699.0, + 1404.0, + 699.0, + 1404.0, + 731.0, + 296.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 727.0, + 991.0, + 727.0, + 991.0, + 760.0, + 293.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1849.0, + 1404.0, + 1849.0, + 1404.0, + 1885.0, + 296.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1881.0, + 1403.0, + 1881.0, + 1403.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1974.0, + 1339.0, + 1974.0, + 1339.0, + 2005.0, + 296.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1974.0, + 1404.0, + 1974.0, + 1404.0, + 2005.0, + 1394.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1407.0, + 2003.0, + 1407.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1285.0, + 1403.0, + 1285.0, + 1403.0, + 1318.0, + 296.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1315.0, + 1403.0, + 1315.0, + 1403.0, + 1348.0, + 295.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1344.0, + 1406.0, + 1344.0, + 1406.0, + 1381.0, + 294.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1377.0, + 1403.0, + 1377.0, + 1403.0, + 1410.0, + 296.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1407.0, + 1402.0, + 1407.0, + 1402.0, + 1440.0, + 295.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1436.0, + 1405.0, + 1436.0, + 1405.0, + 1473.0, + 295.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1469.0, + 1403.0, + 1469.0, + 1403.0, + 1501.0, + 296.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1496.0, + 1408.0, + 1496.0, + 1408.0, + 1534.0, + 291.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1529.0, + 487.0, + 1529.0, + 487.0, + 1562.0, + 295.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1682.0, + 1404.0, + 1682.0, + 1404.0, + 1715.0, + 296.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1714.0, + 1404.0, + 1714.0, + 1404.0, + 1747.0, + 296.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1742.0, + 1406.0, + 1742.0, + 1406.0, + 1779.0, + 294.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1772.0, + 1404.0, + 1772.0, + 1404.0, + 1809.0, + 296.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1805.0, + 826.0, + 1805.0, + 826.0, + 1838.0, + 294.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 514.0, + 1121.0, + 514.0, + 1121.0, + 554.0, + 293.0, + 554.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 338, + 1404, + 338, + 1404, + 491, + 299, + 491 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1399, + 229, + 1399, + 322, + 298, + 322 + ], + "score": 0.944 + }, + { + "category_id": 0, + "poly": [ + 299, + 538, + 488, + 538, + 488, + 571, + 299, + 571 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.886 + }, + { + "category_id": 1, + "poly": [ + 293, + 588, + 1403, + 588, + 1403, + 652, + 293, + 652 + ], + "score": 0.831 + }, + { + "category_id": 1, + "poly": [ + 300, + 1089, + 1399, + 1089, + 1399, + 1153, + 300, + 1153 + ], + "score": 0.828 + }, + { + "category_id": 1, + "poly": [ + 292, + 1396, + 1397, + 1396, + 1397, + 1430, + 292, + 1430 + ], + "score": 0.794 + }, + { + "category_id": 1, + "poly": [ + 299, + 1725, + 1398, + 1725, + 1398, + 1790, + 299, + 1790 + ], + "score": 0.793 + }, + { + "category_id": 1, + "poly": [ + 300, + 1806, + 1403, + 1806, + 1403, + 1871, + 300, + 1871 + ], + "score": 0.792 + }, + { + "category_id": 1, + "poly": [ + 298, + 1283, + 1405, + 1283, + 1405, + 1379, + 298, + 1379 + ], + "score": 0.789 + }, + { + "category_id": 1, + "poly": [ + 299, + 1448, + 1400, + 1448, + 1400, + 1511, + 299, + 1511 + ], + "score": 0.781 + }, + { + "category_id": 1, + "poly": [ + 298, + 1171, + 1405, + 1171, + 1405, + 1264, + 298, + 1264 + ], + "score": 0.78 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.773 + }, + { + "category_id": 1, + "poly": [ + 298, + 1006, + 1399, + 1006, + 1399, + 1071, + 298, + 1071 + ], + "score": 0.765 + }, + { + "category_id": 1, + "poly": [ + 298, + 1889, + 1209, + 1889, + 1209, + 1923, + 298, + 1923 + ], + "score": 0.737 + }, + { + "category_id": 1, + "poly": [ + 298, + 1530, + 1398, + 1530, + 1398, + 1594, + 298, + 1594 + ], + "score": 0.732 + }, + { + "category_id": 1, + "poly": [ + 299, + 894, + 1400, + 894, + 1400, + 989, + 299, + 989 + ], + "score": 0.722 + }, + { + "category_id": 1, + "poly": [ + 297, + 1612, + 1404, + 1612, + 1404, + 1706, + 297, + 1706 + ], + "score": 0.69 + }, + { + "category_id": 1, + "poly": [ + 300, + 1942, + 1403, + 1942, + 1403, + 2034, + 300, + 2034 + ], + "score": 0.675 + }, + { + "category_id": 1, + "poly": [ + 303, + 781, + 1399, + 781, + 1399, + 877, + 303, + 877 + ], + "score": 0.647 + }, + { + "category_id": 1, + "poly": [ + 304, + 668, + 1402, + 668, + 1402, + 764, + 304, + 764 + ], + "score": 0.623 + }, + { + "category_id": 13, + "poly": [ + 329, + 230, + 403, + 230, + 403, + 260, + 329, + 260 + ], + "score": 0.87, + "latex": "1 . 2 7 \\%" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 538.0, + 490.0, + 538.0, + 490.0, + 575.0, + 297.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 335.0, + 1406.0, + 335.0, + 1406.0, + 374.0, + 295.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 366.0, + 1404.0, + 366.0, + 1404.0, + 404.0, + 294.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 400.0, + 1405.0, + 400.0, + 1405.0, + 433.0, + 294.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 424.0, + 1405.0, + 424.0, + 1405.0, + 469.0, + 291.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 460.0, + 381.0, + 460.0, + 381.0, + 492.0, + 291.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 227.0, + 328.0, + 227.0, + 328.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 227.0, + 1403.0, + 227.0, + 1403.0, + 266.0, + 404.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1401.0, + 261.0, + 1401.0, + 295.0, + 294.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 295.0, + 374.0, + 295.0, + 374.0, + 328.0, + 291.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 588.0, + 1404.0, + 588.0, + 1404.0, + 624.0, + 296.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 619.0, + 989.0, + 619.0, + 989.0, + 652.0, + 322.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1089.0, + 1404.0, + 1089.0, + 1404.0, + 1125.0, + 297.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1122.0, + 707.0, + 1122.0, + 707.0, + 1153.0, + 320.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1395.0, + 1401.0, + 1395.0, + 1401.0, + 1434.0, + 296.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1724.0, + 1403.0, + 1724.0, + 1403.0, + 1763.0, + 295.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1756.0, + 1012.0, + 1756.0, + 1012.0, + 1790.0, + 321.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1806.0, + 1405.0, + 1806.0, + 1405.0, + 1844.0, + 295.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1837.0, + 1278.0, + 1837.0, + 1278.0, + 1872.0, + 321.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1281.0, + 1405.0, + 1281.0, + 1405.0, + 1322.0, + 292.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1314.0, + 1405.0, + 1314.0, + 1405.0, + 1351.0, + 322.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1346.0, + 1312.0, + 1346.0, + 1312.0, + 1382.0, + 321.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1446.0, + 1404.0, + 1446.0, + 1404.0, + 1485.0, + 294.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1478.0, + 637.0, + 1478.0, + 637.0, + 1511.0, + 323.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1169.0, + 1403.0, + 1169.0, + 1403.0, + 1209.0, + 293.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1202.0, + 1403.0, + 1202.0, + 1403.0, + 1238.0, + 320.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1228.0, + 399.0, + 1228.0, + 399.0, + 1269.0, + 320.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1008.0, + 1404.0, + 1008.0, + 1404.0, + 1044.0, + 297.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1039.0, + 707.0, + 1039.0, + 707.0, + 1072.0, + 319.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1887.0, + 1205.0, + 1887.0, + 1205.0, + 1927.0, + 294.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1530.0, + 1403.0, + 1530.0, + 1403.0, + 1566.0, + 298.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1561.0, + 1160.0, + 1561.0, + 1160.0, + 1595.0, + 323.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 893.0, + 1404.0, + 893.0, + 1404.0, + 933.0, + 294.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 929.0, + 1402.0, + 929.0, + 1402.0, + 960.0, + 324.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 957.0, + 928.0, + 957.0, + 928.0, + 990.0, + 324.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1610.0, + 1404.0, + 1610.0, + 1404.0, + 1649.0, + 294.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1644.0, + 1404.0, + 1644.0, + 1404.0, + 1678.0, + 324.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1674.0, + 545.0, + 1674.0, + 545.0, + 1707.0, + 322.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1939.0, + 1404.0, + 1939.0, + 1404.0, + 1980.0, + 294.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1974.0, + 1405.0, + 1974.0, + 1405.0, + 2007.0, + 325.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2002.0, + 1318.0, + 2002.0, + 1318.0, + 2040.0, + 321.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 781.0, + 1403.0, + 781.0, + 1403.0, + 818.0, + 295.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 813.0, + 1403.0, + 813.0, + 1403.0, + 850.0, + 323.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 846.0, + 805.0, + 846.0, + 805.0, + 878.0, + 323.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 668.0, + 1406.0, + 668.0, + 1406.0, + 706.0, + 298.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 700.0, + 1407.0, + 700.0, + 1407.0, + 738.0, + 321.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 731.0, + 1327.0, + 731.0, + 1327.0, + 769.0, + 321.0, + 769.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 290, + 592, + 1401, + 592, + 1401, + 655, + 290, + 655 + ], + "score": 0.822 + }, + { + "category_id": 1, + "poly": [ + 292, + 508, + 1400, + 508, + 1400, + 572, + 292, + 572 + ], + "score": 0.809 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.777 + }, + { + "category_id": 1, + "poly": [ + 301, + 229, + 1401, + 229, + 1401, + 324, + 301, + 324 + ], + "score": 0.739 + }, + { + "category_id": 1, + "poly": [ + 293, + 426, + 1402, + 426, + 1402, + 489, + 293, + 489 + ], + "score": 0.738 + }, + { + "category_id": 1, + "poly": [ + 295, + 343, + 1399, + 343, + 1399, + 406, + 295, + 406 + ], + "score": 0.684 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 857, + 74, + 857, + 105, + 298, + 105 + ], + "score": 0.407 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 857, + 74, + 857, + 105, + 298, + 105 + ], + "score": 0.286 + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 588.0, + 1406.0, + 588.0, + 1406.0, + 630.0, + 294.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 621.0, + 612.0, + 621.0, + 612.0, + 655.0, + 321.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 508.0, + 1402.0, + 508.0, + 1402.0, + 544.0, + 296.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 540.0, + 774.0, + 540.0, + 774.0, + 572.0, + 323.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 267.0, + 297.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 262.0, + 1403.0, + 262.0, + 1403.0, + 296.0, + 320.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 292.0, + 856.0, + 292.0, + 856.0, + 325.0, + 322.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 427.0, + 1404.0, + 427.0, + 1404.0, + 463.0, + 295.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 458.0, + 842.0, + 458.0, + 842.0, + 489.0, + 324.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 341.0, + 1403.0, + 341.0, + 1403.0, + 380.0, + 295.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 375.0, + 1316.0, + 375.0, + 1316.0, + 408.0, + 321.0, + 408.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 486, + 966, + 1215, + 966, + 1215, + 1617, + 486, + 1617 + ], + "score": 0.985, + "html": "
ParametersValues
CIFAR10Total Epochs Batch Size Learning Rate Learning Rate Scale Factor95 16 0.01
Learning Rate Schedule Momentum Optimizer margin0.5 25,50,75,100 0.9 SGD
Weight Decay Weights of Losses Total Epochs Batch Size Learning Rate Learning1 gRate Scale Factor Birdsnap34 Learning Rate Schedule2.0 0.0005 N/A 195 16 0.01 0.5
" + }, + { + "category_id": 1, + "poly": [ + 299, + 295, + 1404, + 295, + 1404, + 419, + 299, + 419 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 297, + 724, + 1400, + 724, + 1400, + 788, + 297, + 788 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 294, + 646, + 1401, + 646, + 1401, + 710, + 294, + 710 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 369, + 444, + 1403, + 444, + 1403, + 622, + 369, + 622 + ], + "score": 0.927 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.909 + }, + { + "category_id": 6, + "poly": [ + 529, + 909, + 1170, + 909, + 1170, + 942, + 529, + 942 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 302, + 225, + 692, + 225, + 692, + 262, + 302, + 262 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 301, + 831, + 685, + 831, + 685, + 867, + 301, + 867 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.846 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 908.0, + 1171.0, + 908.0, + 1171.0, + 943.0, + 528.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 223.0, + 695.0, + 223.0, + 695.0, + 267.0, + 294.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 826.0, + 689.0, + 826.0, + 689.0, + 871.0, + 291.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 296.0, + 1404.0, + 296.0, + 1404.0, + 329.0, + 297.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 325.0, + 1404.0, + 325.0, + 1404.0, + 361.0, + 294.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 355.0, + 1405.0, + 355.0, + 1405.0, + 394.0, + 292.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 384.0, + 445.0, + 384.0, + 445.0, + 423.0, + 294.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 723.0, + 1404.0, + 723.0, + 1404.0, + 760.0, + 293.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 756.0, + 1213.0, + 756.0, + 1213.0, + 788.0, + 296.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 643.0, + 1404.0, + 643.0, + 1404.0, + 682.0, + 293.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 677.0, + 594.0, + 677.0, + 594.0, + 713.0, + 294.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 447.0, + 797.0, + 447.0, + 797.0, + 478.0, + 370.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 485.0, + 1403.0, + 485.0, + 1403.0, + 523.0, + 366.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 528.0, + 1405.0, + 528.0, + 1405.0, + 563.0, + 367.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 559.0, + 1405.0, + 559.0, + 1405.0, + 594.0, + 393.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 589.0, + 605.0, + 589.0, + 605.0, + 624.0, + 396.0, + 624.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 485, + 1284, + 1214, + 1284, + 1214, + 1936, + 485, + 1936 + ], + "score": 0.984, + "html": "
Parameters Total EpochsValues
CIFAR10Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay0.5 25,50,75,100 0.9 SGD 2.0
Weights of Losses Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of Losses0.0005 0.2, 0.8 195 128 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.0005 0.2, 0.8
" + }, + { + "category_id": 5, + "poly": [ + 486, + 373, + 1214, + 373, + 1214, + 1024, + 486, + 1024 + ], + "score": 0.983, + "html": "
ParametersValues
CIFAR10Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay Weights of Losses0.5 25,50,75,100 0.9 SGD 2.0 0.0005
Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of LossesN/A 195 1024 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.001 N/A
" + }, + { + "category_id": 6, + "poly": [ + 548, + 317, + 1152, + 317, + 1152, + 350, + 548, + 350 + ], + "score": 0.927 + }, + { + "category_id": 6, + "poly": [ + 549, + 1228, + 1151, + 1228, + 1151, + 1261, + 549, + 1261 + ], + "score": 0.926 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 857, + 75, + 857, + 105, + 297, + 105 + ], + "score": 0.914 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.862 + }, + { + "category_id": 15, + "poly": [ + 546.0, + 316.0, + 1151.0, + 316.0, + 1151.0, + 352.0, + 546.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 1227.0, + 1152.0, + 1227.0, + 1152.0, + 1263.0, + 547.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 485, + 1284, + 1214, + 1284, + 1214, + 1936, + 485, + 1936 + ], + "score": 0.984, + "html": "
Parameters Total EpochsValues
CIFAR10Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay0.5 25,50,75,100 0.9 SGD 2.0
Weights of Losses Total Epochs Batch Size Learning Rate Learning Rate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of Losses0.0005 0.4, 0.2, 0.4 195 128 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.0005 0.4, 0.2, 0.4
" + }, + { + "category_id": 5, + "poly": [ + 486, + 373, + 1215, + 373, + 1215, + 1024, + 486, + 1024 + ], + "score": 0.983, + "html": "
Parameters Total EpochsValues
CIFAR10Batch Size Learning Rate Learning Rate Scale Factor Learning Rate Schedule95 16 0.01
Momentum Optimizer margin Weight Decay Weights of Losses0.5 25,50, 75,100 0.9 SGD 2.0 0.0005
Total Epochs Batch Size Learning Rate Learning gRate Scale Factor Birdsnap34 Learning Rate Schedule Momentum Optimizer margin Weight Decay Weights of Losses0.4, 0.2, 0.4 195 128 0.01 0.5 50,100,150,200 0.9 SGD 2.0 0.0005 0.4, 0.2, 0.4
" + }, + { + "category_id": 6, + "poly": [ + 548, + 317, + 1152, + 317, + 1152, + 350, + 548, + 350 + ], + "score": 0.927 + }, + { + "category_id": 6, + "poly": [ + 540, + 1228, + 1159, + 1228, + 1159, + 1261, + 540, + 1261 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 857, + 75, + 857, + 105, + 297, + 105 + ], + "score": 0.914 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.861 + }, + { + "category_id": 15, + "poly": [ + 547.0, + 316.0, + 1151.0, + 316.0, + 1151.0, + 352.0, + 547.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 1227.0, + 1160.0, + 1227.0, + 1160.0, + 1263.0, + 539.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/r17lFgZ0Z/r17lFgZ0Z.md b/parse/train/r17lFgZ0Z/r17lFgZ0Z.md new file mode 100644 index 0000000000000000000000000000000000000000..fdd23c65cd8c8e9d4ce0e4c4a066fdf4ebf5914f --- /dev/null +++ b/parse/train/r17lFgZ0Z/r17lFgZ0Z.md @@ -0,0 +1,280 @@ +# RELEVANCE OF UNSUPERVISED METRICS IN TASKORIENTED DIALOGUE FOR EVALUATING NATURAL LANGUAGE GENERATION + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +Automated metrics such as BLEU are widely used in the machine translation literature. They have also been used recently in the dialogue community for evaluating dialogue response generation. However, previous work in dialogue response generation has shown that these metrics do not correlate strongly with human judgment in the non task-oriented dialogue setting. Task-oriented dialogue responses are expressed on narrower domains and exhibit lower diversity. It is thus reasonable to think that these automated metrics would correlate well with human judgment in the task-oriented setting where the generation task consists of translating dialogue acts into a sentence. We conduct an empirical study to confirm whether this is the case. Our findings indicate that these automated metrics have stronger correlation with human judgments in the task-oriented setting compared to what has been observed in the non task-oriented setting. We also observe that these metrics correlate even better for datasets which provide multiple ground truth reference sentences. In addition, we show that some of the currently available corpora for task-oriented language generation can be solved with simple models and advocate for more challenging datasets. + +# 1 INTRODUCTION + +Rule-based and template-based dialogue response generation systems have been around for a long time (Axelrod, 2000; Elhadad, 1992). Even today, many task-oriented dialogue systems deployed in production are rule-based and template-based. These systems do not scale with increasing domain complexity and maintaining the increasing number of templates becomes cumbersome. In the past, Oh & Rudnicky (2000) proposed a corpus-based approach for Natural Language Generation (NLG) for task-oriented dialogue systems. Other statistical approaches were proposed using tree-based models and reinforcement learning (Walker et al., 2007; Rieser & Lemon, 2009). Recently, deeplearning based approaches (Wen et al., 2015b; Sharma et al., 2017; Lowe et al., 2015; Serban et al., 2016) have shown promising results for dialogue response generation. + +The automated evaluation of machine-generated language is challenging and an important problem for the natural language processing community. The most widely used automated metrics currently are word-overlap based metrics such as BLEU (Papineni et al., 2002), METEOR (Banerjee & Lavie, 2005) which were proposed originally for machine translation. While these metrics were shown to correlate well with manual human evaluation in machine translation tasks, previous studies showed that this is not the case in non-task oriented dialogue (Liu et al., 2016). This is explained by the fact that for the same context (e.g. a user utterance), responses in dialogue have more diversity. Word-overlap metrics are unable to capture semantics and thus, can lead to poor scores even for appropriate responses. Human evaluation in this case is the most reliable metric. However, human judgments are expensive to obtain and not readily available at all times. + +Task-oriented dialogue systems are employed in narrower domains (e.g. booking a restaurant) and responses do not have as much diversity as in the non-task oriented setting. Another important difference is that in the non-task oriented setting, response generation is often performed end-toend, which means that the model takes as input the last user utterance and potentially the dialogue history and it outputs the next system answer. In the task-oriented setting, on the other hand, the language generation task is often seen as a translation step from an abstract representation of a sentence to the sentence itself. As a consequence, automated metrics which compare a generated sentence to a reference sentence might be more appropriate and correlate with human judgments. In this paper, we: + +• study the correlation between human judgments and several unsupervised automated metrics on two popular task-oriented dialogue datasets, • introduce variants of existing models and evaluate their performance on these metrics + +We find that the automated metrics have stronger correlation with human judgments in the taskoriented setting than what has been observed in the non task-oriented setting. We also observe that these metrics correlate even more in the presence of multiple reference sentences. + +# 2 RELATED WORK + +Liu et al. (2016) did an empirical study to evaluate the correlation between human scores and several automated word-overlap metrics as well as embedding-based metrics for dialogue response generation. They observed that these metrics, though widely used in the literature, had only weak correlation with human judgments in the non task-oriented dialogue NLG setting. + +In terms of supervised NLG evaluation metrics, Lowe et al. (2017) proposed the ADEM model which trains a hierarchical recurrent neural network in a supervised manner to predict human-like scores. This learned score was shown to correlate better with human judgments than any other automated metric. However, the drawback of this approach is the requirement for expensive human ratings. + +Li et al. (2016) proposed to use reinforcement learning to train an end-to-end dialogue system. They simulate a dialogue between two agents and use a policy gradient algorithm with a reward function which evaluates specific properties of the responses generated by the dialogue system. + +In the adversarial setting, Kannan & Vinyals (2016) train a recurrent neural network discriminator to differentiate human-generated responses from model-generated responses. However, an extensive analysis of the viability and the ease of standardization of this approach is yet to be conducted. Li et al. (2017), apart from adversarially training dialogue response models, propose an independent adversarial evaluation metric AdverSuc and a measure of the model’s reliability called evaluator reliability error. Drawbacks of these approaches are that they are model-dependent. Adversarial methods might be promising for task-oriented dialogue systems but more research needs to be conducted on their account. + +Most of the work described so far has been done in the non task-oriented dialogue setting as there has been prior work indicating that automated metrics do not correlate well with humans in that setting. There has not yet been any empirical validation that these conclusions also apply to the task oriented setting. Research in the task oriented setting has mostly made use of automated metrics such as BLEU and human evaluation (Wen et al., 2015b; Sharma et al., 2017; Dusek & Jurcicek, ˇ 2016). + +# 3 METRICS + +This section describes the set of automatic metrics whose correlation with human evaluation is studied. We consider first word-overlap metrics and then embedding-based metrics. In all that follows, when multiple references are provided, we compute the similarity between the prediction and all the references one-by-one, and then select the maximum value. We then average the scores across the entire corpus. + +# 3.1 WORD-OVERLAP BASED METRICS + +# 3.1.1 BLEU + +The BLEU metric (Papineni et al., 2002) compares $n$ -grams between the candidate utterance and the reference utterance. The BLEU score is computed at the corpus-level and relies on the following + +modified precision: + +$$ +p _ { n } = \frac { \displaystyle \sum _ { C \in \{ C a n d i d a t e s \} } \sum _ { n - g r a m \in C } C t _ { c l i p } ( n - g r a m ) } { \displaystyle \sum _ { C ^ { \prime } \in \{ C a n d i d a t e s ^ { \prime } \} } \sum _ { n - g r a m ^ { \prime } \in C ^ { \prime } } C t _ { c l i p } ( n - g r a m ^ { \prime } ) } +$$ + +where {Candidates} are the candidate answers generated by the model and $C t _ { c l i p }$ is the clipped count for the $n$ -gram which is the number of times the $n$ -gram is common to the candidate answer and the reference answer clipped by the maximum number of occurrences of the $n$ -gram in the reference answer. The BLEU-N score is defined as: + +$$ +\begin{array} { l } { { \displaystyle \forall \mathrm { ~ s c u r e ~ 1 s ~ u c t 1 1 1 e u ~ a s . } } } \\ { { \displaystyle \mathsf { B L E U - N } = \mathsf { B P } \exp ( \sum _ { n } ^ { N } \omega _ { n } \log ( p _ { n } ) ) } } \end{array} +$$ + +where $N$ is the maximum length of the $n$ -grams (in this paper, we compute BLEU-1 to BLEU-4), $\omega$ is a weighting that is often uniform and BP is a brevity penalty. In this paper we report the BLEU score at the corpus level but we also compute this score at the sentence level to analyze its correlation with human evaluation. + +# 3.1.2 METEOR + +The METEOR metric (Banerjee & Lavie, 2005) was proposed as a metric which correlates better at the sentence level with human evaluation. To compute the METEOR score, first, an alignment between the candidate and the reference sentences is created by mapping each unigram in the candidate sentence to 0 or 1 unigram in the reference sentence. The alignment is not only based on exact matches but also stem, synonym, and paraphrase matches. Based on this alignment, unigram precision and recall are computed and the METEOR score is: + +$$ +\mathbf { M E T E O R } = F _ { m e a n } ( 1 - p ) +$$ + +where $F _ { m e a n }$ is the harmonic mean between precision and recall with the weight for recall 9 times a high as the weight for precision, and $p$ is a penalty. + +# 3.1.3 ROUGE + +ROUGE (Lin, 2004) is a set of metrics that was first introduced for summarization. We compute ROUGE-L which is an F-measure based on the Longest Common Subsequence (LCS) between the candidate and reference utterances. + +# 3.2 EMBEDDING BASED METRICS + +We consider another set of metrics which compute the cosine similarity between the embeddings of the predicted and the reference sentence instead of relying on word overlaps. + +# 3.2.1 SKIP-THOUGHT + +The Skip-Thought model (Kiros et al., 2015) is trained in an unsupervised fashion and uses a recurrent network to encode a given sentence into an embedding and then decode it to predict the preceding and following sentences. The model was trained on the BookCorpus dataset (Zhu et al., 2015). The embeddings produced by the encoder have a robust performance on semantic relatedness tasks. We use the pre-trained Skip-Thought encoder provided by the authors1. + +We also compute other embedding-based methods which have been used as evaluation metrics for measuring human correlation in recent literature (Liu et al., 2016) for non task-oriented dialogue in Sections 3.2.2, 3.2.3, and 3.2.4. + +# 3.2.2 EMBEDDING AVERAGE + +This metric computes a sentence-level embedding by averaging the embeddings of the words composing this sentence: + +$$ +\bar { e } _ { C } = \frac { \sum _ { w \in C } e _ { w } } { \lvert \sum _ { w ^ { \prime } \in C } e _ { w ^ { \prime } } \rvert } . +$$ + +In this equation, the vectors $e _ { w }$ are embeddings for the words $w$ in the candidate sentence $C$ . + +# 3.2.3 VECTOR EXTREMA + +Vector extrema (Forgues et al., 2014) computes a sentence-level embedding by taking the most extreme value of the embeddings of the words composing the sentence for each dimension of the + +$$ +e _ { r d } = \left\{ \begin{array} { l l } { \operatorname* { m a x } _ { w \in C } e _ { w d } } & { \mathrm { i f } e _ { w d } > \vert \operatorname* { m i n } _ { w ^ { \prime } \in C } e _ { w ^ { \prime } d } \vert } \\ { \operatorname* { m i n } _ { w \in C } e _ { w d } } & { \mathrm { o t h e r w i s e } \mathrm { . } } \end{array} \right. +$$ + +In this equation, $d$ is an index over the dimensions of the embedding and $C$ is the candidate sentence. + +# 3.2.4 GREEDY MATCHING + +Greedy matching does not compute a sentence embedding but directly a similarity score between a candidate $C$ and a reference $r$ (Rus & Lintean, 2012). This similarity score is computed as follows: + +$$ +\begin{array} { c } { \displaystyle { G ( C , r ) = \frac { \sum w \in C \operatorname* { m a x } _ { \hat { w } \in r } c o s _ { - } s i m ( e _ { w } , w _ { \hat { w } } ) } { | C | } } } \\ { \displaystyle { G { M ( C , r ) } = \frac { G ( C , r ) + G ( r , C ) } { 2 } . } } \end{array} +$$ + +Each word in the candidate sentence is greedily matched to a word in the reference sentence based on their embeddings’ cosine similarity. The score is an average of these similarities over the number of words in the candidate sentence. The same score is computed by reversing the roles of the candidate and reference sentences and the average of the two scores gives the final similarity score. + +# 4 RESPONSE GENERATION MODELS + +This section presents the different natural language generation models that we use in this study. All of these models take as input a set of dialogue acts (Austin, 1962) potentially with slot types and slot values and translate this input into an utterance. An example input is inform(food $=$ Chinese) and a corresponding output would be “I am looking for a Chinese restaurant.”. In this example, the dialogue act is inform, the slot type is food, and the slot value is Chinese. + +# 4.1 RANDOM + +Given a dialogue act with one or more slot types, the random model finds all the examples in the training set with the same dialogue act and slots (while ignoring slot values) and it randomly selects its output from this set of reference sentences. The datasets that we experiment on have some special slot values such as “yes”, “no”, and “don’t care”. Since the model ignores all slot values, these special cases are not properly handled, which results in slightly lower performance than what we could get by spending more time hand-engineering the model’s behavior for these values. + +# 4.2 LSTM + +This model consists of a recurrent LSTM (Hochreiter & Schmidhuber, 1997) decoder. The dialogue acts and slot types are first encoded as a binary vector whose length is the number of possible combinations of dialogue acts and slot types in the dataset. We refer to this binary vector as the Dialogue Act $( D A )$ vector. The $D A$ vector for a given set of dialogue acts is a binary vector over the fused dialogue act-slot types, e.g., INFORM-FOOD, INFORM-COUNT, etc. + +This binary vector is given as input to the decoder at each time-step of the LSTM. The decoder then outputs a delexicalized sentence. A delexicalized sentence contains placeholders for the slot values. An example is “I am looking for a FOOD restaurant.”. The values for the delexicalized slots (the type of food in this example) are then directly copied from the input. + +# 4.3 DELEX-SC-LSTM + +This model uses the same architecture as the LSTM model presented in the previous section except that it uses sc-LSTM (Wen et al., 2015b) units in the decoder instead of LSTM units. We call this + +model the “delex-sc-LSTM”2. As in the previous model, the input $D A$ vector only encodes acts and delexicalized slots. It does not contain any information about the slot value. + +By providing this model the same $D A$ vector input as the one given to the LSTM model, we can directly study if the additional complexity of the sc-LSTM unit’s reading gate provides significant improvement over the small-sized task-oriented dialogue datasets which are currently available. + +# 4.4 HIERARCHICAL-LEX-DELEX-SC-LSTM + +This model is a variant of the “ld-sc-LSTM” model proposed by Sharma et al. (2017) which is based on an encoder-decoder framework. We call our model “hierarchical-lex-delex-sc-LSTM”3. + +We present the encoder in Figure 1. The encoder consists of a hierarchical LSTM with $N _ { e }$ time-steps, where $N _ { e }$ is the number of non-zero entries in the $D A$ vector. Each time-step of the encoder encodes one dialogue act’s delexicalized and lexicalized slot-value pair (e.g. (INFORM-FOOD, ‘Chinese’)). The delexicalized act-slot part is encoded as a one-hot vector which we refer to as $D A _ { t }$ . $D A _ { t }$ is constructed by masking all except the $t ^ { \mathrm { { t h } } }$ dialogue act in the DA vector4. The lexicalized value part is en + +![](images/17acec8a69179f00d7d76eae7dbbeb4746f29c87169dd83f9d9aae7969b3e28f.jpg) +Figure 1: Encoder of the hld-scLSTM model + +coded by an LSTM encoder which shares parameters across all time-steps and operates over the word-embeddings of the lexicalized values $v _ { t , i }$ . Our model differs from the “ld-sc-LSTM” model in that we use an LSTM encoder over the word-embeddings instead of computing the mean of the word-embeddings. The final hidden state of this LSTM is concatenated with $D A _ { t }$ and is given as input to the upper LSTM (see Figure 1). The final hidden state of the upper LSTM is then provided to the decoder as input. This is another difference from the “ld-sc-LSTM” as that uses the mean of all the hidden states of the encoder instead, which, in our experiments, did not perform as well as using just the final hidden state $E$ . + +The decoder is described in Figure 2. It is the same as in the “ld-sc-LSTM” model. At each time-step, it takes as input the encoder output $E$ , the $D A$ vector, and the word-embedding of the word generated at the previous time-step. The $D A$ vector is also additionally provided to the scLSTM cell in order for it to be regulated by its reading gate as described in Wen et al. (2015b). + +![](images/114dbf8875d6790d50ee41019e8384e6fe26efeea5c18f9566f2fbec598cb400.jpg) +Figure 2: Decoder of the hld-scLSTM model + +# 5 EXPERIMENTS + +# 5.1 DECODING + +During training, at each time-step, we use the ground truth word from the previous time-step. The model thus learns to generate the next word given the previous one. On the other hand, to generate sentences during test time, we use beam search. The first word input to the generator is a special token $< b o s >$ which indicates the beginning of the sequence. Decoding is stopped if we reach a specified maximum number of time-steps or if the model outputs a special token $< e o s >$ which indicates the end of the sequence. We also use a slot error rate penalty, similarly to Wen et al. (2015b), to re-rank the sentences generated with beam search. We use this method for all three of the LSTM, d-scLSTM, and hld-scLSTM models for fairness. + +Table 1: Performance comparison across models on word-overlap based automated metrics + +
DSTC2Restaurants
B-1B-2B-3B-4MRLB-1B-2B-3B-4MRL
Gold1.001.001.001.001.001.001.001.001.001.001.001.00
Random0.8750.8430.8220.8070.5640.8520.8720.8130.7650.7210.5040.796
LSTM0.9000.8790.8630.8510.6100.8880.9820.9660.9490.9320.6520.944
d-scLSTM0.8800.8500.8280.8120.5780.8740.9800.9640.9480.9310.6540.945
hld-scLSTM0.9090.8900.8780.8700.6240.8990.9850.9780.9700.9620.7040.965
+ +Table 2: Performance comparison across models on sentence-embedding based automated metrics + +
DSTC2Restaurants
SkipThoughtEmbeddingAverageVectorExtremaGreedyMatchingSkipThoughtEmbeddingAverageVectorExtremaGreedyMatching
Gold1.001.001.001.001.001.001.001.00
Random0.9060.9810.9100.9470.8430.9570.9050.930
LSTM0.9460.9850.9350.9620.9450.9970.9860.991
d-scLSTM0.9250.9840.9260.9570.9480.9970.9860.991
hld-scLSTM0.9320.9870.9420.9640.9680.9970.9890.993
+ +Similarly to the LSTM model, the d-scLSTM and hld-scLSTM generate delexicalized sentences, i.e., they generate slot tokens instead of slot values directly. These slot tokens are replaced with slot values in a post-processing step which is a fairly common step in task-oriented dialogue NLG literature. + +# 5.2 EVALUATION + +In NLG tasks, improvements in automated metric scores are most commonly used to demonstrate improvement in the generation task. However, these metrics have been shown to only weakly correlate with human evaluation in the non task-oriented dialogue setting (Liu et al., 2016) and hence are not considered reliable measures of improvement. Human evaluation is considered the metric of choice, but human ratings are expensive to obtain. The ease of computing these automated metrics and their availability for rapid prototyping has lead to their widespread adoption. + +We evaluate the models described in the previous section on the DSTC2 (Henderson et al., 2014) and the Restaurants datasets (Wen et al., 2015a) using these automated metrics. These datasets are some of the only available resources for studying NLG for task-oriented dialogue. The DSTC2 dataset contains dialogues between human users and a dialogue system in a restaurant domain. The dataset is annotated with dialogue acts, slot type, and slot values. The NLG component of the dialogue system used for data collection is templated. The Restaurants dataset was specifically proposed for NLG and provides, for a set of dialogue acts with slot types and slot values, two sentences generated by humans. + +We present the results of our experiments in Table 1 and Table 2. The scores of all the models on these automated metrics are very high. This indicates that there is significant word overlap between the generated and the reference sentences and that the NLG task on these datasets can be solved with a simple model such as the LSTM model. In effect, Table 1 shows that the LSTM model performs comparably to the d-scLSTM model based on the word-overlap metrics. This can be explained by the fact that the d-scLSTM model has more parameters and might suffer from overfitting issues on these relatively small datasets. + +The hld-scLSTM is considered to consistently outperform the other models based on the wordoverlap metrics. As explained by Sharma et al. (2017), this improvement results from the model’s access to the lexicalized slot values, due to which it can take into account the grammatical associations of the generated words near the output tokens, thereby generating higher quality sentences. However, Table 2 shows that sentence-embedding based metrics judge all the models except the random one to perform quite similarly with again, very high performance scores. + +In the next section, we add human evaluation for these models on these datasets. + +# 5.3 HUMAN RATING COLLECTION + +We randomly selected 20 dialogue acts from the test set of each dataset. For each of these contexts, we presented 5 sentences to the evaluators: the gold response provided in the test set and the responses generated by the four models described in Section 4. These sentences were randomly shuffled and not presented in the same order. We invited 18 human users to score each of these 100 sentences on a Likert-type scale of 1 to 5. The users were asked to rate the responses depending on how appropriate they were for the specified dialogue acts. + +![](images/18289be4190cbbf67f923162b61f526162ab7aca3e0d7abbab61623c5036e6d9.jpg) +Figure 3: Scatter plots for correlation of some automated metrics with human evaluation for (a) the DSTC2 dataset, and (b) the Restaurants dataset. Random gaussian noise $\mathcal { N } ( 0 , 0 . 1 )$ has been added to data points along the human score axis and $\mathcal { N } ( 0 , 0 . 0 2 )$ has been added to the automated metric score’s axis to aid visualization of overlapping data points. Transparency has been added for the same effect. + +A score of 1 was the lowest score, meaning that the response was not appropriate at all whereas a score of 5 meant that the sentence was highly appropriate. + +Table 3: Pairwise Cohen’s kappa scores for the 11 human users + +
K# pairs% pairs
>0.155/55100.0 %
>0.240/5572.7 %
>0.328/5550.9 %
>0.419/5534.5 %
>0.58/5514.5 %
>0.60/550.0 %
+ +We computed Cohen’s kappa scores (Cohen, 1960) between the human users in pairs of two. We removed 7 users who had kappa scores less than 0.1 and used the remaining 11 users for the correlation study. The kappa scores are presented in Table 3. Most of the user pairs have a Cohen’s $\kappa > 0 . 3$ which indicates fair agreement between users (Viera et al., 2005). + +# 5.4 CORRELATION BETWEEN AUTOMATED METRICS AND HUMAN SCORES + +We present the correlation between the automated metrics and our collected human ratings in Table 4. We measure human v.s. human correlation by randomly splitting the human users into two groups. The results indicate that in most cases, human scores correlate the best with other human scores. Except in the case of the Spearman correlation for BLEU-N scores, we can see that there is a positive correlation between the automated metrics and the human scores for these task-oriented datasets, which contrasts with the non task-oriented dialogue setting where Liu et al. (2016) observed no strong correlation trends. + +A likely explanation for the negative Spearman correlation values for BLEU-N is that there is only one gold reference per context in the DSTC2 dataset. The Restaurants dataset, on the other hand, provides two gold references per context. Having multiple gold references increases the likelihood that the generated response will have significant word-overlap with one of the reference responses. + +We present scatter plots for some of the metrics presented in Table 4 in Figure 3. We observe that all the metrics correlate very well with humans on high scoring examples. As it can be seen in the scatter plots, most of the sentences are given the maximal score of 5 by the human evaluators. This confirms our previous observation that the available corpora for task-oriented dialogue NLG task are not very challenging and a simple LSTM-based model can output high-quality responses. + +Table 4: Correlation of automated metrics with human evaluations scores + +
MetricDSTC2Restaurants
Spearmanp-valuePearsonp-valueSpearmanp-valuePearsonp-value
Bleu1Bleu 2-0.317<0.0050.583<0.0050.0690.4940.2770.005
-0.318<0.0050.526<0.0050.0910.3660.1660.099
Bleu 3-0.318<0.0050.500<0.0050.1090.2800.2230.026
Bleu 4METEORROUGELSkip Thought-0.318<0.0050.461<0.0050.1050.2960.2550.010
0.295<0.0050.582<0.0050.353<0.0050.489<0.005
0.294<0.0050.448<0.0050.346<0.0050.382<0.005
0.528<0.0050.0860.3970.284<0.0050.364<0.005
Embedding Average0.295<0.0050.485<0.0050.423<0.0050.2600.009
Vector ExtremaGreedy Matching0.299<0.0050.624<0.0050.446<0.0050.287<0.005
0.295<0.0050.572<0.0050.446<0.0050.325<0.005
Human0.810<0.0050.984<0.0050.653<0.0050.857<0.005
+ +Overall, among the word overlap based automated metrics, METEOR consistently correlates with human evaluation on both datasets. These results confirm the original findings by Banerjee & Lavie (2005) who showed that METEOR had good correlation with human evaluation in the machine translation task. The comparison with machine translation is highly relevant in the task-oriented setting because the NLG model essentially learns to translate the abstract representation of a sentence into a sentence. It is a translation task contrary to the non task-oriented setting where the NLG model needs to decide and output a new sentence based on the last sentence typed by a user and dialogue history. Therefore, automated metrics coming from the machine translation literature are more adequate in our case than in the non-task oriented case as shown by Liu et al. (2016). + +It is interesting to see that METEOR correlates well with human evaluation consistently. This can be explained by the fact that even though METEOR does not rely on word embeddings, it includes notions of synonymy and paraphrasing when computing the alignment between the candidate and reference utterances. + +# 6 DISCUSSION + +We evaluated several natural language generation models trained on the DSTC2 and the Restaurants datasets based on several automated metrics. We also performed human evaluation on the model-generated responses and our study shows that human evaluation is a much more reliable metric compared to the others. Among the word-overlap based automated metrics, we found that the METEOR score correlates the most with human judgments and we suggest using METEOR for taskoriented dialogue natural language generation instead of BLEU. We also observe that these metrics are more reliable in the task-oriented dialogue setting compared to the general, non task-oriented one due to the limited possible diversity in the task-oriented setting. Also, as observed by Galley et al. (2015), we can see that word-overlap based metrics correlate better with human evaluation when multiple references are provided, as in the Restaurants dataset. Otherwise, as in the case of DSTC2 which only provides one reference sentence per example, we observe that all the BLEU-N metrics negatively correlate with human evaluation on Spearman correlation. + +As has been observed in the machine translation literature, using beam search improves the quality of generated sentences significantly compared to stochastic sampling. For similar models, our results show improvement in the automated metrics’ scores compared to Wen et al. (2015b) who used stochastic sampling for decoding instead of beam search. + +Wen et al. (2015b) did not use the slot error rate penalty with the vanilla LSTM model in their experiments. After adding the penalty in our case, we observe that the vanilla LSTM-based model performs as well as the delexicalized semantically-controlled LSTM model. This suggests that the added complexity introduced by the sc-LSTM unit does not offer a significant advantage for these two datasets. + +High performance on automated metrics, achieved by our models on the DSTC2 and the Restaurants datasets lead us to conclude that these datasets are not very challenging for the NLG task. The taskoriented dialogue community should move towards using larger and more complex datasets, which have been recently announced, such as the Frames dataset (El Asri et al., 2017) or the E2E NLG Challenge dataset (Novikova et al., 2016). + +# REFERENCES + +John Langshaw Austin. How to do things with words. Oxford University Press, 1962. + +Scott Axelrod. Natural language generation in the ibm flight information system. In Proceedings of the 2000 ANLP/NAACL Workshop on Conversational Systems - Volume 3, ANLP/NAACLConvSyst ’00, pp. 21–26, Stroudsburg, PA, USA, 2000. Association for Computational Linguistics. doi: 10.3115/1117562.1117567. + +Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proc. ACL workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, 2005. + +Jacob Cohen. A coefficient of agreement for nominal scales. Educational and psychological measurement, 20(1):37–46, 1960. + +Ondˇrej Dusek and Filip Jurcicek. Sequence-to-sequence generation for spoken dialogue via deep ˇ syntax trees and strings. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pp. 45–51, Berlin, Germany, August 2016. Association for Computational Linguistics. + +Layla El Asri, Hannes Schulz, Shikhar Sharma, Jeremie Zumer, Justin Harris, Emery Fine, Rahul Mehrotra, and Kaheer Suleman. Frames: A corpus for adding memory to goal-oriented dialogue systems. arXiv preprint arXiv:1704.00057, 2017. + +Michael Elhadad. Generating coherent argumentative paragraphs. In In Proceedings of COLING ’92, volume II, pp. 638–644, 1992. + +Gabriel Forgues, Joelle Pineau, Jean-Marie Larcheveque, and Real Tremblay. Bootstrapping dialog systems with word embeddings. 2014. + +Michel Galley, Chris Brockett, Alessandro Sordoni, Yangfeng Ji, Michael Auli, Chris Quirk, Margaret Mitchell, Jianfeng Gao, and Bill Dolan. deltableu: A discriminative metric for generation tasks with intrinsically diverse targets. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pp. 445–450, Beijing, China, July 2015. Association for Computational Linguistics. + +Matthew Henderson, Blaise Thomson, and Jason D. Williams. The second dialog state tracking challenge. In Proceedings of the 15th Annual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL), pp. 263–272, Philadelphia, PA, U.S.A., June 2014. Association for Computational Linguistics. + +Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8): 1735–1780, 1997. + +Anjuli Kannan and Oriol Vinyals. Adversarial evaluation of dialogue models. In NIPS 2016 Workshop on Adversarial Training, 2016. + +Ryan Kiros, Yukun Zhu, Ruslan R Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Skip-thought vectors. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett (eds.), Advances in Neural Information Processing Systems 28, pp. 3294–3302. Curran Associates, Inc., 2015. + +Jiwei Li, Will Monroe, Alan Ritter, Dan Jurafsky, Michel Galley, and Jianfeng Gao. Deep reinforcement learning for dialogue generation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pp. 1192–1202, Austin, Texas, November 2016. Association for Computational Linguistics. + +Jiwei Li, Will Monroe, Tianlin Shi, Alan Ritter, and Dan Jurafsky. Adversarial learning for neural dialogue generation. CoRR, abs/1701.06547, 2017. + +Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Proc. ACL workshop on Text Summarization Branches Out, 2004. + +Chia-Wei Liu, Ryan Lowe, Iulian Serban, Mike Noseworthy, Laurent Charlin, and Joelle Pineau. How not to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics for dialogue response generation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pp. 2122–2132. Association for Computational Linguistics, 2016. + +Ryan Lowe, Nissan Pow, IV Serban, Laurent Charlin, and Joelle Pineau. Incorporating unstructured textual knowledge sources into neural dialogue systems. In Neural Information Processing Systems Workshop on Machine Learning for Spoken Language Understanding, 2015. + +Ryan Lowe, Michael Noseworthy, Iulian V. Serban, Nicolas Angelard-Gontier, Yoshua Bengio, and Joelle Pineau. Towards an automatic turing test: Learning to evaluate dialogue responses. In Proceedings of the 5th International Conference on Learning Representations (ICLR) Workshop, Toulon, France, April 2017. + +Jekaterina Novikova, Oliver Lemon, and Verena Rieser. Crowd-sourcing nlg data: Pictures elicit better data. In Proceedings of the 9th International Natural Language Generation conference, pp. 265–273, Edinburgh, UK, September 5-8 2016. Association for Computational Linguistics. + +Alice H. Oh and Alexander I. Rudnicky. Stochastic language generation for spoken dialogue systems. In Proceedings of the 2000 ANLP/NAACL Workshop on Conversational Systems - Volume 3, ANLP/NAACL-ConvSyst ’00, pp. 27–32, Stroudsburg, PA, USA, 2000. Association for Computational Linguistics. doi: 10.3115/1117562.1117568. + +Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proc. of ACL, 2002. + +Verena Rieser and Oliver Lemon. Natural language generation as planning under uncertainty for spoken dialogue systems. In Proceedings of the 12th Conference of the European Chapter of the ACL (EACL 2009), pp. 683–691, Athens, Greece, March 2009. Association for Computational Linguistics. + +Vasile Rus and Mihai C. Lintean. A comparison of greedy and optimal assessment of natural language student input using word-to-word similarity metrics. In Proceedings of the Seventh Workshop on Building Educational Applications Using NLP, 2012. + +Iulian Vlad Serban, Alessandro Sordoni, Ryan Lowe, Laurent Charlin, Joelle Pineau, Aaron C. Courville, and Yoshua Bengio. A hierarchical latent variable encoder-decoder model for generating dialogues. CoRR, abs/1605.06069, 2016. + +Shikhar Sharma, Jing He, Kaheer Suleman, Hannes Schulz, and Philip Bachman. Natural language generation in dialogue using lexicalized and delexicalized data. In Proceedings of the 5th International Conference on Learning Representations (ICLR) Workshop, 2017. + +Anthony J Viera, Joanne M Garrett, et al. Understanding interobserver agreement: the kappa statistic. Family Medicine, 37(5):360–363, 2005. + +Marilyn Walker, Amanda Stent, Franc¸ois Mairesse, and Rashmi Prasad. Individual and domain adaptation in sentence planning for dialogue. J. Artif. Int. Res., 30(1):413–456, November 2007. ISSN 1076-9757. + +Tsung-Hsien Wen, Milica Gasiˇ c, Dongho Kim, Nikola Mrk ´ siˇ c, Pei-Hao Su, David Vandyke, and ´ Steve Young. Stochastic Language Generation in Dialogue using Recurrent Neural Networks with Convolutional Sentence Reranking. In Proceedings of the 16th Annual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL). Association for Computational Linguistics, September 2015a. + +Tsung-Hsien Wen, Milica Gasiˇ c, Nikola Mrk ´ siˇ c, Pei-Hao Su, David Vandyke, and Steve Young. ´ Semantically conditioned LSTM-based natural language generation for spoken dialogue systems. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, September 2015b. + +Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In The IEEE International Conference on Computer Vision (ICCV), December 2015. \ No newline at end of file diff --git a/parse/train/r17lFgZ0Z/r17lFgZ0Z_content_list.json b/parse/train/r17lFgZ0Z/r17lFgZ0Z_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8ec57515a4ff229e08d79cf3c66dbdb57b4fdab9 --- /dev/null +++ b/parse/train/r17lFgZ0Z/r17lFgZ0Z_content_list.json @@ -0,0 +1,1524 @@ +[ + { + "type": "text", + "text": "RELEVANCE OF UNSUPERVISED METRICS IN TASKORIENTED DIALOGUE FOR EVALUATING NATURAL LANGUAGE GENERATION ", + "text_level": 1, + "bbox": [ + 174, + 98, + 823, + 171 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 195, + 398, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 260, + 544, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Automated metrics such as BLEU are widely used in the machine translation literature. They have also been used recently in the dialogue community for evaluating dialogue response generation. However, previous work in dialogue response generation has shown that these metrics do not correlate strongly with human judgment in the non task-oriented dialogue setting. Task-oriented dialogue responses are expressed on narrower domains and exhibit lower diversity. It is thus reasonable to think that these automated metrics would correlate well with human judgment in the task-oriented setting where the generation task consists of translating dialogue acts into a sentence. We conduct an empirical study to confirm whether this is the case. Our findings indicate that these automated metrics have stronger correlation with human judgments in the task-oriented setting compared to what has been observed in the non task-oriented setting. We also observe that these metrics correlate even better for datasets which provide multiple ground truth reference sentences. In addition, we show that some of the currently available corpora for task-oriented language generation can be solved with simple models and advocate for more challenging datasets. ", + "bbox": [ + 233, + 291, + 764, + 515 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 542, + 334, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rule-based and template-based dialogue response generation systems have been around for a long time (Axelrod, 2000; Elhadad, 1992). Even today, many task-oriented dialogue systems deployed in production are rule-based and template-based. These systems do not scale with increasing domain complexity and maintaining the increasing number of templates becomes cumbersome. In the past, Oh & Rudnicky (2000) proposed a corpus-based approach for Natural Language Generation (NLG) for task-oriented dialogue systems. Other statistical approaches were proposed using tree-based models and reinforcement learning (Walker et al., 2007; Rieser & Lemon, 2009). Recently, deeplearning based approaches (Wen et al., 2015b; Sharma et al., 2017; Lowe et al., 2015; Serban et al., 2016) have shown promising results for dialogue response generation. ", + "bbox": [ + 174, + 575, + 825, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The automated evaluation of machine-generated language is challenging and an important problem for the natural language processing community. The most widely used automated metrics currently are word-overlap based metrics such as BLEU (Papineni et al., 2002), METEOR (Banerjee & Lavie, 2005) which were proposed originally for machine translation. While these metrics were shown to correlate well with manual human evaluation in machine translation tasks, previous studies showed that this is not the case in non-task oriented dialogue (Liu et al., 2016). This is explained by the fact that for the same context (e.g. a user utterance), responses in dialogue have more diversity. Word-overlap metrics are unable to capture semantics and thus, can lead to poor scores even for appropriate responses. Human evaluation in this case is the most reliable metric. However, human judgments are expensive to obtain and not readily available at all times. ", + "bbox": [ + 174, + 708, + 825, + 847 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Task-oriented dialogue systems are employed in narrower domains (e.g. booking a restaurant) and responses do not have as much diversity as in the non-task oriented setting. Another important difference is that in the non-task oriented setting, response generation is often performed end-toend, which means that the model takes as input the last user utterance and potentially the dialogue history and it outputs the next system answer. In the task-oriented setting, on the other hand, the language generation task is often seen as a translation step from an abstract representation of a sentence to the sentence itself. As a consequence, automated metrics which compare a generated sentence to a reference sentence might be more appropriate and correlate with human judgments. In this paper, we: ", + "bbox": [ + 174, + 854, + 823, + 924 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 103, + 825, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "• study the correlation between human judgments and several unsupervised automated metrics on two popular task-oriented dialogue datasets, • introduce variants of existing models and evaluate their performance on these metrics ", + "bbox": [ + 215, + 169, + 823, + 217 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We find that the automated metrics have stronger correlation with human judgments in the taskoriented setting than what has been observed in the non task-oriented setting. We also observe that these metrics correlate even more in the presence of multiple reference sentences. ", + "bbox": [ + 176, + 227, + 821, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 287, + 344, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Liu et al. (2016) did an empirical study to evaluate the correlation between human scores and several automated word-overlap metrics as well as embedding-based metrics for dialogue response generation. They observed that these metrics, though widely used in the literature, had only weak correlation with human judgments in the non task-oriented dialogue NLG setting. ", + "bbox": [ + 174, + 319, + 823, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In terms of supervised NLG evaluation metrics, Lowe et al. (2017) proposed the ADEM model which trains a hierarchical recurrent neural network in a supervised manner to predict human-like scores. This learned score was shown to correlate better with human judgments than any other automated metric. However, the drawback of this approach is the requirement for expensive human ratings. ", + "bbox": [ + 174, + 381, + 825, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Li et al. (2016) proposed to use reinforcement learning to train an end-to-end dialogue system. They simulate a dialogue between two agents and use a policy gradient algorithm with a reward function which evaluates specific properties of the responses generated by the dialogue system. ", + "bbox": [ + 176, + 458, + 825, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the adversarial setting, Kannan & Vinyals (2016) train a recurrent neural network discriminator to differentiate human-generated responses from model-generated responses. However, an extensive analysis of the viability and the ease of standardization of this approach is yet to be conducted. Li et al. (2017), apart from adversarially training dialogue response models, propose an independent adversarial evaluation metric AdverSuc and a measure of the model’s reliability called evaluator reliability error. Drawbacks of these approaches are that they are model-dependent. Adversarial methods might be promising for task-oriented dialogue systems but more research needs to be conducted on their account. ", + "bbox": [ + 174, + 507, + 825, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most of the work described so far has been done in the non task-oriented dialogue setting as there has been prior work indicating that automated metrics do not correlate well with humans in that setting. There has not yet been any empirical validation that these conclusions also apply to the task oriented setting. Research in the task oriented setting has mostly made use of automated metrics such as BLEU and human evaluation (Wen et al., 2015b; Sharma et al., 2017; Dusek & Jurcicek, ˇ 2016). ", + "bbox": [ + 174, + 626, + 825, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 METRICS ", + "text_level": 1, + "bbox": [ + 174, + 728, + 284, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section describes the set of automatic metrics whose correlation with human evaluation is studied. We consider first word-overlap metrics and then embedding-based metrics. In all that follows, when multiple references are provided, we compute the similarity between the prediction and all the references one-by-one, and then select the maximum value. We then average the scores across the entire corpus. ", + "bbox": [ + 174, + 758, + 825, + 829 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1 WORD-OVERLAP BASED METRICS ", + "text_level": 1, + "bbox": [ + 176, + 845, + 447, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1.1 BLEU ", + "text_level": 1, + "bbox": [ + 176, + 871, + 276, + 886 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The BLEU metric (Papineni et al., 2002) compares $n$ -grams between the candidate utterance and the reference utterance. The BLEU score is computed at the corpus-level and relies on the following ", + "bbox": [ + 174, + 895, + 823, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "modified precision: ", + "bbox": [ + 173, + 103, + 302, + 118 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/d32f40b1ba7b2dc150402592d5e78756627a8322215e60ec96ffe2cc4f87826c.jpg", + "text": "$$\np _ { n } = \\frac { \\displaystyle \\sum _ { C \\in \\{ C a n d i d a t e s \\} } \\sum _ { n - g r a m \\in C } C t _ { c l i p } ( n - g r a m ) } { \\displaystyle \\sum _ { C ^ { \\prime } \\in \\{ C a n d i d a t e s ^ { \\prime } \\} } \\sum _ { n - g r a m ^ { \\prime } \\in C ^ { \\prime } } C t _ { c l i p } ( n - g r a m ^ { \\prime } ) }\n$$", + "text_format": "latex", + "bbox": [ + 315, + 101, + 681, + 174 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where {Candidates} are the candidate answers generated by the model and $C t _ { c l i p }$ is the clipped count for the $n$ -gram which is the number of times the $n$ -gram is common to the candidate answer and the reference answer clipped by the maximum number of occurrences of the $n$ -gram in the reference answer. The BLEU-N score is defined as: ", + "bbox": [ + 176, + 176, + 821, + 232 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/ca3d22fd182470d656e2c0e4bfc007d7b99cc9009ca2a43f3d9f26749f273029.jpg", + "text": "$$\n\\begin{array} { l } { { \\displaystyle \\forall \\mathrm { ~ s c u r e ~ 1 s ~ u c t 1 1 1 e u ~ a s . } } } \\\\ { { \\displaystyle \\mathsf { B L E U - N } = \\mathsf { B P } \\exp ( \\sum _ { n } ^ { N } \\omega _ { n } \\log ( p _ { n } ) ) } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 374, + 224, + 622, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N$ is the maximum length of the $n$ -grams (in this paper, we compute BLEU-1 to BLEU-4), $\\omega$ is a weighting that is often uniform and BP is a brevity penalty. In this paper we report the BLEU score at the corpus level but we also compute this score at the sentence level to analyze its correlation with human evaluation. ", + "bbox": [ + 174, + 267, + 825, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2 METEOR ", + "text_level": 1, + "bbox": [ + 174, + 337, + 300, + 352 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The METEOR metric (Banerjee & Lavie, 2005) was proposed as a metric which correlates better at the sentence level with human evaluation. To compute the METEOR score, first, an alignment between the candidate and the reference sentences is created by mapping each unigram in the candidate sentence to 0 or 1 unigram in the reference sentence. The alignment is not only based on exact matches but also stem, synonym, and paraphrase matches. Based on this alignment, unigram precision and recall are computed and the METEOR score is: ", + "bbox": [ + 173, + 361, + 825, + 445 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/790b2f6cd10cf8720ca62ace23883c3da41ec9fb6871fd7e5a7b5e241ce282ef.jpg", + "text": "$$\n\\mathbf { M E T E O R } = F _ { m e a n } ( 1 - p )\n$$", + "text_format": "latex", + "bbox": [ + 405, + 448, + 591, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F _ { m e a n }$ is the harmonic mean between precision and recall with the weight for recall 9 times a high as the weight for precision, and $p$ is a penalty. ", + "bbox": [ + 173, + 467, + 823, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.3 ROUGE ", + "text_level": 1, + "bbox": [ + 174, + 510, + 289, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ROUGE (Lin, 2004) is a set of metrics that was first introduced for summarization. We compute ROUGE-L which is an F-measure based on the Longest Common Subsequence (LCS) between the candidate and reference utterances. ", + "bbox": [ + 174, + 534, + 825, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 EMBEDDING BASED METRICS ", + "text_level": 1, + "bbox": [ + 176, + 592, + 419, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider another set of metrics which compute the cosine similarity between the embeddings of the predicted and the reference sentence instead of relying on word overlaps. ", + "bbox": [ + 173, + 618, + 825, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 SKIP-THOUGHT ", + "text_level": 1, + "bbox": [ + 174, + 661, + 339, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Skip-Thought model (Kiros et al., 2015) is trained in an unsupervised fashion and uses a recurrent network to encode a given sentence into an embedding and then decode it to predict the preceding and following sentences. The model was trained on the BookCorpus dataset (Zhu et al., 2015). The embeddings produced by the encoder have a robust performance on semantic relatedness tasks. We use the pre-trained Skip-Thought encoder provided by the authors1. ", + "bbox": [ + 174, + 684, + 825, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We also compute other embedding-based methods which have been used as evaluation metrics for measuring human correlation in recent literature (Liu et al., 2016) for non task-oriented dialogue in Sections 3.2.2, 3.2.3, and 3.2.4. ", + "bbox": [ + 174, + 762, + 825, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.2 EMBEDDING AVERAGE ", + "text_level": 1, + "bbox": [ + 174, + 819, + 387, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This metric computes a sentence-level embedding by averaging the embeddings of the words composing this sentence: ", + "bbox": [ + 171, + 843, + 823, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/de135268afb6f0d9437350ee94d7386b11e424c2ef5bbbb687ca2870eacc206c.jpg", + "text": "$$\n\\bar { e } _ { C } = \\frac { \\sum _ { w \\in C } e _ { w } } { \\lvert \\sum _ { w ^ { \\prime } \\in C } e _ { w ^ { \\prime } } \\rvert } .\n$$", + "text_format": "latex", + "bbox": [ + 428, + 866, + 568, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this equation, the vectors $e _ { w }$ are embeddings for the words $w$ in the candidate sentence $C$ . ", + "bbox": [ + 168, + 103, + 781, + 118 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3 VECTOR EXTREMA ", + "text_level": 1, + "bbox": [ + 176, + 132, + 361, + 147 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vector extrema (Forgues et al., 2014) computes a sentence-level embedding by taking the most extreme value of the embeddings of the words composing the sentence for each dimension of the ", + "bbox": [ + 174, + 156, + 823, + 186 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/30d01bb57faa794509b59b892e9176188487ae8ffa33c33096cc55efab4b0df8.jpg", + "text": "$$\ne _ { r d } = \\left\\{ \\begin{array} { l l } { \\operatorname* { m a x } _ { w \\in C } e _ { w d } } & { \\mathrm { i f } e _ { w d } > \\vert \\operatorname* { m i n } _ { w ^ { \\prime } \\in C } e _ { w ^ { \\prime } d } \\vert } \\\\ { \\operatorname* { m i n } _ { w \\in C } e _ { w d } } & { \\mathrm { o t h e r w i s e } \\mathrm { . } } \\end{array} \\right.\n$$", + "text_format": "latex", + "bbox": [ + 333, + 188, + 663, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this equation, $d$ is an index over the dimensions of the embedding and $C$ is the candidate sentence. ", + "bbox": [ + 171, + 228, + 823, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.4 GREEDY MATCHING ", + "text_level": 1, + "bbox": [ + 176, + 251, + 367, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Greedy matching does not compute a sentence embedding but directly a similarity score between a candidate $C$ and a reference $r$ (Rus & Lintean, 2012). This similarity score is computed as follows: ", + "bbox": [ + 173, + 275, + 821, + 304 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/0858831b0e7f88249c5bf3756d1aa9dd3d4bb4205809b4735f8f61b484b3b299.jpg", + "text": "$$\n\\begin{array} { c } { \\displaystyle { G ( C , r ) = \\frac { \\sum w \\in C \\operatorname* { m a x } _ { \\hat { w } \\in r } c o s _ { - } s i m ( e _ { w } , w _ { \\hat { w } } ) } { | C | } } } \\\\ { \\displaystyle { G { M ( C , r ) } = \\frac { G ( C , r ) + G ( r , C ) } { 2 } . } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 336, + 308, + 661, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each word in the candidate sentence is greedily matched to a word in the reference sentence based on their embeddings’ cosine similarity. The score is an average of these similarities over the number of words in the candidate sentence. The same score is computed by reversing the roles of the candidate and reference sentences and the average of the two scores gives the final similarity score. ", + "bbox": [ + 174, + 373, + 825, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 RESPONSE GENERATION MODELS ", + "text_level": 1, + "bbox": [ + 174, + 449, + 490, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section presents the different natural language generation models that we use in this study. All of these models take as input a set of dialogue acts (Austin, 1962) potentially with slot types and slot values and translate this input into an utterance. An example input is inform(food $=$ Chinese) and a corresponding output would be “I am looking for a Chinese restaurant.”. In this example, the dialogue act is inform, the slot type is food, and the slot value is Chinese. ", + "bbox": [ + 174, + 481, + 825, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1 RANDOM ", + "text_level": 1, + "bbox": [ + 174, + 568, + 281, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a dialogue act with one or more slot types, the random model finds all the examples in the training set with the same dialogue act and slots (while ignoring slot values) and it randomly selects its output from this set of reference sentences. The datasets that we experiment on have some special slot values such as “yes”, “no”, and “don’t care”. Since the model ignores all slot values, these special cases are not properly handled, which results in slightly lower performance than what we could get by spending more time hand-engineering the model’s behavior for these values. ", + "bbox": [ + 174, + 593, + 825, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2 LSTM ", + "text_level": 1, + "bbox": [ + 174, + 694, + 261, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This model consists of a recurrent LSTM (Hochreiter & Schmidhuber, 1997) decoder. The dialogue acts and slot types are first encoded as a binary vector whose length is the number of possible combinations of dialogue acts and slot types in the dataset. We refer to this binary vector as the Dialogue Act $( D A )$ vector. The $D A$ vector for a given set of dialogue acts is a binary vector over the fused dialogue act-slot types, e.g., INFORM-FOOD, INFORM-COUNT, etc. ", + "bbox": [ + 174, + 719, + 825, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This binary vector is given as input to the decoder at each time-step of the LSTM. The decoder then outputs a delexicalized sentence. A delexicalized sentence contains placeholders for the slot values. An example is “I am looking for a FOOD restaurant.”. The values for the delexicalized slots (the type of food in this example) are then directly copied from the input. ", + "bbox": [ + 174, + 796, + 825, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3 DELEX-SC-LSTM ", + "text_level": 1, + "bbox": [ + 174, + 869, + 339, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This model uses the same architecture as the LSTM model presented in the previous section except that it uses sc-LSTM (Wen et al., 2015b) units in the decoder instead of LSTM units. We call this ", + "bbox": [ + 174, + 895, + 823, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model the “delex-sc-LSTM”2. As in the previous model, the input $D A$ vector only encodes acts and delexicalized slots. It does not contain any information about the slot value. ", + "bbox": [ + 169, + 103, + 823, + 132 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By providing this model the same $D A$ vector input as the one given to the LSTM model, we can directly study if the additional complexity of the sc-LSTM unit’s reading gate provides significant improvement over the small-sized task-oriented dialogue datasets which are currently available. ", + "bbox": [ + 174, + 138, + 823, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4 HIERARCHICAL-LEX-DELEX-SC-LSTM ", + "text_level": 1, + "bbox": [ + 176, + 204, + 485, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This model is a variant of the “ld-sc-LSTM” model proposed by Sharma et al. (2017) which is based on an encoder-decoder framework. We call our model “hierarchical-lex-delex-sc-LSTM”3. ", + "bbox": [ + 174, + 231, + 823, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present the encoder in Figure 1. The encoder consists of a hierarchical LSTM with $N _ { e }$ time-steps, where $N _ { e }$ is the number of non-zero entries in the $D A$ vector. Each time-step of the encoder encodes one dialogue act’s delexicalized and lexicalized slot-value pair (e.g. (INFORM-FOOD, ‘Chinese’)). The delexicalized act-slot part is encoded as a one-hot vector which we refer to as $D A _ { t }$ . $D A _ { t }$ is constructed by masking all except the $t ^ { \\mathrm { { t h } } }$ dialogue act in the DA vector4. The lexicalized value part is en", + "bbox": [ + 174, + 281, + 483, + 433 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/17acec8a69179f00d7d76eae7dbbeb4746f29c87169dd83f9d9aae7969b3e28f.jpg", + "image_caption": [ + "Figure 1: Encoder of the hld-scLSTM model " + ], + "image_footnote": [], + "bbox": [ + 496, + 276, + 823, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "coded by an LSTM encoder which shares parameters across all time-steps and operates over the word-embeddings of the lexicalized values $v _ { t , i }$ . Our model differs from the “ld-sc-LSTM” model in that we use an LSTM encoder over the word-embeddings instead of computing the mean of the word-embeddings. The final hidden state of this LSTM is concatenated with $D A _ { t }$ and is given as input to the upper LSTM (see Figure 1). The final hidden state of the upper LSTM is then provided to the decoder as input. This is another difference from the “ld-sc-LSTM” as that uses the mean of all the hidden states of the encoder instead, which, in our experiments, did not perform as well as using just the final hidden state $E$ . ", + "bbox": [ + 174, + 434, + 825, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The decoder is described in Figure 2. It is the same as in the “ld-sc-LSTM” model. At each time-step, it takes as input the encoder output $E$ , the $D A$ vector, and the word-embedding of the word generated at the previous time-step. The $D A$ vector is also additionally provided to the scLSTM cell in order for it to be regulated by its reading gate as described in Wen et al. (2015b). ", + "bbox": [ + 174, + 551, + 500, + 662 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/114dbf8875d6790d50ee41019e8384e6fe26efeea5c18f9566f2fbec598cb400.jpg", + "image_caption": [ + "Figure 2: Decoder of the hld-scLSTM model " + ], + "image_footnote": [], + "bbox": [ + 511, + 553, + 825, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 689, + 326, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1 DECODING ", + "text_level": 1, + "bbox": [ + 174, + 723, + 290, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, at each time-step, we use the ground truth word from the previous time-step. The model thus learns to generate the next word given the previous one. On the other hand, to generate sentences during test time, we use beam search. The first word input to the generator is a special token $< b o s >$ which indicates the beginning of the sequence. Decoding is stopped if we reach a specified maximum number of time-steps or if the model outputs a special token $< e o s >$ which indicates the end of the sequence. We also use a slot error rate penalty, similarly to Wen et al. (2015b), to re-rank the sentences generated with beam search. We use this method for all three of the LSTM, d-scLSTM, and hld-scLSTM models for fairness. ", + "bbox": [ + 173, + 752, + 825, + 863 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4cdc49a5bd65ea5ff860b94e3cce4146a9861fad40d0462bec314d81a7dbfe5e.jpg", + "table_caption": [ + "Table 1: Performance comparison across models on word-overlap based automated metrics " + ], + "table_footnote": [], + "table_body": "
DSTC2Restaurants
B-1B-2B-3B-4MRLB-1B-2B-3B-4MRL
Gold1.001.001.001.001.001.001.001.001.001.001.001.00
Random0.8750.8430.8220.8070.5640.8520.8720.8130.7650.7210.5040.796
LSTM0.9000.8790.8630.8510.6100.8880.9820.9660.9490.9320.6520.944
d-scLSTM0.8800.8500.8280.8120.5780.8740.9800.9640.9480.9310.6540.945
hld-scLSTM0.9090.8900.8780.8700.6240.8990.9850.9780.9700.9620.7040.965
", + "bbox": [ + 186, + 119, + 812, + 196 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/782379d7f8a969d348efbec4caa0e6edf6b2d4c105b13cb0daa619db4478039d.jpg", + "table_caption": [ + "Table 2: Performance comparison across models on sentence-embedding based automated metrics " + ], + "table_footnote": [], + "table_body": "
DSTC2Restaurants
SkipThoughtEmbeddingAverageVectorExtremaGreedyMatchingSkipThoughtEmbeddingAverageVectorExtremaGreedyMatching
Gold1.001.001.001.001.001.001.001.00
Random0.9060.9810.9100.9470.8430.9570.9050.930
LSTM0.9460.9850.9350.9620.9450.9970.9860.991
d-scLSTM0.9250.9840.9260.9570.9480.9970.9860.991
hld-scLSTM0.9320.9870.9420.9640.9680.9970.9890.993
", + "bbox": [ + 202, + 219, + 792, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similarly to the LSTM model, the d-scLSTM and hld-scLSTM generate delexicalized sentences, i.e., they generate slot tokens instead of slot values directly. These slot tokens are replaced with slot values in a post-processing step which is a fairly common step in task-oriented dialogue NLG literature. ", + "bbox": [ + 174, + 316, + 825, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2 EVALUATION ", + "text_level": 1, + "bbox": [ + 174, + 391, + 305, + 406 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In NLG tasks, improvements in automated metric scores are most commonly used to demonstrate improvement in the generation task. However, these metrics have been shown to only weakly correlate with human evaluation in the non task-oriented dialogue setting (Liu et al., 2016) and hence are not considered reliable measures of improvement. Human evaluation is considered the metric of choice, but human ratings are expensive to obtain. The ease of computing these automated metrics and their availability for rapid prototyping has lead to their widespread adoption. ", + "bbox": [ + 174, + 419, + 825, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the models described in the previous section on the DSTC2 (Henderson et al., 2014) and the Restaurants datasets (Wen et al., 2015a) using these automated metrics. These datasets are some of the only available resources for studying NLG for task-oriented dialogue. The DSTC2 dataset contains dialogues between human users and a dialogue system in a restaurant domain. The dataset is annotated with dialogue acts, slot type, and slot values. The NLG component of the dialogue system used for data collection is templated. The Restaurants dataset was specifically proposed for NLG and provides, for a set of dialogue acts with slot types and slot values, two sentences generated by humans. ", + "bbox": [ + 173, + 508, + 825, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present the results of our experiments in Table 1 and Table 2. The scores of all the models on these automated metrics are very high. This indicates that there is significant word overlap between the generated and the reference sentences and that the NLG task on these datasets can be solved with a simple model such as the LSTM model. In effect, Table 1 shows that the LSTM model performs comparably to the d-scLSTM model based on the word-overlap metrics. This can be explained by the fact that the d-scLSTM model has more parameters and might suffer from overfitting issues on these relatively small datasets. ", + "bbox": [ + 173, + 626, + 825, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The hld-scLSTM is considered to consistently outperform the other models based on the wordoverlap metrics. As explained by Sharma et al. (2017), this improvement results from the model’s access to the lexicalized slot values, due to which it can take into account the grammatical associations of the generated words near the output tokens, thereby generating higher quality sentences. However, Table 2 shows that sentence-embedding based metrics judge all the models except the random one to perform quite similarly with again, very high performance scores. ", + "bbox": [ + 174, + 731, + 823, + 815 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the next section, we add human evaluation for these models on these datasets. ", + "bbox": [ + 174, + 821, + 697, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3 HUMAN RATING COLLECTION ", + "text_level": 1, + "bbox": [ + 176, + 854, + 419, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We randomly selected 20 dialogue acts from the test set of each dataset. For each of these contexts, we presented 5 sentences to the evaluators: the gold response provided in the test set and the responses generated by the four models described in Section 4. These sentences were randomly shuffled and not presented in the same order. We invited 18 human users to score each of these 100 sentences on a Likert-type scale of 1 to 5. The users were asked to rate the responses depending on how appropriate they were for the specified dialogue acts. ", + "bbox": [ + 176, + 882, + 823, + 922 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/18289be4190cbbf67f923162b61f526162ab7aca3e0d7abbab61623c5036e6d9.jpg", + "image_caption": [ + "Figure 3: Scatter plots for correlation of some automated metrics with human evaluation for (a) the DSTC2 dataset, and (b) the Restaurants dataset. Random gaussian noise $\\mathcal { N } ( 0 , 0 . 1 )$ has been added to data points along the human score axis and $\\mathcal { N } ( 0 , 0 . 0 2 )$ has been added to the automated metric score’s axis to aid visualization of overlapping data points. Transparency has been added for the same effect. " + ], + "image_footnote": [], + "bbox": [ + 183, + 106, + 813, + 373 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 486, + 823, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A score of 1 was the lowest score, meaning that the response was not appropriate at all whereas a score of 5 meant that the sentence was highly appropriate. ", + "bbox": [ + 174, + 529, + 612, + 569 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ee87a9e1885b0fe151cbbeb91e4be51300f4ad64919da97839a19c8d62eef0b2.jpg", + "table_caption": [ + "Table 3: Pairwise Cohen’s kappa scores for the 11 human users " + ], + "table_footnote": [], + "table_body": "
K# pairs% pairs
>0.155/55100.0 %
>0.240/5572.7 %
>0.328/5550.9 %
>0.419/5534.5 %
>0.58/5514.5 %
>0.60/550.0 %
", + "bbox": [ + 632, + 577, + 813, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We computed Cohen’s kappa scores (Cohen, 1960) between the human users in pairs of two. We removed 7 users who had kappa scores less than 0.1 and used the remaining 11 users for the correlation study. The kappa scores are presented in Table 3. Most of the user pairs have a Cohen’s $\\kappa > 0 . 3$ which indicates fair agreement between users (Viera et al., 2005). ", + "bbox": [ + 174, + 577, + 614, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4 CORRELATION BETWEEN AUTOMATED METRICS AND HUMAN SCORES ", + "text_level": 1, + "bbox": [ + 176, + 685, + 696, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present the correlation between the automated metrics and our collected human ratings in Table 4. We measure human v.s. human correlation by randomly splitting the human users into two groups. The results indicate that in most cases, human scores correlate the best with other human scores. Except in the case of the Spearman correlation for BLEU-N scores, we can see that there is a positive correlation between the automated metrics and the human scores for these task-oriented datasets, which contrasts with the non task-oriented dialogue setting where Liu et al. (2016) observed no strong correlation trends. ", + "bbox": [ + 174, + 714, + 825, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A likely explanation for the negative Spearman correlation values for BLEU-N is that there is only one gold reference per context in the DSTC2 dataset. The Restaurants dataset, on the other hand, provides two gold references per context. Having multiple gold references increases the likelihood that the generated response will have significant word-overlap with one of the reference responses. ", + "bbox": [ + 174, + 819, + 825, + 875 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present scatter plots for some of the metrics presented in Table 4 in Figure 3. We observe that all the metrics correlate very well with humans on high scoring examples. As it can be seen in the scatter plots, most of the sentences are given the maximal score of 5 by the human evaluators. This confirms our previous observation that the available corpora for task-oriented dialogue NLG task are not very challenging and a simple LSTM-based model can output high-quality responses. ", + "bbox": [ + 176, + 882, + 823, + 924 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f97c5804c3aaa31ce0e6671a6a34b9ae69f2e34375917e374536be3806e7bca6.jpg", + "table_caption": [ + "Table 4: Correlation of automated metrics with human evaluations scores " + ], + "table_footnote": [], + "table_body": "
MetricDSTC2Restaurants
Spearmanp-valuePearsonp-valueSpearmanp-valuePearsonp-value
Bleu1Bleu 2-0.317<0.0050.583<0.0050.0690.4940.2770.005
-0.318<0.0050.526<0.0050.0910.3660.1660.099
Bleu 3-0.318<0.0050.500<0.0050.1090.2800.2230.026
Bleu 4METEORROUGELSkip Thought-0.318<0.0050.461<0.0050.1050.2960.2550.010
0.295<0.0050.582<0.0050.353<0.0050.489<0.005
0.294<0.0050.448<0.0050.346<0.0050.382<0.005
0.528<0.0050.0860.3970.284<0.0050.364<0.005
Embedding Average0.295<0.0050.485<0.0050.423<0.0050.2600.009
Vector ExtremaGreedy Matching0.299<0.0050.624<0.0050.446<0.0050.287<0.005
0.295<0.0050.572<0.0050.446<0.0050.325<0.005
Human0.810<0.0050.984<0.0050.653<0.0050.857<0.005
", + "bbox": [ + 207, + 117, + 789, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 265, + 821, + 294 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, among the word overlap based automated metrics, METEOR consistently correlates with human evaluation on both datasets. These results confirm the original findings by Banerjee & Lavie (2005) who showed that METEOR had good correlation with human evaluation in the machine translation task. The comparison with machine translation is highly relevant in the task-oriented setting because the NLG model essentially learns to translate the abstract representation of a sentence into a sentence. It is a translation task contrary to the non task-oriented setting where the NLG model needs to decide and output a new sentence based on the last sentence typed by a user and dialogue history. Therefore, automated metrics coming from the machine translation literature are more adequate in our case than in the non-task oriented case as shown by Liu et al. (2016). ", + "bbox": [ + 174, + 300, + 825, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It is interesting to see that METEOR correlates well with human evaluation consistently. This can be explained by the fact that even though METEOR does not rely on word embeddings, it includes notions of synonymy and paraphrasing when computing the alignment between the candidate and reference utterances. ", + "bbox": [ + 176, + 433, + 825, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6 DISCUSSION ", + "text_level": 1, + "bbox": [ + 174, + 508, + 310, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluated several natural language generation models trained on the DSTC2 and the Restaurants datasets based on several automated metrics. We also performed human evaluation on the model-generated responses and our study shows that human evaluation is a much more reliable metric compared to the others. Among the word-overlap based automated metrics, we found that the METEOR score correlates the most with human judgments and we suggest using METEOR for taskoriented dialogue natural language generation instead of BLEU. We also observe that these metrics are more reliable in the task-oriented dialogue setting compared to the general, non task-oriented one due to the limited possible diversity in the task-oriented setting. Also, as observed by Galley et al. (2015), we can see that word-overlap based metrics correlate better with human evaluation when multiple references are provided, as in the Restaurants dataset. Otherwise, as in the case of DSTC2 which only provides one reference sentence per example, we observe that all the BLEU-N metrics negatively correlate with human evaluation on Spearman correlation. ", + "bbox": [ + 174, + 540, + 825, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As has been observed in the machine translation literature, using beam search improves the quality of generated sentences significantly compared to stochastic sampling. For similar models, our results show improvement in the automated metrics’ scores compared to Wen et al. (2015b) who used stochastic sampling for decoding instead of beam search. ", + "bbox": [ + 176, + 714, + 825, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Wen et al. (2015b) did not use the slot error rate penalty with the vanilla LSTM model in their experiments. After adding the penalty in our case, we observe that the vanilla LSTM-based model performs as well as the delexicalized semantically-controlled LSTM model. This suggests that the added complexity introduced by the sc-LSTM unit does not offer a significant advantage for these two datasets. ", + "bbox": [ + 174, + 777, + 825, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "High performance on automated metrics, achieved by our models on the DSTC2 and the Restaurants datasets lead us to conclude that these datasets are not very challenging for the NLG task. The taskoriented dialogue community should move towards using larger and more complex datasets, which have been recently announced, such as the Frames dataset (El Asri et al., 2017) or the E2E NLG Challenge dataset (Novikova et al., 2016). ", + "bbox": [ + 174, + 854, + 823, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 102, + 285, + 118 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "John Langshaw Austin. How to do things with words. Oxford University Press, 1962. ", + "bbox": [ + 173, + 126, + 733, + 140 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Scott Axelrod. Natural language generation in the ibm flight information system. In Proceedings of the 2000 ANLP/NAACL Workshop on Conversational Systems - Volume 3, ANLP/NAACLConvSyst ’00, pp. 21–26, Stroudsburg, PA, USA, 2000. Association for Computational Linguistics. doi: 10.3115/1117562.1117567. ", + "bbox": [ + 173, + 148, + 825, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proc. ACL workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, 2005. ", + "bbox": [ + 174, + 214, + 823, + 257 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jacob Cohen. A coefficient of agreement for nominal scales. Educational and psychological measurement, 20(1):37–46, 1960. ", + "bbox": [ + 168, + 265, + 823, + 295 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ondˇrej Dusek and Filip Jurcicek. Sequence-to-sequence generation for spoken dialogue via deep ˇ syntax trees and strings. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pp. 45–51, Berlin, Germany, August 2016. Association for Computational Linguistics. ", + "bbox": [ + 174, + 303, + 825, + 359 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Layla El Asri, Hannes Schulz, Shikhar Sharma, Jeremie Zumer, Justin Harris, Emery Fine, Rahul Mehrotra, and Kaheer Suleman. Frames: A corpus for adding memory to goal-oriented dialogue systems. arXiv preprint arXiv:1704.00057, 2017. ", + "bbox": [ + 173, + 368, + 825, + 411 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Michael Elhadad. Generating coherent argumentative paragraphs. In In Proceedings of COLING ’92, volume II, pp. 638–644, 1992. ", + "bbox": [ + 173, + 420, + 823, + 448 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Gabriel Forgues, Joelle Pineau, Jean-Marie Larcheveque, and Real Tremblay. Bootstrapping dialog systems with word embeddings. 2014. ", + "bbox": [ + 173, + 457, + 821, + 486 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Michel Galley, Chris Brockett, Alessandro Sordoni, Yangfeng Ji, Michael Auli, Chris Quirk, Margaret Mitchell, Jianfeng Gao, and Bill Dolan. deltableu: A discriminative metric for generation tasks with intrinsically diverse targets. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pp. 445–450, Beijing, China, July 2015. Association for Computational Linguistics. ", + "bbox": [ + 173, + 494, + 825, + 579 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Matthew Henderson, Blaise Thomson, and Jason D. Williams. The second dialog state tracking challenge. In Proceedings of the 15th Annual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL), pp. 263–272, Philadelphia, PA, U.S.A., June 2014. Association for Computational Linguistics. ", + "bbox": [ + 174, + 587, + 825, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8): 1735–1780, 1997. ", + "bbox": [ + 171, + 652, + 823, + 681 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Anjuli Kannan and Oriol Vinyals. Adversarial evaluation of dialogue models. In NIPS 2016 Workshop on Adversarial Training, 2016. ", + "bbox": [ + 171, + 689, + 823, + 719 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ryan Kiros, Yukun Zhu, Ruslan R Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Skip-thought vectors. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett (eds.), Advances in Neural Information Processing Systems 28, pp. 3294–3302. Curran Associates, Inc., 2015. ", + "bbox": [ + 174, + 727, + 825, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jiwei Li, Will Monroe, Alan Ritter, Dan Jurafsky, Michel Galley, and Jianfeng Gao. Deep reinforcement learning for dialogue generation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pp. 1192–1202, Austin, Texas, November 2016. Association for Computational Linguistics. ", + "bbox": [ + 173, + 792, + 825, + 849 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jiwei Li, Will Monroe, Tianlin Shi, Alan Ritter, and Dan Jurafsky. Adversarial learning for neural dialogue generation. CoRR, abs/1701.06547, 2017. ", + "bbox": [ + 173, + 858, + 821, + 887 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Proc. ACL workshop on Text Summarization Branches Out, 2004. ", + "bbox": [ + 173, + 895, + 821, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Chia-Wei Liu, Ryan Lowe, Iulian Serban, Mike Noseworthy, Laurent Charlin, and Joelle Pineau. How not to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics for dialogue response generation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pp. 2122–2132. Association for Computational Linguistics, 2016. ", + "bbox": [ + 174, + 103, + 825, + 160 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ryan Lowe, Nissan Pow, IV Serban, Laurent Charlin, and Joelle Pineau. Incorporating unstructured textual knowledge sources into neural dialogue systems. In Neural Information Processing Systems Workshop on Machine Learning for Spoken Language Understanding, 2015. ", + "bbox": [ + 176, + 166, + 823, + 210 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ryan Lowe, Michael Noseworthy, Iulian V. Serban, Nicolas Angelard-Gontier, Yoshua Bengio, and Joelle Pineau. Towards an automatic turing test: Learning to evaluate dialogue responses. In Proceedings of the 5th International Conference on Learning Representations (ICLR) Workshop, Toulon, France, April 2017. ", + "bbox": [ + 173, + 215, + 825, + 272 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jekaterina Novikova, Oliver Lemon, and Verena Rieser. Crowd-sourcing nlg data: Pictures elicit better data. In Proceedings of the 9th International Natural Language Generation conference, pp. 265–273, Edinburgh, UK, September 5-8 2016. Association for Computational Linguistics. ", + "bbox": [ + 174, + 279, + 821, + 323 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alice H. Oh and Alexander I. Rudnicky. Stochastic language generation for spoken dialogue systems. In Proceedings of the 2000 ANLP/NAACL Workshop on Conversational Systems - Volume 3, ANLP/NAACL-ConvSyst ’00, pp. 27–32, Stroudsburg, PA, USA, 2000. Association for Computational Linguistics. doi: 10.3115/1117562.1117568. ", + "bbox": [ + 174, + 329, + 825, + 386 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proc. of ACL, 2002. ", + "bbox": [ + 169, + 392, + 823, + 422 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Verena Rieser and Oliver Lemon. Natural language generation as planning under uncertainty for spoken dialogue systems. In Proceedings of the 12th Conference of the European Chapter of the ACL (EACL 2009), pp. 683–691, Athens, Greece, March 2009. Association for Computational Linguistics. ", + "bbox": [ + 174, + 429, + 825, + 486 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Vasile Rus and Mihai C. Lintean. A comparison of greedy and optimal assessment of natural language student input using word-to-word similarity metrics. In Proceedings of the Seventh Workshop on Building Educational Applications Using NLP, 2012. ", + "bbox": [ + 174, + 492, + 823, + 535 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Iulian Vlad Serban, Alessandro Sordoni, Ryan Lowe, Laurent Charlin, Joelle Pineau, Aaron C. Courville, and Yoshua Bengio. A hierarchical latent variable encoder-decoder model for generating dialogues. CoRR, abs/1605.06069, 2016. ", + "bbox": [ + 176, + 541, + 821, + 585 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Shikhar Sharma, Jing He, Kaheer Suleman, Hannes Schulz, and Philip Bachman. Natural language generation in dialogue using lexicalized and delexicalized data. In Proceedings of the 5th International Conference on Learning Representations (ICLR) Workshop, 2017. ", + "bbox": [ + 173, + 590, + 823, + 635 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Anthony J Viera, Joanne M Garrett, et al. Understanding interobserver agreement: the kappa statistic. Family Medicine, 37(5):360–363, 2005. ", + "bbox": [ + 169, + 641, + 823, + 670 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Marilyn Walker, Amanda Stent, Franc¸ois Mairesse, and Rashmi Prasad. Individual and domain adaptation in sentence planning for dialogue. J. Artif. Int. Res., 30(1):413–456, November 2007. ISSN 1076-9757. ", + "bbox": [ + 173, + 676, + 823, + 719 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tsung-Hsien Wen, Milica Gasiˇ c, Dongho Kim, Nikola Mrk ´ siˇ c, Pei-Hao Su, David Vandyke, and ´ Steve Young. Stochastic Language Generation in Dialogue using Recurrent Neural Networks with Convolutional Sentence Reranking. In Proceedings of the 16th Annual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL). Association for Computational Linguistics, September 2015a. ", + "bbox": [ + 174, + 726, + 825, + 796 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tsung-Hsien Wen, Milica Gasiˇ c, Nikola Mrk ´ siˇ c, Pei-Hao Su, David Vandyke, and Steve Young. ´ Semantically conditioned LSTM-based natural language generation for spoken dialogue systems. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, September 2015b. ", + "bbox": [ + 174, + 803, + 825, + 859 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In The IEEE International Conference on Computer Vision (ICCV), December 2015. ", + "bbox": [ + 176, + 867, + 825, + 922 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/parse/train/r17lFgZ0Z/r17lFgZ0Z_middle.json b/parse/train/r17lFgZ0Z/r17lFgZ0Z_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..fdb67ef0dcb1c1d736a6785f7a99471ae49c0a4f --- /dev/null +++ b/parse/train/r17lFgZ0Z/r17lFgZ0Z_middle.json @@ -0,0 +1,27015 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 504, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 504, + 96 + ], + "score": 1.0, + "content": "RELEVANCE OF UNSUPERVISED METRICS IN TASK-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 505, + 117 + ], + "score": 1.0, + "content": "ORIENTED DIALOGUE FOR EVALUATING NATURAL", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 297, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 297, + 137 + ], + "score": 1.0, + "content": "LANGUAGE GENERATION", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 112, + 155, + 244, + 177 + ], + "lines": [ + { + "bbox": [ + 113, + 155, + 201, + 167 + ], + "spans": [ + { + "bbox": [ + 113, + 155, + 201, + 167 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 166, + 245, + 178 + ], + "spans": [ + { + "bbox": [ + 111, + 166, + 245, + 178 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 278, + 206, + 333, + 218 + ], + "lines": [ + { + "bbox": [ + 277, + 206, + 335, + 219 + ], + "spans": [ + { + "bbox": [ + 277, + 206, + 335, + 219 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 231, + 468, + 408 + ], + "lines": [ + { + "bbox": [ + 142, + 232, + 469, + 244 + ], + "spans": [ + { + "bbox": [ + 142, + 232, + 469, + 244 + ], + "score": 1.0, + "content": "Automated metrics such as BLEU are widely used in the machine translation liter-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 243, + 469, + 256 + ], + "spans": [ + { + "bbox": [ + 141, + 243, + 469, + 256 + ], + "score": 1.0, + "content": "ature. They have also been used recently in the dialogue community for evaluating", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 254, + 469, + 267 + ], + "spans": [ + { + "bbox": [ + 141, + 254, + 469, + 267 + ], + "score": 1.0, + "content": "dialogue response generation. However, previous work in dialogue response gen-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 263, + 469, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 263, + 469, + 279 + ], + "score": 1.0, + "content": "eration has shown that these metrics do not correlate strongly with human judg-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 276, + 469, + 288 + ], + "spans": [ + { + "bbox": [ + 141, + 276, + 469, + 288 + ], + "score": 1.0, + "content": "ment in the non task-oriented dialogue setting. Task-oriented dialogue responses", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 288, + 469, + 299 + ], + "spans": [ + { + "bbox": [ + 141, + 288, + 469, + 299 + ], + "score": 1.0, + "content": "are expressed on narrower domains and exhibit lower diversity. It is thus reason-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 296, + 469, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 296, + 469, + 311 + ], + "score": 1.0, + "content": "able to think that these automated metrics would correlate well with human judg-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 308, + 470, + 322 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 470, + 322 + ], + "score": 1.0, + "content": "ment in the task-oriented setting where the generation task consists of translating", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 320, + 470, + 332 + ], + "spans": [ + { + "bbox": [ + 141, + 320, + 470, + 332 + ], + "score": 1.0, + "content": "dialogue acts into a sentence. We conduct an empirical study to confirm whether", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 330, + 470, + 343 + ], + "spans": [ + { + "bbox": [ + 141, + 330, + 470, + 343 + ], + "score": 1.0, + "content": "this is the case. Our findings indicate that these automated metrics have stronger", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 343, + 470, + 354 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 470, + 354 + ], + "score": 1.0, + "content": "correlation with human judgments in the task-oriented setting compared to what", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 352, + 470, + 364 + ], + "spans": [ + { + "bbox": [ + 141, + 352, + 470, + 364 + ], + "score": 1.0, + "content": "has been observed in the non task-oriented setting. We also observe that these", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 364, + 469, + 375 + ], + "spans": [ + { + "bbox": [ + 141, + 364, + 469, + 375 + ], + "score": 1.0, + "content": "metrics correlate even better for datasets which provide multiple ground truth ref-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 375, + 469, + 386 + ], + "spans": [ + { + "bbox": [ + 142, + 375, + 469, + 386 + ], + "score": 1.0, + "content": "erence sentences. In addition, we show that some of the currently available cor-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 385, + 470, + 398 + ], + "spans": [ + { + "bbox": [ + 141, + 385, + 470, + 398 + ], + "score": 1.0, + "content": "pora for task-oriented language generation can be solved with simple models and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 397, + 302, + 408 + ], + "spans": [ + { + "bbox": [ + 141, + 397, + 302, + 408 + ], + "score": 1.0, + "content": "advocate for more challenging datasets.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 13.5 + }, + { + "type": "title", + "bbox": [ + 108, + 430, + 205, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 208, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 208, + 446 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 456, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 104, + 455, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 104, + 455, + 506, + 470 + ], + "score": 1.0, + "content": "Rule-based and template-based dialogue response generation systems have been around for a long", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "score": 1.0, + "content": "time (Axelrod, 2000; Elhadad, 1992). Even today, many task-oriented dialogue systems deployed in", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 478, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 104, + 478, + 506, + 492 + ], + "score": 1.0, + "content": "production are rule-based and template-based. These systems do not scale with increasing domain", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "complexity and maintaining the increasing number of templates becomes cumbersome. In the past,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 500, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 514 + ], + "score": 1.0, + "content": "Oh & Rudnicky (2000) proposed a corpus-based approach for Natural Language Generation (NLG)", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "for task-oriented dialogue systems. Other statistical approaches were proposed using tree-based", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 521, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 104, + 521, + 506, + 537 + ], + "score": 1.0, + "content": "models and reinforcement learning (Walker et al., 2007; Rieser & Lemon, 2009). Recently, deep-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 532, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 104, + 532, + 506, + 547 + ], + "score": 1.0, + "content": "learning based approaches (Wen et al., 2015b; Sharma et al., 2017; Lowe et al., 2015; Serban et al.,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 545, + 388, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 388, + 558 + ], + "score": 1.0, + "content": "2016) have shown promising results for dialogue response generation.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "The automated evaluation of machine-generated language is challenging and an important problem", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "score": 1.0, + "content": "for the natural language processing community. The most widely used automated metrics currently", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 582, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 104, + 582, + 506, + 596 + ], + "score": 1.0, + "content": "are word-overlap based metrics such as BLEU (Papineni et al., 2002), METEOR (Banerjee & Lavie,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 593, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 506, + 606 + ], + "score": 1.0, + "content": "2005) which were proposed originally for machine translation. While these metrics were shown to", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "correlate well with manual human evaluation in machine translation tasks, previous studies showed", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "that this is not the case in non-task oriented dialogue (Liu et al., 2016). This is explained by the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 104, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "fact that for the same context (e.g. a user utterance), responses in dialogue have more diversity.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "Word-overlap metrics are unable to capture semantics and thus, can lead to poor scores even for", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "appropriate responses. Human evaluation in this case is the most reliable metric. However, human", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 660, + 394, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 394, + 672 + ], + "score": 1.0, + "content": "judgments are expensive to obtain and not readily available at all times.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 504, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 504, + 689 + ], + "score": 1.0, + "content": "Task-oriented dialogue systems are employed in narrower domains (e.g. booking a restaurant) and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 687, + 504, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 504, + 699 + ], + "score": 1.0, + "content": "responses do not have as much diversity as in the non-task oriented setting. Another important", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "difference is that in the non-task oriented setting, response generation is often performed end-to-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "end, which means that the model takes as input the last user utterance and potentially the dialogue", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "history and it outputs the next system answer. In the task-oriented setting, on the other hand, the", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 504, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 504, + 96 + ], + "score": 1.0, + "content": "RELEVANCE OF UNSUPERVISED METRICS IN TASK-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 505, + 117 + ], + "score": 1.0, + "content": "ORIENTED DIALOGUE FOR EVALUATING NATURAL", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 297, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 297, + 137 + ], + "score": 1.0, + "content": "LANGUAGE GENERATION", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 112, + 155, + 244, + 177 + ], + "lines": [ + { + "bbox": [ + 113, + 155, + 201, + 167 + ], + "spans": [ + { + "bbox": [ + 113, + 155, + 201, + 167 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 166, + 245, + 178 + ], + "spans": [ + { + "bbox": [ + 111, + 166, + 245, + 178 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 111, + 155, + 245, + 178 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 206, + 333, + 218 + ], + "lines": [ + { + "bbox": [ + 277, + 206, + 335, + 219 + ], + "spans": [ + { + "bbox": [ + 277, + 206, + 335, + 219 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 231, + 468, + 408 + ], + "lines": [ + { + "bbox": [ + 142, + 232, + 469, + 244 + ], + "spans": [ + { + "bbox": [ + 142, + 232, + 469, + 244 + ], + "score": 1.0, + "content": "Automated metrics such as BLEU are widely used in the machine translation liter-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 243, + 469, + 256 + ], + "spans": [ + { + "bbox": [ + 141, + 243, + 469, + 256 + ], + "score": 1.0, + "content": "ature. They have also been used recently in the dialogue community for evaluating", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 254, + 469, + 267 + ], + "spans": [ + { + "bbox": [ + 141, + 254, + 469, + 267 + ], + "score": 1.0, + "content": "dialogue response generation. However, previous work in dialogue response gen-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 263, + 469, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 263, + 469, + 279 + ], + "score": 1.0, + "content": "eration has shown that these metrics do not correlate strongly with human judg-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 276, + 469, + 288 + ], + "spans": [ + { + "bbox": [ + 141, + 276, + 469, + 288 + ], + "score": 1.0, + "content": "ment in the non task-oriented dialogue setting. Task-oriented dialogue responses", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 288, + 469, + 299 + ], + "spans": [ + { + "bbox": [ + 141, + 288, + 469, + 299 + ], + "score": 1.0, + "content": "are expressed on narrower domains and exhibit lower diversity. It is thus reason-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 296, + 469, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 296, + 469, + 311 + ], + "score": 1.0, + "content": "able to think that these automated metrics would correlate well with human judg-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 308, + 470, + 322 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 470, + 322 + ], + "score": 1.0, + "content": "ment in the task-oriented setting where the generation task consists of translating", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 320, + 470, + 332 + ], + "spans": [ + { + "bbox": [ + 141, + 320, + 470, + 332 + ], + "score": 1.0, + "content": "dialogue acts into a sentence. We conduct an empirical study to confirm whether", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 330, + 470, + 343 + ], + "spans": [ + { + "bbox": [ + 141, + 330, + 470, + 343 + ], + "score": 1.0, + "content": "this is the case. Our findings indicate that these automated metrics have stronger", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 343, + 470, + 354 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 470, + 354 + ], + "score": 1.0, + "content": "correlation with human judgments in the task-oriented setting compared to what", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 352, + 470, + 364 + ], + "spans": [ + { + "bbox": [ + 141, + 352, + 470, + 364 + ], + "score": 1.0, + "content": "has been observed in the non task-oriented setting. We also observe that these", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 364, + 469, + 375 + ], + "spans": [ + { + "bbox": [ + 141, + 364, + 469, + 375 + ], + "score": 1.0, + "content": "metrics correlate even better for datasets which provide multiple ground truth ref-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 375, + 469, + 386 + ], + "spans": [ + { + "bbox": [ + 142, + 375, + 469, + 386 + ], + "score": 1.0, + "content": "erence sentences. In addition, we show that some of the currently available cor-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 385, + 470, + 398 + ], + "spans": [ + { + "bbox": [ + 141, + 385, + 470, + 398 + ], + "score": 1.0, + "content": "pora for task-oriented language generation can be solved with simple models and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 397, + 302, + 408 + ], + "spans": [ + { + "bbox": [ + 141, + 397, + 302, + 408 + ], + "score": 1.0, + "content": "advocate for more challenging datasets.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 13.5, + "bbox_fs": [ + 141, + 232, + 470, + 408 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 430, + 205, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 208, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 208, + 446 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 456, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 104, + 455, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 104, + 455, + 506, + 470 + ], + "score": 1.0, + "content": "Rule-based and template-based dialogue response generation systems have been around for a long", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "score": 1.0, + "content": "time (Axelrod, 2000; Elhadad, 1992). Even today, many task-oriented dialogue systems deployed in", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 478, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 104, + 478, + 506, + 492 + ], + "score": 1.0, + "content": "production are rule-based and template-based. These systems do not scale with increasing domain", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "complexity and maintaining the increasing number of templates becomes cumbersome. In the past,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 500, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 514 + ], + "score": 1.0, + "content": "Oh & Rudnicky (2000) proposed a corpus-based approach for Natural Language Generation (NLG)", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "for task-oriented dialogue systems. Other statistical approaches were proposed using tree-based", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 521, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 104, + 521, + 506, + 537 + ], + "score": 1.0, + "content": "models and reinforcement learning (Walker et al., 2007; Rieser & Lemon, 2009). Recently, deep-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 532, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 104, + 532, + 506, + 547 + ], + "score": 1.0, + "content": "learning based approaches (Wen et al., 2015b; Sharma et al., 2017; Lowe et al., 2015; Serban et al.,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 545, + 388, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 388, + 558 + ], + "score": 1.0, + "content": "2016) have shown promising results for dialogue response generation.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27, + "bbox_fs": [ + 104, + 455, + 506, + 558 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "The automated evaluation of machine-generated language is challenging and an important problem", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "score": 1.0, + "content": "for the natural language processing community. The most widely used automated metrics currently", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 582, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 104, + 582, + 506, + 596 + ], + "score": 1.0, + "content": "are word-overlap based metrics such as BLEU (Papineni et al., 2002), METEOR (Banerjee & Lavie,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 593, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 506, + 606 + ], + "score": 1.0, + "content": "2005) which were proposed originally for machine translation. While these metrics were shown to", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "correlate well with manual human evaluation in machine translation tasks, previous studies showed", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "that this is not the case in non-task oriented dialogue (Liu et al., 2016). This is explained by the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 104, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "fact that for the same context (e.g. a user utterance), responses in dialogue have more diversity.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "Word-overlap metrics are unable to capture semantics and thus, can lead to poor scores even for", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "appropriate responses. Human evaluation in this case is the most reliable metric. However, human", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 660, + 394, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 394, + 672 + ], + "score": 1.0, + "content": "judgments are expensive to obtain and not readily available at all times.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 36.5, + "bbox_fs": [ + 104, + 560, + 506, + 672 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 504, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 504, + 689 + ], + "score": 1.0, + "content": "Task-oriented dialogue systems are employed in narrower domains (e.g. booking a restaurant) and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 687, + 504, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 504, + 699 + ], + "score": 1.0, + "content": "responses do not have as much diversity as in the non-task oriented setting. Another important", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "difference is that in the non-task oriented setting, response generation is often performed end-to-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "end, which means that the model takes as input the last user utterance and potentially the dialogue", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "history and it outputs the next system answer. In the task-oriented setting, on the other hand, the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 83, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 506, + 95 + ], + "score": 1.0, + "content": "language generation task is often seen as a translation step from an abstract representation of a", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "sentence to the sentence itself. As a consequence, automated metrics which compare a generated", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "sentence to a reference sentence might be more appropriate and correlate with human judgments. In", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 168, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 168, + 129 + ], + "score": 1.0, + "content": "this paper, we:", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 677, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 506, + 95 + ], + "score": 1.0, + "content": "language generation task is often seen as a translation step from an abstract representation of a", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "sentence to the sentence itself. As a consequence, automated metrics which compare a generated", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "sentence to a reference sentence might be more appropriate and correlate with human judgments. In", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 168, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 168, + 129 + ], + "score": 1.0, + "content": "this paper, we:", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 132, + 134, + 504, + 172 + ], + "lines": [ + { + "bbox": [ + 132, + 134, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 134, + 505, + 147 + ], + "score": 1.0, + "content": "• study the correlation between human judgments and several unsupervised automated met-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 140, + 146, + 350, + 159 + ], + "spans": [ + { + "bbox": [ + 140, + 146, + 350, + 159 + ], + "score": 1.0, + "content": "rics on two popular task-oriented dialogue datasets,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 160, + 484, + 173 + ], + "spans": [ + { + "bbox": [ + 131, + 160, + 484, + 173 + ], + "score": 1.0, + "content": "• introduce variants of existing models and evaluate their performance on these metrics", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 108, + 180, + 503, + 213 + ], + "lines": [ + { + "bbox": [ + 106, + 179, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 106, + 179, + 505, + 192 + ], + "score": 1.0, + "content": "We find that the automated metrics have stronger correlation with human judgments in the task-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 191, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 505, + 202 + ], + "score": 1.0, + "content": "oriented setting than what has been observed in the non task-oriented setting. We also observe that", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 201, + 434, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 434, + 214 + ], + "score": 1.0, + "content": "these metrics correlate even more in the presence of multiple reference sentences.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "title", + "bbox": [ + 108, + 228, + 211, + 241 + ], + "lines": [ + { + "bbox": [ + 104, + 227, + 213, + 244 + ], + "spans": [ + { + "bbox": [ + 104, + 227, + 213, + 244 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 253, + 504, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 253, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 264 + ], + "score": 1.0, + "content": "Liu et al. (2016) did an empirical study to evaluate the correlation between human scores and sev-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 264, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 276 + ], + "score": 1.0, + "content": "eral automated word-overlap metrics as well as embedding-based metrics for dialogue response", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 275, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 287 + ], + "score": 1.0, + "content": "generation. They observed that these metrics, though widely used in the literature, had only weak", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 285, + 434, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 434, + 299 + ], + "score": 1.0, + "content": "correlation with human judgments in the non task-oriented dialogue NLG setting.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 505, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "In terms of supervised NLG evaluation metrics, Lowe et al. (2017) proposed the ADEM model", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "which trains a hierarchical recurrent neural network in a supervised manner to predict human-like", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "score": 1.0, + "content": "scores. This learned score was shown to correlate better with human judgments than any other", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "score": 1.0, + "content": "automated metric. However, the drawback of this approach is the requirement for expensive human", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 345, + 140, + 361 + ], + "spans": [ + { + "bbox": [ + 104, + 345, + 140, + 361 + ], + "score": 1.0, + "content": "ratings.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 108, + 363, + 505, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 362, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 376 + ], + "score": 1.0, + "content": "Li et al. (2016) proposed to use reinforcement learning to train an end-to-end dialogue system. They", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "score": 1.0, + "content": "simulate a dialogue between two agents and use a policy gradient algorithm with a reward function", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 385, + 452, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 452, + 398 + ], + "score": 1.0, + "content": "which evaluates specific properties of the responses generated by the dialogue system.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 402, + 505, + 490 + ], + "lines": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "In the adversarial setting, Kannan & Vinyals (2016) train a recurrent neural network discriminator to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 412, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 506, + 425 + ], + "score": 1.0, + "content": "differentiate human-generated responses from model-generated responses. However, an extensive", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "analysis of the viability and the ease of standardization of this approach is yet to be conducted. Li", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "et al. (2017), apart from adversarially training dialogue response models, propose an independent", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 445, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 459 + ], + "score": 1.0, + "content": "adversarial evaluation metric AdverSuc and a measure of the model’s reliability called evaluator", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 456, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 470 + ], + "score": 1.0, + "content": "reliability error. Drawbacks of these approaches are that they are model-dependent. Adversarial", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "methods might be promising for task-oriented dialogue systems but more research needs to be con-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 478, + 204, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 204, + 491 + ], + "score": 1.0, + "content": "ducted on their account.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 496, + 505, + 562 + ], + "lines": [ + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "Most of the work described so far has been done in the non task-oriented dialogue setting as there", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "has been prior work indicating that automated metrics do not correlate well with humans in that", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 518, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 505, + 530 + ], + "score": 1.0, + "content": "setting. There has not yet been any empirical validation that these conclusions also apply to the task", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "oriented setting. Research in the task oriented setting has mostly made use of automated metrics", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "such as BLEU and human evaluation (Wen et al., 2015b; Sharma et al., 2017; Dusek & Jurcicek, ˇ", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 549, + 136, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 136, + 564 + ], + "score": 1.0, + "content": "2016).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + }, + { + "type": "title", + "bbox": [ + 107, + 577, + 174, + 590 + ], + "lines": [ + { + "bbox": [ + 104, + 575, + 176, + 593 + ], + "spans": [ + { + "bbox": [ + 104, + 575, + 176, + 593 + ], + "score": 1.0, + "content": "3 METRICS", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 601, + 505, + 657 + ], + "lines": [ + { + "bbox": [ + 106, + 601, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 505, + 614 + ], + "score": 1.0, + "content": "This section describes the set of automatic metrics whose correlation with human evaluation is", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 613, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 505, + 624 + ], + "score": 1.0, + "content": "studied. We consider first word-overlap metrics and then embedding-based metrics. In all that", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 623, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 505, + 636 + ], + "score": 1.0, + "content": "follows, when multiple references are provided, we compute the similarity between the prediction", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 634, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 505, + 648 + ], + "score": 1.0, + "content": "and all the references one-by-one, and then select the maximum value. We then average the scores", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 645, + 205, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 205, + 659 + ], + "score": 1.0, + "content": "across the entire corpus.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + }, + { + "type": "title", + "bbox": [ + 108, + 670, + 274, + 681 + ], + "lines": [ + { + "bbox": [ + 106, + 669, + 276, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 276, + 682 + ], + "score": 1.0, + "content": "3.1 WORD-OVERLAP BASED METRICS", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "title", + "bbox": [ + 108, + 690, + 169, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 690, + 169, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 169, + 703 + ], + "score": 1.0, + "content": "3.1.1 BLEU", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 318, + 722 + ], + "score": 1.0, + "content": "The BLEU metric (Papineni et al., 2002) compares", + "type": "text" + }, + { + "bbox": [ + 319, + 714, + 325, + 720 + ], + "score": 0.8, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "-grams between the candidate utterance and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "the reference utterance. The BLEU score is computed at the corpus-level and relies on the following", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 505, + 127 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 106, + 83, + 506, + 129 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 132, + 134, + 504, + 172 + ], + "lines": [ + { + "bbox": [ + 132, + 134, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 134, + 505, + 147 + ], + "score": 1.0, + "content": "• study the correlation between human judgments and several unsupervised automated met-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 140, + 146, + 350, + 159 + ], + "spans": [ + { + "bbox": [ + 140, + 146, + 350, + 159 + ], + "score": 1.0, + "content": "rics on two popular task-oriented dialogue datasets,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 160, + 484, + 173 + ], + "spans": [ + { + "bbox": [ + 131, + 160, + 484, + 173 + ], + "score": 1.0, + "content": "• introduce variants of existing models and evaluate their performance on these metrics", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 131, + 134, + 505, + 173 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 180, + 503, + 213 + ], + "lines": [ + { + "bbox": [ + 106, + 179, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 106, + 179, + 505, + 192 + ], + "score": 1.0, + "content": "We find that the automated metrics have stronger correlation with human judgments in the task-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 191, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 505, + 202 + ], + "score": 1.0, + "content": "oriented setting than what has been observed in the non task-oriented setting. We also observe that", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 201, + 434, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 434, + 214 + ], + "score": 1.0, + "content": "these metrics correlate even more in the presence of multiple reference sentences.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 106, + 179, + 505, + 214 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 228, + 211, + 241 + ], + "lines": [ + { + "bbox": [ + 104, + 227, + 213, + 244 + ], + "spans": [ + { + "bbox": [ + 104, + 227, + 213, + 244 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 253, + 504, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 253, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 264 + ], + "score": 1.0, + "content": "Liu et al. (2016) did an empirical study to evaluate the correlation between human scores and sev-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 264, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 276 + ], + "score": 1.0, + "content": "eral automated word-overlap metrics as well as embedding-based metrics for dialogue response", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 275, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 287 + ], + "score": 1.0, + "content": "generation. They observed that these metrics, though widely used in the literature, had only weak", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 285, + 434, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 434, + 299 + ], + "score": 1.0, + "content": "correlation with human judgments in the non task-oriented dialogue NLG setting.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 253, + 506, + 299 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 505, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "In terms of supervised NLG evaluation metrics, Lowe et al. (2017) proposed the ADEM model", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "which trains a hierarchical recurrent neural network in a supervised manner to predict human-like", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "score": 1.0, + "content": "scores. This learned score was shown to correlate better with human judgments than any other", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "score": 1.0, + "content": "automated metric. However, the drawback of this approach is the requirement for expensive human", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 345, + 140, + 361 + ], + "spans": [ + { + "bbox": [ + 104, + 345, + 140, + 361 + ], + "score": 1.0, + "content": "ratings.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17, + "bbox_fs": [ + 104, + 303, + 505, + 361 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 363, + 505, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 362, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 376 + ], + "score": 1.0, + "content": "Li et al. (2016) proposed to use reinforcement learning to train an end-to-end dialogue system. They", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "score": 1.0, + "content": "simulate a dialogue between two agents and use a policy gradient algorithm with a reward function", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 385, + 452, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 452, + 398 + ], + "score": 1.0, + "content": "which evaluates specific properties of the responses generated by the dialogue system.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 362, + 505, + 398 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 402, + 505, + 490 + ], + "lines": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "In the adversarial setting, Kannan & Vinyals (2016) train a recurrent neural network discriminator to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 412, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 506, + 425 + ], + "score": 1.0, + "content": "differentiate human-generated responses from model-generated responses. However, an extensive", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "analysis of the viability and the ease of standardization of this approach is yet to be conducted. Li", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "et al. (2017), apart from adversarially training dialogue response models, propose an independent", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 445, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 459 + ], + "score": 1.0, + "content": "adversarial evaluation metric AdverSuc and a measure of the model’s reliability called evaluator", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 456, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 470 + ], + "score": 1.0, + "content": "reliability error. Drawbacks of these approaches are that they are model-dependent. Adversarial", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "methods might be promising for task-oriented dialogue systems but more research needs to be con-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 478, + 204, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 204, + 491 + ], + "score": 1.0, + "content": "ducted on their account.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 402, + 506, + 491 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 496, + 505, + 562 + ], + "lines": [ + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "Most of the work described so far has been done in the non task-oriented dialogue setting as there", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "has been prior work indicating that automated metrics do not correlate well with humans in that", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 518, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 505, + 530 + ], + "score": 1.0, + "content": "setting. There has not yet been any empirical validation that these conclusions also apply to the task", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "oriented setting. Research in the task oriented setting has mostly made use of automated metrics", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "such as BLEU and human evaluation (Wen et al., 2015b; Sharma et al., 2017; Dusek & Jurcicek, ˇ", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 549, + 136, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 136, + 564 + ], + "score": 1.0, + "content": "2016).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 495, + 505, + 564 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 577, + 174, + 590 + ], + "lines": [ + { + "bbox": [ + 104, + 575, + 176, + 593 + ], + "spans": [ + { + "bbox": [ + 104, + 575, + 176, + 593 + ], + "score": 1.0, + "content": "3 METRICS", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 601, + 505, + 657 + ], + "lines": [ + { + "bbox": [ + 106, + 601, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 505, + 614 + ], + "score": 1.0, + "content": "This section describes the set of automatic metrics whose correlation with human evaluation is", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 613, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 505, + 624 + ], + "score": 1.0, + "content": "studied. We consider first word-overlap metrics and then embedding-based metrics. In all that", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 623, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 505, + 636 + ], + "score": 1.0, + "content": "follows, when multiple references are provided, we compute the similarity between the prediction", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 634, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 505, + 648 + ], + "score": 1.0, + "content": "and all the references one-by-one, and then select the maximum value. We then average the scores", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 645, + 205, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 205, + 659 + ], + "score": 1.0, + "content": "across the entire corpus.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 601, + 505, + 659 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 670, + 274, + 681 + ], + "lines": [ + { + "bbox": [ + 106, + 669, + 276, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 276, + 682 + ], + "score": 1.0, + "content": "3.1 WORD-OVERLAP BASED METRICS", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "title", + "bbox": [ + 108, + 690, + 169, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 690, + 169, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 169, + 703 + ], + "score": 1.0, + "content": "3.1.1 BLEU", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 318, + 722 + ], + "score": 1.0, + "content": "The BLEU metric (Papineni et al., 2002) compares", + "type": "text" + }, + { + "bbox": [ + 319, + 714, + 325, + 720 + ], + "score": 0.8, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "-grams between the candidate utterance and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "the reference utterance. The BLEU score is computed at the corpus-level and relies on the following", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 185, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 187, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 187, + 95 + ], + "score": 1.0, + "content": "modified precision:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 193, + 80, + 417, + 138 + ], + "lines": [ + { + "bbox": [ + 193, + 80, + 417, + 138 + ], + "spans": [ + { + "bbox": [ + 193, + 80, + 417, + 138 + ], + "score": 0.94, + "content": "p _ { n } = \\frac { \\displaystyle \\sum _ { C \\in \\{ C a n d i d a t e s \\} } \\sum _ { n - g r a m \\in C } C t _ { c l i p } ( n - g r a m ) } { \\displaystyle \\sum _ { C ^ { \\prime } \\in \\{ C a n d i d a t e s ^ { \\prime } \\} } \\sum _ { n - g r a m ^ { \\prime } \\in C ^ { \\prime } } C t _ { c l i p } ( n - g r a m ^ { \\prime } ) }", + "type": "interline_equation", + "image_path": "d32f40b1ba7b2dc150402592d5e78756627a8322215e60ec96ffe2cc4f87826c.jpg" + } + ] + } + ], + "index": 2.5, + "virtual_lines": [ + { + "bbox": [ + 193, + 80, + 417, + 94.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 193, + 94.5, + 417, + 109.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 193, + 109.0, + 417, + 123.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 193, + 123.5, + 417, + 138.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 140, + 503, + 184 + ], + "lines": [ + { + "bbox": [ + 106, + 140, + 506, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 419, + 154 + ], + "score": 1.0, + "content": "where {Candidates} are the candidate answers generated by the model and", + "type": "text" + }, + { + "bbox": [ + 419, + 141, + 445, + 153 + ], + "score": 0.92, + "content": "C t _ { c l i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 140, + 506, + 154 + ], + "score": 1.0, + "content": "is the clipped", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 152, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 152, + 160, + 163 + ], + "score": 1.0, + "content": "count for the", + "type": "text" + }, + { + "bbox": [ + 161, + 154, + 168, + 161 + ], + "score": 0.76, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 152, + 328, + 163 + ], + "score": 1.0, + "content": "-gram which is the number of times the", + "type": "text" + }, + { + "bbox": [ + 328, + 153, + 335, + 161 + ], + "score": 0.79, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 152, + 505, + 163 + ], + "score": 1.0, + "content": "-gram is common to the candidate answer", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 163, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 445, + 174 + ], + "score": 1.0, + "content": "and the reference answer clipped by the maximum number of occurrences of the", + "type": "text" + }, + { + "bbox": [ + 445, + 164, + 452, + 172 + ], + "score": 0.76, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 163, + 505, + 174 + ], + "score": 1.0, + "content": "-gram in the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 173, + 316, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 316, + 185 + ], + "score": 1.0, + "content": "reference answer. The BLEU-N score is defined as:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "interline_equation", + "bbox": [ + 229, + 178, + 381, + 212 + ], + "lines": [ + { + "bbox": [ + 229, + 178, + 381, + 212 + ], + "spans": [ + { + "bbox": [ + 229, + 178, + 381, + 212 + ], + "score": 0.93, + "content": "\\begin{array} { l } { { \\displaystyle \\forall \\mathrm { ~ s c u r e ~ 1 s ~ u c t 1 1 1 e u ~ a s . } } } \\\\ { { \\displaystyle \\mathsf { B L E U - N } = \\mathsf { B P } \\exp ( \\sum _ { n } ^ { N } \\omega _ { n } \\log ( p _ { n } ) ) } } \\end{array}", + "type": "interline_equation", + "image_path": "ca3d22fd182470d656e2c0e4bfc007d7b99cc9009ca2a43f3d9f26749f273029.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 229, + 178, + 381, + 195.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 229, + 195.0, + 381, + 212.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 212, + 505, + 256 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 504, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 133, + 225 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 213, + 144, + 222 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 212, + 263, + 225 + ], + "score": 1.0, + "content": "is the maximum length of the", + "type": "text" + }, + { + "bbox": [ + 263, + 214, + 271, + 222 + ], + "score": 0.79, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 212, + 496, + 225 + ], + "score": 1.0, + "content": "-grams (in this paper, we compute BLEU-1 to BLEU-4),", + "type": "text" + }, + { + "bbox": [ + 496, + 214, + 504, + 222 + ], + "score": 0.67, + "content": "\\omega", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 223, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 236 + ], + "score": 1.0, + "content": "is a weighting that is often uniform and BP is a brevity penalty. In this paper we report the BLEU", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 235, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 505, + 246 + ], + "score": 1.0, + "content": "score at the corpus level but we also compute this score at the sentence level to analyze its correlation", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 245, + 202, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 202, + 257 + ], + "score": 1.0, + "content": "with human evaluation.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 107, + 267, + 184, + 279 + ], + "lines": [ + { + "bbox": [ + 105, + 266, + 186, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 186, + 281 + ], + "score": 1.0, + "content": "3.1.2 METEOR", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 286, + 505, + 353 + ], + "lines": [ + { + "bbox": [ + 105, + 286, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 300 + ], + "score": 1.0, + "content": "The METEOR metric (Banerjee & Lavie, 2005) was proposed as a metric which correlates better", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 299, + 505, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 310 + ], + "score": 1.0, + "content": "at the sentence level with human evaluation. To compute the METEOR score, first, an alignment", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "between the candidate and the reference sentences is created by mapping each unigram in the can-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 320, + 504, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 504, + 332 + ], + "score": 1.0, + "content": "didate sentence to 0 or 1 unigram in the reference sentence. The alignment is not only based on", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "exact matches but also stem, synonym, and paraphrase matches. Based on this alignment, unigram", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 343, + 354, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 354, + 353 + ], + "score": 1.0, + "content": "precision and recall are computed and the METEOR score is:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5 + }, + { + "type": "interline_equation", + "bbox": [ + 248, + 355, + 362, + 369 + ], + "lines": [ + { + "bbox": [ + 248, + 355, + 362, + 369 + ], + "spans": [ + { + "bbox": [ + 248, + 355, + 362, + 369 + ], + "score": 0.89, + "content": "\\mathbf { M E T E O R } = F _ { m e a n } ( 1 - p )", + "type": "interline_equation", + "image_path": "790b2f6cd10cf8720ca62ace23883c3da41ec9fb6871fd7e5a7b5e241ce282ef.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 248, + 355, + 362, + 369 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 370, + 504, + 393 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 133, + 384 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 371, + 162, + 382 + ], + "score": 0.92, + "content": "F _ { m e a n }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 370, + 505, + 384 + ], + "score": 1.0, + "content": "is the harmonic mean between precision and recall with the weight for recall 9 times", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 381, + 318, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 261, + 394 + ], + "score": 1.0, + "content": "a high as the weight for precision, and", + "type": "text" + }, + { + "bbox": [ + 261, + 383, + 268, + 393 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 381, + 318, + 394 + ], + "score": 1.0, + "content": "is a penalty.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "title", + "bbox": [ + 107, + 404, + 177, + 415 + ], + "lines": [ + { + "bbox": [ + 106, + 404, + 178, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 178, + 417 + ], + "score": 1.0, + "content": "3.1.3 ROUGE", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 505, + 457 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 505, + 436 + ], + "score": 1.0, + "content": "ROUGE (Lin, 2004) is a set of metrics that was first introduced for summarization. We compute", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "ROUGE-L which is an F-measure based on the Longest Common Subsequence (LCS) between the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 445, + 249, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 249, + 458 + ], + "score": 1.0, + "content": "candidate and reference utterances.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 108, + 469, + 257, + 481 + ], + "lines": [ + { + "bbox": [ + 105, + 469, + 258, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 258, + 482 + ], + "score": 1.0, + "content": "3.2 EMBEDDING BASED METRICS", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 490, + 505, + 513 + ], + "lines": [ + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "We consider another set of metrics which compute the cosine similarity between the embeddings of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 500, + 415, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 415, + 514 + ], + "score": 1.0, + "content": "the predicted and the reference sentence instead of relying on word overlaps.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "title", + "bbox": [ + 107, + 524, + 208, + 536 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 209, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 209, + 536 + ], + "score": 1.0, + "content": "3.2.1 SKIP-THOUGHT", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 542, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 542, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 556 + ], + "score": 1.0, + "content": "The Skip-Thought model (Kiros et al., 2015) is trained in an unsupervised fashion and uses a re-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 554, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 567 + ], + "score": 1.0, + "content": "current network to encode a given sentence into an embedding and then decode it to predict the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "preceding and following sentences. The model was trained on the BookCorpus dataset (Zhu et al.,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 575, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 505, + 589 + ], + "score": 1.0, + "content": "2015). The embeddings produced by the encoder have a robust performance on semantic relatedness", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 587, + 419, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 419, + 599 + ], + "score": 1.0, + "content": "tasks. We use the pre-trained Skip-Thought encoder provided by the authors1.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 616 + ], + "score": 1.0, + "content": "We also compute other embedding-based methods which have been used as evaluation metrics for", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "measuring human correlation in recent literature (Liu et al., 2016) for non task-oriented dialogue in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 626, + 235, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 235, + 637 + ], + "score": 1.0, + "content": "Sections 3.2.2, 3.2.3, and 3.2.4.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "title", + "bbox": [ + 107, + 649, + 237, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 648, + 237, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 237, + 662 + ], + "score": 1.0, + "content": "3.2.2 EMBEDDING AVERAGE", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 105, + 668, + 504, + 689 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 681 + ], + "score": 1.0, + "content": "This metric computes a sentence-level embedding by averaging the embeddings of the words com-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 679, + 192, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 192, + 692 + ], + "score": 1.0, + "content": "posing this sentence:", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + }, + { + "type": "interline_equation", + "bbox": [ + 262, + 686, + 348, + 715 + ], + "lines": [ + { + "bbox": [ + 262, + 686, + 348, + 715 + ], + "spans": [ + { + "bbox": [ + 262, + 686, + 348, + 715 + ], + "score": 0.94, + "content": "\\bar { e } _ { C } = \\frac { \\sum _ { w \\in C } e _ { w } } { \\lvert \\sum _ { w ^ { \\prime } \\in C } e _ { w ^ { \\prime } } \\rvert } .", + "type": "interline_equation", + "image_path": "de135268afb6f0d9437350ee94d7386b11e424c2ef5bbbb687ca2870eacc206c.jpg" + } + ] + } + ], + "index": 44.5, + "virtual_lines": [ + { + "bbox": [ + 262, + 686, + 348, + 700.5 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 262, + 700.5, + 348, + 715.0 + ], + "spans": [], + "index": 45 + } + ] + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 120, + 721, + 277, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 719, + 278, + 734 + ], + "spans": [ + { + "bbox": [ + 119, + 719, + 278, + 734 + ], + "score": 1.0, + "content": "1https://github.com/ryankiros/skip-thoughts", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 185, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 187, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 187, + 95 + ], + "score": 1.0, + "content": "modified precision:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 82, + 187, + 95 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 193, + 80, + 417, + 138 + ], + "lines": [ + { + "bbox": [ + 193, + 80, + 417, + 138 + ], + "spans": [ + { + "bbox": [ + 193, + 80, + 417, + 138 + ], + "score": 0.94, + "content": "p _ { n } = \\frac { \\displaystyle \\sum _ { C \\in \\{ C a n d i d a t e s \\} } \\sum _ { n - g r a m \\in C } C t _ { c l i p } ( n - g r a m ) } { \\displaystyle \\sum _ { C ^ { \\prime } \\in \\{ C a n d i d a t e s ^ { \\prime } \\} } \\sum _ { n - g r a m ^ { \\prime } \\in C ^ { \\prime } } C t _ { c l i p } ( n - g r a m ^ { \\prime } ) }", + "type": "interline_equation", + "image_path": "d32f40b1ba7b2dc150402592d5e78756627a8322215e60ec96ffe2cc4f87826c.jpg" + } + ] + } + ], + "index": 2.5, + "virtual_lines": [ + { + "bbox": [ + 193, + 80, + 417, + 94.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 193, + 94.5, + 417, + 109.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 193, + 109.0, + 417, + 123.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 193, + 123.5, + 417, + 138.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 140, + 503, + 184 + ], + "lines": [ + { + "bbox": [ + 106, + 140, + 506, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 419, + 154 + ], + "score": 1.0, + "content": "where {Candidates} are the candidate answers generated by the model and", + "type": "text" + }, + { + "bbox": [ + 419, + 141, + 445, + 153 + ], + "score": 0.92, + "content": "C t _ { c l i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 140, + 506, + 154 + ], + "score": 1.0, + "content": "is the clipped", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 152, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 152, + 160, + 163 + ], + "score": 1.0, + "content": "count for the", + "type": "text" + }, + { + "bbox": [ + 161, + 154, + 168, + 161 + ], + "score": 0.76, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 152, + 328, + 163 + ], + "score": 1.0, + "content": "-gram which is the number of times the", + "type": "text" + }, + { + "bbox": [ + 328, + 153, + 335, + 161 + ], + "score": 0.79, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 152, + 505, + 163 + ], + "score": 1.0, + "content": "-gram is common to the candidate answer", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 163, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 445, + 174 + ], + "score": 1.0, + "content": "and the reference answer clipped by the maximum number of occurrences of the", + "type": "text" + }, + { + "bbox": [ + 445, + 164, + 452, + 172 + ], + "score": 0.76, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 163, + 505, + 174 + ], + "score": 1.0, + "content": "-gram in the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 173, + 316, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 316, + 185 + ], + "score": 1.0, + "content": "reference answer. The BLEU-N score is defined as:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 106, + 140, + 506, + 185 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 229, + 178, + 381, + 212 + ], + "lines": [ + { + "bbox": [ + 229, + 178, + 381, + 212 + ], + "spans": [ + { + "bbox": [ + 229, + 178, + 381, + 212 + ], + "score": 0.93, + "content": "\\begin{array} { l } { { \\displaystyle \\forall \\mathrm { ~ s c u r e ~ 1 s ~ u c t 1 1 1 e u ~ a s . } } } \\\\ { { \\displaystyle \\mathsf { B L E U - N } = \\mathsf { B P } \\exp ( \\sum _ { n } ^ { N } \\omega _ { n } \\log ( p _ { n } ) ) } } \\end{array}", + "type": "interline_equation", + "image_path": "ca3d22fd182470d656e2c0e4bfc007d7b99cc9009ca2a43f3d9f26749f273029.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 229, + 178, + 381, + 195.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 229, + 195.0, + 381, + 212.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 212, + 505, + 256 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 504, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 133, + 225 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 213, + 144, + 222 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 212, + 263, + 225 + ], + "score": 1.0, + "content": "is the maximum length of the", + "type": "text" + }, + { + "bbox": [ + 263, + 214, + 271, + 222 + ], + "score": 0.79, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 212, + 496, + 225 + ], + "score": 1.0, + "content": "-grams (in this paper, we compute BLEU-1 to BLEU-4),", + "type": "text" + }, + { + "bbox": [ + 496, + 214, + 504, + 222 + ], + "score": 0.67, + "content": "\\omega", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 223, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 236 + ], + "score": 1.0, + "content": "is a weighting that is often uniform and BP is a brevity penalty. In this paper we report the BLEU", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 235, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 505, + 246 + ], + "score": 1.0, + "content": "score at the corpus level but we also compute this score at the sentence level to analyze its correlation", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 245, + 202, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 202, + 257 + ], + "score": 1.0, + "content": "with human evaluation.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 212, + 505, + 257 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 267, + 184, + 279 + ], + "lines": [ + { + "bbox": [ + 105, + 266, + 186, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 186, + 281 + ], + "score": 1.0, + "content": "3.1.2 METEOR", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 286, + 505, + 353 + ], + "lines": [ + { + "bbox": [ + 105, + 286, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 300 + ], + "score": 1.0, + "content": "The METEOR metric (Banerjee & Lavie, 2005) was proposed as a metric which correlates better", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 299, + 505, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 310 + ], + "score": 1.0, + "content": "at the sentence level with human evaluation. To compute the METEOR score, first, an alignment", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "between the candidate and the reference sentences is created by mapping each unigram in the can-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 320, + 504, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 504, + 332 + ], + "score": 1.0, + "content": "didate sentence to 0 or 1 unigram in the reference sentence. The alignment is not only based on", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "exact matches but also stem, synonym, and paraphrase matches. Based on this alignment, unigram", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 343, + 354, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 354, + 353 + ], + "score": 1.0, + "content": "precision and recall are computed and the METEOR score is:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 286, + 506, + 353 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 248, + 355, + 362, + 369 + ], + "lines": [ + { + "bbox": [ + 248, + 355, + 362, + 369 + ], + "spans": [ + { + "bbox": [ + 248, + 355, + 362, + 369 + ], + "score": 0.89, + "content": "\\mathbf { M E T E O R } = F _ { m e a n } ( 1 - p )", + "type": "interline_equation", + "image_path": "790b2f6cd10cf8720ca62ace23883c3da41ec9fb6871fd7e5a7b5e241ce282ef.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 248, + 355, + 362, + 369 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 370, + 504, + 393 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 133, + 384 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 371, + 162, + 382 + ], + "score": 0.92, + "content": "F _ { m e a n }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 370, + 505, + 384 + ], + "score": 1.0, + "content": "is the harmonic mean between precision and recall with the weight for recall 9 times", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 381, + 318, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 261, + 394 + ], + "score": 1.0, + "content": "a high as the weight for precision, and", + "type": "text" + }, + { + "bbox": [ + 261, + 383, + 268, + 393 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 381, + 318, + 394 + ], + "score": 1.0, + "content": "is a penalty.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 370, + 505, + 394 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 404, + 177, + 415 + ], + "lines": [ + { + "bbox": [ + 106, + 404, + 178, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 178, + 417 + ], + "score": 1.0, + "content": "3.1.3 ROUGE", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 505, + 457 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 505, + 436 + ], + "score": 1.0, + "content": "ROUGE (Lin, 2004) is a set of metrics that was first introduced for summarization. We compute", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "ROUGE-L which is an F-measure based on the Longest Common Subsequence (LCS) between the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 445, + 249, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 249, + 458 + ], + "score": 1.0, + "content": "candidate and reference utterances.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 422, + 505, + 458 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 469, + 257, + 481 + ], + "lines": [ + { + "bbox": [ + 105, + 469, + 258, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 258, + 482 + ], + "score": 1.0, + "content": "3.2 EMBEDDING BASED METRICS", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 490, + 505, + 513 + ], + "lines": [ + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "We consider another set of metrics which compute the cosine similarity between the embeddings of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 500, + 415, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 415, + 514 + ], + "score": 1.0, + "content": "the predicted and the reference sentence instead of relying on word overlaps.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 490, + 505, + 514 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 524, + 208, + 536 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 209, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 209, + 536 + ], + "score": 1.0, + "content": "3.2.1 SKIP-THOUGHT", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 542, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 542, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 556 + ], + "score": 1.0, + "content": "The Skip-Thought model (Kiros et al., 2015) is trained in an unsupervised fashion and uses a re-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 554, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 567 + ], + "score": 1.0, + "content": "current network to encode a given sentence into an embedding and then decode it to predict the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "preceding and following sentences. The model was trained on the BookCorpus dataset (Zhu et al.,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 575, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 505, + 589 + ], + "score": 1.0, + "content": "2015). The embeddings produced by the encoder have a robust performance on semantic relatedness", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 587, + 419, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 419, + 599 + ], + "score": 1.0, + "content": "tasks. We use the pre-trained Skip-Thought encoder provided by the authors1.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 542, + 505, + 599 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 616 + ], + "score": 1.0, + "content": "We also compute other embedding-based methods which have been used as evaluation metrics for", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "measuring human correlation in recent literature (Liu et al., 2016) for non task-oriented dialogue in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 626, + 235, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 235, + 637 + ], + "score": 1.0, + "content": "Sections 3.2.2, 3.2.3, and 3.2.4.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 604, + 505, + 637 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 649, + 237, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 648, + 237, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 237, + 662 + ], + "score": 1.0, + "content": "3.2.2 EMBEDDING AVERAGE", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 105, + 668, + 504, + 689 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 681 + ], + "score": 1.0, + "content": "This metric computes a sentence-level embedding by averaging the embeddings of the words com-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 679, + 192, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 192, + 692 + ], + "score": 1.0, + "content": "posing this sentence:", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 667, + 505, + 692 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 262, + 686, + 348, + 715 + ], + "lines": [ + { + "bbox": [ + 262, + 686, + 348, + 715 + ], + "spans": [ + { + "bbox": [ + 262, + 686, + 348, + 715 + ], + "score": 0.94, + "content": "\\bar { e } _ { C } = \\frac { \\sum _ { w \\in C } e _ { w } } { \\lvert \\sum _ { w ^ { \\prime } \\in C } e _ { w ^ { \\prime } } \\rvert } .", + "type": "interline_equation", + "image_path": "de135268afb6f0d9437350ee94d7386b11e424c2ef5bbbb687ca2870eacc206c.jpg" + } + ] + } + ], + "index": 44.5, + "virtual_lines": [ + { + "bbox": [ + 262, + 686, + 348, + 700.5 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 262, + 700.5, + 348, + 715.0 + ], + "spans": [], + "index": 45 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 103, + 82, + 478, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 478, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 219, + 96 + ], + "score": 1.0, + "content": "In this equation, the vectors", + "type": "text" + }, + { + "bbox": [ + 219, + 85, + 231, + 93 + ], + "score": 0.86, + "content": "e _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 81, + 353, + 96 + ], + "score": 1.0, + "content": "are embeddings for the words", + "type": "text" + }, + { + "bbox": [ + 354, + 85, + 362, + 92 + ], + "score": 0.73, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 81, + 466, + 96 + ], + "score": 1.0, + "content": "in the candidate sentence", + "type": "text" + }, + { + "bbox": [ + 466, + 83, + 475, + 92 + ], + "score": 0.79, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 81, + 478, + 96 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 105, + 221, + 117 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 222, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 222, + 118 + ], + "score": 1.0, + "content": "3.2.3 VECTOR EXTREMA", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 124, + 504, + 148 + ], + "lines": [ + { + "bbox": [ + 106, + 124, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 124, + 505, + 138 + ], + "score": 1.0, + "content": "Vector extrema (Forgues et al., 2014) computes a sentence-level embedding by taking the most", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 136, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 149 + ], + "score": 1.0, + "content": "extreme value of the embeddings of the words composing the sentence for each dimension of the", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "interline_equation", + "bbox": [ + 204, + 149, + 406, + 177 + ], + "lines": [ + { + "bbox": [ + 204, + 149, + 406, + 177 + ], + "spans": [ + { + "bbox": [ + 204, + 149, + 406, + 177 + ], + "score": 0.91, + "content": "e _ { r d } = \\left\\{ \\begin{array} { l l } { \\operatorname* { m a x } _ { w \\in C } e _ { w d } } & { \\mathrm { i f } e _ { w d } > \\vert \\operatorname* { m i n } _ { w ^ { \\prime } \\in C } e _ { w ^ { \\prime } d } \\vert } \\\\ { \\operatorname* { m i n } _ { w \\in C } e _ { w d } } & { \\mathrm { o t h e r w i s e } \\mathrm { . } } \\end{array} \\right.", + "type": "interline_equation", + "image_path": "30d01bb57faa794509b59b892e9176188487ae8ffa33c33096cc55efab4b0df8.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 204, + 149, + 406, + 163.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 204, + 163.0, + 406, + 177.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 181, + 504, + 194 + ], + "lines": [ + { + "bbox": [ + 108, + 181, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 108, + 181, + 174, + 195 + ], + "score": 1.0, + "content": "In this equation,", + "type": "text" + }, + { + "bbox": [ + 174, + 183, + 180, + 192 + ], + "score": 0.83, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 181, + 393, + 195 + ], + "score": 1.0, + "content": "is an index over the dimensions of the embedding and", + "type": "text" + }, + { + "bbox": [ + 393, + 182, + 402, + 192 + ], + "score": 0.85, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 181, + 505, + 195 + ], + "score": 1.0, + "content": "is the candidate sentence.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "title", + "bbox": [ + 108, + 199, + 225, + 211 + ], + "lines": [ + { + "bbox": [ + 106, + 198, + 226, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 226, + 212 + ], + "score": 1.0, + "content": "3.2.4 GREEDY MATCHING", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 218, + 503, + 241 + ], + "lines": [ + { + "bbox": [ + 106, + 218, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 505, + 232 + ], + "score": 1.0, + "content": "Greedy matching does not compute a sentence embedding but directly a similarity score between a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 229, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 147, + 243 + ], + "score": 1.0, + "content": "candidate", + "type": "text" + }, + { + "bbox": [ + 147, + 230, + 156, + 240 + ], + "score": 0.84, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 229, + 220, + 243 + ], + "score": 1.0, + "content": "and a reference", + "type": "text" + }, + { + "bbox": [ + 221, + 232, + 227, + 240 + ], + "score": 0.76, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 229, + 505, + 243 + ], + "score": 1.0, + "content": "(Rus & Lintean, 2012). This similarity score is computed as follows:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "interline_equation", + "bbox": [ + 206, + 244, + 405, + 294 + ], + "lines": [ + { + "bbox": [ + 206, + 244, + 405, + 294 + ], + "spans": [ + { + "bbox": [ + 206, + 244, + 405, + 294 + ], + "score": 0.93, + "content": "\\begin{array} { c } { \\displaystyle { G ( C , r ) = \\frac { \\sum w \\in C \\operatorname* { m a x } _ { \\hat { w } \\in r } c o s _ { - } s i m ( e _ { w } , w _ { \\hat { w } } ) } { | C | } } } \\\\ { \\displaystyle { G { M ( C , r ) } = \\frac { G ( C , r ) + G ( r , C ) } { 2 } . } } \\end{array}", + "type": "interline_equation", + "image_path": "0858831b0e7f88249c5bf3756d1aa9dd3d4bb4205809b4735f8f61b484b3b299.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 206, + 244, + 405, + 260.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 206, + 260.6666666666667, + 405, + 277.33333333333337 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 206, + 277.33333333333337, + 405, + 294.00000000000006 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 296, + 505, + 341 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 506, + 309 + ], + "score": 1.0, + "content": "Each word in the candidate sentence is greedily matched to a word in the reference sentence based on", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 308, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 506, + 320 + ], + "score": 1.0, + "content": "their embeddings’ cosine similarity. The score is an average of these similarities over the number of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "words in the candidate sentence. The same score is computed by reversing the roles of the candidate", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 330, + 463, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 463, + 342 + ], + "score": 1.0, + "content": "and reference sentences and the average of the two scores gives the final similarity score.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 107, + 356, + 300, + 369 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 300, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 300, + 371 + ], + "score": 1.0, + "content": "4 RESPONSE GENERATION MODELS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 381, + 505, + 437 + ], + "lines": [ + { + "bbox": [ + 106, + 380, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 505, + 393 + ], + "score": 1.0, + "content": "This section presents the different natural language generation models that we use in this study.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 390, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 505, + 406 + ], + "score": 1.0, + "content": "All of these models take as input a set of dialogue acts (Austin, 1962) potentially with slot types", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 495, + 416 + ], + "score": 1.0, + "content": "and slot values and translate this input into an utterance. An example input is inform(food", + "type": "text" + }, + { + "bbox": [ + 495, + 404, + 505, + 414 + ], + "score": 0.76, + "content": "=", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "score": 1.0, + "content": "Chinese) and a corresponding output would be “I am looking for a Chinese restaurant.”. In this", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 426, + 462, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 462, + 438 + ], + "score": 1.0, + "content": "example, the dialogue act is inform, the slot type is food, and the slot value is Chinese.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 107, + 450, + 172, + 461 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 173, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 173, + 463 + ], + "score": 1.0, + "content": "4.1 RANDOM", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 470, + 505, + 537 + ], + "lines": [ + { + "bbox": [ + 106, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "Given a dialogue act with one or more slot types, the random model finds all the examples in the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 481, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 106, + 481, + 505, + 494 + ], + "score": 1.0, + "content": "training set with the same dialogue act and slots (while ignoring slot values) and it randomly selects", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "its output from this set of reference sentences. The datasets that we experiment on have some", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 503, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 506, + 516 + ], + "score": 1.0, + "content": "special slot values such as “yes”, “no”, and “don’t care”. Since the model ignores all slot values,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 514, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 506, + 527 + ], + "score": 1.0, + "content": "these special cases are not properly handled, which results in slightly lower performance than what", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 525, + 479, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 479, + 538 + ], + "score": 1.0, + "content": "we could get by spending more time hand-engineering the model’s behavior for these values.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5 + }, + { + "type": "title", + "bbox": [ + 107, + 550, + 160, + 561 + ], + "lines": [ + { + "bbox": [ + 106, + 549, + 162, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 162, + 562 + ], + "score": 1.0, + "content": "4.2 LSTM", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 570, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 571, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 505, + 583 + ], + "score": 1.0, + "content": "This model consists of a recurrent LSTM (Hochreiter & Schmidhuber, 1997) decoder. The dialogue", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "score": 1.0, + "content": "acts and slot types are first encoded as a binary vector whose length is the number of possible", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "score": 1.0, + "content": "combinations of dialogue acts and slot types in the dataset. We refer to this binary vector as the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 603, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 163, + 616 + ], + "score": 1.0, + "content": "Dialogue Act", + "type": "text" + }, + { + "bbox": [ + 164, + 604, + 187, + 615 + ], + "score": 0.73, + "content": "( D A )", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 603, + 238, + 616 + ], + "score": 1.0, + "content": "vector. The", + "type": "text" + }, + { + "bbox": [ + 238, + 604, + 256, + 614 + ], + "score": 0.58, + "content": "D A", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 603, + 505, + 616 + ], + "score": 1.0, + "content": "vector for a given set of dialogue acts is a binary vector over", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 614, + 423, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 423, + 628 + ], + "score": 1.0, + "content": "the fused dialogue act-slot types, e.g., INFORM-FOOD, INFORM-COUNT, etc.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 631, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "This binary vector is given as input to the decoder at each time-step of the LSTM. The decoder then", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 642, + 504, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 504, + 654 + ], + "score": 1.0, + "content": "outputs a delexicalized sentence. A delexicalized sentence contains placeholders for the slot values.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 654, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 666 + ], + "score": 1.0, + "content": "An example is “I am looking for a FOOD restaurant.”. The values for the delexicalized slots (the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 664, + 382, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 382, + 678 + ], + "score": 1.0, + "content": "type of food in this example) are then directly copied from the input.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + }, + { + "type": "title", + "bbox": [ + 107, + 689, + 208, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 209, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 209, + 702 + ], + "score": 1.0, + "content": "4.3 DELEX-SC-LSTM", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "This model uses the same architecture as the LSTM model presented in the previous section except", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "that it uses sc-LSTM (Wen et al., 2015b) units in the decoder instead of LSTM units. We call this", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 103, + 82, + 478, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 478, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 219, + 96 + ], + "score": 1.0, + "content": "In this equation, the vectors", + "type": "text" + }, + { + "bbox": [ + 219, + 85, + 231, + 93 + ], + "score": 0.86, + "content": "e _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 81, + 353, + 96 + ], + "score": 1.0, + "content": "are embeddings for the words", + "type": "text" + }, + { + "bbox": [ + 354, + 85, + 362, + 92 + ], + "score": 0.73, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 81, + 466, + 96 + ], + "score": 1.0, + "content": "in the candidate sentence", + "type": "text" + }, + { + "bbox": [ + 466, + 83, + 475, + 92 + ], + "score": 0.79, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 81, + 478, + 96 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 81, + 478, + 96 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 105, + 221, + 117 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 222, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 222, + 118 + ], + "score": 1.0, + "content": "3.2.3 VECTOR EXTREMA", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 124, + 504, + 148 + ], + "lines": [ + { + "bbox": [ + 106, + 124, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 124, + 505, + 138 + ], + "score": 1.0, + "content": "Vector extrema (Forgues et al., 2014) computes a sentence-level embedding by taking the most", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 136, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 149 + ], + "score": 1.0, + "content": "extreme value of the embeddings of the words composing the sentence for each dimension of the", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 124, + 505, + 149 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 204, + 149, + 406, + 177 + ], + "lines": [ + { + "bbox": [ + 204, + 149, + 406, + 177 + ], + "spans": [ + { + "bbox": [ + 204, + 149, + 406, + 177 + ], + "score": 0.91, + "content": "e _ { r d } = \\left\\{ \\begin{array} { l l } { \\operatorname* { m a x } _ { w \\in C } e _ { w d } } & { \\mathrm { i f } e _ { w d } > \\vert \\operatorname* { m i n } _ { w ^ { \\prime } \\in C } e _ { w ^ { \\prime } d } \\vert } \\\\ { \\operatorname* { m i n } _ { w \\in C } e _ { w d } } & { \\mathrm { o t h e r w i s e } \\mathrm { . } } \\end{array} \\right.", + "type": "interline_equation", + "image_path": "30d01bb57faa794509b59b892e9176188487ae8ffa33c33096cc55efab4b0df8.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 204, + 149, + 406, + 163.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 204, + 163.0, + 406, + 177.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 181, + 504, + 194 + ], + "lines": [ + { + "bbox": [ + 108, + 181, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 108, + 181, + 174, + 195 + ], + "score": 1.0, + "content": "In this equation,", + "type": "text" + }, + { + "bbox": [ + 174, + 183, + 180, + 192 + ], + "score": 0.83, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 181, + 393, + 195 + ], + "score": 1.0, + "content": "is an index over the dimensions of the embedding and", + "type": "text" + }, + { + "bbox": [ + 393, + 182, + 402, + 192 + ], + "score": 0.85, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 181, + 505, + 195 + ], + "score": 1.0, + "content": "is the candidate sentence.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6, + "bbox_fs": [ + 108, + 181, + 505, + 195 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 199, + 225, + 211 + ], + "lines": [ + { + "bbox": [ + 106, + 198, + 226, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 226, + 212 + ], + "score": 1.0, + "content": "3.2.4 GREEDY MATCHING", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 218, + 503, + 241 + ], + "lines": [ + { + "bbox": [ + 106, + 218, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 505, + 232 + ], + "score": 1.0, + "content": "Greedy matching does not compute a sentence embedding but directly a similarity score between a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 229, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 147, + 243 + ], + "score": 1.0, + "content": "candidate", + "type": "text" + }, + { + "bbox": [ + 147, + 230, + 156, + 240 + ], + "score": 0.84, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 229, + 220, + 243 + ], + "score": 1.0, + "content": "and a reference", + "type": "text" + }, + { + "bbox": [ + 221, + 232, + 227, + 240 + ], + "score": 0.76, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 229, + 505, + 243 + ], + "score": 1.0, + "content": "(Rus & Lintean, 2012). This similarity score is computed as follows:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 218, + 505, + 243 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 206, + 244, + 405, + 294 + ], + "lines": [ + { + "bbox": [ + 206, + 244, + 405, + 294 + ], + "spans": [ + { + "bbox": [ + 206, + 244, + 405, + 294 + ], + "score": 0.93, + "content": "\\begin{array} { c } { \\displaystyle { G ( C , r ) = \\frac { \\sum w \\in C \\operatorname* { m a x } _ { \\hat { w } \\in r } c o s _ { - } s i m ( e _ { w } , w _ { \\hat { w } } ) } { | C | } } } \\\\ { \\displaystyle { G { M ( C , r ) } = \\frac { G ( C , r ) + G ( r , C ) } { 2 } . } } \\end{array}", + "type": "interline_equation", + "image_path": "0858831b0e7f88249c5bf3756d1aa9dd3d4bb4205809b4735f8f61b484b3b299.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 206, + 244, + 405, + 260.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 206, + 260.6666666666667, + 405, + 277.33333333333337 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 206, + 277.33333333333337, + 405, + 294.00000000000006 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 296, + 505, + 341 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 506, + 309 + ], + "score": 1.0, + "content": "Each word in the candidate sentence is greedily matched to a word in the reference sentence based on", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 308, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 506, + 320 + ], + "score": 1.0, + "content": "their embeddings’ cosine similarity. The score is an average of these similarities over the number of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "words in the candidate sentence. The same score is computed by reversing the roles of the candidate", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 330, + 463, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 463, + 342 + ], + "score": 1.0, + "content": "and reference sentences and the average of the two scores gives the final similarity score.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 295, + 506, + 342 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 356, + 300, + 369 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 300, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 300, + 371 + ], + "score": 1.0, + "content": "4 RESPONSE GENERATION MODELS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 381, + 505, + 437 + ], + "lines": [ + { + "bbox": [ + 106, + 380, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 505, + 393 + ], + "score": 1.0, + "content": "This section presents the different natural language generation models that we use in this study.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 390, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 505, + 406 + ], + "score": 1.0, + "content": "All of these models take as input a set of dialogue acts (Austin, 1962) potentially with slot types", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 495, + 416 + ], + "score": 1.0, + "content": "and slot values and translate this input into an utterance. An example input is inform(food", + "type": "text" + }, + { + "bbox": [ + 495, + 404, + 505, + 414 + ], + "score": 0.76, + "content": "=", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "score": 1.0, + "content": "Chinese) and a corresponding output would be “I am looking for a Chinese restaurant.”. In this", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 426, + 462, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 462, + 438 + ], + "score": 1.0, + "content": "example, the dialogue act is inform, the slot type is food, and the slot value is Chinese.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 380, + 506, + 438 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 450, + 172, + 461 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 173, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 173, + 463 + ], + "score": 1.0, + "content": "4.1 RANDOM", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 470, + 505, + 537 + ], + "lines": [ + { + "bbox": [ + 106, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "Given a dialogue act with one or more slot types, the random model finds all the examples in the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 481, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 106, + 481, + 505, + 494 + ], + "score": 1.0, + "content": "training set with the same dialogue act and slots (while ignoring slot values) and it randomly selects", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "its output from this set of reference sentences. The datasets that we experiment on have some", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 503, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 506, + 516 + ], + "score": 1.0, + "content": "special slot values such as “yes”, “no”, and “don’t care”. Since the model ignores all slot values,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 514, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 506, + 527 + ], + "score": 1.0, + "content": "these special cases are not properly handled, which results in slightly lower performance than what", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 525, + 479, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 479, + 538 + ], + "score": 1.0, + "content": "we could get by spending more time hand-engineering the model’s behavior for these values.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 470, + 506, + 538 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 550, + 160, + 561 + ], + "lines": [ + { + "bbox": [ + 106, + 549, + 162, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 162, + 562 + ], + "score": 1.0, + "content": "4.2 LSTM", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 570, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 571, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 505, + 583 + ], + "score": 1.0, + "content": "This model consists of a recurrent LSTM (Hochreiter & Schmidhuber, 1997) decoder. The dialogue", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "score": 1.0, + "content": "acts and slot types are first encoded as a binary vector whose length is the number of possible", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "score": 1.0, + "content": "combinations of dialogue acts and slot types in the dataset. We refer to this binary vector as the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 603, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 163, + 616 + ], + "score": 1.0, + "content": "Dialogue Act", + "type": "text" + }, + { + "bbox": [ + 164, + 604, + 187, + 615 + ], + "score": 0.73, + "content": "( D A )", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 603, + 238, + 616 + ], + "score": 1.0, + "content": "vector. The", + "type": "text" + }, + { + "bbox": [ + 238, + 604, + 256, + 614 + ], + "score": 0.58, + "content": "D A", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 603, + 505, + 616 + ], + "score": 1.0, + "content": "vector for a given set of dialogue acts is a binary vector over", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 614, + 423, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 423, + 628 + ], + "score": 1.0, + "content": "the fused dialogue act-slot types, e.g., INFORM-FOOD, INFORM-COUNT, etc.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 571, + 505, + 628 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 631, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "This binary vector is given as input to the decoder at each time-step of the LSTM. The decoder then", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 642, + 504, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 504, + 654 + ], + "score": 1.0, + "content": "outputs a delexicalized sentence. A delexicalized sentence contains placeholders for the slot values.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 654, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 666 + ], + "score": 1.0, + "content": "An example is “I am looking for a FOOD restaurant.”. The values for the delexicalized slots (the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 664, + 382, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 382, + 678 + ], + "score": 1.0, + "content": "type of food in this example) are then directly copied from the input.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 631, + 505, + 678 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 689, + 208, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 209, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 209, + 702 + ], + "score": 1.0, + "content": "4.3 DELEX-SC-LSTM", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "This model uses the same architecture as the LSTM model presented in the previous section except", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "that it uses sc-LSTM (Wen et al., 2015b) units in the decoder instead of LSTM units. We call this", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 106, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 370, + 95 + ], + "score": 1.0, + "content": "model the “delex-sc-LSTM”2. As in the previous model, the input", + "type": "text" + }, + { + "bbox": [ + 370, + 83, + 388, + 93 + ], + "score": 0.51, + "content": "D A", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "vector only encodes acts and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 411, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 411, + 106 + ], + "score": 1.0, + "content": "delexicalized slots. It does not contain any information about the slot value.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 504, + 143 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 247, + 123 + ], + "score": 1.0, + "content": "By providing this model the same", + "type": "text" + }, + { + "bbox": [ + 248, + 111, + 265, + 121 + ], + "score": 0.57, + "content": "D A", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 110, + 506, + 123 + ], + "score": 1.0, + "content": "vector input as the one given to the LSTM model, we can", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "score": 1.0, + "content": "directly study if the additional complexity of the sc-LSTM unit’s reading gate provides significant", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 489, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 489, + 144 + ], + "score": 1.0, + "content": "improvement over the small-sized task-oriented dialogue datasets which are currently available.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 108, + 162, + 297, + 173 + ], + "lines": [ + { + "bbox": [ + 106, + 161, + 298, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 298, + 174 + ], + "score": 1.0, + "content": "4.4 HIERARCHICAL-LEX-DELEX-SC-LSTM", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 183, + 504, + 206 + ], + "lines": [ + { + "bbox": [ + 106, + 183, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 505, + 196 + ], + "score": 1.0, + "content": "This model is a variant of the “ld-sc-LSTM” model proposed by Sharma et al. (2017) which is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 194, + 504, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 504, + 207 + ], + "score": 1.0, + "content": "based on an encoder-decoder framework. We call our model “hierarchical-lex-delex-sc-LSTM”3.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 223, + 296, + 343 + ], + "lines": [ + { + "bbox": [ + 106, + 222, + 297, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 297, + 235 + ], + "score": 1.0, + "content": "We present the encoder in Figure 1. The en-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 233, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 282, + 246 + ], + "score": 1.0, + "content": "coder consists of a hierarchical LSTM with", + "type": "text" + }, + { + "bbox": [ + 282, + 234, + 296, + 245 + ], + "score": 0.82, + "content": "N _ { e }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 245, + 297, + 256 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 178, + 256 + ], + "score": 1.0, + "content": "time-steps, where", + "type": "text" + }, + { + "bbox": [ + 179, + 245, + 192, + 256 + ], + "score": 0.89, + "content": "N _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 245, + 297, + 256 + ], + "score": 1.0, + "content": "is the number of non-zero", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 256, + 298, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 256, + 165, + 268 + ], + "score": 1.0, + "content": "entries in the", + "type": "text" + }, + { + "bbox": [ + 165, + 256, + 183, + 266 + ], + "score": 0.59, + "content": "D A", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 256, + 298, + 268 + ], + "score": 1.0, + "content": "vector. Each time-step of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 266, + 297, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 297, + 278 + ], + "score": 1.0, + "content": "the encoder encodes one dialogue act’s delex-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 276, + 297, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 297, + 291 + ], + "score": 1.0, + "content": "icalized and lexicalized slot-value pair (e.g.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 289, + 297, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 297, + 300 + ], + "score": 1.0, + "content": "(INFORM-FOOD, ‘Chinese’)). The delexical-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 299, + 298, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 298, + 311 + ], + "score": 1.0, + "content": "ized act-slot part is encoded as a one-hot vector", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 310, + 297, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 191, + 322 + ], + "score": 1.0, + "content": "which we refer to as", + "type": "text" + }, + { + "bbox": [ + 191, + 311, + 212, + 322 + ], + "score": 0.87, + "content": "D A _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 310, + 217, + 322 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 217, + 311, + 237, + 322 + ], + "score": 0.88, + "content": "D A _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 310, + 297, + 322 + ], + "score": 1.0, + "content": "is constructed", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 321, + 297, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 209, + 334 + ], + "score": 1.0, + "content": "by masking all except the", + "type": "text" + }, + { + "bbox": [ + 210, + 322, + 221, + 332 + ], + "score": 0.85, + "content": "t ^ { \\mathrm { { t h } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 321, + 297, + 334 + ], + "score": 1.0, + "content": "dialogue act in the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 332, + 297, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 297, + 345 + ], + "score": 1.0, + "content": "DA vector4. The lexicalized value part is en-", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 304, + 219, + 504, + 302 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 304, + 219, + 504, + 302 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 504, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 504, + 302 + ], + "score": 0.967, + "type": "image", + "image_path": "17acec8a69179f00d7d76eae7dbbeb4746f29c87169dd83f9d9aae7969b3e28f.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 304, + 219, + 504, + 232.83333333333334 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 304, + 232.83333333333334, + 504, + 246.66666666666669 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 304, + 246.66666666666669, + 504, + 260.5 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 304, + 260.5, + 504, + 274.3333333333333 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 304, + 274.3333333333333, + 504, + 288.16666666666663 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 304, + 288.16666666666663, + 504, + 301.99999999999994 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 315, + 312, + 494, + 324 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 496, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 496, + 325 + ], + "score": 1.0, + "content": "Figure 1: Encoder of the hld-scLSTM model", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + } + ], + "index": 23.25 + }, + { + "type": "text", + "bbox": [ + 107, + 344, + 505, + 431 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 357 + ], + "score": 1.0, + "content": "coded by an LSTM encoder which shares parameters across all time-steps and operates over the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 282, + 367 + ], + "score": 1.0, + "content": "word-embeddings of the lexicalized values", + "type": "text" + }, + { + "bbox": [ + 282, + 356, + 297, + 367 + ], + "score": 0.87, + "content": "v _ { t , i }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 354, + 505, + 367 + ], + "score": 1.0, + "content": ". Our model differs from the “ld-sc-LSTM” model", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "in that we use an LSTM encoder over the word-embeddings instead of computing the mean of the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 375, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 420, + 389 + ], + "score": 1.0, + "content": "word-embeddings. The final hidden state of this LSTM is concatenated with", + "type": "text" + }, + { + "bbox": [ + 420, + 377, + 441, + 388 + ], + "score": 0.9, + "content": "D A _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 375, + 506, + 389 + ], + "score": 1.0, + "content": "and is given as", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 388, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 506, + 399 + ], + "score": 1.0, + "content": "input to the upper LSTM (see Figure 1). The final hidden state of the upper LSTM is then provided", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 397, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 506, + 411 + ], + "score": 1.0, + "content": "to the decoder as input. This is another difference from the “ld-sc-LSTM” as that uses the mean of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 410, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 505, + 421 + ], + "score": 1.0, + "content": "all the hidden states of the encoder instead, which, in our experiments, did not perform as well as", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 420, + 245, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 232, + 432 + ], + "score": 1.0, + "content": "using just the final hidden state", + "type": "text" + }, + { + "bbox": [ + 233, + 421, + 242, + 430 + ], + "score": 0.78, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 420, + 245, + 432 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 437, + 306, + 525 + ], + "lines": [ + { + "bbox": [ + 106, + 437, + 307, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 307, + 448 + ], + "score": 1.0, + "content": "The decoder is described in Figure 2. It is the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 448, + 307, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 307, + 459 + ], + "score": 1.0, + "content": "same as in the “ld-sc-LSTM” model. At each", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 459, + 307, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 294, + 471 + ], + "score": 1.0, + "content": "time-step, it takes as input the encoder output", + "type": "text" + }, + { + "bbox": [ + 294, + 460, + 303, + 469 + ], + "score": 0.71, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 459, + 307, + 471 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 470, + 307, + 482 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 122, + 482 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 123, + 470, + 140, + 480 + ], + "score": 0.6, + "content": "D A", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 470, + 307, + 482 + ], + "score": 1.0, + "content": "vector, and the word-embedding of the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 481, + 307, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 481, + 307, + 493 + ], + "score": 1.0, + "content": "word generated at the previous time-step. The", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 491, + 307, + 504 + ], + "spans": [ + { + "bbox": [ + 107, + 492, + 124, + 502 + ], + "score": 0.64, + "content": "D A", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 491, + 307, + 504 + ], + "score": 1.0, + "content": "vector is also additionally provided to the sc-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 502, + 307, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 307, + 515 + ], + "score": 1.0, + "content": "LSTM cell in order for it to be regulated by its", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 514, + 298, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 298, + 525 + ], + "score": 1.0, + "content": "reading gate as described in Wen et al. (2015b).", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 37.5 + }, + { + "type": "image", + "bbox": [ + 313, + 438, + 505, + 524 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 313, + 438, + 505, + 524 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 313, + 438, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 505, + 524 + ], + "score": 0.969, + "type": "image", + "image_path": "114dbf8875d6790d50ee41019e8384e6fe26efeea5c18f9566f2fbec598cb400.jpg" + } + ] + } + ], + "index": 44.5, + "virtual_lines": [ + { + "bbox": [ + 313, + 438, + 505, + 452.3333333333333 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 313, + 452.3333333333333, + 505, + 466.66666666666663 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 313, + 466.66666666666663, + 505, + 480.99999999999994 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 313, + 480.99999999999994, + 505, + 495.33333333333326 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 313, + 495.33333333333326, + 505, + 509.6666666666666 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 313, + 509.6666666666666, + 505, + 523.9999999999999 + ], + "spans": [], + "index": 47 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 320, + 534, + 500, + 546 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 317, + 533, + 501, + 546 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 501, + 546 + ], + "score": 1.0, + "content": "Figure 2: Decoder of the hld-scLSTM model", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + } + ], + "index": 46.25 + }, + { + "type": "title", + "bbox": [ + 108, + 546, + 200, + 558 + ], + "lines": [ + { + "bbox": [ + 105, + 545, + 201, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 201, + 560 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 49 + }, + { + "type": "title", + "bbox": [ + 107, + 573, + 178, + 585 + ], + "lines": [ + { + "bbox": [ + 106, + 573, + 180, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 180, + 586 + ], + "score": 1.0, + "content": "5.1 DECODING", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 50 + }, + { + "type": "text", + "bbox": [ + 106, + 596, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "During training, at each time-step, we use the ground truth word from the previous time-step. The", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "score": 1.0, + "content": "model thus learns to generate the next word given the previous one. On the other hand, to generate", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "sentences during test time, we use beam search. The first word input to the generator is a special", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 131, + 641 + ], + "score": 1.0, + "content": "token", + "type": "text" + }, + { + "bbox": [ + 131, + 629, + 170, + 639 + ], + "score": 0.87, + "content": "< b o s >", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "which indicates the beginning of the sequence. Decoding is stopped if we reach a", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 640, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 437, + 652 + ], + "score": 1.0, + "content": "specified maximum number of time-steps or if the model outputs a special token", + "type": "text" + }, + { + "bbox": [ + 438, + 641, + 477, + 650 + ], + "score": 0.88, + "content": "< e o s >", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 640, + 506, + 652 + ], + "score": 1.0, + "content": "which", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 650, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 662 + ], + "score": 1.0, + "content": "indicates the end of the sequence. We also use a slot error rate penalty, similarly to Wen et al.", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 661, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 506, + 674 + ], + "score": 1.0, + "content": "(2015b), to re-rank the sentences generated with beam search. We use this method for all three of", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 671, + 353, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 353, + 685 + ], + "score": 1.0, + "content": "the LSTM, d-scLSTM, and hld-scLSTM models for fairness.", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 54.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 699, + 270, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 697, + 267, + 711 + ], + "spans": [ + { + "bbox": [ + 118, + 697, + 267, + 711 + ], + "score": 1.0, + "content": "2We will also refer to it as “d-scLSTM”.", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 708, + 271, + 722 + ], + "spans": [ + { + "bbox": [ + 118, + 708, + 271, + 722 + ], + "score": 1.0, + "content": "3We will also refer to it as“hld-scLSTM”.", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 720, + 230, + 732 + ], + "spans": [ + { + "bbox": [ + 118, + 720, + 230, + 732 + ], + "score": 1.0, + "content": "4also referred to as DAt=1:N", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 370, + 95 + ], + "score": 1.0, + "content": "model the “delex-sc-LSTM”2. As in the previous model, the input", + "type": "text" + }, + { + "bbox": [ + 370, + 83, + 388, + 93 + ], + "score": 0.51, + "content": "D A", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "vector only encodes acts and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 411, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 411, + 106 + ], + "score": 1.0, + "content": "delexicalized slots. It does not contain any information about the slot value.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 504, + 143 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 247, + 123 + ], + "score": 1.0, + "content": "By providing this model the same", + "type": "text" + }, + { + "bbox": [ + 248, + 111, + 265, + 121 + ], + "score": 0.57, + "content": "D A", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 110, + 506, + 123 + ], + "score": 1.0, + "content": "vector input as the one given to the LSTM model, we can", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "score": 1.0, + "content": "directly study if the additional complexity of the sc-LSTM unit’s reading gate provides significant", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 489, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 489, + 144 + ], + "score": 1.0, + "content": "improvement over the small-sized task-oriented dialogue datasets which are currently available.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 110, + 506, + 144 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 162, + 297, + 173 + ], + "lines": [ + { + "bbox": [ + 106, + 161, + 298, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 298, + 174 + ], + "score": 1.0, + "content": "4.4 HIERARCHICAL-LEX-DELEX-SC-LSTM", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 183, + 504, + 206 + ], + "lines": [ + { + "bbox": [ + 106, + 183, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 505, + 196 + ], + "score": 1.0, + "content": "This model is a variant of the “ld-sc-LSTM” model proposed by Sharma et al. (2017) which is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 194, + 504, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 504, + 207 + ], + "score": 1.0, + "content": "based on an encoder-decoder framework. We call our model “hierarchical-lex-delex-sc-LSTM”3.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 183, + 505, + 207 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 223, + 296, + 343 + ], + "lines": [ + { + "bbox": [ + 106, + 222, + 297, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 297, + 235 + ], + "score": 1.0, + "content": "We present the encoder in Figure 1. The en-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 233, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 282, + 246 + ], + "score": 1.0, + "content": "coder consists of a hierarchical LSTM with", + "type": "text" + }, + { + "bbox": [ + 282, + 234, + 296, + 245 + ], + "score": 0.82, + "content": "N _ { e }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 245, + 297, + 256 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 178, + 256 + ], + "score": 1.0, + "content": "time-steps, where", + "type": "text" + }, + { + "bbox": [ + 179, + 245, + 192, + 256 + ], + "score": 0.89, + "content": "N _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 245, + 297, + 256 + ], + "score": 1.0, + "content": "is the number of non-zero", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 256, + 298, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 256, + 165, + 268 + ], + "score": 1.0, + "content": "entries in the", + "type": "text" + }, + { + "bbox": [ + 165, + 256, + 183, + 266 + ], + "score": 0.59, + "content": "D A", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 256, + 298, + 268 + ], + "score": 1.0, + "content": "vector. Each time-step of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 266, + 297, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 297, + 278 + ], + "score": 1.0, + "content": "the encoder encodes one dialogue act’s delex-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 276, + 297, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 297, + 291 + ], + "score": 1.0, + "content": "icalized and lexicalized slot-value pair (e.g.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 289, + 297, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 297, + 300 + ], + "score": 1.0, + "content": "(INFORM-FOOD, ‘Chinese’)). The delexical-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 299, + 298, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 298, + 311 + ], + "score": 1.0, + "content": "ized act-slot part is encoded as a one-hot vector", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 310, + 297, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 191, + 322 + ], + "score": 1.0, + "content": "which we refer to as", + "type": "text" + }, + { + "bbox": [ + 191, + 311, + 212, + 322 + ], + "score": 0.87, + "content": "D A _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 310, + 217, + 322 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 217, + 311, + 237, + 322 + ], + "score": 0.88, + "content": "D A _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 310, + 297, + 322 + ], + "score": 1.0, + "content": "is constructed", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 321, + 297, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 209, + 334 + ], + "score": 1.0, + "content": "by masking all except the", + "type": "text" + }, + { + "bbox": [ + 210, + 322, + 221, + 332 + ], + "score": 0.85, + "content": "t ^ { \\mathrm { { t h } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 321, + 297, + 334 + ], + "score": 1.0, + "content": "dialogue act in the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 332, + 297, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 297, + 345 + ], + "score": 1.0, + "content": "DA vector4. The lexicalized value part is en-", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 222, + 298, + 345 + ] + }, + { + "type": "image", + "bbox": [ + 304, + 219, + 504, + 302 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 304, + 219, + 504, + 302 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 504, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 504, + 302 + ], + "score": 0.967, + "type": "image", + "image_path": "17acec8a69179f00d7d76eae7dbbeb4746f29c87169dd83f9d9aae7969b3e28f.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 304, + 219, + 504, + 232.83333333333334 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 304, + 232.83333333333334, + 504, + 246.66666666666669 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 304, + 246.66666666666669, + 504, + 260.5 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 304, + 260.5, + 504, + 274.3333333333333 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 304, + 274.3333333333333, + 504, + 288.16666666666663 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 304, + 288.16666666666663, + 504, + 301.99999999999994 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 315, + 312, + 494, + 324 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 496, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 496, + 325 + ], + "score": 1.0, + "content": "Figure 1: Encoder of the hld-scLSTM model", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + } + ], + "index": 23.25 + }, + { + "type": "text", + "bbox": [ + 107, + 344, + 505, + 431 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 357 + ], + "score": 1.0, + "content": "coded by an LSTM encoder which shares parameters across all time-steps and operates over the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 282, + 367 + ], + "score": 1.0, + "content": "word-embeddings of the lexicalized values", + "type": "text" + }, + { + "bbox": [ + 282, + 356, + 297, + 367 + ], + "score": 0.87, + "content": "v _ { t , i }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 354, + 505, + 367 + ], + "score": 1.0, + "content": ". Our model differs from the “ld-sc-LSTM” model", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "in that we use an LSTM encoder over the word-embeddings instead of computing the mean of the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 375, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 420, + 389 + ], + "score": 1.0, + "content": "word-embeddings. The final hidden state of this LSTM is concatenated with", + "type": "text" + }, + { + "bbox": [ + 420, + 377, + 441, + 388 + ], + "score": 0.9, + "content": "D A _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 375, + 506, + 389 + ], + "score": 1.0, + "content": "and is given as", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 388, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 506, + 399 + ], + "score": 1.0, + "content": "input to the upper LSTM (see Figure 1). The final hidden state of the upper LSTM is then provided", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 397, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 506, + 411 + ], + "score": 1.0, + "content": "to the decoder as input. This is another difference from the “ld-sc-LSTM” as that uses the mean of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 410, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 505, + 421 + ], + "score": 1.0, + "content": "all the hidden states of the encoder instead, which, in our experiments, did not perform as well as", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 420, + 245, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 232, + 432 + ], + "score": 1.0, + "content": "using just the final hidden state", + "type": "text" + }, + { + "bbox": [ + 233, + 421, + 242, + 430 + ], + "score": 0.78, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 420, + 245, + 432 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 342, + 506, + 432 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 437, + 306, + 525 + ], + "lines": [ + { + "bbox": [ + 106, + 437, + 307, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 307, + 448 + ], + "score": 1.0, + "content": "The decoder is described in Figure 2. It is the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 448, + 307, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 307, + 459 + ], + "score": 1.0, + "content": "same as in the “ld-sc-LSTM” model. At each", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 459, + 307, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 294, + 471 + ], + "score": 1.0, + "content": "time-step, it takes as input the encoder output", + "type": "text" + }, + { + "bbox": [ + 294, + 460, + 303, + 469 + ], + "score": 0.71, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 459, + 307, + 471 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 470, + 307, + 482 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 122, + 482 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 123, + 470, + 140, + 480 + ], + "score": 0.6, + "content": "D A", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 470, + 307, + 482 + ], + "score": 1.0, + "content": "vector, and the word-embedding of the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 481, + 307, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 481, + 307, + 493 + ], + "score": 1.0, + "content": "word generated at the previous time-step. The", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 491, + 307, + 504 + ], + "spans": [ + { + "bbox": [ + 107, + 492, + 124, + 502 + ], + "score": 0.64, + "content": "D A", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 491, + 307, + 504 + ], + "score": 1.0, + "content": "vector is also additionally provided to the sc-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 502, + 307, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 307, + 515 + ], + "score": 1.0, + "content": "LSTM cell in order for it to be regulated by its", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 514, + 298, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 298, + 525 + ], + "score": 1.0, + "content": "reading gate as described in Wen et al. (2015b).", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 437, + 307, + 525 + ] + }, + { + "type": "image", + "bbox": [ + 313, + 438, + 505, + 524 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 313, + 438, + 505, + 524 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 313, + 438, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 505, + 524 + ], + "score": 0.969, + "type": "image", + "image_path": "114dbf8875d6790d50ee41019e8384e6fe26efeea5c18f9566f2fbec598cb400.jpg" + } + ] + } + ], + "index": 44.5, + "virtual_lines": [ + { + "bbox": [ + 313, + 438, + 505, + 452.3333333333333 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 313, + 452.3333333333333, + 505, + 466.66666666666663 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 313, + 466.66666666666663, + 505, + 480.99999999999994 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 313, + 480.99999999999994, + 505, + 495.33333333333326 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 313, + 495.33333333333326, + 505, + 509.6666666666666 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 313, + 509.6666666666666, + 505, + 523.9999999999999 + ], + "spans": [], + "index": 47 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 320, + 534, + 500, + 546 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 317, + 533, + 501, + 546 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 501, + 546 + ], + "score": 1.0, + "content": "Figure 2: Decoder of the hld-scLSTM model", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + } + ], + "index": 46.25 + }, + { + "type": "title", + "bbox": [ + 108, + 546, + 200, + 558 + ], + "lines": [ + { + "bbox": [ + 105, + 545, + 201, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 201, + 560 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 49 + }, + { + "type": "title", + "bbox": [ + 107, + 573, + 178, + 585 + ], + "lines": [ + { + "bbox": [ + 106, + 573, + 180, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 180, + 586 + ], + "score": 1.0, + "content": "5.1 DECODING", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 50 + }, + { + "type": "text", + "bbox": [ + 106, + 596, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "During training, at each time-step, we use the ground truth word from the previous time-step. The", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "score": 1.0, + "content": "model thus learns to generate the next word given the previous one. On the other hand, to generate", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "sentences during test time, we use beam search. The first word input to the generator is a special", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 131, + 641 + ], + "score": 1.0, + "content": "token", + "type": "text" + }, + { + "bbox": [ + 131, + 629, + 170, + 639 + ], + "score": 0.87, + "content": "< b o s >", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "which indicates the beginning of the sequence. Decoding is stopped if we reach a", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 640, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 437, + 652 + ], + "score": 1.0, + "content": "specified maximum number of time-steps or if the model outputs a special token", + "type": "text" + }, + { + "bbox": [ + 438, + 641, + 477, + 650 + ], + "score": 0.88, + "content": "< e o s >", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 640, + 506, + 652 + ], + "score": 1.0, + "content": "which", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 650, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 662 + ], + "score": 1.0, + "content": "indicates the end of the sequence. We also use a slot error rate penalty, similarly to Wen et al.", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 661, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 506, + 674 + ], + "score": 1.0, + "content": "(2015b), to re-rank the sentences generated with beam search. We use this method for all three of", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 671, + 353, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 353, + 685 + ], + "score": 1.0, + "content": "the LSTM, d-scLSTM, and hld-scLSTM models for fairness.", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 54.5, + "bbox_fs": [ + 105, + 596, + 506, + 685 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 114, + 95, + 497, + 156 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 120, + 84, + 488, + 94 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 83, + 489, + 96 + ], + "spans": [ + { + "bbox": [ + 122, + 83, + 489, + 96 + ], + "score": 1.0, + "content": "Table 1: Performance comparison across models on word-overlap based automated metrics", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 114, + 95, + 497, + 156 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 95, + 497, + 156 + ], + "spans": [ + { + "bbox": [ + 114, + 95, + 497, + 156 + ], + "score": 0.977, + "html": "
DSTC2Restaurants
B-1B-2B-3B-4MRLB-1B-2B-3B-4MRL
Gold1.001.001.001.001.001.001.001.001.001.001.001.00
Random0.8750.8430.8220.8070.5640.8520.8720.8130.7650.7210.5040.796
LSTM0.9000.8790.8630.8510.6100.8880.9820.9660.9490.9320.6520.944
d-scLSTM0.8800.8500.8280.8120.5780.8740.9800.9640.9480.9310.6540.945
hld-scLSTM0.9090.8900.8780.8700.6240.8990.9850.9780.9700.9620.7040.965
", + "type": "table", + "image_path": "4cdc49a5bd65ea5ff860b94e3cce4146a9861fad40d0462bec314d81a7dbfe5e.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 114, + 95, + 497, + 115.33333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 114, + 115.33333333333333, + 497, + 135.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 114, + 135.66666666666666, + 497, + 156.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "table", + "bbox": [ + 124, + 174, + 485, + 244 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 110, + 164, + 501, + 173 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 162, + 503, + 176 + ], + "spans": [ + { + "bbox": [ + 108, + 162, + 503, + 176 + ], + "score": 1.0, + "content": "Table 2: Performance comparison across models on sentence-embedding based automated metrics", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 124, + 174, + 485, + 244 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 174, + 485, + 244 + ], + "spans": [ + { + "bbox": [ + 124, + 174, + 485, + 244 + ], + "score": 0.979, + "html": "
DSTC2Restaurants
SkipThoughtEmbeddingAverageVectorExtremaGreedyMatchingSkipThoughtEmbeddingAverageVectorExtremaGreedyMatching
Gold1.001.001.001.001.001.001.001.00
Random0.9060.9810.9100.9470.8430.9570.9050.930
LSTM0.9460.9850.9350.9620.9450.9970.9860.991
d-scLSTM0.9250.9840.9260.9570.9480.9970.9860.991
hld-scLSTM0.9320.9870.9420.9640.9680.9970.9890.993
", + "type": "table", + "image_path": "782379d7f8a969d348efbec4caa0e6edf6b2d4c105b13cb0daa619db4478039d.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 124, + 174, + 485, + 197.33333333333334 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 124, + 197.33333333333334, + 485, + 220.66666666666669 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 124, + 220.66666666666669, + 485, + 244.00000000000003 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 5.0 + }, + { + "type": "text", + "bbox": [ + 107, + 251, + 505, + 295 + ], + "lines": [ + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "score": 1.0, + "content": "Similarly to the LSTM model, the d-scLSTM and hld-scLSTM generate delexicalized sentences,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 263, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 274 + ], + "score": 1.0, + "content": "i.e., they generate slot tokens instead of slot values directly. These slot tokens are replaced with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 273, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 104, + 273, + 505, + 286 + ], + "score": 1.0, + "content": "slot values in a post-processing step which is a fairly common step in task-oriented dialogue NLG", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 284, + 148, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 148, + 297 + ], + "score": 1.0, + "content": "literature.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 107, + 310, + 187, + 322 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 189, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 189, + 324 + ], + "score": 1.0, + "content": "5.2 EVALUATION", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 332, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 332, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 505, + 343 + ], + "score": 1.0, + "content": "In NLG tasks, improvements in automated metric scores are most commonly used to demonstrate", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 355 + ], + "score": 1.0, + "content": "improvement in the generation task. However, these metrics have been shown to only weakly cor-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 354, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 365 + ], + "score": 1.0, + "content": "relate with human evaluation in the non task-oriented dialogue setting (Liu et al., 2016) and hence", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 365, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 376 + ], + "score": 1.0, + "content": "are not considered reliable measures of improvement. Human evaluation is considered the metric of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "choice, but human ratings are expensive to obtain. The ease of computing these automated metrics", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 387, + 430, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 430, + 399 + ], + "score": 1.0, + "content": "and their availability for rapid prototyping has lead to their widespread adoption.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 403, + 505, + 491 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "We evaluate the models described in the previous section on the DSTC2 (Henderson et al., 2014) and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "score": 1.0, + "content": "the Restaurants datasets (Wen et al., 2015a) using these automated metrics. These datasets are some", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 425, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 506, + 437 + ], + "score": 1.0, + "content": "of the only available resources for studying NLG for task-oriented dialogue. The DSTC2 dataset", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 436, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 506, + 449 + ], + "score": 1.0, + "content": "contains dialogues between human users and a dialogue system in a restaurant domain. The dataset", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "score": 1.0, + "content": "is annotated with dialogue acts, slot type, and slot values. The NLG component of the dialogue", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 458, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 104, + 458, + 505, + 470 + ], + "score": 1.0, + "content": "system used for data collection is templated. The Restaurants dataset was specifically proposed for", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 468, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 104, + 468, + 506, + 483 + ], + "score": 1.0, + "content": "NLG and provides, for a set of dialogue acts with slot types and slot values, two sentences generated", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 479, + 155, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 155, + 493 + ], + "score": 1.0, + "content": "by humans.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 496, + 505, + 574 + ], + "lines": [ + { + "bbox": [ + 106, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "We present the results of our experiments in Table 1 and Table 2. The scores of all the models on", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 507, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 506, + 521 + ], + "score": 1.0, + "content": "these automated metrics are very high. This indicates that there is significant word overlap between", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "score": 1.0, + "content": "the generated and the reference sentences and that the NLG task on these datasets can be solved with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "a simple model such as the LSTM model. In effect, Table 1 shows that the LSTM model performs", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 540, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 104, + 540, + 505, + 554 + ], + "score": 1.0, + "content": "comparably to the d-scLSTM model based on the word-overlap metrics. This can be explained by", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "score": 1.0, + "content": "the fact that the d-scLSTM model has more parameters and might suffer from overfitting issues on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 563, + 228, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 228, + 574 + ], + "score": 1.0, + "content": "these relatively small datasets.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 579, + 504, + 646 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 504, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 504, + 591 + ], + "score": 1.0, + "content": "The hld-scLSTM is considered to consistently outperform the other models based on the word-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 590, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 602 + ], + "score": 1.0, + "content": "overlap metrics. As explained by Sharma et al. (2017), this improvement results from the model’s", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 602, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 602, + 505, + 614 + ], + "score": 1.0, + "content": "access to the lexicalized slot values, due to which it can take into account the grammatical associ-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "ations of the generated words near the output tokens, thereby generating higher quality sentences.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 623, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 506, + 636 + ], + "score": 1.0, + "content": "However, Table 2 shows that sentence-embedding based metrics judge all the models except the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 635, + 431, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 431, + 647 + ], + "score": 1.0, + "content": "random one to perform quite similarly with again, very high performance scores.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 651, + 427, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 429, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 429, + 664 + ], + "score": 1.0, + "content": "In the next section, we add human evaluation for these models on these datasets.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "title", + "bbox": [ + 108, + 677, + 257, + 688 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 259, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 259, + 690 + ], + "score": 1.0, + "content": "5.3 HUMAN RATING COLLECTION", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "We randomly selected 20 dialogue acts from the test set of each dataset. For each of these", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "contexts, we presented 5 sentences to the evaluators: the gold response provided in the test", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "set and the responses generated by the four models described in Section 4. These sentences", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 114, + 95, + 497, + 156 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 120, + 84, + 488, + 94 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 83, + 489, + 96 + ], + "spans": [ + { + "bbox": [ + 122, + 83, + 489, + 96 + ], + "score": 1.0, + "content": "Table 1: Performance comparison across models on word-overlap based automated metrics", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 114, + 95, + 497, + 156 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 95, + 497, + 156 + ], + "spans": [ + { + "bbox": [ + 114, + 95, + 497, + 156 + ], + "score": 0.977, + "html": "
DSTC2Restaurants
B-1B-2B-3B-4MRLB-1B-2B-3B-4MRL
Gold1.001.001.001.001.001.001.001.001.001.001.001.00
Random0.8750.8430.8220.8070.5640.8520.8720.8130.7650.7210.5040.796
LSTM0.9000.8790.8630.8510.6100.8880.9820.9660.9490.9320.6520.944
d-scLSTM0.8800.8500.8280.8120.5780.8740.9800.9640.9480.9310.6540.945
hld-scLSTM0.9090.8900.8780.8700.6240.8990.9850.9780.9700.9620.7040.965
", + "type": "table", + "image_path": "4cdc49a5bd65ea5ff860b94e3cce4146a9861fad40d0462bec314d81a7dbfe5e.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 114, + 95, + 497, + 115.33333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 114, + 115.33333333333333, + 497, + 135.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 114, + 135.66666666666666, + 497, + 156.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "table", + "bbox": [ + 124, + 174, + 485, + 244 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 110, + 164, + 501, + 173 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 162, + 503, + 176 + ], + "spans": [ + { + "bbox": [ + 108, + 162, + 503, + 176 + ], + "score": 1.0, + "content": "Table 2: Performance comparison across models on sentence-embedding based automated metrics", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 124, + 174, + 485, + 244 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 174, + 485, + 244 + ], + "spans": [ + { + "bbox": [ + 124, + 174, + 485, + 244 + ], + "score": 0.979, + "html": "
DSTC2Restaurants
SkipThoughtEmbeddingAverageVectorExtremaGreedyMatchingSkipThoughtEmbeddingAverageVectorExtremaGreedyMatching
Gold1.001.001.001.001.001.001.001.00
Random0.9060.9810.9100.9470.8430.9570.9050.930
LSTM0.9460.9850.9350.9620.9450.9970.9860.991
d-scLSTM0.9250.9840.9260.9570.9480.9970.9860.991
hld-scLSTM0.9320.9870.9420.9640.9680.9970.9890.993
", + "type": "table", + "image_path": "782379d7f8a969d348efbec4caa0e6edf6b2d4c105b13cb0daa619db4478039d.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 124, + 174, + 485, + 197.33333333333334 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 124, + 197.33333333333334, + 485, + 220.66666666666669 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 124, + 220.66666666666669, + 485, + 244.00000000000003 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 5.0 + }, + { + "type": "text", + "bbox": [ + 107, + 251, + 505, + 295 + ], + "lines": [ + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "score": 1.0, + "content": "Similarly to the LSTM model, the d-scLSTM and hld-scLSTM generate delexicalized sentences,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 263, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 274 + ], + "score": 1.0, + "content": "i.e., they generate slot tokens instead of slot values directly. These slot tokens are replaced with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 273, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 104, + 273, + 505, + 286 + ], + "score": 1.0, + "content": "slot values in a post-processing step which is a fairly common step in task-oriented dialogue NLG", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 284, + 148, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 148, + 297 + ], + "score": 1.0, + "content": "literature.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5, + "bbox_fs": [ + 104, + 250, + 506, + 297 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 310, + 187, + 322 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 189, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 189, + 324 + ], + "score": 1.0, + "content": "5.2 EVALUATION", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 332, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 332, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 505, + 343 + ], + "score": 1.0, + "content": "In NLG tasks, improvements in automated metric scores are most commonly used to demonstrate", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 355 + ], + "score": 1.0, + "content": "improvement in the generation task. However, these metrics have been shown to only weakly cor-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 354, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 365 + ], + "score": 1.0, + "content": "relate with human evaluation in the non task-oriented dialogue setting (Liu et al., 2016) and hence", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 365, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 376 + ], + "score": 1.0, + "content": "are not considered reliable measures of improvement. Human evaluation is considered the metric of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "choice, but human ratings are expensive to obtain. The ease of computing these automated metrics", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 387, + 430, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 430, + 399 + ], + "score": 1.0, + "content": "and their availability for rapid prototyping has lead to their widespread adoption.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 332, + 505, + 399 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 403, + 505, + 491 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "We evaluate the models described in the previous section on the DSTC2 (Henderson et al., 2014) and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "score": 1.0, + "content": "the Restaurants datasets (Wen et al., 2015a) using these automated metrics. These datasets are some", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 425, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 506, + 437 + ], + "score": 1.0, + "content": "of the only available resources for studying NLG for task-oriented dialogue. The DSTC2 dataset", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 436, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 506, + 449 + ], + "score": 1.0, + "content": "contains dialogues between human users and a dialogue system in a restaurant domain. The dataset", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "score": 1.0, + "content": "is annotated with dialogue acts, slot type, and slot values. The NLG component of the dialogue", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 458, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 104, + 458, + 505, + 470 + ], + "score": 1.0, + "content": "system used for data collection is templated. The Restaurants dataset was specifically proposed for", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 468, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 104, + 468, + 506, + 483 + ], + "score": 1.0, + "content": "NLG and provides, for a set of dialogue acts with slot types and slot values, two sentences generated", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 479, + 155, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 155, + 493 + ], + "score": 1.0, + "content": "by humans.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5, + "bbox_fs": [ + 104, + 403, + 506, + 493 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 496, + 505, + 574 + ], + "lines": [ + { + "bbox": [ + 106, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "We present the results of our experiments in Table 1 and Table 2. The scores of all the models on", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 507, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 506, + 521 + ], + "score": 1.0, + "content": "these automated metrics are very high. This indicates that there is significant word overlap between", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "score": 1.0, + "content": "the generated and the reference sentences and that the NLG task on these datasets can be solved with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "a simple model such as the LSTM model. In effect, Table 1 shows that the LSTM model performs", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 540, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 104, + 540, + 505, + 554 + ], + "score": 1.0, + "content": "comparably to the d-scLSTM model based on the word-overlap metrics. This can be explained by", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "score": 1.0, + "content": "the fact that the d-scLSTM model has more parameters and might suffer from overfitting issues on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 563, + 228, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 228, + 574 + ], + "score": 1.0, + "content": "these relatively small datasets.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30, + "bbox_fs": [ + 104, + 496, + 506, + 574 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 579, + 504, + 646 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 504, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 504, + 591 + ], + "score": 1.0, + "content": "The hld-scLSTM is considered to consistently outperform the other models based on the word-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 590, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 602 + ], + "score": 1.0, + "content": "overlap metrics. As explained by Sharma et al. (2017), this improvement results from the model’s", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 602, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 602, + 505, + 614 + ], + "score": 1.0, + "content": "access to the lexicalized slot values, due to which it can take into account the grammatical associ-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "ations of the generated words near the output tokens, thereby generating higher quality sentences.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 623, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 506, + 636 + ], + "score": 1.0, + "content": "However, Table 2 shows that sentence-embedding based metrics judge all the models except the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 635, + 431, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 431, + 647 + ], + "score": 1.0, + "content": "random one to perform quite similarly with again, very high performance scores.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 579, + 506, + 647 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 651, + 427, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 429, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 429, + 664 + ], + "score": 1.0, + "content": "In the next section, we add human evaluation for these models on these datasets.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 650, + 429, + 664 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 677, + 257, + 688 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 259, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 259, + 690 + ], + "score": 1.0, + "content": "5.3 HUMAN RATING COLLECTION", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "We randomly selected 20 dialogue acts from the test set of each dataset. For each of these", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "contexts, we presented 5 sentences to the evaluators: the gold response provided in the test", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "set and the responses generated by the four models described in Section 4. These sentences", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 385, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 506, + 397 + ], + "score": 1.0, + "content": "were randomly shuffled and not presented in the same order. We invited 18 human users to", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "score each of these 100 sentences on a Likert-type scale of 1 to 5. The users were asked", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 406, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 420 + ], + "score": 1.0, + "content": "to rate the responses depending on how appropriate they were for the specified dialogue acts.", + "type": "text", + "cross_page": true + } + ], + "index": 10 + } + ], + "index": 43, + "bbox_fs": [ + 106, + 699, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 84, + 498, + 296 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 84, + 498, + 296 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 84, + 498, + 296 + ], + "spans": [ + { + "bbox": [ + 112, + 84, + 498, + 296 + ], + "score": 0.973, + "type": "image", + "image_path": "18289be4190cbbf67f923162b61f526162ab7aca3e0d7abbab61623c5036e6d9.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 84, + 498, + 154.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 154.66666666666669, + 498, + 225.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 225.33333333333337, + 498, + 296.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 306, + 505, + 360 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 305, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 506, + 318 + ], + "score": 1.0, + "content": "Figure 3: Scatter plots for correlation of some automated metrics with human evaluation for (a) the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 315, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 399, + 329 + ], + "score": 1.0, + "content": "DSTC2 dataset, and (b) the Restaurants dataset. Random gaussian noise", + "type": "text" + }, + { + "bbox": [ + 399, + 316, + 439, + 329 + ], + "score": 0.92, + "content": "\\mathcal { N } ( 0 , 0 . 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 315, + 505, + 329 + ], + "score": 1.0, + "content": "has been added", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 327, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 294, + 340 + ], + "score": 1.0, + "content": "to data points along the human score axis and", + "type": "text" + }, + { + "bbox": [ + 295, + 327, + 340, + 339 + ], + "score": 0.93, + "content": "\\mathcal { N } ( 0 , 0 . 0 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 327, + 505, + 340 + ], + "score": 1.0, + "content": "has been added to the automated metric", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "score’s axis to aid visualization of overlapping data points. Transparency has been added for the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 350, + 156, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 156, + 361 + ], + "score": 1.0, + "content": "same effect.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 108, + 385, + 504, + 418 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 506, + 397 + ], + "score": 1.0, + "content": "were randomly shuffled and not presented in the same order. We invited 18 human users to", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "score each of these 100 sentences on a Likert-type scale of 1 to 5. The users were asked", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 406, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 420 + ], + "score": 1.0, + "content": "to rate the responses depending on how appropriate they were for the specified dialogue acts.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 419, + 375, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 376, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 376, + 430 + ], + "score": 1.0, + "content": "A score of 1 was the lowest score, meaning that the response was", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 428, + 377, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 377, + 441 + ], + "score": 1.0, + "content": "not appropriate at all whereas a score of 5 meant that the sentence", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 440, + 203, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 203, + 452 + ], + "score": 1.0, + "content": "was highly appropriate.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 387, + 457, + 498, + 529 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 383, + 427, + 504, + 457 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 383, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 383, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "Table 3: Pairwise Cohen’s", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 383, + 437, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 383, + 437, + 505, + 448 + ], + "score": 1.0, + "content": "kappa scores for the 11 hu-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 383, + 450, + 427, + 460 + ], + "spans": [ + { + "bbox": [ + 383, + 450, + 427, + 460 + ], + "score": 1.0, + "content": "man users", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "table_body", + "bbox": [ + 387, + 457, + 498, + 529 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 387, + 457, + 498, + 529 + ], + "spans": [ + { + "bbox": [ + 387, + 457, + 498, + 529 + ], + "score": 0.972, + "html": "
K# pairs% pairs
>0.155/55100.0 %
>0.240/5572.7 %
>0.328/5550.9 %
>0.419/5534.5 %
>0.58/5514.5 %
>0.60/550.0 %
", + "type": "table", + "image_path": "ee87a9e1885b0fe151cbbeb91e4be51300f4ad64919da97839a19c8d62eef0b2.jpg" + } + ] + } + ], + "index": 22.0, + "virtual_lines": [ + { + "bbox": [ + 387, + 457, + 498, + 493.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 387, + 493.0, + 498, + 529.0 + ], + "spans": [], + "index": 24 + } + ] + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 457, + 376, + 523 + ], + "lines": [ + { + "bbox": [ + 106, + 457, + 376, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 376, + 469 + ], + "score": 1.0, + "content": "We computed Cohen’s kappa scores (Cohen, 1960) between the hu-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 468, + 376, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 376, + 480 + ], + "score": 1.0, + "content": "man users in pairs of two. We removed 7 users who had kappa", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 479, + 376, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 376, + 490 + ], + "score": 1.0, + "content": "scores less than 0.1 and used the remaining 11 users for the corre-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 489, + 376, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 376, + 501 + ], + "score": 1.0, + "content": "lation study. The kappa scores are presented in Table 3. Most of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 500, + 377, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 213, + 514 + ], + "score": 1.0, + "content": "user pairs have a Cohen’s", + "type": "text" + }, + { + "bbox": [ + 213, + 501, + 248, + 511 + ], + "score": 0.89, + "content": "\\kappa > 0 . 3", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 500, + 377, + 514 + ], + "score": 1.0, + "content": "which indicates fair agreement", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 512, + 245, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 245, + 523 + ], + "score": 1.0, + "content": "between users (Viera et al., 2005).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.0 + }, + { + "type": "title", + "bbox": [ + 108, + 543, + 426, + 554 + ], + "lines": [ + { + "bbox": [ + 105, + 542, + 428, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 428, + 555 + ], + "score": 1.0, + "content": "5.4 CORRELATION BETWEEN AUTOMATED METRICS AND HUMAN SCORES", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 566, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "We present the correlation between the automated metrics and our collected human ratings in Ta-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 577, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 590 + ], + "score": 1.0, + "content": "ble 4. We measure human v.s. human correlation by randomly splitting the human users into two", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "score": 1.0, + "content": "groups. The results indicate that in most cases, human scores correlate the best with other human", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 599, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 611 + ], + "score": 1.0, + "content": "scores. Except in the case of the Spearman correlation for BLEU-N scores, we can see that there is", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 610, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 622 + ], + "score": 1.0, + "content": "a positive correlation between the automated metrics and the human scores for these task-oriented", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 621, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 633 + ], + "score": 1.0, + "content": "datasets, which contrasts with the non task-oriented dialogue setting where Liu et al. (2016) observed", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 633, + 221, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 221, + 644 + ], + "score": 1.0, + "content": "no strong correlation trends.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 649, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "A likely explanation for the negative Spearman correlation values for BLEU-N is that there is only", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "one gold reference per context in the DSTC2 dataset. The Restaurants dataset, on the other hand,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "provides two gold references per context. Having multiple gold references increases the likelihood", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 681, + 502, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 502, + 695 + ], + "score": 1.0, + "content": "that the generated response will have significant word-overlap with one of the reference responses.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "We present scatter plots for some of the metrics presented in Table 4 in Figure 3. We observe that", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "all the metrics correlate very well with humans on high scoring examples. As it can be seen in the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "scatter plots, most of the sentences are given the maximal score of 5 by the human evaluators. This", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 84, + 498, + 296 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 84, + 498, + 296 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 84, + 498, + 296 + ], + "spans": [ + { + "bbox": [ + 112, + 84, + 498, + 296 + ], + "score": 0.973, + "type": "image", + "image_path": "18289be4190cbbf67f923162b61f526162ab7aca3e0d7abbab61623c5036e6d9.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 84, + 498, + 154.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 154.66666666666669, + 498, + 225.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 225.33333333333337, + 498, + 296.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 306, + 505, + 360 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 305, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 506, + 318 + ], + "score": 1.0, + "content": "Figure 3: Scatter plots for correlation of some automated metrics with human evaluation for (a) the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 315, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 399, + 329 + ], + "score": 1.0, + "content": "DSTC2 dataset, and (b) the Restaurants dataset. Random gaussian noise", + "type": "text" + }, + { + "bbox": [ + 399, + 316, + 439, + 329 + ], + "score": 0.92, + "content": "\\mathcal { N } ( 0 , 0 . 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 315, + 505, + 329 + ], + "score": 1.0, + "content": "has been added", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 327, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 294, + 340 + ], + "score": 1.0, + "content": "to data points along the human score axis and", + "type": "text" + }, + { + "bbox": [ + 295, + 327, + 340, + 339 + ], + "score": 0.93, + "content": "\\mathcal { N } ( 0 , 0 . 0 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 327, + 505, + 340 + ], + "score": 1.0, + "content": "has been added to the automated metric", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "score’s axis to aid visualization of overlapping data points. Transparency has been added for the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 350, + 156, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 156, + 361 + ], + "score": 1.0, + "content": "same effect.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 108, + 385, + 504, + 418 + ], + "lines": [], + "index": 9, + "bbox_fs": [ + 105, + 385, + 506, + 420 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 419, + 375, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 376, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 376, + 430 + ], + "score": 1.0, + "content": "A score of 1 was the lowest score, meaning that the response was", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 428, + 377, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 377, + 441 + ], + "score": 1.0, + "content": "not appropriate at all whereas a score of 5 meant that the sentence", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 440, + 203, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 203, + 452 + ], + "score": 1.0, + "content": "was highly appropriate.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 417, + 377, + 452 + ] + }, + { + "type": "table", + "bbox": [ + 387, + 457, + 498, + 529 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 383, + 427, + 504, + 457 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 383, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 383, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "Table 3: Pairwise Cohen’s", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 383, + 437, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 383, + 437, + 505, + 448 + ], + "score": 1.0, + "content": "kappa scores for the 11 hu-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 383, + 450, + 427, + 460 + ], + "spans": [ + { + "bbox": [ + 383, + 450, + 427, + 460 + ], + "score": 1.0, + "content": "man users", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "table_body", + "bbox": [ + 387, + 457, + 498, + 529 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 387, + 457, + 498, + 529 + ], + "spans": [ + { + "bbox": [ + 387, + 457, + 498, + 529 + ], + "score": 0.972, + "html": "
K# pairs% pairs
>0.155/55100.0 %
>0.240/5572.7 %
>0.328/5550.9 %
>0.419/5534.5 %
>0.58/5514.5 %
>0.60/550.0 %
", + "type": "table", + "image_path": "ee87a9e1885b0fe151cbbeb91e4be51300f4ad64919da97839a19c8d62eef0b2.jpg" + } + ] + } + ], + "index": 22.0, + "virtual_lines": [ + { + "bbox": [ + 387, + 457, + 498, + 493.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 387, + 493.0, + 498, + 529.0 + ], + "spans": [], + "index": 24 + } + ] + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 457, + 376, + 523 + ], + "lines": [ + { + "bbox": [ + 106, + 457, + 376, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 376, + 469 + ], + "score": 1.0, + "content": "We computed Cohen’s kappa scores (Cohen, 1960) between the hu-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 468, + 376, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 376, + 480 + ], + "score": 1.0, + "content": "man users in pairs of two. We removed 7 users who had kappa", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 479, + 376, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 376, + 490 + ], + "score": 1.0, + "content": "scores less than 0.1 and used the remaining 11 users for the corre-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 489, + 376, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 376, + 501 + ], + "score": 1.0, + "content": "lation study. The kappa scores are presented in Table 3. Most of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 500, + 377, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 213, + 514 + ], + "score": 1.0, + "content": "user pairs have a Cohen’s", + "type": "text" + }, + { + "bbox": [ + 213, + 501, + 248, + 511 + ], + "score": 0.89, + "content": "\\kappa > 0 . 3", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 500, + 377, + 514 + ], + "score": 1.0, + "content": "which indicates fair agreement", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 512, + 245, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 245, + 523 + ], + "score": 1.0, + "content": "between users (Viera et al., 2005).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.0, + "bbox_fs": [ + 105, + 457, + 377, + 523 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 543, + 426, + 554 + ], + "lines": [ + { + "bbox": [ + 105, + 542, + 428, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 428, + 555 + ], + "score": 1.0, + "content": "5.4 CORRELATION BETWEEN AUTOMATED METRICS AND HUMAN SCORES", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 566, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "We present the correlation between the automated metrics and our collected human ratings in Ta-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 577, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 590 + ], + "score": 1.0, + "content": "ble 4. We measure human v.s. human correlation by randomly splitting the human users into two", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "score": 1.0, + "content": "groups. The results indicate that in most cases, human scores correlate the best with other human", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 599, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 611 + ], + "score": 1.0, + "content": "scores. Except in the case of the Spearman correlation for BLEU-N scores, we can see that there is", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 610, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 622 + ], + "score": 1.0, + "content": "a positive correlation between the automated metrics and the human scores for these task-oriented", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 621, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 633 + ], + "score": 1.0, + "content": "datasets, which contrasts with the non task-oriented dialogue setting where Liu et al. (2016) observed", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 633, + 221, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 221, + 644 + ], + "score": 1.0, + "content": "no strong correlation trends.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 566, + 506, + 644 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 649, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "A likely explanation for the negative Spearman correlation values for BLEU-N is that there is only", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "one gold reference per context in the DSTC2 dataset. The Restaurants dataset, on the other hand,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "provides two gold references per context. Having multiple gold references increases the likelihood", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 681, + 502, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 502, + 695 + ], + "score": 1.0, + "content": "that the generated response will have significant word-overlap with one of the reference responses.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 649, + 505, + 695 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "We present scatter plots for some of the metrics presented in Table 4 in Figure 3. We observe that", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "all the metrics correlate very well with humans on high scoring examples. As it can be seen in the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "scatter plots, most of the sentences are given the maximal score of 5 by the human evaluators. This", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 210, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 505, + 223 + ], + "score": 1.0, + "content": "confirms our previous observation that the available corpora for task-oriented dialogue NLG task", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 221, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 480, + 235 + ], + "score": 1.0, + "content": "are not very challenging and a simple LSTM-based model can output high-quality responses.", + "type": "text", + "cross_page": true + } + ], + "index": 5 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 699, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 127, + 93, + 483, + 201 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 163, + 84, + 449, + 93 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 160, + 82, + 452, + 96 + ], + "spans": [ + { + "bbox": [ + 160, + 82, + 452, + 96 + ], + "score": 1.0, + "content": "Table 4: Correlation of automated metrics with human evaluations scores", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 127, + 93, + 483, + 201 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 93, + 483, + 201 + ], + "spans": [ + { + "bbox": [ + 127, + 93, + 483, + 201 + ], + "score": 0.985, + "html": "
MetricDSTC2Restaurants
Spearmanp-valuePearsonp-valueSpearmanp-valuePearsonp-value
Bleu1Bleu 2-0.317<0.0050.583<0.0050.0690.4940.2770.005
-0.318<0.0050.526<0.0050.0910.3660.1660.099
Bleu 3-0.318<0.0050.500<0.0050.1090.2800.2230.026
Bleu 4METEORROUGELSkip Thought-0.318<0.0050.461<0.0050.1050.2960.2550.010
0.295<0.0050.582<0.0050.353<0.0050.489<0.005
0.294<0.0050.448<0.0050.346<0.0050.382<0.005
0.528<0.0050.0860.3970.284<0.0050.364<0.005
Embedding Average0.295<0.0050.485<0.0050.423<0.0050.2600.009
Vector ExtremaGreedy Matching0.299<0.0050.624<0.0050.446<0.0050.287<0.005
0.295<0.0050.572<0.0050.446<0.0050.325<0.005
Human0.810<0.0050.984<0.0050.653<0.0050.857<0.005
", + "type": "table", + "image_path": "f97c5804c3aaa31ce0e6671a6a34b9ae69f2e34375917e374536be3806e7bca6.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 127, + 93, + 483, + 129.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 127, + 129.0, + 483, + 165.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 127, + 165.0, + 483, + 201.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 108, + 210, + 503, + 233 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 505, + 223 + ], + "score": 1.0, + "content": "confirms our previous observation that the available corpora for task-oriented dialogue NLG task", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 221, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 480, + 235 + ], + "score": 1.0, + "content": "are not very challenging and a simple LSTM-based model can output high-quality responses.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 238, + 505, + 338 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 505, + 250 + ], + "score": 1.0, + "content": "Overall, among the word overlap based automated metrics, METEOR consistently correlates with", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "human evaluation on both datasets. These results confirm the original findings by Banerjee & Lavie", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "(2005) who showed that METEOR had good correlation with human evaluation in the machine", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "translation task. The comparison with machine translation is highly relevant in the task-oriented", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 282, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 506, + 295 + ], + "score": 1.0, + "content": "setting because the NLG model essentially learns to translate the abstract representation of a sen-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 293, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 307 + ], + "score": 1.0, + "content": "tence into a sentence. It is a translation task contrary to the non task-oriented setting where the NLG", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "model needs to decide and output a new sentence based on the last sentence typed by a user and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 315, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 328 + ], + "score": 1.0, + "content": "dialogue history. Therefore, automated metrics coming from the machine translation literature are", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 327, + 469, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 469, + 339 + ], + "score": 1.0, + "content": "more adequate in our case than in the non-task oriented case as shown by Liu et al. (2016).", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 108, + 343, + 505, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "score": 1.0, + "content": "It is interesting to see that METEOR correlates well with human evaluation consistently. This can", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "be explained by the fact that even though METEOR does not rely on word embeddings, it includes", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "notions of synonymy and paraphrasing when computing the alignment between the candidate and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 376, + 191, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 191, + 388 + ], + "score": 1.0, + "content": "reference utterances.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 107, + 403, + 190, + 416 + ], + "lines": [ + { + "bbox": [ + 105, + 402, + 192, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 192, + 419 + ], + "score": 1.0, + "content": "6 DISCUSSION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 428, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "score": 1.0, + "content": "We evaluated several natural language generation models trained on the DSTC2 and the Restau-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 441, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 451 + ], + "score": 1.0, + "content": "rants datasets based on several automated metrics. We also performed human evaluation on the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 451, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 463 + ], + "score": 1.0, + "content": "model-generated responses and our study shows that human evaluation is a much more reliable met-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 462, + 504, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 504, + 473 + ], + "score": 1.0, + "content": "ric compared to the others. Among the word-overlap based automated metrics, we found that the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "score": 1.0, + "content": "METEOR score correlates the most with human judgments and we suggest using METEOR for task-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "score": 1.0, + "content": "oriented dialogue natural language generation instead of BLEU. We also observe that these metrics", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "are more reliable in the task-oriented dialogue setting compared to the general, non task-oriented", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 504, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 519 + ], + "score": 1.0, + "content": "one due to the limited possible diversity in the task-oriented setting. Also, as observed by Galley", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 517, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 528 + ], + "score": 1.0, + "content": "et al. (2015), we can see that word-overlap based metrics correlate better with human evaluation", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 528, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 540 + ], + "score": 1.0, + "content": "when multiple references are provided, as in the Restaurants dataset. Otherwise, as in the case of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 538, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 550 + ], + "score": 1.0, + "content": "DSTC2 which only provides one reference sentence per example, we observe that all the BLEU-N", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 550, + 414, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 414, + 561 + ], + "score": 1.0, + "content": "metrics negatively correlate with human evaluation on Spearman correlation.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 108, + 566, + 505, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 566, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 506, + 578 + ], + "score": 1.0, + "content": "As has been observed in the machine translation literature, using beam search improves the quality of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "generated sentences significantly compared to stochastic sampling. For similar models, our results", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "show improvement in the automated metrics’ scores compared to Wen et al. (2015b) who used", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 599, + 336, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 336, + 610 + ], + "score": 1.0, + "content": "stochastic sampling for decoding instead of beam search.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "Wen et al. (2015b) did not use the slot error rate penalty with the vanilla LSTM model in their", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "experiments. After adding the penalty in our case, we observe that the vanilla LSTM-based model", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "performs as well as the delexicalized semantically-controlled LSTM model. This suggests that the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "added complexity introduced by the sc-LSTM unit does not offer a significant advantage for these", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 659, + 160, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 160, + 672 + ], + "score": 1.0, + "content": "two datasets.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "High performance on automated metrics, achieved by our models on the DSTC2 and the Restaurants", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "datasets lead us to conclude that these datasets are not very challenging for the NLG task. The task-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "oriented dialogue community should move towards using larger and more complex datasets, which", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 709, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 721 + ], + "score": 1.0, + "content": "have been recently announced, such as the Frames dataset (El Asri et al., 2017) or the E2E NLG", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 720, + 276, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 276, + 733 + ], + "score": 1.0, + "content": "Challenge dataset (Novikova et al., 2016).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 127, + 93, + 483, + 201 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 163, + 84, + 449, + 93 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 160, + 82, + 452, + 96 + ], + "spans": [ + { + "bbox": [ + 160, + 82, + 452, + 96 + ], + "score": 1.0, + "content": "Table 4: Correlation of automated metrics with human evaluations scores", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 127, + 93, + 483, + 201 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 93, + 483, + 201 + ], + "spans": [ + { + "bbox": [ + 127, + 93, + 483, + 201 + ], + "score": 0.985, + "html": "
MetricDSTC2Restaurants
Spearmanp-valuePearsonp-valueSpearmanp-valuePearsonp-value
Bleu1Bleu 2-0.317<0.0050.583<0.0050.0690.4940.2770.005
-0.318<0.0050.526<0.0050.0910.3660.1660.099
Bleu 3-0.318<0.0050.500<0.0050.1090.2800.2230.026
Bleu 4METEORROUGELSkip Thought-0.318<0.0050.461<0.0050.1050.2960.2550.010
0.295<0.0050.582<0.0050.353<0.0050.489<0.005
0.294<0.0050.448<0.0050.346<0.0050.382<0.005
0.528<0.0050.0860.3970.284<0.0050.364<0.005
Embedding Average0.295<0.0050.485<0.0050.423<0.0050.2600.009
Vector ExtremaGreedy Matching0.299<0.0050.624<0.0050.446<0.0050.287<0.005
0.295<0.0050.572<0.0050.446<0.0050.325<0.005
Human0.810<0.0050.984<0.0050.653<0.0050.857<0.005
", + "type": "table", + "image_path": "f97c5804c3aaa31ce0e6671a6a34b9ae69f2e34375917e374536be3806e7bca6.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 127, + 93, + 483, + 129.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 127, + 129.0, + 483, + 165.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 127, + 165.0, + 483, + 201.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 108, + 210, + 503, + 233 + ], + "lines": [], + "index": 4.5, + "bbox_fs": [ + 105, + 210, + 505, + 235 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 238, + 505, + 338 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 505, + 250 + ], + "score": 1.0, + "content": "Overall, among the word overlap based automated metrics, METEOR consistently correlates with", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "human evaluation on both datasets. These results confirm the original findings by Banerjee & Lavie", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "(2005) who showed that METEOR had good correlation with human evaluation in the machine", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "translation task. The comparison with machine translation is highly relevant in the task-oriented", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 282, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 506, + 295 + ], + "score": 1.0, + "content": "setting because the NLG model essentially learns to translate the abstract representation of a sen-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 293, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 307 + ], + "score": 1.0, + "content": "tence into a sentence. It is a translation task contrary to the non task-oriented setting where the NLG", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "model needs to decide and output a new sentence based on the last sentence typed by a user and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 315, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 328 + ], + "score": 1.0, + "content": "dialogue history. Therefore, automated metrics coming from the machine translation literature are", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 327, + 469, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 469, + 339 + ], + "score": 1.0, + "content": "more adequate in our case than in the non-task oriented case as shown by Liu et al. (2016).", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 239, + 506, + 339 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 343, + 505, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "score": 1.0, + "content": "It is interesting to see that METEOR correlates well with human evaluation consistently. This can", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "be explained by the fact that even though METEOR does not rely on word embeddings, it includes", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "notions of synonymy and paraphrasing when computing the alignment between the candidate and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 376, + 191, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 191, + 388 + ], + "score": 1.0, + "content": "reference utterances.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 343, + 505, + 388 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 403, + 190, + 416 + ], + "lines": [ + { + "bbox": [ + 105, + 402, + 192, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 192, + 419 + ], + "score": 1.0, + "content": "6 DISCUSSION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 428, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "score": 1.0, + "content": "We evaluated several natural language generation models trained on the DSTC2 and the Restau-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 441, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 451 + ], + "score": 1.0, + "content": "rants datasets based on several automated metrics. We also performed human evaluation on the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 451, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 463 + ], + "score": 1.0, + "content": "model-generated responses and our study shows that human evaluation is a much more reliable met-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 462, + 504, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 504, + 473 + ], + "score": 1.0, + "content": "ric compared to the others. Among the word-overlap based automated metrics, we found that the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "score": 1.0, + "content": "METEOR score correlates the most with human judgments and we suggest using METEOR for task-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "score": 1.0, + "content": "oriented dialogue natural language generation instead of BLEU. We also observe that these metrics", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "are more reliable in the task-oriented dialogue setting compared to the general, non task-oriented", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 504, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 519 + ], + "score": 1.0, + "content": "one due to the limited possible diversity in the task-oriented setting. Also, as observed by Galley", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 517, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 528 + ], + "score": 1.0, + "content": "et al. (2015), we can see that word-overlap based metrics correlate better with human evaluation", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 528, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 540 + ], + "score": 1.0, + "content": "when multiple references are provided, as in the Restaurants dataset. Otherwise, as in the case of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 538, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 550 + ], + "score": 1.0, + "content": "DSTC2 which only provides one reference sentence per example, we observe that all the BLEU-N", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 550, + 414, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 414, + 561 + ], + "score": 1.0, + "content": "metrics negatively correlate with human evaluation on Spearman correlation.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 428, + 506, + 561 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 566, + 505, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 566, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 506, + 578 + ], + "score": 1.0, + "content": "As has been observed in the machine translation literature, using beam search improves the quality of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "generated sentences significantly compared to stochastic sampling. For similar models, our results", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "show improvement in the automated metrics’ scores compared to Wen et al. (2015b) who used", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 599, + 336, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 336, + 610 + ], + "score": 1.0, + "content": "stochastic sampling for decoding instead of beam search.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 566, + 506, + 610 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "Wen et al. (2015b) did not use the slot error rate penalty with the vanilla LSTM model in their", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "experiments. After adding the penalty in our case, we observe that the vanilla LSTM-based model", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "performs as well as the delexicalized semantically-controlled LSTM model. This suggests that the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "added complexity introduced by the sc-LSTM unit does not offer a significant advantage for these", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 659, + 160, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 160, + 672 + ], + "score": 1.0, + "content": "two datasets.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 615, + 506, + 672 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "High performance on automated metrics, achieved by our models on the DSTC2 and the Restaurants", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "datasets lead us to conclude that these datasets are not very challenging for the NLG task. The task-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "oriented dialogue community should move towards using larger and more complex datasets, which", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 709, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 721 + ], + "score": 1.0, + "content": "have been recently announced, such as the Frames dataset (El Asri et al., 2017) or the E2E NLG", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 720, + 276, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 276, + 733 + ], + "score": 1.0, + "content": "Challenge dataset (Novikova et al., 2016).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 676, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 175, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 449, + 111 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 451, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 451, + 113 + ], + "score": 1.0, + "content": "John Langshaw Austin. How to do things with words. Oxford University Press, 1962.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 118, + 505, + 163 + ], + "lines": [ + { + "bbox": [ + 105, + 117, + 505, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 132 + ], + "score": 1.0, + "content": "Scott Axelrod. Natural language generation in the ibm flight information system. In Proceedings", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 129, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 115, + 129, + 505, + 141 + ], + "score": 1.0, + "content": "of the 2000 ANLP/NAACL Workshop on Conversational Systems - Volume 3, ANLP/NAACL-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 141, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 115, + 141, + 505, + 154 + ], + "score": 1.0, + "content": "ConvSyst ’00, pp. 21–26, Stroudsburg, PA, USA, 2000. Association for Computational Linguis-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 151, + 268, + 163 + ], + "spans": [ + { + "bbox": [ + 115, + 151, + 268, + 163 + ], + "score": 1.0, + "content": "tics. doi: 10.3115/1117562.1117567.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 170, + 504, + 204 + ], + "lines": [ + { + "bbox": [ + 106, + 170, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 505, + 182 + ], + "score": 1.0, + "content": "Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 116, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "correlation with human judgments. In Proc. ACL workshop on intrinsic and extrinsic evaluation", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 192, + 371, + 204 + ], + "spans": [ + { + "bbox": [ + 116, + 192, + 371, + 204 + ], + "score": 1.0, + "content": "measures for machine translation and/or summarization, 2005.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 103, + 210, + 504, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 504, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 504, + 224 + ], + "score": 1.0, + "content": "Jacob Cohen. A coefficient of agreement for nominal scales. Educational and psychological mea-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 222, + 236, + 233 + ], + "spans": [ + { + "bbox": [ + 115, + 222, + 236, + 233 + ], + "score": 1.0, + "content": "surement, 20(1):37–46, 1960.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 240, + 505, + 285 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 505, + 254 + ], + "score": 1.0, + "content": "Ondˇrej Dusek and Filip Jurcicek. Sequence-to-sequence generation for spoken dialogue via deep ˇ", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 115, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "syntax trees and strings. In Proceedings of the 54th Annual Meeting of the Association for Com-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 263, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 115, + 263, + 505, + 275 + ], + "score": 1.0, + "content": "putational Linguistics (Volume 2: Short Papers), pp. 45–51, Berlin, Germany, August 2016. As-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 274, + 280, + 286 + ], + "spans": [ + { + "bbox": [ + 115, + 274, + 280, + 286 + ], + "score": 1.0, + "content": "sociation for Computational Linguistics.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 292, + 505, + 326 + ], + "lines": [ + { + "bbox": [ + 105, + 291, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 305 + ], + "score": 1.0, + "content": "Layla El Asri, Hannes Schulz, Shikhar Sharma, Jeremie Zumer, Justin Harris, Emery Fine, Rahul", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 302, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 115, + 302, + 505, + 316 + ], + "score": 1.0, + "content": "Mehrotra, and Kaheer Suleman. Frames: A corpus for adding memory to goal-oriented dialogue", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 315, + 316, + 326 + ], + "spans": [ + { + "bbox": [ + 115, + 315, + 316, + 326 + ], + "score": 1.0, + "content": "systems. arXiv preprint arXiv:1704.00057, 2017.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 333, + 504, + 355 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 346 + ], + "score": 1.0, + "content": "Michael Elhadad. Generating coherent argumentative paragraphs. In In Proceedings of COLING", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 118, + 343, + 257, + 356 + ], + "spans": [ + { + "bbox": [ + 118, + 343, + 257, + 356 + ], + "score": 1.0, + "content": "’92, volume II, pp. 638–644, 1992.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 362, + 503, + 385 + ], + "lines": [ + { + "bbox": [ + 105, + 360, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 505, + 376 + ], + "score": 1.0, + "content": "Gabriel Forgues, Joelle Pineau, Jean-Marie Larcheveque, and Real Tremblay. Bootstrapping dialog", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 374, + 273, + 386 + ], + "spans": [ + { + "bbox": [ + 115, + 374, + 273, + 386 + ], + "score": 1.0, + "content": "systems with word embeddings. 2014.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 392, + 505, + 459 + ], + "lines": [ + { + "bbox": [ + 106, + 392, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 505, + 404 + ], + "score": 1.0, + "content": "Michel Galley, Chris Brockett, Alessandro Sordoni, Yangfeng Ji, Michael Auli, Chris Quirk, Mar-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 402, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 115, + 402, + 506, + 416 + ], + "score": 1.0, + "content": "garet Mitchell, Jianfeng Gao, and Bill Dolan. deltableu: A discriminative metric for generation", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 115, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "tasks with intrinsically diverse targets. In Proceedings of the 53rd Annual Meeting of the As-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 115, + 425, + 505, + 437 + ], + "score": 1.0, + "content": "sociation for Computational Linguistics and the 7th International Joint Conference on Natural", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 436, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 116, + 436, + 505, + 449 + ], + "score": 1.0, + "content": "Language Processing (Volume 2: Short Papers), pp. 445–450, Beijing, China, July 2015. Associ-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 447, + 263, + 460 + ], + "spans": [ + { + "bbox": [ + 116, + 447, + 263, + 460 + ], + "score": 1.0, + "content": "ation for Computational Linguistics.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 465, + 505, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 479 + ], + "score": 1.0, + "content": "Matthew Henderson, Blaise Thomson, and Jason D. Williams. The second dialog state tracking", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 116, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "challenge. In Proceedings of the 15th Annual Meeting of the Special Interest Group on Discourse", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 487, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 115, + 487, + 506, + 501 + ], + "score": 1.0, + "content": "and Dialogue (SIGDIAL), pp. 263–272, Philadelphia, PA, U.S.A., June 2014. Association for", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 117, + 499, + 227, + 511 + ], + "spans": [ + { + "bbox": [ + 117, + 499, + 227, + 511 + ], + "score": 1.0, + "content": "Computational Linguistics.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 105, + 517, + 504, + 540 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 117, + 528, + 191, + 540 + ], + "spans": [ + { + "bbox": [ + 117, + 528, + 191, + 540 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 105, + 546, + 504, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "Anjuli Kannan and Oriol Vinyals. Adversarial evaluation of dialogue models. In NIPS 2016 Work-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 558, + 264, + 570 + ], + "spans": [ + { + "bbox": [ + 115, + 558, + 264, + 570 + ], + "score": 1.0, + "content": "shop on Adversarial Training, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 576, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "Ryan Kiros, Yukun Zhu, Ruslan R Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio Tor-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 114, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "ralba, and Sanja Fidler. Skip-thought vectors. In C. Cortes, N. D. Lawrence, D. D. Lee,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 598, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 115, + 598, + 505, + 612 + ], + "score": 1.0, + "content": "M. Sugiyama, and R. Garnett (eds.), Advances in Neural Information Processing Systems 28,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 609, + 304, + 622 + ], + "spans": [ + { + "bbox": [ + 115, + 609, + 304, + 622 + ], + "score": 1.0, + "content": "pp. 3294–3302. Curran Associates, Inc., 2015.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 106, + 628, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 105, + 627, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 641 + ], + "score": 1.0, + "content": "Jiwei Li, Will Monroe, Alan Ritter, Dan Jurafsky, Michel Galley, and Jianfeng Gao. Deep rein-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 116, + 639, + 505, + 651 + ], + "score": 1.0, + "content": "forcement learning for dialogue generation. In Proceedings of the 2016 Conference on Empirical", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 115, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "Methods in Natural Language Processing, pp. 1192–1202, Austin, Texas, November 2016. Asso-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 662, + 271, + 675 + ], + "spans": [ + { + "bbox": [ + 115, + 662, + 271, + 675 + ], + "score": 1.0, + "content": "ciation for Computational Linguistics.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 106, + 680, + 503, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "Jiwei Li, Will Monroe, Tianlin Shi, Alan Ritter, and Dan Jurafsky. Adversarial learning for neural", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 690, + 324, + 703 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 324, + 703 + ], + "score": 1.0, + "content": "dialogue generation. CoRR, abs/1701.06547, 2017.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Proc. ACL workshop", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 721, + 294, + 731 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 294, + 731 + ], + "score": 1.0, + "content": "on Text Summarization Branches Out, 2004.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 175, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 449, + 111 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 451, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 451, + 113 + ], + "score": 1.0, + "content": "John Langshaw Austin. How to do things with words. Oxford University Press, 1962.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 99, + 451, + 113 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 118, + 505, + 163 + ], + "lines": [ + { + "bbox": [ + 105, + 117, + 505, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 132 + ], + "score": 1.0, + "content": "Scott Axelrod. Natural language generation in the ibm flight information system. In Proceedings", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 129, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 115, + 129, + 505, + 141 + ], + "score": 1.0, + "content": "of the 2000 ANLP/NAACL Workshop on Conversational Systems - Volume 3, ANLP/NAACL-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 141, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 115, + 141, + 505, + 154 + ], + "score": 1.0, + "content": "ConvSyst ’00, pp. 21–26, Stroudsburg, PA, USA, 2000. Association for Computational Linguis-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 151, + 268, + 163 + ], + "spans": [ + { + "bbox": [ + 115, + 151, + 268, + 163 + ], + "score": 1.0, + "content": "tics. doi: 10.3115/1117562.1117567.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 117, + 505, + 163 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 170, + 504, + 204 + ], + "lines": [ + { + "bbox": [ + 106, + 170, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 505, + 182 + ], + "score": 1.0, + "content": "Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 116, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "correlation with human judgments. In Proc. ACL workshop on intrinsic and extrinsic evaluation", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 192, + 371, + 204 + ], + "spans": [ + { + "bbox": [ + 116, + 192, + 371, + 204 + ], + "score": 1.0, + "content": "measures for machine translation and/or summarization, 2005.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 170, + 505, + 204 + ] + }, + { + "type": "text", + "bbox": [ + 103, + 210, + 504, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 504, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 504, + 224 + ], + "score": 1.0, + "content": "Jacob Cohen. A coefficient of agreement for nominal scales. Educational and psychological mea-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 222, + 236, + 233 + ], + "spans": [ + { + "bbox": [ + 115, + 222, + 236, + 233 + ], + "score": 1.0, + "content": "surement, 20(1):37–46, 1960.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 209, + 504, + 233 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 240, + 505, + 285 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 505, + 254 + ], + "score": 1.0, + "content": "Ondˇrej Dusek and Filip Jurcicek. Sequence-to-sequence generation for spoken dialogue via deep ˇ", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 115, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "syntax trees and strings. In Proceedings of the 54th Annual Meeting of the Association for Com-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 263, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 115, + 263, + 505, + 275 + ], + "score": 1.0, + "content": "putational Linguistics (Volume 2: Short Papers), pp. 45–51, Berlin, Germany, August 2016. As-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 274, + 280, + 286 + ], + "spans": [ + { + "bbox": [ + 115, + 274, + 280, + 286 + ], + "score": 1.0, + "content": "sociation for Computational Linguistics.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 106, + 240, + 505, + 286 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 292, + 505, + 326 + ], + "lines": [ + { + "bbox": [ + 105, + 291, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 305 + ], + "score": 1.0, + "content": "Layla El Asri, Hannes Schulz, Shikhar Sharma, Jeremie Zumer, Justin Harris, Emery Fine, Rahul", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 302, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 115, + 302, + 505, + 316 + ], + "score": 1.0, + "content": "Mehrotra, and Kaheer Suleman. Frames: A corpus for adding memory to goal-oriented dialogue", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 315, + 316, + 326 + ], + "spans": [ + { + "bbox": [ + 115, + 315, + 316, + 326 + ], + "score": 1.0, + "content": "systems. arXiv preprint arXiv:1704.00057, 2017.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 291, + 505, + 326 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 333, + 504, + 355 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 346 + ], + "score": 1.0, + "content": "Michael Elhadad. Generating coherent argumentative paragraphs. In In Proceedings of COLING", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 118, + 343, + 257, + 356 + ], + "spans": [ + { + "bbox": [ + 118, + 343, + 257, + 356 + ], + "score": 1.0, + "content": "’92, volume II, pp. 638–644, 1992.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 331, + 505, + 356 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 362, + 503, + 385 + ], + "lines": [ + { + "bbox": [ + 105, + 360, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 505, + 376 + ], + "score": 1.0, + "content": "Gabriel Forgues, Joelle Pineau, Jean-Marie Larcheveque, and Real Tremblay. Bootstrapping dialog", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 374, + 273, + 386 + ], + "spans": [ + { + "bbox": [ + 115, + 374, + 273, + 386 + ], + "score": 1.0, + "content": "systems with word embeddings. 2014.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 360, + 505, + 386 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 392, + 505, + 459 + ], + "lines": [ + { + "bbox": [ + 106, + 392, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 505, + 404 + ], + "score": 1.0, + "content": "Michel Galley, Chris Brockett, Alessandro Sordoni, Yangfeng Ji, Michael Auli, Chris Quirk, Mar-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 402, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 115, + 402, + 506, + 416 + ], + "score": 1.0, + "content": "garet Mitchell, Jianfeng Gao, and Bill Dolan. deltableu: A discriminative metric for generation", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 115, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "tasks with intrinsically diverse targets. In Proceedings of the 53rd Annual Meeting of the As-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 115, + 425, + 505, + 437 + ], + "score": 1.0, + "content": "sociation for Computational Linguistics and the 7th International Joint Conference on Natural", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 436, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 116, + 436, + 505, + 449 + ], + "score": 1.0, + "content": "Language Processing (Volume 2: Short Papers), pp. 445–450, Beijing, China, July 2015. Associ-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 447, + 263, + 460 + ], + "spans": [ + { + "bbox": [ + 116, + 447, + 263, + 460 + ], + "score": 1.0, + "content": "ation for Computational Linguistics.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5, + "bbox_fs": [ + 106, + 392, + 506, + 460 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 465, + 505, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 479 + ], + "score": 1.0, + "content": "Matthew Henderson, Blaise Thomson, and Jason D. Williams. The second dialog state tracking", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 116, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "challenge. In Proceedings of the 15th Annual Meeting of the Special Interest Group on Discourse", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 487, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 115, + 487, + 506, + 501 + ], + "score": 1.0, + "content": "and Dialogue (SIGDIAL), pp. 263–272, Philadelphia, PA, U.S.A., June 2014. Association for", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 117, + 499, + 227, + 511 + ], + "spans": [ + { + "bbox": [ + 117, + 499, + 227, + 511 + ], + "score": 1.0, + "content": "Computational Linguistics.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 464, + 506, + 511 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 517, + 504, + 540 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 117, + 528, + 191, + 540 + ], + "spans": [ + { + "bbox": [ + 117, + 528, + 191, + 540 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 516, + 505, + 540 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 546, + 504, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "Anjuli Kannan and Oriol Vinyals. Adversarial evaluation of dialogue models. In NIPS 2016 Work-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 558, + 264, + 570 + ], + "spans": [ + { + "bbox": [ + 115, + 558, + 264, + 570 + ], + "score": 1.0, + "content": "shop on Adversarial Training, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 106, + 546, + 505, + 570 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 576, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "Ryan Kiros, Yukun Zhu, Ruslan R Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio Tor-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 114, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "ralba, and Sanja Fidler. Skip-thought vectors. In C. Cortes, N. D. Lawrence, D. D. Lee,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 598, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 115, + 598, + 505, + 612 + ], + "score": 1.0, + "content": "M. Sugiyama, and R. Garnett (eds.), Advances in Neural Information Processing Systems 28,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 609, + 304, + 622 + ], + "spans": [ + { + "bbox": [ + 115, + 609, + 304, + 622 + ], + "score": 1.0, + "content": "pp. 3294–3302. Curran Associates, Inc., 2015.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 576, + 505, + 622 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 628, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 105, + 627, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 641 + ], + "score": 1.0, + "content": "Jiwei Li, Will Monroe, Alan Ritter, Dan Jurafsky, Michel Galley, and Jianfeng Gao. Deep rein-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 116, + 639, + 505, + 651 + ], + "score": 1.0, + "content": "forcement learning for dialogue generation. In Proceedings of the 2016 Conference on Empirical", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 115, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "Methods in Natural Language Processing, pp. 1192–1202, Austin, Texas, November 2016. Asso-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 662, + 271, + 675 + ], + "spans": [ + { + "bbox": [ + 115, + 662, + 271, + 675 + ], + "score": 1.0, + "content": "ciation for Computational Linguistics.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 627, + 505, + 675 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 680, + 503, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "Jiwei Li, Will Monroe, Tianlin Shi, Alan Ritter, and Dan Jurafsky. Adversarial learning for neural", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 690, + 324, + 703 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 324, + 703 + ], + "score": 1.0, + "content": "dialogue generation. CoRR, abs/1701.06547, 2017.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 106, + 680, + 505, + 703 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Proc. ACL workshop", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 721, + 294, + 731 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 294, + 731 + ], + "score": 1.0, + "content": "on Text Summarization Branches Out, 2004.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5, + "bbox_fs": [ + 106, + 708, + 505, + 731 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Chia-Wei Liu, Ryan Lowe, Iulian Serban, Mike Noseworthy, Laurent Charlin, and Joelle Pineau.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "How not to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "for dialogue response generation. In Proceedings of the 2016 Conference on Empirical Methods in", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 115, + 116, + 505, + 128 + ], + "score": 1.0, + "content": "Natural Language Processing, pp. 2122–2132. Association for Computational Linguistics, 2016.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 108, + 132, + 504, + 167 + ], + "lines": [ + { + "bbox": [ + 105, + 132, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 146 + ], + "score": 1.0, + "content": "Ryan Lowe, Nissan Pow, IV Serban, Laurent Charlin, and Joelle Pineau. Incorporating unstruc-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 143, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 505, + 157 + ], + "score": 1.0, + "content": "tured textual knowledge sources into neural dialogue systems. In Neural Information Processing", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 155, + 459, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 155, + 459, + 168 + ], + "score": 1.0, + "content": "Systems Workshop on Machine Learning for Spoken Language Understanding, 2015.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 171, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 106, + 172, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 505, + 184 + ], + "score": 1.0, + "content": "Ryan Lowe, Michael Noseworthy, Iulian V. Serban, Nicolas Angelard-Gontier, Yoshua Bengio, and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 182, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 115, + 182, + 505, + 195 + ], + "score": 1.0, + "content": "Joelle Pineau. Towards an automatic turing test: Learning to evaluate dialogue responses. In", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 193, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 115, + 193, + 505, + 207 + ], + "score": 1.0, + "content": "Proceedings of the 5th International Conference on Learning Representations (ICLR) Workshop,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 204, + 230, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 204, + 230, + 217 + ], + "score": 1.0, + "content": "Toulon, France, April 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 221, + 503, + 256 + ], + "lines": [ + { + "bbox": [ + 105, + 221, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 505, + 235 + ], + "score": 1.0, + "content": "Jekaterina Novikova, Oliver Lemon, and Verena Rieser. Crowd-sourcing nlg data: Pictures elicit", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 232, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 232, + 505, + 248 + ], + "score": 1.0, + "content": "better data. In Proceedings of the 9th International Natural Language Generation conference, pp.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 244, + 483, + 258 + ], + "spans": [ + { + "bbox": [ + 116, + 244, + 483, + 258 + ], + "score": 1.0, + "content": "265–273, Edinburgh, UK, September 5-8 2016. Association for Computational Linguistics.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 105, + 261, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 275 + ], + "score": 1.0, + "content": "Alice H. Oh and Alexander I. Rudnicky. Stochastic language generation for spoken dialogue sys-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 272, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 115, + 272, + 505, + 285 + ], + "score": 1.0, + "content": "tems. In Proceedings of the 2000 ANLP/NAACL Workshop on Conversational Systems - Volume", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 284, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 115, + 284, + 505, + 296 + ], + "score": 1.0, + "content": "3, ANLP/NAACL-ConvSyst ’00, pp. 27–32, Stroudsburg, PA, USA, 2000. Association for Com-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 294, + 340, + 307 + ], + "spans": [ + { + "bbox": [ + 115, + 294, + 340, + 307 + ], + "score": 1.0, + "content": "putational Linguistics. doi: 10.3115/1117562.1117568.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 104, + 311, + 504, + 335 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 323, + 348, + 335 + ], + "spans": [ + { + "bbox": [ + 116, + 323, + 348, + 335 + ], + "score": 1.0, + "content": "evaluation of machine translation. In Proc. of ACL, 2002.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 340, + 505, + 385 + ], + "lines": [ + { + "bbox": [ + 106, + 339, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 505, + 353 + ], + "score": 1.0, + "content": "Verena Rieser and Oliver Lemon. Natural language generation as planning under uncertainty for", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 351, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 116, + 351, + 505, + 363 + ], + "score": 1.0, + "content": "spoken dialogue systems. In Proceedings of the 12th Conference of the European Chapter of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 361, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 115, + 361, + 505, + 375 + ], + "score": 1.0, + "content": "ACL (EACL 2009), pp. 683–691, Athens, Greece, March 2009. Association for Computational", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 114, + 372, + 167, + 387 + ], + "spans": [ + { + "bbox": [ + 114, + 372, + 167, + 387 + ], + "score": 1.0, + "content": "Linguistics.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 390, + 504, + 424 + ], + "lines": [ + { + "bbox": [ + 106, + 390, + 504, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 504, + 403 + ], + "score": 1.0, + "content": "Vasile Rus and Mihai C. Lintean. A comparison of greedy and optimal assessment of natural lan-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 402, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 116, + 402, + 505, + 413 + ], + "score": 1.0, + "content": "guage student input using word-to-word similarity metrics. In Proceedings of the Seventh Work-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 412, + 366, + 425 + ], + "spans": [ + { + "bbox": [ + 116, + 412, + 366, + 425 + ], + "score": 1.0, + "content": "shop on Building Educational Applications Using NLP, 2012.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 108, + 429, + 503, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "score": 1.0, + "content": "Iulian Vlad Serban, Alessandro Sordoni, Ryan Lowe, Laurent Charlin, Joelle Pineau, Aaron C.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 116, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "Courville, and Yoshua Bengio. A hierarchical latent variable encoder-decoder model for generat-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 452, + 298, + 464 + ], + "spans": [ + { + "bbox": [ + 115, + 452, + 298, + 464 + ], + "score": 1.0, + "content": "ing dialogues. CoRR, abs/1605.06069, 2016.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 468, + 504, + 503 + ], + "lines": [ + { + "bbox": [ + 105, + 466, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 506, + 483 + ], + "score": 1.0, + "content": "Shikhar Sharma, Jing He, Kaheer Suleman, Hannes Schulz, and Philip Bachman. Natural language", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 480, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 115, + 480, + 506, + 493 + ], + "score": 1.0, + "content": "generation in dialogue using lexicalized and delexicalized data. In Proceedings of the 5th Inter-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 491, + 418, + 504 + ], + "spans": [ + { + "bbox": [ + 116, + 491, + 418, + 504 + ], + "score": 1.0, + "content": "national Conference on Learning Representations (ICLR) Workshop, 2017.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 104, + 508, + 504, + 531 + ], + "lines": [ + { + "bbox": [ + 106, + 508, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 505, + 521 + ], + "score": 1.0, + "content": "Anthony J Viera, Joanne M Garrett, et al. Understanding interobserver agreement: the kappa statis-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 519, + 293, + 531 + ], + "spans": [ + { + "bbox": [ + 115, + 519, + 293, + 531 + ], + "score": 1.0, + "content": "tic. Family Medicine, 37(5):360–363, 2005.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 536, + 504, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "Marilyn Walker, Amanda Stent, Franc¸ois Mairesse, and Rashmi Prasad. Individual and domain", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 115, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "adaptation in sentence planning for dialogue. J. Artif. Int. Res., 30(1):413–456, November 2007.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 559, + 189, + 570 + ], + "spans": [ + { + "bbox": [ + 116, + 559, + 189, + 570 + ], + "score": 1.0, + "content": "ISSN 1076-9757.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 575, + 505, + 631 + ], + "lines": [ + { + "bbox": [ + 106, + 575, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 505, + 588 + ], + "score": 1.0, + "content": "Tsung-Hsien Wen, Milica Gasiˇ c, Dongho Kim, Nikola Mrk ´ siˇ c, Pei-Hao Su, David Vandyke, and ´", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 586, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 116, + 586, + 505, + 600 + ], + "score": 1.0, + "content": "Steve Young. Stochastic Language Generation in Dialogue using Recurrent Neural Networks with", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 597, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 115, + 597, + 505, + 611 + ], + "score": 1.0, + "content": "Convolutional Sentence Reranking. In Proceedings of the 16th Annual Meeting of the Special In-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 114, + 607, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 114, + 607, + 506, + 623 + ], + "score": 1.0, + "content": "terest Group on Discourse and Dialogue (SIGDIAL). Association for Computational Linguistics,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 619, + 191, + 633 + ], + "spans": [ + { + "bbox": [ + 115, + 619, + 191, + 633 + ], + "score": 1.0, + "content": "September 2015a.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 636, + 505, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 505, + 651 + ], + "score": 1.0, + "content": "Tsung-Hsien Wen, Milica Gasiˇ c, Nikola Mrk ´ siˇ c, Pei-Hao Su, David Vandyke, and Steve Young. ´", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 114, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 114, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "Semantically conditioned LSTM-based natural language generation for spoken dialogue systems.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 114, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 114, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 670, + 410, + 683 + ], + "spans": [ + { + "bbox": [ + 116, + 670, + 410, + 683 + ], + "score": 1.0, + "content": "(EMNLP). Association for Computational Linguistics, September 2015b.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 697, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 115, + 697, + 505, + 712 + ], + "score": 1.0, + "content": "Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 115, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "movies and reading books. In The IEEE International Conference on Computer Vision (ICCV),", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 116, + 721, + 185, + 731 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 185, + 731 + ], + "score": 1.0, + "content": "December 2015.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Chia-Wei Liu, Ryan Lowe, Iulian Serban, Mike Noseworthy, Laurent Charlin, and Joelle Pineau.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "How not to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "for dialogue response generation. In Proceedings of the 2016 Conference on Empirical Methods in", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 115, + 116, + 505, + 128 + ], + "score": 1.0, + "content": "Natural Language Processing, pp. 2122–2132. Association for Computational Linguistics, 2016.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 82, + 506, + 128 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 132, + 504, + 167 + ], + "lines": [ + { + "bbox": [ + 105, + 132, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 146 + ], + "score": 1.0, + "content": "Ryan Lowe, Nissan Pow, IV Serban, Laurent Charlin, and Joelle Pineau. Incorporating unstruc-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 143, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 505, + 157 + ], + "score": 1.0, + "content": "tured textual knowledge sources into neural dialogue systems. In Neural Information Processing", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 155, + 459, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 155, + 459, + 168 + ], + "score": 1.0, + "content": "Systems Workshop on Machine Learning for Spoken Language Understanding, 2015.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 132, + 505, + 168 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 171, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 106, + 172, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 505, + 184 + ], + "score": 1.0, + "content": "Ryan Lowe, Michael Noseworthy, Iulian V. Serban, Nicolas Angelard-Gontier, Yoshua Bengio, and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 182, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 115, + 182, + 505, + 195 + ], + "score": 1.0, + "content": "Joelle Pineau. Towards an automatic turing test: Learning to evaluate dialogue responses. In", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 193, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 115, + 193, + 505, + 207 + ], + "score": 1.0, + "content": "Proceedings of the 5th International Conference on Learning Representations (ICLR) Workshop,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 204, + 230, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 204, + 230, + 217 + ], + "score": 1.0, + "content": "Toulon, France, April 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5, + "bbox_fs": [ + 106, + 172, + 505, + 217 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 221, + 503, + 256 + ], + "lines": [ + { + "bbox": [ + 105, + 221, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 505, + 235 + ], + "score": 1.0, + "content": "Jekaterina Novikova, Oliver Lemon, and Verena Rieser. Crowd-sourcing nlg data: Pictures elicit", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 232, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 232, + 505, + 248 + ], + "score": 1.0, + "content": "better data. In Proceedings of the 9th International Natural Language Generation conference, pp.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 244, + 483, + 258 + ], + "spans": [ + { + "bbox": [ + 116, + 244, + 483, + 258 + ], + "score": 1.0, + "content": "265–273, Edinburgh, UK, September 5-8 2016. Association for Computational Linguistics.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 221, + 505, + 258 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 105, + 261, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 275 + ], + "score": 1.0, + "content": "Alice H. Oh and Alexander I. Rudnicky. Stochastic language generation for spoken dialogue sys-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 272, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 115, + 272, + 505, + 285 + ], + "score": 1.0, + "content": "tems. In Proceedings of the 2000 ANLP/NAACL Workshop on Conversational Systems - Volume", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 284, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 115, + 284, + 505, + 296 + ], + "score": 1.0, + "content": "3, ANLP/NAACL-ConvSyst ’00, pp. 27–32, Stroudsburg, PA, USA, 2000. Association for Com-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 294, + 340, + 307 + ], + "spans": [ + { + "bbox": [ + 115, + 294, + 340, + 307 + ], + "score": 1.0, + "content": "putational Linguistics. doi: 10.3115/1117562.1117568.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 261, + 505, + 307 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 311, + 504, + 335 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 323, + 348, + 335 + ], + "spans": [ + { + "bbox": [ + 116, + 323, + 348, + 335 + ], + "score": 1.0, + "content": "evaluation of machine translation. In Proc. of ACL, 2002.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 311, + 505, + 335 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 340, + 505, + 385 + ], + "lines": [ + { + "bbox": [ + 106, + 339, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 505, + 353 + ], + "score": 1.0, + "content": "Verena Rieser and Oliver Lemon. Natural language generation as planning under uncertainty for", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 351, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 116, + 351, + 505, + 363 + ], + "score": 1.0, + "content": "spoken dialogue systems. In Proceedings of the 12th Conference of the European Chapter of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 361, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 115, + 361, + 505, + 375 + ], + "score": 1.0, + "content": "ACL (EACL 2009), pp. 683–691, Athens, Greece, March 2009. Association for Computational", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 114, + 372, + 167, + 387 + ], + "spans": [ + { + "bbox": [ + 114, + 372, + 167, + 387 + ], + "score": 1.0, + "content": "Linguistics.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5, + "bbox_fs": [ + 106, + 339, + 505, + 387 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 390, + 504, + 424 + ], + "lines": [ + { + "bbox": [ + 106, + 390, + 504, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 504, + 403 + ], + "score": 1.0, + "content": "Vasile Rus and Mihai C. Lintean. A comparison of greedy and optimal assessment of natural lan-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 402, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 116, + 402, + 505, + 413 + ], + "score": 1.0, + "content": "guage student input using word-to-word similarity metrics. In Proceedings of the Seventh Work-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 412, + 366, + 425 + ], + "spans": [ + { + "bbox": [ + 116, + 412, + 366, + 425 + ], + "score": 1.0, + "content": "shop on Building Educational Applications Using NLP, 2012.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 390, + 505, + 425 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 429, + 503, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "score": 1.0, + "content": "Iulian Vlad Serban, Alessandro Sordoni, Ryan Lowe, Laurent Charlin, Joelle Pineau, Aaron C.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 116, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "Courville, and Yoshua Bengio. A hierarchical latent variable encoder-decoder model for generat-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 452, + 298, + 464 + ], + "spans": [ + { + "bbox": [ + 115, + 452, + 298, + 464 + ], + "score": 1.0, + "content": "ing dialogues. CoRR, abs/1605.06069, 2016.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 429, + 505, + 464 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 468, + 504, + 503 + ], + "lines": [ + { + "bbox": [ + 105, + 466, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 506, + 483 + ], + "score": 1.0, + "content": "Shikhar Sharma, Jing He, Kaheer Suleman, Hannes Schulz, and Philip Bachman. Natural language", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 480, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 115, + 480, + 506, + 493 + ], + "score": 1.0, + "content": "generation in dialogue using lexicalized and delexicalized data. In Proceedings of the 5th Inter-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 491, + 418, + 504 + ], + "spans": [ + { + "bbox": [ + 116, + 491, + 418, + 504 + ], + "score": 1.0, + "content": "national Conference on Learning Representations (ICLR) Workshop, 2017.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 466, + 506, + 504 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 508, + 504, + 531 + ], + "lines": [ + { + "bbox": [ + 106, + 508, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 505, + 521 + ], + "score": 1.0, + "content": "Anthony J Viera, Joanne M Garrett, et al. Understanding interobserver agreement: the kappa statis-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 519, + 293, + 531 + ], + "spans": [ + { + "bbox": [ + 115, + 519, + 293, + 531 + ], + "score": 1.0, + "content": "tic. Family Medicine, 37(5):360–363, 2005.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 106, + 508, + 505, + 531 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 536, + 504, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "Marilyn Walker, Amanda Stent, Franc¸ois Mairesse, and Rashmi Prasad. Individual and domain", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 115, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "adaptation in sentence planning for dialogue. J. Artif. Int. Res., 30(1):413–456, November 2007.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 559, + 189, + 570 + ], + "spans": [ + { + "bbox": [ + 116, + 559, + 189, + 570 + ], + "score": 1.0, + "content": "ISSN 1076-9757.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 106, + 537, + 505, + 570 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 575, + 505, + 631 + ], + "lines": [ + { + "bbox": [ + 106, + 575, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 505, + 588 + ], + "score": 1.0, + "content": "Tsung-Hsien Wen, Milica Gasiˇ c, Dongho Kim, Nikola Mrk ´ siˇ c, Pei-Hao Su, David Vandyke, and ´", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 586, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 116, + 586, + 505, + 600 + ], + "score": 1.0, + "content": "Steve Young. Stochastic Language Generation in Dialogue using Recurrent Neural Networks with", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 597, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 115, + 597, + 505, + 611 + ], + "score": 1.0, + "content": "Convolutional Sentence Reranking. In Proceedings of the 16th Annual Meeting of the Special In-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 114, + 607, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 114, + 607, + 506, + 623 + ], + "score": 1.0, + "content": "terest Group on Discourse and Dialogue (SIGDIAL). Association for Computational Linguistics,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 619, + 191, + 633 + ], + "spans": [ + { + "bbox": [ + 115, + 619, + 191, + 633 + ], + "score": 1.0, + "content": "September 2015a.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 106, + 575, + 506, + 633 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 636, + 505, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 505, + 651 + ], + "score": 1.0, + "content": "Tsung-Hsien Wen, Milica Gasiˇ c, Nikola Mrk ´ siˇ c, Pei-Hao Su, David Vandyke, and Steve Young. ´", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 114, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 114, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "Semantically conditioned LSTM-based natural language generation for spoken dialogue systems.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 114, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 114, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 670, + 410, + 683 + ], + "spans": [ + { + "bbox": [ + 116, + 670, + 410, + 683 + ], + "score": 1.0, + "content": "(EMNLP). Association for Computational Linguistics, September 2015b.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 635, + 505, + 683 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 697, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 115, + 697, + 505, + 712 + ], + "score": 1.0, + "content": "Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 115, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "movies and reading books. In The IEEE International Conference on Computer Vision (ICCV),", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 116, + 721, + 185, + 731 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 185, + 731 + ], + "score": 1.0, + "content": "December 2015.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48.5, + "bbox_fs": [ + 106, + 687, + 505, + 731 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/r17lFgZ0Z/r17lFgZ0Z_model.json b/parse/train/r17lFgZ0Z/r17lFgZ0Z_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5eb5e42df26df978fba503f3d8787306fd46081d --- /dev/null +++ b/parse/train/r17lFgZ0Z/r17lFgZ0Z_model.json @@ -0,0 +1,13157 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 644, + 1302, + 644, + 1302, + 1134, + 398, + 1134 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1560, + 1403, + 1560, + 1403, + 1865, + 298, + 1865 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1268, + 1403, + 1268, + 1403, + 1545, + 298, + 1545 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 299, + 1881, + 1401, + 1881, + 1401, + 2034, + 299, + 2034 + ], + "score": 0.976 + }, + { + "category_id": 0, + "poly": [ + 299, + 218, + 1400, + 218, + 1400, + 379, + 299, + 379 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 313, + 432, + 680, + 432, + 680, + 493, + 313, + 493 + ], + "score": 0.939 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.901 + }, + { + "category_id": 0, + "poly": [ + 302, + 1197, + 572, + 1197, + 572, + 1232, + 302, + 1232 + ], + "score": 0.886 + }, + { + "category_id": 0, + "poly": [ + 773, + 574, + 927, + 574, + 927, + 608, + 773, + 608 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 842, + 2089, + 857, + 2089, + 857, + 2112, + 842, + 2112 + ], + "score": 0.624 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 218.0, + 1402.0, + 218.0, + 1402.0, + 269.0, + 295.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 275.0, + 1404.0, + 275.0, + 1404.0, + 326.0, + 297.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 327.0, + 825.0, + 327.0, + 825.0, + 382.0, + 293.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1194.0, + 579.0, + 1194.0, + 579.0, + 1241.0, + 293.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 573.0, + 932.0, + 573.0, + 932.0, + 611.0, + 770.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 860.0, + 2088.0, + 860.0, + 2118.0, + 841.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 646.0, + 1303.0, + 646.0, + 1303.0, + 678.0, + 396.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 675.0, + 1305.0, + 675.0, + 1305.0, + 713.0, + 393.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 706.0, + 1304.0, + 706.0, + 1304.0, + 744.0, + 393.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 732.0, + 1305.0, + 732.0, + 1305.0, + 775.0, + 392.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 767.0, + 1305.0, + 767.0, + 1305.0, + 802.0, + 393.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 800.0, + 1304.0, + 800.0, + 1304.0, + 832.0, + 394.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 824.0, + 1304.0, + 824.0, + 1304.0, + 864.0, + 393.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 857.0, + 1306.0, + 857.0, + 1306.0, + 895.0, + 392.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 889.0, + 1306.0, + 889.0, + 1306.0, + 923.0, + 394.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 919.0, + 1306.0, + 919.0, + 1306.0, + 955.0, + 394.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 953.0, + 1306.0, + 953.0, + 1306.0, + 984.0, + 394.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 980.0, + 1306.0, + 980.0, + 1306.0, + 1013.0, + 393.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1012.0, + 1305.0, + 1012.0, + 1305.0, + 1044.0, + 393.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1043.0, + 1305.0, + 1043.0, + 1305.0, + 1074.0, + 395.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1072.0, + 1306.0, + 1072.0, + 1306.0, + 1107.0, + 394.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1103.0, + 841.0, + 1103.0, + 841.0, + 1136.0, + 394.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1558.0, + 1405.0, + 1558.0, + 1405.0, + 1595.0, + 294.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1588.0, + 1403.0, + 1588.0, + 1403.0, + 1625.0, + 293.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1618.0, + 1407.0, + 1618.0, + 1407.0, + 1658.0, + 291.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1649.0, + 1407.0, + 1649.0, + 1407.0, + 1686.0, + 293.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1681.0, + 1405.0, + 1681.0, + 1405.0, + 1715.0, + 293.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1710.0, + 1405.0, + 1710.0, + 1405.0, + 1748.0, + 292.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1739.0, + 1407.0, + 1739.0, + 1407.0, + 1780.0, + 291.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 1405.0, + 1773.0, + 1405.0, + 1809.0, + 294.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1804.0, + 1405.0, + 1804.0, + 1405.0, + 1838.0, + 293.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1835.0, + 1095.0, + 1835.0, + 1095.0, + 1869.0, + 292.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1264.0, + 1407.0, + 1264.0, + 1407.0, + 1308.0, + 291.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1300.0, + 1404.0, + 1300.0, + 1404.0, + 1333.0, + 296.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1329.0, + 1407.0, + 1329.0, + 1407.0, + 1367.0, + 291.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1361.0, + 1405.0, + 1361.0, + 1405.0, + 1397.0, + 293.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1389.0, + 1407.0, + 1389.0, + 1407.0, + 1430.0, + 293.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1421.0, + 1405.0, + 1421.0, + 1405.0, + 1458.0, + 293.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1448.0, + 1407.0, + 1448.0, + 1407.0, + 1492.0, + 291.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1478.0, + 1407.0, + 1478.0, + 1407.0, + 1520.0, + 291.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1514.0, + 1079.0, + 1514.0, + 1079.0, + 1550.0, + 294.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1881.0, + 1402.0, + 1881.0, + 1402.0, + 1914.0, + 297.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1911.0, + 1402.0, + 1911.0, + 1402.0, + 1944.0, + 295.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1940.0, + 1403.0, + 1940.0, + 1403.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2009.0, + 293.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2004.0, + 1403.0, + 2004.0, + 1403.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 432.0, + 560.0, + 432.0, + 560.0, + 465.0, + 315.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 462.0, + 682.0, + 462.0, + 682.0, + 495.0, + 311.0, + 495.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1117, + 1403, + 1117, + 1403, + 1362, + 299, + 1362 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1378, + 1403, + 1378, + 1403, + 1562, + 298, + 1562 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1672, + 1403, + 1672, + 1403, + 1827, + 299, + 1827 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 703, + 1402, + 703, + 1402, + 825, + 298, + 825 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 840, + 1403, + 840, + 1403, + 995, + 299, + 995 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 300, + 1010, + 1403, + 1010, + 1403, + 1103, + 300, + 1103 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 302, + 229, + 1403, + 229, + 1403, + 353, + 302, + 353 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 300, + 500, + 1398, + 500, + 1398, + 592, + 300, + 592 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 368, + 374, + 1401, + 374, + 1401, + 478, + 368, + 478 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 298, + 1972, + 1400, + 1972, + 1400, + 2034, + 298, + 2034 + ], + "score": 0.939 + }, + { + "category_id": 0, + "poly": [ + 299, + 1603, + 485, + 1603, + 485, + 1640, + 299, + 1640 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 300, + 1918, + 470, + 1918, + 470, + 1950, + 300, + 1950 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.884 + }, + { + "category_id": 0, + "poly": [ + 301, + 1863, + 763, + 1863, + 763, + 1893, + 301, + 1893 + ], + "score": 0.864 + }, + { + "category_id": 0, + "poly": [ + 300, + 634, + 587, + 634, + 587, + 670, + 300, + 670 + ], + "score": 0.835 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.698 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.202 + }, + { + "category_id": 0, + "poly": [ + 300, + 634, + 587, + 634, + 587, + 670, + 300, + 670 + ], + "score": 0.15 + }, + { + "category_id": 13, + "poly": [ + 887, + 1986, + 905, + 1986, + 905, + 2000, + 887, + 2000 + ], + "score": 0.8, + "latex": "n" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1599.0, + 490.0, + 1599.0, + 490.0, + 1649.0, + 291.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1918.0, + 471.0, + 1918.0, + 471.0, + 1954.0, + 295.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1861.0, + 768.0, + 1861.0, + 768.0, + 1896.0, + 295.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 631.0, + 594.0, + 631.0, + 594.0, + 678.0, + 291.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 631.0, + 594.0, + 631.0, + 594.0, + 678.0, + 291.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1117.0, + 1407.0, + 1117.0, + 1407.0, + 1154.0, + 293.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1147.0, + 1407.0, + 1147.0, + 1407.0, + 1183.0, + 294.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1180.0, + 1404.0, + 1180.0, + 1404.0, + 1213.0, + 294.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1208.0, + 1405.0, + 1208.0, + 1405.0, + 1244.0, + 293.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1237.0, + 1407.0, + 1237.0, + 1407.0, + 1275.0, + 292.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1269.0, + 1407.0, + 1269.0, + 1407.0, + 1307.0, + 293.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1302.0, + 1405.0, + 1302.0, + 1405.0, + 1335.0, + 293.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1330.0, + 569.0, + 1330.0, + 569.0, + 1364.0, + 294.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1375.0, + 1405.0, + 1375.0, + 1405.0, + 1413.0, + 293.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1408.0, + 1404.0, + 1408.0, + 1404.0, + 1440.0, + 296.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1441.0, + 1405.0, + 1441.0, + 1405.0, + 1473.0, + 296.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1471.0, + 1404.0, + 1471.0, + 1404.0, + 1503.0, + 296.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1498.0, + 1405.0, + 1498.0, + 1405.0, + 1535.0, + 292.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1525.0, + 379.0, + 1525.0, + 379.0, + 1567.0, + 292.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1672.0, + 1404.0, + 1672.0, + 1404.0, + 1706.0, + 297.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1703.0, + 1404.0, + 1703.0, + 1404.0, + 1736.0, + 295.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1732.0, + 1404.0, + 1732.0, + 1404.0, + 1768.0, + 295.0, + 1768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1763.0, + 1405.0, + 1763.0, + 1405.0, + 1801.0, + 293.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1793.0, + 571.0, + 1793.0, + 571.0, + 1831.0, + 294.0, + 1831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 704.0, + 1403.0, + 704.0, + 1403.0, + 736.0, + 294.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 735.0, + 1406.0, + 735.0, + 1406.0, + 767.0, + 294.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 764.0, + 1407.0, + 764.0, + 1407.0, + 799.0, + 293.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 793.0, + 1207.0, + 793.0, + 1207.0, + 832.0, + 293.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 842.0, + 1404.0, + 842.0, + 1404.0, + 875.0, + 295.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 870.0, + 1404.0, + 870.0, + 1404.0, + 904.0, + 295.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 900.0, + 1405.0, + 900.0, + 1405.0, + 939.0, + 292.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 931.0, + 1405.0, + 931.0, + 1405.0, + 969.0, + 293.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 961.0, + 389.0, + 961.0, + 389.0, + 1003.0, + 291.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1008.0, + 1404.0, + 1008.0, + 1404.0, + 1047.0, + 293.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1040.0, + 1405.0, + 1040.0, + 1405.0, + 1076.0, + 294.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1070.0, + 1258.0, + 1070.0, + 1258.0, + 1108.0, + 295.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 231.0, + 1407.0, + 231.0, + 1407.0, + 264.0, + 297.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 263.0, + 1405.0, + 263.0, + 1405.0, + 296.0, + 296.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 293.0, + 1404.0, + 293.0, + 1404.0, + 326.0, + 296.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 321.0, + 467.0, + 321.0, + 467.0, + 359.0, + 295.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 498.0, + 1403.0, + 498.0, + 1403.0, + 535.0, + 295.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 533.0, + 1403.0, + 533.0, + 1403.0, + 563.0, + 296.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 559.0, + 1207.0, + 559.0, + 1207.0, + 595.0, + 295.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 374.0, + 1403.0, + 374.0, + 1403.0, + 411.0, + 368.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 406.0, + 973.0, + 406.0, + 973.0, + 443.0, + 391.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 445.0, + 1347.0, + 445.0, + 1347.0, + 481.0, + 366.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1970.0, + 886.0, + 1970.0, + 886.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1970.0, + 1405.0, + 1970.0, + 1405.0, + 2007.0, + 906.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2000.0, + 1404.0, + 2000.0, + 1404.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 797, + 1404, + 797, + 1404, + 982, + 297, + 982 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1508, + 1404, + 1508, + 1404, + 1666, + 298, + 1666 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1176, + 1405, + 1176, + 1405, + 1270, + 298, + 1270 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 591, + 1404, + 591, + 1404, + 713, + 298, + 713 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1678, + 1404, + 1678, + 1404, + 1772, + 298, + 1772 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 300, + 391, + 1398, + 391, + 1398, + 512, + 300, + 512 + ], + "score": 0.97 + }, + { + "category_id": 8, + "poly": [ + 535, + 225, + 1159, + 225, + 1159, + 384, + 535, + 384 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 296, + 1362, + 1403, + 1362, + 1403, + 1426, + 296, + 1426 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 295, + 1030, + 1402, + 1030, + 1402, + 1093, + 295, + 1093 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 729, + 1912, + 966, + 1912, + 966, + 1989, + 729, + 1989 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 637, + 510, + 1062, + 510, + 1062, + 588, + 637, + 588 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 692, + 988, + 1007, + 988, + 1007, + 1025, + 692, + 1025 + ], + "score": 0.938 + }, + { + "category_id": 1, + "poly": [ + 293, + 1856, + 1400, + 1856, + 1400, + 1916, + 293, + 1916 + ], + "score": 0.935 + }, + { + "category_id": 2, + "poly": [ + 335, + 2005, + 771, + 2005, + 771, + 2035, + 335, + 2035 + ], + "score": 0.924 + }, + { + "category_id": 0, + "poly": [ + 299, + 743, + 513, + 743, + 513, + 776, + 299, + 776 + ], + "score": 0.923 + }, + { + "category_id": 0, + "poly": [ + 298, + 1123, + 493, + 1123, + 493, + 1155, + 298, + 1155 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 297, + 74, + 857, + 74, + 857, + 106, + 297, + 106 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 297, + 230, + 516, + 230, + 516, + 262, + 297, + 262 + ], + "score": 0.918 + }, + { + "category_id": 0, + "poly": [ + 299, + 1456, + 578, + 1456, + 578, + 1489, + 299, + 1489 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 299, + 1803, + 659, + 1803, + 659, + 1835, + 299, + 1835 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 301, + 1305, + 715, + 1305, + 715, + 1337, + 301, + 1337 + ], + "score": 0.912 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.688 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.383 + }, + { + "category_id": 14, + "poly": [ + 537, + 224, + 1160, + 224, + 1160, + 386, + 537, + 386 + ], + "score": 0.94, + "latex": "p _ { n } = \\frac { \\displaystyle \\sum _ { C \\in \\{ C a n d i d a t e s \\} } \\sum _ { n - g r a m \\in C } C t _ { c l i p } ( n - g r a m ) } { \\displaystyle \\sum _ { C ^ { \\prime } \\in \\{ C a n d i d a t e s ^ { \\prime } \\} } \\sum _ { n - g r a m ^ { \\prime } \\in C ^ { \\prime } } C t _ { c l i p } ( n - g r a m ^ { \\prime } ) }" + }, + { + "category_id": 14, + "poly": [ + 730, + 1907, + 968, + 1907, + 968, + 1987, + 730, + 1987 + ], + "score": 0.94, + "latex": "\\bar { e } _ { C } = \\frac { \\sum _ { w \\in C } e _ { w } } { \\lvert \\sum _ { w ^ { \\prime } \\in C } e _ { w ^ { \\prime } } \\rvert } ." + }, + { + "category_id": 14, + "poly": [ + 638, + 497, + 1060, + 497, + 1060, + 591, + 638, + 591 + ], + "score": 0.93, + "latex": "\\begin{array} { l } { { \\displaystyle \\forall \\mathrm { ~ s c u r e ~ 1 s ~ u c t 1 1 1 e u ~ a s . } } } \\\\ { { \\displaystyle \\mathsf { B L E U - N } = \\mathsf { B P } \\exp ( \\sum _ { n } ^ { N } \\omega _ { n } \\log ( p _ { n } ) ) } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 373, + 1032, + 450, + 1032, + 450, + 1062, + 373, + 1062 + ], + "score": 0.92, + "latex": "F _ { m e a n }" + }, + { + "category_id": 13, + "poly": [ + 1165, + 393, + 1237, + 393, + 1237, + 426, + 1165, + 426 + ], + "score": 0.92, + "latex": "C t _ { c l i p }" + }, + { + "category_id": 14, + "poly": [ + 690, + 988, + 1007, + 988, + 1007, + 1025, + 690, + 1025 + ], + "score": 0.89, + "latex": "\\mathbf { M E T E O R } = F _ { m e a n } ( 1 - p )" + }, + { + "category_id": 13, + "poly": [ + 372, + 593, + 400, + 593, + 400, + 618, + 372, + 618 + ], + "score": 0.83, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 727, + 1066, + 745, + 1066, + 745, + 1092, + 727, + 1092 + ], + "score": 0.81, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 733, + 597, + 753, + 597, + 753, + 618, + 733, + 618 + ], + "score": 0.79, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 913, + 427, + 933, + 427, + 933, + 449, + 913, + 449 + ], + "score": 0.79, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 1238, + 458, + 1258, + 458, + 1258, + 480, + 1238, + 480 + ], + "score": 0.76, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 448, + 428, + 467, + 428, + 467, + 449, + 448, + 449 + ], + "score": 0.76, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 1379, + 595, + 1402, + 595, + 1402, + 619, + 1379, + 619 + ], + "score": 0.67, + "latex": "\\omega" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 1998.0, + 773.0, + 1998.0, + 773.0, + 2041.0, + 333.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 741.0, + 517.0, + 741.0, + 517.0, + 783.0, + 293.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1123.0, + 495.0, + 1123.0, + 495.0, + 1159.0, + 295.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 109.0, + 297.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1455.0, + 582.0, + 1455.0, + 582.0, + 1491.0, + 295.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1800.0, + 661.0, + 1800.0, + 661.0, + 1839.0, + 295.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1305.0, + 717.0, + 1305.0, + 717.0, + 1341.0, + 294.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 797.0, + 1406.0, + 797.0, + 1406.0, + 834.0, + 294.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 831.0, + 1404.0, + 831.0, + 1404.0, + 863.0, + 295.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 860.0, + 1404.0, + 860.0, + 1404.0, + 896.0, + 294.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 889.0, + 1402.0, + 889.0, + 1402.0, + 924.0, + 295.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 920.0, + 1405.0, + 920.0, + 1405.0, + 956.0, + 294.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 954.0, + 984.0, + 954.0, + 984.0, + 982.0, + 293.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1508.0, + 1404.0, + 1508.0, + 1404.0, + 1545.0, + 296.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1540.0, + 1405.0, + 1540.0, + 1405.0, + 1576.0, + 292.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1571.0, + 1404.0, + 1571.0, + 1404.0, + 1608.0, + 293.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1599.0, + 1405.0, + 1599.0, + 1405.0, + 1638.0, + 293.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1631.0, + 1166.0, + 1631.0, + 1166.0, + 1666.0, + 293.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1174.0, + 1403.0, + 1174.0, + 1403.0, + 1212.0, + 292.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1208.0, + 1403.0, + 1208.0, + 1403.0, + 1242.0, + 295.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1238.0, + 693.0, + 1238.0, + 693.0, + 1273.0, + 295.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 590.0, + 371.0, + 590.0, + 371.0, + 625.0, + 294.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 590.0, + 732.0, + 590.0, + 732.0, + 625.0, + 401.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 590.0, + 1378.0, + 590.0, + 1378.0, + 625.0, + 754.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 621.0, + 1405.0, + 621.0, + 1405.0, + 657.0, + 292.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 653.0, + 1404.0, + 653.0, + 1404.0, + 686.0, + 294.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 681.0, + 563.0, + 681.0, + 563.0, + 716.0, + 293.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1679.0, + 1405.0, + 1679.0, + 1405.0, + 1713.0, + 296.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1709.0, + 1404.0, + 1709.0, + 1404.0, + 1743.0, + 293.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1741.0, + 653.0, + 1741.0, + 653.0, + 1771.0, + 296.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 389.0, + 1164.0, + 389.0, + 1164.0, + 428.0, + 295.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 389.0, + 1407.0, + 389.0, + 1407.0, + 428.0, + 1238.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 423.0, + 447.0, + 423.0, + 447.0, + 455.0, + 296.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 423.0, + 912.0, + 423.0, + 912.0, + 455.0, + 468.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 423.0, + 1404.0, + 423.0, + 1404.0, + 455.0, + 934.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 454.0, + 1237.0, + 454.0, + 1237.0, + 486.0, + 295.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 454.0, + 1403.0, + 454.0, + 1403.0, + 486.0, + 1259.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 481.0, + 878.0, + 481.0, + 878.0, + 516.0, + 295.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1362.0, + 1405.0, + 1362.0, + 1405.0, + 1398.0, + 296.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1391.0, + 1153.0, + 1391.0, + 1153.0, + 1430.0, + 295.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1028.0, + 372.0, + 1028.0, + 372.0, + 1068.0, + 294.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1028.0, + 1405.0, + 1028.0, + 1405.0, + 1068.0, + 451.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1060.0, + 726.0, + 1060.0, + 726.0, + 1096.0, + 295.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 1060.0, + 885.0, + 1060.0, + 885.0, + 1096.0, + 746.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1855.0, + 1403.0, + 1855.0, + 1403.0, + 1894.0, + 293.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1888.0, + 536.0, + 1888.0, + 536.0, + 1924.0, + 295.0, + 1924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 520.0, + 228.0, + 520.0, + 264.0, + 294.0, + 264.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1307, + 1405, + 1307, + 1405, + 1492, + 298, + 1492 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1059, + 1404, + 1059, + 1404, + 1215, + 298, + 1215 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 824, + 1405, + 824, + 1405, + 949, + 298, + 949 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1585, + 1403, + 1585, + 1403, + 1740, + 298, + 1740 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1754, + 1404, + 1754, + 1404, + 1879, + 298, + 1879 + ], + "score": 0.975 + }, + { + "category_id": 8, + "poly": [ + 574, + 681, + 1122, + 681, + 1122, + 815, + 574, + 815 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 296, + 608, + 1399, + 608, + 1399, + 671, + 296, + 671 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 299, + 1972, + 1401, + 1972, + 1401, + 2034, + 299, + 2034 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 298, + 347, + 1400, + 347, + 1400, + 413, + 298, + 413 + ], + "score": 0.943 + }, + { + "category_id": 8, + "poly": [ + 566, + 420, + 1125, + 420, + 1125, + 492, + 566, + 492 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 294, + 505, + 1400, + 505, + 1400, + 540, + 294, + 540 + ], + "score": 0.917 + }, + { + "category_id": 1, + "poly": [ + 288, + 229, + 1328, + 229, + 1328, + 263, + 288, + 263 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 300, + 294, + 615, + 294, + 615, + 326, + 300, + 326 + ], + "score": 0.91 + }, + { + "category_id": 0, + "poly": [ + 300, + 555, + 627, + 555, + 627, + 587, + 300, + 587 + ], + "score": 0.903 + }, + { + "category_id": 0, + "poly": [ + 299, + 1528, + 447, + 1528, + 447, + 1559, + 299, + 1559 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 299, + 1914, + 578, + 1914, + 578, + 1946, + 299, + 1946 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 299, + 1250, + 478, + 1250, + 478, + 1282, + 299, + 1282 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 299, + 990, + 835, + 990, + 835, + 1027, + 299, + 1027 + ], + "score": 0.851 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.767 + }, + { + "category_id": 0, + "poly": [ + 301, + 991, + 834, + 991, + 834, + 1027, + 301, + 1027 + ], + "score": 0.101 + }, + { + "category_id": 14, + "poly": [ + 573, + 679, + 1125, + 679, + 1125, + 818, + 573, + 818 + ], + "score": 0.93, + "latex": "\\begin{array} { c } { \\displaystyle { G ( C , r ) = \\frac { \\sum w \\in C \\operatorname* { m a x } _ { \\hat { w } \\in r } c o s _ { - } s i m ( e _ { w } , w _ { \\hat { w } } ) } { | C | } } } \\\\ { \\displaystyle { G { M ( C , r ) } = \\frac { G ( C , r ) + G ( r , C ) } { 2 } . } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 568, + 416, + 1128, + 416, + 1128, + 493, + 568, + 493 + ], + "score": 0.91, + "latex": "e _ { r d } = \\left\\{ \\begin{array} { l l } { \\operatorname* { m a x } _ { w \\in C } e _ { w d } } & { \\mathrm { i f } e _ { w d } > \\vert \\operatorname* { m i n } _ { w ^ { \\prime } \\in C } e _ { w ^ { \\prime } d } \\vert } \\\\ { \\operatorname* { m i n } _ { w \\in C } e _ { w d } } & { \\mathrm { o t h e r w i s e } \\mathrm { . } } \\end{array} \\right." + }, + { + "category_id": 13, + "poly": [ + 611, + 237, + 644, + 237, + 644, + 261, + 611, + 261 + ], + "score": 0.86, + "latex": "e _ { w }" + }, + { + "category_id": 13, + "poly": [ + 1094, + 508, + 1118, + 508, + 1118, + 534, + 1094, + 534 + ], + "score": 0.85, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 411, + 641, + 435, + 641, + 435, + 667, + 411, + 667 + ], + "score": 0.84, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 485, + 509, + 502, + 509, + 502, + 534, + 485, + 534 + ], + "score": 0.83, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1296, + 232, + 1320, + 232, + 1320, + 258, + 1296, + 258 + ], + "score": 0.79, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 614, + 646, + 631, + 646, + 631, + 667, + 614, + 667 + ], + "score": 0.76, + "latex": "r" + }, + { + "category_id": 13, + "poly": [ + 1377, + 1124, + 1403, + 1124, + 1403, + 1150, + 1377, + 1150 + ], + "score": 0.76, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 456, + 1678, + 521, + 1678, + 521, + 1709, + 456, + 1709 + ], + "score": 0.73, + "latex": "( D A )" + }, + { + "category_id": 13, + "poly": [ + 984, + 237, + 1008, + 237, + 1008, + 258, + 984, + 258 + ], + "score": 0.73, + "latex": "w" + }, + { + "category_id": 13, + "poly": [ + 663, + 1678, + 712, + 1678, + 712, + 1706, + 663, + 1706 + ], + "score": 0.58, + "latex": "D A" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 618.0, + 291.0, + 618.0, + 330.0, + 294.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 552.0, + 629.0, + 552.0, + 629.0, + 590.0, + 295.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1526.0, + 451.0, + 1526.0, + 451.0, + 1563.0, + 295.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1912.0, + 581.0, + 1912.0, + 581.0, + 1952.0, + 293.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1247.0, + 483.0, + 1247.0, + 483.0, + 1287.0, + 294.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 986.0, + 836.0, + 986.0, + 836.0, + 1033.0, + 292.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 987.0, + 837.0, + 987.0, + 837.0, + 1032.0, + 292.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1308.0, + 1405.0, + 1308.0, + 1405.0, + 1344.0, + 295.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1338.0, + 1405.0, + 1338.0, + 1405.0, + 1374.0, + 295.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1368.0, + 1405.0, + 1368.0, + 1405.0, + 1405.0, + 293.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1398.0, + 1406.0, + 1398.0, + 1406.0, + 1436.0, + 292.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1430.0, + 1406.0, + 1430.0, + 1406.0, + 1466.0, + 295.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1461.0, + 1333.0, + 1461.0, + 1333.0, + 1496.0, + 295.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1058.0, + 1403.0, + 1058.0, + 1403.0, + 1094.0, + 296.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1086.0, + 1404.0, + 1086.0, + 1404.0, + 1129.0, + 295.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1121.0, + 1376.0, + 1121.0, + 1376.0, + 1158.0, + 294.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1150.0, + 1406.0, + 1150.0, + 1406.0, + 1188.0, + 294.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1184.0, + 1286.0, + 1184.0, + 1286.0, + 1217.0, + 296.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 822.0, + 1406.0, + 822.0, + 1406.0, + 860.0, + 292.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 856.0, + 1406.0, + 856.0, + 1406.0, + 889.0, + 296.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 886.0, + 1405.0, + 886.0, + 1405.0, + 919.0, + 296.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 918.0, + 1288.0, + 918.0, + 1288.0, + 951.0, + 295.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1587.0, + 1404.0, + 1587.0, + 1404.0, + 1620.0, + 297.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1616.0, + 1405.0, + 1616.0, + 1405.0, + 1651.0, + 293.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1647.0, + 1404.0, + 1647.0, + 1404.0, + 1682.0, + 293.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1676.0, + 455.0, + 1676.0, + 455.0, + 1713.0, + 294.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1676.0, + 662.0, + 1676.0, + 662.0, + 1713.0, + 522.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 1676.0, + 1405.0, + 1676.0, + 1405.0, + 1713.0, + 713.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1706.0, + 1177.0, + 1706.0, + 1177.0, + 1745.0, + 293.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1754.0, + 1405.0, + 1754.0, + 1405.0, + 1791.0, + 293.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1785.0, + 1402.0, + 1785.0, + 1402.0, + 1818.0, + 294.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1818.0, + 1404.0, + 1818.0, + 1404.0, + 1850.0, + 294.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1846.0, + 1063.0, + 1846.0, + 1063.0, + 1884.0, + 293.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 606.0, + 1405.0, + 606.0, + 1405.0, + 645.0, + 295.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 637.0, + 410.0, + 637.0, + 410.0, + 675.0, + 294.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 637.0, + 613.0, + 637.0, + 613.0, + 675.0, + 436.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 637.0, + 1403.0, + 637.0, + 1403.0, + 675.0, + 632.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1970.0, + 1405.0, + 1970.0, + 1405.0, + 2008.0, + 295.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 2003.0, + 1403.0, + 2003.0, + 1403.0, + 2035.0, + 297.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 346.0, + 1405.0, + 346.0, + 1405.0, + 384.0, + 296.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 379.0, + 1404.0, + 379.0, + 1404.0, + 415.0, + 294.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 503.0, + 484.0, + 503.0, + 484.0, + 542.0, + 301.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 503.0, + 1093.0, + 503.0, + 1093.0, + 542.0, + 503.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 503.0, + 1405.0, + 503.0, + 1405.0, + 542.0, + 1119.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 227.0, + 610.0, + 227.0, + 610.0, + 267.0, + 294.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 227.0, + 983.0, + 227.0, + 983.0, + 267.0, + 645.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 227.0, + 1295.0, + 227.0, + 1295.0, + 267.0, + 1009.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1321.0, + 227.0, + 1330.0, + 227.0, + 1330.0, + 267.0, + 1321.0, + 267.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1215, + 852, + 1215, + 852, + 1459, + 298, + 1459 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1656, + 1405, + 1656, + 1405, + 1900, + 297, + 1900 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 957, + 1405, + 957, + 1405, + 1198, + 298, + 1198 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 872, + 1219, + 1403, + 1219, + 1403, + 1456, + 872, + 1456 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 847, + 611, + 1401, + 611, + 1401, + 839, + 847, + 839 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 298, + 621, + 824, + 621, + 824, + 955, + 298, + 955 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 299, + 308, + 1402, + 308, + 1402, + 399, + 299, + 399 + ], + "score": 0.956 + }, + { + "category_id": 2, + "poly": [ + 331, + 1943, + 751, + 1943, + 751, + 2036, + 331, + 2036 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 291, + 229, + 1402, + 229, + 1402, + 293, + 291, + 293 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 298, + 511, + 1402, + 511, + 1402, + 574, + 298, + 574 + ], + "score": 0.94 + }, + { + "category_id": 4, + "poly": [ + 877, + 869, + 1373, + 869, + 1373, + 901, + 877, + 901 + ], + "score": 0.92 + }, + { + "category_id": 4, + "poly": [ + 890, + 1485, + 1391, + 1485, + 1391, + 1518, + 890, + 1518 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 300, + 1517, + 557, + 1517, + 557, + 1552, + 300, + 1552 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 299, + 1593, + 497, + 1593, + 497, + 1625, + 299, + 1625 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 104, + 298, + 104 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.732 + }, + { + "category_id": 0, + "poly": [ + 301, + 450, + 826, + 450, + 826, + 481, + 301, + 481 + ], + "score": 0.56 + }, + { + "category_id": 1, + "poly": [ + 301, + 450, + 826, + 450, + 826, + 481, + 301, + 481 + ], + "score": 0.331 + }, + { + "category_id": 13, + "poly": [ + 1168, + 1048, + 1225, + 1048, + 1225, + 1078, + 1168, + 1078 + ], + "score": 0.9, + "latex": "D A _ { t }" + }, + { + "category_id": 13, + "poly": [ + 526, + 2005, + 635, + 2005, + 635, + 2035, + 526, + 2035 + ], + "score": 0.89, + "latex": "D A _ { t = 1 : N _ { e } }" + }, + { + "category_id": 13, + "poly": [ + 498, + 683, + 534, + 683, + 534, + 712, + 498, + 712 + ], + "score": 0.89, + "latex": "N _ { e }" + }, + { + "category_id": 13, + "poly": [ + 604, + 865, + 661, + 865, + 661, + 895, + 604, + 895 + ], + "score": 0.88, + "latex": "D A _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1217, + 1781, + 1326, + 1781, + 1326, + 1808, + 1217, + 1808 + ], + "score": 0.88, + "latex": "< e o s >" + }, + { + "category_id": 13, + "poly": [ + 366, + 1748, + 473, + 1748, + 473, + 1777, + 366, + 1777 + ], + "score": 0.87, + "latex": "< b o s >" + }, + { + "category_id": 13, + "poly": [ + 532, + 865, + 589, + 865, + 589, + 895, + 532, + 895 + ], + "score": 0.87, + "latex": "D A _ { t }" + }, + { + "category_id": 13, + "poly": [ + 785, + 990, + 826, + 990, + 826, + 1020, + 785, + 1020 + ], + "score": 0.87, + "latex": "v _ { t , i }" + }, + { + "category_id": 13, + "poly": [ + 584, + 895, + 615, + 895, + 615, + 923, + 584, + 923 + ], + "score": 0.85, + "latex": "t ^ { \\mathrm { { t h } } }" + }, + { + "category_id": 13, + "poly": [ + 785, + 652, + 823, + 652, + 823, + 682, + 785, + 682 + ], + "score": 0.82, + "latex": "N _ { e }" + }, + { + "category_id": 13, + "poly": [ + 648, + 1170, + 673, + 1170, + 673, + 1196, + 648, + 1196 + ], + "score": 0.78, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 819, + 1278, + 844, + 1278, + 844, + 1305, + 819, + 1305 + ], + "score": 0.71, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 298, + 1368, + 345, + 1368, + 345, + 1396, + 298, + 1396 + ], + "score": 0.64, + "latex": "D A" + }, + { + "category_id": 13, + "poly": [ + 342, + 1308, + 391, + 1308, + 391, + 1336, + 342, + 1336 + ], + "score": 0.6, + "latex": "D A" + }, + { + "category_id": 13, + "poly": [ + 461, + 713, + 510, + 713, + 510, + 741, + 461, + 741 + ], + "score": 0.59, + "latex": "D A" + }, + { + "category_id": 13, + "poly": [ + 689, + 309, + 738, + 309, + 738, + 337, + 689, + 337 + ], + "score": 0.57, + "latex": "D A" + }, + { + "category_id": 13, + "poly": [ + 1030, + 231, + 1079, + 231, + 1079, + 259, + 1030, + 259 + ], + "score": 0.51, + "latex": "D A" + }, + { + "category_id": 15, + "poly": [ + 1026.0, + 1215.0, + 1077.0, + 1215.0, + 1077.0, + 1244.0, + 1026.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1218.0, + 1332.0, + 1218.0, + 1332.0, + 1244.0, + 1256.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 1243.0, + 1059.0, + 1243.0, + 1059.0, + 1256.0, + 1046.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 1243.0, + 1298.0, + 1243.0, + 1298.0, + 1256.0, + 1287.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1262.0, + 1093.0, + 1262.0, + 1093.0, + 1293.0, + 1011.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 1260.0, + 1334.0, + 1260.0, + 1334.0, + 1294.0, + 1255.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1284.0, + 969.0, + 1284.0, + 969.0, + 1322.0, + 877.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 1290.0, + 1065.0, + 1290.0, + 1065.0, + 1316.0, + 1040.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 1292.0, + 1304.0, + 1292.0, + 1304.0, + 1315.0, + 1284.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 1315.0, + 1096.0, + 1315.0, + 1096.0, + 1344.0, + 1007.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 1315.0, + 1338.0, + 1315.0, + 1338.0, + 1344.0, + 1249.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 1349.0, + 1059.0, + 1349.0, + 1059.0, + 1360.0, + 1046.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 1389.0, + 883.0, + 1389.0, + 883.0, + 1399.0, + 875.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 1384.0, + 915.0, + 1384.0, + 915.0, + 1412.0, + 886.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 1378.0, + 1124.0, + 1378.0, + 1124.0, + 1413.0, + 1017.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1387.0, + 1164.0, + 1387.0, + 1164.0, + 1400.0, + 1152.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 1383.0, + 1194.0, + 1383.0, + 1194.0, + 1412.0, + 1165.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 1379.0, + 1295.0, + 1379.0, + 1295.0, + 1416.0, + 1203.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 1378.0, + 1396.0, + 1378.0, + 1396.0, + 1413.0, + 1296.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 1395.0, + 1124.0, + 1395.0, + 1124.0, + 1407.0, + 1116.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1396.0, + 1397.0, + 1404.0, + 1397.0, + 1404.0, + 1405.0, + 1396.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1405.0, + 883.0, + 1405.0, + 883.0, + 1413.0, + 876.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1408.0, + 1107.0, + 1408.0, + 1107.0, + 1438.0, + 1041.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 1412.0, + 1124.0, + 1412.0, + 1124.0, + 1420.0, + 1116.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1403.0, + 1164.0, + 1403.0, + 1164.0, + 1415.0, + 1153.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 1411.0, + 1382.0, + 1411.0, + 1382.0, + 1431.0, + 1335.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 1430.0, + 1105.0, + 1430.0, + 1105.0, + 1458.0, + 1030.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 1427.0, + 1387.0, + 1427.0, + 1387.0, + 1458.0, + 1305.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1314.0, + 969.0, + 1314.0, + 969.0, + 1345.0, + 877.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1381.5, + 1013.0, + 1381.5, + 1013.0, + 1413.0, + 922.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 609.0, + 1316.0, + 609.0, + 1316.0, + 626.0, + 1177.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 621.0, + 979.0, + 621.0, + 979.0, + 649.0, + 921.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 624.0, + 1275.0, + 624.0, + 1275.0, + 648.0, + 1219.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 652.0, + 1253.0, + 652.0, + 1253.0, + 661.0, + 1243.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 675.0, + 948.0, + 675.0, + 948.0, + 707.0, + 889.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 709.0, + 1109.0, + 709.0, + 1109.0, + 734.0, + 1003.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 710.0, + 1404.0, + 710.0, + 1404.0, + 731.0, + 1303.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 733.0, + 961.0, + 733.0, + 961.0, + 761.0, + 900.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 740.0, + 981.0, + 740.0, + 981.0, + 756.0, + 962.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 735.0, + 1041.0, + 735.0, + 1041.0, + 760.0, + 983.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 733.0, + 1178.0, + 733.0, + 1178.0, + 761.0, + 1117.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 742.0, + 1197.0, + 742.0, + 1197.0, + 754.0, + 1182.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 733.0, + 1259.0, + 733.0, + 1259.0, + 761.0, + 1199.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 743.0, + 1275.0, + 743.0, + 1275.0, + 752.0, + 1265.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 734.0, + 1340.0, + 734.0, + 1340.0, + 762.0, + 1279.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 760.0, + 940.0, + 760.0, + 940.0, + 778.0, + 923.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 759.0, + 1022.0, + 759.0, + 1022.0, + 780.0, + 1001.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 759.0, + 1157.0, + 759.0, + 1157.0, + 778.0, + 1139.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 760.0, + 1237.0, + 760.0, + 1237.0, + 779.0, + 1220.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1300.0, + 758.0, + 1320.0, + 758.0, + 1320.0, + 780.0, + 1300.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 774.0, + 1053.0, + 774.0, + 1053.0, + 805.0, + 891.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 774.0, + 1350.0, + 774.0, + 1350.0, + 805.0, + 1108.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 803.0, + 940.0, + 803.0, + 940.0, + 822.0, + 924.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 802.0, + 1022.0, + 802.0, + 1022.0, + 824.0, + 1002.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 803.0, + 1157.0, + 803.0, + 1157.0, + 823.0, + 1140.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 804.0, + 1237.0, + 804.0, + 1237.0, + 822.0, + 1221.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 804.0, + 1317.0, + 804.0, + 1317.0, + 822.0, + 1302.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 812.0, + 957.0, + 812.0, + 957.0, + 842.0, + 904.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 814.0, + 1259.0, + 814.0, + 1259.0, + 842.0, + 1199.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 814.0, + 1340.0, + 814.0, + 1340.0, + 842.0, + 1279.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 679.0, + 1246.0, + 679.0, + 1246.0, + 703.5, + 1184.0, + 703.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 815.0, + 1035.0, + 815.0, + 1035.0, + 840.0, + 987.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 814.0, + 1176.0, + 814.0, + 1176.0, + 842.0, + 1118.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1937.0, + 743.0, + 1937.0, + 743.0, + 1977.0, + 329.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1967.0, + 755.0, + 1967.0, + 755.0, + 2007.0, + 329.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 2001.0, + 640.0, + 2001.0, + 640.0, + 2036.0, + 329.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 866.0, + 1378.0, + 866.0, + 1378.0, + 905.0, + 871.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 1483.0, + 1392.0, + 1483.0, + 1392.0, + 1519.0, + 883.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1514.0, + 561.0, + 1514.0, + 561.0, + 1558.0, + 292.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1593.0, + 501.0, + 1593.0, + 501.0, + 1629.0, + 295.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 449.0, + 830.0, + 449.0, + 830.0, + 485.0, + 295.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1214.0, + 854.0, + 1214.0, + 854.0, + 1247.0, + 295.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1245.0, + 854.0, + 1245.0, + 854.0, + 1277.0, + 293.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1276.0, + 818.0, + 1276.0, + 818.0, + 1310.0, + 295.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1276.0, + 854.0, + 1276.0, + 854.0, + 1310.0, + 845.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1307.0, + 341.0, + 1307.0, + 341.0, + 1339.0, + 295.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1307.0, + 854.0, + 1307.0, + 854.0, + 1339.0, + 392.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1337.0, + 855.0, + 1337.0, + 855.0, + 1370.0, + 295.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1364.0, + 854.0, + 1364.0, + 854.0, + 1401.0, + 346.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1396.0, + 854.0, + 1396.0, + 854.0, + 1431.0, + 293.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1428.0, + 830.0, + 1428.0, + 830.0, + 1461.0, + 294.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1657.0, + 1405.0, + 1657.0, + 1405.0, + 1691.0, + 296.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1688.0, + 1405.0, + 1688.0, + 1405.0, + 1722.0, + 295.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1718.0, + 1405.0, + 1718.0, + 1405.0, + 1752.0, + 294.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1748.0, + 365.0, + 1748.0, + 365.0, + 1782.0, + 294.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 1748.0, + 1405.0, + 1748.0, + 1405.0, + 1782.0, + 474.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1778.0, + 1216.0, + 1778.0, + 1216.0, + 1813.0, + 292.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1327.0, + 1778.0, + 1406.0, + 1778.0, + 1406.0, + 1813.0, + 1327.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1808.0, + 1405.0, + 1808.0, + 1405.0, + 1841.0, + 294.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1838.0, + 1407.0, + 1838.0, + 1407.0, + 1874.0, + 294.0, + 1874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1866.0, + 981.0, + 1866.0, + 981.0, + 1903.0, + 292.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 952.0, + 1405.0, + 952.0, + 1405.0, + 992.0, + 293.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 984.0, + 784.0, + 984.0, + 784.0, + 1020.0, + 295.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 984.0, + 1403.0, + 984.0, + 1403.0, + 1020.0, + 827.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1014.0, + 1405.0, + 1014.0, + 1405.0, + 1052.0, + 293.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1043.0, + 1167.0, + 1043.0, + 1167.0, + 1083.0, + 293.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 1043.0, + 1407.0, + 1043.0, + 1407.0, + 1083.0, + 1226.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1078.0, + 1406.0, + 1078.0, + 1406.0, + 1111.0, + 295.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1105.0, + 1407.0, + 1105.0, + 1407.0, + 1143.0, + 293.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1139.0, + 1403.0, + 1139.0, + 1403.0, + 1172.0, + 296.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1168.0, + 647.0, + 1168.0, + 647.0, + 1201.0, + 294.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 1168.0, + 683.0, + 1168.0, + 683.0, + 1201.0, + 674.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 619.0, + 827.0, + 619.0, + 827.0, + 653.0, + 297.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 649.0, + 784.0, + 649.0, + 784.0, + 684.0, + 295.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 681.0, + 497.0, + 681.0, + 497.0, + 713.0, + 295.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 681.0, + 826.0, + 681.0, + 826.0, + 713.0, + 535.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 712.0, + 460.0, + 712.0, + 460.0, + 745.0, + 295.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 712.0, + 828.0, + 712.0, + 828.0, + 745.0, + 511.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 741.0, + 826.0, + 741.0, + 826.0, + 773.0, + 295.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 767.0, + 827.0, + 767.0, + 827.0, + 809.0, + 293.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 803.0, + 827.0, + 803.0, + 827.0, + 835.0, + 295.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 833.0, + 828.0, + 833.0, + 828.0, + 865.0, + 295.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 862.0, + 531.0, + 862.0, + 531.0, + 897.0, + 294.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 862.0, + 603.0, + 862.0, + 603.0, + 897.0, + 590.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 862.0, + 827.0, + 862.0, + 827.0, + 897.0, + 662.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 892.0, + 583.0, + 892.0, + 583.0, + 928.0, + 293.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 892.0, + 827.0, + 892.0, + 827.0, + 928.0, + 616.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 923.0, + 827.0, + 923.0, + 827.0, + 959.0, + 295.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 307.0, + 688.0, + 307.0, + 688.0, + 344.0, + 295.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 307.0, + 1406.0, + 307.0, + 1406.0, + 344.0, + 739.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 338.0, + 1406.0, + 338.0, + 1406.0, + 375.0, + 294.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 369.0, + 1361.0, + 369.0, + 1361.0, + 402.0, + 297.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 226.0, + 1029.0, + 226.0, + 1029.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 226.0, + 1404.0, + 226.0, + 1404.0, + 265.0, + 1080.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 258.0, + 1143.0, + 258.0, + 1143.0, + 296.0, + 294.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 509.0, + 1404.0, + 509.0, + 1404.0, + 545.0, + 296.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 540.0, + 1402.0, + 540.0, + 1402.0, + 576.0, + 293.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 449.0, + 830.0, + 449.0, + 830.0, + 485.0, + 295.0, + 485.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1121, + 1404, + 1121, + 1404, + 1366, + 297, + 1366 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1380, + 1403, + 1380, + 1403, + 1595, + 297, + 1595 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 923, + 1403, + 923, + 1403, + 1106, + 298, + 1106 + ], + "score": 0.979 + }, + { + "category_id": 5, + "poly": [ + 346, + 486, + 1349, + 486, + 1349, + 680, + 346, + 680 + ], + "score": 0.979, + "html": "
DSTC2Restaurants
SkipThoughtEmbeddingAverageVectorExtremaGreedyMatchingSkipThoughtEmbeddingAverageVectorExtremaGreedyMatching
Gold1.001.001.001.001.001.001.001.00
Random0.9060.9810.9100.9470.8430.9570.9050.930
LSTM0.9460.9850.9350.9620.9450.9970.9860.991
d-scLSTM0.9250.9840.9260.9570.9480.9970.9860.991
hld-scLSTM0.9320.9870.9420.9640.9680.9970.9890.993
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1611, + 1402, + 1611, + 1402, + 1796, + 299, + 1796 + ], + "score": 0.978 + }, + { + "category_id": 5, + "poly": [ + 318, + 264, + 1381, + 264, + 1381, + 436, + 318, + 436 + ], + "score": 0.977, + "html": "
DSTC2Restaurants
B-1B-2B-3B-4MRLB-1B-2B-3B-4MRL
Gold1.001.001.001.001.001.001.001.001.001.001.001.00
Random0.8750.8430.8220.8070.5640.8520.8720.8130.7650.7210.5040.796
LSTM0.9000.8790.8630.8510.6100.8880.9820.9660.9490.9320.6520.944
d-scLSTM0.8800.8500.8280.8120.5780.8740.9800.9640.9480.9310.6540.945
hld-scLSTM0.9090.8900.8780.8700.6240.8990.9850.9780.9700.9620.7040.965
" + }, + { + "category_id": 1, + "poly": [ + 301, + 1942, + 1401, + 1942, + 1401, + 2033, + 301, + 2033 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 699, + 1405, + 699, + 1405, + 822, + 298, + 822 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 298, + 1810, + 1188, + 1810, + 1188, + 1841, + 298, + 1841 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 299, + 863, + 521, + 863, + 521, + 895, + 299, + 895 + ], + "score": 0.895 + }, + { + "category_id": 6, + "poly": [ + 335, + 234, + 1357, + 234, + 1357, + 263, + 335, + 263 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 301, + 1883, + 716, + 1883, + 716, + 1913, + 301, + 1913 + ], + "score": 0.88 + }, + { + "category_id": 6, + "poly": [ + 306, + 456, + 1393, + 456, + 1393, + 483, + 306, + 483 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.765 + }, + { + "category_id": 6, + "poly": [ + 329, + 235, + 1354, + 235, + 1354, + 263, + 329, + 263 + ], + "score": 0.116 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 859.0, + 527.0, + 859.0, + 527.0, + 901.0, + 293.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 231.0, + 1359.0, + 231.0, + 1359.0, + 269.0, + 341.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 722.0, + 1882.0, + 722.0, + 1917.0, + 294.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 452.0, + 1398.0, + 452.0, + 1398.0, + 489.0, + 302.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 229.0, + 1358.0, + 229.0, + 1358.0, + 269.0, + 340.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1121.0, + 1405.0, + 1121.0, + 1405.0, + 1155.0, + 296.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1151.0, + 1406.0, + 1151.0, + 1406.0, + 1187.0, + 294.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1182.0, + 1406.0, + 1182.0, + 1406.0, + 1216.0, + 295.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1212.0, + 1406.0, + 1212.0, + 1406.0, + 1249.0, + 294.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1241.0, + 1404.0, + 1241.0, + 1404.0, + 1279.0, + 292.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1273.0, + 1405.0, + 1273.0, + 1405.0, + 1308.0, + 291.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1301.0, + 1406.0, + 1301.0, + 1406.0, + 1342.0, + 291.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1332.0, + 433.0, + 1332.0, + 433.0, + 1371.0, + 293.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1380.0, + 1405.0, + 1380.0, + 1405.0, + 1415.0, + 296.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1409.0, + 1407.0, + 1409.0, + 1407.0, + 1448.0, + 292.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1442.0, + 1404.0, + 1442.0, + 1404.0, + 1475.0, + 293.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1473.0, + 1405.0, + 1473.0, + 1405.0, + 1507.0, + 292.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1501.0, + 1404.0, + 1501.0, + 1404.0, + 1540.0, + 291.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1533.0, + 1405.0, + 1533.0, + 1405.0, + 1570.0, + 292.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1566.0, + 636.0, + 1566.0, + 636.0, + 1597.0, + 296.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 924.0, + 1404.0, + 924.0, + 1404.0, + 955.0, + 294.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 951.0, + 1405.0, + 951.0, + 1405.0, + 988.0, + 293.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 984.0, + 1404.0, + 984.0, + 1404.0, + 1016.0, + 294.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1015.0, + 1405.0, + 1015.0, + 1405.0, + 1046.0, + 294.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1046.0, + 1404.0, + 1046.0, + 1404.0, + 1078.0, + 296.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1075.0, + 1196.0, + 1075.0, + 1196.0, + 1111.0, + 294.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1611.0, + 1402.0, + 1611.0, + 1402.0, + 1642.0, + 297.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1641.0, + 1406.0, + 1641.0, + 1406.0, + 1674.0, + 293.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1674.0, + 1403.0, + 1674.0, + 1403.0, + 1706.0, + 295.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1702.0, + 1404.0, + 1702.0, + 1404.0, + 1737.0, + 294.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1731.0, + 1406.0, + 1731.0, + 1406.0, + 1768.0, + 293.0, + 1768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1766.0, + 1198.0, + 1766.0, + 1198.0, + 1798.0, + 295.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1976.0, + 297.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2005.0, + 295.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 1405.0, + 2003.0, + 1405.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 695.0, + 1406.0, + 695.0, + 1406.0, + 735.0, + 293.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 731.0, + 1405.0, + 731.0, + 1405.0, + 762.0, + 293.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 759.0, + 1405.0, + 759.0, + 1405.0, + 795.0, + 291.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 790.0, + 413.0, + 790.0, + 413.0, + 825.0, + 293.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1808.0, + 1194.0, + 1808.0, + 1194.0, + 1845.0, + 293.0, + 1845.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1573, + 1404, + 1573, + 1404, + 1787, + 298, + 1787 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1271, + 1045, + 1271, + 1045, + 1454, + 298, + 1454 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 313, + 234, + 1385, + 234, + 1385, + 824, + 313, + 824 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1803, + 1403, + 1803, + 1403, + 1926, + 299, + 1926 + ], + "score": 0.972 + }, + { + "category_id": 5, + "poly": [ + 1077, + 1272, + 1385, + 1272, + 1385, + 1471, + 1077, + 1471 + ], + "score": 0.972, + "html": "
K# pairs% pairs
>0.155/55100.0 %
>0.240/5572.7 %
>0.328/5550.9 %
>0.419/5534.5 %
>0.58/5514.5 %
>0.60/550.0 %
" + }, + { + "category_id": 1, + "poly": [ + 300, + 1942, + 1402, + 1942, + 1402, + 2034, + 300, + 2034 + ], + "score": 0.969 + }, + { + "category_id": 4, + "poly": [ + 297, + 850, + 1405, + 850, + 1405, + 1002, + 297, + 1002 + ], + "score": 0.964 + }, + { + "category_id": 6, + "poly": [ + 1065, + 1187, + 1401, + 1187, + 1401, + 1272, + 1065, + 1272 + ], + "score": 0.916 + }, + { + "category_id": 1, + "poly": [ + 298, + 1165, + 1042, + 1165, + 1042, + 1255, + 298, + 1255 + ], + "score": 0.864 + }, + { + "category_id": 0, + "poly": [ + 300, + 1511, + 1184, + 1511, + 1184, + 1540, + 300, + 1540 + ], + "score": 0.849 + }, + { + "category_id": 1, + "poly": [ + 302, + 1071, + 1402, + 1071, + 1402, + 1163, + 302, + 1163 + ], + "score": 0.802 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.793 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2111, + 842, + 2111 + ], + "score": 0.635 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 855, + 76, + 855, + 104, + 299, + 104 + ], + "score": 0.312 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2111, + 842, + 2111 + ], + "score": 0.263 + }, + { + "category_id": 13, + "poly": [ + 820, + 911, + 946, + 911, + 946, + 944, + 820, + 944 + ], + "score": 0.93, + "latex": "\\mathcal { N } ( 0 , 0 . 0 2 )" + }, + { + "category_id": 13, + "poly": [ + 1110, + 880, + 1222, + 880, + 1222, + 914, + 1110, + 914 + ], + "score": 0.92, + "latex": "\\mathcal { N } ( 0 , 0 . 1 )" + }, + { + "category_id": 13, + "poly": [ + 593, + 1394, + 691, + 1394, + 691, + 1422, + 593, + 1422 + ], + "score": 0.89, + "latex": "\\kappa > 0 . 3" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1330, + 1151, + 1330, + 1151, + 1355, + 1095, + 1355 + ], + "score": 0.73, + "latex": "{ > } 0 . 2" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1358, + 1150, + 1358, + 1150, + 1382, + 1095, + 1382 + ], + "score": 0.72, + "latex": "{ > } 0 . 3" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1441, + 1151, + 1441, + 1151, + 1466, + 1095, + 1466 + ], + "score": 0.71, + "latex": "{ > } 0 . 6" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1386, + 1151, + 1386, + 1151, + 1410, + 1095, + 1410 + ], + "score": 0.7, + "latex": "{ > } 0 . 4" + }, + { + "category_id": 13, + "poly": [ + 1281, + 1301, + 1366, + 1301, + 1366, + 1327, + 1281, + 1327 + ], + "score": 0.68, + "latex": "1 0 0 . 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 1096, + 1303, + 1149, + 1303, + 1149, + 1327, + 1096, + 1327 + ], + "score": 0.67, + "latex": "\\overline { { > 0 . 1 } }" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1414, + 1150, + 1414, + 1150, + 1438, + 1095, + 1438 + ], + "score": 0.67, + "latex": "{ > } 0 . 5" + }, + { + "category_id": 13, + "poly": [ + 1285, + 1274, + 1307, + 1274, + 1307, + 1297, + 1285, + 1297 + ], + "score": 0.6, + "latex": "\\overline { { \\% } }" + }, + { + "category_id": 13, + "poly": [ + 1113, + 1278, + 1132, + 1278, + 1132, + 1297, + 1113, + 1297 + ], + "score": 0.52, + "latex": "\\kappa" + }, + { + "category_id": 13, + "poly": [ + 1292, + 1440, + 1354, + 1440, + 1354, + 1466, + 1292, + 1466 + ], + "score": 0.43, + "latex": "0 . 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 1285, + 1329, + 1361, + 1329, + 1361, + 1354, + 1285, + 1354 + ], + "score": 0.4, + "latex": "7 2 . 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 1286, + 1384, + 1360, + 1384, + 1360, + 1410, + 1286, + 1410 + ], + "score": 0.38, + "latex": "3 4 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1285, + 1357, + 1360, + 1357, + 1360, + 1382, + 1285, + 1382 + ], + "score": 0.33, + "latex": "5 0 . 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 1287, + 1412, + 1360, + 1412, + 1360, + 1437, + 1287, + 1437 + ], + "score": 0.31, + "latex": "1 4 . 5 \\%" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 256.0, + 384.0, + 256.0, + 384.0, + 287.0, + 342.0, + 287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 234.0, + 882.0, + 234.0, + 882.0, + 443.0, + 852.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 249.0, + 920.0, + 249.0, + 920.0, + 275.0, + 883.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 234.0, + 1185.0, + 234.0, + 1185.0, + 444.0, + 1136.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 265.0, + 611.0, + 265.0, + 611.0, + 413.0, + 583.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 265.0, + 650.0, + 265.0, + 650.0, + 288.0, + 615.0, + 288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 265.0, + 815.0, + 265.0, + 815.0, + 286.0, + 778.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 268.0, + 1009.0, + 268.0, + 1009.0, + 277.0, + 993.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 277.0, + 383.0, + 277.0, + 383.0, + 401.0, + 316.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 294.0, + 651.0, + 294.0, + 651.0, + 321.0, + 612.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 280.0, + 921.0, + 280.0, + 921.0, + 311.0, + 881.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 300.0, + 1185.0, + 300.0, + 1185.0, + 316.0, + 1170.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 310.0, + 383.0, + 310.0, + 383.0, + 368.0, + 341.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 327.0, + 651.0, + 327.0, + 651.0, + 353.0, + 612.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 331.0, + 735.0, + 331.0, + 735.0, + 347.0, + 720.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 315.0, + 921.0, + 315.0, + 921.0, + 346.0, + 880.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 341.0, + 1185.0, + 341.0, + 1185.0, + 358.0, + 1171.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 364.0, + 384.0, + 364.0, + 384.0, + 420.0, + 342.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 356.0, + 652.0, + 356.0, + 652.0, + 416.0, + 612.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 353.0, + 815.0, + 353.0, + 815.0, + 377.0, + 788.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 351.0, + 920.0, + 351.0, + 920.0, + 413.0, + 881.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 384.0, + 1185.0, + 384.0, + 1185.0, + 402.0, + 1171.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 420.0, + 651.0, + 420.0, + 651.0, + 448.0, + 612.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 422.0, + 919.0, + 422.0, + 919.0, + 445.0, + 883.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 438.0, + 523.0, + 438.0, + 523.0, + 466.0, + 405.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 440.0, + 549.0, + 440.0, + 549.0, + 462.0, + 528.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 438.0, + 822.0, + 438.0, + 822.0, + 466.0, + 673.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 438.0, + 1059.0, + 438.0, + 1059.0, + 466.0, + 943.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 442.0, + 1086.0, + 442.0, + 1086.0, + 462.0, + 1068.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 444.0, + 1234.0, + 444.0, + 1234.0, + 460.0, + 1216.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 438.0, + 1356.0, + 438.0, + 1356.0, + 466.0, + 1237.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 464.0, + 542.0, + 464.0, + 542.0, + 494.0, + 411.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 464.0, + 811.0, + 464.0, + 811.0, + 496.0, + 680.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 464.0, + 1080.0, + 464.0, + 1080.0, + 496.0, + 950.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 463.0, + 1386.0, + 463.0, + 1386.0, + 496.0, + 1180.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 493.0, + 910.0, + 493.0, + 910.0, + 525.0, + 786.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 554.0, + 384.0, + 554.0, + 384.0, + 699.0, + 313.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 590.0, + 424.0, + 590.0, + 424.0, + 607.0, + 388.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 565.0, + 611.0, + 565.0, + 611.0, + 711.0, + 583.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 557.0, + 651.0, + 557.0, + 651.0, + 610.0, + 612.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 532.0, + 920.0, + 532.0, + 920.0, + 742.0, + 852.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 555.0, + 954.0, + 555.0, + 954.0, + 570.0, + 944.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 532.0, + 1185.0, + 532.0, + 1185.0, + 742.0, + 1136.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 597.0, + 1185.0, + 597.0, + 1185.0, + 614.0, + 1169.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 609.0, + 384.0, + 609.0, + 384.0, + 694.0, + 341.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 610.0, + 651.0, + 610.0, + 651.0, + 692.0, + 612.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 685.0, + 787.0, + 685.0, + 787.0, + 698.0, + 767.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 630.0, + 817.0, + 630.0, + 817.0, + 673.0, + 797.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 615.0, + 921.0, + 615.0, + 921.0, + 713.0, + 881.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 656.0, + 1049.0, + 656.0, + 1049.0, + 670.0, + 1039.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 639.0, + 1185.0, + 639.0, + 1185.0, + 699.0, + 1171.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 691.0, + 383.0, + 691.0, + 383.0, + 718.0, + 342.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 702.0, + 516.0, + 702.0, + 516.0, + 717.0, + 500.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 696.0, + 561.0, + 696.0, + 561.0, + 719.0, + 530.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 693.0, + 651.0, + 693.0, + 651.0, + 718.0, + 612.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 706.0, + 1252.0, + 706.0, + 1252.0, + 721.0, + 1220.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 718.0, + 920.0, + 718.0, + 920.0, + 745.0, + 882.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 740.0, + 395.0, + 740.0, + 395.0, + 760.0, + 375.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 739.0, + 435.0, + 739.0, + 435.0, + 761.0, + 414.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 737.0, + 479.0, + 737.0, + 479.0, + 763.0, + 453.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 740.0, + 517.0, + 740.0, + 517.0, + 761.0, + 497.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 739.0, + 560.0, + 739.0, + 560.0, + 761.0, + 540.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 740.0, + 698.0, + 740.0, + 698.0, + 759.0, + 677.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 734.0, + 822.0, + 734.0, + 822.0, + 766.0, + 703.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 741.0, + 930.0, + 741.0, + 930.0, + 758.0, + 915.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 737.0, + 976.0, + 737.0, + 976.0, + 763.0, + 952.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 739.0, + 1015.0, + 739.0, + 1015.0, + 761.0, + 994.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 740.0, + 1057.0, + 740.0, + 1057.0, + 760.0, + 1035.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 739.0, + 1098.0, + 739.0, + 1098.0, + 761.0, + 1079.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 743.0, + 1227.0, + 743.0, + 1227.0, + 755.0, + 1218.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 740.0, + 1264.0, + 740.0, + 1264.0, + 760.0, + 1242.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 736.0, + 1356.0, + 736.0, + 1356.0, + 764.0, + 1265.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 764.0, + 540.0, + 764.0, + 540.0, + 793.0, + 410.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 765.0, + 811.0, + 765.0, + 811.0, + 791.0, + 681.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 764.0, + 1079.0, + 764.0, + 1079.0, + 793.0, + 949.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 765.0, + 1385.0, + 765.0, + 1385.0, + 791.0, + 1182.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 794.0, + 932.0, + 794.0, + 932.0, + 825.0, + 770.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 587.0, + 938.0, + 587.0, + 938.0, + 596.5, + 917.0, + 596.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.25, + 618.5, + 812.25, + 618.5, + 812.25, + 636.0, + 772.25, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 849.0, + 1406.0, + 849.0, + 1406.0, + 884.0, + 294.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 877.0, + 1109.0, + 877.0, + 1109.0, + 914.0, + 295.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1223.0, + 877.0, + 1405.0, + 877.0, + 1405.0, + 914.0, + 1223.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 910.0, + 819.0, + 910.0, + 819.0, + 947.0, + 295.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 910.0, + 1405.0, + 910.0, + 1405.0, + 947.0, + 947.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 940.0, + 1405.0, + 940.0, + 1405.0, + 977.0, + 294.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 973.0, + 436.0, + 973.0, + 436.0, + 1003.0, + 295.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 1183.0, + 1404.0, + 1183.0, + 1404.0, + 1217.0, + 1065.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 1215.0, + 1404.0, + 1215.0, + 1404.0, + 1247.0, + 1064.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 1250.0, + 1187.0, + 1250.0, + 1187.0, + 1278.0, + 1065.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1507.0, + 1189.0, + 1507.0, + 1189.0, + 1544.0, + 294.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1573.0, + 1404.0, + 1573.0, + 1404.0, + 1608.0, + 296.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1603.0, + 1406.0, + 1603.0, + 1406.0, + 1640.0, + 293.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1633.0, + 1407.0, + 1633.0, + 1407.0, + 1670.0, + 292.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1665.0, + 1406.0, + 1665.0, + 1406.0, + 1698.0, + 292.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1696.0, + 1405.0, + 1696.0, + 1405.0, + 1728.0, + 292.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1727.0, + 1405.0, + 1727.0, + 1405.0, + 1761.0, + 294.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1759.0, + 615.0, + 1759.0, + 615.0, + 1789.0, + 294.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1270.0, + 1045.0, + 1270.0, + 1045.0, + 1303.0, + 296.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1300.0, + 1047.0, + 1300.0, + 1047.0, + 1334.0, + 295.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1332.0, + 1045.0, + 1332.0, + 1045.0, + 1362.0, + 295.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1361.0, + 1046.0, + 1361.0, + 1046.0, + 1394.0, + 293.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1390.0, + 592.0, + 1390.0, + 592.0, + 1428.0, + 292.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 692.0, + 1390.0, + 1048.0, + 1390.0, + 1048.0, + 1428.0, + 692.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1423.0, + 683.0, + 1423.0, + 683.0, + 1455.0, + 296.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1803.0, + 1403.0, + 1803.0, + 1403.0, + 1835.0, + 293.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1835.0, + 1404.0, + 1835.0, + 1404.0, + 1868.0, + 295.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1866.0, + 1405.0, + 1866.0, + 1405.0, + 1898.0, + 293.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1892.0, + 1397.0, + 1892.0, + 1397.0, + 1931.0, + 293.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2004.0, + 1404.0, + 2004.0, + 1404.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1159.0, + 1047.0, + 1159.0, + 1047.0, + 1197.0, + 294.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1190.0, + 1048.0, + 1190.0, + 1048.0, + 1227.0, + 293.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1224.0, + 564.0, + 1224.0, + 564.0, + 1258.0, + 296.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1071.0, + 1406.0, + 1071.0, + 1406.0, + 1105.0, + 297.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1101.0, + 1404.0, + 1101.0, + 1404.0, + 1134.0, + 294.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1130.0, + 1403.0, + 1130.0, + 1403.0, + 1169.0, + 295.0, + 1169.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 353, + 261, + 1343, + 261, + 1343, + 561, + 353, + 561 + ], + "score": 0.985, + "html": "
MetricDSTC2Restaurants
Spearmanp-valuePearsonp-valueSpearmanp-valuePearsonp-value
Bleu1Bleu 2-0.317<0.0050.583<0.0050.0690.4940.2770.005
-0.318<0.0050.526<0.0050.0910.3660.1660.099
Bleu 3-0.318<0.0050.500<0.0050.1090.2800.2230.026
Bleu 4METEORROUGELSkip Thought-0.318<0.0050.461<0.0050.1050.2960.2550.010
0.295<0.0050.582<0.0050.353<0.0050.489<0.005
0.294<0.0050.448<0.0050.346<0.0050.382<0.005
0.528<0.0050.0860.3970.284<0.0050.364<0.005
Embedding Average0.295<0.0050.485<0.0050.423<0.0050.2600.009
Vector ExtremaGreedy Matching0.299<0.0050.624<0.0050.446<0.0050.287<0.005
0.295<0.0050.572<0.0050.446<0.0050.325<0.005
Human0.810<0.0050.984<0.0050.653<0.0050.857<0.005
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1191, + 1404, + 1191, + 1404, + 1558, + 298, + 1558 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 662, + 1403, + 662, + 1403, + 939, + 298, + 939 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 300, + 954, + 1403, + 954, + 1403, + 1076, + 300, + 1076 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1712, + 1403, + 1712, + 1403, + 1865, + 299, + 1865 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 300, + 1573, + 1404, + 1573, + 1404, + 1696, + 300, + 1696 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1881, + 1402, + 1881, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 301, + 586, + 1399, + 586, + 1399, + 648, + 301, + 648 + ], + "score": 0.951 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 299, + 1122, + 530, + 1122, + 530, + 1158, + 299, + 1158 + ], + "score": 0.904 + }, + { + "category_id": 6, + "poly": [ + 453, + 235, + 1249, + 235, + 1249, + 261, + 453, + 261 + ], + "score": 0.834 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.791 + }, + { + "category_id": 13, + "poly": [ + 1260, + 506, + 1323, + 506, + 1323, + 527, + 1260, + 527 + ], + "score": 0.63, + "latex": "< 0 . 0 0 5" + }, + { + "category_id": 13, + "poly": [ + 1260, + 483, + 1324, + 483, + 1324, + 504, + 1260, + 504 + ], + "score": 0.58, + "latex": "< 0 . 0 0 5" + }, + { + "category_id": 13, + "poly": [ + 1072, + 506, + 1135, + 506, + 1135, + 527, + 1072, + 527 + ], + "score": 0.57, + "latex": "< 0 . 0 0 5" + }, + { + "category_id": 13, + "poly": [ + 1072, + 535, + 1135, + 535, + 1135, + 555, + 1072, + 555 + ], + "score": 0.49, + "latex": "- 0 . 0 0 5" + }, + { + "category_id": 13, + "poly": [ + 1072, + 461, + 1135, + 461, + 1135, + 482, + 1072, + 482 + ], + "score": 0.45, + "latex": "< 0 . 0 0 5" + }, + { + "category_id": 13, + "poly": [ + 1072, + 484, + 1135, + 484, + 1135, + 503, + 1072, + 503 + ], + "score": 0.37, + "latex": "< 0 . 0 0 5" + }, + { + "category_id": 13, + "poly": [ + 1072, + 395, + 1135, + 395, + 1135, + 416, + 1072, + 416 + ], + "score": 0.35, + "latex": "< 0 . 0 0 5" + }, + { + "category_id": 13, + "poly": [ + 1260, + 394, + 1324, + 394, + 1324, + 416, + 1260, + 416 + ], + "score": 0.35, + "latex": "< 0 . 0 0 5" + }, + { + "category_id": 13, + "poly": [ + 1260, + 439, + 1324, + 439, + 1324, + 460, + 1260, + 460 + ], + "score": 0.35, + "latex": "< 0 . 0 0 5" + }, + { + "category_id": 13, + "poly": [ + 1072, + 417, + 1135, + 417, + 1135, + 437, + 1072, + 437 + ], + "score": 0.3, + "latex": "< 0 . 0 0 5" + }, + { + "category_id": 13, + "poly": [ + 1071, + 439, + 1135, + 439, + 1135, + 459, + 1071, + 459 + ], + "score": 0.29, + "latex": "< 0 . 0 0 5" + }, + { + "category_id": 13, + "poly": [ + 861, + 506, + 924, + 506, + 924, + 527, + 861, + 527 + ], + "score": 0.26, + "latex": "< 0 . 0 0 5" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1118.0, + 535.0, + 1118.0, + 535.0, + 1166.0, + 292.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 230.0, + 1256.0, + 230.0, + 1256.0, + 269.0, + 446.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1191.0, + 1405.0, + 1191.0, + 1405.0, + 1229.0, + 294.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1225.0, + 1404.0, + 1225.0, + 1404.0, + 1255.0, + 294.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1253.0, + 1404.0, + 1253.0, + 1404.0, + 1287.0, + 294.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1286.0, + 1402.0, + 1286.0, + 1402.0, + 1316.0, + 296.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1312.0, + 1405.0, + 1312.0, + 1405.0, + 1350.0, + 293.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1344.0, + 1406.0, + 1344.0, + 1406.0, + 1382.0, + 293.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1376.0, + 1405.0, + 1376.0, + 1405.0, + 1410.0, + 294.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1402.0, + 1405.0, + 1402.0, + 1405.0, + 1443.0, + 292.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1438.0, + 1405.0, + 1438.0, + 1405.0, + 1468.0, + 296.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1467.0, + 1406.0, + 1467.0, + 1406.0, + 1501.0, + 294.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1496.0, + 1405.0, + 1496.0, + 1405.0, + 1530.0, + 293.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1529.0, + 1152.0, + 1529.0, + 1152.0, + 1559.0, + 297.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 664.0, + 1404.0, + 664.0, + 1404.0, + 697.0, + 297.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 696.0, + 1404.0, + 696.0, + 1404.0, + 729.0, + 294.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 725.0, + 1404.0, + 725.0, + 1404.0, + 759.0, + 294.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 756.0, + 1404.0, + 756.0, + 1404.0, + 789.0, + 296.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 785.0, + 1407.0, + 785.0, + 1407.0, + 821.0, + 293.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 816.0, + 1405.0, + 816.0, + 1405.0, + 853.0, + 294.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 846.0, + 1405.0, + 846.0, + 1405.0, + 882.0, + 293.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 875.0, + 1405.0, + 875.0, + 1405.0, + 913.0, + 293.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 909.0, + 1305.0, + 909.0, + 1305.0, + 942.0, + 296.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 953.0, + 1403.0, + 953.0, + 1403.0, + 988.0, + 293.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 985.0, + 1405.0, + 985.0, + 1405.0, + 1017.0, + 295.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1015.0, + 1405.0, + 1015.0, + 1405.0, + 1051.0, + 293.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1046.0, + 533.0, + 1046.0, + 533.0, + 1078.0, + 295.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1711.0, + 1407.0, + 1711.0, + 1407.0, + 1746.0, + 295.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1741.0, + 1405.0, + 1741.0, + 1405.0, + 1777.0, + 294.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1771.0, + 1407.0, + 1771.0, + 1407.0, + 1811.0, + 292.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1802.0, + 1405.0, + 1802.0, + 1405.0, + 1840.0, + 293.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1832.0, + 446.0, + 1832.0, + 446.0, + 1869.0, + 293.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1574.0, + 1406.0, + 1574.0, + 1406.0, + 1607.0, + 296.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1604.0, + 1405.0, + 1604.0, + 1405.0, + 1638.0, + 293.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1636.0, + 1405.0, + 1636.0, + 1405.0, + 1668.0, + 295.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1666.0, + 935.0, + 1666.0, + 935.0, + 1697.0, + 294.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1880.0, + 1406.0, + 1880.0, + 1406.0, + 1917.0, + 294.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1910.0, + 1404.0, + 1910.0, + 1404.0, + 1947.0, + 295.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2005.0, + 295.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2001.0, + 769.0, + 2001.0, + 769.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 585.0, + 1404.0, + 585.0, + 1404.0, + 621.0, + 296.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 616.0, + 1334.0, + 616.0, + 1334.0, + 654.0, + 293.0, + 654.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1294, + 1404, + 1294, + 1404, + 1418, + 298, + 1418 + ], + "score": 0.92 + }, + { + "category_id": 1, + "poly": [ + 297, + 1089, + 1405, + 1089, + 1405, + 1275, + 297, + 1275 + ], + "score": 0.912 + }, + { + "category_id": 1, + "poly": [ + 298, + 1601, + 1404, + 1601, + 1404, + 1726, + 298, + 1726 + ], + "score": 0.911 + }, + { + "category_id": 1, + "poly": [ + 296, + 1745, + 1404, + 1745, + 1404, + 1871, + 296, + 1871 + ], + "score": 0.905 + }, + { + "category_id": 1, + "poly": [ + 297, + 812, + 1404, + 812, + 1404, + 907, + 297, + 907 + ], + "score": 0.903 + }, + { + "category_id": 1, + "poly": [ + 293, + 1519, + 1401, + 1519, + 1401, + 1584, + 293, + 1584 + ], + "score": 0.902 + }, + { + "category_id": 1, + "poly": [ + 296, + 1006, + 1398, + 1006, + 1398, + 1072, + 296, + 1072 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.895 + }, + { + "category_id": 1, + "poly": [ + 293, + 1437, + 1400, + 1437, + 1400, + 1500, + 293, + 1500 + ], + "score": 0.89 + }, + { + "category_id": 1, + "poly": [ + 298, + 669, + 1404, + 669, + 1404, + 793, + 298, + 793 + ], + "score": 0.889 + }, + { + "category_id": 1, + "poly": [ + 288, + 586, + 1400, + 586, + 1400, + 650, + 288, + 650 + ], + "score": 0.885 + }, + { + "category_id": 1, + "poly": [ + 295, + 925, + 1401, + 925, + 1401, + 988, + 295, + 988 + ], + "score": 0.882 + }, + { + "category_id": 1, + "poly": [ + 297, + 330, + 1404, + 330, + 1404, + 454, + 297, + 454 + ], + "score": 0.882 + }, + { + "category_id": 1, + "poly": [ + 297, + 1972, + 1399, + 1972, + 1399, + 2034, + 297, + 2034 + ], + "score": 0.868 + }, + { + "category_id": 1, + "poly": [ + 298, + 473, + 1401, + 473, + 1401, + 568, + 298, + 568 + ], + "score": 0.865 + }, + { + "category_id": 0, + "poly": [ + 300, + 227, + 488, + 227, + 488, + 262, + 300, + 262 + ], + "score": 0.855 + }, + { + "category_id": 1, + "poly": [ + 295, + 278, + 1249, + 278, + 1249, + 311, + 295, + 311 + ], + "score": 0.849 + }, + { + "category_id": 1, + "poly": [ + 296, + 1889, + 1399, + 1889, + 1399, + 1953, + 296, + 1953 + ], + "score": 0.841 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.776 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 227.0, + 490.0, + 227.0, + 490.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1291.0, + 1404.0, + 1291.0, + 1404.0, + 1332.0, + 292.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1324.0, + 1404.0, + 1324.0, + 1404.0, + 1360.0, + 323.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1355.0, + 1406.0, + 1355.0, + 1406.0, + 1394.0, + 321.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1388.0, + 632.0, + 1388.0, + 632.0, + 1421.0, + 325.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1091.0, + 1403.0, + 1091.0, + 1403.0, + 1123.0, + 296.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1119.0, + 1406.0, + 1119.0, + 1406.0, + 1158.0, + 321.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1151.0, + 1405.0, + 1151.0, + 1405.0, + 1186.0, + 321.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1182.0, + 1405.0, + 1182.0, + 1405.0, + 1216.0, + 321.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1213.0, + 1403.0, + 1213.0, + 1403.0, + 1248.0, + 323.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1243.0, + 733.0, + 1243.0, + 733.0, + 1279.0, + 323.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1601.0, + 1405.0, + 1601.0, + 1405.0, + 1638.0, + 292.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1631.0, + 1405.0, + 1631.0, + 1405.0, + 1669.0, + 318.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1662.0, + 1405.0, + 1662.0, + 1405.0, + 1700.0, + 320.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1694.0, + 846.0, + 1694.0, + 846.0, + 1728.0, + 320.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1743.0, + 1405.0, + 1743.0, + 1405.0, + 1782.0, + 293.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1777.0, + 1404.0, + 1777.0, + 1404.0, + 1810.0, + 323.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1807.0, + 1405.0, + 1807.0, + 1405.0, + 1843.0, + 322.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1839.0, + 754.0, + 1839.0, + 754.0, + 1875.0, + 322.0, + 1875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 810.0, + 1405.0, + 810.0, + 1405.0, + 849.0, + 292.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 840.0, + 1404.0, + 840.0, + 1404.0, + 880.0, + 321.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 875.0, + 879.0, + 875.0, + 879.0, + 907.0, + 322.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1519.0, + 1404.0, + 1519.0, + 1404.0, + 1555.0, + 296.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1550.0, + 734.0, + 1550.0, + 734.0, + 1585.0, + 322.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1001.0, + 1404.0, + 1001.0, + 1404.0, + 1047.0, + 294.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1039.0, + 760.0, + 1039.0, + 760.0, + 1073.0, + 322.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1434.0, + 1403.0, + 1434.0, + 1403.0, + 1475.0, + 293.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1467.0, + 532.0, + 1467.0, + 532.0, + 1500.0, + 325.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 668.0, + 1405.0, + 668.0, + 1405.0, + 707.0, + 295.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 701.0, + 1404.0, + 701.0, + 1404.0, + 736.0, + 321.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 731.0, + 1405.0, + 731.0, + 1405.0, + 766.0, + 320.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 763.0, + 778.0, + 763.0, + 778.0, + 795.0, + 322.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 583.0, + 1402.0, + 583.0, + 1402.0, + 623.0, + 292.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 619.0, + 658.0, + 619.0, + 658.0, + 649.0, + 321.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 922.0, + 1405.0, + 922.0, + 1405.0, + 962.0, + 293.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 955.0, + 715.0, + 955.0, + 715.0, + 991.0, + 328.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 327.0, + 1405.0, + 327.0, + 1405.0, + 368.0, + 294.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 361.0, + 1404.0, + 361.0, + 1404.0, + 394.0, + 321.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 392.0, + 1405.0, + 392.0, + 1405.0, + 428.0, + 322.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 422.0, + 745.0, + 422.0, + 745.0, + 454.0, + 321.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1968.0, + 1405.0, + 1968.0, + 1405.0, + 2009.0, + 295.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2004.0, + 819.0, + 2004.0, + 819.0, + 2033.0, + 323.0, + 2033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 474.0, + 1405.0, + 474.0, + 1405.0, + 508.0, + 296.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 506.0, + 1405.0, + 506.0, + 1405.0, + 540.0, + 324.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 535.0, + 1031.0, + 535.0, + 1031.0, + 569.0, + 323.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 277.0, + 1253.0, + 277.0, + 1253.0, + 315.0, + 293.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1889.0, + 1404.0, + 1889.0, + 1404.0, + 1925.0, + 295.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1918.0, + 901.0, + 1918.0, + 901.0, + 1954.0, + 322.0, + 1954.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 301, + 1910, + 1404, + 1910, + 1404, + 2033, + 301, + 2033 + ], + "score": 0.912 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.886 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1403, + 229, + 1403, + 354, + 298, + 354 + ], + "score": 0.883 + }, + { + "category_id": 1, + "poly": [ + 298, + 727, + 1404, + 727, + 1404, + 851, + 298, + 851 + ], + "score": 0.875 + }, + { + "category_id": 1, + "poly": [ + 295, + 1491, + 1401, + 1491, + 1401, + 1584, + 295, + 1584 + ], + "score": 0.869 + }, + { + "category_id": 1, + "poly": [ + 299, + 616, + 1399, + 616, + 1399, + 712, + 299, + 712 + ], + "score": 0.865 + }, + { + "category_id": 1, + "poly": [ + 297, + 477, + 1405, + 477, + 1405, + 602, + 297, + 602 + ], + "score": 0.861 + }, + { + "category_id": 1, + "poly": [ + 296, + 1302, + 1401, + 1302, + 1401, + 1398, + 296, + 1398 + ], + "score": 0.861 + }, + { + "category_id": 1, + "poly": [ + 299, + 1769, + 1403, + 1769, + 1403, + 1894, + 299, + 1894 + ], + "score": 0.857 + }, + { + "category_id": 1, + "poly": [ + 301, + 1194, + 1399, + 1194, + 1399, + 1289, + 301, + 1289 + ], + "score": 0.853 + }, + { + "category_id": 1, + "poly": [ + 300, + 369, + 1400, + 369, + 1400, + 464, + 300, + 464 + ], + "score": 0.85 + }, + { + "category_id": 1, + "poly": [ + 298, + 1599, + 1404, + 1599, + 1404, + 1754, + 298, + 1754 + ], + "score": 0.836 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 865, + 2088, + 865, + 2112, + 837, + 2112 + ], + "score": 0.835 + }, + { + "category_id": 1, + "poly": [ + 291, + 1412, + 1402, + 1412, + 1402, + 1476, + 291, + 1476 + ], + "score": 0.826 + }, + { + "category_id": 1, + "poly": [ + 291, + 866, + 1401, + 866, + 1401, + 931, + 291, + 931 + ], + "score": 0.819 + }, + { + "category_id": 1, + "poly": [ + 299, + 1085, + 1400, + 1085, + 1400, + 1179, + 299, + 1179 + ], + "score": 0.816 + }, + { + "category_id": 1, + "poly": [ + 298, + 945, + 1404, + 945, + 1404, + 1070, + 298, + 1070 + ], + "score": 0.811 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1909.0, + 1405.0, + 1909.0, + 1405.0, + 1946.0, + 295.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1938.0, + 1405.0, + 1938.0, + 1405.0, + 1979.0, + 320.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2009.0, + 321.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 2003.0, + 514.0, + 2003.0, + 514.0, + 2033.0, + 324.0, + 2033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 295.0, + 323.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 289.0, + 1407.0, + 289.0, + 1407.0, + 328.0, + 321.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 323.0, + 1403.0, + 323.0, + 1403.0, + 356.0, + 322.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 726.0, + 1405.0, + 726.0, + 1405.0, + 765.0, + 292.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 757.0, + 1405.0, + 757.0, + 1405.0, + 794.0, + 321.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 790.0, + 1404.0, + 790.0, + 1404.0, + 823.0, + 321.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 819.0, + 946.0, + 819.0, + 946.0, + 853.0, + 321.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1492.0, + 1403.0, + 1492.0, + 1403.0, + 1525.0, + 295.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1522.0, + 1404.0, + 1522.0, + 1404.0, + 1556.0, + 322.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1553.0, + 526.0, + 1553.0, + 526.0, + 1585.0, + 324.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 615.0, + 1404.0, + 615.0, + 1404.0, + 654.0, + 293.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 646.0, + 1404.0, + 646.0, + 1404.0, + 689.0, + 322.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 679.0, + 1344.0, + 679.0, + 1344.0, + 717.0, + 323.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 480.0, + 1405.0, + 480.0, + 1405.0, + 513.0, + 295.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 508.0, + 1405.0, + 508.0, + 1405.0, + 543.0, + 320.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 538.0, + 1405.0, + 538.0, + 1405.0, + 577.0, + 320.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 569.0, + 641.0, + 569.0, + 641.0, + 605.0, + 321.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1297.0, + 1406.0, + 1297.0, + 1406.0, + 1344.0, + 293.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1334.0, + 1406.0, + 1334.0, + 1406.0, + 1370.0, + 321.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1366.0, + 1162.0, + 1366.0, + 1162.0, + 1400.0, + 323.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1766.0, + 1404.0, + 1766.0, + 1404.0, + 1810.0, + 292.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1801.0, + 1404.0, + 1801.0, + 1404.0, + 1840.0, + 319.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1831.0, + 1404.0, + 1831.0, + 1404.0, + 1870.0, + 319.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1862.0, + 1141.0, + 1862.0, + 1141.0, + 1898.0, + 323.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1192.0, + 1403.0, + 1192.0, + 1403.0, + 1231.0, + 292.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1226.0, + 1403.0, + 1226.0, + 1403.0, + 1260.0, + 323.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1256.0, + 829.0, + 1256.0, + 829.0, + 1290.0, + 322.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 369.0, + 1404.0, + 369.0, + 1404.0, + 407.0, + 294.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 398.0, + 1405.0, + 398.0, + 1405.0, + 438.0, + 320.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 431.0, + 1275.0, + 431.0, + 1275.0, + 469.0, + 322.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1599.0, + 1404.0, + 1599.0, + 1404.0, + 1635.0, + 296.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1630.0, + 1405.0, + 1630.0, + 1405.0, + 1667.0, + 323.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1659.0, + 1404.0, + 1659.0, + 1404.0, + 1699.0, + 321.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1688.0, + 1406.0, + 1688.0, + 1406.0, + 1732.0, + 318.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1722.0, + 532.0, + 1722.0, + 532.0, + 1760.0, + 322.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1412.0, + 1404.0, + 1412.0, + 1404.0, + 1448.0, + 295.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1444.0, + 816.0, + 1444.0, + 816.0, + 1477.0, + 322.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 866.0, + 1404.0, + 866.0, + 1404.0, + 902.0, + 295.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 898.0, + 968.0, + 898.0, + 968.0, + 931.0, + 324.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1086.0, + 1402.0, + 1086.0, + 1402.0, + 1120.0, + 297.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1117.0, + 1404.0, + 1117.0, + 1404.0, + 1149.0, + 323.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1145.0, + 1018.0, + 1145.0, + 1018.0, + 1183.0, + 323.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 943.0, + 1405.0, + 943.0, + 1405.0, + 982.0, + 295.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 976.0, + 1404.0, + 976.0, + 1404.0, + 1011.0, + 323.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1005.0, + 1405.0, + 1005.0, + 1405.0, + 1043.0, + 320.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1034.0, + 466.0, + 1034.0, + 466.0, + 1076.0, + 319.0, + 1076.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/r1l-VeSKwS/r1l-VeSKwS.md b/parse/train/r1l-VeSKwS/r1l-VeSKwS.md new file mode 100644 index 0000000000000000000000000000000000000000..4dc764fed02067146f4e8d88fcd95e5ca8b4631e --- /dev/null +++ b/parse/train/r1l-VeSKwS/r1l-VeSKwS.md @@ -0,0 +1,432 @@ +# SEMANTICADV: GENERATING ADVERSARIAL EXAMPLES VIA ATTRIBUTE-CONDITIONAL IMAGE EDITING + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +Deep neural networks (DNNs) have achieved great success in various applications due to their strong expressive power. However, recent studies have shown that DNNs are vulnerable to adversarial examples which are manipulated instances targeting to mislead DNNs to make incorrect predictions. Currently, most such adversarial examples try to guarantee “subtle perturbation" by limiting the $L _ { p }$ norm of the perturbation. In this paper, we aim to explore the impact of semantic manipulation on DNNs predictions by manipulating the semantic attributes of images and generate “unrestricted adversarial examples". In particular, we propose an algorithm SemanticAdv which leverages disentangled semantic factors to generate adversarial perturbation by altering controlled semantic attributes to fool the learner towards various “adversarial" targets. We conduct extensive experiments to show that the semantic based adversarial examples can not only fool different learning tasks such as face verification and landmark detection, but also achieve high targeted attack success rate against real-world black-box services such as Azure face verification service based on transferability. To further demonstrate the applicability of SemanticAdv beyond face recognition domain, we also generate semantic perturbations on street-view images. Such adversarial examples with controlled semantic manipulation can shed light on further understanding about vulnerabilities of DNNs as well as potential defensive approaches. + +# 1 INTRODUCTION + +Deep neural networks (DNNs) have demonstrated great successes in advancing the state-of-the-art performance of discriminative tasks (Krizhevsky et al., 2012; Goodfellow et al., 2016; He et al., 2016; Collobert & Weston, 2008; Deng et al., 2013; Silver et al., 2016). However, recent research found that DNNs are vulnerable to adversarial examples which are carefully crafted instances aiming to induce arbitrary prediction errors for learning systems. Such adversarial examples containing small magnitude of perturbation have shed light on understanding and discovering potential vulnerabilities of DNNs (Szegedy et al., 2013; Goodfellow et al., 2014b; Moosavi-Dezfooli et al., 2016; Papernot et al., 2016; Carlini & Wagner, 2017; Xiao et al., 2018b;c;a; 2019). Most existing work focused on constructing adversarial examples by adding ${ \mathcal { L } } _ { p }$ bounded pixel-wise perturbations (Goodfellow et al., 2014b) or spatially transforming the image (Xiao et al., 2018c; Engstrom et al., 2017) (e.g., in-plane rotation or out-of-plane rotation). Generating unrestricted perturbations with semantically meaningful patterns is an important yet under-explored field. + +At the same time, deep generative models have demonstrated impressive performance in learning disentangled semantic factors through data generation in an unsupervised (Radford et al., 2015; Karras et al., 2018; Brock et al., 2019) or weakly-supervised manner based on semantic attributes (Yan et al., 2016; Choi et al., 2018). Empirical findings in (Yan et al., 2016; Zhu et al., 2016a; Radford et al., 2015) demonstrated that a simple linear interpolation on the learned image manifold can produce smooth visual transitions between a pair of input images. + +In this paper, we introduce a novel attack SemanticAdv which generates unrestricted perturbations with semantically meaningful patterns. Motivated by the findings mentioned above, we leverage an attribute-conditional image editing model (Choi et al., 2018) to synthesize adversarial examples by interpolating between source and target images in the feature-map space. Here, we focus on changing a single attribute dimension to achieve adversarial goals while keeping the generated adversarial image reasonably-looking (e.g., see Figure 1). To validate the effectiveness of the proposed attack method, we consider two tasks, namely, face verification and landmark detection, as face recognition field has been extensively explored and the commercially used face models are relatively robust since they require a low false positive rate. We conduct both qualitative and quantitative evaluations on CelebA dataset (Liu et al., 2015). To demonstrate the applicability of SemanticAdv beyond face domain, we further extend SemanticAdv to generate adversarial street-view images. We treat semantic layouts as input attributes and use the image editing model (Hong et al., 2018) pre-trained on Cityscape dataset (Cordts et al., 2016). Please find more visualization results on the anonymous website: https://sites.google.com/view/generate-semantic-adv-example. + +![](images/464ae493fc35cd3e8384b15d0c1e64cf82fc55148d4ff781b38e7a7264ac13e9.jpg) +Figure 1: Left: Overview of the proposed SemanticAdv. Right: Illustration of our SemanticAdv in the real world face verification platform. Note that the confidence denotes the likelihood that two faces belong to the same person. + +The contributions of the proposed SemanticAdv are three-folds. First, we propose a novel semanticbased attack method to generate unrestricted adversarial examples by feature-space interpolation. Second, the proposed method is able to generate semantically-controllable perturbations due to the attribute-conditioned modeling. This allows us to analyze the robustness of a recognition system against different types of semantic attacks. Third, as a side benefit, the proposed attack exhibits high transferability and leads to a $65 \%$ query-free black-box attack success rate on a real-world face verification platform, which outperforms the pixel-wise perturbations in attacking existing defense methods. + +# 2 RELATED WORK + +Semantic image editing. Semantic image synthesis and manipulation is a popular research topic in machine learning, graphics and vision. Thanks to recent advances in deep generative models (Kingma & Welling, 2014; Goodfellow et al., 2014a; Oord et al., 2016) and the empirical analysis of deep classification networks (Krizhevsky et al., 2012; Simonyan & Zisserman, 2014; Szegedy et al., 2015), past few years have witnessed tremendous breakthroughs towards high-fidelity pure image generation (Radford et al., 2015; Karras et al., 2018; Brock et al., 2019), attribute-to-image generation (Yan et al., 2016; Choi et al., 2018), text-to-image generation (Mansimov et al., 2015; Reed et al., 2016; Van den Oord et al., 2016; Odena et al., 2017; Zhang et al., 2017; Johnson et al., 2018), and imageto-image translation (Isola et al., 2017; Zhu et al., 2017; Liu et al., 2017; Wang et al., 2018b; Hong et al., 2018). + +Adversarial examples. Generating $L _ { p }$ bounded adversarial perturbation has been extensively studied recently (Szegedy et al., 2013; Goodfellow et al., 2014b; Moosavi-Dezfooli et al., 2016; Papernot et al., 2016; Carlini & Wagner, 2017; Xiao et al., 2018b). To further explore diverse adversarial attacks and potentially help inspire defense mechanisms, it is important to generate the socalled “unrestricted" adversarial examples which contain unrestricted magnitude of perturbation while still preserve perceptual realism Brown et al. (2018). Recently, Xiao et al. (2018c); Engstrom et al. (2017) propose to spatially transform the image patches instead of adding pixel-wise perturbation, while such spatial transformation does not consider semantic information. Our proposed semanticAdv focuses on generating unrestricted perturbation with semantically meaningful patterns guided by visual attributes. + +Relevant to our work, Song et al. (2018) proposed to synthesize adversarial examples with an unconditional generative model. Bhattad et al. (2019) studied semantic transformation in only the color or texture space. Compared to these works, semanticAdv is able to generate adversarial examples in a controllable fashion using specific visual attributes by performing manipulation in the feature space. We further analyze the robustness of the recognition system by generating adversarial examples guided by different visual attributes. Concurrent to our work, Joshi et al. (2019) proposed to generate semantic-based attacks against a restricted binary classifier, while our attack is able to mislead the model towards arbitrary adversarial targets. They conduct the manipulation within the attribution space which is less flexible and effective than our proposed feature-space interpolation. + +# 3 SEMANTIC ADVERSARIAL EXAMPLES + +# 3.1 PROBLEM DEFINITION + +Let $\mathcal { M }$ be a machine learning model trained on a dataset $\mathcal { D } = \{ ( { \bf x } , { \bf y } ) \}$ consisting of image-label pairs, where $\mathbf { x } \in \mathbb { R } ^ { H \times W \times D _ { I } }$ and $\mathbf { y } \in \mathbb { R } ^ { D _ { L } }$ denote the image and the ground-truth label, respectively. Here, $H , W$ , $D _ { I }$ , and $D _ { L }$ denote the image height, image width, number of image channels, and label dimensions, respectively. For each image $\mathbf { x }$ , our model $\mathcal { M }$ makes a prediction $\hat { \mathbf { y } } = \mathcal { M } ( \mathbf { x } ) \in \mathbb { R } ^ { D _ { L } }$ . Given a target image-label pair $\left( \mathbf { x } ^ { \mathrm { { t g t } } } , \mathbf { y } ^ { \mathrm { { t g t } } } \right)$ and $\mathbf { y } \neq \mathbf { y } ^ { \mathrm { t g t } }$ , a traditional attacker aims to synthesize adversarial examples $\{ { \bf x } ^ { \mathrm { a d v } } \}$ by adding pixel-wise perturbations to or spatially transforming the original image x such that $\dot { \mathcal { M } } ( \mathbf { x } ^ { \mathrm { { a d v } } } ) = \mathbf { y } ^ { \mathrm { { t g t } } }$ . + +In this work, we introduce the concept of semantic attacker that aims at generating adversarial examples by adding semantically meaningful perturbation with a conditional generative model $\mathcal { G }$ . Compared to traditional attacker that usually produces pixel-wise perturbations, the proposed method is able to produce semantically meaningful perturbations. + +Semantic image editing. For simplicity, we start with the formulation where the input attribute is represented as a compact vector. This formulation can be directly extended to other input attribute formats including semantic layouts. Let $\mathbf { c } \in \mathbb { R } ^ { D _ { c } }$ be an attribute representation reflecting the semantic factors (e.g., expression or hair color of a portrait image) of image $\mathbf { x }$ , where $D _ { C }$ indicates the attribute dimension and $c _ { i } \in \{ 0 , 1 \}$ indicates the appearance of $i$ -th attribute. Here, our goal is to use the conditional generator for semantic image editing. For example, given a portrait image of a girl with black hair and blonde hair as the new attribute, our generator is supposed to synthesize a new image that turns the girl’s hair from black to blonde. More specifically, we denote the new attribute as $\mathbf { c } ^ { \mathrm { { n e w } } } \in \mathbb { R } ^ { D _ { C } }$ such that the synthesized image is given by $\mathbf { x } ^ { \mathrm { n e w } } = \mathcal { G } ( \bar { \mathbf { x } , \mathbf { c } ^ { \mathrm { n e w } } } )$ . In the special case when there is no attribute change $( \mathbf { c } = \mathbf { c } ^ { \mathrm { n e w } }$ ), the generator simply reconstructs the input: $\mathbf { x } = \mathcal { G } ( \mathbf { x } , \mathbf { c } )$ . Supported by the findings mentioned in (Bengio et al., 2013; Reed et al., 2014), our synthesized image $\mathbf { x } ^ { \mathrm { n e w } }$ should fall close to the data manifold if we constrain the change of attribute values to be sufficiently small (e.g., we only update one semantic attribute at a time). In addition, we can potentially generate many such images by linearly interpolating between the semantic embeddings of the conditional generator $\mathcal { G }$ using original image $\mathbf { x }$ and the synthesized image $\mathbf { x } ^ { \mathrm { n e w } }$ with the augmented attribute. + +Attribute-space interpolation. We start with a simple solution (detailed in Eq. 1) assuming the adversarial example can be found by directly interpolating in the attribute-space. Given a pair of attributes c and $\mathbf { c } ^ { \mathrm { n e w } }$ , we introduce an interpolation parameter $\alpha \in ( 0 , 1 )$ to generate the augmented attribute vector $\mathbf { c } ^ { \ast } \in \mathbb { R } ^ { D _ { C } }$ (see Eq. 1). Given augmented attribute $\mathbf { c } ^ { * }$ and original image $\mathbf { x }$ , we produce the synthesized image by the generator $\mathcal { G }$ . For our notation purpose, we also introduce a delegated function $\tau _ { \mathscr { G } }$ as a re-parametrization for the generator $\mathcal { G }$ . Our formulation is also supported by the empirical results on attribute-conditioned image progression (Yan et al., 2016; Radford et al., 2015) that a well-trained generative model has the capability to synthesize a sequence of images with smooth attribute transitions. + +$$ +\begin{array} { r l } & { \mathbf { x } ^ { \mathrm { a d v } } = \arg \operatorname* { m i n } _ { \alpha } \mathcal { L } \big ( \mathcal { T } _ { \mathcal { G } } ( \alpha ; \mathbf { x } , \mathbf { c } , \mathbf { c } ^ { \mathrm { n e w } } ) \big ) } \\ & { \mathrm { w h e r e ~ } \mathcal { T } _ { \mathcal { G } } \big ( \alpha ; \mathbf { x } , \mathbf { c } , \mathbf { c } ^ { \mathrm { n e w } } \big ) = \mathcal { G } \big ( \mathbf { x } , \mathbf { c } ^ { \ast } \big ) \mathrm { ~ a n d ~ } \mathbf { c } ^ { \ast } = \alpha \cdot \mathbf { c } + ( 1 - \alpha ) \cdot \mathbf { c } ^ { \mathrm { n e w } } } \end{array} +$$ + +Feature-map interpolation. Alternatively, we propose to interpolate using the feature map produced by the generator $\mathcal { G } = \mathcal { G } _ { \mathrm { d e c } } \circ \mathcal { G } _ { \mathrm { e n c } }$ . Here, $\mathcal { G } _ { \mathrm { e n c } }$ is the encoder module that takes the image as input and outputs the feature map. Similarly, $\mathcal { G } _ { \mathrm { d e c } }$ is the decoder module that takes the feature map as input and outputs the synthesized image. Let $\mathbf { f } ^ { * } = { \mathcal { G } } _ { \mathrm { e n c } } ( \mathbf { x } , \mathbf { c } ) \in \mathbb { R } ^ { H _ { F } \times W _ { F } \times C _ { F } }$ be the feature map of an intermediate layer in the generator, where $H _ { F }$ , $W _ { F }$ and $C _ { F }$ indicate the height, width, and + +number of channels in the feature map. + +$$ +\begin{array} { r l } & { \mathbf { x } ^ { \mathrm { a d v } } = \arg \operatorname* { m i n } _ { \alpha } \mathcal { L } ( \mathcal { T } _ { \mathcal { G } } ( \alpha ; \mathbf { x } , \mathbf { c } , \mathbf { c } ^ { \mathrm { n e w } } ) ) } \\ & { \mathrm { w h e r e ~ } \mathcal { T } _ { \mathcal { G } } ( \alpha ; \mathbf { x } , \mathbf { c } , \mathbf { c } ^ { \mathrm { n e w } } ) = \mathcal { G } _ { \mathrm { d e c } } ( \mathbf { f } ^ { * } ) , } \\ & { \mathbf { f } ^ { * } = \alpha \odot \mathcal { G } _ { \mathrm { e n c } } ( \mathbf { x } , \mathbf { c } ) + ( \mathbf { 1 } - \alpha ) \odot \mathcal { G } _ { \mathrm { e n c } } ( \mathbf { x } , \mathbf { c } ^ { \mathrm { n e w } } ) } \end{array} +$$ + +Compared to attribute-space interpolation which is parameterized by a scalar, we parameterize feature-map interpolation by a tensor $\alpha \in \mathbb { R } ^ { H _ { F } \times W _ { F } \times C _ { F } }$ $( \alpha _ { h , w , k } \in ( 0 , 1 )$ , where $1 \leq h \leq H _ { F }$ , $1 \leq w \leq W _ { F }$ , and $1 \leq k \leq C _ { F } ,$ ) with the same shape as the feature map. Compared to linear interpolation over attribute-space, such design introduces more flexibility when interpolating between the original image and the synthesized image. Empirical results in Section 4.2 show our design is critical to the adversarial attack success rate. + +# 3.2 ADVERSARIAL OPTIMIZATION OBJECTIVES + +As we see in Eq. 3, we obtain the adversarial image $\mathbf { x } ^ { \mathrm { a d v } }$ by minimizing the objective $\mathcal { L } ( \cdot )$ with respect to the synthesized image $\mathcal { T } _ { \mathcal { G } } ( \alpha ; { \mathbf { x } } , { \mathbf { c } } , { \mathbf { c } } ^ { \mathrm { n e w } } )$ , which is defined in Eq.(1) and Eq.(2) respectively. Here, each synthesized image $\mathcal { T } _ { \mathcal { G } } ( \alpha ; { \mathbf { x } } , { \mathbf { c } } , { \mathbf { c } } ^ { \mathrm { n e w } } )$ is produced by the interpolation using the conditional generator $\mathcal { G }$ . In our objective function, the first term is the adversarial metric, the second term is a smoothness constraint, and $\lambda$ is used to control the balance between the two terms. The adversarial metric is minimized once the model $\mathcal { M }$ has been successfully attacked towards the target image-label pair $\left( \mathbf { x } ^ { \mathrm { { t g t } } } , \mathbf { y } ^ { \mathrm { { t g t } } } \right)$ . For identify verification, $\mathbf { y } ^ { \mathrm { t g t } }$ is the identity representation of the target image; For structured prediction tasks in our paper, $\mathbf { y } ^ { \mathrm { t g t } }$ either represents certain coordinates (landmark detection) or semantic label maps (semantic segmentation). + +$$ +\begin{array} { r } { \begin{array} { c } { \mathbf { x } ^ { \mathrm { a d v } } = \mathop { \mathrm { a r g m i n } _ { \alpha } } \mathcal { L } ( \mathcal { T } _ { \mathcal { G } } ( \alpha ; \mathbf { x } , \mathbf { c } , \mathbf { c } ^ { \mathrm { n e w } } ) ) } \\ { \mathcal { L } ( \mathcal { T } _ { \mathcal { G } } ( \alpha ; \mathbf { x } , \mathbf { c } , \mathbf { c } ^ { \mathrm { n e w } } ) ) = \mathcal { L } _ { \mathrm { a d v } } ( \mathcal { T } _ { \mathcal { G } } ( \alpha ; \mathbf { x } , \mathbf { c } , \mathbf { c } ^ { \mathrm { n e w } } ) ; \mathcal { M } , \mathbf { y } ^ { \mathrm { t g t } } ) + \lambda \cdot \mathcal { L } _ { \mathrm { s m o o t h } } ( \alpha ) } \end{array} } \end{array} +$$ + +Identity verification. In the identity verification task, two images are considered to be the same identity if the corresponding identity embeddings from the verification model $\mathcal { M }$ are reasonably close. + +$$ +\mathcal { L } _ { \mathrm { a d v } } \big ( \mathcal { T } _ { \mathcal { G } } ( \alpha ; \mathbf { x } , \mathbf { c } , \mathbf { c } ^ { \mathrm { n e w } } ) ; \mathcal { M } , \mathbf { y } ^ { \mathrm { t g t } } \big ) = \operatorname* { m a x } \big ( \kappa , \Phi _ { \mathcal { M } } ^ { \mathrm { i d } } ( \mathcal { T } _ { \mathcal { G } } ( \alpha ; \mathbf { x } , \mathbf { c } , \mathbf { c } ^ { \mathrm { n e w } } ) , \mathbf { x } ^ { \mathrm { t g t } } ) \big ) +$$ + +As we see in Eq. 4, $\Phi _ { \mathcal { M } } ^ { \mathrm { i d } } ( \cdot , \cdot )$ measures the distance between two identity embeddings from the model $\mathcal { M }$ , where the normalized $L _ { 2 }$ distance is used in our setting. In addition, we introduce the parameter $\kappa$ representing the constant related to the false positive rate (FPR) threshold computed from the development set. + +Structured prediction. For structured prediction tasks such as landmark detection and semantic segmentation, we use Houdini objective proposed in Cisse et al. (2017) as our adversarial metric and select the target landmark (semantic segmentation) target as $\mathbf { y } ^ { \mathrm { t g t } }$ . In addition, $\Phi _ { \mathcal { M } } ( \cdot , \cdot )$ is a scoring function for each image-label pair and $\gamma$ is the threshold. + +$$ +\begin{array} { r l } & { \mathcal { L } _ { \mathrm { a d v } } \big ( \mathcal { T } _ { \mathcal { G } } \big ( \alpha ; \mathbf { x } , \mathbf { c } , \mathbf { c } ^ { \mathrm { n e w } } \big ) ; \mathcal { M } , \mathbf { y } ^ { \mathrm { t g t } } \big ) = P _ { \gamma \sim \mathcal { N } ( 0 , 1 ) } \big [ \Phi _ { \mathcal { M } } \big ( \mathcal { T } _ { \mathcal { G } } \big ( \alpha ; \mathbf { x } , \mathbf { c } , \mathbf { c } ^ { \mathrm { n e w } } \big ) , \mathbf { y } ^ { * } \big ) } \\ & { \qquad - \Phi _ { \mathcal { M } } \big ( \mathcal { T } _ { \mathcal { G } } \big ( \alpha ; \mathbf { x } , \mathbf { c } , \mathbf { c } ^ { \mathrm { n e w } } \big ) , \mathbf { y } ^ { \mathrm { t g t } } \big ) < \gamma \big ] \cdot l \big ( \mathbf { y } ^ { * } , \mathbf { y } ^ { \mathrm { t g t } } \big ) } \end{array} +$$ + +where $\mathbf { y } ^ { * } = \mathcal { M } ( \mathcal { T } _ { \mathcal { G } } ( \alpha ; \mathbf { x } , \mathbf { c } , \mathbf { c } ^ { \mathrm { n e w } } ) )$ and $l ( \mathbf { y } ^ { * } , \mathbf { y } ^ { \mathrm { t g t } } )$ is task loss decided by the specific adversarial target. + +Interpolation smoothness $\mathcal { L } _ { \mathrm { s m o o t h } }$ . As the tensor to be interpolated in the feature-map space has far more parameters compared to the attribute itself, we propose to enforce a smoothness constraint on the tensor $\alpha$ used in feature-map interpolation. As we see in Eq. 6, the smoothness loss encourages the interpolation tensors to consist of piece-wise constant patches spatially, which has been widely used as a pixel-wise de-noising objective for natural image processing (Mahendran & Vedaldi, 2015; Johnson et al., 2016). + +$$ +\mathcal { L } _ { \mathrm { s m o o t h } } ( \alpha ) = \sum _ { h = 1 } ^ { H _ { F } - 1 } \sum _ { w = 1 } ^ { W _ { F } } \| \alpha _ { h + 1 , w } - \alpha _ { h , w } \| _ { 2 } ^ { 2 } + \sum _ { h = 1 } ^ { H _ { F } } \sum _ { w = 1 } ^ { W _ { F } - 1 } \| \alpha _ { h , w + 1 } - \alpha _ { h , w } \| _ { 2 } ^ { 2 } +$$ + +# 4 EXPERIMENTS + +In the experimental section, we mainly focus on analyzing the proposed SemanticAdv in attacking state-of-the-art face recognition systems on CelebA (Liu et al., 2015) due to its wide applicability (e.g., identification for mobile payment) in the real world. In addition, we extend our attack to urban street scenes with semantic label maps as the condition. We attack the semantic segmentation model DRN-D-22 (Yu et al., 2017) previously trained on Cityscape (Cordts et al., 2016) by generating adversarial examples with dynamic objects manipulated (e.g., insert a car into the scene). + +The experimental section is organized as follows. First, we analyze the quality of generated adversarial examples and qualitatively compare our method with ${ \mathcal { L } } _ { p }$ bounded pixel-wise optimization-based method (Carlini & Wagner, 2017; Dong et al., 2018; Xie et al., 2019). Second, we provide both qualitative and quantitative results by controlling each of the semantic attributes at a time. In terms of attack transferability, we evaluate our proposed SemanticAdv on various settings and further demonstrate the effectiveness of our method via query-free black-box attacks against online face verification platforms. Third, we compare our method with the baseline against different defense methods on the face verification task. Fourth, we demonstrate that the proposed SemanticAdv also applies to the face landmark detection and street-view semantic segmentation. + +# 4.1 EXPERIMENTAL SETUP + +Face identity verification. We select ResNet-50 and ResNet-101 (He et al., 2016) trained on MS-Celeb-1M (Guo et al., 2016) as our face verification models. The models are trained using two different objectives, namely, softmax loss (Sun et al., 2014; Zhang et al., 2018) and cosine loss (Wang et al., 2018a). For simplicity, we use the notation “R-N-S” to indicate the model with $N$ -layer residual blocks as backbone trained using softmax loss, while “R-N-C” indicates the same backbone trained using cosine loss. The distance between face features is measured by normalized L2 distance. For R-101-S model, we decide the parameter $\kappa$ based on the false positive rate (FPR) for the identity verification task. Three different FPRs have been used: $1 0 ^ { - 3 }$ (with $\kappa = 1 . 2 4 ,$ ), $3 \times 1 0 ^ { - 4 }$ (with $\kappa = 1 . 0 5$ ), and $1 0 ^ { - 4 }$ (with $\kappa = 0 . 6 0 $ ). The distance metrics and selected thresholds are commonly used when evaluating the performance of face recognition model Klare et al. (2015); Kemelmacher-Shlizerman et al. (2016). Please check the Appendix (see Table B) for more details. To distinguish between the FPR we used in generating adversarial examples and the other FPR used in evaluation, we introduce two notations “Generating FPR (G-FPR)” and “Test FPR (T-FPR)”. For the experiment with query-free black-box API attacks, we use the online face verification services provided by ${ \mathrm { F a c e } } + +$ (fac) and AliYun (ali). + +Face landmark detection. We select Face Alignment Network (FAN) (Bulat & Tzimiropoulos, 2017b) trained on 300W-LP (Zhu et al., 2016b) and fine-tuned on 300-W (Sagonas et al., 2013) for 2D landmark detection. The network is constructed by stacking Hour-Glass network (Newell et al., 2016) with hierarchical block (Bulat & Tzimiropoulos, 2017a). Given a portrait image as input, FAN outputs 2D heatmaps which can be subsequently leveraged to yield 68 2D landmarks. + +Semantic attacks on face images. In our experiments, we randomly sample 1, 280 distinct identities form CelebA (Liu et al., 2015). To reduce the reconstruction error brought by the generator (e.g., $\mathbf { x } \neq \mathcal { G } ( \mathbf { x } , \mathbf { c } ) )$ in practice, we take one more step to obtain the updated feature map $\mathbf { f } ^ { \prime } = \mathcal { G } _ { \mathrm { e n c } } ( \mathbf { x } ^ { \prime } , \mathbf { c } )$ , where $\mathbf { x } ^ { \prime } = \mathrm { a r g m i n } _ { \mathbf { x } ^ { \prime } }$ $\| \mathcal { G } ( \mathbf { x } ^ { \prime } , \mathbf { c } ) - \mathbf { x } \|$ in feature-map interpolation. In our experiments, we use the last conv layer before upsampling in the generator as our as feature-map f given by the attack effectiveness. We also fix the parameter $\lambda$ (e.g., balances the adversarial loss and smoothness constraint in Eq. 3) to be 0.01 for both face verification and landmark detection. + +We used the StarGAN (Choi et al., 2018) for attribute-conditional image editing. In particular, we re-trained model on CelebA dataset (Liu et al., 2015) by aligning the face landmarks and then resizing images to resolution $1 1 2 \times 1 1 2$ . In addition, we select 17 identity-preserving attributes as our input condition, as such attributes related to facial expression and hair color. + +For each distinct identity pair $( \mathbf { x } , \mathbf { x } ^ { \mathrm { t g t } } )$ , we perform semanticAdv guided by each of the 17 attributes (e.g., we intentionally add or remove one specific attribute while keeping the rest unchanged). In total, for each image $\mathbf { x }$ , we generate 17 adversarial images with different augmented attributes. In the experiments, we select a pixel-wise adversarial attack method (Carlini & Wagner, 2017) (referred as CW) as our baseline for comparison. Compared to our proposed method, CW does not require visual attributes as part of the system, as it only generates one adversarial example for each instance. We refer the corresponding attack success rate as the instance-wise success rate in which the attack success rate is calculated for each instance. For each instance with 17 adversarial images using different augmented attributes, if one of the 17 resulting images can attack successfully, we count the attack of this instance as one success, vice verse. + +Semantic attacks on street-view images. We select DRN-D-22 (Yu et al., 2017) as our semantic segmentation model and fine-tune the model on image regions with resolution $2 5 6 \times 2 5 6$ . To synthesize semantic adversarial perturbations, we consider semantic label maps as the input attribute and leverage a generative image manipulation model (Hong et al., 2018) pre-trained on CityScape (Cordts et al., 2016) dataset. Given an input semantic label map at resolution $2 5 6 \times 2 5 6$ , we select a target object instance (e.g., a pedestrian) to attack. Then, we create a manipulated semantic label map by inserting another object instance (e.g., a car) in the vicinity of the target object. Similar to the experiments in the face domain, for both semantic label maps, we used the image manipulation encoder to extract features (with 1, 024 channels at spatial resolution $1 6 \times 1 6$ ) and conducted featurespace interpolation. We synthesized the final image by feeding the interpolated features to the image manipulation decoder. By searching the interpolation coefficient that maximizes the attack rate, we are able to fool the segmentation model with the synthesized final image. + +# 4.2 SemanticAdv ON FACE IDENTITY VERIFICATION + +Attribute-space vs. feature-space interpolation. First, we found that both attribute-space and feature-space interpolation could generate reasonable samples (see Figure I in Appendix). Compared to attribute-space interpolation, generating adversarial examples with feature-space interpolation produced much better quantitative results (see Table E in Appendix). We measured the attack success rate of attribute-space interpolation (with $\mathrm { G \mathrm { - } F P R = T \mathrm { - } F P R = 1 0 ^ { - 3 } , }$ ): $0 . 0 8 \%$ on R-101-S, $0 . 3 1 \%$ on R-101-C, and $0 . 1 6 \%$ on both R-50-S and R-50-C. While feature-space interpolation achieves almost $100 \%$ success rate on all those models (see Figure 3). We conjecture that this is because the high dimensional feature space can provide more manipulation freedom. + +![](images/bc14263add8b7333aa1ab335a48d188a488851f0eeec7d268fb45170e51410fe.jpg) +Figure 2: Qualitative comparisons between our proposed SemanticAdv and pixel-wise adversarial examples generated by CW (Carlini & Wagner, 2017). Along with the adversarial examples, we also provide the corresponding perturbations (residual) on the right. Perturbations generated by our SemanticAdv $\mathrm { \bf G } \mathrm { - } \mathrm { \bf F P R } = 1 0 ^ { - 3 } ,$ are unrestricted with semantically meaningful patterns. More results are shown in Appendix (see Figure N). + +Overall analysis. Figure 2 shows the generated adversarial images and corresponding perturbations against R-101-S of SemanticAdv and CW respectively. The text below each figure is the name of augmented attribute, the sign before the name represents “adding” (in red) or “removing” (in blue) the corresponding attribute from the original image. We see that SemanticAdv is able to generate perceptually reasonable examples guided by the corresponding attribute. In particular, SemanticAdv is able to generate perturbations on the corresponding regions correlated with the augmented attribute, while the perturbations of CW have no specific pattern and are evenly distributed across the image. + +Analysis: controlling single attribute. One of the key advantages of SemanticAdv is that we can generate adversarial perturbations in a more controllable fashion guided by the semantic attributes. This allows analyzing the robustness of a recognition system against different types of semantic attacks. We group the adversarial examples by augmented attributes in various settings. In Figure 3, we present the attack success rate against two face verification models, namely, R-101-S and R-101-C, guided by different attributes. We highlight the bar with light blue for G-FPR equals to $1 0 ^ { - 3 }$ and blue for G-FPR equals to $1 0 ^ { - 4 }$ , respectively. As we see in this figure, with a larger T-FPR $1 0 ^ { - 3 }$ , our SemanticAdv can achieve almost $100 \%$ attack success rate across different attributes. With a smaller T-FPR $1 0 ^ { - 4 }$ , we find that SemanticAdv guided by some attributes such as Mouth Slightly Open and Arched Eyebrows achieve less than $50 \%$ attack success rate, while the other attributes such as Pale Skin and Eyeglasses are relatively less affected. In summary, we found that SemanticAdv guided by attributes describing the local shape (e.g., mouth, earrings) achieve a relatively lower attack success rate compared to attributes relevant to the color (e.g., hair color) or entire face region (e.g., skin). This suggests that the face verification models used in our experiments are more robustly trained in terms of detecting local shapes compared to colors. Please note that in practice we have the flexibility to select attributes for attacking an image based on the perceptual quality and attack success rate. + +Figure 4 shows the adversarial examples with augmented semantic attributes against R-101-S model. The attribute names are shown in the bottom. The upper images are $\mathcal { G } ( \mathbf { x } , \mathbf { c } ^ { \mathrm { n e w } } )$ generated by StarGAN with augmented attribute $\mathbf { c } ^ { \mathrm { { n e w } } }$ where the lower images are the corresponding adversarial images with the same augmented attribute. + +![](images/f425e3990a1d398f6b0c0ca3d32c623c42afee2547c5a6fc767eb289e27fcc47.jpg) + +![](images/b9a1a0608405ea4229a059b5e220b91f0651c5175ef5d100359c433825165714.jpg) +Figure 3: Quantitative analysis on the attack success rate with different single-attribute attacks. In each figure, we show the results correspond to a larger FPR $( \mathrm { G } \mathrm { - } \mathrm { F P R } = \mathrm { T } \mathrm { - } \mathrm { F P R } = 1 0 ^ { - 3 } )$ in skyblue and the results correspond to a smaller FPR $\mathrm { \langle G \mathrm { - } F P R = T \mathrm { - } F P R = 1 0 ^ { - 4 } \rangle }$ ) in blue, respectively. + +![](images/7b1364cfc82980abf6870d20ac5cdf4f6c60172679d0b02e3a93fb132610f5c2.jpg) +Figure 4: Qualitative analysis on single-attribute adversarial attack $\mathrm { ( G { - } F P R = 1 0 ^ { - 3 } }$ ). More results are shown in Appendix (see Figure K, Figure L and Figure M). + +Analysis: semantic attack transferability. To further understand the property of SemanticAdv, we analyze the transferability of SemanticAdv on various settings. For each model with different FPRs, we select the successfully attacked adversarial examples from Section 4.1 to construct our evaluation dataset and evaluate these adversarial samples across different models. Table 1a illustrates the transferability of SemanticAdv among different models by using the same FPRs $\begin{array} { r } { ( \mathrm { G - F P R } = \mathrm { T } . } \end{array}$ FPR $= 1 0 ^ { - 3 }$ ). Table 1b illustrates the result with different FPRs $\mathbf { \bar { ( G - F P R = 1 0 ^ { - 4 } } }$ and $\mathrm { \Delta T - F P R } = 1 0 ^ { - 3 }$ ) for generation and evaluation. As shown in Table 1a, adversarial examples generated against models trained with softmax loss exhibit certain transferability compared to models trained with cosine loss. We conduct the same experiment by generating adversarial examples with CW and found it has weaker transferability compared to our SemanticAdv (results in brackets of Table 1). + +As Table 1b illustrates, the adversarial examples generated against the model with smaller G-FPR $= 1 0 ^ { - 4 }$ exhibit strong attack success rate when evaluating on the model with larger T-FPR $. = 1 0 ^ { - 3 }$ Especially, we found the adversarial examples generated against R-101-S have the best attack performance on other models. These findings motivate the analysis of query-free black-box API attack detailed in the following paragraph. + +
Mtest/MoptR-50-SR-101-S
R-50-S1.000 (1.000)0.862 (0.530)
R-101-S0.874 (0.422)1.000 (1.000)
R-50-C0.693(0.347)0.837 (0.579)
R-101-C0.617 (0.218)0.888 (0.617)
+ +Table 1: Transferability of SemanticAdv: cell $( i , j )$ shows attack success rate of adversarial examples generated against $j$ -th model and evaluate on $i$ -th model. Results of CW are listed in brackets. Left: Results generated with G-FPR $= 1 0 ^ { - 3 }$ and T-FPR $= 1 0 ^ { - 3 }$ ; Right: Results generated with G-FPR $=$ $1 0 ^ { - 4 }$ and $\mathrm { T } \mathrm { \mathrm { - F P R } } = 1 0 ^ { - 3 }$ . + +
MtestMoptR-50-SR-101-SR-50-CR-101-C
R-50-S1.000 (1.000)0.108 (0.007)0.023(0.002)0.018 (0.002)
R-101-S0.169 (0.006)1.000 (1.000)0.030 (0.002)0.032(0.003)
R-50-C0.166 (0.019)0.202(0.025)1.000 (1.000)0.048(0.007)
R-101-C0.120 (0.015)0.236 (0.029)0.040 (0.006)1.000(1.000)
+ +(a) + +Query-free black-box API attack. In this experiment, we generate adversarial examples against R101-S with $\mathrm { G } \mathrm { - } \mathrm { F P R } = 1 0 ^ { - 3 } ( \kappa = 1 . 2 4 )$ , $\mathbf { G } \mathbf { - } \mathbf { F P R } \mathbf { \bar { = } } 1 0 ^ { - 4 } ( \kappa = \mathbf { \bar { 0 } } . 6 0 )$ , and G-FPR $< 1 0 ^ { - 4 } ( \kappa = 0 . 2 0 )$ , respectively. We evaluate our algorithm on two industry level APIs, namely, ${ \mathrm { F a c e } } + +$ and AliYun face verification platform. Since attack transferability has never been explored in concurrent work that generates semantic adversarial examples, we use ${ \mathcal { L } } _ { p }$ bounded pixel-wise methods (Carlini & Wagner, 2017; Dong et al., 2018; Xie et al., 2019) as our baselines. As we see Table 2, which shows the best performance of each method, our SemanticAdv achieves much higher attack success rate than CW in both APIs with all FPR thresholds (e.g., our adversarial examples generated with G-FPR $< 1 0 ^ { - 4 }$ achieves $6 4 . 6 3 \%$ attack success rate on Face++ platform with $\mathrm { T } \mathrm { \bar { F } P R } = 1 0 ^ { - 3 }$ ). In addition, we found that lower G-FPR can achieve higher attack success rate on APIs within the same T-FPR (see Table I in Appendix). + +
API name Attacker/Evaluation MetricFace++AliYun
T-FPR=10 -3T-FPR=10 -4T-FPR=10 -3T-FPR=10 -4
Dong et al. (2018)30.7721.0318.006.50
Xie et al. (2019)37.9525.6421.5011.00
CW (G-FPR<10-4)41.6224.3719.0012.00
SemanticAdv (G-FPR <10-4)64.6342.6935.5022.17
+ +Table 2: Quantitative analysis on query-free black-box attack. We use ResNet-101 optimized with softmax loss for evaluation and report the attack success rate $( \% )$ on two online face verification platforms. Note that for PGD-based attacks, we adopt MI-FGSM $\epsilon = 8$ ) in Dong et al. (2018) and $\mathrm { \dot { \mathbf { M } } \mathbf { - } D I ^ { 2 } }$ -FGSM $\epsilon = 8$ ) in Xie et al. (2019), respectively. + +User study. To measure the perceptual quality of the adversarial images generated by SemanticAdv, we conduct a user study on Amazon Mechanical Turk (AMT). We use the adversarial examples generated with $\mathrm { G } \mathrm { - } \mathrm { F P R } < 1 0 ^ { - 4 }$ , which is the most strict setting in our experiment, to conduct the user study for both CW and SemanticAdv. In total, we collect 2, 620 annotations from 77 participants. In $3 9 . 1 \dot { 4 } \pm 1 . 9 6 \%$ (close to random guess $5 0 \%$ ) of trials the adversarial images generated by SemanticAdv are selected as reasonably-looking images and in $3 0 . 2 7 \pm 1 . 9 6 \%$ of trials, the adversarial images generated by CW are selected as reasonably-looking. It indicates that SemanticAdv can generate more reasonable-looking adversarial examples compared with CW under the most strict setting with G-FPR $< 1 0 ^ { - 4 }$ . Qualitative comparisons are shown in Appendix (see Figure H). + +![](images/e038d18d452da8c19a0e36bbebaafe417ea1f9ce207ae878fb70f59f617d2e80.jpg) +Figure 5: Quantitative analysis on attacking several defense methods including JPEG (Dziugaite et al., 2016), Blurring (Li & Li, 2017), and Feature Squeezing (Xu et al., 2017). + +SemanticAdv against defense methods. We evaluate the strength of the proposed attack by testing against four existing defense methods, namely, Feature squeezing (Xu et al., 2017), Blurring (Li & Li, 2017), JPEG (Dziugaite et al., 2016) and AMI (Tao et al., 2018). For AMI (Tao et al., 2018), we first extract attribute witnesses with our aligned face images and then leverage them to construct attribute-steered model. We use fc7 of pretrained VGG (Parkhi et al., 2015) as the face representation. AMI yields a consistency score for each face image to indicate whether it is a benign image. The score is measured by the cosine similarity between the representations from original model and attribute-steered model. With $1 0 \%$ false positives on benign inputs, it only achieves $8 \%$ detection accuracy for SemanticAdv and $1 2 \%$ detection accuracy for CW. + +![](images/2514972a90994ba8bb3e4bf4a64b6780f5d758436abdb8a62c957825f90e25ed.jpg) +Figure 6: Qualitative results on attacking face landmark detection model. + +![](images/136eee23c85c9f369b82e0445973320cf8e16d085aa0e61850f470aa410f04ff.jpg) +Figure 7: Qualitative results on attacking street-view semantic segmentation model. + +Figure 5 illustrates SemanticAdv is more robust against these defense methods comparing with CW. The same G-FPR and T-FPR are used for evaluation. Under the condition that T-FPR is $\bar { 1 } 0 ^ { - 3 }$ , both SemanticAdv and CW achieve high attack success rate, while SemanticAdv marginally outperforms CW when FPR goes down to $1 0 ^ { - 4 }$ . While defense methods have proven to be effective against CW attacks on classifiers trained with ImageNet (Krizhevsky et al., 2012), our results indicate that these methods are still vulnerable in face verification system with small T-FPR. + +# 4.3 SemanticAdv ON FACE LANDMARK DETECTION + +We also evaluate the effectiveness of SemanticAdv on face landmark detection. We select two attack tasks, namely, “Rotating Eyes” and “Out of Region”. For the “Rotating Eyes” task, we rotate the coordinates of the eyes in the image counter-clockwise by $9 0 ^ { \circ }$ . For the “Out of Region” task, we set a target bounding box and attempt to push all points out of the box. We summarize the experimental setup and quantitative results in the Appendix (see Table D). As we see in Figure 6, our method is applicable to attack landmark detection models. + +# 4.4 SemanticAdv ON STREET-VIEW SEMANTIC SEGMENTATION + +We further demonstrate the applicability of our SemanticAdv beyond face domain by generating adversarial perturbations on street-view images. Figure 7 illustrates the adversarial examples on semantic segmentation. In the first example, we select the leftmost the pedestrian as the target object instance and insert another car into the scene to attack it. The segmentation model has been successfully attacked to neglect the pedestrian (see last column), while it does exist in the scene (see second-to-last column). In the second example, we insert an adversarial car in the scene by SemanticAdv and the cyclist has been recognized as a pedestrian by the segmentation model. + +# 5 CONCLUSIONS + +Overall we presented a novel attack method SemanticAdv, which is capable of generating unrestricted adversarial perturbations guided by semantic attributes edition. Compared to existing methods, SemanticAdv works in a more controllable fashion. Experimental evaluations on face verification and landmark detection demonstrate several unique properties including attack transferability. We believe this work would open up new research opportunities and challenges in the field of adversarial learning. For instance, how to leverage semantic information to defend against such attacks will lead to potential new discussion. + +# REFERENCES + +Alibaba Cloud Computing Co. Ltd. https://help.aliyun.com/knowledge_detail/ 53535.html. +Megvii Technology Co. Ltd. https://console.faceplusplus.com/documents/ 5679308. +Yoshua Bengio, Grégoire Mesnil, Yann Dauphin, and Salah Rifai. Better mixing via deep representations. In ICML, 2013. +Anand Bhattad, Min Jin Chong, Kaizhao Liang, Bo Li, and David A Forsyth. Big but imperceptible adversarial perturbations via semantic manipulation. arXiv preprint arXiv:1904.06347, 2019. +Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. In ICLR, 2019. +Tom B Brown, Nicholas Carlini, Chiyuan Zhang, Catherine Olsson, Paul Christiano, and Ian Goodfellow. Unrestricted adversarial examples. arXiv preprint arXiv:1809.08352, 2018. +Adrian Bulat and Georgios Tzimiropoulos. Binarized convolutional landmark localizers for human pose estimation and face alignment with limited resources. In Proceedings of the IEEE International Conference on Computer Vision, pp. 3706–3714, 2017a. +Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem?(and a dataset of 230,000 3d facial landmarks). In ICCV, 2017b. +Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy (S&P). IEEE, 2017. +Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In CVPR, 2018. +Moustapha Cisse, Yossi Adi, Natalia Neverova, and Joseph Keshet. Houdini: Fooling deep structured prediction models. In NIPS, 2017. +Ronan Collobert and Jason Weston. A unified architecture for natural language processing: Deep neural networks with multitask learning. In Proceedings of the 25th international conference on Machine learning, pp. 160–167. ACM, 2008. +Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 3213–3223, 2016. +Ltsc Deng, Jinyu Li, Jui-Ting Huang, Kaisheng Yao, Dong Yu, Frank Seide, Michael L Seltzer, Geoffrey Zweig, Xiaodong He, Jason D Williams, et al. Recent advances in deep learning for speech research at microsoft. In ICASSP, volume 26, pp. 64, 2013. +Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 9185–9193, 2018. + +Gintare Karolina Dziugaite, Zoubin Ghahramani, and Daniel M Roy. A study of the effect of jpg compression on adversarial images. arXiv preprint arXiv:1608.00853, 2016. + +Logan Engstrom, Brandon Tran, Dimitris Tsipras, Ludwig Schmidt, and Aleksander Madry. A rotation and a translation suffice: Fooling cnns with simple transformations. arXiv preprint arXiv:1712.02779, 2017. + +Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014a. + +Ian Goodfellow, Yoshua Bengio, Aaron Courville, and Yoshua Bengio. Deep learning, volume 1. MIT Press, 2016. + +Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2014b. + +Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In ECCV. Springer, 2016. + +Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. + +Seunghoon Hong, Xinchen Yan, Thomas S Huang, and Honglak Lee. Learning hierarchical semantic image manipulation through structured representations. In NeurIPS, 2018. + +Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database forstudying face recognition in unconstrained environments. In Workshop on faces in’Real-Life’Images: detection, alignment, and recognition, 2008. + +Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, pp. 1125–1134, 2017. + +Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV. Springer, 2016. + +Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In CVPR, pp. 1219–1228, 2018. + +Ameya Joshi, Amitangshu Mukherjee, Soumik Sarkar, and Chinmay Hegde. Semantic adversarial attacks: Parametric transformations that fool deep classifiers. arXiv preprint arXiv:1904.08489, 2019. + +Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In ICLR, 2018. + +Ira Kemelmacher-Shlizerman, Steven M Seitz, Daniel Miller, and Evan Brossard. The megaface benchmark: 1 million faces for recognition at scale. In CVPR, pp. 4873–4882, 2016. + +Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. + +Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. + +Brendan F Klare, Ben Klein, Emma Taborsky, Austin Blanton, Jordan Cheney, Kristen Allen, Patrick Grother, Alan Mah, and Anil K Jain. Pushing the frontiers of unconstrained face detection and recognition: Iarpa janus benchmark a. In CVPR, 2015. + +Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NIPS, 2012. + +Xin Li and Fuxin Li. Adversarial examples detection in deep networks with convolutional filter statistics. In Proceedings of the IEEE International Conference on Computer Vision, pp. 5764– 5772, 2017. + +Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In NIPS, 2017. + +Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In ICCV, 2015. + +Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them. In CVPR, 2015. + +Elman Mansimov, Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Generating images from captions with attention. In ICLR, 2015. + +Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2574–2582, 2016. + +Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In CVPR Workshops, pp. 51–59, 2017. + +Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In European Conference on Computer Vision, pp. 483–499. Springer, 2016. + +Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In ICML. JMLR, 2017. + +Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In ICML, 2016. + +Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In Security and Privacy (EuroS&P), 2016 IEEE European Symposium on, 2016. + +Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, et al. Deep face recognition. In bmvc, volume 1, pp. 6, 2015. + +Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. In ICLR, 2015. + +Scott Reed, Kihyuk Sohn, Yuting Zhang, and Honglak Lee. Learning to disentangle factors of variation with manifold interaction. In ICML, 2014. + +Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. + +Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-thewild challenge: The first facial landmark localization challenge. In ICCV Workshop, 2013. + +Soumyadip Sengupta, Jun-Cheng Chen, Carlos Castillo, Vishal M Patel, Rama Chellappa, and David W Jacobs. Frontal to profile face verification in the wild. In WACV, pp. 1–9. IEEE, 2016. + +David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. nature, 529(7587):484, 2016. + +Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. + +Yang Song, Rui Shu, Nate Kushman, and Stefano Ermon. Constructing unrestricted adversarial examples with generative models. In Advances in Neural Information Processing Systems, pp. 8312–8323, 2018. + +Yi Sun, Xiaogang Wang, and Xiaoou Tang. Deep learning face representation from predicting 10,000 classes. In CVPR, 2014. + +Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. + +Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, Andrew Rabinovich, et al. Going deeper with convolutions. In CVPR, 2015. + +Guanhong Tao, Shiqing Ma, Yingqi Liu, and Xiangyu Zhang. Attacks meet interpretability: Attributesteered detection of adversarial samples. In NeurIPS, 2018. + +Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelcnn decoders. In NIPS, 2016. + +Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In CVPR, 2018a. + +Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Highresolution image synthesis and semantic manipulation with conditional gans. In CVPR, 2018b. + +Chaowei Xiao, Ruizhi Deng, Bo Li, Fisher Yu, Mingyan Liu, and Dawn Song. Characterizing adversarial examples based on spatial consistency information for semantic segmentation. In ECCV, 2018a. + +Chaowei Xiao, Bo Li, Jun-Yan Zhu, Warren He, Mingyan Liu, and Dawn Song. Generating adversarial examples with adversarial networks. In IJCAI, 2018b. + +Chaowei Xiao, Jun-Yan Zhu, Bo Li, Warren He, Mingyan Liu, and Dawn Song. Spatially transformed adversarial examples. In ICLR, 2018c. + +Chaowei Xiao, Dawei Yang, Bo Li, Jia Deng, and Mingyan Liu. Meshadv: Adversarial meshes for visual recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6898–6907, 2019. + +Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2730–2739, 2019. + +Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. arXiv preprint arXiv:1704.01155, 2017. + +Xinchen Yan, Jimei Yang, Kihyuk Sohn, and Honglak Lee. Attribute2image: Conditional image generation from visual attributes. In ECCV. Springer, 2016. + +Fisher Yu, Vladlen Koltun, and Thomas Funkhouser. Dilated residual networks. In Computer Vision and Pattern Recognition (CVPR), 2017. + +Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In ICCV, 2017. + +Xingcheng Zhang, Lei Yang, Junjie Yan, and Dahua Lin. Accelerated training for massive classification via dynamic class selection. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018. + +Jun-Yan Zhu, Philipp Krähenbühl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the natural image manifold. In ECCV. Springer, 2016a. + +Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. + +Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z Li. Face alignment across large poses: A 3d solution. In CVPR, 2016b. + +# A FACE IDENTITY VERIFICATION + +Benchmark performance. First, we provide additional information about the ResNet models we used in the experiments. We summarize in Table A the performance on several face identity verification benchmarks including Labeled Face in the Wild (LFW) dataset (Huang et al., 2008), AgeDB-30 dataset (Moschoglou et al., 2017), and Celebrities in Frontal-Profile (CFP) dataset (Sengupta et al., 2016). + +
M benchmarksLFWAgeDB-30CFP-FFCFP-FP
R-50-S99.2794.1599.2691.49
R-101-S99.4295.9399.5795.07
R-50-C99.3895.0899.2490.24
R-101-C99.6795.5899.5792.71
+ +Table A: The performnace of ResNet models on several benchmark datasets. + +Identity verification thresholds. Table B shows the threshold values used in our experiments when determining whether two portrait images belong to the same identity or not. The selected FPR thresholds and normalized L2 distance between face features are commonly used when evaluating the performance of face recognition model Klare et al. (2015); Kemelmacher-Shlizerman et al. (2016). + +
FPR/MR-50-SR-101-SR-50-CR-101-C
10-33×10-410-41.1811.2441.4471.469
1.0580.6571.0481.2931.2420.809
0.5970.864
+ +Table B: The threshold values for face identity verification. + +Implementation details. We use Adam (Kingma & Ba, 2015) as the optimizer to generate adversarial examples for both our SemanticAdv and CW. More specifically, we run optimization for up to 200 steps with a fixed learning rate 0.05 for cases when $\mathbf { \bar { G } } \mathbf { - } \mathbf { F P R } \leq 1 0 ^ { - 4 }$ . Otherwise, we run optimization for up to 500 steps with a fixed learning rate 0.01. For pixel-wise attack method CW, we use additional pixel reconstruction loss with corresponding loss weight to 5. We run optimization for up to $1 , 0 0 0$ steps with a fixed learning rate $1 0 ^ { - 3 }$ . + +Evaluation metrics. To evaluate the performance of semanticAdv under different attributes, we consider three metrics as follows: + +• Best: if there is one attribute among 17 attributes that can be successfully attacked, we count the attack success rate for this face identity as 1; +• Average: we calculate the average attack success rate among 17 attributes for the same face identity; +• Worst: only when all of 17 attributes can be successfully attacked, we count the attack success rate for this person as 1; + +Note that, for a fair comparison with CW, we should use the Best metric for our SemanticAdv, as CW is the traditional pixel-wise attack method works regardless of the attribute. In addition, we report the performance using the average and worst metric, which actually provides additional insights into the robustness of face verification models across different attributes. For instance, combining the results from Table C and Figure 3, we understand that the face verification models used in our experiments have different levels of robustness across attributes. For example, face verification models are more robust against local shape variations than color variations, e.g., pale skin has higher attack success rate than mouth open. We believe these discoveries will help the community further understand the properties of face verification models. + +Table C shows the overall performance (accuracy) of face verification model and attack success rate of SemanticAdv and CW. As we can see from Table C, although the face model trained with cosine loss achieves higher face recognition performance, it is more vulnerable to adversarial attack compared with the model trained with softmax loss. + +Table C: Quantitative result of identity verification $( \% )$ . It shows accuracy of face verification model and attack success rate of SemanticAdv and CW. $\mathbf { x } ^ { \prime }$ , $G ( \mathbf { x } ^ { \prime } , \mathbf { c } )$ and $G ( \mathbf { x } ^ { \prime } , \mathbf { c } ^ { \mathrm { n e w } } )$ are the intermediate results of our method before adversarial perturbation. + +
G-FPRMetricsMR-50-SR-101-SR-50-CR-101-C
10-3Verification Accuracy98.3698.7898.6398.84
xG(x',c)G(x',cnew)W)(Best)0.000.000.160.000.000.080.080.000.160.000.230.31
SemanticAdv (Best)SemanticAdv (Worst)SemanticAdv (Average)100.0091.9598.98100.0093.9899.29100.0099.5399.97100.0099.7799.99
CW100.00100.00100.00100.00
3 ×10-4Verification Accuracy97.7397.9797.9197.85
xG(x',c)G(x',cnew)(Best)0.000.000.000.000.000.000.000.000.000.000.000.00
SemanticAdv (Best)SemanticAdv (Worst)SemanticAdv (Average)100.0083.7597.72100.0079.0697.35100.0098.9899.92100.0096.6499.72
CW100.00100.00100.00100.00
10-4Verification Accuracy93.2592.8093.4392.98
XG(x',c)G(x',cnew)(Best)0.000.000.000.000.000.000.000.000.000.000.000.00
SemanticAdv (Best)SemanticAdv (Worst)SemanticAdv (Average)100.0033.5983.53100.0019.8476.64100.0067.0395.57100.0048.6791.13
CW100.00100.00100.00100.00
+ +Defense methods. Feature squeezing ( $\mathrm { { X u } }$ et al., 2017) is a simple but effective method by reducing color bit depth to remove the adversarial effects. We compress the image represented by 8 bits for each channel to 4 bits for each channel to evaluate the effectiveness. For Blurring (Li & Li, 2017), we use a $3 \times 3$ Gaussian kernel with standard deviation 1 to smooth the adversarial perturbations. For JPEG Dziugaite et al. (2016), it leverages the compression and decompression to remove the adversarial perturbation. We set the compression ratio as 0.75 in our experiment. + +# B FACE LANDMARK DETECTION + +Implementation details. To perform attack on the face landmark detection model, we run optimization for up to 2, 000 steps with a fixed learning rate 0.05. We set the balancing factor $\lambda$ (see Eq. 3) to value 0.01 for this experiment. + +Evaluation Metrics. We apply two different metrics for two adversarial attack tasks respectively. For “Rotating Eyes” task, we use a well-adopted metric Normalized Mean Error (NME) (Bulat & Tzimiropoulos, 2017b) for experimental evaluation. + +$$ +r _ { \mathrm { N M E } } = \frac { 1 } { N } \sum _ { k = 1 } ^ { N } \frac { | | \mathbf { p } _ { k } - \hat { \mathbf { p } } _ { k } | | _ { 2 } } { \sqrt { W _ { B } * H _ { B } } } , +$$ + +where√ $\mathbf { p } _ { k }$ denotes the $k$ -th ground-truth landmark, $\hat { \mathbf { p } } _ { \mathbf { k } }$ denotes the $k$ -th predicted landmark and $\sqrt { W _ { B } * H _ { B } }$ is the square-root area of ground-truth bounding box, where $W _ { B }$ and $H _ { B }$ represents the width and height of the box. + +For “Out of Region” task, we consider the attack is successful if the landmark predictions fall outside a pre-defined centering region on the portrait image. Thus, we introduce a metric that reflects the portion of landmarks outside of the pre-defined centering region: rOUT = NoutNtotal , where Nout denotes the number of predicted landmarks outside the pre-defined bounding box and $N _ { \mathrm { t o t a l } }$ denotes the total number of landmarks. + +
Tasks (Metrics)PristineAugmented Attributes
BlondHairYoungEyeglassesRosyCheeksSmilingArchedEyebrowsBangsPaleSkin
Rotating eyes (rNME)↓28.0414.0317.288.5813.2419.2123.4215.9910.72
Out-of-region (rouT)↓45.9817.4223.047.5116.6525.4433.8520.0313.51
+ +Table D: Quantitative results on face landmark detection $( \% )$ ) The two row shows the measured ratios (lower is better) for “Rotating Eyes” and “Out Of Region” task, respectively. + +We present the quantitative results of SemanticAdv on face landmark detection model in Table D including two adversarial tasks, namely, “Out of Region” and “Rotating Eyes”. We observe that our method is efficient to perform attacking on landmark detection models. For certain attributes such as “Eyeglasses”, “Plae Skin”, SemanticAdv is able to achieve reasonably-good performance. + +# C USER STUDY + +We conducted a user study on the adversarial images of SemanticAdv and CW used in the experiment of API-attack and the original images. The adversarial images are generated with $\mathrm { { G } \mathrm { { - F P R } \mathrm { { < } \ 1 0 ^ { - 4 } } } }$ for both methods. We present a pair of original image and adversarial image to participants and ask them to rank the two options. The order of these two images is randomized and the images are displayed for 2 seconds in the screen during each trial. After the images disappear, the participants have unlimited time to select the more reasonably-looking image according to their perception. For each participant, one could only conduct at most 50 trials, and each adversarial image was shown to 5 different participants. Some qualitative results are shown in Figure H. In total, we collect 2, 620 annotations from 77 participants. In $3 9 . 1 4 \pm 1 . 9 6 \%$ of trials the adversarial images generated by SemanticAdv are selected as reasonably-looking images and in $3 0 . 2 7 \pm 1 . 9 6 \%$ of trails, the adversarial images generated by CW are selected as reasonably-looking images. It indicates that our semantic adversarial examples are more perceptual reasonably-looking than CW. Additionally, we also conduct the user study with larger $\mathrm { { G } { - } \mathrm { { F P R } } \mathrm { { = } \ i 0 ^ { - 3 } } }$ . In $4 5 . 4 2 \pm 1 . 9 6 \%$ of trials, the adversarial images generated by SemanticAdv are selected as reasonably-looking images, which is very close to the random guess $( \dot { 5 } 0 \% )$ . + +![](images/e11a9c182edab33d61fc6b803c35b0b754579590f5fdb181259f75a0c14af2ed.jpg) +Figure H: Qualitative comparisons among ground truth, pixel-wise adversarial examples generated by CW, and our proposed SemanticAdv. Here, we present the results from G-FPR $\dot { < } 1 0 ^ { { - } 4 }$ so that perturbations are visible. + +# D ABLATION STUDY: FEATURE-SPACE INTERPOLATION + +We conduct an ablation study on feature-space interpolation by analyzing attack success rates with different feature-maps in the StarGAN network. Table E shows the attack success rate on R-101-S. Here, we use $\mathbf { f } _ { i }$ to represent the feature-map after $i$ -th up-sampling operation. $\mathbf { f } _ { 0 }$ denotes the featuremap before applying up-sampling operation. The result demonstrates that samples generated by interpolating on $\mathbf { f } _ { 0 }$ achieve the highest success rate. Since $\mathbf { f } _ { 0 }$ is the feature-map before decoder, it still well embeds semantic information in the feature space. We adopt $\mathbf { f } _ { 0 }$ for interpolation in our experiments. + +We also conduct a qualitative comparison between attribute-space and feature-space interpolation. As shown in Figure I, images generated by attribute-space and feature-space interpolation are both reasonably-looking. + +
T-FPR (G-FPR)10-3 (10-3)3× 10-4 (3×10-4)10-4 (10-4)
Layer (f)ff1f2ff1f2ff1f
Attack SuccessRate99.2998.3275.6297.3594.1057.1576.6467.4019.63
+ +Table E: Attack success rate by selecting different layer’s feature-map for interpolation on $\mathrm { R - } 1 0 1 { - } \mathrm { S } ( \% )$ $\mathbf { f } _ { i }$ indicates the feature-map after $i$ -th up-sampling operation. + +
T-FPR (G-FPR)10-3 (10-3)3×10)-4(3×10-4)10-4(10-4)
Layer (f)f-2f-1f-2f-1f-2f-1
Attack Success Rate49.4092.0930.4481.876.6645.46
+ +Table F: Attack success rate by selecting different layer’s feature-map for interpolation on R-101- $S ( \% )$ . $\mathbf { f } _ { - 2 }$ indicates the feature-map after the last down-sampling operation and $\mathbf { f } _ { - 2 }$ indicates the feature-map after $\mathbf { f } _ { - 2 }$ . + +![](images/d2f867cde088d9d2eedec96d6e11f67a1803de628ed29a100713bd1b68d74171.jpg) +Figure I: Qualitative comparisons between attribute-space and feature-space interpolation. + +# E SEMANTIC ATTACK TRANSFERABILITY + +In Table G, we present the quantitative results of the transferability with $\mathrm { { G } \mathrm { { - } \mathrm { { F P R } = 1 0 ^ { - 4 } } } }$ , T-FPR $= 1 0 ^ { - 4 }$ . We observe that with the strict criterion (Lower T-FPR) of the verification model, the transferability becomes lower cross different models. + +To explore whether the improvement of transferability in SemanticAdv is introduced by semantic editing rather than optimized semantic perturbation in feature space. We add a StrawMan $\mathrm { [ C W + }$ $\mathbf { x } ^ { \mathrm { n e w } } ]$ baseline which uses a controllable semantic-attribute-based generator to generate semantically different images without any notion of an adversarial attack, and then applies standard $L _ { p }$ CW attacks on that generated image. The results are shown in Table H. The performance of the StrawMan $\mathrm { [ C W + }$ $\mathbf { x } ^ { \mathrm { n e w } } ]$ is worse than SemanticAdv. This result justifies that our SemanticAdv is able to produce novel adversarial examples which cannot be simply achieved by combining attribute-conditional image editing model with $L _ { p }$ bounded perturbation. + +
Mtest/MoptR-50-SR-101-SR-50-CR-101-C
R-50-S1.0000.0050.0000.000
R-101-S0.0001.0000.0000.000
R-50-C0.0000.0001.0000.000
R-101-C0.0000.0000.0001.000
+ +Table G: Transferability of SemanticAdv: cell $( i , j )$ shows attack success rate of adversarial examples generated against $j$ -th model and evaluate on $i$ -th model. Results generated with $\mathrm { { G } { - } \mathrm { { F P R } } = 1 \bar { 0 } ^ { - 4 } }$ , T-FPR $= 1 0 ^ { \overline { { - 4 } } }$ . + +
MtestMoptR-101-S
R-50-S0.035 (0.108)
R-101-S1.000 (1.000)
R-50-C0.145 (0.202)
R-101-C0.085 (0.236)
+ +(a) $\mathrm { G } \mathrm { - } \mathrm { F P R } { = } 1 0 ^ { - 3 }$ , $\mathrm { T } \mathrm { - } \mathrm { F P R } { = } 1 0 ^ { - 3 }$ + +
MtestMoptR-101-S
R-50-S0.615 (0.862)
R-101-S1.000(1.000)
R-50-C0.570(0.837)
R-101-C0.6955(0.888)
+ +(b) G-FPR=10−4, $\mathrm { T } { \cdot } \mathrm { F P R } { = } 1 0 ^ { - 3 }$ + +Table H: Transferability of StrawMan: cell $( i , j )$ shows attack success rate of adversarial examples generated against $j$ -th model and evaluate on $i$ -th model. Results of SemanticAdv are listed in brackets. + +# F QUERY-FREE BLACK-BOX API ATTACK + +
API nameAttacker/Evaluation MetricFace++AliYun
T-FPR= 10-3T-FPR=10-4T-FPR=10-3T-FPR=10
Original x2.040.510.500.00
Generated xnew4.210.530.500.00
CW (G-FPR=10-)StrawMan [CW + xnew ](G-FPR = 10-3)SemanticAdv (G-FPR =10-3)16.2410.7127.323.554.089.794.500.00
3.000.00
7.502.00
CW (G-FPR= 10-4)StrawMan [CW+ xnew ](G-FPR = 10-4)SemanticAdv (G-FPR =10-4)30.6121.3257.2215.829.1438.6612.504.50
7.5029.501.5017.50
Dong et al. (2018)Xie et al. (2019)CW (G-FPR<10-4)StrawMan [CW + xnew](G-FPR<10-4)SemanticAdv (G-FPR<10-4)30.7737.9521.0318.006.50
25.6421.5011.00
41.6227.6964.6324.3719.0012.00
15.3842.6910.003.00
35.5022.17
+ +Table I: Quantitative analysis on query-free black-box attack. We use ResNet-101 optimized with softmax loss for evaluation and report the attack success rate $( \% )$ . + +In Table I, we present the results of SemanticAdv performing query-free black-box attack on two online face verification platforms. SemanticAdv outperforms CW and StawnMan in both APIs under all FPR thresholds. In addition, we achieve higher attack success rate on APIs using samples generated using lower G-FPR compared to samples generated using higher G-FPR with the same T-FPR. Original $\mathbf { x }$ and generated $\mathbf { x } ^ { \mathrm { n e w } }$ are regarded as reference point of the performance of online face verification platforms. + +Figure J illustrates our SemanticAdv on attacking Microsoft Azure face verification system, which further demonstrate the effectiveness of our approach. + +![](images/e030ab5d5338f1ba2bcc57782a19fec5f232a173459480ed89b123b6102bfa5c.jpg) +Figure J: Illustration of our SemanticAdv in the real world face verification platform (editing on pale skin). Note that the confidence denotes the likelihood that two faces belong to the same person. + +# G MORE VISUALIZATIONS + +![](images/3a7d8c713d580a52035bd642fd776679a6153075581bff76f96b228a53b76fa7.jpg) +Figure K: Qualitative analysis on single-attribute adversarial attack $\mathrm { ( G { - } F P R = 1 0 ^ { - 3 } }$ ). + +![](images/de33f8bb372e2f4abdf4f6769428fdaa8a0623516671ba763ae2fab7c70afb7c.jpg) +Figure L: Qualitative analysis on single-attribute adversarial attack $\mathrm { ( G - F P R = 1 0 ^ { - 3 } }$ ). + +![](images/7c8afb5499aa1945df2aa8c49f2ce5827b76cbafd081d867c0cbe11df573a7e7.jpg) +Figure M: Qualitative analysis on single-attribute adversarial attack $\mathrm { \bf G } \mathrm { - } \mathrm { \bf F P R } = 1 0 ^ { - 3 } ,$ ). + +![](images/bc2242ce0987cedaa8c20dea5c4a3437513f800c661325f11f4134300127ee38.jpg) +Figure N: Qualitative comparisons between our proposed SemanticAdv $\mathrm { ( G { - } F P R = 1 0 ^ { - 3 } }$ ) and pixelwise adversarial examples generated by CW. Along with the adversarial examples, we also provide the corresponding perturbations (residual) on the right. + +![](images/0b62987d5a2bc9e11b6691551e91f0da0480e64b18a63bf77468e026095b805e.jpg) +Figure O: Qualitative analysis on single-attribute adversarial attack (SemanticAdv with $\mathrm { { \bf G } { - } F P R = }$ $1 0 ^ { - 3 }$ ) by each other. Along with the adversarial examples, we also provide the corresponding perturbations (residual) on the right. \ No newline at end of file diff --git a/parse/train/r1l-VeSKwS/r1l-VeSKwS_content_list.json b/parse/train/r1l-VeSKwS/r1l-VeSKwS_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..861a2e20314e733cc0f6e837979256fd8ef5d2f5 --- /dev/null +++ b/parse/train/r1l-VeSKwS/r1l-VeSKwS_content_list.json @@ -0,0 +1,2264 @@ +[ + { + "type": "text", + "text": "SEMANTICADV: GENERATING ADVERSARIAL EXAMPLES VIA ATTRIBUTE-CONDITIONAL IMAGE EDITING ", + "text_level": 1, + "bbox": [ + 176, + 99, + 826, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 170, + 398, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 217, + 544, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) have achieved great success in various applications due to their strong expressive power. However, recent studies have shown that DNNs are vulnerable to adversarial examples which are manipulated instances targeting to mislead DNNs to make incorrect predictions. Currently, most such adversarial examples try to guarantee “subtle perturbation\" by limiting the $L _ { p }$ norm of the perturbation. In this paper, we aim to explore the impact of semantic manipulation on DNNs predictions by manipulating the semantic attributes of images and generate “unrestricted adversarial examples\". In particular, we propose an algorithm SemanticAdv which leverages disentangled semantic factors to generate adversarial perturbation by altering controlled semantic attributes to fool the learner towards various “adversarial\" targets. We conduct extensive experiments to show that the semantic based adversarial examples can not only fool different learning tasks such as face verification and landmark detection, but also achieve high targeted attack success rate against real-world black-box services such as Azure face verification service based on transferability. To further demonstrate the applicability of SemanticAdv beyond face recognition domain, we also generate semantic perturbations on street-view images. Such adversarial examples with controlled semantic manipulation can shed light on further understanding about vulnerabilities of DNNs as well as potential defensive approaches. ", + "bbox": [ + 233, + 248, + 766, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 527, + 336, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) have demonstrated great successes in advancing the state-of-the-art performance of discriminative tasks (Krizhevsky et al., 2012; Goodfellow et al., 2016; He et al., 2016; Collobert & Weston, 2008; Deng et al., 2013; Silver et al., 2016). However, recent research found that DNNs are vulnerable to adversarial examples which are carefully crafted instances aiming to induce arbitrary prediction errors for learning systems. Such adversarial examples containing small magnitude of perturbation have shed light on understanding and discovering potential vulnerabilities of DNNs (Szegedy et al., 2013; Goodfellow et al., 2014b; Moosavi-Dezfooli et al., 2016; Papernot et al., 2016; Carlini & Wagner, 2017; Xiao et al., 2018b;c;a; 2019). Most existing work focused on constructing adversarial examples by adding ${ \\mathcal { L } } _ { p }$ bounded pixel-wise perturbations (Goodfellow et al., 2014b) or spatially transforming the image (Xiao et al., 2018c; Engstrom et al., 2017) (e.g., in-plane rotation or out-of-plane rotation). Generating unrestricted perturbations with semantically meaningful patterns is an important yet under-explored field. ", + "bbox": [ + 174, + 547, + 825, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "At the same time, deep generative models have demonstrated impressive performance in learning disentangled semantic factors through data generation in an unsupervised (Radford et al., 2015; Karras et al., 2018; Brock et al., 2019) or weakly-supervised manner based on semantic attributes (Yan et al., 2016; Choi et al., 2018). Empirical findings in (Yan et al., 2016; Zhu et al., 2016a; Radford et al., 2015) demonstrated that a simple linear interpolation on the learned image manifold can produce smooth visual transitions between a pair of input images. ", + "bbox": [ + 174, + 722, + 825, + 805 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel attack SemanticAdv which generates unrestricted perturbations with semantically meaningful patterns. Motivated by the findings mentioned above, we leverage an attribute-conditional image editing model (Choi et al., 2018) to synthesize adversarial examples by interpolating between source and target images in the feature-map space. Here, we focus on changing a single attribute dimension to achieve adversarial goals while keeping the generated adversarial image reasonably-looking (e.g., see Figure 1). To validate the effectiveness of the proposed attack method, we consider two tasks, namely, face verification and landmark detection, as face recognition field has been extensively explored and the commercially used face models are relatively robust since they require a low false positive rate. We conduct both qualitative and quantitative evaluations on CelebA dataset (Liu et al., 2015). To demonstrate the applicability of SemanticAdv beyond face domain, we further extend SemanticAdv to generate adversarial street-view images. We treat semantic layouts as input attributes and use the image editing model (Hong et al., 2018) pre-trained on Cityscape dataset (Cordts et al., 2016). Please find more visualization results on the anonymous website: https://sites.google.com/view/generate-semantic-adv-example. ", + "bbox": [ + 174, + 813, + 823, + 924 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/464ae493fc35cd3e8384b15d0c1e64cf82fc55148d4ff781b38e7a7264ac13e9.jpg", + "image_caption": [ + "Figure 1: Left: Overview of the proposed SemanticAdv. Right: Illustration of our SemanticAdv in the real world face verification platform. Note that the confidence denotes the likelihood that two faces belong to the same person. " + ], + "image_footnote": [], + "bbox": [ + 173, + 99, + 825, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 353, + 825, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of the proposed SemanticAdv are three-folds. First, we propose a novel semanticbased attack method to generate unrestricted adversarial examples by feature-space interpolation. Second, the proposed method is able to generate semantically-controllable perturbations due to the attribute-conditioned modeling. This allows us to analyze the robustness of a recognition system against different types of semantic attacks. Third, as a side benefit, the proposed attack exhibits high transferability and leads to a $65 \\%$ query-free black-box attack success rate on a real-world face verification platform, which outperforms the pixel-wise perturbations in attacking existing defense methods. ", + "bbox": [ + 174, + 444, + 825, + 555 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 558, + 346, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic image editing. Semantic image synthesis and manipulation is a popular research topic in machine learning, graphics and vision. Thanks to recent advances in deep generative models (Kingma & Welling, 2014; Goodfellow et al., 2014a; Oord et al., 2016) and the empirical analysis of deep classification networks (Krizhevsky et al., 2012; Simonyan & Zisserman, 2014; Szegedy et al., 2015), past few years have witnessed tremendous breakthroughs towards high-fidelity pure image generation (Radford et al., 2015; Karras et al., 2018; Brock et al., 2019), attribute-to-image generation (Yan et al., 2016; Choi et al., 2018), text-to-image generation (Mansimov et al., 2015; Reed et al., 2016; Van den Oord et al., 2016; Odena et al., 2017; Zhang et al., 2017; Johnson et al., 2018), and imageto-image translation (Isola et al., 2017; Zhu et al., 2017; Liu et al., 2017; Wang et al., 2018b; Hong et al., 2018). ", + "bbox": [ + 173, + 577, + 826, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial examples. Generating $L _ { p }$ bounded adversarial perturbation has been extensively studied recently (Szegedy et al., 2013; Goodfellow et al., 2014b; Moosavi-Dezfooli et al., 2016; Papernot et al., 2016; Carlini & Wagner, 2017; Xiao et al., 2018b). To further explore diverse adversarial attacks and potentially help inspire defense mechanisms, it is important to generate the socalled “unrestricted\" adversarial examples which contain unrestricted magnitude of perturbation while still preserve perceptual realism Brown et al. (2018). Recently, Xiao et al. (2018c); Engstrom et al. (2017) propose to spatially transform the image patches instead of adding pixel-wise perturbation, while such spatial transformation does not consider semantic information. Our proposed semanticAdv focuses on generating unrestricted perturbation with semantically meaningful patterns guided by visual attributes. ", + "bbox": [ + 174, + 722, + 825, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Relevant to our work, Song et al. (2018) proposed to synthesize adversarial examples with an unconditional generative model. Bhattad et al. (2019) studied semantic transformation in only the color or texture space. Compared to these works, semanticAdv is able to generate adversarial examples in a controllable fashion using specific visual attributes by performing manipulation in the feature space. We further analyze the robustness of the recognition system by generating adversarial examples guided by different visual attributes. Concurrent to our work, Joshi et al. (2019) proposed to generate semantic-based attacks against a restricted binary classifier, while our attack is able to mislead the model towards arbitrary adversarial targets. They conduct the manipulation within the attribution space which is less flexible and effective than our proposed feature-space interpolation. ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 174 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 SEMANTIC ADVERSARIAL EXAMPLES ", + "text_level": 1, + "bbox": [ + 176, + 181, + 522, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 PROBLEM DEFINITION ", + "text_level": 1, + "bbox": [ + 176, + 205, + 370, + 219 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal { M }$ be a machine learning model trained on a dataset $\\mathcal { D } = \\{ ( { \\bf x } , { \\bf y } ) \\}$ consisting of image-label pairs, where $\\mathbf { x } \\in \\mathbb { R } ^ { H \\times W \\times D _ { I } }$ and $\\mathbf { y } \\in \\mathbb { R } ^ { D _ { L } }$ denote the image and the ground-truth label, respectively. Here, $H , W$ , $D _ { I }$ , and $D _ { L }$ denote the image height, image width, number of image channels, and label dimensions, respectively. For each image $\\mathbf { x }$ , our model $\\mathcal { M }$ makes a prediction $\\hat { \\mathbf { y } } = \\mathcal { M } ( \\mathbf { x } ) \\in \\mathbb { R } ^ { D _ { L } }$ . Given a target image-label pair $\\left( \\mathbf { x } ^ { \\mathrm { { t g t } } } , \\mathbf { y } ^ { \\mathrm { { t g t } } } \\right)$ and $\\mathbf { y } \\neq \\mathbf { y } ^ { \\mathrm { t g t } }$ , a traditional attacker aims to synthesize adversarial examples $\\{ { \\bf x } ^ { \\mathrm { a d v } } \\}$ by adding pixel-wise perturbations to or spatially transforming the original image x such that $\\dot { \\mathcal { M } } ( \\mathbf { x } ^ { \\mathrm { { a d v } } } ) = \\mathbf { y } ^ { \\mathrm { { t g t } } }$ . ", + "bbox": [ + 173, + 227, + 825, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we introduce the concept of semantic attacker that aims at generating adversarial examples by adding semantically meaningful perturbation with a conditional generative model $\\mathcal { G }$ . Compared to traditional attacker that usually produces pixel-wise perturbations, the proposed method is able to produce semantically meaningful perturbations. ", + "bbox": [ + 174, + 332, + 825, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semantic image editing. For simplicity, we start with the formulation where the input attribute is represented as a compact vector. This formulation can be directly extended to other input attribute formats including semantic layouts. Let $\\mathbf { c } \\in \\mathbb { R } ^ { D _ { c } }$ be an attribute representation reflecting the semantic factors (e.g., expression or hair color of a portrait image) of image $\\mathbf { x }$ , where $D _ { C }$ indicates the attribute dimension and $c _ { i } \\in \\{ 0 , 1 \\}$ indicates the appearance of $i$ -th attribute. Here, our goal is to use the conditional generator for semantic image editing. For example, given a portrait image of a girl with black hair and blonde hair as the new attribute, our generator is supposed to synthesize a new image that turns the girl’s hair from black to blonde. More specifically, we denote the new attribute as $\\mathbf { c } ^ { \\mathrm { { n e w } } } \\in \\mathbb { R } ^ { D _ { C } }$ such that the synthesized image is given by $\\mathbf { x } ^ { \\mathrm { n e w } } = \\mathcal { G } ( \\bar { \\mathbf { x } , \\mathbf { c } ^ { \\mathrm { n e w } } } )$ . In the special case when there is no attribute change $( \\mathbf { c } = \\mathbf { c } ^ { \\mathrm { n e w } }$ ), the generator simply reconstructs the input: $\\mathbf { x } = \\mathcal { G } ( \\mathbf { x } , \\mathbf { c } )$ . Supported by the findings mentioned in (Bengio et al., 2013; Reed et al., 2014), our synthesized image $\\mathbf { x } ^ { \\mathrm { n e w } }$ should fall close to the data manifold if we constrain the change of attribute values to be sufficiently small (e.g., we only update one semantic attribute at a time). In addition, we can potentially generate many such images by linearly interpolating between the semantic embeddings of the conditional generator $\\mathcal { G }$ using original image $\\mathbf { x }$ and the synthesized image $\\mathbf { x } ^ { \\mathrm { n e w } }$ with the augmented attribute. ", + "bbox": [ + 173, + 401, + 825, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attribute-space interpolation. We start with a simple solution (detailed in Eq. 1) assuming the adversarial example can be found by directly interpolating in the attribute-space. Given a pair of attributes c and $\\mathbf { c } ^ { \\mathrm { n e w } }$ , we introduce an interpolation parameter $\\alpha \\in ( 0 , 1 )$ to generate the augmented attribute vector $\\mathbf { c } ^ { \\ast } \\in \\mathbb { R } ^ { D _ { C } }$ (see Eq. 1). Given augmented attribute $\\mathbf { c } ^ { * }$ and original image $\\mathbf { x }$ , we produce the synthesized image by the generator $\\mathcal { G }$ . For our notation purpose, we also introduce a delegated function $\\tau _ { \\mathscr { G } }$ as a re-parametrization for the generator $\\mathcal { G }$ . Our formulation is also supported by the empirical results on attribute-conditioned image progression (Yan et al., 2016; Radford et al., 2015) that a well-trained generative model has the capability to synthesize a sequence of images with smooth attribute transitions. ", + "bbox": [ + 174, + 635, + 825, + 761 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/2cf3fa0d742819a17cd71c8f1c9444d6cdea0b5cdcfc852b7c0a1acebf2c9b24.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathbf { x } ^ { \\mathrm { a d v } } = \\arg \\operatorname* { m i n } _ { \\alpha } \\mathcal { L } \\big ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) \\big ) } \\\\ & { \\mathrm { w h e r e ~ } \\mathcal { T } _ { \\mathcal { G } } \\big ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } \\big ) = \\mathcal { G } \\big ( \\mathbf { x } , \\mathbf { c } ^ { \\ast } \\big ) \\mathrm { ~ a n d ~ } \\mathbf { c } ^ { \\ast } = \\alpha \\cdot \\mathbf { c } + ( 1 - \\alpha ) \\cdot \\mathbf { c } ^ { \\mathrm { n e w } } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 276, + 792, + 722, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature-map interpolation. Alternatively, we propose to interpolate using the feature map produced by the generator $\\mathcal { G } = \\mathcal { G } _ { \\mathrm { d e c } } \\circ \\mathcal { G } _ { \\mathrm { e n c } }$ . Here, $\\mathcal { G } _ { \\mathrm { e n c } }$ is the encoder module that takes the image as input and outputs the feature map. Similarly, $\\mathcal { G } _ { \\mathrm { d e c } }$ is the decoder module that takes the feature map as input and outputs the synthesized image. Let $\\mathbf { f } ^ { * } = { \\mathcal { G } } _ { \\mathrm { e n c } } ( \\mathbf { x } , \\mathbf { c } ) \\in \\mathbb { R } ^ { H _ { F } \\times W _ { F } \\times C _ { F } }$ be the feature map of an intermediate layer in the generator, where $H _ { F }$ , $W _ { F }$ and $C _ { F }$ indicate the height, width, and ", + "bbox": [ + 174, + 853, + 825, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "number of channels in the feature map. ", + "bbox": [ + 174, + 103, + 431, + 118 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/37bf53fe103985a1cf5894bd65e445513384b94a20c44c690d6ccc10495428c6.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathbf { x } ^ { \\mathrm { a d v } } = \\arg \\operatorname* { m i n } _ { \\alpha } \\mathcal { L } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ) } \\\\ & { \\mathrm { w h e r e ~ } \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) = \\mathcal { G } _ { \\mathrm { d e c } } ( \\mathbf { f } ^ { * } ) , } \\\\ & { \\mathbf { f } ^ { * } = \\alpha \\odot \\mathcal { G } _ { \\mathrm { e n c } } ( \\mathbf { x } , \\mathbf { c } ) + ( \\mathbf { 1 } - \\alpha ) \\odot \\mathcal { G } _ { \\mathrm { e n c } } ( \\mathbf { x } , \\mathbf { c } ^ { \\mathrm { n e w } } ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 336, + 125, + 658, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to attribute-space interpolation which is parameterized by a scalar, we parameterize feature-map interpolation by a tensor $\\alpha \\in \\mathbb { R } ^ { H _ { F } \\times W _ { F } \\times C _ { F } }$ $( \\alpha _ { h , w , k } \\in ( 0 , 1 )$ , where $1 \\leq h \\leq H _ { F }$ , $1 \\leq w \\leq W _ { F }$ , and $1 \\leq k \\leq C _ { F } ,$ ) with the same shape as the feature map. Compared to linear interpolation over attribute-space, such design introduces more flexibility when interpolating between the original image and the synthesized image. Empirical results in Section 4.2 show our design is critical to the adversarial attack success rate. ", + "bbox": [ + 173, + 191, + 826, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 ADVERSARIAL OPTIMIZATION OBJECTIVES ", + "text_level": 1, + "bbox": [ + 174, + 284, + 517, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As we see in Eq. 3, we obtain the adversarial image $\\mathbf { x } ^ { \\mathrm { a d v } }$ by minimizing the objective $\\mathcal { L } ( \\cdot )$ with respect to the synthesized image $\\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; { \\mathbf { x } } , { \\mathbf { c } } , { \\mathbf { c } } ^ { \\mathrm { n e w } } )$ , which is defined in Eq.(1) and Eq.(2) respectively. Here, each synthesized image $\\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; { \\mathbf { x } } , { \\mathbf { c } } , { \\mathbf { c } } ^ { \\mathrm { n e w } } )$ is produced by the interpolation using the conditional generator $\\mathcal { G }$ . In our objective function, the first term is the adversarial metric, the second term is a smoothness constraint, and $\\lambda$ is used to control the balance between the two terms. The adversarial metric is minimized once the model $\\mathcal { M }$ has been successfully attacked towards the target image-label pair $\\left( \\mathbf { x } ^ { \\mathrm { { t g t } } } , \\mathbf { y } ^ { \\mathrm { { t g t } } } \\right)$ . For identify verification, $\\mathbf { y } ^ { \\mathrm { t g t } }$ is the identity representation of the target image; For structured prediction tasks in our paper, $\\mathbf { y } ^ { \\mathrm { t g t } }$ either represents certain coordinates (landmark detection) or semantic label maps (semantic segmentation). ", + "bbox": [ + 173, + 303, + 825, + 429 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/1efed1d7a46b0aca0463c4579d9e699f884e9c5703308feb90b77c9b1a7a9eb5.jpg", + "text": "$$\n\\begin{array} { r } { \\begin{array} { c } { \\mathbf { x } ^ { \\mathrm { a d v } } = \\mathop { \\mathrm { a r g m i n } _ { \\alpha } } \\mathcal { L } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ) } \\\\ { \\mathcal { L } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ) = \\mathcal { L } _ { \\mathrm { a d v } } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ; \\mathcal { M } , \\mathbf { y } ^ { \\mathrm { t g t } } ) + \\lambda \\cdot \\mathcal { L } _ { \\mathrm { s m o o t h } } ( \\alpha ) } \\end{array} } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 259, + 454, + 738, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Identity verification. In the identity verification task, two images are considered to be the same identity if the corresponding identity embeddings from the verification model $\\mathcal { M }$ are reasonably close. ", + "bbox": [ + 174, + 505, + 826, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/8c4c82a4f6f71c490a457568b88e3a93f1f8097911b5a01ce8b2bd52defa3fba.jpg", + "text": "$$\n\\mathcal { L } _ { \\mathrm { a d v } } \\big ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ; \\mathcal { M } , \\mathbf { y } ^ { \\mathrm { t g t } } \\big ) = \\operatorname* { m a x } \\big ( \\kappa , \\Phi _ { \\mathcal { M } } ^ { \\mathrm { i d } } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) , \\mathbf { x } ^ { \\mathrm { t g t } } ) \\big )\n$$", + "text_format": "latex", + "bbox": [ + 259, + 544, + 736, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As we see in Eq. 4, $\\Phi _ { \\mathcal { M } } ^ { \\mathrm { i d } } ( \\cdot , \\cdot )$ measures the distance between two identity embeddings from the model $\\mathcal { M }$ , where the normalized $L _ { 2 }$ distance is used in our setting. In addition, we introduce the parameter $\\kappa$ representing the constant related to the false positive rate (FPR) threshold computed from the development set. ", + "bbox": [ + 173, + 568, + 825, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Structured prediction. For structured prediction tasks such as landmark detection and semantic segmentation, we use Houdini objective proposed in Cisse et al. (2017) as our adversarial metric and select the target landmark (semantic segmentation) target as $\\mathbf { y } ^ { \\mathrm { t g t } }$ . In addition, $\\Phi _ { \\mathcal { M } } ( \\cdot , \\cdot )$ is a scoring function for each image-label pair and $\\gamma$ is the threshold. ", + "bbox": [ + 174, + 630, + 825, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/1dd7fd2a01f78abd1aecaa21efc8865ed4c8013a46ddad5fd92dd920a1eb9697.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathcal { L } _ { \\mathrm { a d v } } \\big ( \\mathcal { T } _ { \\mathcal { G } } \\big ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } \\big ) ; \\mathcal { M } , \\mathbf { y } ^ { \\mathrm { t g t } } \\big ) = P _ { \\gamma \\sim \\mathcal { N } ( 0 , 1 ) } \\big [ \\Phi _ { \\mathcal { M } } \\big ( \\mathcal { T } _ { \\mathcal { G } } \\big ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } \\big ) , \\mathbf { y } ^ { * } \\big ) } \\\\ & { \\qquad - \\Phi _ { \\mathcal { M } } \\big ( \\mathcal { T } _ { \\mathcal { G } } \\big ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } \\big ) , \\mathbf { y } ^ { \\mathrm { t g t } } \\big ) < \\gamma \\big ] \\cdot l \\big ( \\mathbf { y } ^ { * } , \\mathbf { y } ^ { \\mathrm { t g t } } \\big ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 227, + 712, + 769, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf { y } ^ { * } = \\mathcal { M } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) )$ and $l ( \\mathbf { y } ^ { * } , \\mathbf { y } ^ { \\mathrm { t g t } } )$ is task loss decided by the specific adversarial target. ", + "bbox": [ + 174, + 762, + 825, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interpolation smoothness $\\mathcal { L } _ { \\mathrm { s m o o t h } }$ . As the tensor to be interpolated in the feature-map space has far more parameters compared to the attribute itself, we propose to enforce a smoothness constraint on the tensor $\\alpha$ used in feature-map interpolation. As we see in Eq. 6, the smoothness loss encourages the interpolation tensors to consist of piece-wise constant patches spatially, which has been widely used as a pixel-wise de-noising objective for natural image processing (Mahendran & Vedaldi, 2015; Johnson et al., 2016). ", + "bbox": [ + 173, + 797, + 825, + 883 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/9fb15c11a095b5f8eecd931cbd767537b05d0b949e50e726cb865aac1ce83152.jpg", + "text": "$$\n\\mathcal { L } _ { \\mathrm { s m o o t h } } ( \\alpha ) = \\sum _ { h = 1 } ^ { H _ { F } - 1 } \\sum _ { w = 1 } ^ { W _ { F } } \\| \\alpha _ { h + 1 , w } - \\alpha _ { h , w } \\| _ { 2 } ^ { 2 } + \\sum _ { h = 1 } ^ { H _ { F } } \\sum _ { w = 1 } ^ { W _ { F } - 1 } \\| \\alpha _ { h , w + 1 } - \\alpha _ { h , w } \\| _ { 2 } ^ { 2 }\n$$", + "text_format": "latex", + "bbox": [ + 248, + 887, + 750, + 933 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 174, + 102, + 326, + 117 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the experimental section, we mainly focus on analyzing the proposed SemanticAdv in attacking state-of-the-art face recognition systems on CelebA (Liu et al., 2015) due to its wide applicability (e.g., identification for mobile payment) in the real world. In addition, we extend our attack to urban street scenes with semantic label maps as the condition. We attack the semantic segmentation model DRN-D-22 (Yu et al., 2017) previously trained on Cityscape (Cordts et al., 2016) by generating adversarial examples with dynamic objects manipulated (e.g., insert a car into the scene). ", + "bbox": [ + 173, + 123, + 825, + 208 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The experimental section is organized as follows. First, we analyze the quality of generated adversarial examples and qualitatively compare our method with ${ \\mathcal { L } } _ { p }$ bounded pixel-wise optimization-based method (Carlini & Wagner, 2017; Dong et al., 2018; Xie et al., 2019). Second, we provide both qualitative and quantitative results by controlling each of the semantic attributes at a time. In terms of attack transferability, we evaluate our proposed SemanticAdv on various settings and further demonstrate the effectiveness of our method via query-free black-box attacks against online face verification platforms. Third, we compare our method with the baseline against different defense methods on the face verification task. Fourth, we demonstrate that the proposed SemanticAdv also applies to the face landmark detection and street-view semantic segmentation. ", + "bbox": [ + 174, + 213, + 825, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 EXPERIMENTAL SETUP ", + "text_level": 1, + "bbox": [ + 174, + 349, + 375, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Face identity verification. We select ResNet-50 and ResNet-101 (He et al., 2016) trained on MS-Celeb-1M (Guo et al., 2016) as our face verification models. The models are trained using two different objectives, namely, softmax loss (Sun et al., 2014; Zhang et al., 2018) and cosine loss (Wang et al., 2018a). For simplicity, we use the notation “R-N-S” to indicate the model with $N$ -layer residual blocks as backbone trained using softmax loss, while “R-N-C” indicates the same backbone trained using cosine loss. The distance between face features is measured by normalized L2 distance. For R-101-S model, we decide the parameter $\\kappa$ based on the false positive rate (FPR) for the identity verification task. Three different FPRs have been used: $1 0 ^ { - 3 }$ (with $\\kappa = 1 . 2 4 ,$ ), $3 \\times 1 0 ^ { - 4 }$ (with $\\kappa = 1 . 0 5$ ), and $1 0 ^ { - 4 }$ (with $\\kappa = 0 . 6 0 $ ). The distance metrics and selected thresholds are commonly used when evaluating the performance of face recognition model Klare et al. (2015); Kemelmacher-Shlizerman et al. (2016). Please check the Appendix (see Table B) for more details. To distinguish between the FPR we used in generating adversarial examples and the other FPR used in evaluation, we introduce two notations “Generating FPR (G-FPR)” and “Test FPR (T-FPR)”. For the experiment with query-free black-box API attacks, we use the online face verification services provided by ${ \\mathrm { F a c e } } + +$ (fac) and AliYun (ali). ", + "bbox": [ + 173, + 366, + 825, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Face landmark detection. We select Face Alignment Network (FAN) (Bulat & Tzimiropoulos, 2017b) trained on 300W-LP (Zhu et al., 2016b) and fine-tuned on 300-W (Sagonas et al., 2013) for 2D landmark detection. The network is constructed by stacking Hour-Glass network (Newell et al., 2016) with hierarchical block (Bulat & Tzimiropoulos, 2017a). Given a portrait image as input, FAN outputs 2D heatmaps which can be subsequently leveraged to yield 68 2D landmarks. ", + "bbox": [ + 174, + 582, + 825, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic attacks on face images. In our experiments, we randomly sample 1, 280 distinct identities form CelebA (Liu et al., 2015). To reduce the reconstruction error brought by the generator (e.g., $\\mathbf { x } \\neq \\mathcal { G } ( \\mathbf { x } , \\mathbf { c } ) )$ in practice, we take one more step to obtain the updated feature map $\\mathbf { f } ^ { \\prime } = \\mathcal { G } _ { \\mathrm { e n c } } ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } )$ , where $\\mathbf { x } ^ { \\prime } = \\mathrm { a r g m i n } _ { \\mathbf { x } ^ { \\prime } }$ $\\| \\mathcal { G } ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } ) - \\mathbf { x } \\|$ in feature-map interpolation. In our experiments, we use the last conv layer before upsampling in the generator as our as feature-map f given by the attack effectiveness. We also fix the parameter $\\lambda$ (e.g., balances the adversarial loss and smoothness constraint in Eq. 3) to be 0.01 for both face verification and landmark detection. ", + "bbox": [ + 174, + 659, + 825, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We used the StarGAN (Choi et al., 2018) for attribute-conditional image editing. In particular, we re-trained model on CelebA dataset (Liu et al., 2015) by aligning the face landmarks and then resizing images to resolution $1 1 2 \\times 1 1 2$ . In addition, we select 17 identity-preserving attributes as our input condition, as such attributes related to facial expression and hair color. ", + "bbox": [ + 174, + 762, + 825, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each distinct identity pair $( \\mathbf { x } , \\mathbf { x } ^ { \\mathrm { t g t } } )$ , we perform semanticAdv guided by each of the 17 attributes (e.g., we intentionally add or remove one specific attribute while keeping the rest unchanged). In total, for each image $\\mathbf { x }$ , we generate 17 adversarial images with different augmented attributes. In the experiments, we select a pixel-wise adversarial attack method (Carlini & Wagner, 2017) (referred as CW) as our baseline for comparison. Compared to our proposed method, CW does not require visual attributes as part of the system, as it only generates one adversarial example for each instance. We refer the corresponding attack success rate as the instance-wise success rate in which the attack success rate is calculated for each instance. For each instance with 17 adversarial images using different augmented attributes, if one of the 17 resulting images can attack successfully, we count the attack of this instance as one success, vice verse. ", + "bbox": [ + 173, + 825, + 825, + 924 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 103, + 823, + 145 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semantic attacks on street-view images. We select DRN-D-22 (Yu et al., 2017) as our semantic segmentation model and fine-tune the model on image regions with resolution $2 5 6 \\times 2 5 6$ . To synthesize semantic adversarial perturbations, we consider semantic label maps as the input attribute and leverage a generative image manipulation model (Hong et al., 2018) pre-trained on CityScape (Cordts et al., 2016) dataset. Given an input semantic label map at resolution $2 5 6 \\times 2 5 6$ , we select a target object instance (e.g., a pedestrian) to attack. Then, we create a manipulated semantic label map by inserting another object instance (e.g., a car) in the vicinity of the target object. Similar to the experiments in the face domain, for both semantic label maps, we used the image manipulation encoder to extract features (with 1, 024 channels at spatial resolution $1 6 \\times 1 6$ ) and conducted featurespace interpolation. We synthesized the final image by feeding the interpolated features to the image manipulation decoder. By searching the interpolation coefficient that maximizes the attack rate, we are able to fool the segmentation model with the synthesized final image. ", + "bbox": [ + 173, + 151, + 825, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 SemanticAdv ON FACE IDENTITY VERIFICATION ", + "text_level": 1, + "bbox": [ + 174, + 327, + 544, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attribute-space vs. feature-space interpolation. First, we found that both attribute-space and feature-space interpolation could generate reasonable samples (see Figure I in Appendix). Compared to attribute-space interpolation, generating adversarial examples with feature-space interpolation produced much better quantitative results (see Table E in Appendix). We measured the attack success rate of attribute-space interpolation (with $\\mathrm { G \\mathrm { - } F P R = T \\mathrm { - } F P R = 1 0 ^ { - 3 } , }$ ): $0 . 0 8 \\%$ on R-101-S, $0 . 3 1 \\%$ on R-101-C, and $0 . 1 6 \\%$ on both R-50-S and R-50-C. While feature-space interpolation achieves almost $100 \\%$ success rate on all those models (see Figure 3). We conjecture that this is because the high dimensional feature space can provide more manipulation freedom. ", + "bbox": [ + 173, + 344, + 825, + 455 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bc14263add8b7333aa1ab335a48d188a488851f0eeec7d268fb45170e51410fe.jpg", + "image_caption": [ + "Figure 2: Qualitative comparisons between our proposed SemanticAdv and pixel-wise adversarial examples generated by CW (Carlini & Wagner, 2017). Along with the adversarial examples, we also provide the corresponding perturbations (residual) on the right. Perturbations generated by our SemanticAdv $\\mathrm { \\bf G } \\mathrm { - } \\mathrm { \\bf F P R } = 1 0 ^ { - 3 } ,$ are unrestricted with semantically meaningful patterns. More results are shown in Appendix (see Figure N). " + ], + "image_footnote": [], + "bbox": [ + 220, + 470, + 769, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall analysis. Figure 2 shows the generated adversarial images and corresponding perturbations against R-101-S of SemanticAdv and CW respectively. The text below each figure is the name of augmented attribute, the sign before the name represents “adding” (in red) or “removing” (in blue) the corresponding attribute from the original image. We see that SemanticAdv is able to generate perceptually reasonable examples guided by the corresponding attribute. In particular, SemanticAdv is able to generate perturbations on the corresponding regions correlated with the augmented attribute, while the perturbations of CW have no specific pattern and are evenly distributed across the image. ", + "bbox": [ + 174, + 680, + 825, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis: controlling single attribute. One of the key advantages of SemanticAdv is that we can generate adversarial perturbations in a more controllable fashion guided by the semantic attributes. This allows analyzing the robustness of a recognition system against different types of semantic attacks. We group the adversarial examples by augmented attributes in various settings. In Figure 3, we present the attack success rate against two face verification models, namely, R-101-S and R-101-C, guided by different attributes. We highlight the bar with light blue for G-FPR equals to $1 0 ^ { - 3 }$ and blue for G-FPR equals to $1 0 ^ { - 4 }$ , respectively. As we see in this figure, with a larger T-FPR $1 0 ^ { - 3 }$ , our SemanticAdv can achieve almost $100 \\%$ attack success rate across different attributes. With a smaller T-FPR $1 0 ^ { - 4 }$ , we find that SemanticAdv guided by some attributes such as Mouth Slightly Open and Arched Eyebrows achieve less than $50 \\%$ attack success rate, while the other attributes such as Pale Skin and Eyeglasses are relatively less affected. In summary, we found that SemanticAdv guided by attributes describing the local shape (e.g., mouth, earrings) achieve a relatively lower attack success rate compared to attributes relevant to the color (e.g., hair color) or entire face region (e.g., skin). This suggests that the face verification models used in our experiments are more robustly trained in terms of detecting local shapes compared to colors. Please note that in practice we have the flexibility to select attributes for attacking an image based on the perceptual quality and attack success rate. ", + "bbox": [ + 174, + 785, + 825, + 924 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 825, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 shows the adversarial examples with augmented semantic attributes against R-101-S model. The attribute names are shown in the bottom. The upper images are $\\mathcal { G } ( \\mathbf { x } , \\mathbf { c } ^ { \\mathrm { n e w } } )$ generated by StarGAN with augmented attribute $\\mathbf { c } ^ { \\mathrm { { n e w } } }$ where the lower images are the corresponding adversarial images with the same augmented attribute. ", + "bbox": [ + 173, + 208, + 825, + 263 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f425e3990a1d398f6b0c0ca3d32c623c42afee2547c5a6fc767eb289e27fcc47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 280, + 818, + 426 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9a1a0608405ea4229a059b5e220b91f0651c5175ef5d100359c433825165714.jpg", + "image_caption": [ + "Figure 3: Quantitative analysis on the attack success rate with different single-attribute attacks. In each figure, we show the results correspond to a larger FPR $( \\mathrm { G } \\mathrm { - } \\mathrm { F P R } = \\mathrm { T } \\mathrm { - } \\mathrm { F P R } = 1 0 ^ { - 3 } )$ in skyblue and the results correspond to a smaller FPR $\\mathrm { \\langle G \\mathrm { - } F P R = T \\mathrm { - } F P R = 1 0 ^ { - 4 } \\rangle }$ ) in blue, respectively. " + ], + "image_footnote": [], + "bbox": [ + 176, + 280, + 491, + 425 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b1364cfc82980abf6870d20ac5cdf4f6c60172679d0b02e3a93fb132610f5c2.jpg", + "image_caption": [ + "Figure 4: Qualitative analysis on single-attribute adversarial attack $\\mathrm { ( G { - } F P R = 1 0 ^ { - 3 } }$ ). More results are shown in Appendix (see Figure K, Figure L and Figure M). " + ], + "image_footnote": [], + "bbox": [ + 241, + 501, + 753, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis: semantic attack transferability. To further understand the property of SemanticAdv, we analyze the transferability of SemanticAdv on various settings. For each model with different FPRs, we select the successfully attacked adversarial examples from Section 4.1 to construct our evaluation dataset and evaluate these adversarial samples across different models. Table 1a illustrates the transferability of SemanticAdv among different models by using the same FPRs $\\begin{array} { r } { ( \\mathrm { G - F P R } = \\mathrm { T } . } \\end{array}$ FPR $= 1 0 ^ { - 3 }$ ). Table 1b illustrates the result with different FPRs $\\mathbf { \\bar { ( G - F P R = 1 0 ^ { - 4 } } }$ and $\\mathrm { \\Delta T - F P R } = 1 0 ^ { - 3 }$ ) for generation and evaluation. As shown in Table 1a, adversarial examples generated against models trained with softmax loss exhibit certain transferability compared to models trained with cosine loss. We conduct the same experiment by generating adversarial examples with CW and found it has weaker transferability compared to our SemanticAdv (results in brackets of Table 1). ", + "bbox": [ + 173, + 714, + 825, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As Table 1b illustrates, the adversarial examples generated against the model with smaller G-FPR $= 1 0 ^ { - 4 }$ exhibit strong attack success rate when evaluating on the model with larger T-FPR $. = 1 0 ^ { - 3 }$ Especially, we found the adversarial examples generated against R-101-S have the best attack performance on other models. These findings motivate the analysis of query-free black-box API attack detailed in the following paragraph. ", + "bbox": [ + 174, + 861, + 825, + 930 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/57469e8d95a46fe820286c74e8b8ee4b6109015ae65e817728e3fc376178968b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Mtest/MoptR-50-SR-101-S
R-50-S1.000 (1.000)0.862 (0.530)
R-101-S0.874 (0.422)1.000 (1.000)
R-50-C0.693(0.347)0.837 (0.579)
R-101-C0.617 (0.218)0.888 (0.617)
", + "bbox": [ + 571, + 101, + 802, + 156 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a3f40dce747cd167588dcf20b5c1fc20e4dff6329527f1a11479082ce178844d.jpg", + "table_caption": [ + "Table 1: Transferability of SemanticAdv: cell $( i , j )$ shows attack success rate of adversarial examples generated against $j$ -th model and evaluate on $i$ -th model. Results of CW are listed in brackets. Left: Results generated with G-FPR $= 1 0 ^ { - 3 }$ and T-FPR $= 1 0 ^ { - 3 }$ ; Right: Results generated with G-FPR $=$ $1 0 ^ { - 4 }$ and $\\mathrm { T } \\mathrm { \\mathrm { - F P R } } = 1 0 ^ { - 3 }$ . " + ], + "table_footnote": [], + "table_body": "
MtestMoptR-50-SR-101-SR-50-CR-101-C
R-50-S1.000 (1.000)0.108 (0.007)0.023(0.002)0.018 (0.002)
R-101-S0.169 (0.006)1.000 (1.000)0.030 (0.002)0.032(0.003)
R-50-C0.166 (0.019)0.202(0.025)1.000 (1.000)0.048(0.007)
R-101-C0.120 (0.015)0.236 (0.029)0.040 (0.006)1.000(1.000)
", + "bbox": [ + 181, + 101, + 563, + 156 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(a) ", + "bbox": [ + 361, + 161, + 382, + 174 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Query-free black-box API attack. In this experiment, we generate adversarial examples against R101-S with $\\mathrm { G } \\mathrm { - } \\mathrm { F P R } = 1 0 ^ { - 3 } ( \\kappa = 1 . 2 4 )$ , $\\mathbf { G } \\mathbf { - } \\mathbf { F P R } \\mathbf { \\bar { = } } 1 0 ^ { - 4 } ( \\kappa = \\mathbf { \\bar { 0 } } . 6 0 )$ , and G-FPR $< 1 0 ^ { - 4 } ( \\kappa = 0 . 2 0 )$ , respectively. We evaluate our algorithm on two industry level APIs, namely, ${ \\mathrm { F a c e } } + +$ and AliYun face verification platform. Since attack transferability has never been explored in concurrent work that generates semantic adversarial examples, we use ${ \\mathcal { L } } _ { p }$ bounded pixel-wise methods (Carlini & Wagner, 2017; Dong et al., 2018; Xie et al., 2019) as our baselines. As we see Table 2, which shows the best performance of each method, our SemanticAdv achieves much higher attack success rate than CW in both APIs with all FPR thresholds (e.g., our adversarial examples generated with G-FPR $< 1 0 ^ { - 4 }$ achieves $6 4 . 6 3 \\%$ attack success rate on Face++ platform with $\\mathrm { T } \\mathrm { \\bar { F } P R } = 1 0 ^ { - 3 }$ ). In addition, we found that lower G-FPR can achieve higher attack success rate on APIs within the same T-FPR (see Table I in Appendix). ", + "bbox": [ + 173, + 250, + 825, + 402 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f7239179d0edfb192fb9f0ec0f1d521cf4a0fa8658cf81e559552e303921b81e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
API name Attacker/Evaluation MetricFace++AliYun
T-FPR=10 -3T-FPR=10 -4T-FPR=10 -3T-FPR=10 -4
Dong et al. (2018)30.7721.0318.006.50
Xie et al. (2019)37.9525.6421.5011.00
CW (G-FPR<10-4)41.6224.3719.0012.00
SemanticAdv (G-FPR <10-4)64.6342.6935.5022.17
", + "bbox": [ + 184, + 415, + 810, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2: Quantitative analysis on query-free black-box attack. We use ResNet-101 optimized with softmax loss for evaluation and report the attack success rate $( \\% )$ on two online face verification platforms. Note that for PGD-based attacks, we adopt MI-FGSM $\\epsilon = 8$ ) in Dong et al. (2018) and $\\mathrm { \\dot { \\mathbf { M } } \\mathbf { - } D I ^ { 2 } }$ -FGSM $\\epsilon = 8$ ) in Xie et al. (2019), respectively. ", + "bbox": [ + 174, + 505, + 825, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User study. To measure the perceptual quality of the adversarial images generated by SemanticAdv, we conduct a user study on Amazon Mechanical Turk (AMT). We use the adversarial examples generated with $\\mathrm { G } \\mathrm { - } \\mathrm { F P R } < 1 0 ^ { - 4 }$ , which is the most strict setting in our experiment, to conduct the user study for both CW and SemanticAdv. In total, we collect 2, 620 annotations from 77 participants. In $3 9 . 1 \\dot { 4 } \\pm 1 . 9 6 \\%$ (close to random guess $5 0 \\%$ ) of trials the adversarial images generated by SemanticAdv are selected as reasonably-looking images and in $3 0 . 2 7 \\pm 1 . 9 6 \\%$ of trials, the adversarial images generated by CW are selected as reasonably-looking. It indicates that SemanticAdv can generate more reasonable-looking adversarial examples compared with CW under the most strict setting with G-FPR $< 1 0 ^ { - 4 }$ . Qualitative comparisons are shown in Appendix (see Figure H). ", + "bbox": [ + 173, + 565, + 825, + 691 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e038d18d452da8c19a0e36bbebaafe417ea1f9ce207ae878fb70f59f617d2e80.jpg", + "image_caption": [ + "Figure 5: Quantitative analysis on attacking several defense methods including JPEG (Dziugaite et al., 2016), Blurring (Li & Li, 2017), and Feature Squeezing (Xu et al., 2017). " + ], + "image_footnote": [], + "bbox": [ + 207, + 705, + 787, + 827 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SemanticAdv against defense methods. We evaluate the strength of the proposed attack by testing against four existing defense methods, namely, Feature squeezing (Xu et al., 2017), Blurring (Li & Li, 2017), JPEG (Dziugaite et al., 2016) and AMI (Tao et al., 2018). For AMI (Tao et al., 2018), we first extract attribute witnesses with our aligned face images and then leverage them to construct attribute-steered model. We use fc7 of pretrained VGG (Parkhi et al., 2015) as the face representation. AMI yields a consistency score for each face image to indicate whether it is a benign image. The score is measured by the cosine similarity between the representations from original model and attribute-steered model. With $1 0 \\%$ false positives on benign inputs, it only achieves $8 \\%$ detection accuracy for SemanticAdv and $1 2 \\%$ detection accuracy for CW. ", + "bbox": [ + 174, + 868, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2514972a90994ba8bb3e4bf4a64b6780f5d758436abdb8a62c957825f90e25ed.jpg", + "image_caption": [ + "Figure 6: Qualitative results on attacking face landmark detection model. " + ], + "image_footnote": [], + "bbox": [ + 176, + 101, + 815, + 231 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/136eee23c85c9f369b82e0445973320cf8e16d085aa0e61850f470aa410f04ff.jpg", + "image_caption": [ + "Figure 7: Qualitative results on attacking street-view semantic segmentation model. " + ], + "image_footnote": [], + "bbox": [ + 179, + 280, + 815, + 448 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 494, + 825, + 564 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Figure 5 illustrates SemanticAdv is more robust against these defense methods comparing with CW. The same G-FPR and T-FPR are used for evaluation. Under the condition that T-FPR is $\\bar { 1 } 0 ^ { - 3 }$ , both SemanticAdv and CW achieve high attack success rate, while SemanticAdv marginally outperforms CW when FPR goes down to $1 0 ^ { - 4 }$ . While defense methods have proven to be effective against CW attacks on classifiers trained with ImageNet (Krizhevsky et al., 2012), our results indicate that these methods are still vulnerable in face verification system with small T-FPR. ", + "bbox": [ + 174, + 571, + 825, + 655 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 SemanticAdv ON FACE LANDMARK DETECTION ", + "text_level": 1, + "bbox": [ + 174, + 674, + 540, + 689 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We also evaluate the effectiveness of SemanticAdv on face landmark detection. We select two attack tasks, namely, “Rotating Eyes” and “Out of Region”. For the “Rotating Eyes” task, we rotate the coordinates of the eyes in the image counter-clockwise by $9 0 ^ { \\circ }$ . For the “Out of Region” task, we set a target bounding box and attempt to push all points out of the box. We summarize the experimental setup and quantitative results in the Appendix (see Table D). As we see in Figure 6, our method is applicable to attack landmark detection models. ", + "bbox": [ + 174, + 699, + 825, + 782 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 SemanticAdv ON STREET-VIEW SEMANTIC SEGMENTATION ", + "text_level": 1, + "bbox": [ + 174, + 801, + 619, + 816 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We further demonstrate the applicability of our SemanticAdv beyond face domain by generating adversarial perturbations on street-view images. Figure 7 illustrates the adversarial examples on semantic segmentation. In the first example, we select the leftmost the pedestrian as the target object instance and insert another car into the scene to attack it. The segmentation model has been successfully attacked to neglect the pedestrian (see last column), while it does exist in the scene (see second-to-last column). In the second example, we insert an adversarial car in the scene by SemanticAdv and the cyclist has been recognized as a pedestrian by the segmentation model. ", + "bbox": [ + 174, + 825, + 825, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 CONCLUSIONS ", + "text_level": 1, + "bbox": [ + 176, + 102, + 328, + 117 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Overall we presented a novel attack method SemanticAdv, which is capable of generating unrestricted adversarial perturbations guided by semantic attributes edition. Compared to existing methods, SemanticAdv works in a more controllable fashion. Experimental evaluations on face verification and landmark detection demonstrate several unique properties including attack transferability. We believe this work would open up new research opportunities and challenges in the field of adversarial learning. For instance, how to leverage semantic information to defend against such attacks will lead to potential new discussion. ", + "bbox": [ + 174, + 122, + 825, + 219 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 241, + 285, + 257 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alibaba Cloud Computing Co. Ltd. https://help.aliyun.com/knowledge_detail/ 53535.html. \nMegvii Technology Co. Ltd. https://console.faceplusplus.com/documents/ 5679308. \nYoshua Bengio, Grégoire Mesnil, Yann Dauphin, and Salah Rifai. Better mixing via deep representations. In ICML, 2013. \nAnand Bhattad, Min Jin Chong, Kaizhao Liang, Bo Li, and David A Forsyth. Big but imperceptible adversarial perturbations via semantic manipulation. arXiv preprint arXiv:1904.06347, 2019. \nAndrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. In ICLR, 2019. \nTom B Brown, Nicholas Carlini, Chiyuan Zhang, Catherine Olsson, Paul Christiano, and Ian Goodfellow. Unrestricted adversarial examples. arXiv preprint arXiv:1809.08352, 2018. \nAdrian Bulat and Georgios Tzimiropoulos. Binarized convolutional landmark localizers for human pose estimation and face alignment with limited resources. In Proceedings of the IEEE International Conference on Computer Vision, pp. 3706–3714, 2017a. \nAdrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem?(and a dataset of 230,000 3d facial landmarks). In ICCV, 2017b. \nNicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy (S&P). IEEE, 2017. \nYunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In CVPR, 2018. \nMoustapha Cisse, Yossi Adi, Natalia Neverova, and Joseph Keshet. Houdini: Fooling deep structured prediction models. In NIPS, 2017. \nRonan Collobert and Jason Weston. A unified architecture for natural language processing: Deep neural networks with multitask learning. In Proceedings of the 25th international conference on Machine learning, pp. 160–167. ACM, 2008. \nMarius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 3213–3223, 2016. \nLtsc Deng, Jinyu Li, Jui-Ting Huang, Kaisheng Yao, Dong Yu, Frank Seide, Michael L Seltzer, Geoffrey Zweig, Xiaodong He, Jason D Williams, et al. Recent advances in deep learning for speech research at microsoft. In ICASSP, volume 26, pp. 64, 2013. \nYinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 9185–9193, 2018. ", + "bbox": [ + 171, + 261, + 828, + 933 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Gintare Karolina Dziugaite, Zoubin Ghahramani, and Daniel M Roy. A study of the effect of jpg compression on adversarial images. arXiv preprint arXiv:1608.00853, 2016. ", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Logan Engstrom, Brandon Tran, Dimitris Tsipras, Ludwig Schmidt, and Aleksander Madry. A rotation and a translation suffice: Fooling cnns with simple transformations. arXiv preprint arXiv:1712.02779, 2017. ", + "bbox": [ + 173, + 140, + 823, + 183 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014a. ", + "bbox": [ + 169, + 191, + 825, + 222 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ian Goodfellow, Yoshua Bengio, Aaron Courville, and Yoshua Bengio. Deep learning, volume 1. MIT Press, 2016. ", + "bbox": [ + 173, + 229, + 825, + 258 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2014b. ", + "bbox": [ + 173, + 267, + 823, + 296 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In ECCV. Springer, 2016. ", + "bbox": [ + 173, + 304, + 823, + 334 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. ", + "bbox": [ + 173, + 342, + 823, + 371 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Seunghoon Hong, Xinchen Yan, Thomas S Huang, and Honglak Lee. Learning hierarchical semantic image manipulation through structured representations. In NeurIPS, 2018. ", + "bbox": [ + 173, + 378, + 823, + 409 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database forstudying face recognition in unconstrained environments. In Workshop on faces in’Real-Life’Images: detection, alignment, and recognition, 2008. ", + "bbox": [ + 176, + 416, + 825, + 460 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, pp. 1125–1134, 2017. ", + "bbox": [ + 169, + 468, + 823, + 497 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV. Springer, 2016. ", + "bbox": [ + 176, + 506, + 821, + 535 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In CVPR, pp. 1219–1228, 2018. ", + "bbox": [ + 173, + 542, + 825, + 573 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ameya Joshi, Amitangshu Mukherjee, Soumik Sarkar, and Chinmay Hegde. Semantic adversarial attacks: Parametric transformations that fool deep classifiers. arXiv preprint arXiv:1904.08489, 2019. ", + "bbox": [ + 174, + 580, + 825, + 623 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In ICLR, 2018. ", + "bbox": [ + 173, + 632, + 821, + 661 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ira Kemelmacher-Shlizerman, Steven M Seitz, Daniel Miller, and Evan Brossard. The megaface benchmark: 1 million faces for recognition at scale. In CVPR, pp. 4873–4882, 2016. ", + "bbox": [ + 174, + 670, + 820, + 699 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. ", + "bbox": [ + 173, + 707, + 810, + 723 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. ", + "bbox": [ + 173, + 731, + 751, + 746 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Brendan F Klare, Ben Klein, Emma Taborsky, Austin Blanton, Jordan Cheney, Kristen Allen, Patrick Grother, Alan Mah, and Anil K Jain. Pushing the frontiers of unconstrained face detection and recognition: Iarpa janus benchmark a. In CVPR, 2015. ", + "bbox": [ + 176, + 755, + 823, + 797 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NIPS, 2012. ", + "bbox": [ + 173, + 805, + 825, + 835 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Xin Li and Fuxin Li. Adversarial examples detection in deep networks with convolutional filter statistics. In Proceedings of the IEEE International Conference on Computer Vision, pp. 5764– 5772, 2017. ", + "bbox": [ + 173, + 843, + 823, + 886 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In NIPS, 2017. ", + "bbox": [ + 171, + 895, + 821, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In ICCV, 2015. ", + "bbox": [ + 171, + 103, + 825, + 133 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them. In CVPR, 2015. ", + "bbox": [ + 174, + 141, + 823, + 170 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Elman Mansimov, Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Generating images from captions with attention. In ICLR, 2015. ", + "bbox": [ + 173, + 179, + 823, + 208 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2574–2582, 2016. ", + "bbox": [ + 176, + 217, + 825, + 261 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In CVPR Workshops, pp. 51–59, 2017. ", + "bbox": [ + 173, + 268, + 825, + 313 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In European Conference on Computer Vision, pp. 483–499. Springer, 2016. ", + "bbox": [ + 171, + 320, + 823, + 351 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In ICML. JMLR, 2017. ", + "bbox": [ + 173, + 358, + 823, + 388 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In ICML, 2016. ", + "bbox": [ + 173, + 397, + 823, + 426 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In Security and Privacy (EuroS&P), 2016 IEEE European Symposium on, 2016. ", + "bbox": [ + 174, + 434, + 825, + 478 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, et al. Deep face recognition. In bmvc, volume 1, pp. 6, 2015. ", + "bbox": [ + 173, + 486, + 825, + 517 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. In ICLR, 2015. ", + "bbox": [ + 174, + 525, + 823, + 554 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Scott Reed, Kihyuk Sohn, Yuting Zhang, and Honglak Lee. Learning to disentangle factors of variation with manifold interaction. In ICML, 2014. ", + "bbox": [ + 173, + 563, + 825, + 593 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. ", + "bbox": [ + 173, + 601, + 825, + 631 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-thewild challenge: The first facial landmark localization challenge. In ICCV Workshop, 2013. ", + "bbox": [ + 171, + 638, + 825, + 669 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Soumyadip Sengupta, Jun-Cheng Chen, Carlos Castillo, Vishal M Patel, Rama Chellappa, and David W Jacobs. Frontal to profile face verification in the wild. In WACV, pp. 1–9. IEEE, 2016. ", + "bbox": [ + 171, + 676, + 823, + 707 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. nature, 529(7587):484, 2016. ", + "bbox": [ + 176, + 714, + 826, + 758 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. ", + "bbox": [ + 171, + 767, + 825, + 796 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yang Song, Rui Shu, Nate Kushman, and Stefano Ermon. Constructing unrestricted adversarial examples with generative models. In Advances in Neural Information Processing Systems, pp. 8312–8323, 2018. ", + "bbox": [ + 174, + 804, + 825, + 848 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yi Sun, Xiaogang Wang, and Xiaoou Tang. Deep learning face representation from predicting 10,000 classes. In CVPR, 2014. ", + "bbox": [ + 174, + 857, + 823, + 886 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. ", + "bbox": [ + 174, + 895, + 823, + 924 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, Andrew Rabinovich, et al. Going deeper with convolutions. In CVPR, 2015. ", + "bbox": [ + 176, + 103, + 825, + 145 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Guanhong Tao, Shiqing Ma, Yingqi Liu, and Xiangyu Zhang. Attacks meet interpretability: Attributesteered detection of adversarial samples. In NeurIPS, 2018. ", + "bbox": [ + 173, + 155, + 825, + 184 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelcnn decoders. In NIPS, 2016. ", + "bbox": [ + 173, + 193, + 823, + 222 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In CVPR, 2018a. ", + "bbox": [ + 174, + 229, + 823, + 260 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Highresolution image synthesis and semantic manipulation with conditional gans. In CVPR, 2018b. ", + "bbox": [ + 173, + 267, + 825, + 297 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Chaowei Xiao, Ruizhi Deng, Bo Li, Fisher Yu, Mingyan Liu, and Dawn Song. Characterizing adversarial examples based on spatial consistency information for semantic segmentation. In ECCV, 2018a. ", + "bbox": [ + 174, + 305, + 823, + 348 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Chaowei Xiao, Bo Li, Jun-Yan Zhu, Warren He, Mingyan Liu, and Dawn Song. Generating adversarial examples with adversarial networks. In IJCAI, 2018b. ", + "bbox": [ + 173, + 357, + 821, + 387 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Chaowei Xiao, Jun-Yan Zhu, Bo Li, Warren He, Mingyan Liu, and Dawn Song. Spatially transformed adversarial examples. In ICLR, 2018c. ", + "bbox": [ + 173, + 395, + 823, + 424 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Chaowei Xiao, Dawei Yang, Bo Li, Jia Deng, and Mingyan Liu. Meshadv: Adversarial meshes for visual recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6898–6907, 2019. ", + "bbox": [ + 176, + 433, + 825, + 476 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2730–2739, 2019. ", + "bbox": [ + 174, + 484, + 823, + 527 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. arXiv preprint arXiv:1704.01155, 2017. ", + "bbox": [ + 173, + 535, + 823, + 565 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Xinchen Yan, Jimei Yang, Kihyuk Sohn, and Honglak Lee. Attribute2image: Conditional image generation from visual attributes. In ECCV. Springer, 2016. ", + "bbox": [ + 174, + 573, + 823, + 603 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fisher Yu, Vladlen Koltun, and Thomas Funkhouser. Dilated residual networks. In Computer Vision and Pattern Recognition (CVPR), 2017. ", + "bbox": [ + 173, + 611, + 823, + 641 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In ICCV, 2017. ", + "bbox": [ + 174, + 648, + 825, + 693 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Xingcheng Zhang, Lei Yang, Junjie Yan, and Dahua Lin. Accelerated training for massive classification via dynamic class selection. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018. ", + "bbox": [ + 176, + 700, + 823, + 742 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Jun-Yan Zhu, Philipp Krähenbühl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the natural image manifold. In ECCV. Springer, 2016a. ", + "bbox": [ + 169, + 752, + 825, + 781 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. ", + "bbox": [ + 169, + 790, + 825, + 820 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z Li. Face alignment across large poses: A 3d solution. In CVPR, 2016b. ", + "bbox": [ + 173, + 828, + 823, + 857 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A FACE IDENTITY VERIFICATION ", + "text_level": 1, + "bbox": [ + 176, + 102, + 467, + 118 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Benchmark performance. First, we provide additional information about the ResNet models we used in the experiments. We summarize in Table A the performance on several face identity verification benchmarks including Labeled Face in the Wild (LFW) dataset (Huang et al., 2008), AgeDB-30 dataset (Moschoglou et al., 2017), and Celebrities in Frontal-Profile (CFP) dataset (Sengupta et al., 2016). ", + "bbox": [ + 173, + 126, + 826, + 196 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/09c7f12c436f5b70809b6349b5b2cc53a92a826f00065e6f26ae9a1543329bdf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
M benchmarksLFWAgeDB-30CFP-FFCFP-FP
R-50-S99.2794.1599.2691.49
R-101-S99.4295.9399.5795.07
R-50-C99.3895.0899.2490.24
R-101-C99.6795.5899.5792.71
", + "bbox": [ + 282, + 212, + 715, + 289 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table A: The performnace of ResNet models on several benchmark datasets. ", + "bbox": [ + 250, + 295, + 748, + 310 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Identity verification thresholds. Table B shows the threshold values used in our experiments when determining whether two portrait images belong to the same identity or not. The selected FPR thresholds and normalized L2 distance between face features are commonly used when evaluating the performance of face recognition model Klare et al. (2015); Kemelmacher-Shlizerman et al. (2016). ", + "bbox": [ + 174, + 319, + 825, + 376 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/60612b66a039fb5769d8d04f005e45f0aa712bca2bab53d6b13d0716fa50ea18.jpg", + "table_caption": [], + "table_footnote": [ + "Table B: The threshold values for face identity verification. " + ], + "table_body": "
FPR/MR-50-SR-101-SR-50-CR-101-C
10-33×10-410-41.1811.2441.4471.469
1.0580.6571.0481.2931.2420.809
0.5970.864
", + "bbox": [ + 308, + 391, + 684, + 457 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Implementation details. We use Adam (Kingma & Ba, 2015) as the optimizer to generate adversarial examples for both our SemanticAdv and CW. More specifically, we run optimization for up to 200 steps with a fixed learning rate 0.05 for cases when $\\mathbf { \\bar { G } } \\mathbf { - } \\mathbf { F P R } \\leq 1 0 ^ { - 4 }$ . Otherwise, we run optimization for up to 500 steps with a fixed learning rate 0.01. For pixel-wise attack method CW, we use additional pixel reconstruction loss with corresponding loss weight to 5. We run optimization for up to $1 , 0 0 0$ steps with a fixed learning rate $1 0 ^ { - 3 }$ . ", + "bbox": [ + 173, + 492, + 826, + 575 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Evaluation metrics. To evaluate the performance of semanticAdv under different attributes, we consider three metrics as follows: ", + "bbox": [ + 171, + 585, + 823, + 614 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "• Best: if there is one attribute among 17 attributes that can be successfully attacked, we count the attack success rate for this face identity as 1; \n• Average: we calculate the average attack success rate among 17 attributes for the same face identity; \n• Worst: only when all of 17 attributes can be successfully attacked, we count the attack success rate for this person as 1; ", + "bbox": [ + 215, + 627, + 825, + 729 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Note that, for a fair comparison with CW, we should use the Best metric for our SemanticAdv, as CW is the traditional pixel-wise attack method works regardless of the attribute. In addition, we report the performance using the average and worst metric, which actually provides additional insights into the robustness of face verification models across different attributes. For instance, combining the results from Table C and Figure 3, we understand that the face verification models used in our experiments have different levels of robustness across attributes. For example, face verification models are more robust against local shape variations than color variations, e.g., pale skin has higher attack success rate than mouth open. We believe these discoveries will help the community further understand the properties of face verification models. ", + "bbox": [ + 173, + 742, + 825, + 867 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table C shows the overall performance (accuracy) of face verification model and attack success rate of SemanticAdv and CW. As we can see from Table C, although the face model trained with cosine loss achieves higher face recognition performance, it is more vulnerable to adversarial attack compared with the model trained with softmax loss. ", + "bbox": [ + 174, + 875, + 823, + 930 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/c723aa96ee2ccc87ec708b7666c69b0e35c1469e2eeecab9e18e4c3f9fb0351f.jpg", + "table_caption": [ + "Table C: Quantitative result of identity verification $( \\% )$ . It shows accuracy of face verification model and attack success rate of SemanticAdv and CW. $\\mathbf { x } ^ { \\prime }$ , $G ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } )$ and $G ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } ^ { \\mathrm { n e w } } )$ are the intermediate results of our method before adversarial perturbation. " + ], + "table_footnote": [], + "table_body": "
G-FPRMetricsMR-50-SR-101-SR-50-CR-101-C
10-3Verification Accuracy98.3698.7898.6398.84
xG(x',c)G(x',cnew)W)(Best)0.000.000.160.000.000.080.080.000.160.000.230.31
SemanticAdv (Best)SemanticAdv (Worst)SemanticAdv (Average)100.0091.9598.98100.0093.9899.29100.0099.5399.97100.0099.7799.99
CW100.00100.00100.00100.00
3 ×10-4Verification Accuracy97.7397.9797.9197.85
xG(x',c)G(x',cnew)(Best)0.000.000.000.000.000.000.000.000.000.000.000.00
SemanticAdv (Best)SemanticAdv (Worst)SemanticAdv (Average)100.0083.7597.72100.0079.0697.35100.0098.9899.92100.0096.6499.72
CW100.00100.00100.00100.00
10-4Verification Accuracy93.2592.8093.4392.98
XG(x',c)G(x',cnew)(Best)0.000.000.000.000.000.000.000.000.000.000.000.00
SemanticAdv (Best)SemanticAdv (Worst)SemanticAdv (Average)100.0033.5983.53100.0019.8476.64100.0067.0395.57100.0048.6791.13
CW100.00100.00100.00100.00
", + "bbox": [ + 223, + 101, + 771, + 455 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Defense methods. Feature squeezing ( $\\mathrm { { X u } }$ et al., 2017) is a simple but effective method by reducing color bit depth to remove the adversarial effects. We compress the image represented by 8 bits for each channel to 4 bits for each channel to evaluate the effectiveness. For Blurring (Li & Li, 2017), we use a $3 \\times 3$ Gaussian kernel with standard deviation 1 to smooth the adversarial perturbations. For JPEG Dziugaite et al. (2016), it leverages the compression and decompression to remove the adversarial perturbation. We set the compression ratio as 0.75 in our experiment. ", + "bbox": [ + 173, + 518, + 825, + 602 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "B FACE LANDMARK DETECTION ", + "text_level": 1, + "bbox": [ + 176, + 625, + 462, + 641 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Implementation details. To perform attack on the face landmark detection model, we run optimization for up to 2, 000 steps with a fixed learning rate 0.05. We set the balancing factor $\\lambda$ (see Eq. 3) to value 0.01 for this experiment. ", + "bbox": [ + 174, + 647, + 826, + 690 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Evaluation Metrics. We apply two different metrics for two adversarial attack tasks respectively. For “Rotating Eyes” task, we use a well-adopted metric Normalized Mean Error (NME) (Bulat & Tzimiropoulos, 2017b) for experimental evaluation. ", + "bbox": [ + 174, + 699, + 826, + 741 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/b4e26125808bdb559bc0b3affbfb84101b7146cbcdd7dc4829f9d6e7f7fdf291.jpg", + "text": "$$\nr _ { \\mathrm { N M E } } = \\frac { 1 } { N } \\sum _ { k = 1 } ^ { N } \\frac { | | \\mathbf { p } _ { k } - \\hat { \\mathbf { p } } _ { k } | | _ { 2 } } { \\sqrt { W _ { B } * H _ { B } } } ,\n$$", + "text_format": "latex", + "bbox": [ + 400, + 750, + 598, + 794 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "where√ $\\mathbf { p } _ { k }$ denotes the $k$ -th ground-truth landmark, $\\hat { \\mathbf { p } } _ { \\mathbf { k } }$ denotes the $k$ -th predicted landmark and $\\sqrt { W _ { B } * H _ { B } }$ is the square-root area of ground-truth bounding box, where $W _ { B }$ and $H _ { B }$ represents the width and height of the box. ", + "bbox": [ + 176, + 801, + 825, + 844 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "For “Out of Region” task, we consider the attack is successful if the landmark predictions fall outside a pre-defined centering region on the portrait image. Thus, we introduce a metric that reflects the portion of landmarks outside of the pre-defined centering region: rOUT = NoutNtotal , where Nout denotes the number of predicted landmarks outside the pre-defined bounding box and $N _ { \\mathrm { t o t a l } }$ denotes the total number of landmarks. ", + "bbox": [ + 173, + 849, + 825, + 924 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/15014ed6d55c113e14271d46bd0e3619d3839cde6544458347bb6b4ba6bb7f93.jpg", + "table_caption": [], + "table_footnote": [ + "Table D: Quantitative results on face landmark detection $( \\% )$ ) The two row shows the measured ratios (lower is better) for “Rotating Eyes” and “Out Of Region” task, respectively. " + ], + "table_body": "
Tasks (Metrics)PristineAugmented Attributes
BlondHairYoungEyeglassesRosyCheeksSmilingArchedEyebrowsBangsPaleSkin
Rotating eyes (rNME)↓28.0414.0317.288.5813.2419.2123.4215.9910.72
Out-of-region (rouT)↓45.9817.4223.047.5116.6525.4433.8520.0313.51
", + "bbox": [ + 184, + 101, + 810, + 159 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "We present the quantitative results of SemanticAdv on face landmark detection model in Table D including two adversarial tasks, namely, “Out of Region” and “Rotating Eyes”. We observe that our method is efficient to perform attacking on landmark detection models. For certain attributes such as “Eyeglasses”, “Plae Skin”, SemanticAdv is able to achieve reasonably-good performance. ", + "bbox": [ + 174, + 213, + 825, + 268 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "C USER STUDY ", + "text_level": 1, + "bbox": [ + 174, + 296, + 318, + 313 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "We conducted a user study on the adversarial images of SemanticAdv and CW used in the experiment of API-attack and the original images. The adversarial images are generated with $\\mathrm { { G } \\mathrm { { - F P R } \\mathrm { { < } \\ 1 0 ^ { - 4 } } } }$ for both methods. We present a pair of original image and adversarial image to participants and ask them to rank the two options. The order of these two images is randomized and the images are displayed for 2 seconds in the screen during each trial. After the images disappear, the participants have unlimited time to select the more reasonably-looking image according to their perception. For each participant, one could only conduct at most 50 trials, and each adversarial image was shown to 5 different participants. Some qualitative results are shown in Figure H. In total, we collect 2, 620 annotations from 77 participants. In $3 9 . 1 4 \\pm 1 . 9 6 \\%$ of trials the adversarial images generated by SemanticAdv are selected as reasonably-looking images and in $3 0 . 2 7 \\pm 1 . 9 6 \\%$ of trails, the adversarial images generated by CW are selected as reasonably-looking images. It indicates that our semantic adversarial examples are more perceptual reasonably-looking than CW. Additionally, we also conduct the user study with larger $\\mathrm { { G } { - } \\mathrm { { F P R } } \\mathrm { { = } \\ i 0 ^ { - 3 } } }$ . In $4 5 . 4 2 \\pm 1 . 9 6 \\%$ of trials, the adversarial images generated by SemanticAdv are selected as reasonably-looking images, which is very close to the random guess $( \\dot { 5 } 0 \\% )$ . ", + "bbox": [ + 173, + 333, + 825, + 541 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/e11a9c182edab33d61fc6b803c35b0b754579590f5fdb181259f75a0c14af2ed.jpg", + "image_caption": [ + "Figure H: Qualitative comparisons among ground truth, pixel-wise adversarial examples generated by CW, and our proposed SemanticAdv. Here, we present the results from G-FPR $\\dot { < } 1 0 ^ { { - } 4 }$ so that perturbations are visible. " + ], + "image_footnote": [], + "bbox": [ + 173, + 560, + 823, + 857 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "D ABLATION STUDY: FEATURE-SPACE INTERPOLATION ", + "text_level": 1, + "bbox": [ + 174, + 102, + 651, + 118 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "We conduct an ablation study on feature-space interpolation by analyzing attack success rates with different feature-maps in the StarGAN network. Table E shows the attack success rate on R-101-S. Here, we use $\\mathbf { f } _ { i }$ to represent the feature-map after $i$ -th up-sampling operation. $\\mathbf { f } _ { 0 }$ denotes the featuremap before applying up-sampling operation. The result demonstrates that samples generated by interpolating on $\\mathbf { f } _ { 0 }$ achieve the highest success rate. Since $\\mathbf { f } _ { 0 }$ is the feature-map before decoder, it still well embeds semantic information in the feature space. We adopt $\\mathbf { f } _ { 0 }$ for interpolation in our experiments. ", + "bbox": [ + 173, + 132, + 826, + 231 + ], + "page_idx": 16 + }, + { + "type": "table", + "img_path": "images/91a2bd6a85284f2649ead1cc0ff2b5c630d90d41b7256375255f9041a4ee43e7.jpg", + "table_caption": [ + "We also conduct a qualitative comparison between attribute-space and feature-space interpolation. As shown in Figure I, images generated by attribute-space and feature-space interpolation are both reasonably-looking. " + ], + "table_footnote": [], + "table_body": "
T-FPR (G-FPR)10-3 (10-3)3× 10-4 (3×10-4)10-4 (10-4)
Layer (f)ff1f2ff1f2ff1f
Attack SuccessRate99.2998.3275.6297.3594.1057.1576.6467.4019.63
", + "bbox": [ + 181, + 290, + 820, + 337 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Table E: Attack success rate by selecting different layer’s feature-map for interpolation on $\\mathrm { R - } 1 0 1 { - } \\mathrm { S } ( \\% )$ $\\mathbf { f } _ { i }$ indicates the feature-map after $i$ -th up-sampling operation. ", + "bbox": [ + 173, + 348, + 825, + 376 + ], + "page_idx": 16 + }, + { + "type": "table", + "img_path": "images/d4093ecc5c81935abb31c8c6f5d35a1f9dcd51757521594078caa223e7ff226b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
T-FPR (G-FPR)10-3 (10-3)3×10)-4(3×10-4)10-4(10-4)
Layer (f)f-2f-1f-2f-1f-2f-1
Attack Success Rate49.4092.0930.4481.876.6645.46
", + "bbox": [ + 245, + 393, + 751, + 441 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Table F: Attack success rate by selecting different layer’s feature-map for interpolation on R-101- $S ( \\% )$ . $\\mathbf { f } _ { - 2 }$ indicates the feature-map after the last down-sampling operation and $\\mathbf { f } _ { - 2 }$ indicates the feature-map after $\\mathbf { f } _ { - 2 }$ . ", + "bbox": [ + 174, + 452, + 825, + 494 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/d2f867cde088d9d2eedec96d6e11f67a1803de628ed29a100713bd1b68d74171.jpg", + "image_caption": [ + "Figure I: Qualitative comparisons between attribute-space and feature-space interpolation. " + ], + "image_footnote": [], + "bbox": [ + 174, + 513, + 813, + 683 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "E SEMANTIC ATTACK TRANSFERABILITY ", + "text_level": 1, + "bbox": [ + 173, + 733, + 534, + 750 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "In Table G, we present the quantitative results of the transferability with $\\mathrm { { G } \\mathrm { { - } \\mathrm { { F P R } = 1 0 ^ { - 4 } } } }$ , T-FPR $= 1 0 ^ { - 4 }$ . We observe that with the strict criterion (Lower T-FPR) of the verification model, the transferability becomes lower cross different models. ", + "bbox": [ + 174, + 763, + 825, + 805 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "To explore whether the improvement of transferability in SemanticAdv is introduced by semantic editing rather than optimized semantic perturbation in feature space. We add a StrawMan $\\mathrm { [ C W + }$ $\\mathbf { x } ^ { \\mathrm { n e w } } ]$ baseline which uses a controllable semantic-attribute-based generator to generate semantically different images without any notion of an adversarial attack, and then applies standard $L _ { p }$ CW attacks on that generated image. The results are shown in Table H. The performance of the StrawMan $\\mathrm { [ C W + }$ $\\mathbf { x } ^ { \\mathrm { n e w } } ]$ is worse than SemanticAdv. This result justifies that our SemanticAdv is able to produce novel adversarial examples which cannot be simply achieved by combining attribute-conditional image editing model with $L _ { p }$ bounded perturbation. ", + "bbox": [ + 173, + 811, + 825, + 924 + ], + "page_idx": 16 + }, + { + "type": "table", + "img_path": "images/b9638e8993d797e1b0aabd5661c4971c38b6b28ec0567c1805c498bb103d705a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Mtest/MoptR-50-SR-101-SR-50-CR-101-C
R-50-S1.0000.0050.0000.000
R-101-S0.0001.0000.0000.000
R-50-C0.0000.0001.0000.000
R-101-C0.0000.0000.0001.000
", + "bbox": [ + 339, + 101, + 679, + 166 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Table G: Transferability of SemanticAdv: cell $( i , j )$ shows attack success rate of adversarial examples generated against $j$ -th model and evaluate on $i$ -th model. Results generated with $\\mathrm { { G } { - } \\mathrm { { F P R } } = 1 \\bar { 0 } ^ { - 4 } }$ , T-FPR $= 1 0 ^ { \\overline { { - 4 } } }$ . ", + "bbox": [ + 176, + 171, + 826, + 213 + ], + "page_idx": 17 + }, + { + "type": "table", + "img_path": "images/15be8ca66a93ac5a78392d7beb4717cbe9393676b91e1e007b467321f3e04dd8.jpg", + "table_caption": [], + "table_footnote": [ + "(a) $\\mathrm { G } \\mathrm { - } \\mathrm { F P R } { = } 1 0 ^ { - 3 }$ , $\\mathrm { T } \\mathrm { - } \\mathrm { F P R } { = } 1 0 ^ { - 3 }$ " + ], + "table_body": "
MtestMoptR-101-S
R-50-S0.035 (0.108)
R-101-S1.000 (1.000)
R-50-C0.145 (0.202)
R-101-C0.085 (0.236)
", + "bbox": [ + 228, + 228, + 449, + 306 + ], + "page_idx": 17 + }, + { + "type": "table", + "img_path": "images/4827389ee367da045ec825b88c18f43da5dc29f9192d061d184b57e4f16201f6.jpg", + "table_caption": [], + "table_footnote": [ + "(b) G-FPR=10−4, $\\mathrm { T } { \\cdot } \\mathrm { F P R } { = } 1 0 ^ { - 3 }$ " + ], + "table_body": "
MtestMoptR-101-S
R-50-S0.615 (0.862)
R-101-S1.000(1.000)
R-50-C0.570(0.837)
R-101-C0.6955(0.888)
", + "bbox": [ + 544, + 228, + 764, + 306 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Table H: Transferability of StrawMan: cell $( i , j )$ shows attack success rate of adversarial examples generated against $j$ -th model and evaluate on $i$ -th model. Results of SemanticAdv are listed in brackets. ", + "bbox": [ + 173, + 332, + 825, + 375 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "F QUERY-FREE BLACK-BOX API ATTACK ", + "text_level": 1, + "bbox": [ + 174, + 400, + 535, + 416 + ], + "page_idx": 17 + }, + { + "type": "table", + "img_path": "images/633977a38163cddb5aed19b8bc3c70a9c7d81355e9906835d31b8c39eb63b23b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
API nameAttacker/Evaluation MetricFace++AliYun
T-FPR= 10-3T-FPR=10-4T-FPR=10-3T-FPR=10
Original x2.040.510.500.00
Generated xnew4.210.530.500.00
CW (G-FPR=10-)StrawMan [CW + xnew ](G-FPR = 10-3)SemanticAdv (G-FPR =10-3)16.2410.7127.323.554.089.794.500.00
3.000.00
7.502.00
CW (G-FPR= 10-4)StrawMan [CW+ xnew ](G-FPR = 10-4)SemanticAdv (G-FPR =10-4)30.6121.3257.2215.829.1438.6612.504.50
7.5029.501.5017.50
Dong et al. (2018)Xie et al. (2019)CW (G-FPR<10-4)StrawMan [CW + xnew](G-FPR<10-4)SemanticAdv (G-FPR<10-4)30.7737.9521.0318.006.50
25.6421.5011.00
41.6227.6964.6324.3719.0012.00
15.3842.6910.003.00
35.5022.17
", + "bbox": [ + 176, + 438, + 825, + 631 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Table I: Quantitative analysis on query-free black-box attack. We use ResNet-101 optimized with softmax loss for evaluation and report the attack success rate $( \\% )$ . ", + "bbox": [ + 174, + 638, + 820, + 666 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "In Table I, we present the results of SemanticAdv performing query-free black-box attack on two online face verification platforms. SemanticAdv outperforms CW and StawnMan in both APIs under all FPR thresholds. In addition, we achieve higher attack success rate on APIs using samples generated using lower G-FPR compared to samples generated using higher G-FPR with the same T-FPR. Original $\\mathbf { x }$ and generated $\\mathbf { x } ^ { \\mathrm { n e w } }$ are regarded as reference point of the performance of online face verification platforms. ", + "bbox": [ + 173, + 669, + 825, + 752 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Figure J illustrates our SemanticAdv on attacking Microsoft Azure face verification system, which further demonstrate the effectiveness of our approach. ", + "bbox": [ + 173, + 758, + 825, + 789 + ], + "page_idx": 17 + }, + { + "type": "image", + "img_path": "images/e030ab5d5338f1ba2bcc57782a19fec5f232a173459480ed89b123b6102bfa5c.jpg", + "image_caption": [ + "Figure J: Illustration of our SemanticAdv in the real world face verification platform (editing on pale skin). Note that the confidence denotes the likelihood that two faces belong to the same person. " + ], + "image_footnote": [], + "bbox": [ + 179, + 343, + 821, + 647 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "G MORE VISUALIZATIONS ", + "text_level": 1, + "bbox": [ + 176, + 103, + 411, + 117 + ], + "page_idx": 19 + }, + { + "type": "image", + "img_path": "images/3a7d8c713d580a52035bd642fd776679a6153075581bff76f96b228a53b76fa7.jpg", + "image_caption": [ + "Figure K: Qualitative analysis on single-attribute adversarial attack $\\mathrm { ( G { - } F P R = 1 0 ^ { - 3 } }$ ). " + ], + "image_footnote": [], + "bbox": [ + 194, + 142, + 803, + 864 + ], + "page_idx": 19 + }, + { + "type": "image", + "img_path": "images/de33f8bb372e2f4abdf4f6769428fdaa8a0623516671ba763ae2fab7c70afb7c.jpg", + "image_caption": [ + "Figure L: Qualitative analysis on single-attribute adversarial attack $\\mathrm { ( G - F P R = 1 0 ^ { - 3 } }$ ). " + ], + "image_footnote": [], + "bbox": [ + 194, + 148, + 803, + 863 + ], + "page_idx": 20 + }, + { + "type": "image", + "img_path": "images/7c8afb5499aa1945df2aa8c49f2ce5827b76cbafd081d867c0cbe11df573a7e7.jpg", + "image_caption": [ + "Figure M: Qualitative analysis on single-attribute adversarial attack $\\mathrm { \\bf G } \\mathrm { - } \\mathrm { \\bf F P R } = 1 0 ^ { - 3 } ,$ ). " + ], + "image_footnote": [], + "bbox": [ + 194, + 334, + 803, + 680 + ], + "page_idx": 21 + }, + { + "type": "image", + "img_path": "images/bc2242ce0987cedaa8c20dea5c4a3437513f800c661325f11f4134300127ee38.jpg", + "image_caption": [ + "Figure N: Qualitative comparisons between our proposed SemanticAdv $\\mathrm { ( G { - } F P R = 1 0 ^ { - 3 } }$ ) and pixelwise adversarial examples generated by CW. Along with the adversarial examples, we also provide the corresponding perturbations (residual) on the right. " + ], + "image_footnote": [], + "bbox": [ + 186, + 104, + 805, + 872 + ], + "page_idx": 22 + }, + { + "type": "image", + "img_path": "images/0b62987d5a2bc9e11b6691551e91f0da0480e64b18a63bf77468e026095b805e.jpg", + "image_caption": [ + "Figure O: Qualitative analysis on single-attribute adversarial attack (SemanticAdv with $\\mathrm { { \\bf G } { - } F P R = }$ $1 0 ^ { - 3 }$ ) by each other. Along with the adversarial examples, we also provide the corresponding perturbations (residual) on the right. " + ], + "image_footnote": [], + "bbox": [ + 179, + 219, + 816, + 743 + ], + "page_idx": 23 + } +] \ No newline at end of file diff --git a/parse/train/r1l-VeSKwS/r1l-VeSKwS_middle.json b/parse/train/r1l-VeSKwS/r1l-VeSKwS_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..7b3697b207e1d1157f359254f218db874d9b0220 --- /dev/null +++ b/parse/train/r1l-VeSKwS/r1l-VeSKwS_middle.json @@ -0,0 +1,52129 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 79, + 506, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 508, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 508, + 97 + ], + "score": 1.0, + "content": "SEMANTICADV: GENERATING ADVERSARIAL EXAM-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 497, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 497, + 117 + ], + "score": 1.0, + "content": "PLES VIA ATTRIBUTE-CONDITIONAL IMAGE EDITING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 135, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 135, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 135, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 146, + 245, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 146, + 245, + 158 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 278, + 172, + 333, + 184 + ], + "lines": [ + { + "bbox": [ + 276, + 171, + 336, + 186 + ], + "spans": [ + { + "bbox": [ + 276, + 171, + 336, + 186 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 197, + 469, + 406 + ], + "lines": [ + { + "bbox": [ + 141, + 197, + 469, + 210 + ], + "spans": [ + { + "bbox": [ + 141, + 197, + 469, + 210 + ], + "score": 1.0, + "content": "Deep neural networks (DNNs) have achieved great success in various applications", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 208, + 469, + 221 + ], + "spans": [ + { + "bbox": [ + 141, + 208, + 469, + 221 + ], + "score": 1.0, + "content": "due to their strong expressive power. However, recent studies have shown that", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 219, + 469, + 231 + ], + "spans": [ + { + "bbox": [ + 141, + 219, + 469, + 231 + ], + "score": 1.0, + "content": "DNNs are vulnerable to adversarial examples which are manipulated instances", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 231, + 469, + 243 + ], + "spans": [ + { + "bbox": [ + 142, + 231, + 469, + 243 + ], + "score": 1.0, + "content": "targeting to mislead DNNs to make incorrect predictions. Currently, most such", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 241, + 468, + 254 + ], + "spans": [ + { + "bbox": [ + 141, + 241, + 455, + 254 + ], + "score": 1.0, + "content": "adversarial examples try to guarantee “subtle perturbation\" by limiting the", + "type": "text" + }, + { + "bbox": [ + 455, + 241, + 468, + 253 + ], + "score": 0.87, + "content": "L _ { p }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 253, + 469, + 264 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 469, + 264 + ], + "score": 1.0, + "content": "norm of the perturbation. In this paper, we aim to explore the impact of semantic", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 263, + 470, + 275 + ], + "spans": [ + { + "bbox": [ + 141, + 263, + 470, + 275 + ], + "score": 1.0, + "content": "manipulation on DNNs predictions by manipulating the semantic attributes of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 274, + 469, + 287 + ], + "spans": [ + { + "bbox": [ + 141, + 274, + 469, + 287 + ], + "score": 1.0, + "content": "images and generate “unrestricted adversarial examples\". In particular, we propose", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 286, + 470, + 298 + ], + "spans": [ + { + "bbox": [ + 142, + 286, + 470, + 298 + ], + "score": 1.0, + "content": "an algorithm SemanticAdv which leverages disentangled semantic factors to gener-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 297, + 469, + 307 + ], + "spans": [ + { + "bbox": [ + 142, + 297, + 469, + 307 + ], + "score": 1.0, + "content": "ate adversarial perturbation by altering controlled semantic attributes to fool the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 307, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 141, + 307, + 469, + 320 + ], + "score": 1.0, + "content": "learner towards various “adversarial\" targets. We conduct extensive experiments", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 318, + 469, + 330 + ], + "spans": [ + { + "bbox": [ + 141, + 318, + 469, + 330 + ], + "score": 1.0, + "content": "to show that the semantic based adversarial examples can not only fool different", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 329, + 469, + 341 + ], + "spans": [ + { + "bbox": [ + 141, + 329, + 469, + 341 + ], + "score": 1.0, + "content": "learning tasks such as face verification and landmark detection, but also achieve", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 340, + 470, + 352 + ], + "spans": [ + { + "bbox": [ + 141, + 340, + 470, + 352 + ], + "score": 1.0, + "content": "high targeted attack success rate against real-world black-box services such as", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 351, + 469, + 362 + ], + "spans": [ + { + "bbox": [ + 142, + 351, + 469, + 362 + ], + "score": 1.0, + "content": "Azure face verification service based on transferability. To further demonstrate the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 362, + 470, + 374 + ], + "spans": [ + { + "bbox": [ + 141, + 362, + 470, + 374 + ], + "score": 1.0, + "content": "applicability of SemanticAdv beyond face recognition domain, we also generate", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 373, + 469, + 385 + ], + "spans": [ + { + "bbox": [ + 141, + 373, + 469, + 385 + ], + "score": 1.0, + "content": "semantic perturbations on street-view images. Such adversarial examples with", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 384, + 470, + 396 + ], + "spans": [ + { + "bbox": [ + 142, + 384, + 470, + 396 + ], + "score": 1.0, + "content": "controlled semantic manipulation can shed light on further understanding about", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 394, + 409, + 407 + ], + "spans": [ + { + "bbox": [ + 141, + 394, + 409, + 407 + ], + "score": 1.0, + "content": "vulnerabilities of DNNs as well as potential defensive approaches.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 14 + }, + { + "type": "title", + "bbox": [ + 108, + 418, + 206, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 208, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 208, + 433 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 434, + 505, + 566 + ], + "lines": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "Deep neural networks (DNNs) have demonstrated great successes in advancing the state-of-the-art", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 446, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 458 + ], + "score": 1.0, + "content": "performance of discriminative tasks (Krizhevsky et al., 2012; Goodfellow et al., 2016; He et al., 2016;", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "Collobert & Weston, 2008; Deng et al., 2013; Silver et al., 2016). However, recent research found", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "that DNNs are vulnerable to adversarial examples which are carefully crafted instances aiming to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "induce arbitrary prediction errors for learning systems. Such adversarial examples containing small", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "score": 1.0, + "content": "magnitude of perturbation have shed light on understanding and discovering potential vulnerabilities", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "score": 1.0, + "content": "of DNNs (Szegedy et al., 2013; Goodfellow et al., 2014b; Moosavi-Dezfooli et al., 2016; Papernot", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "score": 1.0, + "content": "et al., 2016; Carlini & Wagner, 2017; Xiao et al., 2018b;c;a; 2019). Most existing work focused", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 300, + 535 + ], + "score": 1.0, + "content": "on constructing adversarial examples by adding", + "type": "text" + }, + { + "bbox": [ + 301, + 523, + 313, + 534 + ], + "score": 0.89, + "content": "{ \\mathcal { L } } _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "bounded pixel-wise perturbations (Goodfellow", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 532, + 507, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 507, + 547 + ], + "score": 1.0, + "content": "et al., 2014b) or spatially transforming the image (Xiao et al., 2018c; Engstrom et al., 2017) (e.g.,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "score": 1.0, + "content": "in-plane rotation or out-of-plane rotation). Generating unrestricted perturbations with semantically", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 555, + 351, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 351, + 568 + ], + "score": 1.0, + "content": "meaningful patterns is an important yet under-explored field.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 570, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 586 + ], + "score": 1.0, + "content": "At the same time, deep generative models have demonstrated impressive performance in learning", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "disentangled semantic factors through data generation in an unsupervised (Radford et al., 2015; Karras", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 593, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 506, + 606 + ], + "score": 1.0, + "content": "et al., 2018; Brock et al., 2019) or weakly-supervised manner based on semantic attributes (Yan et al.,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "2016; Choi et al., 2018). Empirical findings in (Yan et al., 2016; Zhu et al., 2016a; Radford et al.,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "2015) demonstrated that a simple linear interpolation on the learned image manifold can produce", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 626, + 336, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 336, + 640 + ], + "score": 1.0, + "content": "smooth visual transitions between a pair of input images.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "score": 1.0, + "content": "In this paper, we introduce a novel attack SemanticAdv which generates unrestricted perturbations", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "with semantically meaningful patterns. Motivated by the findings mentioned above, we leverage an", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "attribute-conditional image editing model (Choi et al., 2018) to synthesize adversarial examples by", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "interpolating between source and target images in the feature-map space. Here, we focus on changing", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "a single attribute dimension to achieve adversarial goals while keeping the generated adversarial", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "image reasonably-looking (e.g., see Figure 1). To validate the effectiveness of the proposed attack", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 711, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 721 + ], + "score": 1.0, + "content": "method, we consider two tasks, namely, face verification and landmark detection, as face recognition", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 732 + ], + "score": 1.0, + "content": "field has been extensively explored and the commercially used face models are relatively robust", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 46.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 79, + 506, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 508, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 508, + 97 + ], + "score": 1.0, + "content": "SEMANTICADV: GENERATING ADVERSARIAL EXAM-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 497, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 497, + 117 + ], + "score": 1.0, + "content": "PLES VIA ATTRIBUTE-CONDITIONAL IMAGE EDITING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 135, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 135, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 135, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 146, + 245, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 146, + 245, + 158 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 111, + 135, + 245, + 158 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 172, + 333, + 184 + ], + "lines": [ + { + "bbox": [ + 276, + 171, + 336, + 186 + ], + "spans": [ + { + "bbox": [ + 276, + 171, + 336, + 186 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 197, + 469, + 406 + ], + "lines": [ + { + "bbox": [ + 141, + 197, + 469, + 210 + ], + "spans": [ + { + "bbox": [ + 141, + 197, + 469, + 210 + ], + "score": 1.0, + "content": "Deep neural networks (DNNs) have achieved great success in various applications", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 208, + 469, + 221 + ], + "spans": [ + { + "bbox": [ + 141, + 208, + 469, + 221 + ], + "score": 1.0, + "content": "due to their strong expressive power. However, recent studies have shown that", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 219, + 469, + 231 + ], + "spans": [ + { + "bbox": [ + 141, + 219, + 469, + 231 + ], + "score": 1.0, + "content": "DNNs are vulnerable to adversarial examples which are manipulated instances", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 231, + 469, + 243 + ], + "spans": [ + { + "bbox": [ + 142, + 231, + 469, + 243 + ], + "score": 1.0, + "content": "targeting to mislead DNNs to make incorrect predictions. Currently, most such", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 241, + 468, + 254 + ], + "spans": [ + { + "bbox": [ + 141, + 241, + 455, + 254 + ], + "score": 1.0, + "content": "adversarial examples try to guarantee “subtle perturbation\" by limiting the", + "type": "text" + }, + { + "bbox": [ + 455, + 241, + 468, + 253 + ], + "score": 0.87, + "content": "L _ { p }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 253, + 469, + 264 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 469, + 264 + ], + "score": 1.0, + "content": "norm of the perturbation. In this paper, we aim to explore the impact of semantic", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 263, + 470, + 275 + ], + "spans": [ + { + "bbox": [ + 141, + 263, + 470, + 275 + ], + "score": 1.0, + "content": "manipulation on DNNs predictions by manipulating the semantic attributes of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 274, + 469, + 287 + ], + "spans": [ + { + "bbox": [ + 141, + 274, + 469, + 287 + ], + "score": 1.0, + "content": "images and generate “unrestricted adversarial examples\". In particular, we propose", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 286, + 470, + 298 + ], + "spans": [ + { + "bbox": [ + 142, + 286, + 470, + 298 + ], + "score": 1.0, + "content": "an algorithm SemanticAdv which leverages disentangled semantic factors to gener-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 297, + 469, + 307 + ], + "spans": [ + { + "bbox": [ + 142, + 297, + 469, + 307 + ], + "score": 1.0, + "content": "ate adversarial perturbation by altering controlled semantic attributes to fool the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 307, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 141, + 307, + 469, + 320 + ], + "score": 1.0, + "content": "learner towards various “adversarial\" targets. We conduct extensive experiments", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 318, + 469, + 330 + ], + "spans": [ + { + "bbox": [ + 141, + 318, + 469, + 330 + ], + "score": 1.0, + "content": "to show that the semantic based adversarial examples can not only fool different", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 329, + 469, + 341 + ], + "spans": [ + { + "bbox": [ + 141, + 329, + 469, + 341 + ], + "score": 1.0, + "content": "learning tasks such as face verification and landmark detection, but also achieve", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 340, + 470, + 352 + ], + "spans": [ + { + "bbox": [ + 141, + 340, + 470, + 352 + ], + "score": 1.0, + "content": "high targeted attack success rate against real-world black-box services such as", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 351, + 469, + 362 + ], + "spans": [ + { + "bbox": [ + 142, + 351, + 469, + 362 + ], + "score": 1.0, + "content": "Azure face verification service based on transferability. To further demonstrate the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 362, + 470, + 374 + ], + "spans": [ + { + "bbox": [ + 141, + 362, + 470, + 374 + ], + "score": 1.0, + "content": "applicability of SemanticAdv beyond face recognition domain, we also generate", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 373, + 469, + 385 + ], + "spans": [ + { + "bbox": [ + 141, + 373, + 469, + 385 + ], + "score": 1.0, + "content": "semantic perturbations on street-view images. Such adversarial examples with", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 384, + 470, + 396 + ], + "spans": [ + { + "bbox": [ + 142, + 384, + 470, + 396 + ], + "score": 1.0, + "content": "controlled semantic manipulation can shed light on further understanding about", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 394, + 409, + 407 + ], + "spans": [ + { + "bbox": [ + 141, + 394, + 409, + 407 + ], + "score": 1.0, + "content": "vulnerabilities of DNNs as well as potential defensive approaches.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 14, + "bbox_fs": [ + 141, + 197, + 470, + 407 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 418, + 206, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 208, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 208, + 433 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 434, + 505, + 566 + ], + "lines": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "Deep neural networks (DNNs) have demonstrated great successes in advancing the state-of-the-art", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 446, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 458 + ], + "score": 1.0, + "content": "performance of discriminative tasks (Krizhevsky et al., 2012; Goodfellow et al., 2016; He et al., 2016;", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "Collobert & Weston, 2008; Deng et al., 2013; Silver et al., 2016). However, recent research found", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "that DNNs are vulnerable to adversarial examples which are carefully crafted instances aiming to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "induce arbitrary prediction errors for learning systems. Such adversarial examples containing small", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "score": 1.0, + "content": "magnitude of perturbation have shed light on understanding and discovering potential vulnerabilities", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "score": 1.0, + "content": "of DNNs (Szegedy et al., 2013; Goodfellow et al., 2014b; Moosavi-Dezfooli et al., 2016; Papernot", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "score": 1.0, + "content": "et al., 2016; Carlini & Wagner, 2017; Xiao et al., 2018b;c;a; 2019). Most existing work focused", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 300, + 535 + ], + "score": 1.0, + "content": "on constructing adversarial examples by adding", + "type": "text" + }, + { + "bbox": [ + 301, + 523, + 313, + 534 + ], + "score": 0.89, + "content": "{ \\mathcal { L } } _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "bounded pixel-wise perturbations (Goodfellow", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 532, + 507, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 507, + 547 + ], + "score": 1.0, + "content": "et al., 2014b) or spatially transforming the image (Xiao et al., 2018c; Engstrom et al., 2017) (e.g.,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "score": 1.0, + "content": "in-plane rotation or out-of-plane rotation). Generating unrestricted perturbations with semantically", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 555, + 351, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 351, + 568 + ], + "score": 1.0, + "content": "meaningful patterns is an important yet under-explored field.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 435, + 507, + 568 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 570, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 586 + ], + "score": 1.0, + "content": "At the same time, deep generative models have demonstrated impressive performance in learning", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "disentangled semantic factors through data generation in an unsupervised (Radford et al., 2015; Karras", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 593, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 506, + 606 + ], + "score": 1.0, + "content": "et al., 2018; Brock et al., 2019) or weakly-supervised manner based on semantic attributes (Yan et al.,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "2016; Choi et al., 2018). Empirical findings in (Yan et al., 2016; Zhu et al., 2016a; Radford et al.,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "2015) demonstrated that a simple linear interpolation on the learned image manifold can produce", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 626, + 336, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 336, + 640 + ], + "score": 1.0, + "content": "smooth visual transitions between a pair of input images.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 570, + 506, + 640 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "score": 1.0, + "content": "In this paper, we introduce a novel attack SemanticAdv which generates unrestricted perturbations", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "with semantically meaningful patterns. Motivated by the findings mentioned above, we leverage an", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "attribute-conditional image editing model (Choi et al., 2018) to synthesize adversarial examples by", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "interpolating between source and target images in the feature-map space. Here, we focus on changing", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "a single attribute dimension to achieve adversarial goals while keeping the generated adversarial", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "image reasonably-looking (e.g., see Figure 1). To validate the effectiveness of the proposed attack", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 711, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 721 + ], + "score": 1.0, + "content": "method, we consider two tasks, namely, face verification and landmark detection, as face recognition", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 732 + ], + "score": 1.0, + "content": "field has been extensively explored and the commercially used face models are relatively robust", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 280, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 293 + ], + "score": 1.0, + "content": "since they require a low false positive rate. We conduct both qualitative and quantitative evaluations", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "on CelebA dataset (Liu et al., 2015). To demonstrate the applicability of SemanticAdv beyond", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 303, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 314 + ], + "score": 1.0, + "content": "face domain, we further extend SemanticAdv to generate adversarial street-view images. We treat", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "semantic layouts as input attributes and use the image editing model (Hong et al., 2018) pre-trained", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 324, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 337 + ], + "score": 1.0, + "content": "on Cityscape dataset (Cordts et al., 2016). Please find more visualization results on the anonymous", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 335, + 502, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 502, + 347 + ], + "score": 1.0, + "content": "website: https://sites.google.com/view/generate-semantic-adv-example.", + "type": "text", + "cross_page": true + } + ], + "index": 11 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 644, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 79, + 505, + 222 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 79, + 505, + 222 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 79, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 505, + 222 + ], + "score": 0.973, + "type": "image", + "image_path": "464ae493fc35cd3e8384b15d0c1e64cf82fc55148d4ff781b38e7a7264ac13e9.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 79, + 505, + 126.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 126.66666666666666, + 505, + 174.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 174.33333333333331, + 505, + 221.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 237, + 506, + 270 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 236, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 506, + 250 + ], + "score": 1.0, + "content": "Figure 1: Left: Overview of the proposed SemanticAdv. Right: Illustration of our SemanticAdv in the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "real world face verification platform. Note that the confidence denotes the likelihood that two faces", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 259, + 216, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 216, + 273 + ], + "score": 1.0, + "content": "belong to the same person.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 280, + 505, + 346 + ], + "lines": [ + { + "bbox": [ + 105, + 280, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 293 + ], + "score": 1.0, + "content": "since they require a low false positive rate. We conduct both qualitative and quantitative evaluations", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "on CelebA dataset (Liu et al., 2015). To demonstrate the applicability of SemanticAdv beyond", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 303, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 314 + ], + "score": 1.0, + "content": "face domain, we further extend SemanticAdv to generate adversarial street-view images. We treat", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "semantic layouts as input attributes and use the image editing model (Hong et al., 2018) pre-trained", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 324, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 337 + ], + "score": 1.0, + "content": "on Cityscape dataset (Cordts et al., 2016). Please find more visualization results on the anonymous", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 335, + 502, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 502, + 347 + ], + "score": 1.0, + "content": "website: https://sites.google.com/view/generate-semantic-adv-example.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 505, + 440 + ], + "lines": [ + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "score": 1.0, + "content": "The contributions of the proposed SemanticAdv are three-folds. First, we propose a novel semantic-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "score": 1.0, + "content": "based attack method to generate unrestricted adversarial examples by feature-space interpolation.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 373, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 505, + 387 + ], + "score": 1.0, + "content": "Second, the proposed method is able to generate semantically-controllable perturbations due to the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "attribute-conditioned modeling. This allows us to analyze the robustness of a recognition system", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 396, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 408 + ], + "score": 1.0, + "content": "against different types of semantic attacks. Third, as a side benefit, the proposed attack exhibits", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 241, + 419 + ], + "score": 1.0, + "content": "high transferability and leads to a", + "type": "text" + }, + { + "bbox": [ + 241, + 407, + 261, + 417 + ], + "score": 0.86, + "content": "65 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "query-free black-box attack success rate on a real-world face", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "score": 1.0, + "content": "verification platform, which outperforms the pixel-wise perturbations in attacking existing defense", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 429, + 145, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 145, + 440 + ], + "score": 1.0, + "content": "methods.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 108, + 442, + 212, + 454 + ], + "lines": [ + { + "bbox": [ + 104, + 440, + 213, + 456 + ], + "spans": [ + { + "bbox": [ + 104, + 440, + 213, + 456 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 457, + 506, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "score": 1.0, + "content": "Semantic image editing. Semantic image synthesis and manipulation is a popular research topic in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 468, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 481 + ], + "score": 1.0, + "content": "machine learning, graphics and vision. Thanks to recent advances in deep generative models (Kingma", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 478, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 104, + 478, + 506, + 493 + ], + "score": 1.0, + "content": "& Welling, 2014; Goodfellow et al., 2014a; Oord et al., 2016) and the empirical analysis of deep", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 490, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 506, + 503 + ], + "score": 1.0, + "content": "classification networks (Krizhevsky et al., 2012; Simonyan & Zisserman, 2014; Szegedy et al., 2015),", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 500, + 507, + 515 + ], + "spans": [ + { + "bbox": [ + 104, + 500, + 507, + 515 + ], + "score": 1.0, + "content": "past few years have witnessed tremendous breakthroughs towards high-fidelity pure image genera-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 512, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 506, + 525 + ], + "score": 1.0, + "content": "tion (Radford et al., 2015; Karras et al., 2018; Brock et al., 2019), attribute-to-image generation (Yan", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 522, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 536 + ], + "score": 1.0, + "content": "et al., 2016; Choi et al., 2018), text-to-image generation (Mansimov et al., 2015; Reed et al., 2016;", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 532, + 507, + 548 + ], + "spans": [ + { + "bbox": [ + 104, + 532, + 507, + 548 + ], + "score": 1.0, + "content": "Van den Oord et al., 2016; Odena et al., 2017; Zhang et al., 2017; Johnson et al., 2018), and image-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 544, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 104, + 544, + 506, + 559 + ], + "score": 1.0, + "content": "to-image translation (Isola et al., 2017; Zhu et al., 2017; Liu et al., 2017; Wang et al., 2018b; Hong", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 556, + 158, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 158, + 568 + ], + "score": 1.0, + "content": "et al., 2018).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 261, + 585 + ], + "score": 1.0, + "content": "Adversarial examples. Generating", + "type": "text" + }, + { + "bbox": [ + 261, + 573, + 273, + 584 + ], + "score": 0.87, + "content": "L _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "bounded adversarial perturbation has been extensively", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "studied recently (Szegedy et al., 2013; Goodfellow et al., 2014b; Moosavi-Dezfooli et al., 2016;", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 593, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 104, + 593, + 506, + 608 + ], + "score": 1.0, + "content": "Papernot et al., 2016; Carlini & Wagner, 2017; Xiao et al., 2018b). To further explore diverse", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "adversarial attacks and potentially help inspire defense mechanisms, it is important to generate the so-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "called “unrestricted\" adversarial examples which contain unrestricted magnitude of perturbation while", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 626, + 507, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 507, + 640 + ], + "score": 1.0, + "content": "still preserve perceptual realism Brown et al. (2018). Recently, Xiao et al. (2018c); Engstrom et al.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 638, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 506, + 651 + ], + "score": 1.0, + "content": "(2017) propose to spatially transform the image patches instead of adding pixel-wise perturbation,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "while such spatial transformation does not consider semantic information. Our proposed semanticAdv", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 659, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 104, + 659, + 506, + 674 + ], + "score": 1.0, + "content": "focuses on generating unrestricted perturbation with semantically meaningful patterns guided by", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 671, + 174, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 174, + 683 + ], + "score": 1.0, + "content": "visual attributes.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "Relevant to our work, Song et al. (2018) proposed to synthesize adversarial examples with an", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "unconditional generative model. Bhattad et al. (2019) studied semantic transformation in only", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "the color or texture space. Compared to these works, semanticAdv is able to generate adversarial", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 721, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 734 + ], + "score": 1.0, + "content": "examples in a controllable fashion using specific visual attributes by performing manipulation in the", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 79, + 505, + 222 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 79, + 505, + 222 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 79, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 505, + 222 + ], + "score": 0.973, + "type": "image", + "image_path": "464ae493fc35cd3e8384b15d0c1e64cf82fc55148d4ff781b38e7a7264ac13e9.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 79, + 505, + 126.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 126.66666666666666, + 505, + 174.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 174.33333333333331, + 505, + 221.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 237, + 506, + 270 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 236, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 506, + 250 + ], + "score": 1.0, + "content": "Figure 1: Left: Overview of the proposed SemanticAdv. Right: Illustration of our SemanticAdv in the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "real world face verification platform. Note that the confidence denotes the likelihood that two faces", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 259, + 216, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 216, + 273 + ], + "score": 1.0, + "content": "belong to the same person.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 280, + 505, + 346 + ], + "lines": [], + "index": 8.5, + "bbox_fs": [ + 105, + 280, + 505, + 347 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 505, + 440 + ], + "lines": [ + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "score": 1.0, + "content": "The contributions of the proposed SemanticAdv are three-folds. First, we propose a novel semantic-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "score": 1.0, + "content": "based attack method to generate unrestricted adversarial examples by feature-space interpolation.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 373, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 505, + 387 + ], + "score": 1.0, + "content": "Second, the proposed method is able to generate semantically-controllable perturbations due to the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "attribute-conditioned modeling. This allows us to analyze the robustness of a recognition system", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 396, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 408 + ], + "score": 1.0, + "content": "against different types of semantic attacks. Third, as a side benefit, the proposed attack exhibits", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 241, + 419 + ], + "score": 1.0, + "content": "high transferability and leads to a", + "type": "text" + }, + { + "bbox": [ + 241, + 407, + 261, + 417 + ], + "score": 0.86, + "content": "65 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "query-free black-box attack success rate on a real-world face", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "score": 1.0, + "content": "verification platform, which outperforms the pixel-wise perturbations in attacking existing defense", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 429, + 145, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 145, + 440 + ], + "score": 1.0, + "content": "methods.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 351, + 506, + 440 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 442, + 212, + 454 + ], + "lines": [ + { + "bbox": [ + 104, + 440, + 213, + 456 + ], + "spans": [ + { + "bbox": [ + 104, + 440, + 213, + 456 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 457, + 506, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "score": 1.0, + "content": "Semantic image editing. Semantic image synthesis and manipulation is a popular research topic in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 468, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 481 + ], + "score": 1.0, + "content": "machine learning, graphics and vision. Thanks to recent advances in deep generative models (Kingma", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 478, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 104, + 478, + 506, + 493 + ], + "score": 1.0, + "content": "& Welling, 2014; Goodfellow et al., 2014a; Oord et al., 2016) and the empirical analysis of deep", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 490, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 506, + 503 + ], + "score": 1.0, + "content": "classification networks (Krizhevsky et al., 2012; Simonyan & Zisserman, 2014; Szegedy et al., 2015),", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 500, + 507, + 515 + ], + "spans": [ + { + "bbox": [ + 104, + 500, + 507, + 515 + ], + "score": 1.0, + "content": "past few years have witnessed tremendous breakthroughs towards high-fidelity pure image genera-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 512, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 506, + 525 + ], + "score": 1.0, + "content": "tion (Radford et al., 2015; Karras et al., 2018; Brock et al., 2019), attribute-to-image generation (Yan", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 522, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 536 + ], + "score": 1.0, + "content": "et al., 2016; Choi et al., 2018), text-to-image generation (Mansimov et al., 2015; Reed et al., 2016;", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 532, + 507, + 548 + ], + "spans": [ + { + "bbox": [ + 104, + 532, + 507, + 548 + ], + "score": 1.0, + "content": "Van den Oord et al., 2016; Odena et al., 2017; Zhang et al., 2017; Johnson et al., 2018), and image-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 544, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 104, + 544, + 506, + 559 + ], + "score": 1.0, + "content": "to-image translation (Isola et al., 2017; Zhu et al., 2017; Liu et al., 2017; Wang et al., 2018b; Hong", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 556, + 158, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 158, + 568 + ], + "score": 1.0, + "content": "et al., 2018).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25.5, + "bbox_fs": [ + 104, + 456, + 507, + 568 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 261, + 585 + ], + "score": 1.0, + "content": "Adversarial examples. Generating", + "type": "text" + }, + { + "bbox": [ + 261, + 573, + 273, + 584 + ], + "score": 0.87, + "content": "L _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "bounded adversarial perturbation has been extensively", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "studied recently (Szegedy et al., 2013; Goodfellow et al., 2014b; Moosavi-Dezfooli et al., 2016;", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 593, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 104, + 593, + 506, + 608 + ], + "score": 1.0, + "content": "Papernot et al., 2016; Carlini & Wagner, 2017; Xiao et al., 2018b). To further explore diverse", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "adversarial attacks and potentially help inspire defense mechanisms, it is important to generate the so-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "called “unrestricted\" adversarial examples which contain unrestricted magnitude of perturbation while", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 626, + 507, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 507, + 640 + ], + "score": 1.0, + "content": "still preserve perceptual realism Brown et al. (2018). Recently, Xiao et al. (2018c); Engstrom et al.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 638, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 506, + 651 + ], + "score": 1.0, + "content": "(2017) propose to spatially transform the image patches instead of adding pixel-wise perturbation,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "while such spatial transformation does not consider semantic information. Our proposed semanticAdv", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 659, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 104, + 659, + 506, + 674 + ], + "score": 1.0, + "content": "focuses on generating unrestricted perturbation with semantically meaningful patterns guided by", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 671, + 174, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 174, + 683 + ], + "score": 1.0, + "content": "visual attributes.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 35.5, + "bbox_fs": [ + 104, + 572, + 507, + 683 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "Relevant to our work, Song et al. (2018) proposed to synthesize adversarial examples with an", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "unconditional generative model. Bhattad et al. (2019) studied semantic transformation in only", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "the color or texture space. Compared to these works, semanticAdv is able to generate adversarial", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 721, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 734 + ], + "score": 1.0, + "content": "examples in a controllable fashion using specific visual attributes by performing manipulation in the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "feature space. We further analyze the robustness of the recognition system by generating adversarial", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "examples guided by different visual attributes. Concurrent to our work, Joshi et al. (2019) proposed", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "to generate semantic-based attacks against a restricted binary classifier, while our attack is able to", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 504, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 504, + 127 + ], + "score": 1.0, + "content": "mislead the model towards arbitrary adversarial targets. They conduct the manipulation within the", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 500, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 500, + 140 + ], + "score": 1.0, + "content": "attribution space which is less flexible and effective than our proposed feature-space interpolation.", + "type": "text", + "cross_page": true + } + ], + "index": 4 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 687, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "feature space. We further analyze the robustness of the recognition system by generating adversarial", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "examples guided by different visual attributes. Concurrent to our work, Joshi et al. (2019) proposed", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "to generate semantic-based attacks against a restricted binary classifier, while our attack is able to", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 504, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 504, + 127 + ], + "score": 1.0, + "content": "mislead the model towards arbitrary adversarial targets. They conduct the manipulation within the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 500, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 500, + 140 + ], + "score": 1.0, + "content": "attribution space which is less flexible and effective than our proposed feature-space interpolation.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 108, + 144, + 320, + 156 + ], + "lines": [ + { + "bbox": [ + 104, + 142, + 320, + 159 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 320, + 159 + ], + "score": 1.0, + "content": "3 SEMANTIC ADVERSARIAL EXAMPLES", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 108, + 163, + 227, + 174 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 228, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 228, + 176 + ], + "score": 1.0, + "content": "3.1 PROBLEM DEFINITION", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 180, + 505, + 259 + ], + "lines": [ + { + "bbox": [ + 105, + 180, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 123, + 194 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 182, + 136, + 191 + ], + "score": 0.78, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 180, + 340, + 194 + ], + "score": 1.0, + "content": "be a machine learning model trained on a dataset", + "type": "text" + }, + { + "bbox": [ + 340, + 181, + 398, + 193 + ], + "score": 0.94, + "content": "\\mathcal { D } = \\{ ( { \\bf x } , { \\bf y } ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 180, + 506, + 194 + ], + "score": 1.0, + "content": "consisting of image-label", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 188, + 508, + 207 + ], + "spans": [ + { + "bbox": [ + 104, + 188, + 156, + 207 + ], + "score": 1.0, + "content": "pairs, where", + "type": "text" + }, + { + "bbox": [ + 157, + 191, + 222, + 202 + ], + "score": 0.92, + "content": "\\mathbf { x } \\in \\mathbb { R } ^ { H \\times W \\times D _ { I } }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 188, + 240, + 207 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 240, + 191, + 278, + 203 + ], + "score": 0.93, + "content": "\\mathbf { y } \\in \\mathbb { R } ^ { D _ { L } }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 188, + 508, + 207 + ], + "score": 1.0, + "content": "denote the image and the ground-truth label, respectively.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 203, + 506, + 215 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 130, + 215 + ], + "score": 1.0, + "content": "Here,", + "type": "text" + }, + { + "bbox": [ + 130, + 203, + 156, + 214 + ], + "score": 0.28, + "content": "H , W", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 203, + 159, + 215 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 159, + 203, + 173, + 214 + ], + "score": 0.8, + "content": "D _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 203, + 193, + 215 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 193, + 203, + 209, + 214 + ], + "score": 0.89, + "content": "D _ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 203, + 506, + 215 + ], + "score": 1.0, + "content": "denote the image height, image width, number of image channels, and label", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 212, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 104, + 212, + 276, + 227 + ], + "score": 1.0, + "content": "dimensions, respectively. For each image", + "type": "text" + }, + { + "bbox": [ + 276, + 216, + 284, + 224 + ], + "score": 0.39, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 212, + 331, + 227 + ], + "score": 1.0, + "content": ", our model", + "type": "text" + }, + { + "bbox": [ + 332, + 214, + 344, + 224 + ], + "score": 0.76, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 212, + 425, + 227 + ], + "score": 1.0, + "content": "makes a prediction", + "type": "text" + }, + { + "bbox": [ + 425, + 213, + 503, + 226 + ], + "score": 0.93, + "content": "\\hat { \\mathbf { y } } = \\mathcal { M } ( \\mathbf { x } ) \\in \\mathbb { R } ^ { D _ { L } }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 212, + 506, + 227 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 223, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 237, + 238 + ], + "score": 1.0, + "content": "Given a target image-label pair", + "type": "text" + }, + { + "bbox": [ + 237, + 225, + 277, + 237 + ], + "score": 0.92, + "content": "\\left( \\mathbf { x } ^ { \\mathrm { { t g t } } } , \\mathbf { y } ^ { \\mathrm { { t g t } } } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 223, + 297, + 238 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 297, + 225, + 333, + 236 + ], + "score": 0.92, + "content": "\\mathbf { y } \\neq \\mathbf { y } ^ { \\mathrm { t g t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 223, + 506, + 238 + ], + "score": 1.0, + "content": ", a traditional attacker aims to synthesize", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 234, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 104, + 234, + 196, + 250 + ], + "score": 1.0, + "content": "adversarial examples", + "type": "text" + }, + { + "bbox": [ + 196, + 236, + 223, + 248 + ], + "score": 0.92, + "content": "\\{ { \\bf x } ^ { \\mathrm { a d v } } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 234, + 506, + 250 + ], + "score": 1.0, + "content": "by adding pixel-wise perturbations to or spatially transforming the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 246, + 282, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 214, + 259 + ], + "score": 1.0, + "content": "original image x such that", + "type": "text" + }, + { + "bbox": [ + 214, + 247, + 278, + 259 + ], + "score": 0.92, + "content": "\\dot { \\mathcal { M } } ( \\mathbf { x } ^ { \\mathrm { { a d v } } } ) = \\mathbf { y } ^ { \\mathrm { { t g t } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 246, + 282, + 259 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 505, + 308 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "In this work, we introduce the concept of semantic attacker that aims at generating adversarial", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 275, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 495, + 286 + ], + "score": 1.0, + "content": "examples by adding semantically meaningful perturbation with a conditional generative model", + "type": "text" + }, + { + "bbox": [ + 495, + 275, + 503, + 285 + ], + "score": 0.81, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 275, + 506, + 286 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 285, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 506, + 298 + ], + "score": 1.0, + "content": "Compared to traditional attacker that usually produces pixel-wise perturbations, the proposed method", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 297, + 337, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 337, + 308 + ], + "score": 1.0, + "content": "is able to produce semantically meaningful perturbations.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 318, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 106, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "Semantic image editing. For simplicity, we start with the formulation where the input attribute is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 329, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 341 + ], + "score": 1.0, + "content": "represented as a compact vector. This formulation can be directly extended to other input attribute", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 338, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 104, + 338, + 278, + 353 + ], + "score": 1.0, + "content": "formats including semantic layouts. Let", + "type": "text" + }, + { + "bbox": [ + 278, + 339, + 320, + 350 + ], + "score": 0.91, + "content": "\\mathbf { c } \\in \\mathbb { R } ^ { D _ { c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 338, + 506, + 353 + ], + "score": 1.0, + "content": "be an attribute representation reflecting the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 411, + 364 + ], + "score": 1.0, + "content": "semantic factors (e.g., expression or hair color of a portrait image) of image", + "type": "text" + }, + { + "bbox": [ + 411, + 353, + 419, + 361 + ], + "score": 0.44, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 351, + 449, + 364 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 450, + 351, + 466, + 362 + ], + "score": 0.89, + "content": "D _ { C }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 351, + 505, + 364 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 361, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 216, + 375 + ], + "score": 1.0, + "content": "the attribute dimension and", + "type": "text" + }, + { + "bbox": [ + 216, + 361, + 262, + 374 + ], + "score": 0.93, + "content": "c _ { i } \\in \\{ 0 , 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 361, + 372, + 375 + ], + "score": 1.0, + "content": "indicates the appearance of", + "type": "text" + }, + { + "bbox": [ + 372, + 362, + 376, + 371 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 361, + 506, + 375 + ], + "score": 1.0, + "content": "-th attribute. Here, our goal is to", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 372, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 506, + 386 + ], + "score": 1.0, + "content": "use the conditional generator for semantic image editing. For example, given a portrait image of a girl", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 382, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 506, + 397 + ], + "score": 1.0, + "content": "with black hair and blonde hair as the new attribute, our generator is supposed to synthesize a new", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 394, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 407 + ], + "score": 1.0, + "content": "image that turns the girl’s hair from black to blonde. More specifically, we denote the new attribute as", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 403, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 156, + 416 + ], + "score": 0.92, + "content": "\\mathbf { c } ^ { \\mathrm { { n e w } } } \\in \\mathbb { R } ^ { D _ { C } }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 403, + 329, + 419 + ], + "score": 1.0, + "content": "such that the synthesized image is given by", + "type": "text" + }, + { + "bbox": [ + 329, + 405, + 403, + 417 + ], + "score": 0.93, + "content": "\\mathbf { x } ^ { \\mathrm { n e w } } = \\mathcal { G } ( \\bar { \\mathbf { x } , \\mathbf { c } ^ { \\mathrm { n e w } } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 403, + 506, + 419 + ], + "score": 1.0, + "content": ". In the special case when", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 415, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 222, + 430 + ], + "score": 1.0, + "content": "there is no attribute change", + "type": "text" + }, + { + "bbox": [ + 223, + 417, + 262, + 428 + ], + "score": 0.85, + "content": "( \\mathbf { c } = \\mathbf { c } ^ { \\mathrm { n e w } }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 415, + 451, + 430 + ], + "score": 1.0, + "content": "), the generator simply reconstructs the input:", + "type": "text" + }, + { + "bbox": [ + 451, + 416, + 503, + 428 + ], + "score": 0.93, + "content": "\\mathbf { x } = \\mathcal { G } ( \\mathbf { x } , \\mathbf { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 415, + 506, + 430 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 428, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 506, + 440 + ], + "score": 1.0, + "content": "Supported by the findings mentioned in (Bengio et al., 2013; Reed et al., 2014), our synthesized", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 437, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 104, + 437, + 133, + 451 + ], + "score": 1.0, + "content": "image", + "type": "text" + }, + { + "bbox": [ + 134, + 438, + 153, + 449 + ], + "score": 0.87, + "content": "\\mathbf { x } ^ { \\mathrm { n e w } }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 437, + 506, + 451 + ], + "score": 1.0, + "content": "should fall close to the data manifold if we constrain the change of attribute values to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "score": 1.0, + "content": "be sufficiently small (e.g., we only update one semantic attribute at a time). In addition, we can", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 460, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 474 + ], + "score": 1.0, + "content": "potentially generate many such images by linearly interpolating between the semantic embeddings", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 225, + 484 + ], + "score": 1.0, + "content": "of the conditional generator", + "type": "text" + }, + { + "bbox": [ + 225, + 472, + 234, + 482 + ], + "score": 0.82, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 470, + 324, + 484 + ], + "score": 1.0, + "content": "using original image", + "type": "text" + }, + { + "bbox": [ + 324, + 473, + 332, + 482 + ], + "score": 0.54, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 470, + 446, + 484 + ], + "score": 1.0, + "content": "and the synthesized image", + "type": "text" + }, + { + "bbox": [ + 447, + 471, + 466, + 482 + ], + "score": 0.87, + "content": "\\mathbf { x } ^ { \\mathrm { n e w } }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 470, + 506, + 484 + ], + "score": 1.0, + "content": "with the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 482, + 191, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 191, + 495 + ], + "score": 1.0, + "content": "augmented attribute.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 503, + 505, + 603 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 515 + ], + "score": 1.0, + "content": "Attribute-space interpolation. We start with a simple solution (detailed in Eq. 1) assuming the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "score": 1.0, + "content": "adversarial example can be found by directly interpolating in the attribute-space. Given a pair of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 523, + 506, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 170, + 539 + ], + "score": 1.0, + "content": "attributes c and", + "type": "text" + }, + { + "bbox": [ + 171, + 526, + 188, + 536 + ], + "score": 0.86, + "content": "\\mathbf { c } ^ { \\mathrm { n e w } }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 523, + 354, + 539 + ], + "score": 1.0, + "content": ", we introduce an interpolation parameter", + "type": "text" + }, + { + "bbox": [ + 355, + 525, + 396, + 537 + ], + "score": 0.93, + "content": "\\alpha \\in ( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 523, + 506, + 539 + ], + "score": 1.0, + "content": "to generate the augmented", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 534, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 171, + 550 + ], + "score": 1.0, + "content": "attribute vector", + "type": "text" + }, + { + "bbox": [ + 172, + 536, + 217, + 547 + ], + "score": 0.92, + "content": "\\mathbf { c } ^ { \\ast } \\in \\mathbb { R } ^ { D _ { C } }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 534, + 384, + 550 + ], + "score": 1.0, + "content": "(see Eq. 1). Given augmented attribute", + "type": "text" + }, + { + "bbox": [ + 385, + 537, + 396, + 547 + ], + "score": 0.84, + "content": "\\mathbf { c } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 534, + 478, + 550 + ], + "score": 1.0, + "content": "and original image", + "type": "text" + }, + { + "bbox": [ + 478, + 538, + 486, + 547 + ], + "score": 0.4, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 534, + 506, + 550 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 547, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 304, + 561 + ], + "score": 1.0, + "content": "produce the synthesized image by the generator", + "type": "text" + }, + { + "bbox": [ + 304, + 548, + 312, + 558 + ], + "score": 0.83, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 547, + 506, + 561 + ], + "score": 1.0, + "content": ". For our notation purpose, we also introduce a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 558, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 182, + 572 + ], + "score": 1.0, + "content": "delegated function", + "type": "text" + }, + { + "bbox": [ + 182, + 559, + 194, + 570 + ], + "score": 0.89, + "content": "\\tau _ { \\mathscr { G } }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 558, + 357, + 572 + ], + "score": 1.0, + "content": "as a re-parametrization for the generator", + "type": "text" + }, + { + "bbox": [ + 357, + 559, + 365, + 569 + ], + "score": 0.82, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 558, + 506, + 572 + ], + "score": 1.0, + "content": ". Our formulation is also supported", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "by the empirical results on attribute-conditioned image progression (Yan et al., 2016; Radford et al.,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "score": 1.0, + "content": "2015) that a well-trained generative model has the capability to synthesize a sequence of images with", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 591, + 219, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 219, + 603 + ], + "score": 1.0, + "content": "smooth attribute transitions.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38 + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 628, + 442, + 658 + ], + "lines": [ + { + "bbox": [ + 169, + 628, + 442, + 658 + ], + "spans": [ + { + "bbox": [ + 169, + 628, + 442, + 658 + ], + "score": 0.69, + "content": "\\begin{array} { r l } & { \\mathbf { x } ^ { \\mathrm { a d v } } = \\arg \\operatorname* { m i n } _ { \\alpha } \\mathcal { L } \\big ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) \\big ) } \\\\ & { \\mathrm { w h e r e ~ } \\mathcal { T } _ { \\mathcal { G } } \\big ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } \\big ) = \\mathcal { G } \\big ( \\mathbf { x } , \\mathbf { c } ^ { \\ast } \\big ) \\mathrm { ~ a n d ~ } \\mathbf { c } ^ { \\ast } = \\alpha \\cdot \\mathbf { c } + ( 1 - \\alpha ) \\cdot \\mathbf { c } ^ { \\mathrm { n e w } } } \\end{array}", + "type": "interline_equation", + "image_path": "2cf3fa0d742819a17cd71c8f1c9444d6cdea0b5cdcfc852b7c0a1acebf2c9b24.jpg" + } + ] + } + ], + "index": 44, + "virtual_lines": [ + { + "bbox": [ + 169, + 628, + 442, + 638.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 169, + 638.0, + 442, + 648.0 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 169, + 648.0, + 442, + 658.0 + ], + "spans": [], + "index": 45 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 675, + 507, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 507, + 691 + ], + "score": 1.0, + "content": "Feature-map interpolation. Alternatively, we propose to interpolate using the feature map pro-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 686, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 686, + 203, + 701 + ], + "score": 1.0, + "content": "duced by the generator", + "type": "text" + }, + { + "bbox": [ + 203, + 688, + 267, + 699 + ], + "score": 0.91, + "content": "\\mathcal { G } = \\mathcal { G } _ { \\mathrm { d e c } } \\circ \\mathcal { G } _ { \\mathrm { e n c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 686, + 297, + 701 + ], + "score": 1.0, + "content": ". Here,", + "type": "text" + }, + { + "bbox": [ + 298, + 688, + 315, + 699 + ], + "score": 0.89, + "content": "\\mathcal { G } _ { \\mathrm { e n c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 686, + 506, + 701 + ], + "score": 1.0, + "content": "is the encoder module that takes the image as", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 284, + 711 + ], + "score": 1.0, + "content": "input and outputs the feature map. Similarly,", + "type": "text" + }, + { + "bbox": [ + 284, + 699, + 301, + 710 + ], + "score": 0.9, + "content": "\\mathcal { G } _ { \\mathrm { d e c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "is the decoder module that takes the feature map as", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 103, + 705, + 507, + 725 + ], + "spans": [ + { + "bbox": [ + 103, + 705, + 293, + 725 + ], + "score": 1.0, + "content": "input and outputs the synthesized image. Let", + "type": "text" + }, + { + "bbox": [ + 293, + 709, + 425, + 721 + ], + "score": 0.91, + "content": "\\mathbf { f } ^ { * } = { \\mathcal { G } } _ { \\mathrm { e n c } } ( \\mathbf { x } , \\mathbf { c } ) \\in \\mathbb { R } ^ { H _ { F } \\times W _ { F } \\times C _ { F } }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 705, + 507, + 725 + ], + "score": 1.0, + "content": "be the feature map", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 304, + 734 + ], + "score": 1.0, + "content": "of an intermediate layer in the generator, where", + "type": "text" + }, + { + "bbox": [ + 304, + 721, + 320, + 732 + ], + "score": 0.85, + "content": "H _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 720, + 325, + 734 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 325, + 721, + 342, + 732 + ], + "score": 0.82, + "content": "W _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 720, + 361, + 734 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 362, + 721, + 376, + 732 + ], + "score": 0.86, + "content": "C _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "indicate the height, width, and", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [], + "index": 2, + "bbox_fs": [ + 105, + 83, + 506, + 140 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 144, + 320, + 156 + ], + "lines": [ + { + "bbox": [ + 104, + 142, + 320, + 159 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 320, + 159 + ], + "score": 1.0, + "content": "3 SEMANTIC ADVERSARIAL EXAMPLES", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 108, + 163, + 227, + 174 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 228, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 228, + 176 + ], + "score": 1.0, + "content": "3.1 PROBLEM DEFINITION", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 180, + 505, + 259 + ], + "lines": [ + { + "bbox": [ + 105, + 180, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 123, + 194 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 182, + 136, + 191 + ], + "score": 0.78, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 180, + 340, + 194 + ], + "score": 1.0, + "content": "be a machine learning model trained on a dataset", + "type": "text" + }, + { + "bbox": [ + 340, + 181, + 398, + 193 + ], + "score": 0.94, + "content": "\\mathcal { D } = \\{ ( { \\bf x } , { \\bf y } ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 180, + 506, + 194 + ], + "score": 1.0, + "content": "consisting of image-label", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 188, + 508, + 207 + ], + "spans": [ + { + "bbox": [ + 104, + 188, + 156, + 207 + ], + "score": 1.0, + "content": "pairs, where", + "type": "text" + }, + { + "bbox": [ + 157, + 191, + 222, + 202 + ], + "score": 0.92, + "content": "\\mathbf { x } \\in \\mathbb { R } ^ { H \\times W \\times D _ { I } }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 188, + 240, + 207 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 240, + 191, + 278, + 203 + ], + "score": 0.93, + "content": "\\mathbf { y } \\in \\mathbb { R } ^ { D _ { L } }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 188, + 508, + 207 + ], + "score": 1.0, + "content": "denote the image and the ground-truth label, respectively.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 203, + 506, + 215 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 130, + 215 + ], + "score": 1.0, + "content": "Here,", + "type": "text" + }, + { + "bbox": [ + 130, + 203, + 156, + 214 + ], + "score": 0.28, + "content": "H , W", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 203, + 159, + 215 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 159, + 203, + 173, + 214 + ], + "score": 0.8, + "content": "D _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 203, + 193, + 215 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 193, + 203, + 209, + 214 + ], + "score": 0.89, + "content": "D _ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 203, + 506, + 215 + ], + "score": 1.0, + "content": "denote the image height, image width, number of image channels, and label", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 212, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 104, + 212, + 276, + 227 + ], + "score": 1.0, + "content": "dimensions, respectively. For each image", + "type": "text" + }, + { + "bbox": [ + 276, + 216, + 284, + 224 + ], + "score": 0.39, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 212, + 331, + 227 + ], + "score": 1.0, + "content": ", our model", + "type": "text" + }, + { + "bbox": [ + 332, + 214, + 344, + 224 + ], + "score": 0.76, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 212, + 425, + 227 + ], + "score": 1.0, + "content": "makes a prediction", + "type": "text" + }, + { + "bbox": [ + 425, + 213, + 503, + 226 + ], + "score": 0.93, + "content": "\\hat { \\mathbf { y } } = \\mathcal { M } ( \\mathbf { x } ) \\in \\mathbb { R } ^ { D _ { L } }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 212, + 506, + 227 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 223, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 237, + 238 + ], + "score": 1.0, + "content": "Given a target image-label pair", + "type": "text" + }, + { + "bbox": [ + 237, + 225, + 277, + 237 + ], + "score": 0.92, + "content": "\\left( \\mathbf { x } ^ { \\mathrm { { t g t } } } , \\mathbf { y } ^ { \\mathrm { { t g t } } } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 223, + 297, + 238 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 297, + 225, + 333, + 236 + ], + "score": 0.92, + "content": "\\mathbf { y } \\neq \\mathbf { y } ^ { \\mathrm { t g t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 223, + 506, + 238 + ], + "score": 1.0, + "content": ", a traditional attacker aims to synthesize", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 234, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 104, + 234, + 196, + 250 + ], + "score": 1.0, + "content": "adversarial examples", + "type": "text" + }, + { + "bbox": [ + 196, + 236, + 223, + 248 + ], + "score": 0.92, + "content": "\\{ { \\bf x } ^ { \\mathrm { a d v } } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 234, + 506, + 250 + ], + "score": 1.0, + "content": "by adding pixel-wise perturbations to or spatially transforming the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 246, + 282, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 214, + 259 + ], + "score": 1.0, + "content": "original image x such that", + "type": "text" + }, + { + "bbox": [ + 214, + 247, + 278, + 259 + ], + "score": 0.92, + "content": "\\dot { \\mathcal { M } } ( \\mathbf { x } ^ { \\mathrm { { a d v } } } ) = \\mathbf { y } ^ { \\mathrm { { t g t } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 246, + 282, + 259 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10, + "bbox_fs": [ + 104, + 180, + 508, + 259 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 505, + 308 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "In this work, we introduce the concept of semantic attacker that aims at generating adversarial", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 275, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 495, + 286 + ], + "score": 1.0, + "content": "examples by adding semantically meaningful perturbation with a conditional generative model", + "type": "text" + }, + { + "bbox": [ + 495, + 275, + 503, + 285 + ], + "score": 0.81, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 275, + 506, + 286 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 285, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 506, + 298 + ], + "score": 1.0, + "content": "Compared to traditional attacker that usually produces pixel-wise perturbations, the proposed method", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 297, + 337, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 337, + 308 + ], + "score": 1.0, + "content": "is able to produce semantically meaningful perturbations.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 264, + 506, + 308 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 318, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 106, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "Semantic image editing. For simplicity, we start with the formulation where the input attribute is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 329, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 341 + ], + "score": 1.0, + "content": "represented as a compact vector. This formulation can be directly extended to other input attribute", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 338, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 104, + 338, + 278, + 353 + ], + "score": 1.0, + "content": "formats including semantic layouts. Let", + "type": "text" + }, + { + "bbox": [ + 278, + 339, + 320, + 350 + ], + "score": 0.91, + "content": "\\mathbf { c } \\in \\mathbb { R } ^ { D _ { c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 338, + 506, + 353 + ], + "score": 1.0, + "content": "be an attribute representation reflecting the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 411, + 364 + ], + "score": 1.0, + "content": "semantic factors (e.g., expression or hair color of a portrait image) of image", + "type": "text" + }, + { + "bbox": [ + 411, + 353, + 419, + 361 + ], + "score": 0.44, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 351, + 449, + 364 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 450, + 351, + 466, + 362 + ], + "score": 0.89, + "content": "D _ { C }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 351, + 505, + 364 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 361, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 216, + 375 + ], + "score": 1.0, + "content": "the attribute dimension and", + "type": "text" + }, + { + "bbox": [ + 216, + 361, + 262, + 374 + ], + "score": 0.93, + "content": "c _ { i } \\in \\{ 0 , 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 361, + 372, + 375 + ], + "score": 1.0, + "content": "indicates the appearance of", + "type": "text" + }, + { + "bbox": [ + 372, + 362, + 376, + 371 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 361, + 506, + 375 + ], + "score": 1.0, + "content": "-th attribute. Here, our goal is to", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 372, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 506, + 386 + ], + "score": 1.0, + "content": "use the conditional generator for semantic image editing. For example, given a portrait image of a girl", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 382, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 506, + 397 + ], + "score": 1.0, + "content": "with black hair and blonde hair as the new attribute, our generator is supposed to synthesize a new", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 394, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 407 + ], + "score": 1.0, + "content": "image that turns the girl’s hair from black to blonde. More specifically, we denote the new attribute as", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 403, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 156, + 416 + ], + "score": 0.92, + "content": "\\mathbf { c } ^ { \\mathrm { { n e w } } } \\in \\mathbb { R } ^ { D _ { C } }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 403, + 329, + 419 + ], + "score": 1.0, + "content": "such that the synthesized image is given by", + "type": "text" + }, + { + "bbox": [ + 329, + 405, + 403, + 417 + ], + "score": 0.93, + "content": "\\mathbf { x } ^ { \\mathrm { n e w } } = \\mathcal { G } ( \\bar { \\mathbf { x } , \\mathbf { c } ^ { \\mathrm { n e w } } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 403, + 506, + 419 + ], + "score": 1.0, + "content": ". In the special case when", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 415, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 222, + 430 + ], + "score": 1.0, + "content": "there is no attribute change", + "type": "text" + }, + { + "bbox": [ + 223, + 417, + 262, + 428 + ], + "score": 0.85, + "content": "( \\mathbf { c } = \\mathbf { c } ^ { \\mathrm { n e w } }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 415, + 451, + 430 + ], + "score": 1.0, + "content": "), the generator simply reconstructs the input:", + "type": "text" + }, + { + "bbox": [ + 451, + 416, + 503, + 428 + ], + "score": 0.93, + "content": "\\mathbf { x } = \\mathcal { G } ( \\mathbf { x } , \\mathbf { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 415, + 506, + 430 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 428, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 506, + 440 + ], + "score": 1.0, + "content": "Supported by the findings mentioned in (Bengio et al., 2013; Reed et al., 2014), our synthesized", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 437, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 104, + 437, + 133, + 451 + ], + "score": 1.0, + "content": "image", + "type": "text" + }, + { + "bbox": [ + 134, + 438, + 153, + 449 + ], + "score": 0.87, + "content": "\\mathbf { x } ^ { \\mathrm { n e w } }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 437, + 506, + 451 + ], + "score": 1.0, + "content": "should fall close to the data manifold if we constrain the change of attribute values to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "score": 1.0, + "content": "be sufficiently small (e.g., we only update one semantic attribute at a time). In addition, we can", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 460, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 474 + ], + "score": 1.0, + "content": "potentially generate many such images by linearly interpolating between the semantic embeddings", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 225, + 484 + ], + "score": 1.0, + "content": "of the conditional generator", + "type": "text" + }, + { + "bbox": [ + 225, + 472, + 234, + 482 + ], + "score": 0.82, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 470, + 324, + 484 + ], + "score": 1.0, + "content": "using original image", + "type": "text" + }, + { + "bbox": [ + 324, + 473, + 332, + 482 + ], + "score": 0.54, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 470, + 446, + 484 + ], + "score": 1.0, + "content": "and the synthesized image", + "type": "text" + }, + { + "bbox": [ + 447, + 471, + 466, + 482 + ], + "score": 0.87, + "content": "\\mathbf { x } ^ { \\mathrm { n e w } }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 470, + 506, + 484 + ], + "score": 1.0, + "content": "with the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 482, + 191, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 191, + 495 + ], + "score": 1.0, + "content": "augmented attribute.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 25.5, + "bbox_fs": [ + 104, + 318, + 506, + 495 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 503, + 505, + 603 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 515 + ], + "score": 1.0, + "content": "Attribute-space interpolation. We start with a simple solution (detailed in Eq. 1) assuming the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "score": 1.0, + "content": "adversarial example can be found by directly interpolating in the attribute-space. Given a pair of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 523, + 506, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 170, + 539 + ], + "score": 1.0, + "content": "attributes c and", + "type": "text" + }, + { + "bbox": [ + 171, + 526, + 188, + 536 + ], + "score": 0.86, + "content": "\\mathbf { c } ^ { \\mathrm { n e w } }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 523, + 354, + 539 + ], + "score": 1.0, + "content": ", we introduce an interpolation parameter", + "type": "text" + }, + { + "bbox": [ + 355, + 525, + 396, + 537 + ], + "score": 0.93, + "content": "\\alpha \\in ( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 523, + 506, + 539 + ], + "score": 1.0, + "content": "to generate the augmented", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 534, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 171, + 550 + ], + "score": 1.0, + "content": "attribute vector", + "type": "text" + }, + { + "bbox": [ + 172, + 536, + 217, + 547 + ], + "score": 0.92, + "content": "\\mathbf { c } ^ { \\ast } \\in \\mathbb { R } ^ { D _ { C } }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 534, + 384, + 550 + ], + "score": 1.0, + "content": "(see Eq. 1). Given augmented attribute", + "type": "text" + }, + { + "bbox": [ + 385, + 537, + 396, + 547 + ], + "score": 0.84, + "content": "\\mathbf { c } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 534, + 478, + 550 + ], + "score": 1.0, + "content": "and original image", + "type": "text" + }, + { + "bbox": [ + 478, + 538, + 486, + 547 + ], + "score": 0.4, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 534, + 506, + 550 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 547, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 304, + 561 + ], + "score": 1.0, + "content": "produce the synthesized image by the generator", + "type": "text" + }, + { + "bbox": [ + 304, + 548, + 312, + 558 + ], + "score": 0.83, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 547, + 506, + 561 + ], + "score": 1.0, + "content": ". For our notation purpose, we also introduce a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 558, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 182, + 572 + ], + "score": 1.0, + "content": "delegated function", + "type": "text" + }, + { + "bbox": [ + 182, + 559, + 194, + 570 + ], + "score": 0.89, + "content": "\\tau _ { \\mathscr { G } }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 558, + 357, + 572 + ], + "score": 1.0, + "content": "as a re-parametrization for the generator", + "type": "text" + }, + { + "bbox": [ + 357, + 559, + 365, + 569 + ], + "score": 0.82, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 558, + 506, + 572 + ], + "score": 1.0, + "content": ". Our formulation is also supported", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "by the empirical results on attribute-conditioned image progression (Yan et al., 2016; Radford et al.,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "score": 1.0, + "content": "2015) that a well-trained generative model has the capability to synthesize a sequence of images with", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 591, + 219, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 219, + 603 + ], + "score": 1.0, + "content": "smooth attribute transitions.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 503, + 506, + 603 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 628, + 442, + 658 + ], + "lines": [ + { + "bbox": [ + 169, + 628, + 442, + 658 + ], + "spans": [ + { + "bbox": [ + 169, + 628, + 442, + 658 + ], + "score": 0.69, + "content": "\\begin{array} { r l } & { \\mathbf { x } ^ { \\mathrm { a d v } } = \\arg \\operatorname* { m i n } _ { \\alpha } \\mathcal { L } \\big ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) \\big ) } \\\\ & { \\mathrm { w h e r e ~ } \\mathcal { T } _ { \\mathcal { G } } \\big ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } \\big ) = \\mathcal { G } \\big ( \\mathbf { x } , \\mathbf { c } ^ { \\ast } \\big ) \\mathrm { ~ a n d ~ } \\mathbf { c } ^ { \\ast } = \\alpha \\cdot \\mathbf { c } + ( 1 - \\alpha ) \\cdot \\mathbf { c } ^ { \\mathrm { n e w } } } \\end{array}", + "type": "interline_equation", + "image_path": "2cf3fa0d742819a17cd71c8f1c9444d6cdea0b5cdcfc852b7c0a1acebf2c9b24.jpg" + } + ] + } + ], + "index": 44, + "virtual_lines": [ + { + "bbox": [ + 169, + 628, + 442, + 638.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 169, + 638.0, + 442, + 648.0 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 169, + 648.0, + 442, + 658.0 + ], + "spans": [], + "index": 45 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 675, + 507, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 507, + 691 + ], + "score": 1.0, + "content": "Feature-map interpolation. Alternatively, we propose to interpolate using the feature map pro-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 686, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 686, + 203, + 701 + ], + "score": 1.0, + "content": "duced by the generator", + "type": "text" + }, + { + "bbox": [ + 203, + 688, + 267, + 699 + ], + "score": 0.91, + "content": "\\mathcal { G } = \\mathcal { G } _ { \\mathrm { d e c } } \\circ \\mathcal { G } _ { \\mathrm { e n c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 686, + 297, + 701 + ], + "score": 1.0, + "content": ". Here,", + "type": "text" + }, + { + "bbox": [ + 298, + 688, + 315, + 699 + ], + "score": 0.89, + "content": "\\mathcal { G } _ { \\mathrm { e n c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 686, + 506, + 701 + ], + "score": 1.0, + "content": "is the encoder module that takes the image as", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 284, + 711 + ], + "score": 1.0, + "content": "input and outputs the feature map. Similarly,", + "type": "text" + }, + { + "bbox": [ + 284, + 699, + 301, + 710 + ], + "score": 0.9, + "content": "\\mathcal { G } _ { \\mathrm { d e c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "is the decoder module that takes the feature map as", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 103, + 705, + 507, + 725 + ], + "spans": [ + { + "bbox": [ + 103, + 705, + 293, + 725 + ], + "score": 1.0, + "content": "input and outputs the synthesized image. Let", + "type": "text" + }, + { + "bbox": [ + 293, + 709, + 425, + 721 + ], + "score": 0.91, + "content": "\\mathbf { f } ^ { * } = { \\mathcal { G } } _ { \\mathrm { e n c } } ( \\mathbf { x } , \\mathbf { c } ) \\in \\mathbb { R } ^ { H _ { F } \\times W _ { F } \\times C _ { F } }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 705, + 507, + 725 + ], + "score": 1.0, + "content": "be the feature map", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 304, + 734 + ], + "score": 1.0, + "content": "of an intermediate layer in the generator, where", + "type": "text" + }, + { + "bbox": [ + 304, + 721, + 320, + 732 + ], + "score": 0.85, + "content": "H _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 720, + 325, + 734 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 325, + 721, + 342, + 732 + ], + "score": 0.82, + "content": "W _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 720, + 361, + 734 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 362, + 721, + 376, + 732 + ], + "score": 0.86, + "content": "C _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "indicate the height, width, and", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48, + "bbox_fs": [ + 103, + 675, + 507, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 264, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 265, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 265, + 96 + ], + "score": 1.0, + "content": "number of channels in the feature map.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 206, + 99, + 403, + 143 + ], + "lines": [ + { + "bbox": [ + 206, + 99, + 403, + 143 + ], + "spans": [ + { + "bbox": [ + 206, + 99, + 403, + 143 + ], + "score": 0.89, + "content": "\\begin{array} { r l } & { \\mathbf { x } ^ { \\mathrm { a d v } } = \\arg \\operatorname* { m i n } _ { \\alpha } \\mathcal { L } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ) } \\\\ & { \\mathrm { w h e r e ~ } \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) = \\mathcal { G } _ { \\mathrm { d e c } } ( \\mathbf { f } ^ { * } ) , } \\\\ & { \\mathbf { f } ^ { * } = \\alpha \\odot \\mathcal { G } _ { \\mathrm { e n c } } ( \\mathbf { x } , \\mathbf { c } ) + ( \\mathbf { 1 } - \\alpha ) \\odot \\mathcal { G } _ { \\mathrm { e n c } } ( \\mathbf { x } , \\mathbf { c } ^ { \\mathrm { n e w } } ) } \\end{array}", + "type": "interline_equation", + "image_path": "37bf53fe103985a1cf5894bd65e445513384b94a20c44c690d6ccc10495428c6.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 206, + 99, + 403, + 113.66666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 206, + 113.66666666666667, + 403, + 128.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 206, + 128.33333333333334, + 403, + 143.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 152, + 506, + 218 + ], + "lines": [ + { + "bbox": [ + 106, + 152, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 152, + 506, + 165 + ], + "score": 1.0, + "content": "Compared to attribute-space interpolation which is parameterized by a scalar, we parameterize", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 160, + 508, + 178 + ], + "spans": [ + { + "bbox": [ + 104, + 160, + 263, + 178 + ], + "score": 1.0, + "content": "feature-map interpolation by a tensor", + "type": "text" + }, + { + "bbox": [ + 263, + 163, + 342, + 174 + ], + "score": 0.85, + "content": "\\alpha \\in \\mathbb { R } ^ { H _ { F } \\times W _ { F } \\times C _ { F } }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 160, + 345, + 178 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 346, + 163, + 411, + 176 + ], + "score": 0.85, + "content": "( \\alpha _ { h , w , k } \\in ( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 160, + 444, + 178 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 444, + 163, + 502, + 175 + ], + "score": 0.92, + "content": "1 \\leq h \\leq H _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 160, + 508, + 178 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 174, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 167, + 185 + ], + "score": 0.91, + "content": "1 \\leq w \\leq W _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 174, + 190, + 187 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 190, + 174, + 248, + 185 + ], + "score": 0.91, + "content": "1 \\leq k \\leq C _ { F } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 174, + 505, + 187 + ], + "score": 1.0, + "content": ") with the same shape as the feature map. Compared to linear", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 185, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 505, + 198 + ], + "score": 1.0, + "content": "interpolation over attribute-space, such design introduces more flexibility when interpolating between", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 196, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 505, + 208 + ], + "score": 1.0, + "content": "the original image and the synthesized image. Empirical results in Section 4.2 show our design is", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 207, + 286, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 286, + 219 + ], + "score": 1.0, + "content": "critical to the adversarial attack success rate.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 107, + 225, + 317, + 236 + ], + "lines": [ + { + "bbox": [ + 106, + 224, + 317, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 224, + 317, + 237 + ], + "score": 1.0, + "content": "3.2 ADVERSARIAL OPTIMIZATION OBJECTIVES", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 240, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 104, + 238, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 104, + 238, + 325, + 254 + ], + "score": 1.0, + "content": "As we see in Eq. 3, we obtain the adversarial image", + "type": "text" + }, + { + "bbox": [ + 325, + 240, + 343, + 251 + ], + "score": 0.86, + "content": "\\mathbf { x } ^ { \\mathrm { a d v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 238, + 464, + 254 + ], + "score": 1.0, + "content": "by minimizing the objective", + "type": "text" + }, + { + "bbox": [ + 464, + 240, + 483, + 253 + ], + "score": 0.91, + "content": "\\mathcal { L } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 238, + 506, + 254 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 249, + 507, + 265 + ], + "spans": [ + { + "bbox": [ + 104, + 249, + 236, + 265 + ], + "score": 1.0, + "content": "respect to the synthesized image", + "type": "text" + }, + { + "bbox": [ + 236, + 252, + 303, + 263 + ], + "score": 0.91, + "content": "\\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; { \\mathbf { x } } , { \\mathbf { c } } , { \\mathbf { c } } ^ { \\mathrm { n e w } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 249, + 507, + 265 + ], + "score": 1.0, + "content": ", which is defined in Eq.(1) and Eq.(2) respectively.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 261, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 104, + 261, + 225, + 275 + ], + "score": 1.0, + "content": "Here, each synthesized image", + "type": "text" + }, + { + "bbox": [ + 225, + 263, + 292, + 274 + ], + "score": 0.9, + "content": "\\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; { \\mathbf { x } } , { \\mathbf { c } } , { \\mathbf { c } } ^ { \\mathrm { n e w } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 261, + 506, + 275 + ], + "score": 1.0, + "content": "is produced by the interpolation using the conditional", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 273, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 147, + 286 + ], + "score": 1.0, + "content": "generator", + "type": "text" + }, + { + "bbox": [ + 147, + 274, + 155, + 284 + ], + "score": 0.77, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 273, + 506, + 286 + ], + "score": 1.0, + "content": ". In our objective function, the first term is the adversarial metric, the second term is a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 284, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 217, + 296 + ], + "score": 1.0, + "content": "smoothness constraint, and", + "type": "text" + }, + { + "bbox": [ + 217, + 285, + 225, + 294 + ], + "score": 0.8, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 284, + 505, + 296 + ], + "score": 1.0, + "content": "is used to control the balance between the two terms. The adversarial", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 295, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 249, + 308 + ], + "score": 1.0, + "content": "metric is minimized once the model", + "type": "text" + }, + { + "bbox": [ + 249, + 296, + 262, + 305 + ], + "score": 0.8, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 295, + 506, + 308 + ], + "score": 1.0, + "content": "has been successfully attacked towards the target image-label", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 304, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 104, + 304, + 125, + 320 + ], + "score": 1.0, + "content": "pair", + "type": "text" + }, + { + "bbox": [ + 126, + 306, + 165, + 318 + ], + "score": 0.91, + "content": "\\left( \\mathbf { x } ^ { \\mathrm { { t g t } } } , \\mathbf { y } ^ { \\mathrm { { t g t } } } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 304, + 272, + 320 + ], + "score": 1.0, + "content": ". For identify verification,", + "type": "text" + }, + { + "bbox": [ + 272, + 306, + 288, + 317 + ], + "score": 0.88, + "content": "\\mathbf { y } ^ { \\mathrm { t g t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 304, + 506, + 320 + ], + "score": 1.0, + "content": "is the identity representation of the target image; For", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 317, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 262, + 330 + ], + "score": 1.0, + "content": "structured prediction tasks in our paper,", + "type": "text" + }, + { + "bbox": [ + 263, + 318, + 278, + 329 + ], + "score": 0.89, + "content": "\\mathbf { y } ^ { \\mathrm { t g t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 317, + 506, + 330 + ], + "score": 1.0, + "content": "either represents certain coordinates (landmark detection)", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 328, + 303, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 303, + 341 + ], + "score": 1.0, + "content": "or semantic label maps (semantic segmentation).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 159, + 360, + 452, + 393 + ], + "lines": [ + { + "bbox": [ + 159, + 360, + 452, + 393 + ], + "spans": [ + { + "bbox": [ + 159, + 360, + 452, + 393 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\begin{array} { c } { \\mathbf { x } ^ { \\mathrm { a d v } } = \\mathop { \\mathrm { a r g m i n } _ { \\alpha } } \\mathcal { L } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ) } \\\\ { \\mathcal { L } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ) = \\mathcal { L } _ { \\mathrm { a d v } } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ; \\mathcal { M } , \\mathbf { y } ^ { \\mathrm { t g t } } ) + \\lambda \\cdot \\mathcal { L } _ { \\mathrm { s m o o t h } } ( \\alpha ) } \\end{array} } \\end{array}", + "type": "interline_equation", + "image_path": "1efed1d7a46b0aca0463c4579d9e699f884e9c5703308feb90b77c9b1a7a9eb5.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 159, + 360, + 452, + 371.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 159, + 371.0, + 452, + 382.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 159, + 382.0, + 452, + 393.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 506, + 433 + ], + "lines": [ + { + "bbox": [ + 106, + 401, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 505, + 413 + ], + "score": 1.0, + "content": "Identity verification. In the identity verification task, two images are considered to be the same", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 428, + 424 + ], + "score": 1.0, + "content": "identity if the corresponding identity embeddings from the verification model", + "type": "text" + }, + { + "bbox": [ + 429, + 412, + 442, + 421 + ], + "score": 0.84, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "are reasonably", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 421, + 134, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 134, + 435 + ], + "score": 1.0, + "content": "close.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "interline_equation", + "bbox": [ + 159, + 431, + 451, + 447 + ], + "lines": [ + { + "bbox": [ + 159, + 431, + 451, + 447 + ], + "spans": [ + { + "bbox": [ + 159, + 431, + 451, + 447 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { \\mathrm { a d v } } \\big ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ; \\mathcal { M } , \\mathbf { y } ^ { \\mathrm { t g t } } \\big ) = \\operatorname* { m a x } \\big ( \\kappa , \\Phi _ { \\mathcal { M } } ^ { \\mathrm { i d } } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) , \\mathbf { x } ^ { \\mathrm { t g t } } ) \\big )", + "type": "interline_equation", + "image_path": "8c4c82a4f6f71c490a457568b88e3a93f1f8097911b5a01ce8b2bd52defa3fba.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 159, + 431, + 451, + 447 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 450, + 505, + 494 + ], + "lines": [ + { + "bbox": [ + 104, + 448, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 183, + 464 + ], + "score": 1.0, + "content": "As we see in Eq. 4,", + "type": "text" + }, + { + "bbox": [ + 184, + 449, + 219, + 462 + ], + "score": 0.91, + "content": "\\Phi _ { \\mathcal { M } } ^ { \\mathrm { i d } } ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 448, + 506, + 464 + ], + "score": 1.0, + "content": "measures the distance between two identity embeddings from the model", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 461, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 107, + 462, + 119, + 471 + ], + "score": 0.83, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 461, + 211, + 473 + ], + "score": 1.0, + "content": ", where the normalized", + "type": "text" + }, + { + "bbox": [ + 212, + 462, + 224, + 472 + ], + "score": 0.89, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 461, + 506, + 473 + ], + "score": 1.0, + "content": "distance is used in our setting. In addition, we introduce the parameter", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 107, + 475, + 114, + 482 + ], + "score": 0.74, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 472, + 505, + 484 + ], + "score": 1.0, + "content": "representing the constant related to the false positive rate (FPR) threshold computed from the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 482, + 177, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 177, + 496 + ], + "score": 1.0, + "content": "development set.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 499, + 505, + 544 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 504, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 504, + 511 + ], + "score": 1.0, + "content": "Structured prediction. For structured prediction tasks such as landmark detection and semantic", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 510, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 522 + ], + "score": 1.0, + "content": "segmentation, we use Houdini objective proposed in Cisse et al. (2017) as our adversarial metric and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 350, + 534 + ], + "score": 1.0, + "content": "select the target landmark (semantic segmentation) target as", + "type": "text" + }, + { + "bbox": [ + 350, + 522, + 366, + 533 + ], + "score": 0.86, + "content": "\\mathbf { y } ^ { \\mathrm { t g t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 520, + 419, + 534 + ], + "score": 1.0, + "content": ". In addition,", + "type": "text" + }, + { + "bbox": [ + 420, + 521, + 455, + 533 + ], + "score": 0.93, + "content": "\\Phi _ { \\mathcal { M } } ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 520, + 505, + 534 + ], + "score": 1.0, + "content": "is a scoring", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 532, + 335, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 261, + 544 + ], + "score": 1.0, + "content": "function for each image-label pair and", + "type": "text" + }, + { + "bbox": [ + 261, + 534, + 269, + 544 + ], + "score": 0.82, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 532, + 335, + 544 + ], + "score": 1.0, + "content": "is the threshold.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5 + }, + { + "type": "interline_equation", + "bbox": [ + 139, + 564, + 471, + 596 + ], + "lines": [ + { + "bbox": [ + 139, + 564, + 471, + 596 + ], + "spans": [ + { + "bbox": [ + 139, + 564, + 471, + 596 + ], + "score": 0.9, + "content": "\\begin{array} { r l } & { \\mathcal { L } _ { \\mathrm { a d v } } \\big ( \\mathcal { T } _ { \\mathcal { G } } \\big ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } \\big ) ; \\mathcal { M } , \\mathbf { y } ^ { \\mathrm { t g t } } \\big ) = P _ { \\gamma \\sim \\mathcal { N } ( 0 , 1 ) } \\big [ \\Phi _ { \\mathcal { M } } \\big ( \\mathcal { T } _ { \\mathcal { G } } \\big ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } \\big ) , \\mathbf { y } ^ { * } \\big ) } \\\\ & { \\qquad - \\Phi _ { \\mathcal { M } } \\big ( \\mathcal { T } _ { \\mathcal { G } } \\big ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } \\big ) , \\mathbf { y } ^ { \\mathrm { t g t } } \\big ) < \\gamma \\big ] \\cdot l \\big ( \\mathbf { y } ^ { * } , \\mathbf { y } ^ { \\mathrm { t g t } } \\big ) } \\end{array}", + "type": "interline_equation", + "image_path": "1dd7fd2a01f78abd1aecaa21efc8865ed4c8013a46ddad5fd92dd920a1eb9697.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 139, + 564, + 471, + 574.6666666666666 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 139, + 574.6666666666666, + 471, + 585.3333333333333 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 139, + 585.3333333333333, + 471, + 595.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 628 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 133, + 619 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 605, + 248, + 618 + ], + "score": 0.9, + "content": "\\mathbf { y } ^ { * } = \\mathcal { M } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 603, + 268, + 619 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 268, + 605, + 309, + 618 + ], + "score": 0.93, + "content": "l ( \\mathbf { y } ^ { * } , \\mathbf { y } ^ { \\mathrm { t g t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 603, + 506, + 619 + ], + "score": 1.0, + "content": "is task loss decided by the specific adversarial", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 615, + 136, + 631 + ], + "spans": [ + { + "bbox": [ + 104, + 615, + 136, + 631 + ], + "score": 1.0, + "content": "target.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 631, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 217, + 646 + ], + "score": 1.0, + "content": "Interpolation smoothness", + "type": "text" + }, + { + "bbox": [ + 217, + 633, + 246, + 644 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { \\mathrm { s m o o t h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 631, + 506, + 646 + ], + "score": 1.0, + "content": ". As the tensor to be interpolated in the feature-map space has", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "far more parameters compared to the attribute itself, we propose to enforce a smoothness constraint", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 159, + 668 + ], + "score": 1.0, + "content": "on the tensor", + "type": "text" + }, + { + "bbox": [ + 159, + 657, + 167, + 665 + ], + "score": 0.77, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "used in feature-map interpolation. As we see in Eq. 6, the smoothness loss encourages", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "the interpolation tensors to consist of piece-wise constant patches spatially, which has been widely", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "score": 1.0, + "content": "used as a pixel-wise de-noising objective for natural image processing (Mahendran & Vedaldi, 2015;", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 195, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 195, + 700 + ], + "score": 1.0, + "content": "Johnson et al., 2016).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42.5 + }, + { + "type": "interline_equation", + "bbox": [ + 152, + 703, + 459, + 739 + ], + "lines": [ + { + "bbox": [ + 152, + 703, + 459, + 739 + ], + "spans": [ + { + "bbox": [ + 152, + 703, + 459, + 739 + ], + "score": 0.95, + "content": "\\mathcal { L } _ { \\mathrm { s m o o t h } } ( \\alpha ) = \\sum _ { h = 1 } ^ { H _ { F } - 1 } \\sum _ { w = 1 } ^ { W _ { F } } \\| \\alpha _ { h + 1 , w } - \\alpha _ { h , w } \\| _ { 2 } ^ { 2 } + \\sum _ { h = 1 } ^ { H _ { F } } \\sum _ { w = 1 } ^ { W _ { F } - 1 } \\| \\alpha _ { h , w + 1 } - \\alpha _ { h , w } \\| _ { 2 } ^ { 2 }", + "type": "interline_equation", + "image_path": "9fb15c11a095b5f8eecd931cbd767537b05d0b949e50e726cb865aac1ce83152.jpg" + } + ] + } + ], + "index": 47, + "virtual_lines": [ + { + "bbox": [ + 152, + 703, + 459, + 715.0 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 152, + 715.0, + 459, + 727.0 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 152, + 727.0, + 459, + 739.0 + ], + "spans": [], + "index": 48 + } + ] + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 264, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 265, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 265, + 96 + ], + "score": 1.0, + "content": "number of channels in the feature map.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 81, + 265, + 96 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 206, + 99, + 403, + 143 + ], + "lines": [ + { + "bbox": [ + 206, + 99, + 403, + 143 + ], + "spans": [ + { + "bbox": [ + 206, + 99, + 403, + 143 + ], + "score": 0.89, + "content": "\\begin{array} { r l } & { \\mathbf { x } ^ { \\mathrm { a d v } } = \\arg \\operatorname* { m i n } _ { \\alpha } \\mathcal { L } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ) } \\\\ & { \\mathrm { w h e r e ~ } \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) = \\mathcal { G } _ { \\mathrm { d e c } } ( \\mathbf { f } ^ { * } ) , } \\\\ & { \\mathbf { f } ^ { * } = \\alpha \\odot \\mathcal { G } _ { \\mathrm { e n c } } ( \\mathbf { x } , \\mathbf { c } ) + ( \\mathbf { 1 } - \\alpha ) \\odot \\mathcal { G } _ { \\mathrm { e n c } } ( \\mathbf { x } , \\mathbf { c } ^ { \\mathrm { n e w } } ) } \\end{array}", + "type": "interline_equation", + "image_path": "37bf53fe103985a1cf5894bd65e445513384b94a20c44c690d6ccc10495428c6.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 206, + 99, + 403, + 113.66666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 206, + 113.66666666666667, + 403, + 128.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 206, + 128.33333333333334, + 403, + 143.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 152, + 506, + 218 + ], + "lines": [ + { + "bbox": [ + 106, + 152, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 152, + 506, + 165 + ], + "score": 1.0, + "content": "Compared to attribute-space interpolation which is parameterized by a scalar, we parameterize", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 160, + 508, + 178 + ], + "spans": [ + { + "bbox": [ + 104, + 160, + 263, + 178 + ], + "score": 1.0, + "content": "feature-map interpolation by a tensor", + "type": "text" + }, + { + "bbox": [ + 263, + 163, + 342, + 174 + ], + "score": 0.85, + "content": "\\alpha \\in \\mathbb { R } ^ { H _ { F } \\times W _ { F } \\times C _ { F } }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 160, + 345, + 178 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 346, + 163, + 411, + 176 + ], + "score": 0.85, + "content": "( \\alpha _ { h , w , k } \\in ( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 160, + 444, + 178 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 444, + 163, + 502, + 175 + ], + "score": 0.92, + "content": "1 \\leq h \\leq H _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 160, + 508, + 178 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 174, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 167, + 185 + ], + "score": 0.91, + "content": "1 \\leq w \\leq W _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 174, + 190, + 187 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 190, + 174, + 248, + 185 + ], + "score": 0.91, + "content": "1 \\leq k \\leq C _ { F } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 174, + 505, + 187 + ], + "score": 1.0, + "content": ") with the same shape as the feature map. Compared to linear", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 185, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 505, + 198 + ], + "score": 1.0, + "content": "interpolation over attribute-space, such design introduces more flexibility when interpolating between", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 196, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 505, + 208 + ], + "score": 1.0, + "content": "the original image and the synthesized image. Empirical results in Section 4.2 show our design is", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 207, + 286, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 286, + 219 + ], + "score": 1.0, + "content": "critical to the adversarial attack success rate.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5, + "bbox_fs": [ + 104, + 152, + 508, + 219 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 225, + 317, + 236 + ], + "lines": [ + { + "bbox": [ + 106, + 224, + 317, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 224, + 317, + 237 + ], + "score": 1.0, + "content": "3.2 ADVERSARIAL OPTIMIZATION OBJECTIVES", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 240, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 104, + 238, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 104, + 238, + 325, + 254 + ], + "score": 1.0, + "content": "As we see in Eq. 3, we obtain the adversarial image", + "type": "text" + }, + { + "bbox": [ + 325, + 240, + 343, + 251 + ], + "score": 0.86, + "content": "\\mathbf { x } ^ { \\mathrm { a d v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 238, + 464, + 254 + ], + "score": 1.0, + "content": "by minimizing the objective", + "type": "text" + }, + { + "bbox": [ + 464, + 240, + 483, + 253 + ], + "score": 0.91, + "content": "\\mathcal { L } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 238, + 506, + 254 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 249, + 507, + 265 + ], + "spans": [ + { + "bbox": [ + 104, + 249, + 236, + 265 + ], + "score": 1.0, + "content": "respect to the synthesized image", + "type": "text" + }, + { + "bbox": [ + 236, + 252, + 303, + 263 + ], + "score": 0.91, + "content": "\\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; { \\mathbf { x } } , { \\mathbf { c } } , { \\mathbf { c } } ^ { \\mathrm { n e w } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 249, + 507, + 265 + ], + "score": 1.0, + "content": ", which is defined in Eq.(1) and Eq.(2) respectively.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 261, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 104, + 261, + 225, + 275 + ], + "score": 1.0, + "content": "Here, each synthesized image", + "type": "text" + }, + { + "bbox": [ + 225, + 263, + 292, + 274 + ], + "score": 0.9, + "content": "\\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; { \\mathbf { x } } , { \\mathbf { c } } , { \\mathbf { c } } ^ { \\mathrm { n e w } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 261, + 506, + 275 + ], + "score": 1.0, + "content": "is produced by the interpolation using the conditional", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 273, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 147, + 286 + ], + "score": 1.0, + "content": "generator", + "type": "text" + }, + { + "bbox": [ + 147, + 274, + 155, + 284 + ], + "score": 0.77, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 273, + 506, + 286 + ], + "score": 1.0, + "content": ". In our objective function, the first term is the adversarial metric, the second term is a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 284, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 217, + 296 + ], + "score": 1.0, + "content": "smoothness constraint, and", + "type": "text" + }, + { + "bbox": [ + 217, + 285, + 225, + 294 + ], + "score": 0.8, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 284, + 505, + 296 + ], + "score": 1.0, + "content": "is used to control the balance between the two terms. The adversarial", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 295, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 249, + 308 + ], + "score": 1.0, + "content": "metric is minimized once the model", + "type": "text" + }, + { + "bbox": [ + 249, + 296, + 262, + 305 + ], + "score": 0.8, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 295, + 506, + 308 + ], + "score": 1.0, + "content": "has been successfully attacked towards the target image-label", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 304, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 104, + 304, + 125, + 320 + ], + "score": 1.0, + "content": "pair", + "type": "text" + }, + { + "bbox": [ + 126, + 306, + 165, + 318 + ], + "score": 0.91, + "content": "\\left( \\mathbf { x } ^ { \\mathrm { { t g t } } } , \\mathbf { y } ^ { \\mathrm { { t g t } } } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 304, + 272, + 320 + ], + "score": 1.0, + "content": ". For identify verification,", + "type": "text" + }, + { + "bbox": [ + 272, + 306, + 288, + 317 + ], + "score": 0.88, + "content": "\\mathbf { y } ^ { \\mathrm { t g t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 304, + 506, + 320 + ], + "score": 1.0, + "content": "is the identity representation of the target image; For", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 317, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 262, + 330 + ], + "score": 1.0, + "content": "structured prediction tasks in our paper,", + "type": "text" + }, + { + "bbox": [ + 263, + 318, + 278, + 329 + ], + "score": 0.89, + "content": "\\mathbf { y } ^ { \\mathrm { t g t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 317, + 506, + 330 + ], + "score": 1.0, + "content": "either represents certain coordinates (landmark detection)", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 328, + 303, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 303, + 341 + ], + "score": 1.0, + "content": "or semantic label maps (semantic segmentation).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15, + "bbox_fs": [ + 104, + 238, + 507, + 341 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 159, + 360, + 452, + 393 + ], + "lines": [ + { + "bbox": [ + 159, + 360, + 452, + 393 + ], + "spans": [ + { + "bbox": [ + 159, + 360, + 452, + 393 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\begin{array} { c } { \\mathbf { x } ^ { \\mathrm { a d v } } = \\mathop { \\mathrm { a r g m i n } _ { \\alpha } } \\mathcal { L } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ) } \\\\ { \\mathcal { L } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ) = \\mathcal { L } _ { \\mathrm { a d v } } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ; \\mathcal { M } , \\mathbf { y } ^ { \\mathrm { t g t } } ) + \\lambda \\cdot \\mathcal { L } _ { \\mathrm { s m o o t h } } ( \\alpha ) } \\end{array} } \\end{array}", + "type": "interline_equation", + "image_path": "1efed1d7a46b0aca0463c4579d9e699f884e9c5703308feb90b77c9b1a7a9eb5.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 159, + 360, + 452, + 371.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 159, + 371.0, + 452, + 382.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 159, + 382.0, + 452, + 393.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 506, + 433 + ], + "lines": [ + { + "bbox": [ + 106, + 401, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 505, + 413 + ], + "score": 1.0, + "content": "Identity verification. In the identity verification task, two images are considered to be the same", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 428, + 424 + ], + "score": 1.0, + "content": "identity if the corresponding identity embeddings from the verification model", + "type": "text" + }, + { + "bbox": [ + 429, + 412, + 442, + 421 + ], + "score": 0.84, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "are reasonably", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 421, + 134, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 134, + 435 + ], + "score": 1.0, + "content": "close.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 401, + 505, + 435 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 159, + 431, + 451, + 447 + ], + "lines": [ + { + "bbox": [ + 159, + 431, + 451, + 447 + ], + "spans": [ + { + "bbox": [ + 159, + 431, + 451, + 447 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { \\mathrm { a d v } } \\big ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ; \\mathcal { M } , \\mathbf { y } ^ { \\mathrm { t g t } } \\big ) = \\operatorname* { m a x } \\big ( \\kappa , \\Phi _ { \\mathcal { M } } ^ { \\mathrm { i d } } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) , \\mathbf { x } ^ { \\mathrm { t g t } } ) \\big )", + "type": "interline_equation", + "image_path": "8c4c82a4f6f71c490a457568b88e3a93f1f8097911b5a01ce8b2bd52defa3fba.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 159, + 431, + 451, + 447 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 450, + 505, + 494 + ], + "lines": [ + { + "bbox": [ + 104, + 448, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 183, + 464 + ], + "score": 1.0, + "content": "As we see in Eq. 4,", + "type": "text" + }, + { + "bbox": [ + 184, + 449, + 219, + 462 + ], + "score": 0.91, + "content": "\\Phi _ { \\mathcal { M } } ^ { \\mathrm { i d } } ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 448, + 506, + 464 + ], + "score": 1.0, + "content": "measures the distance between two identity embeddings from the model", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 461, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 107, + 462, + 119, + 471 + ], + "score": 0.83, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 461, + 211, + 473 + ], + "score": 1.0, + "content": ", where the normalized", + "type": "text" + }, + { + "bbox": [ + 212, + 462, + 224, + 472 + ], + "score": 0.89, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 461, + 506, + 473 + ], + "score": 1.0, + "content": "distance is used in our setting. In addition, we introduce the parameter", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 107, + 475, + 114, + 482 + ], + "score": 0.74, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 472, + 505, + 484 + ], + "score": 1.0, + "content": "representing the constant related to the false positive rate (FPR) threshold computed from the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 482, + 177, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 177, + 496 + ], + "score": 1.0, + "content": "development set.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 104, + 448, + 506, + 496 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 499, + 505, + 544 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 504, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 504, + 511 + ], + "score": 1.0, + "content": "Structured prediction. For structured prediction tasks such as landmark detection and semantic", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 510, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 522 + ], + "score": 1.0, + "content": "segmentation, we use Houdini objective proposed in Cisse et al. (2017) as our adversarial metric and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 350, + 534 + ], + "score": 1.0, + "content": "select the target landmark (semantic segmentation) target as", + "type": "text" + }, + { + "bbox": [ + 350, + 522, + 366, + 533 + ], + "score": 0.86, + "content": "\\mathbf { y } ^ { \\mathrm { t g t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 520, + 419, + 534 + ], + "score": 1.0, + "content": ". In addition,", + "type": "text" + }, + { + "bbox": [ + 420, + 521, + 455, + 533 + ], + "score": 0.93, + "content": "\\Phi _ { \\mathcal { M } } ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 520, + 505, + 534 + ], + "score": 1.0, + "content": "is a scoring", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 532, + 335, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 261, + 544 + ], + "score": 1.0, + "content": "function for each image-label pair and", + "type": "text" + }, + { + "bbox": [ + 261, + 534, + 269, + 544 + ], + "score": 0.82, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 532, + 335, + 544 + ], + "score": 1.0, + "content": "is the threshold.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 500, + 505, + 544 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 139, + 564, + 471, + 596 + ], + "lines": [ + { + "bbox": [ + 139, + 564, + 471, + 596 + ], + "spans": [ + { + "bbox": [ + 139, + 564, + 471, + 596 + ], + "score": 0.9, + "content": "\\begin{array} { r l } & { \\mathcal { L } _ { \\mathrm { a d v } } \\big ( \\mathcal { T } _ { \\mathcal { G } } \\big ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } \\big ) ; \\mathcal { M } , \\mathbf { y } ^ { \\mathrm { t g t } } \\big ) = P _ { \\gamma \\sim \\mathcal { N } ( 0 , 1 ) } \\big [ \\Phi _ { \\mathcal { M } } \\big ( \\mathcal { T } _ { \\mathcal { G } } \\big ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } \\big ) , \\mathbf { y } ^ { * } \\big ) } \\\\ & { \\qquad - \\Phi _ { \\mathcal { M } } \\big ( \\mathcal { T } _ { \\mathcal { G } } \\big ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } \\big ) , \\mathbf { y } ^ { \\mathrm { t g t } } \\big ) < \\gamma \\big ] \\cdot l \\big ( \\mathbf { y } ^ { * } , \\mathbf { y } ^ { \\mathrm { t g t } } \\big ) } \\end{array}", + "type": "interline_equation", + "image_path": "1dd7fd2a01f78abd1aecaa21efc8865ed4c8013a46ddad5fd92dd920a1eb9697.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 139, + 564, + 471, + 574.6666666666666 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 139, + 574.6666666666666, + 471, + 585.3333333333333 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 139, + 585.3333333333333, + 471, + 595.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 628 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 133, + 619 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 605, + 248, + 618 + ], + "score": 0.9, + "content": "\\mathbf { y } ^ { * } = \\mathcal { M } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 603, + 268, + 619 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 268, + 605, + 309, + 618 + ], + "score": 0.93, + "content": "l ( \\mathbf { y } ^ { * } , \\mathbf { y } ^ { \\mathrm { t g t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 603, + 506, + 619 + ], + "score": 1.0, + "content": "is task loss decided by the specific adversarial", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 615, + 136, + 631 + ], + "spans": [ + { + "bbox": [ + 104, + 615, + 136, + 631 + ], + "score": 1.0, + "content": "target.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 104, + 603, + 506, + 631 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 631, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 217, + 646 + ], + "score": 1.0, + "content": "Interpolation smoothness", + "type": "text" + }, + { + "bbox": [ + 217, + 633, + 246, + 644 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { \\mathrm { s m o o t h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 631, + 506, + 646 + ], + "score": 1.0, + "content": ". As the tensor to be interpolated in the feature-map space has", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "far more parameters compared to the attribute itself, we propose to enforce a smoothness constraint", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 159, + 668 + ], + "score": 1.0, + "content": "on the tensor", + "type": "text" + }, + { + "bbox": [ + 159, + 657, + 167, + 665 + ], + "score": 0.77, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "used in feature-map interpolation. As we see in Eq. 6, the smoothness loss encourages", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "the interpolation tensors to consist of piece-wise constant patches spatially, which has been widely", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "score": 1.0, + "content": "used as a pixel-wise de-noising objective for natural image processing (Mahendran & Vedaldi, 2015;", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 195, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 195, + 700 + ], + "score": 1.0, + "content": "Johnson et al., 2016).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 631, + 506, + 700 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 152, + 703, + 459, + 739 + ], + "lines": [ + { + "bbox": [ + 152, + 703, + 459, + 739 + ], + "spans": [ + { + "bbox": [ + 152, + 703, + 459, + 739 + ], + "score": 0.95, + "content": "\\mathcal { L } _ { \\mathrm { s m o o t h } } ( \\alpha ) = \\sum _ { h = 1 } ^ { H _ { F } - 1 } \\sum _ { w = 1 } ^ { W _ { F } } \\| \\alpha _ { h + 1 , w } - \\alpha _ { h , w } \\| _ { 2 } ^ { 2 } + \\sum _ { h = 1 } ^ { H _ { F } } \\sum _ { w = 1 } ^ { W _ { F } - 1 } \\| \\alpha _ { h , w + 1 } - \\alpha _ { h , w } \\| _ { 2 } ^ { 2 }", + "type": "interline_equation", + "image_path": "9fb15c11a095b5f8eecd931cbd767537b05d0b949e50e726cb865aac1ce83152.jpg" + } + ] + } + ], + "index": 47, + "virtual_lines": [ + { + "bbox": [ + 152, + 703, + 459, + 715.0 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 152, + 715.0, + 459, + 727.0 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 152, + 727.0, + 459, + 739.0 + ], + "spans": [], + "index": 48 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 200, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 201, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 201, + 95 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 98, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 96, + 506, + 111 + ], + "spans": [ + { + "bbox": [ + 105, + 96, + 506, + 111 + ], + "score": 1.0, + "content": "In the experimental section, we mainly focus on analyzing the proposed SemanticAdv in attacking", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 108, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 108, + 505, + 123 + ], + "score": 1.0, + "content": "state-of-the-art face recognition systems on CelebA (Liu et al., 2015) due to its wide applicability", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 505, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 505, + 132 + ], + "score": 1.0, + "content": "(e.g., identification for mobile payment) in the real world. In addition, we extend our attack to urban", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 142 + ], + "score": 1.0, + "content": "street scenes with semantic label maps as the condition. We attack the semantic segmentation model", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 140, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 505, + 155 + ], + "score": 1.0, + "content": "DRN-D-22 (Yu et al., 2017) previously trained on Cityscape (Cordts et al., 2016) by generating", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 464, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 464, + 165 + ], + "score": 1.0, + "content": "adversarial examples with dynamic objects manipulated (e.g., insert a car into the scene).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 169, + 505, + 269 + ], + "lines": [ + { + "bbox": [ + 106, + 170, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 505, + 182 + ], + "score": 1.0, + "content": "The experimental section is organized as follows. First, we analyze the quality of generated adversarial", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 180, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 327, + 194 + ], + "score": 1.0, + "content": "examples and qualitatively compare our method with", + "type": "text" + }, + { + "bbox": [ + 327, + 181, + 340, + 193 + ], + "score": 0.87, + "content": "{ \\mathcal { L } } _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 180, + 506, + 194 + ], + "score": 1.0, + "content": "bounded pixel-wise optimization-based", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "score": 1.0, + "content": "method (Carlini & Wagner, 2017; Dong et al., 2018; Xie et al., 2019). Second, we provide both", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "score": 1.0, + "content": "qualitative and quantitative results by controlling each of the semantic attributes at a time. In terms", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 214, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 506, + 227 + ], + "score": 1.0, + "content": "of attack transferability, we evaluate our proposed SemanticAdv on various settings and further", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 224, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 505, + 238 + ], + "score": 1.0, + "content": "demonstrate the effectiveness of our method via query-free black-box attacks against online face", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 235, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 506, + 249 + ], + "score": 1.0, + "content": "verification platforms. Third, we compare our method with the baseline against different defense", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "methods on the face verification task. Fourth, we demonstrate that the proposed SemanticAdv also", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 257, + 419, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 419, + 271 + ], + "score": 1.0, + "content": "applies to the face landmark detection and street-view semantic segmentation.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 277, + 230, + 287 + ], + "lines": [ + { + "bbox": [ + 106, + 276, + 231, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 231, + 289 + ], + "score": 1.0, + "content": "4.1 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 290, + 505, + 455 + ], + "lines": [ + { + "bbox": [ + 106, + 290, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 505, + 302 + ], + "score": 1.0, + "content": "Face identity verification. We select ResNet-50 and ResNet-101 (He et al., 2016) trained", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 300, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 506, + 315 + ], + "score": 1.0, + "content": "on MS-Celeb-1M (Guo et al., 2016) as our face verification models. The models are trained using", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "two different objectives, namely, softmax loss (Sun et al., 2014; Zhang et al., 2018) and cosine", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 323, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 505, + 335 + ], + "score": 1.0, + "content": "loss (Wang et al., 2018a). For simplicity, we use the notation “R-N-S” to indicate the model with", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 107, + 334, + 117, + 344 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 333, + 506, + 347 + ], + "score": 1.0, + "content": "-layer residual blocks as backbone trained using softmax loss, while “R-N-C” indicates the same", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 345, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 357 + ], + "score": 1.0, + "content": "backbone trained using cosine loss. The distance between face features is measured by normalized", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 355, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 343, + 368 + ], + "score": 1.0, + "content": "L2 distance. For R-101-S model, we decide the parameter", + "type": "text" + }, + { + "bbox": [ + 344, + 358, + 351, + 366 + ], + "score": 0.75, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 355, + 506, + 368 + ], + "score": 1.0, + "content": "based on the false positive rate (FPR)", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 365, + 507, + 380 + ], + "spans": [ + { + "bbox": [ + 104, + 365, + 408, + 380 + ], + "score": 1.0, + "content": "for the identity verification task. Three different FPRs have been used:", + "type": "text" + }, + { + "bbox": [ + 408, + 366, + 430, + 378 + ], + "score": 0.89, + "content": "1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 365, + 457, + 380 + ], + "score": 1.0, + "content": "(with", + "type": "text" + }, + { + "bbox": [ + 457, + 367, + 500, + 378 + ], + "score": 0.87, + "content": "\\kappa = 1 . 2 4 ,", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 365, + 507, + 380 + ], + "score": 1.0, + "content": "),", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 376, + 507, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 145, + 389 + ], + "score": 0.91, + "content": "3 \\times 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 376, + 170, + 390 + ], + "score": 1.0, + "content": "(with", + "type": "text" + }, + { + "bbox": [ + 170, + 378, + 209, + 389 + ], + "score": 0.88, + "content": "\\kappa = 1 . 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 376, + 232, + 390 + ], + "score": 1.0, + "content": "), and", + "type": "text" + }, + { + "bbox": [ + 232, + 377, + 254, + 388 + ], + "score": 0.91, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 376, + 279, + 390 + ], + "score": 1.0, + "content": "(with", + "type": "text" + }, + { + "bbox": [ + 279, + 378, + 317, + 389 + ], + "score": 0.87, + "content": "\\kappa = 0 . 6 0 ", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 376, + 507, + 390 + ], + "score": 1.0, + "content": "). The distance metrics and selected thresholds", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 389, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 506, + 401 + ], + "score": 1.0, + "content": "are commonly used when evaluating the performance of face recognition model Klare et al. (2015);", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 399, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 506, + 412 + ], + "score": 1.0, + "content": "Kemelmacher-Shlizerman et al. (2016). Please check the Appendix (see Table B) for more details.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 411, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 506, + 423 + ], + "score": 1.0, + "content": "To distinguish between the FPR we used in generating adversarial examples and the other FPR used", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 421, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 434 + ], + "score": 1.0, + "content": "in evaluation, we introduce two notations “Generating FPR (G-FPR)” and “Test FPR (T-FPR)”. For", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 433, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 505, + 444 + ], + "score": 1.0, + "content": "the experiment with query-free black-box API attacks, we use the online face verification services", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 444, + 280, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 156, + 456 + ], + "score": 1.0, + "content": "provided by", + "type": "text" + }, + { + "bbox": [ + 156, + 444, + 189, + 455 + ], + "score": 0.32, + "content": "{ \\mathrm { F a c e } } + +", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 444, + 280, + 456 + ], + "score": 1.0, + "content": "(fac) and AliYun (ali).", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 461, + 505, + 516 + ], + "lines": [ + { + "bbox": [ + 106, + 459, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 506, + 474 + ], + "score": 1.0, + "content": "Face landmark detection. We select Face Alignment Network (FAN) (Bulat & Tzimiropoulos,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 471, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 505, + 484 + ], + "score": 1.0, + "content": "2017b) trained on 300W-LP (Zhu et al., 2016b) and fine-tuned on 300-W (Sagonas et al., 2013) for", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 481, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 506, + 496 + ], + "score": 1.0, + "content": "2D landmark detection. The network is constructed by stacking Hour-Glass network (Newell et al.,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 493, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 506, + 507 + ], + "score": 1.0, + "content": "2016) with hierarchical block (Bulat & Tzimiropoulos, 2017a). Given a portrait image as input, FAN", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 505, + 449, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 449, + 518 + ], + "score": 1.0, + "content": "outputs 2D heatmaps which can be subsequently leveraged to yield 68 2D landmarks.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 522, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 506, + 535 + ], + "score": 1.0, + "content": "Semantic attacks on face images. In our experiments, we randomly sample 1, 280 distinct identi-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 531, + 507, + 548 + ], + "spans": [ + { + "bbox": [ + 104, + 531, + 507, + 548 + ], + "score": 1.0, + "content": "ties form CelebA (Liu et al., 2015). To reduce the reconstruction error brought by the generator (e.g.,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 544, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 159, + 556 + ], + "score": 0.9, + "content": "\\mathbf { x } \\neq \\mathcal { G } ( \\mathbf { x } , \\mathbf { c } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 544, + 438, + 558 + ], + "score": 1.0, + "content": "in practice, we take one more step to obtain the updated feature map", + "type": "text" + }, + { + "bbox": [ + 439, + 544, + 502, + 556 + ], + "score": 0.92, + "content": "\\mathbf { f } ^ { \\prime } = \\mathcal { G } _ { \\mathrm { e n c } } ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 544, + 506, + 558 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 555, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 134, + 568 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 556, + 198, + 567 + ], + "score": 0.81, + "content": "\\mathbf { x } ^ { \\prime } = \\mathrm { a r g m i n } _ { \\mathbf { x } ^ { \\prime } }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 555, + 263, + 567 + ], + "score": 0.91, + "content": "\\| \\mathcal { G } ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } ) - \\mathbf { x } \\|", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 555, + 506, + 568 + ], + "score": 1.0, + "content": "in feature-map interpolation. In our experiments, we use", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "the last conv layer before upsampling in the generator as our as feature-map f given by the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 578, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 293, + 588 + ], + "score": 1.0, + "content": "attack effectiveness. We also fix the parameter", + "type": "text" + }, + { + "bbox": [ + 293, + 578, + 301, + 587 + ], + "score": 0.71, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 578, + 505, + 588 + ], + "score": 1.0, + "content": "(e.g., balances the adversarial loss and smoothness", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 588, + 427, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 427, + 600 + ], + "score": 1.0, + "content": "constraint in Eq. 3) to be 0.01 for both face verification and landmark detection.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "We used the StarGAN (Choi et al., 2018) for attribute-conditional image editing. In particular, we", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "re-trained model on CelebA dataset (Liu et al., 2015) by aligning the face landmarks and then resizing", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 190, + 639 + ], + "score": 1.0, + "content": "images to resolution", + "type": "text" + }, + { + "bbox": [ + 190, + 627, + 233, + 637 + ], + "score": 0.89, + "content": "1 1 2 \\times 1 1 2", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 627, + 505, + 639 + ], + "score": 1.0, + "content": ". In addition, we select 17 identity-preserving attributes as our input", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 639, + 389, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 389, + 650 + ], + "score": 1.0, + "content": "condition, as such attributes related to facial expression and hair color.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45.5 + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 226, + 668 + ], + "score": 1.0, + "content": "For each distinct identity pair", + "type": "text" + }, + { + "bbox": [ + 227, + 655, + 259, + 667 + ], + "score": 0.92, + "content": "( \\mathbf { x } , \\mathbf { x } ^ { \\mathrm { t g t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 654, + 506, + 668 + ], + "score": 1.0, + "content": ", we perform semanticAdv guided by each of the 17 attributes", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "(e.g., we intentionally add or remove one specific attribute while keeping the rest unchanged). In", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 189, + 690 + ], + "score": 1.0, + "content": "total, for each image", + "type": "text" + }, + { + "bbox": [ + 189, + 679, + 197, + 687 + ], + "score": 0.27, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 677, + 505, + 690 + ], + "score": 1.0, + "content": ", we generate 17 adversarial images with different augmented attributes. In the", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "experiments, we select a pixel-wise adversarial attack method (Carlini & Wagner, 2017) (referred", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "as CW) as our baseline for comparison. Compared to our proposed method, CW does not require", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 710, + 507, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 507, + 722 + ], + "score": 1.0, + "content": "visual attributes as part of the system, as it only generates one adversarial example for each instance.", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 721, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 506, + 733 + ], + "score": 1.0, + "content": "We refer the corresponding attack success rate as the instance-wise success rate in which the attack", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 51 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 200, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 201, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 201, + 95 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 98, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 96, + 506, + 111 + ], + "spans": [ + { + "bbox": [ + 105, + 96, + 506, + 111 + ], + "score": 1.0, + "content": "In the experimental section, we mainly focus on analyzing the proposed SemanticAdv in attacking", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 108, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 108, + 505, + 123 + ], + "score": 1.0, + "content": "state-of-the-art face recognition systems on CelebA (Liu et al., 2015) due to its wide applicability", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 505, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 505, + 132 + ], + "score": 1.0, + "content": "(e.g., identification for mobile payment) in the real world. In addition, we extend our attack to urban", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 142 + ], + "score": 1.0, + "content": "street scenes with semantic label maps as the condition. We attack the semantic segmentation model", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 140, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 505, + 155 + ], + "score": 1.0, + "content": "DRN-D-22 (Yu et al., 2017) previously trained on Cityscape (Cordts et al., 2016) by generating", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 464, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 464, + 165 + ], + "score": 1.0, + "content": "adversarial examples with dynamic objects manipulated (e.g., insert a car into the scene).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 96, + 506, + 165 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 169, + 505, + 269 + ], + "lines": [ + { + "bbox": [ + 106, + 170, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 505, + 182 + ], + "score": 1.0, + "content": "The experimental section is organized as follows. First, we analyze the quality of generated adversarial", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 180, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 327, + 194 + ], + "score": 1.0, + "content": "examples and qualitatively compare our method with", + "type": "text" + }, + { + "bbox": [ + 327, + 181, + 340, + 193 + ], + "score": 0.87, + "content": "{ \\mathcal { L } } _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 180, + 506, + 194 + ], + "score": 1.0, + "content": "bounded pixel-wise optimization-based", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "score": 1.0, + "content": "method (Carlini & Wagner, 2017; Dong et al., 2018; Xie et al., 2019). Second, we provide both", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "score": 1.0, + "content": "qualitative and quantitative results by controlling each of the semantic attributes at a time. In terms", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 214, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 506, + 227 + ], + "score": 1.0, + "content": "of attack transferability, we evaluate our proposed SemanticAdv on various settings and further", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 224, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 505, + 238 + ], + "score": 1.0, + "content": "demonstrate the effectiveness of our method via query-free black-box attacks against online face", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 235, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 506, + 249 + ], + "score": 1.0, + "content": "verification platforms. Third, we compare our method with the baseline against different defense", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "methods on the face verification task. Fourth, we demonstrate that the proposed SemanticAdv also", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 257, + 419, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 419, + 271 + ], + "score": 1.0, + "content": "applies to the face landmark detection and street-view semantic segmentation.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 170, + 506, + 271 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 277, + 230, + 287 + ], + "lines": [ + { + "bbox": [ + 106, + 276, + 231, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 231, + 289 + ], + "score": 1.0, + "content": "4.1 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 290, + 505, + 455 + ], + "lines": [ + { + "bbox": [ + 106, + 290, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 505, + 302 + ], + "score": 1.0, + "content": "Face identity verification. We select ResNet-50 and ResNet-101 (He et al., 2016) trained", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 300, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 506, + 315 + ], + "score": 1.0, + "content": "on MS-Celeb-1M (Guo et al., 2016) as our face verification models. The models are trained using", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "two different objectives, namely, softmax loss (Sun et al., 2014; Zhang et al., 2018) and cosine", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 323, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 505, + 335 + ], + "score": 1.0, + "content": "loss (Wang et al., 2018a). For simplicity, we use the notation “R-N-S” to indicate the model with", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 107, + 334, + 117, + 344 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 333, + 506, + 347 + ], + "score": 1.0, + "content": "-layer residual blocks as backbone trained using softmax loss, while “R-N-C” indicates the same", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 345, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 357 + ], + "score": 1.0, + "content": "backbone trained using cosine loss. The distance between face features is measured by normalized", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 355, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 343, + 368 + ], + "score": 1.0, + "content": "L2 distance. For R-101-S model, we decide the parameter", + "type": "text" + }, + { + "bbox": [ + 344, + 358, + 351, + 366 + ], + "score": 0.75, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 355, + 506, + 368 + ], + "score": 1.0, + "content": "based on the false positive rate (FPR)", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 365, + 507, + 380 + ], + "spans": [ + { + "bbox": [ + 104, + 365, + 408, + 380 + ], + "score": 1.0, + "content": "for the identity verification task. Three different FPRs have been used:", + "type": "text" + }, + { + "bbox": [ + 408, + 366, + 430, + 378 + ], + "score": 0.89, + "content": "1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 365, + 457, + 380 + ], + "score": 1.0, + "content": "(with", + "type": "text" + }, + { + "bbox": [ + 457, + 367, + 500, + 378 + ], + "score": 0.87, + "content": "\\kappa = 1 . 2 4 ,", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 365, + 507, + 380 + ], + "score": 1.0, + "content": "),", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 376, + 507, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 145, + 389 + ], + "score": 0.91, + "content": "3 \\times 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 376, + 170, + 390 + ], + "score": 1.0, + "content": "(with", + "type": "text" + }, + { + "bbox": [ + 170, + 378, + 209, + 389 + ], + "score": 0.88, + "content": "\\kappa = 1 . 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 376, + 232, + 390 + ], + "score": 1.0, + "content": "), and", + "type": "text" + }, + { + "bbox": [ + 232, + 377, + 254, + 388 + ], + "score": 0.91, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 376, + 279, + 390 + ], + "score": 1.0, + "content": "(with", + "type": "text" + }, + { + "bbox": [ + 279, + 378, + 317, + 389 + ], + "score": 0.87, + "content": "\\kappa = 0 . 6 0 ", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 376, + 507, + 390 + ], + "score": 1.0, + "content": "). The distance metrics and selected thresholds", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 389, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 506, + 401 + ], + "score": 1.0, + "content": "are commonly used when evaluating the performance of face recognition model Klare et al. (2015);", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 399, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 506, + 412 + ], + "score": 1.0, + "content": "Kemelmacher-Shlizerman et al. (2016). Please check the Appendix (see Table B) for more details.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 411, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 506, + 423 + ], + "score": 1.0, + "content": "To distinguish between the FPR we used in generating adversarial examples and the other FPR used", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 421, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 434 + ], + "score": 1.0, + "content": "in evaluation, we introduce two notations “Generating FPR (G-FPR)” and “Test FPR (T-FPR)”. For", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 433, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 505, + 444 + ], + "score": 1.0, + "content": "the experiment with query-free black-box API attacks, we use the online face verification services", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 444, + 280, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 156, + 456 + ], + "score": 1.0, + "content": "provided by", + "type": "text" + }, + { + "bbox": [ + 156, + 444, + 189, + 455 + ], + "score": 0.32, + "content": "{ \\mathrm { F a c e } } + +", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 444, + 280, + 456 + ], + "score": 1.0, + "content": "(fac) and AliYun (ali).", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 24, + "bbox_fs": [ + 104, + 290, + 507, + 456 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 461, + 505, + 516 + ], + "lines": [ + { + "bbox": [ + 106, + 459, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 506, + 474 + ], + "score": 1.0, + "content": "Face landmark detection. We select Face Alignment Network (FAN) (Bulat & Tzimiropoulos,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 471, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 505, + 484 + ], + "score": 1.0, + "content": "2017b) trained on 300W-LP (Zhu et al., 2016b) and fine-tuned on 300-W (Sagonas et al., 2013) for", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 481, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 506, + 496 + ], + "score": 1.0, + "content": "2D landmark detection. The network is constructed by stacking Hour-Glass network (Newell et al.,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 493, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 506, + 507 + ], + "score": 1.0, + "content": "2016) with hierarchical block (Bulat & Tzimiropoulos, 2017a). Given a portrait image as input, FAN", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 505, + 449, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 449, + 518 + ], + "score": 1.0, + "content": "outputs 2D heatmaps which can be subsequently leveraged to yield 68 2D landmarks.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 459, + 506, + 518 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 522, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 506, + 535 + ], + "score": 1.0, + "content": "Semantic attacks on face images. In our experiments, we randomly sample 1, 280 distinct identi-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 531, + 507, + 548 + ], + "spans": [ + { + "bbox": [ + 104, + 531, + 507, + 548 + ], + "score": 1.0, + "content": "ties form CelebA (Liu et al., 2015). To reduce the reconstruction error brought by the generator (e.g.,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 544, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 159, + 556 + ], + "score": 0.9, + "content": "\\mathbf { x } \\neq \\mathcal { G } ( \\mathbf { x } , \\mathbf { c } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 544, + 438, + 558 + ], + "score": 1.0, + "content": "in practice, we take one more step to obtain the updated feature map", + "type": "text" + }, + { + "bbox": [ + 439, + 544, + 502, + 556 + ], + "score": 0.92, + "content": "\\mathbf { f } ^ { \\prime } = \\mathcal { G } _ { \\mathrm { e n c } } ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 544, + 506, + 558 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 555, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 134, + 568 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 556, + 198, + 567 + ], + "score": 0.81, + "content": "\\mathbf { x } ^ { \\prime } = \\mathrm { a r g m i n } _ { \\mathbf { x } ^ { \\prime } }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 555, + 263, + 567 + ], + "score": 0.91, + "content": "\\| \\mathcal { G } ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } ) - \\mathbf { x } \\|", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 555, + 506, + 568 + ], + "score": 1.0, + "content": "in feature-map interpolation. In our experiments, we use", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "the last conv layer before upsampling in the generator as our as feature-map f given by the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 578, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 293, + 588 + ], + "score": 1.0, + "content": "attack effectiveness. We also fix the parameter", + "type": "text" + }, + { + "bbox": [ + 293, + 578, + 301, + 587 + ], + "score": 0.71, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 578, + 505, + 588 + ], + "score": 1.0, + "content": "(e.g., balances the adversarial loss and smoothness", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 588, + 427, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 427, + 600 + ], + "score": 1.0, + "content": "constraint in Eq. 3) to be 0.01 for both face verification and landmark detection.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40, + "bbox_fs": [ + 104, + 522, + 507, + 600 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "We used the StarGAN (Choi et al., 2018) for attribute-conditional image editing. In particular, we", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "re-trained model on CelebA dataset (Liu et al., 2015) by aligning the face landmarks and then resizing", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 190, + 639 + ], + "score": 1.0, + "content": "images to resolution", + "type": "text" + }, + { + "bbox": [ + 190, + 627, + 233, + 637 + ], + "score": 0.89, + "content": "1 1 2 \\times 1 1 2", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 627, + 505, + 639 + ], + "score": 1.0, + "content": ". In addition, we select 17 identity-preserving attributes as our input", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 639, + 389, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 389, + 650 + ], + "score": 1.0, + "content": "condition, as such attributes related to facial expression and hair color.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 604, + 505, + 650 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 226, + 668 + ], + "score": 1.0, + "content": "For each distinct identity pair", + "type": "text" + }, + { + "bbox": [ + 227, + 655, + 259, + 667 + ], + "score": 0.92, + "content": "( \\mathbf { x } , \\mathbf { x } ^ { \\mathrm { t g t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 654, + 506, + 668 + ], + "score": 1.0, + "content": ", we perform semanticAdv guided by each of the 17 attributes", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "(e.g., we intentionally add or remove one specific attribute while keeping the rest unchanged). In", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 189, + 690 + ], + "score": 1.0, + "content": "total, for each image", + "type": "text" + }, + { + "bbox": [ + 189, + 679, + 197, + 687 + ], + "score": 0.27, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 677, + 505, + 690 + ], + "score": 1.0, + "content": ", we generate 17 adversarial images with different augmented attributes. In the", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "experiments, we select a pixel-wise adversarial attack method (Carlini & Wagner, 2017) (referred", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "as CW) as our baseline for comparison. Compared to our proposed method, CW does not require", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 710, + 507, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 507, + 722 + ], + "score": 1.0, + "content": "visual attributes as part of the system, as it only generates one adversarial example for each instance.", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 721, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 506, + 733 + ], + "score": 1.0, + "content": "We refer the corresponding attack success rate as the instance-wise success rate in which the attack", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "success rate is calculated for each instance. For each instance with 17 adversarial images using", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "different augmented attributes, if one of the 17 resulting images can attack successfully, we count the", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 303, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 303, + 117 + ], + "score": 1.0, + "content": "attack of this instance as one success, vice verse.", + "type": "text", + "cross_page": true + } + ], + "index": 2 + } + ], + "index": 51, + "bbox_fs": [ + 105, + 654, + 507, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "success rate is calculated for each instance. For each instance with 17 adversarial images using", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "different augmented attributes, if one of the 17 resulting images can attack successfully, we count the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 303, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 303, + 117 + ], + "score": 1.0, + "content": "attack of this instance as one success, vice verse.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 120, + 505, + 253 + ], + "lines": [ + { + "bbox": [ + 106, + 121, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 121, + 505, + 133 + ], + "score": 1.0, + "content": "Semantic attacks on street-view images. We select DRN-D-22 (Yu et al., 2017) as our semantic", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 415, + 144 + ], + "score": 1.0, + "content": "segmentation model and fine-tune the model on image regions with resolution", + "type": "text" + }, + { + "bbox": [ + 415, + 132, + 459, + 142 + ], + "score": 0.9, + "content": "2 5 6 \\times 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 132, + 506, + 144 + ], + "score": 1.0, + "content": ". To synthe-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 505, + 155 + ], + "score": 1.0, + "content": "size semantic adversarial perturbations, we consider semantic label maps as the input attribute and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "leverage a generative image manipulation model (Hong et al., 2018) pre-trained on CityScape (Cordts", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 386, + 177 + ], + "score": 1.0, + "content": "et al., 2016) dataset. Given an input semantic label map at resolution", + "type": "text" + }, + { + "bbox": [ + 386, + 165, + 429, + 175 + ], + "score": 0.9, + "content": "2 5 6 \\times 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 164, + 505, + 177 + ], + "score": 1.0, + "content": ", we select a target", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 506, + 189 + ], + "score": 1.0, + "content": "object instance (e.g., a pedestrian) to attack. Then, we create a manipulated semantic label map", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "by inserting another object instance (e.g., a car) in the vicinity of the target object. Similar to the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 210 + ], + "score": 1.0, + "content": "experiments in the face domain, for both semantic label maps, we used the image manipulation", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 209, + 506, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 377, + 220 + ], + "score": 1.0, + "content": "encoder to extract features (with 1, 024 channels at spatial resolution", + "type": "text" + }, + { + "bbox": [ + 378, + 209, + 411, + 219 + ], + "score": 0.88, + "content": "1 6 \\times 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 209, + 506, + 220 + ], + "score": 1.0, + "content": ") and conducted feature-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 218, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 506, + 234 + ], + "score": 1.0, + "content": "space interpolation. We synthesized the final image by feeding the interpolated features to the image", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "manipulation decoder. By searching the interpolation coefficient that maximizes the attack rate, we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 240, + 400, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 400, + 255 + ], + "score": 1.0, + "content": "are able to fool the segmentation model with the synthesized final image.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 107, + 259, + 333, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 259, + 333, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 333, + 271 + ], + "score": 1.0, + "content": "4.2 SemanticAdv ON FACE IDENTITY VERIFICATION", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 273, + 505, + 361 + ], + "lines": [ + { + "bbox": [ + 106, + 273, + 504, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 504, + 285 + ], + "score": 1.0, + "content": "Attribute-space vs. feature-space interpolation. First, we found that both attribute-space and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 283, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 506, + 297 + ], + "score": 1.0, + "content": "feature-space interpolation could generate reasonable samples (see Figure I in Appendix). Compared", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 295, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 505, + 307 + ], + "score": 1.0, + "content": "to attribute-space interpolation, generating adversarial examples with feature-space interpolation", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 305, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 506, + 318 + ], + "score": 1.0, + "content": "produced much better quantitative results (see Table E in Appendix). We measured the attack success", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 315, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 104, + 315, + 275, + 328 + ], + "score": 1.0, + "content": "rate of attribute-space interpolation (with", + "type": "text" + }, + { + "bbox": [ + 276, + 316, + 376, + 327 + ], + "score": 0.64, + "content": "\\mathrm { G \\mathrm { - } F P R = T \\mathrm { - } F P R = 1 0 ^ { - 3 } , }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 315, + 383, + 328 + ], + "score": 1.0, + "content": "):", + "type": "text" + }, + { + "bbox": [ + 383, + 316, + 411, + 327 + ], + "score": 0.87, + "content": "0 . 0 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 315, + 463, + 328 + ], + "score": 1.0, + "content": "on R-101-S,", + "type": "text" + }, + { + "bbox": [ + 464, + 316, + 491, + 327 + ], + "score": 0.87, + "content": "0 . 3 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 315, + 505, + 328 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 163, + 339 + ], + "score": 1.0, + "content": "R-101-C, and", + "type": "text" + }, + { + "bbox": [ + 163, + 327, + 190, + 338 + ], + "score": 0.89, + "content": "0 . 1 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "on both R-50-S and R-50-C. While feature-space interpolation achieves almost", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 131, + 349 + ], + "score": 0.86, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "success rate on all those models (see Figure 3). We conjecture that this is because the high", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 349, + 377, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 377, + 362 + ], + "score": 1.0, + "content": "dimensional feature space can provide more manipulation freedom.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5 + }, + { + "type": "image", + "bbox": [ + 135, + 373, + 471, + 469 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 135, + 373, + 471, + 469 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 373, + 471, + 469 + ], + "spans": [ + { + "bbox": [ + 135, + 373, + 471, + 469 + ], + "score": 0.973, + "type": "image", + "image_path": "bc14263add8b7333aa1ab335a48d188a488851f0eeec7d268fb45170e51410fe.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 135, + 373, + 471, + 405.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 135, + 405.0, + 471, + 437.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 135, + 437.0, + 471, + 469.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 476, + 505, + 531 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "Figure 2: Qualitative comparisons between our proposed SemanticAdv and pixel-wise adversarial", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 486, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 501 + ], + "score": 1.0, + "content": "examples generated by CW (Carlini & Wagner, 2017). Along with the adversarial examples, we", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "also provide the corresponding perturbations (residual) on the right. Perturbations generated by our", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 509, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 163, + 522 + ], + "score": 1.0, + "content": "SemanticAdv", + "type": "text" + }, + { + "bbox": [ + 164, + 509, + 226, + 520 + ], + "score": 0.67, + "content": "\\mathrm { \\bf G } \\mathrm { - } \\mathrm { \\bf F P R } = 1 0 ^ { - 3 } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 509, + 505, + 522 + ], + "score": 1.0, + "content": "are unrestricted with semantically meaningful patterns. More results", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 520, + 263, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 263, + 533 + ], + "score": 1.0, + "content": "are shown in Appendix (see Figure N).", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + } + ], + "index": 27.0 + }, + { + "type": "text", + "bbox": [ + 107, + 539, + 505, + 617 + ], + "lines": [ + { + "bbox": [ + 106, + 540, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 553 + ], + "score": 1.0, + "content": "Overall analysis. Figure 2 shows the generated adversarial images and corresponding perturbations", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "against R-101-S of SemanticAdv and CW respectively. The text below each figure is the name of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 562, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 574 + ], + "score": 1.0, + "content": "augmented attribute, the sign before the name represents “adding” (in red) or “removing” (in blue)", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "score": 1.0, + "content": "the corresponding attribute from the original image. We see that SemanticAdv is able to generate", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "perceptually reasonable examples guided by the corresponding attribute. In particular, SemanticAdv", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "score": 1.0, + "content": "is able to generate perturbations on the corresponding regions correlated with the augmented attribute,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 605, + 502, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 502, + 619 + ], + "score": 1.0, + "content": "while the perturbations of CW have no specific pattern and are evenly distributed across the image.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 622, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 506, + 634 + ], + "score": 1.0, + "content": "Analysis: controlling single attribute. One of the key advantages of SemanticAdv is that we can", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 632, + 507, + 646 + ], + "spans": [ + { + "bbox": [ + 104, + 632, + 507, + 646 + ], + "score": 1.0, + "content": "generate adversarial perturbations in a more controllable fashion guided by the semantic attributes.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "This allows analyzing the robustness of a recognition system against different types of semantic", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "attacks. We group the adversarial examples by augmented attributes in various settings. In Figure 3,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "we present the attack success rate against two face verification models, namely, R-101-S and R-101-C,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 464, + 689 + ], + "score": 1.0, + "content": "guided by different attributes. We highlight the bar with light blue for G-FPR equals to", + "type": "text" + }, + { + "bbox": [ + 465, + 676, + 487, + 687 + ], + "score": 0.9, + "content": "1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 676, + 506, + 689 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 104, + 685, + 507, + 702 + ], + "spans": [ + { + "bbox": [ + 104, + 685, + 208, + 702 + ], + "score": 1.0, + "content": "blue for G-FPR equals to", + "type": "text" + }, + { + "bbox": [ + 208, + 687, + 230, + 698 + ], + "score": 0.9, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 685, + 464, + 702 + ], + "score": 1.0, + "content": ", respectively. As we see in this figure, with a larger T-FPR", + "type": "text" + }, + { + "bbox": [ + 464, + 687, + 486, + 698 + ], + "score": 0.89, + "content": "1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 685, + 507, + 702 + ], + "score": 1.0, + "content": ", our", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 238, + 711 + ], + "score": 1.0, + "content": "SemanticAdv can achieve almost", + "type": "text" + }, + { + "bbox": [ + 238, + 699, + 262, + 709 + ], + "score": 0.86, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "attack success rate across different attributes. With a smaller", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 104, + 708, + 136, + 723 + ], + "score": 1.0, + "content": "T-FPR", + "type": "text" + }, + { + "bbox": [ + 136, + 709, + 158, + 720 + ], + "score": 0.91, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 708, + 506, + 723 + ], + "score": 1.0, + "content": ", we find that SemanticAdv guided by some attributes such as Mouth Slightly", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 309, + 733 + ], + "score": 1.0, + "content": "Open and Arched Eyebrows achieve less than", + "type": "text" + }, + { + "bbox": [ + 309, + 721, + 328, + 731 + ], + "score": 0.87, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "attack success rate, while the other attributes", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 43.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 115 + ], + "lines": [], + "index": 1, + "bbox_fs": [ + 105, + 81, + 505, + 117 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 120, + 505, + 253 + ], + "lines": [ + { + "bbox": [ + 106, + 121, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 121, + 505, + 133 + ], + "score": 1.0, + "content": "Semantic attacks on street-view images. We select DRN-D-22 (Yu et al., 2017) as our semantic", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 415, + 144 + ], + "score": 1.0, + "content": "segmentation model and fine-tune the model on image regions with resolution", + "type": "text" + }, + { + "bbox": [ + 415, + 132, + 459, + 142 + ], + "score": 0.9, + "content": "2 5 6 \\times 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 132, + 506, + 144 + ], + "score": 1.0, + "content": ". To synthe-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 505, + 155 + ], + "score": 1.0, + "content": "size semantic adversarial perturbations, we consider semantic label maps as the input attribute and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "leverage a generative image manipulation model (Hong et al., 2018) pre-trained on CityScape (Cordts", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 386, + 177 + ], + "score": 1.0, + "content": "et al., 2016) dataset. Given an input semantic label map at resolution", + "type": "text" + }, + { + "bbox": [ + 386, + 165, + 429, + 175 + ], + "score": 0.9, + "content": "2 5 6 \\times 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 164, + 505, + 177 + ], + "score": 1.0, + "content": ", we select a target", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 506, + 189 + ], + "score": 1.0, + "content": "object instance (e.g., a pedestrian) to attack. Then, we create a manipulated semantic label map", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "by inserting another object instance (e.g., a car) in the vicinity of the target object. Similar to the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 210 + ], + "score": 1.0, + "content": "experiments in the face domain, for both semantic label maps, we used the image manipulation", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 209, + 506, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 377, + 220 + ], + "score": 1.0, + "content": "encoder to extract features (with 1, 024 channels at spatial resolution", + "type": "text" + }, + { + "bbox": [ + 378, + 209, + 411, + 219 + ], + "score": 0.88, + "content": "1 6 \\times 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 209, + 506, + 220 + ], + "score": 1.0, + "content": ") and conducted feature-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 218, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 506, + 234 + ], + "score": 1.0, + "content": "space interpolation. We synthesized the final image by feeding the interpolated features to the image", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "manipulation decoder. By searching the interpolation coefficient that maximizes the attack rate, we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 240, + 400, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 400, + 255 + ], + "score": 1.0, + "content": "are able to fool the segmentation model with the synthesized final image.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 121, + 506, + 255 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 259, + 333, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 259, + 333, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 333, + 271 + ], + "score": 1.0, + "content": "4.2 SemanticAdv ON FACE IDENTITY VERIFICATION", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 273, + 505, + 361 + ], + "lines": [ + { + "bbox": [ + 106, + 273, + 504, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 504, + 285 + ], + "score": 1.0, + "content": "Attribute-space vs. feature-space interpolation. First, we found that both attribute-space and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 283, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 506, + 297 + ], + "score": 1.0, + "content": "feature-space interpolation could generate reasonable samples (see Figure I in Appendix). Compared", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 295, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 505, + 307 + ], + "score": 1.0, + "content": "to attribute-space interpolation, generating adversarial examples with feature-space interpolation", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 305, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 506, + 318 + ], + "score": 1.0, + "content": "produced much better quantitative results (see Table E in Appendix). We measured the attack success", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 315, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 104, + 315, + 275, + 328 + ], + "score": 1.0, + "content": "rate of attribute-space interpolation (with", + "type": "text" + }, + { + "bbox": [ + 276, + 316, + 376, + 327 + ], + "score": 0.64, + "content": "\\mathrm { G \\mathrm { - } F P R = T \\mathrm { - } F P R = 1 0 ^ { - 3 } , }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 315, + 383, + 328 + ], + "score": 1.0, + "content": "):", + "type": "text" + }, + { + "bbox": [ + 383, + 316, + 411, + 327 + ], + "score": 0.87, + "content": "0 . 0 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 315, + 463, + 328 + ], + "score": 1.0, + "content": "on R-101-S,", + "type": "text" + }, + { + "bbox": [ + 464, + 316, + 491, + 327 + ], + "score": 0.87, + "content": "0 . 3 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 315, + 505, + 328 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 163, + 339 + ], + "score": 1.0, + "content": "R-101-C, and", + "type": "text" + }, + { + "bbox": [ + 163, + 327, + 190, + 338 + ], + "score": 0.89, + "content": "0 . 1 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "on both R-50-S and R-50-C. While feature-space interpolation achieves almost", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 131, + 349 + ], + "score": 0.86, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "success rate on all those models (see Figure 3). We conjecture that this is because the high", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 349, + 377, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 377, + 362 + ], + "score": 1.0, + "content": "dimensional feature space can provide more manipulation freedom.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5, + "bbox_fs": [ + 104, + 273, + 506, + 362 + ] + }, + { + "type": "image", + "bbox": [ + 135, + 373, + 471, + 469 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 135, + 373, + 471, + 469 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 373, + 471, + 469 + ], + "spans": [ + { + "bbox": [ + 135, + 373, + 471, + 469 + ], + "score": 0.973, + "type": "image", + "image_path": "bc14263add8b7333aa1ab335a48d188a488851f0eeec7d268fb45170e51410fe.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 135, + 373, + 471, + 405.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 135, + 405.0, + 471, + 437.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 135, + 437.0, + 471, + 469.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 476, + 505, + 531 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "Figure 2: Qualitative comparisons between our proposed SemanticAdv and pixel-wise adversarial", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 486, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 501 + ], + "score": 1.0, + "content": "examples generated by CW (Carlini & Wagner, 2017). Along with the adversarial examples, we", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "also provide the corresponding perturbations (residual) on the right. Perturbations generated by our", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 509, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 163, + 522 + ], + "score": 1.0, + "content": "SemanticAdv", + "type": "text" + }, + { + "bbox": [ + 164, + 509, + 226, + 520 + ], + "score": 0.67, + "content": "\\mathrm { \\bf G } \\mathrm { - } \\mathrm { \\bf F P R } = 1 0 ^ { - 3 } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 509, + 505, + 522 + ], + "score": 1.0, + "content": "are unrestricted with semantically meaningful patterns. More results", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 520, + 263, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 263, + 533 + ], + "score": 1.0, + "content": "are shown in Appendix (see Figure N).", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + } + ], + "index": 27.0 + }, + { + "type": "text", + "bbox": [ + 107, + 539, + 505, + 617 + ], + "lines": [ + { + "bbox": [ + 106, + 540, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 553 + ], + "score": 1.0, + "content": "Overall analysis. Figure 2 shows the generated adversarial images and corresponding perturbations", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "against R-101-S of SemanticAdv and CW respectively. The text below each figure is the name of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 562, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 574 + ], + "score": 1.0, + "content": "augmented attribute, the sign before the name represents “adding” (in red) or “removing” (in blue)", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "score": 1.0, + "content": "the corresponding attribute from the original image. We see that SemanticAdv is able to generate", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "perceptually reasonable examples guided by the corresponding attribute. In particular, SemanticAdv", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "score": 1.0, + "content": "is able to generate perturbations on the corresponding regions correlated with the augmented attribute,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 605, + 502, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 502, + 619 + ], + "score": 1.0, + "content": "while the perturbations of CW have no specific pattern and are evenly distributed across the image.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 540, + 506, + 619 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 622, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 506, + 634 + ], + "score": 1.0, + "content": "Analysis: controlling single attribute. One of the key advantages of SemanticAdv is that we can", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 632, + 507, + 646 + ], + "spans": [ + { + "bbox": [ + 104, + 632, + 507, + 646 + ], + "score": 1.0, + "content": "generate adversarial perturbations in a more controllable fashion guided by the semantic attributes.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "This allows analyzing the robustness of a recognition system against different types of semantic", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "attacks. We group the adversarial examples by augmented attributes in various settings. In Figure 3,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "we present the attack success rate against two face verification models, namely, R-101-S and R-101-C,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 464, + 689 + ], + "score": 1.0, + "content": "guided by different attributes. We highlight the bar with light blue for G-FPR equals to", + "type": "text" + }, + { + "bbox": [ + 465, + 676, + 487, + 687 + ], + "score": 0.9, + "content": "1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 676, + 506, + 689 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 104, + 685, + 507, + 702 + ], + "spans": [ + { + "bbox": [ + 104, + 685, + 208, + 702 + ], + "score": 1.0, + "content": "blue for G-FPR equals to", + "type": "text" + }, + { + "bbox": [ + 208, + 687, + 230, + 698 + ], + "score": 0.9, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 685, + 464, + 702 + ], + "score": 1.0, + "content": ", respectively. As we see in this figure, with a larger T-FPR", + "type": "text" + }, + { + "bbox": [ + 464, + 687, + 486, + 698 + ], + "score": 0.89, + "content": "1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 685, + 507, + 702 + ], + "score": 1.0, + "content": ", our", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 238, + 711 + ], + "score": 1.0, + "content": "SemanticAdv can achieve almost", + "type": "text" + }, + { + "bbox": [ + 238, + 699, + 262, + 709 + ], + "score": 0.86, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "attack success rate across different attributes. With a smaller", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 104, + 708, + 136, + 723 + ], + "score": 1.0, + "content": "T-FPR", + "type": "text" + }, + { + "bbox": [ + 136, + 709, + 158, + 720 + ], + "score": 0.91, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 708, + 506, + 723 + ], + "score": 1.0, + "content": ", we find that SemanticAdv guided by some attributes such as Mouth Slightly", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 309, + 733 + ], + "score": 1.0, + "content": "Open and Arched Eyebrows achieve less than", + "type": "text" + }, + { + "bbox": [ + 309, + 721, + 328, + 731 + ], + "score": 0.87, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "attack success rate, while the other attributes", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "such as Pale Skin and Eyeglasses are relatively less affected. In summary, we found that", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 104, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "SemanticAdv guided by attributes describing the local shape (e.g., mouth, earrings) achieve a relatively", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "lower attack success rate compared to attributes relevant to the color (e.g., hair color) or entire face", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "region (e.g., skin). This suggests that the face verification models used in our experiments are more", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "robustly trained in terms of detecting local shapes compared to colors. Please note that in practice", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 151 + ], + "score": 1.0, + "content": "we have the flexibility to select attributes for attacking an image based on the perceptual quality and", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 148, + 185, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 185, + 161 + ], + "score": 1.0, + "content": "attack success rate.", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 43.5, + "bbox_fs": [ + 104, + 621, + 507, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "such as Pale Skin and Eyeglasses are relatively less affected. In summary, we found that", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 104, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "SemanticAdv guided by attributes describing the local shape (e.g., mouth, earrings) achieve a relatively", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "lower attack success rate compared to attributes relevant to the color (e.g., hair color) or entire face", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "region (e.g., skin). This suggests that the face verification models used in our experiments are more", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "robustly trained in terms of detecting local shapes compared to colors. Please note that in practice", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 151 + ], + "score": 1.0, + "content": "we have the flexibility to select attributes for attacking an image based on the perceptual quality and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 148, + 185, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 185, + 161 + ], + "score": 1.0, + "content": "attack success rate.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 165, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "Figure 4 shows the adversarial examples with augmented semantic attributes against R-101-S model.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 369, + 189 + ], + "score": 1.0, + "content": "The attribute names are shown in the bottom. The upper images are", + "type": "text" + }, + { + "bbox": [ + 370, + 176, + 412, + 188 + ], + "score": 0.93, + "content": "\\mathcal { G } ( \\mathbf { x } , \\mathbf { c } ^ { \\mathrm { n e w } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 175, + 506, + 189 + ], + "score": 1.0, + "content": "generated by StarGAN", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 206, + 201 + ], + "score": 1.0, + "content": "with augmented attribute", + "type": "text" + }, + { + "bbox": [ + 207, + 187, + 225, + 198 + ], + "score": 0.85, + "content": "\\mathbf { c } ^ { \\mathrm { { n e w } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 186, + 506, + 201 + ], + "score": 1.0, + "content": "where the lower images are the corresponding adversarial images with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 199, + 228, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 228, + 210 + ], + "score": 1.0, + "content": "the same augmented attribute.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "image", + "bbox": [ + 318, + 222, + 501, + 338 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 318, + 222, + 501, + 338 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 318, + 222, + 501, + 338 + ], + "spans": [ + { + "bbox": [ + 318, + 222, + 501, + 338 + ], + "score": 0.613, + "type": "image", + "image_path": "f425e3990a1d398f6b0c0ca3d32c623c42afee2547c5a6fc767eb289e27fcc47.jpg" + } + ] + } + ], + "index": 19.0, + "virtual_lines": [ + { + "bbox": [ + 318, + 222, + 501, + 236.5 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 318, + 236.5, + 501, + 251.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 318, + 251.0, + 501, + 265.5 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 318, + 265.5, + 501, + 280.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 318, + 280.0, + 501, + 294.5 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 318, + 294.5, + 501, + 309.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 318, + 309.0, + 501, + 323.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 318, + 323.5, + 501, + 338.0 + ], + "spans": [], + "index": 26 + } + ] + } + ], + "index": 19.0 + }, + { + "type": "image", + "bbox": [ + 108, + 222, + 301, + 337 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 222, + 301, + 337 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 222, + 301, + 337 + ], + "spans": [ + { + "bbox": [ + 108, + 222, + 301, + 337 + ], + "score": 0.724, + "type": "image", + "image_path": "b9a1a0608405ea4229a059b5e220b91f0651c5175ef5d100359c433825165714.jpg" + } + ] + } + ], + "index": 18.0, + "virtual_lines": [ + { + "bbox": [ + 108, + 222, + 301, + 236.375 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 236.375, + 301, + 250.75 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 108, + 250.75, + 301, + 265.125 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 108, + 265.125, + 301, + 279.5 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 108, + 279.5, + 301, + 293.875 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 108, + 293.875, + 301, + 308.25 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 108, + 308.25, + 301, + 322.625 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 108, + 322.625, + 301, + 337.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 354, + 506, + 388 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "Figure 3: Quantitative analysis on the attack success rate with different single-attribute attacks. In", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 365, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 347, + 378 + ], + "score": 1.0, + "content": "each figure, we show the results correspond to a larger FPR", + "type": "text" + }, + { + "bbox": [ + 347, + 365, + 447, + 376 + ], + "score": 0.83, + "content": "( \\mathrm { G } \\mathrm { - } \\mathrm { F P R } = \\mathrm { T } \\mathrm { - } \\mathrm { F P R } = 1 0 ^ { - 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 365, + 506, + 378 + ], + "score": 1.0, + "content": "in skyblue", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 376, + 479, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 285, + 389 + ], + "score": 1.0, + "content": "and the results correspond to a smaller FPR", + "type": "text" + }, + { + "bbox": [ + 285, + 376, + 385, + 387 + ], + "score": 0.53, + "content": "\\mathrm { \\langle G \\mathrm { - } F P R = T \\mathrm { - } F P R = 1 0 ^ { - 4 } \\rangle }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 376, + 479, + 389 + ], + "score": 1.0, + "content": ") in blue, respectively.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + } + ], + "index": 23.0 + }, + { + "type": "image", + "bbox": [ + 148, + 397, + 461, + 523 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 148, + 397, + 461, + 523 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 148, + 397, + 461, + 523 + ], + "spans": [ + { + "bbox": [ + 148, + 397, + 461, + 523 + ], + "score": 0.969, + "type": "image", + "image_path": "7b1364cfc82980abf6870d20ac5cdf4f6c60172679d0b02e3a93fb132610f5c2.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 148, + 397, + 461, + 439.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 148, + 439.0, + 461, + 481.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 148, + 481.0, + 461, + 523.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 530, + 504, + 552 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 529, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 370, + 543 + ], + "score": 1.0, + "content": "Figure 4: Qualitative analysis on single-attribute adversarial attack", + "type": "text" + }, + { + "bbox": [ + 370, + 529, + 432, + 541 + ], + "score": 0.8, + "content": "\\mathrm { ( G { - } F P R = 1 0 ^ { - 3 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 529, + 505, + 543 + ], + "score": 1.0, + "content": "). More results are", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 541, + 345, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 345, + 552 + ], + "score": 1.0, + "content": "shown in Appendix (see Figure K, Figure L and Figure M).", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + } + ], + "index": 32.25 + }, + { + "type": "text", + "bbox": [ + 106, + 566, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 106, + 567, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 506, + 578 + ], + "score": 1.0, + "content": "Analysis: semantic attack transferability. To further understand the property of SemanticAdv,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 577, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 589 + ], + "score": 1.0, + "content": "we analyze the transferability of SemanticAdv on various settings. For each model with different", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "score": 1.0, + "content": "FPRs, we select the successfully attacked adversarial examples from Section 4.1 to construct our", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 600, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 505, + 611 + ], + "score": 1.0, + "content": "evaluation dataset and evaluate these adversarial samples across different models. Table 1a illustrates", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 609, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 437, + 622 + ], + "score": 1.0, + "content": "the transferability of SemanticAdv among different models by using the same FPRs", + "type": "text" + }, + { + "bbox": [ + 437, + 610, + 485, + 620 + ], + "score": 0.32, + "content": "\\begin{array} { r } { ( \\mathrm { G - F P R } = \\mathrm { T } . } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 609, + 505, + 622 + ], + "score": 1.0, + "content": "FPR", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 620, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 137, + 632 + ], + "score": 0.88, + "content": "= 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 620, + 347, + 633 + ], + "score": 1.0, + "content": "). Table 1b illustrates the result with different FPRs", + "type": "text" + }, + { + "bbox": [ + 347, + 621, + 409, + 632 + ], + "score": 0.86, + "content": "\\mathbf { \\bar { ( G - F P R = 1 0 ^ { - 4 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 620, + 427, + 633 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 427, + 621, + 487, + 632 + ], + "score": 0.88, + "content": "\\mathrm { \\Delta T - F P R } = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 620, + 506, + 633 + ], + "score": 1.0, + "content": ") for", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 631, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 104, + 631, + 506, + 645 + ], + "score": 1.0, + "content": "generation and evaluation. As shown in Table 1a, adversarial examples generated against models", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "score": 1.0, + "content": "trained with softmax loss exhibit certain transferability compared to models trained with cosine", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 654, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 666 + ], + "score": 1.0, + "content": "loss. We conduct the same experiment by generating adversarial examples with CW and found it has", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 665, + 444, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 444, + 677 + ], + "score": 1.0, + "content": "weaker transferability compared to our SemanticAdv (results in brackets of Table 1).", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 107, + 682, + 505, + 737 + ], + "lines": [ + { + "bbox": [ + 106, + 682, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 505, + 694 + ], + "score": 1.0, + "content": "As Table 1b illustrates, the adversarial examples generated against the model with smaller G-FPR", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 691, + 503, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 139, + 703 + ], + "score": 0.88, + "content": "= 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 691, + 468, + 705 + ], + "score": 1.0, + "content": "exhibit strong attack success rate when evaluating on the model with larger T-FPR", + "type": "text" + }, + { + "bbox": [ + 468, + 692, + 503, + 703 + ], + "score": 0.6, + "content": ". = 1 0 ^ { - 3 }", + "type": "inline_equation" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 703, + 506, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 703, + 506, + 717 + ], + "score": 1.0, + "content": "Especially, we found the adversarial examples generated against R-101-S have the best attack", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 714, + 506, + 727 + ], + "spans": [ + { + "bbox": [ + 105, + 714, + 506, + 727 + ], + "score": 1.0, + "content": "performance on other models. These findings motivate the analysis of query-free black-box API", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 724, + 277, + 739 + ], + "spans": [ + { + "bbox": [ + 105, + 724, + 277, + 739 + ], + "score": 1.0, + "content": "attack detailed in the following paragraph.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 159 + ], + "lines": [], + "index": 3, + "bbox_fs": [ + 104, + 83, + 505, + 161 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 165, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "Figure 4 shows the adversarial examples with augmented semantic attributes against R-101-S model.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 369, + 189 + ], + "score": 1.0, + "content": "The attribute names are shown in the bottom. The upper images are", + "type": "text" + }, + { + "bbox": [ + 370, + 176, + 412, + 188 + ], + "score": 0.93, + "content": "\\mathcal { G } ( \\mathbf { x } , \\mathbf { c } ^ { \\mathrm { n e w } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 175, + 506, + 189 + ], + "score": 1.0, + "content": "generated by StarGAN", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 206, + 201 + ], + "score": 1.0, + "content": "with augmented attribute", + "type": "text" + }, + { + "bbox": [ + 207, + 187, + 225, + 198 + ], + "score": 0.85, + "content": "\\mathbf { c } ^ { \\mathrm { { n e w } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 186, + 506, + 201 + ], + "score": 1.0, + "content": "where the lower images are the corresponding adversarial images with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 199, + 228, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 228, + 210 + ], + "score": 1.0, + "content": "the same augmented attribute.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 165, + 506, + 210 + ] + }, + { + "type": "image", + "bbox": [ + 318, + 222, + 501, + 338 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 318, + 222, + 501, + 338 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 318, + 222, + 501, + 338 + ], + "spans": [ + { + "bbox": [ + 318, + 222, + 501, + 338 + ], + "score": 0.613, + "type": "image", + "image_path": "f425e3990a1d398f6b0c0ca3d32c623c42afee2547c5a6fc767eb289e27fcc47.jpg" + } + ] + } + ], + "index": 19.0, + "virtual_lines": [ + { + "bbox": [ + 318, + 222, + 501, + 236.5 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 318, + 236.5, + 501, + 251.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 318, + 251.0, + 501, + 265.5 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 318, + 265.5, + 501, + 280.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 318, + 280.0, + 501, + 294.5 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 318, + 294.5, + 501, + 309.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 318, + 309.0, + 501, + 323.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 318, + 323.5, + 501, + 338.0 + ], + "spans": [], + "index": 26 + } + ] + } + ], + "index": 19.0 + }, + { + "type": "image", + "bbox": [ + 108, + 222, + 301, + 337 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 222, + 301, + 337 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 222, + 301, + 337 + ], + "spans": [ + { + "bbox": [ + 108, + 222, + 301, + 337 + ], + "score": 0.724, + "type": "image", + "image_path": "b9a1a0608405ea4229a059b5e220b91f0651c5175ef5d100359c433825165714.jpg" + } + ] + } + ], + "index": 18.0, + "virtual_lines": [ + { + "bbox": [ + 108, + 222, + 301, + 236.375 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 236.375, + 301, + 250.75 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 108, + 250.75, + 301, + 265.125 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 108, + 265.125, + 301, + 279.5 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 108, + 279.5, + 301, + 293.875 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 108, + 293.875, + 301, + 308.25 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 108, + 308.25, + 301, + 322.625 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 108, + 322.625, + 301, + 337.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 354, + 506, + 388 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "Figure 3: Quantitative analysis on the attack success rate with different single-attribute attacks. In", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 365, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 347, + 378 + ], + "score": 1.0, + "content": "each figure, we show the results correspond to a larger FPR", + "type": "text" + }, + { + "bbox": [ + 347, + 365, + 447, + 376 + ], + "score": 0.83, + "content": "( \\mathrm { G } \\mathrm { - } \\mathrm { F P R } = \\mathrm { T } \\mathrm { - } \\mathrm { F P R } = 1 0 ^ { - 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 365, + 506, + 378 + ], + "score": 1.0, + "content": "in skyblue", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 376, + 479, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 285, + 389 + ], + "score": 1.0, + "content": "and the results correspond to a smaller FPR", + "type": "text" + }, + { + "bbox": [ + 285, + 376, + 385, + 387 + ], + "score": 0.53, + "content": "\\mathrm { \\langle G \\mathrm { - } F P R = T \\mathrm { - } F P R = 1 0 ^ { - 4 } \\rangle }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 376, + 479, + 389 + ], + "score": 1.0, + "content": ") in blue, respectively.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + } + ], + "index": 23.0 + }, + { + "type": "image", + "bbox": [ + 148, + 397, + 461, + 523 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 148, + 397, + 461, + 523 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 148, + 397, + 461, + 523 + ], + "spans": [ + { + "bbox": [ + 148, + 397, + 461, + 523 + ], + "score": 0.969, + "type": "image", + "image_path": "7b1364cfc82980abf6870d20ac5cdf4f6c60172679d0b02e3a93fb132610f5c2.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 148, + 397, + 461, + 439.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 148, + 439.0, + 461, + 481.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 148, + 481.0, + 461, + 523.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 530, + 504, + 552 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 529, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 370, + 543 + ], + "score": 1.0, + "content": "Figure 4: Qualitative analysis on single-attribute adversarial attack", + "type": "text" + }, + { + "bbox": [ + 370, + 529, + 432, + 541 + ], + "score": 0.8, + "content": "\\mathrm { ( G { - } F P R = 1 0 ^ { - 3 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 529, + 505, + 543 + ], + "score": 1.0, + "content": "). More results are", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 541, + 345, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 345, + 552 + ], + "score": 1.0, + "content": "shown in Appendix (see Figure K, Figure L and Figure M).", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + } + ], + "index": 32.25 + }, + { + "type": "text", + "bbox": [ + 106, + 566, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 106, + 567, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 506, + 578 + ], + "score": 1.0, + "content": "Analysis: semantic attack transferability. To further understand the property of SemanticAdv,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 577, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 589 + ], + "score": 1.0, + "content": "we analyze the transferability of SemanticAdv on various settings. For each model with different", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "score": 1.0, + "content": "FPRs, we select the successfully attacked adversarial examples from Section 4.1 to construct our", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 600, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 505, + 611 + ], + "score": 1.0, + "content": "evaluation dataset and evaluate these adversarial samples across different models. Table 1a illustrates", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 609, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 437, + 622 + ], + "score": 1.0, + "content": "the transferability of SemanticAdv among different models by using the same FPRs", + "type": "text" + }, + { + "bbox": [ + 437, + 610, + 485, + 620 + ], + "score": 0.32, + "content": "\\begin{array} { r } { ( \\mathrm { G - F P R } = \\mathrm { T } . } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 609, + 505, + 622 + ], + "score": 1.0, + "content": "FPR", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 620, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 137, + 632 + ], + "score": 0.88, + "content": "= 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 620, + 347, + 633 + ], + "score": 1.0, + "content": "). Table 1b illustrates the result with different FPRs", + "type": "text" + }, + { + "bbox": [ + 347, + 621, + 409, + 632 + ], + "score": 0.86, + "content": "\\mathbf { \\bar { ( G - F P R = 1 0 ^ { - 4 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 620, + 427, + 633 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 427, + 621, + 487, + 632 + ], + "score": 0.88, + "content": "\\mathrm { \\Delta T - F P R } = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 620, + 506, + 633 + ], + "score": 1.0, + "content": ") for", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 631, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 104, + 631, + 506, + 645 + ], + "score": 1.0, + "content": "generation and evaluation. As shown in Table 1a, adversarial examples generated against models", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "score": 1.0, + "content": "trained with softmax loss exhibit certain transferability compared to models trained with cosine", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 654, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 666 + ], + "score": 1.0, + "content": "loss. We conduct the same experiment by generating adversarial examples with CW and found it has", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 665, + 444, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 444, + 677 + ], + "score": 1.0, + "content": "weaker transferability compared to our SemanticAdv (results in brackets of Table 1).", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 39.5, + "bbox_fs": [ + 104, + 567, + 506, + 677 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 682, + 505, + 737 + ], + "lines": [ + { + "bbox": [ + 106, + 682, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 505, + 694 + ], + "score": 1.0, + "content": "As Table 1b illustrates, the adversarial examples generated against the model with smaller G-FPR", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 691, + 503, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 139, + 703 + ], + "score": 0.88, + "content": "= 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 691, + 468, + 705 + ], + "score": 1.0, + "content": "exhibit strong attack success rate when evaluating on the model with larger T-FPR", + "type": "text" + }, + { + "bbox": [ + 468, + 692, + 503, + 703 + ], + "score": 0.6, + "content": ". = 1 0 ^ { - 3 }", + "type": "inline_equation" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 703, + 506, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 703, + 506, + 717 + ], + "score": 1.0, + "content": "Especially, we found the adversarial examples generated against R-101-S have the best attack", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 714, + 506, + 727 + ], + "spans": [ + { + "bbox": [ + 105, + 714, + 506, + 727 + ], + "score": 1.0, + "content": "performance on other models. These findings motivate the analysis of query-free black-box API", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 724, + 277, + 739 + ], + "spans": [ + { + "bbox": [ + 105, + 724, + 277, + 739 + ], + "score": 1.0, + "content": "attack detailed in the following paragraph.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 682, + 506, + 739 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 350, + 80, + 491, + 124 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 350, + 80, + 491, + 124 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 350, + 80, + 491, + 124 + ], + "spans": [ + { + "bbox": [ + 350, + 80, + 491, + 124 + ], + "score": 0.797, + "html": "
Mtest/MoptR-50-SR-101-S
R-50-S1.000 (1.000)0.862 (0.530)
R-101-S0.874 (0.422)1.000 (1.000)
R-50-C0.693(0.347)0.837 (0.579)
R-101-C0.617 (0.218)0.888 (0.617)
", + "type": "table", + "image_path": "57469e8d95a46fe820286c74e8b8ee4b6109015ae65e817728e3fc376178968b.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 350, + 80, + 491, + 102.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 350, + 102.0, + 491, + 124.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 3.5 + }, + { + "type": "table", + "bbox": [ + 111, + 80, + 345, + 124 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 111, + 80, + 345, + 124 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 80, + 345, + 124 + ], + "spans": [ + { + "bbox": [ + 111, + 80, + 345, + 124 + ], + "score": 0.963, + "html": "
MtestMoptR-50-SR-101-SR-50-CR-101-C
R-50-S1.000 (1.000)0.108 (0.007)0.023(0.002)0.018 (0.002)
R-101-S0.169 (0.006)1.000 (1.000)0.030 (0.002)0.032(0.003)
R-50-C0.166 (0.019)0.202(0.025)1.000 (1.000)0.048(0.007)
R-101-C0.120 (0.015)0.236 (0.029)0.040 (0.006)1.000(1.000)
", + "type": "table", + "image_path": "a3f40dce747cd167588dcf20b5c1fc20e4dff6329527f1a11479082ce178844d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 80, + 345, + 94.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 94.66666666666667, + 345, + 109.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 109.33333333333334, + 345, + 124.00000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 144, + 505, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 144, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 285, + 156 + ], + "score": 1.0, + "content": "Table 1: Transferability of SemanticAdv: cell", + "type": "text" + }, + { + "bbox": [ + 286, + 144, + 307, + 156 + ], + "score": 0.91, + "content": "( i , j )", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 144, + 506, + 156 + ], + "score": 1.0, + "content": "shows attack success rate of adversarial examples", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 155, + 507, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 178, + 168 + ], + "score": 1.0, + "content": "generated against", + "type": "text" + }, + { + "bbox": [ + 178, + 156, + 184, + 167 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 155, + 288, + 168 + ], + "score": 1.0, + "content": "-th model and evaluate on", + "type": "text" + }, + { + "bbox": [ + 289, + 156, + 293, + 165 + ], + "score": 0.7, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 155, + 507, + 168 + ], + "score": 1.0, + "content": "-th model. Results of CW are listed in brackets. Left:", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 231, + 178 + ], + "score": 1.0, + "content": "Results generated with G-FPR", + "type": "text" + }, + { + "bbox": [ + 231, + 165, + 262, + 176 + ], + "score": 0.48, + "content": "= 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 165, + 308, + 178 + ], + "score": 1.0, + "content": "and T-FPR", + "type": "text" + }, + { + "bbox": [ + 308, + 165, + 339, + 176 + ], + "score": 0.52, + "content": "= 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 165, + 496, + 178 + ], + "score": 1.0, + "content": "; Right: Results generated with G-FPR", + "type": "text" + }, + { + "bbox": [ + 496, + 167, + 505, + 176 + ], + "score": 0.68, + "content": "=", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 176, + 209, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 128, + 188 + ], + "score": 0.89, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 176, + 146, + 188 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 146, + 177, + 204, + 187 + ], + "score": 0.77, + "content": "\\mathrm { T } \\mathrm { \\mathrm { - F P R } } = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 176, + 209, + 188 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + } + ], + "index": 4.25 + }, + { + "type": "text", + "bbox": [ + 221, + 128, + 234, + 138 + ], + "lines": [ + { + "bbox": [ + 220, + 128, + 235, + 141 + ], + "spans": [ + { + "bbox": [ + 220, + 128, + 235, + 141 + ], + "score": 1.0, + "content": "(a)", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 505, + 319 + ], + "lines": [ + { + "bbox": [ + 106, + 198, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 506, + 211 + ], + "score": 1.0, + "content": "Query-free black-box API attack. In this experiment, we generate adversarial examples against R-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 153, + 222 + ], + "score": 1.0, + "content": "101-S with", + "type": "text" + }, + { + "bbox": [ + 153, + 209, + 261, + 221 + ], + "score": 0.77, + "content": "\\mathrm { G } \\mathrm { - } \\mathrm { F P R } = 1 0 ^ { - 3 } ( \\kappa = 1 . 2 4 )", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 208, + 265, + 222 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 265, + 209, + 373, + 221 + ], + "score": 0.8, + "content": "\\mathbf { G } \\mathbf { - } \\mathbf { F P R } \\mathbf { \\bar { = } } 1 0 ^ { - 4 } ( \\kappa = \\mathbf { \\bar { 0 } } . 6 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 208, + 426, + 222 + ], + "score": 1.0, + "content": ", and G-FPR", + "type": "text" + }, + { + "bbox": [ + 426, + 209, + 502, + 221 + ], + "score": 0.82, + "content": "< 1 0 ^ { - 4 } ( \\kappa = 0 . 2 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 208, + 506, + 222 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 407, + 232 + ], + "score": 1.0, + "content": "respectively. We evaluate our algorithm on two industry level APIs, namely,", + "type": "text" + }, + { + "bbox": [ + 407, + 221, + 438, + 231 + ], + "score": 0.48, + "content": "{ \\mathrm { F a c e } } + +", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 220, + 506, + 232 + ], + "score": 1.0, + "content": "and AliYun face", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 506, + 244 + ], + "score": 1.0, + "content": "verification platform. Since attack transferability has never been explored in concurrent work that", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 507, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 299, + 255 + ], + "score": 1.0, + "content": "generates semantic adversarial examples, we use", + "type": "text" + }, + { + "bbox": [ + 299, + 243, + 312, + 254 + ], + "score": 0.9, + "content": "{ \\mathcal { L } } _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 241, + 507, + 255 + ], + "score": 1.0, + "content": "bounded pixel-wise methods (Carlini & Wagner,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 253, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 506, + 266 + ], + "score": 1.0, + "content": "2017; Dong et al., 2018; Xie et al., 2019) as our baselines. As we see Table 2, which shows the best", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "performance of each method, our SemanticAdv achieves much higher attack success rate than CW in", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 274, + 504, + 288 + ], + "spans": [ + { + "bbox": [ + 104, + 274, + 471, + 288 + ], + "score": 1.0, + "content": "both APIs with all FPR thresholds (e.g., our adversarial examples generated with G-FPR", + "type": "text" + }, + { + "bbox": [ + 471, + 275, + 504, + 286 + ], + "score": 0.82, + "content": "< 1 0 ^ { - 4 }", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 285, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 142, + 298 + ], + "score": 1.0, + "content": "achieves", + "type": "text" + }, + { + "bbox": [ + 142, + 286, + 175, + 297 + ], + "score": 0.89, + "content": "6 4 . 6 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 285, + 349, + 298 + ], + "score": 1.0, + "content": "attack success rate on Face++ platform with", + "type": "text" + }, + { + "bbox": [ + 350, + 286, + 411, + 297 + ], + "score": 0.87, + "content": "\\mathrm { T } \\mathrm { \\bar { F } P R } = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 285, + 505, + 298 + ], + "score": 1.0, + "content": "). In addition, we found", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 297, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 506, + 308 + ], + "score": 1.0, + "content": "that lower G-FPR can achieve higher attack success rate on APIs within the same T-FPR (see Table I", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 307, + 164, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 164, + 320 + ], + "score": 1.0, + "content": "in Appendix).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 113, + 329, + 496, + 396 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 113, + 329, + 496, + 396 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 113, + 329, + 496, + 396 + ], + "spans": [ + { + "bbox": [ + 113, + 329, + 496, + 396 + ], + "score": 0.979, + "html": "
API name Attacker/Evaluation MetricFace++AliYun
T-FPR=10 -3T-FPR=10 -4T-FPR=10 -3T-FPR=10 -4
Dong et al. (2018)30.7721.0318.006.50
Xie et al. (2019)37.9525.6421.5011.00
CW (G-FPR<10-4)41.6224.3719.0012.00
SemanticAdv (G-FPR <10-4)64.6342.6935.5022.17
", + "type": "table", + "image_path": "f7239179d0edfb192fb9f0ec0f1d521cf4a0fa8658cf81e559552e303921b81e.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 113, + 329, + 496, + 351.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 113, + 351.3333333333333, + 496, + 373.66666666666663 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 113, + 373.66666666666663, + 496, + 395.99999999999994 + ], + "spans": [], + "index": 23 + } + ] + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 505, + 445 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 505, + 414 + ], + "score": 1.0, + "content": "Table 2: Quantitative analysis on query-free black-box attack. We use ResNet-101 optimized with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 412, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 362, + 424 + ], + "score": 1.0, + "content": "softmax loss for evaluation and report the attack success rate", + "type": "text" + }, + { + "bbox": [ + 363, + 412, + 378, + 423 + ], + "score": 0.75, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 412, + 505, + 424 + ], + "score": 1.0, + "content": "on two online face verification", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 373, + 435 + ], + "score": 1.0, + "content": "platforms. Note that for PGD-based attacks, we adopt MI-FGSM", + "type": "text" + }, + { + "bbox": [ + 373, + 423, + 398, + 433 + ], + "score": 0.86, + "content": "\\epsilon = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 423, + 505, + 435 + ], + "score": 1.0, + "content": ") in Dong et al. (2018) and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 432, + 330, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 135, + 444 + ], + "score": 0.34, + "content": "\\mathrm { \\dot { \\mathbf { M } } \\mathbf { - } D I ^ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 432, + 169, + 447 + ], + "score": 1.0, + "content": "-FGSM", + "type": "text" + }, + { + "bbox": [ + 170, + 434, + 194, + 444 + ], + "score": 0.84, + "content": "\\epsilon = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 432, + 330, + 447 + ], + "score": 1.0, + "content": ") in Xie et al. (2019), respectively.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 448, + 505, + 548 + ], + "lines": [ + { + "bbox": [ + 107, + 449, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 107, + 449, + 506, + 461 + ], + "score": 1.0, + "content": "User study. To measure the perceptual quality of the adversarial images generated by SemanticAdv,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 460, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 505, + 472 + ], + "score": 1.0, + "content": "we conduct a user study on Amazon Mechanical Turk (AMT). We use the adversarial examples", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 469, + 507, + 485 + ], + "spans": [ + { + "bbox": [ + 104, + 469, + 166, + 485 + ], + "score": 1.0, + "content": "generated with", + "type": "text" + }, + { + "bbox": [ + 167, + 471, + 229, + 482 + ], + "score": 0.67, + "content": "\\mathrm { G } \\mathrm { - } \\mathrm { F P R } < 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 469, + 507, + 485 + ], + "score": 1.0, + "content": ", which is the most strict setting in our experiment, to conduct the user", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 481, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 104, + 481, + 506, + 495 + ], + "score": 1.0, + "content": "study for both CW and SemanticAdv. In total, we collect 2, 620 annotations from 77 participants. In", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 492, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 165, + 504 + ], + "score": 0.91, + "content": "3 9 . 1 \\dot { 4 } \\pm 1 . 9 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 492, + 256, + 507 + ], + "score": 1.0, + "content": "(close to random guess", + "type": "text" + }, + { + "bbox": [ + 256, + 493, + 276, + 504 + ], + "score": 0.88, + "content": "5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 492, + 506, + 507 + ], + "score": 1.0, + "content": ") of trials the adversarial images generated by SemanticAdv", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 504, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 309, + 517 + ], + "score": 1.0, + "content": "are selected as reasonably-looking images and in", + "type": "text" + }, + { + "bbox": [ + 310, + 504, + 373, + 515 + ], + "score": 0.92, + "content": "3 0 . 2 7 \\pm 1 . 9 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 504, + 505, + 517 + ], + "score": 1.0, + "content": "of trials, the adversarial images", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 514, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 104, + 514, + 506, + 529 + ], + "score": 1.0, + "content": "generated by CW are selected as reasonably-looking. It indicates that SemanticAdv can generate", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "more reasonable-looking adversarial examples compared with CW under the most strict setting with", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 536, + 430, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 137, + 550 + ], + "score": 1.0, + "content": "G-FPR", + "type": "text" + }, + { + "bbox": [ + 137, + 536, + 169, + 547 + ], + "score": 0.84, + "content": "< 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 536, + 430, + 550 + ], + "score": 1.0, + "content": ". Qualitative comparisons are shown in Appendix (see Figure H).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 127, + 559, + 482, + 655 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 127, + 559, + 482, + 655 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 559, + 482, + 655 + ], + "spans": [ + { + "bbox": [ + 127, + 559, + 482, + 655 + ], + "score": 0.967, + "type": "image", + "image_path": "e038d18d452da8c19a0e36bbebaafe417ea1f9ce207ae878fb70f59f617d2e80.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 127, + 559, + 482, + 591.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 127, + 591.0, + 482, + 623.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 127, + 623.0, + 482, + 655.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 662, + 504, + 684 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 660, + 504, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 504, + 673 + ], + "score": 1.0, + "content": "Figure 5: Quantitative analysis on attacking several defense methods including JPEG (Dziugaite", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 672, + 468, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 468, + 685 + ], + "score": 1.0, + "content": "et al., 2016), Blurring (Li & Li, 2017), and Feature Squeezing (Xu et al., 2017).", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + } + ], + "index": 39.25 + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "SemanticAdv against defense methods. We evaluate the strength of the proposed attack by test-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "ing against four existing defense methods, namely, Feature squeezing (Xu et al., 2017),", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "Blurring (Li & Li, 2017), JPEG (Dziugaite et al., 2016) and AMI (Tao et al., 2018). For AMI (Tao", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "et al., 2018), we first extract attribute witnesses with our aligned face images and then leverage them", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 350, + 80, + 491, + 124 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 350, + 80, + 491, + 124 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 350, + 80, + 491, + 124 + ], + "spans": [ + { + "bbox": [ + 350, + 80, + 491, + 124 + ], + "score": 0.797, + "html": "
Mtest/MoptR-50-SR-101-S
R-50-S1.000 (1.000)0.862 (0.530)
R-101-S0.874 (0.422)1.000 (1.000)
R-50-C0.693(0.347)0.837 (0.579)
R-101-C0.617 (0.218)0.888 (0.617)
", + "type": "table", + "image_path": "57469e8d95a46fe820286c74e8b8ee4b6109015ae65e817728e3fc376178968b.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 350, + 80, + 491, + 102.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 350, + 102.0, + 491, + 124.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 3.5 + }, + { + "type": "table", + "bbox": [ + 111, + 80, + 345, + 124 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 111, + 80, + 345, + 124 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 80, + 345, + 124 + ], + "spans": [ + { + "bbox": [ + 111, + 80, + 345, + 124 + ], + "score": 0.963, + "html": "
MtestMoptR-50-SR-101-SR-50-CR-101-C
R-50-S1.000 (1.000)0.108 (0.007)0.023(0.002)0.018 (0.002)
R-101-S0.169 (0.006)1.000 (1.000)0.030 (0.002)0.032(0.003)
R-50-C0.166 (0.019)0.202(0.025)1.000 (1.000)0.048(0.007)
R-101-C0.120 (0.015)0.236 (0.029)0.040 (0.006)1.000(1.000)
", + "type": "table", + "image_path": "a3f40dce747cd167588dcf20b5c1fc20e4dff6329527f1a11479082ce178844d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 80, + 345, + 94.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 94.66666666666667, + 345, + 109.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 109.33333333333334, + 345, + 124.00000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 144, + 505, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 144, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 285, + 156 + ], + "score": 1.0, + "content": "Table 1: Transferability of SemanticAdv: cell", + "type": "text" + }, + { + "bbox": [ + 286, + 144, + 307, + 156 + ], + "score": 0.91, + "content": "( i , j )", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 144, + 506, + 156 + ], + "score": 1.0, + "content": "shows attack success rate of adversarial examples", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 155, + 507, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 178, + 168 + ], + "score": 1.0, + "content": "generated against", + "type": "text" + }, + { + "bbox": [ + 178, + 156, + 184, + 167 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 155, + 288, + 168 + ], + "score": 1.0, + "content": "-th model and evaluate on", + "type": "text" + }, + { + "bbox": [ + 289, + 156, + 293, + 165 + ], + "score": 0.7, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 155, + 507, + 168 + ], + "score": 1.0, + "content": "-th model. Results of CW are listed in brackets. Left:", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 231, + 178 + ], + "score": 1.0, + "content": "Results generated with G-FPR", + "type": "text" + }, + { + "bbox": [ + 231, + 165, + 262, + 176 + ], + "score": 0.48, + "content": "= 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 165, + 308, + 178 + ], + "score": 1.0, + "content": "and T-FPR", + "type": "text" + }, + { + "bbox": [ + 308, + 165, + 339, + 176 + ], + "score": 0.52, + "content": "= 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 165, + 496, + 178 + ], + "score": 1.0, + "content": "; Right: Results generated with G-FPR", + "type": "text" + }, + { + "bbox": [ + 496, + 167, + 505, + 176 + ], + "score": 0.68, + "content": "=", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 176, + 209, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 128, + 188 + ], + "score": 0.89, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 176, + 146, + 188 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 146, + 177, + 204, + 187 + ], + "score": 0.77, + "content": "\\mathrm { T } \\mathrm { \\mathrm { - F P R } } = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 176, + 209, + 188 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + } + ], + "index": 4.25 + }, + { + "type": "text", + "bbox": [ + 221, + 128, + 234, + 138 + ], + "lines": [ + { + "bbox": [ + 220, + 128, + 235, + 141 + ], + "spans": [ + { + "bbox": [ + 220, + 128, + 235, + 141 + ], + "score": 1.0, + "content": "(a)", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5, + "bbox_fs": [ + 220, + 128, + 235, + 141 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 505, + 319 + ], + "lines": [ + { + "bbox": [ + 106, + 198, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 506, + 211 + ], + "score": 1.0, + "content": "Query-free black-box API attack. In this experiment, we generate adversarial examples against R-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 153, + 222 + ], + "score": 1.0, + "content": "101-S with", + "type": "text" + }, + { + "bbox": [ + 153, + 209, + 261, + 221 + ], + "score": 0.77, + "content": "\\mathrm { G } \\mathrm { - } \\mathrm { F P R } = 1 0 ^ { - 3 } ( \\kappa = 1 . 2 4 )", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 208, + 265, + 222 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 265, + 209, + 373, + 221 + ], + "score": 0.8, + "content": "\\mathbf { G } \\mathbf { - } \\mathbf { F P R } \\mathbf { \\bar { = } } 1 0 ^ { - 4 } ( \\kappa = \\mathbf { \\bar { 0 } } . 6 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 208, + 426, + 222 + ], + "score": 1.0, + "content": ", and G-FPR", + "type": "text" + }, + { + "bbox": [ + 426, + 209, + 502, + 221 + ], + "score": 0.82, + "content": "< 1 0 ^ { - 4 } ( \\kappa = 0 . 2 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 208, + 506, + 222 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 407, + 232 + ], + "score": 1.0, + "content": "respectively. We evaluate our algorithm on two industry level APIs, namely,", + "type": "text" + }, + { + "bbox": [ + 407, + 221, + 438, + 231 + ], + "score": 0.48, + "content": "{ \\mathrm { F a c e } } + +", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 220, + 506, + 232 + ], + "score": 1.0, + "content": "and AliYun face", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 506, + 244 + ], + "score": 1.0, + "content": "verification platform. Since attack transferability has never been explored in concurrent work that", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 507, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 299, + 255 + ], + "score": 1.0, + "content": "generates semantic adversarial examples, we use", + "type": "text" + }, + { + "bbox": [ + 299, + 243, + 312, + 254 + ], + "score": 0.9, + "content": "{ \\mathcal { L } } _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 241, + 507, + 255 + ], + "score": 1.0, + "content": "bounded pixel-wise methods (Carlini & Wagner,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 253, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 506, + 266 + ], + "score": 1.0, + "content": "2017; Dong et al., 2018; Xie et al., 2019) as our baselines. As we see Table 2, which shows the best", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "performance of each method, our SemanticAdv achieves much higher attack success rate than CW in", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 274, + 504, + 288 + ], + "spans": [ + { + "bbox": [ + 104, + 274, + 471, + 288 + ], + "score": 1.0, + "content": "both APIs with all FPR thresholds (e.g., our adversarial examples generated with G-FPR", + "type": "text" + }, + { + "bbox": [ + 471, + 275, + 504, + 286 + ], + "score": 0.82, + "content": "< 1 0 ^ { - 4 }", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 285, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 142, + 298 + ], + "score": 1.0, + "content": "achieves", + "type": "text" + }, + { + "bbox": [ + 142, + 286, + 175, + 297 + ], + "score": 0.89, + "content": "6 4 . 6 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 285, + 349, + 298 + ], + "score": 1.0, + "content": "attack success rate on Face++ platform with", + "type": "text" + }, + { + "bbox": [ + 350, + 286, + 411, + 297 + ], + "score": 0.87, + "content": "\\mathrm { T } \\mathrm { \\bar { F } P R } = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 285, + 505, + 298 + ], + "score": 1.0, + "content": "). In addition, we found", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 297, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 506, + 308 + ], + "score": 1.0, + "content": "that lower G-FPR can achieve higher attack success rate on APIs within the same T-FPR (see Table I", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 307, + 164, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 164, + 320 + ], + "score": 1.0, + "content": "in Appendix).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15, + "bbox_fs": [ + 104, + 198, + 507, + 320 + ] + }, + { + "type": "table", + "bbox": [ + 113, + 329, + 496, + 396 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 113, + 329, + 496, + 396 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 113, + 329, + 496, + 396 + ], + "spans": [ + { + "bbox": [ + 113, + 329, + 496, + 396 + ], + "score": 0.979, + "html": "
API name Attacker/Evaluation MetricFace++AliYun
T-FPR=10 -3T-FPR=10 -4T-FPR=10 -3T-FPR=10 -4
Dong et al. (2018)30.7721.0318.006.50
Xie et al. (2019)37.9525.6421.5011.00
CW (G-FPR<10-4)41.6224.3719.0012.00
SemanticAdv (G-FPR <10-4)64.6342.6935.5022.17
", + "type": "table", + "image_path": "f7239179d0edfb192fb9f0ec0f1d521cf4a0fa8658cf81e559552e303921b81e.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 113, + 329, + 496, + 351.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 113, + 351.3333333333333, + 496, + 373.66666666666663 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 113, + 373.66666666666663, + 496, + 395.99999999999994 + ], + "spans": [], + "index": 23 + } + ] + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 505, + 445 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 505, + 414 + ], + "score": 1.0, + "content": "Table 2: Quantitative analysis on query-free black-box attack. We use ResNet-101 optimized with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 412, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 362, + 424 + ], + "score": 1.0, + "content": "softmax loss for evaluation and report the attack success rate", + "type": "text" + }, + { + "bbox": [ + 363, + 412, + 378, + 423 + ], + "score": 0.75, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 412, + 505, + 424 + ], + "score": 1.0, + "content": "on two online face verification", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 373, + 435 + ], + "score": 1.0, + "content": "platforms. Note that for PGD-based attacks, we adopt MI-FGSM", + "type": "text" + }, + { + "bbox": [ + 373, + 423, + 398, + 433 + ], + "score": 0.86, + "content": "\\epsilon = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 423, + 505, + 435 + ], + "score": 1.0, + "content": ") in Dong et al. (2018) and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 432, + 330, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 135, + 444 + ], + "score": 0.34, + "content": "\\mathrm { \\dot { \\mathbf { M } } \\mathbf { - } D I ^ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 432, + 169, + 447 + ], + "score": 1.0, + "content": "-FGSM", + "type": "text" + }, + { + "bbox": [ + 170, + 434, + 194, + 444 + ], + "score": 0.84, + "content": "\\epsilon = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 432, + 330, + 447 + ], + "score": 1.0, + "content": ") in Xie et al. (2019), respectively.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 400, + 505, + 447 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 448, + 505, + 548 + ], + "lines": [ + { + "bbox": [ + 107, + 449, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 107, + 449, + 506, + 461 + ], + "score": 1.0, + "content": "User study. To measure the perceptual quality of the adversarial images generated by SemanticAdv,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 460, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 505, + 472 + ], + "score": 1.0, + "content": "we conduct a user study on Amazon Mechanical Turk (AMT). We use the adversarial examples", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 469, + 507, + 485 + ], + "spans": [ + { + "bbox": [ + 104, + 469, + 166, + 485 + ], + "score": 1.0, + "content": "generated with", + "type": "text" + }, + { + "bbox": [ + 167, + 471, + 229, + 482 + ], + "score": 0.67, + "content": "\\mathrm { G } \\mathrm { - } \\mathrm { F P R } < 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 469, + 507, + 485 + ], + "score": 1.0, + "content": ", which is the most strict setting in our experiment, to conduct the user", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 481, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 104, + 481, + 506, + 495 + ], + "score": 1.0, + "content": "study for both CW and SemanticAdv. In total, we collect 2, 620 annotations from 77 participants. In", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 492, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 165, + 504 + ], + "score": 0.91, + "content": "3 9 . 1 \\dot { 4 } \\pm 1 . 9 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 492, + 256, + 507 + ], + "score": 1.0, + "content": "(close to random guess", + "type": "text" + }, + { + "bbox": [ + 256, + 493, + 276, + 504 + ], + "score": 0.88, + "content": "5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 492, + 506, + 507 + ], + "score": 1.0, + "content": ") of trials the adversarial images generated by SemanticAdv", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 504, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 309, + 517 + ], + "score": 1.0, + "content": "are selected as reasonably-looking images and in", + "type": "text" + }, + { + "bbox": [ + 310, + 504, + 373, + 515 + ], + "score": 0.92, + "content": "3 0 . 2 7 \\pm 1 . 9 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 504, + 505, + 517 + ], + "score": 1.0, + "content": "of trials, the adversarial images", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 514, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 104, + 514, + 506, + 529 + ], + "score": 1.0, + "content": "generated by CW are selected as reasonably-looking. It indicates that SemanticAdv can generate", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "more reasonable-looking adversarial examples compared with CW under the most strict setting with", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 536, + 430, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 137, + 550 + ], + "score": 1.0, + "content": "G-FPR", + "type": "text" + }, + { + "bbox": [ + 137, + 536, + 169, + 547 + ], + "score": 0.84, + "content": "< 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 536, + 430, + 550 + ], + "score": 1.0, + "content": ". Qualitative comparisons are shown in Appendix (see Figure H).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32, + "bbox_fs": [ + 104, + 449, + 507, + 550 + ] + }, + { + "type": "image", + "bbox": [ + 127, + 559, + 482, + 655 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 127, + 559, + 482, + 655 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 559, + 482, + 655 + ], + "spans": [ + { + "bbox": [ + 127, + 559, + 482, + 655 + ], + "score": 0.967, + "type": "image", + "image_path": "e038d18d452da8c19a0e36bbebaafe417ea1f9ce207ae878fb70f59f617d2e80.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 127, + 559, + 482, + 591.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 127, + 591.0, + 482, + 623.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 127, + 623.0, + 482, + 655.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 662, + 504, + 684 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 660, + 504, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 504, + 673 + ], + "score": 1.0, + "content": "Figure 5: Quantitative analysis on attacking several defense methods including JPEG (Dziugaite", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 672, + 468, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 468, + 685 + ], + "score": 1.0, + "content": "et al., 2016), Blurring (Li & Li, 2017), and Feature Squeezing (Xu et al., 2017).", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + } + ], + "index": 39.25 + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "SemanticAdv against defense methods. We evaluate the strength of the proposed attack by test-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "ing against four existing defense methods, namely, Feature squeezing (Xu et al., 2017),", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "Blurring (Li & Li, 2017), JPEG (Dziugaite et al., 2016) and AMI (Tao et al., 2018). For AMI (Tao", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "et al., 2018), we first extract attribute witnesses with our aligned face images and then leverage them", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 392, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 505, + 404 + ], + "score": 1.0, + "content": "to construct attribute-steered model. We use fc7 of pretrained VGG (Parkhi et al., 2015) as the face", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "representation. AMI yields a consistency score for each face image to indicate whether it is a benign", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 415, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 505, + 426 + ], + "score": 1.0, + "content": "image. The score is measured by the cosine similarity between the representations from original", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 272, + 438 + ], + "score": 1.0, + "content": "model and attribute-steered model. With", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 272, + 425, + 292, + 435 + ], + "score": 0.87, + "content": "1 0 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 292, + 425, + 489, + 438 + ], + "score": 1.0, + "content": "false positives on benign inputs, it only achieves", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 489, + 425, + 505, + 436 + ], + "score": 0.84, + "content": "8 \\%", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 435, + 401, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 269, + 448 + ], + "score": 1.0, + "content": "detection accuracy for SemanticAdv and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 269, + 436, + 289, + 447 + ], + "score": 0.87, + "content": "1 2 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 289, + 435, + 401, + 448 + ], + "score": 1.0, + "content": "detection accuracy for CW.", + "type": "text", + "cross_page": true + } + ], + "index": 12 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 687, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 80, + 499, + 183 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 80, + 499, + 183 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 80, + 499, + 183 + ], + "spans": [ + { + "bbox": [ + 108, + 80, + 499, + 183 + ], + "score": 0.965, + "type": "image", + "image_path": "2514972a90994ba8bb3e4bf4a64b6780f5d758436abdb8a62c957825f90e25ed.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 80, + 499, + 114.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 114.33333333333334, + 499, + 148.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 148.66666666666669, + 499, + 183.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 158, + 201, + 451, + 212 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 158, + 199, + 453, + 214 + ], + "spans": [ + { + "bbox": [ + 158, + 199, + 453, + 214 + ], + "score": 1.0, + "content": "Figure 6: Qualitative results on attacking face landmark detection model.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 110, + 222, + 499, + 355 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 222, + 499, + 355 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 222, + 499, + 355 + ], + "spans": [ + { + "bbox": [ + 110, + 222, + 499, + 355 + ], + "score": 0.969, + "type": "image", + "image_path": "136eee23c85c9f369b82e0445973320cf8e16d085aa0e61850f470aa410f04ff.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 110, + 222, + 499, + 266.3333333333333 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 110, + 266.3333333333333, + 499, + 310.66666666666663 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 110, + 310.66666666666663, + 499, + 354.99999999999994 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 137, + 362, + 472, + 374 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 136, + 361, + 474, + 375 + ], + "spans": [ + { + "bbox": [ + 136, + 361, + 474, + 375 + ], + "score": 1.0, + "content": "Figure 7: Qualitative results on attacking street-view semantic segmentation model.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 505, + 447 + ], + "lines": [ + { + "bbox": [ + 106, + 392, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 505, + 404 + ], + "score": 1.0, + "content": "to construct attribute-steered model. We use fc7 of pretrained VGG (Parkhi et al., 2015) as the face", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "representation. AMI yields a consistency score for each face image to indicate whether it is a benign", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 415, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 505, + 426 + ], + "score": 1.0, + "content": "image. The score is measured by the cosine similarity between the representations from original", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 272, + 438 + ], + "score": 1.0, + "content": "model and attribute-steered model. With", + "type": "text" + }, + { + "bbox": [ + 272, + 425, + 292, + 435 + ], + "score": 0.87, + "content": "1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 425, + 489, + 438 + ], + "score": 1.0, + "content": "false positives on benign inputs, it only achieves", + "type": "text" + }, + { + "bbox": [ + 489, + 425, + 505, + 436 + ], + "score": 0.84, + "content": "8 \\%", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 435, + 401, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 269, + 448 + ], + "score": 1.0, + "content": "detection accuracy for SemanticAdv and", + "type": "text" + }, + { + "bbox": [ + 269, + 436, + 289, + 447 + ], + "score": 0.87, + "content": "1 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 435, + 401, + 448 + ], + "score": 1.0, + "content": "detection accuracy for CW.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 505, + 519 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 506, + 464 + ], + "score": 1.0, + "content": "Figure 5 illustrates SemanticAdv is more robust against these defense methods comparing with CW.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 459, + 476 + ], + "score": 1.0, + "content": "The same G-FPR and T-FPR are used for evaluation. Under the condition that T-FPR is", + "type": "text" + }, + { + "bbox": [ + 459, + 464, + 480, + 475 + ], + "score": 0.89, + "content": "\\bar { 1 } 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 464, + 505, + 476 + ], + "score": 1.0, + "content": ", both", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 474, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 488 + ], + "score": 1.0, + "content": "SemanticAdv and CW achieve high attack success rate, while SemanticAdv marginally outperforms", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 224, + 498 + ], + "score": 1.0, + "content": "CW when FPR goes down to", + "type": "text" + }, + { + "bbox": [ + 225, + 485, + 246, + 496 + ], + "score": 0.9, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 484, + 506, + 498 + ], + "score": 1.0, + "content": ". While defense methods have proven to be effective against CW", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "attacks on classifiers trained with ImageNet (Krizhevsky et al., 2012), our results indicate that these", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 507, + 401, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 401, + 519 + ], + "score": 1.0, + "content": "methods are still vulnerable in face verification system with small T-FPR.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 107, + 534, + 331, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 534, + 332, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 332, + 547 + ], + "score": 1.0, + "content": "4.3 SemanticAdv ON FACE LANDMARK DETECTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 505, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 554, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 505, + 565 + ], + "score": 1.0, + "content": "We also evaluate the effectiveness of SemanticAdv on face landmark detection. We select two attack", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 565, + 504, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 504, + 577 + ], + "score": 1.0, + "content": "tasks, namely, “Rotating Eyes” and “Out of Region”. For the “Rotating Eyes” task, we rotate the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 576, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 338, + 588 + ], + "score": 1.0, + "content": "coordinates of the eyes in the image counter-clockwise by", + "type": "text" + }, + { + "bbox": [ + 338, + 576, + 353, + 586 + ], + "score": 0.84, + "content": "9 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 576, + 506, + 588 + ], + "score": 1.0, + "content": ". For the “Out of Region” task, we set", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 587, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 600 + ], + "score": 1.0, + "content": "a target bounding box and attempt to push all points out of the box. We summarize the experimental", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "setup and quantitative results in the Appendix (see Table D). As we see in Figure 6, our method is", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 609, + 299, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 299, + 621 + ], + "score": 1.0, + "content": "applicable to attack landmark detection models.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5 + }, + { + "type": "title", + "bbox": [ + 107, + 635, + 379, + 647 + ], + "lines": [ + { + "bbox": [ + 106, + 635, + 380, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 380, + 648 + ], + "score": 1.0, + "content": "4.4 SemanticAdv ON STREET-VIEW SEMANTIC SEGMENTATION", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 653, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 505, + 669 + ], + "score": 1.0, + "content": "We further demonstrate the applicability of our SemanticAdv beyond face domain by generating", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "adversarial perturbations on street-view images. Figure 7 illustrates the adversarial examples on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "semantic segmentation. In the first example, we select the leftmost the pedestrian as the target", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "object instance and insert another car into the scene to attack it. The segmentation model has been", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "successfully attacked to neglect the pedestrian (see last column), while it does exist in the scene", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "(see second-to-last column). In the second example, we insert an adversarial car in the scene by", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 721, + 477, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 477, + 732 + ], + "score": 1.0, + "content": "SemanticAdv and the cyclist has been recognized as a pedestrian by the segmentation model.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 80, + 499, + 183 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 80, + 499, + 183 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 80, + 499, + 183 + ], + "spans": [ + { + "bbox": [ + 108, + 80, + 499, + 183 + ], + "score": 0.965, + "type": "image", + "image_path": "2514972a90994ba8bb3e4bf4a64b6780f5d758436abdb8a62c957825f90e25ed.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 80, + 499, + 114.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 114.33333333333334, + 499, + 148.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 148.66666666666669, + 499, + 183.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 158, + 201, + 451, + 212 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 158, + 199, + 453, + 214 + ], + "spans": [ + { + "bbox": [ + 158, + 199, + 453, + 214 + ], + "score": 1.0, + "content": "Figure 6: Qualitative results on attacking face landmark detection model.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 110, + 222, + 499, + 355 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 222, + 499, + 355 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 222, + 499, + 355 + ], + "spans": [ + { + "bbox": [ + 110, + 222, + 499, + 355 + ], + "score": 0.969, + "type": "image", + "image_path": "136eee23c85c9f369b82e0445973320cf8e16d085aa0e61850f470aa410f04ff.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 110, + 222, + 499, + 266.3333333333333 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 110, + 266.3333333333333, + 499, + 310.66666666666663 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 110, + 310.66666666666663, + 499, + 354.99999999999994 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 137, + 362, + 472, + 374 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 136, + 361, + 474, + 375 + ], + "spans": [ + { + "bbox": [ + 136, + 361, + 474, + 375 + ], + "score": 1.0, + "content": "Figure 7: Qualitative results on attacking street-view semantic segmentation model.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 505, + 447 + ], + "lines": [], + "index": 10, + "bbox_fs": [ + 105, + 392, + 505, + 448 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 505, + 519 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 506, + 464 + ], + "score": 1.0, + "content": "Figure 5 illustrates SemanticAdv is more robust against these defense methods comparing with CW.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 459, + 476 + ], + "score": 1.0, + "content": "The same G-FPR and T-FPR are used for evaluation. Under the condition that T-FPR is", + "type": "text" + }, + { + "bbox": [ + 459, + 464, + 480, + 475 + ], + "score": 0.89, + "content": "\\bar { 1 } 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 464, + 505, + 476 + ], + "score": 1.0, + "content": ", both", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 474, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 488 + ], + "score": 1.0, + "content": "SemanticAdv and CW achieve high attack success rate, while SemanticAdv marginally outperforms", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 224, + 498 + ], + "score": 1.0, + "content": "CW when FPR goes down to", + "type": "text" + }, + { + "bbox": [ + 225, + 485, + 246, + 496 + ], + "score": 0.9, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 484, + 506, + 498 + ], + "score": 1.0, + "content": ". While defense methods have proven to be effective against CW", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "attacks on classifiers trained with ImageNet (Krizhevsky et al., 2012), our results indicate that these", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 507, + 401, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 401, + 519 + ], + "score": 1.0, + "content": "methods are still vulnerable in face verification system with small T-FPR.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 453, + 506, + 519 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 534, + 331, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 534, + 332, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 332, + 547 + ], + "score": 1.0, + "content": "4.3 SemanticAdv ON FACE LANDMARK DETECTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 505, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 554, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 505, + 565 + ], + "score": 1.0, + "content": "We also evaluate the effectiveness of SemanticAdv on face landmark detection. We select two attack", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 565, + 504, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 504, + 577 + ], + "score": 1.0, + "content": "tasks, namely, “Rotating Eyes” and “Out of Region”. For the “Rotating Eyes” task, we rotate the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 576, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 338, + 588 + ], + "score": 1.0, + "content": "coordinates of the eyes in the image counter-clockwise by", + "type": "text" + }, + { + "bbox": [ + 338, + 576, + 353, + 586 + ], + "score": 0.84, + "content": "9 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 576, + 506, + 588 + ], + "score": 1.0, + "content": ". For the “Out of Region” task, we set", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 587, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 600 + ], + "score": 1.0, + "content": "a target bounding box and attempt to push all points out of the box. We summarize the experimental", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "setup and quantitative results in the Appendix (see Table D). As we see in Figure 6, our method is", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 609, + 299, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 299, + 621 + ], + "score": 1.0, + "content": "applicable to attack landmark detection models.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 554, + 506, + 621 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 635, + 379, + 647 + ], + "lines": [ + { + "bbox": [ + 106, + 635, + 380, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 380, + 648 + ], + "score": 1.0, + "content": "4.4 SemanticAdv ON STREET-VIEW SEMANTIC SEGMENTATION", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 653, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 505, + 669 + ], + "score": 1.0, + "content": "We further demonstrate the applicability of our SemanticAdv beyond face domain by generating", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "adversarial perturbations on street-view images. Figure 7 illustrates the adversarial examples on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "semantic segmentation. In the first example, we select the leftmost the pedestrian as the target", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "object instance and insert another car into the scene to attack it. The segmentation model has been", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "successfully attacked to neglect the pedestrian (see last column), while it does exist in the scene", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "(see second-to-last column). In the second example, we insert an adversarial car in the scene by", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 721, + 477, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 477, + 732 + ], + "score": 1.0, + "content": "SemanticAdv and the cyclist has been recognized as a pedestrian by the segmentation model.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 653, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 201, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 78, + 203, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 78, + 203, + 97 + ], + "score": 1.0, + "content": "5 CONCLUSIONS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 97, + 505, + 174 + ], + "lines": [ + { + "bbox": [ + 106, + 97, + 505, + 110 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 505, + 110 + ], + "score": 1.0, + "content": "Overall we presented a novel attack method SemanticAdv, which is capable of generating unrestricted", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 109, + 506, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 109, + 506, + 122 + ], + "score": 1.0, + "content": "adversarial perturbations guided by semantic attributes edition. Compared to existing methods,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 120, + 504, + 131 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 504, + 131 + ], + "score": 1.0, + "content": "SemanticAdv works in a more controllable fashion. Experimental evaluations on face verification", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 131, + 505, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 505, + 143 + ], + "score": 1.0, + "content": "and landmark detection demonstrate several unique properties including attack transferability. We", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 140, + 506, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 506, + 155 + ], + "score": 1.0, + "content": "believe this work would open up new research opportunities and challenges in the field of adversarial", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 152, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 506, + 165 + ], + "score": 1.0, + "content": "learning. For instance, how to leverage semantic information to defend against such attacks will lead", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 164, + 219, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 219, + 177 + ], + "score": 1.0, + "content": "to potential new discussion.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 107, + 191, + 175, + 204 + ], + "lines": [ + { + "bbox": [ + 106, + 191, + 176, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 176, + 205 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 105, + 207, + 507, + 739 + ], + "lines": [ + { + "bbox": [ + 105, + 210, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 505, + 224 + ], + "score": 1.0, + "content": "Alibaba Cloud Computing Co. Ltd. https://help.aliyun.com/knowledge_detail/", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 221, + 181, + 235 + ], + "spans": [ + { + "bbox": [ + 115, + 221, + 181, + 235 + ], + "score": 1.0, + "content": "53535.html.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "Megvii Technology Co. Ltd. https://console.faceplusplus.com/documents/", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 251, + 164, + 264 + ], + "spans": [ + { + "bbox": [ + 116, + 251, + 164, + 264 + ], + "score": 1.0, + "content": "5679308.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 270, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 506, + 286 + ], + "score": 1.0, + "content": "Yoshua Bengio, Grégoire Mesnil, Yann Dauphin, and Salah Rifai. Better mixing via deep representa-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 281, + 207, + 294 + ], + "spans": [ + { + "bbox": [ + 115, + 281, + 207, + 294 + ], + "score": 1.0, + "content": "tions. In ICML, 2013.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 301, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 505, + 314 + ], + "score": 1.0, + "content": "Anand Bhattad, Min Jin Chong, Kaizhao Liang, Bo Li, and David A Forsyth. Big but imperceptible", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 312, + 490, + 325 + ], + "spans": [ + { + "bbox": [ + 115, + 312, + 490, + 325 + ], + "score": 1.0, + "content": "adversarial perturbations via semantic manipulation. arXiv preprint arXiv:1904.06347, 2019.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "score": 1.0, + "content": "Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 343, + 249, + 356 + ], + "spans": [ + { + "bbox": [ + 115, + 343, + 249, + 356 + ], + "score": 1.0, + "content": "image synthesis. In ICLR, 2019.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "Tom B Brown, Nicholas Carlini, Chiyuan Zhang, Catherine Olsson, Paul Christiano, and Ian Good-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 372, + 451, + 387 + ], + "spans": [ + { + "bbox": [ + 115, + 372, + 451, + 387 + ], + "score": 1.0, + "content": "fellow. Unrestricted adversarial examples. arXiv preprint arXiv:1809.08352, 2018.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 391, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 406 + ], + "score": 1.0, + "content": "Adrian Bulat and Georgios Tzimiropoulos. Binarized convolutional landmark localizers for human", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 402, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 115, + 402, + 506, + 416 + ], + "score": 1.0, + "content": "pose estimation and face alignment with limited resources. In Proceedings of the IEEE International", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 117, + 414, + 343, + 427 + ], + "spans": [ + { + "bbox": [ + 117, + 414, + 343, + 427 + ], + "score": 1.0, + "content": "Conference on Computer Vision, pp. 3706–3714, 2017a.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 431, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 506, + 447 + ], + "score": 1.0, + "content": "Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 443, + 414, + 457 + ], + "spans": [ + { + "bbox": [ + 115, + 443, + 414, + 457 + ], + "score": 1.0, + "content": "problem?(and a dataset of 230,000 3d facial landmarks). In ICCV, 2017b.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 463, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 506, + 476 + ], + "score": 1.0, + "content": "Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 474, + 369, + 487 + ], + "spans": [ + { + "bbox": [ + 115, + 474, + 369, + 487 + ], + "score": 1.0, + "content": "IEEE Symposium on Security and Privacy (S&P). IEEE, 2017.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 493, + 507, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 507, + 507 + ], + "score": 1.0, + "content": "Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan:", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 114, + 503, + 507, + 518 + ], + "spans": [ + { + "bbox": [ + 114, + 503, + 507, + 518 + ], + "score": 1.0, + "content": "Unified generative adversarial networks for multi-domain image-to-image translation. In CVPR,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 514, + 143, + 528 + ], + "spans": [ + { + "bbox": [ + 115, + 514, + 143, + 528 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 533, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 104, + 533, + 506, + 548 + ], + "score": 1.0, + "content": "Moustapha Cisse, Yossi Adi, Natalia Neverova, and Joseph Keshet. Houdini: Fooling deep structured", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 545, + 256, + 558 + ], + "spans": [ + { + "bbox": [ + 115, + 545, + 256, + 558 + ], + "score": 1.0, + "content": "prediction models. In NIPS, 2017.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 562, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 579 + ], + "score": 1.0, + "content": "Ronan Collobert and Jason Weston. A unified architecture for natural language processing: Deep", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 114, + 574, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 114, + 574, + 506, + 590 + ], + "score": 1.0, + "content": "neural networks with multitask learning. In Proceedings of the 25th international conference on", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 586, + 298, + 599 + ], + "spans": [ + { + "bbox": [ + 115, + 586, + 298, + 599 + ], + "score": 1.0, + "content": "Machine learning, pp. 160–167. ACM, 2008.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 604, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 620 + ], + "score": 1.0, + "content": "Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 617, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 506, + 630 + ], + "score": 1.0, + "content": "Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 113, + 626, + 506, + 643 + ], + "spans": [ + { + "bbox": [ + 113, + 626, + 506, + 643 + ], + "score": 1.0, + "content": "scene understanding. In Proceedings of the IEEE conference on computer vision and pattern", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 638, + 257, + 652 + ], + "spans": [ + { + "bbox": [ + 115, + 638, + 257, + 652 + ], + "score": 1.0, + "content": "recognition, pp. 3213–3223, 2016.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 657, + 507, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 507, + 672 + ], + "score": 1.0, + "content": "Ltsc Deng, Jinyu Li, Jui-Ting Huang, Kaisheng Yao, Dong Yu, Frank Seide, Michael L Seltzer,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 668, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 115, + 668, + 506, + 683 + ], + "score": 1.0, + "content": "Geoffrey Zweig, Xiaodong He, Jason D Williams, et al. Recent advances in deep learning for", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 679, + 388, + 694 + ], + "spans": [ + { + "bbox": [ + 115, + 679, + 388, + 694 + ], + "score": 1.0, + "content": "speech research at microsoft. In ICASSP, volume 26, pp. 64, 2013.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 697, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 506, + 714 + ], + "score": 1.0, + "content": "Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 710, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 506, + 725 + ], + "score": 1.0, + "content": "adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 114, + 720, + 306, + 735 + ], + "spans": [ + { + "bbox": [ + 114, + 720, + 306, + 735 + ], + "score": 1.0, + "content": "and pattern recognition, pp. 9185–9193, 2018.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 27 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 201, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 78, + 203, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 78, + 203, + 97 + ], + "score": 1.0, + "content": "5 CONCLUSIONS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 97, + 505, + 174 + ], + "lines": [ + { + "bbox": [ + 106, + 97, + 505, + 110 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 505, + 110 + ], + "score": 1.0, + "content": "Overall we presented a novel attack method SemanticAdv, which is capable of generating unrestricted", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 109, + 506, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 109, + 506, + 122 + ], + "score": 1.0, + "content": "adversarial perturbations guided by semantic attributes edition. Compared to existing methods,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 120, + 504, + 131 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 504, + 131 + ], + "score": 1.0, + "content": "SemanticAdv works in a more controllable fashion. Experimental evaluations on face verification", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 131, + 505, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 505, + 143 + ], + "score": 1.0, + "content": "and landmark detection demonstrate several unique properties including attack transferability. We", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 140, + 506, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 506, + 155 + ], + "score": 1.0, + "content": "believe this work would open up new research opportunities and challenges in the field of adversarial", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 152, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 506, + 165 + ], + "score": 1.0, + "content": "learning. For instance, how to leverage semantic information to defend against such attacks will lead", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 164, + 219, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 219, + 177 + ], + "score": 1.0, + "content": "to potential new discussion.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 97, + 506, + 177 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 191, + 175, + 204 + ], + "lines": [ + { + "bbox": [ + 106, + 191, + 176, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 176, + 205 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "list", + "bbox": [ + 105, + 207, + 507, + 739 + ], + "lines": [ + { + "bbox": [ + 105, + 210, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 505, + 224 + ], + "score": 1.0, + "content": "Alibaba Cloud Computing Co. Ltd. https://help.aliyun.com/knowledge_detail/", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 221, + 181, + 235 + ], + "spans": [ + { + "bbox": [ + 115, + 221, + 181, + 235 + ], + "score": 1.0, + "content": "53535.html.", + "type": "text" + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "Megvii Technology Co. Ltd. https://console.faceplusplus.com/documents/", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 251, + 164, + 264 + ], + "spans": [ + { + "bbox": [ + 116, + 251, + 164, + 264 + ], + "score": 1.0, + "content": "5679308.", + "type": "text" + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 270, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 506, + 286 + ], + "score": 1.0, + "content": "Yoshua Bengio, Grégoire Mesnil, Yann Dauphin, and Salah Rifai. Better mixing via deep representa-", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 281, + 207, + 294 + ], + "spans": [ + { + "bbox": [ + 115, + 281, + 207, + 294 + ], + "score": 1.0, + "content": "tions. In ICML, 2013.", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 301, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 505, + 314 + ], + "score": 1.0, + "content": "Anand Bhattad, Min Jin Chong, Kaizhao Liang, Bo Li, and David A Forsyth. Big but imperceptible", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 312, + 490, + 325 + ], + "spans": [ + { + "bbox": [ + 115, + 312, + 490, + 325 + ], + "score": 1.0, + "content": "adversarial perturbations via semantic manipulation. arXiv preprint arXiv:1904.06347, 2019.", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "score": 1.0, + "content": "Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 343, + 249, + 356 + ], + "spans": [ + { + "bbox": [ + 115, + 343, + 249, + 356 + ], + "score": 1.0, + "content": "image synthesis. In ICLR, 2019.", + "type": "text" + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "Tom B Brown, Nicholas Carlini, Chiyuan Zhang, Catherine Olsson, Paul Christiano, and Ian Good-", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 372, + 451, + 387 + ], + "spans": [ + { + "bbox": [ + 115, + 372, + 451, + 387 + ], + "score": 1.0, + "content": "fellow. Unrestricted adversarial examples. arXiv preprint arXiv:1809.08352, 2018.", + "type": "text" + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 391, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 406 + ], + "score": 1.0, + "content": "Adrian Bulat and Georgios Tzimiropoulos. Binarized convolutional landmark localizers for human", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 402, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 115, + 402, + 506, + 416 + ], + "score": 1.0, + "content": "pose estimation and face alignment with limited resources. In Proceedings of the IEEE International", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 117, + 414, + 343, + 427 + ], + "spans": [ + { + "bbox": [ + 117, + 414, + 343, + 427 + ], + "score": 1.0, + "content": "Conference on Computer Vision, pp. 3706–3714, 2017a.", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 431, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 506, + 447 + ], + "score": 1.0, + "content": "Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 443, + 414, + 457 + ], + "spans": [ + { + "bbox": [ + 115, + 443, + 414, + 457 + ], + "score": 1.0, + "content": "problem?(and a dataset of 230,000 3d facial landmarks). In ICCV, 2017b.", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 463, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 506, + 476 + ], + "score": 1.0, + "content": "Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 474, + 369, + 487 + ], + "spans": [ + { + "bbox": [ + 115, + 474, + 369, + 487 + ], + "score": 1.0, + "content": "IEEE Symposium on Security and Privacy (S&P). IEEE, 2017.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 493, + 507, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 507, + 507 + ], + "score": 1.0, + "content": "Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan:", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 503, + 507, + 518 + ], + "spans": [ + { + "bbox": [ + 114, + 503, + 507, + 518 + ], + "score": 1.0, + "content": "Unified generative adversarial networks for multi-domain image-to-image translation. In CVPR,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 514, + 143, + 528 + ], + "spans": [ + { + "bbox": [ + 115, + 514, + 143, + 528 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 533, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 104, + 533, + 506, + 548 + ], + "score": 1.0, + "content": "Moustapha Cisse, Yossi Adi, Natalia Neverova, and Joseph Keshet. Houdini: Fooling deep structured", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 545, + 256, + 558 + ], + "spans": [ + { + "bbox": [ + 115, + 545, + 256, + 558 + ], + "score": 1.0, + "content": "prediction models. In NIPS, 2017.", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 562, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 579 + ], + "score": 1.0, + "content": "Ronan Collobert and Jason Weston. A unified architecture for natural language processing: Deep", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 574, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 114, + 574, + 506, + 590 + ], + "score": 1.0, + "content": "neural networks with multitask learning. In Proceedings of the 25th international conference on", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 586, + 298, + 599 + ], + "spans": [ + { + "bbox": [ + 115, + 586, + 298, + 599 + ], + "score": 1.0, + "content": "Machine learning, pp. 160–167. ACM, 2008.", + "type": "text" + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 604, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 620 + ], + "score": 1.0, + "content": "Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 617, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 506, + 630 + ], + "score": 1.0, + "content": "Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 113, + 626, + 506, + 643 + ], + "spans": [ + { + "bbox": [ + 113, + 626, + 506, + 643 + ], + "score": 1.0, + "content": "scene understanding. In Proceedings of the IEEE conference on computer vision and pattern", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 638, + 257, + 652 + ], + "spans": [ + { + "bbox": [ + 115, + 638, + 257, + 652 + ], + "score": 1.0, + "content": "recognition, pp. 3213–3223, 2016.", + "type": "text" + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 657, + 507, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 507, + 672 + ], + "score": 1.0, + "content": "Ltsc Deng, Jinyu Li, Jui-Ting Huang, Kaisheng Yao, Dong Yu, Frank Seide, Michael L Seltzer,", + "type": "text" + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 668, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 115, + 668, + 506, + 683 + ], + "score": 1.0, + "content": "Geoffrey Zweig, Xiaodong He, Jason D Williams, et al. Recent advances in deep learning for", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 679, + 388, + 694 + ], + "spans": [ + { + "bbox": [ + 115, + 679, + 388, + 694 + ], + "score": 1.0, + "content": "speech research at microsoft. In ICASSP, volume 26, pp. 64, 2013.", + "type": "text" + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 697, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 506, + 714 + ], + "score": 1.0, + "content": "Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting", + "type": "text" + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 710, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 506, + 725 + ], + "score": 1.0, + "content": "adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 114, + 720, + 306, + 735 + ], + "spans": [ + { + "bbox": [ + 114, + 720, + 306, + 735 + ], + "score": 1.0, + "content": "and pattern recognition, pp. 9185–9193, 2018.", + "type": "text" + } + ], + "index": 45, + "is_list_end_line": true + } + ], + "index": 27, + "bbox_fs": [ + 104, + 210, + 507, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 505, + 97 + ], + "score": 1.0, + "content": "Gintare Karolina Dziugaite, Zoubin Ghahramani, and Daniel M Roy. A study of the effect of jpg", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 424, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 424, + 106 + ], + "score": 1.0, + "content": "compression on adversarial images. arXiv preprint arXiv:1608.00853, 2016.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 504, + 145 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "score": 1.0, + "content": "Logan Engstrom, Brandon Tran, Dimitris Tsipras, Ludwig Schmidt, and Aleksander Madry. A", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "rotation and a translation suffice: Fooling cnns with simple transformations. arXiv preprint", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 219, + 145 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 219, + 145 + ], + "score": 1.0, + "content": "arXiv:1712.02779, 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 104, + 152, + 505, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 152, + 506, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 506, + 166 + ], + "score": 1.0, + "content": "Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 163, + 450, + 176 + ], + "spans": [ + { + "bbox": [ + 116, + 163, + 450, + 176 + ], + "score": 1.0, + "content": "Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014a.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 182, + 505, + 205 + ], + "lines": [ + { + "bbox": [ + 104, + 180, + 507, + 196 + ], + "spans": [ + { + "bbox": [ + 104, + 180, + 507, + 196 + ], + "score": 1.0, + "content": "Ian Goodfellow, Yoshua Bengio, Aaron Courville, and Yoshua Bengio. Deep learning, volume 1.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 193, + 188, + 205 + ], + "spans": [ + { + "bbox": [ + 116, + 193, + 188, + 205 + ], + "score": 1.0, + "content": "MIT Press, 2016.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 212, + 504, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 212, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 505, + 225 + ], + "score": 1.0, + "content": "Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 223, + 227, + 235 + ], + "spans": [ + { + "bbox": [ + 115, + 223, + 227, + 235 + ], + "score": 1.0, + "content": "examples. In ICLR, 2014b.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 241, + 504, + 265 + ], + "lines": [ + { + "bbox": [ + 107, + 241, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 107, + 241, + 506, + 254 + ], + "score": 1.0, + "content": "Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 253, + 398, + 265 + ], + "spans": [ + { + "bbox": [ + 116, + 253, + 398, + 265 + ], + "score": 1.0, + "content": "benchmark for large-scale face recognition. In ECCV. Springer, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 271, + 504, + 294 + ], + "lines": [ + { + "bbox": [ + 105, + 269, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 506, + 286 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 283, + 234, + 294 + ], + "spans": [ + { + "bbox": [ + 115, + 283, + 234, + 294 + ], + "score": 1.0, + "content": "recognition. In CVPR, 2016.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 300, + 504, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 300, + 504, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 504, + 313 + ], + "score": 1.0, + "content": "Seunghoon Hong, Xinchen Yan, Thomas S Huang, and Honglak Lee. Learning hierarchical semantic", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 312, + 416, + 325 + ], + "spans": [ + { + "bbox": [ + 115, + 312, + 416, + 325 + ], + "score": 1.0, + "content": "image manipulation through structured representations. In NeurIPS, 2018.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 108, + 330, + 505, + 365 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 506, + 343 + ], + "score": 1.0, + "content": "Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild:", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 115, + 342, + 505, + 355 + ], + "score": 1.0, + "content": "A database forstudying face recognition in unconstrained environments. In Workshop on faces", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 353, + 381, + 365 + ], + "spans": [ + { + "bbox": [ + 115, + 353, + 381, + 365 + ], + "score": 1.0, + "content": "in’Real-Life’Images: detection, alignment, and recognition, 2008.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 104, + 371, + 504, + 394 + ], + "lines": [ + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "score": 1.0, + "content": "Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 382, + 383, + 395 + ], + "spans": [ + { + "bbox": [ + 115, + 382, + 383, + 395 + ], + "score": 1.0, + "content": "conditional adversarial networks. In CVPR, pp. 1125–1134, 2017.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 108, + 401, + 503, + 424 + ], + "lines": [ + { + "bbox": [ + 106, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 505, + 414 + ], + "score": 1.0, + "content": "Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 412, + 292, + 425 + ], + "spans": [ + { + "bbox": [ + 115, + 412, + 292, + 425 + ], + "score": 1.0, + "content": "super-resolution. In ECCV. Springer, 2016.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 430, + 505, + 454 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 506, + 446 + ], + "score": 1.0, + "content": "Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In CVPR, pp.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 442, + 190, + 454 + ], + "spans": [ + { + "bbox": [ + 116, + 442, + 190, + 454 + ], + "score": 1.0, + "content": "1219–1228, 2018.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 460, + 505, + 494 + ], + "lines": [ + { + "bbox": [ + 106, + 460, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 506, + 474 + ], + "score": 1.0, + "content": "Ameya Joshi, Amitangshu Mukherjee, Soumik Sarkar, and Chinmay Hegde. Semantic adversarial", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 473, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 116, + 473, + 506, + 484 + ], + "score": 1.0, + "content": "attacks: Parametric transformations that fool deep classifiers. arXiv preprint arXiv:1904.08489,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 482, + 143, + 495 + ], + "spans": [ + { + "bbox": [ + 115, + 482, + 143, + 495 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 501, + 503, + 524 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 505, + 515 + ], + "score": 1.0, + "content": "Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 513, + 346, + 525 + ], + "spans": [ + { + "bbox": [ + 116, + 513, + 346, + 525 + ], + "score": 1.0, + "content": "improved quality, stability, and variation. In ICLR, 2018.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 531, + 502, + 554 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 544 + ], + "score": 1.0, + "content": "Ira Kemelmacher-Shlizerman, Steven M Seitz, Daniel Miller, and Evan Brossard. The megaface", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 542, + 457, + 555 + ], + "spans": [ + { + "bbox": [ + 116, + 542, + 457, + 555 + ], + "score": 1.0, + "content": "benchmark: 1 million faces for recognition at scale. In CVPR, pp. 4873–4882, 2016.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 496, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 560, + 497, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 497, + 574 + ], + "score": 1.0, + "content": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 579, + 460, + 591 + ], + "lines": [ + { + "bbox": [ + 106, + 578, + 460, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 460, + 593 + ], + "score": 1.0, + "content": "Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 108, + 598, + 504, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 598, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 505, + 610 + ], + "score": 1.0, + "content": "Brendan F Klare, Ben Klein, Emma Taborsky, Austin Blanton, Jordan Cheney, Kristen Allen, Patrick", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 117, + 610, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 117, + 610, + 505, + 621 + ], + "score": 1.0, + "content": "Grother, Alan Mah, and Anil K Jain. Pushing the frontiers of unconstrained face detection and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 620, + 337, + 632 + ], + "spans": [ + { + "bbox": [ + 115, + 620, + 337, + 632 + ], + "score": 1.0, + "content": "recognition: Iarpa janus benchmark a. In CVPR, 2015.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 638, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 106, + 639, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 506, + 651 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolu-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 650, + 273, + 662 + ], + "spans": [ + { + "bbox": [ + 116, + 650, + 273, + 662 + ], + "score": 1.0, + "content": "tional neural networks. In NIPS, 2012.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 106, + 668, + 504, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "Xin Li and Fuxin Li. Adversarial examples detection in deep networks with convolutional filter", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 680, + 505, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 505, + 692 + ], + "score": 1.0, + "content": "statistics. In Proceedings of the IEEE International Conference on Computer Vision, pp. 5764–", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 690, + 167, + 703 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 167, + 703 + ], + "score": 1.0, + "content": "5772, 2017.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 105, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 720, + 168, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 168, + 732 + ], + "score": 1.0, + "content": "NIPS, 2017.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 505, + 97 + ], + "score": 1.0, + "content": "Gintare Karolina Dziugaite, Zoubin Ghahramani, and Daniel M Roy. A study of the effect of jpg", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 424, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 424, + 106 + ], + "score": 1.0, + "content": "compression on adversarial images. arXiv preprint arXiv:1608.00853, 2016.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 80, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 504, + 145 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "score": 1.0, + "content": "Logan Engstrom, Brandon Tran, Dimitris Tsipras, Ludwig Schmidt, and Aleksander Madry. A", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "rotation and a translation suffice: Fooling cnns with simple transformations. arXiv preprint", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 219, + 145 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 219, + 145 + ], + "score": 1.0, + "content": "arXiv:1712.02779, 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 106, + 112, + 505, + 145 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 152, + 505, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 152, + 506, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 506, + 166 + ], + "score": 1.0, + "content": "Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 163, + 450, + 176 + ], + "spans": [ + { + "bbox": [ + 116, + 163, + 450, + 176 + ], + "score": 1.0, + "content": "Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014a.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 152, + 506, + 176 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 182, + 505, + 205 + ], + "lines": [ + { + "bbox": [ + 104, + 180, + 507, + 196 + ], + "spans": [ + { + "bbox": [ + 104, + 180, + 507, + 196 + ], + "score": 1.0, + "content": "Ian Goodfellow, Yoshua Bengio, Aaron Courville, and Yoshua Bengio. Deep learning, volume 1.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 193, + 188, + 205 + ], + "spans": [ + { + "bbox": [ + 116, + 193, + 188, + 205 + ], + "score": 1.0, + "content": "MIT Press, 2016.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 104, + 180, + 507, + 205 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 212, + 504, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 212, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 505, + 225 + ], + "score": 1.0, + "content": "Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 223, + 227, + 235 + ], + "spans": [ + { + "bbox": [ + 115, + 223, + 227, + 235 + ], + "score": 1.0, + "content": "examples. In ICLR, 2014b.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 212, + 505, + 235 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 241, + 504, + 265 + ], + "lines": [ + { + "bbox": [ + 107, + 241, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 107, + 241, + 506, + 254 + ], + "score": 1.0, + "content": "Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 253, + 398, + 265 + ], + "spans": [ + { + "bbox": [ + 116, + 253, + 398, + 265 + ], + "score": 1.0, + "content": "benchmark for large-scale face recognition. In ECCV. Springer, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 107, + 241, + 506, + 265 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 271, + 504, + 294 + ], + "lines": [ + { + "bbox": [ + 105, + 269, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 506, + 286 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 283, + 234, + 294 + ], + "spans": [ + { + "bbox": [ + 115, + 283, + 234, + 294 + ], + "score": 1.0, + "content": "recognition. In CVPR, 2016.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 269, + 506, + 294 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 300, + 504, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 300, + 504, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 504, + 313 + ], + "score": 1.0, + "content": "Seunghoon Hong, Xinchen Yan, Thomas S Huang, and Honglak Lee. Learning hierarchical semantic", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 312, + 416, + 325 + ], + "spans": [ + { + "bbox": [ + 115, + 312, + 416, + 325 + ], + "score": 1.0, + "content": "image manipulation through structured representations. In NeurIPS, 2018.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 106, + 300, + 504, + 325 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 330, + 505, + 365 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 506, + 343 + ], + "score": 1.0, + "content": "Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild:", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 115, + 342, + 505, + 355 + ], + "score": 1.0, + "content": "A database forstudying face recognition in unconstrained environments. In Workshop on faces", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 353, + 381, + 365 + ], + "spans": [ + { + "bbox": [ + 115, + 353, + 381, + 365 + ], + "score": 1.0, + "content": "in’Real-Life’Images: detection, alignment, and recognition, 2008.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 106, + 331, + 506, + 365 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 371, + 504, + 394 + ], + "lines": [ + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "score": 1.0, + "content": "Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 382, + 383, + 395 + ], + "spans": [ + { + "bbox": [ + 115, + 382, + 383, + 395 + ], + "score": 1.0, + "content": "conditional adversarial networks. In CVPR, pp. 1125–1134, 2017.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 371, + 505, + 395 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 401, + 503, + 424 + ], + "lines": [ + { + "bbox": [ + 106, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 505, + 414 + ], + "score": 1.0, + "content": "Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 412, + 292, + 425 + ], + "spans": [ + { + "bbox": [ + 115, + 412, + 292, + 425 + ], + "score": 1.0, + "content": "super-resolution. In ECCV. Springer, 2016.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 106, + 401, + 505, + 425 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 430, + 505, + 454 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 506, + 446 + ], + "score": 1.0, + "content": "Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In CVPR, pp.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 442, + 190, + 454 + ], + "spans": [ + { + "bbox": [ + 116, + 442, + 190, + 454 + ], + "score": 1.0, + "content": "1219–1228, 2018.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 429, + 506, + 454 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 460, + 505, + 494 + ], + "lines": [ + { + "bbox": [ + 106, + 460, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 506, + 474 + ], + "score": 1.0, + "content": "Ameya Joshi, Amitangshu Mukherjee, Soumik Sarkar, and Chinmay Hegde. Semantic adversarial", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 473, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 116, + 473, + 506, + 484 + ], + "score": 1.0, + "content": "attacks: Parametric transformations that fool deep classifiers. arXiv preprint arXiv:1904.08489,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 482, + 143, + 495 + ], + "spans": [ + { + "bbox": [ + 115, + 482, + 143, + 495 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 460, + 506, + 495 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 501, + 503, + 524 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 505, + 515 + ], + "score": 1.0, + "content": "Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 513, + 346, + 525 + ], + "spans": [ + { + "bbox": [ + 116, + 513, + 346, + 525 + ], + "score": 1.0, + "content": "improved quality, stability, and variation. In ICLR, 2018.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 500, + 505, + 525 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 531, + 502, + 554 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 544 + ], + "score": 1.0, + "content": "Ira Kemelmacher-Shlizerman, Steven M Seitz, Daniel Miller, and Evan Brossard. The megaface", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 542, + 457, + 555 + ], + "spans": [ + { + "bbox": [ + 116, + 542, + 457, + 555 + ], + "score": 1.0, + "content": "benchmark: 1 million faces for recognition at scale. In CVPR, pp. 4873–4882, 2016.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 531, + 505, + 555 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 496, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 560, + 497, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 497, + 574 + ], + "score": 1.0, + "content": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33, + "bbox_fs": [ + 106, + 560, + 497, + 574 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 579, + 460, + 591 + ], + "lines": [ + { + "bbox": [ + 106, + 578, + 460, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 460, + 593 + ], + "score": 1.0, + "content": "Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34, + "bbox_fs": [ + 106, + 578, + 460, + 593 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 598, + 504, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 598, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 505, + 610 + ], + "score": 1.0, + "content": "Brendan F Klare, Ben Klein, Emma Taborsky, Austin Blanton, Jordan Cheney, Kristen Allen, Patrick", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 117, + 610, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 117, + 610, + 505, + 621 + ], + "score": 1.0, + "content": "Grother, Alan Mah, and Anil K Jain. Pushing the frontiers of unconstrained face detection and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 620, + 337, + 632 + ], + "spans": [ + { + "bbox": [ + 115, + 620, + 337, + 632 + ], + "score": 1.0, + "content": "recognition: Iarpa janus benchmark a. In CVPR, 2015.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 106, + 598, + 505, + 632 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 638, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 106, + 639, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 506, + 651 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolu-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 650, + 273, + 662 + ], + "spans": [ + { + "bbox": [ + 116, + 650, + 273, + 662 + ], + "score": 1.0, + "content": "tional neural networks. In NIPS, 2012.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 106, + 639, + 506, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 668, + 504, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "Xin Li and Fuxin Li. Adversarial examples detection in deep networks with convolutional filter", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 680, + 505, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 505, + 692 + ], + "score": 1.0, + "content": "statistics. In Proceedings of the IEEE International Conference on Computer Vision, pp. 5764–", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 690, + 167, + 703 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 167, + 703 + ], + "score": 1.0, + "content": "5772, 2017.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 106, + 668, + 505, + 703 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 720, + 168, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 168, + 732 + ], + "score": 1.0, + "content": "NIPS, 2017.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 106, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 169, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 169, + 105 + ], + "score": 1.0, + "content": "ICCV, 2015.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 112, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "score": 1.0, + "content": "Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 124, + 209, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 124, + 209, + 135 + ], + "score": 1.0, + "content": "them. In CVPR, 2015.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 142, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 140, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 505, + 156 + ], + "score": 1.0, + "content": "Elman Mansimov, Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Generating images", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 154, + 297, + 165 + ], + "spans": [ + { + "bbox": [ + 116, + 154, + 297, + 165 + ], + "score": 1.0, + "content": "from captions with attention. In ICLR, 2015.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 108, + 172, + 505, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 184, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 115, + 184, + 505, + 197 + ], + "score": 1.0, + "content": "accurate method to fool deep neural networks. In Proceedings of the IEEE Conference on Computer", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 194, + 335, + 208 + ], + "spans": [ + { + "bbox": [ + 115, + 194, + 335, + 208 + ], + "score": 1.0, + "content": "Vision and Pattern Recognition, pp. 2574–2582, 2016.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 505, + 248 + ], + "lines": [ + { + "bbox": [ + 106, + 212, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 506, + 227 + ], + "score": 1.0, + "content": "Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 225, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 115, + 225, + 505, + 237 + ], + "score": 1.0, + "content": "and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In CVPR", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 235, + 235, + 249 + ], + "spans": [ + { + "bbox": [ + 116, + 235, + 235, + 249 + ], + "score": 1.0, + "content": "Workshops, pp. 51–59, 2017.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 105, + 254, + 504, + 278 + ], + "lines": [ + { + "bbox": [ + 106, + 254, + 504, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 504, + 267 + ], + "score": 1.0, + "content": "Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 266, + 420, + 279 + ], + "spans": [ + { + "bbox": [ + 115, + 266, + 420, + 279 + ], + "score": 1.0, + "content": "In European Conference on Computer Vision, pp. 483–499. Springer, 2016.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 284, + 504, + 308 + ], + "lines": [ + { + "bbox": [ + 105, + 283, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 505, + 299 + ], + "score": 1.0, + "content": "Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 296, + 273, + 307 + ], + "spans": [ + { + "bbox": [ + 116, + 296, + 273, + 307 + ], + "score": 1.0, + "content": "classifier gans. In ICML. JMLR, 2017.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 315, + 504, + 338 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 326, + 181, + 338 + ], + "spans": [ + { + "bbox": [ + 116, + 326, + 181, + 338 + ], + "score": 1.0, + "content": "In ICML, 2016.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 344, + 505, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 344, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 358 + ], + "score": 1.0, + "content": "Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 356, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 116, + 356, + 506, + 369 + ], + "score": 1.0, + "content": "Swami. The limitations of deep learning in adversarial settings. In Security and Privacy (EuroS&P),", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 367, + 292, + 380 + ], + "spans": [ + { + "bbox": [ + 116, + 367, + 292, + 380 + ], + "score": 1.0, + "content": "2016 IEEE European Symposium on, 2016.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 385, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 383, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 506, + 400 + ], + "score": 1.0, + "content": "Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, et al. Deep face recognition. In bmvc,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 397, + 211, + 410 + ], + "spans": [ + { + "bbox": [ + 115, + 397, + 211, + 410 + ], + "score": 1.0, + "content": "volume 1, pp. 6, 2015.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 416, + 504, + 439 + ], + "lines": [ + { + "bbox": [ + 105, + 415, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 506, + 430 + ], + "score": 1.0, + "content": "Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 428, + 368, + 440 + ], + "spans": [ + { + "bbox": [ + 116, + 428, + 368, + 440 + ], + "score": 1.0, + "content": "convolutional generative adversarial networks. In ICLR, 2015.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 446, + 505, + 470 + ], + "lines": [ + { + "bbox": [ + 106, + 446, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 506, + 460 + ], + "score": 1.0, + "content": "Scott Reed, Kihyuk Sohn, Yuting Zhang, and Honglak Lee. Learning to disentangle factors of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 457, + 326, + 469 + ], + "spans": [ + { + "bbox": [ + 115, + 457, + 326, + 469 + ], + "score": 1.0, + "content": "variation with manifold interaction. In ICML, 2014.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 476, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 105, + 475, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 489 + ], + "score": 1.0, + "content": "Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 117, + 488, + 370, + 500 + ], + "spans": [ + { + "bbox": [ + 117, + 488, + 370, + 500 + ], + "score": 1.0, + "content": "Generative adversarial text to image synthesis. In ICML, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 105, + 506, + 505, + 530 + ], + "lines": [ + { + "bbox": [ + 106, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 506, + 519 + ], + "score": 1.0, + "content": "Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 518, + 479, + 531 + ], + "spans": [ + { + "bbox": [ + 115, + 518, + 479, + 531 + ], + "score": 1.0, + "content": "wild challenge: The first facial landmark localization challenge. In ICCV Workshop, 2013.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 105, + 536, + 504, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 504, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 504, + 549 + ], + "score": 1.0, + "content": "Soumyadip Sengupta, Jun-Cheng Chen, Carlos Castillo, Vishal M Patel, Rama Chellappa, and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 548, + 501, + 560 + ], + "spans": [ + { + "bbox": [ + 116, + 548, + 501, + 560 + ], + "score": 1.0, + "content": "David W Jacobs. Frontal to profile face verification in the wild. In WACV, pp. 1–9. IEEE, 2016.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 108, + 566, + 506, + 601 + ], + "lines": [ + { + "bbox": [ + 106, + 566, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 506, + 580 + ], + "score": 1.0, + "content": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 577, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 506, + 592 + ], + "score": 1.0, + "content": "Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 589, + 472, + 601 + ], + "spans": [ + { + "bbox": [ + 115, + 589, + 472, + 601 + ], + "score": 1.0, + "content": "the game of go with deep neural networks and tree search. nature, 529(7587):484, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 105, + 608, + 505, + 631 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 505, + 623 + ], + "score": 1.0, + "content": "Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 619, + 325, + 631 + ], + "spans": [ + { + "bbox": [ + 115, + 619, + 325, + 631 + ], + "score": 1.0, + "content": "recognition. arXiv preprint arXiv:1409.1556, 2014.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "Yang Song, Rui Shu, Nate Kushman, and Stefano Ermon. Constructing unrestricted adversarial", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 648, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 115, + 648, + 506, + 664 + ], + "score": 1.0, + "content": "examples with generative models. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 660, + 192, + 672 + ], + "spans": [ + { + "bbox": [ + 116, + 660, + 192, + 672 + ], + "score": 1.0, + "content": "8312–8323, 2018.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 679, + 504, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 693 + ], + "score": 1.0, + "content": "Yi Sun, Xiaogang Wang, and Xiaoou Tang. Deep learning face representation from predicting 10,000", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 691, + 216, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 691, + 216, + 702 + ], + "score": 1.0, + "content": "classes. In CVPR, 2014.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 721, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 506, + 732 + ], + "score": 1.0, + "content": "and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 169, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 169, + 105 + ], + "score": 1.0, + "content": "ICCV, 2015.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 506, + 105 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 112, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "score": 1.0, + "content": "Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 124, + 209, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 124, + 209, + 135 + ], + "score": 1.0, + "content": "them. In CVPR, 2015.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 111, + 505, + 135 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 142, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 140, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 505, + 156 + ], + "score": 1.0, + "content": "Elman Mansimov, Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Generating images", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 154, + 297, + 165 + ], + "spans": [ + { + "bbox": [ + 116, + 154, + 297, + 165 + ], + "score": 1.0, + "content": "from captions with attention. In ICLR, 2015.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 140, + 505, + 165 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 172, + 505, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 184, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 115, + 184, + 505, + 197 + ], + "score": 1.0, + "content": "accurate method to fool deep neural networks. In Proceedings of the IEEE Conference on Computer", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 194, + 335, + 208 + ], + "spans": [ + { + "bbox": [ + 115, + 194, + 335, + 208 + ], + "score": 1.0, + "content": "Vision and Pattern Recognition, pp. 2574–2582, 2016.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 172, + 505, + 208 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 505, + 248 + ], + "lines": [ + { + "bbox": [ + 106, + 212, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 506, + 227 + ], + "score": 1.0, + "content": "Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 225, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 115, + 225, + 505, + 237 + ], + "score": 1.0, + "content": "and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In CVPR", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 235, + 235, + 249 + ], + "spans": [ + { + "bbox": [ + 116, + 235, + 235, + 249 + ], + "score": 1.0, + "content": "Workshops, pp. 51–59, 2017.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 106, + 212, + 506, + 249 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 254, + 504, + 278 + ], + "lines": [ + { + "bbox": [ + 106, + 254, + 504, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 504, + 267 + ], + "score": 1.0, + "content": "Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 266, + 420, + 279 + ], + "spans": [ + { + "bbox": [ + 115, + 266, + 420, + 279 + ], + "score": 1.0, + "content": "In European Conference on Computer Vision, pp. 483–499. Springer, 2016.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 106, + 254, + 504, + 279 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 284, + 504, + 308 + ], + "lines": [ + { + "bbox": [ + 105, + 283, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 505, + 299 + ], + "score": 1.0, + "content": "Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 296, + 273, + 307 + ], + "spans": [ + { + "bbox": [ + 116, + 296, + 273, + 307 + ], + "score": 1.0, + "content": "classifier gans. In ICML. JMLR, 2017.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 283, + 505, + 307 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 315, + 504, + 338 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 326, + 181, + 338 + ], + "spans": [ + { + "bbox": [ + 116, + 326, + 181, + 338 + ], + "score": 1.0, + "content": "In ICML, 2016.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 315, + 505, + 338 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 344, + 505, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 344, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 358 + ], + "score": 1.0, + "content": "Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 356, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 116, + 356, + 506, + 369 + ], + "score": 1.0, + "content": "Swami. The limitations of deep learning in adversarial settings. In Security and Privacy (EuroS&P),", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 367, + 292, + 380 + ], + "spans": [ + { + "bbox": [ + 116, + 367, + 292, + 380 + ], + "score": 1.0, + "content": "2016 IEEE European Symposium on, 2016.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 344, + 506, + 380 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 385, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 383, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 506, + 400 + ], + "score": 1.0, + "content": "Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, et al. Deep face recognition. In bmvc,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 397, + 211, + 410 + ], + "spans": [ + { + "bbox": [ + 115, + 397, + 211, + 410 + ], + "score": 1.0, + "content": "volume 1, pp. 6, 2015.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 383, + 506, + 410 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 416, + 504, + 439 + ], + "lines": [ + { + "bbox": [ + 105, + 415, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 506, + 430 + ], + "score": 1.0, + "content": "Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 428, + 368, + 440 + ], + "spans": [ + { + "bbox": [ + 116, + 428, + 368, + 440 + ], + "score": 1.0, + "content": "convolutional generative adversarial networks. In ICLR, 2015.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 415, + 506, + 440 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 446, + 505, + 470 + ], + "lines": [ + { + "bbox": [ + 106, + 446, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 506, + 460 + ], + "score": 1.0, + "content": "Scott Reed, Kihyuk Sohn, Yuting Zhang, and Honglak Lee. Learning to disentangle factors of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 457, + 326, + 469 + ], + "spans": [ + { + "bbox": [ + 115, + 457, + 326, + 469 + ], + "score": 1.0, + "content": "variation with manifold interaction. In ICML, 2014.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 446, + 506, + 469 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 476, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 105, + 475, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 489 + ], + "score": 1.0, + "content": "Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 117, + 488, + 370, + 500 + ], + "spans": [ + { + "bbox": [ + 117, + 488, + 370, + 500 + ], + "score": 1.0, + "content": "Generative adversarial text to image synthesis. In ICML, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 475, + 506, + 500 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 506, + 505, + 530 + ], + "lines": [ + { + "bbox": [ + 106, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 506, + 519 + ], + "score": 1.0, + "content": "Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 518, + 479, + 531 + ], + "spans": [ + { + "bbox": [ + 115, + 518, + 479, + 531 + ], + "score": 1.0, + "content": "wild challenge: The first facial landmark localization challenge. In ICCV Workshop, 2013.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 506, + 506, + 531 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 536, + 504, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 504, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 504, + 549 + ], + "score": 1.0, + "content": "Soumyadip Sengupta, Jun-Cheng Chen, Carlos Castillo, Vishal M Patel, Rama Chellappa, and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 548, + 501, + 560 + ], + "spans": [ + { + "bbox": [ + 116, + 548, + 501, + 560 + ], + "score": 1.0, + "content": "David W Jacobs. Frontal to profile face verification in the wild. In WACV, pp. 1–9. IEEE, 2016.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 106, + 537, + 504, + 560 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 566, + 506, + 601 + ], + "lines": [ + { + "bbox": [ + 106, + 566, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 506, + 580 + ], + "score": 1.0, + "content": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 577, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 506, + 592 + ], + "score": 1.0, + "content": "Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 589, + 472, + 601 + ], + "spans": [ + { + "bbox": [ + 115, + 589, + 472, + 601 + ], + "score": 1.0, + "content": "the game of go with deep neural networks and tree search. nature, 529(7587):484, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 106, + 566, + 506, + 601 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 608, + 505, + 631 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 505, + 623 + ], + "score": 1.0, + "content": "Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 619, + 325, + 631 + ], + "spans": [ + { + "bbox": [ + 115, + 619, + 325, + 631 + ], + "score": 1.0, + "content": "recognition. arXiv preprint arXiv:1409.1556, 2014.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 606, + 505, + 631 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "Yang Song, Rui Shu, Nate Kushman, and Stefano Ermon. Constructing unrestricted adversarial", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 648, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 115, + 648, + 506, + 664 + ], + "score": 1.0, + "content": "examples with generative models. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 660, + 192, + 672 + ], + "spans": [ + { + "bbox": [ + 116, + 660, + 192, + 672 + ], + "score": 1.0, + "content": "8312–8323, 2018.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 106, + 637, + 506, + 672 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 679, + 504, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 693 + ], + "score": 1.0, + "content": "Yi Sun, Xiaogang Wang, and Xiaoou Tang. Deep learning face representation from predicting 10,000", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 691, + 216, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 691, + 216, + 702 + ], + "score": 1.0, + "content": "classes. In CVPR, 2014.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 677, + 505, + 702 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 721, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 506, + 732 + ], + "score": 1.0, + "content": "and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 106, + 709, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 505, + 115 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "Erhan, Vincent Vanhoucke, Andrew Rabinovich, et al. Going deeper with convolutions. In CVPR,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 102, + 144, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 102, + 144, + 117 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 505, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 123, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 506, + 136 + ], + "score": 1.0, + "content": "Guanhong Tao, Shiqing Ma, Yingqi Liu, and Xiangyu Zhang. Attacks meet interpretability: Attribute-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 356, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 356, + 146 + ], + "score": 1.0, + "content": "steered detection of adversarial samples. In NeurIPS, 2018.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 153, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 164, + 348, + 176 + ], + "spans": [ + { + "bbox": [ + 116, + 164, + 348, + 176 + ], + "score": 1.0, + "content": "image generation with pixelcnn decoders. In NIPS, 2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 182, + 504, + 206 + ], + "lines": [ + { + "bbox": [ + 106, + 183, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 505, + 196 + ], + "score": 1.0, + "content": "Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 193, + 452, + 207 + ], + "spans": [ + { + "bbox": [ + 115, + 193, + 452, + 207 + ], + "score": 1.0, + "content": "Liu. Cosface: Large margin cosine loss for deep face recognition. In CVPR, 2018a.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 212, + 505, + 236 + ], + "lines": [ + { + "bbox": [ + 107, + 213, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 107, + 213, + 506, + 226 + ], + "score": 1.0, + "content": "Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 224, + 497, + 236 + ], + "spans": [ + { + "bbox": [ + 115, + 224, + 497, + 236 + ], + "score": 1.0, + "content": "resolution image synthesis and semantic manipulation with conditional gans. In CVPR, 2018b.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 242, + 504, + 276 + ], + "lines": [ + { + "bbox": [ + 105, + 241, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 257 + ], + "score": 1.0, + "content": "Chaowei Xiao, Ruizhi Deng, Bo Li, Fisher Yu, Mingyan Liu, and Dawn Song. Characterizing", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 116, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "adversarial examples based on spatial consistency information for semantic segmentation. In", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 265, + 176, + 276 + ], + "spans": [ + { + "bbox": [ + 116, + 265, + 176, + 276 + ], + "score": 1.0, + "content": "ECCV, 2018a.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 503, + 307 + ], + "lines": [ + { + "bbox": [ + 106, + 283, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 505, + 296 + ], + "score": 1.0, + "content": "Chaowei Xiao, Bo Li, Jun-Yan Zhu, Warren He, Mingyan Liu, and Dawn Song. Generating adversarial", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 295, + 334, + 307 + ], + "spans": [ + { + "bbox": [ + 115, + 295, + 334, + 307 + ], + "score": 1.0, + "content": "examples with adversarial networks. In IJCAI, 2018b.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 313, + 504, + 336 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "Chaowei Xiao, Jun-Yan Zhu, Bo Li, Warren He, Mingyan Liu, and Dawn Song. Spatially transformed", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 117, + 325, + 273, + 337 + ], + "spans": [ + { + "bbox": [ + 117, + 325, + 273, + 337 + ], + "score": 1.0, + "content": "adversarial examples. In ICLR, 2018c.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 108, + 343, + 505, + 377 + ], + "lines": [ + { + "bbox": [ + 106, + 343, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 505, + 356 + ], + "score": 1.0, + "content": "Chaowei Xiao, Dawei Yang, Bo Li, Jia Deng, and Mingyan Liu. Meshadv: Adversarial meshes", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 116, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "for visual recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 366, + 259, + 378 + ], + "spans": [ + { + "bbox": [ + 116, + 366, + 259, + 378 + ], + "score": 1.0, + "content": "Recognition, pp. 6898–6907, 2019.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 504, + 418 + ], + "lines": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 395, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 115, + 395, + 505, + 407 + ], + "score": 1.0, + "content": "Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 406, + 439, + 419 + ], + "spans": [ + { + "bbox": [ + 116, + 406, + 439, + 419 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition, pp. 2730–2739, 2019.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 424, + 504, + 448 + ], + "lines": [ + { + "bbox": [ + 105, + 423, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 439 + ], + "score": 1.0, + "content": "Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 436, + 349, + 447 + ], + "spans": [ + { + "bbox": [ + 116, + 436, + 349, + 447 + ], + "score": 1.0, + "content": "neural networks. arXiv preprint arXiv:1704.01155, 2017.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 454, + 504, + 478 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 469 + ], + "score": 1.0, + "content": "Xinchen Yan, Jimei Yang, Kihyuk Sohn, and Honglak Lee. Attribute2image: Conditional image", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 466, + 357, + 479 + ], + "spans": [ + { + "bbox": [ + 115, + 466, + 357, + 479 + ], + "score": 1.0, + "content": "generation from visual attributes. In ECCV. Springer, 2016.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 484, + 504, + 508 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "score": 1.0, + "content": "Fisher Yu, Vladlen Koltun, and Thomas Funkhouser. Dilated residual networks. In Computer Vision", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 496, + 277, + 508 + ], + "spans": [ + { + "bbox": [ + 116, + 496, + 277, + 508 + ], + "score": 1.0, + "content": "and Pattern Recognition (CVPR), 2017.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 514, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 514, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 528 + ], + "score": 1.0, + "content": "Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 116, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 537, + 222, + 547 + ], + "spans": [ + { + "bbox": [ + 116, + 537, + 222, + 547 + ], + "score": 1.0, + "content": "networks. In ICCV, 2017.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 108, + 555, + 504, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "score": 1.0, + "content": "Xingcheng Zhang, Lei Yang, Junjie Yan, and Dahua Lin. Accelerated training for massive classifi-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 566, + 507, + 580 + ], + "spans": [ + { + "bbox": [ + 115, + 566, + 507, + 580 + ], + "score": 1.0, + "content": "cation via dynamic class selection. In Thirty-Second AAAI Conference on Artificial Intelligence,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 577, + 143, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 143, + 590 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 104, + 596, + 505, + 619 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 504, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 504, + 609 + ], + "score": 1.0, + "content": "Jun-Yan Zhu, Philipp Krähenbühl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 607, + 353, + 620 + ], + "spans": [ + { + "bbox": [ + 115, + 607, + 353, + 620 + ], + "score": 1.0, + "content": "on the natural image manifold. In ECCV. Springer, 2016a.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 104, + 626, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 625, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 505, + 640 + ], + "score": 1.0, + "content": "Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 638, + 359, + 649 + ], + "spans": [ + { + "bbox": [ + 115, + 638, + 359, + 649 + ], + "score": 1.0, + "content": "using cycle-consistent adversarial networks. In ICCV, 2017.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 106, + 656, + 504, + 679 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 670 + ], + "score": 1.0, + "content": "Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z Li. Face alignment across large poses:", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 667, + 246, + 678 + ], + "spans": [ + { + "bbox": [ + 115, + 667, + 246, + 678 + ], + "score": 1.0, + "content": "A 3d solution. In CVPR, 2016b.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 505, + 115 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "Erhan, Vincent Vanhoucke, Andrew Rabinovich, et al. Going deeper with convolutions. In CVPR,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 102, + 144, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 102, + 144, + 117 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 82, + 506, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 505, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 123, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 506, + 136 + ], + "score": 1.0, + "content": "Guanhong Tao, Shiqing Ma, Yingqi Liu, and Xiangyu Zhang. Attacks meet interpretability: Attribute-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 356, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 356, + 146 + ], + "score": 1.0, + "content": "steered detection of adversarial samples. In NeurIPS, 2018.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 106, + 123, + 506, + 146 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 153, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 164, + 348, + 176 + ], + "spans": [ + { + "bbox": [ + 116, + 164, + 348, + 176 + ], + "score": 1.0, + "content": "image generation with pixelcnn decoders. In NIPS, 2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 106, + 153, + 505, + 176 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 182, + 504, + 206 + ], + "lines": [ + { + "bbox": [ + 106, + 183, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 505, + 196 + ], + "score": 1.0, + "content": "Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 193, + 452, + 207 + ], + "spans": [ + { + "bbox": [ + 115, + 193, + 452, + 207 + ], + "score": 1.0, + "content": "Liu. Cosface: Large margin cosine loss for deep face recognition. In CVPR, 2018a.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 183, + 505, + 207 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 212, + 505, + 236 + ], + "lines": [ + { + "bbox": [ + 107, + 213, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 107, + 213, + 506, + 226 + ], + "score": 1.0, + "content": "Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 224, + 497, + 236 + ], + "spans": [ + { + "bbox": [ + 115, + 224, + 497, + 236 + ], + "score": 1.0, + "content": "resolution image synthesis and semantic manipulation with conditional gans. In CVPR, 2018b.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 107, + 213, + 506, + 236 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 242, + 504, + 276 + ], + "lines": [ + { + "bbox": [ + 105, + 241, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 257 + ], + "score": 1.0, + "content": "Chaowei Xiao, Ruizhi Deng, Bo Li, Fisher Yu, Mingyan Liu, and Dawn Song. Characterizing", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 116, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "adversarial examples based on spatial consistency information for semantic segmentation. In", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 265, + 176, + 276 + ], + "spans": [ + { + "bbox": [ + 116, + 265, + 176, + 276 + ], + "score": 1.0, + "content": "ECCV, 2018a.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 241, + 506, + 276 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 503, + 307 + ], + "lines": [ + { + "bbox": [ + 106, + 283, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 505, + 296 + ], + "score": 1.0, + "content": "Chaowei Xiao, Bo Li, Jun-Yan Zhu, Warren He, Mingyan Liu, and Dawn Song. Generating adversarial", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 295, + 334, + 307 + ], + "spans": [ + { + "bbox": [ + 115, + 295, + 334, + 307 + ], + "score": 1.0, + "content": "examples with adversarial networks. In IJCAI, 2018b.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 283, + 505, + 307 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 313, + 504, + 336 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "Chaowei Xiao, Jun-Yan Zhu, Bo Li, Warren He, Mingyan Liu, and Dawn Song. Spatially transformed", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 117, + 325, + 273, + 337 + ], + "spans": [ + { + "bbox": [ + 117, + 325, + 273, + 337 + ], + "score": 1.0, + "content": "adversarial examples. In ICLR, 2018c.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 313, + 505, + 337 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 343, + 505, + 377 + ], + "lines": [ + { + "bbox": [ + 106, + 343, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 505, + 356 + ], + "score": 1.0, + "content": "Chaowei Xiao, Dawei Yang, Bo Li, Jia Deng, and Mingyan Liu. Meshadv: Adversarial meshes", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 116, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "for visual recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 366, + 259, + 378 + ], + "spans": [ + { + "bbox": [ + 116, + 366, + 259, + 378 + ], + "score": 1.0, + "content": "Recognition, pp. 6898–6907, 2019.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 106, + 343, + 505, + 378 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 504, + 418 + ], + "lines": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 395, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 115, + 395, + 505, + 407 + ], + "score": 1.0, + "content": "Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 406, + 439, + 419 + ], + "spans": [ + { + "bbox": [ + 116, + 406, + 439, + 419 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition, pp. 2730–2739, 2019.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 384, + 505, + 419 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 424, + 504, + 448 + ], + "lines": [ + { + "bbox": [ + 105, + 423, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 439 + ], + "score": 1.0, + "content": "Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 436, + 349, + 447 + ], + "spans": [ + { + "bbox": [ + 116, + 436, + 349, + 447 + ], + "score": 1.0, + "content": "neural networks. arXiv preprint arXiv:1704.01155, 2017.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 423, + 505, + 447 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 454, + 504, + 478 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 469 + ], + "score": 1.0, + "content": "Xinchen Yan, Jimei Yang, Kihyuk Sohn, and Honglak Lee. Attribute2image: Conditional image", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 466, + 357, + 479 + ], + "spans": [ + { + "bbox": [ + 115, + 466, + 357, + 479 + ], + "score": 1.0, + "content": "generation from visual attributes. In ECCV. Springer, 2016.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 453, + 505, + 479 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 484, + 504, + 508 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "score": 1.0, + "content": "Fisher Yu, Vladlen Koltun, and Thomas Funkhouser. Dilated residual networks. In Computer Vision", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 496, + 277, + 508 + ], + "spans": [ + { + "bbox": [ + 116, + 496, + 277, + 508 + ], + "score": 1.0, + "content": "and Pattern Recognition (CVPR), 2017.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 484, + 505, + 508 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 514, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 514, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 528 + ], + "score": 1.0, + "content": "Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 116, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 537, + 222, + 547 + ], + "spans": [ + { + "bbox": [ + 116, + 537, + 222, + 547 + ], + "score": 1.0, + "content": "networks. In ICCV, 2017.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 514, + 505, + 547 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 555, + 504, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "score": 1.0, + "content": "Xingcheng Zhang, Lei Yang, Junjie Yan, and Dahua Lin. Accelerated training for massive classifi-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 566, + 507, + 580 + ], + "spans": [ + { + "bbox": [ + 115, + 566, + 507, + 580 + ], + "score": 1.0, + "content": "cation via dynamic class selection. In Thirty-Second AAAI Conference on Artificial Intelligence,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 577, + 143, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 143, + 590 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 555, + 507, + 590 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 596, + 505, + 619 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 504, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 504, + 609 + ], + "score": 1.0, + "content": "Jun-Yan Zhu, Philipp Krähenbühl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 607, + 353, + 620 + ], + "spans": [ + { + "bbox": [ + 115, + 607, + 353, + 620 + ], + "score": 1.0, + "content": "on the natural image manifold. In ECCV. Springer, 2016a.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 106, + 597, + 504, + 620 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 626, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 625, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 505, + 640 + ], + "score": 1.0, + "content": "Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 638, + 359, + 649 + ], + "spans": [ + { + "bbox": [ + 115, + 638, + 359, + 649 + ], + "score": 1.0, + "content": "using cycle-consistent adversarial networks. In ICCV, 2017.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 625, + 505, + 649 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 656, + 504, + 679 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 670 + ], + "score": 1.0, + "content": "Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z Li. Face alignment across large poses:", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 667, + 246, + 678 + ], + "spans": [ + { + "bbox": [ + 115, + 667, + 246, + 678 + ], + "score": 1.0, + "content": "A 3d solution. In CVPR, 2016b.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 654, + 506, + 678 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 286, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 288, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 288, + 96 + ], + "score": 1.0, + "content": "A FACE IDENTITY VERIFICATION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 506, + 156 + ], + "lines": [ + { + "bbox": [ + 106, + 101, + 506, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 101, + 506, + 113 + ], + "score": 1.0, + "content": "Benchmark performance. First, we provide additional information about the ResNet models we", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 112, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 506, + 124 + ], + "score": 1.0, + "content": "used in the experiments. We summarize in Table A the performance on several face identity verifica-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 123, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 505, + 135 + ], + "score": 1.0, + "content": "tion benchmarks including Labeled Face in the Wild (LFW) dataset (Huang et al., 2008), AgeDB-30", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 507, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 507, + 147 + ], + "score": 1.0, + "content": "dataset (Moschoglou et al., 2017), and Celebrities in Frontal-Profile (CFP) dataset (Sengupta et al.,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 136, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 136, + 158 + ], + "score": 1.0, + "content": "2016).", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 173, + 168, + 438, + 229 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 173, + 168, + 438, + 229 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 173, + 168, + 438, + 229 + ], + "spans": [ + { + "bbox": [ + 173, + 168, + 438, + 229 + ], + "score": 0.98, + "html": "
M benchmarksLFWAgeDB-30CFP-FFCFP-FP
R-50-S99.2794.1599.2691.49
R-101-S99.4295.9399.5795.07
R-50-C99.3895.0899.2490.24
R-101-C99.6795.5899.5792.71
", + "type": "table", + "image_path": "09c7f12c436f5b70809b6349b5b2cc53a92a826f00065e6f26ae9a1543329bdf.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 173, + 168, + 438, + 188.33333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 173, + 188.33333333333334, + 438, + 208.66666666666669 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 173, + 208.66666666666669, + 438, + 229.00000000000003 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 153, + 234, + 458, + 246 + ], + "lines": [ + { + "bbox": [ + 151, + 233, + 460, + 247 + ], + "spans": [ + { + "bbox": [ + 151, + 233, + 460, + 247 + ], + "score": 1.0, + "content": "Table A: The performnace of ResNet models on several benchmark datasets.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 253, + 505, + 298 + ], + "lines": [ + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "Identity verification thresholds. Table B shows the threshold values used in our experiments", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 264, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 277 + ], + "score": 1.0, + "content": "when determining whether two portrait images belong to the same identity or not. The selected FPR", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "thresholds and normalized L2 distance between face features are commonly used when evaluating the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 286, + 502, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 502, + 299 + ], + "score": 1.0, + "content": "performance of face recognition model Klare et al. (2015); Kemelmacher-Shlizerman et al. (2016).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "table", + "bbox": [ + 189, + 310, + 419, + 362 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 189, + 310, + 419, + 362 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 189, + 310, + 419, + 362 + ], + "spans": [ + { + "bbox": [ + 189, + 310, + 419, + 362 + ], + "score": 0.976, + "html": "
FPR/MR-50-SR-101-SR-50-CR-101-C
10-33×10-410-41.1811.2441.4471.469
1.0580.6571.0481.2931.2420.809
0.5970.864
", + "type": "table", + "image_path": "60612b66a039fb5769d8d04f005e45f0aa712bca2bab53d6b13d0716fa50ea18.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 189, + 310, + 419, + 323.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 189, + 323.0, + 419, + 336.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 189, + 336.0, + 419, + 349.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 189, + 349.0, + 419, + 362.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 187, + 366, + 423, + 377 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 185, + 364, + 424, + 379 + ], + "spans": [ + { + "bbox": [ + 185, + 364, + 424, + 379 + ], + "score": 1.0, + "content": "Table B: The threshold values for face identity verification.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + } + ], + "index": 16.75 + }, + { + "type": "text", + "bbox": [ + 106, + 390, + 506, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 507, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 507, + 403 + ], + "score": 1.0, + "content": "Implementation details. We use Adam (Kingma & Ba, 2015) as the optimizer to generate ad-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "score": 1.0, + "content": "versarial examples for both our SemanticAdv and CW. More specifically, we run optimization for", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 410, + 507, + 426 + ], + "spans": [ + { + "bbox": [ + 104, + 410, + 358, + 426 + ], + "score": 1.0, + "content": "up to 200 steps with a fixed learning rate 0.05 for cases when", + "type": "text" + }, + { + "bbox": [ + 358, + 412, + 423, + 423 + ], + "score": 0.69, + "content": "\\mathbf { \\bar { G } } \\mathbf { - } \\mathbf { F P R } \\leq 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 410, + 507, + 426 + ], + "score": 1.0, + "content": ". Otherwise, we run", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 423, + 507, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 507, + 435 + ], + "score": 1.0, + "content": "optimization for up to 500 steps with a fixed learning rate 0.01. For pixel-wise attack method CW,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "we use additional pixel reconstruction loss with corresponding loss weight to 5. We run optimization", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 444, + 320, + 458 + ], + "spans": [ + { + "bbox": [ + 104, + 444, + 143, + 458 + ], + "score": 1.0, + "content": "for up to", + "type": "text" + }, + { + "bbox": [ + 143, + 445, + 169, + 456 + ], + "score": 0.26, + "content": "1 , 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 444, + 294, + 458 + ], + "score": 1.0, + "content": "steps with a fixed learning rate", + "type": "text" + }, + { + "bbox": [ + 295, + 444, + 316, + 455 + ], + "score": 0.9, + "content": "1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 444, + 320, + 458 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 105, + 464, + 504, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 478 + ], + "score": 1.0, + "content": "Evaluation metrics. To evaluate the performance of semanticAdv under different attributes, we", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 475, + 242, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 242, + 487 + ], + "score": 1.0, + "content": "consider three metrics as follows:", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 132, + 497, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 131, + 497, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 497, + 505, + 511 + ], + "score": 1.0, + "content": "• Best: if there is one attribute among 17 attributes that can be successfully attacked, we count", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 508, + 338, + 521 + ], + "spans": [ + { + "bbox": [ + 141, + 508, + 338, + 521 + ], + "score": 1.0, + "content": "the attack success rate for this face identity as 1;", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "• Average: we calculate the average attack success rate among 17 attributes for the same face", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 141, + 536, + 179, + 550 + ], + "spans": [ + { + "bbox": [ + 141, + 536, + 179, + 550 + ], + "score": 1.0, + "content": "identity;", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 132, + 554, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 505, + 567 + ], + "score": 1.0, + "content": "• Worst: only when all of 17 attributes can be successfully attacked, we count the attack", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 141, + 566, + 274, + 579 + ], + "spans": [ + { + "bbox": [ + 141, + 566, + 274, + 579 + ], + "score": 1.0, + "content": "success rate for this person as 1;", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 588, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "Note that, for a fair comparison with CW, we should use the Best metric for our SemanticAdv, as CW", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "is the traditional pixel-wise attack method works regardless of the attribute. In addition, we report the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 610, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 622 + ], + "score": 1.0, + "content": "performance using the average and worst metric, which actually provides additional insights into the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 620, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 505, + 634 + ], + "score": 1.0, + "content": "robustness of face verification models across different attributes. For instance, combining the results", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "from Table C and Figure 3, we understand that the face verification models used in our experiments", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 642, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 505, + 655 + ], + "score": 1.0, + "content": "have different levels of robustness across attributes. For example, face verification models are more", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 654, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 505, + 666 + ], + "score": 1.0, + "content": "robust against local shape variations than color variations, e.g., pale skin has higher attack success", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 665, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 677 + ], + "score": 1.0, + "content": "rate than mouth open. We believe these discoveries will help the community further understand the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 676, + 260, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 260, + 688 + ], + "score": 1.0, + "content": "properties of face verification models.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 693, + 504, + 737 + ], + "lines": [ + { + "bbox": [ + 105, + 693, + 506, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 506, + 705 + ], + "score": 1.0, + "content": "Table C shows the overall performance (accuracy) of face verification model and attack success", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 704, + 505, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 704, + 505, + 716 + ], + "score": 1.0, + "content": "rate of SemanticAdv and CW. As we can see from Table C, although the face model trained with", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 715, + 506, + 726 + ], + "spans": [ + { + "bbox": [ + 105, + 715, + 506, + 726 + ], + "score": 1.0, + "content": "cosine loss achieves higher face recognition performance, it is more vulnerable to adversarial attack", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 726, + 316, + 737 + ], + "spans": [ + { + "bbox": [ + 105, + 726, + 316, + 737 + ], + "score": 1.0, + "content": "compared with the model trained with softmax loss.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 286, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 288, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 288, + 96 + ], + "score": 1.0, + "content": "A FACE IDENTITY VERIFICATION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 506, + 156 + ], + "lines": [ + { + "bbox": [ + 106, + 101, + 506, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 101, + 506, + 113 + ], + "score": 1.0, + "content": "Benchmark performance. First, we provide additional information about the ResNet models we", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 112, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 506, + 124 + ], + "score": 1.0, + "content": "used in the experiments. We summarize in Table A the performance on several face identity verifica-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 123, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 505, + 135 + ], + "score": 1.0, + "content": "tion benchmarks including Labeled Face in the Wild (LFW) dataset (Huang et al., 2008), AgeDB-30", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 507, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 507, + 147 + ], + "score": 1.0, + "content": "dataset (Moschoglou et al., 2017), and Celebrities in Frontal-Profile (CFP) dataset (Sengupta et al.,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 136, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 136, + 158 + ], + "score": 1.0, + "content": "2016).", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 101, + 507, + 158 + ] + }, + { + "type": "table", + "bbox": [ + 173, + 168, + 438, + 229 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 173, + 168, + 438, + 229 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 173, + 168, + 438, + 229 + ], + "spans": [ + { + "bbox": [ + 173, + 168, + 438, + 229 + ], + "score": 0.98, + "html": "
M benchmarksLFWAgeDB-30CFP-FFCFP-FP
R-50-S99.2794.1599.2691.49
R-101-S99.4295.9399.5795.07
R-50-C99.3895.0899.2490.24
R-101-C99.6795.5899.5792.71
", + "type": "table", + "image_path": "09c7f12c436f5b70809b6349b5b2cc53a92a826f00065e6f26ae9a1543329bdf.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 173, + 168, + 438, + 188.33333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 173, + 188.33333333333334, + 438, + 208.66666666666669 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 173, + 208.66666666666669, + 438, + 229.00000000000003 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 153, + 234, + 458, + 246 + ], + "lines": [ + { + "bbox": [ + 151, + 233, + 460, + 247 + ], + "spans": [ + { + "bbox": [ + 151, + 233, + 460, + 247 + ], + "score": 1.0, + "content": "Table A: The performnace of ResNet models on several benchmark datasets.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 151, + 233, + 460, + 247 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 253, + 505, + 298 + ], + "lines": [ + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "Identity verification thresholds. Table B shows the threshold values used in our experiments", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 264, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 277 + ], + "score": 1.0, + "content": "when determining whether two portrait images belong to the same identity or not. The selected FPR", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "thresholds and normalized L2 distance between face features are commonly used when evaluating the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 286, + 502, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 502, + 299 + ], + "score": 1.0, + "content": "performance of face recognition model Klare et al. (2015); Kemelmacher-Shlizerman et al. (2016).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 253, + 506, + 299 + ] + }, + { + "type": "table", + "bbox": [ + 189, + 310, + 419, + 362 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 189, + 310, + 419, + 362 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 189, + 310, + 419, + 362 + ], + "spans": [ + { + "bbox": [ + 189, + 310, + 419, + 362 + ], + "score": 0.976, + "html": "
FPR/MR-50-SR-101-SR-50-CR-101-C
10-33×10-410-41.1811.2441.4471.469
1.0580.6571.0481.2931.2420.809
0.5970.864
", + "type": "table", + "image_path": "60612b66a039fb5769d8d04f005e45f0aa712bca2bab53d6b13d0716fa50ea18.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 189, + 310, + 419, + 323.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 189, + 323.0, + 419, + 336.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 189, + 336.0, + 419, + 349.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 189, + 349.0, + 419, + 362.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 187, + 366, + 423, + 377 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 185, + 364, + 424, + 379 + ], + "spans": [ + { + "bbox": [ + 185, + 364, + 424, + 379 + ], + "score": 1.0, + "content": "Table B: The threshold values for face identity verification.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + } + ], + "index": 16.75 + }, + { + "type": "text", + "bbox": [ + 106, + 390, + 506, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 507, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 507, + 403 + ], + "score": 1.0, + "content": "Implementation details. We use Adam (Kingma & Ba, 2015) as the optimizer to generate ad-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "score": 1.0, + "content": "versarial examples for both our SemanticAdv and CW. More specifically, we run optimization for", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 410, + 507, + 426 + ], + "spans": [ + { + "bbox": [ + 104, + 410, + 358, + 426 + ], + "score": 1.0, + "content": "up to 200 steps with a fixed learning rate 0.05 for cases when", + "type": "text" + }, + { + "bbox": [ + 358, + 412, + 423, + 423 + ], + "score": 0.69, + "content": "\\mathbf { \\bar { G } } \\mathbf { - } \\mathbf { F P R } \\leq 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 410, + 507, + 426 + ], + "score": 1.0, + "content": ". Otherwise, we run", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 423, + 507, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 507, + 435 + ], + "score": 1.0, + "content": "optimization for up to 500 steps with a fixed learning rate 0.01. For pixel-wise attack method CW,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "we use additional pixel reconstruction loss with corresponding loss weight to 5. We run optimization", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 444, + 320, + 458 + ], + "spans": [ + { + "bbox": [ + 104, + 444, + 143, + 458 + ], + "score": 1.0, + "content": "for up to", + "type": "text" + }, + { + "bbox": [ + 143, + 445, + 169, + 456 + ], + "score": 0.26, + "content": "1 , 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 444, + 294, + 458 + ], + "score": 1.0, + "content": "steps with a fixed learning rate", + "type": "text" + }, + { + "bbox": [ + 295, + 444, + 316, + 455 + ], + "score": 0.9, + "content": "1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 444, + 320, + 458 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5, + "bbox_fs": [ + 104, + 389, + 507, + 458 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 464, + 504, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 478 + ], + "score": 1.0, + "content": "Evaluation metrics. To evaluate the performance of semanticAdv under different attributes, we", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 475, + 242, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 242, + 487 + ], + "score": 1.0, + "content": "consider three metrics as follows:", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 462, + 506, + 487 + ] + }, + { + "type": "list", + "bbox": [ + 132, + 497, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 131, + 497, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 497, + 505, + 511 + ], + "score": 1.0, + "content": "• Best: if there is one attribute among 17 attributes that can be successfully attacked, we count", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 508, + 338, + 521 + ], + "spans": [ + { + "bbox": [ + 141, + 508, + 338, + 521 + ], + "score": 1.0, + "content": "the attack success rate for this face identity as 1;", + "type": "text" + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "• Average: we calculate the average attack success rate among 17 attributes for the same face", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 536, + 179, + 550 + ], + "spans": [ + { + "bbox": [ + 141, + 536, + 179, + 550 + ], + "score": 1.0, + "content": "identity;", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 554, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 505, + 567 + ], + "score": 1.0, + "content": "• Worst: only when all of 17 attributes can be successfully attacked, we count the attack", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 566, + 274, + 579 + ], + "spans": [ + { + "bbox": [ + 141, + 566, + 274, + 579 + ], + "score": 1.0, + "content": "success rate for this person as 1;", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + } + ], + "index": 29.5, + "bbox_fs": [ + 131, + 497, + 505, + 579 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 588, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "Note that, for a fair comparison with CW, we should use the Best metric for our SemanticAdv, as CW", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "is the traditional pixel-wise attack method works regardless of the attribute. In addition, we report the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 610, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 622 + ], + "score": 1.0, + "content": "performance using the average and worst metric, which actually provides additional insights into the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 620, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 505, + 634 + ], + "score": 1.0, + "content": "robustness of face verification models across different attributes. For instance, combining the results", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "from Table C and Figure 3, we understand that the face verification models used in our experiments", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 642, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 505, + 655 + ], + "score": 1.0, + "content": "have different levels of robustness across attributes. For example, face verification models are more", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 654, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 505, + 666 + ], + "score": 1.0, + "content": "robust against local shape variations than color variations, e.g., pale skin has higher attack success", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 665, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 677 + ], + "score": 1.0, + "content": "rate than mouth open. We believe these discoveries will help the community further understand the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 676, + 260, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 260, + 688 + ], + "score": 1.0, + "content": "properties of face verification models.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 588, + 505, + 688 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 693, + 504, + 737 + ], + "lines": [ + { + "bbox": [ + 105, + 693, + 506, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 506, + 705 + ], + "score": 1.0, + "content": "Table C shows the overall performance (accuracy) of face verification model and attack success", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 704, + 505, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 704, + 505, + 716 + ], + "score": 1.0, + "content": "rate of SemanticAdv and CW. As we can see from Table C, although the face model trained with", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 715, + 506, + 726 + ], + "spans": [ + { + "bbox": [ + 105, + 715, + 506, + 726 + ], + "score": 1.0, + "content": "cosine loss achieves higher face recognition performance, it is more vulnerable to adversarial attack", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 726, + 316, + 737 + ], + "spans": [ + { + "bbox": [ + 105, + 726, + 316, + 737 + ], + "score": 1.0, + "content": "compared with the model trained with softmax loss.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 693, + 506, + 737 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 80, + 472, + 361 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 137, + 80, + 472, + 361 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 137, + 80, + 472, + 361 + ], + "spans": [ + { + "bbox": [ + 137, + 80, + 472, + 361 + ], + "score": 0.986, + "html": "
G-FPRMetricsMR-50-SR-101-SR-50-CR-101-C
10-3Verification Accuracy98.3698.7898.6398.84
xG(x',c)G(x',cnew)W)(Best)0.000.000.160.000.000.080.080.000.160.000.230.31
SemanticAdv (Best)SemanticAdv (Worst)SemanticAdv (Average)100.0091.9598.98100.0093.9899.29100.0099.5399.97100.0099.7799.99
CW100.00100.00100.00100.00
3 ×10-4Verification Accuracy97.7397.9797.9197.85
xG(x',c)G(x',cnew)(Best)0.000.000.000.000.000.000.000.000.000.000.000.00
SemanticAdv (Best)SemanticAdv (Worst)SemanticAdv (Average)100.0083.7597.72100.0079.0697.35100.0098.9899.92100.0096.6499.72
CW100.00100.00100.00100.00
10-4Verification Accuracy93.2592.8093.4392.98
XG(x',c)G(x',cnew)(Best)0.000.000.000.000.000.000.000.000.000.000.000.00
SemanticAdv (Best)SemanticAdv (Worst)SemanticAdv (Average)100.0033.5983.53100.0019.8476.64100.0067.0395.57100.0048.6791.13
CW100.00100.00100.00100.00
", + "type": "table", + "image_path": "c723aa96ee2ccc87ec708b7666c69b0e35c1469e2eeecab9e18e4c3f9fb0351f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 137, + 80, + 472, + 173.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 137, + 173.66666666666669, + 472, + 267.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 137, + 267.33333333333337, + 472, + 361.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 365, + 504, + 398 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 364, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 308, + 378 + ], + "score": 1.0, + "content": "Table C: Quantitative result of identity verification", + "type": "text" + }, + { + "bbox": [ + 308, + 366, + 323, + 376 + ], + "score": 0.81, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 364, + 506, + 378 + ], + "score": 1.0, + "content": ". It shows accuracy of face verification model", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 374, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 305, + 389 + ], + "score": 1.0, + "content": "and attack success rate of SemanticAdv and CW.", + "type": "text" + }, + { + "bbox": [ + 305, + 376, + 316, + 387 + ], + "score": 0.79, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 374, + 319, + 389 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 320, + 376, + 354, + 388 + ], + "score": 0.93, + "content": "G ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 374, + 373, + 389 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 374, + 376, + 420, + 388 + ], + "score": 0.94, + "content": "G ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } ^ { \\mathrm { n e w } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 374, + 506, + 389 + ], + "score": 1.0, + "content": "are the intermediate", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 387, + 321, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 321, + 399 + ], + "score": 1.0, + "content": "results of our method before adversarial perturbation.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 411, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 105, + 409, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 298, + 424 + ], + "score": 1.0, + "content": "Defense methods. Feature squeezing (", + "type": "text" + }, + { + "bbox": [ + 298, + 411, + 312, + 421 + ], + "score": 0.5, + "content": "\\mathrm { { X u } }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 409, + 506, + 424 + ], + "score": 1.0, + "content": "et al., 2017) is a simple but effective method by", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "score": 1.0, + "content": "reducing color bit depth to remove the adversarial effects. We compress the image represented by", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 432, + 506, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 506, + 445 + ], + "score": 1.0, + "content": "8 bits for each channel to 4 bits for each channel to evaluate the effectiveness. For Blurring (Li", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 443, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 199, + 455 + ], + "score": 1.0, + "content": "& Li, 2017), we use a", + "type": "text" + }, + { + "bbox": [ + 199, + 444, + 223, + 454 + ], + "score": 0.9, + "content": "3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 443, + 505, + 455 + ], + "score": 1.0, + "content": "Gaussian kernel with standard deviation 1 to smooth the adversarial", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 455, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 467 + ], + "score": 1.0, + "content": "perturbations. For JPEG Dziugaite et al. (2016), it leverages the compression and decompression to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 465, + 477, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 477, + 478 + ], + "score": 1.0, + "content": "remove the adversarial perturbation. We set the compression ratio as 0.75 in our experiment.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 108, + 495, + 283, + 508 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 285, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 285, + 510 + ], + "score": 1.0, + "content": "B FACE LANDMARK DETECTION", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 513, + 506, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 514, + 506, + 526 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 506, + 526 + ], + "score": 1.0, + "content": "Implementation details. To perform attack on the face landmark detection model, we run opti-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 524, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 464, + 538 + ], + "score": 1.0, + "content": "mization for up to 2, 000 steps with a fixed learning rate 0.05. We set the balancing factor", + "type": "text" + }, + { + "bbox": [ + 464, + 525, + 471, + 535 + ], + "score": 0.77, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 524, + 506, + 538 + ], + "score": 1.0, + "content": "(see Eq.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 536, + 252, + 548 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 252, + 548 + ], + "score": 1.0, + "content": "3) to value 0.01 for this experiment.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 506, + 587 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 506, + 567 + ], + "score": 1.0, + "content": "Evaluation Metrics. We apply two different metrics for two adversarial attack tasks respectively.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 565, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 506, + 577 + ], + "score": 1.0, + "content": "For “Rotating Eyes” task, we use a well-adopted metric Normalized Mean Error (NME) (Bulat &", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 576, + 315, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 315, + 588 + ], + "score": 1.0, + "content": "Tzimiropoulos, 2017b) for experimental evaluation.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 245, + 594, + 366, + 629 + ], + "lines": [ + { + "bbox": [ + 245, + 594, + 366, + 629 + ], + "spans": [ + { + "bbox": [ + 245, + 594, + 366, + 629 + ], + "score": 0.95, + "content": "r _ { \\mathrm { N M E } } = \\frac { 1 } { N } \\sum _ { k = 1 } ^ { N } \\frac { | | \\mathbf { p } _ { k } - \\hat { \\mathbf { p } } _ { k } | | _ { 2 } } { \\sqrt { W _ { B } * H _ { B } } } ,", + "type": "interline_equation", + "image_path": "b4e26125808bdb559bc0b3affbfb84101b7146cbcdd7dc4829f9d6e7f7fdf291.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 245, + 594, + 366, + 611.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 245, + 611.5, + 366, + 629.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 635, + 505, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 134, + 648 + ], + "score": 1.0, + "content": "where√", + "type": "text" + }, + { + "bbox": [ + 134, + 637, + 147, + 646 + ], + "score": 0.83, + "content": "\\mathbf { p } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 635, + 199, + 648 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + }, + { + "bbox": [ + 199, + 636, + 206, + 645 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 635, + 318, + 648 + ], + "score": 1.0, + "content": "-th ground-truth landmark,", + "type": "text" + }, + { + "bbox": [ + 318, + 636, + 332, + 647 + ], + "score": 0.88, + "content": "\\hat { \\mathbf { p } } _ { \\mathbf { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 635, + 384, + 648 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + }, + { + "bbox": [ + 384, + 636, + 391, + 645 + ], + "score": 0.84, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 635, + 505, + 648 + ], + "score": 1.0, + "content": "-th predicted landmark and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 645, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 107, + 646, + 157, + 658 + ], + "score": 0.92, + "content": "\\sqrt { W _ { B } * H _ { B } }", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 645, + 395, + 659 + ], + "score": 1.0, + "content": "is the square-root area of ground-truth bounding box, where", + "type": "text" + }, + { + "bbox": [ + 396, + 646, + 413, + 658 + ], + "score": 0.9, + "content": "W _ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 645, + 430, + 659 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 431, + 646, + 447, + 658 + ], + "score": 0.9, + "content": "H _ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 645, + 505, + 659 + ], + "score": 1.0, + "content": "represents the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 656, + 222, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 222, + 670 + ], + "score": 1.0, + "content": "width and height of the box.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 673, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "score": 1.0, + "content": "For “Out of Region” task, we consider the attack is successful if the landmark predictions fall outside", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 685, + 505, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 505, + 698 + ], + "score": 1.0, + "content": "a pre-defined centering region on the portrait image. Thus, we introduce a metric that reflects the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 103, + 693, + 506, + 715 + ], + "spans": [ + { + "bbox": [ + 103, + 693, + 426, + 715 + ], + "score": 1.0, + "content": "portion of landmarks outside of the pre-defined centering region: rOUT = NoutNtotal ,", + "type": "text" + }, + { + "bbox": [ + 401, + 695, + 506, + 711 + ], + "score": 1.0, + "content": "where Nout denotes", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 414, + 723 + ], + "score": 1.0, + "content": "the number of predicted landmarks outside the pre-defined bounding box and", + "type": "text" + }, + { + "bbox": [ + 415, + 711, + 437, + 721 + ], + "score": 0.9, + "content": "N _ { \\mathrm { t o t a l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "denotes the total", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 719, + 196, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 196, + 732 + ], + "score": 1.0, + "content": "number of landmarks.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 80, + 472, + 361 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 137, + 80, + 472, + 361 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 137, + 80, + 472, + 361 + ], + "spans": [ + { + "bbox": [ + 137, + 80, + 472, + 361 + ], + "score": 0.986, + "html": "
G-FPRMetricsMR-50-SR-101-SR-50-CR-101-C
10-3Verification Accuracy98.3698.7898.6398.84
xG(x',c)G(x',cnew)W)(Best)0.000.000.160.000.000.080.080.000.160.000.230.31
SemanticAdv (Best)SemanticAdv (Worst)SemanticAdv (Average)100.0091.9598.98100.0093.9899.29100.0099.5399.97100.0099.7799.99
CW100.00100.00100.00100.00
3 ×10-4Verification Accuracy97.7397.9797.9197.85
xG(x',c)G(x',cnew)(Best)0.000.000.000.000.000.000.000.000.000.000.000.00
SemanticAdv (Best)SemanticAdv (Worst)SemanticAdv (Average)100.0083.7597.72100.0079.0697.35100.0098.9899.92100.0096.6499.72
CW100.00100.00100.00100.00
10-4Verification Accuracy93.2592.8093.4392.98
XG(x',c)G(x',cnew)(Best)0.000.000.000.000.000.000.000.000.000.000.000.00
SemanticAdv (Best)SemanticAdv (Worst)SemanticAdv (Average)100.0033.5983.53100.0019.8476.64100.0067.0395.57100.0048.6791.13
CW100.00100.00100.00100.00
", + "type": "table", + "image_path": "c723aa96ee2ccc87ec708b7666c69b0e35c1469e2eeecab9e18e4c3f9fb0351f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 137, + 80, + 472, + 173.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 137, + 173.66666666666669, + 472, + 267.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 137, + 267.33333333333337, + 472, + 361.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 365, + 504, + 398 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 364, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 308, + 378 + ], + "score": 1.0, + "content": "Table C: Quantitative result of identity verification", + "type": "text" + }, + { + "bbox": [ + 308, + 366, + 323, + 376 + ], + "score": 0.81, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 364, + 506, + 378 + ], + "score": 1.0, + "content": ". It shows accuracy of face verification model", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 374, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 305, + 389 + ], + "score": 1.0, + "content": "and attack success rate of SemanticAdv and CW.", + "type": "text" + }, + { + "bbox": [ + 305, + 376, + 316, + 387 + ], + "score": 0.79, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 374, + 319, + 389 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 320, + 376, + 354, + 388 + ], + "score": 0.93, + "content": "G ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 374, + 373, + 389 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 374, + 376, + 420, + 388 + ], + "score": 0.94, + "content": "G ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } ^ { \\mathrm { n e w } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 374, + 506, + 389 + ], + "score": 1.0, + "content": "are the intermediate", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 387, + 321, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 321, + 399 + ], + "score": 1.0, + "content": "results of our method before adversarial perturbation.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 411, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 105, + 409, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 298, + 424 + ], + "score": 1.0, + "content": "Defense methods. Feature squeezing (", + "type": "text" + }, + { + "bbox": [ + 298, + 411, + 312, + 421 + ], + "score": 0.5, + "content": "\\mathrm { { X u } }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 409, + 506, + 424 + ], + "score": 1.0, + "content": "et al., 2017) is a simple but effective method by", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "score": 1.0, + "content": "reducing color bit depth to remove the adversarial effects. We compress the image represented by", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 432, + 506, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 506, + 445 + ], + "score": 1.0, + "content": "8 bits for each channel to 4 bits for each channel to evaluate the effectiveness. For Blurring (Li", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 443, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 199, + 455 + ], + "score": 1.0, + "content": "& Li, 2017), we use a", + "type": "text" + }, + { + "bbox": [ + 199, + 444, + 223, + 454 + ], + "score": 0.9, + "content": "3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 443, + 505, + 455 + ], + "score": 1.0, + "content": "Gaussian kernel with standard deviation 1 to smooth the adversarial", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 455, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 467 + ], + "score": 1.0, + "content": "perturbations. For JPEG Dziugaite et al. (2016), it leverages the compression and decompression to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 465, + 477, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 477, + 478 + ], + "score": 1.0, + "content": "remove the adversarial perturbation. We set the compression ratio as 0.75 in our experiment.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 409, + 506, + 478 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 495, + 283, + 508 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 285, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 285, + 510 + ], + "score": 1.0, + "content": "B FACE LANDMARK DETECTION", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 513, + 506, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 514, + 506, + 526 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 506, + 526 + ], + "score": 1.0, + "content": "Implementation details. To perform attack on the face landmark detection model, we run opti-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 524, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 464, + 538 + ], + "score": 1.0, + "content": "mization for up to 2, 000 steps with a fixed learning rate 0.05. We set the balancing factor", + "type": "text" + }, + { + "bbox": [ + 464, + 525, + 471, + 535 + ], + "score": 0.77, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 524, + 506, + 538 + ], + "score": 1.0, + "content": "(see Eq.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 536, + 252, + 548 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 252, + 548 + ], + "score": 1.0, + "content": "3) to value 0.01 for this experiment.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 514, + 506, + 548 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 506, + 587 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 506, + 567 + ], + "score": 1.0, + "content": "Evaluation Metrics. We apply two different metrics for two adversarial attack tasks respectively.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 565, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 506, + 577 + ], + "score": 1.0, + "content": "For “Rotating Eyes” task, we use a well-adopted metric Normalized Mean Error (NME) (Bulat &", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 576, + 315, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 315, + 588 + ], + "score": 1.0, + "content": "Tzimiropoulos, 2017b) for experimental evaluation.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 553, + 506, + 588 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 245, + 594, + 366, + 629 + ], + "lines": [ + { + "bbox": [ + 245, + 594, + 366, + 629 + ], + "spans": [ + { + "bbox": [ + 245, + 594, + 366, + 629 + ], + "score": 0.95, + "content": "r _ { \\mathrm { N M E } } = \\frac { 1 } { N } \\sum _ { k = 1 } ^ { N } \\frac { | | \\mathbf { p } _ { k } - \\hat { \\mathbf { p } } _ { k } | | _ { 2 } } { \\sqrt { W _ { B } * H _ { B } } } ,", + "type": "interline_equation", + "image_path": "b4e26125808bdb559bc0b3affbfb84101b7146cbcdd7dc4829f9d6e7f7fdf291.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 245, + 594, + 366, + 611.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 245, + 611.5, + 366, + 629.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 635, + 505, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 134, + 648 + ], + "score": 1.0, + "content": "where√", + "type": "text" + }, + { + "bbox": [ + 134, + 637, + 147, + 646 + ], + "score": 0.83, + "content": "\\mathbf { p } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 635, + 199, + 648 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + }, + { + "bbox": [ + 199, + 636, + 206, + 645 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 635, + 318, + 648 + ], + "score": 1.0, + "content": "-th ground-truth landmark,", + "type": "text" + }, + { + "bbox": [ + 318, + 636, + 332, + 647 + ], + "score": 0.88, + "content": "\\hat { \\mathbf { p } } _ { \\mathbf { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 635, + 384, + 648 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + }, + { + "bbox": [ + 384, + 636, + 391, + 645 + ], + "score": 0.84, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 635, + 505, + 648 + ], + "score": 1.0, + "content": "-th predicted landmark and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 645, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 107, + 646, + 157, + 658 + ], + "score": 0.92, + "content": "\\sqrt { W _ { B } * H _ { B } }", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 645, + 395, + 659 + ], + "score": 1.0, + "content": "is the square-root area of ground-truth bounding box, where", + "type": "text" + }, + { + "bbox": [ + 396, + 646, + 413, + 658 + ], + "score": 0.9, + "content": "W _ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 645, + 430, + 659 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 431, + 646, + 447, + 658 + ], + "score": 0.9, + "content": "H _ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 645, + 505, + 659 + ], + "score": 1.0, + "content": "represents the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 656, + 222, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 222, + 670 + ], + "score": 1.0, + "content": "width and height of the box.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 635, + 505, + 670 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 673, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "score": 1.0, + "content": "For “Out of Region” task, we consider the attack is successful if the landmark predictions fall outside", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 685, + 505, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 505, + 698 + ], + "score": 1.0, + "content": "a pre-defined centering region on the portrait image. Thus, we introduce a metric that reflects the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 103, + 693, + 506, + 715 + ], + "spans": [ + { + "bbox": [ + 103, + 693, + 426, + 715 + ], + "score": 1.0, + "content": "portion of landmarks outside of the pre-defined centering region: rOUT = NoutNtotal ,", + "type": "text" + }, + { + "bbox": [ + 401, + 695, + 506, + 711 + ], + "score": 1.0, + "content": "where Nout denotes", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 414, + 723 + ], + "score": 1.0, + "content": "the number of predicted landmarks outside the pre-defined bounding box and", + "type": "text" + }, + { + "bbox": [ + 415, + 711, + 437, + 721 + ], + "score": 0.9, + "content": "N _ { \\mathrm { t o t a l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "denotes the total", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 719, + 196, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 196, + 732 + ], + "score": 1.0, + "content": "number of landmarks.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26, + "bbox_fs": [ + 103, + 674, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 113, + 80, + 496, + 126 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 113, + 80, + 496, + 126 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 80, + 496, + 126 + ], + "spans": [ + { + "bbox": [ + 113, + 80, + 496, + 126 + ], + "score": 0.976, + "html": "
Tasks (Metrics)PristineAugmented Attributes
BlondHairYoungEyeglassesRosyCheeksSmilingArchedEyebrowsBangsPaleSkin
Rotating eyes (rNME)↓28.0414.0317.288.5813.2419.2123.4215.9910.72
Out-of-region (rouT)↓45.9817.4223.047.5116.6525.4433.8520.0313.51
", + "type": "table", + "image_path": "15014ed6d55c113e14271d46bd0e3619d3839cde6544458347bb6b4ba6bb7f93.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 113, + 80, + 496, + 95.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 113, + 95.33333333333333, + 496, + 110.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 113, + 110.66666666666666, + 496, + 125.99999999999999 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 107, + 131, + 504, + 153 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 129, + 505, + 142 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 331, + 142 + ], + "score": 1.0, + "content": "Table D: Quantitative results on face landmark detection", + "type": "text" + }, + { + "bbox": [ + 332, + 131, + 344, + 141 + ], + "score": 0.63, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 129, + 505, + 142 + ], + "score": 1.0, + "content": ") The two row shows the measured ratios", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 141, + 414, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 414, + 154 + ], + "score": 1.0, + "content": "(lower is better) for “Rotating Eyes” and “Out Of Region” task, respectively.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 169, + 505, + 213 + ], + "lines": [ + { + "bbox": [ + 105, + 169, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 505, + 180 + ], + "score": 1.0, + "content": "We present the quantitative results of SemanticAdv on face landmark detection model in Table D", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 180, + 506, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 506, + 192 + ], + "score": 1.0, + "content": "including two adversarial tasks, namely, “Out of Region” and “Rotating Eyes”. We observe that our", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 191, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 203 + ], + "score": 1.0, + "content": "method is efficient to perform attacking on landmark detection models. For certain attributes such as", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 202, + 464, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 464, + 215 + ], + "score": 1.0, + "content": "“Eyeglasses”, “Plae Skin”, SemanticAdv is able to achieve reasonably-good performance.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 107, + 235, + 195, + 248 + ], + "lines": [ + { + "bbox": [ + 105, + 233, + 197, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 197, + 251 + ], + "score": 1.0, + "content": "C USER STUDY", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 264, + 505, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 506, + 276 + ], + "score": 1.0, + "content": "We conducted a user study on the adversarial images of SemanticAdv and CW used in the experiment", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 273, + 504, + 288 + ], + "spans": [ + { + "bbox": [ + 104, + 273, + 442, + 288 + ], + "score": 1.0, + "content": "of API-attack and the original images. The adversarial images are generated with", + "type": "text" + }, + { + "bbox": [ + 442, + 275, + 504, + 286 + ], + "score": 0.84, + "content": "\\mathrm { { G } \\mathrm { { - F P R } \\mathrm { { < } \\ 1 0 ^ { - 4 } } } }", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "score": 1.0, + "content": "for both methods. We present a pair of original image and adversarial image to participants and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 297, + 505, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 310 + ], + "score": 1.0, + "content": "ask them to rank the two options. The order of these two images is randomized and the images are", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 307, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 505, + 321 + ], + "score": 1.0, + "content": "displayed for 2 seconds in the screen during each trial. After the images disappear, the participants", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 319, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 505, + 331 + ], + "score": 1.0, + "content": "have unlimited time to select the more reasonably-looking image according to their perception. For", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 330, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 505, + 342 + ], + "score": 1.0, + "content": "each participant, one could only conduct at most 50 trials, and each adversarial image was shown to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 341, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 505, + 352 + ], + "score": 1.0, + "content": "5 different participants. Some qualitative results are shown in Figure H. In total, we collect 2, 620", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 351, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 257, + 365 + ], + "score": 1.0, + "content": "annotations from 77 participants. In", + "type": "text" + }, + { + "bbox": [ + 257, + 352, + 320, + 363 + ], + "score": 0.91, + "content": "3 9 . 1 4 \\pm 1 . 9 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 351, + 505, + 365 + ], + "score": 1.0, + "content": "of trials the adversarial images generated by", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 362, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 104, + 362, + 351, + 375 + ], + "score": 1.0, + "content": "SemanticAdv are selected as reasonably-looking images and in", + "type": "text" + }, + { + "bbox": [ + 351, + 362, + 410, + 374 + ], + "score": 0.92, + "content": "3 0 . 2 7 \\pm 1 . 9 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 362, + 506, + 375 + ], + "score": 1.0, + "content": "of trails, the adversarial", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "score": 1.0, + "content": "images generated by CW are selected as reasonably-looking images. It indicates that our semantic", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "adversarial examples are more perceptual reasonably-looking than CW. Additionally, we also conduct", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 207, + 408 + ], + "score": 1.0, + "content": "the user study with larger", + "type": "text" + }, + { + "bbox": [ + 207, + 395, + 266, + 406 + ], + "score": 0.67, + "content": "\\mathrm { { G } { - } \\mathrm { { F P R } } \\mathrm { { = } \\ i 0 ^ { - 3 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 394, + 281, + 408 + ], + "score": 1.0, + "content": ". In", + "type": "text" + }, + { + "bbox": [ + 281, + 395, + 341, + 406 + ], + "score": 0.91, + "content": "4 5 . 4 2 \\pm 1 . 9 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "of trials, the adversarial images generated", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "by SemanticAdv are selected as reasonably-looking images, which is very close to the random guess", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 417, + 137, + 430 + ], + "spans": [ + { + "bbox": [ + 107, + 417, + 131, + 429 + ], + "score": 0.82, + "content": "( \\dot { 5 } 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 417, + 137, + 430 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 106, + 444, + 504, + 679 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 444, + 504, + 679 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 444, + 504, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 504, + 679 + ], + "score": 0.955, + "type": "image", + "image_path": "e11a9c182edab33d61fc6b803c35b0b754579590f5fdb181259f75a0c14af2ed.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 106, + 444, + 504, + 522.3333333333334 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 106, + 522.3333333333334, + 504, + 600.6666666666667 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 106, + 600.6666666666667, + 504, + 679.0000000000001 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 695, + 504, + 728 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 694, + 506, + 708 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 506, + 708 + ], + "score": 1.0, + "content": "Figure H: Qualitative comparisons among ground truth, pixel-wise adversarial examples generated", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 705, + 506, + 718 + ], + "spans": [ + { + "bbox": [ + 105, + 705, + 443, + 718 + ], + "score": 1.0, + "content": "by CW, and our proposed SemanticAdv. Here, we present the results from G-FPR", + "type": "text" + }, + { + "bbox": [ + 443, + 705, + 474, + 717 + ], + "score": 0.8, + "content": "\\dot { < } 1 0 ^ { { - } 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 705, + 506, + 718 + ], + "score": 1.0, + "content": "so that", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 717, + 208, + 728 + ], + "spans": [ + { + "bbox": [ + 105, + 717, + 208, + 728 + ], + "score": 1.0, + "content": "perturbations are visible.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + } + ], + "index": 27.5 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 752, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 113, + 80, + 496, + 126 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 113, + 80, + 496, + 126 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 80, + 496, + 126 + ], + "spans": [ + { + "bbox": [ + 113, + 80, + 496, + 126 + ], + "score": 0.976, + "html": "
Tasks (Metrics)PristineAugmented Attributes
BlondHairYoungEyeglassesRosyCheeksSmilingArchedEyebrowsBangsPaleSkin
Rotating eyes (rNME)↓28.0414.0317.288.5813.2419.2123.4215.9910.72
Out-of-region (rouT)↓45.9817.4223.047.5116.6525.4433.8520.0313.51
", + "type": "table", + "image_path": "15014ed6d55c113e14271d46bd0e3619d3839cde6544458347bb6b4ba6bb7f93.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 113, + 80, + 496, + 95.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 113, + 95.33333333333333, + 496, + 110.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 113, + 110.66666666666666, + 496, + 125.99999999999999 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 107, + 131, + 504, + 153 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 129, + 505, + 142 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 331, + 142 + ], + "score": 1.0, + "content": "Table D: Quantitative results on face landmark detection", + "type": "text" + }, + { + "bbox": [ + 332, + 131, + 344, + 141 + ], + "score": 0.63, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 129, + 505, + 142 + ], + "score": 1.0, + "content": ") The two row shows the measured ratios", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 141, + 414, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 414, + 154 + ], + "score": 1.0, + "content": "(lower is better) for “Rotating Eyes” and “Out Of Region” task, respectively.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 169, + 505, + 213 + ], + "lines": [ + { + "bbox": [ + 105, + 169, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 505, + 180 + ], + "score": 1.0, + "content": "We present the quantitative results of SemanticAdv on face landmark detection model in Table D", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 180, + 506, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 506, + 192 + ], + "score": 1.0, + "content": "including two adversarial tasks, namely, “Out of Region” and “Rotating Eyes”. We observe that our", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 191, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 203 + ], + "score": 1.0, + "content": "method is efficient to perform attacking on landmark detection models. For certain attributes such as", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 202, + 464, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 464, + 215 + ], + "score": 1.0, + "content": "“Eyeglasses”, “Plae Skin”, SemanticAdv is able to achieve reasonably-good performance.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 169, + 506, + 215 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 235, + 195, + 248 + ], + "lines": [ + { + "bbox": [ + 105, + 233, + 197, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 197, + 251 + ], + "score": 1.0, + "content": "C USER STUDY", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 264, + 505, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 506, + 276 + ], + "score": 1.0, + "content": "We conducted a user study on the adversarial images of SemanticAdv and CW used in the experiment", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 273, + 504, + 288 + ], + "spans": [ + { + "bbox": [ + 104, + 273, + 442, + 288 + ], + "score": 1.0, + "content": "of API-attack and the original images. The adversarial images are generated with", + "type": "text" + }, + { + "bbox": [ + 442, + 275, + 504, + 286 + ], + "score": 0.84, + "content": "\\mathrm { { G } \\mathrm { { - F P R } \\mathrm { { < } \\ 1 0 ^ { - 4 } } } }", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "score": 1.0, + "content": "for both methods. We present a pair of original image and adversarial image to participants and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 297, + 505, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 310 + ], + "score": 1.0, + "content": "ask them to rank the two options. The order of these two images is randomized and the images are", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 307, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 505, + 321 + ], + "score": 1.0, + "content": "displayed for 2 seconds in the screen during each trial. After the images disappear, the participants", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 319, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 505, + 331 + ], + "score": 1.0, + "content": "have unlimited time to select the more reasonably-looking image according to their perception. For", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 330, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 505, + 342 + ], + "score": 1.0, + "content": "each participant, one could only conduct at most 50 trials, and each adversarial image was shown to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 341, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 505, + 352 + ], + "score": 1.0, + "content": "5 different participants. Some qualitative results are shown in Figure H. In total, we collect 2, 620", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 351, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 257, + 365 + ], + "score": 1.0, + "content": "annotations from 77 participants. In", + "type": "text" + }, + { + "bbox": [ + 257, + 352, + 320, + 363 + ], + "score": 0.91, + "content": "3 9 . 1 4 \\pm 1 . 9 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 351, + 505, + 365 + ], + "score": 1.0, + "content": "of trials the adversarial images generated by", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 362, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 104, + 362, + 351, + 375 + ], + "score": 1.0, + "content": "SemanticAdv are selected as reasonably-looking images and in", + "type": "text" + }, + { + "bbox": [ + 351, + 362, + 410, + 374 + ], + "score": 0.92, + "content": "3 0 . 2 7 \\pm 1 . 9 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 362, + 506, + 375 + ], + "score": 1.0, + "content": "of trails, the adversarial", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "score": 1.0, + "content": "images generated by CW are selected as reasonably-looking images. It indicates that our semantic", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "adversarial examples are more perceptual reasonably-looking than CW. Additionally, we also conduct", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 207, + 408 + ], + "score": 1.0, + "content": "the user study with larger", + "type": "text" + }, + { + "bbox": [ + 207, + 395, + 266, + 406 + ], + "score": 0.67, + "content": "\\mathrm { { G } { - } \\mathrm { { F P R } } \\mathrm { { = } \\ i 0 ^ { - 3 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 394, + 281, + 408 + ], + "score": 1.0, + "content": ". In", + "type": "text" + }, + { + "bbox": [ + 281, + 395, + 341, + 406 + ], + "score": 0.91, + "content": "4 5 . 4 2 \\pm 1 . 9 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "of trials, the adversarial images generated", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "by SemanticAdv are selected as reasonably-looking images, which is very close to the random guess", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 417, + 137, + 430 + ], + "spans": [ + { + "bbox": [ + 107, + 417, + 131, + 429 + ], + "score": 0.82, + "content": "( \\dot { 5 } 0 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 417, + 137, + 430 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 17, + "bbox_fs": [ + 104, + 264, + 506, + 430 + ] + }, + { + "type": "image", + "bbox": [ + 106, + 444, + 504, + 679 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 444, + 504, + 679 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 444, + 504, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 504, + 679 + ], + "score": 0.955, + "type": "image", + "image_path": "e11a9c182edab33d61fc6b803c35b0b754579590f5fdb181259f75a0c14af2ed.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 106, + 444, + 504, + 522.3333333333334 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 106, + 522.3333333333334, + 504, + 600.6666666666667 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 106, + 600.6666666666667, + 504, + 679.0000000000001 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 695, + 504, + 728 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 694, + 506, + 708 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 506, + 708 + ], + "score": 1.0, + "content": "Figure H: Qualitative comparisons among ground truth, pixel-wise adversarial examples generated", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 705, + 506, + 718 + ], + "spans": [ + { + "bbox": [ + 105, + 705, + 443, + 718 + ], + "score": 1.0, + "content": "by CW, and our proposed SemanticAdv. Here, we present the results from G-FPR", + "type": "text" + }, + { + "bbox": [ + 443, + 705, + 474, + 717 + ], + "score": 0.8, + "content": "\\dot { < } 1 0 ^ { { - } 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 705, + 506, + 718 + ], + "score": 1.0, + "content": "so that", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 717, + 208, + 728 + ], + "spans": [ + { + "bbox": [ + 105, + 717, + 208, + 728 + ], + "score": 1.0, + "content": "perturbations are visible.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + } + ], + "index": 27.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 399, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 401, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 401, + 95 + ], + "score": 1.0, + "content": "D ABLATION STUDY: FEATURE-SPACE INTERPOLATION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 105, + 506, + 183 + ], + "lines": [ + { + "bbox": [ + 105, + 105, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 506, + 119 + ], + "score": 1.0, + "content": "We conduct an ablation study on feature-space interpolation by analyzing attack success rates with", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 117, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 117, + 506, + 128 + ], + "score": 1.0, + "content": "different feature-maps in the StarGAN network. Table E shows the attack success rate on R-101-S.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 127, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 160, + 140 + ], + "score": 1.0, + "content": "Here, we use", + "type": "text" + }, + { + "bbox": [ + 160, + 128, + 168, + 138 + ], + "score": 0.86, + "content": "\\mathbf { f } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 127, + 304, + 140 + ], + "score": 1.0, + "content": "to represent the feature-map after", + "type": "text" + }, + { + "bbox": [ + 304, + 128, + 309, + 137 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 127, + 415, + 140 + ], + "score": 1.0, + "content": "-th up-sampling operation.", + "type": "text" + }, + { + "bbox": [ + 415, + 128, + 425, + 139 + ], + "score": 0.76, + "content": "\\mathbf { f } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 127, + 506, + 140 + ], + "score": 1.0, + "content": "denotes the feature-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 138, + 506, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 506, + 152 + ], + "score": 1.0, + "content": "map before applying up-sampling operation. The result demonstrates that samples generated by", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 173, + 162 + ], + "score": 1.0, + "content": "interpolating on", + "type": "text" + }, + { + "bbox": [ + 174, + 150, + 183, + 161 + ], + "score": 0.86, + "content": "\\mathbf { f } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 149, + 344, + 162 + ], + "score": 1.0, + "content": "achieve the highest success rate. Since", + "type": "text" + }, + { + "bbox": [ + 344, + 150, + 354, + 161 + ], + "score": 0.84, + "content": "\\mathbf { f } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 149, + 506, + 162 + ], + "score": 1.0, + "content": "is the feature-map before decoder, it", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 160, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 397, + 173 + ], + "score": 1.0, + "content": "still well embeds semantic information in the feature space. We adopt", + "type": "text" + }, + { + "bbox": [ + 397, + 160, + 407, + 172 + ], + "score": 0.84, + "content": "\\mathbf { f } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 160, + 506, + 173 + ], + "score": 1.0, + "content": "for interpolation in our", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 172, + 160, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 160, + 185 + ], + "score": 1.0, + "content": "experiments.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 111, + 230, + 502, + 267 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 188, + 505, + 222 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 187, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 201 + ], + "score": 1.0, + "content": "We also conduct a qualitative comparison between attribute-space and feature-space interpolation.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 199, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 199, + 505, + 212 + ], + "score": 1.0, + "content": "As shown in Figure I, images generated by attribute-space and feature-space interpolation are both", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 208, + 189, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 189, + 224 + ], + "score": 1.0, + "content": "reasonably-looking.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "table_body", + "bbox": [ + 111, + 230, + 502, + 267 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 230, + 502, + 267 + ], + "spans": [ + { + "bbox": [ + 111, + 230, + 502, + 267 + ], + "score": 0.976, + "html": "
T-FPR (G-FPR)10-3 (10-3)3× 10-4 (3×10-4)10-4 (10-4)
Layer (f)ff1f2ff1f2ff1f
Attack SuccessRate99.2998.3275.6297.3594.1057.1576.6467.4019.63
", + "type": "table", + "image_path": "91a2bd6a85284f2649ead1cc0ff2b5c630d90d41b7256375255f9041a4ee43e7.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 111, + 230, + 502, + 242.33333333333334 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 111, + 242.33333333333334, + 502, + 254.66666666666669 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 111, + 254.66666666666669, + 502, + 267.0 + ], + "spans": [], + "index": 13 + } + ] + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 276, + 505, + 298 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 503, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 454, + 289 + ], + "score": 1.0, + "content": "Table E: Attack success rate by selecting different layer’s feature-map for interpolation on", + "type": "text" + }, + { + "bbox": [ + 454, + 276, + 503, + 287 + ], + "score": 0.52, + "content": "\\mathrm { R - } 1 0 1 { - } \\mathrm { S } ( \\% )", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 286, + 351, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 114, + 298 + ], + "score": 0.83, + "content": "\\mathbf { f } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 286, + 240, + 299 + ], + "score": 1.0, + "content": "indicates the feature-map after", + "type": "text" + }, + { + "bbox": [ + 240, + 288, + 244, + 297 + ], + "score": 0.61, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 286, + 351, + 299 + ], + "score": 1.0, + "content": "-th up-sampling operation.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "table", + "bbox": [ + 150, + 312, + 460, + 350 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 150, + 312, + 460, + 350 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 150, + 312, + 460, + 350 + ], + "spans": [ + { + "bbox": [ + 150, + 312, + 460, + 350 + ], + "score": 0.973, + "html": "
T-FPR (G-FPR)10-3 (10-3)3×10)-4(3×10-4)10-4(10-4)
Layer (f)f-2f-1f-2f-1f-2f-1
Attack Success Rate49.4092.0930.4481.876.6645.46
", + "type": "table", + "image_path": "d4093ecc5c81935abb31c8c6f5d35a1f9dcd51757521594078caa223e7ff226b.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 150, + 312, + 460, + 324.6666666666667 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 150, + 324.6666666666667, + 460, + 337.33333333333337 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 150, + 337.33333333333337, + 460, + 350.00000000000006 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 358, + 505, + 392 + ], + "lines": [ + { + "bbox": [ + 105, + 357, + 507, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 507, + 370 + ], + "score": 1.0, + "content": "Table F: Attack success rate by selecting different layer’s feature-map for interpolation on R-101-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 128, + 380 + ], + "score": 0.84, + "content": "S ( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 369, + 133, + 381 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 133, + 369, + 149, + 381 + ], + "score": 0.77, + "content": "\\mathbf { f } _ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 369, + 434, + 381 + ], + "score": 1.0, + "content": "indicates the feature-map after the last down-sampling operation and", + "type": "text" + }, + { + "bbox": [ + 434, + 369, + 450, + 380 + ], + "score": 0.86, + "content": "\\mathbf { f } _ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "indicates the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 378, + 198, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 177, + 394 + ], + "score": 1.0, + "content": "feature-map after", + "type": "text" + }, + { + "bbox": [ + 178, + 380, + 193, + 391 + ], + "score": 0.79, + "content": "\\mathbf { f } _ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 378, + 198, + 394 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 107, + 407, + 498, + 541 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 407, + 498, + 541 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 407, + 498, + 541 + ], + "spans": [ + { + "bbox": [ + 107, + 407, + 498, + 541 + ], + "score": 0.968, + "type": "image", + "image_path": "d2f867cde088d9d2eedec96d6e11f67a1803de628ed29a100713bd1b68d74171.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 107, + 407, + 498, + 451.6666666666667 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 107, + 451.6666666666667, + 498, + 496.33333333333337 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 107, + 496.33333333333337, + 498, + 541.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 125, + 559, + 485, + 571 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 558, + 486, + 572 + ], + "spans": [ + { + "bbox": [ + 124, + 558, + 486, + 572 + ], + "score": 1.0, + "content": "Figure I: Qualitative comparisons between attribute-space and feature-space interpolation.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + } + ], + "index": 24.0 + }, + { + "type": "title", + "bbox": [ + 106, + 581, + 327, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 580, + 329, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 329, + 595 + ], + "score": 1.0, + "content": "E SEMANTIC ATTACK TRANSFERABILITY", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 605, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 408, + 617 + ], + "score": 1.0, + "content": "In Table G, we present the quantitative results of the transferability with", + "type": "text" + }, + { + "bbox": [ + 409, + 604, + 471, + 615 + ], + "score": 0.9, + "content": "\\mathrm { { G } \\mathrm { { - } \\mathrm { { F P R } = 1 0 ^ { - 4 } } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 604, + 505, + 617 + ], + "score": 1.0, + "content": ", T-FPR", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 614, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 137, + 627 + ], + "score": 0.9, + "content": "= 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 614, + 505, + 629 + ], + "score": 1.0, + "content": ". We observe that with the strict criterion (Lower T-FPR) of the verification model, the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 627, + 321, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 321, + 639 + ], + "score": 1.0, + "content": "transferability becomes lower cross different models.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 642, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 506, + 657 + ], + "score": 1.0, + "content": "To explore whether the improvement of transferability in SemanticAdv is introduced by semantic", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 476, + 667 + ], + "score": 1.0, + "content": "editing rather than optimized semantic perturbation in feature space. We add a StrawMan", + "type": "text" + }, + { + "bbox": [ + 477, + 655, + 506, + 666 + ], + "score": 0.63, + "content": "\\mathrm { [ C W + }", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 664, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 128, + 677 + ], + "score": 0.8, + "content": "\\mathbf { x } ^ { \\mathrm { n e w } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 664, + 505, + 679 + ], + "score": 1.0, + "content": "baseline which uses a controllable semantic-attribute-based generator to generate semantically", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 443, + 689 + ], + "score": 1.0, + "content": "different images without any notion of an adversarial attack, and then applies standard", + "type": "text" + }, + { + "bbox": [ + 444, + 677, + 456, + 689 + ], + "score": 0.88, + "content": "L _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "CW attacks", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 477, + 700 + ], + "score": 1.0, + "content": "on that generated image. The results are shown in Table H. The performance of the StrawMan", + "type": "text" + }, + { + "bbox": [ + 478, + 688, + 506, + 699 + ], + "score": 0.65, + "content": "\\mathrm { [ C W + }", + "type": "inline_equation" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 696, + 507, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 128, + 710 + ], + "score": 0.87, + "content": "\\mathbf { x } ^ { \\mathrm { n e w } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 696, + 507, + 712 + ], + "score": 1.0, + "content": "is worse than SemanticAdv. This result justifies that our SemanticAdv is able to produce novel", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "adversarial examples which cannot be simply achieved by combining attribute-conditional image", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 720, + 289, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 184, + 733 + ], + "score": 1.0, + "content": "editing model with", + "type": "text" + }, + { + "bbox": [ + 184, + 721, + 197, + 733 + ], + "score": 0.9, + "content": "L _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 720, + 289, + 733 + ], + "score": 1.0, + "content": "bounded perturbation.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 399, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 401, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 401, + 95 + ], + "score": 1.0, + "content": "D ABLATION STUDY: FEATURE-SPACE INTERPOLATION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 105, + 506, + 183 + ], + "lines": [ + { + "bbox": [ + 105, + 105, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 506, + 119 + ], + "score": 1.0, + "content": "We conduct an ablation study on feature-space interpolation by analyzing attack success rates with", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 117, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 117, + 506, + 128 + ], + "score": 1.0, + "content": "different feature-maps in the StarGAN network. Table E shows the attack success rate on R-101-S.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 127, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 160, + 140 + ], + "score": 1.0, + "content": "Here, we use", + "type": "text" + }, + { + "bbox": [ + 160, + 128, + 168, + 138 + ], + "score": 0.86, + "content": "\\mathbf { f } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 127, + 304, + 140 + ], + "score": 1.0, + "content": "to represent the feature-map after", + "type": "text" + }, + { + "bbox": [ + 304, + 128, + 309, + 137 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 127, + 415, + 140 + ], + "score": 1.0, + "content": "-th up-sampling operation.", + "type": "text" + }, + { + "bbox": [ + 415, + 128, + 425, + 139 + ], + "score": 0.76, + "content": "\\mathbf { f } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 127, + 506, + 140 + ], + "score": 1.0, + "content": "denotes the feature-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 138, + 506, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 506, + 152 + ], + "score": 1.0, + "content": "map before applying up-sampling operation. The result demonstrates that samples generated by", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 173, + 162 + ], + "score": 1.0, + "content": "interpolating on", + "type": "text" + }, + { + "bbox": [ + 174, + 150, + 183, + 161 + ], + "score": 0.86, + "content": "\\mathbf { f } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 149, + 344, + 162 + ], + "score": 1.0, + "content": "achieve the highest success rate. Since", + "type": "text" + }, + { + "bbox": [ + 344, + 150, + 354, + 161 + ], + "score": 0.84, + "content": "\\mathbf { f } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 149, + 506, + 162 + ], + "score": 1.0, + "content": "is the feature-map before decoder, it", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 160, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 397, + 173 + ], + "score": 1.0, + "content": "still well embeds semantic information in the feature space. We adopt", + "type": "text" + }, + { + "bbox": [ + 397, + 160, + 407, + 172 + ], + "score": 0.84, + "content": "\\mathbf { f } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 160, + 506, + 173 + ], + "score": 1.0, + "content": "for interpolation in our", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 172, + 160, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 160, + 185 + ], + "score": 1.0, + "content": "experiments.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 105, + 506, + 185 + ] + }, + { + "type": "table", + "bbox": [ + 111, + 230, + 502, + 267 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 188, + 505, + 222 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 187, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 201 + ], + "score": 1.0, + "content": "We also conduct a qualitative comparison between attribute-space and feature-space interpolation.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 199, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 199, + 505, + 212 + ], + "score": 1.0, + "content": "As shown in Figure I, images generated by attribute-space and feature-space interpolation are both", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 208, + 189, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 189, + 224 + ], + "score": 1.0, + "content": "reasonably-looking.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "table_body", + "bbox": [ + 111, + 230, + 502, + 267 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 230, + 502, + 267 + ], + "spans": [ + { + "bbox": [ + 111, + 230, + 502, + 267 + ], + "score": 0.976, + "html": "
T-FPR (G-FPR)10-3 (10-3)3× 10-4 (3×10-4)10-4 (10-4)
Layer (f)ff1f2ff1f2ff1f
Attack SuccessRate99.2998.3275.6297.3594.1057.1576.6467.4019.63
", + "type": "table", + "image_path": "91a2bd6a85284f2649ead1cc0ff2b5c630d90d41b7256375255f9041a4ee43e7.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 111, + 230, + 502, + 242.33333333333334 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 111, + 242.33333333333334, + 502, + 254.66666666666669 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 111, + 254.66666666666669, + 502, + 267.0 + ], + "spans": [], + "index": 13 + } + ] + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 276, + 505, + 298 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 503, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 454, + 289 + ], + "score": 1.0, + "content": "Table E: Attack success rate by selecting different layer’s feature-map for interpolation on", + "type": "text" + }, + { + "bbox": [ + 454, + 276, + 503, + 287 + ], + "score": 0.52, + "content": "\\mathrm { R - } 1 0 1 { - } \\mathrm { S } ( \\% )", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 286, + 351, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 114, + 298 + ], + "score": 0.83, + "content": "\\mathbf { f } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 286, + 240, + 299 + ], + "score": 1.0, + "content": "indicates the feature-map after", + "type": "text" + }, + { + "bbox": [ + 240, + 288, + 244, + 297 + ], + "score": 0.61, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 286, + 351, + 299 + ], + "score": 1.0, + "content": "-th up-sampling operation.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 274, + 503, + 299 + ] + }, + { + "type": "table", + "bbox": [ + 150, + 312, + 460, + 350 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 150, + 312, + 460, + 350 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 150, + 312, + 460, + 350 + ], + "spans": [ + { + "bbox": [ + 150, + 312, + 460, + 350 + ], + "score": 0.973, + "html": "
T-FPR (G-FPR)10-3 (10-3)3×10)-4(3×10-4)10-4(10-4)
Layer (f)f-2f-1f-2f-1f-2f-1
Attack Success Rate49.4092.0930.4481.876.6645.46
", + "type": "table", + "image_path": "d4093ecc5c81935abb31c8c6f5d35a1f9dcd51757521594078caa223e7ff226b.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 150, + 312, + 460, + 324.6666666666667 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 150, + 324.6666666666667, + 460, + 337.33333333333337 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 150, + 337.33333333333337, + 460, + 350.00000000000006 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 358, + 505, + 392 + ], + "lines": [ + { + "bbox": [ + 105, + 357, + 507, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 507, + 370 + ], + "score": 1.0, + "content": "Table F: Attack success rate by selecting different layer’s feature-map for interpolation on R-101-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 128, + 380 + ], + "score": 0.84, + "content": "S ( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 369, + 133, + 381 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 133, + 369, + 149, + 381 + ], + "score": 0.77, + "content": "\\mathbf { f } _ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 369, + 434, + 381 + ], + "score": 1.0, + "content": "indicates the feature-map after the last down-sampling operation and", + "type": "text" + }, + { + "bbox": [ + 434, + 369, + 450, + 380 + ], + "score": 0.86, + "content": "\\mathbf { f } _ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "indicates the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 378, + 198, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 177, + 394 + ], + "score": 1.0, + "content": "feature-map after", + "type": "text" + }, + { + "bbox": [ + 178, + 380, + 193, + 391 + ], + "score": 0.79, + "content": "\\mathbf { f } _ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 378, + 198, + 394 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 357, + 507, + 394 + ] + }, + { + "type": "image", + "bbox": [ + 107, + 407, + 498, + 541 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 407, + 498, + 541 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 407, + 498, + 541 + ], + "spans": [ + { + "bbox": [ + 107, + 407, + 498, + 541 + ], + "score": 0.968, + "type": "image", + "image_path": "d2f867cde088d9d2eedec96d6e11f67a1803de628ed29a100713bd1b68d74171.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 107, + 407, + 498, + 451.6666666666667 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 107, + 451.6666666666667, + 498, + 496.33333333333337 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 107, + 496.33333333333337, + 498, + 541.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 125, + 559, + 485, + 571 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 558, + 486, + 572 + ], + "spans": [ + { + "bbox": [ + 124, + 558, + 486, + 572 + ], + "score": 1.0, + "content": "Figure I: Qualitative comparisons between attribute-space and feature-space interpolation.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + } + ], + "index": 24.0 + }, + { + "type": "title", + "bbox": [ + 106, + 581, + 327, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 580, + 329, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 329, + 595 + ], + "score": 1.0, + "content": "E SEMANTIC ATTACK TRANSFERABILITY", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 605, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 408, + 617 + ], + "score": 1.0, + "content": "In Table G, we present the quantitative results of the transferability with", + "type": "text" + }, + { + "bbox": [ + 409, + 604, + 471, + 615 + ], + "score": 0.9, + "content": "\\mathrm { { G } \\mathrm { { - } \\mathrm { { F P R } = 1 0 ^ { - 4 } } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 604, + 505, + 617 + ], + "score": 1.0, + "content": ", T-FPR", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 614, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 137, + 627 + ], + "score": 0.9, + "content": "= 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 614, + 505, + 629 + ], + "score": 1.0, + "content": ". We observe that with the strict criterion (Lower T-FPR) of the verification model, the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 627, + 321, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 321, + 639 + ], + "score": 1.0, + "content": "transferability becomes lower cross different models.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 604, + 505, + 639 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 642, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 506, + 657 + ], + "score": 1.0, + "content": "To explore whether the improvement of transferability in SemanticAdv is introduced by semantic", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 476, + 667 + ], + "score": 1.0, + "content": "editing rather than optimized semantic perturbation in feature space. We add a StrawMan", + "type": "text" + }, + { + "bbox": [ + 477, + 655, + 506, + 666 + ], + "score": 0.63, + "content": "\\mathrm { [ C W + }", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 664, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 128, + 677 + ], + "score": 0.8, + "content": "\\mathbf { x } ^ { \\mathrm { n e w } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 664, + 505, + 679 + ], + "score": 1.0, + "content": "baseline which uses a controllable semantic-attribute-based generator to generate semantically", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 443, + 689 + ], + "score": 1.0, + "content": "different images without any notion of an adversarial attack, and then applies standard", + "type": "text" + }, + { + "bbox": [ + 444, + 677, + 456, + 689 + ], + "score": 0.88, + "content": "L _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "CW attacks", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 477, + 700 + ], + "score": 1.0, + "content": "on that generated image. The results are shown in Table H. The performance of the StrawMan", + "type": "text" + }, + { + "bbox": [ + 478, + 688, + 506, + 699 + ], + "score": 0.65, + "content": "\\mathrm { [ C W + }", + "type": "inline_equation" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 696, + 507, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 128, + 710 + ], + "score": 0.87, + "content": "\\mathbf { x } ^ { \\mathrm { n e w } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 696, + 507, + 712 + ], + "score": 1.0, + "content": "is worse than SemanticAdv. This result justifies that our SemanticAdv is able to produce novel", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "adversarial examples which cannot be simply achieved by combining attribute-conditional image", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 720, + 289, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 184, + 733 + ], + "score": 1.0, + "content": "editing model with", + "type": "text" + }, + { + "bbox": [ + 184, + 721, + 197, + 733 + ], + "score": 0.9, + "content": "L _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 720, + 289, + 733 + ], + "score": 1.0, + "content": "bounded perturbation.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 642, + 507, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 208, + 80, + 416, + 132 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 208, + 80, + 416, + 132 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 208, + 80, + 416, + 132 + ], + "spans": [ + { + "bbox": [ + 208, + 80, + 416, + 132 + ], + "score": 0.974, + "html": "
Mtest/MoptR-50-SR-101-SR-50-CR-101-C
R-50-S1.0000.0050.0000.000
R-101-S0.0001.0000.0000.000
R-50-C0.0000.0001.0000.000
R-101-C0.0000.0000.0001.000
", + "type": "table", + "image_path": "b9638e8993d797e1b0aabd5661c4971c38b6b28ec0567c1805c498bb103d705a.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 80, + 416, + 93.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 208, + 93.0, + 416, + 106.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 208, + 106.0, + 416, + 119.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 208, + 119.0, + 416, + 132.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 108, + 136, + 506, + 169 + ], + "lines": [ + { + "bbox": [ + 105, + 135, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 286, + 149 + ], + "score": 1.0, + "content": "Table G: Transferability of SemanticAdv: cell", + "type": "text" + }, + { + "bbox": [ + 287, + 136, + 307, + 148 + ], + "score": 0.91, + "content": "( i , j )", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 135, + 505, + 149 + ], + "score": 1.0, + "content": "shows attack success rate of adversarial examples", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 145, + 507, + 160 + ], + "spans": [ + { + "bbox": [ + 104, + 145, + 180, + 160 + ], + "score": 1.0, + "content": "generated against", + "type": "text" + }, + { + "bbox": [ + 181, + 148, + 186, + 159 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 145, + 294, + 160 + ], + "score": 1.0, + "content": "-th model and evaluate on", + "type": "text" + }, + { + "bbox": [ + 294, + 149, + 299, + 157 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 145, + 440, + 160 + ], + "score": 1.0, + "content": "-th model. Results generated with", + "type": "text" + }, + { + "bbox": [ + 440, + 147, + 502, + 158 + ], + "score": 0.71, + "content": "\\mathrm { { G } { - } \\mathrm { { F P R } } = 1 \\bar { 0 } ^ { - 4 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 145, + 507, + 160 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 157, + 168, + 169 + ], + "spans": [ + { + "bbox": [ + 104, + 157, + 133, + 169 + ], + "score": 1.0, + "content": "T-FPR", + "type": "text" + }, + { + "bbox": [ + 134, + 158, + 164, + 168 + ], + "score": 0.68, + "content": "= 1 0 ^ { \\overline { { - 4 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 157, + 168, + 169 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 140, + 181, + 275, + 243 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 140, + 181, + 275, + 243 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 140, + 181, + 275, + 243 + ], + "spans": [ + { + "bbox": [ + 140, + 181, + 275, + 243 + ], + "score": 0.965, + "html": "
MtestMoptR-101-S
R-50-S0.035 (0.108)
R-101-S1.000 (1.000)
R-50-C0.145 (0.202)
R-101-C0.085 (0.236)
", + "type": "table", + "image_path": "15be8ca66a93ac5a78392d7beb4717cbe9393676b91e1e007b467321f3e04dd8.jpg" + } + ] + } + ], + "index": 8.0, + "virtual_lines": [ + { + "bbox": [ + 140, + 181, + 275, + 212.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 140, + 212.0, + 275, + 243.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 149, + 246, + 265, + 258 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 148, + 245, + 265, + 258 + ], + "spans": [ + { + "bbox": [ + 148, + 245, + 162, + 258 + ], + "score": 1.0, + "content": "(a)", + "type": "text" + }, + { + "bbox": [ + 162, + 246, + 212, + 258 + ], + "score": 0.34, + "content": "\\mathrm { G } \\mathrm { - } \\mathrm { F P R } { = } 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 245, + 216, + 258 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 216, + 246, + 265, + 258 + ], + "score": 0.69, + "content": "\\mathrm { T } \\mathrm { - } \\mathrm { F P R } { = } 1 0 ^ { - 3 }", + "type": "inline_equation" + } + ], + "index": 11 + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "table", + "bbox": [ + 333, + 181, + 468, + 243 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 333, + 181, + 468, + 243 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 333, + 181, + 468, + 243 + ], + "spans": [ + { + "bbox": [ + 333, + 181, + 468, + 243 + ], + "score": 0.968, + "html": "
MtestMoptR-101-S
R-50-S0.615 (0.862)
R-101-S1.000(1.000)
R-50-C0.570(0.837)
R-101-C0.6955(0.888)
", + "type": "table", + "image_path": "4827389ee367da045ec825b88c18f43da5dc29f9192d061d184b57e4f16201f6.jpg" + } + ] + } + ], + "index": 9.0, + "virtual_lines": [ + { + "bbox": [ + 333, + 181, + 468, + 212.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 333, + 212.0, + 468, + 243.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 341, + 246, + 458, + 258 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 341, + 245, + 458, + 258 + ], + "spans": [ + { + "bbox": [ + 341, + 245, + 407, + 258 + ], + "score": 1.0, + "content": "(b) G-FPR=10−4,", + "type": "text" + }, + { + "bbox": [ + 408, + 246, + 458, + 258 + ], + "score": 0.49, + "content": "\\mathrm { T } { \\cdot } \\mathrm { F P R } { = } 1 0 ^ { - 3 }", + "type": "inline_equation" + } + ], + "index": 12 + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 263, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 281, + 276 + ], + "score": 1.0, + "content": "Table H: Transferability of StrawMan: cell", + "type": "text" + }, + { + "bbox": [ + 281, + 263, + 302, + 276 + ], + "score": 0.91, + "content": "( i , j )", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 263, + 505, + 276 + ], + "score": 1.0, + "content": "shows attack success rate of adversarial examples", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 181, + 287 + ], + "score": 1.0, + "content": "generated against", + "type": "text" + }, + { + "bbox": [ + 182, + 276, + 188, + 286 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 275, + 298, + 287 + ], + "score": 1.0, + "content": "-th model and evaluate on", + "type": "text" + }, + { + "bbox": [ + 299, + 276, + 303, + 285 + ], + "score": 0.68, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "-th model. Results of SemanticAdv are listed in", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 285, + 145, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 145, + 298 + ], + "score": 1.0, + "content": "brackets.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "title", + "bbox": [ + 107, + 317, + 328, + 330 + ], + "lines": [ + { + "bbox": [ + 104, + 315, + 329, + 334 + ], + "spans": [ + { + "bbox": [ + 104, + 315, + 329, + 334 + ], + "score": 1.0, + "content": "F QUERY-FREE BLACK-BOX API ATTACK", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 108, + 347, + 505, + 500 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 347, + 505, + 500 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 108, + 347, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 108, + 347, + 505, + 500 + ], + "score": 0.985, + "html": "
API nameAttacker/Evaluation MetricFace++AliYun
T-FPR= 10-3T-FPR=10-4T-FPR=10-3T-FPR=10
Original x2.040.510.500.00
Generated xnew4.210.530.500.00
CW (G-FPR=10-)StrawMan [CW + xnew ](G-FPR = 10-3)SemanticAdv (G-FPR =10-3)16.2410.7127.323.554.089.794.500.00
3.000.00
7.502.00
CW (G-FPR= 10-4)StrawMan [CW+ xnew ](G-FPR = 10-4)SemanticAdv (G-FPR =10-4)30.6121.3257.2215.829.1438.6612.504.50
7.5029.501.5017.50
Dong et al. (2018)Xie et al. (2019)CW (G-FPR<10-4)StrawMan [CW + xnew](G-FPR<10-4)SemanticAdv (G-FPR<10-4)30.7737.9521.0318.006.50
25.6421.5011.00
41.6227.6964.6324.3719.0012.00
15.3842.6910.003.00
35.5022.17
", + "type": "table", + "image_path": "633977a38163cddb5aed19b8bc3c70a9c7d81355e9906835d31b8c39eb63b23b.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 108, + 347, + 505, + 398.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 108, + 398.0, + 505, + 449.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 108, + 449.0, + 505, + 500.0 + ], + "spans": [], + "index": 19 + } + ] + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 506, + 502, + 528 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 504, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 504, + 518 + ], + "score": 1.0, + "content": "Table I: Quantitative analysis on query-free black-box attack. We use ResNet-101 optimized with", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 517, + 377, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 359, + 528 + ], + "score": 1.0, + "content": "softmax loss for evaluation and report the attack success rate", + "type": "text" + }, + { + "bbox": [ + 360, + 517, + 374, + 528 + ], + "score": 0.68, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 517, + 377, + 528 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 530, + 505, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "score": 1.0, + "content": "In Table I, we present the results of SemanticAdv performing query-free black-box attack on two", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "score": 1.0, + "content": "online face verification platforms. SemanticAdv outperforms CW and StawnMan in both APIs", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 552, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 505, + 564 + ], + "score": 1.0, + "content": "under all FPR thresholds. In addition, we achieve higher attack success rate on APIs using samples", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 104, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "generated using lower G-FPR compared to samples generated using higher G-FPR with the same", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 173, + 587 + ], + "score": 1.0, + "content": "T-FPR. Original", + "type": "text" + }, + { + "bbox": [ + 173, + 576, + 181, + 584 + ], + "score": 0.32, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 573, + 240, + 587 + ], + "score": 1.0, + "content": "and generated", + "type": "text" + }, + { + "bbox": [ + 241, + 574, + 260, + 584 + ], + "score": 0.84, + "content": "\\mathbf { x } ^ { \\mathrm { n e w } }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 573, + 506, + 587 + ], + "score": 1.0, + "content": "are regarded as reference point of the performance of online", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 584, + 217, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 217, + 598 + ], + "score": 1.0, + "content": "face verification platforms.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 601, + 505, + 625 + ], + "lines": [ + { + "bbox": [ + 106, + 601, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 505, + 614 + ], + "score": 1.0, + "content": "Figure J illustrates our SemanticAdv on attacking Microsoft Azure face verification system, which", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 612, + 323, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 323, + 625 + ], + "score": 1.0, + "content": "further demonstrate the effectiveness of our approach.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 309, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 208, + 80, + 416, + 132 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 208, + 80, + 416, + 132 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 208, + 80, + 416, + 132 + ], + "spans": [ + { + "bbox": [ + 208, + 80, + 416, + 132 + ], + "score": 0.974, + "html": "
Mtest/MoptR-50-SR-101-SR-50-CR-101-C
R-50-S1.0000.0050.0000.000
R-101-S0.0001.0000.0000.000
R-50-C0.0000.0001.0000.000
R-101-C0.0000.0000.0001.000
", + "type": "table", + "image_path": "b9638e8993d797e1b0aabd5661c4971c38b6b28ec0567c1805c498bb103d705a.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 80, + 416, + 93.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 208, + 93.0, + 416, + 106.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 208, + 106.0, + 416, + 119.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 208, + 119.0, + 416, + 132.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 108, + 136, + 506, + 169 + ], + "lines": [ + { + "bbox": [ + 105, + 135, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 286, + 149 + ], + "score": 1.0, + "content": "Table G: Transferability of SemanticAdv: cell", + "type": "text" + }, + { + "bbox": [ + 287, + 136, + 307, + 148 + ], + "score": 0.91, + "content": "( i , j )", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 135, + 505, + 149 + ], + "score": 1.0, + "content": "shows attack success rate of adversarial examples", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 145, + 507, + 160 + ], + "spans": [ + { + "bbox": [ + 104, + 145, + 180, + 160 + ], + "score": 1.0, + "content": "generated against", + "type": "text" + }, + { + "bbox": [ + 181, + 148, + 186, + 159 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 145, + 294, + 160 + ], + "score": 1.0, + "content": "-th model and evaluate on", + "type": "text" + }, + { + "bbox": [ + 294, + 149, + 299, + 157 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 145, + 440, + 160 + ], + "score": 1.0, + "content": "-th model. Results generated with", + "type": "text" + }, + { + "bbox": [ + 440, + 147, + 502, + 158 + ], + "score": 0.71, + "content": "\\mathrm { { G } { - } \\mathrm { { F P R } } = 1 \\bar { 0 } ^ { - 4 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 145, + 507, + 160 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 157, + 168, + 169 + ], + "spans": [ + { + "bbox": [ + 104, + 157, + 133, + 169 + ], + "score": 1.0, + "content": "T-FPR", + "type": "text" + }, + { + "bbox": [ + 134, + 158, + 164, + 168 + ], + "score": 0.68, + "content": "= 1 0 ^ { \\overline { { - 4 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 157, + 168, + 169 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 104, + 135, + 507, + 169 + ] + }, + { + "type": "table", + "bbox": [ + 140, + 181, + 275, + 243 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 140, + 181, + 275, + 243 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 140, + 181, + 275, + 243 + ], + "spans": [ + { + "bbox": [ + 140, + 181, + 275, + 243 + ], + "score": 0.965, + "html": "
MtestMoptR-101-S
R-50-S0.035 (0.108)
R-101-S1.000 (1.000)
R-50-C0.145 (0.202)
R-101-C0.085 (0.236)
", + "type": "table", + "image_path": "15be8ca66a93ac5a78392d7beb4717cbe9393676b91e1e007b467321f3e04dd8.jpg" + } + ] + } + ], + "index": 8.0, + "virtual_lines": [ + { + "bbox": [ + 140, + 181, + 275, + 212.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 140, + 212.0, + 275, + 243.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 149, + 246, + 265, + 258 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 148, + 245, + 265, + 258 + ], + "spans": [ + { + "bbox": [ + 148, + 245, + 162, + 258 + ], + "score": 1.0, + "content": "(a)", + "type": "text" + }, + { + "bbox": [ + 162, + 246, + 212, + 258 + ], + "score": 0.34, + "content": "\\mathrm { G } \\mathrm { - } \\mathrm { F P R } { = } 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 245, + 216, + 258 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 216, + 246, + 265, + 258 + ], + "score": 0.69, + "content": "\\mathrm { T } \\mathrm { - } \\mathrm { F P R } { = } 1 0 ^ { - 3 }", + "type": "inline_equation" + } + ], + "index": 11 + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "table", + "bbox": [ + 333, + 181, + 468, + 243 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 333, + 181, + 468, + 243 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 333, + 181, + 468, + 243 + ], + "spans": [ + { + "bbox": [ + 333, + 181, + 468, + 243 + ], + "score": 0.968, + "html": "
MtestMoptR-101-S
R-50-S0.615 (0.862)
R-101-S1.000(1.000)
R-50-C0.570(0.837)
R-101-C0.6955(0.888)
", + "type": "table", + "image_path": "4827389ee367da045ec825b88c18f43da5dc29f9192d061d184b57e4f16201f6.jpg" + } + ] + } + ], + "index": 9.0, + "virtual_lines": [ + { + "bbox": [ + 333, + 181, + 468, + 212.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 333, + 212.0, + 468, + 243.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 341, + 246, + 458, + 258 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 341, + 245, + 458, + 258 + ], + "spans": [ + { + "bbox": [ + 341, + 245, + 407, + 258 + ], + "score": 1.0, + "content": "(b) G-FPR=10−4,", + "type": "text" + }, + { + "bbox": [ + 408, + 246, + 458, + 258 + ], + "score": 0.49, + "content": "\\mathrm { T } { \\cdot } \\mathrm { F P R } { = } 1 0 ^ { - 3 }", + "type": "inline_equation" + } + ], + "index": 12 + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 263, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 281, + 276 + ], + "score": 1.0, + "content": "Table H: Transferability of StrawMan: cell", + "type": "text" + }, + { + "bbox": [ + 281, + 263, + 302, + 276 + ], + "score": 0.91, + "content": "( i , j )", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 263, + 505, + 276 + ], + "score": 1.0, + "content": "shows attack success rate of adversarial examples", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 181, + 287 + ], + "score": 1.0, + "content": "generated against", + "type": "text" + }, + { + "bbox": [ + 182, + 276, + 188, + 286 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 275, + 298, + 287 + ], + "score": 1.0, + "content": "-th model and evaluate on", + "type": "text" + }, + { + "bbox": [ + 299, + 276, + 303, + 285 + ], + "score": 0.68, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "-th model. Results of SemanticAdv are listed in", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 285, + 145, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 145, + 298 + ], + "score": 1.0, + "content": "brackets.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 263, + 505, + 298 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 317, + 328, + 330 + ], + "lines": [ + { + "bbox": [ + 104, + 315, + 329, + 334 + ], + "spans": [ + { + "bbox": [ + 104, + 315, + 329, + 334 + ], + "score": 1.0, + "content": "F QUERY-FREE BLACK-BOX API ATTACK", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 108, + 347, + 505, + 500 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 347, + 505, + 500 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 108, + 347, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 108, + 347, + 505, + 500 + ], + "score": 0.985, + "html": "
API nameAttacker/Evaluation MetricFace++AliYun
T-FPR= 10-3T-FPR=10-4T-FPR=10-3T-FPR=10
Original x2.040.510.500.00
Generated xnew4.210.530.500.00
CW (G-FPR=10-)StrawMan [CW + xnew ](G-FPR = 10-3)SemanticAdv (G-FPR =10-3)16.2410.7127.323.554.089.794.500.00
3.000.00
7.502.00
CW (G-FPR= 10-4)StrawMan [CW+ xnew ](G-FPR = 10-4)SemanticAdv (G-FPR =10-4)30.6121.3257.2215.829.1438.6612.504.50
7.5029.501.5017.50
Dong et al. (2018)Xie et al. (2019)CW (G-FPR<10-4)StrawMan [CW + xnew](G-FPR<10-4)SemanticAdv (G-FPR<10-4)30.7737.9521.0318.006.50
25.6421.5011.00
41.6227.6964.6324.3719.0012.00
15.3842.6910.003.00
35.5022.17
", + "type": "table", + "image_path": "633977a38163cddb5aed19b8bc3c70a9c7d81355e9906835d31b8c39eb63b23b.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 108, + 347, + 505, + 398.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 108, + 398.0, + 505, + 449.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 108, + 449.0, + 505, + 500.0 + ], + "spans": [], + "index": 19 + } + ] + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 506, + 502, + 528 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 504, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 504, + 518 + ], + "score": 1.0, + "content": "Table I: Quantitative analysis on query-free black-box attack. We use ResNet-101 optimized with", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 517, + 377, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 359, + 528 + ], + "score": 1.0, + "content": "softmax loss for evaluation and report the attack success rate", + "type": "text" + }, + { + "bbox": [ + 360, + 517, + 374, + 528 + ], + "score": 0.68, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 517, + 377, + 528 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 505, + 504, + 528 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 530, + 505, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "score": 1.0, + "content": "In Table I, we present the results of SemanticAdv performing query-free black-box attack on two", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "score": 1.0, + "content": "online face verification platforms. SemanticAdv outperforms CW and StawnMan in both APIs", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 552, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 505, + 564 + ], + "score": 1.0, + "content": "under all FPR thresholds. In addition, we achieve higher attack success rate on APIs using samples", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 104, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "generated using lower G-FPR compared to samples generated using higher G-FPR with the same", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 173, + 587 + ], + "score": 1.0, + "content": "T-FPR. Original", + "type": "text" + }, + { + "bbox": [ + 173, + 576, + 181, + 584 + ], + "score": 0.32, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 573, + 240, + 587 + ], + "score": 1.0, + "content": "and generated", + "type": "text" + }, + { + "bbox": [ + 241, + 574, + 260, + 584 + ], + "score": 0.84, + "content": "\\mathbf { x } ^ { \\mathrm { n e w } }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 573, + 506, + 587 + ], + "score": 1.0, + "content": "are regarded as reference point of the performance of online", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 584, + 217, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 217, + 598 + ], + "score": 1.0, + "content": "face verification platforms.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5, + "bbox_fs": [ + 104, + 530, + 506, + 598 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 601, + 505, + 625 + ], + "lines": [ + { + "bbox": [ + 106, + 601, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 505, + 614 + ], + "score": 1.0, + "content": "Figure J illustrates our SemanticAdv on attacking Microsoft Azure face verification system, which", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 612, + 323, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 323, + 625 + ], + "score": 1.0, + "content": "further demonstrate the effectiveness of our approach.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 601, + 505, + 625 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 272, + 503, + 513 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 272, + 503, + 513 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 272, + 503, + 513 + ], + "spans": [ + { + "bbox": [ + 110, + 272, + 503, + 513 + ], + "score": 0.976, + "type": "image", + "image_path": "e030ab5d5338f1ba2bcc57782a19fec5f232a173459480ed89b123b6102bfa5c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 272, + 503, + 352.3333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 352.3333333333333, + 503, + 432.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 432.66666666666663, + 503, + 513.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 527, + 505, + 551 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "score": 1.0, + "content": "Figure J: Illustration of our SemanticAdv in the real world face verification platform (editing on pale", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 538, + 488, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 488, + 551 + ], + "score": 1.0, + "content": "skin). Note that the confidence denotes the likelihood that two faces belong to the same person.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 309, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "19", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 272, + 503, + 513 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 272, + 503, + 513 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 272, + 503, + 513 + ], + "spans": [ + { + "bbox": [ + 110, + 272, + 503, + 513 + ], + "score": 0.976, + "type": "image", + "image_path": "e030ab5d5338f1ba2bcc57782a19fec5f232a173459480ed89b123b6102bfa5c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 272, + 503, + 352.3333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 352.3333333333333, + 503, + 432.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 432.66666666666663, + 503, + 513.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 527, + 505, + 551 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "score": 1.0, + "content": "Figure J: Illustration of our SemanticAdv in the real world face verification platform (editing on pale", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 538, + 488, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 488, + 551 + ], + "score": 1.0, + "content": "skin). Note that the confidence denotes the likelihood that two faces belong to the same person.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 252, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 253, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 253, + 96 + ], + "score": 1.0, + "content": "G MORE VISUALIZATIONS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 119, + 113, + 492, + 685 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 113, + 492, + 685 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 119, + 113, + 492, + 685 + ], + "spans": [ + { + "bbox": [ + 119, + 113, + 492, + 685 + ], + "score": 0.97, + "type": "image", + "image_path": "3a7d8c713d580a52035bd642fd776679a6153075581bff76f96b228a53b76fa7.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 119, + 113, + 492, + 303.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 303.66666666666663, + 492, + 494.33333333333326 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 119, + 494.33333333333326, + 492, + 684.9999999999999 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 135, + 685, + 475, + 697 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 134, + 684, + 476, + 698 + ], + "spans": [ + { + "bbox": [ + 134, + 684, + 407, + 698 + ], + "score": 1.0, + "content": "Figure K: Qualitative analysis on single-attribute adversarial attack", + "type": "text" + }, + { + "bbox": [ + 407, + 685, + 469, + 696 + ], + "score": 0.68, + "content": "\\mathrm { ( G { - } F P R = 1 0 ^ { - 3 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 684, + 476, + 698 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + } + ], + "index": 3.0 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 252, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 253, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 253, + 96 + ], + "score": 1.0, + "content": "G MORE VISUALIZATIONS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 119, + 113, + 492, + 685 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 113, + 492, + 685 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 119, + 113, + 492, + 685 + ], + "spans": [ + { + "bbox": [ + 119, + 113, + 492, + 685 + ], + "score": 0.97, + "type": "image", + "image_path": "3a7d8c713d580a52035bd642fd776679a6153075581bff76f96b228a53b76fa7.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 119, + 113, + 492, + 303.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 303.66666666666663, + 492, + 494.33333333333326 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 119, + 494.33333333333326, + 492, + 684.9999999999999 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 135, + 685, + 475, + 697 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 134, + 684, + 476, + 698 + ], + "spans": [ + { + "bbox": [ + 134, + 684, + 407, + 698 + ], + "score": 1.0, + "content": "Figure K: Qualitative analysis on single-attribute adversarial attack", + "type": "text" + }, + { + "bbox": [ + 407, + 685, + 469, + 696 + ], + "score": 0.68, + "content": "\\mathrm { ( G { - } F P R = 1 0 ^ { - 3 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 684, + 476, + 698 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + } + ], + "index": 3.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 119, + 118, + 492, + 684 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 118, + 492, + 684 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 119, + 118, + 492, + 684 + ], + "spans": [ + { + "bbox": [ + 119, + 118, + 492, + 684 + ], + "score": 0.967, + "type": "image", + "image_path": "de33f8bb372e2f4abdf4f6769428fdaa8a0623516671ba763ae2fab7c70afb7c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 119, + 118, + 492, + 306.66666666666663 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 119, + 306.66666666666663, + 492, + 495.33333333333326 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 495.33333333333326, + 492, + 683.9999999999999 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 135, + 685, + 474, + 697 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 683, + 475, + 698 + ], + "spans": [ + { + "bbox": [ + 135, + 683, + 406, + 698 + ], + "score": 1.0, + "content": "Figure L: Qualitative analysis on single-attribute adversarial attack", + "type": "text" + }, + { + "bbox": [ + 406, + 684, + 469, + 696 + ], + "score": 0.53, + "content": "\\mathrm { ( G - F P R = 1 0 ^ { - 3 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 683, + 475, + 698 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ], + "page_idx": 20, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 119, + 118, + 492, + 684 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 118, + 492, + 684 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 119, + 118, + 492, + 684 + ], + "spans": [ + { + "bbox": [ + 119, + 118, + 492, + 684 + ], + "score": 0.967, + "type": "image", + "image_path": "de33f8bb372e2f4abdf4f6769428fdaa8a0623516671ba763ae2fab7c70afb7c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 119, + 118, + 492, + 306.66666666666663 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 119, + 306.66666666666663, + 492, + 495.33333333333326 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 495.33333333333326, + 492, + 683.9999999999999 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 135, + 685, + 474, + 697 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 683, + 475, + 698 + ], + "spans": [ + { + "bbox": [ + 135, + 683, + 406, + 698 + ], + "score": 1.0, + "content": "Figure L: Qualitative analysis on single-attribute adversarial attack", + "type": "text" + }, + { + "bbox": [ + 406, + 684, + 469, + 696 + ], + "score": 0.53, + "content": "\\mathrm { ( G - F P R = 1 0 ^ { - 3 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 683, + 475, + 698 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 119, + 265, + 492, + 539 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 265, + 492, + 539 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 119, + 265, + 492, + 539 + ], + "spans": [ + { + "bbox": [ + 119, + 265, + 492, + 539 + ], + "score": 0.974, + "type": "image", + "image_path": "7c8afb5499aa1945df2aa8c49f2ce5827b76cbafd081d867c0cbe11df573a7e7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 119, + 265, + 492, + 356.3333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 119, + 356.3333333333333, + 492, + 447.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 447.66666666666663, + 492, + 539.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 133, + 546, + 475, + 558 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 477, + 558 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 408, + 558 + ], + "score": 1.0, + "content": "Figure M: Qualitative analysis on single-attribute adversarial attack", + "type": "text" + }, + { + "bbox": [ + 409, + 545, + 471, + 556 + ], + "score": 0.75, + "content": "\\mathrm { \\bf G } \\mathrm { - } \\mathrm { \\bf F P R } = 1 0 ^ { - 3 } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 545, + 477, + 558 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ], + "page_idx": 21, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 309, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 119, + 265, + 492, + 539 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 265, + 492, + 539 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 119, + 265, + 492, + 539 + ], + "spans": [ + { + "bbox": [ + 119, + 265, + 492, + 539 + ], + "score": 0.974, + "type": "image", + "image_path": "7c8afb5499aa1945df2aa8c49f2ce5827b76cbafd081d867c0cbe11df573a7e7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 119, + 265, + 492, + 356.3333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 119, + 356.3333333333333, + 492, + 447.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 447.66666666666663, + 492, + 539.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 133, + 546, + 475, + 558 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 477, + 558 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 408, + 558 + ], + "score": 1.0, + "content": "Figure M: Qualitative analysis on single-attribute adversarial attack", + "type": "text" + }, + { + "bbox": [ + 409, + 545, + 471, + 556 + ], + "score": 0.75, + "content": "\\mathrm { \\bf G } \\mathrm { - } \\mathrm { \\bf F P R } = 1 0 ^ { - 3 } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 545, + 477, + 558 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 114, + 83, + 493, + 691 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 83, + 493, + 691 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 83, + 493, + 691 + ], + "spans": [ + { + "bbox": [ + 114, + 83, + 493, + 691 + ], + "score": 0.962, + "type": "image", + "image_path": "bc2242ce0987cedaa8c20dea5c4a3437513f800c661325f11f4134300127ee38.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 114, + 83, + 493, + 285.66666666666663 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 114, + 285.66666666666663, + 493, + 488.33333333333326 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 114, + 488.33333333333326, + 493, + 690.9999999999999 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 701, + 505, + 735 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 701, + 396, + 713 + ], + "score": 1.0, + "content": "Figure N: Qualitative comparisons between our proposed SemanticAdv", + "type": "text" + }, + { + "bbox": [ + 397, + 700, + 460, + 712 + ], + "score": 0.8, + "content": "\\mathrm { ( G { - } F P R = 1 0 ^ { - 3 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 701, + 506, + 713 + ], + "score": 1.0, + "content": ") and pixel-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 712, + 505, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 712, + 505, + 725 + ], + "score": 1.0, + "content": "wise adversarial examples generated by CW. Along with the adversarial examples, we also provide", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 723, + 327, + 736 + ], + "spans": [ + { + "bbox": [ + 105, + 723, + 327, + 736 + ], + "score": 1.0, + "content": "the corresponding perturbations (residual) on the right.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + } + ], + "page_idx": 22, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 114, + 83, + 493, + 691 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 83, + 493, + 691 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 83, + 493, + 691 + ], + "spans": [ + { + "bbox": [ + 114, + 83, + 493, + 691 + ], + "score": 0.962, + "type": "image", + "image_path": "bc2242ce0987cedaa8c20dea5c4a3437513f800c661325f11f4134300127ee38.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 114, + 83, + 493, + 285.66666666666663 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 114, + 285.66666666666663, + 493, + 488.33333333333326 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 114, + 488.33333333333326, + 493, + 690.9999999999999 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 701, + 505, + 735 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 701, + 396, + 713 + ], + "score": 1.0, + "content": "Figure N: Qualitative comparisons between our proposed SemanticAdv", + "type": "text" + }, + { + "bbox": [ + 397, + 700, + 460, + 712 + ], + "score": 0.8, + "content": "\\mathrm { ( G { - } F P R = 1 0 ^ { - 3 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 701, + 506, + 713 + ], + "score": 1.0, + "content": ") and pixel-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 712, + 505, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 712, + 505, + 725 + ], + "score": 1.0, + "content": "wise adversarial examples generated by CW. Along with the adversarial examples, we also provide", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 723, + 327, + 736 + ], + "spans": [ + { + "bbox": [ + 105, + 723, + 327, + 736 + ], + "score": 1.0, + "content": "the corresponding perturbations (residual) on the right.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 174, + 500, + 589 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 174, + 500, + 589 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 174, + 500, + 589 + ], + "spans": [ + { + "bbox": [ + 110, + 174, + 500, + 589 + ], + "score": 0.978, + "type": "image", + "image_path": "0b62987d5a2bc9e11b6691551e91f0da0480e64b18a63bf77468e026095b805e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 174, + 500, + 312.33333333333337 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 312.33333333333337, + 500, + 450.66666666666674 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 450.66666666666674, + 500, + 589.0000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 608, + 505, + 641 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 465, + 620 + ], + "score": 1.0, + "content": "Figure O: Qualitative analysis on single-attribute adversarial attack (SemanticAdv with", + "type": "text" + }, + { + "bbox": [ + 465, + 608, + 506, + 619 + ], + "score": 0.27, + "content": "\\mathrm { { \\bf G } { - } F P R = }", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 617, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 128, + 630 + ], + "score": 0.85, + "content": "1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 617, + 505, + 632 + ], + "score": 1.0, + "content": ") by each other. Along with the adversarial examples, we also provide the corresponding", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 629, + 253, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 253, + 643 + ], + "score": 1.0, + "content": "perturbations (residual) on the right.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + } + ], + "page_idx": 23, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 309, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 174, + 500, + 589 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 174, + 500, + 589 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 174, + 500, + 589 + ], + "spans": [ + { + "bbox": [ + 110, + 174, + 500, + 589 + ], + "score": 0.978, + "type": "image", + "image_path": "0b62987d5a2bc9e11b6691551e91f0da0480e64b18a63bf77468e026095b805e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 174, + 500, + 312.33333333333337 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 312.33333333333337, + 500, + 450.66666666666674 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 450.66666666666674, + 500, + 589.0000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 608, + 505, + 641 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 465, + 620 + ], + "score": 1.0, + "content": "Figure O: Qualitative analysis on single-attribute adversarial attack (SemanticAdv with", + "type": "text" + }, + { + "bbox": [ + 465, + 608, + 506, + 619 + ], + "score": 0.27, + "content": "\\mathrm { { \\bf G } { - } F P R = }", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 617, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 128, + 630 + ], + "score": 0.85, + "content": "1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 617, + 505, + 632 + ], + "score": 1.0, + "content": ") by each other. Along with the adversarial examples, we also provide the corresponding", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 629, + 253, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 253, + 643 + ], + "score": 1.0, + "content": "perturbations (residual) on the right.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/r1l-VeSKwS/r1l-VeSKwS_model.json b/parse/train/r1l-VeSKwS/r1l-VeSKwS_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d5c7245305f5a667f951cbf7b6b969ae659e97e2 --- /dev/null +++ b/parse/train/r1l-VeSKwS/r1l-VeSKwS_model.json @@ -0,0 +1,40322 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 549, + 1303, + 549, + 1303, + 1128, + 398, + 1128 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1208, + 1404, + 1208, + 1404, + 1574, + 298, + 1574 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1590, + 1404, + 1590, + 1404, + 1774, + 298, + 1774 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1789, + 1402, + 1789, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 0, + "poly": [ + 301, + 220, + 1407, + 220, + 1407, + 324, + 301, + 324 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 313, + 377, + 680, + 377, + 680, + 438, + 313, + 438 + ], + "score": 0.933 + }, + { + "category_id": 0, + "poly": [ + 301, + 1163, + 573, + 1163, + 573, + 1197, + 301, + 1197 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 773, + 478, + 927, + 478, + 927, + 513, + 773, + 513 + ], + "score": 0.868 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 856, + 2088, + 856, + 2112, + 841, + 2112 + ], + "score": 0.733 + }, + { + "category_id": 13, + "poly": [ + 837, + 1454, + 871, + 1454, + 871, + 1486, + 837, + 1486 + ], + "score": 0.89, + "latex": "{ \\mathcal { L } } _ { p }" + }, + { + "category_id": 13, + "poly": [ + 1266, + 672, + 1301, + 672, + 1301, + 705, + 1266, + 705 + ], + "score": 0.87, + "latex": "L _ { p }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 217.0, + 1412.0, + 217.0, + 1412.0, + 271.0, + 295.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 277.0, + 1382.0, + 277.0, + 1382.0, + 326.0, + 295.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1159.0, + 579.0, + 1159.0, + 579.0, + 1205.0, + 293.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 475.0, + 934.0, + 475.0, + 934.0, + 518.0, + 768.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2088.0, + 857.0, + 2088.0, + 857.0, + 2116.0, + 840.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 549.0, + 1305.0, + 549.0, + 1305.0, + 585.0, + 394.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 580.0, + 1305.0, + 580.0, + 1305.0, + 615.0, + 394.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 609.0, + 1305.0, + 609.0, + 1305.0, + 644.0, + 393.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 642.0, + 1305.0, + 642.0, + 1305.0, + 675.0, + 395.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 672.0, + 1265.0, + 672.0, + 1265.0, + 708.0, + 394.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 703.0, + 1304.0, + 703.0, + 1304.0, + 736.0, + 393.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 733.0, + 1306.0, + 733.0, + 1306.0, + 766.0, + 393.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 763.0, + 1305.0, + 763.0, + 1305.0, + 798.0, + 394.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 795.0, + 1307.0, + 795.0, + 1307.0, + 828.0, + 395.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 826.0, + 1304.0, + 826.0, + 1304.0, + 855.0, + 395.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 853.0, + 1305.0, + 853.0, + 1305.0, + 889.0, + 393.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 885.0, + 1305.0, + 885.0, + 1305.0, + 918.0, + 394.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 914.0, + 1304.0, + 914.0, + 1304.0, + 948.0, + 393.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 946.0, + 1306.0, + 946.0, + 1306.0, + 979.0, + 393.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 977.0, + 1304.0, + 977.0, + 1304.0, + 1007.0, + 395.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1007.0, + 1306.0, + 1007.0, + 1306.0, + 1041.0, + 393.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1038.0, + 1304.0, + 1038.0, + 1304.0, + 1071.0, + 394.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1069.0, + 1306.0, + 1069.0, + 1306.0, + 1102.0, + 395.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1096.0, + 1137.0, + 1096.0, + 1137.0, + 1132.0, + 394.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1209.0, + 1405.0, + 1209.0, + 1405.0, + 1243.0, + 296.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1239.0, + 1407.0, + 1239.0, + 1407.0, + 1273.0, + 293.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1270.0, + 1404.0, + 1270.0, + 1404.0, + 1304.0, + 294.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1300.0, + 1405.0, + 1300.0, + 1405.0, + 1334.0, + 294.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1332.0, + 1405.0, + 1332.0, + 1405.0, + 1366.0, + 294.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1359.0, + 1406.0, + 1359.0, + 1406.0, + 1396.0, + 292.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1392.0, + 1406.0, + 1392.0, + 1406.0, + 1426.0, + 294.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1421.0, + 1405.0, + 1421.0, + 1405.0, + 1455.0, + 294.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1453.0, + 836.0, + 1453.0, + 836.0, + 1487.0, + 292.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 1453.0, + 1404.0, + 1453.0, + 1404.0, + 1487.0, + 872.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1480.0, + 1409.0, + 1480.0, + 1409.0, + 1521.0, + 292.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1515.0, + 1405.0, + 1515.0, + 1405.0, + 1549.0, + 294.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1544.0, + 977.0, + 1544.0, + 977.0, + 1578.0, + 293.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1586.0, + 1405.0, + 1586.0, + 1405.0, + 1629.0, + 292.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1620.0, + 1406.0, + 1620.0, + 1406.0, + 1655.0, + 294.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1649.0, + 1408.0, + 1649.0, + 1408.0, + 1686.0, + 292.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1679.0, + 1408.0, + 1679.0, + 1408.0, + 1716.0, + 293.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1710.0, + 1405.0, + 1710.0, + 1405.0, + 1747.0, + 293.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1740.0, + 935.0, + 1740.0, + 935.0, + 1778.0, + 293.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1791.0, + 1406.0, + 1791.0, + 1406.0, + 1824.0, + 294.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1822.0, + 1406.0, + 1822.0, + 1406.0, + 1855.0, + 295.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1852.0, + 1404.0, + 1852.0, + 1404.0, + 1885.0, + 295.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 1407.0, + 1880.0, + 1407.0, + 1919.0, + 293.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1913.0, + 1406.0, + 1913.0, + 1406.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1975.0, + 1404.0, + 1975.0, + 1404.0, + 2005.0, + 295.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2001.0, + 1406.0, + 2001.0, + 1406.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 377.0, + 560.0, + 377.0, + 560.0, + 410.0, + 315.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 407.0, + 682.0, + 407.0, + 682.0, + 440.0, + 311.0, + 440.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1589, + 1405, + 1589, + 1405, + 1895, + 298, + 1895 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1271, + 1406, + 1271, + 1406, + 1577, + 297, + 1577 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 978, + 1405, + 978, + 1405, + 1223, + 298, + 1223 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 779, + 1403, + 779, + 1403, + 963, + 298, + 963 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 295, + 222, + 1405, + 222, + 1405, + 617, + 295, + 617 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1402, + 1911, + 1402, + 2035, + 299, + 2035 + ], + "score": 0.973 + }, + { + "category_id": 4, + "poly": [ + 297, + 659, + 1406, + 659, + 1406, + 752, + 297, + 752 + ], + "score": 0.956 + }, + { + "category_id": 0, + "poly": [ + 300, + 1228, + 589, + 1228, + 589, + 1262, + 300, + 1262 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 858, + 76, + 858, + 104, + 298, + 104 + ], + "score": 0.825 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.711 + }, + { + "category_id": 2, + "poly": [ + 300, + 77, + 856, + 77, + 856, + 104, + 300, + 104 + ], + "score": 0.258 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.137 + }, + { + "category_id": 13, + "poly": [ + 727, + 1592, + 761, + 1592, + 761, + 1624, + 727, + 1624 + ], + "score": 0.87, + "latex": "L _ { p }" + }, + { + "category_id": 13, + "poly": [ + 671, + 1131, + 726, + 1131, + 726, + 1160, + 671, + 1160 + ], + "score": 0.86, + "latex": "65 \\%" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 227.0, + 522.0, + 227.0, + 522.0, + 248.0, + 350.0, + 248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 231.0, + 632.0, + 231.0, + 632.0, + 253.0, + 576.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 225.0, + 840.0, + 225.0, + 840.0, + 252.0, + 719.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 234.0, + 1010.0, + 234.0, + 1010.0, + 260.0, + 904.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1348.0, + 239.0, + 1394.0, + 239.0, + 1394.0, + 256.0, + 1348.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 247.0, + 626.0, + 247.0, + 626.0, + 265.0, + 583.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 262.0, + 1369.0, + 262.0, + 1369.0, + 282.0, + 922.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 298.0, + 884.0, + 298.0, + 884.0, + 315.0, + 821.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 319.0, + 860.0, + 319.0, + 860.0, + 325.0, + 854.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 332.0, + 356.0, + 332.0, + 356.0, + 365.0, + 297.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 343.0, + 647.0, + 343.0, + 647.0, + 377.0, + 574.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 334.0, + 1018.0, + 334.0, + 1018.0, + 369.0, + 959.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 337.0, + 1322.0, + 337.0, + 1322.0, + 375.0, + 1272.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 393.0, + 518.0, + 393.0, + 518.0, + 417.0, + 438.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 390.0, + 874.0, + 390.0, + 874.0, + 410.0, + 822.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 389.0, + 962.0, + 389.0, + 962.0, + 409.0, + 908.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 386.0, + 1118.0, + 386.0, + 1118.0, + 412.0, + 1069.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 389.0, + 1199.0, + 389.0, + 1199.0, + 409.0, + 1145.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 389.0, + 1274.0, + 389.0, + 1274.0, + 410.0, + 1218.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 409.0, + 370.0, + 409.0, + 370.0, + 444.0, + 298.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 411.0, + 518.0, + 411.0, + 518.0, + 434.0, + 384.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1275.0, + 418.0, + 1391.0, + 418.0, + 1391.0, + 443.0, + 1275.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 434.0, + 1073.0, + 434.0, + 1073.0, + 454.0, + 1010.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 484.0, + 909.0, + 484.0, + 909.0, + 621.0, + 893.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 494.0, + 1029.0, + 494.0, + 1029.0, + 515.0, + 950.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 503.0, + 873.0, + 503.0, + 873.0, + 520.0, + 822.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 505.0, + 1013.0, + 505.0, + 1013.0, + 532.0, + 966.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 525.0, + 864.0, + 525.0, + 864.0, + 537.0, + 853.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 542.0, + 962.0, + 542.0, + 962.0, + 563.0, + 911.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 539.0, + 1115.0, + 539.0, + 1115.0, + 563.0, + 1071.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 543.0, + 1195.0, + 543.0, + 1195.0, + 560.0, + 1146.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 542.0, + 1272.0, + 542.0, + 1272.0, + 563.0, + 1219.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 557.0, + 694.0, + 557.0, + 694.0, + 579.0, + 568.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 581.0, + 550.0, + 581.0, + 550.0, + 608.0, + 345.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 569.0, + 630.0, + 569.0, + 630.0, + 594.0, + 583.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 569.0, + 693.0, + 569.0, + 693.0, + 595.0, + 647.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 573.0, + 1389.0, + 573.0, + 1389.0, + 595.0, + 1032.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 597.0, + 493.0, + 597.0, + 493.0, + 618.0, + 403.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 415.0, + 1287.0, + 415.0, + 1287.0, + 443.5, + 940.0, + 443.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 426.0, + 424.0, + 426.0, + 424.0, + 445.0, + 394.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 658.0, + 1406.0, + 658.0, + 1406.0, + 695.0, + 294.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 691.0, + 1404.0, + 691.0, + 1404.0, + 725.0, + 294.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 720.0, + 600.0, + 720.0, + 600.0, + 759.0, + 294.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1223.0, + 593.0, + 1223.0, + 593.0, + 1269.0, + 291.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 70.0, + 859.0, + 70.0, + 859.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 107.0, + 297.0, + 107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1590.0, + 726.0, + 1590.0, + 726.0, + 1625.0, + 295.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 1590.0, + 1405.0, + 1590.0, + 1405.0, + 1625.0, + 762.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1620.0, + 1407.0, + 1620.0, + 1407.0, + 1656.0, + 295.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1648.0, + 1406.0, + 1648.0, + 1406.0, + 1689.0, + 291.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1681.0, + 1408.0, + 1681.0, + 1408.0, + 1717.0, + 295.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1712.0, + 1406.0, + 1712.0, + 1406.0, + 1746.0, + 292.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1741.0, + 1410.0, + 1741.0, + 1410.0, + 1780.0, + 293.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1774.0, + 1408.0, + 1774.0, + 1408.0, + 1809.0, + 295.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1805.0, + 1405.0, + 1805.0, + 1405.0, + 1837.0, + 295.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1831.0, + 1406.0, + 1831.0, + 1406.0, + 1873.0, + 291.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1866.0, + 486.0, + 1866.0, + 486.0, + 1898.0, + 295.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1268.0, + 1406.0, + 1268.0, + 1406.0, + 1309.0, + 292.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1302.0, + 1404.0, + 1302.0, + 1404.0, + 1338.0, + 294.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1328.0, + 1406.0, + 1328.0, + 1406.0, + 1371.0, + 291.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1363.0, + 1406.0, + 1363.0, + 1406.0, + 1399.0, + 294.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1391.0, + 1411.0, + 1391.0, + 1411.0, + 1432.0, + 291.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1423.0, + 1407.0, + 1423.0, + 1407.0, + 1461.0, + 292.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1451.0, + 1408.0, + 1451.0, + 1408.0, + 1491.0, + 292.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1478.0, + 1409.0, + 1478.0, + 1409.0, + 1523.0, + 291.0, + 1523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1512.0, + 1407.0, + 1512.0, + 1407.0, + 1553.0, + 291.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1545.0, + 441.0, + 1545.0, + 441.0, + 1578.0, + 294.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 977.0, + 1407.0, + 977.0, + 1407.0, + 1013.0, + 293.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1007.0, + 1408.0, + 1007.0, + 1408.0, + 1045.0, + 292.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1038.0, + 1405.0, + 1038.0, + 1405.0, + 1075.0, + 293.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1069.0, + 1405.0, + 1069.0, + 1405.0, + 1105.0, + 293.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1101.0, + 1406.0, + 1101.0, + 1406.0, + 1134.0, + 292.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1129.0, + 670.0, + 1129.0, + 670.0, + 1165.0, + 292.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 1129.0, + 1406.0, + 1129.0, + 1406.0, + 1165.0, + 727.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1163.0, + 1405.0, + 1163.0, + 1405.0, + 1196.0, + 295.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1192.0, + 405.0, + 1192.0, + 405.0, + 1223.0, + 293.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 778.0, + 1405.0, + 778.0, + 1405.0, + 814.0, + 294.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 808.0, + 1405.0, + 808.0, + 1405.0, + 844.0, + 293.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 842.0, + 1404.0, + 842.0, + 1404.0, + 874.0, + 296.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 872.0, + 1403.0, + 872.0, + 1403.0, + 904.0, + 296.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 901.0, + 1405.0, + 901.0, + 1405.0, + 937.0, + 294.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 933.0, + 1395.0, + 933.0, + 1395.0, + 965.0, + 296.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1911.0, + 1406.0, + 1911.0, + 1406.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1406.0, + 2003.0, + 1406.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 884, + 1405, + 884, + 1405, + 1372, + 297, + 1372 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1398, + 1405, + 1398, + 1405, + 1675, + 298, + 1675 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 502, + 1405, + 502, + 1405, + 720, + 297, + 720 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 733, + 1405, + 733, + 1405, + 857, + 298, + 857 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1880, + 1404, + 1880, + 1404, + 2035, + 298, + 2035 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 1404, + 228, + 1404, + 384, + 298, + 384 + ], + "score": 0.975 + }, + { + "category_id": 8, + "poly": [ + 469, + 1748, + 1227, + 1748, + 1227, + 1826, + 469, + 1826 + ], + "score": 0.932 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.91 + }, + { + "category_id": 0, + "poly": [ + 300, + 402, + 889, + 402, + 889, + 436, + 300, + 436 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 300, + 454, + 633, + 454, + 633, + 486, + 300, + 486 + ], + "score": 0.895 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1792, + 1400, + 1792, + 1400, + 1821, + 1366, + 1821 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.588 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.569 + }, + { + "category_id": 13, + "poly": [ + 946, + 504, + 1107, + 504, + 1107, + 538, + 946, + 538 + ], + "score": 0.94, + "latex": "\\mathcal { D } = \\{ ( { \\bf x } , { \\bf y } ) \\}" + }, + { + "category_id": 13, + "poly": [ + 602, + 1005, + 728, + 1005, + 728, + 1039, + 602, + 1039 + ], + "score": 0.93, + "latex": "c _ { i } \\in \\{ 0 , 1 \\}" + }, + { + "category_id": 13, + "poly": [ + 668, + 533, + 774, + 533, + 774, + 566, + 668, + 566 + ], + "score": 0.93, + "latex": "\\mathbf { y } \\in \\mathbb { R } ^ { D _ { L } }" + }, + { + "category_id": 13, + "poly": [ + 987, + 1461, + 1101, + 1461, + 1101, + 1494, + 987, + 1494 + ], + "score": 0.93, + "latex": "\\alpha \\in ( 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 1255, + 1158, + 1398, + 1158, + 1398, + 1191, + 1255, + 1191 + ], + "score": 0.93, + "latex": "\\mathbf { x } = \\mathcal { G } ( \\mathbf { x } , \\mathbf { c } )" + }, + { + "category_id": 13, + "poly": [ + 1182, + 593, + 1398, + 593, + 1398, + 628, + 1182, + 628 + ], + "score": 0.93, + "latex": "\\hat { \\mathbf { y } } = \\mathcal { M } ( \\mathbf { x } ) \\in \\mathbb { R } ^ { D _ { L } }" + }, + { + "category_id": 13, + "poly": [ + 915, + 1127, + 1122, + 1127, + 1122, + 1161, + 915, + 1161 + ], + "score": 0.93, + "latex": "\\mathbf { x } ^ { \\mathrm { n e w } } = \\mathcal { G } ( \\bar { \\mathbf { x } , \\mathbf { c } ^ { \\mathrm { n e w } } } )" + }, + { + "category_id": 13, + "poly": [ + 546, + 656, + 622, + 656, + 622, + 689, + 546, + 689 + ], + "score": 0.92, + "latex": "\\{ { \\bf x } ^ { \\mathrm { a d v } } \\}" + }, + { + "category_id": 13, + "poly": [ + 660, + 626, + 772, + 626, + 772, + 659, + 660, + 659 + ], + "score": 0.92, + "latex": "\\left( \\mathbf { x } ^ { \\mathrm { { t g t } } } , \\mathbf { y } ^ { \\mathrm { { t g t } } } \\right)" + }, + { + "category_id": 13, + "poly": [ + 437, + 532, + 617, + 532, + 617, + 563, + 437, + 563 + ], + "score": 0.92, + "latex": "\\mathbf { x } \\in \\mathbb { R } ^ { H \\times W \\times D _ { I } }" + }, + { + "category_id": 13, + "poly": [ + 596, + 687, + 774, + 687, + 774, + 720, + 596, + 720 + ], + "score": 0.92, + "latex": "\\dot { \\mathcal { M } } ( \\mathbf { x } ^ { \\mathrm { { a d v } } } ) = \\mathbf { y } ^ { \\mathrm { { t g t } } }" + }, + { + "category_id": 13, + "poly": [ + 826, + 626, + 926, + 626, + 926, + 658, + 826, + 658 + ], + "score": 0.92, + "latex": "\\mathbf { y } \\neq \\mathbf { y } ^ { \\mathrm { t g t } }" + }, + { + "category_id": 13, + "poly": [ + 478, + 1490, + 604, + 1490, + 604, + 1520, + 478, + 1520 + ], + "score": 0.92, + "latex": "\\mathbf { c } ^ { \\ast } \\in \\mathbb { R } ^ { D _ { C } }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1126, + 435, + 1126, + 435, + 1156, + 297, + 1156 + ], + "score": 0.92, + "latex": "\\mathbf { c } ^ { \\mathrm { { n e w } } } \\in \\mathbb { R } ^ { D _ { C } }" + }, + { + "category_id": 13, + "poly": [ + 816, + 1971, + 1182, + 1971, + 1182, + 2005, + 816, + 2005 + ], + "score": 0.91, + "latex": "\\mathbf { f } ^ { * } = { \\mathcal { G } } _ { \\mathrm { e n c } } ( \\mathbf { x } , \\mathbf { c } ) \\in \\mathbb { R } ^ { H _ { F } \\times W _ { F } \\times C _ { F } }" + }, + { + "category_id": 13, + "poly": [ + 566, + 1912, + 743, + 1912, + 743, + 1943, + 566, + 1943 + ], + "score": 0.91, + "latex": "\\mathcal { G } = \\mathcal { G } _ { \\mathrm { d e c } } \\circ \\mathcal { G } _ { \\mathrm { e n c } }" + }, + { + "category_id": 13, + "poly": [ + 774, + 943, + 890, + 943, + 890, + 974, + 774, + 974 + ], + "score": 0.91, + "latex": "\\mathbf { c } \\in \\mathbb { R } ^ { D _ { c } }" + }, + { + "category_id": 13, + "poly": [ + 791, + 1944, + 838, + 1944, + 838, + 1973, + 791, + 1973 + ], + "score": 0.9, + "latex": "\\mathcal { G } _ { \\mathrm { d e c } }" + }, + { + "category_id": 13, + "poly": [ + 828, + 1913, + 876, + 1913, + 876, + 1942, + 828, + 1942 + ], + "score": 0.89, + "latex": "\\mathcal { G } _ { \\mathrm { e n c } }" + }, + { + "category_id": 13, + "poly": [ + 507, + 1553, + 541, + 1553, + 541, + 1585, + 507, + 1585 + ], + "score": 0.89, + "latex": "\\tau _ { \\mathscr { G } }" + }, + { + "category_id": 13, + "poly": [ + 1250, + 976, + 1295, + 976, + 1295, + 1007, + 1250, + 1007 + ], + "score": 0.89, + "latex": "D _ { C }" + }, + { + "category_id": 13, + "poly": [ + 538, + 566, + 581, + 566, + 581, + 596, + 538, + 596 + ], + "score": 0.89, + "latex": "D _ { L }" + }, + { + "category_id": 13, + "poly": [ + 373, + 1219, + 427, + 1219, + 427, + 1248, + 373, + 1248 + ], + "score": 0.87, + "latex": "\\mathbf { x } ^ { \\mathrm { n e w } }" + }, + { + "category_id": 13, + "poly": [ + 1242, + 1311, + 1296, + 1311, + 1296, + 1340, + 1242, + 1340 + ], + "score": 0.87, + "latex": "\\mathbf { x } ^ { \\mathrm { n e w } }" + }, + { + "category_id": 13, + "poly": [ + 475, + 1462, + 524, + 1462, + 524, + 1489, + 475, + 1489 + ], + "score": 0.86, + "latex": "\\mathbf { c } ^ { \\mathrm { n e w } }" + }, + { + "category_id": 13, + "poly": [ + 1006, + 2005, + 1046, + 2005, + 1046, + 2034, + 1006, + 2034 + ], + "score": 0.86, + "latex": "C _ { F }" + }, + { + "category_id": 13, + "poly": [ + 847, + 2004, + 890, + 2004, + 890, + 2034, + 847, + 2034 + ], + "score": 0.85, + "latex": "H _ { F }" + }, + { + "category_id": 13, + "poly": [ + 620, + 1160, + 728, + 1160, + 728, + 1189, + 620, + 1189 + ], + "score": 0.85, + "latex": "( \\mathbf { c } = \\mathbf { c } ^ { \\mathrm { n e w } }" + }, + { + "category_id": 13, + "poly": [ + 1070, + 1494, + 1101, + 1494, + 1101, + 1520, + 1070, + 1520 + ], + "score": 0.84, + "latex": "\\mathbf { c } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 846, + 1523, + 867, + 1523, + 867, + 1551, + 846, + 1551 + ], + "score": 0.83, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 904, + 2004, + 951, + 2004, + 951, + 2034, + 904, + 2034 + ], + "score": 0.82, + "latex": "W _ { F }" + }, + { + "category_id": 13, + "poly": [ + 627, + 1312, + 650, + 1312, + 650, + 1340, + 627, + 1340 + ], + "score": 0.82, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 994, + 1554, + 1014, + 1554, + 1014, + 1582, + 994, + 1582 + ], + "score": 0.82, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 1377, + 766, + 1398, + 766, + 1398, + 794, + 1377, + 794 + ], + "score": 0.81, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 444, + 566, + 482, + 566, + 482, + 596, + 444, + 596 + ], + "score": 0.8, + "latex": "D _ { I }" + }, + { + "category_id": 13, + "poly": [ + 343, + 506, + 379, + 506, + 379, + 533, + 343, + 533 + ], + "score": 0.78, + "latex": "\\mathcal { M }" + }, + { + "category_id": 13, + "poly": [ + 923, + 597, + 958, + 597, + 958, + 623, + 923, + 623 + ], + "score": 0.76, + "latex": "\\mathcal { M }" + }, + { + "category_id": 13, + "poly": [ + 1035, + 1008, + 1047, + 1008, + 1047, + 1033, + 1035, + 1033 + ], + "score": 0.75, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 547, + 1789, + 870, + 1789, + 870, + 1824, + 547, + 1824 + ], + "score": 0.69, + "latex": "\\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) = \\mathcal { G } ( \\mathbf { x } , \\mathbf { c } ^ { * } )" + }, + { + "category_id": 14, + "poly": [ + 470, + 1747, + 1230, + 1747, + 1230, + 1828, + 470, + 1828 + ], + "score": 0.69, + "latex": "\\begin{array} { r l } & { \\mathbf { x } ^ { \\mathrm { a d v } } = \\arg \\operatorname* { m i n } _ { \\alpha } \\mathcal { L } \\big ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) \\big ) } \\\\ & { \\mathrm { w h e r e ~ } \\mathcal { T } _ { \\mathcal { G } } \\big ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } \\big ) = \\mathcal { G } \\big ( \\mathbf { x } , \\mathbf { c } ^ { \\ast } \\big ) \\mathrm { ~ a n d ~ } \\mathbf { c } ^ { \\ast } = \\alpha \\cdot \\mathbf { c } + ( 1 - \\alpha ) \\cdot \\mathbf { c } ^ { \\mathrm { n e w } } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 470, + 1749, + 885, + 1749, + 885, + 1787, + 470, + 1787 + ], + "score": 0.59, + "latex": "\\begin{array} { r } { { \\mathbf { x } } ^ { \\mathrm { a d v } } = \\operatorname * { a r g m i n } _ { \\alpha } \\mathcal { L } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; { \\mathbf { x } } , { \\mathbf { c } } , { \\mathbf { c } } ^ { \\mathrm { n e w } } ) ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 901, + 1316, + 924, + 1316, + 924, + 1339, + 901, + 1339 + ], + "score": 0.54, + "latex": "\\mathbf { x }" + }, + { + "category_id": 13, + "poly": [ + 364, + 565, + 392, + 565, + 392, + 592, + 364, + 592 + ], + "score": 0.47, + "latex": "H" + }, + { + "category_id": 13, + "poly": [ + 1144, + 981, + 1165, + 981, + 1165, + 1004, + 1144, + 1004 + ], + "score": 0.44, + "latex": "\\mathbf { x }" + }, + { + "category_id": 13, + "poly": [ + 1330, + 1497, + 1351, + 1497, + 1351, + 1520, + 1330, + 1520 + ], + "score": 0.4, + "latex": "\\mathbf { x }" + }, + { + "category_id": 13, + "poly": [ + 769, + 601, + 789, + 601, + 789, + 623, + 769, + 623 + ], + "score": 0.39, + "latex": "\\mathbf { x }" + }, + { + "category_id": 13, + "poly": [ + 400, + 565, + 434, + 565, + 434, + 593, + 400, + 593 + ], + "score": 0.37, + "latex": "W" + }, + { + "category_id": 13, + "poly": [ + 363, + 565, + 435, + 565, + 435, + 595, + 363, + 595 + ], + "score": 0.28, + "latex": "H , W" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 397.0, + 891.0, + 397.0, + 891.0, + 442.0, + 291.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 451.0, + 636.0, + 451.0, + 636.0, + 490.0, + 294.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 884.0, + 1403.0, + 884.0, + 1403.0, + 919.0, + 295.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 915.0, + 1405.0, + 915.0, + 1405.0, + 949.0, + 294.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 939.0, + 773.0, + 939.0, + 773.0, + 983.0, + 291.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 939.0, + 1407.0, + 939.0, + 1407.0, + 983.0, + 891.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 975.0, + 1143.0, + 975.0, + 1143.0, + 1012.0, + 292.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 975.0, + 1249.0, + 975.0, + 1249.0, + 1012.0, + 1166.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 975.0, + 1405.0, + 975.0, + 1405.0, + 1012.0, + 1296.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1004.0, + 601.0, + 1004.0, + 601.0, + 1042.0, + 294.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 1004.0, + 1034.0, + 1004.0, + 1034.0, + 1042.0, + 729.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 1004.0, + 1406.0, + 1004.0, + 1406.0, + 1042.0, + 1048.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1036.0, + 1406.0, + 1036.0, + 1406.0, + 1074.0, + 294.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1063.0, + 1406.0, + 1063.0, + 1406.0, + 1105.0, + 294.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1096.0, + 1406.0, + 1096.0, + 1406.0, + 1133.0, + 294.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1121.0, + 296.0, + 1121.0, + 296.0, + 1166.0, + 291.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 1121.0, + 914.0, + 1121.0, + 914.0, + 1166.0, + 436.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 1121.0, + 1407.0, + 1121.0, + 1407.0, + 1166.0, + 1123.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1155.0, + 619.0, + 1155.0, + 619.0, + 1195.0, + 292.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 1155.0, + 1254.0, + 1155.0, + 1254.0, + 1195.0, + 729.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1155.0, + 1408.0, + 1155.0, + 1408.0, + 1195.0, + 1399.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1189.0, + 1407.0, + 1189.0, + 1407.0, + 1224.0, + 295.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1214.0, + 372.0, + 1214.0, + 372.0, + 1255.0, + 291.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 1214.0, + 1407.0, + 1214.0, + 1407.0, + 1255.0, + 428.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1248.0, + 1406.0, + 1248.0, + 1406.0, + 1286.0, + 294.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1279.0, + 1405.0, + 1279.0, + 1405.0, + 1317.0, + 294.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1308.0, + 626.0, + 1308.0, + 626.0, + 1346.0, + 292.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 1308.0, + 900.0, + 1308.0, + 900.0, + 1346.0, + 651.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1308.0, + 1241.0, + 1308.0, + 1241.0, + 1346.0, + 925.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 1308.0, + 1406.0, + 1308.0, + 1406.0, + 1346.0, + 1297.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1341.0, + 533.0, + 1341.0, + 533.0, + 1375.0, + 293.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1398.0, + 1403.0, + 1398.0, + 1403.0, + 1433.0, + 296.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1429.0, + 1406.0, + 1429.0, + 1406.0, + 1464.0, + 293.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1455.0, + 474.0, + 1455.0, + 474.0, + 1499.0, + 292.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 1455.0, + 986.0, + 1455.0, + 986.0, + 1499.0, + 525.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 1455.0, + 1407.0, + 1455.0, + 1407.0, + 1499.0, + 1102.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1485.0, + 477.0, + 1485.0, + 477.0, + 1529.0, + 292.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 1485.0, + 1069.0, + 1485.0, + 1069.0, + 1529.0, + 605.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 1485.0, + 1329.0, + 1485.0, + 1329.0, + 1529.0, + 1102.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 1485.0, + 1407.0, + 1485.0, + 1407.0, + 1529.0, + 1352.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1522.0, + 845.0, + 1522.0, + 845.0, + 1559.0, + 293.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1522.0, + 1407.0, + 1522.0, + 1407.0, + 1559.0, + 868.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1552.0, + 506.0, + 1552.0, + 506.0, + 1589.0, + 295.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 1552.0, + 993.0, + 1552.0, + 993.0, + 1589.0, + 542.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 1552.0, + 1406.0, + 1552.0, + 1406.0, + 1589.0, + 1015.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1582.0, + 1407.0, + 1582.0, + 1407.0, + 1619.0, + 295.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1610.0, + 1405.0, + 1610.0, + 1405.0, + 1649.0, + 293.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1644.0, + 611.0, + 1644.0, + 611.0, + 1677.0, + 295.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 502.0, + 342.0, + 502.0, + 342.0, + 540.0, + 294.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 502.0, + 945.0, + 502.0, + 945.0, + 540.0, + 380.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 502.0, + 1406.0, + 502.0, + 1406.0, + 540.0, + 1108.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 524.0, + 436.0, + 524.0, + 436.0, + 575.0, + 290.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 524.0, + 667.0, + 524.0, + 667.0, + 575.0, + 618.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 524.0, + 1412.0, + 524.0, + 1412.0, + 575.0, + 775.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 564.0, + 362.0, + 564.0, + 362.0, + 599.0, + 295.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 564.0, + 443.0, + 564.0, + 443.0, + 599.0, + 436.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 564.0, + 537.0, + 564.0, + 537.0, + 599.0, + 483.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 564.0, + 1406.0, + 564.0, + 1406.0, + 599.0, + 582.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 591.0, + 768.0, + 591.0, + 768.0, + 631.0, + 291.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 591.0, + 922.0, + 591.0, + 922.0, + 631.0, + 790.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 591.0, + 1181.0, + 591.0, + 1181.0, + 631.0, + 959.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 591.0, + 1407.0, + 591.0, + 1407.0, + 631.0, + 1399.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 621.0, + 659.0, + 621.0, + 659.0, + 662.0, + 294.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 621.0, + 825.0, + 621.0, + 825.0, + 662.0, + 773.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 621.0, + 1406.0, + 621.0, + 1406.0, + 662.0, + 927.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 651.0, + 545.0, + 651.0, + 545.0, + 697.0, + 291.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 651.0, + 1407.0, + 651.0, + 1407.0, + 697.0, + 623.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 684.0, + 595.0, + 684.0, + 595.0, + 722.0, + 293.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 684.0, + 785.0, + 684.0, + 785.0, + 722.0, + 775.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 734.0, + 1405.0, + 734.0, + 1405.0, + 767.0, + 295.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 765.0, + 1376.0, + 765.0, + 1376.0, + 797.0, + 293.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 765.0, + 1407.0, + 765.0, + 1407.0, + 797.0, + 1399.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 794.0, + 1406.0, + 794.0, + 1406.0, + 830.0, + 295.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 825.0, + 938.0, + 825.0, + 938.0, + 858.0, + 295.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1875.0, + 1409.0, + 1875.0, + 1409.0, + 1920.0, + 292.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1908.0, + 565.0, + 1908.0, + 565.0, + 1948.0, + 295.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 1908.0, + 827.0, + 1908.0, + 827.0, + 1948.0, + 744.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1908.0, + 1406.0, + 1908.0, + 1406.0, + 1948.0, + 877.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1940.0, + 790.0, + 1940.0, + 790.0, + 1977.0, + 293.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 1940.0, + 1405.0, + 1940.0, + 1405.0, + 1977.0, + 839.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1960.0, + 815.0, + 1960.0, + 815.0, + 2016.0, + 288.0, + 2016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1960.0, + 1409.0, + 1960.0, + 1409.0, + 2016.0, + 1183.0, + 2016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2001.0, + 846.0, + 2001.0, + 846.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 2001.0, + 903.0, + 2001.0, + 903.0, + 2039.0, + 891.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 2001.0, + 1005.0, + 2001.0, + 1005.0, + 2039.0, + 952.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 2001.0, + 1405.0, + 2001.0, + 1405.0, + 2039.0, + 1047.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 231.0, + 1404.0, + 231.0, + 1404.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 262.0, + 1404.0, + 262.0, + 1404.0, + 295.0, + 296.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 291.0, + 1406.0, + 291.0, + 1406.0, + 326.0, + 293.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 322.0, + 1402.0, + 322.0, + 1402.0, + 355.0, + 296.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 350.0, + 1390.0, + 350.0, + 1390.0, + 389.0, + 293.0, + 389.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 667, + 1405, + 667, + 1405, + 946, + 296, + 946 + ], + "score": 0.986 + }, + { + "category_id": 1, + "poly": [ + 296, + 423, + 1407, + 423, + 1407, + 608, + 296, + 608 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1758, + 1403, + 1758, + 1403, + 1945, + 297, + 1945 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1387, + 1404, + 1387, + 1404, + 1512, + 298, + 1512 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1250, + 1404, + 1250, + 1404, + 1374, + 297, + 1374 + ], + "score": 0.977 + }, + { + "category_id": 8, + "poly": [ + 577, + 277, + 1120, + 277, + 1120, + 397, + 577, + 397 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 422, + 1957, + 1278, + 1957, + 1278, + 2053, + 422, + 2053 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 443, + 1003, + 1255, + 1003, + 1255, + 1088, + 443, + 1088 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 299, + 1113, + 1406, + 1113, + 1406, + 1204, + 299, + 1204 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 389, + 1570, + 1295, + 1570, + 1295, + 1656, + 389, + 1656 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 299, + 1680, + 1404, + 1680, + 1404, + 1747, + 299, + 1747 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 443, + 1203, + 1253, + 1203, + 1253, + 1242, + 443, + 1242 + ], + "score": 0.932 + }, + { + "category_id": 0, + "poly": [ + 299, + 626, + 882, + 626, + 882, + 658, + 299, + 658 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 298, + 230, + 734, + 230, + 734, + 262, + 298, + 262 + ], + "score": 0.926 + }, + { + "category_id": 2, + "poly": [ + 297, + 74, + 858, + 74, + 858, + 106, + 297, + 106 + ], + "score": 0.92 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1207, + 1400, + 1207, + 1400, + 1236, + 1366, + 1236 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1051, + 1400, + 1051, + 1400, + 1081, + 1366, + 1081 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1365, + 360, + 1401, + 360, + 1401, + 391, + 1365, + 391 + ], + "score": 0.882 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1619, + 1400, + 1619, + 1400, + 1650, + 1366, + 1650 + ], + "score": 0.88 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1990, + 1401, + 1990, + 1401, + 2020, + 1365, + 2020 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 857, + 2089, + 857, + 2111, + 840, + 2111 + ], + "score": 0.801 + }, + { + "category_id": 14, + "poly": [ + 423, + 1954, + 1276, + 1954, + 1276, + 2054, + 423, + 2054 + ], + "score": 0.95, + "latex": "\\mathcal { L } _ { \\mathrm { s m o o t h } } ( \\alpha ) = \\sum _ { h = 1 } ^ { H _ { F } - 1 } \\sum _ { w = 1 } ^ { W _ { F } } \\| \\alpha _ { h + 1 , w } - \\alpha _ { h , w } \\| _ { 2 } ^ { 2 } + \\sum _ { h = 1 } ^ { H _ { F } } \\sum _ { w = 1 } ^ { W _ { F } - 1 } \\| \\alpha _ { h , w + 1 } - \\alpha _ { h , w } \\| _ { 2 } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 747, + 1683, + 861, + 1683, + 861, + 1717, + 747, + 1717 + ], + "score": 0.93, + "latex": "l ( \\mathbf { y } ^ { * } , \\mathbf { y } ^ { \\mathrm { t g t } } )" + }, + { + "category_id": 13, + "poly": [ + 1167, + 1449, + 1265, + 1449, + 1265, + 1483, + 1167, + 1483 + ], + "score": 0.93, + "latex": "\\Phi _ { \\mathcal { M } } ( \\cdot , \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1236, + 455, + 1397, + 455, + 1397, + 487, + 1236, + 487 + ], + "score": 0.92, + "latex": "1 \\leq h \\leq H _ { F }" + }, + { + "category_id": 13, + "poly": [ + 296, + 486, + 465, + 486, + 465, + 516, + 296, + 516 + ], + "score": 0.91, + "latex": "1 \\leq w \\leq W _ { F }" + }, + { + "category_id": 13, + "poly": [ + 658, + 700, + 844, + 700, + 844, + 732, + 658, + 732 + ], + "score": 0.91, + "latex": "\\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; { \\mathbf { x } } , { \\mathbf { c } } , { \\mathbf { c } } ^ { \\mathrm { n e w } } )" + }, + { + "category_id": 13, + "poly": [ + 512, + 1249, + 610, + 1249, + 610, + 1286, + 512, + 1286 + ], + "score": 0.91, + "latex": "\\Phi _ { \\mathcal { M } } ^ { \\mathrm { i d } } ( \\cdot , \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 529, + 486, + 690, + 486, + 690, + 516, + 529, + 516 + ], + "score": 0.91, + "latex": "1 \\leq k \\leq C _ { F } ," + }, + { + "category_id": 13, + "poly": [ + 350, + 851, + 461, + 851, + 461, + 885, + 350, + 885 + ], + "score": 0.91, + "latex": "\\left( \\mathbf { x } ^ { \\mathrm { { t g t } } } , \\mathbf { y } ^ { \\mathrm { { t g t } } } \\right)" + }, + { + "category_id": 13, + "poly": [ + 1290, + 669, + 1342, + 669, + 1342, + 703, + 1290, + 703 + ], + "score": 0.91, + "latex": "\\mathcal { L } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 627, + 732, + 813, + 732, + 813, + 763, + 627, + 763 + ], + "score": 0.9, + "latex": "\\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; { \\mathbf { x } } , { \\mathbf { c } } , { \\mathbf { c } } ^ { \\mathrm { n e w } } )" + }, + { + "category_id": 14, + "poly": [ + 388, + 1568, + 1309, + 1568, + 1309, + 1658, + 388, + 1658 + ], + "score": 0.9, + "latex": "\\begin{array} { r l } & { \\mathcal { L } _ { \\mathrm { a d v } } \\big ( \\mathcal { T } _ { \\mathcal { G } } \\big ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } \\big ) ; \\mathcal { M } , \\mathbf { y } ^ { \\mathrm { t g t } } \\big ) = P _ { \\gamma \\sim \\mathcal { N } ( 0 , 1 ) } \\big [ \\Phi _ { \\mathcal { M } } \\big ( \\mathcal { T } _ { \\mathcal { G } } \\big ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } \\big ) , \\mathbf { y } ^ { * } \\big ) } \\\\ & { \\qquad - \\Phi _ { \\mathcal { M } } \\big ( \\mathcal { T } _ { \\mathcal { G } } \\big ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } \\big ) , \\mathbf { y } ^ { \\mathrm { t g t } } \\big ) < \\gamma \\big ] \\cdot l \\big ( \\mathbf { y } ^ { * } , \\mathbf { y } ^ { \\mathrm { t g t } } \\big ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 373, + 1683, + 689, + 1683, + 689, + 1718, + 373, + 1718 + ], + "score": 0.9, + "latex": "\\mathbf { y } ^ { * } = \\mathcal { M } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) )" + }, + { + "category_id": 14, + "poly": [ + 443, + 1001, + 1256, + 1001, + 1256, + 1092, + 443, + 1092 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { \\begin{array} { c } { \\mathbf { x } ^ { \\mathrm { a d v } } = \\mathop { \\mathrm { a r g m i n } _ { \\alpha } } \\mathcal { L } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ) } \\\\ { \\mathcal { L } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ) = \\mathcal { L } _ { \\mathrm { a d v } } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ; \\mathcal { M } , \\mathbf { y } ^ { \\mathrm { t g t } } ) + \\lambda \\cdot \\mathcal { L } _ { \\mathrm { s m o o t h } } ( \\alpha ) } \\end{array} } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 443, + 1199, + 1254, + 1199, + 1254, + 1243, + 443, + 1243 + ], + "score": 0.89, + "latex": "\\mathcal { L } _ { \\mathrm { a d v } } \\big ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ; \\mathcal { M } , \\mathbf { y } ^ { \\mathrm { t g t } } \\big ) = \\operatorname* { m a x } \\big ( \\kappa , \\Phi _ { \\mathcal { M } } ^ { \\mathrm { i d } } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) , \\mathbf { x } ^ { \\mathrm { t g t } } ) \\big )" + }, + { + "category_id": 13, + "poly": [ + 604, + 1761, + 686, + 1761, + 686, + 1791, + 604, + 1791 + ], + "score": 0.89, + "latex": "\\mathcal { L } _ { \\mathrm { s m o o t h } }" + }, + { + "category_id": 13, + "poly": [ + 731, + 884, + 774, + 884, + 774, + 915, + 731, + 915 + ], + "score": 0.89, + "latex": "\\mathbf { y } ^ { \\mathrm { t g t } }" + }, + { + "category_id": 13, + "poly": [ + 589, + 1285, + 623, + 1285, + 623, + 1313, + 589, + 1313 + ], + "score": 0.89, + "latex": "L _ { 2 }" + }, + { + "category_id": 14, + "poly": [ + 574, + 275, + 1122, + 275, + 1122, + 399, + 574, + 399 + ], + "score": 0.89, + "latex": "\\begin{array} { r l } & { \\mathbf { x } ^ { \\mathrm { a d v } } = \\arg \\operatorname* { m i n } _ { \\alpha } \\mathcal { L } ( \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) ) } \\\\ & { \\mathrm { w h e r e ~ } \\mathcal { T } _ { \\mathcal { G } } ( \\alpha ; \\mathbf { x } , \\mathbf { c } , \\mathbf { c } ^ { \\mathrm { n e w } } ) = \\mathcal { G } _ { \\mathrm { d e c } } ( \\mathbf { f } ^ { * } ) , } \\\\ & { \\mathbf { f } ^ { * } = \\alpha \\odot \\mathcal { G } _ { \\mathrm { e n c } } ( \\mathbf { x } , \\mathbf { c } ) + ( \\mathbf { 1 } - \\alpha ) \\odot \\mathcal { G } _ { \\mathrm { e n c } } ( \\mathbf { x } , \\mathbf { c } ^ { \\mathrm { n e w } } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 758, + 852, + 801, + 852, + 801, + 883, + 758, + 883 + ], + "score": 0.88, + "latex": "\\mathbf { y } ^ { \\mathrm { t g t } }" + }, + { + "category_id": 13, + "poly": [ + 904, + 669, + 954, + 669, + 954, + 698, + 904, + 698 + ], + "score": 0.86, + "latex": "\\mathbf { x } ^ { \\mathrm { a d v } }" + }, + { + "category_id": 13, + "poly": [ + 974, + 1450, + 1017, + 1450, + 1017, + 1482, + 974, + 1482 + ], + "score": 0.86, + "latex": "\\mathbf { y } ^ { \\mathrm { t g t } }" + }, + { + "category_id": 13, + "poly": [ + 962, + 455, + 1144, + 455, + 1144, + 489, + 962, + 489 + ], + "score": 0.85, + "latex": "( \\alpha _ { h , w , k } \\in ( 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 732, + 453, + 950, + 453, + 950, + 485, + 732, + 485 + ], + "score": 0.85, + "latex": "\\alpha \\in \\mathbb { R } ^ { H _ { F } \\times W _ { F } \\times C _ { F } }" + }, + { + "category_id": 13, + "poly": [ + 1192, + 1146, + 1228, + 1146, + 1228, + 1172, + 1192, + 1172 + ], + "score": 0.84, + "latex": "\\mathcal { M }" + }, + { + "category_id": 13, + "poly": [ + 299, + 1284, + 333, + 1284, + 333, + 1310, + 299, + 1310 + ], + "score": 0.83, + "latex": "\\mathcal { M }" + }, + { + "category_id": 13, + "poly": [ + 727, + 1484, + 748, + 1484, + 748, + 1512, + 727, + 1512 + ], + "score": 0.82, + "latex": "\\gamma" + }, + { + "category_id": 13, + "poly": [ + 605, + 792, + 625, + 792, + 625, + 818, + 605, + 818 + ], + "score": 0.8, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 694, + 823, + 729, + 823, + 729, + 849, + 694, + 849 + ], + "score": 0.8, + "latex": "\\mathcal { M }" + }, + { + "category_id": 13, + "poly": [ + 411, + 762, + 432, + 762, + 432, + 790, + 411, + 790 + ], + "score": 0.77, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 443, + 1826, + 464, + 1826, + 464, + 1848, + 443, + 1848 + ], + "score": 0.77, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 298, + 1320, + 317, + 1320, + 317, + 1340, + 298, + 1340 + ], + "score": 0.74, + "latex": "\\kappa" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 624.0, + 883.0, + 624.0, + 883.0, + 661.0, + 295.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 109.0, + 297.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 663.0, + 903.0, + 663.0, + 903.0, + 707.0, + 291.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 663.0, + 1289.0, + 663.0, + 1289.0, + 707.0, + 955.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1343.0, + 663.0, + 1406.0, + 663.0, + 1406.0, + 707.0, + 1343.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 693.0, + 657.0, + 693.0, + 657.0, + 738.0, + 291.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 693.0, + 1411.0, + 693.0, + 1411.0, + 738.0, + 845.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 726.0, + 626.0, + 726.0, + 626.0, + 766.0, + 291.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 726.0, + 1407.0, + 726.0, + 1407.0, + 766.0, + 814.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 759.0, + 410.0, + 759.0, + 410.0, + 795.0, + 293.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 759.0, + 1407.0, + 759.0, + 1407.0, + 795.0, + 433.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 791.0, + 604.0, + 791.0, + 604.0, + 824.0, + 294.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 791.0, + 1405.0, + 791.0, + 1405.0, + 824.0, + 626.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 821.0, + 693.0, + 821.0, + 693.0, + 856.0, + 293.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 821.0, + 1406.0, + 821.0, + 1406.0, + 856.0, + 730.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 847.0, + 349.0, + 847.0, + 349.0, + 889.0, + 291.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 847.0, + 757.0, + 847.0, + 757.0, + 889.0, + 462.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 802.0, + 847.0, + 1408.0, + 847.0, + 1408.0, + 889.0, + 802.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 882.0, + 730.0, + 882.0, + 730.0, + 918.0, + 294.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 882.0, + 1407.0, + 882.0, + 1407.0, + 918.0, + 775.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 913.0, + 843.0, + 913.0, + 843.0, + 949.0, + 293.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 424.0, + 1407.0, + 424.0, + 1407.0, + 460.0, + 296.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 446.0, + 731.0, + 446.0, + 731.0, + 495.0, + 289.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 446.0, + 961.0, + 446.0, + 961.0, + 495.0, + 951.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 446.0, + 1235.0, + 446.0, + 1235.0, + 495.0, + 1145.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 446.0, + 1412.0, + 446.0, + 1412.0, + 495.0, + 1398.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 485.0, + 528.0, + 485.0, + 528.0, + 521.0, + 466.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 485.0, + 1405.0, + 485.0, + 1405.0, + 521.0, + 691.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 514.0, + 1405.0, + 514.0, + 1405.0, + 551.0, + 293.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 545.0, + 1405.0, + 545.0, + 1405.0, + 580.0, + 295.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 575.0, + 795.0, + 575.0, + 795.0, + 610.0, + 295.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1755.0, + 603.0, + 1755.0, + 603.0, + 1796.0, + 292.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1755.0, + 1407.0, + 1755.0, + 1407.0, + 1796.0, + 687.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1791.0, + 1405.0, + 1791.0, + 1405.0, + 1823.0, + 296.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1818.0, + 442.0, + 1818.0, + 442.0, + 1857.0, + 292.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1818.0, + 1407.0, + 1818.0, + 1407.0, + 1857.0, + 465.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1849.0, + 1405.0, + 1849.0, + 1405.0, + 1886.0, + 292.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 1408.0, + 1880.0, + 1408.0, + 1916.0, + 293.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1910.0, + 542.0, + 1910.0, + 542.0, + 1946.0, + 292.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1389.0, + 1402.0, + 1389.0, + 1402.0, + 1421.0, + 296.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1419.0, + 1405.0, + 1419.0, + 1405.0, + 1452.0, + 293.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1446.0, + 973.0, + 1446.0, + 973.0, + 1486.0, + 292.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 1446.0, + 1166.0, + 1446.0, + 1166.0, + 1486.0, + 1018.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 1446.0, + 1404.0, + 1446.0, + 1404.0, + 1486.0, + 1266.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1480.0, + 726.0, + 1480.0, + 726.0, + 1512.0, + 293.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 1480.0, + 933.0, + 1480.0, + 933.0, + 1512.0, + 749.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1247.0, + 511.0, + 1247.0, + 511.0, + 1290.0, + 291.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 1247.0, + 1407.0, + 1247.0, + 1407.0, + 1290.0, + 611.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1283.0, + 298.0, + 1283.0, + 298.0, + 1316.0, + 295.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 1283.0, + 588.0, + 1283.0, + 588.0, + 1316.0, + 334.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 1283.0, + 1406.0, + 1283.0, + 1406.0, + 1316.0, + 624.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1312.0, + 297.0, + 1312.0, + 297.0, + 1347.0, + 291.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1312.0, + 1405.0, + 1312.0, + 1405.0, + 1347.0, + 318.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1341.0, + 492.0, + 1341.0, + 492.0, + 1379.0, + 294.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1114.0, + 1404.0, + 1114.0, + 1404.0, + 1148.0, + 296.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1142.0, + 1191.0, + 1142.0, + 1191.0, + 1180.0, + 292.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 1142.0, + 1404.0, + 1142.0, + 1404.0, + 1180.0, + 1229.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1172.0, + 373.0, + 1172.0, + 373.0, + 1211.0, + 292.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1677.0, + 372.0, + 1677.0, + 372.0, + 1721.0, + 293.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 1677.0, + 746.0, + 1677.0, + 746.0, + 1721.0, + 690.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 1677.0, + 1406.0, + 1677.0, + 1406.0, + 1721.0, + 862.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1711.0, + 378.0, + 1711.0, + 378.0, + 1755.0, + 290.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 225.0, + 737.0, + 225.0, + 737.0, + 268.0, + 294.0, + 268.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 808, + 1405, + 808, + 1405, + 1265, + 297, + 1265 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 472, + 1404, + 472, + 1404, + 748, + 298, + 748 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 273, + 1403, + 273, + 1403, + 459, + 297, + 459 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1452, + 1405, + 1452, + 1405, + 1666, + 298, + 1666 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1819, + 1403, + 1819, + 1403, + 2034, + 297, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1281, + 1405, + 1281, + 1405, + 1435, + 298, + 1435 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1680, + 1403, + 1680, + 1403, + 1804, + 299, + 1804 + ], + "score": 0.977 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 299, + 770, + 641, + 770, + 641, + 799, + 299, + 799 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 299, + 226, + 558, + 226, + 558, + 261, + 299, + 261 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.718 + }, + { + "category_id": 13, + "poly": [ + 1220, + 1513, + 1397, + 1513, + 1397, + 1547, + 1220, + 1547 + ], + "score": 0.92, + "latex": "\\mathbf { f } ^ { \\prime } = \\mathcal { G } _ { \\mathrm { e n c } } ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } )" + }, + { + "category_id": 13, + "poly": [ + 631, + 1820, + 720, + 1820, + 720, + 1854, + 631, + 1854 + ], + "score": 0.92, + "latex": "( \\mathbf { x } , \\mathbf { x } ^ { \\mathrm { t g t } } )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1049, + 405, + 1049, + 405, + 1081, + 297, + 1081 + ], + "score": 0.91, + "latex": "3 \\times 1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 559, + 1543, + 732, + 1543, + 732, + 1577, + 559, + 1577 + ], + "score": 0.91, + "latex": "\\| \\mathcal { G } ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } ) - \\mathbf { x } \\|" + }, + { + "category_id": 13, + "poly": [ + 646, + 1049, + 707, + 1049, + 707, + 1080, + 646, + 1080 + ], + "score": 0.91, + "latex": "1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1513, + 442, + 1513, + 442, + 1545, + 297, + 1545 + ], + "score": 0.9, + "latex": "\\mathbf { x } \\neq \\mathcal { G } ( \\mathbf { x } , \\mathbf { c } ) )" + }, + { + "category_id": 13, + "poly": [ + 529, + 1743, + 649, + 1743, + 649, + 1772, + 529, + 1772 + ], + "score": 0.89, + "latex": "1 1 2 \\times 1 1 2" + }, + { + "category_id": 13, + "poly": [ + 1136, + 1019, + 1197, + 1019, + 1197, + 1050, + 1136, + 1050 + ], + "score": 0.89, + "latex": "1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 474, + 1051, + 582, + 1051, + 582, + 1081, + 474, + 1081 + ], + "score": 0.88, + "latex": "\\kappa = 1 . 0 5" + }, + { + "category_id": 13, + "poly": [ + 910, + 505, + 946, + 505, + 946, + 537, + 910, + 537 + ], + "score": 0.87, + "latex": "{ \\mathcal { L } } _ { p }" + }, + { + "category_id": 13, + "poly": [ + 1272, + 1021, + 1391, + 1021, + 1391, + 1050, + 1272, + 1050 + ], + "score": 0.87, + "latex": "\\kappa = 1 . 2 4 ," + }, + { + "category_id": 13, + "poly": [ + 776, + 1051, + 883, + 1051, + 883, + 1081, + 776, + 1081 + ], + "score": 0.87, + "latex": "\\kappa = 0 . 6 0 " + }, + { + "category_id": 13, + "poly": [ + 374, + 1545, + 550, + 1545, + 550, + 1577, + 374, + 1577 + ], + "score": 0.81, + "latex": "\\mathbf { x } ^ { \\prime } = \\mathrm { a r g m i n } _ { \\mathbf { x } ^ { \\prime } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 930, + 325, + 930, + 325, + 956, + 298, + 956 + ], + "score": 0.8, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 956, + 995, + 976, + 995, + 976, + 1017, + 956, + 1017 + ], + "score": 0.75, + "latex": "\\kappa" + }, + { + "category_id": 13, + "poly": [ + 816, + 1606, + 837, + 1606, + 837, + 1632, + 816, + 1632 + ], + "score": 0.71, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 436, + 1235, + 525, + 1235, + 525, + 1264, + 436, + 1264 + ], + "score": 0.32, + "latex": "{ \\mathrm { F a c e } } + +" + }, + { + "category_id": 13, + "poly": [ + 527, + 1887, + 548, + 1887, + 548, + 1910, + 527, + 1910 + ], + "score": 0.27, + "latex": "\\mathbf { x }" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 767.0, + 644.0, + 767.0, + 644.0, + 803.0, + 295.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 226.0, + 561.0, + 226.0, + 561.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 808.0, + 1405.0, + 808.0, + 1405.0, + 839.0, + 296.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 835.0, + 1406.0, + 835.0, + 1406.0, + 875.0, + 292.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 869.0, + 1405.0, + 869.0, + 1405.0, + 903.0, + 295.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 898.0, + 1405.0, + 898.0, + 1405.0, + 933.0, + 295.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 926.0, + 297.0, + 926.0, + 297.0, + 964.0, + 294.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 926.0, + 1406.0, + 926.0, + 1406.0, + 964.0, + 326.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 959.0, + 1406.0, + 959.0, + 1406.0, + 994.0, + 294.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 987.0, + 955.0, + 987.0, + 955.0, + 1024.0, + 292.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 987.0, + 1406.0, + 987.0, + 1406.0, + 1024.0, + 977.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1015.0, + 1135.0, + 1015.0, + 1135.0, + 1056.0, + 291.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 1015.0, + 1271.0, + 1015.0, + 1271.0, + 1056.0, + 1198.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1392.0, + 1015.0, + 1409.0, + 1015.0, + 1409.0, + 1056.0, + 1392.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1045.0, + 296.0, + 1045.0, + 296.0, + 1086.0, + 289.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1045.0, + 473.0, + 1045.0, + 473.0, + 1086.0, + 406.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 1045.0, + 645.0, + 1045.0, + 645.0, + 1086.0, + 583.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 1045.0, + 775.0, + 1045.0, + 775.0, + 1086.0, + 708.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 1045.0, + 1410.0, + 1045.0, + 1410.0, + 1086.0, + 884.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1081.0, + 1407.0, + 1081.0, + 1407.0, + 1116.0, + 292.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1111.0, + 1408.0, + 1111.0, + 1408.0, + 1146.0, + 295.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1142.0, + 1406.0, + 1142.0, + 1406.0, + 1177.0, + 295.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1172.0, + 1405.0, + 1172.0, + 1405.0, + 1207.0, + 294.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1204.0, + 1405.0, + 1204.0, + 1405.0, + 1235.0, + 296.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1234.0, + 435.0, + 1234.0, + 435.0, + 1268.0, + 293.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 1234.0, + 778.0, + 1234.0, + 778.0, + 1268.0, + 526.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 474.0, + 1404.0, + 474.0, + 1404.0, + 507.0, + 297.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 502.0, + 909.0, + 502.0, + 909.0, + 539.0, + 292.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 502.0, + 1406.0, + 502.0, + 1406.0, + 539.0, + 947.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 533.0, + 1405.0, + 533.0, + 1405.0, + 570.0, + 294.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 562.0, + 1408.0, + 562.0, + 1408.0, + 602.0, + 293.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 596.0, + 1406.0, + 596.0, + 1406.0, + 632.0, + 293.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 623.0, + 1404.0, + 623.0, + 1404.0, + 662.0, + 293.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 654.0, + 1406.0, + 654.0, + 1406.0, + 692.0, + 293.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 684.0, + 1405.0, + 684.0, + 1405.0, + 722.0, + 293.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 714.0, + 1166.0, + 714.0, + 1166.0, + 754.0, + 292.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 269.0, + 1407.0, + 269.0, + 1407.0, + 311.0, + 292.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 301.0, + 1405.0, + 301.0, + 1405.0, + 342.0, + 292.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 334.0, + 1405.0, + 334.0, + 1405.0, + 369.0, + 292.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 367.0, + 1404.0, + 367.0, + 1404.0, + 397.0, + 293.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 391.0, + 1405.0, + 391.0, + 1405.0, + 432.0, + 292.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 429.0, + 1290.0, + 429.0, + 1290.0, + 460.0, + 296.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1452.0, + 1408.0, + 1452.0, + 1408.0, + 1487.0, + 295.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1476.0, + 1411.0, + 1476.0, + 1411.0, + 1523.0, + 291.0, + 1523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1512.0, + 296.0, + 1512.0, + 296.0, + 1550.0, + 292.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 1512.0, + 1219.0, + 1512.0, + 1219.0, + 1550.0, + 443.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1512.0, + 1407.0, + 1512.0, + 1407.0, + 1550.0, + 1398.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1544.0, + 373.0, + 1544.0, + 373.0, + 1579.0, + 295.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 1544.0, + 558.0, + 1544.0, + 558.0, + 1579.0, + 551.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 1544.0, + 1406.0, + 1544.0, + 1406.0, + 1579.0, + 733.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1574.0, + 1405.0, + 1574.0, + 1405.0, + 1608.0, + 295.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1606.0, + 815.0, + 1606.0, + 815.0, + 1636.0, + 296.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 1606.0, + 1403.0, + 1606.0, + 1403.0, + 1636.0, + 838.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1634.0, + 1187.0, + 1634.0, + 1187.0, + 1668.0, + 293.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1818.0, + 630.0, + 1818.0, + 630.0, + 1856.0, + 293.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 1818.0, + 1407.0, + 1818.0, + 1407.0, + 1856.0, + 721.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1850.0, + 1405.0, + 1850.0, + 1405.0, + 1886.0, + 292.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1882.0, + 526.0, + 1882.0, + 526.0, + 1917.0, + 295.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1917.0, + 549.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1946.0, + 293.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1977.0, + 295.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1973.0, + 1409.0, + 1973.0, + 1409.0, + 2008.0, + 295.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 1407.0, + 2003.0, + 1407.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1277.0, + 1407.0, + 1277.0, + 1407.0, + 1317.0, + 295.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1310.0, + 1405.0, + 1310.0, + 1405.0, + 1346.0, + 295.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1338.0, + 1407.0, + 1338.0, + 1407.0, + 1379.0, + 293.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1370.0, + 1406.0, + 1370.0, + 1406.0, + 1410.0, + 293.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1403.0, + 1249.0, + 1403.0, + 1249.0, + 1440.0, + 295.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1678.0, + 1404.0, + 1678.0, + 1404.0, + 1716.0, + 294.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1709.0, + 1405.0, + 1709.0, + 1405.0, + 1748.0, + 292.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1742.0, + 528.0, + 1742.0, + 528.0, + 1777.0, + 294.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 1742.0, + 1405.0, + 1742.0, + 1405.0, + 1777.0, + 650.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1775.0, + 1082.0, + 1775.0, + 1082.0, + 1807.0, + 295.0, + 1807.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 336, + 1405, + 336, + 1405, + 703, + 297, + 703 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1728, + 1405, + 1728, + 1405, + 2034, + 298, + 2034 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 759, + 1404, + 759, + 1404, + 1003, + 297, + 1003 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1499, + 1404, + 1499, + 1404, + 1715, + 298, + 1715 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 377, + 1038, + 1311, + 1038, + 1311, + 1305, + 377, + 1305 + ], + "score": 0.973 + }, + { + "category_id": 4, + "poly": [ + 296, + 1324, + 1405, + 1324, + 1405, + 1477, + 296, + 1477 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 301, + 229, + 1400, + 229, + 1400, + 322, + 301, + 322 + ], + "score": 0.952 + }, + { + "category_id": 0, + "poly": [ + 298, + 721, + 926, + 721, + 926, + 751, + 298, + 751 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.781 + }, + { + "category_id": 13, + "poly": [ + 380, + 1971, + 440, + 1971, + 440, + 2002, + 380, + 2002 + ], + "score": 0.91, + "latex": "1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 580, + 1910, + 639, + 1910, + 639, + 1941, + 580, + 1941 + ], + "score": 0.9, + "latex": "1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 1074, + 459, + 1194, + 459, + 1194, + 488, + 1074, + 488 + ], + "score": 0.9, + "latex": "2 5 6 \\times 2 5 6" + }, + { + "category_id": 13, + "poly": [ + 1292, + 1880, + 1353, + 1880, + 1353, + 1909, + 1292, + 1909 + ], + "score": 0.9, + "latex": "1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 1155, + 368, + 1275, + 368, + 1275, + 397, + 1155, + 397 + ], + "score": 0.9, + "latex": "2 5 6 \\times 2 5 6" + }, + { + "category_id": 13, + "poly": [ + 1290, + 1911, + 1350, + 1911, + 1350, + 1941, + 1290, + 1941 + ], + "score": 0.89, + "latex": "1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 454, + 910, + 529, + 910, + 529, + 941, + 454, + 941 + ], + "score": 0.89, + "latex": "0 . 1 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 1050, + 582, + 1143, + 582, + 1143, + 610, + 1050, + 610 + ], + "score": 0.88, + "latex": "1 6 \\times 1 6" + }, + { + "category_id": 13, + "poly": [ + 1065, + 879, + 1142, + 879, + 1142, + 910, + 1065, + 910 + ], + "score": 0.87, + "latex": "0 . 0 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 860, + 2003, + 913, + 2003, + 913, + 2031, + 860, + 2031 + ], + "score": 0.87, + "latex": "50 \\%" + }, + { + "category_id": 13, + "poly": [ + 1289, + 879, + 1366, + 879, + 1366, + 910, + 1289, + 910 + ], + "score": 0.87, + "latex": "0 . 3 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 663, + 1943, + 730, + 1943, + 730, + 1972, + 663, + 1972 + ], + "score": 0.86, + "latex": "100 \\%" + }, + { + "category_id": 13, + "poly": [ + 296, + 941, + 365, + 941, + 365, + 970, + 296, + 970 + ], + "score": 0.86, + "latex": "100 \\%" + }, + { + "category_id": 13, + "poly": [ + 456, + 1414, + 630, + 1414, + 630, + 1445, + 456, + 1445 + ], + "score": 0.67, + "latex": "\\mathrm { \\bf G } \\mathrm { - } \\mathrm { \\bf F P R } = 1 0 ^ { - 3 } ," + }, + { + "category_id": 13, + "poly": [ + 767, + 879, + 1045, + 879, + 1045, + 910, + 767, + 910 + ], + "score": 0.64, + "latex": "\\mathrm { G \\mathrm { - } F P R = T \\mathrm { - } F P R = 1 0 ^ { - 3 } , }" + }, + { + "category_id": 13, + "poly": [ + 851, + 882, + 877, + 882, + 877, + 905, + 851, + 905 + ], + "score": 0.34, + "latex": "=" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 1040.0, + 483.0, + 1040.0, + 483.0, + 1060.0, + 386.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 1041.0, + 674.0, + 1041.0, + 674.0, + 1061.0, + 546.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 1042.0, + 901.0, + 1042.0, + 901.0, + 1062.0, + 718.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 1041.0, + 954.0, + 1041.0, + 954.0, + 1062.0, + 909.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1043.0, + 1139.0, + 1043.0, + 1139.0, + 1060.0, + 1012.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1042.0, + 1316.0, + 1042.0, + 1316.0, + 1062.0, + 1185.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1054.0, + 431.0, + 1054.0, + 431.0, + 1075.0, + 388.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 1058.0, + 640.0, + 1058.0, + 640.0, + 1076.0, + 580.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 1057.0, + 814.0, + 1057.0, + 814.0, + 1075.0, + 752.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1058.0, + 1108.0, + 1058.0, + 1108.0, + 1076.0, + 1045.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 1058.0, + 1280.0, + 1058.0, + 1280.0, + 1076.0, + 1221.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 1094.0, + 904.0, + 1094.0, + 904.0, + 1126.0, + 856.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 1135.0, + 606.0, + 1135.0, + 606.0, + 1152.0, + 496.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 1135.0, + 718.0, + 1135.0, + 718.0, + 1152.0, + 616.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 1134.0, + 1058.0, + 1134.0, + 1058.0, + 1154.0, + 979.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 1166.0, + 487.0, + 1166.0, + 487.0, + 1213.0, + 383.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1169.0, + 551.0, + 1169.0, + 551.0, + 1208.0, + 497.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1174.0, + 658.0, + 1174.0, + 658.0, + 1195.0, + 621.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1155.0, + 1020.0, + 1155.0, + 1020.0, + 1222.0, + 716.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1170.0, + 1245.0, + 1170.0, + 1245.0, + 1213.0, + 1205.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1213.0, + 574.0, + 1213.0, + 574.0, + 1231.0, + 528.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 1212.0, + 719.0, + 1212.0, + 719.0, + 1231.0, + 611.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 1213.0, + 1054.0, + 1213.0, + 1054.0, + 1230.0, + 984.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 1213.0, + 1188.0, + 1213.0, + 1188.0, + 1230.0, + 1084.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 1233.0, + 495.0, + 1233.0, + 495.0, + 1289.0, + 376.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1246.0, + 780.0, + 1246.0, + 780.0, + 1272.0, + 734.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 1230.0, + 1029.0, + 1230.0, + 1029.0, + 1292.0, + 836.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 1256.0, + 1050.0, + 1256.0, + 1050.0, + 1266.0, + 1030.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 1283.0, + 578.0, + 1283.0, + 578.0, + 1307.0, + 521.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1287.0, + 721.0, + 1287.0, + 721.0, + 1305.0, + 610.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 1287.0, + 1195.0, + 1287.0, + 1195.0, + 1304.0, + 1071.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 1056.0, + 481.0, + 1056.0, + 481.0, + 1074.0, + 438.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 1056.0, + 898.0, + 1056.0, + 898.0, + 1074.0, + 856.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1056.0, + 952.0, + 1056.0, + 952.0, + 1075.0, + 912.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.75, + 1132.5, + 1159.75, + 1132.5, + 1159.75, + 1153.0, + 1106.75, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.75, + 1179.5, + 1041.75, + 1179.5, + 1041.75, + 1200.0, + 1012.75, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 1287.0, + 1041.0, + 1287.0, + 1041.0, + 1305.5, + 996.0, + 1305.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1323.0, + 1405.0, + 1323.0, + 1405.0, + 1360.0, + 295.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1350.0, + 1405.0, + 1350.0, + 1405.0, + 1393.0, + 293.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1385.0, + 1405.0, + 1385.0, + 1405.0, + 1422.0, + 294.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1414.0, + 455.0, + 1414.0, + 455.0, + 1451.0, + 294.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 1414.0, + 1403.0, + 1414.0, + 1403.0, + 1451.0, + 631.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1445.0, + 733.0, + 1445.0, + 733.0, + 1482.0, + 294.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 720.0, + 927.0, + 720.0, + 927.0, + 755.0, + 294.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 337.0, + 1405.0, + 337.0, + 1405.0, + 371.0, + 295.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 369.0, + 1154.0, + 369.0, + 1154.0, + 400.0, + 292.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 369.0, + 1407.0, + 369.0, + 1407.0, + 400.0, + 1276.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 398.0, + 1405.0, + 398.0, + 1405.0, + 432.0, + 294.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 430.0, + 1405.0, + 430.0, + 1405.0, + 464.0, + 295.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 457.0, + 1073.0, + 457.0, + 1073.0, + 494.0, + 292.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 457.0, + 1405.0, + 457.0, + 1405.0, + 494.0, + 1195.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 487.0, + 1406.0, + 487.0, + 1406.0, + 527.0, + 292.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 521.0, + 1405.0, + 521.0, + 1405.0, + 555.0, + 295.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 550.0, + 1406.0, + 550.0, + 1406.0, + 586.0, + 292.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 583.0, + 1049.0, + 583.0, + 1049.0, + 613.0, + 296.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 583.0, + 1407.0, + 583.0, + 1407.0, + 613.0, + 1144.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 608.0, + 1407.0, + 608.0, + 1407.0, + 650.0, + 292.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 642.0, + 1405.0, + 642.0, + 1405.0, + 676.0, + 295.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 669.0, + 1113.0, + 669.0, + 1113.0, + 709.0, + 292.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1727.0, + 1406.0, + 1727.0, + 1406.0, + 1763.0, + 296.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1758.0, + 1410.0, + 1758.0, + 1410.0, + 1795.0, + 291.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1787.0, + 1406.0, + 1787.0, + 1406.0, + 1825.0, + 293.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1820.0, + 1407.0, + 1820.0, + 1407.0, + 1856.0, + 295.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1851.0, + 1408.0, + 1851.0, + 1408.0, + 1884.0, + 292.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1880.0, + 1291.0, + 1880.0, + 1291.0, + 1916.0, + 292.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 1880.0, + 1408.0, + 1880.0, + 1408.0, + 1916.0, + 1354.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1904.0, + 579.0, + 1904.0, + 579.0, + 1950.0, + 291.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 1904.0, + 1289.0, + 1904.0, + 1289.0, + 1950.0, + 640.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 1904.0, + 1411.0, + 1904.0, + 1411.0, + 1950.0, + 1351.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 662.0, + 1943.0, + 662.0, + 1975.0, + 295.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 1943.0, + 1406.0, + 1943.0, + 1406.0, + 1975.0, + 731.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1967.0, + 379.0, + 1967.0, + 379.0, + 2010.0, + 291.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 1967.0, + 1406.0, + 1967.0, + 1406.0, + 2010.0, + 441.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 859.0, + 2002.0, + 859.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2037.0, + 914.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 759.0, + 1402.0, + 759.0, + 1402.0, + 793.0, + 295.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 788.0, + 1406.0, + 788.0, + 1406.0, + 825.0, + 294.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 820.0, + 1404.0, + 820.0, + 1404.0, + 854.0, + 295.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 849.0, + 1407.0, + 849.0, + 1407.0, + 886.0, + 292.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 877.0, + 766.0, + 877.0, + 766.0, + 913.0, + 291.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 877.0, + 1064.0, + 877.0, + 1064.0, + 913.0, + 1046.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 877.0, + 1288.0, + 877.0, + 1288.0, + 913.0, + 1143.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1367.0, + 877.0, + 1405.0, + 877.0, + 1405.0, + 913.0, + 1367.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 908.0, + 453.0, + 908.0, + 453.0, + 944.0, + 292.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 908.0, + 1406.0, + 908.0, + 1406.0, + 944.0, + 530.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 940.0, + 1404.0, + 940.0, + 1404.0, + 975.0, + 366.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 971.0, + 1049.0, + 971.0, + 1049.0, + 1006.0, + 295.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1500.0, + 1405.0, + 1500.0, + 1405.0, + 1538.0, + 296.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1530.0, + 1407.0, + 1530.0, + 1407.0, + 1566.0, + 293.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1563.0, + 1406.0, + 1563.0, + 1406.0, + 1597.0, + 294.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1593.0, + 1406.0, + 1593.0, + 1406.0, + 1628.0, + 294.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1622.0, + 1407.0, + 1622.0, + 1407.0, + 1658.0, + 292.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1650.0, + 1408.0, + 1650.0, + 1408.0, + 1690.0, + 292.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1681.0, + 1397.0, + 1681.0, + 1397.0, + 1721.0, + 293.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 267.0, + 292.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 262.0, + 1404.0, + 262.0, + 1404.0, + 296.0, + 296.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 291.0, + 843.0, + 291.0, + 843.0, + 325.0, + 296.0, + 325.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1573, + 1405, + 1573, + 1405, + 1880, + 297, + 1880 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1404, + 229, + 1404, + 444, + 297, + 444 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1895, + 1404, + 1895, + 1404, + 2049, + 299, + 2049 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 459, + 1405, + 459, + 1405, + 582, + 297, + 582 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 412, + 1105, + 1283, + 1105, + 1283, + 1454, + 412, + 1454 + ], + "score": 0.969 + }, + { + "category_id": 4, + "poly": [ + 295, + 1473, + 1400, + 1473, + 1400, + 1536, + 295, + 1536 + ], + "score": 0.928 + }, + { + "category_id": 4, + "poly": [ + 297, + 985, + 1406, + 985, + 1406, + 1078, + 297, + 1078 + ], + "score": 0.925 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 858, + 76, + 858, + 104, + 298, + 104 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 857, + 2088, + 857, + 2111, + 841, + 2111 + ], + "score": 0.762 + }, + { + "category_id": 3, + "poly": [ + 300, + 618, + 837, + 618, + 837, + 938, + 300, + 938 + ], + "score": 0.724 + }, + { + "category_id": 3, + "poly": [ + 884, + 618, + 1394, + 618, + 1394, + 939, + 884, + 939 + ], + "score": 0.613 + }, + { + "category_id": 13, + "poly": [ + 1028, + 491, + 1146, + 491, + 1146, + 524, + 1028, + 524 + ], + "score": 0.93, + "latex": "\\mathcal { G } ( \\mathbf { x } , \\mathbf { c } ^ { \\mathrm { n e w } } )" + }, + { + "category_id": 13, + "poly": [ + 294, + 1924, + 387, + 1924, + 387, + 1955, + 294, + 1955 + ], + "score": 0.88, + "latex": "= 1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 294, + 1725, + 382, + 1725, + 382, + 1757, + 294, + 1757 + ], + "score": 0.88, + "latex": "= 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 1188, + 1726, + 1354, + 1726, + 1354, + 1756, + 1188, + 1756 + ], + "score": 0.88, + "latex": "\\mathrm { \\Delta T - F P R } = 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 965, + 1725, + 1137, + 1725, + 1137, + 1756, + 965, + 1756 + ], + "score": 0.86, + "latex": "\\mathbf { \\bar { ( G - F P R = 1 0 ^ { - 4 } } }" + }, + { + "category_id": 13, + "poly": [ + 575, + 522, + 626, + 522, + 626, + 550, + 575, + 550 + ], + "score": 0.85, + "latex": "\\mathbf { c } ^ { \\mathrm { { n e w } } }" + }, + { + "category_id": 13, + "poly": [ + 966, + 1015, + 1244, + 1015, + 1244, + 1046, + 966, + 1046 + ], + "score": 0.83, + "latex": "( \\mathrm { G } \\mathrm { - } \\mathrm { F P R } = \\mathrm { T } \\mathrm { - } \\mathrm { F P R } = 1 0 ^ { - 3 } )" + }, + { + "category_id": 13, + "poly": [ + 1030, + 1472, + 1200, + 1472, + 1200, + 1503, + 1030, + 1503 + ], + "score": 0.8, + "latex": "\\mathrm { ( G { - } F P R = 1 0 ^ { - 3 } }" + }, + { + "category_id": 13, + "poly": [ + 1302, + 1924, + 1398, + 1924, + 1398, + 1955, + 1302, + 1955 + ], + "score": 0.6, + "latex": ". = 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 793, + 1047, + 1070, + 1047, + 1070, + 1076, + 793, + 1076 + ], + "score": 0.53, + "latex": "\\mathrm { \\langle G \\mathrm { - } F P R = T \\mathrm { - } F P R = 1 0 ^ { - 4 } \\rangle }" + }, + { + "category_id": 13, + "poly": [ + 877, + 1049, + 902, + 1049, + 902, + 1072, + 877, + 1072 + ], + "score": 0.47, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1216, + 1695, + 1348, + 1695, + 1348, + 1724, + 1216, + 1724 + ], + "score": 0.32, + "latex": "\\begin{array} { r } { ( \\mathrm { G - F P R } = \\mathrm { T } . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1301, + 1696, + 1342, + 1696, + 1342, + 1722, + 1301, + 1722 + ], + "score": 0.27, + "latex": "= \\mathrm { T }" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1108.0, + 474.0, + 1108.0, + 474.0, + 1128.0, + 423.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 1109.0, + 533.0, + 1109.0, + 533.0, + 1129.0, + 485.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1110.0, + 1079.0, + 1110.0, + 1079.0, + 1134.0, + 757.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 1123.0, + 474.0, + 1123.0, + 474.0, + 1146.0, + 424.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 1123.0, + 533.0, + 1123.0, + 533.0, + 1147.0, + 485.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 1166.0, + 467.0, + 1166.0, + 467.0, + 1179.0, + 429.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 1160.0, + 542.0, + 1160.0, + 542.0, + 1203.0, + 487.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1168.0, + 797.0, + 1168.0, + 797.0, + 1175.0, + 789.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 1156.0, + 1271.0, + 1156.0, + 1271.0, + 1176.0, + 1119.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 1172.0, + 1278.0, + 1172.0, + 1278.0, + 1191.0, + 1109.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 1205.0, + 1265.0, + 1205.0, + 1265.0, + 1230.0, + 1124.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 1221.0, + 1276.0, + 1221.0, + 1276.0, + 1246.0, + 1113.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 1238.0, + 1279.0, + 1238.0, + 1279.0, + 1260.0, + 1107.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 1265.0, + 620.0, + 1265.0, + 620.0, + 1282.0, + 567.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1265.0, + 678.0, + 1265.0, + 678.0, + 1283.0, + 634.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1263.0, + 891.0, + 1263.0, + 891.0, + 1286.0, + 703.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1265.0, + 957.0, + 1265.0, + 957.0, + 1283.0, + 916.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1264.0, + 1038.0, + 1264.0, + 1038.0, + 1284.0, + 974.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 1265.0, + 1104.0, + 1265.0, + 1104.0, + 1282.0, + 1052.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1281.0, + 608.0, + 1281.0, + 608.0, + 1298.0, + 575.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 1279.0, + 674.0, + 1279.0, + 674.0, + 1298.0, + 639.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 1282.0, + 895.0, + 1282.0, + 895.0, + 1299.0, + 838.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1281.0, + 962.0, + 1281.0, + 962.0, + 1298.0, + 912.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 1278.0, + 1123.0, + 1278.0, + 1123.0, + 1301.0, + 1033.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.0, + 1417.0, + 617.0, + 1417.0, + 617.0, + 1441.0, + 566.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 1416.0, + 698.0, + 1416.0, + 698.0, + 1440.0, + 632.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 1420.0, + 827.0, + 1420.0, + 827.0, + 1437.0, + 773.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1418.0, + 970.0, + 1418.0, + 970.0, + 1439.0, + 904.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 1419.0, + 1027.0, + 1419.0, + 1027.0, + 1439.0, + 984.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 1415.0, + 1114.0, + 1415.0, + 1114.0, + 1439.0, + 1044.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 1420.0, + 1165.0, + 1420.0, + 1165.0, + 1437.0, + 1128.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 1437.0, + 625.0, + 1437.0, + 625.0, + 1456.0, + 556.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1433.0, + 693.0, + 1433.0, + 693.0, + 1455.0, + 635.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 1433.0, + 835.0, + 1433.0, + 835.0, + 1454.0, + 763.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1429.0, + 1110.0, + 1429.0, + 1110.0, + 1454.0, + 904.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 1432.0, + 1164.0, + 1432.0, + 1164.0, + 1453.0, + 1128.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 1416.5, + 760.0, + 1416.5, + 760.0, + 1439.5, + 698.0, + 1439.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 1418.5, + 889.0, + 1418.5, + 889.0, + 1439.0, + 838.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1470.0, + 1029.0, + 1470.0, + 1029.0, + 1510.0, + 294.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 1470.0, + 1405.0, + 1470.0, + 1405.0, + 1510.0, + 1201.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1505.0, + 959.0, + 1505.0, + 959.0, + 1536.0, + 296.0, + 1536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 986.0, + 1404.0, + 986.0, + 1404.0, + 1019.0, + 296.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1015.0, + 965.0, + 1015.0, + 965.0, + 1052.0, + 295.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 1015.0, + 1406.0, + 1015.0, + 1406.0, + 1052.0, + 1245.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1045.0, + 792.0, + 1045.0, + 792.0, + 1082.0, + 295.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 1045.0, + 1333.0, + 1045.0, + 1333.0, + 1082.0, + 1071.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 70.0, + 859.0, + 70.0, + 859.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 616.0, + 326.0, + 616.0, + 326.0, + 820.0, + 298.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 617.0, + 710.0, + 617.0, + 710.0, + 647.0, + 485.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 633.0, + 361.0, + 633.0, + 361.0, + 662.0, + 321.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 647.0, + 363.0, + 647.0, + 363.0, + 709.0, + 329.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 691.0, + 364.0, + 691.0, + 364.0, + 740.0, + 328.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 741.0, + 360.0, + 741.0, + 360.0, + 769.0, + 331.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 782.0, + 356.0, + 782.0, + 356.0, + 797.0, + 345.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 789.0, + 639.0, + 789.0, + 639.0, + 868.0, + 609.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 790.0, + 689.0, + 790.0, + 689.0, + 924.0, + 660.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 790.0, + 747.0, + 790.0, + 747.0, + 871.0, + 721.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 790.0, + 765.0, + 790.0, + 765.0, + 858.0, + 738.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 790.0, + 824.0, + 790.0, + 824.0, + 912.0, + 795.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 791.0, + 408.0, + 791.0, + 408.0, + 877.0, + 382.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 791.0, + 436.0, + 791.0, + 436.0, + 891.0, + 410.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 791.0, + 454.0, + 791.0, + 454.0, + 881.0, + 425.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 791.0, + 486.0, + 791.0, + 486.0, + 905.0, + 457.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 791.0, + 655.0, + 791.0, + 655.0, + 870.0, + 628.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 791.0, + 788.0, + 791.0, + 788.0, + 942.0, + 762.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 792.0, + 509.0, + 792.0, + 509.0, + 887.0, + 483.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 792.0, + 535.0, + 792.0, + 535.0, + 867.0, + 509.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 792.0, + 610.0, + 792.0, + 610.0, + 886.0, + 584.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 792.0, + 712.0, + 792.0, + 712.0, + 939.0, + 686.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 792.0, + 730.0, + 792.0, + 730.0, + 870.0, + 704.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 793.0, + 562.0, + 793.0, + 562.0, + 859.0, + 536.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 793.0, + 585.0, + 793.0, + 585.0, + 933.0, + 560.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 794.0, + 467.0, + 794.0, + 467.0, + 878.0, + 441.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 794.0, + 668.0, + 794.0, + 668.0, + 866.0, + 645.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 814.0, + 803.0, + 814.0, + 803.0, + 884.0, + 778.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 614.0, + 1249.0, + 614.0, + 1249.0, + 649.0, + 1065.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 632.0, + 921.0, + 632.0, + 921.0, + 663.0, + 882.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 669.0, + 920.0, + 669.0, + 920.0, + 700.0, + 891.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 703.0, + 922.0, + 703.0, + 922.0, + 738.0, + 889.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 739.0, + 920.0, + 739.0, + 920.0, + 769.0, + 890.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 778.0, + 918.0, + 778.0, + 918.0, + 799.0, + 902.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1323.0, + 788.0, + 1350.0, + 788.0, + 1350.0, + 860.0, + 1323.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 789.0, + 1149.0, + 789.0, + 1149.0, + 868.0, + 1119.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 790.0, + 1121.0, + 790.0, + 1121.0, + 860.0, + 1095.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 790.0, + 1296.0, + 790.0, + 1296.0, + 941.0, + 1270.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 790.0, + 1317.0, + 790.0, + 1317.0, + 877.0, + 1287.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 790.0, + 1383.0, + 790.0, + 1383.0, + 910.0, + 1354.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 791.0, + 995.0, + 791.0, + 995.0, + 891.0, + 969.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 791.0, + 1245.0, + 791.0, + 1245.0, + 938.0, + 1216.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 792.0, + 967.0, + 792.0, + 967.0, + 877.0, + 941.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 792.0, + 1012.0, + 792.0, + 1012.0, + 880.0, + 985.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 792.0, + 1067.0, + 792.0, + 1067.0, + 886.0, + 1040.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 792.0, + 1095.0, + 792.0, + 1095.0, + 867.0, + 1069.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 792.0, + 1197.0, + 792.0, + 1197.0, + 901.0, + 1170.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 792.0, + 1264.0, + 792.0, + 1264.0, + 871.0, + 1238.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 792.0, + 1332.0, + 792.0, + 1332.0, + 875.0, + 1306.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 793.0, + 1044.0, + 793.0, + 1044.0, + 905.0, + 1018.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 793.0, + 1170.0, + 793.0, + 1170.0, + 888.0, + 1144.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 794.0, + 1219.0, + 794.0, + 1219.0, + 932.0, + 1194.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 795.0, + 1278.0, + 795.0, + 1278.0, + 867.0, + 1255.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 797.0, + 1026.0, + 797.0, + 1026.0, + 872.0, + 1003.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1337.0, + 816.0, + 1363.0, + 816.0, + 1363.0, + 889.0, + 1337.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1575.0, + 1406.0, + 1575.0, + 1406.0, + 1607.0, + 296.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1604.0, + 1407.0, + 1604.0, + 1407.0, + 1637.0, + 292.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1634.0, + 1407.0, + 1634.0, + 1407.0, + 1671.0, + 292.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1667.0, + 1405.0, + 1667.0, + 1405.0, + 1699.0, + 295.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1693.0, + 1215.0, + 1693.0, + 1215.0, + 1729.0, + 295.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1349.0, + 1693.0, + 1405.0, + 1693.0, + 1405.0, + 1729.0, + 1349.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 1724.0, + 964.0, + 1724.0, + 964.0, + 1760.0, + 383.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 1724.0, + 1187.0, + 1724.0, + 1187.0, + 1760.0, + 1138.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 1724.0, + 1407.0, + 1724.0, + 1407.0, + 1760.0, + 1355.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1755.0, + 1406.0, + 1755.0, + 1406.0, + 1793.0, + 291.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1789.0, + 1403.0, + 1789.0, + 1403.0, + 1821.0, + 295.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1817.0, + 1406.0, + 1817.0, + 1406.0, + 1851.0, + 292.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1849.0, + 1236.0, + 1849.0, + 1236.0, + 1882.0, + 292.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 231.0, + 1405.0, + 231.0, + 1405.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 259.0, + 1405.0, + 259.0, + 1405.0, + 298.0, + 291.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 292.0, + 1404.0, + 292.0, + 1404.0, + 327.0, + 295.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 322.0, + 1405.0, + 322.0, + 1405.0, + 356.0, + 292.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 352.0, + 1404.0, + 352.0, + 1404.0, + 387.0, + 294.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 382.0, + 1405.0, + 382.0, + 1405.0, + 420.0, + 292.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 413.0, + 515.0, + 413.0, + 515.0, + 449.0, + 295.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1895.0, + 1404.0, + 1895.0, + 1404.0, + 1928.0, + 295.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1920.0, + 1301.0, + 1920.0, + 1301.0, + 1960.0, + 388.0, + 1960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1920.0, + 1409.0, + 1920.0, + 1409.0, + 1960.0, + 1399.0, + 1960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1955.0, + 1408.0, + 1955.0, + 1408.0, + 1992.0, + 294.0, + 1992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1985.0, + 1406.0, + 1985.0, + 1406.0, + 2021.0, + 293.0, + 2021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2013.0, + 770.0, + 2013.0, + 770.0, + 2053.0, + 294.0, + 2053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 459.0, + 1408.0, + 459.0, + 1408.0, + 495.0, + 294.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 487.0, + 1027.0, + 487.0, + 1027.0, + 526.0, + 292.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 487.0, + 1407.0, + 487.0, + 1407.0, + 526.0, + 1147.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 519.0, + 574.0, + 519.0, + 574.0, + 559.0, + 292.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 519.0, + 1406.0, + 519.0, + 1406.0, + 559.0, + 627.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 553.0, + 636.0, + 553.0, + 636.0, + 586.0, + 295.0, + 586.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1247, + 1405, + 1247, + 1405, + 1524, + 296, + 1524 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 551, + 1405, + 551, + 1405, + 888, + 297, + 888 + ], + "score": 0.98 + }, + { + "category_id": 5, + "poly": [ + 314, + 914, + 1380, + 914, + 1380, + 1102, + 314, + 1102 + ], + "score": 0.979, + "html": "
API name Attacker/Evaluation MetricFace++AliYun
T-FPR=10 -3T-FPR=10 -4T-FPR=10 -3T-FPR=10 -4
Dong et al. (2018)30.7721.0318.006.50
Xie et al. (2019)37.9525.6421.5011.00
CW (G-FPR<10-4)41.6224.3719.0012.00
SemanticAdv (G-FPR <10-4)64.6342.6935.5022.17
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1912, + 1405, + 1912, + 1405, + 2034, + 298, + 2034 + ], + "score": 0.971 + }, + { + "category_id": 3, + "poly": [ + 353, + 1554, + 1341, + 1554, + 1341, + 1820, + 353, + 1820 + ], + "score": 0.967 + }, + { + "category_id": 5, + "poly": [ + 310, + 224, + 960, + 224, + 960, + 345, + 310, + 345 + ], + "score": 0.963, + "html": "
MtestMoptR-50-SR-101-SR-50-CR-101-C
R-50-S1.000 (1.000)0.108 (0.007)0.023(0.002)0.018 (0.002)
R-101-S0.169 (0.006)1.000 (1.000)0.030 (0.002)0.032(0.003)
R-50-C0.166 (0.019)0.202(0.025)1.000 (1.000)0.048(0.007)
R-101-C0.120 (0.015)0.236 (0.029)0.040 (0.006)1.000(1.000)
" + }, + { + "category_id": 4, + "poly": [ + 295, + 1839, + 1400, + 1839, + 1400, + 1901, + 295, + 1901 + ], + "score": 0.945 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.905 + }, + { + "category_id": 5, + "poly": [ + 974, + 224, + 1364, + 224, + 1364, + 345, + 974, + 345 + ], + "score": 0.797, + "html": "
Mtest/MoptR-50-SR-101-S
R-50-S1.000 (1.000)0.862 (0.530)
R-101-S0.874 (0.422)1.000 (1.000)
R-50-C0.693(0.347)0.837 (0.579)
R-101-C0.617 (0.218)0.888 (0.617)
" + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.792 + }, + { + "category_id": 1, + "poly": [ + 299, + 1113, + 1403, + 1113, + 1403, + 1238, + 299, + 1238 + ], + "score": 0.71 + }, + { + "category_id": 6, + "poly": [ + 297, + 400, + 1405, + 400, + 1405, + 524, + 297, + 524 + ], + "score": 0.675 + }, + { + "category_id": 1, + "poly": [ + 297, + 400, + 1405, + 400, + 1405, + 524, + 297, + 524 + ], + "score": 0.394 + }, + { + "category_id": 6, + "poly": [ + 299, + 1113, + 1403, + 1113, + 1403, + 1238, + 299, + 1238 + ], + "score": 0.368 + }, + { + "category_id": 1, + "poly": [ + 616, + 358, + 650, + 358, + 650, + 385, + 616, + 385 + ], + "score": 0.169 + }, + { + "category_id": 13, + "poly": [ + 862, + 1401, + 1037, + 1401, + 1037, + 1431, + 862, + 1431 + ], + "score": 0.92, + "latex": "3 0 . 2 7 \\pm 1 . 9 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 795, + 402, + 853, + 402, + 853, + 434, + 795, + 434 + ], + "score": 0.91, + "latex": "( i , j )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1370, + 460, + 1370, + 460, + 1401, + 297, + 1401 + ], + "score": 0.91, + "latex": "3 9 . 1 \\dot { 4 } \\pm 1 . 9 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 833, + 675, + 868, + 675, + 868, + 707, + 833, + 707 + ], + "score": 0.9, + "latex": "{ \\mathcal { L } } _ { p }" + }, + { + "category_id": 13, + "poly": [ + 396, + 795, + 487, + 795, + 487, + 825, + 396, + 825 + ], + "score": 0.89, + "latex": "6 4 . 6 3 \\%" + }, + { + "category_id": 13, + "poly": [ + 296, + 491, + 357, + 491, + 357, + 523, + 296, + 523 + ], + "score": 0.89, + "latex": "1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 713, + 1370, + 768, + 1370, + 768, + 1401, + 713, + 1401 + ], + "score": 0.88, + "latex": "5 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 973, + 795, + 1143, + 795, + 1143, + 825, + 973, + 825 + ], + "score": 0.87, + "latex": "\\mathrm { T } \\mathrm { \\bar { F } P R } = 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 1038, + 1177, + 1107, + 1177, + 1107, + 1205, + 1038, + 1205 + ], + "score": 0.86, + "latex": "\\epsilon = 8" + }, + { + "category_id": 13, + "poly": [ + 473, + 1207, + 541, + 1207, + 541, + 1235, + 473, + 1235 + ], + "score": 0.84, + "latex": "\\epsilon = 8" + }, + { + "category_id": 13, + "poly": [ + 382, + 1491, + 471, + 1491, + 471, + 1522, + 382, + 1522 + ], + "score": 0.84, + "latex": "< 1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 497, + 435, + 513, + 435, + 513, + 464, + 497, + 464 + ], + "score": 0.82, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 1310, + 764, + 1401, + 764, + 1401, + 795, + 1310, + 795 + ], + "score": 0.82, + "latex": "< 1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 1185, + 581, + 1396, + 581, + 1396, + 615, + 1185, + 615 + ], + "score": 0.82, + "latex": "< 1 0 ^ { - 4 } ( \\kappa = 0 . 2 0 )" + }, + { + "category_id": 13, + "poly": [ + 738, + 582, + 1037, + 582, + 1037, + 615, + 738, + 615 + ], + "score": 0.8, + "latex": "\\mathbf { G } \\mathbf { - } \\mathbf { F P R } \\mathbf { \\bar { = } } 1 0 ^ { - 4 } ( \\kappa = \\mathbf { \\bar { 0 } } . 6 0 )" + }, + { + "category_id": 13, + "poly": [ + 408, + 492, + 569, + 492, + 569, + 522, + 408, + 522 + ], + "score": 0.77, + "latex": "\\mathrm { T } \\mathrm { \\mathrm { - F P R } } = 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 427, + 581, + 726, + 581, + 726, + 616, + 427, + 616 + ], + "score": 0.77, + "latex": "\\mathrm { G } \\mathrm { - } \\mathrm { F P R } = 1 0 ^ { - 3 } ( \\kappa = 1 . 2 4 )" + }, + { + "category_id": 13, + "poly": [ + 1009, + 1147, + 1051, + 1147, + 1051, + 1176, + 1009, + 1176 + ], + "score": 0.75, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 803, + 436, + 815, + 436, + 815, + 460, + 803, + 460 + ], + "score": 0.7, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1379, + 465, + 1404, + 465, + 1404, + 490, + 1379, + 490 + ], + "score": 0.68, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 464, + 1309, + 637, + 1309, + 637, + 1339, + 464, + 1339 + ], + "score": 0.67, + "latex": "\\mathrm { G } \\mathrm { - } \\mathrm { F P R } < 1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 857, + 461, + 943, + 461, + 943, + 491, + 857, + 491 + ], + "score": 0.52, + "latex": "= 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 1133, + 616, + 1219, + 616, + 1219, + 642, + 1133, + 642 + ], + "score": 0.48, + "latex": "{ \\mathrm { F a c e } } + +" + }, + { + "category_id": 13, + "poly": [ + 643, + 461, + 728, + 461, + 728, + 491, + 643, + 491 + ], + "score": 0.48, + "latex": "= 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 296, + 1204, + 375, + 1204, + 375, + 1235, + 296, + 1235 + ], + "score": 0.34, + "latex": "\\mathrm { \\dot { \\mathbf { M } } \\mathbf { - } D I ^ { 2 } }" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1558.0, + 765.0, + 1558.0, + 765.0, + 1576.0, + 389.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 1554.0, + 940.0, + 1554.0, + 940.0, + 1577.0, + 790.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 1558.0, + 1137.0, + 1558.0, + 1137.0, + 1576.0, + 982.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1576.0, + 397.0, + 1576.0, + 397.0, + 1597.0, + 369.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1577.0, + 592.0, + 1577.0, + 592.0, + 1594.0, + 564.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 1577.0, + 784.0, + 1577.0, + 784.0, + 1594.0, + 756.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1577.0, + 976.0, + 1577.0, + 976.0, + 1594.0, + 951.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1616.0, + 377.0, + 1616.0, + 377.0, + 1746.0, + 352.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1616.0, + 589.0, + 1616.0, + 589.0, + 1632.0, + 564.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 1614.0, + 784.0, + 1614.0, + 784.0, + 1634.0, + 756.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1618.0, + 974.0, + 1618.0, + 974.0, + 1631.0, + 952.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 1628.0, + 442.0, + 1628.0, + 442.0, + 1642.0, + 427.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1629.0, + 1312.0, + 1629.0, + 1312.0, + 1647.0, + 1178.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1644.0, + 1309.0, + 1644.0, + 1309.0, + 1665.0, + 1178.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1653.0, + 397.0, + 1653.0, + 397.0, + 1673.0, + 368.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 1656.0, + 587.0, + 1656.0, + 587.0, + 1669.0, + 565.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 1655.0, + 784.0, + 1655.0, + 784.0, + 1672.0, + 756.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 1653.0, + 978.0, + 1653.0, + 978.0, + 1673.0, + 950.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1658.0, + 1334.0, + 1658.0, + 1334.0, + 1680.0, + 1177.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1677.0, + 1257.0, + 1677.0, + 1257.0, + 1698.0, + 1177.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1694.0, + 395.0, + 1694.0, + 395.0, + 1711.0, + 370.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1693.0, + 589.0, + 1693.0, + 589.0, + 1710.0, + 564.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 1693.0, + 784.0, + 1693.0, + 784.0, + 1710.0, + 756.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1693.0, + 983.0, + 1693.0, + 983.0, + 1709.0, + 951.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1693.0, + 1254.0, + 1693.0, + 1254.0, + 1716.0, + 1177.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 1717.0, + 1165.0, + 1717.0, + 1165.0, + 1726.0, + 1155.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1712.0, + 1278.0, + 1712.0, + 1278.0, + 1730.0, + 1178.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 1725.0, + 623.0, + 1725.0, + 623.0, + 1735.0, + 614.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1729.0, + 396.0, + 1729.0, + 396.0, + 1750.0, + 368.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1730.0, + 589.0, + 1730.0, + 589.0, + 1748.0, + 564.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 1733.0, + 782.0, + 1733.0, + 782.0, + 1749.0, + 756.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1729.0, + 976.0, + 1729.0, + 976.0, + 1746.0, + 951.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1768.0, + 404.0, + 1768.0, + 404.0, + 1789.0, + 369.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 1751.0, + 612.0, + 1751.0, + 612.0, + 1788.0, + 561.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 1770.0, + 785.0, + 1770.0, + 785.0, + 1788.0, + 756.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1767.0, + 978.0, + 1767.0, + 978.0, + 1784.0, + 951.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1783.0, + 422.0, + 1783.0, + 422.0, + 1804.0, + 387.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 1784.0, + 505.0, + 1784.0, + 505.0, + 1804.0, + 453.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 1785.0, + 569.0, + 1785.0, + 569.0, + 1802.0, + 533.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 1785.0, + 615.0, + 1785.0, + 615.0, + 1801.0, + 580.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 1785.0, + 698.0, + 1785.0, + 698.0, + 1801.0, + 645.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 1785.0, + 810.0, + 1785.0, + 810.0, + 1801.0, + 773.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 1784.0, + 892.0, + 1784.0, + 892.0, + 1801.0, + 838.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1785.0, + 956.0, + 1785.0, + 956.0, + 1801.0, + 920.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1785.0, + 1002.0, + 1785.0, + 1002.0, + 1802.0, + 967.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 1784.0, + 1086.0, + 1784.0, + 1086.0, + 1803.0, + 1033.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1800.0, + 535.0, + 1800.0, + 535.0, + 1819.0, + 421.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 1800.0, + 729.0, + 1800.0, + 729.0, + 1819.0, + 615.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 1800.0, + 924.0, + 1800.0, + 924.0, + 1819.0, + 808.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 1800.0, + 1117.0, + 1800.0, + 1117.0, + 1819.0, + 1002.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1617.5, + 397.0, + 1617.5, + 397.0, + 1633.0, + 366.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 1614.0, + 1075.0, + 1614.0, + 1075.0, + 1641.5, + 987.0, + 1641.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.75, + 1636.0, + 842.75, + 1636.0, + 842.75, + 1651.0, + 794.75, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 1618.5, + 701.0, + 1618.5, + 701.0, + 1660.5, + 567.0, + 1660.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.75, + 1717.5, + 471.75, + 1717.5, + 471.75, + 1731.0, + 382.75, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 1783.5, + 765.0, + 1783.5, + 765.0, + 1803.5, + 726.0, + 1803.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 1688.0, + 888.0, + 1688.0, + 888.0, + 1715.0, + 745.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.75, + 1782.5, + 1151.75, + 1782.5, + 1151.75, + 1803.5, + 1113.75, + 1803.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1836.0, + 1402.0, + 1836.0, + 1402.0, + 1871.0, + 295.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1867.0, + 1302.0, + 1867.0, + 1302.0, + 1903.0, + 294.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 400.0, + 794.0, + 400.0, + 794.0, + 436.0, + 294.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 400.0, + 1406.0, + 400.0, + 1406.0, + 436.0, + 854.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 431.0, + 496.0, + 431.0, + 496.0, + 467.0, + 292.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 431.0, + 802.0, + 431.0, + 802.0, + 467.0, + 514.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 431.0, + 1410.0, + 431.0, + 1410.0, + 467.0, + 816.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 461.0, + 642.0, + 461.0, + 642.0, + 497.0, + 294.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 461.0, + 856.0, + 461.0, + 856.0, + 497.0, + 729.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 461.0, + 1378.0, + 461.0, + 1378.0, + 497.0, + 944.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 489.0, + 295.0, + 489.0, + 295.0, + 524.0, + 291.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 489.0, + 407.0, + 489.0, + 407.0, + 524.0, + 358.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 489.0, + 583.0, + 489.0, + 583.0, + 524.0, + 570.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1113.0, + 1405.0, + 1113.0, + 1405.0, + 1150.0, + 293.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1145.0, + 1008.0, + 1145.0, + 1008.0, + 1178.0, + 295.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 1145.0, + 1404.0, + 1145.0, + 1404.0, + 1178.0, + 1052.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1177.0, + 1037.0, + 1177.0, + 1037.0, + 1209.0, + 294.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 1177.0, + 1404.0, + 1177.0, + 1404.0, + 1209.0, + 1108.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 1202.0, + 472.0, + 1202.0, + 472.0, + 1242.0, + 376.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 1202.0, + 919.0, + 1202.0, + 919.0, + 1242.0, + 542.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1249.0, + 1407.0, + 1249.0, + 1407.0, + 1282.0, + 298.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1279.0, + 1403.0, + 1279.0, + 1403.0, + 1312.0, + 295.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1304.0, + 463.0, + 1304.0, + 463.0, + 1348.0, + 290.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1304.0, + 1410.0, + 1304.0, + 1410.0, + 1348.0, + 638.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1337.0, + 1407.0, + 1337.0, + 1407.0, + 1376.0, + 291.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1368.0, + 296.0, + 1368.0, + 296.0, + 1409.0, + 293.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 1368.0, + 712.0, + 1368.0, + 712.0, + 1409.0, + 461.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 1368.0, + 1407.0, + 1368.0, + 1407.0, + 1409.0, + 769.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1400.0, + 861.0, + 1400.0, + 861.0, + 1438.0, + 293.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 1400.0, + 1405.0, + 1400.0, + 1405.0, + 1438.0, + 1038.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1429.0, + 1406.0, + 1429.0, + 1406.0, + 1471.0, + 291.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1460.0, + 1405.0, + 1460.0, + 1405.0, + 1497.0, + 293.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1489.0, + 381.0, + 1489.0, + 381.0, + 1530.0, + 294.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 1489.0, + 1197.0, + 1489.0, + 1197.0, + 1530.0, + 472.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 552.0, + 1408.0, + 552.0, + 1408.0, + 587.0, + 296.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 579.0, + 426.0, + 579.0, + 426.0, + 618.0, + 294.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 579.0, + 737.0, + 579.0, + 737.0, + 618.0, + 727.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 579.0, + 1184.0, + 579.0, + 1184.0, + 618.0, + 1038.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 579.0, + 1408.0, + 579.0, + 1408.0, + 618.0, + 1397.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 612.0, + 1132.0, + 612.0, + 1132.0, + 647.0, + 292.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 612.0, + 1407.0, + 612.0, + 1407.0, + 647.0, + 1220.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 644.0, + 1406.0, + 644.0, + 1406.0, + 679.0, + 295.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 672.0, + 832.0, + 672.0, + 832.0, + 711.0, + 292.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 672.0, + 1410.0, + 672.0, + 1410.0, + 711.0, + 869.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 705.0, + 1407.0, + 705.0, + 1407.0, + 739.0, + 295.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 734.0, + 1405.0, + 734.0, + 1405.0, + 767.0, + 294.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 762.0, + 1309.0, + 762.0, + 1309.0, + 801.0, + 291.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 762.0, + 1408.0, + 762.0, + 1408.0, + 801.0, + 1402.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 794.0, + 395.0, + 794.0, + 395.0, + 829.0, + 295.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 794.0, + 972.0, + 794.0, + 972.0, + 829.0, + 488.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 794.0, + 1405.0, + 794.0, + 1405.0, + 829.0, + 1144.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 826.0, + 1406.0, + 826.0, + 1406.0, + 857.0, + 296.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 855.0, + 456.0, + 855.0, + 456.0, + 891.0, + 292.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1911.0, + 1407.0, + 1911.0, + 1407.0, + 1945.0, + 295.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1942.0, + 1408.0, + 1942.0, + 1408.0, + 1976.0, + 292.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1974.0, + 1405.0, + 1974.0, + 1405.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 1403.0, + 2001.0, + 1403.0, + 2038.0, + 293.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1113.0, + 1405.0, + 1113.0, + 1405.0, + 1150.0, + 293.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1145.0, + 1008.0, + 1145.0, + 1008.0, + 1178.0, + 295.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 1145.0, + 1404.0, + 1145.0, + 1404.0, + 1178.0, + 1052.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1177.0, + 1037.0, + 1177.0, + 1037.0, + 1209.0, + 294.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 1177.0, + 1404.0, + 1177.0, + 1404.0, + 1209.0, + 1108.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 1202.0, + 472.0, + 1202.0, + 472.0, + 1242.0, + 376.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 1202.0, + 919.0, + 1202.0, + 919.0, + 1242.0, + 542.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 400.0, + 794.0, + 400.0, + 794.0, + 436.0, + 294.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 400.0, + 1406.0, + 400.0, + 1406.0, + 436.0, + 854.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 431.0, + 496.0, + 431.0, + 496.0, + 467.0, + 292.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 431.0, + 802.0, + 431.0, + 802.0, + 467.0, + 514.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 431.0, + 1410.0, + 431.0, + 1410.0, + 467.0, + 816.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 461.0, + 642.0, + 461.0, + 642.0, + 497.0, + 294.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 461.0, + 856.0, + 461.0, + 856.0, + 497.0, + 729.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 461.0, + 1378.0, + 461.0, + 1378.0, + 497.0, + 944.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 489.0, + 295.0, + 489.0, + 295.0, + 524.0, + 291.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 489.0, + 407.0, + 489.0, + 407.0, + 524.0, + 358.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 489.0, + 583.0, + 489.0, + 583.0, + 524.0, + 570.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 356.0, + 654.0, + 356.0, + 654.0, + 392.0, + 613.0, + 392.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1819, + 1403, + 1819, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1539, + 1404, + 1539, + 1404, + 1724, + 298, + 1724 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1259, + 1405, + 1259, + 1405, + 1442, + 298, + 1442 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1089, + 1403, + 1089, + 1403, + 1243, + 298, + 1243 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 308, + 619, + 1387, + 619, + 1387, + 987, + 308, + 987 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 301, + 224, + 1388, + 224, + 1388, + 509, + 301, + 509 + ], + "score": 0.965 + }, + { + "category_id": 4, + "poly": [ + 383, + 1007, + 1313, + 1007, + 1313, + 1039, + 383, + 1039 + ], + "score": 0.926 + }, + { + "category_id": 4, + "poly": [ + 441, + 559, + 1255, + 559, + 1255, + 591, + 441, + 591 + ], + "score": 0.92 + }, + { + "category_id": 0, + "poly": [ + 298, + 1766, + 1053, + 1766, + 1053, + 1798, + 298, + 1798 + ], + "score": 0.904 + }, + { + "category_id": 0, + "poly": [ + 299, + 1485, + 920, + 1485, + 920, + 1517, + 299, + 1517 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 858, + 76, + 858, + 104, + 298, + 104 + ], + "score": 0.84 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.79 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 856, + 76, + 856, + 104, + 300, + 104 + ], + "score": 0.094 + }, + { + "category_id": 13, + "poly": [ + 625, + 1349, + 685, + 1349, + 685, + 1379, + 625, + 1379 + ], + "score": 0.9, + "latex": "1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 1277, + 1289, + 1336, + 1289, + 1336, + 1320, + 1277, + 1320 + ], + "score": 0.89, + "latex": "\\bar { 1 } 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 757, + 1182, + 812, + 1182, + 812, + 1211, + 757, + 1211 + ], + "score": 0.87, + "latex": "1 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 749, + 1213, + 803, + 1213, + 803, + 1242, + 749, + 1242 + ], + "score": 0.87, + "latex": "1 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 940, + 1601, + 983, + 1601, + 983, + 1630, + 940, + 1630 + ], + "score": 0.84, + "latex": "9 0 ^ { \\circ }" + }, + { + "category_id": 13, + "poly": [ + 1361, + 1182, + 1403, + 1182, + 1403, + 1212, + 1361, + 1212 + ], + "score": 0.84, + "latex": "8 \\%" + }, + { + "category_id": 13, + "poly": [ + 414, + 807, + 433, + 807, + 433, + 822, + 414, + 822 + ], + "score": 0.37, + "latex": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 620.0, + 560.0, + 620.0, + 560.0, + 643.0, + 325.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 620.0, + 1102.0, + 620.0, + 1102.0, + 640.0, + 838.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 634.0, + 1330.0, + 634.0, + 1330.0, + 653.0, + 1207.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 649.0, + 528.0, + 649.0, + 528.0, + 669.0, + 442.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.0, + 647.0, + 653.0, + 647.0, + 653.0, + 666.0, + 566.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 647.0, + 774.0, + 647.0, + 774.0, + 669.0, + 690.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 648.0, + 1003.0, + 648.0, + 1003.0, + 668.0, + 926.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 648.0, + 1140.0, + 648.0, + 1140.0, + 668.0, + 1023.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 646.0, + 1350.0, + 646.0, + 1350.0, + 669.0, + 1186.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 657.0, + 414.0, + 657.0, + 414.0, + 680.0, + 332.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 665.0, + 520.0, + 665.0, + 520.0, + 681.0, + 450.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.0, + 660.0, + 652.0, + 660.0, + 652.0, + 680.0, + 566.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 661.0, + 775.0, + 661.0, + 775.0, + 681.0, + 689.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 659.0, + 904.0, + 659.0, + 904.0, + 679.0, + 790.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 662.0, + 1019.0, + 662.0, + 1019.0, + 681.0, + 908.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 662.0, + 1125.0, + 662.0, + 1125.0, + 681.0, + 1038.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 803.0, + 413.0, + 803.0, + 413.0, + 826.0, + 324.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 803.0, + 563.0, + 803.0, + 563.0, + 826.0, + 434.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 808.0, + 1104.0, + 808.0, + 1104.0, + 827.0, + 839.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 823.0, + 1330.0, + 823.0, + 1330.0, + 842.0, + 1207.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 841.0, + 411.0, + 841.0, + 411.0, + 864.0, + 327.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 835.0, + 525.0, + 835.0, + 525.0, + 868.0, + 439.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 835.0, + 647.0, + 835.0, + 647.0, + 868.0, + 560.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 835.0, + 769.0, + 835.0, + 769.0, + 868.0, + 683.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 847.0, + 905.0, + 847.0, + 905.0, + 866.0, + 791.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 835.0, + 1023.0, + 835.0, + 1023.0, + 868.0, + 909.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 837.0, + 1140.0, + 837.0, + 1140.0, + 868.0, + 1024.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 835.0, + 1350.0, + 835.0, + 1350.0, + 858.0, + 1188.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 877.0, + 333.0, + 877.0, + 333.0, + 887.0, + 321.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 874.0, + 812.0, + 874.0, + 812.0, + 890.0, + 795.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 875.0, + 1170.0, + 875.0, + 1170.0, + 890.0, + 1153.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 228.0, + 488.0, + 228.0, + 488.0, + 252.0, + 301.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 232.0, + 586.0, + 232.0, + 586.0, + 262.0, + 502.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 232.0, + 701.0, + 232.0, + 701.0, + 262.0, + 599.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 232.0, + 801.0, + 232.0, + 801.0, + 259.0, + 720.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 232.0, + 916.0, + 232.0, + 916.0, + 262.0, + 814.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 232.0, + 1049.0, + 232.0, + 1049.0, + 263.0, + 966.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 232.0, + 1165.0, + 232.0, + 1165.0, + 263.0, + 1062.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 234.0, + 1265.0, + 234.0, + 1265.0, + 263.0, + 1181.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 234.0, + 1380.0, + 234.0, + 1380.0, + 260.0, + 1277.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 296.0, + 1002.0, + 296.0, + 1002.0, + 305.0, + 989.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 348.0, + 487.0, + 348.0, + 487.0, + 371.0, + 300.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 349.0, + 894.0, + 349.0, + 894.0, + 366.0, + 837.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 373.0, + 411.0, + 373.0, + 411.0, + 404.0, + 326.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 371.0, + 525.0, + 371.0, + 525.0, + 404.0, + 423.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 376.0, + 621.0, + 376.0, + 621.0, + 405.0, + 538.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 375.0, + 732.0, + 375.0, + 732.0, + 405.0, + 648.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 376.0, + 845.0, + 376.0, + 845.0, + 405.0, + 742.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 372.0, + 953.0, + 372.0, + 953.0, + 405.0, + 872.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 373.0, + 1070.0, + 373.0, + 1070.0, + 406.0, + 967.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 377.0, + 1165.0, + 377.0, + 1165.0, + 406.0, + 1084.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 377.0, + 1275.0, + 377.0, + 1275.0, + 406.0, + 1192.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 377.0, + 1389.0, + 377.0, + 1389.0, + 406.0, + 1286.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 446.0, + 1134.0, + 446.0, + 1134.0, + 464.0, + 1119.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 489.0, + 821.0, + 489.0, + 821.0, + 505.0, + 768.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 347.5, + 1355.0, + 347.5, + 1355.0, + 364.5, + 1316.0, + 364.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1323.0, + 490.5, + 1360.0, + 490.5, + 1360.0, + 506.0, + 1323.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 1004.0, + 1317.0, + 1004.0, + 1317.0, + 1043.0, + 380.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 555.0, + 1259.0, + 555.0, + 1259.0, + 596.0, + 439.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1764.0, + 1057.0, + 1764.0, + 1057.0, + 1802.0, + 295.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1485.0, + 923.0, + 1485.0, + 923.0, + 1521.0, + 295.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 70.0, + 859.0, + 70.0, + 859.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1816.0, + 1404.0, + 1816.0, + 1404.0, + 1859.0, + 293.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1851.0, + 1405.0, + 1851.0, + 1405.0, + 1886.0, + 294.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1970.0, + 1407.0, + 1970.0, + 1407.0, + 2010.0, + 292.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2005.0, + 1326.0, + 2005.0, + 1326.0, + 2036.0, + 296.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1540.0, + 1405.0, + 1540.0, + 1405.0, + 1572.0, + 297.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1572.0, + 1402.0, + 1572.0, + 1402.0, + 1604.0, + 296.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1600.0, + 939.0, + 1600.0, + 939.0, + 1635.0, + 294.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 1600.0, + 1406.0, + 1600.0, + 1406.0, + 1635.0, + 984.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1631.0, + 1406.0, + 1631.0, + 1406.0, + 1667.0, + 293.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1661.0, + 1405.0, + 1661.0, + 1405.0, + 1695.0, + 293.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1693.0, + 833.0, + 1693.0, + 833.0, + 1726.0, + 294.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1260.0, + 1407.0, + 1260.0, + 1407.0, + 1291.0, + 296.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1289.0, + 1276.0, + 1289.0, + 1276.0, + 1323.0, + 293.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1337.0, + 1289.0, + 1405.0, + 1289.0, + 1405.0, + 1323.0, + 1337.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1318.0, + 1405.0, + 1318.0, + 1405.0, + 1356.0, + 292.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1347.0, + 624.0, + 1347.0, + 624.0, + 1386.0, + 293.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1347.0, + 1406.0, + 1347.0, + 1406.0, + 1386.0, + 686.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1381.0, + 1405.0, + 1381.0, + 1405.0, + 1415.0, + 295.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1411.0, + 1116.0, + 1411.0, + 1116.0, + 1443.0, + 296.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1091.0, + 1405.0, + 1091.0, + 1405.0, + 1124.0, + 296.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1121.0, + 1405.0, + 1121.0, + 1405.0, + 1157.0, + 293.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1153.0, + 1404.0, + 1153.0, + 1404.0, + 1186.0, + 296.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1181.0, + 756.0, + 1181.0, + 756.0, + 1218.0, + 294.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1181.0, + 1360.0, + 1181.0, + 1360.0, + 1218.0, + 813.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1211.0, + 748.0, + 1211.0, + 748.0, + 1247.0, + 293.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 1211.0, + 1114.0, + 1211.0, + 1114.0, + 1247.0, + 804.0, + 1247.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 272, + 1405, + 272, + 1405, + 486, + 298, + 486 + ], + "score": 0.972 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.904 + }, + { + "category_id": 0, + "poly": [ + 300, + 226, + 561, + 226, + 561, + 261, + 300, + 261 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 299, + 533, + 488, + 533, + 488, + 567, + 299, + 567 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.823 + }, + { + "category_id": 1, + "poly": [ + 292, + 577, + 1409, + 577, + 1409, + 2054, + 292, + 2054 + ], + "score": 0.614 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 219.0, + 566.0, + 219.0, + 566.0, + 270.0, + 290.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 533.0, + 490.0, + 533.0, + 490.0, + 570.0, + 296.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 868.0, + 2084.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 272.0, + 1405.0, + 272.0, + 1405.0, + 307.0, + 296.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 304.0, + 1407.0, + 304.0, + 1407.0, + 339.0, + 295.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 335.0, + 1402.0, + 335.0, + 1402.0, + 366.0, + 296.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 364.0, + 1405.0, + 364.0, + 1405.0, + 399.0, + 295.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 391.0, + 1406.0, + 391.0, + 1406.0, + 432.0, + 292.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 423.0, + 1406.0, + 423.0, + 1406.0, + 461.0, + 293.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 456.0, + 610.0, + 456.0, + 610.0, + 492.0, + 295.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 584.0, + 1404.0, + 584.0, + 1404.0, + 624.0, + 293.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 614.0, + 505.0, + 614.0, + 505.0, + 654.0, + 322.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 668.0, + 1406.0, + 668.0, + 1406.0, + 708.0, + 292.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 699.0, + 456.0, + 699.0, + 456.0, + 736.0, + 324.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 750.0, + 1408.0, + 750.0, + 1408.0, + 795.0, + 293.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 783.0, + 576.0, + 783.0, + 576.0, + 819.0, + 321.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 837.0, + 1403.0, + 837.0, + 1403.0, + 874.0, + 295.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 869.0, + 1363.0, + 869.0, + 1363.0, + 905.0, + 321.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 920.0, + 1408.0, + 920.0, + 1408.0, + 961.0, + 293.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 953.0, + 692.0, + 953.0, + 692.0, + 989.0, + 320.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1005.0, + 1408.0, + 1005.0, + 1408.0, + 1041.0, + 297.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1035.0, + 1254.0, + 1035.0, + 1254.0, + 1076.0, + 321.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1087.0, + 1408.0, + 1087.0, + 1408.0, + 1128.0, + 293.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1117.0, + 1406.0, + 1117.0, + 1406.0, + 1158.0, + 321.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1151.0, + 955.0, + 1151.0, + 955.0, + 1187.0, + 325.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1199.0, + 1408.0, + 1199.0, + 1408.0, + 1243.0, + 292.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1232.0, + 1150.0, + 1232.0, + 1150.0, + 1271.0, + 320.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1288.0, + 1406.0, + 1288.0, + 1406.0, + 1324.0, + 295.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1319.0, + 1026.0, + 1319.0, + 1026.0, + 1355.0, + 321.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1370.0, + 1409.0, + 1370.0, + 1409.0, + 1411.0, + 295.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1398.0, + 1409.0, + 1398.0, + 1409.0, + 1440.0, + 318.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1429.0, + 399.0, + 1429.0, + 399.0, + 1468.0, + 320.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1481.0, + 1406.0, + 1481.0, + 1406.0, + 1524.0, + 290.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1516.0, + 713.0, + 1516.0, + 713.0, + 1552.0, + 321.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1563.0, + 1406.0, + 1563.0, + 1406.0, + 1611.0, + 292.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1596.0, + 1408.0, + 1596.0, + 1408.0, + 1641.0, + 318.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1629.0, + 829.0, + 1629.0, + 829.0, + 1665.0, + 320.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1678.0, + 1404.0, + 1678.0, + 1404.0, + 1723.0, + 292.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1715.0, + 1406.0, + 1715.0, + 1406.0, + 1751.0, + 321.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1739.0, + 1408.0, + 1739.0, + 1408.0, + 1787.0, + 316.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1774.0, + 716.0, + 1774.0, + 716.0, + 1813.0, + 320.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1826.0, + 1409.0, + 1826.0, + 1409.0, + 1867.0, + 293.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1858.0, + 1408.0, + 1858.0, + 1408.0, + 1899.0, + 321.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1887.0, + 1080.0, + 1887.0, + 1080.0, + 1928.0, + 320.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1938.0, + 1406.0, + 1938.0, + 1406.0, + 1984.0, + 293.0, + 1984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1973.0, + 1408.0, + 1973.0, + 1408.0, + 2014.0, + 320.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 2002.0, + 852.0, + 2002.0, + 852.0, + 2043.0, + 318.0, + 2043.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 861, + 2088, + 861, + 2113, + 835, + 2113 + ], + "score": 0.821 + }, + { + "category_id": 1, + "poly": [ + 295, + 1197, + 1403, + 1197, + 1403, + 1262, + 295, + 1262 + ], + "score": 0.806 + }, + { + "category_id": 1, + "poly": [ + 297, + 1558, + 1378, + 1558, + 1378, + 1592, + 297, + 1592 + ], + "score": 0.804 + }, + { + "category_id": 1, + "poly": [ + 300, + 1114, + 1398, + 1114, + 1398, + 1180, + 300, + 1180 + ], + "score": 0.777 + }, + { + "category_id": 1, + "poly": [ + 297, + 589, + 1402, + 589, + 1402, + 653, + 297, + 653 + ], + "score": 0.765 + }, + { + "category_id": 1, + "poly": [ + 296, + 754, + 1401, + 754, + 1401, + 818, + 296, + 818 + ], + "score": 0.757 + }, + { + "category_id": 1, + "poly": [ + 298, + 1280, + 1405, + 1280, + 1405, + 1373, + 298, + 1373 + ], + "score": 0.749 + }, + { + "category_id": 1, + "poly": [ + 296, + 506, + 1404, + 506, + 1404, + 571, + 296, + 571 + ], + "score": 0.746 + }, + { + "category_id": 1, + "poly": [ + 296, + 671, + 1402, + 671, + 1402, + 738, + 296, + 738 + ], + "score": 0.735 + }, + { + "category_id": 1, + "poly": [ + 296, + 836, + 1401, + 836, + 1401, + 902, + 296, + 902 + ], + "score": 0.727 + }, + { + "category_id": 1, + "poly": [ + 294, + 1971, + 1399, + 1971, + 1399, + 2034, + 294, + 2034 + ], + "score": 0.711 + }, + { + "category_id": 1, + "poly": [ + 299, + 1475, + 1397, + 1475, + 1397, + 1540, + 299, + 1540 + ], + "score": 0.708 + }, + { + "category_id": 1, + "poly": [ + 295, + 1610, + 1278, + 1610, + 1278, + 1644, + 295, + 1644 + ], + "score": 0.703 + }, + { + "category_id": 1, + "poly": [ + 301, + 919, + 1403, + 919, + 1403, + 1015, + 301, + 1015 + ], + "score": 0.698 + }, + { + "category_id": 1, + "poly": [ + 297, + 1393, + 1399, + 1393, + 1399, + 1458, + 297, + 1458 + ], + "score": 0.694 + }, + { + "category_id": 1, + "poly": [ + 291, + 1032, + 1402, + 1032, + 1402, + 1097, + 291, + 1097 + ], + "score": 0.689 + }, + { + "category_id": 1, + "poly": [ + 291, + 424, + 1404, + 424, + 1404, + 489, + 291, + 489 + ], + "score": 0.68 + }, + { + "category_id": 1, + "poly": [ + 292, + 228, + 1402, + 228, + 1402, + 294, + 292, + 294 + ], + "score": 0.679 + }, + { + "category_id": 1, + "poly": [ + 297, + 1858, + 1400, + 1858, + 1400, + 1952, + 297, + 1952 + ], + "score": 0.643 + }, + { + "category_id": 1, + "poly": [ + 302, + 1662, + 1402, + 1662, + 1402, + 1757, + 302, + 1757 + ], + "score": 0.64 + }, + { + "category_id": 1, + "poly": [ + 295, + 1774, + 1403, + 1774, + 1403, + 1840, + 295, + 1840 + ], + "score": 0.634 + }, + { + "category_id": 1, + "poly": [ + 297, + 311, + 1400, + 311, + 1400, + 405, + 297, + 405 + ], + "score": 0.629 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1193.0, + 1408.0, + 1193.0, + 1408.0, + 1239.0, + 293.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1228.0, + 529.0, + 1228.0, + 529.0, + 1262.0, + 323.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1557.0, + 1381.0, + 1557.0, + 1381.0, + 1596.0, + 295.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1115.0, + 1403.0, + 1115.0, + 1403.0, + 1151.0, + 295.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1147.0, + 813.0, + 1147.0, + 813.0, + 1181.0, + 321.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 590.0, + 1403.0, + 590.0, + 1403.0, + 626.0, + 295.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 621.0, + 632.0, + 621.0, + 632.0, + 654.0, + 322.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 748.0, + 1406.0, + 748.0, + 1406.0, + 796.0, + 293.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 787.0, + 650.0, + 787.0, + 650.0, + 817.0, + 321.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1279.0, + 1406.0, + 1279.0, + 1406.0, + 1317.0, + 295.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1314.0, + 1406.0, + 1314.0, + 1406.0, + 1345.0, + 323.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1340.0, + 399.0, + 1340.0, + 399.0, + 1376.0, + 320.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 502.0, + 1410.0, + 502.0, + 1410.0, + 546.0, + 291.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 538.0, + 523.0, + 538.0, + 523.0, + 571.0, + 323.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 672.0, + 1406.0, + 672.0, + 1406.0, + 708.0, + 298.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 703.0, + 1108.0, + 703.0, + 1108.0, + 738.0, + 323.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 835.0, + 1402.0, + 835.0, + 1402.0, + 872.0, + 296.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 869.0, + 1157.0, + 869.0, + 1157.0, + 903.0, + 321.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2001.0, + 467.0, + 2001.0, + 467.0, + 2034.0, + 321.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1475.0, + 1403.0, + 1475.0, + 1403.0, + 1513.0, + 294.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1506.0, + 1270.0, + 1506.0, + 1270.0, + 1542.0, + 324.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1608.0, + 1279.0, + 1608.0, + 1279.0, + 1648.0, + 295.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 920.0, + 1408.0, + 920.0, + 1408.0, + 955.0, + 297.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 950.0, + 1405.0, + 950.0, + 1405.0, + 987.0, + 320.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 982.0, + 1061.0, + 982.0, + 1061.0, + 1016.0, + 322.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1391.0, + 1403.0, + 1391.0, + 1403.0, + 1431.0, + 296.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1426.0, + 962.0, + 1426.0, + 962.0, + 1459.0, + 323.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1031.0, + 1404.0, + 1031.0, + 1404.0, + 1068.0, + 294.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1063.0, + 1066.0, + 1063.0, + 1066.0, + 1098.0, + 322.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 423.0, + 1407.0, + 423.0, + 1407.0, + 463.0, + 293.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 455.0, + 1251.0, + 455.0, + 1251.0, + 491.0, + 323.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 223.0, + 1405.0, + 223.0, + 1405.0, + 270.0, + 295.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 261.0, + 1180.0, + 261.0, + 1180.0, + 295.0, + 323.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1856.0, + 1405.0, + 1856.0, + 1405.0, + 1893.0, + 295.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1889.0, + 1404.0, + 1889.0, + 1404.0, + 1923.0, + 322.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1918.0, + 466.0, + 1918.0, + 466.0, + 1953.0, + 322.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1663.0, + 1405.0, + 1663.0, + 1405.0, + 1697.0, + 297.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1696.0, + 1403.0, + 1696.0, + 1403.0, + 1727.0, + 326.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1724.0, + 938.0, + 1724.0, + 938.0, + 1758.0, + 322.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1775.0, + 1406.0, + 1775.0, + 1406.0, + 1811.0, + 297.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1807.0, + 760.0, + 1807.0, + 760.0, + 1839.0, + 323.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 313.0, + 1405.0, + 313.0, + 1405.0, + 347.0, + 295.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 342.0, + 1405.0, + 342.0, + 1405.0, + 379.0, + 320.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 374.0, + 611.0, + 374.0, + 611.0, + 403.0, + 322.0, + 403.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.832 + }, + { + "category_id": 1, + "poly": [ + 296, + 1324, + 1405, + 1324, + 1405, + 1390, + 296, + 1390 + ], + "score": 0.718 + }, + { + "category_id": 1, + "poly": [ + 294, + 1407, + 1404, + 1407, + 1404, + 1474, + 294, + 1474 + ], + "score": 0.713 + }, + { + "category_id": 1, + "poly": [ + 293, + 228, + 1403, + 228, + 1403, + 295, + 293, + 295 + ], + "score": 0.711 + }, + { + "category_id": 1, + "poly": [ + 293, + 1491, + 1402, + 1491, + 1402, + 1557, + 293, + 1557 + ], + "score": 0.701 + }, + { + "category_id": 1, + "poly": [ + 296, + 875, + 1401, + 875, + 1401, + 940, + 296, + 940 + ], + "score": 0.688 + }, + { + "category_id": 1, + "poly": [ + 297, + 1240, + 1404, + 1240, + 1404, + 1306, + 297, + 1306 + ], + "score": 0.67 + }, + { + "category_id": 1, + "poly": [ + 295, + 791, + 1400, + 791, + 1400, + 856, + 295, + 856 + ], + "score": 0.667 + }, + { + "category_id": 1, + "poly": [ + 302, + 1574, + 1406, + 1574, + 1406, + 1671, + 302, + 1671 + ], + "score": 0.66 + }, + { + "category_id": 1, + "poly": [ + 298, + 958, + 1404, + 958, + 1404, + 1054, + 298, + 1054 + ], + "score": 0.659 + }, + { + "category_id": 1, + "poly": [ + 299, + 312, + 1402, + 312, + 1402, + 377, + 299, + 377 + ], + "score": 0.658 + }, + { + "category_id": 1, + "poly": [ + 297, + 1072, + 1403, + 1072, + 1403, + 1139, + 297, + 1139 + ], + "score": 0.658 + }, + { + "category_id": 1, + "poly": [ + 294, + 1689, + 1403, + 1689, + 1403, + 1755, + 294, + 1755 + ], + "score": 0.648 + }, + { + "category_id": 1, + "poly": [ + 298, + 1156, + 1402, + 1156, + 1402, + 1222, + 298, + 1222 + ], + "score": 0.647 + }, + { + "category_id": 1, + "poly": [ + 299, + 1971, + 1401, + 1971, + 1401, + 2035, + 299, + 2035 + ], + "score": 0.616 + }, + { + "category_id": 1, + "poly": [ + 293, + 707, + 1400, + 707, + 1400, + 774, + 293, + 774 + ], + "score": 0.616 + }, + { + "category_id": 1, + "poly": [ + 297, + 395, + 1401, + 395, + 1401, + 461, + 297, + 461 + ], + "score": 0.588 + }, + { + "category_id": 1, + "poly": [ + 301, + 479, + 1403, + 479, + 1403, + 575, + 301, + 575 + ], + "score": 0.568 + }, + { + "category_id": 1, + "poly": [ + 299, + 1887, + 1401, + 1887, + 1401, + 1952, + 299, + 1952 + ], + "score": 0.56 + }, + { + "category_id": 1, + "poly": [ + 297, + 593, + 1403, + 593, + 1403, + 690, + 297, + 690 + ], + "score": 0.551 + }, + { + "category_id": 1, + "poly": [ + 299, + 1772, + 1403, + 1772, + 1403, + 1868, + 299, + 1868 + ], + "score": 0.549 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1322.0, + 1407.0, + 1322.0, + 1407.0, + 1361.0, + 294.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1357.0, + 1028.0, + 1357.0, + 1028.0, + 1389.0, + 325.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1407.0, + 1406.0, + 1407.0, + 1406.0, + 1443.0, + 296.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1439.0, + 1333.0, + 1439.0, + 1333.0, + 1476.0, + 322.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 225.0, + 1406.0, + 225.0, + 1406.0, + 267.0, + 292.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 260.0, + 472.0, + 260.0, + 472.0, + 294.0, + 322.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1494.0, + 1402.0, + 1494.0, + 1402.0, + 1527.0, + 297.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1524.0, + 1394.0, + 1524.0, + 1394.0, + 1556.0, + 324.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 875.0, + 1405.0, + 875.0, + 1405.0, + 911.0, + 296.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 906.0, + 503.0, + 906.0, + 503.0, + 940.0, + 323.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1239.0, + 1406.0, + 1239.0, + 1406.0, + 1279.0, + 295.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1272.0, + 906.0, + 1272.0, + 906.0, + 1304.0, + 322.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 787.0, + 1405.0, + 787.0, + 1405.0, + 831.0, + 294.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 824.0, + 761.0, + 824.0, + 761.0, + 854.0, + 323.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1573.0, + 1407.0, + 1573.0, + 1407.0, + 1613.0, + 295.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1603.0, + 1406.0, + 1603.0, + 1406.0, + 1646.0, + 320.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1637.0, + 1312.0, + 1637.0, + 1312.0, + 1672.0, + 322.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 957.0, + 1404.0, + 957.0, + 1404.0, + 996.0, + 292.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 989.0, + 1406.0, + 989.0, + 1406.0, + 1027.0, + 323.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1022.0, + 812.0, + 1022.0, + 812.0, + 1056.0, + 323.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 310.0, + 1404.0, + 310.0, + 1404.0, + 351.0, + 294.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 345.0, + 581.0, + 345.0, + 581.0, + 375.0, + 322.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1066.0, + 1408.0, + 1066.0, + 1408.0, + 1113.0, + 294.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1103.0, + 587.0, + 1103.0, + 587.0, + 1139.0, + 321.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1684.0, + 1405.0, + 1684.0, + 1405.0, + 1731.0, + 294.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1722.0, + 904.0, + 1722.0, + 904.0, + 1753.0, + 322.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1153.0, + 1406.0, + 1153.0, + 1406.0, + 1197.0, + 293.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1189.0, + 1023.0, + 1189.0, + 1023.0, + 1223.0, + 323.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1971.0, + 1407.0, + 1971.0, + 1407.0, + 2007.0, + 297.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2003.0, + 1406.0, + 2003.0, + 1406.0, + 2036.0, + 323.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 707.0, + 1401.0, + 707.0, + 1401.0, + 743.0, + 296.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 739.0, + 1167.0, + 739.0, + 1167.0, + 776.0, + 322.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 390.0, + 1405.0, + 390.0, + 1405.0, + 436.0, + 293.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 429.0, + 826.0, + 429.0, + 826.0, + 459.0, + 323.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 479.0, + 1405.0, + 479.0, + 1405.0, + 516.0, + 294.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 512.0, + 1405.0, + 512.0, + 1405.0, + 548.0, + 321.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 539.0, + 932.0, + 539.0, + 932.0, + 578.0, + 321.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1883.0, + 1405.0, + 1883.0, + 1405.0, + 1925.0, + 294.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1920.0, + 602.0, + 1920.0, + 602.0, + 1950.0, + 322.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 591.0, + 1408.0, + 591.0, + 1408.0, + 633.0, + 295.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 627.0, + 1405.0, + 627.0, + 1405.0, + 661.0, + 322.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 654.0, + 653.0, + 654.0, + 653.0, + 692.0, + 323.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1772.0, + 1405.0, + 1772.0, + 1405.0, + 1810.0, + 295.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1802.0, + 1408.0, + 1802.0, + 1408.0, + 1845.0, + 322.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1835.0, + 534.0, + 1835.0, + 534.0, + 1867.0, + 324.0, + 1867.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.841 + }, + { + "category_id": 1, + "poly": [ + 300, + 229, + 1403, + 229, + 1403, + 322, + 300, + 322 + ], + "score": 0.765 + }, + { + "category_id": 1, + "poly": [ + 296, + 591, + 1405, + 591, + 1405, + 656, + 296, + 656 + ], + "score": 0.714 + }, + { + "category_id": 1, + "poly": [ + 296, + 1347, + 1402, + 1347, + 1402, + 1412, + 296, + 1412 + ], + "score": 0.707 + }, + { + "category_id": 1, + "poly": [ + 297, + 1823, + 1402, + 1823, + 1402, + 1888, + 297, + 1888 + ], + "score": 0.701 + }, + { + "category_id": 1, + "poly": [ + 298, + 674, + 1402, + 674, + 1402, + 769, + 298, + 769 + ], + "score": 0.696 + }, + { + "category_id": 1, + "poly": [ + 291, + 1740, + 1403, + 1740, + 1403, + 1806, + 291, + 1806 + ], + "score": 0.689 + }, + { + "category_id": 1, + "poly": [ + 298, + 508, + 1402, + 508, + 1402, + 573, + 298, + 573 + ], + "score": 0.686 + }, + { + "category_id": 1, + "poly": [ + 300, + 1543, + 1402, + 1543, + 1402, + 1636, + 300, + 1636 + ], + "score": 0.669 + }, + { + "category_id": 1, + "poly": [ + 296, + 870, + 1400, + 870, + 1400, + 936, + 296, + 936 + ], + "score": 0.667 + }, + { + "category_id": 1, + "poly": [ + 296, + 426, + 1402, + 426, + 1402, + 490, + 296, + 490 + ], + "score": 0.664 + }, + { + "category_id": 1, + "poly": [ + 296, + 787, + 1399, + 787, + 1399, + 853, + 296, + 853 + ], + "score": 0.648 + }, + { + "category_id": 1, + "poly": [ + 298, + 1263, + 1400, + 1263, + 1400, + 1330, + 298, + 1330 + ], + "score": 0.647 + }, + { + "category_id": 1, + "poly": [ + 290, + 1656, + 1404, + 1656, + 1404, + 1722, + 290, + 1722 + ], + "score": 0.646 + }, + { + "category_id": 1, + "poly": [ + 297, + 343, + 1404, + 343, + 1404, + 406, + 297, + 406 + ], + "score": 0.644 + }, + { + "category_id": 1, + "poly": [ + 299, + 1429, + 1403, + 1429, + 1403, + 1525, + 299, + 1525 + ], + "score": 0.624 + }, + { + "category_id": 1, + "poly": [ + 299, + 1067, + 1400, + 1067, + 1400, + 1163, + 299, + 1163 + ], + "score": 0.62 + }, + { + "category_id": 1, + "poly": [ + 297, + 1180, + 1400, + 1180, + 1400, + 1246, + 297, + 1246 + ], + "score": 0.617 + }, + { + "category_id": 1, + "poly": [ + 300, + 953, + 1403, + 953, + 1403, + 1049, + 300, + 1049 + ], + "score": 0.565 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 260.0, + 1408.0, + 260.0, + 1408.0, + 296.0, + 321.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 286.0, + 400.0, + 286.0, + 400.0, + 327.0, + 320.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 593.0, + 1408.0, + 593.0, + 1408.0, + 629.0, + 298.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 624.0, + 1381.0, + 624.0, + 1381.0, + 657.0, + 322.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1345.0, + 1404.0, + 1345.0, + 1404.0, + 1385.0, + 294.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1378.0, + 772.0, + 1378.0, + 772.0, + 1413.0, + 323.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1818.0, + 1407.0, + 1818.0, + 1407.0, + 1862.0, + 294.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1854.0, + 686.0, + 1854.0, + 686.0, + 1886.0, + 322.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 670.0, + 1407.0, + 670.0, + 1407.0, + 715.0, + 293.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 706.0, + 1404.0, + 706.0, + 1404.0, + 740.0, + 323.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 737.0, + 490.0, + 737.0, + 490.0, + 769.0, + 323.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1737.0, + 1405.0, + 1737.0, + 1405.0, + 1780.0, + 293.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1773.0, + 998.0, + 1773.0, + 998.0, + 1805.0, + 322.0, + 1805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 510.0, + 1404.0, + 510.0, + 1404.0, + 546.0, + 296.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 538.0, + 1258.0, + 538.0, + 1258.0, + 577.0, + 322.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1544.0, + 1407.0, + 1544.0, + 1407.0, + 1580.0, + 294.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1573.0, + 1409.0, + 1573.0, + 1409.0, + 1612.0, + 320.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1603.0, + 399.0, + 1603.0, + 399.0, + 1639.0, + 320.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 871.0, + 1404.0, + 871.0, + 1404.0, + 907.0, + 296.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 904.0, + 759.0, + 904.0, + 759.0, + 937.0, + 325.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 426.0, + 1404.0, + 426.0, + 1404.0, + 462.0, + 296.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 458.0, + 968.0, + 458.0, + 968.0, + 491.0, + 323.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 788.0, + 1403.0, + 788.0, + 1403.0, + 824.0, + 297.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 820.0, + 930.0, + 820.0, + 930.0, + 854.0, + 321.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1259.0, + 1404.0, + 1259.0, + 1404.0, + 1304.0, + 293.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1297.0, + 994.0, + 1297.0, + 994.0, + 1332.0, + 322.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1659.0, + 1402.0, + 1659.0, + 1402.0, + 1692.0, + 296.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1688.0, + 981.0, + 1688.0, + 981.0, + 1724.0, + 322.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 344.0, + 1407.0, + 344.0, + 1407.0, + 380.0, + 296.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 375.0, + 990.0, + 375.0, + 990.0, + 406.0, + 324.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1429.0, + 1404.0, + 1429.0, + 1404.0, + 1467.0, + 294.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1462.0, + 1404.0, + 1462.0, + 1404.0, + 1497.0, + 323.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1493.0, + 618.0, + 1493.0, + 618.0, + 1522.0, + 323.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1067.0, + 1405.0, + 1067.0, + 1405.0, + 1105.0, + 295.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1099.0, + 1404.0, + 1099.0, + 1404.0, + 1133.0, + 322.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1128.0, + 1220.0, + 1128.0, + 1220.0, + 1166.0, + 324.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1177.0, + 1405.0, + 1177.0, + 1405.0, + 1221.0, + 293.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1213.0, + 970.0, + 1213.0, + 970.0, + 1244.0, + 323.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 953.0, + 1405.0, + 953.0, + 1405.0, + 991.0, + 295.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 986.0, + 1404.0, + 986.0, + 1404.0, + 1021.0, + 324.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1017.0, + 720.0, + 1017.0, + 720.0, + 1051.0, + 323.0, + 1051.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 483, + 469, + 1217, + 469, + 1217, + 638, + 483, + 638 + ], + "score": 0.98, + "html": "
M benchmarksLFWAgeDB-30CFP-FFCFP-FP
R-50-S99.2794.1599.2691.49
R-101-S99.4295.9399.5795.07
R-50-C99.3895.0899.2490.24
R-101-C99.6795.5899.5792.71
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1084, + 1407, + 1084, + 1407, + 1269, + 297, + 1269 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1634, + 1405, + 1634, + 1405, + 1910, + 297, + 1910 + ], + "score": 0.979 + }, + { + "category_id": 5, + "poly": [ + 526, + 863, + 1164, + 863, + 1164, + 1006, + 526, + 1006 + ], + "score": 0.976, + "html": "
FPR/MR-50-SR-101-SR-50-CR-101-C
10-33×10-410-41.1811.2441.4471.469
1.0580.6571.0481.2931.2420.809
0.5970.864
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1925, + 1402, + 1925, + 1402, + 2048, + 298, + 2048 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 368, + 1382, + 1404, + 1382, + 1404, + 1606, + 368, + 1606 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 295, + 280, + 1408, + 280, + 1408, + 434, + 295, + 434 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 299, + 704, + 1405, + 704, + 1405, + 829, + 299, + 829 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 293, + 1290, + 1402, + 1290, + 1402, + 1354, + 293, + 1354 + ], + "score": 0.946 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 302, + 225, + 796, + 225, + 796, + 262, + 302, + 262 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.859 + }, + { + "category_id": 7, + "poly": [ + 520, + 1017, + 1177, + 1017, + 1177, + 1049, + 520, + 1049 + ], + "score": 0.701 + }, + { + "category_id": 1, + "poly": [ + 425, + 652, + 1273, + 652, + 1273, + 684, + 425, + 684 + ], + "score": 0.508 + }, + { + "category_id": 1, + "poly": [ + 520, + 1017, + 1177, + 1017, + 1177, + 1049, + 520, + 1049 + ], + "score": 0.173 + }, + { + "category_id": 7, + "poly": [ + 425, + 652, + 1273, + 652, + 1273, + 684, + 425, + 684 + ], + "score": 0.13 + }, + { + "category_id": 13, + "poly": [ + 820, + 1236, + 879, + 1236, + 879, + 1266, + 820, + 1266 + ], + "score": 0.9, + "latex": "1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 997, + 1145, + 1175, + 1145, + 1175, + 1177, + 997, + 1177 + ], + "score": 0.69, + "latex": "\\mathbf { \\bar { G } } \\mathbf { - } \\mathbf { F P R } \\leq 1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 399, + 1238, + 472, + 1238, + 472, + 1269, + 399, + 1269 + ], + "score": 0.26, + "latex": "1 , 0 0 0" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 221.0, + 801.0, + 221.0, + 801.0, + 268.0, + 294.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 868.0, + 2084.0, + 868.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 1012.0, + 1179.0, + 1012.0, + 1179.0, + 1055.0, + 516.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 649.0, + 1279.0, + 649.0, + 1279.0, + 688.0, + 420.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1082.0, + 1409.0, + 1082.0, + 1409.0, + 1121.0, + 294.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1115.0, + 1407.0, + 1115.0, + 1407.0, + 1151.0, + 294.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1141.0, + 996.0, + 1141.0, + 996.0, + 1184.0, + 291.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 1141.0, + 1409.0, + 1141.0, + 1409.0, + 1184.0, + 1176.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1175.0, + 1409.0, + 1175.0, + 1409.0, + 1211.0, + 292.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1207.0, + 1405.0, + 1207.0, + 1405.0, + 1243.0, + 294.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1235.0, + 398.0, + 1235.0, + 398.0, + 1273.0, + 290.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 1235.0, + 819.0, + 1235.0, + 819.0, + 1273.0, + 473.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 1235.0, + 891.0, + 1235.0, + 891.0, + 1273.0, + 880.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1634.0, + 1405.0, + 1634.0, + 1405.0, + 1669.0, + 294.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1666.0, + 1403.0, + 1666.0, + 1403.0, + 1699.0, + 295.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1696.0, + 1405.0, + 1696.0, + 1405.0, + 1730.0, + 292.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1724.0, + 1405.0, + 1724.0, + 1405.0, + 1762.0, + 292.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1755.0, + 1405.0, + 1755.0, + 1405.0, + 1790.0, + 294.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1786.0, + 1405.0, + 1786.0, + 1405.0, + 1822.0, + 294.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1819.0, + 1405.0, + 1819.0, + 1405.0, + 1852.0, + 295.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1849.0, + 1403.0, + 1849.0, + 1403.0, + 1882.0, + 295.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1879.0, + 724.0, + 1879.0, + 724.0, + 1913.0, + 292.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1926.0, + 1406.0, + 1926.0, + 1406.0, + 1959.0, + 294.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1956.0, + 1404.0, + 1956.0, + 1404.0, + 1989.0, + 293.0, + 1989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1987.0, + 1406.0, + 1987.0, + 1406.0, + 2019.0, + 293.0, + 2019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2018.0, + 879.0, + 2018.0, + 879.0, + 2049.0, + 294.0, + 2049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 1381.0, + 1405.0, + 1381.0, + 1405.0, + 1420.0, + 365.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1413.0, + 939.0, + 1413.0, + 939.0, + 1448.0, + 394.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1463.0, + 1405.0, + 1463.0, + 1405.0, + 1497.0, + 368.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1491.0, + 498.0, + 1491.0, + 498.0, + 1529.0, + 392.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1541.0, + 1405.0, + 1541.0, + 1405.0, + 1576.0, + 368.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1574.0, + 762.0, + 1574.0, + 762.0, + 1609.0, + 393.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 282.0, + 1406.0, + 282.0, + 1406.0, + 315.0, + 296.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 312.0, + 1406.0, + 312.0, + 1406.0, + 345.0, + 296.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 342.0, + 1405.0, + 342.0, + 1405.0, + 376.0, + 296.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 369.0, + 1409.0, + 369.0, + 1409.0, + 409.0, + 292.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 398.0, + 378.0, + 398.0, + 378.0, + 441.0, + 292.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 703.0, + 1405.0, + 703.0, + 1405.0, + 738.0, + 294.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 735.0, + 1406.0, + 735.0, + 1406.0, + 771.0, + 294.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 766.0, + 1403.0, + 766.0, + 1403.0, + 799.0, + 295.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 796.0, + 1397.0, + 796.0, + 1397.0, + 832.0, + 293.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1285.0, + 1407.0, + 1285.0, + 1407.0, + 1328.0, + 292.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1322.0, + 674.0, + 1322.0, + 674.0, + 1354.0, + 296.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 649.0, + 1279.0, + 649.0, + 1279.0, + 688.0, + 420.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 1012.0, + 1179.0, + 1012.0, + 1179.0, + 1055.0, + 516.0, + 1055.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 381, + 223, + 1313, + 223, + 1313, + 1005, + 381, + 1005 + ], + "score": 0.986, + "html": "
G-FPRMetricsMR-50-SR-101-SR-50-CR-101-C
10-3Verification Accuracy98.3698.7898.6398.84
xG(x',c)G(x',cnew)W)(Best)0.000.000.160.000.000.080.080.000.160.000.230.31
SemanticAdv (Best)SemanticAdv (Worst)SemanticAdv (Average)100.0091.9598.98100.0093.9899.29100.0099.5399.97100.0099.7799.99
CW100.00100.00100.00100.00
3 ×10-4Verification Accuracy97.7397.9797.9197.85
xG(x',c)G(x',cnew)(Best)0.000.000.000.000.000.000.000.000.000.000.000.00
SemanticAdv (Best)SemanticAdv (Worst)SemanticAdv (Average)100.0083.7597.72100.0079.0697.35100.0098.9899.92100.0096.6499.72
CW100.00100.00100.00100.00
10-4Verification Accuracy93.2592.8093.4392.98
XG(x',c)G(x',cnew)(Best)0.000.000.000.000.000.000.000.000.000.000.000.00
SemanticAdv (Best)SemanticAdv (Worst)SemanticAdv (Average)100.0033.5983.53100.0019.8476.64100.0067.0395.57100.0048.6791.13
CW100.00100.00100.00100.00
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1142, + 1405, + 1142, + 1405, + 1327, + 297, + 1327 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1872, + 1404, + 1872, + 1404, + 2034, + 297, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1539, + 1406, + 1539, + 1406, + 1632, + 298, + 1632 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1427, + 1406, + 1427, + 1406, + 1521, + 299, + 1521 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 682, + 1657, + 1017, + 1657, + 1017, + 1748, + 682, + 1748 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 300, + 1765, + 1403, + 1765, + 1403, + 1859, + 300, + 1859 + ], + "score": 0.951 + }, + { + "category_id": 0, + "poly": [ + 301, + 1375, + 788, + 1375, + 788, + 1412, + 301, + 1412 + ], + "score": 0.92 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.908 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1685, + 1400, + 1685, + 1400, + 1715, + 1366, + 1715 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 863, + 2088, + 863, + 2112, + 835, + 2112 + ], + "score": 0.857 + }, + { + "category_id": 6, + "poly": [ + 298, + 1015, + 1402, + 1015, + 1402, + 1108, + 298, + 1108 + ], + "score": 0.435 + }, + { + "category_id": 7, + "poly": [ + 298, + 1015, + 1402, + 1015, + 1402, + 1108, + 298, + 1108 + ], + "score": 0.261 + }, + { + "category_id": 14, + "poly": [ + 681, + 1652, + 1017, + 1652, + 1017, + 1748, + 681, + 1748 + ], + "score": 0.95, + "latex": "r _ { \\mathrm { N M E } } = \\frac { 1 } { N } \\sum _ { k = 1 } ^ { N } \\frac { | | \\mathbf { p } _ { k } - \\hat { \\mathbf { p } } _ { k } | | _ { 2 } } { \\sqrt { W _ { B } * H _ { B } } } ," + }, + { + "category_id": 13, + "poly": [ + 1039, + 1045, + 1169, + 1045, + 1169, + 1079, + 1039, + 1079 + ], + "score": 0.94, + "latex": "G ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } ^ { \\mathrm { n e w } } )" + }, + { + "category_id": 13, + "poly": [ + 1028, + 1934, + 1170, + 1934, + 1170, + 1976, + 1028, + 1976 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { r _ { \\mathrm { O U T } } = \\frac { N _ { \\mathrm { o u t } } } { N _ { \\mathrm { t o t a l } } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 889, + 1045, + 986, + 1045, + 986, + 1079, + 889, + 1079 + ], + "score": 0.93, + "latex": "G ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1796, + 437, + 1796, + 437, + 1829, + 298, + 1829 + ], + "score": 0.92, + "latex": "\\sqrt { W _ { B } * H _ { B } }" + }, + { + "category_id": 13, + "poly": [ + 1198, + 1797, + 1243, + 1797, + 1243, + 1828, + 1198, + 1828 + ], + "score": 0.9, + "latex": "H _ { B }" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1797, + 1148, + 1797, + 1148, + 1828, + 1100, + 1828 + ], + "score": 0.9, + "latex": "W _ { B }" + }, + { + "category_id": 13, + "poly": [ + 1257, + 1938, + 1309, + 1938, + 1309, + 1970, + 1257, + 1970 + ], + "score": 0.9, + "latex": "N _ { \\mathrm { o u t } }" + }, + { + "category_id": 13, + "poly": [ + 554, + 1234, + 621, + 1234, + 621, + 1263, + 554, + 1263 + ], + "score": 0.9, + "latex": "3 \\times 3" + }, + { + "category_id": 13, + "poly": [ + 1153, + 1975, + 1214, + 1975, + 1214, + 2003, + 1153, + 2003 + ], + "score": 0.9, + "latex": "N _ { \\mathrm { t o t a l } }" + }, + { + "category_id": 13, + "poly": [ + 886, + 1768, + 924, + 1768, + 924, + 1799, + 886, + 1799 + ], + "score": 0.88, + "latex": "\\hat { \\mathbf { p } } _ { \\mathbf { k } }" + }, + { + "category_id": 13, + "poly": [ + 1069, + 1768, + 1088, + 1768, + 1088, + 1794, + 1069, + 1794 + ], + "score": 0.84, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 374, + 1771, + 409, + 1771, + 409, + 1797, + 374, + 1797 + ], + "score": 0.83, + "latex": "\\mathbf { p } _ { k }" + }, + { + "category_id": 13, + "poly": [ + 555, + 1768, + 574, + 1768, + 574, + 1794, + 555, + 1794 + ], + "score": 0.83, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 858, + 1017, + 899, + 1017, + 899, + 1045, + 858, + 1045 + ], + "score": 0.81, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 540, + 788, + 566, + 788, + 566, + 810, + 540, + 810 + ], + "score": 0.8, + "latex": "\\mathbf { x } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 849, + 1047, + 878, + 1047, + 878, + 1075, + 849, + 1075 + ], + "score": 0.79, + "latex": "\\mathbf { x } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 423, + 373, + 482, + 373, + 482, + 398, + 423, + 398 + ], + "score": 0.77, + "latex": "1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 1291, + 1461, + 1311, + 1461, + 1311, + 1487, + 1291, + 1487 + ], + "score": 0.77, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 539, + 543, + 565, + 543, + 565, + 565, + 539, + 565 + ], + "score": 0.74, + "latex": "\\mathbf { x } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 539, + 573, + 635, + 573, + 635, + 602, + 539, + 602 + ], + "score": 0.68, + "latex": "G ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } )" + }, + { + "category_id": 13, + "poly": [ + 399, + 617, + 506, + 617, + 506, + 645, + 399, + 645 + ], + "score": 0.65, + "latex": "3 \\times 1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 424, + 861, + 482, + 861, + 482, + 887, + 424, + 887 + ], + "score": 0.65, + "latex": "1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 539, + 602, + 735, + 602, + 735, + 632, + 539, + 632 + ], + "score": 0.63, + "latex": "G ( { \\bf x } ^ { \\prime } , { \\bf c } ^ { \\mathrm { n e w } } ) ( B e s t )" + }, + { + "category_id": 13, + "poly": [ + 539, + 817, + 635, + 817, + 635, + 846, + 539, + 846 + ], + "score": 0.62, + "latex": "G ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } )" + }, + { + "category_id": 13, + "poly": [ + 539, + 356, + 736, + 356, + 736, + 388, + 539, + 388 + ], + "score": 0.58, + "latex": "G ( { \\bf x } ^ { \\prime } , { \\bf c } ^ { \\mathrm { n e w } } ) ( B e s t )" + }, + { + "category_id": 13, + "poly": [ + 830, + 1144, + 868, + 1144, + 868, + 1171, + 830, + 1171 + ], + "score": 0.5, + "latex": "\\mathrm { { X u } }" + }, + { + "category_id": 13, + "poly": [ + 539, + 325, + 636, + 325, + 636, + 357, + 539, + 357 + ], + "score": 0.48, + "latex": "G ( \\mathbf { x } ^ { \\prime } , \\mathbf { c } )" + }, + { + "category_id": 13, + "poly": [ + 539, + 846, + 735, + 846, + 735, + 877, + 539, + 877 + ], + "score": 0.44, + "latex": "G ( { \\bf x } ^ { \\prime } , { \\bf c } ^ { \\mathrm { n e w } } ) ( B e s t )" + }, + { + "category_id": 13, + "poly": [ + 539, + 297, + 566, + 297, + 566, + 320, + 539, + 320 + ], + "score": 0.41, + "latex": "\\overline { { \\mathbf { x } ^ { \\prime } } }" + }, + { + "category_id": 13, + "poly": [ + 652, + 226, + 686, + 226, + 686, + 253, + 652, + 253 + ], + "score": 0.35, + "latex": "\\overline { { \\mathcal { M } } }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1372.0, + 792.0, + 1372.0, + 792.0, + 1417.0, + 293.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1013.0, + 857.0, + 1013.0, + 857.0, + 1051.0, + 293.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1013.0, + 1406.0, + 1013.0, + 1406.0, + 1051.0, + 900.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1041.0, + 848.0, + 1041.0, + 848.0, + 1082.0, + 292.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 1041.0, + 888.0, + 1041.0, + 888.0, + 1082.0, + 879.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 1041.0, + 1038.0, + 1041.0, + 1038.0, + 1082.0, + 987.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 1041.0, + 1407.0, + 1041.0, + 1407.0, + 1082.0, + 1170.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1075.0, + 894.0, + 1075.0, + 894.0, + 1110.0, + 293.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1013.0, + 857.0, + 1013.0, + 857.0, + 1051.0, + 293.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1013.0, + 1406.0, + 1013.0, + 1406.0, + 1051.0, + 900.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1041.0, + 848.0, + 1041.0, + 848.0, + 1082.0, + 292.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 1041.0, + 888.0, + 1041.0, + 888.0, + 1082.0, + 879.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 1041.0, + 1038.0, + 1041.0, + 1038.0, + 1082.0, + 987.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 1041.0, + 1407.0, + 1041.0, + 1407.0, + 1082.0, + 1170.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1075.0, + 894.0, + 1075.0, + 894.0, + 1110.0, + 293.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1138.0, + 829.0, + 1138.0, + 829.0, + 1180.0, + 292.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1138.0, + 1407.0, + 1138.0, + 1407.0, + 1180.0, + 869.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1172.0, + 1406.0, + 1172.0, + 1406.0, + 1209.0, + 292.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1201.0, + 1406.0, + 1201.0, + 1406.0, + 1238.0, + 292.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1233.0, + 553.0, + 1233.0, + 553.0, + 1265.0, + 295.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 1233.0, + 1405.0, + 1233.0, + 1405.0, + 1265.0, + 622.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1265.0, + 1407.0, + 1265.0, + 1407.0, + 1298.0, + 292.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1294.0, + 1327.0, + 1294.0, + 1327.0, + 1330.0, + 294.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1874.0, + 1404.0, + 1874.0, + 1404.0, + 1908.0, + 295.0, + 1908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1904.0, + 1404.0, + 1904.0, + 1404.0, + 1939.0, + 292.0, + 1939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1152.0, + 1971.0, + 1152.0, + 2009.0, + 295.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2009.0, + 1215.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1999.0, + 546.0, + 1999.0, + 546.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.25, + 1926.5, + 1184.25, + 1926.5, + 1184.25, + 1986.5, + 286.25, + 1986.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 1932.0, + 1407.0, + 1932.0, + 1407.0, + 1975.5, + 1115.0, + 1975.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1538.0, + 1408.0, + 1538.0, + 1408.0, + 1576.0, + 293.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1570.0, + 1407.0, + 1570.0, + 1407.0, + 1604.0, + 295.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1601.0, + 876.0, + 1601.0, + 876.0, + 1636.0, + 293.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1428.0, + 1406.0, + 1428.0, + 1406.0, + 1462.0, + 296.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1457.0, + 1290.0, + 1457.0, + 1290.0, + 1495.0, + 294.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 1457.0, + 1408.0, + 1457.0, + 1408.0, + 1495.0, + 1312.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1490.0, + 700.0, + 1490.0, + 700.0, + 1524.0, + 296.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1764.0, + 373.0, + 1764.0, + 373.0, + 1800.0, + 293.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 1764.0, + 554.0, + 1764.0, + 554.0, + 1800.0, + 410.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1764.0, + 885.0, + 1764.0, + 885.0, + 1800.0, + 575.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1764.0, + 1068.0, + 1764.0, + 1068.0, + 1800.0, + 925.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 1764.0, + 1405.0, + 1764.0, + 1405.0, + 1800.0, + 1089.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 1794.0, + 1099.0, + 1794.0, + 1099.0, + 1831.0, + 438.0, + 1831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 1794.0, + 1197.0, + 1794.0, + 1197.0, + 1831.0, + 1149.0, + 1831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 1794.0, + 1404.0, + 1794.0, + 1404.0, + 1831.0, + 1244.0, + 1831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1824.0, + 617.0, + 1824.0, + 617.0, + 1862.0, + 294.0, + 1862.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 734, + 1405, + 734, + 1405, + 1192, + 296, + 1192 + ], + "score": 0.983 + }, + { + "category_id": 5, + "poly": [ + 315, + 224, + 1380, + 224, + 1380, + 351, + 315, + 351 + ], + "score": 0.976, + "html": "
Tasks (Metrics)PristineAugmented Attributes
BlondHairYoungEyeglassesRosyCheeksSmilingArchedEyebrowsBangsPaleSkin
Rotating eyes (rNME)↓28.0414.0317.288.5813.2419.2123.4215.9910.72
Out-of-region (rouT)↓45.9817.4223.047.5116.6525.4433.8520.0313.51
" + }, + { + "category_id": 1, + "poly": [ + 298, + 470, + 1404, + 470, + 1404, + 594, + 298, + 594 + ], + "score": 0.971 + }, + { + "category_id": 3, + "poly": [ + 297, + 1234, + 1400, + 1234, + 1400, + 1887, + 297, + 1887 + ], + "score": 0.955 + }, + { + "category_id": 4, + "poly": [ + 296, + 1931, + 1400, + 1931, + 1400, + 2024, + 296, + 2024 + ], + "score": 0.936 + }, + { + "category_id": 0, + "poly": [ + 299, + 654, + 542, + 654, + 542, + 691, + 299, + 691 + ], + "score": 0.91 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 836, + 2089, + 864, + 2089, + 864, + 2113, + 836, + 2113 + ], + "score": 0.857 + }, + { + "category_id": 7, + "poly": [ + 299, + 364, + 1400, + 364, + 1400, + 427, + 299, + 427 + ], + "score": 0.741 + }, + { + "category_id": 6, + "poly": [ + 299, + 364, + 1400, + 364, + 1400, + 427, + 299, + 427 + ], + "score": 0.148 + }, + { + "category_id": 13, + "poly": [ + 976, + 1008, + 1141, + 1008, + 1141, + 1039, + 976, + 1039 + ], + "score": 0.92, + "latex": "3 0 . 2 7 \\pm 1 . 9 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 783, + 1099, + 948, + 1099, + 948, + 1130, + 783, + 1130 + ], + "score": 0.91, + "latex": "4 5 . 4 2 \\pm 1 . 9 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 716, + 978, + 890, + 978, + 890, + 1009, + 716, + 1009 + ], + "score": 0.91, + "latex": "3 9 . 1 4 \\pm 1 . 9 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 1230, + 764, + 1401, + 764, + 1401, + 795, + 1230, + 795 + ], + "score": 0.84, + "latex": "\\mathrm { { G } \\mathrm { { - F P R } \\mathrm { { < } \\ 1 0 ^ { - 4 } } } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1160, + 366, + 1160, + 366, + 1192, + 298, + 1192 + ], + "score": 0.82, + "latex": "( \\dot { 5 } 0 \\% )" + }, + { + "category_id": 13, + "poly": [ + 1232, + 1960, + 1318, + 1960, + 1318, + 1992, + 1232, + 1992 + ], + "score": 0.8, + "latex": "\\dot { < } 1 0 ^ { { - } 4 }" + }, + { + "category_id": 13, + "poly": [ + 576, + 1098, + 741, + 1098, + 741, + 1129, + 576, + 1129 + ], + "score": 0.67, + "latex": "\\mathrm { { G } { - } \\mathrm { { F P R } } \\mathrm { { = } \\ i 0 ^ { - 3 } } }" + }, + { + "category_id": 13, + "poly": [ + 923, + 365, + 956, + 365, + 956, + 393, + 923, + 393 + ], + "score": 0.63, + "latex": "( \\% )" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 1242.0, + 367.0, + 1242.0, + 367.0, + 1267.0, + 336.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 1242.0, + 480.0, + 1242.0, + 480.0, + 1268.0, + 439.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1241.0, + 624.0, + 1241.0, + 624.0, + 1269.0, + 505.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 1243.0, + 760.0, + 1243.0, + 760.0, + 1267.0, + 727.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 1241.0, + 874.0, + 1241.0, + 874.0, + 1268.0, + 831.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 1240.0, + 1016.0, + 1240.0, + 1016.0, + 1272.0, + 895.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 1240.0, + 1144.0, + 1240.0, + 1144.0, + 1268.0, + 1112.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 1241.0, + 1257.0, + 1241.0, + 1257.0, + 1268.0, + 1217.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 1241.0, + 1401.0, + 1241.0, + 1401.0, + 1269.0, + 1282.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1262.0, + 629.0, + 1262.0, + 629.0, + 1399.0, + 291.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1284.0, + 993.0, + 1284.0, + 993.0, + 1387.0, + 704.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 1262.0, + 1407.0, + 1262.0, + 1407.0, + 1399.0, + 1072.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1381.0, + 627.0, + 1381.0, + 627.0, + 1518.0, + 290.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1387.0, + 1013.0, + 1387.0, + 1013.0, + 1513.0, + 684.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 1388.0, + 1392.0, + 1388.0, + 1392.0, + 1515.0, + 1084.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1504.0, + 1015.0, + 1504.0, + 1015.0, + 1632.0, + 684.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 1520.0, + 1388.0, + 1520.0, + 1388.0, + 1625.0, + 1082.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 1632.0, + 615.0, + 1632.0, + 615.0, + 1743.0, + 312.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 1617.0, + 1018.0, + 1617.0, + 1018.0, + 1750.0, + 681.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 1627.0, + 1392.0, + 1627.0, + 1392.0, + 1751.0, + 1080.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1738.0, + 628.0, + 1738.0, + 628.0, + 1866.0, + 292.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1738.0, + 1014.0, + 1738.0, + 1014.0, + 1866.0, + 684.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 1738.0, + 1405.0, + 1738.0, + 1405.0, + 1866.0, + 1072.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 1853.0, + 624.0, + 1853.0, + 624.0, + 1889.0, + 500.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1854.0, + 1001.0, + 1854.0, + 1001.0, + 1888.0, + 920.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 1859.0, + 1395.0, + 1859.0, + 1395.0, + 1883.0, + 1306.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1930.0, + 1406.0, + 1930.0, + 1406.0, + 1968.0, + 294.0, + 1968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1959.0, + 1231.0, + 1959.0, + 1231.0, + 1997.0, + 292.0, + 1997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1319.0, + 1959.0, + 1406.0, + 1959.0, + 1406.0, + 1997.0, + 1319.0, + 1997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1993.0, + 579.0, + 1993.0, + 579.0, + 2024.0, + 293.0, + 2024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 648.0, + 548.0, + 648.0, + 548.0, + 699.0, + 293.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 361.0, + 922.0, + 361.0, + 922.0, + 397.0, + 295.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 361.0, + 1404.0, + 361.0, + 1404.0, + 397.0, + 957.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 394.0, + 1151.0, + 394.0, + 1151.0, + 430.0, + 295.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 361.0, + 922.0, + 361.0, + 922.0, + 397.0, + 295.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 361.0, + 1404.0, + 361.0, + 1404.0, + 397.0, + 957.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 394.0, + 1151.0, + 394.0, + 1151.0, + 430.0, + 295.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 734.0, + 1406.0, + 734.0, + 1406.0, + 769.0, + 295.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 761.0, + 1229.0, + 761.0, + 1229.0, + 802.0, + 290.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 761.0, + 1407.0, + 761.0, + 1407.0, + 802.0, + 1402.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 794.0, + 1405.0, + 794.0, + 1405.0, + 831.0, + 293.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 825.0, + 1405.0, + 825.0, + 1405.0, + 862.0, + 293.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 855.0, + 1405.0, + 855.0, + 1405.0, + 892.0, + 293.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 887.0, + 1405.0, + 887.0, + 1405.0, + 922.0, + 294.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 919.0, + 1403.0, + 919.0, + 1403.0, + 951.0, + 295.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 949.0, + 1405.0, + 949.0, + 1405.0, + 980.0, + 295.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 975.0, + 715.0, + 975.0, + 715.0, + 1015.0, + 293.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 975.0, + 1405.0, + 975.0, + 1405.0, + 1015.0, + 891.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1007.0, + 975.0, + 1007.0, + 975.0, + 1044.0, + 291.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 1007.0, + 1407.0, + 1007.0, + 1407.0, + 1044.0, + 1142.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1039.0, + 1405.0, + 1039.0, + 1405.0, + 1074.0, + 293.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1070.0, + 1405.0, + 1070.0, + 1405.0, + 1105.0, + 294.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1095.0, + 575.0, + 1095.0, + 575.0, + 1136.0, + 293.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 1095.0, + 782.0, + 1095.0, + 782.0, + 1136.0, + 742.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1095.0, + 1406.0, + 1095.0, + 1406.0, + 1136.0, + 949.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1129.0, + 1405.0, + 1129.0, + 1405.0, + 1166.0, + 293.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1159.0, + 297.0, + 1159.0, + 297.0, + 1196.0, + 293.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 1159.0, + 382.0, + 1159.0, + 382.0, + 1196.0, + 367.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 470.0, + 1405.0, + 470.0, + 1405.0, + 502.0, + 294.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 502.0, + 1406.0, + 502.0, + 1406.0, + 534.0, + 294.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 532.0, + 1405.0, + 532.0, + 1405.0, + 565.0, + 294.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 562.0, + 1291.0, + 562.0, + 1291.0, + 598.0, + 293.0, + 598.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 293, + 1406, + 293, + 1406, + 509, + 297, + 509 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1788, + 1404, + 1788, + 1404, + 2035, + 297, + 2035 + ], + "score": 0.979 + }, + { + "category_id": 5, + "poly": [ + 309, + 641, + 1395, + 641, + 1395, + 744, + 309, + 744 + ], + "score": 0.976, + "html": "
T-FPR (G-FPR)10-3 (10-3)3× 10-4 (3×10-4)10-4 (10-4)
Layer (f)ff1f2ff1f2ff1f
Attack SuccessRate99.2998.3275.6297.3594.1057.1576.6467.4019.63
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1681, + 1403, + 1681, + 1403, + 1773, + 298, + 1773 + ], + "score": 0.974 + }, + { + "category_id": 5, + "poly": [ + 417, + 868, + 1280, + 868, + 1280, + 973, + 417, + 973 + ], + "score": 0.973, + "html": "
T-FPR (G-FPR)10-3 (10-3)3×10)-4(3×10-4)10-4(10-4)
Layer (f)f-2f-1f-2f-1f-2f-1
Attack Success Rate49.4092.0930.4481.876.6645.46
" + }, + { + "category_id": 3, + "poly": [ + 299, + 1132, + 1386, + 1132, + 1386, + 1505, + 299, + 1505 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 298, + 996, + 1404, + 996, + 1404, + 1089, + 298, + 1089 + ], + "score": 0.93 + }, + { + "category_id": 4, + "poly": [ + 348, + 1554, + 1349, + 1554, + 1349, + 1587, + 348, + 1587 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 295, + 767, + 1405, + 767, + 1405, + 830, + 295, + 830 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.862 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.846 + }, + { + "category_id": 0, + "poly": [ + 299, + 225, + 1111, + 225, + 1111, + 262, + 299, + 262 + ], + "score": 0.795 + }, + { + "category_id": 0, + "poly": [ + 297, + 1614, + 911, + 1614, + 911, + 1650, + 297, + 1650 + ], + "score": 0.752 + }, + { + "category_id": 6, + "poly": [ + 298, + 524, + 1404, + 524, + 1404, + 617, + 298, + 617 + ], + "score": 0.572 + }, + { + "category_id": 1, + "poly": [ + 298, + 524, + 1404, + 524, + 1404, + 617, + 298, + 617 + ], + "score": 0.355 + }, + { + "category_id": 1, + "poly": [ + 297, + 1614, + 911, + 1614, + 911, + 1650, + 297, + 1650 + ], + "score": 0.163 + }, + { + "category_id": 1, + "poly": [ + 299, + 225, + 1111, + 225, + 1111, + 262, + 299, + 262 + ], + "score": 0.093 + }, + { + "category_id": 13, + "poly": [ + 1137, + 1679, + 1309, + 1679, + 1309, + 1711, + 1137, + 1711 + ], + "score": 0.9, + "latex": "\\mathrm { { G } \\mathrm { { - } \\mathrm { { F P R } = 1 0 ^ { - 4 } } } }" + }, + { + "category_id": 13, + "poly": [ + 294, + 1710, + 383, + 1710, + 383, + 1742, + 294, + 1742 + ], + "score": 0.9, + "latex": "= 1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 513, + 2004, + 548, + 2004, + 548, + 2037, + 513, + 2037 + ], + "score": 0.9, + "latex": "L _ { p }" + }, + { + "category_id": 13, + "poly": [ + 1234, + 1882, + 1269, + 1882, + 1269, + 1915, + 1234, + 1915 + ], + "score": 0.88, + "latex": "L _ { p }" + }, + { + "category_id": 13, + "poly": [ + 295, + 1943, + 358, + 1943, + 358, + 1974, + 295, + 1974 + ], + "score": 0.87, + "latex": "\\mathbf { x } ^ { \\mathrm { n e w } } ]" + }, + { + "category_id": 13, + "poly": [ + 1207, + 1027, + 1250, + 1027, + 1250, + 1058, + 1207, + 1058 + ], + "score": 0.86, + "latex": "\\mathbf { f } _ { - 2 }" + }, + { + "category_id": 13, + "poly": [ + 446, + 356, + 469, + 356, + 469, + 386, + 446, + 386 + ], + "score": 0.86, + "latex": "\\mathbf { f } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 484, + 418, + 510, + 418, + 510, + 448, + 484, + 448 + ], + "score": 0.86, + "latex": "\\mathbf { f } _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 957, + 417, + 984, + 417, + 984, + 448, + 957, + 448 + ], + "score": 0.84, + "latex": "\\mathbf { f } _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1026, + 357, + 1026, + 357, + 1057, + 297, + 1057 + ], + "score": 0.84, + "latex": "S ( \\% )" + }, + { + "category_id": 13, + "poly": [ + 1105, + 447, + 1132, + 447, + 1132, + 478, + 1105, + 478 + ], + "score": 0.84, + "latex": "\\mathbf { f } _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 296, + 798, + 319, + 798, + 319, + 828, + 296, + 828 + ], + "score": 0.83, + "latex": "\\mathbf { f } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 296, + 1851, + 357, + 1851, + 357, + 1882, + 296, + 1882 + ], + "score": 0.8, + "latex": "\\mathbf { x } ^ { \\mathrm { n e w } } ]" + }, + { + "category_id": 13, + "poly": [ + 495, + 1057, + 537, + 1057, + 537, + 1088, + 495, + 1088 + ], + "score": 0.79, + "latex": "\\mathbf { f } _ { - 2 }" + }, + { + "category_id": 13, + "poly": [ + 692, + 871, + 826, + 871, + 826, + 901, + 692, + 901 + ], + "score": 0.77, + "latex": "\\overline { { 1 0 ^ { - 3 } \\ : ( 1 0 ^ { - 3 } ) } }" + }, + { + "category_id": 13, + "poly": [ + 372, + 1026, + 416, + 1026, + 416, + 1059, + 372, + 1059 + ], + "score": 0.77, + "latex": "\\mathbf { f } _ { - 2 }" + }, + { + "category_id": 13, + "poly": [ + 1155, + 356, + 1182, + 356, + 1182, + 387, + 1155, + 387 + ], + "score": 0.76, + "latex": "\\mathbf { f } _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 846, + 358, + 859, + 358, + 859, + 383, + 846, + 383 + ], + "score": 0.74, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1328, + 1912, + 1407, + 1912, + 1407, + 1943, + 1328, + 1943 + ], + "score": 0.65, + "latex": "\\mathrm { [ C W + }" + }, + { + "category_id": 13, + "poly": [ + 868, + 870, + 1095, + 870, + 1095, + 901, + 868, + 901 + ], + "score": 0.64, + "latex": "\\overline { { 3 \\times 1 0 ^ { - 4 } ( 3 \\times 1 0 ^ { - 4 } ) } }" + }, + { + "category_id": 13, + "poly": [ + 1325, + 1821, + 1406, + 1821, + 1406, + 1852, + 1325, + 1852 + ], + "score": 0.63, + "latex": "\\mathrm { [ C W + }" + }, + { + "category_id": 13, + "poly": [ + 668, + 800, + 680, + 800, + 680, + 825, + 668, + 825 + ], + "score": 0.61, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 860, + 643, + 1085, + 643, + 1085, + 673, + 860, + 673 + ], + "score": 0.6, + "latex": "\\overline { { 3 \\times 1 0 ^ { - 4 } ( 3 \\times 1 0 ^ { - 4 } ) } }" + }, + { + "category_id": 13, + "poly": [ + 785, + 905, + 823, + 905, + 823, + 932, + 785, + 932 + ], + "score": 0.54, + "latex": "{ \\bf f } _ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1263, + 767, + 1398, + 767, + 1398, + 799, + 1263, + 799 + ], + "score": 0.52, + "latex": "\\mathrm { R - } 1 0 1 { - } \\mathrm { S } ( \\% )" + }, + { + "category_id": 13, + "poly": [ + 1181, + 643, + 1316, + 643, + 1316, + 674, + 1181, + 674 + ], + "score": 0.49, + "latex": "\\overline { { 1 0 ^ { - 4 } \\left( 1 0 ^ { - 4 } \\right) } }" + }, + { + "category_id": 13, + "poly": [ + 1125, + 872, + 1260, + 872, + 1260, + 901, + 1125, + 901 + ], + "score": 0.39, + "latex": "\\overline { { 1 0 ^ { - 4 } \\left( 1 0 ^ { - 4 } \\right) } }" + }, + { + "category_id": 13, + "poly": [ + 1004, + 905, + 1044, + 905, + 1044, + 932, + 1004, + 932 + ], + "score": 0.34, + "latex": "{ \\bf f } _ { - 1 }" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1133.0, + 709.0, + 1133.0, + 709.0, + 1156.0, + 465.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 1133.0, + 1210.0, + 1133.0, + 1210.0, + 1156.0, + 965.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1166.0, + 386.0, + 1166.0, + 386.0, + 1185.0, + 298.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 1166.0, + 430.0, + 1166.0, + 430.0, + 1184.0, + 405.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 1166.0, + 495.0, + 1166.0, + 495.0, + 1184.0, + 469.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 1166.0, + 566.0, + 1166.0, + 566.0, + 1184.0, + 540.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1166.0, + 632.0, + 1166.0, + 632.0, + 1184.0, + 606.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 1166.0, + 704.0, + 1166.0, + 704.0, + 1184.0, + 679.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 1166.0, + 765.0, + 1166.0, + 765.0, + 1184.0, + 739.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 1165.0, + 886.0, + 1165.0, + 886.0, + 1184.0, + 797.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1166.0, + 928.0, + 1166.0, + 928.0, + 1183.0, + 902.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1165.0, + 998.0, + 1165.0, + 998.0, + 1183.0, + 972.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 1165.0, + 1063.0, + 1165.0, + 1063.0, + 1183.0, + 1038.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 1165.0, + 1131.0, + 1165.0, + 1131.0, + 1183.0, + 1105.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 1165.0, + 1202.0, + 1165.0, + 1202.0, + 1183.0, + 1176.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 1166.0, + 1268.0, + 1166.0, + 1268.0, + 1183.0, + 1243.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 1176.0, + 1299.0, + 1176.0, + 1299.0, + 1264.0, + 300.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1301.0, + 1210.0, + 1357.0, + 1210.0, + 1357.0, + 1226.0, + 1301.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 1225.0, + 1372.0, + 1225.0, + 1372.0, + 1241.0, + 1305.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1248.0, + 731.0, + 1248.0, + 731.0, + 1333.0, + 332.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 1252.0, + 1229.0, + 1252.0, + 1229.0, + 1336.0, + 830.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 1282.0, + 1333.0, + 1282.0, + 1333.0, + 1290.0, + 1307.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 1295.0, + 1330.0, + 1295.0, + 1330.0, + 1304.0, + 1307.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1340.0, + 1300.0, + 1340.0, + 1300.0, + 1438.0, + 297.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1417.0, + 733.0, + 1417.0, + 733.0, + 1509.0, + 406.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1413.0, + 1226.0, + 1413.0, + 1226.0, + 1504.0, + 900.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1552.0, + 1352.0, + 1552.0, + 1352.0, + 1591.0, + 346.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 225.0, + 1115.0, + 225.0, + 1115.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1613.0, + 914.0, + 1613.0, + 914.0, + 1653.0, + 294.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 521.0, + 1405.0, + 521.0, + 1405.0, + 560.0, + 293.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 555.0, + 1404.0, + 555.0, + 1404.0, + 589.0, + 294.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 580.0, + 526.0, + 580.0, + 526.0, + 624.0, + 292.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 292.0, + 1406.0, + 292.0, + 1406.0, + 331.0, + 294.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 326.0, + 1407.0, + 326.0, + 1407.0, + 357.0, + 296.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 354.0, + 445.0, + 354.0, + 445.0, + 391.0, + 294.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 354.0, + 845.0, + 354.0, + 845.0, + 391.0, + 470.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 354.0, + 1154.0, + 354.0, + 1154.0, + 391.0, + 860.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 354.0, + 1408.0, + 354.0, + 1408.0, + 391.0, + 1183.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 385.0, + 1406.0, + 385.0, + 1406.0, + 423.0, + 294.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 415.0, + 483.0, + 415.0, + 483.0, + 452.0, + 292.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 415.0, + 956.0, + 415.0, + 956.0, + 452.0, + 511.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 415.0, + 1407.0, + 415.0, + 1407.0, + 452.0, + 985.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 447.0, + 1104.0, + 447.0, + 1104.0, + 481.0, + 295.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 447.0, + 1406.0, + 447.0, + 1406.0, + 481.0, + 1133.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 479.0, + 445.0, + 479.0, + 445.0, + 515.0, + 294.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1786.0, + 1406.0, + 1786.0, + 1406.0, + 1825.0, + 292.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1821.0, + 1324.0, + 1821.0, + 1324.0, + 1855.0, + 295.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1846.0, + 295.0, + 1846.0, + 295.0, + 1887.0, + 291.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1846.0, + 1405.0, + 1846.0, + 1405.0, + 1887.0, + 358.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1880.0, + 1233.0, + 1880.0, + 1233.0, + 1915.0, + 294.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 1880.0, + 1405.0, + 1880.0, + 1405.0, + 1915.0, + 1270.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1911.0, + 1327.0, + 1911.0, + 1327.0, + 1945.0, + 295.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1936.0, + 294.0, + 1936.0, + 294.0, + 1979.0, + 290.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1936.0, + 1409.0, + 1936.0, + 1409.0, + 1979.0, + 359.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 512.0, + 2002.0, + 512.0, + 2038.0, + 295.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 2002.0, + 803.0, + 2002.0, + 803.0, + 2038.0, + 549.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1680.0, + 1136.0, + 1680.0, + 1136.0, + 1714.0, + 292.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1310.0, + 1680.0, + 1404.0, + 1680.0, + 1404.0, + 1714.0, + 1310.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 1708.0, + 1404.0, + 1708.0, + 1404.0, + 1749.0, + 384.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1743.0, + 893.0, + 1743.0, + 893.0, + 1776.0, + 296.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 993.0, + 1409.0, + 993.0, + 1409.0, + 1030.0, + 293.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1026.0, + 371.0, + 1026.0, + 371.0, + 1059.0, + 358.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 1026.0, + 1206.0, + 1026.0, + 1206.0, + 1059.0, + 417.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 1026.0, + 1404.0, + 1026.0, + 1404.0, + 1059.0, + 1251.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1051.0, + 494.0, + 1051.0, + 494.0, + 1095.0, + 293.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 1051.0, + 550.0, + 1051.0, + 550.0, + 1095.0, + 538.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 762.0, + 1262.0, + 762.0, + 1262.0, + 805.0, + 292.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 762.0, + 1410.0, + 762.0, + 1410.0, + 805.0, + 1399.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 796.0, + 667.0, + 796.0, + 667.0, + 832.0, + 320.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 796.0, + 976.0, + 796.0, + 976.0, + 832.0, + 681.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 521.0, + 1405.0, + 521.0, + 1405.0, + 560.0, + 293.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 555.0, + 1404.0, + 555.0, + 1404.0, + 589.0, + 294.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 580.0, + 526.0, + 580.0, + 526.0, + 624.0, + 292.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1613.0, + 914.0, + 1613.0, + 914.0, + 1653.0, + 294.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 225.0, + 1115.0, + 225.0, + 1115.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 300, + 965, + 1404, + 965, + 1404, + 1391, + 300, + 1391 + ], + "score": 0.985, + "html": "
API nameAttacker/Evaluation MetricFace++AliYun
T-FPR= 10-3T-FPR=10-4T-FPR=10-3T-FPR=10
Original x2.040.510.500.00
Generated xnew4.210.530.500.00
CW (G-FPR=10-)StrawMan [CW + xnew ](G-FPR = 10-3)SemanticAdv (G-FPR =10-3)16.2410.7127.323.554.089.794.500.00
3.000.00
7.502.00
CW (G-FPR= 10-4)StrawMan [CW+ xnew ](G-FPR = 10-4)SemanticAdv (G-FPR =10-4)30.6121.3257.2215.829.1438.6612.504.50
7.5029.501.5017.50
Dong et al. (2018)Xie et al. (2019)CW (G-FPR<10-4)StrawMan [CW + xnew](G-FPR<10-4)SemanticAdv (G-FPR<10-4)30.7737.9521.0318.006.50
25.6421.5011.00
41.6227.6964.6324.3719.0012.00
15.3842.6910.003.00
35.5022.17
" + }, + { + "category_id": 5, + "poly": [ + 578, + 223, + 1157, + 223, + 1157, + 368, + 578, + 368 + ], + "score": 0.974, + "html": "
Mtest/MoptR-50-SR-101-SR-50-CR-101-C
R-50-S1.0000.0050.0000.000
R-101-S0.0001.0000.0000.000
R-50-C0.0000.0001.0000.000
R-101-C0.0000.0000.0001.000
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1474, + 1404, + 1474, + 1404, + 1658, + 297, + 1658 + ], + "score": 0.973 + }, + { + "category_id": 5, + "poly": [ + 926, + 503, + 1300, + 503, + 1300, + 675, + 926, + 675 + ], + "score": 0.968, + "html": "
MtestMoptR-101-S
R-50-S0.615 (0.862)
R-101-S1.000(1.000)
R-50-C0.570(0.837)
R-101-C0.6955(0.888)
" + }, + { + "category_id": 5, + "poly": [ + 391, + 503, + 766, + 503, + 766, + 675, + 391, + 675 + ], + "score": 0.965, + "html": "
MtestMoptR-101-S
R-50-S0.035 (0.108)
R-101-S1.000 (1.000)
R-50-C0.145 (0.202)
R-101-C0.085 (0.236)
" + }, + { + "category_id": 0, + "poly": [ + 298, + 881, + 912, + 881, + 912, + 919, + 298, + 919 + ], + "score": 0.911 + }, + { + "category_id": 2, + "poly": [ + 297, + 74, + 859, + 74, + 859, + 105, + 297, + 105 + ], + "score": 0.911 + }, + { + "category_id": 7, + "poly": [ + 415, + 686, + 737, + 686, + 737, + 718, + 415, + 718 + ], + "score": 0.887 + }, + { + "category_id": 1, + "poly": [ + 297, + 1672, + 1403, + 1672, + 1403, + 1737, + 297, + 1737 + ], + "score": 0.864 + }, + { + "category_id": 7, + "poly": [ + 949, + 686, + 1274, + 686, + 1274, + 718, + 949, + 718 + ], + "score": 0.86 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.851 + }, + { + "category_id": 1, + "poly": [ + 299, + 1406, + 1397, + 1406, + 1397, + 1467, + 299, + 1467 + ], + "score": 0.663 + }, + { + "category_id": 1, + "poly": [ + 297, + 733, + 1404, + 733, + 1404, + 826, + 297, + 826 + ], + "score": 0.635 + }, + { + "category_id": 1, + "poly": [ + 300, + 379, + 1407, + 379, + 1407, + 471, + 300, + 471 + ], + "score": 0.569 + }, + { + "category_id": 6, + "poly": [ + 300, + 379, + 1407, + 379, + 1407, + 471, + 300, + 471 + ], + "score": 0.425 + }, + { + "category_id": 6, + "poly": [ + 297, + 733, + 1404, + 733, + 1404, + 826, + 297, + 826 + ], + "score": 0.138 + }, + { + "category_id": 6, + "poly": [ + 299, + 1406, + 1397, + 1406, + 1397, + 1467, + 299, + 1467 + ], + "score": 0.119 + }, + { + "category_id": 13, + "poly": [ + 782, + 733, + 840, + 733, + 840, + 767, + 782, + 767 + ], + "score": 0.91, + "latex": "( i , j )" + }, + { + "category_id": 13, + "poly": [ + 798, + 379, + 855, + 379, + 855, + 412, + 798, + 412 + ], + "score": 0.91, + "latex": "( i , j )" + }, + { + "category_id": 13, + "poly": [ + 670, + 1597, + 723, + 1597, + 723, + 1624, + 670, + 1624 + ], + "score": 0.84, + "latex": "\\mathbf { x } ^ { \\mathrm { n e w } }" + }, + { + "category_id": 13, + "poly": [ + 503, + 413, + 519, + 413, + 519, + 443, + 503, + 443 + ], + "score": 0.82, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 506, + 767, + 523, + 767, + 523, + 797, + 506, + 797 + ], + "score": 0.82, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 819, + 414, + 832, + 414, + 832, + 438, + 819, + 438 + ], + "score": 0.75, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1224, + 409, + 1395, + 409, + 1395, + 440, + 1224, + 440 + ], + "score": 0.71, + "latex": "\\mathrm { { G } { - } \\mathrm { { F P R } } = 1 \\bar { 0 } ^ { - 4 } }" + }, + { + "category_id": 13, + "poly": [ + 601, + 684, + 737, + 684, + 737, + 717, + 601, + 717 + ], + "score": 0.69, + "latex": "\\mathrm { T } \\mathrm { - } \\mathrm { F P R } { = } 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 373, + 439, + 458, + 439, + 458, + 469, + 373, + 469 + ], + "score": 0.68, + "latex": "= 1 0 ^ { \\overline { { - 4 } } }" + }, + { + "category_id": 13, + "poly": [ + 1000, + 1438, + 1039, + 1438, + 1039, + 1467, + 1000, + 1467 + ], + "score": 0.68, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 831, + 768, + 844, + 768, + 844, + 792, + 831, + 792 + ], + "score": 0.68, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 367, + 1302, + 514, + 1302, + 514, + 1329, + 367, + 1329 + ], + "score": 0.64, + "latex": "\\mathrm { ( G - F P R < 1 0 ^ { - 4 } }" + }, + { + "category_id": 13, + "poly": [ + 367, + 1167, + 515, + 1167, + 515, + 1193, + 367, + 1193 + ], + "score": 0.63, + "latex": "\\overline { { \\mathrm { ( G - F P R = 1 0 ^ { - 4 } } } } ." + }, + { + "category_id": 13, + "poly": [ + 741, + 993, + 881, + 993, + 881, + 1020, + 741, + 1020 + ], + "score": 0.55, + "latex": "\\overline { { \\mathrm { T } \\mathrm { F P R } = 1 0 ^ { - 3 } } }" + }, + { + "category_id": 13, + "poly": [ + 453, + 1222, + 600, + 1222, + 600, + 1249, + 453, + 1249 + ], + "score": 0.52, + "latex": "\\mathrm { ( G - F P R = 1 0 ^ { - 4 } } ." + }, + { + "category_id": 13, + "poly": [ + 450, + 1138, + 602, + 1138, + 602, + 1165, + 450, + 1165 + ], + "score": 0.51, + "latex": "( \\mathbf { G } \\mathbf { - } \\mathbf { F P R } = 1 0 ^ { - 3 } )" + }, + { + "category_id": 13, + "poly": [ + 1134, + 684, + 1273, + 684, + 1273, + 717, + 1134, + 717 + ], + "score": 0.49, + "latex": "\\mathrm { T } { \\cdot } \\mathrm { F P R } { = } 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 1077, + 992, + 1217, + 992, + 1217, + 1020, + 1077, + 1020 + ], + "score": 0.46, + "latex": "\\overline { { \\mathrm { T } \\mathrm { F P R } = 1 0 ^ { - 3 } } }" + }, + { + "category_id": 13, + "poly": [ + 552, + 1108, + 709, + 1108, + 709, + 1137, + 552, + 1137 + ], + "score": 0.44, + "latex": "( \\mathrm { G } \\mathrm { - F P R } = 1 0 ^ { - 3 } )" + }, + { + "category_id": 13, + "poly": [ + 368, + 1082, + 516, + 1082, + 516, + 1109, + 368, + 1109 + ], + "score": 0.38, + "latex": "\\overline { { \\mathrm { ( G - F P R = 1 0 ^ { - 3 } } } } ," + }, + { + "category_id": 13, + "poly": [ + 909, + 993, + 1050, + 993, + 1050, + 1021, + 909, + 1021 + ], + "score": 0.35, + "latex": "\\overline { { \\mathrm { T } \\mathrm { F P R } = 1 0 ^ { - 4 } } }" + }, + { + "category_id": 13, + "poly": [ + 451, + 684, + 591, + 684, + 591, + 718, + 451, + 718 + ], + "score": 0.34, + "latex": "\\mathrm { G } \\mathrm { - } \\mathrm { F P R } { = } 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 483, + 1600, + 505, + 1600, + 505, + 1624, + 483, + 1624 + ], + "score": 0.32, + "latex": "\\mathbf { x }" + }, + { + "category_id": 13, + "poly": [ + 554, + 1195, + 707, + 1195, + 707, + 1222, + 554, + 1222 + ], + "score": 0.27, + "latex": "\\mathrm { ( G - F P R = 1 0 ^ { - 4 } } ," + }, + { + "category_id": 15, + "poly": [ + 290.0, + 877.0, + 916.0, + 877.0, + 916.0, + 928.0, + 290.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 683.0, + 450.0, + 683.0, + 450.0, + 719.0, + 413.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 683.0, + 600.0, + 683.0, + 600.0, + 719.0, + 592.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 683.0, + 742.0, + 683.0, + 742.0, + 719.0, + 738.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 683.0, + 1133.0, + 683.0, + 1133.0, + 719.0, + 948.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 683.0, + 1278.0, + 683.0, + 1278.0, + 719.0, + 1274.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2122.0, + 831.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 377.0, + 797.0, + 377.0, + 797.0, + 414.0, + 293.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 377.0, + 1405.0, + 377.0, + 1405.0, + 414.0, + 856.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 405.0, + 502.0, + 405.0, + 502.0, + 445.0, + 290.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 405.0, + 818.0, + 405.0, + 818.0, + 445.0, + 520.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 405.0, + 1223.0, + 405.0, + 1223.0, + 445.0, + 833.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1396.0, + 405.0, + 1409.0, + 405.0, + 1409.0, + 445.0, + 1396.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 437.0, + 372.0, + 437.0, + 372.0, + 472.0, + 291.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 437.0, + 467.0, + 437.0, + 467.0, + 472.0, + 459.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 731.0, + 781.0, + 731.0, + 781.0, + 769.0, + 294.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 731.0, + 1405.0, + 731.0, + 1405.0, + 769.0, + 841.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 764.0, + 505.0, + 764.0, + 505.0, + 798.0, + 292.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 764.0, + 830.0, + 764.0, + 830.0, + 798.0, + 524.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 764.0, + 1405.0, + 764.0, + 1405.0, + 798.0, + 845.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 793.0, + 404.0, + 793.0, + 404.0, + 828.0, + 294.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1405.0, + 1402.0, + 1405.0, + 1402.0, + 1440.0, + 295.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1438.0, + 999.0, + 1438.0, + 999.0, + 1469.0, + 296.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 1438.0, + 1049.0, + 1438.0, + 1049.0, + 1469.0, + 1040.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1473.0, + 1406.0, + 1473.0, + 1406.0, + 1509.0, + 294.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1505.0, + 1405.0, + 1505.0, + 1405.0, + 1537.0, + 296.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1536.0, + 1405.0, + 1536.0, + 1405.0, + 1567.0, + 296.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1565.0, + 1406.0, + 1565.0, + 1406.0, + 1600.0, + 291.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1593.0, + 482.0, + 1593.0, + 482.0, + 1632.0, + 292.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 1593.0, + 669.0, + 1593.0, + 669.0, + 1632.0, + 506.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 1593.0, + 1406.0, + 1593.0, + 1406.0, + 1632.0, + 724.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1624.0, + 604.0, + 1624.0, + 604.0, + 1662.0, + 295.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1672.0, + 1404.0, + 1672.0, + 1404.0, + 1708.0, + 295.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1701.0, + 898.0, + 1701.0, + 898.0, + 1737.0, + 295.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1405.0, + 1402.0, + 1405.0, + 1402.0, + 1440.0, + 295.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1438.0, + 999.0, + 1438.0, + 999.0, + 1469.0, + 296.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 1438.0, + 1049.0, + 1438.0, + 1049.0, + 1469.0, + 1040.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 731.0, + 781.0, + 731.0, + 781.0, + 769.0, + 294.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 731.0, + 1405.0, + 731.0, + 1405.0, + 769.0, + 841.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 764.0, + 505.0, + 764.0, + 505.0, + 798.0, + 292.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 764.0, + 830.0, + 764.0, + 830.0, + 798.0, + 524.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 764.0, + 1405.0, + 764.0, + 1405.0, + 798.0, + 845.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 793.0, + 404.0, + 793.0, + 404.0, + 828.0, + 294.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 377.0, + 797.0, + 377.0, + 797.0, + 414.0, + 293.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 377.0, + 1405.0, + 377.0, + 1405.0, + 414.0, + 856.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 405.0, + 502.0, + 405.0, + 502.0, + 445.0, + 290.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 405.0, + 818.0, + 405.0, + 818.0, + 445.0, + 520.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 405.0, + 1223.0, + 405.0, + 1223.0, + 445.0, + 833.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1396.0, + 405.0, + 1409.0, + 405.0, + 1409.0, + 445.0, + 1396.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 437.0, + 372.0, + 437.0, + 372.0, + 472.0, + 291.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 437.0, + 467.0, + 437.0, + 467.0, + 472.0, + 459.0, + 472.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 307, + 757, + 1399, + 757, + 1399, + 1425, + 307, + 1425 + ], + "score": 0.976 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 859, + 75, + 859, + 105, + 297, + 105 + ], + "score": 0.853 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.835 + }, + { + "category_id": 4, + "poly": [ + 294, + 1465, + 1404, + 1465, + 1404, + 1531, + 294, + 1531 + ], + "score": 0.765 + }, + { + "category_id": 1, + "poly": [ + 294, + 1465, + 1404, + 1465, + 1404, + 1531, + 294, + 1531 + ], + "score": 0.233 + }, + { + "category_id": 15, + "poly": [ + 306.0, + 759.0, + 400.0, + 759.0, + 400.0, + 782.0, + 306.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 766.0, + 790.0, + 766.0, + 790.0, + 776.0, + 781.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 760.0, + 843.0, + 760.0, + 843.0, + 782.0, + 800.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 757.0, + 948.0, + 757.0, + 948.0, + 780.0, + 855.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 766.0, + 1335.0, + 766.0, + 1335.0, + 773.0, + 1329.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1348.0, + 761.0, + 1389.0, + 761.0, + 1389.0, + 779.0, + 1348.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 783.0, + 725.0, + 783.0, + 725.0, + 803.0, + 309.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 782.0, + 1273.0, + 782.0, + 1273.0, + 802.0, + 857.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 819.0, + 479.0, + 819.0, + 479.0, + 899.0, + 408.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 819.0, + 752.0, + 819.0, + 752.0, + 898.0, + 680.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 819.0, + 1027.0, + 819.0, + 1027.0, + 896.0, + 954.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1227.0, + 818.0, + 1299.0, + 818.0, + 1299.0, + 895.0, + 1227.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 898.0, + 358.0, + 898.0, + 358.0, + 915.0, + 314.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 891.0, + 505.0, + 891.0, + 505.0, + 921.0, + 451.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 896.0, + 566.0, + 896.0, + 566.0, + 914.0, + 521.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 898.0, + 636.0, + 898.0, + 636.0, + 917.0, + 587.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 893.0, + 777.0, + 893.0, + 777.0, + 919.0, + 725.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 896.0, + 841.0, + 896.0, + 841.0, + 914.0, + 796.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 895.0, + 909.0, + 895.0, + 909.0, + 914.0, + 860.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 891.0, + 1052.0, + 891.0, + 1052.0, + 917.0, + 996.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 894.0, + 1117.0, + 894.0, + 1117.0, + 914.0, + 1067.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 894.0, + 1182.0, + 894.0, + 1182.0, + 915.0, + 1134.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 891.0, + 1325.0, + 891.0, + 1325.0, + 918.0, + 1271.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1343.0, + 895.0, + 1392.0, + 895.0, + 1392.0, + 914.0, + 1343.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 924.0, + 736.0, + 924.0, + 736.0, + 944.0, + 422.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 923.0, + 1283.0, + 923.0, + 1283.0, + 943.0, + 971.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 961.0, + 476.0, + 961.0, + 476.0, + 1024.0, + 412.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 958.0, + 750.0, + 958.0, + 750.0, + 1027.0, + 680.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 947.0, + 1031.0, + 947.0, + 1031.0, + 1028.0, + 952.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 963.0, + 1293.0, + 963.0, + 1293.0, + 1019.0, + 1234.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 1025.0, + 360.0, + 1025.0, + 360.0, + 1047.0, + 310.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1025.0, + 502.0, + 1025.0, + 502.0, + 1047.0, + 449.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 1027.0, + 570.0, + 1027.0, + 570.0, + 1046.0, + 521.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1027.0, + 633.0, + 1027.0, + 633.0, + 1044.0, + 587.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 1025.0, + 775.0, + 1025.0, + 775.0, + 1047.0, + 728.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 1027.0, + 841.0, + 1027.0, + 841.0, + 1044.0, + 795.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1025.0, + 906.0, + 1025.0, + 906.0, + 1043.0, + 860.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1022.0, + 1049.0, + 1022.0, + 1049.0, + 1047.0, + 1000.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 1025.0, + 1115.0, + 1025.0, + 1115.0, + 1043.0, + 1069.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1026.0, + 1181.0, + 1026.0, + 1181.0, + 1044.0, + 1133.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 1024.0, + 1321.0, + 1024.0, + 1321.0, + 1046.0, + 1276.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 1025.0, + 1386.0, + 1025.0, + 1386.0, + 1042.0, + 1345.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1053.0, + 735.0, + 1053.0, + 735.0, + 1073.0, + 420.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 1053.0, + 1284.0, + 1053.0, + 1284.0, + 1073.0, + 968.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1085.0, + 402.0, + 1085.0, + 402.0, + 1109.0, + 307.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 1088.0, + 842.0, + 1088.0, + 842.0, + 1106.0, + 801.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 1083.0, + 949.0, + 1083.0, + 949.0, + 1110.0, + 855.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1366.0, + 1091.0, + 1388.0, + 1091.0, + 1388.0, + 1104.0, + 1366.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 1114.0, + 335.0, + 1114.0, + 335.0, + 1126.0, + 313.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 1112.0, + 572.0, + 1112.0, + 572.0, + 1126.0, + 397.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 1112.0, + 971.0, + 1112.0, + 971.0, + 1125.0, + 936.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 1114.0, + 1058.0, + 1114.0, + 1058.0, + 1126.0, + 978.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1151.0, + 480.0, + 1151.0, + 480.0, + 1228.0, + 408.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 1155.0, + 752.0, + 1155.0, + 752.0, + 1227.0, + 682.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1150.0, + 1027.0, + 1150.0, + 1027.0, + 1230.0, + 956.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 1158.0, + 1299.0, + 1158.0, + 1299.0, + 1226.0, + 1229.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1227.0, + 359.0, + 1227.0, + 359.0, + 1245.0, + 317.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 1224.0, + 503.0, + 1224.0, + 503.0, + 1249.0, + 453.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 1225.0, + 571.0, + 1225.0, + 571.0, + 1246.0, + 520.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 1227.0, + 633.0, + 1227.0, + 633.0, + 1245.0, + 589.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 1225.0, + 775.0, + 1225.0, + 775.0, + 1246.0, + 728.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1223.0, + 845.0, + 1223.0, + 845.0, + 1248.0, + 794.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 1227.0, + 906.0, + 1227.0, + 906.0, + 1245.0, + 863.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1222.0, + 1049.0, + 1222.0, + 1049.0, + 1249.0, + 1000.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 1225.0, + 1117.0, + 1225.0, + 1117.0, + 1246.0, + 1067.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 1227.0, + 1180.0, + 1227.0, + 1180.0, + 1245.0, + 1135.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 1224.0, + 1319.0, + 1224.0, + 1319.0, + 1246.0, + 1276.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1343.0, + 1223.0, + 1391.0, + 1223.0, + 1391.0, + 1248.0, + 1343.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1254.0, + 736.0, + 1254.0, + 736.0, + 1274.0, + 423.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 1254.0, + 1284.0, + 1254.0, + 1284.0, + 1274.0, + 970.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1284.0, + 481.0, + 1284.0, + 481.0, + 1368.0, + 404.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 1289.0, + 752.0, + 1289.0, + 752.0, + 1366.0, + 679.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 1288.0, + 1027.0, + 1288.0, + 1027.0, + 1368.0, + 957.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 1287.0, + 1302.0, + 1287.0, + 1302.0, + 1366.0, + 1225.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 1365.0, + 359.0, + 1365.0, + 359.0, + 1383.0, + 313.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1361.0, + 505.0, + 1361.0, + 505.0, + 1388.0, + 448.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 1364.0, + 567.0, + 1364.0, + 567.0, + 1382.0, + 521.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 1365.0, + 634.0, + 1365.0, + 634.0, + 1383.0, + 589.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 1361.0, + 781.0, + 1361.0, + 781.0, + 1388.0, + 725.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 1363.0, + 843.0, + 1363.0, + 843.0, + 1384.0, + 795.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 1364.0, + 906.0, + 1364.0, + 906.0, + 1382.0, + 863.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1359.0, + 1049.0, + 1359.0, + 1049.0, + 1385.0, + 1000.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 1364.0, + 1106.0, + 1364.0, + 1106.0, + 1382.0, + 1069.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 1365.0, + 1181.0, + 1365.0, + 1181.0, + 1382.0, + 1142.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 1359.0, + 1322.0, + 1359.0, + 1322.0, + 1385.0, + 1273.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 1364.0, + 1389.0, + 1364.0, + 1389.0, + 1382.0, + 1345.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1393.0, + 735.0, + 1393.0, + 735.0, + 1413.0, + 423.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 1392.0, + 1286.0, + 1392.0, + 1286.0, + 1412.0, + 968.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1464.0, + 1404.0, + 1464.0, + 1404.0, + 1502.0, + 293.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1495.0, + 1358.0, + 1495.0, + 1358.0, + 1533.0, + 294.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1464.0, + 1404.0, + 1464.0, + 1404.0, + 1502.0, + 293.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1495.0, + 1358.0, + 1495.0, + 1358.0, + 1533.0, + 294.0, + 1533.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 18, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 331, + 315, + 1367, + 315, + 1367, + 1903, + 331, + 1903 + ], + "score": 0.97 + }, + { + "category_id": 4, + "poly": [ + 375, + 1904, + 1321, + 1904, + 1321, + 1937, + 375, + 1937 + ], + "score": 0.909 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 866, + 2088, + 866, + 2113, + 835, + 2113 + ], + "score": 0.856 + }, + { + "category_id": 0, + "poly": [ + 300, + 228, + 701, + 228, + 701, + 261, + 300, + 261 + ], + "score": 0.851 + }, + { + "category_id": 3, + "poly": [ + 331, + 315, + 1368, + 315, + 1368, + 1900, + 331, + 1900 + ], + "score": 0.105 + }, + { + "category_id": 13, + "poly": [ + 1133, + 1903, + 1305, + 1903, + 1305, + 1934, + 1133, + 1934 + ], + "score": 0.68, + "latex": "\\mathrm { ( G { - } F P R = 1 0 ^ { - 3 } }" + }, + { + "category_id": 13, + "poly": [ + 1326, + 774, + 1363, + 774, + 1363, + 792, + 1326, + 792 + ], + "score": 0.33, + "latex": "\\scriptstyle ( \\mathbf { a } = 0 )" + }, + { + "category_id": 13, + "poly": [ + 1327, + 389, + 1363, + 389, + 1363, + 408, + 1327, + 408 + ], + "score": 0.3, + "latex": "\\scriptstyle ( \\alpha = 0 )" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 307.0, + 411.0, + 307.0, + 411.0, + 356.0, + 343.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 305.0, + 482.0, + 305.0, + 482.0, + 361.0, + 418.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 314.0, + 1131.0, + 314.0, + 1131.0, + 342.0, + 743.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 363.0, + 1364.0, + 363.0, + 1364.0, + 397.0, + 1170.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 383.0, + 1326.0, + 383.0, + 1326.0, + 412.0, + 1159.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 383.0, + 1372.0, + 383.0, + 1372.0, + 412.0, + 1364.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 427.0, + 1353.0, + 427.0, + 1353.0, + 455.0, + 1181.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 453.0, + 567.0, + 453.0, + 567.0, + 467.0, + 554.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 451.0, + 629.0, + 451.0, + 629.0, + 465.0, + 609.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 453.0, + 650.0, + 453.0, + 650.0, + 465.0, + 632.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 450.0, + 812.0, + 450.0, + 812.0, + 469.0, + 771.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 451.0, + 874.0, + 451.0, + 874.0, + 467.0, + 856.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 453.0, + 895.0, + 453.0, + 895.0, + 465.0, + 881.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 450.0, + 980.0, + 450.0, + 980.0, + 469.0, + 938.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 453.0, + 1039.0, + 453.0, + 1039.0, + 467.0, + 1023.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 444.0, + 1367.0, + 444.0, + 1367.0, + 478.0, + 1169.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 462.0, + 1372.0, + 462.0, + 1372.0, + 493.0, + 1158.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 491.0, + 587.0, + 491.0, + 587.0, + 526.0, + 509.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 497.0, + 654.0, + 497.0, + 654.0, + 522.0, + 597.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 495.0, + 842.0, + 495.0, + 842.0, + 523.0, + 684.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 497.0, + 908.0, + 497.0, + 908.0, + 522.0, + 847.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 497.0, + 984.0, + 497.0, + 984.0, + 522.0, + 932.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 495.0, + 1080.0, + 495.0, + 1080.0, + 523.0, + 1003.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 494.0, + 1163.0, + 494.0, + 1163.0, + 523.0, + 1096.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 515.0, + 569.0, + 515.0, + 569.0, + 539.0, + 526.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 513.0, + 648.0, + 513.0, + 648.0, + 539.0, + 604.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 513.0, + 915.0, + 513.0, + 915.0, + 541.0, + 838.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 515.0, + 989.0, + 515.0, + 989.0, + 539.0, + 927.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 513.0, + 1183.0, + 513.0, + 1183.0, + 541.0, + 1074.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 566.0, + 569.0, + 566.0, + 569.0, + 580.0, + 554.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 566.0, + 625.0, + 566.0, + 625.0, + 578.0, + 611.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 564.0, + 711.0, + 564.0, + 711.0, + 578.0, + 691.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 566.0, + 734.0, + 566.0, + 734.0, + 578.0, + 718.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 562.0, + 817.0, + 562.0, + 817.0, + 585.0, + 771.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 562.0, + 899.0, + 562.0, + 899.0, + 581.0, + 872.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 562.0, + 984.0, + 562.0, + 984.0, + 583.0, + 936.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 564.0, + 1042.0, + 564.0, + 1042.0, + 583.0, + 1018.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 567.0, + 1062.0, + 567.0, + 1062.0, + 581.0, + 1044.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 618.0, + 888.0, + 618.0, + 888.0, + 636.0, + 860.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 638.0, + 789.0, + 638.0, + 789.0, + 652.0, + 774.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 636.0, + 874.0, + 636.0, + 874.0, + 652.0, + 856.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 638.0, + 959.0, + 638.0, + 959.0, + 652.0, + 938.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 640.0, + 980.0, + 640.0, + 980.0, + 652.0, + 966.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 640.0, + 1037.0, + 640.0, + 1037.0, + 652.0, + 1023.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 680.0, + 583.0, + 680.0, + 583.0, + 710.0, + 514.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 676.0, + 754.0, + 676.0, + 754.0, + 710.0, + 674.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 682.0, + 831.0, + 682.0, + 831.0, + 705.0, + 760.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 680.0, + 1002.0, + 680.0, + 1002.0, + 708.0, + 918.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 676.0, + 1068.0, + 676.0, + 1068.0, + 709.0, + 1012.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 680.0, + 1234.0, + 680.0, + 1234.0, + 705.0, + 1184.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 699.0, + 592.0, + 699.0, + 592.0, + 727.0, + 505.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 698.0, + 668.0, + 698.0, + 668.0, + 727.0, + 599.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 698.0, + 840.0, + 698.0, + 840.0, + 727.0, + 753.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 696.0, + 1167.0, + 696.0, + 1167.0, + 729.0, + 918.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 696.0, + 1231.0, + 696.0, + 1231.0, + 722.0, + 1188.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 752.0, + 411.0, + 752.0, + 411.0, + 800.0, + 347.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 756.0, + 473.0, + 756.0, + 473.0, + 789.0, + 423.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 764.0, + 810.0, + 764.0, + 810.0, + 784.0, + 773.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 768.0, + 874.0, + 768.0, + 874.0, + 782.0, + 858.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 768.0, + 897.0, + 768.0, + 897.0, + 780.0, + 881.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 768.0, + 980.0, + 768.0, + 980.0, + 780.0, + 963.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 749.0, + 1325.0, + 749.0, + 1325.0, + 798.0, + 1163.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 749.0, + 1369.0, + 749.0, + 1369.0, + 798.0, + 1364.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 812.0, + 1353.0, + 812.0, + 1353.0, + 840.0, + 1181.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 840.0, + 794.0, + 840.0, + 794.0, + 852.0, + 776.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 840.0, + 874.0, + 840.0, + 874.0, + 852.0, + 856.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 840.0, + 899.0, + 840.0, + 899.0, + 852.0, + 881.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 838.0, + 984.0, + 838.0, + 984.0, + 858.0, + 961.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 840.0, + 1039.0, + 840.0, + 1039.0, + 852.0, + 1019.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 840.0, + 1062.0, + 840.0, + 1062.0, + 852.0, + 1044.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 831.0, + 1366.0, + 831.0, + 1366.0, + 859.0, + 1169.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 847.0, + 1372.0, + 847.0, + 1372.0, + 879.0, + 1158.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 886.0, + 588.0, + 886.0, + 588.0, + 928.0, + 519.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 887.0, + 659.0, + 887.0, + 659.0, + 928.0, + 604.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 887.0, + 744.0, + 887.0, + 744.0, + 912.0, + 684.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 884.0, + 840.0, + 884.0, + 840.0, + 907.0, + 746.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 882.0, + 918.0, + 882.0, + 918.0, + 928.0, + 844.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 884.0, + 987.0, + 884.0, + 987.0, + 926.0, + 925.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 880.0, + 1179.0, + 880.0, + 1179.0, + 930.0, + 998.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 958.0, + 645.0, + 958.0, + 645.0, + 967.0, + 636.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 956.0, + 732.0, + 956.0, + 732.0, + 970.0, + 716.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 956.0, + 817.0, + 956.0, + 817.0, + 970.0, + 796.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 954.0, + 897.0, + 954.0, + 897.0, + 974.0, + 856.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 956.0, + 959.0, + 956.0, + 959.0, + 970.0, + 938.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 956.0, + 980.0, + 956.0, + 980.0, + 970.0, + 961.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 970.0, + 886.0, + 970.0, + 886.0, + 984.0, + 870.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 1032.0, + 787.0, + 1032.0, + 787.0, + 1040.0, + 778.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 1032.0, + 872.0, + 1032.0, + 872.0, + 1040.0, + 863.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 1074.0, + 572.0, + 1074.0, + 572.0, + 1098.0, + 515.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 1074.0, + 670.0, + 1074.0, + 670.0, + 1102.0, + 588.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 1070.0, + 835.0, + 1070.0, + 835.0, + 1098.0, + 760.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 1074.0, + 1066.0, + 1074.0, + 1066.0, + 1098.0, + 1019.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 1091.0, + 586.0, + 1091.0, + 586.0, + 1120.0, + 499.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 1090.0, + 659.0, + 1090.0, + 659.0, + 1120.0, + 592.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 1090.0, + 840.0, + 1090.0, + 840.0, + 1118.0, + 755.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1090.0, + 1165.0, + 1090.0, + 1165.0, + 1120.0, + 920.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 1090.0, + 1229.0, + 1090.0, + 1229.0, + 1116.0, + 1184.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 1135.0, + 1362.0, + 1135.0, + 1362.0, + 1169.0, + 1167.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 1156.0, + 1367.0, + 1156.0, + 1367.0, + 1185.0, + 1160.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 1193.0, + 1352.0, + 1193.0, + 1352.0, + 1231.0, + 1174.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 1225.0, + 893.0, + 1225.0, + 893.0, + 1237.0, + 874.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 1225.0, + 952.0, + 1225.0, + 952.0, + 1237.0, + 938.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 1218.0, + 1362.0, + 1218.0, + 1362.0, + 1246.0, + 1165.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1234.0, + 1367.0, + 1234.0, + 1367.0, + 1267.0, + 1156.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 1265.0, + 581.0, + 1265.0, + 581.0, + 1294.0, + 508.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1263.0, + 653.0, + 1263.0, + 653.0, + 1296.0, + 590.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1263.0, + 838.0, + 1263.0, + 838.0, + 1298.0, + 678.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1267.0, + 904.0, + 1267.0, + 904.0, + 1292.0, + 842.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 1269.0, + 982.0, + 1269.0, + 982.0, + 1292.0, + 931.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1265.0, + 1074.0, + 1265.0, + 1074.0, + 1294.0, + 1000.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1267.0, + 1160.0, + 1267.0, + 1160.0, + 1294.0, + 1090.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1287.0, + 565.0, + 1287.0, + 565.0, + 1311.0, + 522.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1285.0, + 643.0, + 1285.0, + 643.0, + 1311.0, + 601.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 1281.0, + 912.0, + 1281.0, + 912.0, + 1315.0, + 836.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1287.0, + 986.0, + 1287.0, + 986.0, + 1311.0, + 925.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 1279.0, + 1178.0, + 1279.0, + 1178.0, + 1319.0, + 1069.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 1338.0, + 975.0, + 1338.0, + 975.0, + 1352.0, + 959.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 1411.0, + 789.0, + 1411.0, + 789.0, + 1424.0, + 774.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 1452.0, + 577.0, + 1452.0, + 577.0, + 1482.0, + 512.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1450.0, + 676.0, + 1450.0, + 676.0, + 1482.0, + 591.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1452.0, + 829.0, + 1452.0, + 829.0, + 1480.0, + 757.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1446.0, + 999.0, + 1446.0, + 999.0, + 1481.0, + 914.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1452.0, + 1064.0, + 1452.0, + 1064.0, + 1476.0, + 1012.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 1446.0, + 1169.0, + 1446.0, + 1169.0, + 1481.0, + 1083.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1450.0, + 1232.0, + 1450.0, + 1232.0, + 1480.0, + 1179.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 1471.0, + 586.0, + 1471.0, + 586.0, + 1499.0, + 501.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 1469.0, + 666.0, + 1469.0, + 666.0, + 1499.0, + 599.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 1470.0, + 834.0, + 1470.0, + 834.0, + 1497.0, + 749.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1465.0, + 1085.0, + 1465.0, + 1085.0, + 1498.0, + 919.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 1468.0, + 1161.0, + 1468.0, + 1161.0, + 1496.0, + 1087.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 1466.0, + 1229.0, + 1466.0, + 1229.0, + 1498.0, + 1181.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 1508.0, + 498.0, + 1508.0, + 498.0, + 1593.0, + 336.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1522.0, + 1358.0, + 1522.0, + 1358.0, + 1550.0, + 1169.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1540.0, + 1367.0, + 1540.0, + 1367.0, + 1569.0, + 1156.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 1584.0, + 1350.0, + 1584.0, + 1350.0, + 1612.0, + 1176.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 1601.0, + 1364.0, + 1601.0, + 1364.0, + 1635.0, + 1163.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1619.0, + 1367.0, + 1619.0, + 1367.0, + 1652.0, + 1158.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 1654.0, + 586.0, + 1654.0, + 586.0, + 1702.0, + 514.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1652.0, + 658.0, + 1652.0, + 658.0, + 1700.0, + 597.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 1654.0, + 915.0, + 1654.0, + 915.0, + 1700.0, + 680.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1672.0, + 982.0, + 1672.0, + 982.0, + 1696.0, + 922.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 1652.0, + 1177.0, + 1652.0, + 1177.0, + 1702.0, + 995.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1726.0, + 808.0, + 1726.0, + 808.0, + 1740.0, + 792.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 1724.0, + 952.0, + 1724.0, + 952.0, + 1740.0, + 938.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 1728.0, + 973.0, + 1728.0, + 973.0, + 1742.0, + 959.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1798.0, + 725.0, + 1798.0, + 725.0, + 1811.0, + 711.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 1793.0, + 812.0, + 1793.0, + 812.0, + 1816.0, + 787.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 1800.0, + 973.0, + 1800.0, + 973.0, + 1812.0, + 959.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 1842.0, + 574.0, + 1842.0, + 574.0, + 1872.0, + 506.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 1844.0, + 666.0, + 1844.0, + 666.0, + 1874.0, + 583.0, + 1874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 1838.0, + 747.0, + 1838.0, + 747.0, + 1873.0, + 671.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1842.0, + 831.0, + 1842.0, + 831.0, + 1870.0, + 757.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1844.0, + 998.0, + 1844.0, + 998.0, + 1874.0, + 915.0, + 1874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 1836.0, + 1171.0, + 1836.0, + 1171.0, + 1876.0, + 1081.0, + 1876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1844.0, + 1227.0, + 1844.0, + 1227.0, + 1869.0, + 1179.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 1862.0, + 583.0, + 1862.0, + 583.0, + 1891.0, + 496.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1863.0, + 656.0, + 1863.0, + 656.0, + 1888.0, + 590.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 1860.0, + 837.0, + 1860.0, + 837.0, + 1890.0, + 750.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 1857.0, + 1085.0, + 1857.0, + 1085.0, + 1892.0, + 918.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 1857.0, + 1162.0, + 1857.0, + 1162.0, + 1892.0, + 1086.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 1858.0, + 1227.0, + 1858.0, + 1227.0, + 1890.0, + 1181.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.25, + 677.5, + 673.25, + 677.5, + 673.25, + 707.5, + 595.25, + 707.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 679.5, + 906.0, + 679.5, + 906.0, + 708.5, + 837.0, + 708.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.25, + 677.5, + 1172.25, + 677.5, + 1172.25, + 705.5, + 1085.25, + 705.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 958.5, + 796.0, + 958.5, + 796.0, + 983.5, + 783.0, + 983.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.75, + 1069.5, + 749.75, + 1069.5, + 749.75, + 1099.5, + 674.75, + 1099.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.75, + 1071.0, + 910.75, + 1071.0, + 910.75, + 1101.0, + 847.75, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.25, + 1071.5, + 998.25, + 1071.5, + 998.25, + 1101.0, + 918.25, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.75, + 1071.5, + 1170.75, + 1071.5, + 1170.75, + 1099.5, + 1086.75, + 1099.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.25, + 1072.5, + 1231.25, + 1072.5, + 1231.25, + 1096.5, + 1180.25, + 1096.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 1216.5, + 813.0, + 1216.5, + 813.0, + 1248.5, + 763.0, + 1248.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.75, + 1449.5, + 747.75, + 1449.5, + 747.75, + 1479.5, + 671.75, + 1479.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1451.0, + 900.0, + 1451.0, + 900.0, + 1480.5, + 837.0, + 1480.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 1652.5, + 977.0, + 1652.5, + 977.0, + 1676.5, + 927.0, + 1676.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1841.5, + 909.0, + 1841.5, + 909.0, + 1871.5, + 842.0, + 1871.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.75, + 1843.0, + 1062.75, + 1843.0, + 1062.75, + 1873.0, + 1011.75, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1902.0, + 1132.0, + 1902.0, + 1132.0, + 1939.0, + 373.0, + 1939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 1902.0, + 1323.0, + 1902.0, + 1323.0, + 1939.0, + 1306.0, + 1939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2083.0, + 872.0, + 2083.0, + 872.0, + 2124.0, + 829.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 222.0, + 704.0, + 222.0, + 704.0, + 269.0, + 295.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 311.0, + 409.0, + 311.0, + 409.0, + 356.0, + 345.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 305.0, + 483.0, + 305.0, + 483.0, + 361.0, + 418.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 314.0, + 1132.0, + 314.0, + 1132.0, + 342.0, + 743.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 363.0, + 1365.0, + 363.0, + 1365.0, + 397.0, + 1171.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 383.0, + 1326.0, + 383.0, + 1326.0, + 412.0, + 1158.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 383.0, + 1371.0, + 383.0, + 1371.0, + 412.0, + 1364.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 421.0, + 1354.0, + 421.0, + 1354.0, + 458.0, + 1177.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 453.0, + 626.0, + 453.0, + 626.0, + 465.0, + 610.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 449.0, + 798.0, + 449.0, + 798.0, + 469.0, + 771.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 451.0, + 874.0, + 451.0, + 874.0, + 467.0, + 857.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 453.0, + 958.0, + 453.0, + 958.0, + 467.0, + 940.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 453.0, + 979.0, + 453.0, + 979.0, + 467.0, + 960.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 453.0, + 1144.0, + 453.0, + 1144.0, + 467.0, + 1130.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 444.0, + 1366.0, + 444.0, + 1366.0, + 477.0, + 1168.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 465.0, + 1370.0, + 465.0, + 1370.0, + 493.0, + 1162.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 491.0, + 587.0, + 491.0, + 587.0, + 525.0, + 511.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 497.0, + 652.0, + 497.0, + 652.0, + 521.0, + 597.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 495.0, + 842.0, + 495.0, + 842.0, + 523.0, + 683.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 497.0, + 908.0, + 497.0, + 908.0, + 521.0, + 846.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 497.0, + 985.0, + 497.0, + 985.0, + 521.0, + 933.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 495.0, + 1080.0, + 495.0, + 1080.0, + 523.0, + 1002.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 497.0, + 1160.0, + 497.0, + 1160.0, + 521.0, + 1095.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 514.0, + 569.0, + 514.0, + 569.0, + 539.0, + 526.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 514.0, + 647.0, + 514.0, + 647.0, + 539.0, + 603.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 508.0, + 916.0, + 508.0, + 916.0, + 543.0, + 838.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 514.0, + 990.0, + 514.0, + 990.0, + 539.0, + 928.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 510.0, + 1182.0, + 510.0, + 1182.0, + 543.0, + 1072.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 565.0, + 569.0, + 565.0, + 569.0, + 579.0, + 555.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 565.0, + 711.0, + 565.0, + 711.0, + 577.0, + 693.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 565.0, + 732.0, + 565.0, + 732.0, + 579.0, + 718.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 565.0, + 789.0, + 565.0, + 789.0, + 579.0, + 775.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 567.0, + 816.0, + 567.0, + 816.0, + 579.0, + 796.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 562.0, + 876.0, + 562.0, + 876.0, + 586.0, + 853.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 565.0, + 896.0, + 565.0, + 896.0, + 579.0, + 881.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 563.0, + 983.0, + 563.0, + 983.0, + 584.0, + 937.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 563.0, + 1043.0, + 563.0, + 1043.0, + 584.0, + 1018.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 567.0, + 1063.0, + 567.0, + 1063.0, + 583.0, + 1045.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 621.0, + 881.0, + 621.0, + 881.0, + 634.0, + 862.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 637.0, + 791.0, + 637.0, + 791.0, + 653.0, + 775.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 639.0, + 956.0, + 639.0, + 956.0, + 651.0, + 938.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 639.0, + 1038.0, + 639.0, + 1038.0, + 651.0, + 1024.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 675.0, + 591.0, + 675.0, + 591.0, + 725.0, + 503.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 671.0, + 907.0, + 671.0, + 907.0, + 727.0, + 595.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 673.0, + 1173.0, + 673.0, + 1173.0, + 725.0, + 917.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 678.0, + 1235.0, + 678.0, + 1235.0, + 726.0, + 1179.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 742.0, + 491.0, + 742.0, + 491.0, + 811.0, + 340.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 748.0, + 1365.0, + 748.0, + 1365.0, + 781.0, + 1171.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 767.0, + 1325.0, + 767.0, + 1325.0, + 797.0, + 1158.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 767.0, + 1371.0, + 767.0, + 1371.0, + 797.0, + 1364.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 807.0, + 1354.0, + 807.0, + 1354.0, + 844.0, + 1179.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 839.0, + 874.0, + 839.0, + 874.0, + 851.0, + 857.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 839.0, + 897.0, + 839.0, + 897.0, + 851.0, + 881.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 832.0, + 1365.0, + 832.0, + 1365.0, + 860.0, + 1169.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 848.0, + 1372.0, + 848.0, + 1372.0, + 881.0, + 1160.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 883.0, + 588.0, + 883.0, + 588.0, + 911.0, + 516.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 886.0, + 659.0, + 886.0, + 659.0, + 911.0, + 603.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 880.0, + 844.0, + 880.0, + 844.0, + 915.0, + 682.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 883.0, + 912.0, + 883.0, + 912.0, + 907.0, + 848.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 885.0, + 983.0, + 885.0, + 983.0, + 907.0, + 931.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 881.0, + 1073.0, + 881.0, + 1073.0, + 909.0, + 999.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 881.0, + 1162.0, + 881.0, + 1162.0, + 909.0, + 1088.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 902.0, + 574.0, + 902.0, + 574.0, + 928.0, + 531.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 902.0, + 651.0, + 902.0, + 651.0, + 928.0, + 608.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 899.0, + 917.0, + 899.0, + 917.0, + 927.0, + 842.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 900.0, + 988.0, + 900.0, + 988.0, + 925.0, + 926.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 899.0, + 1178.0, + 899.0, + 1178.0, + 927.0, + 1070.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 955.0, + 649.0, + 955.0, + 649.0, + 969.0, + 633.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 953.0, + 798.0, + 953.0, + 798.0, + 974.0, + 771.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 946.0, + 906.0, + 946.0, + 906.0, + 992.0, + 846.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 953.0, + 983.0, + 953.0, + 983.0, + 972.0, + 937.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 958.0, + 1056.0, + 958.0, + 1056.0, + 967.0, + 1047.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 1029.0, + 791.0, + 1029.0, + 791.0, + 1044.0, + 773.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 1029.0, + 896.0, + 1029.0, + 896.0, + 1041.0, + 881.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 1068.0, + 573.0, + 1068.0, + 573.0, + 1103.0, + 514.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1072.0, + 668.0, + 1072.0, + 668.0, + 1102.0, + 587.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 1072.0, + 752.0, + 1072.0, + 752.0, + 1100.0, + 675.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 1071.0, + 835.0, + 1071.0, + 835.0, + 1099.0, + 761.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 1068.0, + 914.0, + 1068.0, + 914.0, + 1103.0, + 847.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1072.0, + 1001.0, + 1072.0, + 1001.0, + 1100.0, + 919.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 1074.0, + 1066.0, + 1074.0, + 1066.0, + 1099.0, + 1016.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1072.0, + 1231.0, + 1072.0, + 1231.0, + 1097.0, + 1182.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1090.0, + 586.0, + 1090.0, + 586.0, + 1118.0, + 497.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 1093.0, + 658.0, + 1093.0, + 658.0, + 1118.0, + 592.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 1090.0, + 839.0, + 1090.0, + 839.0, + 1118.0, + 755.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1088.0, + 1166.0, + 1088.0, + 1166.0, + 1122.0, + 919.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1090.0, + 1230.0, + 1090.0, + 1230.0, + 1116.0, + 1185.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 1127.0, + 491.0, + 1127.0, + 491.0, + 1195.0, + 345.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 1151.0, + 816.0, + 1151.0, + 816.0, + 1169.0, + 775.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1136.0, + 1368.0, + 1136.0, + 1368.0, + 1185.0, + 1157.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 1199.0, + 1349.0, + 1199.0, + 1349.0, + 1227.0, + 1176.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 1218.0, + 1361.0, + 1218.0, + 1361.0, + 1246.0, + 1164.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1236.0, + 1368.0, + 1236.0, + 1368.0, + 1269.0, + 1157.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 1263.0, + 582.0, + 1263.0, + 582.0, + 1298.0, + 506.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1267.0, + 652.0, + 1267.0, + 652.0, + 1297.0, + 590.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 1269.0, + 835.0, + 1269.0, + 835.0, + 1294.0, + 679.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 1263.0, + 905.0, + 1263.0, + 905.0, + 1298.0, + 839.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1269.0, + 981.0, + 1269.0, + 981.0, + 1294.0, + 929.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 1267.0, + 1075.0, + 1267.0, + 1075.0, + 1295.0, + 1001.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1269.0, + 1160.0, + 1269.0, + 1160.0, + 1295.0, + 1091.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1287.0, + 565.0, + 1287.0, + 565.0, + 1311.0, + 523.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 1285.0, + 643.0, + 1285.0, + 643.0, + 1311.0, + 599.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1287.0, + 910.0, + 1287.0, + 910.0, + 1315.0, + 837.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 1287.0, + 986.0, + 1287.0, + 986.0, + 1311.0, + 926.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1283.0, + 1180.0, + 1283.0, + 1180.0, + 1316.0, + 1068.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 1452.0, + 576.0, + 1452.0, + 576.0, + 1481.0, + 514.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 1447.0, + 751.0, + 1447.0, + 751.0, + 1482.0, + 593.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1452.0, + 830.0, + 1452.0, + 830.0, + 1480.0, + 757.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1452.0, + 999.0, + 1452.0, + 999.0, + 1480.0, + 915.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 1452.0, + 1231.0, + 1452.0, + 1231.0, + 1476.0, + 1180.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 1471.0, + 587.0, + 1471.0, + 587.0, + 1499.0, + 501.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1469.0, + 665.0, + 1469.0, + 665.0, + 1499.0, + 597.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 1471.0, + 835.0, + 1471.0, + 835.0, + 1499.0, + 750.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1469.0, + 1086.0, + 1469.0, + 1086.0, + 1497.0, + 919.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 1467.0, + 1162.0, + 1467.0, + 1162.0, + 1495.0, + 1088.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 1466.0, + 1230.0, + 1466.0, + 1230.0, + 1495.0, + 1180.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1498.0, + 506.0, + 1498.0, + 506.0, + 1600.0, + 324.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1539.0, + 725.0, + 1539.0, + 725.0, + 1553.0, + 711.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 1536.0, + 896.0, + 1536.0, + 896.0, + 1555.0, + 873.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 1539.0, + 974.0, + 1539.0, + 974.0, + 1555.0, + 958.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1520.0, + 1368.0, + 1520.0, + 1368.0, + 1569.0, + 1157.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 1578.0, + 1351.0, + 1578.0, + 1351.0, + 1615.0, + 1173.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 1602.0, + 1363.0, + 1602.0, + 1363.0, + 1631.0, + 1164.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1622.0, + 1366.0, + 1622.0, + 1366.0, + 1650.0, + 1157.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 1655.0, + 587.0, + 1655.0, + 587.0, + 1683.0, + 514.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1655.0, + 659.0, + 1655.0, + 659.0, + 1685.0, + 597.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 1655.0, + 908.0, + 1655.0, + 908.0, + 1683.0, + 679.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1655.0, + 979.0, + 1655.0, + 979.0, + 1680.0, + 928.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 1652.0, + 1070.0, + 1652.0, + 1070.0, + 1681.0, + 995.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1653.0, + 1153.0, + 1653.0, + 1153.0, + 1678.0, + 1091.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1674.0, + 571.0, + 1674.0, + 571.0, + 1701.0, + 528.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 1674.0, + 649.0, + 1674.0, + 649.0, + 1701.0, + 604.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 1671.0, + 915.0, + 1671.0, + 915.0, + 1699.0, + 839.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1673.0, + 985.0, + 1673.0, + 985.0, + 1697.0, + 922.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1671.0, + 1175.0, + 1671.0, + 1175.0, + 1699.0, + 1068.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1729.0, + 810.0, + 1729.0, + 810.0, + 1741.0, + 789.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1799.0, + 810.0, + 1799.0, + 810.0, + 1813.0, + 789.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 1839.0, + 575.0, + 1839.0, + 575.0, + 1873.0, + 506.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 1839.0, + 747.0, + 1839.0, + 747.0, + 1873.0, + 671.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 1843.0, + 832.0, + 1843.0, + 832.0, + 1871.0, + 755.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 1839.0, + 911.0, + 1839.0, + 911.0, + 1873.0, + 840.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1839.0, + 999.0, + 1839.0, + 999.0, + 1873.0, + 915.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 1839.0, + 1170.0, + 1839.0, + 1170.0, + 1873.0, + 1083.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1845.0, + 1228.0, + 1845.0, + 1228.0, + 1869.0, + 1178.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 1862.0, + 581.0, + 1862.0, + 581.0, + 1890.0, + 496.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1864.0, + 656.0, + 1864.0, + 656.0, + 1889.0, + 590.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 1861.0, + 837.0, + 1861.0, + 837.0, + 1888.0, + 748.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1862.0, + 1086.0, + 1862.0, + 1086.0, + 1890.0, + 917.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 1862.0, + 1160.0, + 1862.0, + 1160.0, + 1890.0, + 1088.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1860.0, + 1226.0, + 1860.0, + 1226.0, + 1887.0, + 1182.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 676.5, + 1065.0, + 676.5, + 1065.0, + 706.5, + 1012.0, + 706.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 948.0, + 808.0, + 948.0, + 808.0, + 986.0, + 795.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 1069.5, + 1169.0, + 1069.5, + 1169.0, + 1099.5, + 1086.0, + 1099.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1101.0, + 1339.0, + 1132.0, + 1339.0, + 1132.0, + 1357.0, + 1101.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.75, + 1450.5, + 898.75, + 1450.5, + 898.75, + 1480.5, + 837.75, + 1480.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1450.5, + 1060.0, + 1450.5, + 1060.0, + 1476.5, + 1012.0, + 1476.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.75, + 1447.5, + 1167.75, + 1447.5, + 1167.75, + 1477.0, + 1083.75, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.75, + 1842.0, + 664.75, + 1842.0, + 664.75, + 1871.5, + 582.75, + 1871.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1843.0, + 1064.0, + 1843.0, + 1064.0, + 1871.0, + 1011.0, + 1871.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 19, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 331, + 329, + 1367, + 329, + 1367, + 1902, + 331, + 1902 + ], + "score": 0.967 + }, + { + "category_id": 4, + "poly": [ + 376, + 1903, + 1319, + 1903, + 1319, + 1937, + 376, + 1937 + ], + "score": 0.912 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 105, + 297, + 105 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 863, + 2088, + 863, + 2113, + 834, + 2113 + ], + "score": 0.833 + }, + { + "category_id": 13, + "poly": [ + 1130, + 1901, + 1305, + 1901, + 1305, + 1934, + 1130, + 1934 + ], + "score": 0.53, + "latex": "\\mathrm { ( G - F P R = 1 0 ^ { - 3 } }" + }, + { + "category_id": 13, + "poly": [ + 1322, + 1545, + 1359, + 1545, + 1359, + 1563, + 1322, + 1563 + ], + "score": 0.32, + "latex": "\\scriptstyle ( \\mathbf { a } = \\mathbf { 0 } )" + }, + { + "category_id": 13, + "poly": [ + 1326, + 774, + 1363, + 774, + 1363, + 792, + 1326, + 792 + ], + "score": 0.26, + "latex": "\\scriptstyle ( \\mathbf { a } = \\mathbf { 0 } )" + }, + { + "category_id": 13, + "poly": [ + 1326, + 389, + 1363, + 389, + 1363, + 407, + 1326, + 407 + ], + "score": 0.26, + "latex": "\\scriptstyle ( \\mathbf { a } = 0 )" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 366.0, + 1362.0, + 366.0, + 1362.0, + 394.0, + 1172.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 384.0, + 1325.0, + 384.0, + 1325.0, + 411.0, + 1161.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 384.0, + 1371.0, + 384.0, + 1371.0, + 411.0, + 1364.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 427.0, + 1353.0, + 427.0, + 1353.0, + 455.0, + 1181.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 446.0, + 1366.0, + 446.0, + 1366.0, + 474.0, + 1170.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 465.0, + 1371.0, + 465.0, + 1371.0, + 493.0, + 1163.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 497.0, + 585.0, + 497.0, + 585.0, + 519.0, + 514.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 497.0, + 909.0, + 497.0, + 909.0, + 523.0, + 678.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 497.0, + 1080.0, + 497.0, + 1080.0, + 519.0, + 1005.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 497.0, + 1161.0, + 497.0, + 1161.0, + 519.0, + 1092.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 518.0, + 569.0, + 518.0, + 569.0, + 537.0, + 528.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 514.0, + 647.0, + 514.0, + 647.0, + 539.0, + 604.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 508.0, + 917.0, + 508.0, + 917.0, + 543.0, + 838.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 514.0, + 989.0, + 514.0, + 989.0, + 539.0, + 927.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 516.0, + 1179.0, + 516.0, + 1179.0, + 539.0, + 1074.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 682.0, + 576.0, + 682.0, + 576.0, + 706.0, + 519.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 680.0, + 675.0, + 680.0, + 675.0, + 708.0, + 597.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 678.0, + 833.0, + 678.0, + 833.0, + 706.0, + 758.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 680.0, + 1234.0, + 680.0, + 1234.0, + 702.0, + 1186.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 699.0, + 590.0, + 699.0, + 590.0, + 727.0, + 505.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 700.0, + 666.0, + 700.0, + 666.0, + 724.0, + 600.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 699.0, + 838.0, + 699.0, + 838.0, + 727.0, + 753.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 696.0, + 1165.0, + 696.0, + 1165.0, + 723.0, + 918.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 693.0, + 1232.0, + 693.0, + 1232.0, + 724.0, + 1189.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 750.0, + 1363.0, + 750.0, + 1363.0, + 779.0, + 1168.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 769.0, + 1325.0, + 769.0, + 1325.0, + 797.0, + 1161.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 769.0, + 1371.0, + 769.0, + 1371.0, + 797.0, + 1364.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 812.0, + 1353.0, + 812.0, + 1353.0, + 840.0, + 1181.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 840.0, + 895.0, + 840.0, + 895.0, + 854.0, + 881.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 831.0, + 1366.0, + 831.0, + 1366.0, + 859.0, + 1169.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 847.0, + 1371.0, + 847.0, + 1371.0, + 880.0, + 1160.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 885.0, + 588.0, + 885.0, + 588.0, + 927.0, + 522.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 887.0, + 659.0, + 887.0, + 659.0, + 927.0, + 604.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 880.0, + 918.0, + 880.0, + 918.0, + 927.0, + 746.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 884.0, + 986.0, + 884.0, + 986.0, + 924.0, + 927.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 880.0, + 1179.0, + 880.0, + 1179.0, + 929.0, + 1000.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 959.0, + 647.0, + 959.0, + 647.0, + 971.0, + 632.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 955.0, + 813.0, + 955.0, + 813.0, + 971.0, + 796.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 957.0, + 979.0, + 957.0, + 979.0, + 971.0, + 959.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1028.0, + 813.0, + 1028.0, + 813.0, + 1042.0, + 796.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 1074.0, + 570.0, + 1074.0, + 570.0, + 1098.0, + 515.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 1074.0, + 670.0, + 1074.0, + 670.0, + 1102.0, + 588.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 1070.0, + 835.0, + 1070.0, + 835.0, + 1098.0, + 762.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 1070.0, + 914.0, + 1070.0, + 914.0, + 1102.0, + 847.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 1072.0, + 1000.0, + 1072.0, + 1000.0, + 1100.0, + 918.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 1074.0, + 1066.0, + 1074.0, + 1066.0, + 1098.0, + 1019.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1074.0, + 1231.0, + 1074.0, + 1231.0, + 1096.0, + 1183.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1089.0, + 586.0, + 1089.0, + 586.0, + 1117.0, + 497.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1089.0, + 659.0, + 1089.0, + 659.0, + 1119.0, + 590.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 1089.0, + 840.0, + 1089.0, + 840.0, + 1117.0, + 755.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1089.0, + 1165.0, + 1089.0, + 1165.0, + 1119.0, + 920.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1089.0, + 1231.0, + 1089.0, + 1231.0, + 1115.0, + 1186.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1128.0, + 494.0, + 1128.0, + 494.0, + 1199.0, + 341.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 1135.0, + 1360.0, + 1135.0, + 1360.0, + 1168.0, + 1167.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1154.0, + 1367.0, + 1154.0, + 1367.0, + 1183.0, + 1156.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1199.0, + 1350.0, + 1199.0, + 1350.0, + 1227.0, + 1177.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 1213.0, + 1362.0, + 1213.0, + 1362.0, + 1250.0, + 1164.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1236.0, + 1367.0, + 1236.0, + 1367.0, + 1265.0, + 1156.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 1267.0, + 581.0, + 1267.0, + 581.0, + 1293.0, + 512.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 1265.0, + 838.0, + 1265.0, + 838.0, + 1295.0, + 679.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1267.0, + 904.0, + 1267.0, + 904.0, + 1292.0, + 842.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 1265.0, + 984.0, + 1265.0, + 984.0, + 1295.0, + 927.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1265.0, + 1076.0, + 1265.0, + 1076.0, + 1293.0, + 1000.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 1266.0, + 1156.0, + 1266.0, + 1156.0, + 1291.0, + 1088.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1286.0, + 565.0, + 1286.0, + 565.0, + 1311.0, + 524.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1285.0, + 643.0, + 1285.0, + 643.0, + 1311.0, + 601.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1285.0, + 911.0, + 1285.0, + 911.0, + 1312.0, + 837.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 1283.0, + 987.0, + 1283.0, + 987.0, + 1312.0, + 924.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 1285.0, + 1177.0, + 1285.0, + 1177.0, + 1312.0, + 1069.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 1448.0, + 575.0, + 1448.0, + 575.0, + 1482.0, + 515.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 1446.0, + 750.0, + 1446.0, + 750.0, + 1480.0, + 592.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 1450.0, + 829.0, + 1450.0, + 829.0, + 1478.0, + 755.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 1448.0, + 903.0, + 1448.0, + 903.0, + 1482.0, + 836.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1446.0, + 999.0, + 1446.0, + 999.0, + 1479.0, + 916.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1452.0, + 1064.0, + 1452.0, + 1064.0, + 1476.0, + 1012.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 1450.0, + 1169.0, + 1450.0, + 1169.0, + 1478.0, + 1085.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1450.0, + 1231.0, + 1450.0, + 1231.0, + 1474.0, + 1183.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 1469.0, + 588.0, + 1469.0, + 588.0, + 1497.0, + 503.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1468.0, + 666.0, + 1468.0, + 666.0, + 1497.0, + 597.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 1469.0, + 835.0, + 1469.0, + 835.0, + 1497.0, + 750.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 1466.0, + 1161.0, + 1466.0, + 1161.0, + 1495.0, + 913.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1468.0, + 1225.0, + 1468.0, + 1225.0, + 1494.0, + 1183.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 1518.0, + 483.0, + 1518.0, + 483.0, + 1586.0, + 334.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1522.0, + 1358.0, + 1522.0, + 1358.0, + 1549.0, + 1169.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1539.0, + 1321.0, + 1539.0, + 1321.0, + 1568.0, + 1156.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1360.0, + 1539.0, + 1367.0, + 1539.0, + 1367.0, + 1568.0, + 1360.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1583.0, + 1348.0, + 1583.0, + 1348.0, + 1610.0, + 1177.0, + 1610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 1600.0, + 1362.0, + 1600.0, + 1362.0, + 1628.0, + 1165.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 1621.0, + 1366.0, + 1621.0, + 1366.0, + 1649.0, + 1160.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 1656.0, + 583.0, + 1656.0, + 583.0, + 1699.0, + 519.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 1657.0, + 656.0, + 1657.0, + 656.0, + 1699.0, + 599.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 1654.0, + 915.0, + 1654.0, + 915.0, + 1700.0, + 679.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1654.0, + 984.0, + 1654.0, + 984.0, + 1694.0, + 922.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 1651.0, + 1176.0, + 1651.0, + 1176.0, + 1701.0, + 996.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 1844.0, + 572.0, + 1844.0, + 572.0, + 1868.0, + 508.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 1844.0, + 664.0, + 1844.0, + 664.0, + 1874.0, + 585.0, + 1874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 1838.0, + 747.0, + 1838.0, + 747.0, + 1872.0, + 671.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 1844.0, + 828.0, + 1844.0, + 828.0, + 1867.0, + 758.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1838.0, + 997.0, + 1838.0, + 997.0, + 1872.0, + 916.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 1842.0, + 1065.0, + 1842.0, + 1065.0, + 1871.0, + 1010.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 1861.0, + 583.0, + 1861.0, + 583.0, + 1889.0, + 498.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 1860.0, + 657.0, + 1860.0, + 657.0, + 1889.0, + 588.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 1860.0, + 837.0, + 1860.0, + 837.0, + 1888.0, + 751.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 1861.0, + 1085.0, + 1861.0, + 1085.0, + 1889.0, + 918.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 1857.0, + 1160.0, + 1857.0, + 1160.0, + 1890.0, + 1086.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.75, + 496.5, + 652.75, + 496.5, + 652.75, + 519.5, + 595.75, + 519.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 495.5, + 982.0, + 495.5, + 982.0, + 519.5, + 936.0, + 519.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 675.5, + 751.0, + 675.5, + 751.0, + 706.5, + 675.0, + 706.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.75, + 678.5, + 901.75, + 678.5, + 901.75, + 708.5, + 840.75, + 708.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.25, + 675.5, + 998.25, + 675.5, + 998.25, + 705.0, + 918.25, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 679.5, + 1066.0, + 679.5, + 1066.0, + 702.5, + 1012.0, + 702.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 675.5, + 1170.0, + 675.5, + 1170.0, + 705.0, + 1087.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 882.5, + 749.0, + 882.5, + 749.0, + 912.5, + 681.0, + 912.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1069.5, + 748.0, + 1069.5, + 748.0, + 1098.5, + 677.0, + 1098.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 1069.5, + 1170.0, + 1069.5, + 1170.0, + 1099.0, + 1087.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.25, + 1264.5, + 649.25, + 1264.5, + 649.25, + 1294.0, + 592.25, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1841.5, + 907.0, + 1841.5, + 907.0, + 1870.5, + 842.0, + 1870.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 1840.5, + 1165.0, + 1840.5, + 1165.0, + 1870.5, + 1082.0, + 1870.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 1899.0, + 1129.0, + 1899.0, + 1129.0, + 1940.0, + 375.0, + 1940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 1899.0, + 1322.0, + 1899.0, + 1322.0, + 1940.0, + 1306.0, + 1940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 830.0, + 2124.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 20, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 331, + 738, + 1367, + 738, + 1367, + 1499, + 331, + 1499 + ], + "score": 0.974 + }, + { + "category_id": 4, + "poly": [ + 372, + 1517, + 1322, + 1517, + 1322, + 1550, + 372, + 1550 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 859, + 75, + 859, + 106, + 297, + 106 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2113, + 834, + 2113 + ], + "score": 0.831 + }, + { + "category_id": 3, + "poly": [ + 331, + 732, + 1367, + 732, + 1367, + 1499, + 331, + 1499 + ], + "score": 0.11 + }, + { + "category_id": 13, + "poly": [ + 1137, + 1516, + 1310, + 1516, + 1310, + 1547, + 1137, + 1547 + ], + "score": 0.75, + "latex": "\\mathrm { \\bf G } \\mathrm { - } \\mathrm { \\bf F P R } = 1 0 ^ { - 3 } ," + }, + { + "category_id": 13, + "poly": [ + 1323, + 774, + 1358, + 774, + 1358, + 792, + 1323, + 792 + ], + "score": 0.32, + "latex": "\\scriptstyle ( { \\mathrm { a } } = 0 )" + }, + { + "category_id": 13, + "poly": [ + 1322, + 1159, + 1358, + 1159, + 1358, + 1177, + 1322, + 1177 + ], + "score": 0.28, + "latex": "\\scriptstyle ( { \\mathrm { a } } = 0 )" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 745.0, + 409.0, + 745.0, + 409.0, + 807.0, + 339.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 740.0, + 484.0, + 740.0, + 484.0, + 811.0, + 414.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 752.0, + 1356.0, + 752.0, + 1356.0, + 777.0, + 1171.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 771.0, + 1322.0, + 771.0, + 1322.0, + 796.0, + 1161.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1359.0, + 771.0, + 1365.0, + 771.0, + 1365.0, + 796.0, + 1359.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 809.0, + 1347.0, + 809.0, + 1347.0, + 843.0, + 1178.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 832.0, + 1360.0, + 832.0, + 1360.0, + 859.0, + 1167.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 853.0, + 1364.0, + 853.0, + 1364.0, + 876.0, + 1159.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 883.0, + 578.0, + 883.0, + 578.0, + 903.0, + 512.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 879.0, + 650.0, + 879.0, + 650.0, + 909.0, + 595.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 882.0, + 837.0, + 882.0, + 837.0, + 908.0, + 680.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 883.0, + 980.0, + 883.0, + 980.0, + 907.0, + 931.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 882.0, + 1074.0, + 882.0, + 1074.0, + 906.0, + 1002.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 885.0, + 1155.0, + 885.0, + 1155.0, + 904.0, + 1094.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 902.0, + 564.0, + 902.0, + 564.0, + 923.0, + 526.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 900.0, + 643.0, + 900.0, + 643.0, + 924.0, + 601.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 899.0, + 909.0, + 899.0, + 909.0, + 926.0, + 838.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 901.0, + 986.0, + 901.0, + 986.0, + 924.0, + 925.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 897.0, + 1176.0, + 897.0, + 1176.0, + 928.0, + 1070.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 935.0, + 1235.0, + 935.0, + 1235.0, + 1005.0, + 1176.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 989.0, + 1215.0, + 989.0, + 1215.0, + 1000.0, + 1191.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 1065.0, + 574.0, + 1065.0, + 574.0, + 1092.0, + 515.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 1066.0, + 672.0, + 1066.0, + 672.0, + 1092.0, + 592.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 1064.0, + 748.0, + 1064.0, + 748.0, + 1092.0, + 674.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 1067.0, + 828.0, + 1067.0, + 828.0, + 1090.0, + 759.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1063.0, + 995.0, + 1063.0, + 995.0, + 1091.0, + 915.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1064.0, + 1060.0, + 1064.0, + 1060.0, + 1091.0, + 1012.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 1061.0, + 1168.0, + 1061.0, + 1168.0, + 1092.0, + 1082.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1067.0, + 1229.0, + 1067.0, + 1229.0, + 1087.0, + 1183.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 1087.0, + 583.0, + 1087.0, + 583.0, + 1108.0, + 502.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1086.0, + 663.0, + 1086.0, + 663.0, + 1109.0, + 601.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 1087.0, + 834.0, + 1087.0, + 834.0, + 1109.0, + 753.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 1082.0, + 1082.0, + 1082.0, + 1082.0, + 1110.0, + 918.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1085.0, + 1158.0, + 1085.0, + 1158.0, + 1104.0, + 1090.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1085.0, + 1224.0, + 1085.0, + 1224.0, + 1106.0, + 1185.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1154.0, + 366.0, + 1154.0, + 366.0, + 1163.0, + 353.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1142.0, + 477.0, + 1142.0, + 477.0, + 1179.0, + 423.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 1137.0, + 1321.0, + 1137.0, + 1321.0, + 1180.0, + 1161.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1359.0, + 1137.0, + 1364.0, + 1137.0, + 1364.0, + 1180.0, + 1359.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1196.0, + 1346.0, + 1196.0, + 1346.0, + 1225.0, + 1178.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 1218.0, + 1358.0, + 1218.0, + 1358.0, + 1244.0, + 1166.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 1237.0, + 1363.0, + 1237.0, + 1363.0, + 1262.0, + 1160.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 1272.0, + 583.0, + 1272.0, + 583.0, + 1292.0, + 519.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 1269.0, + 655.0, + 1269.0, + 655.0, + 1296.0, + 599.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1267.0, + 908.0, + 1267.0, + 908.0, + 1293.0, + 677.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1271.0, + 976.0, + 1271.0, + 976.0, + 1290.0, + 929.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 1266.0, + 1069.0, + 1266.0, + 1069.0, + 1292.0, + 995.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1271.0, + 1153.0, + 1271.0, + 1153.0, + 1289.0, + 1090.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 1290.0, + 569.0, + 1290.0, + 569.0, + 1310.0, + 531.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1290.0, + 647.0, + 1290.0, + 647.0, + 1310.0, + 606.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1287.0, + 913.0, + 1287.0, + 913.0, + 1310.0, + 842.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1286.0, + 983.0, + 1286.0, + 983.0, + 1309.0, + 921.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1283.0, + 1173.0, + 1283.0, + 1173.0, + 1314.0, + 1068.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 1456.0, + 570.0, + 1456.0, + 570.0, + 1484.0, + 510.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 1460.0, + 663.0, + 1460.0, + 663.0, + 1483.0, + 584.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 1454.0, + 747.0, + 1454.0, + 747.0, + 1485.0, + 670.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 1458.0, + 829.0, + 1458.0, + 829.0, + 1481.0, + 760.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1456.0, + 996.0, + 1456.0, + 996.0, + 1484.0, + 916.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 1457.0, + 1065.0, + 1457.0, + 1065.0, + 1485.0, + 1013.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 1457.0, + 1166.0, + 1457.0, + 1166.0, + 1484.0, + 1081.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1461.0, + 1224.0, + 1461.0, + 1224.0, + 1479.0, + 1178.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 1479.0, + 580.0, + 1479.0, + 580.0, + 1501.0, + 500.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1478.0, + 655.0, + 1478.0, + 655.0, + 1501.0, + 591.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 1476.0, + 834.0, + 1476.0, + 834.0, + 1500.0, + 750.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1476.0, + 1158.0, + 1476.0, + 1158.0, + 1502.0, + 919.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1475.0, + 1224.0, + 1475.0, + 1224.0, + 1500.0, + 1183.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 882.0, + 903.0, + 882.0, + 903.0, + 905.5, + 828.0, + 905.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 1067.0, + 898.0, + 1067.0, + 898.0, + 1090.0, + 839.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.25, + 1458.5, + 906.25, + 1458.5, + 906.25, + 1481.5, + 845.25, + 1481.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 1515.0, + 1136.0, + 1515.0, + 1136.0, + 1552.0, + 371.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 1515.0, + 1325.0, + 1515.0, + 1325.0, + 1552.0, + 1311.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 829.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 744.0, + 486.0, + 744.0, + 486.0, + 808.0, + 340.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 753.0, + 1356.0, + 753.0, + 1356.0, + 777.0, + 1170.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 772.0, + 1322.0, + 772.0, + 1322.0, + 795.0, + 1161.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1359.0, + 772.0, + 1365.0, + 772.0, + 1365.0, + 795.0, + 1359.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 813.0, + 1348.0, + 813.0, + 1348.0, + 842.0, + 1178.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 832.0, + 1359.0, + 832.0, + 1359.0, + 859.0, + 1165.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 852.0, + 1364.0, + 852.0, + 1364.0, + 876.0, + 1160.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 880.0, + 580.0, + 880.0, + 580.0, + 906.0, + 510.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 881.0, + 651.0, + 881.0, + 651.0, + 907.0, + 594.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 882.0, + 837.0, + 882.0, + 837.0, + 906.0, + 677.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 882.0, + 980.0, + 882.0, + 980.0, + 906.0, + 931.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 880.0, + 1074.0, + 880.0, + 1074.0, + 906.0, + 1002.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 883.0, + 1155.0, + 883.0, + 1155.0, + 903.0, + 1094.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 900.0, + 565.0, + 900.0, + 565.0, + 925.0, + 524.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 900.0, + 642.0, + 900.0, + 642.0, + 925.0, + 601.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 901.0, + 908.0, + 901.0, + 908.0, + 925.0, + 838.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 900.0, + 986.0, + 900.0, + 986.0, + 923.0, + 925.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 900.0, + 1176.0, + 900.0, + 1176.0, + 926.0, + 1070.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 1064.0, + 575.0, + 1064.0, + 575.0, + 1092.0, + 514.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1062.0, + 749.0, + 1062.0, + 749.0, + 1094.0, + 591.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 1068.0, + 827.0, + 1068.0, + 827.0, + 1088.0, + 758.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 1064.0, + 900.0, + 1064.0, + 900.0, + 1092.0, + 836.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1067.0, + 996.0, + 1067.0, + 996.0, + 1090.0, + 916.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1064.0, + 1061.0, + 1064.0, + 1061.0, + 1091.0, + 1012.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 1063.0, + 1166.0, + 1063.0, + 1166.0, + 1090.0, + 1084.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 1067.0, + 1229.0, + 1067.0, + 1229.0, + 1087.0, + 1180.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 1086.0, + 585.0, + 1086.0, + 585.0, + 1110.0, + 503.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 1085.0, + 662.0, + 1085.0, + 662.0, + 1107.0, + 599.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 1086.0, + 833.0, + 1086.0, + 833.0, + 1107.0, + 751.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1087.0, + 1081.0, + 1087.0, + 1081.0, + 1106.0, + 921.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1085.0, + 1158.0, + 1085.0, + 1158.0, + 1105.0, + 1090.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1083.0, + 1225.0, + 1083.0, + 1225.0, + 1107.0, + 1183.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 1144.0, + 474.0, + 1144.0, + 474.0, + 1185.0, + 424.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 1137.0, + 1321.0, + 1137.0, + 1321.0, + 1180.0, + 1161.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1359.0, + 1137.0, + 1364.0, + 1137.0, + 1364.0, + 1180.0, + 1359.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 1200.0, + 1345.0, + 1200.0, + 1345.0, + 1223.0, + 1180.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 1218.0, + 1359.0, + 1218.0, + 1359.0, + 1243.0, + 1166.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 1234.0, + 1365.0, + 1234.0, + 1365.0, + 1264.0, + 1160.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 1268.0, + 585.0, + 1268.0, + 585.0, + 1295.0, + 516.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1268.0, + 655.0, + 1268.0, + 655.0, + 1295.0, + 598.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 1270.0, + 910.0, + 1270.0, + 910.0, + 1293.0, + 676.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1268.0, + 977.0, + 1268.0, + 977.0, + 1291.0, + 929.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 1266.0, + 1069.0, + 1266.0, + 1069.0, + 1291.0, + 996.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 1270.0, + 1153.0, + 1270.0, + 1153.0, + 1289.0, + 1089.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 1290.0, + 569.0, + 1290.0, + 569.0, + 1311.0, + 531.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1290.0, + 645.0, + 1290.0, + 645.0, + 1311.0, + 606.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 1284.0, + 912.0, + 1284.0, + 912.0, + 1310.0, + 841.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 1287.0, + 982.0, + 1287.0, + 982.0, + 1308.0, + 924.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1285.0, + 1173.0, + 1285.0, + 1173.0, + 1311.0, + 1068.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 1356.0, + 1043.0, + 1356.0, + 1043.0, + 1375.0, + 1024.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 1456.0, + 570.0, + 1456.0, + 570.0, + 1483.0, + 510.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 1457.0, + 664.0, + 1457.0, + 664.0, + 1485.0, + 585.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 1457.0, + 746.0, + 1457.0, + 746.0, + 1483.0, + 670.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 1457.0, + 829.0, + 1457.0, + 829.0, + 1481.0, + 759.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1457.0, + 908.0, + 1457.0, + 908.0, + 1485.0, + 845.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1456.0, + 996.0, + 1456.0, + 996.0, + 1483.0, + 916.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 1456.0, + 1065.0, + 1456.0, + 1065.0, + 1486.0, + 1013.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 1454.0, + 1168.0, + 1454.0, + 1168.0, + 1486.0, + 1082.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 1460.0, + 1224.0, + 1460.0, + 1224.0, + 1479.0, + 1176.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 1478.0, + 580.0, + 1478.0, + 580.0, + 1502.0, + 499.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1476.0, + 656.0, + 1476.0, + 656.0, + 1503.0, + 590.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 1477.0, + 833.0, + 1477.0, + 833.0, + 1498.0, + 752.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1477.0, + 1160.0, + 1477.0, + 1160.0, + 1503.0, + 919.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1475.0, + 1224.0, + 1475.0, + 1224.0, + 1500.0, + 1183.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.25, + 763.5, + 541.25, + 763.5, + 541.25, + 776.5, + 535.25, + 776.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 881.5, + 903.0, + 881.5, + 903.0, + 905.5, + 842.0, + 905.5 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 21, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 318, + 231, + 1370, + 231, + 1370, + 1921, + 318, + 1921 + ], + "score": 0.962 + }, + { + "category_id": 4, + "poly": [ + 298, + 1948, + 1404, + 1948, + 1404, + 2043, + 298, + 2043 + ], + "score": 0.943 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 858, + 76, + 858, + 104, + 298, + 104 + ], + "score": 0.874 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.861 + }, + { + "category_id": 13, + "poly": [ + 1103, + 1947, + 1278, + 1947, + 1278, + 1979, + 1103, + 1979 + ], + "score": 0.8, + "latex": "\\mathrm { ( G { - } F P R = 1 0 ^ { - 3 } }" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 252.0, + 391.0, + 252.0, + 391.0, + 276.0, + 336.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 252.0, + 467.0, + 252.0, + 467.0, + 276.0, + 416.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 254.0, + 944.0, + 254.0, + 944.0, + 278.0, + 742.0, + 278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 252.0, + 1369.0, + 252.0, + 1369.0, + 276.0, + 1221.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 269.0, + 1329.0, + 269.0, + 1329.0, + 293.0, + 1259.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 366.0, + 639.0, + 366.0, + 639.0, + 390.0, + 533.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 364.0, + 797.0, + 364.0, + 797.0, + 395.0, + 717.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 366.0, + 948.0, + 366.0, + 948.0, + 392.0, + 901.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 364.0, + 1159.0, + 364.0, + 1159.0, + 394.0, + 1043.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 466.0, + 609.0, + 466.0, + 609.0, + 492.0, + 558.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 467.0, + 817.0, + 467.0, + 817.0, + 492.0, + 696.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 464.0, + 967.0, + 464.0, + 967.0, + 494.0, + 878.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 466.0, + 1130.0, + 466.0, + 1130.0, + 492.0, + 1062.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 567.0, + 609.0, + 567.0, + 609.0, + 593.0, + 556.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 692.0, + 567.0, + 812.0, + 567.0, + 812.0, + 591.0, + 692.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 567.0, + 969.0, + 567.0, + 969.0, + 591.0, + 880.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 569.0, + 1123.0, + 569.0, + 1123.0, + 587.0, + 1081.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 671.0, + 613.0, + 671.0, + 613.0, + 684.0, + 548.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 664.0, + 996.0, + 664.0, + 996.0, + 694.0, + 876.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 668.0, + 1159.0, + 668.0, + 1159.0, + 692.0, + 1041.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 768.0, + 622.0, + 768.0, + 622.0, + 793.0, + 547.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 768.0, + 806.0, + 768.0, + 806.0, + 793.0, + 706.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 768.0, + 980.0, + 768.0, + 980.0, + 793.0, + 872.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 770.0, + 1132.0, + 770.0, + 1132.0, + 789.0, + 1071.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 863.0, + 608.0, + 863.0, + 608.0, + 894.0, + 556.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 869.0, + 812.0, + 869.0, + 812.0, + 888.0, + 694.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 867.0, + 954.0, + 867.0, + 954.0, + 891.0, + 895.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 869.0, + 1136.0, + 869.0, + 1136.0, + 888.0, + 1060.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 968.0, + 639.0, + 968.0, + 639.0, + 986.0, + 533.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 966.0, + 814.0, + 966.0, + 814.0, + 990.0, + 700.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 968.0, + 965.0, + 968.0, + 965.0, + 986.0, + 888.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 972.0, + 1124.0, + 972.0, + 1124.0, + 985.0, + 1068.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1070.0, + 781.0, + 1070.0, + 781.0, + 1089.0, + 732.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 1070.0, + 948.0, + 1070.0, + 948.0, + 1089.0, + 895.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1171.0, + 607.0, + 1171.0, + 607.0, + 1192.0, + 564.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 1173.0, + 789.0, + 1173.0, + 789.0, + 1192.0, + 721.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 1173.0, + 944.0, + 1173.0, + 944.0, + 1192.0, + 903.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1173.0, + 1155.0, + 1173.0, + 1155.0, + 1192.0, + 1041.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 1270.0, + 624.0, + 1270.0, + 624.0, + 1289.0, + 526.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 1274.0, + 793.0, + 1274.0, + 793.0, + 1292.0, + 713.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 1272.0, + 965.0, + 1272.0, + 965.0, + 1290.0, + 888.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1372.0, + 637.0, + 1372.0, + 637.0, + 1397.0, + 524.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 1371.0, + 783.0, + 1371.0, + 783.0, + 1402.0, + 723.0, + 1402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 1372.0, + 971.0, + 1372.0, + 971.0, + 1397.0, + 888.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1371.0, + 1162.0, + 1371.0, + 1162.0, + 1395.0, + 1045.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 1477.0, + 613.0, + 1477.0, + 613.0, + 1501.0, + 554.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1475.0, + 783.0, + 1475.0, + 783.0, + 1501.0, + 732.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1473.0, + 986.0, + 1473.0, + 986.0, + 1503.0, + 876.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 1475.0, + 1162.0, + 1475.0, + 1162.0, + 1499.0, + 1047.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1576.0, + 645.0, + 1576.0, + 645.0, + 1600.0, + 522.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1578.0, + 776.0, + 1578.0, + 776.0, + 1596.0, + 732.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1578.0, + 979.0, + 1578.0, + 979.0, + 1596.0, + 876.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 1576.0, + 1134.0, + 1576.0, + 1134.0, + 1600.0, + 1062.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 1635.0, + 1268.0, + 1635.0, + 1268.0, + 1665.0, + 1238.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 1676.0, + 615.0, + 1676.0, + 615.0, + 1695.0, + 550.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 1673.0, + 819.0, + 1673.0, + 819.0, + 1697.0, + 696.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1675.0, + 952.0, + 1675.0, + 952.0, + 1693.0, + 897.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 1676.0, + 1160.0, + 1676.0, + 1160.0, + 1695.0, + 1037.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 1771.0, + 605.0, + 1771.0, + 605.0, + 1790.0, + 562.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 1771.0, + 781.0, + 1771.0, + 781.0, + 1790.0, + 717.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 1771.0, + 979.0, + 1771.0, + 979.0, + 1790.0, + 884.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 1771.0, + 1119.0, + 1771.0, + 1119.0, + 1790.0, + 1058.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 1870.0, + 567.0, + 1870.0, + 567.0, + 1883.0, + 535.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1868.0, + 776.0, + 1868.0, + 776.0, + 1887.0, + 736.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1867.0, + 961.0, + 1867.0, + 961.0, + 1885.0, + 899.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1870.0, + 1134.0, + 1870.0, + 1134.0, + 1883.0, + 1068.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.75, + 266.5, + 387.75, + 266.5, + 387.75, + 291.0, + 335.75, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 267.0, + 465.0, + 267.0, + 465.0, + 293.0, + 418.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.75, + 663.0, + 780.75, + 663.0, + 780.75, + 689.0, + 726.75, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1068.5, + 601.0, + 1068.5, + 601.0, + 1089.0, + 564.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.75, + 1321.0, + 1283.75, + 1321.0, + 1283.75, + 1364.0, + 1218.75, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1948.0, + 1102.0, + 1948.0, + 1102.0, + 1982.0, + 293.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 1948.0, + 1407.0, + 1948.0, + 1407.0, + 1982.0, + 1279.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1978.0, + 1405.0, + 1978.0, + 1405.0, + 2016.0, + 293.0, + 2016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2009.0, + 910.0, + 2009.0, + 910.0, + 2047.0, + 294.0, + 2047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 70.0, + 859.0, + 70.0, + 859.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2124.0, + 828.0, + 2124.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 22, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 306, + 485, + 1391, + 485, + 1391, + 1637, + 306, + 1637 + ], + "score": 0.978 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 859, + 75, + 859, + 105, + 297, + 105 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2113, + 834, + 2113 + ], + "score": 0.851 + }, + { + "category_id": 4, + "poly": [ + 297, + 1689, + 1404, + 1689, + 1404, + 1783, + 297, + 1783 + ], + "score": 0.541 + }, + { + "category_id": 4, + "poly": [ + 297, + 1688, + 1403, + 1688, + 1403, + 1783, + 297, + 1783 + ], + "score": 0.521 + }, + { + "category_id": 13, + "poly": [ + 296, + 1718, + 357, + 1718, + 357, + 1750, + 296, + 1750 + ], + "score": 0.85, + "latex": "1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 1378, + 1692, + 1404, + 1692, + 1404, + 1718, + 1378, + 1718 + ], + "score": 0.67, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1293, + 1689, + 1406, + 1689, + 1406, + 1720, + 1293, + 1720 + ], + "score": 0.27, + "latex": "\\mathrm { { \\bf G } { - } F P R = }" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 499.0, + 481.0, + 499.0, + 481.0, + 524.0, + 384.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 499.0, + 777.0, + 499.0, + 777.0, + 524.0, + 653.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 498.0, + 1035.0, + 498.0, + 1035.0, + 524.0, + 941.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 496.0, + 1317.0, + 496.0, + 1317.0, + 525.0, + 1191.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 530.0, + 458.0, + 530.0, + 458.0, + 588.0, + 317.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 541.0, + 920.0, + 541.0, + 920.0, + 576.0, + 872.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 605.0, + 452.0, + 605.0, + 452.0, + 643.0, + 374.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 654.0, + 568.0, + 654.0, + 568.0, + 722.0, + 341.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 658.0, + 1120.0, + 658.0, + 1120.0, + 724.0, + 866.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 711.0, + 579.0, + 711.0, + 579.0, + 796.0, + 299.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 715.0, + 1127.0, + 715.0, + 1127.0, + 794.0, + 852.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 787.0, + 576.0, + 787.0, + 576.0, + 876.0, + 299.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 786.0, + 1128.0, + 786.0, + 1128.0, + 875.0, + 851.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 852.0, + 555.0, + 852.0, + 555.0, + 933.0, + 306.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 851.0, + 1124.0, + 851.0, + 1124.0, + 941.0, + 855.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1056.0, + 519.0, + 1056.0, + 519.0, + 1082.0, + 359.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1076.0, + 845.0, + 1076.0, + 845.0, + 1170.0, + 635.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 1148.0, + 650.0, + 1148.0, + 650.0, + 1221.0, + 581.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 1164.0, + 703.0, + 1164.0, + 703.0, + 1208.0, + 645.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1217.0, + 536.0, + 1217.0, + 536.0, + 1287.0, + 321.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 1227.0, + 1059.0, + 1227.0, + 1059.0, + 1275.0, + 983.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 1274.0, + 579.0, + 1274.0, + 579.0, + 1357.0, + 302.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 1279.0, + 780.0, + 1279.0, + 780.0, + 1352.0, + 647.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 1282.0, + 1122.0, + 1282.0, + 1122.0, + 1351.0, + 855.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 1348.0, + 580.0, + 1348.0, + 580.0, + 1438.0, + 300.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 1352.0, + 845.0, + 1352.0, + 845.0, + 1439.0, + 632.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 1347.0, + 1128.0, + 1347.0, + 1128.0, + 1438.0, + 851.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 1421.0, + 506.0, + 1421.0, + 506.0, + 1494.0, + 302.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1430.0, + 642.0, + 1430.0, + 642.0, + 1486.0, + 587.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 1426.0, + 706.0, + 1426.0, + 706.0, + 1490.0, + 646.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 1438.0, + 759.0, + 1438.0, + 759.0, + 1477.0, + 718.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 1416.0, + 1120.0, + 1416.0, + 1120.0, + 1502.0, + 858.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 1496.0, + 453.0, + 1496.0, + 453.0, + 1552.0, + 315.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 1514.0, + 490.0, + 1514.0, + 490.0, + 1537.0, + 458.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 1517.0, + 551.0, + 1517.0, + 551.0, + 1534.0, + 532.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 1486.0, + 830.0, + 1486.0, + 830.0, + 1558.0, + 579.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 1559.0, + 458.0, + 1559.0, + 458.0, + 1612.0, + 311.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 1550.0, + 777.0, + 1550.0, + 777.0, + 1623.0, + 577.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 1619.0, + 509.0, + 1619.0, + 509.0, + 1644.0, + 371.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 829.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1688.0, + 1292.0, + 1688.0, + 1292.0, + 1724.0, + 292.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1715.0, + 295.0, + 1715.0, + 295.0, + 1757.0, + 292.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1715.0, + 1404.0, + 1715.0, + 1404.0, + 1757.0, + 358.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1749.0, + 703.0, + 1749.0, + 703.0, + 1787.0, + 294.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1687.0, + 1292.0, + 1687.0, + 1292.0, + 1725.0, + 293.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1716.0, + 295.0, + 1716.0, + 295.0, + 1757.0, + 292.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1716.0, + 1404.0, + 1716.0, + 1404.0, + 1757.0, + 358.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1751.0, + 703.0, + 1751.0, + 703.0, + 1785.0, + 293.0, + 1785.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 23, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/r1xwA34KDB/r1xwA34KDB.md b/parse/train/r1xwA34KDB/r1xwA34KDB.md new file mode 100644 index 0000000000000000000000000000000000000000..54a3b9c738eec555040be45062435f2039213d6a --- /dev/null +++ b/parse/train/r1xwA34KDB/r1xwA34KDB.md @@ -0,0 +1,351 @@ +# LEARNING INVARIANTS THROUGH SOFT UNIFICATION + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +Human reasoning involves recognising common underlying principles across many examples by utilising variables. The by-products of such reasoning are invariants that capture patterns across examples such as “if someone went somewhere then they are there” without mentioning specific people or places. Humans learn what variables are and how to use them at a young age, and the question this paper addresses is whether machines can also learn and use variables solely from examples without requiring human pre-engineering. We propose Unification Networks that incorporate soft unification into neural networks to learn variables and by doing so lift examples into invariants that can then be used to solve a given task. We evaluate our approach on four datasets to demonstrate that learning invariants captures patterns in the data and can improve performance over baselines. + +# 1 INTRODUCTION + +Humans have the ability to process symbolic knowledge and maintain symbolic thought (Unger & Deacon, 1998). When reasoning, humans do not require combinatorial enumeration of examples but instead utilise invariant patterns with placeholders replacing specific entities. Symbolic cognitive models (Lewis, 1999) embrace this perspective with the human mind seen as an information processing system operating on formal symbols such as reading a stream of tokens in natural language. The language of thought hypothesis (Morton & Fodor, 1978) frames human thought as a structural construct with varying sub-components such as $^ { 6 6 } \mathrm { X }$ went to Y”. By recognising what varies across examples, humans are capable of lifting examples into invariant principles that account for other instances. This symbolic thought with variables is learned at a young age through symbolic play (Piaget, 2001). For instance a child learns that a sword can be substituted with a stick (Frost et al., 2004) and engage in pretend play. + +Although variables are inherent in models of computation and symbolic formalisms, as in first-order logic (Russell & Norvig, 2016), they are pre-engineered and used to solve specific tasks by means of unification or assignments that bound variables to given values. However, when learning from data only, being able to recognise when and which symbols should take on different values, i.e. symbols that can act as variables, is crucial for lifting examples into general principles that are invariant across multiple instances. Figure 1 shows the invariant learned by our approach: if someone is the same thing as someone else then they have the same colour. With this invariant, our approach can solve all of the training and test examples in task 16 of the bAbI dataset (Weston et al., 2016). + +Figure 1: Invariant learned for bAbI task 16, basic induction, where $X$ :bernhard denotes a variable with default symbol bernhard. This single invariant accounts for all the training and test examples. + +
X:bernhard is a Y:frog Z:lily is a Y:frog
Z:lily is A:green
what colour is X:bernhard
A:green
+ +In this paper we address the question of whether a machine can learn and use the notion of a variable, i.e. a symbol that can take on different values. For instance, given an example of the form “bernhard is a frog” the machine would learn that the token “bernhard” could be someone else and the token “frog” could be something else. If we consider unification a selection of the most appropriate value for a variable given a choice of values, we can reframe it as a form of attention. Attention models (Bahdanau et al., 2015; Luong et al., 2015; Chaudhari et al., 2019) allow neural networks to focus, attend to certain parts of the input often for the purpose of selecting a relevant portion. Since attention mechanisms are also differentiable they are often jointly learned within a task. This perspective motivates our idea of a unification mechanism that utilises attention and is therefore fully differentiable which we refer to as soft unification. + +Hence, we propose an end-to-end differentiable neural network approach for learning and utilising the notion of a variable that in return can lift examples into invariants used by the network to perform reasoning tasks. Specifically, we (i) propose a novel architecture capable of learning and using variables by lifting a given example through soft unification, (ii) present the empirical results of our approach on four datasets and (iii) analyse the learned invariants that capture the underlying patterns present in the tasks. Our implementation using Chainer (Tokui et al., 2015) is publicly available at [link removed](anonymous link provided with submission). + +# 2 SOFT UNIFICATION + +Reasoning with variables involves identifying what variables are, the setting in which they are used as well as the process by which they are assigned values. When the varying components, i.e. variables, of an example are identified, the remaining structure can be lifted into an invariant which then accounts for multiple other instances. + +Definition 1 (Variable). Given a set of symbols $\mathbb { S }$ , a variable $X$ is defined as a pair $\pmb { X } \triangleq \left( \mathbf { x } , s _ { d } \right)$ where $s _ { d } \in \mathbb { S }$ is the default symbol of the variable and $\mathbf { X }$ is a discrete random variable of which the support is $\mathbb { S }$ . The representation of a variable $\phi _ { V } ( { \pmb X } ; s _ { d } )$ is equal to the expected value of the corresponding random variable $\mathbf { X }$ given the default symbol $s _ { d }$ : + +$$ +\phi _ { V } ( X { : } s _ { d } ) = \mathbb { E } _ { \mathbf { x } \sim P } [ \phi ( x ) ] = \sum _ { x \in \mathbb { S } } P ( \mathbf { x } = x | s _ { d } ) \phi ( x ) +$$ + +where $\phi : \mathbb { S } \mathbb { R } ^ { d }$ is a $d$ -dimensional real-valued feature of a symbol $s$ + +For example, $\phi$ could be an embedding and $\phi _ { V } ( { \pmb X } ; s _ { d } )$ would become a weighted sum of symbol embeddings as in conventional attention models. The default symbol of a variable is intended to capture the variable’s bound meaning following the idea of referants by Frege (1948). We denote variables using $X , Y , A$ etc. such as $X$ :bernhard where $X$ is the name of the variable and bernhard the default symbol as shown in Figure 1. + +Definition 2 (Invariant). Given a structure (e.g. list, grid) $\mathcal { G }$ over $\mathbb { S }$ , an invariant is a pair $I \triangleq ( G , \psi )$ where $G \in { \mathcal { G } }$ is the invariant example such as a tokenised story with tokens as symbols and $\psi : \mathbb { S } \to \quad$ $[ 0 , 1 ]$ is a function representing the degree to which the symbol is considered a variable. Thus, the final representation of a symbol $s$ included in $G$ , $\phi _ { I } ( s )$ is: + +$$ +\phi _ { I } ( s ) = ( 1 - \psi ( s ) ) \phi ( s ) + \psi ( s ) \phi _ { V } ( X { : } s ) +$$ + +the linear interpolation between its representation $\phi ( s )$ and its variable bound value with itself as the default symbol $\phi _ { V } ( \pmb { X } ; s )$ . + +We adhere to the term invariant and refrain from mentioning rules, unground rules, etc. used in logicbased formalisms, e.g. Muggleton & de Raedt (1994), since neither the invariant structure needs to be rule-like nor the variables carry logical semantics. This distinction is clarified in Section 6. + +Definition 3 (Unification). Given an invariant $I$ and an example $K \in \mathcal G$ , unification binds the variables in $I$ to symbols in $K$ . Defined as a function $g : \mathcal { T } \times \mathcal { G } \to \mathcal { G }$ , unification binds variables by computing the probability mass functions, $P$ in equation 1, and returns the unified representation using equation 2. The probability mass function of a variable $X { : } s _ { d }$ is: + +$$ +P ( { \bf x } = x | s _ { d } ) = \mathrm { s o f t m a x } ( \phi _ { U } ( s _ { d } ) \phi _ { U } ( K ) ^ { T } ) , \ : K = \mathrm { s u p p o r t \ o f \ x } +$$ + +where $\phi _ { U } : \mathbb { S } \to \mathbb { R } ^ { d }$ is the unifying feature of a symbol and $\phi _ { U } ( K ) \in \mathbb { R } ^ { | K | \times d }$ is applied element wise to symbols in $K$ . If $g$ is differentiable, it is referred to as soft unification. + +We distinguish $\phi$ from $\phi _ { U }$ to emphasise that the unifying properties of the symbols might be different from their representations. For example, $\phi$ (bernhard) could represent a specific person whereas $\phi _ { U }$ (bernhard) the notion of someone. + +Overall soft unification incorporates 3 learnable components: $\phi , \psi , \phi _ { U }$ which denote the base features, variableness and unifying features of a symbol respectively. Given an upstream, potentially + +task specific, network $f : \mathcal G \to \mathbb S$ , an invariant $I \in \mathcal { Z }$ and an input example $K \in { \mathcal { G } }$ with a corresponding desired output $a \in \mathbb { S }$ , the following holds: + +$$ +f \circ g ( I , K ) = f ( K ) = a +$$ + +where $f$ now predicts based on the unified representation produced by $g$ . In this work, we focus on $g$ , the invariants it produces together with the interaction of $f \circ g$ . + +# 3 UNIFICATION NETWORKS + +Since soft unification is end-to-end differentiable, it can be incorporated into existing task-specific upstream architectures. We present 3 architectures that model $f \circ g$ using multi-layer perceptrons (MLP), convolutional neural networks (CNN) and memory networks (Weston et al., 2015) to demonstrate the flexibility of our approach. In all cases, the $d$ dimensional representation of symbols are learnable embeddings $\phi ( s ) = O [ s ] ^ { T } E$ with $E \in \mathbb { R } ^ { | \mathbb { S } | \times d }$ randomly initialised by $\mathcal { N } ( 0 , 1 )$ and $O [ s ]$ the one-hot encoding of the symbol. The variableness of symbols is a learnable weight $\psi _ { w } ( s ) = \sigma ( w _ { s } )$ where $\boldsymbol { w } \in \mathbb { R } ^ { | \mathbb { S } | }$ and $\sigma$ is the sigmoid function. We consider every symbol independently a variable irrespective of its surrounding context and leave further contextualised formulations as future work. The underlying intuition of this configuration is that a useful symbol for a correct prediction might need to take on other values for different inputs. This usefulness can be viewed as the inbound gradient to the corresponding $w _ { s }$ parameter and $\psi _ { w } ( s )$ acting as a gate. For further model details including the size of the embeddings, please refer to Appendix A. + +Unification MLP (UMLP) ( $f$ : MLP, $g$ : RNN) We combine soft unification into a multi-layer perceptron to process fixed length inputs. In this case, the structure $\mathcal { G }$ is a sequence of symbols with a fixed length $l$ , e.g. a sequence of digits 4234. Given an embedded input $\phi ( \pmb { k } ) \in R ^ { l \times d }$ , the upstream MLP computes the output symbol based on the flattened representations $f ( \phi ( \pmb { k } ) ) =$ softmax $( h E ^ { T } )$ where $\hat { h } \in \mathbb { R } ^ { d }$ is the output of the last layer. However, to compute the unifying features $\phi _ { U }$ , definition 3, $g$ uses a bi-directional GRU (Cho et al., 2014) running over $\phi ( k )$ such that $\phi _ { U } ( k ) = W _ { U } \Phi ( k )$ where $\Phi ( k ) \in \mathbb { R } ^ { d }$ is the hidden state of the GRU at symbol $k$ and $\dot { W _ { U } } \in \mathbb { R } ^ { d \times d }$ is a learnable parameter. This model emphasises the flexibility around the boundary of $f \circ g$ and that the unifying features can be computed in any differentiable manner. + +![](images/ef39205e27e246bc37f3a17ac72fc093d4914d8003b2bd66d967fd9856bb6ce9.jpg) +Figure 2: Graphical overview of soft unification within a memory network. Each sentence is processed by two bi-directional RNNs for memory and unification. At each iteration the context attention selects which sentences to unify and the invariant produces the same answer as the example. + +Unification CNN (UCNN) ( $f$ : CNN, $g$ : CNN) Given a grid of embedded symbols $\phi ( K ) \ \in$ $\mathbb { R } ^ { w \times h \times d }$ where $w$ is the width and $h$ the height, we use a convolutional neural network such that the final prediction is $f ( \phi ( K ) ) = \mathrm { s o f t m a x } ( ( W h + b ) E ^ { T } )$ where $^ { h }$ this time is the result of global max pooling and $W , b$ are learnable parameters. We also model $g$ using a separate convolutional network with the same architecture as $f$ and set $\phi _ { U } ( k ) = c _ { 2 } ( \mathrm { r e l u } ( c _ { 1 } ( k ) )$ where $c _ { 1 } , c _ { 2 }$ are the convolutional layers. The grid is padded with 0s to obtain $w \times h \times d$ after each convolution such that every symbol has a unifying feature. This model conveys how soft unification can be adapted to the specifics of the domain for example by using a convolution in a spatially structured input. + +Unification Memory Networks (UMN) ( $f$ : MemNN, $g$ : RNN) Soft unification does not need to happen prior to $f$ in a $f \circ g$ fashion but can also be incorporated at any intermediate stage multiple times. To demonstrate this ability, we unify the symbols at different memory locations at each iteration of a Memory Network (Weston et al., 2015). Memory networks can handle a list of lists structure such as a tokenised story as shown in Figure 2. The memory network $f$ uses the final hidden state of a bi-directional GRU (outer squares in Figure 2) as the sentence representations to compute a context attention. At each iteration, we unify the words between the attended sentences using the same approach in UMLP with another bi-directional GRU (inner diamonds in Figure 2) for unifying features $\phi _ { U } ( \mathbf { b e r n h a r d } ) = W _ { U } \Phi ($ (bernhard). Following equation 2, the new unified representation of the memory slot is computed and $f$ uses it to perform the next iteration. Concretely, $g$ produces an unification tensor $\pmb { \mathsf { U } } \in \dot { \mathbb { R } } ^ { M \times m \times N \times d }$ where $M$ and $m$ is the number of sentences and words in the invariant respectively, and $N$ is the number of sentences in the example such that after the context attentions are applied over $M$ and $N$ , we obtain $\boldsymbol { \phi } ( \pmb { k } ) \in \mathbb { R } ^ { m \times d }$ as the unified sentence at that iteration. Note that unlike in the UMLP case, the sentences can be of varying length. The prediction is then softmax $( W h _ { I } ^ { J } + b )$ where $h _ { I } ^ { J }$ is the hidden state of the invariant after $J$ iterations. This setup, however, requires pre-training $f$ such that the context attentions match the correct sentences. + +Table 1: Sample context, query and answer triples and their training sizes per task. For distribution of generated number of examples per task on Sequence and Grid data refer to Appendix B. + +
DatasetContextQueryAnswerTraining Size
Sequence8384duplicate8≤1k,≤50
Grid00 3 1 859corner7≤1k,≤50
bAbIMary went to the kitchen. Sandra journeyed to the garden.Where is Mary?kitchen1k,50
Logicp(X)← q(X). q(a).p(a).True1k, 10k,50
+ +A task might contain different questions such as “Where is X?” and “Why did $\mathrm { X }$ go to Y?”. To let the models differentiate between questions and potentially learn different invariants, we extend them with a repository of invariants $I \in \mathbb { I }$ and aggregate the predictions from each invariant. One simple approach is to sum the predictions of the invariants $\textstyle \sum _ { I \in \mathbb { I } } f \circ g ( I , K )$ used in UMLP and UCNN. Another approach could be to use features from the invariants such as memory representations in the case of UMN. For UMN, we weigh the predictions using a bilinear attention $\eta$ based on the hidden states at the first iteration $h _ { I } ^ { 0 }$ and $h _ { K } ^ { 0 }$ such that $\eta = \mathrm { s o f t m a x } ( h _ { I } ^ { 0 } W h _ { K } ^ { 0 ^ { T } } )$ . To initially form the repository of invariants, we use the bag-of-words representation of the questions and find the most dissimilar ones based on their cosine similarity as a heuristic to obtain varied examples. + +# 4 DATASETS + +We use 4 datasets consisting of context, query and an answer $( C , \pmb { q } , a )$ : fixed length sequences of symbols, shapes of symbols in a grid, story based natural language reasoning with the bAbI (Weston et al., 2016) dataset and logical reasoning represented as logic programs, examples shown in Table 1 with further samples in Appendix B. In each case we use an appropriate model: UMLP for fixed length sequences, UCNN for grid and UMN for iterative reasoning. We use synthetic datasets of which the data generating distributions are known to evaluate not only the quantitative performance but also the quality of the invariants learned by our approach. + +Fixed Length Sequences We generate sequences of length $l = 4$ with 8 unique symbols represented as digits to predict (i) a constant, (ii) the head of the sequence, (iii) the tail and (iv) the duplicate symbol. We randomly generate 1000 triples and then only take the unique ones to ensure the test split contains unseen examples. The training is then performed over a 5-fold cross-validation. + +Grid To spatially organise symbols, we generate a grid of size $3 \times 3$ with 8 unique symbols organised into $2 \times 2$ box of identical symbol, a vertical, diagonal or horizontal sequence of length 3, a cross or a plus shape and a triangle. In each task we predict (i) the identical symbol, (ii) the head of the sequence, (iii) the centre of the cross or plus and (iv) the corner of the triangle respectively. We follow the same procedure from sequences and randomly generate 1000 discarding duplicate triples. + +bAbI The bAbI dataset has become a standard benchmark for evaluating memory based networks. +It consists of 20 synthetically generated natural language reasoning tasks (refer to Weston et al. + +(2016) for task details). We take the 1k English set and use 0.1 of the training set as validation. Each token is lower cased and considered a unique symbol. Following previous works (Seo et al., 2017; Sukhbaatar et al., 2015), we take multiple word answers also to be a unique symbol in S. + +Logical Reasoning To demonstrate the flexibility of our approach and distinguish our notion of a variable from that used in logic based formalisms, we generate logical reasoning tasks in the form of logic programs using the procedure by Cingillioglu & Russo (2019). The tasks involve learning $f ( \bar { C } , \bar { Q } ) \stackrel { - } { = } \mathrm { T r u e } \bar { C } \vdash \bar { Q }$ over 12 classes of logic programs exhibiting varying paradigms of logical reasoning including negation by failure (Clark, 1978). We generate 1k and 10k logic programs per task for training with 0.1 as validation and another 1k for testing. We set the arity of literals to 1 or 2 using one random character from the English alphabet for predicates and constants, e.g. $p ( p )$ and an upper case character for logical variables, e.g. $p ( \mathrm { x } )$ . + +# 5 EXPERIMENTS + +We probe three aspects of soft unification: the impact of unification on performance over unseen data, the effect of multiple invariants and data efficiency. To that end, we train UMLP and UCNN with and without unification, UMN with pre-training using 1 or 3 invariants over either the entire training set or only 50 examples. Every model is trained 3 times via back-propagation using Adam (Kingma & Ba, 2015) on an Intel Core i7-6700 CPU using the following objective function: + +$$ +J = \lambda _ { K } \mathcal { L } _ { \mathrm { n l l } } ( f ( K ) , a ) + \lambda _ { U } \left[ \mathcal { L } _ { \mathrm { n l l } } ( f \circ g ( I , K ) , a ) + \tau \sum _ { s \in \mathbb { S } } \psi _ { w } ( s ) \right] +$$ + +where ${ \mathcal { L } } _ { \mathrm { n l l } }$ is the negative log-likelihood with sparsity regularisation over $\psi$ at $\tau = 0 . 1$ to discourage the models from utilising spurious number of variables. For UMLP and UCNN, we set $\lambda _ { K } ~ =$ 0, $\lambda _ { U } = 1$ for training just the unified output and the converse for the non-unifying versions. To pre-train the UMN, we start with $\lambda _ { K } = 1 , \lambda _ { U } = 0$ for 40 epochs then set $\lambda _ { U } = 1$ to jointly train the unified output. For iterative tasks, the mean squared error between hidden states $( { h _ { I } ^ { j } - h _ { K } ^ { j } } ) ^ { 2 }$ at each iteration $j$ and, in the strongly supervised cases, the negative log-likelihood for the context attentions using provided supporting facts are also added to the objective function. Further details such as batch size and total number of epochs are available in Appendix C. + +![](images/4aca829873f131070cc20ca4370a0d6ac4a79c32120a7d5e548cfc6ffbcee87f.jpg) +Figure 3: Test accuracy over iterations for Unification MLP and Unification CNN models with 1 invariant versus no unification. We observe that with soft unification the models achieve higher accuracy with fewer iterations than their plain counterparts on both per task training sizes. + +Figure 3 portrays how soft unification generalises better to unseen examples in test sets - the same sequence or grid never appears in both the training and test sets as outlined in Section 4 - over plain models. Despite $f \circ g$ having more trainable parameters than $f$ alone, the models with unification not only maintain higher accuracy in each iteration and solve the tasks in as few as 250 iterations with $\leq 1 0 0 0$ training examples but also improve accuracy by $\approx 0 . 3$ when trained with only $\leq 5 0$ per task. We believe soft unification architecturally biases the models towards learning structural patterns which in return achieves better results on recognising common patterns of symbols across examples. Results with multiple invariants are identical and the models seem to ignore the extra invariants due to the fact that the tasks can be solved with a single invariant and the regularisation applied on $\psi$ zeroing out unnecessary invariants; further results in Appendix D. The fluctuations in accuracy around iterations 750 to 1000 in UCNN are also caused by penalising $\psi$ which forces the model to relearn the task with less variables half way through training. + +Table 2: Aggregate error rates $( \% )$ on bAbI 1k for UMN and N2N, GN2N, MemNN by Sukhbaatar et al. (2015), Liu & Perez (2017) and Weston et al. (2015) respectively. Full comparison on individual tasks are available in Appendix D. + +
Training Size Supervision #Invs /Model1k501k
Weak 13Strong 3WeakQRNStrong MemNN
31N2N GN2NEntNet
Mean19.120.56.36.027.613.912.729.611.36.7
#>5%810441711101554
+ +Following Tables 2 and 3, we observe a trend of better performance through strong supervision, more data per task and using only 1 invariant. We believe strong supervision aids with selecting the correct sentences to unify and in a weak setting the model attempts to unify arbitrary context sentences often failing to follow the iterative reasoning chain. The increase in performance with more data and strong supervision is consistent with previous work reflecting how $f \circ g$ can be bounded by the efficacy of $f$ modelled as a memory network. As a result, only in the supervised case do we observe a minor improvement over MemNN by 0.7 in Table 2 and no improvement in the weak case over comparable memory based networks in Table 3. This dependency on $f$ also limits the ability of $f \circ g$ to learn from 50 examples per task failing 17/20 and 12/12 of bAbI and logical reasoning tasks respectively. The increase in error rate with 3 invariants, we speculate, stems from having more parameters and more pathways in the model rendering training more difficult. + +Table 3: Aggregate task error rates $( \% )$ on the logical reasoning dataset for UMN and DMN, IMA by Cingillioglu & Russo (2019). Strong supervision, more data and only 1 invariant seem to improve the performance of UMN over plain iterative models. Individual task results are in Appendix D. + +
Training Size Supervision1k5020k Weak
WeakStrongWeakStrong
121212122222
# Invs /Model1313131133DMNIMA
Mean36.439.314.328.921.531.82.412.216.047.121.29.1
#>5%91171171015912115
+ +# 6 ANALYSIS + +After training, we can extract the learned invariants by applying a threshold on $\forall s \in \mathbb { S } : \psi ( s ) > t$ indicating whether a symbol is used as a variable or not. We set $t = 0 . 0$ for all datasets except for bAbI, we use $t = 0 . 1$ . The magnitude of this threshold seems to depend on the amount of regularisation $\tau$ , equation 5, and the number of training steps along with batch size all controlling how much $\psi$ is pushed towards 0. Sample invariants shown in Figure 4 describe the common patterns present in the tasks with parts that contribute towards the final answer becoming variables. Extra symbols such as is or travelled do not emerge as variables, as shown in Figure 4a; we attribute this behaviour to the fact that changing the token travelled to went does not influence the prediction but changing the action, the value of Z:left to ‘picked’ does. However, based on random initialisation, our approach can convert an arbitrary symbol into a variable and let $f$ compensate for the unifications it produces. For example, the invariant $^ { * } X { : } 8 \ 5 \ 2 \ 2 ^ { , }$ could predict the tail of another example by unifying the head with the tail using $\phi _ { U }$ , equation 3, of those symbols. Further examples are shown in Appendix D. Pre-training $f$ as done in UMN seems to produce more robust and consistent invariants compared to immediately training $f \circ g$ since, we speculate, by equation 4 $f$ might encourage $g ( I , K ) \approx K$ . + +Interpretability versus Ability A desired property of interpretable models is transparency (Lipton, 2018). A novel outcome of the learned invariants in our approach is that they provide an approximation of the underlying general principle present in the data such as the structure of multi-hop reasoning shown in Figure 4e. However, certain aspects regarding the ability of the model such as how it performs temporal reasoning, are still hidden inside $f$ . In Figure 4b, although we observe $\mathbf { Z }$ :morning as a variable, the overall learned invariant captures nothing about how changing the value of $\mathbf { Z }$ :morning alters the behaviour of $f$ . The model might look before or after a certain time point + +
Y:john Z:left the X:football Y:john travelled to the A:office
where is the X:football
A:office
+ +this Z:morning $X$ :bill went to the Y:school yesterday $X$ :bill journeyed to the A:park where was $X$ :bill before the Y:school A:park + +(a) bAbI task 2, two supporting facts. The model also learns Z:left since people can also drop or pick up objects potentially affecting the answer. + +(b) bAbI task 14, time reasoning. $X$ :bill and Y:school are recognised as variables alongside Z:morning capturing when someone went which is crucial to this task. + +![](images/2b918647c3c931a6d2d86be39b48a163abe2920f4808362f723da0efbea11058.jpg) + +Figure 4: Invariants learned across the four datasets using the three architectures. For iterative reasoning datasets, bAbI and logical reasoning, they are taken from strongly supervised UMN. + +$X$ :bill went somewhere depending what $\mathbf { Z }$ :morning binds to. Without the regularising term on $\psi ( s )$ , we initially noticed the models using, one might call extra, symbols as variables and binding them to the same value occasionally producing unifications such as “bathroom bathroom to the bathroom” and still $f$ predicting, perhaps unsurprisingly, the correct answer as bathroom. Hence, regularising $\psi$ with the correct amount $\tau$ in equation 5 to reduce the capacity of unification seems critical in extracting not just any invariant but one that represents the common structure. + +Soft unification from equation 3 reveals three main patterns: one-to-one, one-to-many or many-toone bindings as shown in Figure 5; further examples are in Appendix D. Figure 5a captures what one might expect unification to look like where variables unify with their corresponding counterparts, e.g. $X$ :bernhard with brian and Y:frog with lion. However, occasionally the model can optimise to use less variables and squeeze the required information into a single variable, for example by binding Y:bathroom to john and kitchen as shown in Figure 5b. We believe this occurs due to the sparsity constraint on $\psi ( s )$ encouraging the model to be as conservative as possible. Finally, if there are more variables than needed as in Figure 5c, we observe a many-to-one binding with Y:w and Z:e mapping to the same constant $q$ . This behaviour begs the question how does the model differentiate between $p ( q )$ and $p ( q , q )$ . We speculate the model uses the magnitude of $\psi ( w ) = 0 . 0 3 7$ and $\psi ( e ) = 0 . 0 4 2$ to encode the difference despite both variables unifying with the same constant. + +(a) bAbI task 16. A one-to-one mapping is created between variables $X$ :bernhard with brian and Y:frog with lion. + +![](images/87f061c60b0d26309091e0532f96894f1603bdbd0e818eb3a44cf2cee5b59e9e.jpg) +(c) Logical reasoning task 1. An arity 2 predicate is forced to bind with arity 1 creating a many-to-one binding. + +(b) bAbI task 6. Only Y:bathroom is recognised as variable creating a one-to-many binding to capture the same information. + +Figure 5: Variable bindings produced from equation 3. Darker cells indicate higher attention values. + +# 7 RELATED WORK + +Learning an underlying general principle in the form of an invariant is often the means for arguing generalisation in neural networks. For example, Neural Turing Machines (Graves et al., 2014) are tested on previously unseen sequences to support the view that the model might have captured the underlying pattern or algorithm. In fact, Weston et al. (2015) claim “MemNNs can discover simple linguistic patterns based on verbal forms such as (X, dropped, Y), (X, took, Y) or (X, journeyed to, Y) and can successfully generalise the meaning of their instantiations.” However, this claim is based on the output of $f$ and unfortunately it is unknown whether the model has truly learned such a representation or indeed is utilising it. Our approach sheds light to this ambiguity and presents these linguistic patterns explicitly as invariants ensuring their utility through $g$ without solely analysing the output of $f$ on previously unseen symbols. Although we associate these invariants with our existing understanding of the task to mistakenly anthropomorphise the machine, for example by thinking it has learned $X$ :mary as someone, it is important to acknowledge that these are just symbolic patterns. In these cases, our interpretations do not necessarily correspond to any understanding of the machine, relating to the Chinese room argument made by Searle (1980). + +Learning invariants by lifting ground examples is related to least common generalisation (Reynolds, 1970) by which inductive inference is performed on facts (Shapiro, 1981) such as generalising went(mary,kitchen) and went(john,garden) to went(X,Y). Unlike in a predicate logic setting, our approach allows for soft alignment and therefore generalisation between varying length sequences. Existing neuro-symbolic systems (Broda et al., 2002) focus on inducing rules that adhere to given logical semantics of what variables and rules are. For example, $\delta I L P$ (Evans & Grefenstette, 2018) constructs a network by rigidly following the given semantics of first-order logic. Similarly, Lifted Relational Neural Networks (Sourek et al., 2015) ground first-order logic rules into a neural network while Neural Theorem Provers (Rocktschel & Riedel, 2017) build neural networks using backwardchaining (Russell & Norvig, 2016) on a given background knowledge base with templates. However, the notion of a variable is pre-defined rather than learned with a focus on presenting a practical approach to solving certain problems, whereas our motivation stems from a cognitive perspective. + +At first it may seem the learned invariants, Section 6, make the model more interpretable; however, this transparency is not of the model $f$ but of the data. The invariant captures patterns that potentially approximates the data generating distribution but we still do not know how the model $f$ uses them upstream. Thus, from the perspective of explainable artificial intelligence (XAI) (Adadi & Berrada, 2018), learning invariants or interpreting them do not constitute an explanation of the reasoning model $f$ even though “if someone goes somewhere then they are there” might look like one. Instead, it can be perceived as causal attribution (Miller, 2019) in which someone being somewhere is attributed to them going there. This perspective also relates to gradient based model explanation methods such as Layer-Wise Relevance Propagation (Bach et al., 2015) and Grad-CAM (Selvaraju et al., 2017; Chattopadhay et al., 2018). Consequently, a possible view on $\psi$ , equation 2, is a gradient based usefulness measure such that a symbol utilised upstream by $f$ to determine the answer becomes a variable similar to how a group of pixels in an image contribute more to its classification. + +Finally, one can argue that our model maintains a form of counterfactual thinking (Roese, 1997) in which soft unification $g$ creates counterfactuals on the invariant example to alter the output of $f$ towards the desired answer, equation 4. The question where Mary would have been if Mary had gone to the garden instead of the kitchen is the process by which an invariant is learned through multiple examples during training. This view relates to methods of causal inference (Pearl, 2019; Holland, 1986) in which counterfactuals are vital as demonstrated in structured models by Pearl (1999). + +# 8 CONCLUSION + +We presented a new approach for learning variables and lifting examples into invariants through the usage of soft unification. Evaluating on four datasets, we analysed how Unification Networks perform comparatively to existing similar architectures while having the benefit of lifting examples into invariants that capture underlying patterns present in the tasks. Since our approach is end-toend differentiable, we plan to apply this technique to multi-modal tasks in order to yield multi-modal invariants for example in visual question answering. + +# REFERENCES + +Amina Adadi and Mohammed Berrada. Peeking inside the black-box: A survey on explainable artificial intelligence (XAI). IEEE Access, 6:52138–52160, 2018. doi: 10.1109/access.2018. 2870052. + +Sebastian Bach, Alexander Binder, Gregoire Montavon, Frederick Klauschen, Klaus-Robert Mller, ´ and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layerwise relevance propagation. PLOS ONE, 10(7):e0130140, July 2015. doi: 10.1371/journal.pone. 0130140. + +Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. ICLR, September 2015. + +Krysia B. Broda, Artur S. D’Avila Garcez, and Dov M. Gabbay. Neural-Symbolic Learning Systems. Springer London, 2002. ISBN 1852335122. + +Aditya Chattopadhay, Anirban Sarkar, Prantik Howlader, and Vineeth N. Balasubramanian. GradCAM $^ { + + }$ : Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 839–847. IEEE, IEEE, March 2018. doi: 10.1109/wacv.2018.00097. + +Sneha Chaudhari, Gungor Polatkan, Rohan Ramanath, and Varun Mithal. An attentive survey of attention models. IJCAI, April 2019. + +Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder–decoder approaches. In Proceedings of SSST-8, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation. Association for Computational Linguistics, 2014. doi: 10.3115/v1/w14-4012. + +Nuri Cingillioglu and Alessandra Russo. Deeplogic: Towards end-to-end differentiable logical reasoning. AAAI-MAKE, May 2019. + +Keith L. Clark. Negation as failure. In Logic and Data Bases, pp. 293–322. Springer US, 1978. doi: 10.1007/978-1-4684-3384-5 11. + +Richard Evans and Edward Grefenstette. Learning explanatory rules from noisy data. Journal of Artificial Intelligence Research, 61:1–64, January 2018. doi: 10.1613/jair.5714. + +Gottlob Frege. Sense and reference. The Philosophical Review, 57(3):209, May 1948. doi: 10.2307/ 2181485. + +Joe L. Frost, Pei-San Brown, John A. Sutterby, and Candra D. Thornton. The Developmental Benefits Of Playgrounds. Association for Childhood Education International, 2004. ISBN 0871731649. + +Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv:1410.5401, 2014. + +Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka GrabskaBarwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, ´ Adria Puigdom \` enech Badia, Karl Moritz Hermann, Yori Zwols, Georg Ostrovski, Adam Cain, \` Helen King, Christopher Summerfield, Phil Blunsom, Koray Kavukcuoglu, and Demis Hassabis. Hybrid computing using a neural network with dynamic external memory. Nature, 538(7626): 471–476, October 2016. doi: 10.1038/nature20101. + +Mikael Henaff, Jason Weston, Arthur Szlam, Antoine Bordes, and Yann LeCun. Tracking the world state with recurrent entity networks. ICLR, 2017. + +Paul W. Holland. Statistics and causal inference. Journal of the American Statistical Association, 81(396):945–960, December 1986. doi: 10.1080/01621459.1986.10478354. + +Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. ICLR, 2015. + +Ankit Kumar, Ozan Irsoy, Peter Ondruska, Mohit Iyyer, James Bradbury, Ishaan Gulrajani, Victor Zhong, Romain Paulus, and Richard Socher. Ask me anything: Dynamic memory networks for natural language processing. In ICML, pp. 1378–1387, 2016. + +Richard L. Lewis. Cognitive modeling, symbolic. The MIT encyclopedia of the cognitive sciences, pp. 525–527, 1999. + +Zachary C. Lipton. The mythos of model interpretability. Communications of the ACM, 61(10): 36–43, September 2018. doi: 10.1145/3233231. + +Fei Liu and Julien Perez. Gated end-to-end memory networks. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pp. 1–10. Association for Computational Linguistics, 2017. doi: 10.18653/v1/e17-1001. + +Thang Luong, Hieu Pham, and Christopher D. Manning. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 2015. doi: 10.18653/ v1/d15-1166. + +Tim Miller. Explanation in artificial intelligence: Insights from the social sciences. Artificial Intelligence, 267:1–38, February 2019. doi: 10.1016/j.artint.2018.07.007. + +Adam Morton and Jerry A. Fodor. The Language of Thought., volume 75. Philosophy Documentation Center, March 1978. doi: 10.2307/2025426. + +Stephen Muggleton and Luc de Raedt. Inductive logic programming: Theory and methods. The Journal of Logic Programming, 19-20:629–679, May 1994. doi: 10.1016/0743-1066(94) 90035-3. + +Judea Pearl. Probabilities of causation: three counterfactual interpretations and their identification. Synthese, 121(1-2):93–149, 1999. + +Judea Pearl. The seven tools of causal inference with reflections on machine learning. Communications of the ACM, 62(3):54–60, February 2019. doi: 10.1145/3241036. + +Jean Piaget. The Psychology of Intelligence. Routledge, August 2001. ISBN 0415254019. + +John C. Reynolds. Transformational systems and algebraic structure of atomic formulas. Machine intelligence, 5:135–151, 1970. + +Tim Rocktschel and Sebastian Riedel. End-to-end differentiable proving. NIPS, pp. 3791–3803, 2017. + +Neal J. Roese. Counterfactual thinking. Psychological Bulletin, 121(1):133–148, 1997. doi: 10. 1037/0033-2909.121.1.133. + +Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach $\langle 3 ^ { r d }$ Edition). Pearson, November 2016. ISBN 1292153962. + +John R. Searle. Minds, brains, and programs. Behavioral and brain sciences, 3(3):417–424, 1980. doi: 10.1016/b978-1-4832-1446-7.50007-8. + +Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pp. 618–626. IEEE, October 2017. doi: 10.1109/iccv.2017.74. + +Minjoon Seo, Sewon Min, Ali Farhadi, and Hannaneh Hajishirzi. Query-reduction networks for question answering. ICLR, 2017. + +Ehud Y. Shapiro. Inductive inference of theories from facts. Yale University, Department of Computer Science, 1981. + +Gustav Sourek, Vojtech Aschenbrenner, Filip Zelezny, and Ondrej Kuzelka. Lifted relational neural networks. arXiv:1508.05128, 2015. + +Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory networks. In NIPS, pp. 2440–2448, 2015. + +Seiya Tokui, Kenta Oono, Shohei Hido, and Justin Clayton. Chainer: a next-generation open source framework for deep learning. In Proceedings of Workshop on Machine Learning Systems (LearningSys) in The Twenty-ninth Annual Conference on Neural Information Processing Systems (NIPS), 2015. URL http://learningsys.org/papers/LearningSys_2015_ paper_33.pdf. + +J. Marshall Unger and Terrence W. Deacon. The Symbolic Species: The Co-Evolution of Language and the Brain, volume 82. Wiley, 1998. doi: 10.2307/329984. + +Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. ICLR, 2015. + +Jason Weston, Antoine Bordes, Sumit Chopra, Alexander M. Rush, Bart van Merrinboer, Armand Joulin, and Tomas Mikolov. Towards ai-complete question answering: A set of prerequisite toy tasks. ICLR, 2016. + +Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and textual question answering. ICML, pp. 2397–2406, 2016. + +A MODEL DETAILS + +# A.1 UNIFICATION MLP & CNN + +Unification MLP (UMLP) To model $f$ as a multi-layer perceptron, we take symbol embeddings of size $d = 1 6$ and flatten sequences of length $l = 4$ into an input vector of size $\phi ( \pmb { k } ) \in \mathbb { R } ^ { 6 4 }$ . The MLP consists of 2 hidden layers with tanh non-linearity of sizes $2 d$ and $d$ respectively. To process the query, we concatenate the one-hot encoding of the task id to $\phi ( k )$ yielding a final input of size $6 4 + 4 = 6 8$ . For unification features $\phi _ { U }$ , we use a bi-directional GRU with hidden size $d$ and an initial state of 0. The hidden state at each symbol is taken with a linear transformation to give $\phi _ { U } ( s ) = W _ { U } \Phi ( s )$ where $\Phi ( s )$ is the hidden state of the biGRU. The variable assignment is then computed as an attention over the according to equation 3. + +Unification CNN (UCNN) We take symbols embeddings of size $d = 3 2$ to obtain an input grid $\phi ( K ) \in \mathbb { R } ^ { 3 \times 3 \times 3 2 }$ . Similar to UMLP, for each symbol we append the task id as a one-hot vector to get an input of shape $3 \times 3 \times ( 3 2 + 4 )$ . Then $f$ consists of 2 convolutional layers with $d$ filters each, kernel size of 3 and stride 1. We use relu non-linearity in between the layers. We pad the grid with 2 columns and 2 rows to a $5 \times 5$ such that the output of the convolutions yield again a hidden output $\boldsymbol { \mathsf { H } } \in \mathbb { R } ^ { 3 \times 3 \times d }$ of the same shape. As the final hidden output $h$ , we take a global max pool to over $\boldsymbol { \mathsf { H } }$ to obtain $\pmb { h } \in \mathbb { R } ^ { d }$ . Unification function $g$ is modelled identical to $f$ without the max pooling such that $\phi _ { U } ( K _ { i j } ) = \mathsf { H } _ { i j } ^ { \prime }$ where $\boldsymbol { \mathsf { H } } ^ { \prime }$ is the hidden output of the convolutional layers. + +# A.2 UNIFICATION MEMORY NETWORKS + +Unlike previous architectures, with UMN we interleave $g$ into $f$ . We use embedding sizes of $d = 3 2$ and model $f$ with an iterative memory network. We take the final hidden state of a bi-directional GRU, with initial state 0, $\Phi _ { M }$ to represent the sentences of the context $C$ and query $\pmb q$ in a $d$ - dimensional vector $M _ { i } = \Phi _ { M } ( C _ { i } )$ and the query ${ \pmb m } _ { \ b q } = \Phi _ { M } ( { \pmb q } )$ . The initial state of the memory network is $h ^ { 0 } = m _ { q }$ . At each iteration $j$ : + +$$ +\begin{array} { c } { { A _ { i } ^ { j } = \operatorname { t a n h } ( W \rho ( M _ { i } , h ^ { j } ) + b ) } } \\ { { \beta ^ { j } = \operatorname { s o f t m a x } ( W \Phi _ { A } ( A ^ { j } ) + b ) } } \end{array} +$$ + +where $\Phi _ { A }$ is another $d$ -dimensional bi-directional GRU and $\rho ( { \pmb x } , { \pmb y } ) = [ { \pmb x } ; { \pmb y } ; { \pmb x } \odot { \pmb y } ; ( { \pmb x } - { \pmb y } ) ^ { 2 } ]$ with $\odot$ the element-wise multiplication and $[ ; ]$ the concatenation of vectors. Taking $\beta ^ { j }$ as the context attention, we obtain the next state of the memory network: + +$$ +h ^ { j + 1 } = \sum _ { i } \beta _ { i } ^ { j } \operatorname { t a n h } ( W \rho ( M _ { i } , h ^ { j } ) + b ) +$$ + +and iterate $J$ many times in advance. The final prediction becomes $f ( C , \boldsymbol { q } ) = \mathrm { s o f t m a x } ( W h ^ { J } + b )$ . +All weight matrices $W$ and bias vectors $^ { b }$ are independent but are tied across iterations. + +B GENERATED DATASET SAMPLES Table 4: Sample context, query and answer triples from sequences and grid tasks. +Table 5: Training sizes for randomly generated fixed length sequences and grid tasks with 8 unique symbols. The reason for Grid task (i) to be smaller is because there are at most 32 combinations of $2 \times 2$ boxes in a $3 \times 3$ grid with 8 unique symbols. + +
DatasetTaskContextQueryAnswer
Sequencei1488constant2
Sequenceii6157head6
Sequenceii1837tail7
Sequenceiv3563duplicate3
Gridi００８ 022 022box2
Gridii4 -070 0 页 0 8 1head4
Gridii0 673 0 1 2 0 0centre7
Gridiv8 0 0 2 60 41corner2
+ +
Task SequencesGrid
i704.7 ± 12.825.6 ± 1.8
ii709.4 ± 13.8623.7 ± 14.1
ii709.7 ± 14.0768.2 ±12.5
iv624.8 ± 12.4795.2 ± 10.3
+ +# C TRAINING DETAILS + +# C.1 UNIFICATION MLP & CNN + +Both unification models are trained on a 5-fold cross-validation over the generated datasets for 2000 iterations with a batch size of 64. We don’t use any weight decay and save the training and test accuracies every 10 iterations, as presented in Figure 3. + +# C.2 UNIFICATION MEMORY NETWORKS + +We again use a batch size of 64 and pre-train $f$ for 40 epochs then $f$ together with $g$ for 260 epochs. We use epochs for UMN since the dataset sizes are fixed. To learn $g$ alongside $f$ , we combine error signals from the unification of the invariant and the example. Following equation 4, the objective function not only incorporates the negative log-likelihood ${ \mathcal L } _ { \mathrm { n l l } }$ of the answer but also the mean squared error between intermediate states $h _ { I } ^ { j }$ and $h _ { K } ^ { j }$ at each iteration as an auxiliary loss: + +$$ +J = \mathcal { L } _ { \mathrm { n l l } } ( f ( K ) , a ) + \lambda _ { U } \left[ \mathcal { L } _ { \mathrm { n l l } } ( f \circ g ( I , K ) , a ) + \frac { 1 } { J } \sum _ { j = 1 } ^ { J } ( h _ { I } ^ { j } - h _ { K } ^ { j } ) ^ { 2 } + \tau \sum _ { s \in \mathcal { S } } \psi _ { w } ( s ) \right] +$$ + +We pre-train by setting $\lambda _ { U } = 0$ for 40 epochs and then set $\lambda _ { U } = 1$ . For strong supervision we also compute the negative log-likelihood ${ \mathcal { L } } _ { \mathrm { n l l } }$ for the context attention $\beta ^ { j }$ , described in Appendix A, at each iteration using the supporting facts of the tasks. We apply a dropout of 0.1 for all recurrent neural networks used and only for the bAbI dataset weight decay with 0.001 as the coefficient. + +# D FURTHER RESULTS + +![](images/c02ba1f113eff3023fe0e72f8db887cb964bac4e0057d50edcd72e8e1f55364f.jpg) +Figure 6: Results of Unification MLP and CNN on increasing number of invariants. There is no impact on performance when more invariants per task are given. Upon closer inspection, we noticed the models ignore the extra invariants and only use 1. We speculate the regularisation $\psi$ encourages the models to use a single 1 invariant. + +Table 6: Individual task error rates on bAbI tasks for Unification Memory Networks. + +
SupervisionWeakStrong
#Invs Training Size1 1k3 1k1 1k3 1k3 50
10.00.00.00.01.4
265.663.10.30.730.0
367.162.61.02.439.8
40.00.00.00.037.0
53.44.00.81.126.5
60.20.60.00.018.4
722.022.810.711.322.8
810.38.57.47.624.7
90.125.70.00.033.8
100.12.00.00.332.6
110.00.00.00.011.9
120.00.10.00.021.3
132.13.70.00.15.8
1419.713.50.50.154.8
150.00.70.00.00.0
1655.256.20.00.039.7
1739.249.051.149.348.8
184.48.00.60.510.4
1991.889.653.946.790.2
200.00.00.00.02.7
Mean19.120.56.36.027.6
Std27.927.015.614.321.0
#>5%8104417
+ +Table 7: Comparison of individual task error rates $( \% )$ on the bAbI (Weston et al., 2016) dataset of the best run. We preferred 1k results if a model had experiments published on both 1k and 10k for data efficiency. References from left to right: (Sukhbaatar et al., 2015) - (Liu & Perez, 2017) - (Henaff et al., 2017) - (Seo et al., 2017) - Ours - (Xiong et al., 2016) - (Graves et al., 2016) - (Weston et al., 2015) - Ours - (Kumar et al., 2016) + +
Support Size ModelWeakStrong
GN2N1k EntNetQRNUMN10k DMN+ DNC1k MemNN UMN10k DMN
10.00.00.70.00.00.00.00.00.00.0
28.38.156.40.565.60.30.40.00.71.8
340.338.869.71.267.11.11.80.02.44.8
42.80.41.40.70.00.00.00.00.00.0
513.11.04.61.23.40.50.82.01.10.7
67.68.430.01.20.20.00.00.00.00.0
717.317.822.39.422.02.40.615.011.33.1
810.012.519.23.710.30.00.39.07.63.5
913.210.731.50.00.10.00.20.00.00.0
1015.116.515.60.00.10.00.22.00.32.5
110.90.08.00.00.00.00.00.00.00.1
120.20.00.80.00.00.00.00.00.00.0
130.40.09.00.32.10.00.10.00.10.2
141.71.262.93.819.70.20.41.00.10.0
150.00.057.80.00.00.00.00.00.00.0
161.30.153.253.455.245.355.10.00.00.6
1751.041.746.451.839.24.212.035.049.340.6
1811.19.28.88.84.42.10.85.00.54.7
1982.888.590.490.791.80.03.964.046.765.5
200.00.02.60.30.00.00.00.00.00.0
Mean
#>5%13.9 1112.7 1029.6 1511.3 519.1 82.8 13.8 26.7 46.0 46.4 2
+ +Table 8: Individual task error rates $( \% )$ on the logical reasoning dataset. + +
Size1k10k5020k
SupportWeakStrongWeakWeak
Arity1111Strong 22
# Invs /Model12 312 312 312 132 3DMN2 IMA
Facts1.20.90.00.40.00.00.00.00.033.50.00.0
Unification0.010.30.010.80.00.00.00.00.041.313.010.0
1 Step50.349.84.420.01.227.80.11.35.750.226.02.0
2 Steps47.550.05.735.037.247.80.029.728.749.933.05.0
3 Steps47.649.210.438.739.645.60.026.026.148.323.06.0
AND31.337.410.716.429.829.00.20.41.250.020.05.0
OR25.238.121.035.020.530.24.420.617.447.613.03.0
Transitivity50.026.639.65.06.049.250.050.0
1 Step NBF46.438.73.828.81.121.60.11.18.047.621.02.0
2 Steps NBF48.548.97.739.630.448.20.133.428.750.315.04.0
AND NBF51.050.143.148.629.444.20.11.340.149.516.08.0
OR NBF51.448.450.847.347.647.821.327.630.547.325.014.0
Mean36.439.314.328.921.531.82.412.216.047.121.29.1
Std18.715.916.414.117.116.76.113.213.64.712.313.4
#>5%91171171015912115
+ +
Xsandra went back to the Y:bathroom
is X:sandra in the Y:bathroom
yes
+ +Figure 7: bAbI task 6, yes or no questions. The invariant does not variablise the answer. + +$$ +\begin{array} { r l } { X { \mathrm { : m ~ ( ~ Y : e ~ ) ~ } } } & { { } { \vdash X { \mathrm { : m ~ ( ~ Y : e ~ ) ~ } } } } \\ { X { \mathrm { : a ~ ( ~ } Y { \mathrm { : w ~ , ~ } } { \mathrm { ~ Z : e ~ ) ~ } } } } & { { } { \vdash X { \mathrm { : a ~ ( ~ } Y { \mathrm { : w ~ , ~ } } { \mathrm { Z : e ~ ) ~ } } } } } \\ { X { \mathrm { : m ~ ( ~ T ~ ) ~ } } } & { { } { \vdash X { \mathrm { : m ~ ( ~ c ~ ) ~ } } } } \\ { X { \mathrm { : x ~ ( ~ \mathbb { A } ~ ) n o t ~ \mathbf { Z } { \mathrm { : q } } ~ ( ~ \mathbb { A } ~ ) ~ } } } & { { } { \vdash X { \mathrm { : x ~ ( ~ } Y { \mathrm { : z } } ~ ) } } } \end{array} +$$ + +Figure 8: Invariants learned on tasks 1, 2 and 11 with arity 1 and 2 from the logical reasoning dataset. Last invariant on task 11 lifts the example around the negation by failure, denoted as not, capturing its semantics. + +3 X 7 Y head 7 4 X Y head X 3 1 X tail X X 5 Y duplicate + +$$ +\begin{array} { r l } { X \mathrm { ~ 0 ~ ~ 0 ~ ~ } } & { \mathrm { ~ 0 ~ ~ Y ~ 0 ~ ~ 6 ~ 4 ~ } Y } \\ { \mathrm { ~ 0 ~ 1 ~ 0 ~ } } & { \mathrm { ~ 1 ~ } X \mathrm { ~ 8 ~ } } \\ { \mathrm { ~ 0 ~ 0 ~ } Y } & { \mathrm { ~ 0 ~ 2 ~ 0 ~ } } \end{array} +$$ + +(a) Invariants with extra varialbes learned with UMLP. (b) Mismatching invariants learned with UCNN. + +Figure 9: Invariants learned that do not match the data generating distribution from UMLP and UCNN using $\leq 1 0 0 0$ examples to train. In these instances the unification still bind to the the correct symbols in order to predict the desired answer; quantitatively we get the same results. Variable default symbols are omitted for clarity. + +(a) bAbI task 2. When a variable is unused in the next iteration, e.g. Z:football, it unifies with random tokens often biased by position. + +![](images/9073cfbc8cdd1d7408ab2bf49ed3c03468c3850a67e1c59f65e66ccdfccc89e1.jpg) +(c) Logical reasoning task 3. Arity 1 atom forced to bind with arity 2 creates a one-to-many mapping. + +(b) Logical reasoning task 1. A one-to-one alignment is created between predicates and constants. + +Figure 10: Further attention maps for equation 3, darker cells indicate higher attention values. \ No newline at end of file diff --git a/parse/train/r1xwA34KDB/r1xwA34KDB_content_list.json b/parse/train/r1xwA34KDB/r1xwA34KDB_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c441b2f427e1bac53fbc5669c374eea28991d9e1 --- /dev/null +++ b/parse/train/r1xwA34KDB/r1xwA34KDB_content_list.json @@ -0,0 +1,1838 @@ +[ + { + "type": "text", + "text": "LEARNING INVARIANTS THROUGH SOFT UNIFICATION ", + "text_level": 1, + "bbox": [ + 173, + 98, + 820, + 121 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 145, + 398, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 210, + 544, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human reasoning involves recognising common underlying principles across many examples by utilising variables. The by-products of such reasoning are invariants that capture patterns across examples such as “if someone went somewhere then they are there” without mentioning specific people or places. Humans learn what variables are and how to use them at a young age, and the question this paper addresses is whether machines can also learn and use variables solely from examples without requiring human pre-engineering. We propose Unification Networks that incorporate soft unification into neural networks to learn variables and by doing so lift examples into invariants that can then be used to solve a given task. We evaluate our approach on four datasets to demonstrate that learning invariants captures patterns in the data and can improve performance over baselines. ", + "bbox": [ + 233, + 244, + 764, + 397 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 431, + 336, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans have the ability to process symbolic knowledge and maintain symbolic thought (Unger & Deacon, 1998). When reasoning, humans do not require combinatorial enumeration of examples but instead utilise invariant patterns with placeholders replacing specific entities. Symbolic cognitive models (Lewis, 1999) embrace this perspective with the human mind seen as an information processing system operating on formal symbols such as reading a stream of tokens in natural language. The language of thought hypothesis (Morton & Fodor, 1978) frames human thought as a structural construct with varying sub-components such as $^ { 6 6 } \\mathrm { X }$ went to Y”. By recognising what varies across examples, humans are capable of lifting examples into invariant principles that account for other instances. This symbolic thought with variables is learned at a young age through symbolic play (Piaget, 2001). For instance a child learns that a sword can be substituted with a stick (Frost et al., 2004) and engage in pretend play. ", + "bbox": [ + 174, + 465, + 825, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although variables are inherent in models of computation and symbolic formalisms, as in first-order logic (Russell & Norvig, 2016), they are pre-engineered and used to solve specific tasks by means of unification or assignments that bound variables to given values. However, when learning from data only, being able to recognise when and which symbols should take on different values, i.e. symbols that can act as variables, is crucial for lifting examples into general principles that are invariant across multiple instances. Figure 1 shows the invariant learned by our approach: if someone is the same thing as someone else then they have the same colour. With this invariant, our approach can solve all of the training and test examples in task 16 of the bAbI dataset (Weston et al., 2016). ", + "bbox": [ + 176, + 625, + 581, + 804 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/005e03f738687ce974ff771c18081cb699f106f6d11b9f8f18552de03cc0852f.jpg", + "table_caption": [ + "Figure 1: Invariant learned for bAbI task 16, basic induction, where $X$ :bernhard denotes a variable with default symbol bernhard. This single invariant accounts for all the training and test examples. " + ], + "table_footnote": [], + "table_body": "
X:bernhard is a Y:frog Z:lily is a Y:frog
Z:lily is A:green
what colour is X:bernhard
A:green
", + "bbox": [ + 611, + 633, + 803, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper we address the question of whether a machine can learn and use the notion of a variable, i.e. a symbol that can take on different values. For instance, given an example of the form “bernhard is a frog” the machine would learn that the token “bernhard” could be someone else and the token “frog” could be something else. If we consider unification a selection of the most appropriate value for a variable given a choice of values, we can reframe it as a form of attention. Attention models (Bahdanau et al., 2015; Luong et al., 2015; Chaudhari et al., 2019) allow neural networks to focus, attend to certain parts of the input often for the purpose of selecting a relevant portion. Since attention mechanisms are also differentiable they are often jointly learned within a task. This perspective motivates our idea of a unification mechanism that utilises attention and is therefore fully differentiable which we refer to as soft unification. ", + "bbox": [ + 176, + 813, + 581, + 827 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 824, + 825, + 924 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hence, we propose an end-to-end differentiable neural network approach for learning and utilising the notion of a variable that in return can lift examples into invariants used by the network to perform reasoning tasks. Specifically, we (i) propose a novel architecture capable of learning and using variables by lifting a given example through soft unification, (ii) present the empirical results of our approach on four datasets and (iii) analyse the learned invariants that capture the underlying patterns present in the tasks. Our implementation using Chainer (Tokui et al., 2015) is publicly available at [link removed](anonymous link provided with submission). ", + "bbox": [ + 174, + 138, + 825, + 237 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 SOFT UNIFICATION ", + "text_level": 1, + "bbox": [ + 176, + 256, + 367, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reasoning with variables involves identifying what variables are, the setting in which they are used as well as the process by which they are assigned values. When the varying components, i.e. variables, of an example are identified, the remaining structure can be lifted into an invariant which then accounts for multiple other instances. ", + "bbox": [ + 174, + 287, + 825, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Definition 1 (Variable). Given a set of symbols $\\mathbb { S }$ , a variable $X$ is defined as a pair $\\pmb { X } \\triangleq \\left( \\mathbf { x } , s _ { d } \\right)$ where $s _ { d } \\in \\mathbb { S }$ is the default symbol of the variable and $\\mathbf { X }$ is a discrete random variable of which the support is $\\mathbb { S }$ . The representation of a variable $\\phi _ { V } ( { \\pmb X } ; s _ { d } )$ is equal to the expected value of the corresponding random variable $\\mathbf { X }$ given the default symbol $s _ { d }$ : ", + "bbox": [ + 173, + 348, + 825, + 405 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/e6f00329572399e3f3f44e676dabe4812b9f6a7696b28c59eebe9ea3a3fc9948.jpg", + "text": "$$\n\\phi _ { V } ( X { : } s _ { d } ) = \\mathbb { E } _ { \\mathbf { x } \\sim P } [ \\phi ( x ) ] = \\sum _ { x \\in \\mathbb { S } } P ( \\mathbf { x } = x | s _ { d } ) \\phi ( x )\n$$", + "text_format": "latex", + "bbox": [ + 331, + 410, + 665, + 444 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\phi : \\mathbb { S } \\mathbb { R } ^ { d }$ is a $d$ -dimensional real-valued feature of a symbol $s$ ", + "bbox": [ + 176, + 449, + 643, + 464 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For example, $\\phi$ could be an embedding and $\\phi _ { V } ( { \\pmb X } ; s _ { d } )$ would become a weighted sum of symbol embeddings as in conventional attention models. The default symbol of a variable is intended to capture the variable’s bound meaning following the idea of referants by Frege (1948). We denote variables using $X , Y , A$ etc. such as $X$ :bernhard where $X$ is the name of the variable and bernhard the default symbol as shown in Figure 1. ", + "bbox": [ + 174, + 474, + 825, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Definition 2 (Invariant). Given a structure (e.g. list, grid) $\\mathcal { G }$ over $\\mathbb { S }$ , an invariant is a pair $I \\triangleq ( G , \\psi )$ where $G \\in { \\mathcal { G } }$ is the invariant example such as a tokenised story with tokens as symbols and $\\psi : \\mathbb { S } \\to \\quad$ $[ 0 , 1 ]$ is a function representing the degree to which the symbol is considered a variable. Thus, the final representation of a symbol $s$ included in $G$ , $\\phi _ { I } ( s )$ is: ", + "bbox": [ + 173, + 550, + 825, + 607 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/2057226d07d21614572dc84a4398bd6f444905435e014f404eb0581291ecae49.jpg", + "text": "$$\n\\phi _ { I } ( s ) = ( 1 - \\psi ( s ) ) \\phi ( s ) + \\psi ( s ) \\phi _ { V } ( X { : } s )\n$$", + "text_format": "latex", + "bbox": [ + 361, + 611, + 635, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the linear interpolation between its representation $\\phi ( s )$ and its variable bound value with itself as the default symbol $\\phi _ { V } ( \\pmb { X } ; s )$ . ", + "bbox": [ + 173, + 632, + 821, + 661 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We adhere to the term invariant and refrain from mentioning rules, unground rules, etc. used in logicbased formalisms, e.g. Muggleton & de Raedt (1994), since neither the invariant structure needs to be rule-like nor the variables carry logical semantics. This distinction is clarified in Section 6. ", + "bbox": [ + 174, + 671, + 825, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Definition 3 (Unification). Given an invariant $I$ and an example $K \\in \\mathcal G$ , unification binds the variables in $I$ to symbols in $K$ . Defined as a function $g : \\mathcal { T } \\times \\mathcal { G } \\to \\mathcal { G }$ , unification binds variables by computing the probability mass functions, $P$ in equation 1, and returns the unified representation using equation 2. The probability mass function of a variable $X { : } s _ { d }$ is: ", + "bbox": [ + 173, + 718, + 825, + 775 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/3084e894930ef9d610b90cd7e1f867b73f6c7f52c81c81e58d5660b31df3b2ba.jpg", + "text": "$$\nP ( { \\bf x } = x | s _ { d } ) = \\mathrm { s o f t m a x } ( \\phi _ { U } ( s _ { d } ) \\phi _ { U } ( K ) ^ { T } ) , \\ : K = \\mathrm { s u p p o r t \\ o f \\ x }\n$$", + "text_format": "latex", + "bbox": [ + 292, + 779, + 705, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\phi _ { U } : \\mathbb { S } \\to \\mathbb { R } ^ { d }$ is the unifying feature of a symbol and $\\phi _ { U } ( K ) \\in \\mathbb { R } ^ { | K | \\times d }$ is applied element wise to symbols in $K$ . If $g$ is differentiable, it is referred to as soft unification. ", + "bbox": [ + 173, + 809, + 823, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We distinguish $\\phi$ from $\\phi _ { U }$ to emphasise that the unifying properties of the symbols might be different from their representations. For example, $\\phi$ (bernhard) could represent a specific person whereas $\\phi _ { U }$ (bernhard) the notion of someone. ", + "bbox": [ + 174, + 845, + 825, + 888 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall soft unification incorporates 3 learnable components: $\\phi , \\psi , \\phi _ { U }$ which denote the base features, variableness and unifying features of a symbol respectively. Given an upstream, potentially ", + "bbox": [ + 174, + 895, + 821, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "task specific, network $f : \\mathcal G \\to \\mathbb S$ , an invariant $I \\in \\mathcal { Z }$ and an input example $K \\in { \\mathcal { G } }$ with a corresponding desired output $a \\in \\mathbb { S }$ , the following holds: ", + "bbox": [ + 171, + 103, + 821, + 132 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/990f4d6f7bb15d522906b6e6ae32537c039f64d1b9e23568ff08005eaf16924c.jpg", + "text": "$$\nf \\circ g ( I , K ) = f ( K ) = a\n$$", + "text_format": "latex", + "bbox": [ + 413, + 138, + 584, + 156 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f$ now predicts based on the unified representation produced by $g$ . In this work, we focus on $g$ , the invariants it produces together with the interaction of $f \\circ g$ . ", + "bbox": [ + 171, + 162, + 823, + 193 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 UNIFICATION NETWORKS ", + "text_level": 1, + "bbox": [ + 176, + 212, + 419, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since soft unification is end-to-end differentiable, it can be incorporated into existing task-specific upstream architectures. We present 3 architectures that model $f \\circ g$ using multi-layer perceptrons (MLP), convolutional neural networks (CNN) and memory networks (Weston et al., 2015) to demonstrate the flexibility of our approach. In all cases, the $d$ dimensional representation of symbols are learnable embeddings $\\phi ( s ) = O [ s ] ^ { T } E$ with $E \\in \\mathbb { R } ^ { | \\mathbb { S } | \\times d }$ randomly initialised by $\\mathcal { N } ( 0 , 1 )$ and $O [ s ]$ the one-hot encoding of the symbol. The variableness of symbols is a learnable weight $\\psi _ { w } ( s ) = \\sigma ( w _ { s } )$ where $\\boldsymbol { w } \\in \\mathbb { R } ^ { | \\mathbb { S } | }$ and $\\sigma$ is the sigmoid function. We consider every symbol independently a variable irrespective of its surrounding context and leave further contextualised formulations as future work. The underlying intuition of this configuration is that a useful symbol for a correct prediction might need to take on other values for different inputs. This usefulness can be viewed as the inbound gradient to the corresponding $w _ { s }$ parameter and $\\psi _ { w } ( s )$ acting as a gate. For further model details including the size of the embeddings, please refer to Appendix A. ", + "bbox": [ + 173, + 243, + 825, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unification MLP (UMLP) ( $f$ : MLP, $g$ : RNN) We combine soft unification into a multi-layer perceptron to process fixed length inputs. In this case, the structure $\\mathcal { G }$ is a sequence of symbols with a fixed length $l$ , e.g. a sequence of digits 4234. Given an embedded input $\\phi ( \\pmb { k } ) \\in R ^ { l \\times d }$ , the upstream MLP computes the output symbol based on the flattened representations $f ( \\phi ( \\pmb { k } ) ) =$ softmax $( h E ^ { T } )$ where $\\hat { h } \\in \\mathbb { R } ^ { d }$ is the output of the last layer. However, to compute the unifying features $\\phi _ { U }$ , definition 3, $g$ uses a bi-directional GRU (Cho et al., 2014) running over $\\phi ( k )$ such that $\\phi _ { U } ( k ) = W _ { U } \\Phi ( k )$ where $\\Phi ( k ) \\in \\mathbb { R } ^ { d }$ is the hidden state of the GRU at symbol $k$ and $\\dot { W _ { U } } \\in \\mathbb { R } ^ { d \\times d }$ is a learnable parameter. This model emphasises the flexibility around the boundary of $f \\circ g$ and that the unifying features can be computed in any differentiable manner. ", + "bbox": [ + 173, + 420, + 825, + 546 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ef39205e27e246bc37f3a17ac72fc093d4914d8003b2bd66d967fd9856bb6ce9.jpg", + "image_caption": [ + "Figure 2: Graphical overview of soft unification within a memory network. Each sentence is processed by two bi-directional RNNs for memory and unification. At each iteration the context attention selects which sentences to unify and the invariant produces the same answer as the example. " + ], + "image_footnote": [], + "bbox": [ + 222, + 555, + 776, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unification CNN (UCNN) ( $f$ : CNN, $g$ : CNN) Given a grid of embedded symbols $\\phi ( K ) \\ \\in$ $\\mathbb { R } ^ { w \\times h \\times d }$ where $w$ is the width and $h$ the height, we use a convolutional neural network such that the final prediction is $f ( \\phi ( K ) ) = \\mathrm { s o f t m a x } ( ( W h + b ) E ^ { T } )$ where $^ { h }$ this time is the result of global max pooling and $W , b$ are learnable parameters. We also model $g$ using a separate convolutional network with the same architecture as $f$ and set $\\phi _ { U } ( k ) = c _ { 2 } ( \\mathrm { r e l u } ( c _ { 1 } ( k ) )$ where $c _ { 1 } , c _ { 2 }$ are the convolutional layers. The grid is padded with 0s to obtain $w \\times h \\times d$ after each convolution such that every symbol has a unifying feature. This model conveys how soft unification can be adapted to the specifics of the domain for example by using a convolution in a spatially structured input. ", + "bbox": [ + 173, + 748, + 825, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unification Memory Networks (UMN) ( $f$ : MemNN, $g$ : RNN) Soft unification does not need to happen prior to $f$ in a $f \\circ g$ fashion but can also be incorporated at any intermediate stage multiple times. To demonstrate this ability, we unify the symbols at different memory locations at each iteration of a Memory Network (Weston et al., 2015). Memory networks can handle a list of lists structure such as a tokenised story as shown in Figure 2. The memory network $f$ uses the final hidden state of a bi-directional GRU (outer squares in Figure 2) as the sentence representations to compute a context attention. At each iteration, we unify the words between the attended sentences using the same approach in UMLP with another bi-directional GRU (inner diamonds in Figure 2) for unifying features $\\phi _ { U } ( \\mathbf { b e r n h a r d } ) = W _ { U } \\Phi ($ (bernhard). Following equation 2, the new unified representation of the memory slot is computed and $f$ uses it to perform the next iteration. Concretely, $g$ produces an unification tensor $\\pmb { \\mathsf { U } } \\in \\dot { \\mathbb { R } } ^ { M \\times m \\times N \\times d }$ where $M$ and $m$ is the number of sentences and words in the invariant respectively, and $N$ is the number of sentences in the example such that after the context attentions are applied over $M$ and $N$ , we obtain $\\boldsymbol { \\phi } ( \\pmb { k } ) \\in \\mathbb { R } ^ { m \\times d }$ as the unified sentence at that iteration. Note that unlike in the UMLP case, the sentences can be of varying length. The prediction is then softmax $( W h _ { I } ^ { J } + b )$ where $h _ { I } ^ { J }$ is the hidden state of the invariant after $J$ iterations. This setup, however, requires pre-training $f$ such that the context attentions match the correct sentences. ", + "bbox": [ + 174, + 867, + 825, + 924 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/e907e40a0c5d0e4e0f752c42d1e303a4907a55461d9f328e2ce4b4abbfbb4726.jpg", + "table_caption": [ + "Table 1: Sample context, query and answer triples and their training sizes per task. For distribution of generated number of examples per task on Sequence and Grid data refer to Appendix B. " + ], + "table_footnote": [], + "table_body": "
DatasetContextQueryAnswerTraining Size
Sequence8384duplicate8≤1k,≤50
Grid00 3 1 859corner7≤1k,≤50
bAbIMary went to the kitchen. Sandra journeyed to the garden.Where is Mary?kitchen1k,50
Logicp(X)← q(X). q(a).p(a).True1k, 10k,50
", + "bbox": [ + 199, + 140, + 799, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 297, + 825, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A task might contain different questions such as “Where is X?” and “Why did $\\mathrm { X }$ go to Y?”. To let the models differentiate between questions and potentially learn different invariants, we extend them with a repository of invariants $I \\in \\mathbb { I }$ and aggregate the predictions from each invariant. One simple approach is to sum the predictions of the invariants $\\textstyle \\sum _ { I \\in \\mathbb { I } } f \\circ g ( I , K )$ used in UMLP and UCNN. Another approach could be to use features from the invariants such as memory representations in the case of UMN. For UMN, we weigh the predictions using a bilinear attention $\\eta$ based on the hidden states at the first iteration $h _ { I } ^ { 0 }$ and $h _ { K } ^ { 0 }$ such that $\\eta = \\mathrm { s o f t m a x } ( h _ { I } ^ { 0 } W h _ { K } ^ { 0 ^ { T } } )$ . To initially form the repository of invariants, we use the bag-of-words representation of the questions and find the most dissimilar ones based on their cosine similarity as a heuristic to obtain varied examples. ", + "bbox": [ + 173, + 470, + 825, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 DATASETS ", + "text_level": 1, + "bbox": [ + 174, + 619, + 292, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use 4 datasets consisting of context, query and an answer $( C , \\pmb { q } , a )$ : fixed length sequences of symbols, shapes of symbols in a grid, story based natural language reasoning with the bAbI (Weston et al., 2016) dataset and logical reasoning represented as logic programs, examples shown in Table 1 with further samples in Appendix B. In each case we use an appropriate model: UMLP for fixed length sequences, UCNN for grid and UMN for iterative reasoning. We use synthetic datasets of which the data generating distributions are known to evaluate not only the quantitative performance but also the quality of the invariants learned by our approach. ", + "bbox": [ + 174, + 651, + 825, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fixed Length Sequences We generate sequences of length $l = 4$ with 8 unique symbols represented as digits to predict (i) a constant, (ii) the head of the sequence, (iii) the tail and (iv) the duplicate symbol. We randomly generate 1000 triples and then only take the unique ones to ensure the test split contains unseen examples. The training is then performed over a 5-fold cross-validation. ", + "bbox": [ + 174, + 756, + 823, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Grid To spatially organise symbols, we generate a grid of size $3 \\times 3$ with 8 unique symbols organised into $2 \\times 2$ box of identical symbol, a vertical, diagonal or horizontal sequence of length 3, a cross or a plus shape and a triangle. In each task we predict (i) the identical symbol, (ii) the head of the sequence, (iii) the centre of the cross or plus and (iv) the corner of the triangle respectively. We follow the same procedure from sequences and randomly generate 1000 discarding duplicate triples. ", + "bbox": [ + 174, + 819, + 823, + 888 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "bAbI The bAbI dataset has become a standard benchmark for evaluating memory based networks. \nIt consists of 20 synthetically generated natural language reasoning tasks (refer to Weston et al. ", + "bbox": [ + 173, + 895, + 823, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(2016) for task details). We take the 1k English set and use 0.1 of the training set as validation. Each token is lower cased and considered a unique symbol. Following previous works (Seo et al., 2017; Sukhbaatar et al., 2015), we take multiple word answers also to be a unique symbol in S. ", + "bbox": [ + 174, + 103, + 825, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Logical Reasoning To demonstrate the flexibility of our approach and distinguish our notion of a variable from that used in logic based formalisms, we generate logical reasoning tasks in the form of logic programs using the procedure by Cingillioglu & Russo (2019). The tasks involve learning $f ( \\bar { C } , \\bar { Q } ) \\stackrel { - } { = } \\mathrm { T r u e } \\bar { C } \\vdash \\bar { Q }$ over 12 classes of logic programs exhibiting varying paradigms of logical reasoning including negation by failure (Clark, 1978). We generate 1k and 10k logic programs per task for training with 0.1 as validation and another 1k for testing. We set the arity of literals to 1 or 2 using one random character from the English alphabet for predicates and constants, e.g. $p ( p )$ and an upper case character for logical variables, e.g. $p ( \\mathrm { x } )$ . ", + "bbox": [ + 173, + 152, + 825, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 285, + 326, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We probe three aspects of soft unification: the impact of unification on performance over unseen data, the effect of multiple invariants and data efficiency. To that end, we train UMLP and UCNN with and without unification, UMN with pre-training using 1 or 3 invariants over either the entire training set or only 50 examples. Every model is trained 3 times via back-propagation using Adam (Kingma & Ba, 2015) on an Intel Core i7-6700 CPU using the following objective function: ", + "bbox": [ + 174, + 315, + 825, + 386 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/562850468dbca51e031b1d308f05cc93e1a537ef85b27f0952beccb216be018d.jpg", + "text": "$$\nJ = \\lambda _ { K } \\mathcal { L } _ { \\mathrm { n l l } } ( f ( K ) , a ) + \\lambda _ { U } \\left[ \\mathcal { L } _ { \\mathrm { n l l } } ( f \\circ g ( I , K ) , a ) + \\tau \\sum _ { s \\in \\mathbb { S } } \\psi _ { w } ( s ) \\right]\n$$", + "text_format": "latex", + "bbox": [ + 279, + 388, + 718, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where ${ \\mathcal { L } } _ { \\mathrm { n l l } }$ is the negative log-likelihood with sparsity regularisation over $\\psi$ at $\\tau = 0 . 1$ to discourage the models from utilising spurious number of variables. For UMLP and UCNN, we set $\\lambda _ { K } ~ =$ 0, $\\lambda _ { U } = 1$ for training just the unified output and the converse for the non-unifying versions. To pre-train the UMN, we start with $\\lambda _ { K } = 1 , \\lambda _ { U } = 0$ for 40 epochs then set $\\lambda _ { U } = 1$ to jointly train the unified output. For iterative tasks, the mean squared error between hidden states $( { h _ { I } ^ { j } - h _ { K } ^ { j } } ) ^ { 2 }$ at each iteration $j$ and, in the strongly supervised cases, the negative log-likelihood for the context attentions using provided supporting facts are also added to the objective function. Further details such as batch size and total number of epochs are available in Appendix C. ", + "bbox": [ + 173, + 436, + 825, + 551 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4aca829873f131070cc20ca4370a0d6ac4a79c32120a7d5e548cfc6ffbcee87f.jpg", + "image_caption": [ + "Figure 3: Test accuracy over iterations for Unification MLP and Unification CNN models with 1 invariant versus no unification. We observe that with soft unification the models achieve higher accuracy with fewer iterations than their plain counterparts on both per task training sizes. " + ], + "image_footnote": [], + "bbox": [ + 178, + 558, + 818, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 3 portrays how soft unification generalises better to unseen examples in test sets - the same sequence or grid never appears in both the training and test sets as outlined in Section 4 - over plain models. Despite $f \\circ g$ having more trainable parameters than $f$ alone, the models with unification not only maintain higher accuracy in each iteration and solve the tasks in as few as 250 iterations with $\\leq 1 0 0 0$ training examples but also improve accuracy by $\\approx 0 . 3$ when trained with only $\\leq 5 0$ per task. We believe soft unification architecturally biases the models towards learning structural patterns which in return achieves better results on recognising common patterns of symbols across examples. Results with multiple invariants are identical and the models seem to ignore the extra invariants due to the fact that the tasks can be solved with a single invariant and the regularisation applied on $\\psi$ zeroing out unnecessary invariants; further results in Appendix D. The fluctuations in accuracy around iterations 750 to 1000 in UCNN are also caused by penalising $\\psi$ which forces the model to relearn the task with less variables half way through training. ", + "bbox": [ + 173, + 757, + 825, + 924 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/768570d0a06568bcbbcbc3b99176f6ae5e403c8c7a9bb3cef63a6be9c26ee2f2.jpg", + "table_caption": [ + "Table 2: Aggregate error rates $( \\% )$ on bAbI 1k for UMN and N2N, GN2N, MemNN by Sukhbaatar et al. (2015), Liu & Perez (2017) and Weston et al. (2015) respectively. Full comparison on individual tasks are available in Appendix D. " + ], + "table_footnote": [], + "table_body": "
Training Size Supervision #Invs /Model1k501k
Weak 13Strong 3WeakQRNStrong MemNN
31N2N GN2NEntNet
Mean19.120.56.36.027.613.912.729.611.36.7
#>5%810441711101554
", + "bbox": [ + 202, + 154, + 795, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following Tables 2 and 3, we observe a trend of better performance through strong supervision, more data per task and using only 1 invariant. We believe strong supervision aids with selecting the correct sentences to unify and in a weak setting the model attempts to unify arbitrary context sentences often failing to follow the iterative reasoning chain. The increase in performance with more data and strong supervision is consistent with previous work reflecting how $f \\circ g$ can be bounded by the efficacy of $f$ modelled as a memory network. As a result, only in the supervised case do we observe a minor improvement over MemNN by 0.7 in Table 2 and no improvement in the weak case over comparable memory based networks in Table 3. This dependency on $f$ also limits the ability of $f \\circ g$ to learn from 50 examples per task failing 17/20 and 12/12 of bAbI and logical reasoning tasks respectively. The increase in error rate with 3 invariants, we speculate, stems from having more parameters and more pathways in the model rendering training more difficult. ", + "bbox": [ + 173, + 258, + 825, + 411 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1e018f6c32d109a7fac63e2f804566d081fd024bb3fdbeb94c09620e36bd8f82.jpg", + "table_caption": [ + "Table 3: Aggregate task error rates $( \\% )$ on the logical reasoning dataset for UMN and DMN, IMA by Cingillioglu & Russo (2019). Strong supervision, more data and only 1 invariant seem to improve the performance of UMN over plain iterative models. Individual task results are in Appendix D. " + ], + "table_footnote": [], + "table_body": "
Training Size Supervision1k5020k Weak
WeakStrongWeakStrong
121212122222
# Invs /Model1313131133DMNIMA
Mean36.439.314.328.921.531.82.412.216.047.121.29.1
#>5%91171171015912115
", + "bbox": [ + 179, + 469, + 818, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6 ANALYSIS ", + "text_level": 1, + "bbox": [ + 174, + 580, + 290, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After training, we can extract the learned invariants by applying a threshold on $\\forall s \\in \\mathbb { S } : \\psi ( s ) > t$ indicating whether a symbol is used as a variable or not. We set $t = 0 . 0$ for all datasets except for bAbI, we use $t = 0 . 1$ . The magnitude of this threshold seems to depend on the amount of regularisation $\\tau$ , equation 5, and the number of training steps along with batch size all controlling how much $\\psi$ is pushed towards 0. Sample invariants shown in Figure 4 describe the common patterns present in the tasks with parts that contribute towards the final answer becoming variables. Extra symbols such as is or travelled do not emerge as variables, as shown in Figure 4a; we attribute this behaviour to the fact that changing the token travelled to went does not influence the prediction but changing the action, the value of Z:left to ‘picked’ does. However, based on random initialisation, our approach can convert an arbitrary symbol into a variable and let $f$ compensate for the unifications it produces. For example, the invariant $^ { * } X { : } 8 \\ 5 \\ 2 \\ 2 ^ { , }$ could predict the tail of another example by unifying the head with the tail using $\\phi _ { U }$ , equation 3, of those symbols. Further examples are shown in Appendix D. Pre-training $f$ as done in UMN seems to produce more robust and consistent invariants compared to immediately training $f \\circ g$ since, we speculate, by equation 4 $f$ might encourage $g ( I , K ) \\approx K$ . ", + "bbox": [ + 173, + 611, + 825, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Interpretability versus Ability A desired property of interpretable models is transparency (Lipton, 2018). A novel outcome of the learned invariants in our approach is that they provide an approximation of the underlying general principle present in the data such as the structure of multi-hop reasoning shown in Figure 4e. However, certain aspects regarding the ability of the model such as how it performs temporal reasoning, are still hidden inside $f$ . In Figure 4b, although we observe $\\mathbf { Z }$ :morning as a variable, the overall learned invariant captures nothing about how changing the value of $\\mathbf { Z }$ :morning alters the behaviour of $f$ . The model might look before or after a certain time point ", + "bbox": [ + 173, + 825, + 825, + 924 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7747279f4df11d55eb0355fa48db5cdb6952a0dfb81cec0c6489109d22418593.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Y:john Z:left the X:football Y:john travelled to the A:office
where is the X:football
A:office
", + "bbox": [ + 220, + 102, + 446, + 169 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "this Z:morning $X$ :bill went to the Y:school yesterday $X$ :bill journeyed to the A:park where was $X$ :bill before the Y:school A:park ", + "bbox": [ + 511, + 102, + 810, + 169 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(a) bAbI task 2, two supporting facts. The model also learns Z:left since people can also drop or pick up objects potentially affecting the answer. ", + "bbox": [ + 173, + 176, + 493, + 215 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(b) bAbI task 14, time reasoning. $X$ :bill and Y:school are recognised as variables alongside Z:morning capturing when someone went which is crucial to this task. ", + "bbox": [ + 503, + 176, + 823, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2b918647c3c931a6d2d86be39b48a163abe2920f4808362f723da0efbea11058.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 173, + 238, + 818, + 343 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4: Invariants learned across the four datasets using the three architectures. For iterative reasoning datasets, bAbI and logical reasoning, they are taken from strongly supervised UMN. ", + "bbox": [ + 174, + 353, + 823, + 382 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$X$ :bill went somewhere depending what $\\mathbf { Z }$ :morning binds to. Without the regularising term on $\\psi ( s )$ , we initially noticed the models using, one might call extra, symbols as variables and binding them to the same value occasionally producing unifications such as “bathroom bathroom to the bathroom” and still $f$ predicting, perhaps unsurprisingly, the correct answer as bathroom. Hence, regularising $\\psi$ with the correct amount $\\tau$ in equation 5 to reduce the capacity of unification seems critical in extracting not just any invariant but one that represents the common structure. ", + "bbox": [ + 174, + 410, + 825, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Soft unification from equation 3 reveals three main patterns: one-to-one, one-to-many or many-toone bindings as shown in Figure 5; further examples are in Appendix D. Figure 5a captures what one might expect unification to look like where variables unify with their corresponding counterparts, e.g. $X$ :bernhard with brian and Y:frog with lion. However, occasionally the model can optimise to use less variables and squeeze the required information into a single variable, for example by binding Y:bathroom to john and kitchen as shown in Figure 5b. We believe this occurs due to the sparsity constraint on $\\psi ( s )$ encouraging the model to be as conservative as possible. Finally, if there are more variables than needed as in Figure 5c, we observe a many-to-one binding with Y:w and Z:e mapping to the same constant $q$ . This behaviour begs the question how does the model differentiate between $p ( q )$ and $p ( q , q )$ . We speculate the model uses the magnitude of $\\psi ( w ) = 0 . 0 3 7$ and $\\psi ( e ) = 0 . 0 4 2$ to encode the difference despite both variables unifying with the same constant. ", + "bbox": [ + 173, + 501, + 825, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(a) bAbI task 16. A one-to-one mapping is created between variables $X$ :bernhard with brian and Y:frog with lion. ", + "bbox": [ + 173, + 842, + 385, + 892 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87f061c60b0d26309091e0532f96894f1603bdbd0e818eb3a44cf2cee5b59e9e.jpg", + "image_caption": [ + "(c) Logical reasoning task 1. An arity 2 predicate is forced to bind with arity 1 creating a many-to-one binding. " + ], + "image_footnote": [], + "bbox": [ + 181, + 685, + 776, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(b) bAbI task 6. Only Y:bathroom is recognised as variable creating a one-to-many binding to capture the same information. ", + "bbox": [ + 392, + 843, + 604, + 893 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5: Variable bindings produced from equation 3. Darker cells indicate higher attention values. ", + "bbox": [ + 174, + 905, + 821, + 920 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 102, + 343, + 117 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learning an underlying general principle in the form of an invariant is often the means for arguing generalisation in neural networks. For example, Neural Turing Machines (Graves et al., 2014) are tested on previously unseen sequences to support the view that the model might have captured the underlying pattern or algorithm. In fact, Weston et al. (2015) claim “MemNNs can discover simple linguistic patterns based on verbal forms such as (X, dropped, Y), (X, took, Y) or (X, journeyed to, Y) and can successfully generalise the meaning of their instantiations.” However, this claim is based on the output of $f$ and unfortunately it is unknown whether the model has truly learned such a representation or indeed is utilising it. Our approach sheds light to this ambiguity and presents these linguistic patterns explicitly as invariants ensuring their utility through $g$ without solely analysing the output of $f$ on previously unseen symbols. Although we associate these invariants with our existing understanding of the task to mistakenly anthropomorphise the machine, for example by thinking it has learned $X$ :mary as someone, it is important to acknowledge that these are just symbolic patterns. In these cases, our interpretations do not necessarily correspond to any understanding of the machine, relating to the Chinese room argument made by Searle (1980). ", + "bbox": [ + 174, + 140, + 825, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learning invariants by lifting ground examples is related to least common generalisation (Reynolds, 1970) by which inductive inference is performed on facts (Shapiro, 1981) such as generalising went(mary,kitchen) and went(john,garden) to went(X,Y). Unlike in a predicate logic setting, our approach allows for soft alignment and therefore generalisation between varying length sequences. Existing neuro-symbolic systems (Broda et al., 2002) focus on inducing rules that adhere to given logical semantics of what variables and rules are. For example, $\\delta I L P$ (Evans & Grefenstette, 2018) constructs a network by rigidly following the given semantics of first-order logic. Similarly, Lifted Relational Neural Networks (Sourek et al., 2015) ground first-order logic rules into a neural network while Neural Theorem Provers (Rocktschel & Riedel, 2017) build neural networks using backwardchaining (Russell & Norvig, 2016) on a given background knowledge base with templates. However, the notion of a variable is pre-defined rather than learned with a focus on presenting a practical approach to solving certain problems, whereas our motivation stems from a cognitive perspective. ", + "bbox": [ + 174, + 340, + 825, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "At first it may seem the learned invariants, Section 6, make the model more interpretable; however, this transparency is not of the model $f$ but of the data. The invariant captures patterns that potentially approximates the data generating distribution but we still do not know how the model $f$ uses them upstream. Thus, from the perspective of explainable artificial intelligence (XAI) (Adadi & Berrada, 2018), learning invariants or interpreting them do not constitute an explanation of the reasoning model $f$ even though “if someone goes somewhere then they are there” might look like one. Instead, it can be perceived as causal attribution (Miller, 2019) in which someone being somewhere is attributed to them going there. This perspective also relates to gradient based model explanation methods such as Layer-Wise Relevance Propagation (Bach et al., 2015) and Grad-CAM (Selvaraju et al., 2017; Chattopadhay et al., 2018). Consequently, a possible view on $\\psi$ , equation 2, is a gradient based usefulness measure such that a symbol utilised upstream by $f$ to determine the answer becomes a variable similar to how a group of pixels in an image contribute more to its classification. ", + "bbox": [ + 174, + 513, + 825, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, one can argue that our model maintains a form of counterfactual thinking (Roese, 1997) in which soft unification $g$ creates counterfactuals on the invariant example to alter the output of $f$ towards the desired answer, equation 4. The question where Mary would have been if Mary had gone to the garden instead of the kitchen is the process by which an invariant is learned through multiple examples during training. This view relates to methods of causal inference (Pearl, 2019; Holland, 1986) in which counterfactuals are vital as demonstrated in structured models by Pearl (1999). ", + "bbox": [ + 174, + 688, + 825, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8 CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 803, + 318, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a new approach for learning variables and lifting examples into invariants through the usage of soft unification. Evaluating on four datasets, we analysed how Unification Networks perform comparatively to existing similar architectures while having the benefit of lifting examples into invariants that capture underlying patterns present in the tasks. Since our approach is end-toend differentiable, we plan to apply this technique to multi-modal tasks in order to yield multi-modal invariants for example in visual question answering. ", + "bbox": [ + 174, + 840, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 103, + 285, + 117 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Amina Adadi and Mohammed Berrada. Peeking inside the black-box: A survey on explainable artificial intelligence (XAI). IEEE Access, 6:52138–52160, 2018. doi: 10.1109/access.2018. 2870052. ", + "bbox": [ + 174, + 126, + 825, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sebastian Bach, Alexander Binder, Gregoire Montavon, Frederick Klauschen, Klaus-Robert Mller, ´ and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layerwise relevance propagation. PLOS ONE, 10(7):e0130140, July 2015. doi: 10.1371/journal.pone. 0130140. ", + "bbox": [ + 174, + 176, + 825, + 232 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. ICLR, September 2015. ", + "bbox": [ + 171, + 241, + 823, + 270 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Krysia B. Broda, Artur S. D’Avila Garcez, and Dov M. Gabbay. Neural-Symbolic Learning Systems. Springer London, 2002. ISBN 1852335122. ", + "bbox": [ + 171, + 277, + 823, + 308 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Aditya Chattopadhay, Anirban Sarkar, Prantik Howlader, and Vineeth N. Balasubramanian. GradCAM $^ { + + }$ : Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 839–847. IEEE, IEEE, March 2018. doi: 10.1109/wacv.2018.00097. ", + "bbox": [ + 173, + 315, + 825, + 372 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sneha Chaudhari, Gungor Polatkan, Rohan Ramanath, and Varun Mithal. An attentive survey of attention models. IJCAI, April 2019. ", + "bbox": [ + 171, + 380, + 823, + 410 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder–decoder approaches. In Proceedings of SSST-8, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation. Association for Computational Linguistics, 2014. doi: 10.3115/v1/w14-4012. ", + "bbox": [ + 173, + 417, + 825, + 473 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Nuri Cingillioglu and Alessandra Russo. Deeplogic: Towards end-to-end differentiable logical reasoning. AAAI-MAKE, May 2019. ", + "bbox": [ + 174, + 482, + 821, + 511 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Keith L. Clark. Negation as failure. In Logic and Data Bases, pp. 293–322. Springer US, 1978. doi: 10.1007/978-1-4684-3384-5 11. ", + "bbox": [ + 174, + 518, + 821, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Richard Evans and Edward Grefenstette. Learning explanatory rules from noisy data. Journal of Artificial Intelligence Research, 61:1–64, January 2018. doi: 10.1613/jair.5714. ", + "bbox": [ + 171, + 556, + 825, + 585 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Gottlob Frege. Sense and reference. The Philosophical Review, 57(3):209, May 1948. doi: 10.2307/ 2181485. ", + "bbox": [ + 171, + 593, + 825, + 622 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Joe L. Frost, Pei-San Brown, John A. Sutterby, and Candra D. Thornton. The Developmental Benefits Of Playgrounds. Association for Childhood Education International, 2004. ISBN 0871731649. ", + "bbox": [ + 171, + 631, + 823, + 660 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv:1410.5401, 2014. ", + "bbox": [ + 173, + 667, + 803, + 683 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka GrabskaBarwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, ´ Adria Puigdom \\` enech Badia, Karl Moritz Hermann, Yori Zwols, Georg Ostrovski, Adam Cain, \\` Helen King, Christopher Summerfield, Phil Blunsom, Koray Kavukcuoglu, and Demis Hassabis. Hybrid computing using a neural network with dynamic external memory. Nature, 538(7626): 471–476, October 2016. doi: 10.1038/nature20101. ", + "bbox": [ + 174, + 690, + 825, + 775 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Mikael Henaff, Jason Weston, Arthur Szlam, Antoine Bordes, and Yann LeCun. Tracking the world state with recurrent entity networks. ICLR, 2017. ", + "bbox": [ + 173, + 784, + 823, + 813 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Paul W. Holland. Statistics and causal inference. Journal of the American Statistical Association, 81(396):945–960, December 1986. doi: 10.1080/01621459.1986.10478354. ", + "bbox": [ + 173, + 820, + 821, + 849 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. ICLR, 2015. ", + "bbox": [ + 174, + 858, + 795, + 873 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ankit Kumar, Ozan Irsoy, Peter Ondruska, Mohit Iyyer, James Bradbury, Ishaan Gulrajani, Victor Zhong, Romain Paulus, and Richard Socher. Ask me anything: Dynamic memory networks for natural language processing. In ICML, pp. 1378–1387, 2016. ", + "bbox": [ + 176, + 882, + 823, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Richard L. Lewis. Cognitive modeling, symbolic. The MIT encyclopedia of the cognitive sciences, pp. 525–527, 1999. ", + "bbox": [ + 173, + 103, + 823, + 132 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Zachary C. Lipton. The mythos of model interpretability. Communications of the ACM, 61(10): 36–43, September 2018. doi: 10.1145/3233231. ", + "bbox": [ + 171, + 141, + 823, + 170 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Fei Liu and Julien Perez. Gated end-to-end memory networks. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pp. 1–10. Association for Computational Linguistics, 2017. doi: 10.18653/v1/e17-1001. ", + "bbox": [ + 178, + 179, + 823, + 222 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Thang Luong, Hieu Pham, and Christopher D. Manning. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 2015. doi: 10.18653/ v1/d15-1166. ", + "bbox": [ + 173, + 231, + 825, + 287 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tim Miller. Explanation in artificial intelligence: Insights from the social sciences. Artificial Intelligence, 267:1–38, February 2019. doi: 10.1016/j.artint.2018.07.007. ", + "bbox": [ + 173, + 296, + 823, + 327 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Adam Morton and Jerry A. Fodor. The Language of Thought., volume 75. Philosophy Documentation Center, March 1978. doi: 10.2307/2025426. ", + "bbox": [ + 173, + 334, + 823, + 364 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Stephen Muggleton and Luc de Raedt. Inductive logic programming: Theory and methods. The Journal of Logic Programming, 19-20:629–679, May 1994. doi: 10.1016/0743-1066(94) 90035-3. ", + "bbox": [ + 173, + 372, + 825, + 415 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Judea Pearl. Probabilities of causation: three counterfactual interpretations and their identification. Synthese, 121(1-2):93–149, 1999. ", + "bbox": [ + 173, + 424, + 821, + 454 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Judea Pearl. The seven tools of causal inference with reflections on machine learning. Communications of the ACM, 62(3):54–60, February 2019. doi: 10.1145/3241036. ", + "bbox": [ + 173, + 463, + 823, + 492 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jean Piaget. The Psychology of Intelligence. Routledge, August 2001. ISBN 0415254019. ", + "bbox": [ + 169, + 501, + 767, + 517 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "John C. Reynolds. Transformational systems and algebraic structure of atomic formulas. Machine intelligence, 5:135–151, 1970. ", + "bbox": [ + 173, + 525, + 826, + 554 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tim Rocktschel and Sebastian Riedel. End-to-end differentiable proving. NIPS, pp. 3791–3803, 2017. ", + "bbox": [ + 174, + 563, + 825, + 592 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Neal J. Roese. Counterfactual thinking. Psychological Bulletin, 121(1):133–148, 1997. doi: 10. 1037/0033-2909.121.1.133. ", + "bbox": [ + 173, + 601, + 825, + 631 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach $\\langle 3 ^ { r d }$ Edition). Pearson, November 2016. ISBN 1292153962. ", + "bbox": [ + 169, + 638, + 825, + 669 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "John R. Searle. Minds, brains, and programs. Behavioral and brain sciences, 3(3):417–424, 1980. doi: 10.1016/b978-1-4832-1446-7.50007-8. ", + "bbox": [ + 169, + 678, + 823, + 707 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pp. 618–626. IEEE, October 2017. doi: 10.1109/iccv.2017.74. ", + "bbox": [ + 173, + 714, + 825, + 771 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Minjoon Seo, Sewon Min, Ali Farhadi, and Hannaneh Hajishirzi. Query-reduction networks for question answering. ICLR, 2017. ", + "bbox": [ + 168, + 780, + 823, + 810 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ehud Y. Shapiro. Inductive inference of theories from facts. Yale University, Department of Computer Science, 1981. ", + "bbox": [ + 171, + 819, + 823, + 848 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Gustav Sourek, Vojtech Aschenbrenner, Filip Zelezny, and Ondrej Kuzelka. Lifted relational neural networks. arXiv:1508.05128, 2015. ", + "bbox": [ + 173, + 857, + 823, + 886 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory networks. In NIPS, pp. 2440–2448, 2015. ", + "bbox": [ + 173, + 895, + 821, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Seiya Tokui, Kenta Oono, Shohei Hido, and Justin Clayton. Chainer: a next-generation open source framework for deep learning. In Proceedings of Workshop on Machine Learning Systems (LearningSys) in The Twenty-ninth Annual Conference on Neural Information Processing Systems (NIPS), 2015. URL http://learningsys.org/papers/LearningSys_2015_ paper_33.pdf. ", + "bbox": [ + 173, + 103, + 825, + 174 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "J. Marshall Unger and Terrence W. Deacon. The Symbolic Species: The Co-Evolution of Language and the Brain, volume 82. Wiley, 1998. doi: 10.2307/329984. ", + "bbox": [ + 171, + 183, + 821, + 212 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. ICLR, 2015. ", + "bbox": [ + 171, + 219, + 723, + 234 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jason Weston, Antoine Bordes, Sumit Chopra, Alexander M. Rush, Bart van Merrinboer, Armand Joulin, and Tomas Mikolov. Towards ai-complete question answering: A set of prerequisite toy tasks. ICLR, 2016. ", + "bbox": [ + 171, + 244, + 825, + 286 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and textual question answering. ICML, pp. 2397–2406, 2016. ", + "bbox": [ + 171, + 296, + 823, + 324 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A MODEL DETAILS ", + "bbox": [ + 176, + 102, + 352, + 118 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A.1 UNIFICATION MLP & CNN ", + "text_level": 1, + "bbox": [ + 176, + 132, + 410, + 148 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Unification MLP (UMLP) To model $f$ as a multi-layer perceptron, we take symbol embeddings of size $d = 1 6$ and flatten sequences of length $l = 4$ into an input vector of size $\\phi ( \\pmb { k } ) \\in \\mathbb { R } ^ { 6 4 }$ . The MLP consists of 2 hidden layers with tanh non-linearity of sizes $2 d$ and $d$ respectively. To process the query, we concatenate the one-hot encoding of the task id to $\\phi ( k )$ yielding a final input of size $6 4 + 4 = 6 8$ . For unification features $\\phi _ { U }$ , we use a bi-directional GRU with hidden size $d$ and an initial state of 0. The hidden state at each symbol is taken with a linear transformation to give $\\phi _ { U } ( s ) = W _ { U } \\Phi ( s )$ where $\\Phi ( s )$ is the hidden state of the biGRU. The variable assignment is then computed as an attention over the according to equation 3. ", + "bbox": [ + 173, + 159, + 825, + 271 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Unification CNN (UCNN) We take symbols embeddings of size $d = 3 2$ to obtain an input grid $\\phi ( K ) \\in \\mathbb { R } ^ { 3 \\times 3 \\times 3 2 }$ . Similar to UMLP, for each symbol we append the task id as a one-hot vector to get an input of shape $3 \\times 3 \\times ( 3 2 + 4 )$ . Then $f$ consists of 2 convolutional layers with $d$ filters each, kernel size of 3 and stride 1. We use relu non-linearity in between the layers. We pad the grid with 2 columns and 2 rows to a $5 \\times 5$ such that the output of the convolutions yield again a hidden output $\\boldsymbol { \\mathsf { H } } \\in \\mathbb { R } ^ { 3 \\times 3 \\times d }$ of the same shape. As the final hidden output $h$ , we take a global max pool to over $\\boldsymbol { \\mathsf { H } }$ to obtain $\\pmb { h } \\in \\mathbb { R } ^ { d }$ . Unification function $g$ is modelled identical to $f$ without the max pooling such that $\\phi _ { U } ( K _ { i j } ) = \\mathsf { H } _ { i j } ^ { \\prime }$ where $\\boldsymbol { \\mathsf { H } } ^ { \\prime }$ is the hidden output of the convolutional layers. ", + "bbox": [ + 173, + 277, + 825, + 390 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A.2 UNIFICATION MEMORY NETWORKS ", + "text_level": 1, + "bbox": [ + 176, + 405, + 465, + 420 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Unlike previous architectures, with UMN we interleave $g$ into $f$ . We use embedding sizes of $d = 3 2$ and model $f$ with an iterative memory network. We take the final hidden state of a bi-directional GRU, with initial state 0, $\\Phi _ { M }$ to represent the sentences of the context $C$ and query $\\pmb q$ in a $d$ - dimensional vector $M _ { i } = \\Phi _ { M } ( C _ { i } )$ and the query ${ \\pmb m } _ { \\ b q } = \\Phi _ { M } ( { \\pmb q } )$ . The initial state of the memory network is $h ^ { 0 } = m _ { q }$ . At each iteration $j$ : ", + "bbox": [ + 173, + 431, + 825, + 503 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/b17d2332dda7b22d96bbade43e656c75a1713d2c81331c84317246e400303740.jpg", + "text": "$$\n\\begin{array} { c } { { A _ { i } ^ { j } = \\operatorname { t a n h } ( W \\rho ( M _ { i } , h ^ { j } ) + b ) } } \\\\ { { \\beta ^ { j } = \\operatorname { s o f t m a x } ( W \\Phi _ { A } ( A ^ { j } ) + b ) } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 390, + 510, + 606, + 553 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "where $\\Phi _ { A }$ is another $d$ -dimensional bi-directional GRU and $\\rho ( { \\pmb x } , { \\pmb y } ) = [ { \\pmb x } ; { \\pmb y } ; { \\pmb x } \\odot { \\pmb y } ; ( { \\pmb x } - { \\pmb y } ) ^ { 2 } ]$ with $\\odot$ the element-wise multiplication and $[ ; ]$ the concatenation of vectors. Taking $\\beta ^ { j }$ as the context attention, we obtain the next state of the memory network: ", + "bbox": [ + 174, + 558, + 825, + 601 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/85e2e7ec827f8b1efbb28c02265ab15e959abdf242665360146e66fedab5e494.jpg", + "text": "$$\nh ^ { j + 1 } = \\sum _ { i } \\beta _ { i } ^ { j } \\operatorname { t a n h } ( W \\rho ( M _ { i } , h ^ { j } ) + b )\n$$", + "text_format": "latex", + "bbox": [ + 364, + 608, + 633, + 641 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "and iterate $J$ many times in advance. The final prediction becomes $f ( C , \\boldsymbol { q } ) = \\mathrm { s o f t m a x } ( W h ^ { J } + b )$ . \nAll weight matrices $W$ and bias vectors $^ { b }$ are independent but are tied across iterations. ", + "bbox": [ + 173, + 648, + 825, + 678 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0614697dfdc3aacbd0a222b3317ffc3cd4f9611857ee0057c4b330f335467084.jpg", + "table_caption": [ + "B GENERATED DATASET SAMPLES Table 4: Sample context, query and answer triples from sequences and grid tasks. ", + "Table 5: Training sizes for randomly generated fixed length sequences and grid tasks with 8 unique symbols. The reason for Grid task (i) to be smaller is because there are at most 32 combinations of $2 \\times 2$ boxes in a $3 \\times 3$ grid with 8 unique symbols. " + ], + "table_footnote": [], + "table_body": "
DatasetTaskContextQueryAnswer
Sequencei1488constant2
Sequenceii6157head6
Sequenceii1837tail7
Sequenceiv3563duplicate3
Gridi００８ 022 022box2
Gridii4 -070 0 页 0 8 1head4
Gridii0 673 0 1 2 0 0centre7
Gridiv8 0 0 2 60 41corner2
", + "bbox": [ + 313, + 154, + 683, + 498 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/164eed252519d30abfd4208cfebc52987bd6e5948e3125927523777305272436.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Task SequencesGrid
i704.7 ± 12.825.6 ± 1.8
ii709.4 ± 13.8623.7 ± 14.1
ii709.7 ± 14.0768.2 ±12.5
iv624.8 ± 12.4795.2 ± 10.3
", + "bbox": [ + 375, + 555, + 620, + 641 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "C TRAINING DETAILS ", + "text_level": 1, + "bbox": [ + 174, + 101, + 372, + 118 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "C.1 UNIFICATION MLP & CNN ", + "text_level": 1, + "bbox": [ + 176, + 132, + 408, + 148 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Both unification models are trained on a 5-fold cross-validation over the generated datasets for 2000 iterations with a batch size of 64. We don’t use any weight decay and save the training and test accuracies every 10 iterations, as presented in Figure 3. ", + "bbox": [ + 174, + 159, + 825, + 202 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "C.2 UNIFICATION MEMORY NETWORKS ", + "text_level": 1, + "bbox": [ + 176, + 218, + 465, + 233 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We again use a batch size of 64 and pre-train $f$ for 40 epochs then $f$ together with $g$ for 260 epochs. We use epochs for UMN since the dataset sizes are fixed. To learn $g$ alongside $f$ , we combine error signals from the unification of the invariant and the example. Following equation 4, the objective function not only incorporates the negative log-likelihood ${ \\mathcal L } _ { \\mathrm { n l l } }$ of the answer but also the mean squared error between intermediate states $h _ { I } ^ { j }$ and $h _ { K } ^ { j }$ at each iteration as an auxiliary loss: ", + "bbox": [ + 176, + 243, + 823, + 318 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/d642031dec0d0c1fb74310b00fa6ec184944a8e1c66f93234e97a793ae2cb952.jpg", + "text": "$$\nJ = \\mathcal { L } _ { \\mathrm { n l l } } ( f ( K ) , a ) + \\lambda _ { U } \\left[ \\mathcal { L } _ { \\mathrm { n l l } } ( f \\circ g ( I , K ) , a ) + \\frac { 1 } { J } \\sum _ { j = 1 } ^ { J } ( h _ { I } ^ { j } - h _ { K } ^ { j } ) ^ { 2 } + \\tau \\sum _ { s \\in \\mathcal { S } } \\psi _ { w } ( s ) \\right]\n$$", + "text_format": "latex", + "bbox": [ + 217, + 324, + 781, + 375 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We pre-train by setting $\\lambda _ { U } = 0$ for 40 epochs and then set $\\lambda _ { U } = 1$ . For strong supervision we also compute the negative log-likelihood ${ \\mathcal { L } } _ { \\mathrm { n l l } }$ for the context attention $\\beta ^ { j }$ , described in Appendix A, at each iteration using the supporting facts of the tasks. We apply a dropout of 0.1 for all recurrent neural networks used and only for the bAbI dataset weight decay with 0.001 as the coefficient. ", + "bbox": [ + 174, + 380, + 823, + 436 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "D FURTHER RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 102, + 370, + 118 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/c02ba1f113eff3023fe0e72f8db887cb964bac4e0057d50edcd72e8e1f55364f.jpg", + "image_caption": [ + "Figure 6: Results of Unification MLP and CNN on increasing number of invariants. There is no impact on performance when more invariants per task are given. Upon closer inspection, we noticed the models ignore the extra invariants and only use 1. We speculate the regularisation $\\psi$ encourages the models to use a single 1 invariant. " + ], + "image_footnote": [], + "bbox": [ + 209, + 128, + 789, + 329 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/7d93fa930a25e39a8a24ed1d60a6f940986c8670a778209ead567218a03c63c2.jpg", + "table_caption": [ + "Table 6: Individual task error rates on bAbI tasks for Unification Memory Networks. " + ], + "table_footnote": [], + "table_body": "
SupervisionWeakStrong
#Invs Training Size1 1k3 1k1 1k3 1k3 50
10.00.00.00.01.4
265.663.10.30.730.0
367.162.61.02.439.8
40.00.00.00.037.0
53.44.00.81.126.5
60.20.60.00.018.4
722.022.810.711.322.8
810.38.57.47.624.7
90.125.70.00.033.8
100.12.00.00.332.6
110.00.00.00.011.9
120.00.10.00.021.3
132.13.70.00.15.8
1419.713.50.50.154.8
150.00.70.00.00.0
1655.256.20.00.039.7
1739.249.051.149.348.8
184.48.00.60.510.4
1991.889.653.946.790.2
200.00.00.00.02.7
Mean19.120.56.36.027.6
Std27.927.015.614.321.0
#>5%8104417
", + "bbox": [ + 341, + 429, + 653, + 779 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/ecc837386c7d436084e3ba2bb9df9b93eb4bf84050d41176fe6f0c107c83ab6a.jpg", + "table_caption": [ + "Table 7: Comparison of individual task error rates $( \\% )$ on the bAbI (Weston et al., 2016) dataset of the best run. We preferred 1k results if a model had experiments published on both 1k and 10k for data efficiency. References from left to right: (Sukhbaatar et al., 2015) - (Liu & Perez, 2017) - (Henaff et al., 2017) - (Seo et al., 2017) - Ours - (Xiong et al., 2016) - (Graves et al., 2016) - (Weston et al., 2015) - Ours - (Kumar et al., 2016) " + ], + "table_footnote": [], + "table_body": "
Support Size ModelWeakStrong
GN2N1k EntNetQRNUMN10k DMN+ DNC1k MemNN UMN10k DMN
10.00.00.70.00.00.00.00.00.00.0
28.38.156.40.565.60.30.40.00.71.8
340.338.869.71.267.11.11.80.02.44.8
42.80.41.40.70.00.00.00.00.00.0
513.11.04.61.23.40.50.82.01.10.7
67.68.430.01.20.20.00.00.00.00.0
717.317.822.39.422.02.40.615.011.33.1
810.012.519.23.710.30.00.39.07.63.5
913.210.731.50.00.10.00.20.00.00.0
1015.116.515.60.00.10.00.22.00.32.5
110.90.08.00.00.00.00.00.00.00.1
120.20.00.80.00.00.00.00.00.00.0
130.40.09.00.32.10.00.10.00.10.2
141.71.262.93.819.70.20.41.00.10.0
150.00.057.80.00.00.00.00.00.00.0
161.30.153.253.455.245.355.10.00.00.6
1751.041.746.451.839.24.212.035.049.340.6
1811.19.28.88.84.42.10.85.00.54.7
1982.888.590.490.791.80.03.964.046.765.5
200.00.02.60.30.00.00.00.00.00.0
Mean
#>5%13.9 1112.7 1029.6 1511.3 519.1 82.8 13.8 26.7 46.0 46.4 2
", + "bbox": [ + 187, + 208, + 808, + 546 + ], + "page_idx": 15 + }, + { + "type": "table", + "img_path": "images/0ff92c677511d717c131f4e89e502bc5ae32309f23904fe815fac80a615b0be8.jpg", + "table_caption": [ + "Table 8: Individual task error rates $( \\% )$ on the logical reasoning dataset. " + ], + "table_footnote": [], + "table_body": "
Size1k10k5020k
SupportWeakStrongWeakWeak
Arity1111Strong 22
# Invs /Model12 312 312 312 132 3DMN2 IMA
Facts1.20.90.00.40.00.00.00.00.033.50.00.0
Unification0.010.30.010.80.00.00.00.00.041.313.010.0
1 Step50.349.84.420.01.227.80.11.35.750.226.02.0
2 Steps47.550.05.735.037.247.80.029.728.749.933.05.0
3 Steps47.649.210.438.739.645.60.026.026.148.323.06.0
AND31.337.410.716.429.829.00.20.41.250.020.05.0
OR25.238.121.035.020.530.24.420.617.447.613.03.0
Transitivity50.026.639.65.06.049.250.050.0
1 Step NBF46.438.73.828.81.121.60.11.18.047.621.02.0
2 Steps NBF48.548.97.739.630.448.20.133.428.750.315.04.0
AND NBF51.050.143.148.629.444.20.11.340.149.516.08.0
OR NBF51.448.450.847.347.647.821.327.630.547.325.014.0
Mean36.439.314.328.921.531.82.412.216.047.121.29.1
Std18.715.916.414.117.116.76.113.213.64.712.313.4
#>5%91171171015912115
", + "bbox": [ + 178, + 632, + 823, + 895 + ], + "page_idx": 15 + }, + { + "type": "table", + "img_path": "images/d8edacdbc450e3ee1dc957f53d9955390327d1fc6d5530051b7e0a7c2cf7a8b4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Xsandra went back to the Y:bathroom
is X:sandra in the Y:bathroom
yes
", + "bbox": [ + 366, + 103, + 632, + 156 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Figure 7: bAbI task 6, yes or no questions. The invariant does not variablise the answer. ", + "bbox": [ + 210, + 167, + 784, + 183 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/53b241b6490c42e9c05a37ee42eac354b639bb17d0da4793cbec9044f1ca384f.jpg", + "text": "$$\n\\begin{array} { r l } { X { \\mathrm { : m ~ ( ~ Y : e ~ ) ~ } } } & { { } { \\vdash X { \\mathrm { : m ~ ( ~ Y : e ~ ) ~ } } } } \\\\ { X { \\mathrm { : a ~ ( ~ } Y { \\mathrm { : w ~ , ~ } } { \\mathrm { ~ Z : e ~ ) ~ } } } } & { { } { \\vdash X { \\mathrm { : a ~ ( ~ } Y { \\mathrm { : w ~ , ~ } } { \\mathrm { Z : e ~ ) ~ } } } } } \\\\ { X { \\mathrm { : m ~ ( ~ T ~ ) ~ } } } & { { } { \\vdash X { \\mathrm { : m ~ ( ~ c ~ ) ~ } } } } \\\\ { X { \\mathrm { : x ~ ( ~ \\mathbb { A } ~ ) n o t ~ \\mathbf { Z } { \\mathrm { : q } } ~ ( ~ \\mathbb { A } ~ ) ~ } } } & { { } { \\vdash X { \\mathrm { : x ~ ( ~ } Y { \\mathrm { : z } } ~ ) } } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 343, + 188, + 653, + 261 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Figure 8: Invariants learned on tasks 1, 2 and 11 with arity 1 and 2 from the logical reasoning dataset. Last invariant on task 11 lifts the example around the negation by failure, denoted as not, capturing its semantics. ", + "bbox": [ + 174, + 271, + 823, + 313 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "3 X 7 Y head 7 4 X Y head X 3 1 X tail X X 5 Y duplicate ", + "bbox": [ + 274, + 318, + 395, + 375 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/af9a832b2966a56b80d8d6d927634fd72447a59e0433605421d6b017bdc935e8.jpg", + "text": "$$\n\\begin{array} { r l } { X \\mathrm { ~ 0 ~ ~ 0 ~ ~ } } & { \\mathrm { ~ 0 ~ ~ Y ~ 0 ~ ~ 6 ~ 4 ~ } Y } \\\\ { \\mathrm { ~ 0 ~ 1 ~ 0 ~ } } & { \\mathrm { ~ 1 ~ } X \\mathrm { ~ 8 ~ } } \\\\ { \\mathrm { ~ 0 ~ 0 ~ } Y } & { \\mathrm { ~ 0 ~ 2 ~ 0 ~ } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 573, + 315, + 759, + 375 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "(a) Invariants with extra varialbes learned with UMLP. (b) Mismatching invariants learned with UCNN. ", + "bbox": [ + 169, + 381, + 805, + 396 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Figure 9: Invariants learned that do not match the data generating distribution from UMLP and UCNN using $\\leq 1 0 0 0$ examples to train. In these instances the unification still bind to the the correct symbols in order to predict the desired answer; quantitatively we get the same results. Variable default symbols are omitted for clarity. ", + "bbox": [ + 173, + 406, + 826, + 462 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "(a) bAbI task 2. When a variable is unused in the next iteration, e.g. Z:football, it unifies with random tokens often biased by position. ", + "bbox": [ + 173, + 593, + 385, + 643 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/9073cfbc8cdd1d7408ab2bf49ed3c03468c3850a67e1c59f65e66ccdfccc89e1.jpg", + "image_caption": [ + "(c) Logical reasoning task 3. Arity 1 atom forced to bind with arity 2 creates a one-to-many mapping. " + ], + "image_footnote": [], + "bbox": [ + 191, + 472, + 810, + 582 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "(b) Logical reasoning task 1. A one-to-one alignment is created between predicates and constants. ", + "bbox": [ + 392, + 593, + 604, + 631 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Figure 10: Further attention maps for equation 3, darker cells indicate higher attention values. ", + "bbox": [ + 189, + 655, + 805, + 670 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/parse/train/r1xwA34KDB/r1xwA34KDB_middle.json b/parse/train/r1xwA34KDB/r1xwA34KDB_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..07bdf5b85bf8244c6ceb1af4c6c0e15dc9ed327c --- /dev/null +++ b/parse/train/r1xwA34KDB/r1xwA34KDB_middle.json @@ -0,0 +1,44811 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 78, + 502, + 96 + ], + "lines": [ + { + "bbox": [ + 104, + 78, + 506, + 99 + ], + "spans": [ + { + "bbox": [ + 104, + 78, + 506, + 99 + ], + "score": 1.0, + "content": "LEARNING INVARIANTS THROUGH SOFT UNIFICATION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 112, + 115, + 244, + 137 + ], + "lines": [ + { + "bbox": [ + 113, + 115, + 201, + 127 + ], + "spans": [ + { + "bbox": [ + 113, + 115, + 201, + 127 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 111, + 126, + 245, + 138 + ], + "spans": [ + { + "bbox": [ + 111, + 126, + 245, + 138 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 278, + 167, + 333, + 179 + ], + "lines": [ + { + "bbox": [ + 277, + 166, + 335, + 180 + ], + "spans": [ + { + "bbox": [ + 277, + 166, + 335, + 180 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 143, + 194, + 468, + 315 + ], + "lines": [ + { + "bbox": [ + 141, + 194, + 469, + 207 + ], + "spans": [ + { + "bbox": [ + 141, + 194, + 469, + 207 + ], + "score": 1.0, + "content": "Human reasoning involves recognising common underlying principles across", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 205, + 470, + 218 + ], + "spans": [ + { + "bbox": [ + 141, + 205, + 470, + 218 + ], + "score": 1.0, + "content": "many examples by utilising variables. The by-products of such reasoning are", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 216, + 470, + 229 + ], + "spans": [ + { + "bbox": [ + 141, + 216, + 470, + 229 + ], + "score": 1.0, + "content": "invariants that capture patterns across examples such as “if someone went some-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 227, + 469, + 239 + ], + "spans": [ + { + "bbox": [ + 141, + 227, + 469, + 239 + ], + "score": 1.0, + "content": "where then they are there” without mentioning specific people or places. Humans", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 238, + 470, + 251 + ], + "spans": [ + { + "bbox": [ + 141, + 238, + 470, + 251 + ], + "score": 1.0, + "content": "learn what variables are and how to use them at a young age, and the question this", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 249, + 469, + 261 + ], + "spans": [ + { + "bbox": [ + 141, + 249, + 469, + 261 + ], + "score": 1.0, + "content": "paper addresses is whether machines can also learn and use variables solely from", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 260, + 470, + 272 + ], + "spans": [ + { + "bbox": [ + 141, + 260, + 470, + 272 + ], + "score": 1.0, + "content": "examples without requiring human pre-engineering. We propose Unification Net-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 271, + 469, + 283 + ], + "spans": [ + { + "bbox": [ + 141, + 271, + 469, + 283 + ], + "score": 1.0, + "content": "works that incorporate soft unification into neural networks to learn variables and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 281, + 469, + 294 + ], + "spans": [ + { + "bbox": [ + 141, + 281, + 469, + 294 + ], + "score": 1.0, + "content": "by doing so lift examples into invariants that can then be used to solve a given task.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 293, + 469, + 306 + ], + "spans": [ + { + "bbox": [ + 142, + 293, + 469, + 306 + ], + "score": 1.0, + "content": "We evaluate our approach on four datasets to demonstrate that learning invariants", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 304, + 439, + 316 + ], + "spans": [ + { + "bbox": [ + 141, + 304, + 439, + 316 + ], + "score": 1.0, + "content": "captures patterns in the data and can improve performance over baselines.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 9 + }, + { + "type": "title", + "bbox": [ + 108, + 342, + 206, + 354 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 208, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 208, + 357 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 369, + 505, + 490 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 382 + ], + "score": 1.0, + "content": "Humans have the ability to process symbolic knowledge and maintain symbolic thought (Unger", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 380, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 392 + ], + "score": 1.0, + "content": "& Deacon, 1998). When reasoning, humans do not require combinatorial enumeration of exam-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "ples but instead utilise invariant patterns with placeholders replacing specific entities. Symbolic", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 104, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "cognitive models (Lewis, 1999) embrace this perspective with the human mind seen as an informa-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 412, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 506, + 425 + ], + "score": 1.0, + "content": "tion processing system operating on formal symbols such as reading a stream of tokens in natural", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 506, + 437 + ], + "score": 1.0, + "content": "language. The language of thought hypothesis (Morton & Fodor, 1978) frames human thought as a", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 334, + 448 + ], + "score": 1.0, + "content": "structural construct with varying sub-components such as", + "type": "text" + }, + { + "bbox": [ + 334, + 435, + 348, + 445 + ], + "score": 0.36, + "content": "^ { 6 6 } \\mathrm { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 434, + 506, + 448 + ], + "score": 1.0, + "content": "went to Y”. By recognising what varies", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "score": 1.0, + "content": "across examples, humans are capable of lifting examples into invariant principles that account for", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "other instances. This symbolic thought with variables is learned at a young age through symbolic", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "score": 1.0, + "content": "play (Piaget, 2001). For instance a child learns that a sword can be substituted with a stick (Frost", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 478, + 268, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 268, + 492 + ], + "score": 1.0, + "content": "et al., 2004) and engage in pretend play.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 108, + 495, + 356, + 637 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 356, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 356, + 507 + ], + "score": 1.0, + "content": "Although variables are inherent in models of computation and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 505, + 357, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 357, + 520 + ], + "score": 1.0, + "content": "symbolic formalisms, as in first-order logic (Russell & Norvig,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 516, + 357, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 357, + 530 + ], + "score": 1.0, + "content": "2016), they are pre-engineered and used to solve specific tasks", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 528, + 357, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 357, + 541 + ], + "score": 1.0, + "content": "by means of unification or assignments that bound variables to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 539, + 357, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 357, + 552 + ], + "score": 1.0, + "content": "given values. However, when learning from data only, being", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 550, + 356, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 356, + 562 + ], + "score": 1.0, + "content": "able to recognise when and which symbols should take on dif-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 561, + 357, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 357, + 574 + ], + "score": 1.0, + "content": "ferent values, i.e. symbols that can act as variables, is crucial", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 572, + 356, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 356, + 584 + ], + "score": 1.0, + "content": "for lifting examples into general principles that are invariant", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 583, + 357, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 357, + 595 + ], + "score": 1.0, + "content": "across multiple instances. Figure 1 shows the invariant learned", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 594, + 357, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 357, + 607 + ], + "score": 1.0, + "content": "by our approach: if someone is the same thing as someone", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 605, + 357, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 357, + 617 + ], + "score": 1.0, + "content": "else then they have the same colour. With this invariant, our", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 616, + 357, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 357, + 629 + ], + "score": 1.0, + "content": "approach can solve all of the training and test examples in task", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 626, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 286, + 640 + ], + "score": 1.0, + "content": "16 of the bAbI dataset (Weston et al., 2016).", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34 + }, + { + "type": "table", + "bbox": [ + 374, + 502, + 492, + 568 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 374, + 502, + 492, + 568 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 374, + 502, + 492, + 568 + ], + "spans": [ + { + "bbox": [ + 374, + 502, + 492, + 568 + ], + "score": 0.78, + "html": "
X:bernhard is a Y:frog Z:lily is a Y:frog
Z:lily is A:green
what colour is X:bernhard
A:green
", + "type": "table", + "image_path": "005e03f738687ce974ff771c18081cb699f106f6d11b9f8f18552de03cc0852f.jpg" + } + ] + } + ], + "index": 36.5, + "virtual_lines": [ + { + "bbox": [ + 374, + 502, + 492, + 535.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 374, + 535.0, + 492, + 568.0 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 363, + 578, + 505, + 644 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 363, + 578, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 363, + 578, + 505, + 590 + ], + "score": 1.0, + "content": "Figure 1: Invariant learned for", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 363, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 363, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "bAbI task 16, basic induction,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 363, + 600, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 363, + 600, + 391, + 612 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 392, + 600, + 400, + 610 + ], + "score": 0.54, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 600, + 505, + 612 + ], + "score": 1.0, + "content": ":bernhard denotes a vari-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 363, + 612, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 363, + 612, + 505, + 623 + ], + "score": 1.0, + "content": "able with default symbol bernhard.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 363, + 622, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 363, + 622, + 505, + 633 + ], + "score": 1.0, + "content": "This single invariant accounts for", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 363, + 632, + 501, + 646 + ], + "spans": [ + { + "bbox": [ + 363, + 632, + 501, + 646 + ], + "score": 1.0, + "content": "all the training and test examples.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 43.5 + } + ], + "index": 40.0 + }, + { + "type": "text", + "bbox": [ + 108, + 644, + 356, + 655 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 356, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 356, + 656 + ], + "score": 1.0, + "content": "In this paper we address the question of whether a machine can", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + }, + { + "type": "text", + "bbox": [ + 107, + 653, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "learn and use the notion of a variable, i.e. a symbol that can take on different values. For instance,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 665, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 677 + ], + "score": 1.0, + "content": "given an example of the form “bernhard is a frog” the machine would learn that the token “bernhard”", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "could be someone else and the token “frog” could be something else. If we consider unification a", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "selection of the most appropriate value for a variable given a choice of values, we can reframe it as", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "a form of attention. Attention models (Bahdanau et al., 2015; Luong et al., 2015; Chaudhari et al.,", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "2019) allow neural networks to focus, attend to certain parts of the input often for the purpose of", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "selecting a relevant portion. Since attention mechanisms are also differentiable they are often jointly", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 52 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 78, + 502, + 96 + ], + "lines": [ + { + "bbox": [ + 104, + 78, + 506, + 99 + ], + "spans": [ + { + "bbox": [ + 104, + 78, + 506, + 99 + ], + "score": 1.0, + "content": "LEARNING INVARIANTS THROUGH SOFT UNIFICATION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 112, + 115, + 244, + 137 + ], + "lines": [ + { + "bbox": [ + 113, + 115, + 201, + 127 + ], + "spans": [ + { + "bbox": [ + 113, + 115, + 201, + 127 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 111, + 126, + 245, + 138 + ], + "spans": [ + { + "bbox": [ + 111, + 126, + 245, + 138 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 111, + 115, + 245, + 138 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 167, + 333, + 179 + ], + "lines": [ + { + "bbox": [ + 277, + 166, + 335, + 180 + ], + "spans": [ + { + "bbox": [ + 277, + 166, + 335, + 180 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 143, + 194, + 468, + 315 + ], + "lines": [ + { + "bbox": [ + 141, + 194, + 469, + 207 + ], + "spans": [ + { + "bbox": [ + 141, + 194, + 469, + 207 + ], + "score": 1.0, + "content": "Human reasoning involves recognising common underlying principles across", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 205, + 470, + 218 + ], + "spans": [ + { + "bbox": [ + 141, + 205, + 470, + 218 + ], + "score": 1.0, + "content": "many examples by utilising variables. The by-products of such reasoning are", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 216, + 470, + 229 + ], + "spans": [ + { + "bbox": [ + 141, + 216, + 470, + 229 + ], + "score": 1.0, + "content": "invariants that capture patterns across examples such as “if someone went some-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 227, + 469, + 239 + ], + "spans": [ + { + "bbox": [ + 141, + 227, + 469, + 239 + ], + "score": 1.0, + "content": "where then they are there” without mentioning specific people or places. Humans", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 238, + 470, + 251 + ], + "spans": [ + { + "bbox": [ + 141, + 238, + 470, + 251 + ], + "score": 1.0, + "content": "learn what variables are and how to use them at a young age, and the question this", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 249, + 469, + 261 + ], + "spans": [ + { + "bbox": [ + 141, + 249, + 469, + 261 + ], + "score": 1.0, + "content": "paper addresses is whether machines can also learn and use variables solely from", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 260, + 470, + 272 + ], + "spans": [ + { + "bbox": [ + 141, + 260, + 470, + 272 + ], + "score": 1.0, + "content": "examples without requiring human pre-engineering. We propose Unification Net-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 271, + 469, + 283 + ], + "spans": [ + { + "bbox": [ + 141, + 271, + 469, + 283 + ], + "score": 1.0, + "content": "works that incorporate soft unification into neural networks to learn variables and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 281, + 469, + 294 + ], + "spans": [ + { + "bbox": [ + 141, + 281, + 469, + 294 + ], + "score": 1.0, + "content": "by doing so lift examples into invariants that can then be used to solve a given task.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 293, + 469, + 306 + ], + "spans": [ + { + "bbox": [ + 142, + 293, + 469, + 306 + ], + "score": 1.0, + "content": "We evaluate our approach on four datasets to demonstrate that learning invariants", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 304, + 439, + 316 + ], + "spans": [ + { + "bbox": [ + 141, + 304, + 439, + 316 + ], + "score": 1.0, + "content": "captures patterns in the data and can improve performance over baselines.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 9, + "bbox_fs": [ + 141, + 194, + 470, + 316 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 342, + 206, + 354 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 208, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 208, + 357 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 369, + 505, + 490 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 382 + ], + "score": 1.0, + "content": "Humans have the ability to process symbolic knowledge and maintain symbolic thought (Unger", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 380, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 392 + ], + "score": 1.0, + "content": "& Deacon, 1998). When reasoning, humans do not require combinatorial enumeration of exam-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "ples but instead utilise invariant patterns with placeholders replacing specific entities. Symbolic", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 104, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "cognitive models (Lewis, 1999) embrace this perspective with the human mind seen as an informa-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 412, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 506, + 425 + ], + "score": 1.0, + "content": "tion processing system operating on formal symbols such as reading a stream of tokens in natural", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 506, + 437 + ], + "score": 1.0, + "content": "language. The language of thought hypothesis (Morton & Fodor, 1978) frames human thought as a", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 334, + 448 + ], + "score": 1.0, + "content": "structural construct with varying sub-components such as", + "type": "text" + }, + { + "bbox": [ + 334, + 435, + 348, + 445 + ], + "score": 0.36, + "content": "^ { 6 6 } \\mathrm { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 434, + 506, + 448 + ], + "score": 1.0, + "content": "went to Y”. By recognising what varies", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "score": 1.0, + "content": "across examples, humans are capable of lifting examples into invariant principles that account for", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "other instances. This symbolic thought with variables is learned at a young age through symbolic", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "score": 1.0, + "content": "play (Piaget, 2001). For instance a child learns that a sword can be substituted with a stick (Frost", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 478, + 268, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 268, + 492 + ], + "score": 1.0, + "content": "et al., 2004) and engage in pretend play.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 21, + "bbox_fs": [ + 104, + 369, + 506, + 492 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 495, + 356, + 637 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 356, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 356, + 507 + ], + "score": 1.0, + "content": "Although variables are inherent in models of computation and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 505, + 357, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 357, + 520 + ], + "score": 1.0, + "content": "symbolic formalisms, as in first-order logic (Russell & Norvig,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 516, + 357, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 357, + 530 + ], + "score": 1.0, + "content": "2016), they are pre-engineered and used to solve specific tasks", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 528, + 357, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 357, + 541 + ], + "score": 1.0, + "content": "by means of unification or assignments that bound variables to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 539, + 357, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 357, + 552 + ], + "score": 1.0, + "content": "given values. However, when learning from data only, being", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 550, + 356, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 356, + 562 + ], + "score": 1.0, + "content": "able to recognise when and which symbols should take on dif-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 561, + 357, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 357, + 574 + ], + "score": 1.0, + "content": "ferent values, i.e. symbols that can act as variables, is crucial", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 572, + 356, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 356, + 584 + ], + "score": 1.0, + "content": "for lifting examples into general principles that are invariant", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 583, + 357, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 357, + 595 + ], + "score": 1.0, + "content": "across multiple instances. Figure 1 shows the invariant learned", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 594, + 357, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 357, + 607 + ], + "score": 1.0, + "content": "by our approach: if someone is the same thing as someone", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 605, + 357, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 357, + 617 + ], + "score": 1.0, + "content": "else then they have the same colour. With this invariant, our", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 616, + 357, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 357, + 629 + ], + "score": 1.0, + "content": "approach can solve all of the training and test examples in task", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 626, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 286, + 640 + ], + "score": 1.0, + "content": "16 of the bAbI dataset (Weston et al., 2016).", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 495, + 357, + 640 + ] + }, + { + "type": "table", + "bbox": [ + 374, + 502, + 492, + 568 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 374, + 502, + 492, + 568 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 374, + 502, + 492, + 568 + ], + "spans": [ + { + "bbox": [ + 374, + 502, + 492, + 568 + ], + "score": 0.78, + "html": "
X:bernhard is a Y:frog Z:lily is a Y:frog
Z:lily is A:green
what colour is X:bernhard
A:green
", + "type": "table", + "image_path": "005e03f738687ce974ff771c18081cb699f106f6d11b9f8f18552de03cc0852f.jpg" + } + ] + } + ], + "index": 36.5, + "virtual_lines": [ + { + "bbox": [ + 374, + 502, + 492, + 535.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 374, + 535.0, + 492, + 568.0 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 363, + 578, + 505, + 644 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 363, + 578, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 363, + 578, + 505, + 590 + ], + "score": 1.0, + "content": "Figure 1: Invariant learned for", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 363, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 363, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "bAbI task 16, basic induction,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 363, + 600, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 363, + 600, + 391, + 612 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 392, + 600, + 400, + 610 + ], + "score": 0.54, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 600, + 505, + 612 + ], + "score": 1.0, + "content": ":bernhard denotes a vari-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 363, + 612, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 363, + 612, + 505, + 623 + ], + "score": 1.0, + "content": "able with default symbol bernhard.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 363, + 622, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 363, + 622, + 505, + 633 + ], + "score": 1.0, + "content": "This single invariant accounts for", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 363, + 632, + 501, + 646 + ], + "spans": [ + { + "bbox": [ + 363, + 632, + 501, + 646 + ], + "score": 1.0, + "content": "all the training and test examples.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 43.5 + } + ], + "index": 40.0 + }, + { + "type": "text", + "bbox": [ + 108, + 644, + 356, + 655 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 356, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 356, + 656 + ], + "score": 1.0, + "content": "In this paper we address the question of whether a machine can", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "learn and use the notion of a variable, i.e. a symbol that can take on different values. For instance,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 665, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 677 + ], + "score": 1.0, + "content": "given an example of the form “bernhard is a frog” the machine would learn that the token “bernhard”", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "could be someone else and the token “frog” could be something else. If we consider unification a", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "selection of the most appropriate value for a variable given a choice of values, we can reframe it as", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "a form of attention. Attention models (Bahdanau et al., 2015; Luong et al., 2015; Chaudhari et al.,", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "2019) allow neural networks to focus, attend to certain parts of the input often for the purpose of", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "selecting a relevant portion. Since attention mechanisms are also differentiable they are often jointly", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "learned within a task. This perspective motivates our idea of a unification mechanism that utilises", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 433, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 433, + 106 + ], + "score": 1.0, + "content": "attention and is therefore fully differentiable which we refer to as soft unification.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 48, + "bbox_fs": [ + 106, + 644, + 356, + 656 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 653, + 505, + 732 + ], + "lines": [], + "index": 52, + "bbox_fs": [ + 105, + 654, + 506, + 733 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "learned within a task. This perspective motivates our idea of a unification mechanism that utilises", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 433, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 433, + 106 + ], + "score": 1.0, + "content": "attention and is therefore fully differentiable which we refer to as soft unification.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "score": 1.0, + "content": "Hence, we propose an end-to-end differentiable neural network approach for learning and utilising", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "the notion of a variable that in return can lift examples into invariants used by the network to perform", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 505, + 146 + ], + "score": 1.0, + "content": "reasoning tasks. Specifically, we (i) propose a novel architecture capable of learning and using", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "variables by lifting a given example through soft unification, (ii) present the empirical results of our", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 168 + ], + "score": 1.0, + "content": "approach on four datasets and (iii) analyse the learned invariants that capture the underlying patterns", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "present in the tasks. Our implementation using Chainer (Tokui et al., 2015) is publicly available at", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 108, + 177, + 440, + 188 + ], + "spans": [ + { + "bbox": [ + 108, + 177, + 440, + 188 + ], + "score": 1.0, + "content": "[link removed](anonymous link provided with submission).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 108, + 203, + 225, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 201, + 226, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 226, + 218 + ], + "score": 1.0, + "content": "2 SOFT UNIFICATION", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 228, + 505, + 272 + ], + "lines": [ + { + "bbox": [ + 105, + 228, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 505, + 240 + ], + "score": 1.0, + "content": "Reasoning with variables involves identifying what variables are, the setting in which they are used", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 239, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 505, + 251 + ], + "score": 1.0, + "content": "as well as the process by which they are assigned values. When the varying components, i.e. vari-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 250, + 504, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 504, + 262 + ], + "score": 1.0, + "content": "ables, of an example are identified, the remaining structure can be lifted into an invariant which then", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 262, + 258, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 258, + 273 + ], + "score": 1.0, + "content": "accounts for multiple other instances.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 276, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 295, + 289 + ], + "score": 1.0, + "content": "Definition 1 (Variable). Given a set of symbols", + "type": "text" + }, + { + "bbox": [ + 296, + 277, + 302, + 286 + ], + "score": 0.64, + "content": "\\mathbb { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 276, + 346, + 289 + ], + "score": 1.0, + "content": ", a variable", + "type": "text" + }, + { + "bbox": [ + 346, + 277, + 355, + 286 + ], + "score": 0.81, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 276, + 430, + 289 + ], + "score": 1.0, + "content": "is defined as a pair", + "type": "text" + }, + { + "bbox": [ + 430, + 276, + 477, + 289 + ], + "score": 0.92, + "content": "\\pmb { X } \\triangleq \\left( \\mathbf { x } , s _ { d } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 107, + 287, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 107, + 288, + 135, + 299 + ], + "score": 0.89, + "content": "s _ { d } \\in \\mathbb { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 287, + 297, + 300 + ], + "score": 1.0, + "content": "is the default symbol of the variable and", + "type": "text" + }, + { + "bbox": [ + 297, + 290, + 303, + 298 + ], + "score": 0.71, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 287, + 505, + 300 + ], + "score": 1.0, + "content": "is a discrete random variable of which the support", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 297, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 104, + 297, + 115, + 312 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 116, + 299, + 123, + 308 + ], + "score": 0.46, + "content": "\\mathbb { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 123, + 297, + 257, + 312 + ], + "score": 1.0, + "content": ". The representation of a variable", + "type": "text" + }, + { + "bbox": [ + 257, + 299, + 298, + 311 + ], + "score": 0.92, + "content": "\\phi _ { V } ( { \\pmb X } ; s _ { d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 297, + 505, + 312 + ], + "score": 1.0, + "content": "is equal to the expected value of the corresponding", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 309, + 297, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 173, + 323 + ], + "score": 1.0, + "content": "random variable", + "type": "text" + }, + { + "bbox": [ + 174, + 311, + 180, + 319 + ], + "score": 0.66, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 309, + 282, + 323 + ], + "score": 1.0, + "content": "given the default symbol", + "type": "text" + }, + { + "bbox": [ + 282, + 311, + 292, + 321 + ], + "score": 0.84, + "content": "s _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 309, + 297, + 323 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 325, + 407, + 352 + ], + "lines": [ + { + "bbox": [ + 203, + 325, + 407, + 352 + ], + "spans": [ + { + "bbox": [ + 203, + 325, + 407, + 352 + ], + "score": 0.93, + "content": "\\phi _ { V } ( X { : } s _ { d } ) = \\mathbb { E } _ { \\mathbf { x } \\sim P } [ \\phi ( x ) ] = \\sum _ { x \\in \\mathbb { S } } P ( \\mathbf { x } = x | s _ { d } ) \\phi ( x )", + "type": "interline_equation", + "image_path": "e6f00329572399e3f3f44e676dabe4812b9f6a7696b28c59eebe9ea3a3fc9948.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 203, + 325, + 407, + 338.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 203, + 338.5, + 407, + 352.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 356, + 394, + 368 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 391, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 133, + 370 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 356, + 182, + 368 + ], + "score": 0.93, + "content": "\\phi : \\mathbb { S } \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 354, + 199, + 370 + ], + "score": 1.0, + "content": "is a", + "type": "text" + }, + { + "bbox": [ + 199, + 357, + 205, + 366 + ], + "score": 0.81, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 354, + 385, + 370 + ], + "score": 1.0, + "content": "-dimensional real-valued feature of a symbol", + "type": "text" + }, + { + "bbox": [ + 386, + 359, + 391, + 366 + ], + "score": 0.73, + "content": "s", + "type": "inline_equation" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 376, + 505, + 432 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 163, + 390 + ], + "score": 1.0, + "content": "For example,", + "type": "text" + }, + { + "bbox": [ + 163, + 378, + 171, + 389 + ], + "score": 0.83, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 376, + 288, + 390 + ], + "score": 1.0, + "content": "could be an embedding and", + "type": "text" + }, + { + "bbox": [ + 289, + 377, + 329, + 389 + ], + "score": 0.92, + "content": "\\phi _ { V } ( { \\pmb X } ; s _ { d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 376, + 505, + 390 + ], + "score": 1.0, + "content": "would become a weighted sum of symbol", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 388, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 505, + 400 + ], + "score": 1.0, + "content": "embeddings as in conventional attention models. The default symbol of a variable is intended to", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 398, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 506, + 411 + ], + "score": 1.0, + "content": "capture the variable’s bound meaning following the idea of referants by Frege (1948). We denote", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 409, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 169, + 423 + ], + "score": 1.0, + "content": "variables using", + "type": "text" + }, + { + "bbox": [ + 169, + 410, + 201, + 421 + ], + "score": 0.66, + "content": "X , Y , A", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 409, + 254, + 423 + ], + "score": 1.0, + "content": "etc. such as", + "type": "text" + }, + { + "bbox": [ + 254, + 410, + 262, + 420 + ], + "score": 0.77, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 409, + 329, + 423 + ], + "score": 1.0, + "content": ":bernhard where", + "type": "text" + }, + { + "bbox": [ + 329, + 410, + 338, + 420 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 409, + 505, + 423 + ], + "score": 1.0, + "content": "is the name of the variable and bernhard", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 421, + 271, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 271, + 433 + ], + "score": 1.0, + "content": "the default symbol as shown in Figure 1.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 436, + 505, + 481 + ], + "lines": [ + { + "bbox": [ + 104, + 434, + 504, + 450 + ], + "spans": [ + { + "bbox": [ + 104, + 434, + 336, + 450 + ], + "score": 1.0, + "content": "Definition 2 (Invariant). Given a structure (e.g. list, grid)", + "type": "text" + }, + { + "bbox": [ + 336, + 437, + 344, + 447 + ], + "score": 0.83, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 434, + 364, + 450 + ], + "score": 1.0, + "content": "over", + "type": "text" + }, + { + "bbox": [ + 365, + 437, + 372, + 446 + ], + "score": 0.5, + "content": "\\mathbb { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 434, + 457, + 450 + ], + "score": 1.0, + "content": ", an invariant is a pair", + "type": "text" + }, + { + "bbox": [ + 457, + 435, + 504, + 448 + ], + "score": 0.89, + "content": "I \\triangleq ( G , \\psi )", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 133, + 459 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 448, + 161, + 458 + ], + "score": 0.91, + "content": "G \\in { \\mathcal { G } }", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 447, + 469, + 459 + ], + "score": 1.0, + "content": "is the invariant example such as a tokenised story with tokens as symbols and", + "type": "text" + }, + { + "bbox": [ + 470, + 448, + 505, + 459 + ], + "score": 0.87, + "content": "\\psi : \\mathbb { S } \\to \\quad", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 458, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 107, + 458, + 127, + 470 + ], + "score": 0.46, + "content": "[ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 458, + 505, + 471 + ], + "score": 1.0, + "content": "is a function representing the degree to which the symbol is considered a variable. Thus, the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 469, + 340, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 235, + 482 + ], + "score": 1.0, + "content": "final representation of a symbol", + "type": "text" + }, + { + "bbox": [ + 235, + 471, + 241, + 479 + ], + "score": 0.75, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 469, + 289, + 482 + ], + "score": 1.0, + "content": "included in", + "type": "text" + }, + { + "bbox": [ + 289, + 470, + 298, + 479 + ], + "score": 0.65, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 469, + 302, + 482 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 302, + 469, + 326, + 481 + ], + "score": 0.77, + "content": "\\phi _ { I } ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 469, + 340, + 482 + ], + "score": 1.0, + "content": "is:", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + }, + { + "type": "interline_equation", + "bbox": [ + 221, + 484, + 389, + 498 + ], + "lines": [ + { + "bbox": [ + 221, + 484, + 389, + 498 + ], + "spans": [ + { + "bbox": [ + 221, + 484, + 389, + 498 + ], + "score": 0.91, + "content": "\\phi _ { I } ( s ) = ( 1 - \\psi ( s ) ) \\phi ( s ) + \\psi ( s ) \\phi _ { V } ( X { : } s )", + "type": "interline_equation", + "image_path": "2057226d07d21614572dc84a4398bd6f444905435e014f404eb0581291ecae49.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 221, + 484, + 389, + 498 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 501, + 503, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 310, + 515 + ], + "score": 1.0, + "content": "the linear interpolation between its representation", + "type": "text" + }, + { + "bbox": [ + 311, + 501, + 330, + 514 + ], + "score": 0.92, + "content": "\\phi ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 500, + 505, + 515 + ], + "score": 1.0, + "content": "and its variable bound value with itself as", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 511, + 223, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 183, + 526 + ], + "score": 1.0, + "content": "the default symbol", + "type": "text" + }, + { + "bbox": [ + 183, + 513, + 219, + 525 + ], + "score": 0.93, + "content": "\\phi _ { V } ( \\pmb { X } ; s )", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 511, + 223, + 526 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 532, + 505, + 566 + ], + "lines": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "We adhere to the term invariant and refrain from mentioning rules, unground rules, etc. used in logic-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 544, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 505, + 555 + ], + "score": 1.0, + "content": "based formalisms, e.g. Muggleton & de Raedt (1994), since neither the invariant structure needs to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 555, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 482, + 566 + ], + "score": 1.0, + "content": "be rule-like nor the variables carry logical semantics. This distinction is clarified in Section 6.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 569, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 300, + 581 + ], + "score": 1.0, + "content": "Definition 3 (Unification). Given an invariant", + "type": "text" + }, + { + "bbox": [ + 301, + 570, + 308, + 579 + ], + "score": 0.77, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 569, + 379, + 581 + ], + "score": 1.0, + "content": "and an example", + "type": "text" + }, + { + "bbox": [ + 379, + 569, + 413, + 580 + ], + "score": 0.9, + "content": "K \\in \\mathcal G", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 569, + 505, + 581 + ], + "score": 1.0, + "content": ", unification binds the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 580, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 155, + 592 + ], + "score": 1.0, + "content": "variables in", + "type": "text" + }, + { + "bbox": [ + 156, + 581, + 163, + 590 + ], + "score": 0.8, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 580, + 221, + 592 + ], + "score": 1.0, + "content": "to symbols in", + "type": "text" + }, + { + "bbox": [ + 222, + 581, + 232, + 590 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 580, + 327, + 592 + ], + "score": 1.0, + "content": ". Defined as a function", + "type": "text" + }, + { + "bbox": [ + 327, + 580, + 392, + 591 + ], + "score": 0.9, + "content": "g : \\mathcal { T } \\times \\mathcal { G } \\to \\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 580, + 505, + 592 + ], + "score": 1.0, + "content": ", unification binds variables", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 288, + 603 + ], + "score": 1.0, + "content": "by computing the probability mass functions,", + "type": "text" + }, + { + "bbox": [ + 289, + 591, + 297, + 601 + ], + "score": 0.83, + "content": "P", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "in equation 1, and returns the unified representation", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 602, + 387, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 352, + 614 + ], + "score": 1.0, + "content": "using equation 2. The probability mass function of a variable", + "type": "text" + }, + { + "bbox": [ + 352, + 603, + 372, + 613 + ], + "score": 0.84, + "content": "X { : } s _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 602, + 387, + 614 + ], + "score": 1.0, + "content": "is:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + }, + { + "type": "interline_equation", + "bbox": [ + 179, + 617, + 432, + 632 + ], + "lines": [ + { + "bbox": [ + 179, + 617, + 432, + 632 + ], + "spans": [ + { + "bbox": [ + 179, + 617, + 432, + 632 + ], + "score": 0.88, + "content": "P ( { \\bf x } = x | s _ { d } ) = \\mathrm { s o f t m a x } ( \\phi _ { U } ( s _ { d } ) \\phi _ { U } ( K ) ^ { T } ) , \\ : K = \\mathrm { s u p p o r t \\ o f \\ x }", + "type": "interline_equation", + "image_path": "3084e894930ef9d610b90cd7e1f867b73f6c7f52c81c81e58d5660b31df3b2ba.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 179, + 617, + 432, + 632 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 641, + 504, + 666 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 133, + 657 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 642, + 191, + 654 + ], + "score": 0.92, + "content": "\\phi _ { U } : \\mathbb { S } \\to \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 639, + 353, + 657 + ], + "score": 1.0, + "content": "is the unifying feature of a symbol and", + "type": "text" + }, + { + "bbox": [ + 353, + 641, + 428, + 655 + ], + "score": 0.92, + "content": "\\phi _ { U } ( K ) \\in \\mathbb { R } ^ { | K | \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 639, + 506, + 657 + ], + "score": 1.0, + "content": "is applied element", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 653, + 420, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 183, + 666 + ], + "score": 1.0, + "content": "wise to symbols in", + "type": "text" + }, + { + "bbox": [ + 184, + 654, + 194, + 664 + ], + "score": 0.81, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 653, + 207, + 666 + ], + "score": 1.0, + "content": ". If", + "type": "text" + }, + { + "bbox": [ + 208, + 655, + 214, + 666 + ], + "score": 0.82, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 653, + 420, + 666 + ], + "score": 1.0, + "content": "is differentiable, it is referred to as soft unification.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 670, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 670, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 167, + 684 + ], + "score": 1.0, + "content": "We distinguish", + "type": "text" + }, + { + "bbox": [ + 167, + 671, + 174, + 682 + ], + "score": 0.8, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 670, + 196, + 684 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 196, + 671, + 210, + 682 + ], + "score": 0.81, + "content": "\\phi _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 670, + 505, + 684 + ], + "score": 1.0, + "content": "to emphasise that the unifying properties of the symbols might be different", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 278, + 694 + ], + "score": 1.0, + "content": "from their representations. For example,", + "type": "text" + }, + { + "bbox": [ + 279, + 682, + 285, + 693 + ], + "score": 0.75, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "(bernhard) could represent a specific person whereas", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 107, + 693, + 259, + 706 + ], + "spans": [ + { + "bbox": [ + 107, + 694, + 120, + 705 + ], + "score": 0.85, + "content": "\\phi _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 693, + 259, + 706 + ], + "score": 1.0, + "content": "(bernhard) the notion of someone.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 357, + 722 + ], + "score": 1.0, + "content": "Overall soft unification incorporates 3 learnable components:", + "type": "text" + }, + { + "bbox": [ + 357, + 710, + 393, + 721 + ], + "score": 0.91, + "content": "\\phi , \\psi , \\phi _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "which denote the base fea-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "tures, variableness and unifying features of a symbol respectively. Given an upstream, potentially", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "score": 1.0, + "content": "Hence, we propose an end-to-end differentiable neural network approach for learning and utilising", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "the notion of a variable that in return can lift examples into invariants used by the network to perform", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 505, + 146 + ], + "score": 1.0, + "content": "reasoning tasks. Specifically, we (i) propose a novel architecture capable of learning and using", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "variables by lifting a given example through soft unification, (ii) present the empirical results of our", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 168 + ], + "score": 1.0, + "content": "approach on four datasets and (iii) analyse the learned invariants that capture the underlying patterns", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "present in the tasks. Our implementation using Chainer (Tokui et al., 2015) is publicly available at", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 108, + 177, + 440, + 188 + ], + "spans": [ + { + "bbox": [ + 108, + 177, + 440, + 188 + ], + "score": 1.0, + "content": "[link removed](anonymous link provided with submission).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 109, + 506, + 188 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 203, + 225, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 201, + 226, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 226, + 218 + ], + "score": 1.0, + "content": "2 SOFT UNIFICATION", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 228, + 505, + 272 + ], + "lines": [ + { + "bbox": [ + 105, + 228, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 505, + 240 + ], + "score": 1.0, + "content": "Reasoning with variables involves identifying what variables are, the setting in which they are used", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 239, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 505, + 251 + ], + "score": 1.0, + "content": "as well as the process by which they are assigned values. When the varying components, i.e. vari-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 250, + 504, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 504, + 262 + ], + "score": 1.0, + "content": "ables, of an example are identified, the remaining structure can be lifted into an invariant which then", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 262, + 258, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 258, + 273 + ], + "score": 1.0, + "content": "accounts for multiple other instances.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 228, + 505, + 273 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 276, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 295, + 289 + ], + "score": 1.0, + "content": "Definition 1 (Variable). Given a set of symbols", + "type": "text" + }, + { + "bbox": [ + 296, + 277, + 302, + 286 + ], + "score": 0.64, + "content": "\\mathbb { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 276, + 346, + 289 + ], + "score": 1.0, + "content": ", a variable", + "type": "text" + }, + { + "bbox": [ + 346, + 277, + 355, + 286 + ], + "score": 0.81, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 276, + 430, + 289 + ], + "score": 1.0, + "content": "is defined as a pair", + "type": "text" + }, + { + "bbox": [ + 430, + 276, + 477, + 289 + ], + "score": 0.92, + "content": "\\pmb { X } \\triangleq \\left( \\mathbf { x } , s _ { d } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 107, + 287, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 107, + 288, + 135, + 299 + ], + "score": 0.89, + "content": "s _ { d } \\in \\mathbb { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 287, + 297, + 300 + ], + "score": 1.0, + "content": "is the default symbol of the variable and", + "type": "text" + }, + { + "bbox": [ + 297, + 290, + 303, + 298 + ], + "score": 0.71, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 287, + 505, + 300 + ], + "score": 1.0, + "content": "is a discrete random variable of which the support", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 297, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 104, + 297, + 115, + 312 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 116, + 299, + 123, + 308 + ], + "score": 0.46, + "content": "\\mathbb { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 123, + 297, + 257, + 312 + ], + "score": 1.0, + "content": ". The representation of a variable", + "type": "text" + }, + { + "bbox": [ + 257, + 299, + 298, + 311 + ], + "score": 0.92, + "content": "\\phi _ { V } ( { \\pmb X } ; s _ { d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 297, + 505, + 312 + ], + "score": 1.0, + "content": "is equal to the expected value of the corresponding", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 309, + 297, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 173, + 323 + ], + "score": 1.0, + "content": "random variable", + "type": "text" + }, + { + "bbox": [ + 174, + 311, + 180, + 319 + ], + "score": 0.66, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 309, + 282, + 323 + ], + "score": 1.0, + "content": "given the default symbol", + "type": "text" + }, + { + "bbox": [ + 282, + 311, + 292, + 321 + ], + "score": 0.84, + "content": "s _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 309, + 297, + 323 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 104, + 276, + 505, + 323 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 325, + 407, + 352 + ], + "lines": [ + { + "bbox": [ + 203, + 325, + 407, + 352 + ], + "spans": [ + { + "bbox": [ + 203, + 325, + 407, + 352 + ], + "score": 0.93, + "content": "\\phi _ { V } ( X { : } s _ { d } ) = \\mathbb { E } _ { \\mathbf { x } \\sim P } [ \\phi ( x ) ] = \\sum _ { x \\in \\mathbb { S } } P ( \\mathbf { x } = x | s _ { d } ) \\phi ( x )", + "type": "interline_equation", + "image_path": "e6f00329572399e3f3f44e676dabe4812b9f6a7696b28c59eebe9ea3a3fc9948.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 203, + 325, + 407, + 338.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 203, + 338.5, + 407, + 352.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 356, + 394, + 368 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 391, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 133, + 370 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 356, + 182, + 368 + ], + "score": 0.93, + "content": "\\phi : \\mathbb { S } \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 354, + 199, + 370 + ], + "score": 1.0, + "content": "is a", + "type": "text" + }, + { + "bbox": [ + 199, + 357, + 205, + 366 + ], + "score": 0.81, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 354, + 385, + 370 + ], + "score": 1.0, + "content": "-dimensional real-valued feature of a symbol", + "type": "text" + }, + { + "bbox": [ + 386, + 359, + 391, + 366 + ], + "score": 0.73, + "content": "s", + "type": "inline_equation" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 354, + 391, + 370 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 376, + 505, + 432 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 163, + 390 + ], + "score": 1.0, + "content": "For example,", + "type": "text" + }, + { + "bbox": [ + 163, + 378, + 171, + 389 + ], + "score": 0.83, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 376, + 288, + 390 + ], + "score": 1.0, + "content": "could be an embedding and", + "type": "text" + }, + { + "bbox": [ + 289, + 377, + 329, + 389 + ], + "score": 0.92, + "content": "\\phi _ { V } ( { \\pmb X } ; s _ { d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 376, + 505, + 390 + ], + "score": 1.0, + "content": "would become a weighted sum of symbol", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 388, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 505, + 400 + ], + "score": 1.0, + "content": "embeddings as in conventional attention models. The default symbol of a variable is intended to", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 398, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 506, + 411 + ], + "score": 1.0, + "content": "capture the variable’s bound meaning following the idea of referants by Frege (1948). We denote", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 409, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 169, + 423 + ], + "score": 1.0, + "content": "variables using", + "type": "text" + }, + { + "bbox": [ + 169, + 410, + 201, + 421 + ], + "score": 0.66, + "content": "X , Y , A", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 409, + 254, + 423 + ], + "score": 1.0, + "content": "etc. such as", + "type": "text" + }, + { + "bbox": [ + 254, + 410, + 262, + 420 + ], + "score": 0.77, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 409, + 329, + 423 + ], + "score": 1.0, + "content": ":bernhard where", + "type": "text" + }, + { + "bbox": [ + 329, + 410, + 338, + 420 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 409, + 505, + 423 + ], + "score": 1.0, + "content": "is the name of the variable and bernhard", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 421, + 271, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 271, + 433 + ], + "score": 1.0, + "content": "the default symbol as shown in Figure 1.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 376, + 506, + 433 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 436, + 505, + 481 + ], + "lines": [ + { + "bbox": [ + 104, + 434, + 504, + 450 + ], + "spans": [ + { + "bbox": [ + 104, + 434, + 336, + 450 + ], + "score": 1.0, + "content": "Definition 2 (Invariant). Given a structure (e.g. list, grid)", + "type": "text" + }, + { + "bbox": [ + 336, + 437, + 344, + 447 + ], + "score": 0.83, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 434, + 364, + 450 + ], + "score": 1.0, + "content": "over", + "type": "text" + }, + { + "bbox": [ + 365, + 437, + 372, + 446 + ], + "score": 0.5, + "content": "\\mathbb { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 434, + 457, + 450 + ], + "score": 1.0, + "content": ", an invariant is a pair", + "type": "text" + }, + { + "bbox": [ + 457, + 435, + 504, + 448 + ], + "score": 0.89, + "content": "I \\triangleq ( G , \\psi )", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 133, + 459 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 448, + 161, + 458 + ], + "score": 0.91, + "content": "G \\in { \\mathcal { G } }", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 447, + 469, + 459 + ], + "score": 1.0, + "content": "is the invariant example such as a tokenised story with tokens as symbols and", + "type": "text" + }, + { + "bbox": [ + 470, + 448, + 505, + 459 + ], + "score": 0.87, + "content": "\\psi : \\mathbb { S } \\to \\quad", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 458, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 107, + 458, + 127, + 470 + ], + "score": 0.46, + "content": "[ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 458, + 505, + 471 + ], + "score": 1.0, + "content": "is a function representing the degree to which the symbol is considered a variable. Thus, the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 469, + 340, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 235, + 482 + ], + "score": 1.0, + "content": "final representation of a symbol", + "type": "text" + }, + { + "bbox": [ + 235, + 471, + 241, + 479 + ], + "score": 0.75, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 469, + 289, + 482 + ], + "score": 1.0, + "content": "included in", + "type": "text" + }, + { + "bbox": [ + 289, + 470, + 298, + 479 + ], + "score": 0.65, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 469, + 302, + 482 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 302, + 469, + 326, + 481 + ], + "score": 0.77, + "content": "\\phi _ { I } ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 469, + 340, + 482 + ], + "score": 1.0, + "content": "is:", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5, + "bbox_fs": [ + 104, + 434, + 505, + 482 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 221, + 484, + 389, + 498 + ], + "lines": [ + { + "bbox": [ + 221, + 484, + 389, + 498 + ], + "spans": [ + { + "bbox": [ + 221, + 484, + 389, + 498 + ], + "score": 0.91, + "content": "\\phi _ { I } ( s ) = ( 1 - \\psi ( s ) ) \\phi ( s ) + \\psi ( s ) \\phi _ { V } ( X { : } s )", + "type": "interline_equation", + "image_path": "2057226d07d21614572dc84a4398bd6f444905435e014f404eb0581291ecae49.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 221, + 484, + 389, + 498 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 501, + 503, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 310, + 515 + ], + "score": 1.0, + "content": "the linear interpolation between its representation", + "type": "text" + }, + { + "bbox": [ + 311, + 501, + 330, + 514 + ], + "score": 0.92, + "content": "\\phi ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 500, + 505, + 515 + ], + "score": 1.0, + "content": "and its variable bound value with itself as", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 511, + 223, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 183, + 526 + ], + "score": 1.0, + "content": "the default symbol", + "type": "text" + }, + { + "bbox": [ + 183, + 513, + 219, + 525 + ], + "score": 0.93, + "content": "\\phi _ { V } ( \\pmb { X } ; s )", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 511, + 223, + 526 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 500, + 505, + 526 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 532, + 505, + 566 + ], + "lines": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "We adhere to the term invariant and refrain from mentioning rules, unground rules, etc. used in logic-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 544, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 505, + 555 + ], + "score": 1.0, + "content": "based formalisms, e.g. Muggleton & de Raedt (1994), since neither the invariant structure needs to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 555, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 482, + 566 + ], + "score": 1.0, + "content": "be rule-like nor the variables carry logical semantics. This distinction is clarified in Section 6.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 106, + 532, + 505, + 566 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 569, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 300, + 581 + ], + "score": 1.0, + "content": "Definition 3 (Unification). Given an invariant", + "type": "text" + }, + { + "bbox": [ + 301, + 570, + 308, + 579 + ], + "score": 0.77, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 569, + 379, + 581 + ], + "score": 1.0, + "content": "and an example", + "type": "text" + }, + { + "bbox": [ + 379, + 569, + 413, + 580 + ], + "score": 0.9, + "content": "K \\in \\mathcal G", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 569, + 505, + 581 + ], + "score": 1.0, + "content": ", unification binds the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 580, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 155, + 592 + ], + "score": 1.0, + "content": "variables in", + "type": "text" + }, + { + "bbox": [ + 156, + 581, + 163, + 590 + ], + "score": 0.8, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 580, + 221, + 592 + ], + "score": 1.0, + "content": "to symbols in", + "type": "text" + }, + { + "bbox": [ + 222, + 581, + 232, + 590 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 580, + 327, + 592 + ], + "score": 1.0, + "content": ". Defined as a function", + "type": "text" + }, + { + "bbox": [ + 327, + 580, + 392, + 591 + ], + "score": 0.9, + "content": "g : \\mathcal { T } \\times \\mathcal { G } \\to \\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 580, + 505, + 592 + ], + "score": 1.0, + "content": ", unification binds variables", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 288, + 603 + ], + "score": 1.0, + "content": "by computing the probability mass functions,", + "type": "text" + }, + { + "bbox": [ + 289, + 591, + 297, + 601 + ], + "score": 0.83, + "content": "P", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "in equation 1, and returns the unified representation", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 602, + 387, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 352, + 614 + ], + "score": 1.0, + "content": "using equation 2. The probability mass function of a variable", + "type": "text" + }, + { + "bbox": [ + 352, + 603, + 372, + 613 + ], + "score": 0.84, + "content": "X { : } s _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 602, + 387, + 614 + ], + "score": 1.0, + "content": "is:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 569, + 505, + 614 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 179, + 617, + 432, + 632 + ], + "lines": [ + { + "bbox": [ + 179, + 617, + 432, + 632 + ], + "spans": [ + { + "bbox": [ + 179, + 617, + 432, + 632 + ], + "score": 0.88, + "content": "P ( { \\bf x } = x | s _ { d } ) = \\mathrm { s o f t m a x } ( \\phi _ { U } ( s _ { d } ) \\phi _ { U } ( K ) ^ { T } ) , \\ : K = \\mathrm { s u p p o r t \\ o f \\ x }", + "type": "interline_equation", + "image_path": "3084e894930ef9d610b90cd7e1f867b73f6c7f52c81c81e58d5660b31df3b2ba.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 179, + 617, + 432, + 632 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 641, + 504, + 666 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 133, + 657 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 642, + 191, + 654 + ], + "score": 0.92, + "content": "\\phi _ { U } : \\mathbb { S } \\to \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 639, + 353, + 657 + ], + "score": 1.0, + "content": "is the unifying feature of a symbol and", + "type": "text" + }, + { + "bbox": [ + 353, + 641, + 428, + 655 + ], + "score": 0.92, + "content": "\\phi _ { U } ( K ) \\in \\mathbb { R } ^ { | K | \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 639, + 506, + 657 + ], + "score": 1.0, + "content": "is applied element", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 653, + 420, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 183, + 666 + ], + "score": 1.0, + "content": "wise to symbols in", + "type": "text" + }, + { + "bbox": [ + 184, + 654, + 194, + 664 + ], + "score": 0.81, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 653, + 207, + 666 + ], + "score": 1.0, + "content": ". If", + "type": "text" + }, + { + "bbox": [ + 208, + 655, + 214, + 666 + ], + "score": 0.82, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 653, + 420, + 666 + ], + "score": 1.0, + "content": "is differentiable, it is referred to as soft unification.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 639, + 506, + 666 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 670, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 670, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 167, + 684 + ], + "score": 1.0, + "content": "We distinguish", + "type": "text" + }, + { + "bbox": [ + 167, + 671, + 174, + 682 + ], + "score": 0.8, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 670, + 196, + 684 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 196, + 671, + 210, + 682 + ], + "score": 0.81, + "content": "\\phi _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 670, + 505, + 684 + ], + "score": 1.0, + "content": "to emphasise that the unifying properties of the symbols might be different", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 278, + 694 + ], + "score": 1.0, + "content": "from their representations. For example,", + "type": "text" + }, + { + "bbox": [ + 279, + 682, + 285, + 693 + ], + "score": 0.75, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "(bernhard) could represent a specific person whereas", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 107, + 693, + 259, + 706 + ], + "spans": [ + { + "bbox": [ + 107, + 694, + 120, + 705 + ], + "score": 0.85, + "content": "\\phi _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 693, + 259, + 706 + ], + "score": 1.0, + "content": "(bernhard) the notion of someone.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 670, + 505, + 706 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 357, + 722 + ], + "score": 1.0, + "content": "Overall soft unification incorporates 3 learnable components:", + "type": "text" + }, + { + "bbox": [ + 357, + 710, + 393, + 721 + ], + "score": 0.91, + "content": "\\phi , \\psi , \\phi _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "which denote the base fea-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "tures, variableness and unifying features of a symbol respectively. Given an upstream, potentially", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 709, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 197, + 96 + ], + "score": 1.0, + "content": "task specific, network", + "type": "text" + }, + { + "bbox": [ + 198, + 83, + 245, + 94 + ], + "score": 0.91, + "content": "f : \\mathcal G \\to \\mathbb S", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 81, + 299, + 96 + ], + "score": 1.0, + "content": ", an invariant", + "type": "text" + }, + { + "bbox": [ + 299, + 83, + 326, + 93 + ], + "score": 0.91, + "content": "I \\in \\mathcal { Z }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 81, + 417, + 96 + ], + "score": 1.0, + "content": "and an input example", + "type": "text" + }, + { + "bbox": [ + 418, + 83, + 449, + 93 + ], + "score": 0.91, + "content": "K \\in { \\mathcal { G } }", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "with a corre-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 315, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 204, + 106 + ], + "score": 1.0, + "content": "sponding desired output", + "type": "text" + }, + { + "bbox": [ + 205, + 94, + 229, + 104 + ], + "score": 0.89, + "content": "a \\in \\mathbb { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 93, + 315, + 106 + ], + "score": 1.0, + "content": ", the following holds:", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "interline_equation", + "bbox": [ + 253, + 110, + 358, + 124 + ], + "lines": [ + { + "bbox": [ + 253, + 110, + 358, + 124 + ], + "spans": [ + { + "bbox": [ + 253, + 110, + 358, + 124 + ], + "score": 0.93, + "content": "f \\circ g ( I , K ) = f ( K ) = a", + "type": "interline_equation", + "image_path": "990f4d6f7bb15d522906b6e6ae32537c039f64d1b9e23568ff08005eaf16924c.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 253, + 110, + 358, + 124 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 129, + 504, + 153 + ], + "lines": [ + { + "bbox": [ + 105, + 128, + 506, + 143 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 133, + 143 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 130, + 141, + 141 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 128, + 389, + 143 + ], + "score": 1.0, + "content": "now predicts based on the unified representation produced by", + "type": "text" + }, + { + "bbox": [ + 390, + 132, + 396, + 141 + ], + "score": 0.77, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 128, + 506, + 143 + ], + "score": 1.0, + "content": ". In this work, we focus on", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 140, + 371, + 154 + ], + "spans": [ + { + "bbox": [ + 107, + 144, + 113, + 153 + ], + "score": 0.71, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 140, + 345, + 154 + ], + "score": 1.0, + "content": ", the invariants it produces together with the interaction of", + "type": "text" + }, + { + "bbox": [ + 345, + 141, + 366, + 153 + ], + "score": 0.92, + "content": "f \\circ g", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 140, + 371, + 154 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 108, + 168, + 257, + 181 + ], + "lines": [ + { + "bbox": [ + 104, + 167, + 258, + 183 + ], + "spans": [ + { + "bbox": [ + 104, + 167, + 258, + 183 + ], + "score": 1.0, + "content": "3 UNIFICATION NETWORKS", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 193, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 105, + 193, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 505, + 207 + ], + "score": 1.0, + "content": "Since soft unification is end-to-end differentiable, it can be incorporated into existing task-specific", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 204, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 371, + 218 + ], + "score": 1.0, + "content": "upstream architectures. We present 3 architectures that model", + "type": "text" + }, + { + "bbox": [ + 371, + 205, + 396, + 217 + ], + "score": 0.92, + "content": "f \\circ g", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 204, + 505, + 218 + ], + "score": 1.0, + "content": "using multi-layer percep-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "trons (MLP), convolutional neural networks (CNN) and memory networks (Weston et al., 2015)", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 354, + 240 + ], + "score": 1.0, + "content": "to demonstrate the flexibility of our approach. In all cases, the", + "type": "text" + }, + { + "bbox": [ + 355, + 227, + 362, + 236 + ], + "score": 0.78, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 226, + 505, + 240 + ], + "score": 1.0, + "content": "dimensional representation of sym-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 236, + 504, + 252 + ], + "spans": [ + { + "bbox": [ + 104, + 236, + 232, + 252 + ], + "score": 1.0, + "content": "bols are learnable embeddings", + "type": "text" + }, + { + "bbox": [ + 232, + 238, + 300, + 251 + ], + "score": 0.93, + "content": "\\phi ( s ) = O [ s ] ^ { T } E", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 236, + 322, + 252 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 322, + 237, + 373, + 249 + ], + "score": 0.92, + "content": "E \\in \\mathbb { R } ^ { | \\mathbb { S } | \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 236, + 470, + 252 + ], + "score": 1.0, + "content": "randomly initialised by", + "type": "text" + }, + { + "bbox": [ + 471, + 238, + 504, + 250 + ], + "score": 0.92, + "content": "\\mathcal { N } ( 0 , 1 )", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 249, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 124, + 262 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 249, + 144, + 261 + ], + "score": 0.9, + "content": "O [ s ]", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 250, + 506, + 262 + ], + "score": 1.0, + "content": "the one-hot encoding of the symbol. The variableness of symbols is a learnable weight", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 259, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 107, + 262, + 171, + 274 + ], + "score": 0.92, + "content": "\\psi _ { w } ( s ) = \\sigma ( w _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 259, + 199, + 276 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 199, + 261, + 237, + 272 + ], + "score": 0.92, + "content": "\\boldsymbol { w } \\in \\mathbb { R } ^ { | \\mathbb { S } | }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 259, + 255, + 276 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 255, + 263, + 263, + 272 + ], + "score": 0.78, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 259, + 506, + 276 + ], + "score": 1.0, + "content": "is the sigmoid function. We consider every symbol indepen-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 273, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 505, + 285 + ], + "score": 1.0, + "content": "dently a variable irrespective of its surrounding context and leave further contextualised formulations", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 284, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 505, + 296 + ], + "score": 1.0, + "content": "as future work. The underlying intuition of this configuration is that a useful symbol for a correct", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 294, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 104, + 294, + 506, + 307 + ], + "score": 1.0, + "content": "prediction might need to take on other values for different inputs. This usefulness can be viewed", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 306, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 288, + 318 + ], + "score": 1.0, + "content": "as the inbound gradient to the corresponding", + "type": "text" + }, + { + "bbox": [ + 289, + 307, + 301, + 317 + ], + "score": 0.85, + "content": "w _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 306, + 362, + 318 + ], + "score": 1.0, + "content": "parameter and", + "type": "text" + }, + { + "bbox": [ + 362, + 306, + 389, + 317 + ], + "score": 0.92, + "content": "\\psi _ { w } ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 306, + 506, + 318 + ], + "score": 1.0, + "content": "acting as a gate. For further", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 317, + 427, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 427, + 330 + ], + "score": 1.0, + "content": "model details including the size of the embeddings, please refer to Appendix A.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 333, + 505, + 433 + ], + "lines": [ + { + "bbox": [ + 106, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 227, + 347 + ], + "score": 1.0, + "content": "Unification MLP (UMLP) (", + "type": "text" + }, + { + "bbox": [ + 227, + 334, + 234, + 345 + ], + "score": 0.6, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 333, + 267, + 347 + ], + "score": 1.0, + "content": ": MLP,", + "type": "text" + }, + { + "bbox": [ + 267, + 335, + 274, + 345 + ], + "score": 0.68, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 333, + 506, + 347 + ], + "score": 1.0, + "content": ": RNN) We combine soft unification into a multi-layer", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 345, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 389, + 357 + ], + "score": 1.0, + "content": "perceptron to process fixed length inputs. In this case, the structure", + "type": "text" + }, + { + "bbox": [ + 389, + 345, + 397, + 356 + ], + "score": 0.81, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 345, + 506, + 357 + ], + "score": 1.0, + "content": "is a sequence of symbols", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 354, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 104, + 354, + 187, + 368 + ], + "score": 1.0, + "content": "with a fixed length", + "type": "text" + }, + { + "bbox": [ + 187, + 356, + 192, + 366 + ], + "score": 0.6, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 354, + 443, + 368 + ], + "score": 1.0, + "content": ", e.g. a sequence of digits 4234. Given an embedded input", + "type": "text" + }, + { + "bbox": [ + 443, + 355, + 501, + 367 + ], + "score": 0.89, + "content": "\\phi ( \\pmb { k } ) \\in R ^ { l \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 354, + 506, + 368 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 365, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 457, + 380 + ], + "score": 1.0, + "content": "the upstream MLP computes the output symbol based on the flattened representations", + "type": "text" + }, + { + "bbox": [ + 458, + 367, + 505, + 379 + ], + "score": 0.88, + "content": "f ( \\phi ( \\pmb { k } ) ) =", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 375, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 104, + 375, + 142, + 392 + ], + "score": 1.0, + "content": "softmax", + "type": "text" + }, + { + "bbox": [ + 142, + 377, + 171, + 389 + ], + "score": 0.87, + "content": "( h E ^ { T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 375, + 201, + 392 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 201, + 377, + 236, + 388 + ], + "score": 0.92, + "content": "\\hat { h } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 375, + 506, + 392 + ], + "score": 1.0, + "content": "is the output of the last layer. However, to compute the unifying", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 388, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 140, + 401 + ], + "score": 1.0, + "content": "features", + "type": "text" + }, + { + "bbox": [ + 140, + 389, + 154, + 400 + ], + "score": 0.86, + "content": "\\phi _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 388, + 207, + 401 + ], + "score": 1.0, + "content": ", definition 3,", + "type": "text" + }, + { + "bbox": [ + 208, + 390, + 214, + 400 + ], + "score": 0.72, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 388, + 445, + 401 + ], + "score": 1.0, + "content": "uses a bi-directional GRU (Cho et al., 2014) running over", + "type": "text" + }, + { + "bbox": [ + 445, + 389, + 466, + 400 + ], + "score": 0.91, + "content": "\\phi ( k )", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 388, + 505, + 401 + ], + "score": 1.0, + "content": "such that", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 398, + 504, + 413 + ], + "spans": [ + { + "bbox": [ + 107, + 399, + 185, + 411 + ], + "score": 0.92, + "content": "\\phi _ { U } ( k ) = W _ { U } \\Phi ( k )", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 398, + 213, + 413 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 214, + 399, + 259, + 411 + ], + "score": 0.93, + "content": "\\Phi ( k ) \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 398, + 425, + 413 + ], + "score": 1.0, + "content": "is the hidden state of the GRU at symbol", + "type": "text" + }, + { + "bbox": [ + 425, + 400, + 432, + 409 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 398, + 451, + 413 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 451, + 399, + 504, + 410 + ], + "score": 0.92, + "content": "\\dot { W _ { U } } \\in \\mathbb { R } ^ { d \\times d }", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 410, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 450, + 423 + ], + "score": 1.0, + "content": "is a learnable parameter. This model emphasises the flexibility around the boundary of", + "type": "text" + }, + { + "bbox": [ + 450, + 411, + 470, + 422 + ], + "score": 0.91, + "content": "f \\circ g", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 410, + 506, + 423 + ], + "score": 1.0, + "content": "and that", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 420, + 379, + 435 + ], + "spans": [ + { + "bbox": [ + 104, + 420, + 379, + 435 + ], + "score": 1.0, + "content": "the unifying features can be computed in any differentiable manner.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 136, + 440, + 475, + 542 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 136, + 440, + 475, + 542 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 136, + 440, + 475, + 542 + ], + "spans": [ + { + "bbox": [ + 136, + 440, + 475, + 542 + ], + "score": 0.964, + "type": "image", + "image_path": "ef39205e27e246bc37f3a17ac72fc093d4914d8003b2bd66d967fd9856bb6ce9.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 136, + 440, + 475, + 474.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 136, + 474.0, + 475, + 508.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 136, + 508.0, + 475, + 542.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 555, + 505, + 588 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 555, + 504, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 504, + 568 + ], + "score": 1.0, + "content": "Figure 2: Graphical overview of soft unification within a memory network. Each sentence is pro-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 567, + 504, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 504, + 578 + ], + "score": 1.0, + "content": "cessed by two bi-directional RNNs for memory and unification. At each iteration the context atten-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 577, + 494, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 494, + 590 + ], + "score": 1.0, + "content": "tion selects which sentences to unify and the invariant produces the same answer as the example.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 593, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 227, + 607 + ], + "score": 1.0, + "content": "Unification CNN (UCNN) (", + "type": "text" + }, + { + "bbox": [ + 227, + 594, + 234, + 606 + ], + "score": 0.58, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 594, + 269, + 607 + ], + "score": 1.0, + "content": ": CNN,", + "type": "text" + }, + { + "bbox": [ + 270, + 596, + 276, + 606 + ], + "score": 0.69, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 594, + 465, + 607 + ], + "score": 1.0, + "content": ": CNN) Given a grid of embedded symbols", + "type": "text" + }, + { + "bbox": [ + 466, + 594, + 505, + 606 + ], + "score": 0.9, + "content": "\\phi ( K ) \\ \\in", + "type": "inline_equation" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 602, + 507, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 142, + 615 + ], + "score": 0.9, + "content": "\\mathbb { R } ^ { w \\times h \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 602, + 172, + 618 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 172, + 607, + 181, + 615 + ], + "score": 0.75, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 602, + 251, + 618 + ], + "score": 1.0, + "content": "is the width and", + "type": "text" + }, + { + "bbox": [ + 251, + 605, + 258, + 615 + ], + "score": 0.79, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 602, + 507, + 618 + ], + "score": 1.0, + "content": "the height, we use a convolutional neural network such that", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 193, + 628 + ], + "score": 1.0, + "content": "the final prediction is", + "type": "text" + }, + { + "bbox": [ + 193, + 615, + 345, + 628 + ], + "score": 0.89, + "content": "f ( \\phi ( K ) ) = \\mathrm { s o f t m a x } ( ( W h + b ) E ^ { T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 615, + 372, + 628 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 373, + 616, + 381, + 626 + ], + "score": 0.77, + "content": "^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "this time is the result of global", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 177, + 640 + ], + "score": 1.0, + "content": "max pooling and", + "type": "text" + }, + { + "bbox": [ + 178, + 627, + 201, + 638 + ], + "score": 0.82, + "content": "W , b", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 627, + 371, + 640 + ], + "score": 1.0, + "content": "are learnable parameters. We also model", + "type": "text" + }, + { + "bbox": [ + 371, + 629, + 378, + 638 + ], + "score": 0.8, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "using a separate convolutional", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 261, + 650 + ], + "score": 1.0, + "content": "network with the same architecture as", + "type": "text" + }, + { + "bbox": [ + 261, + 639, + 269, + 649 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 638, + 301, + 650 + ], + "score": 1.0, + "content": "and set", + "type": "text" + }, + { + "bbox": [ + 301, + 638, + 401, + 650 + ], + "score": 0.9, + "content": "\\phi _ { U } ( k ) = c _ { 2 } ( \\mathrm { r e l u } ( c _ { 1 } ( k ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 638, + 430, + 650 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 430, + 640, + 453, + 649 + ], + "score": 0.88, + "content": "c _ { 1 } , c _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "are the con-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 328, + 660 + ], + "score": 1.0, + "content": "volutional layers. The grid is padded with 0s to obtain", + "type": "text" + }, + { + "bbox": [ + 329, + 650, + 373, + 659 + ], + "score": 0.91, + "content": "w \\times h \\times d", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "after each convolution such that", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "every symbol has a unifying feature. This model conveys how soft unification can be adapted to the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 671, + 465, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 465, + 684 + ], + "score": 1.0, + "content": "specifics of the domain for example by using a convolution in a spatially structured input.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 277, + 700 + ], + "score": 1.0, + "content": "Unification Memory Networks (UMN) (", + "type": "text" + }, + { + "bbox": [ + 277, + 689, + 284, + 699 + ], + "score": 0.74, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 687, + 331, + 700 + ], + "score": 1.0, + "content": ": MemNN,", + "type": "text" + }, + { + "bbox": [ + 331, + 691, + 337, + 699 + ], + "score": 0.59, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 687, + 506, + 700 + ], + "score": 1.0, + "content": ": RNN) Soft unification does not need to", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 170, + 710 + ], + "score": 1.0, + "content": "happen prior to", + "type": "text" + }, + { + "bbox": [ + 171, + 699, + 178, + 711 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 700, + 196, + 710 + ], + "score": 1.0, + "content": "in a", + "type": "text" + }, + { + "bbox": [ + 196, + 699, + 218, + 711 + ], + "score": 0.91, + "content": "f \\circ g", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 700, + 505, + 710 + ], + "score": 1.0, + "content": "fashion but can also be incorporated at any intermediate stage multiple", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "times. To demonstrate this ability, we unify the symbols at different memory locations at each", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "iteration of a Memory Network (Weston et al., 2015). Memory networks can handle a list of lists", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 197, + 96 + ], + "score": 1.0, + "content": "task specific, network", + "type": "text" + }, + { + "bbox": [ + 198, + 83, + 245, + 94 + ], + "score": 0.91, + "content": "f : \\mathcal G \\to \\mathbb S", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 81, + 299, + 96 + ], + "score": 1.0, + "content": ", an invariant", + "type": "text" + }, + { + "bbox": [ + 299, + 83, + 326, + 93 + ], + "score": 0.91, + "content": "I \\in \\mathcal { Z }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 81, + 417, + 96 + ], + "score": 1.0, + "content": "and an input example", + "type": "text" + }, + { + "bbox": [ + 418, + 83, + 449, + 93 + ], + "score": 0.91, + "content": "K \\in { \\mathcal { G } }", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "with a corre-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 315, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 204, + 106 + ], + "score": 1.0, + "content": "sponding desired output", + "type": "text" + }, + { + "bbox": [ + 205, + 94, + 229, + 104 + ], + "score": 0.89, + "content": "a \\in \\mathbb { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 93, + 315, + 106 + ], + "score": 1.0, + "content": ", the following holds:", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 106 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 253, + 110, + 358, + 124 + ], + "lines": [ + { + "bbox": [ + 253, + 110, + 358, + 124 + ], + "spans": [ + { + "bbox": [ + 253, + 110, + 358, + 124 + ], + "score": 0.93, + "content": "f \\circ g ( I , K ) = f ( K ) = a", + "type": "interline_equation", + "image_path": "990f4d6f7bb15d522906b6e6ae32537c039f64d1b9e23568ff08005eaf16924c.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 253, + 110, + 358, + 124 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 129, + 504, + 153 + ], + "lines": [ + { + "bbox": [ + 105, + 128, + 506, + 143 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 133, + 143 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 130, + 141, + 141 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 128, + 389, + 143 + ], + "score": 1.0, + "content": "now predicts based on the unified representation produced by", + "type": "text" + }, + { + "bbox": [ + 390, + 132, + 396, + 141 + ], + "score": 0.77, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 128, + 506, + 143 + ], + "score": 1.0, + "content": ". In this work, we focus on", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 140, + 371, + 154 + ], + "spans": [ + { + "bbox": [ + 107, + 144, + 113, + 153 + ], + "score": 0.71, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 140, + 345, + 154 + ], + "score": 1.0, + "content": ", the invariants it produces together with the interaction of", + "type": "text" + }, + { + "bbox": [ + 345, + 141, + 366, + 153 + ], + "score": 0.92, + "content": "f \\circ g", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 140, + 371, + 154 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 128, + 506, + 154 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 168, + 257, + 181 + ], + "lines": [ + { + "bbox": [ + 104, + 167, + 258, + 183 + ], + "spans": [ + { + "bbox": [ + 104, + 167, + 258, + 183 + ], + "score": 1.0, + "content": "3 UNIFICATION NETWORKS", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 193, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 105, + 193, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 505, + 207 + ], + "score": 1.0, + "content": "Since soft unification is end-to-end differentiable, it can be incorporated into existing task-specific", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 204, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 371, + 218 + ], + "score": 1.0, + "content": "upstream architectures. We present 3 architectures that model", + "type": "text" + }, + { + "bbox": [ + 371, + 205, + 396, + 217 + ], + "score": 0.92, + "content": "f \\circ g", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 204, + 505, + 218 + ], + "score": 1.0, + "content": "using multi-layer percep-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "trons (MLP), convolutional neural networks (CNN) and memory networks (Weston et al., 2015)", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 354, + 240 + ], + "score": 1.0, + "content": "to demonstrate the flexibility of our approach. In all cases, the", + "type": "text" + }, + { + "bbox": [ + 355, + 227, + 362, + 236 + ], + "score": 0.78, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 226, + 505, + 240 + ], + "score": 1.0, + "content": "dimensional representation of sym-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 236, + 504, + 252 + ], + "spans": [ + { + "bbox": [ + 104, + 236, + 232, + 252 + ], + "score": 1.0, + "content": "bols are learnable embeddings", + "type": "text" + }, + { + "bbox": [ + 232, + 238, + 300, + 251 + ], + "score": 0.93, + "content": "\\phi ( s ) = O [ s ] ^ { T } E", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 236, + 322, + 252 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 322, + 237, + 373, + 249 + ], + "score": 0.92, + "content": "E \\in \\mathbb { R } ^ { | \\mathbb { S } | \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 236, + 470, + 252 + ], + "score": 1.0, + "content": "randomly initialised by", + "type": "text" + }, + { + "bbox": [ + 471, + 238, + 504, + 250 + ], + "score": 0.92, + "content": "\\mathcal { N } ( 0 , 1 )", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 249, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 124, + 262 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 249, + 144, + 261 + ], + "score": 0.9, + "content": "O [ s ]", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 250, + 506, + 262 + ], + "score": 1.0, + "content": "the one-hot encoding of the symbol. The variableness of symbols is a learnable weight", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 259, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 107, + 262, + 171, + 274 + ], + "score": 0.92, + "content": "\\psi _ { w } ( s ) = \\sigma ( w _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 259, + 199, + 276 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 199, + 261, + 237, + 272 + ], + "score": 0.92, + "content": "\\boldsymbol { w } \\in \\mathbb { R } ^ { | \\mathbb { S } | }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 259, + 255, + 276 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 255, + 263, + 263, + 272 + ], + "score": 0.78, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 259, + 506, + 276 + ], + "score": 1.0, + "content": "is the sigmoid function. We consider every symbol indepen-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 273, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 505, + 285 + ], + "score": 1.0, + "content": "dently a variable irrespective of its surrounding context and leave further contextualised formulations", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 284, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 505, + 296 + ], + "score": 1.0, + "content": "as future work. The underlying intuition of this configuration is that a useful symbol for a correct", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 294, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 104, + 294, + 506, + 307 + ], + "score": 1.0, + "content": "prediction might need to take on other values for different inputs. This usefulness can be viewed", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 306, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 288, + 318 + ], + "score": 1.0, + "content": "as the inbound gradient to the corresponding", + "type": "text" + }, + { + "bbox": [ + 289, + 307, + 301, + 317 + ], + "score": 0.85, + "content": "w _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 306, + 362, + 318 + ], + "score": 1.0, + "content": "parameter and", + "type": "text" + }, + { + "bbox": [ + 362, + 306, + 389, + 317 + ], + "score": 0.92, + "content": "\\psi _ { w } ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 306, + 506, + 318 + ], + "score": 1.0, + "content": "acting as a gate. For further", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 317, + 427, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 427, + 330 + ], + "score": 1.0, + "content": "model details including the size of the embeddings, please refer to Appendix A.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11.5, + "bbox_fs": [ + 104, + 193, + 506, + 330 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 333, + 505, + 433 + ], + "lines": [ + { + "bbox": [ + 106, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 227, + 347 + ], + "score": 1.0, + "content": "Unification MLP (UMLP) (", + "type": "text" + }, + { + "bbox": [ + 227, + 334, + 234, + 345 + ], + "score": 0.6, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 333, + 267, + 347 + ], + "score": 1.0, + "content": ": MLP,", + "type": "text" + }, + { + "bbox": [ + 267, + 335, + 274, + 345 + ], + "score": 0.68, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 333, + 506, + 347 + ], + "score": 1.0, + "content": ": RNN) We combine soft unification into a multi-layer", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 345, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 389, + 357 + ], + "score": 1.0, + "content": "perceptron to process fixed length inputs. In this case, the structure", + "type": "text" + }, + { + "bbox": [ + 389, + 345, + 397, + 356 + ], + "score": 0.81, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 345, + 506, + 357 + ], + "score": 1.0, + "content": "is a sequence of symbols", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 354, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 104, + 354, + 187, + 368 + ], + "score": 1.0, + "content": "with a fixed length", + "type": "text" + }, + { + "bbox": [ + 187, + 356, + 192, + 366 + ], + "score": 0.6, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 354, + 443, + 368 + ], + "score": 1.0, + "content": ", e.g. a sequence of digits 4234. Given an embedded input", + "type": "text" + }, + { + "bbox": [ + 443, + 355, + 501, + 367 + ], + "score": 0.89, + "content": "\\phi ( \\pmb { k } ) \\in R ^ { l \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 354, + 506, + 368 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 365, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 457, + 380 + ], + "score": 1.0, + "content": "the upstream MLP computes the output symbol based on the flattened representations", + "type": "text" + }, + { + "bbox": [ + 458, + 367, + 505, + 379 + ], + "score": 0.88, + "content": "f ( \\phi ( \\pmb { k } ) ) =", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 375, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 104, + 375, + 142, + 392 + ], + "score": 1.0, + "content": "softmax", + "type": "text" + }, + { + "bbox": [ + 142, + 377, + 171, + 389 + ], + "score": 0.87, + "content": "( h E ^ { T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 375, + 201, + 392 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 201, + 377, + 236, + 388 + ], + "score": 0.92, + "content": "\\hat { h } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 375, + 506, + 392 + ], + "score": 1.0, + "content": "is the output of the last layer. However, to compute the unifying", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 388, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 140, + 401 + ], + "score": 1.0, + "content": "features", + "type": "text" + }, + { + "bbox": [ + 140, + 389, + 154, + 400 + ], + "score": 0.86, + "content": "\\phi _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 388, + 207, + 401 + ], + "score": 1.0, + "content": ", definition 3,", + "type": "text" + }, + { + "bbox": [ + 208, + 390, + 214, + 400 + ], + "score": 0.72, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 388, + 445, + 401 + ], + "score": 1.0, + "content": "uses a bi-directional GRU (Cho et al., 2014) running over", + "type": "text" + }, + { + "bbox": [ + 445, + 389, + 466, + 400 + ], + "score": 0.91, + "content": "\\phi ( k )", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 388, + 505, + 401 + ], + "score": 1.0, + "content": "such that", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 398, + 504, + 413 + ], + "spans": [ + { + "bbox": [ + 107, + 399, + 185, + 411 + ], + "score": 0.92, + "content": "\\phi _ { U } ( k ) = W _ { U } \\Phi ( k )", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 398, + 213, + 413 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 214, + 399, + 259, + 411 + ], + "score": 0.93, + "content": "\\Phi ( k ) \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 398, + 425, + 413 + ], + "score": 1.0, + "content": "is the hidden state of the GRU at symbol", + "type": "text" + }, + { + "bbox": [ + 425, + 400, + 432, + 409 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 398, + 451, + 413 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 451, + 399, + 504, + 410 + ], + "score": 0.92, + "content": "\\dot { W _ { U } } \\in \\mathbb { R } ^ { d \\times d }", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 410, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 450, + 423 + ], + "score": 1.0, + "content": "is a learnable parameter. This model emphasises the flexibility around the boundary of", + "type": "text" + }, + { + "bbox": [ + 450, + 411, + 470, + 422 + ], + "score": 0.91, + "content": "f \\circ g", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 410, + 506, + 423 + ], + "score": 1.0, + "content": "and that", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 420, + 379, + 435 + ], + "spans": [ + { + "bbox": [ + 104, + 420, + 379, + 435 + ], + "score": 1.0, + "content": "the unifying features can be computed in any differentiable manner.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22, + "bbox_fs": [ + 104, + 333, + 506, + 435 + ] + }, + { + "type": "image", + "bbox": [ + 136, + 440, + 475, + 542 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 136, + 440, + 475, + 542 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 136, + 440, + 475, + 542 + ], + "spans": [ + { + "bbox": [ + 136, + 440, + 475, + 542 + ], + "score": 0.964, + "type": "image", + "image_path": "ef39205e27e246bc37f3a17ac72fc093d4914d8003b2bd66d967fd9856bb6ce9.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 136, + 440, + 475, + 474.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 136, + 474.0, + 475, + 508.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 136, + 508.0, + 475, + 542.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 555, + 505, + 588 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 555, + 504, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 504, + 568 + ], + "score": 1.0, + "content": "Figure 2: Graphical overview of soft unification within a memory network. Each sentence is pro-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 567, + 504, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 504, + 578 + ], + "score": 1.0, + "content": "cessed by two bi-directional RNNs for memory and unification. At each iteration the context atten-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 577, + 494, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 494, + 590 + ], + "score": 1.0, + "content": "tion selects which sentences to unify and the invariant produces the same answer as the example.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 593, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 227, + 607 + ], + "score": 1.0, + "content": "Unification CNN (UCNN) (", + "type": "text" + }, + { + "bbox": [ + 227, + 594, + 234, + 606 + ], + "score": 0.58, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 594, + 269, + 607 + ], + "score": 1.0, + "content": ": CNN,", + "type": "text" + }, + { + "bbox": [ + 270, + 596, + 276, + 606 + ], + "score": 0.69, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 594, + 465, + 607 + ], + "score": 1.0, + "content": ": CNN) Given a grid of embedded symbols", + "type": "text" + }, + { + "bbox": [ + 466, + 594, + 505, + 606 + ], + "score": 0.9, + "content": "\\phi ( K ) \\ \\in", + "type": "inline_equation" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 602, + 507, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 142, + 615 + ], + "score": 0.9, + "content": "\\mathbb { R } ^ { w \\times h \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 602, + 172, + 618 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 172, + 607, + 181, + 615 + ], + "score": 0.75, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 602, + 251, + 618 + ], + "score": 1.0, + "content": "is the width and", + "type": "text" + }, + { + "bbox": [ + 251, + 605, + 258, + 615 + ], + "score": 0.79, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 602, + 507, + 618 + ], + "score": 1.0, + "content": "the height, we use a convolutional neural network such that", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 193, + 628 + ], + "score": 1.0, + "content": "the final prediction is", + "type": "text" + }, + { + "bbox": [ + 193, + 615, + 345, + 628 + ], + "score": 0.89, + "content": "f ( \\phi ( K ) ) = \\mathrm { s o f t m a x } ( ( W h + b ) E ^ { T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 615, + 372, + 628 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 373, + 616, + 381, + 626 + ], + "score": 0.77, + "content": "^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "this time is the result of global", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 177, + 640 + ], + "score": 1.0, + "content": "max pooling and", + "type": "text" + }, + { + "bbox": [ + 178, + 627, + 201, + 638 + ], + "score": 0.82, + "content": "W , b", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 627, + 371, + 640 + ], + "score": 1.0, + "content": "are learnable parameters. We also model", + "type": "text" + }, + { + "bbox": [ + 371, + 629, + 378, + 638 + ], + "score": 0.8, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "using a separate convolutional", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 261, + 650 + ], + "score": 1.0, + "content": "network with the same architecture as", + "type": "text" + }, + { + "bbox": [ + 261, + 639, + 269, + 649 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 638, + 301, + 650 + ], + "score": 1.0, + "content": "and set", + "type": "text" + }, + { + "bbox": [ + 301, + 638, + 401, + 650 + ], + "score": 0.9, + "content": "\\phi _ { U } ( k ) = c _ { 2 } ( \\mathrm { r e l u } ( c _ { 1 } ( k ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 638, + 430, + 650 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 430, + 640, + 453, + 649 + ], + "score": 0.88, + "content": "c _ { 1 } , c _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "are the con-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 328, + 660 + ], + "score": 1.0, + "content": "volutional layers. The grid is padded with 0s to obtain", + "type": "text" + }, + { + "bbox": [ + 329, + 650, + 373, + 659 + ], + "score": 0.91, + "content": "w \\times h \\times d", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "after each convolution such that", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "every symbol has a unifying feature. This model conveys how soft unification can be adapted to the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 671, + 465, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 465, + 684 + ], + "score": 1.0, + "content": "specifics of the domain for example by using a convolution in a spatially structured input.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 594, + 507, + 684 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 277, + 700 + ], + "score": 1.0, + "content": "Unification Memory Networks (UMN) (", + "type": "text" + }, + { + "bbox": [ + 277, + 689, + 284, + 699 + ], + "score": 0.74, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 687, + 331, + 700 + ], + "score": 1.0, + "content": ": MemNN,", + "type": "text" + }, + { + "bbox": [ + 331, + 691, + 337, + 699 + ], + "score": 0.59, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 687, + 506, + 700 + ], + "score": 1.0, + "content": ": RNN) Soft unification does not need to", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 170, + 710 + ], + "score": 1.0, + "content": "happen prior to", + "type": "text" + }, + { + "bbox": [ + 171, + 699, + 178, + 711 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 700, + 196, + 710 + ], + "score": 1.0, + "content": "in a", + "type": "text" + }, + { + "bbox": [ + 196, + 699, + 218, + 711 + ], + "score": 0.91, + "content": "f \\circ g", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 700, + 505, + 710 + ], + "score": 1.0, + "content": "fashion but can also be incorporated at any intermediate stage multiple", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "times. To demonstrate this ability, we unify the symbols at different memory locations at each", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "iteration of a Memory Network (Weston et al., 2015). Memory networks can handle a list of lists", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 236, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 415, + 250 + ], + "score": 1.0, + "content": "structure such as a tokenised story as shown in Figure 2. The memory network", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 415, + 237, + 422, + 248 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 423, + 236, + 506, + 250 + ], + "score": 1.0, + "content": "uses the final hidden", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 248, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 259 + ], + "score": 1.0, + "content": "state of a bi-directional GRU (outer squares in Figure 2) as the sentence representations to compute", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 258, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 505, + 271 + ], + "score": 1.0, + "content": "a context attention. At each iteration, we unify the words between the attended sentences using the", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 267, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 104, + 267, + 506, + 284 + ], + "score": 1.0, + "content": "same approach in UMLP with another bi-directional GRU (inner diamonds in Figure 2) for unifying", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 280, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 140, + 292 + ], + "score": 1.0, + "content": "features", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 140, + 280, + 236, + 292 + ], + "score": 0.79, + "content": "\\phi _ { U } ( \\mathbf { b e r n h a r d } ) = W _ { U } \\Phi (", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 237, + 280, + 505, + 292 + ], + "score": 1.0, + "content": "(bernhard). Following equation 2, the new unified representation of", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 291, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 243, + 304 + ], + "score": 1.0, + "content": "the memory slot is computed and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 243, + 292, + 250, + 302 + ], + "score": 0.83, + "content": "f", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 250, + 291, + 446, + 304 + ], + "score": 1.0, + "content": "uses it to perform the next iteration. Concretely,", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 447, + 293, + 453, + 303 + ], + "score": 0.76, + "content": "g", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 454, + 291, + 506, + 304 + ], + "score": 1.0, + "content": "produces an", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 300, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 179, + 315 + ], + "score": 1.0, + "content": "unification tensor", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 180, + 302, + 254, + 312 + ], + "score": 0.92, + "content": "\\pmb { \\mathsf { U } } \\in \\dot { \\mathbb { R } } ^ { M \\times m \\times N \\times d }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 254, + 300, + 283, + 315 + ], + "score": 1.0, + "content": "where", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 283, + 303, + 295, + 312 + ], + "score": 0.84, + "content": "M", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 295, + 300, + 313, + 315 + ], + "score": 1.0, + "content": "and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 314, + 304, + 324, + 312 + ], + "score": 0.78, + "content": "m", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 324, + 300, + 506, + 315 + ], + "score": 1.0, + "content": "is the number of sentences and words in the", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 313, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 214, + 325 + ], + "score": 1.0, + "content": "invariant respectively, and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 214, + 314, + 225, + 323 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 225, + 313, + 506, + 325 + ], + "score": 1.0, + "content": "is the number of sentences in the example such that after the context", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 322, + 506, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 219, + 337 + ], + "score": 1.0, + "content": "attentions are applied over", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 220, + 325, + 232, + 334 + ], + "score": 0.76, + "content": "M", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 232, + 322, + 252, + 337 + ], + "score": 1.0, + "content": "and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 253, + 325, + 263, + 334 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 263, + 322, + 312, + 337 + ], + "score": 1.0, + "content": ", we obtain", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 313, + 324, + 376, + 336 + ], + "score": 0.93, + "content": "\\boldsymbol { \\phi } ( \\pmb { k } ) \\in \\mathbb { R } ^ { m \\times d }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 376, + 322, + 506, + 337 + ], + "score": 1.0, + "content": "as the unified sentence at that", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 335, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 505, + 347 + ], + "score": 1.0, + "content": "iteration. Note that unlike in the UMLP case, the sentences can be of varying length. The prediction", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 345, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 173, + 359 + ], + "score": 1.0, + "content": "is then softmax", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 173, + 345, + 223, + 358 + ], + "score": 0.9, + "content": "( W h _ { I } ^ { J } + b )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 223, + 345, + 252, + 359 + ], + "score": 1.0, + "content": "where", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 252, + 345, + 266, + 358 + ], + "score": 0.89, + "content": "h _ { I } ^ { J }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 266, + 345, + 430, + 359 + ], + "score": 1.0, + "content": "is the hidden state of the invariant after", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 431, + 347, + 438, + 356 + ], + "score": 0.81, + "content": "J", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 439, + 345, + 506, + 359 + ], + "score": 1.0, + "content": "iterations. This", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 255, + 369 + ], + "score": 1.0, + "content": "setup, however, requires pre-training", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 255, + 358, + 262, + 369 + ], + "score": 0.83, + "content": "f", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 262, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "such that the context attentions match the correct sentences.", + "type": "text", + "cross_page": true + } + ], + "index": 16 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 687, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 122, + 111, + 489, + 216 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 80, + 503, + 103 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 80, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 505, + 93 + ], + "score": 1.0, + "content": "Table 1: Sample context, query and answer triples and their training sizes per task. For distribution", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 91, + 471, + 104 + ], + "spans": [ + { + "bbox": [ + 106, + 91, + 471, + 104 + ], + "score": 1.0, + "content": "of generated number of examples per task on Sequence and Grid data refer to Appendix B.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 122, + 111, + 489, + 216 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 111, + 489, + 216 + ], + "spans": [ + { + "bbox": [ + 122, + 111, + 489, + 216 + ], + "score": 0.985, + "html": "
DatasetContextQueryAnswerTraining Size
Sequence8384duplicate8≤1k,≤50
Grid00 3 1 859corner7≤1k,≤50
bAbIMary went to the kitchen. Sandra journeyed to the garden.Where is Mary?kitchen1k,50
Logicp(X)← q(X). q(a).p(a).True1k, 10k,50
", + "type": "table", + "image_path": "e907e40a0c5d0e4e0f752c42d1e303a4907a55461d9f328e2ce4b4abbfbb4726.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 122, + 111, + 489, + 146.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 122, + 146.0, + 489, + 181.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 122, + 181.0, + 489, + 216.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "text", + "bbox": [ + 107, + 236, + 505, + 368 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 415, + 250 + ], + "score": 1.0, + "content": "structure such as a tokenised story as shown in Figure 2. The memory network", + "type": "text" + }, + { + "bbox": [ + 415, + 237, + 422, + 248 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 236, + 506, + 250 + ], + "score": 1.0, + "content": "uses the final hidden", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 248, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 259 + ], + "score": 1.0, + "content": "state of a bi-directional GRU (outer squares in Figure 2) as the sentence representations to compute", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 258, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 505, + 271 + ], + "score": 1.0, + "content": "a context attention. At each iteration, we unify the words between the attended sentences using the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 267, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 104, + 267, + 506, + 284 + ], + "score": 1.0, + "content": "same approach in UMLP with another bi-directional GRU (inner diamonds in Figure 2) for unifying", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 280, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 140, + 292 + ], + "score": 1.0, + "content": "features", + "type": "text" + }, + { + "bbox": [ + 140, + 280, + 236, + 292 + ], + "score": 0.79, + "content": "\\phi _ { U } ( \\mathbf { b e r n h a r d } ) = W _ { U } \\Phi (", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 280, + 505, + 292 + ], + "score": 1.0, + "content": "(bernhard). Following equation 2, the new unified representation of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 291, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 243, + 304 + ], + "score": 1.0, + "content": "the memory slot is computed and", + "type": "text" + }, + { + "bbox": [ + 243, + 292, + 250, + 302 + ], + "score": 0.83, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 291, + 446, + 304 + ], + "score": 1.0, + "content": "uses it to perform the next iteration. Concretely,", + "type": "text" + }, + { + "bbox": [ + 447, + 293, + 453, + 303 + ], + "score": 0.76, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 291, + 506, + 304 + ], + "score": 1.0, + "content": "produces an", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 300, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 179, + 315 + ], + "score": 1.0, + "content": "unification tensor", + "type": "text" + }, + { + "bbox": [ + 180, + 302, + 254, + 312 + ], + "score": 0.92, + "content": "\\pmb { \\mathsf { U } } \\in \\dot { \\mathbb { R } } ^ { M \\times m \\times N \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 300, + 283, + 315 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 283, + 303, + 295, + 312 + ], + "score": 0.84, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 300, + 313, + 315 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 314, + 304, + 324, + 312 + ], + "score": 0.78, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 300, + 506, + 315 + ], + "score": 1.0, + "content": "is the number of sentences and words in the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 313, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 214, + 325 + ], + "score": 1.0, + "content": "invariant respectively, and", + "type": "text" + }, + { + "bbox": [ + 214, + 314, + 225, + 323 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 313, + 506, + 325 + ], + "score": 1.0, + "content": "is the number of sentences in the example such that after the context", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 322, + 506, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 219, + 337 + ], + "score": 1.0, + "content": "attentions are applied over", + "type": "text" + }, + { + "bbox": [ + 220, + 325, + 232, + 334 + ], + "score": 0.76, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 322, + 252, + 337 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 253, + 325, + 263, + 334 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 322, + 312, + 337 + ], + "score": 1.0, + "content": ", we obtain", + "type": "text" + }, + { + "bbox": [ + 313, + 324, + 376, + 336 + ], + "score": 0.93, + "content": "\\boldsymbol { \\phi } ( \\pmb { k } ) \\in \\mathbb { R } ^ { m \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 322, + 506, + 337 + ], + "score": 1.0, + "content": "as the unified sentence at that", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 335, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 505, + 347 + ], + "score": 1.0, + "content": "iteration. Note that unlike in the UMLP case, the sentences can be of varying length. The prediction", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 345, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 173, + 359 + ], + "score": 1.0, + "content": "is then softmax", + "type": "text" + }, + { + "bbox": [ + 173, + 345, + 223, + 358 + ], + "score": 0.9, + "content": "( W h _ { I } ^ { J } + b )", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 345, + 252, + 359 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 252, + 345, + 266, + 358 + ], + "score": 0.89, + "content": "h _ { I } ^ { J }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 345, + 430, + 359 + ], + "score": 1.0, + "content": "is the hidden state of the invariant after", + "type": "text" + }, + { + "bbox": [ + 431, + 347, + 438, + 356 + ], + "score": 0.81, + "content": "J", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 345, + 506, + 359 + ], + "score": 1.0, + "content": "iterations. This", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 255, + 369 + ], + "score": 1.0, + "content": "setup, however, requires pre-training", + "type": "text" + }, + { + "bbox": [ + 255, + 358, + 262, + 369 + ], + "score": 0.83, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "such that the context attentions match the correct sentences.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 373, + 505, + 475 + ], + "lines": [ + { + "bbox": [ + 105, + 373, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 423, + 386 + ], + "score": 1.0, + "content": "A task might contain different questions such as “Where is X?” and “Why did", + "type": "text" + }, + { + "bbox": [ + 424, + 374, + 433, + 384 + ], + "score": 0.26, + "content": "\\mathrm { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 373, + 506, + 386 + ], + "score": 1.0, + "content": "go to Y?”. To let", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "the models differentiate between questions and potentially learn different invariants, we extend them", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 396, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 229, + 409 + ], + "score": 1.0, + "content": "with a repository of invariants", + "type": "text" + }, + { + "bbox": [ + 229, + 396, + 253, + 406 + ], + "score": 0.9, + "content": "I \\in \\mathbb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 396, + 505, + 409 + ], + "score": 1.0, + "content": "and aggregate the predictions from each invariant. One simple", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 406, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 104, + 406, + 316, + 421 + ], + "score": 1.0, + "content": "approach is to sum the predictions of the invariants", + "type": "text" + }, + { + "bbox": [ + 316, + 407, + 390, + 420 + ], + "score": 0.92, + "content": "\\textstyle \\sum _ { I \\in \\mathbb { I } } f \\circ g ( I , K )", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 406, + 506, + 421 + ], + "score": 1.0, + "content": "used in UMLP and UCNN.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "Another approach could be to use features from the invariants such as memory representations in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 441, + 441 + ], + "score": 1.0, + "content": "the case of UMN. For UMN, we weigh the predictions using a bilinear attention", + "type": "text" + }, + { + "bbox": [ + 442, + 431, + 449, + 441 + ], + "score": 0.78, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "based on the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 102, + 438, + 507, + 458 + ], + "spans": [ + { + "bbox": [ + 102, + 438, + 241, + 458 + ], + "score": 1.0, + "content": "hidden states at the first iteration", + "type": "text" + }, + { + "bbox": [ + 241, + 441, + 254, + 454 + ], + "score": 0.89, + "content": "h _ { I } ^ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 438, + 272, + 458 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 272, + 441, + 288, + 454 + ], + "score": 0.91, + "content": "h _ { K } ^ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 438, + 329, + 458 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 329, + 439, + 430, + 454 + ], + "score": 0.9, + "content": "\\eta = \\mathrm { s o f t m a x } ( h _ { I } ^ { 0 } W h _ { K } ^ { 0 ^ { T } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 438, + 507, + 458 + ], + "score": 1.0, + "content": ". To initially form", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "score": 1.0, + "content": "the repository of invariants, we use the bag-of-words representation of the questions and find the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 464, + 479, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 479, + 476 + ], + "score": 1.0, + "content": "most dissimilar ones based on their cosine similarity as a heuristic to obtain varied examples.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 107, + 491, + 179, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 490, + 181, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 181, + 506 + ], + "score": 1.0, + "content": "4 DATASETS", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 516, + 505, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 515, + 507, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 357, + 530 + ], + "score": 1.0, + "content": "We use 4 datasets consisting of context, query and an answer", + "type": "text" + }, + { + "bbox": [ + 358, + 516, + 394, + 528 + ], + "score": 0.93, + "content": "( C , \\pmb { q } , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 515, + 507, + 530 + ], + "score": 1.0, + "content": ": fixed length sequences of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "score": 1.0, + "content": "symbols, shapes of symbols in a grid, story based natural language reasoning with the bAbI (Weston", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 537, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 104, + 537, + 506, + 552 + ], + "score": 1.0, + "content": "et al., 2016) dataset and logical reasoning represented as logic programs, examples shown in Table 1", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 549, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 506, + 563 + ], + "score": 1.0, + "content": "with further samples in Appendix B. In each case we use an appropriate model: UMLP for fixed", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 559, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 506, + 573 + ], + "score": 1.0, + "content": "length sequences, UCNN for grid and UMN for iterative reasoning. We use synthetic datasets of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 570, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 584 + ], + "score": 1.0, + "content": "which the data generating distributions are known to evaluate not only the quantitative performance", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 582, + 352, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 352, + 595 + ], + "score": 1.0, + "content": "but also the quality of the invariants learned by our approach.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 599, + 504, + 643 + ], + "lines": [ + { + "bbox": [ + 105, + 598, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 341, + 612 + ], + "score": 1.0, + "content": "Fixed Length Sequences We generate sequences of length", + "type": "text" + }, + { + "bbox": [ + 342, + 599, + 365, + 609 + ], + "score": 0.89, + "content": "l = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 598, + 506, + 612 + ], + "score": 1.0, + "content": "with 8 unique symbols represented", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 610, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 622 + ], + "score": 1.0, + "content": "as digits to predict (i) a constant, (ii) the head of the sequence, (iii) the tail and (iv) the duplicate", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 622, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 633 + ], + "score": 1.0, + "content": "symbol. We randomly generate 1000 triples and then only take the unique ones to ensure the test", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 632, + 481, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 481, + 644 + ], + "score": 1.0, + "content": "split contains unseen examples. The training is then performed over a 5-fold cross-validation.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 649, + 504, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 352, + 662 + ], + "score": 1.0, + "content": "Grid To spatially organise symbols, we generate a grid of size", + "type": "text" + }, + { + "bbox": [ + 352, + 649, + 372, + 659 + ], + "score": 0.89, + "content": "3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "with 8 unique symbols organised", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 124, + 673 + ], + "score": 1.0, + "content": "into", + "type": "text" + }, + { + "bbox": [ + 125, + 660, + 149, + 671 + ], + "score": 0.89, + "content": "2 \\times 2", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "box of identical symbol, a vertical, diagonal or horizontal sequence of length 3, a cross", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "score": 1.0, + "content": "or a plus shape and a triangle. In each task we predict (i) the identical symbol, (ii) the head of the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 681, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 695 + ], + "score": 1.0, + "content": "sequence, (iii) the centre of the cross or plus and (iv) the corner of the triangle respectively. We", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 692, + 505, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 505, + 706 + ], + "score": 1.0, + "content": "follow the same procedure from sequences and randomly generate 1000 discarding duplicate triples.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "bAbI The bAbI dataset has become a standard benchmark for evaluating memory based networks.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "It consists of 20 synthetically generated natural language reasoning tasks (refer to Weston et al.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 122, + 111, + 489, + 216 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 80, + 503, + 103 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 80, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 505, + 93 + ], + "score": 1.0, + "content": "Table 1: Sample context, query and answer triples and their training sizes per task. For distribution", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 91, + 471, + 104 + ], + "spans": [ + { + "bbox": [ + 106, + 91, + 471, + 104 + ], + "score": 1.0, + "content": "of generated number of examples per task on Sequence and Grid data refer to Appendix B.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 122, + 111, + 489, + 216 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 111, + 489, + 216 + ], + "spans": [ + { + "bbox": [ + 122, + 111, + 489, + 216 + ], + "score": 0.985, + "html": "
DatasetContextQueryAnswerTraining Size
Sequence8384duplicate8≤1k,≤50
Grid00 3 1 859corner7≤1k,≤50
bAbIMary went to the kitchen. Sandra journeyed to the garden.Where is Mary?kitchen1k,50
Logicp(X)← q(X). q(a).p(a).True1k, 10k,50
", + "type": "table", + "image_path": "e907e40a0c5d0e4e0f752c42d1e303a4907a55461d9f328e2ce4b4abbfbb4726.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 122, + 111, + 489, + 146.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 122, + 146.0, + 489, + 181.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 122, + 181.0, + 489, + 216.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "text", + "bbox": [ + 107, + 236, + 505, + 368 + ], + "lines": [], + "index": 10.5, + "bbox_fs": [ + 104, + 236, + 506, + 369 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 373, + 505, + 475 + ], + "lines": [ + { + "bbox": [ + 105, + 373, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 423, + 386 + ], + "score": 1.0, + "content": "A task might contain different questions such as “Where is X?” and “Why did", + "type": "text" + }, + { + "bbox": [ + 424, + 374, + 433, + 384 + ], + "score": 0.26, + "content": "\\mathrm { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 373, + 506, + 386 + ], + "score": 1.0, + "content": "go to Y?”. To let", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "the models differentiate between questions and potentially learn different invariants, we extend them", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 396, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 229, + 409 + ], + "score": 1.0, + "content": "with a repository of invariants", + "type": "text" + }, + { + "bbox": [ + 229, + 396, + 253, + 406 + ], + "score": 0.9, + "content": "I \\in \\mathbb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 396, + 505, + 409 + ], + "score": 1.0, + "content": "and aggregate the predictions from each invariant. One simple", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 406, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 104, + 406, + 316, + 421 + ], + "score": 1.0, + "content": "approach is to sum the predictions of the invariants", + "type": "text" + }, + { + "bbox": [ + 316, + 407, + 390, + 420 + ], + "score": 0.92, + "content": "\\textstyle \\sum _ { I \\in \\mathbb { I } } f \\circ g ( I , K )", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 406, + 506, + 421 + ], + "score": 1.0, + "content": "used in UMLP and UCNN.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "Another approach could be to use features from the invariants such as memory representations in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 441, + 441 + ], + "score": 1.0, + "content": "the case of UMN. For UMN, we weigh the predictions using a bilinear attention", + "type": "text" + }, + { + "bbox": [ + 442, + 431, + 449, + 441 + ], + "score": 0.78, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "based on the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 102, + 438, + 507, + 458 + ], + "spans": [ + { + "bbox": [ + 102, + 438, + 241, + 458 + ], + "score": 1.0, + "content": "hidden states at the first iteration", + "type": "text" + }, + { + "bbox": [ + 241, + 441, + 254, + 454 + ], + "score": 0.89, + "content": "h _ { I } ^ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 438, + 272, + 458 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 272, + 441, + 288, + 454 + ], + "score": 0.91, + "content": "h _ { K } ^ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 438, + 329, + 458 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 329, + 439, + 430, + 454 + ], + "score": 0.9, + "content": "\\eta = \\mathrm { s o f t m a x } ( h _ { I } ^ { 0 } W h _ { K } ^ { 0 ^ { T } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 438, + 507, + 458 + ], + "score": 1.0, + "content": ". To initially form", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "score": 1.0, + "content": "the repository of invariants, we use the bag-of-words representation of the questions and find the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 464, + 479, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 479, + 476 + ], + "score": 1.0, + "content": "most dissimilar ones based on their cosine similarity as a heuristic to obtain varied examples.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 21, + "bbox_fs": [ + 102, + 373, + 507, + 476 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 491, + 179, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 490, + 181, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 181, + 506 + ], + "score": 1.0, + "content": "4 DATASETS", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 516, + 505, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 515, + 507, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 357, + 530 + ], + "score": 1.0, + "content": "We use 4 datasets consisting of context, query and an answer", + "type": "text" + }, + { + "bbox": [ + 358, + 516, + 394, + 528 + ], + "score": 0.93, + "content": "( C , \\pmb { q } , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 515, + 507, + 530 + ], + "score": 1.0, + "content": ": fixed length sequences of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "score": 1.0, + "content": "symbols, shapes of symbols in a grid, story based natural language reasoning with the bAbI (Weston", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 537, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 104, + 537, + 506, + 552 + ], + "score": 1.0, + "content": "et al., 2016) dataset and logical reasoning represented as logic programs, examples shown in Table 1", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 549, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 506, + 563 + ], + "score": 1.0, + "content": "with further samples in Appendix B. In each case we use an appropriate model: UMLP for fixed", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 559, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 506, + 573 + ], + "score": 1.0, + "content": "length sequences, UCNN for grid and UMN for iterative reasoning. We use synthetic datasets of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 570, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 584 + ], + "score": 1.0, + "content": "which the data generating distributions are known to evaluate not only the quantitative performance", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 582, + 352, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 352, + 595 + ], + "score": 1.0, + "content": "but also the quality of the invariants learned by our approach.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30, + "bbox_fs": [ + 104, + 515, + 507, + 595 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 599, + 504, + 643 + ], + "lines": [ + { + "bbox": [ + 105, + 598, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 341, + 612 + ], + "score": 1.0, + "content": "Fixed Length Sequences We generate sequences of length", + "type": "text" + }, + { + "bbox": [ + 342, + 599, + 365, + 609 + ], + "score": 0.89, + "content": "l = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 598, + 506, + 612 + ], + "score": 1.0, + "content": "with 8 unique symbols represented", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 610, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 622 + ], + "score": 1.0, + "content": "as digits to predict (i) a constant, (ii) the head of the sequence, (iii) the tail and (iv) the duplicate", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 622, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 633 + ], + "score": 1.0, + "content": "symbol. We randomly generate 1000 triples and then only take the unique ones to ensure the test", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 632, + 481, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 481, + 644 + ], + "score": 1.0, + "content": "split contains unseen examples. The training is then performed over a 5-fold cross-validation.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 598, + 506, + 644 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 649, + 504, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 352, + 662 + ], + "score": 1.0, + "content": "Grid To spatially organise symbols, we generate a grid of size", + "type": "text" + }, + { + "bbox": [ + 352, + 649, + 372, + 659 + ], + "score": 0.89, + "content": "3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "with 8 unique symbols organised", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 124, + 673 + ], + "score": 1.0, + "content": "into", + "type": "text" + }, + { + "bbox": [ + 125, + 660, + 149, + 671 + ], + "score": 0.89, + "content": "2 \\times 2", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "box of identical symbol, a vertical, diagonal or horizontal sequence of length 3, a cross", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "score": 1.0, + "content": "or a plus shape and a triangle. In each task we predict (i) the identical symbol, (ii) the head of the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 681, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 695 + ], + "score": 1.0, + "content": "sequence, (iii) the centre of the cross or plus and (iv) the corner of the triangle respectively. We", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 692, + 505, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 505, + 706 + ], + "score": 1.0, + "content": "follow the same procedure from sequences and randomly generate 1000 discarding duplicate triples.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 648, + 506, + 706 + ] + }, + { + "type": "list", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "bAbI The bAbI dataset has become a standard benchmark for evaluating memory based networks.", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "It consists of 20 synthetically generated natural language reasoning tasks (refer to Weston et al.", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "(2016) for task details). We take the 1k English set and use 0.1 of the training set as validation. Each", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "token is lower cased and considered a unique symbol. Following previous works (Seo et al., 2017;", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 464, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 464, + 117 + ], + "score": 1.0, + "content": "Sukhbaatar et al., 2015), we take multiple word answers also to be a unique symbol in S.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 121, + 505, + 210 + ], + "lines": [ + { + "bbox": [ + 105, + 121, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 506, + 134 + ], + "score": 1.0, + "content": "Logical Reasoning To demonstrate the flexibility of our approach and distinguish our notion of", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 104, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "a variable from that used in logic based formalisms, we generate logical reasoning tasks in the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 157 + ], + "score": 1.0, + "content": "form of logic programs using the procedure by Cingillioglu & Russo (2019). The tasks involve", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 141, + 167 + ], + "score": 1.0, + "content": "learning", + "type": "text" + }, + { + "bbox": [ + 142, + 154, + 250, + 166 + ], + "score": 0.91, + "content": "f ( \\bar { C } , \\bar { Q } ) \\stackrel { - } { = } \\mathrm { T r u e } \\bar { C } \\vdash \\bar { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "over 12 classes of logic programs exhibiting varying paradigms", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "of logical reasoning including negation by failure (Clark, 1978). We generate 1k and 10k logic", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 104, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "programs per task for training with 0.1 as validation and another 1k for testing. We set the arity of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "literals to 1 or 2 using one random character from the English alphabet for predicates and constants,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 198, + 383, + 211 + ], + "spans": [ + { + "bbox": [ + 104, + 198, + 124, + 211 + ], + "score": 1.0, + "content": "e.g.", + "type": "text" + }, + { + "bbox": [ + 124, + 198, + 143, + 210 + ], + "score": 0.9, + "content": "p ( p )", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 198, + 358, + 211 + ], + "score": 1.0, + "content": "and an upper case character for logical variables, e.g.", + "type": "text" + }, + { + "bbox": [ + 359, + 198, + 378, + 210 + ], + "score": 0.91, + "content": "p ( \\mathrm { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 198, + 383, + 211 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 108, + 226, + 200, + 238 + ], + "lines": [ + { + "bbox": [ + 104, + 224, + 201, + 241 + ], + "spans": [ + { + "bbox": [ + 104, + 224, + 201, + 241 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 250, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 106, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "We probe three aspects of soft unification: the impact of unification on performance over unseen", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 261, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 272 + ], + "score": 1.0, + "content": "data, the effect of multiple invariants and data efficiency. To that end, we train UMLP and UCNN", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "score": 1.0, + "content": "with and without unification, UMN with pre-training using 1 or 3 invariants over either the en-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 281, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 297 + ], + "score": 1.0, + "content": "tire training set or only 50 examples. Every model is trained 3 times via back-propagation using", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 293, + 502, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 502, + 306 + ], + "score": 1.0, + "content": "Adam (Kingma & Ba, 2015) on an Intel Core i7-6700 CPU using the following objective function:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + }, + { + "type": "interline_equation", + "bbox": [ + 171, + 308, + 440, + 343 + ], + "lines": [ + { + "bbox": [ + 171, + 308, + 440, + 343 + ], + "spans": [ + { + "bbox": [ + 171, + 308, + 440, + 343 + ], + "score": 0.93, + "content": "J = \\lambda _ { K } \\mathcal { L } _ { \\mathrm { n l l } } ( f ( K ) , a ) + \\lambda _ { U } \\left[ \\mathcal { L } _ { \\mathrm { n l l } } ( f \\circ g ( I , K ) , a ) + \\tau \\sum _ { s \\in \\mathbb { S } } \\psi _ { w } ( s ) \\right]", + "type": "interline_equation", + "image_path": "562850468dbca51e031b1d308f05cc93e1a537ef85b27f0952beccb216be018d.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 171, + 308, + 440, + 319.6666666666667 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 171, + 319.6666666666667, + 440, + 331.33333333333337 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 171, + 331.33333333333337, + 440, + 343.00000000000006 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 346, + 505, + 437 + ], + "lines": [ + { + "bbox": [ + 105, + 346, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 133, + 360 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 347, + 149, + 358 + ], + "score": 0.9, + "content": "{ \\mathcal { L } } _ { \\mathrm { n l l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 346, + 397, + 360 + ], + "score": 1.0, + "content": "is the negative log-likelihood with sparsity regularisation over", + "type": "text" + }, + { + "bbox": [ + 397, + 347, + 405, + 358 + ], + "score": 0.86, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 346, + 415, + 360 + ], + "score": 1.0, + "content": "at", + "type": "text" + }, + { + "bbox": [ + 416, + 347, + 448, + 357 + ], + "score": 0.89, + "content": "\\tau = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 346, + 505, + 360 + ], + "score": 1.0, + "content": "to discourage", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 358, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 476, + 370 + ], + "score": 1.0, + "content": "the models from utilising spurious number of variables. For UMLP and UCNN, we set", + "type": "text" + }, + { + "bbox": [ + 476, + 358, + 505, + 369 + ], + "score": 0.89, + "content": "\\lambda _ { K } ~ =", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 117, + 381 + ], + "score": 1.0, + "content": "0,", + "type": "text" + }, + { + "bbox": [ + 118, + 369, + 153, + 380 + ], + "score": 0.81, + "content": "\\lambda _ { U } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "for training just the unified output and the converse for the non-unifying versions. To", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 104, + 379, + 242, + 392 + ], + "score": 1.0, + "content": "pre-train the UMN, we start with", + "type": "text" + }, + { + "bbox": [ + 243, + 380, + 317, + 391 + ], + "score": 0.29, + "content": "\\lambda _ { K } = 1 , \\lambda _ { U } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 379, + 409, + 392 + ], + "score": 1.0, + "content": "for 40 epochs then set", + "type": "text" + }, + { + "bbox": [ + 410, + 380, + 443, + 390 + ], + "score": 0.91, + "content": "\\lambda _ { U } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "to jointly train", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 390, + 504, + 406 + ], + "spans": [ + { + "bbox": [ + 104, + 390, + 452, + 406 + ], + "score": 1.0, + "content": "the unified output. For iterative tasks, the mean squared error between hidden states", + "type": "text" + }, + { + "bbox": [ + 452, + 390, + 504, + 405 + ], + "score": 0.92, + "content": "( { h _ { I } ^ { j } - h _ { K } ^ { j } } ) ^ { 2 }", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 403, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 104, + 403, + 173, + 416 + ], + "score": 1.0, + "content": "at each iteration", + "type": "text" + }, + { + "bbox": [ + 174, + 404, + 180, + 415 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 403, + 506, + 416 + ], + "score": 1.0, + "content": "and, in the strongly supervised cases, the negative log-likelihood for the context", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 414, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 104, + 414, + 506, + 427 + ], + "score": 1.0, + "content": "attentions using provided supporting facts are also added to the objective function. Further details", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 425, + 407, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 407, + 437 + ], + "score": 1.0, + "content": "such as batch size and total number of epochs are available in Appendix C.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + }, + { + "type": "image", + "bbox": [ + 109, + 442, + 501, + 549 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 442, + 501, + 549 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 442, + 501, + 549 + ], + "spans": [ + { + "bbox": [ + 109, + 442, + 501, + 549 + ], + "score": 0.972, + "type": "image", + "image_path": "4aca829873f131070cc20ca4370a0d6ac4a79c32120a7d5e548cfc6ffbcee87f.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 109, + 442, + 501, + 477.6666666666667 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 109, + 477.6666666666667, + 501, + 513.3333333333334 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 109, + 513.3333333333334, + 501, + 549.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 560, + 505, + 595 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "Figure 3: Test accuracy over iterations for Unification MLP and Unification CNN models with 1", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "invariant versus no unification. We observe that with soft unification the models achieve higher", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 583, + 468, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 468, + 595 + ], + "score": 1.0, + "content": "accuracy with fewer iterations than their plain counterparts on both per task training sizes.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 600, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "Figure 3 portrays how soft unification generalises better to unseen examples in test sets - the same", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 612, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 623 + ], + "score": 1.0, + "content": "sequence or grid never appears in both the training and test sets as outlined in Section 4 - over plain", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 175, + 634 + ], + "score": 1.0, + "content": "models. Despite", + "type": "text" + }, + { + "bbox": [ + 175, + 622, + 198, + 633 + ], + "score": 0.91, + "content": "f \\circ g", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 622, + 356, + 634 + ], + "score": 1.0, + "content": "having more trainable parameters than", + "type": "text" + }, + { + "bbox": [ + 357, + 622, + 364, + 633 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "alone, the models with unification", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "score": 1.0, + "content": "not only maintain higher accuracy in each iteration and solve the tasks in as few as 250 iterations", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 643, + 504, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 126, + 656 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 644, + 160, + 655 + ], + "score": 0.87, + "content": "\\leq 1 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 643, + 357, + 656 + ], + "score": 1.0, + "content": "training examples but also improve accuracy by", + "type": "text" + }, + { + "bbox": [ + 357, + 644, + 383, + 654 + ], + "score": 0.85, + "content": "\\approx 0 . 3", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 643, + 480, + 656 + ], + "score": 1.0, + "content": "when trained with only", + "type": "text" + }, + { + "bbox": [ + 481, + 644, + 504, + 655 + ], + "score": 0.86, + "content": "\\leq 5 0", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 104, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "per task. We believe soft unification architecturally biases the models towards learning structural", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "patterns which in return achieves better results on recognising common patterns of symbols across", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "examples. Results with multiple invariants are identical and the models seem to ignore the extra", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "invariants due to the fact that the tasks can be solved with a single invariant and the regularisation", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 151, + 711 + ], + "score": 1.0, + "content": "applied on", + "type": "text" + }, + { + "bbox": [ + 151, + 699, + 159, + 710 + ], + "score": 0.83, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "zeroing out unnecessary invariants; further results in Appendix D. The fluctuations in", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 426, + 722 + ], + "score": 1.0, + "content": "accuracy around iterations 750 to 1000 in UCNN are also caused by penalising", + "type": "text" + }, + { + "bbox": [ + 426, + 710, + 435, + 721 + ], + "score": 0.83, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "which forces the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 720, + 389, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 389, + 734 + ], + "score": 1.0, + "content": "model to relearn the task with less variables half way through training.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 39.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "(2016) for task details). We take the 1k English set and use 0.1 of the training set as validation. Each", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "token is lower cased and considered a unique symbol. Following previous works (Seo et al., 2017;", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 464, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 464, + 117 + ], + "score": 1.0, + "content": "Sukhbaatar et al., 2015), we take multiple word answers also to be a unique symbol in S.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 82, + 505, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 121, + 505, + 210 + ], + "lines": [ + { + "bbox": [ + 105, + 121, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 506, + 134 + ], + "score": 1.0, + "content": "Logical Reasoning To demonstrate the flexibility of our approach and distinguish our notion of", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 104, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "a variable from that used in logic based formalisms, we generate logical reasoning tasks in the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 157 + ], + "score": 1.0, + "content": "form of logic programs using the procedure by Cingillioglu & Russo (2019). The tasks involve", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 141, + 167 + ], + "score": 1.0, + "content": "learning", + "type": "text" + }, + { + "bbox": [ + 142, + 154, + 250, + 166 + ], + "score": 0.91, + "content": "f ( \\bar { C } , \\bar { Q } ) \\stackrel { - } { = } \\mathrm { T r u e } \\bar { C } \\vdash \\bar { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "over 12 classes of logic programs exhibiting varying paradigms", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "of logical reasoning including negation by failure (Clark, 1978). We generate 1k and 10k logic", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 104, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "programs per task for training with 0.1 as validation and another 1k for testing. We set the arity of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "literals to 1 or 2 using one random character from the English alphabet for predicates and constants,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 198, + 383, + 211 + ], + "spans": [ + { + "bbox": [ + 104, + 198, + 124, + 211 + ], + "score": 1.0, + "content": "e.g.", + "type": "text" + }, + { + "bbox": [ + 124, + 198, + 143, + 210 + ], + "score": 0.9, + "content": "p ( p )", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 198, + 358, + 211 + ], + "score": 1.0, + "content": "and an upper case character for logical variables, e.g.", + "type": "text" + }, + { + "bbox": [ + 359, + 198, + 378, + 210 + ], + "score": 0.91, + "content": "p ( \\mathrm { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 198, + 383, + 211 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5, + "bbox_fs": [ + 104, + 121, + 506, + 211 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 226, + 200, + 238 + ], + "lines": [ + { + "bbox": [ + 104, + 224, + 201, + 241 + ], + "spans": [ + { + "bbox": [ + 104, + 224, + 201, + 241 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 250, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 106, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "We probe three aspects of soft unification: the impact of unification on performance over unseen", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 261, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 272 + ], + "score": 1.0, + "content": "data, the effect of multiple invariants and data efficiency. To that end, we train UMLP and UCNN", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "score": 1.0, + "content": "with and without unification, UMN with pre-training using 1 or 3 invariants over either the en-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 281, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 297 + ], + "score": 1.0, + "content": "tire training set or only 50 examples. Every model is trained 3 times via back-propagation using", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 293, + 502, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 502, + 306 + ], + "score": 1.0, + "content": "Adam (Kingma & Ba, 2015) on an Intel Core i7-6700 CPU using the following objective function:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 250, + 505, + 306 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 171, + 308, + 440, + 343 + ], + "lines": [ + { + "bbox": [ + 171, + 308, + 440, + 343 + ], + "spans": [ + { + "bbox": [ + 171, + 308, + 440, + 343 + ], + "score": 0.93, + "content": "J = \\lambda _ { K } \\mathcal { L } _ { \\mathrm { n l l } } ( f ( K ) , a ) + \\lambda _ { U } \\left[ \\mathcal { L } _ { \\mathrm { n l l } } ( f \\circ g ( I , K ) , a ) + \\tau \\sum _ { s \\in \\mathbb { S } } \\psi _ { w } ( s ) \\right]", + "type": "interline_equation", + "image_path": "562850468dbca51e031b1d308f05cc93e1a537ef85b27f0952beccb216be018d.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 171, + 308, + 440, + 319.6666666666667 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 171, + 319.6666666666667, + 440, + 331.33333333333337 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 171, + 331.33333333333337, + 440, + 343.00000000000006 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 346, + 505, + 437 + ], + "lines": [ + { + "bbox": [ + 105, + 346, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 133, + 360 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 347, + 149, + 358 + ], + "score": 0.9, + "content": "{ \\mathcal { L } } _ { \\mathrm { n l l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 346, + 397, + 360 + ], + "score": 1.0, + "content": "is the negative log-likelihood with sparsity regularisation over", + "type": "text" + }, + { + "bbox": [ + 397, + 347, + 405, + 358 + ], + "score": 0.86, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 346, + 415, + 360 + ], + "score": 1.0, + "content": "at", + "type": "text" + }, + { + "bbox": [ + 416, + 347, + 448, + 357 + ], + "score": 0.89, + "content": "\\tau = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 346, + 505, + 360 + ], + "score": 1.0, + "content": "to discourage", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 358, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 476, + 370 + ], + "score": 1.0, + "content": "the models from utilising spurious number of variables. For UMLP and UCNN, we set", + "type": "text" + }, + { + "bbox": [ + 476, + 358, + 505, + 369 + ], + "score": 0.89, + "content": "\\lambda _ { K } ~ =", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 117, + 381 + ], + "score": 1.0, + "content": "0,", + "type": "text" + }, + { + "bbox": [ + 118, + 369, + 153, + 380 + ], + "score": 0.81, + "content": "\\lambda _ { U } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "for training just the unified output and the converse for the non-unifying versions. To", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 104, + 379, + 242, + 392 + ], + "score": 1.0, + "content": "pre-train the UMN, we start with", + "type": "text" + }, + { + "bbox": [ + 243, + 380, + 317, + 391 + ], + "score": 0.29, + "content": "\\lambda _ { K } = 1 , \\lambda _ { U } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 379, + 409, + 392 + ], + "score": 1.0, + "content": "for 40 epochs then set", + "type": "text" + }, + { + "bbox": [ + 410, + 380, + 443, + 390 + ], + "score": 0.91, + "content": "\\lambda _ { U } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "to jointly train", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 390, + 504, + 406 + ], + "spans": [ + { + "bbox": [ + 104, + 390, + 452, + 406 + ], + "score": 1.0, + "content": "the unified output. For iterative tasks, the mean squared error between hidden states", + "type": "text" + }, + { + "bbox": [ + 452, + 390, + 504, + 405 + ], + "score": 0.92, + "content": "( { h _ { I } ^ { j } - h _ { K } ^ { j } } ) ^ { 2 }", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 403, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 104, + 403, + 173, + 416 + ], + "score": 1.0, + "content": "at each iteration", + "type": "text" + }, + { + "bbox": [ + 174, + 404, + 180, + 415 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 403, + 506, + 416 + ], + "score": 1.0, + "content": "and, in the strongly supervised cases, the negative log-likelihood for the context", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 414, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 104, + 414, + 506, + 427 + ], + "score": 1.0, + "content": "attentions using provided supporting facts are also added to the objective function. Further details", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 425, + 407, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 407, + 437 + ], + "score": 1.0, + "content": "such as batch size and total number of epochs are available in Appendix C.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5, + "bbox_fs": [ + 104, + 346, + 506, + 437 + ] + }, + { + "type": "image", + "bbox": [ + 109, + 442, + 501, + 549 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 442, + 501, + 549 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 442, + 501, + 549 + ], + "spans": [ + { + "bbox": [ + 109, + 442, + 501, + 549 + ], + "score": 0.972, + "type": "image", + "image_path": "4aca829873f131070cc20ca4370a0d6ac4a79c32120a7d5e548cfc6ffbcee87f.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 109, + 442, + 501, + 477.6666666666667 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 109, + 477.6666666666667, + 501, + 513.3333333333334 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 109, + 513.3333333333334, + 501, + 549.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 560, + 505, + 595 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "Figure 3: Test accuracy over iterations for Unification MLP and Unification CNN models with 1", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "invariant versus no unification. We observe that with soft unification the models achieve higher", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 583, + 468, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 468, + 595 + ], + "score": 1.0, + "content": "accuracy with fewer iterations than their plain counterparts on both per task training sizes.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 600, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "Figure 3 portrays how soft unification generalises better to unseen examples in test sets - the same", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 612, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 623 + ], + "score": 1.0, + "content": "sequence or grid never appears in both the training and test sets as outlined in Section 4 - over plain", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 175, + 634 + ], + "score": 1.0, + "content": "models. Despite", + "type": "text" + }, + { + "bbox": [ + 175, + 622, + 198, + 633 + ], + "score": 0.91, + "content": "f \\circ g", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 622, + 356, + 634 + ], + "score": 1.0, + "content": "having more trainable parameters than", + "type": "text" + }, + { + "bbox": [ + 357, + 622, + 364, + 633 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "alone, the models with unification", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "score": 1.0, + "content": "not only maintain higher accuracy in each iteration and solve the tasks in as few as 250 iterations", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 643, + 504, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 126, + 656 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 644, + 160, + 655 + ], + "score": 0.87, + "content": "\\leq 1 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 643, + 357, + 656 + ], + "score": 1.0, + "content": "training examples but also improve accuracy by", + "type": "text" + }, + { + "bbox": [ + 357, + 644, + 383, + 654 + ], + "score": 0.85, + "content": "\\approx 0 . 3", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 643, + 480, + 656 + ], + "score": 1.0, + "content": "when trained with only", + "type": "text" + }, + { + "bbox": [ + 481, + 644, + 504, + 655 + ], + "score": 0.86, + "content": "\\leq 5 0", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 104, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "per task. We believe soft unification architecturally biases the models towards learning structural", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "patterns which in return achieves better results on recognising common patterns of symbols across", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "examples. Results with multiple invariants are identical and the models seem to ignore the extra", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "invariants due to the fact that the tasks can be solved with a single invariant and the regularisation", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 151, + 711 + ], + "score": 1.0, + "content": "applied on", + "type": "text" + }, + { + "bbox": [ + 151, + 699, + 159, + 710 + ], + "score": 0.83, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "zeroing out unnecessary invariants; further results in Appendix D. The fluctuations in", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 426, + 722 + ], + "score": 1.0, + "content": "accuracy around iterations 750 to 1000 in UCNN are also caused by penalising", + "type": "text" + }, + { + "bbox": [ + 426, + 710, + 435, + 721 + ], + "score": 0.83, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "which forces the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 720, + 389, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 389, + 734 + ], + "score": 1.0, + "content": "model to relearn the task with less variables half way through training.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 39.5, + "bbox_fs": [ + 104, + 600, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 124, + 122, + 487, + 186 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 80, + 505, + 114 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 79, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 229, + 93 + ], + "score": 1.0, + "content": "Table 2: Aggregate error rates", + "type": "text" + }, + { + "bbox": [ + 229, + 81, + 245, + 91 + ], + "score": 0.72, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 79, + 505, + 93 + ], + "score": 1.0, + "content": "on bAbI 1k for UMN and N2N, GN2N, MemNN by Sukhbaatar", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 90, + 505, + 104 + ], + "spans": [ + { + "bbox": [ + 105, + 90, + 505, + 104 + ], + "score": 1.0, + "content": "et al. (2015), Liu & Perez (2017) and Weston et al. (2015) respectively. Full comparison on individ-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 102, + 260, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 260, + 115 + ], + "score": 1.0, + "content": "ual tasks are available in Appendix D.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 124, + 122, + 487, + 186 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 122, + 487, + 186 + ], + "spans": [ + { + "bbox": [ + 124, + 122, + 487, + 186 + ], + "score": 0.981, + "html": "
Training Size Supervision #Invs /Model1k501k
Weak 13Strong 3WeakQRNStrong MemNN
31N2N GN2NEntNet
Mean19.120.56.36.027.613.912.729.611.36.7
#>5%810441711101554
", + "type": "table", + "image_path": "768570d0a06568bcbbcbc3b99176f6ae5e403c8c7a9bb3cef63a6be9c26ee2f2.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 124, + 122, + 487, + 143.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 124, + 143.33333333333334, + 487, + 164.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 124, + 164.66666666666669, + 487, + 186.00000000000003 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 205, + 505, + 326 + ], + "lines": [ + { + "bbox": [ + 105, + 205, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 505, + 218 + ], + "score": 1.0, + "content": "Following Tables 2 and 3, we observe a trend of better performance through strong supervision,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 216, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 229 + ], + "score": 1.0, + "content": "more data per task and using only 1 invariant. We believe strong supervision aids with selecting", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 227, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 506, + 239 + ], + "score": 1.0, + "content": "the correct sentences to unify and in a weak setting the model attempts to unify arbitrary context", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 238, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 505, + 251 + ], + "score": 1.0, + "content": "sentences often failing to follow the iterative reasoning chain. The increase in performance with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 248, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 448, + 262 + ], + "score": 1.0, + "content": "more data and strong supervision is consistent with previous work reflecting how", + "type": "text" + }, + { + "bbox": [ + 449, + 249, + 473, + 261 + ], + "score": 0.92, + "content": "f \\circ g", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 248, + 505, + 262 + ], + "score": 1.0, + "content": "can be", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 259, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 218, + 273 + ], + "score": 1.0, + "content": "bounded by the efficacy of", + "type": "text" + }, + { + "bbox": [ + 218, + 261, + 226, + 272 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 259, + 505, + 273 + ], + "score": 1.0, + "content": "modelled as a memory network. As a result, only in the supervised", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 271, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 505, + 283 + ], + "score": 1.0, + "content": "case do we observe a minor improvement over MemNN by 0.7 in Table 2 and no improvement in the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 282, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 452, + 294 + ], + "score": 1.0, + "content": "weak case over comparable memory based networks in Table 3. This dependency on", + "type": "text" + }, + { + "bbox": [ + 452, + 282, + 460, + 293 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 282, + 505, + 294 + ], + "score": 1.0, + "content": "also limits", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 292, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 161, + 306 + ], + "score": 1.0, + "content": "the ability of", + "type": "text" + }, + { + "bbox": [ + 161, + 293, + 183, + 304 + ], + "score": 0.92, + "content": "f \\circ g", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 292, + 505, + 306 + ], + "score": 1.0, + "content": "to learn from 50 examples per task failing 17/20 and 12/12 of bAbI and logical", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "reasoning tasks respectively. The increase in error rate with 3 invariants, we speculate, stems from", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 315, + 470, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 470, + 328 + ], + "score": 1.0, + "content": "having more parameters and more pathways in the model rendering training more difficult.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 110, + 372, + 501, + 446 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 329, + 504, + 363 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 328, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 245, + 343 + ], + "score": 1.0, + "content": "Table 3: Aggregate task error rates", + "type": "text" + }, + { + "bbox": [ + 245, + 330, + 261, + 340 + ], + "score": 0.68, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 328, + 505, + 343 + ], + "score": 1.0, + "content": "on the logical reasoning dataset for UMN and DMN, IMA by", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 354 + ], + "score": 1.0, + "content": "Cingillioglu & Russo (2019). Strong supervision, more data and only 1 invariant seem to improve", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 352, + 491, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 491, + 364 + ], + "score": 1.0, + "content": "the performance of UMN over plain iterative models. Individual task results are in Appendix D.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "table_body", + "bbox": [ + 110, + 372, + 501, + 446 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 372, + 501, + 446 + ], + "spans": [ + { + "bbox": [ + 110, + 372, + 501, + 446 + ], + "score": 0.98, + "html": "
Training Size Supervision1k5020k Weak
WeakStrongWeakStrong
121212122222
# Invs /Model1313131133DMNIMA
Mean36.439.314.328.921.531.82.412.216.047.121.29.1
#>5%91171171015912115
", + "type": "table", + "image_path": "1e018f6c32d109a7fac63e2f804566d081fd024bb3fdbeb94c09620e36bd8f82.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 110, + 372, + 501, + 396.6666666666667 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 110, + 396.6666666666667, + 501, + 421.33333333333337 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 110, + 421.33333333333337, + 501, + 446.00000000000006 + ], + "spans": [], + "index": 22 + } + ] + } + ], + "index": 19.5 + }, + { + "type": "title", + "bbox": [ + 107, + 460, + 178, + 472 + ], + "lines": [ + { + "bbox": [ + 105, + 457, + 181, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 181, + 474 + ], + "score": 1.0, + "content": "6 ANALYSIS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 484, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 426, + 497 + ], + "score": 1.0, + "content": "After training, we can extract the learned invariants by applying a threshold on", + "type": "text" + }, + { + "bbox": [ + 426, + 484, + 504, + 497 + ], + "score": 0.91, + "content": "\\forall s \\in \\mathbb { S } : \\psi ( s ) > t", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 495, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 104, + 495, + 376, + 509 + ], + "score": 1.0, + "content": "indicating whether a symbol is used as a variable or not. We set", + "type": "text" + }, + { + "bbox": [ + 376, + 496, + 411, + 506 + ], + "score": 0.9, + "content": "t = 0 . 0", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 495, + 506, + 509 + ], + "score": 1.0, + "content": "for all datasets except", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 180, + 519 + ], + "score": 1.0, + "content": "for bAbI, we use", + "type": "text" + }, + { + "bbox": [ + 181, + 507, + 216, + 516 + ], + "score": 0.9, + "content": "t = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 506, + 506, + 519 + ], + "score": 1.0, + "content": ". The magnitude of this threshold seems to depend on the amount of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 516, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 104, + 516, + 164, + 532 + ], + "score": 1.0, + "content": "regularisation", + "type": "text" + }, + { + "bbox": [ + 164, + 519, + 171, + 528 + ], + "score": 0.71, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 516, + 506, + 532 + ], + "score": 1.0, + "content": ", equation 5, and the number of training steps along with batch size all controlling", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 149, + 541 + ], + "score": 1.0, + "content": "how much", + "type": "text" + }, + { + "bbox": [ + 149, + 529, + 158, + 540 + ], + "score": 0.85, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "is pushed towards 0. Sample invariants shown in Figure 4 describe the common patterns", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "score": 1.0, + "content": "present in the tasks with parts that contribute towards the final answer becoming variables. Extra", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 551, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 562 + ], + "score": 1.0, + "content": "symbols such as is or travelled do not emerge as variables, as shown in Figure 4a; we attribute", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 562, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 505, + 573 + ], + "score": 1.0, + "content": "this behaviour to the fact that changing the token travelled to went does not influence the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 104, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "prediction but changing the action, the value of Z:left to ‘picked’ does. However, based on random", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 446, + 595 + ], + "score": 1.0, + "content": "initialisation, our approach can convert an arbitrary symbol into a variable and let", + "type": "text" + }, + { + "bbox": [ + 446, + 583, + 454, + 595 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "compensate", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 593, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 354, + 607 + ], + "score": 1.0, + "content": "for the unifications it produces. For example, the invariant", + "type": "text" + }, + { + "bbox": [ + 354, + 594, + 403, + 605 + ], + "score": 0.68, + "content": "^ { * } X { : } 8 \\ 5 \\ 2 \\ 2 ^ { , }", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 593, + 506, + 607 + ], + "score": 1.0, + "content": "could predict the tail of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 334, + 618 + ], + "score": 1.0, + "content": "another example by unifying the head with the tail using", + "type": "text" + }, + { + "bbox": [ + 335, + 605, + 348, + 617 + ], + "score": 0.89, + "content": "\\phi _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 604, + 506, + 618 + ], + "score": 1.0, + "content": ", equation 3, of those symbols. Further", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 304, + 629 + ], + "score": 1.0, + "content": "examples are shown in Appendix D. Pre-training", + "type": "text" + }, + { + "bbox": [ + 305, + 617, + 312, + 628 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "as done in UMN seems to produce more robust", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 344, + 640 + ], + "score": 1.0, + "content": "and consistent invariants compared to immediately training", + "type": "text" + }, + { + "bbox": [ + 344, + 627, + 366, + 639 + ], + "score": 0.91, + "content": "f \\circ g", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "since, we speculate, by equation 4", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 638, + 243, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 114, + 650 + ], + "score": 0.82, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 638, + 184, + 650 + ], + "score": 1.0, + "content": "might encourage", + "type": "text" + }, + { + "bbox": [ + 185, + 638, + 239, + 650 + ], + "score": 0.93, + "content": "g ( I , K ) \\approx K", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 638, + 243, + 650 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "Interpretability versus Ability A desired property of interpretable models is transparency (Lipton,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 665, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 679 + ], + "score": 1.0, + "content": "2018). A novel outcome of the learned invariants in our approach is that they provide an approx-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "imation of the underlying general principle present in the data such as the structure of multi-hop", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "reasoning shown in Figure 4e. However, certain aspects regarding the ability of the model such as", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 348, + 712 + ], + "score": 1.0, + "content": "how it performs temporal reasoning, are still hidden inside", + "type": "text" + }, + { + "bbox": [ + 349, + 699, + 356, + 711 + ], + "score": 0.83, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 699, + 505, + 712 + ], + "score": 1.0, + "content": ". In Figure 4b, although we observe", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 114, + 720 + ], + "score": 0.47, + "content": "\\mathbf { Z }", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 709, + 505, + 722 + ], + "score": 1.0, + "content": ":morning as a variable, the overall learned invariant captures nothing about how changing the value", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 117, + 734 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 721, + 125, + 730 + ], + "score": 0.4, + "content": "\\mathbf { Z }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 719, + 257, + 734 + ], + "score": 1.0, + "content": ":morning alters the behaviour of", + "type": "text" + }, + { + "bbox": [ + 257, + 721, + 264, + 732 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 719, + 505, + 734 + ], + "score": 1.0, + "content": ". The model might look before or after a certain time point", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 124, + 122, + 487, + 186 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 80, + 505, + 114 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 79, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 229, + 93 + ], + "score": 1.0, + "content": "Table 2: Aggregate error rates", + "type": "text" + }, + { + "bbox": [ + 229, + 81, + 245, + 91 + ], + "score": 0.72, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 79, + 505, + 93 + ], + "score": 1.0, + "content": "on bAbI 1k for UMN and N2N, GN2N, MemNN by Sukhbaatar", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 90, + 505, + 104 + ], + "spans": [ + { + "bbox": [ + 105, + 90, + 505, + 104 + ], + "score": 1.0, + "content": "et al. (2015), Liu & Perez (2017) and Weston et al. (2015) respectively. Full comparison on individ-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 102, + 260, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 260, + 115 + ], + "score": 1.0, + "content": "ual tasks are available in Appendix D.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 124, + 122, + 487, + 186 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 122, + 487, + 186 + ], + "spans": [ + { + "bbox": [ + 124, + 122, + 487, + 186 + ], + "score": 0.981, + "html": "
Training Size Supervision #Invs /Model1k501k
Weak 13Strong 3WeakQRNStrong MemNN
31N2N GN2NEntNet
Mean19.120.56.36.027.613.912.729.611.36.7
#>5%810441711101554
", + "type": "table", + "image_path": "768570d0a06568bcbbcbc3b99176f6ae5e403c8c7a9bb3cef63a6be9c26ee2f2.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 124, + 122, + 487, + 143.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 124, + 143.33333333333334, + 487, + 164.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 124, + 164.66666666666669, + 487, + 186.00000000000003 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 205, + 505, + 326 + ], + "lines": [ + { + "bbox": [ + 105, + 205, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 505, + 218 + ], + "score": 1.0, + "content": "Following Tables 2 and 3, we observe a trend of better performance through strong supervision,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 216, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 229 + ], + "score": 1.0, + "content": "more data per task and using only 1 invariant. We believe strong supervision aids with selecting", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 227, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 506, + 239 + ], + "score": 1.0, + "content": "the correct sentences to unify and in a weak setting the model attempts to unify arbitrary context", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 238, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 505, + 251 + ], + "score": 1.0, + "content": "sentences often failing to follow the iterative reasoning chain. The increase in performance with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 248, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 448, + 262 + ], + "score": 1.0, + "content": "more data and strong supervision is consistent with previous work reflecting how", + "type": "text" + }, + { + "bbox": [ + 449, + 249, + 473, + 261 + ], + "score": 0.92, + "content": "f \\circ g", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 248, + 505, + 262 + ], + "score": 1.0, + "content": "can be", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 259, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 218, + 273 + ], + "score": 1.0, + "content": "bounded by the efficacy of", + "type": "text" + }, + { + "bbox": [ + 218, + 261, + 226, + 272 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 259, + 505, + 273 + ], + "score": 1.0, + "content": "modelled as a memory network. As a result, only in the supervised", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 271, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 505, + 283 + ], + "score": 1.0, + "content": "case do we observe a minor improvement over MemNN by 0.7 in Table 2 and no improvement in the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 282, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 452, + 294 + ], + "score": 1.0, + "content": "weak case over comparable memory based networks in Table 3. This dependency on", + "type": "text" + }, + { + "bbox": [ + 452, + 282, + 460, + 293 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 282, + 505, + 294 + ], + "score": 1.0, + "content": "also limits", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 292, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 161, + 306 + ], + "score": 1.0, + "content": "the ability of", + "type": "text" + }, + { + "bbox": [ + 161, + 293, + 183, + 304 + ], + "score": 0.92, + "content": "f \\circ g", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 292, + 505, + 306 + ], + "score": 1.0, + "content": "to learn from 50 examples per task failing 17/20 and 12/12 of bAbI and logical", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "reasoning tasks respectively. The increase in error rate with 3 invariants, we speculate, stems from", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 315, + 470, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 470, + 328 + ], + "score": 1.0, + "content": "having more parameters and more pathways in the model rendering training more difficult.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 205, + 506, + 328 + ] + }, + { + "type": "table", + "bbox": [ + 110, + 372, + 501, + 446 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 329, + 504, + 363 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 328, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 245, + 343 + ], + "score": 1.0, + "content": "Table 3: Aggregate task error rates", + "type": "text" + }, + { + "bbox": [ + 245, + 330, + 261, + 340 + ], + "score": 0.68, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 328, + 505, + 343 + ], + "score": 1.0, + "content": "on the logical reasoning dataset for UMN and DMN, IMA by", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 354 + ], + "score": 1.0, + "content": "Cingillioglu & Russo (2019). Strong supervision, more data and only 1 invariant seem to improve", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 352, + 491, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 491, + 364 + ], + "score": 1.0, + "content": "the performance of UMN over plain iterative models. Individual task results are in Appendix D.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "table_body", + "bbox": [ + 110, + 372, + 501, + 446 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 372, + 501, + 446 + ], + "spans": [ + { + "bbox": [ + 110, + 372, + 501, + 446 + ], + "score": 0.98, + "html": "
Training Size Supervision1k5020k Weak
WeakStrongWeakStrong
121212122222
# Invs /Model1313131133DMNIMA
Mean36.439.314.328.921.531.82.412.216.047.121.29.1
#>5%91171171015912115
", + "type": "table", + "image_path": "1e018f6c32d109a7fac63e2f804566d081fd024bb3fdbeb94c09620e36bd8f82.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 110, + 372, + 501, + 396.6666666666667 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 110, + 396.6666666666667, + 501, + 421.33333333333337 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 110, + 421.33333333333337, + 501, + 446.00000000000006 + ], + "spans": [], + "index": 22 + } + ] + } + ], + "index": 19.5 + }, + { + "type": "title", + "bbox": [ + 107, + 460, + 178, + 472 + ], + "lines": [ + { + "bbox": [ + 105, + 457, + 181, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 181, + 474 + ], + "score": 1.0, + "content": "6 ANALYSIS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 484, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 426, + 497 + ], + "score": 1.0, + "content": "After training, we can extract the learned invariants by applying a threshold on", + "type": "text" + }, + { + "bbox": [ + 426, + 484, + 504, + 497 + ], + "score": 0.91, + "content": "\\forall s \\in \\mathbb { S } : \\psi ( s ) > t", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 495, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 104, + 495, + 376, + 509 + ], + "score": 1.0, + "content": "indicating whether a symbol is used as a variable or not. We set", + "type": "text" + }, + { + "bbox": [ + 376, + 496, + 411, + 506 + ], + "score": 0.9, + "content": "t = 0 . 0", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 495, + 506, + 509 + ], + "score": 1.0, + "content": "for all datasets except", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 180, + 519 + ], + "score": 1.0, + "content": "for bAbI, we use", + "type": "text" + }, + { + "bbox": [ + 181, + 507, + 216, + 516 + ], + "score": 0.9, + "content": "t = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 506, + 506, + 519 + ], + "score": 1.0, + "content": ". The magnitude of this threshold seems to depend on the amount of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 516, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 104, + 516, + 164, + 532 + ], + "score": 1.0, + "content": "regularisation", + "type": "text" + }, + { + "bbox": [ + 164, + 519, + 171, + 528 + ], + "score": 0.71, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 516, + 506, + 532 + ], + "score": 1.0, + "content": ", equation 5, and the number of training steps along with batch size all controlling", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 149, + 541 + ], + "score": 1.0, + "content": "how much", + "type": "text" + }, + { + "bbox": [ + 149, + 529, + 158, + 540 + ], + "score": 0.85, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "is pushed towards 0. Sample invariants shown in Figure 4 describe the common patterns", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "score": 1.0, + "content": "present in the tasks with parts that contribute towards the final answer becoming variables. Extra", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 551, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 562 + ], + "score": 1.0, + "content": "symbols such as is or travelled do not emerge as variables, as shown in Figure 4a; we attribute", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 562, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 505, + 573 + ], + "score": 1.0, + "content": "this behaviour to the fact that changing the token travelled to went does not influence the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 104, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "prediction but changing the action, the value of Z:left to ‘picked’ does. However, based on random", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 446, + 595 + ], + "score": 1.0, + "content": "initialisation, our approach can convert an arbitrary symbol into a variable and let", + "type": "text" + }, + { + "bbox": [ + 446, + 583, + 454, + 595 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "compensate", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 593, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 354, + 607 + ], + "score": 1.0, + "content": "for the unifications it produces. For example, the invariant", + "type": "text" + }, + { + "bbox": [ + 354, + 594, + 403, + 605 + ], + "score": 0.68, + "content": "^ { * } X { : } 8 \\ 5 \\ 2 \\ 2 ^ { , }", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 593, + 506, + 607 + ], + "score": 1.0, + "content": "could predict the tail of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 334, + 618 + ], + "score": 1.0, + "content": "another example by unifying the head with the tail using", + "type": "text" + }, + { + "bbox": [ + 335, + 605, + 348, + 617 + ], + "score": 0.89, + "content": "\\phi _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 604, + 506, + 618 + ], + "score": 1.0, + "content": ", equation 3, of those symbols. Further", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 304, + 629 + ], + "score": 1.0, + "content": "examples are shown in Appendix D. Pre-training", + "type": "text" + }, + { + "bbox": [ + 305, + 617, + 312, + 628 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "as done in UMN seems to produce more robust", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 344, + 640 + ], + "score": 1.0, + "content": "and consistent invariants compared to immediately training", + "type": "text" + }, + { + "bbox": [ + 344, + 627, + 366, + 639 + ], + "score": 0.91, + "content": "f \\circ g", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "since, we speculate, by equation 4", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 638, + 243, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 114, + 650 + ], + "score": 0.82, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 638, + 184, + 650 + ], + "score": 1.0, + "content": "might encourage", + "type": "text" + }, + { + "bbox": [ + 185, + 638, + 239, + 650 + ], + "score": 0.93, + "content": "g ( I , K ) \\approx K", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 638, + 243, + 650 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 31, + "bbox_fs": [ + 104, + 484, + 506, + 650 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "Interpretability versus Ability A desired property of interpretable models is transparency (Lipton,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 665, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 679 + ], + "score": 1.0, + "content": "2018). A novel outcome of the learned invariants in our approach is that they provide an approx-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "imation of the underlying general principle present in the data such as the structure of multi-hop", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "reasoning shown in Figure 4e. However, certain aspects regarding the ability of the model such as", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 348, + 712 + ], + "score": 1.0, + "content": "how it performs temporal reasoning, are still hidden inside", + "type": "text" + }, + { + "bbox": [ + 349, + 699, + 356, + 711 + ], + "score": 0.83, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 699, + 505, + 712 + ], + "score": 1.0, + "content": ". In Figure 4b, although we observe", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 114, + 720 + ], + "score": 0.47, + "content": "\\mathbf { Z }", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 709, + 505, + 722 + ], + "score": 1.0, + "content": ":morning as a variable, the overall learned invariant captures nothing about how changing the value", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 117, + 734 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 721, + 125, + 730 + ], + "score": 0.4, + "content": "\\mathbf { Z }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 719, + 257, + 734 + ], + "score": 1.0, + "content": ":morning alters the behaviour of", + "type": "text" + }, + { + "bbox": [ + 257, + 721, + 264, + 732 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 719, + 505, + 734 + ], + "score": 1.0, + "content": ". The model might look before or after a certain time point", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 654, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 81, + 273, + 134 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 135, + 81, + 273, + 134 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 81, + 273, + 134 + ], + "spans": [ + { + "bbox": [ + 135, + 81, + 273, + 134 + ], + "score": 0.343, + "html": "
Y:john Z:left the X:football Y:john travelled to the A:office
where is the X:football
A:office
", + "type": "table", + "image_path": "7747279f4df11d55eb0355fa48db5cdb6952a0dfb81cec0c6489109d22418593.jpg" + } + ] + } + ], + "index": 0.5, + "virtual_lines": [ + { + "bbox": [ + 135, + 81, + 273, + 107.5 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 135, + 107.5, + 273, + 134.0 + ], + "spans": [], + "index": 1 + } + ] + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 313, + 81, + 496, + 134 + ], + "lines": [ + { + "bbox": [ + 320, + 81, + 492, + 93 + ], + "spans": [ + { + "bbox": [ + 320, + 81, + 382, + 93 + ], + "score": 1.0, + "content": "this Z:morning", + "type": "text" + }, + { + "bbox": [ + 382, + 82, + 390, + 92 + ], + "score": 0.33, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 81, + 492, + 93 + ], + "score": 1.0, + "content": ":bill went to the Y:school", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 318, + 95, + 483, + 107 + ], + "spans": [ + { + "bbox": [ + 318, + 95, + 360, + 107 + ], + "score": 1.0, + "content": "yesterday", + "type": "text" + }, + { + "bbox": [ + 361, + 95, + 369, + 105 + ], + "score": 0.6, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 95, + 483, + 107 + ], + "score": 1.0, + "content": ":bill journeyed to the A:park", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 320, + 109, + 471, + 120 + ], + "spans": [ + { + "bbox": [ + 320, + 109, + 364, + 120 + ], + "score": 1.0, + "content": "where was", + "type": "text" + }, + { + "bbox": [ + 365, + 109, + 373, + 119 + ], + "score": 0.38, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 109, + 471, + 120 + ], + "score": 1.0, + "content": ":bill before the Y:school", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 318, + 120, + 352, + 136 + ], + "spans": [ + { + "bbox": [ + 318, + 120, + 352, + 136 + ], + "score": 1.0, + "content": "A:park", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 140, + 302, + 171 + ], + "lines": [ + { + "bbox": [ + 106, + 140, + 303, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 303, + 150 + ], + "score": 1.0, + "content": "(a) bAbI task 2, two supporting facts. The model also", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 149, + 303, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 303, + 161 + ], + "score": 1.0, + "content": "learns Z:left since people can also drop or pick up ob-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 159, + 244, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 244, + 171 + ], + "score": 1.0, + "content": "jects potentially affecting the answer.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 308, + 140, + 504, + 170 + ], + "lines": [ + { + "bbox": [ + 308, + 140, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 431, + 150 + ], + "score": 1.0, + "content": "(b) bAbI task 14, time reasoning.", + "type": "text" + }, + { + "bbox": [ + 432, + 140, + 439, + 149 + ], + "score": 0.36, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 140, + 505, + 150 + ], + "score": 1.0, + "content": ":bill and Y:school", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 148, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 306, + 148, + 505, + 162 + ], + "score": 1.0, + "content": "are recognised as variables alongside Z:morning cap-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 159, + 504, + 170 + ], + "spans": [ + { + "bbox": [ + 307, + 159, + 504, + 170 + ], + "score": 1.0, + "content": "turing when someone went which is crucial to this task.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 106, + 189, + 501, + 272 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 189, + 501, + 272 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 189, + 501, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 501, + 272 + ], + "score": 0.125, + "type": "image", + "image_path": "2b918647c3c931a6d2d86be39b48a163abe2920f4808362f723da0efbea11058.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 106, + 189, + 501, + 216.66666666666666 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 106, + 216.66666666666666, + 501, + 244.33333333333331 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 106, + 244.33333333333331, + 501, + 272.0 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 280, + 504, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 280, + 504, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 504, + 292 + ], + "score": 1.0, + "content": "Figure 4: Invariants learned across the four datasets using the three architectures. For iterative", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 291, + 486, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 486, + 303 + ], + "score": 1.0, + "content": "reasoning datasets, bAbI and logical reasoning, they are taken from strongly supervised UMN.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 325, + 505, + 392 + ], + "lines": [ + { + "bbox": [ + 106, + 326, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 114, + 336 + ], + "score": 0.55, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 326, + 266, + 339 + ], + "score": 1.0, + "content": ":bill went somewhere depending what", + "type": "text" + }, + { + "bbox": [ + 266, + 326, + 274, + 336 + ], + "score": 0.27, + "content": "\\mathbf { Z }", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 326, + 480, + 339 + ], + "score": 1.0, + "content": ":morning binds to. Without the regularising term on", + "type": "text" + }, + { + "bbox": [ + 481, + 326, + 501, + 338 + ], + "score": 0.9, + "content": "\\psi ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 326, + 505, + 339 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 336, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 349 + ], + "score": 1.0, + "content": "we initially noticed the models using, one might call extra, symbols as variables and binding them to", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 348, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 360 + ], + "score": 1.0, + "content": "the same value occasionally producing unifications such as “bathroom bathroom to the bathroom”", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 357, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 141, + 372 + ], + "score": 1.0, + "content": "and still", + "type": "text" + }, + { + "bbox": [ + 141, + 359, + 149, + 370 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 357, + 505, + 372 + ], + "score": 1.0, + "content": "predicting, perhaps unsurprisingly, the correct answer as bathroom. Hence, regularising", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 369, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 107, + 371, + 115, + 381 + ], + "score": 0.83, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 369, + 218, + 382 + ], + "score": 1.0, + "content": "with the correct amount", + "type": "text" + }, + { + "bbox": [ + 218, + 372, + 225, + 380 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 369, + 505, + 382 + ], + "score": 1.0, + "content": "in equation 5 to reduce the capacity of unification seems critical in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 381, + 419, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 419, + 393 + ], + "score": 1.0, + "content": "extracting not just any invariant but one that represents the common structure.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 505, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "score": 1.0, + "content": "Soft unification from equation 3 reveals three main patterns: one-to-one, one-to-many or many-to-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "one bindings as shown in Figure 5; further examples are in Appendix D. Figure 5a captures what one", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 418, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 506, + 433 + ], + "score": 1.0, + "content": "might expect unification to look like where variables unify with their corresponding counterparts,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 430, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 124, + 444 + ], + "score": 1.0, + "content": "e.g.", + "type": "text" + }, + { + "bbox": [ + 124, + 431, + 133, + 441 + ], + "score": 0.71, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 430, + 506, + 444 + ], + "score": 1.0, + "content": ":bernhard with brian and Y:frog with lion. However, occasionally the model can optimise to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 440, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 104, + 440, + 506, + 455 + ], + "score": 1.0, + "content": "use less variables and squeeze the required information into a single variable, for example by binding", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 451, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 466 + ], + "score": 1.0, + "content": "Y:bathroom to john and kitchen as shown in Figure 5b. We believe this occurs due to the sparsity", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 160, + 476 + ], + "score": 1.0, + "content": "constraint on", + "type": "text" + }, + { + "bbox": [ + 160, + 464, + 180, + 475 + ], + "score": 0.92, + "content": "\\psi ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 463, + 505, + 476 + ], + "score": 1.0, + "content": "encouraging the model to be as conservative as possible. Finally, if there are more", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 473, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 506, + 488 + ], + "score": 1.0, + "content": "variables than needed as in Figure 5c, we observe a many-to-one binding with Y:w and Z:e mapping", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 190, + 498 + ], + "score": 1.0, + "content": "to the same constant", + "type": "text" + }, + { + "bbox": [ + 191, + 487, + 197, + 497 + ], + "score": 0.7, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 485, + 505, + 498 + ], + "score": 1.0, + "content": ". This behaviour begs the question how does the model differentiate between", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 496, + 504, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 125, + 508 + ], + "score": 0.91, + "content": "p ( q )", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 496, + 144, + 509 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 144, + 496, + 172, + 508 + ], + "score": 0.92, + "content": "p ( q , q )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 496, + 367, + 509 + ], + "score": 1.0, + "content": ". We speculate the model uses the magnitude of", + "type": "text" + }, + { + "bbox": [ + 367, + 496, + 428, + 508 + ], + "score": 0.91, + "content": "\\psi ( w ) = 0 . 0 3 7", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 496, + 446, + 509 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 446, + 496, + 504, + 508 + ], + "score": 0.91, + "content": "\\psi ( e ) = 0 . 0 4 2", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 507, + 426, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 426, + 520 + ], + "score": 1.0, + "content": "to encode the difference despite both variables unifying with the same constant.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 667, + 236, + 707 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 237, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 237, + 677 + ], + "score": 1.0, + "content": "(a) bAbI task 16. A one-to-one", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 677, + 236, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 236, + 686 + ], + "score": 1.0, + "content": "mapping is created between vari-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 686, + 236, + 697 + ], + "spans": [ + { + "bbox": [ + 106, + 686, + 129, + 697 + ], + "score": 1.0, + "content": "ables", + "type": "text" + }, + { + "bbox": [ + 130, + 687, + 138, + 696 + ], + "score": 0.69, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 686, + 236, + 697 + ], + "score": 1.0, + "content": ":bernhard with brian and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 695, + 171, + 708 + ], + "spans": [ + { + "bbox": [ + 105, + 695, + 171, + 708 + ], + "score": 1.0, + "content": "Y:frog with lion.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5 + }, + { + "type": "image", + "bbox": [ + 111, + 543, + 475, + 665 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 543, + 475, + 665 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 543, + 475, + 665 + ], + "spans": [ + { + "bbox": [ + 111, + 543, + 475, + 665 + ], + "score": 0.882, + "type": "image", + "image_path": "87f061c60b0d26309091e0532f96894f1603bdbd0e818eb3a44cf2cee5b59e9e.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 111, + 543, + 475, + 583.6666666666666 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 111, + 583.6666666666666, + 475, + 624.3333333333333 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 111, + 624.3333333333333, + 475, + 664.9999999999999 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 374, + 667, + 504, + 707 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 374, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 374, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "(c) Logical reasoning task 1. An", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 374, + 677, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 374, + 677, + 505, + 687 + ], + "score": 1.0, + "content": "arity 2 predicate is forced to bind", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 375, + 686, + 505, + 698 + ], + "spans": [ + { + "bbox": [ + 375, + 686, + 505, + 698 + ], + "score": 1.0, + "content": "with arity 1 creating a many-to-one", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 374, + 694, + 409, + 709 + ], + "spans": [ + { + "bbox": [ + 374, + 694, + 409, + 709 + ], + "score": 1.0, + "content": "binding.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5 + } + ], + "index": 40.75 + }, + { + "type": "text", + "bbox": [ + 240, + 668, + 370, + 708 + ], + "lines": [ + { + "bbox": [ + 240, + 667, + 371, + 678 + ], + "spans": [ + { + "bbox": [ + 240, + 667, + 371, + 678 + ], + "score": 1.0, + "content": "(b) bAbI task 6. Only Y:bathroom", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 240, + 677, + 371, + 689 + ], + "spans": [ + { + "bbox": [ + 240, + 677, + 371, + 689 + ], + "score": 1.0, + "content": "is recognised as variable creating a", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 240, + 688, + 370, + 698 + ], + "spans": [ + { + "bbox": [ + 240, + 688, + 370, + 698 + ], + "score": 1.0, + "content": "one-to-many binding to capture the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 240, + 698, + 309, + 708 + ], + "spans": [ + { + "bbox": [ + 240, + 698, + 309, + 708 + ], + "score": 1.0, + "content": "same information.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 107, + 717, + 503, + 729 + ], + "lines": [ + { + "bbox": [ + 106, + 716, + 505, + 730 + ], + "spans": [ + { + "bbox": [ + 106, + 716, + 505, + 730 + ], + "score": 1.0, + "content": "Figure 5: Variable bindings produced from equation 3. Darker cells indicate higher attention values.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 49 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 81, + 273, + 134 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 135, + 81, + 273, + 134 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 81, + 273, + 134 + ], + "spans": [ + { + "bbox": [ + 135, + 81, + 273, + 134 + ], + "score": 0.343, + "html": "
Y:john Z:left the X:football Y:john travelled to the A:office
where is the X:football
A:office
", + "type": "table", + "image_path": "7747279f4df11d55eb0355fa48db5cdb6952a0dfb81cec0c6489109d22418593.jpg" + } + ] + } + ], + "index": 0.5, + "virtual_lines": [ + { + "bbox": [ + 135, + 81, + 273, + 107.5 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 135, + 107.5, + 273, + 134.0 + ], + "spans": [], + "index": 1 + } + ] + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 313, + 81, + 496, + 134 + ], + "lines": [ + { + "bbox": [ + 320, + 81, + 492, + 93 + ], + "spans": [ + { + "bbox": [ + 320, + 81, + 382, + 93 + ], + "score": 1.0, + "content": "this Z:morning", + "type": "text" + }, + { + "bbox": [ + 382, + 82, + 390, + 92 + ], + "score": 0.33, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 81, + 492, + 93 + ], + "score": 1.0, + "content": ":bill went to the Y:school", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 318, + 95, + 483, + 107 + ], + "spans": [ + { + "bbox": [ + 318, + 95, + 360, + 107 + ], + "score": 1.0, + "content": "yesterday", + "type": "text" + }, + { + "bbox": [ + 361, + 95, + 369, + 105 + ], + "score": 0.6, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 95, + 483, + 107 + ], + "score": 1.0, + "content": ":bill journeyed to the A:park", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 320, + 109, + 471, + 120 + ], + "spans": [ + { + "bbox": [ + 320, + 109, + 364, + 120 + ], + "score": 1.0, + "content": "where was", + "type": "text" + }, + { + "bbox": [ + 365, + 109, + 373, + 119 + ], + "score": 0.38, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 109, + 471, + 120 + ], + "score": 1.0, + "content": ":bill before the Y:school", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 318, + 120, + 352, + 136 + ], + "spans": [ + { + "bbox": [ + 318, + 120, + 352, + 136 + ], + "score": 1.0, + "content": "A:park", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 318, + 81, + 492, + 136 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 140, + 302, + 171 + ], + "lines": [ + { + "bbox": [ + 106, + 140, + 303, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 303, + 150 + ], + "score": 1.0, + "content": "(a) bAbI task 2, two supporting facts. The model also", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 149, + 303, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 303, + 161 + ], + "score": 1.0, + "content": "learns Z:left since people can also drop or pick up ob-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 159, + 244, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 244, + 171 + ], + "score": 1.0, + "content": "jects potentially affecting the answer.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 140, + 303, + 171 + ] + }, + { + "type": "text", + "bbox": [ + 308, + 140, + 504, + 170 + ], + "lines": [ + { + "bbox": [ + 308, + 140, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 431, + 150 + ], + "score": 1.0, + "content": "(b) bAbI task 14, time reasoning.", + "type": "text" + }, + { + "bbox": [ + 432, + 140, + 439, + 149 + ], + "score": 0.36, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 140, + 505, + 150 + ], + "score": 1.0, + "content": ":bill and Y:school", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 148, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 306, + 148, + 505, + 162 + ], + "score": 1.0, + "content": "are recognised as variables alongside Z:morning cap-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 159, + 504, + 170 + ], + "spans": [ + { + "bbox": [ + 307, + 159, + 504, + 170 + ], + "score": 1.0, + "content": "turing when someone went which is crucial to this task.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 306, + 140, + 505, + 170 + ] + }, + { + "type": "image", + "bbox": [ + 106, + 189, + 501, + 272 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 189, + 501, + 272 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 189, + 501, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 501, + 272 + ], + "score": 0.125, + "type": "image", + "image_path": "2b918647c3c931a6d2d86be39b48a163abe2920f4808362f723da0efbea11058.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 106, + 189, + 501, + 216.66666666666666 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 106, + 216.66666666666666, + 501, + 244.33333333333331 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 106, + 244.33333333333331, + 501, + 272.0 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 280, + 504, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 280, + 504, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 504, + 292 + ], + "score": 1.0, + "content": "Figure 4: Invariants learned across the four datasets using the three architectures. For iterative", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 291, + 486, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 486, + 303 + ], + "score": 1.0, + "content": "reasoning datasets, bAbI and logical reasoning, they are taken from strongly supervised UMN.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 280, + 504, + 303 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 325, + 505, + 392 + ], + "lines": [ + { + "bbox": [ + 106, + 326, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 114, + 336 + ], + "score": 0.55, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 326, + 266, + 339 + ], + "score": 1.0, + "content": ":bill went somewhere depending what", + "type": "text" + }, + { + "bbox": [ + 266, + 326, + 274, + 336 + ], + "score": 0.27, + "content": "\\mathbf { Z }", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 326, + 480, + 339 + ], + "score": 1.0, + "content": ":morning binds to. Without the regularising term on", + "type": "text" + }, + { + "bbox": [ + 481, + 326, + 501, + 338 + ], + "score": 0.9, + "content": "\\psi ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 326, + 505, + 339 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 336, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 349 + ], + "score": 1.0, + "content": "we initially noticed the models using, one might call extra, symbols as variables and binding them to", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 348, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 360 + ], + "score": 1.0, + "content": "the same value occasionally producing unifications such as “bathroom bathroom to the bathroom”", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 357, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 141, + 372 + ], + "score": 1.0, + "content": "and still", + "type": "text" + }, + { + "bbox": [ + 141, + 359, + 149, + 370 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 357, + 505, + 372 + ], + "score": 1.0, + "content": "predicting, perhaps unsurprisingly, the correct answer as bathroom. Hence, regularising", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 369, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 107, + 371, + 115, + 381 + ], + "score": 0.83, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 369, + 218, + 382 + ], + "score": 1.0, + "content": "with the correct amount", + "type": "text" + }, + { + "bbox": [ + 218, + 372, + 225, + 380 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 369, + 505, + 382 + ], + "score": 1.0, + "content": "in equation 5 to reduce the capacity of unification seems critical in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 381, + 419, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 419, + 393 + ], + "score": 1.0, + "content": "extracting not just any invariant but one that represents the common structure.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 326, + 506, + 393 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 505, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "score": 1.0, + "content": "Soft unification from equation 3 reveals three main patterns: one-to-one, one-to-many or many-to-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "one bindings as shown in Figure 5; further examples are in Appendix D. Figure 5a captures what one", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 418, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 506, + 433 + ], + "score": 1.0, + "content": "might expect unification to look like where variables unify with their corresponding counterparts,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 430, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 124, + 444 + ], + "score": 1.0, + "content": "e.g.", + "type": "text" + }, + { + "bbox": [ + 124, + 431, + 133, + 441 + ], + "score": 0.71, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 430, + 506, + 444 + ], + "score": 1.0, + "content": ":bernhard with brian and Y:frog with lion. However, occasionally the model can optimise to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 440, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 104, + 440, + 506, + 455 + ], + "score": 1.0, + "content": "use less variables and squeeze the required information into a single variable, for example by binding", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 451, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 466 + ], + "score": 1.0, + "content": "Y:bathroom to john and kitchen as shown in Figure 5b. We believe this occurs due to the sparsity", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 160, + 476 + ], + "score": 1.0, + "content": "constraint on", + "type": "text" + }, + { + "bbox": [ + 160, + 464, + 180, + 475 + ], + "score": 0.92, + "content": "\\psi ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 463, + 505, + 476 + ], + "score": 1.0, + "content": "encouraging the model to be as conservative as possible. Finally, if there are more", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 473, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 506, + 488 + ], + "score": 1.0, + "content": "variables than needed as in Figure 5c, we observe a many-to-one binding with Y:w and Z:e mapping", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 190, + 498 + ], + "score": 1.0, + "content": "to the same constant", + "type": "text" + }, + { + "bbox": [ + 191, + 487, + 197, + 497 + ], + "score": 0.7, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 485, + 505, + 498 + ], + "score": 1.0, + "content": ". This behaviour begs the question how does the model differentiate between", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 496, + 504, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 125, + 508 + ], + "score": 0.91, + "content": "p ( q )", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 496, + 144, + 509 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 144, + 496, + 172, + 508 + ], + "score": 0.92, + "content": "p ( q , q )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 496, + 367, + 509 + ], + "score": 1.0, + "content": ". We speculate the model uses the magnitude of", + "type": "text" + }, + { + "bbox": [ + 367, + 496, + 428, + 508 + ], + "score": 0.91, + "content": "\\psi ( w ) = 0 . 0 3 7", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 496, + 446, + 509 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 446, + 496, + 504, + 508 + ], + "score": 0.91, + "content": "\\psi ( e ) = 0 . 0 4 2", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 507, + 426, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 426, + 520 + ], + "score": 1.0, + "content": "to encode the difference despite both variables unifying with the same constant.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 28, + "bbox_fs": [ + 104, + 397, + 506, + 520 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 667, + 236, + 707 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 237, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 237, + 677 + ], + "score": 1.0, + "content": "(a) bAbI task 16. A one-to-one", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 677, + 236, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 236, + 686 + ], + "score": 1.0, + "content": "mapping is created between vari-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 686, + 236, + 697 + ], + "spans": [ + { + "bbox": [ + 106, + 686, + 129, + 697 + ], + "score": 1.0, + "content": "ables", + "type": "text" + }, + { + "bbox": [ + 130, + 687, + 138, + 696 + ], + "score": 0.69, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 686, + 236, + 697 + ], + "score": 1.0, + "content": ":bernhard with brian and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 695, + 171, + 708 + ], + "spans": [ + { + "bbox": [ + 105, + 695, + 171, + 708 + ], + "score": 1.0, + "content": "Y:frog with lion.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 666, + 237, + 708 + ] + }, + { + "type": "image", + "bbox": [ + 111, + 543, + 475, + 665 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 543, + 475, + 665 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 543, + 475, + 665 + ], + "spans": [ + { + "bbox": [ + 111, + 543, + 475, + 665 + ], + "score": 0.882, + "type": "image", + "image_path": "87f061c60b0d26309091e0532f96894f1603bdbd0e818eb3a44cf2cee5b59e9e.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 111, + 543, + 475, + 583.6666666666666 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 111, + 583.6666666666666, + 475, + 624.3333333333333 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 111, + 624.3333333333333, + 475, + 664.9999999999999 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 374, + 667, + 504, + 707 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 374, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 374, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "(c) Logical reasoning task 1. An", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 374, + 677, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 374, + 677, + 505, + 687 + ], + "score": 1.0, + "content": "arity 2 predicate is forced to bind", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 375, + 686, + 505, + 698 + ], + "spans": [ + { + "bbox": [ + 375, + 686, + 505, + 698 + ], + "score": 1.0, + "content": "with arity 1 creating a many-to-one", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 374, + 694, + 409, + 709 + ], + "spans": [ + { + "bbox": [ + 374, + 694, + 409, + 709 + ], + "score": 1.0, + "content": "binding.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5 + } + ], + "index": 40.75 + }, + { + "type": "text", + "bbox": [ + 240, + 668, + 370, + 708 + ], + "lines": [ + { + "bbox": [ + 240, + 667, + 371, + 678 + ], + "spans": [ + { + "bbox": [ + 240, + 667, + 371, + 678 + ], + "score": 1.0, + "content": "(b) bAbI task 6. Only Y:bathroom", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 240, + 677, + 371, + 689 + ], + "spans": [ + { + "bbox": [ + 240, + 677, + 371, + 689 + ], + "score": 1.0, + "content": "is recognised as variable creating a", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 240, + 688, + 370, + 698 + ], + "spans": [ + { + "bbox": [ + 240, + 688, + 370, + 698 + ], + "score": 1.0, + "content": "one-to-many binding to capture the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 240, + 698, + 309, + 708 + ], + "spans": [ + { + "bbox": [ + 240, + 698, + 309, + 708 + ], + "score": 1.0, + "content": "same information.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5, + "bbox_fs": [ + 240, + 667, + 371, + 708 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 717, + 503, + 729 + ], + "lines": [ + { + "bbox": [ + 106, + 716, + 505, + 730 + ], + "spans": [ + { + "bbox": [ + 106, + 716, + 505, + 730 + ], + "score": 1.0, + "content": "Figure 5: Variable bindings produced from equation 3. Darker cells indicate higher attention values.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 49, + "bbox_fs": [ + 106, + 716, + 505, + 730 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 210, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 213, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 213, + 95 + ], + "score": 1.0, + "content": "7 RELATED WORK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 111, + 505, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 506, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 506, + 125 + ], + "score": 1.0, + "content": "Learning an underlying general principle in the form of an invariant is often the means for arguing", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 122, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 506, + 135 + ], + "score": 1.0, + "content": "generalisation in neural networks. For example, Neural Turing Machines (Graves et al., 2014) are", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 133, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 506, + 145 + ], + "score": 1.0, + "content": "tested on previously unseen sequences to support the view that the model might have captured the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 145, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 505, + 156 + ], + "score": 1.0, + "content": "underlying pattern or algorithm. In fact, Weston et al. (2015) claim “MemNNs can discover simple", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 155, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 506, + 168 + ], + "score": 1.0, + "content": "linguistic patterns based on verbal forms such as (X, dropped, Y), (X, took, Y) or (X, journeyed", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 166, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 506, + 179 + ], + "score": 1.0, + "content": "to, Y) and can successfully generalise the meaning of their instantiations.” However, this claim is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 176, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 196, + 190 + ], + "score": 1.0, + "content": "based on the output of", + "type": "text" + }, + { + "bbox": [ + 197, + 177, + 204, + 188 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 176, + 506, + 190 + ], + "score": 1.0, + "content": "and unfortunately it is unknown whether the model has truly learned such a", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 187, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 104, + 187, + 506, + 201 + ], + "score": 1.0, + "content": "representation or indeed is utilising it. Our approach sheds light to this ambiguity and presents these", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 198, + 506, + 213 + ], + "spans": [ + { + "bbox": [ + 104, + 198, + 395, + 213 + ], + "score": 1.0, + "content": "linguistic patterns explicitly as invariants ensuring their utility through", + "type": "text" + }, + { + "bbox": [ + 395, + 200, + 402, + 210 + ], + "score": 0.74, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 198, + 506, + 213 + ], + "score": 1.0, + "content": "without solely analysing", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 159, + 222 + ], + "score": 1.0, + "content": "the output of", + "type": "text" + }, + { + "bbox": [ + 159, + 210, + 166, + 221 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "on previously unseen symbols. Although we associate these invariants with our exist-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 219, + 506, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 235 + ], + "score": 1.0, + "content": "ing understanding of the task to mistakenly anthropomorphise the machine, for example by thinking", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 163, + 245 + ], + "score": 1.0, + "content": "it has learned", + "type": "text" + }, + { + "bbox": [ + 163, + 232, + 172, + 241 + ], + "score": 0.74, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 231, + 505, + 245 + ], + "score": 1.0, + "content": ":mary as someone, it is important to acknowledge that these are just symbolic pat-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 242, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 506, + 255 + ], + "score": 1.0, + "content": "terns. In these cases, our interpretations do not necessarily correspond to any understanding of the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 253, + 396, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 396, + 266 + ], + "score": 1.0, + "content": "machine, relating to the Chinese room argument made by Searle (1980).", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 270, + 505, + 402 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 284 + ], + "score": 1.0, + "content": "Learning invariants by lifting ground examples is related to least common generalisation (Reynolds,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 281, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 295 + ], + "score": 1.0, + "content": "1970) by which inductive inference is performed on facts (Shapiro, 1981) such as generalising", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 293, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 305 + ], + "score": 1.0, + "content": "went(mary,kitchen) and went(john,garden) to went(X,Y). Unlike in a predicate logic setting, our", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 303, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 317 + ], + "score": 1.0, + "content": "approach allows for soft alignment and therefore generalisation between varying length sequences.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "Existing neuro-symbolic systems (Broda et al., 2002) focus on inducing rules that adhere to given", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 359, + 338 + ], + "score": 1.0, + "content": "logical semantics of what variables and rules are. For example,", + "type": "text" + }, + { + "bbox": [ + 359, + 325, + 385, + 335 + ], + "score": 0.36, + "content": "\\delta I L P", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "(Evans & Grefenstette, 2018)", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "score": 1.0, + "content": "constructs a network by rigidly following the given semantics of first-order logic. Similarly, Lifted", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 347, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 505, + 359 + ], + "score": 1.0, + "content": "Relational Neural Networks (Sourek et al., 2015) ground first-order logic rules into a neural network", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 358, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 505, + 371 + ], + "score": 1.0, + "content": "while Neural Theorem Provers (Rocktschel & Riedel, 2017) build neural networks using backward-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 368, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 383 + ], + "score": 1.0, + "content": "chaining (Russell & Norvig, 2016) on a given background knowledge base with templates. However,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 379, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 505, + 394 + ], + "score": 1.0, + "content": "the notion of a variable is pre-defined rather than learned with a focus on presenting a practical ap-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 391, + 488, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 488, + 404 + ], + "score": 1.0, + "content": "proach to solving certain problems, whereas our motivation stems from a cognitive perspective.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 105, + 407, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 421 + ], + "score": 1.0, + "content": "At first it may seem the learned invariants, Section 6, make the model more interpretable; however,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 418, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 257, + 432 + ], + "score": 1.0, + "content": "this transparency is not of the model", + "type": "text" + }, + { + "bbox": [ + 257, + 419, + 265, + 430 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 418, + 506, + 432 + ], + "score": 1.0, + "content": "but of the data. The invariant captures patterns that poten-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 476, + 443 + ], + "score": 1.0, + "content": "tially approximates the data generating distribution but we still do not know how the model", + "type": "text" + }, + { + "bbox": [ + 477, + 430, + 484, + 441 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 429, + 505, + 443 + ], + "score": 1.0, + "content": "uses", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "them upstream. Thus, from the perspective of explainable artificial intelligence (XAI) (Adadi &", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "Berrada, 2018), learning invariants or interpreting them do not constitute an explanation of the rea-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 163, + 475 + ], + "score": 1.0, + "content": "soning model", + "type": "text" + }, + { + "bbox": [ + 163, + 463, + 171, + 474 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "even though “if someone goes somewhere then they are there” might look like one.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "Instead, it can be perceived as causal attribution (Miller, 2019) in which someone being somewhere", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "is attributed to them going there. This perspective also relates to gradient based model explanation", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "methods such as Layer-Wise Relevance Propagation (Bach et al., 2015) and Grad-CAM (Selvaraju", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 409, + 520 + ], + "score": 1.0, + "content": "et al., 2017; Chattopadhay et al., 2018). Consequently, a possible view on", + "type": "text" + }, + { + "bbox": [ + 409, + 507, + 417, + 518 + ], + "score": 0.84, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 506, + 506, + 520 + ], + "score": 1.0, + "content": ", equation 2, is a gra-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 516, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 397, + 530 + ], + "score": 1.0, + "content": "dient based usefulness measure such that a symbol utilised upstream by", + "type": "text" + }, + { + "bbox": [ + 397, + 518, + 405, + 529 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 516, + 505, + 530 + ], + "score": 1.0, + "content": "to determine the answer", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "becomes a variable similar to how a group of pixels in an image contribute more to its classification.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 545, + 505, + 611 + ], + "lines": [ + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "Finally, one can argue that our model maintains a form of counterfactual thinking (Roese, 1997)", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 556, + 504, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 207, + 568 + ], + "score": 1.0, + "content": "in which soft unification", + "type": "text" + }, + { + "bbox": [ + 207, + 559, + 214, + 568 + ], + "score": 0.79, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 556, + 496, + 568 + ], + "score": 1.0, + "content": "creates counterfactuals on the invariant example to alter the output of", + "type": "text" + }, + { + "bbox": [ + 497, + 557, + 504, + 568 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "towards the desired answer, equation 4. The question where Mary would have been if Mary had gone", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "to the garden instead of the kitchen is the process by which an invariant is learned through multiple", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 589, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 505, + 602 + ], + "score": 1.0, + "content": "examples during training. This view relates to methods of causal inference (Pearl, 2019; Holland,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 600, + 485, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 485, + 613 + ], + "score": 1.0, + "content": "1986) in which counterfactuals are vital as demonstrated in structured models by Pearl (1999).", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + }, + { + "type": "title", + "bbox": [ + 108, + 636, + 195, + 648 + ], + "lines": [ + { + "bbox": [ + 104, + 633, + 197, + 652 + ], + "spans": [ + { + "bbox": [ + 104, + 633, + 197, + 652 + ], + "score": 1.0, + "content": "8 CONCLUSION", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "We presented a new approach for learning variables and lifting examples into invariants through", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "the usage of soft unification. Evaluating on four datasets, we analysed how Unification Networks", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "perform comparatively to existing similar architectures while having the benefit of lifting examples", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "into invariants that capture underlying patterns present in the tasks. Since our approach is end-to-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "end differentiable, we plan to apply this technique to multi-modal tasks in order to yield multi-modal", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 720, + 316, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 316, + 734 + ], + "score": 1.0, + "content": "invariants for example in visual question answering.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 48.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 210, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 213, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 213, + 95 + ], + "score": 1.0, + "content": "7 RELATED WORK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 111, + 505, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 506, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 506, + 125 + ], + "score": 1.0, + "content": "Learning an underlying general principle in the form of an invariant is often the means for arguing", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 122, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 506, + 135 + ], + "score": 1.0, + "content": "generalisation in neural networks. For example, Neural Turing Machines (Graves et al., 2014) are", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 133, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 506, + 145 + ], + "score": 1.0, + "content": "tested on previously unseen sequences to support the view that the model might have captured the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 145, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 505, + 156 + ], + "score": 1.0, + "content": "underlying pattern or algorithm. In fact, Weston et al. (2015) claim “MemNNs can discover simple", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 155, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 506, + 168 + ], + "score": 1.0, + "content": "linguistic patterns based on verbal forms such as (X, dropped, Y), (X, took, Y) or (X, journeyed", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 166, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 506, + 179 + ], + "score": 1.0, + "content": "to, Y) and can successfully generalise the meaning of their instantiations.” However, this claim is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 176, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 196, + 190 + ], + "score": 1.0, + "content": "based on the output of", + "type": "text" + }, + { + "bbox": [ + 197, + 177, + 204, + 188 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 176, + 506, + 190 + ], + "score": 1.0, + "content": "and unfortunately it is unknown whether the model has truly learned such a", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 187, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 104, + 187, + 506, + 201 + ], + "score": 1.0, + "content": "representation or indeed is utilising it. Our approach sheds light to this ambiguity and presents these", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 198, + 506, + 213 + ], + "spans": [ + { + "bbox": [ + 104, + 198, + 395, + 213 + ], + "score": 1.0, + "content": "linguistic patterns explicitly as invariants ensuring their utility through", + "type": "text" + }, + { + "bbox": [ + 395, + 200, + 402, + 210 + ], + "score": 0.74, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 198, + 506, + 213 + ], + "score": 1.0, + "content": "without solely analysing", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 159, + 222 + ], + "score": 1.0, + "content": "the output of", + "type": "text" + }, + { + "bbox": [ + 159, + 210, + 166, + 221 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "on previously unseen symbols. Although we associate these invariants with our exist-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 219, + 506, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 235 + ], + "score": 1.0, + "content": "ing understanding of the task to mistakenly anthropomorphise the machine, for example by thinking", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 163, + 245 + ], + "score": 1.0, + "content": "it has learned", + "type": "text" + }, + { + "bbox": [ + 163, + 232, + 172, + 241 + ], + "score": 0.74, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 231, + 505, + 245 + ], + "score": 1.0, + "content": ":mary as someone, it is important to acknowledge that these are just symbolic pat-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 242, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 506, + 255 + ], + "score": 1.0, + "content": "terns. In these cases, our interpretations do not necessarily correspond to any understanding of the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 253, + 396, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 396, + 266 + ], + "score": 1.0, + "content": "machine, relating to the Chinese room argument made by Searle (1980).", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 7.5, + "bbox_fs": [ + 104, + 110, + 506, + 266 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 270, + 505, + 402 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 284 + ], + "score": 1.0, + "content": "Learning invariants by lifting ground examples is related to least common generalisation (Reynolds,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 281, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 295 + ], + "score": 1.0, + "content": "1970) by which inductive inference is performed on facts (Shapiro, 1981) such as generalising", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 293, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 305 + ], + "score": 1.0, + "content": "went(mary,kitchen) and went(john,garden) to went(X,Y). Unlike in a predicate logic setting, our", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 303, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 317 + ], + "score": 1.0, + "content": "approach allows for soft alignment and therefore generalisation between varying length sequences.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "Existing neuro-symbolic systems (Broda et al., 2002) focus on inducing rules that adhere to given", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 359, + 338 + ], + "score": 1.0, + "content": "logical semantics of what variables and rules are. For example,", + "type": "text" + }, + { + "bbox": [ + 359, + 325, + 385, + 335 + ], + "score": 0.36, + "content": "\\delta I L P", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "(Evans & Grefenstette, 2018)", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "score": 1.0, + "content": "constructs a network by rigidly following the given semantics of first-order logic. Similarly, Lifted", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 347, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 505, + 359 + ], + "score": 1.0, + "content": "Relational Neural Networks (Sourek et al., 2015) ground first-order logic rules into a neural network", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 358, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 505, + 371 + ], + "score": 1.0, + "content": "while Neural Theorem Provers (Rocktschel & Riedel, 2017) build neural networks using backward-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 368, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 383 + ], + "score": 1.0, + "content": "chaining (Russell & Norvig, 2016) on a given background knowledge base with templates. However,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 379, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 505, + 394 + ], + "score": 1.0, + "content": "the notion of a variable is pre-defined rather than learned with a focus on presenting a practical ap-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 391, + 488, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 488, + 404 + ], + "score": 1.0, + "content": "proach to solving certain problems, whereas our motivation stems from a cognitive perspective.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 270, + 505, + 404 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 105, + 407, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 421 + ], + "score": 1.0, + "content": "At first it may seem the learned invariants, Section 6, make the model more interpretable; however,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 418, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 257, + 432 + ], + "score": 1.0, + "content": "this transparency is not of the model", + "type": "text" + }, + { + "bbox": [ + 257, + 419, + 265, + 430 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 418, + 506, + 432 + ], + "score": 1.0, + "content": "but of the data. The invariant captures patterns that poten-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 476, + 443 + ], + "score": 1.0, + "content": "tially approximates the data generating distribution but we still do not know how the model", + "type": "text" + }, + { + "bbox": [ + 477, + 430, + 484, + 441 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 429, + 505, + 443 + ], + "score": 1.0, + "content": "uses", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "them upstream. Thus, from the perspective of explainable artificial intelligence (XAI) (Adadi &", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "Berrada, 2018), learning invariants or interpreting them do not constitute an explanation of the rea-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 163, + 475 + ], + "score": 1.0, + "content": "soning model", + "type": "text" + }, + { + "bbox": [ + 163, + 463, + 171, + 474 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "even though “if someone goes somewhere then they are there” might look like one.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "Instead, it can be perceived as causal attribution (Miller, 2019) in which someone being somewhere", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "is attributed to them going there. This perspective also relates to gradient based model explanation", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "methods such as Layer-Wise Relevance Propagation (Bach et al., 2015) and Grad-CAM (Selvaraju", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 409, + 520 + ], + "score": 1.0, + "content": "et al., 2017; Chattopadhay et al., 2018). Consequently, a possible view on", + "type": "text" + }, + { + "bbox": [ + 409, + 507, + 417, + 518 + ], + "score": 0.84, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 506, + 506, + 520 + ], + "score": 1.0, + "content": ", equation 2, is a gra-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 516, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 397, + 530 + ], + "score": 1.0, + "content": "dient based usefulness measure such that a symbol utilised upstream by", + "type": "text" + }, + { + "bbox": [ + 397, + 518, + 405, + 529 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 516, + 505, + 530 + ], + "score": 1.0, + "content": "to determine the answer", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "becomes a variable similar to how a group of pixels in an image contribute more to its classification.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 407, + 506, + 541 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 545, + 505, + 611 + ], + "lines": [ + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "Finally, one can argue that our model maintains a form of counterfactual thinking (Roese, 1997)", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 556, + 504, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 207, + 568 + ], + "score": 1.0, + "content": "in which soft unification", + "type": "text" + }, + { + "bbox": [ + 207, + 559, + 214, + 568 + ], + "score": 0.79, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 556, + 496, + 568 + ], + "score": 1.0, + "content": "creates counterfactuals on the invariant example to alter the output of", + "type": "text" + }, + { + "bbox": [ + 497, + 557, + 504, + 568 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "towards the desired answer, equation 4. The question where Mary would have been if Mary had gone", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "to the garden instead of the kitchen is the process by which an invariant is learned through multiple", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 589, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 505, + 602 + ], + "score": 1.0, + "content": "examples during training. This view relates to methods of causal inference (Pearl, 2019; Holland,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 600, + 485, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 485, + 613 + ], + "score": 1.0, + "content": "1986) in which counterfactuals are vital as demonstrated in structured models by Pearl (1999).", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 545, + 505, + 613 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 636, + 195, + 648 + ], + "lines": [ + { + "bbox": [ + 104, + 633, + 197, + 652 + ], + "spans": [ + { + "bbox": [ + 104, + 633, + 197, + 652 + ], + "score": 1.0, + "content": "8 CONCLUSION", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "We presented a new approach for learning variables and lifting examples into invariants through", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "the usage of soft unification. Evaluating on four datasets, we analysed how Unification Networks", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "perform comparatively to existing similar architectures while having the benefit of lifting examples", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "into invariants that capture underlying patterns present in the tasks. Since our approach is end-to-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "end differentiable, we plan to apply this technique to multi-modal tasks in order to yield multi-modal", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 720, + 316, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 316, + 734 + ], + "score": 1.0, + "content": "invariants for example in visual question answering.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 48.5, + "bbox_fs": [ + 105, + 665, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 100, + 505, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "Amina Adadi and Mohammed Berrada. Peeking inside the black-box: A survey on explainable", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "artificial intelligence (XAI). IEEE Access, 6:52138–52160, 2018. doi: 10.1109/access.2018.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 120, + 157, + 134 + ], + "spans": [ + { + "bbox": [ + 115, + 120, + 157, + 134 + ], + "score": 1.0, + "content": "2870052.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 140, + 505, + 184 + ], + "lines": [ + { + "bbox": [ + 105, + 140, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 505, + 153 + ], + "score": 1.0, + "content": "Sebastian Bach, Alexander Binder, Gregoire Montavon, Frederick Klauschen, Klaus-Robert Mller, ´", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 151, + 504, + 163 + ], + "spans": [ + { + "bbox": [ + 116, + 151, + 504, + 163 + ], + "score": 1.0, + "content": "and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 162, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 162, + 505, + 176 + ], + "score": 1.0, + "content": "wise relevance propagation. PLOS ONE, 10(7):e0130140, July 2015. doi: 10.1371/journal.pone.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 173, + 156, + 184 + ], + "spans": [ + { + "bbox": [ + 115, + 173, + 156, + 184 + ], + "score": 1.0, + "content": "0130140.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 105, + 191, + 504, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 190, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 205 + ], + "score": 1.0, + "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 202, + 338, + 214 + ], + "spans": [ + { + "bbox": [ + 115, + 202, + 338, + 214 + ], + "score": 1.0, + "content": "learning to align and translate. ICLR, September 2015.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 105, + 220, + 504, + 244 + ], + "lines": [ + { + "bbox": [ + 105, + 219, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 234 + ], + "score": 1.0, + "content": "Krysia B. Broda, Artur S. D’Avila Garcez, and Dov M. Gabbay. Neural-Symbolic Learning Systems.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 232, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 116, + 232, + 294, + 243 + ], + "score": 1.0, + "content": "Springer London, 2002. ISBN 1852335122.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 505, + 295 + ], + "lines": [ + { + "bbox": [ + 106, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "Aditya Chattopadhay, Anirban Sarkar, Prantik Howlader, and Vineeth N. Balasubramanian. Grad-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 261, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 116, + 261, + 140, + 274 + ], + "score": 1.0, + "content": "CAM", + "type": "text" + }, + { + "bbox": [ + 140, + 263, + 151, + 271 + ], + "score": 0.52, + "content": "^ { + + }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 261, + 505, + 274 + ], + "score": 1.0, + "content": ": Generalized gradient-based visual explanations for deep convolutional networks. In", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 272, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 115, + 272, + 506, + 285 + ], + "score": 1.0, + "content": "2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 839–847. IEEE,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 284, + 326, + 294 + ], + "spans": [ + { + "bbox": [ + 116, + 284, + 326, + 294 + ], + "score": 1.0, + "content": "IEEE, March 2018. doi: 10.1109/wacv.2018.00097.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 105, + 301, + 504, + 325 + ], + "lines": [ + { + "bbox": [ + 105, + 300, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 506, + 315 + ], + "score": 1.0, + "content": "Sneha Chaudhari, Gungor Polatkan, Rohan Ramanath, and Varun Mithal. An attentive survey of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 312, + 265, + 325 + ], + "spans": [ + { + "bbox": [ + 116, + 312, + 265, + 325 + ], + "score": 1.0, + "content": "attention models. IJCAI, April 2019.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 331, + 505, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 115, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "of neural machine translation: Encoder–decoder approaches. In Proceedings of SSST-8, Eighth", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 353, + 504, + 366 + ], + "spans": [ + { + "bbox": [ + 116, + 353, + 504, + 366 + ], + "score": 1.0, + "content": "Workshop on Syntax, Semantics and Structure in Statistical Translation. Association for Compu-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 364, + 336, + 376 + ], + "spans": [ + { + "bbox": [ + 115, + 364, + 336, + 376 + ], + "score": 1.0, + "content": "tational Linguistics, 2014. doi: 10.3115/v1/w14-4012.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 107, + 382, + 503, + 405 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 396 + ], + "score": 1.0, + "content": "Nuri Cingillioglu and Alessandra Russo. Deeplogic: Towards end-to-end differentiable logical rea-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 394, + 252, + 406 + ], + "spans": [ + { + "bbox": [ + 115, + 394, + 252, + 406 + ], + "score": 1.0, + "content": "soning. AAAI-MAKE, May 2019.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 411, + 503, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "Keith L. Clark. Negation as failure. In Logic and Data Bases, pp. 293–322. Springer US, 1978. doi:", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 117, + 423, + 250, + 434 + ], + "spans": [ + { + "bbox": [ + 117, + 423, + 250, + 434 + ], + "score": 1.0, + "content": "10.1007/978-1-4684-3384-5 11.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 105, + 441, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 506, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 506, + 454 + ], + "score": 1.0, + "content": "Richard Evans and Edward Grefenstette. Learning explanatory rules from noisy data. Journal of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 451, + 437, + 464 + ], + "spans": [ + { + "bbox": [ + 115, + 451, + 437, + 464 + ], + "score": 1.0, + "content": "Artificial Intelligence Research, 61:1–64, January 2018. doi: 10.1613/jair.5714.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 105, + 470, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 106, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "Gottlob Frege. Sense and reference. The Philosophical Review, 57(3):209, May 1948. doi: 10.2307/", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 480, + 158, + 494 + ], + "spans": [ + { + "bbox": [ + 115, + 480, + 158, + 494 + ], + "score": 1.0, + "content": "2181485.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 105, + 500, + 504, + 523 + ], + "lines": [ + { + "bbox": [ + 106, + 499, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 505, + 512 + ], + "score": 1.0, + "content": "Joe L. Frost, Pei-San Brown, John A. Sutterby, and Candra D. Thornton. The Developmental Benefits", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 117, + 511, + 498, + 523 + ], + "spans": [ + { + "bbox": [ + 117, + 511, + 498, + 523 + ], + "score": 1.0, + "content": "Of Playgrounds. Association for Childhood Education International, 2004. ISBN 0871731649.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 529, + 492, + 541 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 492, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 492, + 542 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv:1410.5401, 2014.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 547, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 106, + 547, + 504, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 504, + 560 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka Grabska-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 558, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 115, + 558, + 505, + 573 + ], + "score": 1.0, + "content": "Barwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, ´", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 569, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 115, + 569, + 505, + 583 + ], + "score": 1.0, + "content": "Adria Puigdom ` enech Badia, Karl Moritz Hermann, Yori Zwols, Georg Ostrovski, Adam Cain, `", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 580, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 115, + 580, + 505, + 594 + ], + "score": 1.0, + "content": "Helen King, Christopher Summerfield, Phil Blunsom, Koray Kavukcuoglu, and Demis Hassabis.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 590, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 115, + 590, + 505, + 605 + ], + "score": 1.0, + "content": "Hybrid computing using a neural network with dynamic external memory. Nature, 538(7626):", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 603, + 325, + 614 + ], + "spans": [ + { + "bbox": [ + 115, + 603, + 325, + 614 + ], + "score": 1.0, + "content": "471–476, October 2016. doi: 10.1038/nature20101.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 106, + 621, + 504, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "Mikael Henaff, Jason Weston, Arthur Szlam, Antoine Bordes, and Yann LeCun. Tracking the world", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 632, + 316, + 644 + ], + "spans": [ + { + "bbox": [ + 115, + 632, + 316, + 644 + ], + "score": 1.0, + "content": "state with recurrent entity networks. ICLR, 2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 106, + 650, + 503, + 673 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 664 + ], + "score": 1.0, + "content": "Paul W. Holland. Statistics and causal inference. Journal of the American Statistical Association,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 662, + 423, + 673 + ], + "spans": [ + { + "bbox": [ + 115, + 662, + 423, + 673 + ], + "score": 1.0, + "content": "81(396):945–960, December 1986. doi: 10.1080/01621459.1986.10478354.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 680, + 487, + 692 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 487, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 487, + 693 + ], + "score": 1.0, + "content": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. ICLR, 2015.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Ankit Kumar, Ozan Irsoy, Peter Ondruska, Mohit Iyyer, James Bradbury, Ishaan Gulrajani, Victor", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "Zhong, Romain Paulus, and Richard Socher. Ask me anything: Dynamic memory networks for", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 721, + 363, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 363, + 732 + ], + "score": 1.0, + "content": "natural language processing. In ICML, pp. 1378–1387, 2016.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 100, + 505, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "Amina Adadi and Mohammed Berrada. Peeking inside the black-box: A survey on explainable", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "artificial intelligence (XAI). IEEE Access, 6:52138–52160, 2018. doi: 10.1109/access.2018.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 120, + 157, + 134 + ], + "spans": [ + { + "bbox": [ + 115, + 120, + 157, + 134 + ], + "score": 1.0, + "content": "2870052.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 100, + 505, + 134 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 140, + 505, + 184 + ], + "lines": [ + { + "bbox": [ + 105, + 140, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 505, + 153 + ], + "score": 1.0, + "content": "Sebastian Bach, Alexander Binder, Gregoire Montavon, Frederick Klauschen, Klaus-Robert Mller, ´", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 151, + 504, + 163 + ], + "spans": [ + { + "bbox": [ + 116, + 151, + 504, + 163 + ], + "score": 1.0, + "content": "and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 162, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 162, + 505, + 176 + ], + "score": 1.0, + "content": "wise relevance propagation. PLOS ONE, 10(7):e0130140, July 2015. doi: 10.1371/journal.pone.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 173, + 156, + 184 + ], + "spans": [ + { + "bbox": [ + 115, + 173, + 156, + 184 + ], + "score": 1.0, + "content": "0130140.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 140, + 505, + 184 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 191, + 504, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 190, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 205 + ], + "score": 1.0, + "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 202, + 338, + 214 + ], + "spans": [ + { + "bbox": [ + 115, + 202, + 338, + 214 + ], + "score": 1.0, + "content": "learning to align and translate. ICLR, September 2015.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 190, + 505, + 214 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 220, + 504, + 244 + ], + "lines": [ + { + "bbox": [ + 105, + 219, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 234 + ], + "score": 1.0, + "content": "Krysia B. Broda, Artur S. D’Avila Garcez, and Dov M. Gabbay. Neural-Symbolic Learning Systems.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 232, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 116, + 232, + 294, + 243 + ], + "score": 1.0, + "content": "Springer London, 2002. ISBN 1852335122.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 219, + 505, + 243 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 505, + 295 + ], + "lines": [ + { + "bbox": [ + 106, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "Aditya Chattopadhay, Anirban Sarkar, Prantik Howlader, and Vineeth N. Balasubramanian. Grad-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 261, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 116, + 261, + 140, + 274 + ], + "score": 1.0, + "content": "CAM", + "type": "text" + }, + { + "bbox": [ + 140, + 263, + 151, + 271 + ], + "score": 0.52, + "content": "^ { + + }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 261, + 505, + 274 + ], + "score": 1.0, + "content": ": Generalized gradient-based visual explanations for deep convolutional networks. In", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 272, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 115, + 272, + 506, + 285 + ], + "score": 1.0, + "content": "2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 839–847. IEEE,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 284, + 326, + 294 + ], + "spans": [ + { + "bbox": [ + 116, + 284, + 326, + 294 + ], + "score": 1.0, + "content": "IEEE, March 2018. doi: 10.1109/wacv.2018.00097.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 250, + 506, + 294 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 301, + 504, + 325 + ], + "lines": [ + { + "bbox": [ + 105, + 300, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 506, + 315 + ], + "score": 1.0, + "content": "Sneha Chaudhari, Gungor Polatkan, Rohan Ramanath, and Varun Mithal. An attentive survey of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 312, + 265, + 325 + ], + "spans": [ + { + "bbox": [ + 116, + 312, + 265, + 325 + ], + "score": 1.0, + "content": "attention models. IJCAI, April 2019.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 300, + 506, + 325 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 331, + 505, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 115, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "of neural machine translation: Encoder–decoder approaches. In Proceedings of SSST-8, Eighth", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 353, + 504, + 366 + ], + "spans": [ + { + "bbox": [ + 116, + 353, + 504, + 366 + ], + "score": 1.0, + "content": "Workshop on Syntax, Semantics and Structure in Statistical Translation. Association for Compu-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 364, + 336, + 376 + ], + "spans": [ + { + "bbox": [ + 115, + 364, + 336, + 376 + ], + "score": 1.0, + "content": "tational Linguistics, 2014. doi: 10.3115/v1/w14-4012.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 331, + 505, + 376 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 382, + 503, + 405 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 396 + ], + "score": 1.0, + "content": "Nuri Cingillioglu and Alessandra Russo. Deeplogic: Towards end-to-end differentiable logical rea-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 394, + 252, + 406 + ], + "spans": [ + { + "bbox": [ + 115, + 394, + 252, + 406 + ], + "score": 1.0, + "content": "soning. AAAI-MAKE, May 2019.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 381, + 505, + 406 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 411, + 503, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "Keith L. Clark. Negation as failure. In Logic and Data Bases, pp. 293–322. Springer US, 1978. doi:", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 117, + 423, + 250, + 434 + ], + "spans": [ + { + "bbox": [ + 117, + 423, + 250, + 434 + ], + "score": 1.0, + "content": "10.1007/978-1-4684-3384-5 11.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 106, + 412, + 505, + 434 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 441, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 506, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 506, + 454 + ], + "score": 1.0, + "content": "Richard Evans and Edward Grefenstette. Learning explanatory rules from noisy data. Journal of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 451, + 437, + 464 + ], + "spans": [ + { + "bbox": [ + 115, + 451, + 437, + 464 + ], + "score": 1.0, + "content": "Artificial Intelligence Research, 61:1–64, January 2018. doi: 10.1613/jair.5714.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 441, + 506, + 464 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 470, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 106, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "Gottlob Frege. Sense and reference. The Philosophical Review, 57(3):209, May 1948. doi: 10.2307/", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 480, + 158, + 494 + ], + "spans": [ + { + "bbox": [ + 115, + 480, + 158, + 494 + ], + "score": 1.0, + "content": "2181485.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 470, + 505, + 494 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 500, + 504, + 523 + ], + "lines": [ + { + "bbox": [ + 106, + 499, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 505, + 512 + ], + "score": 1.0, + "content": "Joe L. Frost, Pei-San Brown, John A. Sutterby, and Candra D. Thornton. The Developmental Benefits", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 117, + 511, + 498, + 523 + ], + "spans": [ + { + "bbox": [ + 117, + 511, + 498, + 523 + ], + "score": 1.0, + "content": "Of Playgrounds. Association for Childhood Education International, 2004. ISBN 0871731649.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 499, + 505, + 523 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 529, + 492, + 541 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 492, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 492, + 542 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv:1410.5401, 2014.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 528, + 492, + 542 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 547, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 106, + 547, + 504, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 504, + 560 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka Grabska-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 558, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 115, + 558, + 505, + 573 + ], + "score": 1.0, + "content": "Barwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, ´", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 569, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 115, + 569, + 505, + 583 + ], + "score": 1.0, + "content": "Adria Puigdom ` enech Badia, Karl Moritz Hermann, Yori Zwols, Georg Ostrovski, Adam Cain, `", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 580, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 115, + 580, + 505, + 594 + ], + "score": 1.0, + "content": "Helen King, Christopher Summerfield, Phil Blunsom, Koray Kavukcuoglu, and Demis Hassabis.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 590, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 115, + 590, + 505, + 605 + ], + "score": 1.0, + "content": "Hybrid computing using a neural network with dynamic external memory. Nature, 538(7626):", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 603, + 325, + 614 + ], + "spans": [ + { + "bbox": [ + 115, + 603, + 325, + 614 + ], + "score": 1.0, + "content": "471–476, October 2016. doi: 10.1038/nature20101.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5, + "bbox_fs": [ + 106, + 547, + 505, + 614 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 621, + 504, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "Mikael Henaff, Jason Weston, Arthur Szlam, Antoine Bordes, and Yann LeCun. Tracking the world", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 632, + 316, + 644 + ], + "spans": [ + { + "bbox": [ + 115, + 632, + 316, + 644 + ], + "score": 1.0, + "content": "state with recurrent entity networks. ICLR, 2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 106, + 621, + 505, + 644 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 650, + 503, + 673 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 664 + ], + "score": 1.0, + "content": "Paul W. Holland. Statistics and causal inference. Journal of the American Statistical Association,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 662, + 423, + 673 + ], + "spans": [ + { + "bbox": [ + 115, + 662, + 423, + 673 + ], + "score": 1.0, + "content": "81(396):945–960, December 1986. doi: 10.1080/01621459.1986.10478354.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 650, + 505, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 680, + 487, + 692 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 487, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 487, + 693 + ], + "score": 1.0, + "content": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. ICLR, 2015.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43, + "bbox_fs": [ + 106, + 678, + 487, + 693 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Ankit Kumar, Ozan Irsoy, Peter Ondruska, Mohit Iyyer, James Bradbury, Ishaan Gulrajani, Victor", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "Zhong, Romain Paulus, and Richard Socher. Ask me anything: Dynamic memory networks for", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 721, + 363, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 363, + 732 + ], + "score": 1.0, + "content": "natural language processing. In ICML, pp. 1378–1387, 2016.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 698, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "Richard L. Lewis. Cognitive modeling, symbolic. The MIT encyclopedia of the cognitive sciences,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 196, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 196, + 105 + ], + "score": 1.0, + "content": "pp. 525–527, 1999.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 105, + 112, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 112, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 112, + 505, + 125 + ], + "score": 1.0, + "content": "Zachary C. Lipton. The mythos of model interpretability. Communications of the ACM, 61(10):", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 123, + 311, + 135 + ], + "spans": [ + { + "bbox": [ + 116, + 123, + 311, + 135 + ], + "score": 1.0, + "content": "36–43, September 2018. doi: 10.1145/3233231.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 109, + 142, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 141, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 505, + 156 + ], + "score": 1.0, + "content": "Fei Liu and Julien Perez. Gated end-to-end memory networks. In Proceedings of the 15th Confer-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 153, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 115, + 153, + 505, + 167 + ], + "score": 1.0, + "content": "ence of the European Chapter of the Association for Computational Linguistics: Volume 1, Long", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 164, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 116, + 164, + 505, + 178 + ], + "score": 1.0, + "content": "Papers, pp. 1–10. Association for Computational Linguistics, 2017. doi: 10.18653/v1/e17-1001.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 183, + 505, + 228 + ], + "lines": [ + { + "bbox": [ + 106, + 184, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 505, + 196 + ], + "score": 1.0, + "content": "Thang Luong, Hieu Pham, and Christopher D. Manning. Effective approaches to attention-based", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 194, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 115, + 194, + 506, + 207 + ], + "score": 1.0, + "content": "neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 205, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 114, + 205, + 506, + 219 + ], + "score": 1.0, + "content": "Natural Language Processing. Association for Computational Linguistics, 2015. doi: 10.18653/", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 217, + 172, + 228 + ], + "spans": [ + { + "bbox": [ + 116, + 217, + 172, + 228 + ], + "score": 1.0, + "content": "v1/d15-1166.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 235, + 504, + 259 + ], + "lines": [ + { + "bbox": [ + 106, + 235, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 505, + 248 + ], + "score": 1.0, + "content": "Tim Miller. Explanation in artificial intelligence: Insights from the social sciences. Artificial Intel-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 247, + 395, + 259 + ], + "spans": [ + { + "bbox": [ + 115, + 247, + 395, + 259 + ], + "score": 1.0, + "content": "ligence, 267:1–38, February 2019. doi: 10.1016/j.artint.2018.07.007.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 265, + 504, + 289 + ], + "lines": [ + { + "bbox": [ + 106, + 266, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 505, + 279 + ], + "score": 1.0, + "content": "Adam Morton and Jerry A. Fodor. The Language of Thought., volume 75. Philosophy Documenta-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 276, + 313, + 288 + ], + "spans": [ + { + "bbox": [ + 116, + 276, + 313, + 288 + ], + "score": 1.0, + "content": "tion Center, March 1978. doi: 10.2307/2025426.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 295, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 106, + 295, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 505, + 309 + ], + "score": 1.0, + "content": "Stephen Muggleton and Luc de Raedt. Inductive logic programming: Theory and methods.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 306, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 115, + 306, + 506, + 321 + ], + "score": 1.0, + "content": "The Journal of Logic Programming, 19-20:629–679, May 1994. doi: 10.1016/0743-1066(94)", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 316, + 156, + 330 + ], + "spans": [ + { + "bbox": [ + 115, + 316, + 156, + 330 + ], + "score": 1.0, + "content": "90035-3.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 336, + 503, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 336, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 505, + 349 + ], + "score": 1.0, + "content": "Judea Pearl. Probabilities of causation: three counterfactual interpretations and their identification.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 348, + 254, + 360 + ], + "spans": [ + { + "bbox": [ + 115, + 348, + 254, + 360 + ], + "score": 1.0, + "content": "Synthese, 121(1-2):93–149, 1999.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 367, + 504, + 390 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 380 + ], + "score": 1.0, + "content": "Judea Pearl. The seven tools of causal inference with reflections on machine learning. Communica-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 378, + 401, + 391 + ], + "spans": [ + { + "bbox": [ + 116, + 378, + 401, + 391 + ], + "score": 1.0, + "content": "tions of the ACM, 62(3):54–60, February 2019. doi: 10.1145/3241036.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 104, + 397, + 470, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 396, + 470, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 470, + 410 + ], + "score": 1.0, + "content": "Jean Piaget. The Psychology of Intelligence. Routledge, August 2001. ISBN 0415254019.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 416, + 506, + 439 + ], + "lines": [ + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "John C. Reynolds. Transformational systems and algebraic structure of atomic formulas. Machine", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 427, + 240, + 440 + ], + "spans": [ + { + "bbox": [ + 116, + 427, + 240, + 440 + ], + "score": 1.0, + "content": "intelligence, 5:135–151, 1970.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 446, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 445, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 460 + ], + "score": 1.0, + "content": "Tim Rocktschel and Sebastian Riedel. End-to-end differentiable proving. NIPS, pp. 3791–3803,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 457, + 142, + 469 + ], + "spans": [ + { + "bbox": [ + 115, + 457, + 142, + 469 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 476, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "Neal J. Roese. Counterfactual thinking. Psychological Bulletin, 121(1):133–148, 1997. doi: 10.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 117, + 488, + 231, + 499 + ], + "spans": [ + { + "bbox": [ + 117, + 488, + 231, + 499 + ], + "score": 1.0, + "content": "1037/0033-2909.121.1.133.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 104, + 506, + 505, + 530 + ], + "lines": [ + { + "bbox": [ + 105, + 504, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 415, + 521 + ], + "score": 1.0, + "content": "Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach", + "type": "text" + }, + { + "bbox": [ + 415, + 506, + 429, + 517 + ], + "score": 0.73, + "content": "\\langle 3 ^ { r d }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 504, + 506, + 521 + ], + "score": 1.0, + "content": "Edition). Pearson,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 518, + 266, + 529 + ], + "spans": [ + { + "bbox": [ + 116, + 518, + 266, + 529 + ], + "score": 1.0, + "content": "November 2016. ISBN 1292153962.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 104, + 537, + 504, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 504, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 504, + 549 + ], + "score": 1.0, + "content": "John R. Searle. Minds, brains, and programs. Behavioral and brain sciences, 3(3):417–424, 1980.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 548, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 115, + 548, + 296, + 559 + ], + "score": 1.0, + "content": "doi: 10.1016/b978-1-4832-1446-7.50007-8.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 566, + 505, + 611 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 580 + ], + "score": 1.0, + "content": "Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "and Dhruv Batra. Grad-CAM: Visual explanations from deep networks via gradient-based local-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 114, + 588, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 114, + 588, + 506, + 602 + ], + "score": 1.0, + "content": "ization. In 2017 IEEE International Conference on Computer Vision (ICCV), pp. 618–626. IEEE,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 600, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 116, + 600, + 286, + 612 + ], + "score": 1.0, + "content": "October 2017. doi: 10.1109/iccv.2017.74.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 103, + 618, + 504, + 642 + ], + "lines": [ + { + "bbox": [ + 106, + 619, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 505, + 632 + ], + "score": 1.0, + "content": "Minjoon Seo, Sewon Min, Ali Farhadi, and Hannaneh Hajishirzi. Query-reduction networks for", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 630, + 252, + 642 + ], + "spans": [ + { + "bbox": [ + 116, + 630, + 252, + 642 + ], + "score": 1.0, + "content": "question answering. ICLR, 2017.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 105, + 649, + 504, + 672 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 663 + ], + "score": 1.0, + "content": "Ehud Y. Shapiro. Inductive inference of theories from facts. Yale University, Department of Com-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 660, + 200, + 672 + ], + "spans": [ + { + "bbox": [ + 115, + 660, + 200, + 672 + ], + "score": 1.0, + "content": "puter Science, 1981.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 106, + 679, + 504, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 679, + 505, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 505, + 692 + ], + "score": 1.0, + "content": "Gustav Sourek, Vojtech Aschenbrenner, Filip Zelezny, and Ondrej Kuzelka. Lifted relational neural", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 691, + 262, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 691, + 262, + 702 + ], + "score": 1.0, + "content": "networks. arXiv:1508.05128, 2015.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 504, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 504, + 723 + ], + "score": 1.0, + "content": "Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory networks.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 720, + 243, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 243, + 732 + ], + "score": 1.0, + "content": "In NIPS, pp. 2440–2448, 2015.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 107, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "Richard L. Lewis. Cognitive modeling, symbolic. The MIT encyclopedia of the cognitive sciences,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 196, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 196, + 105 + ], + "score": 1.0, + "content": "pp. 525–527, 1999.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 506, + 105 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 112, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 112, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 112, + 505, + 125 + ], + "score": 1.0, + "content": "Zachary C. Lipton. The mythos of model interpretability. Communications of the ACM, 61(10):", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 123, + 311, + 135 + ], + "spans": [ + { + "bbox": [ + 116, + 123, + 311, + 135 + ], + "score": 1.0, + "content": "36–43, September 2018. doi: 10.1145/3233231.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 112, + 505, + 135 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 142, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 141, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 505, + 156 + ], + "score": 1.0, + "content": "Fei Liu and Julien Perez. Gated end-to-end memory networks. In Proceedings of the 15th Confer-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 153, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 115, + 153, + 505, + 167 + ], + "score": 1.0, + "content": "ence of the European Chapter of the Association for Computational Linguistics: Volume 1, Long", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 164, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 116, + 164, + 505, + 178 + ], + "score": 1.0, + "content": "Papers, pp. 1–10. Association for Computational Linguistics, 2017. doi: 10.18653/v1/e17-1001.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 141, + 505, + 178 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 183, + 505, + 228 + ], + "lines": [ + { + "bbox": [ + 106, + 184, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 505, + 196 + ], + "score": 1.0, + "content": "Thang Luong, Hieu Pham, and Christopher D. Manning. Effective approaches to attention-based", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 194, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 115, + 194, + 506, + 207 + ], + "score": 1.0, + "content": "neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 205, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 114, + 205, + 506, + 219 + ], + "score": 1.0, + "content": "Natural Language Processing. Association for Computational Linguistics, 2015. doi: 10.18653/", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 217, + 172, + 228 + ], + "spans": [ + { + "bbox": [ + 116, + 217, + 172, + 228 + ], + "score": 1.0, + "content": "v1/d15-1166.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5, + "bbox_fs": [ + 106, + 184, + 506, + 228 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 235, + 504, + 259 + ], + "lines": [ + { + "bbox": [ + 106, + 235, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 505, + 248 + ], + "score": 1.0, + "content": "Tim Miller. Explanation in artificial intelligence: Insights from the social sciences. Artificial Intel-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 247, + 395, + 259 + ], + "spans": [ + { + "bbox": [ + 115, + 247, + 395, + 259 + ], + "score": 1.0, + "content": "ligence, 267:1–38, February 2019. doi: 10.1016/j.artint.2018.07.007.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 106, + 235, + 505, + 259 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 265, + 504, + 289 + ], + "lines": [ + { + "bbox": [ + 106, + 266, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 505, + 279 + ], + "score": 1.0, + "content": "Adam Morton and Jerry A. Fodor. The Language of Thought., volume 75. Philosophy Documenta-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 276, + 313, + 288 + ], + "spans": [ + { + "bbox": [ + 116, + 276, + 313, + 288 + ], + "score": 1.0, + "content": "tion Center, March 1978. doi: 10.2307/2025426.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 266, + 505, + 288 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 295, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 106, + 295, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 505, + 309 + ], + "score": 1.0, + "content": "Stephen Muggleton and Luc de Raedt. Inductive logic programming: Theory and methods.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 306, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 115, + 306, + 506, + 321 + ], + "score": 1.0, + "content": "The Journal of Logic Programming, 19-20:629–679, May 1994. doi: 10.1016/0743-1066(94)", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 316, + 156, + 330 + ], + "spans": [ + { + "bbox": [ + 115, + 316, + 156, + 330 + ], + "score": 1.0, + "content": "90035-3.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 106, + 295, + 506, + 330 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 336, + 503, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 336, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 505, + 349 + ], + "score": 1.0, + "content": "Judea Pearl. Probabilities of causation: three counterfactual interpretations and their identification.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 348, + 254, + 360 + ], + "spans": [ + { + "bbox": [ + 115, + 348, + 254, + 360 + ], + "score": 1.0, + "content": "Synthese, 121(1-2):93–149, 1999.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 336, + 505, + 360 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 367, + 504, + 390 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 380 + ], + "score": 1.0, + "content": "Judea Pearl. The seven tools of causal inference with reflections on machine learning. Communica-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 378, + 401, + 391 + ], + "spans": [ + { + "bbox": [ + 116, + 378, + 401, + 391 + ], + "score": 1.0, + "content": "tions of the ACM, 62(3):54–60, February 2019. doi: 10.1145/3241036.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 366, + 505, + 391 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 397, + 470, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 396, + 470, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 470, + 410 + ], + "score": 1.0, + "content": "Jean Piaget. The Psychology of Intelligence. Routledge, August 2001. ISBN 0415254019.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 396, + 470, + 410 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 416, + 506, + 439 + ], + "lines": [ + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "John C. Reynolds. Transformational systems and algebraic structure of atomic formulas. Machine", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 427, + 240, + 440 + ], + "spans": [ + { + "bbox": [ + 116, + 427, + 240, + 440 + ], + "score": 1.0, + "content": "intelligence, 5:135–151, 1970.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 106, + 416, + 505, + 440 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 446, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 445, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 460 + ], + "score": 1.0, + "content": "Tim Rocktschel and Sebastian Riedel. End-to-end differentiable proving. NIPS, pp. 3791–3803,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 457, + 142, + 469 + ], + "spans": [ + { + "bbox": [ + 115, + 457, + 142, + 469 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 445, + 505, + 469 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 476, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "Neal J. Roese. Counterfactual thinking. Psychological Bulletin, 121(1):133–148, 1997. doi: 10.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 117, + 488, + 231, + 499 + ], + "spans": [ + { + "bbox": [ + 117, + 488, + 231, + 499 + ], + "score": 1.0, + "content": "1037/0033-2909.121.1.133.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 476, + 505, + 499 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 506, + 505, + 530 + ], + "lines": [ + { + "bbox": [ + 105, + 504, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 415, + 521 + ], + "score": 1.0, + "content": "Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach", + "type": "text" + }, + { + "bbox": [ + 415, + 506, + 429, + 517 + ], + "score": 0.73, + "content": "\\langle 3 ^ { r d }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 504, + 506, + 521 + ], + "score": 1.0, + "content": "Edition). Pearson,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 518, + 266, + 529 + ], + "spans": [ + { + "bbox": [ + 116, + 518, + 266, + 529 + ], + "score": 1.0, + "content": "November 2016. ISBN 1292153962.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 504, + 506, + 529 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 537, + 504, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 504, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 504, + 549 + ], + "score": 1.0, + "content": "John R. Searle. Minds, brains, and programs. Behavioral and brain sciences, 3(3):417–424, 1980.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 548, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 115, + 548, + 296, + 559 + ], + "score": 1.0, + "content": "doi: 10.1016/b978-1-4832-1446-7.50007-8.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 106, + 536, + 504, + 559 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 566, + 505, + 611 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 580 + ], + "score": 1.0, + "content": "Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "and Dhruv Batra. Grad-CAM: Visual explanations from deep networks via gradient-based local-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 114, + 588, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 114, + 588, + 506, + 602 + ], + "score": 1.0, + "content": "ization. In 2017 IEEE International Conference on Computer Vision (ICCV), pp. 618–626. IEEE,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 600, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 116, + 600, + 286, + 612 + ], + "score": 1.0, + "content": "October 2017. doi: 10.1109/iccv.2017.74.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 567, + 506, + 612 + ] + }, + { + "type": "text", + "bbox": [ + 103, + 618, + 504, + 642 + ], + "lines": [ + { + "bbox": [ + 106, + 619, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 505, + 632 + ], + "score": 1.0, + "content": "Minjoon Seo, Sewon Min, Ali Farhadi, and Hannaneh Hajishirzi. Query-reduction networks for", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 630, + 252, + 642 + ], + "spans": [ + { + "bbox": [ + 116, + 630, + 252, + 642 + ], + "score": 1.0, + "content": "question answering. ICLR, 2017.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 106, + 619, + 505, + 642 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 649, + 504, + 672 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 663 + ], + "score": 1.0, + "content": "Ehud Y. Shapiro. Inductive inference of theories from facts. Yale University, Department of Com-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 660, + 200, + 672 + ], + "spans": [ + { + "bbox": [ + 115, + 660, + 200, + 672 + ], + "score": 1.0, + "content": "puter Science, 1981.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 648, + 505, + 672 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 679, + 504, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 679, + 505, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 505, + 692 + ], + "score": 1.0, + "content": "Gustav Sourek, Vojtech Aschenbrenner, Filip Zelezny, and Ondrej Kuzelka. Lifted relational neural", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 691, + 262, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 691, + 262, + 702 + ], + "score": 1.0, + "content": "networks. arXiv:1508.05128, 2015.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 106, + 679, + 505, + 702 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 504, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 504, + 723 + ], + "score": 1.0, + "content": "Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory networks.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 720, + 243, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 243, + 732 + ], + "score": 1.0, + "content": "In NIPS, pp. 2440–2448, 2015.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 106, + 708, + 504, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Seiya Tokui, Kenta Oono, Shohei Hido, and Justin Clayton. Chainer: a next-generation open", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "source framework for deep learning. In Proceedings of Workshop on Machine Learning Systems", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "(LearningSys) in The Twenty-ninth Annual Conference on Neural Information Processing Sys-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 114, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 115, + 114, + 505, + 129 + ], + "score": 1.0, + "content": "tems (NIPS), 2015. URL http://learningsys.org/papers/LearningSys_2015_", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 127, + 193, + 139 + ], + "spans": [ + { + "bbox": [ + 115, + 127, + 193, + 139 + ], + "score": 1.0, + "content": "paper_33.pdf.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 105, + 145, + 503, + 168 + ], + "lines": [ + { + "bbox": [ + 104, + 143, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 104, + 143, + 505, + 160 + ], + "score": 1.0, + "content": "J. Marshall Unger and Terrence W. Deacon. The Symbolic Species: The Co-Evolution of Language", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 156, + 367, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 367, + 168 + ], + "score": 1.0, + "content": "and the Brain, volume 82. Wiley, 1998. doi: 10.2307/329984.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 105, + 174, + 443, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 174, + 443, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 443, + 188 + ], + "score": 1.0, + "content": "Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. ICLR, 2015.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 105, + 194, + 505, + 227 + ], + "lines": [ + { + "bbox": [ + 105, + 194, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 206 + ], + "score": 1.0, + "content": "Jason Weston, Antoine Bordes, Sumit Chopra, Alexander M. Rush, Bart van Merrinboer, Armand", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 203, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 115, + 203, + 505, + 218 + ], + "score": 1.0, + "content": "Joulin, and Tomas Mikolov. Towards ai-complete question answering: A set of prerequisite toy", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 216, + 194, + 227 + ], + "spans": [ + { + "bbox": [ + 115, + 216, + 194, + 227 + ], + "score": 1.0, + "content": "tasks. ICLR, 2016.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 105, + 235, + 504, + 257 + ], + "lines": [ + { + "bbox": [ + 106, + 235, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 505, + 247 + ], + "score": 1.0, + "content": "Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 246, + 348, + 258 + ], + "spans": [ + { + "bbox": [ + 115, + 246, + 348, + 258 + ], + "score": 1.0, + "content": "textual question answering. ICML, pp. 2397–2406, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Seiya Tokui, Kenta Oono, Shohei Hido, and Justin Clayton. Chainer: a next-generation open", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "source framework for deep learning. In Proceedings of Workshop on Machine Learning Systems", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "(LearningSys) in The Twenty-ninth Annual Conference on Neural Information Processing Sys-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 114, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 115, + 114, + 505, + 129 + ], + "score": 1.0, + "content": "tems (NIPS), 2015. URL http://learningsys.org/papers/LearningSys_2015_", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 127, + 193, + 139 + ], + "spans": [ + { + "bbox": [ + 115, + 127, + 193, + 139 + ], + "score": 1.0, + "content": "paper_33.pdf.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 82, + 505, + 139 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 145, + 503, + 168 + ], + "lines": [ + { + "bbox": [ + 104, + 143, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 104, + 143, + 505, + 160 + ], + "score": 1.0, + "content": "J. Marshall Unger and Terrence W. Deacon. The Symbolic Species: The Co-Evolution of Language", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 156, + 367, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 367, + 168 + ], + "score": 1.0, + "content": "and the Brain, volume 82. Wiley, 1998. doi: 10.2307/329984.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 104, + 143, + 505, + 168 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 174, + 443, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 174, + 443, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 443, + 188 + ], + "score": 1.0, + "content": "Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. ICLR, 2015.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 174, + 443, + 188 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 194, + 505, + 227 + ], + "lines": [ + { + "bbox": [ + 105, + 194, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 206 + ], + "score": 1.0, + "content": "Jason Weston, Antoine Bordes, Sumit Chopra, Alexander M. Rush, Bart van Merrinboer, Armand", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 203, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 115, + 203, + 505, + 218 + ], + "score": 1.0, + "content": "Joulin, and Tomas Mikolov. Towards ai-complete question answering: A set of prerequisite toy", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 216, + 194, + 227 + ], + "spans": [ + { + "bbox": [ + 115, + 216, + 194, + 227 + ], + "score": 1.0, + "content": "tasks. ICLR, 2016.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 194, + 505, + 227 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 235, + 504, + 257 + ], + "lines": [ + { + "bbox": [ + 106, + 235, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 505, + 247 + ], + "score": 1.0, + "content": "Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 246, + 348, + 258 + ], + "spans": [ + { + "bbox": [ + 115, + 246, + 348, + 258 + ], + "score": 1.0, + "content": "textual question answering. ICML, pp. 2397–2406, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 106, + 235, + 505, + 258 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 81, + 216, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 217, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 217, + 96 + ], + "score": 1.0, + "content": "A MODEL DETAILS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 105, + 251, + 118 + ], + "lines": [ + { + "bbox": [ + 105, + 105, + 252, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 252, + 119 + ], + "score": 1.0, + "content": "A.1 UNIFICATION MLP & CNN", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 126, + 505, + 215 + ], + "lines": [ + { + "bbox": [ + 106, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 258, + 140 + ], + "score": 1.0, + "content": "Unification MLP (UMLP) To model", + "type": "text" + }, + { + "bbox": [ + 259, + 127, + 266, + 138 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "as a multi-layer perceptron, we take symbol embeddings of", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 104, + 137, + 125, + 150 + ], + "score": 1.0, + "content": "size", + "type": "text" + }, + { + "bbox": [ + 125, + 138, + 158, + 148 + ], + "score": 0.9, + "content": "d = 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 137, + 288, + 150 + ], + "score": 1.0, + "content": "and flatten sequences of length", + "type": "text" + }, + { + "bbox": [ + 289, + 138, + 314, + 148 + ], + "score": 0.89, + "content": "l = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 137, + 429, + 150 + ], + "score": 1.0, + "content": "into an input vector of size", + "type": "text" + }, + { + "bbox": [ + 429, + 137, + 480, + 149 + ], + "score": 0.92, + "content": "\\phi ( \\pmb { k } ) \\in \\mathbb { R } ^ { 6 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 137, + 506, + 150 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 147, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 104, + 147, + 369, + 162 + ], + "score": 1.0, + "content": "MLP consists of 2 hidden layers with tanh non-linearity of sizes", + "type": "text" + }, + { + "bbox": [ + 369, + 149, + 380, + 159 + ], + "score": 0.81, + "content": "2 d", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 147, + 398, + 162 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 399, + 149, + 406, + 159 + ], + "score": 0.8, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 147, + 506, + 162 + ], + "score": 1.0, + "content": "respectively. To process", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 160, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 366, + 172 + ], + "score": 1.0, + "content": "the query, we concatenate the one-hot encoding of the task id to", + "type": "text" + }, + { + "bbox": [ + 367, + 160, + 388, + 171 + ], + "score": 0.92, + "content": "\\phi ( k )", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 160, + 505, + 172 + ], + "score": 1.0, + "content": "yielding a final input of size", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 170, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 163, + 181 + ], + "score": 0.91, + "content": "6 4 + 4 = 6 8", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 170, + 268, + 184 + ], + "score": 1.0, + "content": ". For unification features", + "type": "text" + }, + { + "bbox": [ + 269, + 171, + 282, + 182 + ], + "score": 0.89, + "content": "\\phi _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 170, + 479, + 184 + ], + "score": 1.0, + "content": ", we use a bi-directional GRU with hidden size", + "type": "text" + }, + { + "bbox": [ + 479, + 171, + 486, + 181 + ], + "score": 0.77, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 170, + 506, + 184 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 180, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 506, + 195 + ], + "score": 1.0, + "content": "an initial state of 0. The hidden state at each symbol is taken with a linear transformation to give", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 191, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 107, + 192, + 185, + 204 + ], + "score": 0.92, + "content": "\\phi _ { U } ( s ) = W _ { U } \\Phi ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 191, + 214, + 205 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 215, + 192, + 235, + 204 + ], + "score": 0.92, + "content": "\\Phi ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 191, + 506, + 205 + ], + "score": 1.0, + "content": "is the hidden state of the biGRU. The variable assignment is then", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 204, + 342, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 342, + 216 + ], + "score": 1.0, + "content": "computed as an attention over the according to equation 3.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 220, + 505, + 309 + ], + "lines": [ + { + "bbox": [ + 106, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 375, + 232 + ], + "score": 1.0, + "content": "Unification CNN (UCNN) We take symbols embeddings of size", + "type": "text" + }, + { + "bbox": [ + 375, + 221, + 409, + 231 + ], + "score": 0.9, + "content": "d = 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "to obtain an input grid", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 227, + 507, + 245 + ], + "spans": [ + { + "bbox": [ + 107, + 231, + 181, + 243 + ], + "score": 0.92, + "content": "\\phi ( K ) \\in \\mathbb { R } ^ { 3 \\times 3 \\times 3 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 227, + 507, + 245 + ], + "score": 1.0, + "content": ". Similar to UMLP, for each symbol we append the task id as a one-hot vector to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 191, + 254 + ], + "score": 1.0, + "content": "get an input of shape", + "type": "text" + }, + { + "bbox": [ + 192, + 243, + 259, + 254 + ], + "score": 0.92, + "content": "3 \\times 3 \\times ( 3 2 + 4 )", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 242, + 286, + 254 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + }, + { + "bbox": [ + 286, + 243, + 293, + 254 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 242, + 449, + 254 + ], + "score": 1.0, + "content": "consists of 2 convolutional layers with", + "type": "text" + }, + { + "bbox": [ + 449, + 243, + 456, + 252 + ], + "score": 0.8, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "filters each,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "kernel size of 3 and stride 1. We use relu non-linearity in between the layers. We pad the grid with", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 264, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 212, + 277 + ], + "score": 1.0, + "content": "2 columns and 2 rows to a", + "type": "text" + }, + { + "bbox": [ + 212, + 264, + 235, + 275 + ], + "score": 0.89, + "content": "5 \\times 5", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 264, + 505, + 277 + ], + "score": 1.0, + "content": "such that the output of the convolutions yield again a hidden output", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 272, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 160, + 285 + ], + "score": 0.91, + "content": "\\boldsymbol { \\mathsf { H } } \\in \\mathbb { R } ^ { 3 \\times 3 \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 272, + 345, + 289 + ], + "score": 1.0, + "content": "of the same shape. As the final hidden output", + "type": "text" + }, + { + "bbox": [ + 346, + 276, + 353, + 285 + ], + "score": 0.77, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 272, + 495, + 289 + ], + "score": 1.0, + "content": ", we take a global max pool to over", + "type": "text" + }, + { + "bbox": [ + 495, + 275, + 505, + 285 + ], + "score": 0.28, + "content": "\\boldsymbol { \\mathsf { H } }", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 285, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 104, + 285, + 145, + 299 + ], + "score": 1.0, + "content": "to obtain", + "type": "text" + }, + { + "bbox": [ + 145, + 286, + 178, + 296 + ], + "score": 0.9, + "content": "\\pmb { h } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 285, + 268, + 299 + ], + "score": 1.0, + "content": ". Unification function", + "type": "text" + }, + { + "bbox": [ + 268, + 288, + 275, + 298 + ], + "score": 0.8, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 285, + 373, + 299 + ], + "score": 1.0, + "content": "is modelled identical to", + "type": "text" + }, + { + "bbox": [ + 374, + 286, + 381, + 298 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 285, + 506, + 299 + ], + "score": 1.0, + "content": "without the max pooling such", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 296, + 423, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 124, + 310 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 124, + 297, + 190, + 311 + ], + "score": 0.93, + "content": "\\phi _ { U } ( K _ { i j } ) = \\mathsf { H } _ { i j } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 296, + 217, + 310 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 218, + 297, + 229, + 307 + ], + "score": 0.86, + "content": "\\boldsymbol { \\mathsf { H } } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 296, + 423, + 310 + ], + "score": 1.0, + "content": "is the hidden output of the convolutional layers.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13.5 + }, + { + "type": "title", + "bbox": [ + 108, + 321, + 285, + 333 + ], + "lines": [ + { + "bbox": [ + 106, + 320, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 287, + 335 + ], + "score": 1.0, + "content": "A.2 UNIFICATION MEMORY NETWORKS", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 342, + 505, + 399 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 504, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 328, + 355 + ], + "score": 1.0, + "content": "Unlike previous architectures, with UMN we interleave", + "type": "text" + }, + { + "bbox": [ + 328, + 344, + 335, + 354 + ], + "score": 0.8, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 342, + 353, + 355 + ], + "score": 1.0, + "content": "into", + "type": "text" + }, + { + "bbox": [ + 353, + 343, + 360, + 354 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 342, + 474, + 355 + ], + "score": 1.0, + "content": ". We use embedding sizes of", + "type": "text" + }, + { + "bbox": [ + 474, + 343, + 504, + 353 + ], + "score": 0.88, + "content": "d = 3 2", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 152, + 365 + ], + "score": 1.0, + "content": "and model", + "type": "text" + }, + { + "bbox": [ + 152, + 354, + 159, + 365 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "with an iterative memory network. We take the final hidden state of a bi-directional", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 364, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 216, + 378 + ], + "score": 1.0, + "content": "GRU, with initial state 0,", + "type": "text" + }, + { + "bbox": [ + 217, + 365, + 234, + 375 + ], + "score": 0.78, + "content": "\\Phi _ { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 364, + 407, + 378 + ], + "score": 1.0, + "content": "to represent the sentences of the context", + "type": "text" + }, + { + "bbox": [ + 407, + 365, + 418, + 375 + ], + "score": 0.81, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 364, + 464, + 378 + ], + "score": 1.0, + "content": "and query", + "type": "text" + }, + { + "bbox": [ + 465, + 366, + 472, + 376 + ], + "score": 0.76, + "content": "\\pmb q", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 364, + 494, + 378 + ], + "score": 1.0, + "content": "in a", + "type": "text" + }, + { + "bbox": [ + 494, + 365, + 501, + 374 + ], + "score": 0.75, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 364, + 505, + 378 + ], + "score": 1.0, + "content": "-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 374, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 186, + 390 + ], + "score": 1.0, + "content": "dimensional vector", + "type": "text" + }, + { + "bbox": [ + 186, + 376, + 252, + 388 + ], + "score": 0.92, + "content": "M _ { i } = \\Phi _ { M } ( C _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 374, + 311, + 390 + ], + "score": 1.0, + "content": "and the query", + "type": "text" + }, + { + "bbox": [ + 312, + 375, + 371, + 388 + ], + "score": 0.94, + "content": "{ \\pmb m } _ { \\ b q } = \\Phi _ { M } ( { \\pmb q } )", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 374, + 505, + 390 + ], + "score": 1.0, + "content": ". The initial state of the memory", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 385, + 275, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 151, + 402 + ], + "score": 1.0, + "content": "network is", + "type": "text" + }, + { + "bbox": [ + 151, + 387, + 191, + 400 + ], + "score": 0.92, + "content": "h ^ { 0 } = m _ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 385, + 264, + 402 + ], + "score": 1.0, + "content": ". At each iteration", + "type": "text" + }, + { + "bbox": [ + 264, + 388, + 270, + 399 + ], + "score": 0.81, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 385, + 275, + 402 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21 + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 404, + 371, + 438 + ], + "lines": [ + { + "bbox": [ + 239, + 404, + 371, + 438 + ], + "spans": [ + { + "bbox": [ + 239, + 404, + 371, + 438 + ], + "score": 0.91, + "content": "\\begin{array} { c } { { A _ { i } ^ { j } = \\operatorname { t a n h } ( W \\rho ( M _ { i } , h ^ { j } ) + b ) } } \\\\ { { \\beta ^ { j } = \\operatorname { s o f t m a x } ( W \\Phi _ { A } ( A ^ { j } ) + b ) } } \\end{array}", + "type": "interline_equation", + "image_path": "b17d2332dda7b22d96bbade43e656c75a1713d2c81331c84317246e400303740.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 239, + 404, + 371, + 421.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 239, + 421.0, + 371, + 438.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 442, + 505, + 476 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 133, + 456 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 443, + 148, + 454 + ], + "score": 0.89, + "content": "\\Phi _ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 441, + 190, + 456 + ], + "score": 1.0, + "content": "is another", + "type": "text" + }, + { + "bbox": [ + 191, + 443, + 197, + 453 + ], + "score": 0.82, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 441, + 346, + 456 + ], + "score": 1.0, + "content": "-dimensional bi-directional GRU and", + "type": "text" + }, + { + "bbox": [ + 347, + 442, + 484, + 455 + ], + "score": 0.91, + "content": "\\rho ( { \\pmb x } , { \\pmb y } ) = [ { \\pmb x } ; { \\pmb y } ; { \\pmb x } \\odot { \\pmb y } ; ( { \\pmb x } - { \\pmb y } ) ^ { 2 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 441, + 505, + 456 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 454, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 107, + 455, + 116, + 464 + ], + "score": 0.82, + "content": "\\odot", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 454, + 266, + 466 + ], + "score": 1.0, + "content": "the element-wise multiplication and", + "type": "text" + }, + { + "bbox": [ + 267, + 454, + 277, + 466 + ], + "score": 0.64, + "content": "[ ; ]", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 454, + 433, + 466 + ], + "score": 1.0, + "content": "the concatenation of vectors. Taking", + "type": "text" + }, + { + "bbox": [ + 433, + 455, + 444, + 466 + ], + "score": 0.87, + "content": "\\beta ^ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 454, + 505, + 466 + ], + "score": 1.0, + "content": "as the context", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 465, + 341, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 341, + 476 + ], + "score": 1.0, + "content": "attention, we obtain the next state of the memory network:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 482, + 388, + 508 + ], + "lines": [ + { + "bbox": [ + 223, + 482, + 388, + 508 + ], + "spans": [ + { + "bbox": [ + 223, + 482, + 388, + 508 + ], + "score": 0.94, + "content": "h ^ { j + 1 } = \\sum _ { i } \\beta _ { i } ^ { j } \\operatorname { t a n h } ( W \\rho ( M _ { i } , h ^ { j } ) + b )", + "type": "interline_equation", + "image_path": "85e2e7ec827f8b1efbb28c02265ab15e959abdf242665360146e66fedab5e494.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 223, + 482, + 388, + 508 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 514, + 505, + 537 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 150, + 527 + ], + "score": 1.0, + "content": "and iterate", + "type": "text" + }, + { + "bbox": [ + 150, + 515, + 158, + 524 + ], + "score": 0.82, + "content": "J", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 513, + 372, + 527 + ], + "score": 1.0, + "content": "many times in advance. The final prediction becomes", + "type": "text" + }, + { + "bbox": [ + 372, + 514, + 501, + 527 + ], + "score": 0.92, + "content": "f ( C , \\boldsymbol { q } ) = \\mathrm { s o f t m a x } ( W h ^ { J } + b )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 513, + 505, + 527 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 525, + 459, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 187, + 538 + ], + "score": 1.0, + "content": "All weight matrices", + "type": "text" + }, + { + "bbox": [ + 188, + 526, + 201, + 536 + ], + "score": 0.63, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 525, + 269, + 538 + ], + "score": 1.0, + "content": "and bias vectors", + "type": "text" + }, + { + "bbox": [ + 270, + 526, + 276, + 536 + ], + "score": 0.8, + "content": "^ { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 525, + 459, + 538 + ], + "score": 1.0, + "content": "are independent but are tied across iterations.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 81, + 216, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 217, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 217, + 96 + ], + "score": 1.0, + "content": "A MODEL DETAILS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 106, + 80, + 217, + 96 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 105, + 251, + 118 + ], + "lines": [ + { + "bbox": [ + 105, + 105, + 252, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 252, + 119 + ], + "score": 1.0, + "content": "A.1 UNIFICATION MLP & CNN", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 126, + 505, + 215 + ], + "lines": [ + { + "bbox": [ + 106, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 258, + 140 + ], + "score": 1.0, + "content": "Unification MLP (UMLP) To model", + "type": "text" + }, + { + "bbox": [ + 259, + 127, + 266, + 138 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "as a multi-layer perceptron, we take symbol embeddings of", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 104, + 137, + 125, + 150 + ], + "score": 1.0, + "content": "size", + "type": "text" + }, + { + "bbox": [ + 125, + 138, + 158, + 148 + ], + "score": 0.9, + "content": "d = 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 137, + 288, + 150 + ], + "score": 1.0, + "content": "and flatten sequences of length", + "type": "text" + }, + { + "bbox": [ + 289, + 138, + 314, + 148 + ], + "score": 0.89, + "content": "l = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 137, + 429, + 150 + ], + "score": 1.0, + "content": "into an input vector of size", + "type": "text" + }, + { + "bbox": [ + 429, + 137, + 480, + 149 + ], + "score": 0.92, + "content": "\\phi ( \\pmb { k } ) \\in \\mathbb { R } ^ { 6 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 137, + 506, + 150 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 147, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 104, + 147, + 369, + 162 + ], + "score": 1.0, + "content": "MLP consists of 2 hidden layers with tanh non-linearity of sizes", + "type": "text" + }, + { + "bbox": [ + 369, + 149, + 380, + 159 + ], + "score": 0.81, + "content": "2 d", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 147, + 398, + 162 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 399, + 149, + 406, + 159 + ], + "score": 0.8, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 147, + 506, + 162 + ], + "score": 1.0, + "content": "respectively. To process", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 160, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 366, + 172 + ], + "score": 1.0, + "content": "the query, we concatenate the one-hot encoding of the task id to", + "type": "text" + }, + { + "bbox": [ + 367, + 160, + 388, + 171 + ], + "score": 0.92, + "content": "\\phi ( k )", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 160, + 505, + 172 + ], + "score": 1.0, + "content": "yielding a final input of size", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 170, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 163, + 181 + ], + "score": 0.91, + "content": "6 4 + 4 = 6 8", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 170, + 268, + 184 + ], + "score": 1.0, + "content": ". For unification features", + "type": "text" + }, + { + "bbox": [ + 269, + 171, + 282, + 182 + ], + "score": 0.89, + "content": "\\phi _ { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 170, + 479, + 184 + ], + "score": 1.0, + "content": ", we use a bi-directional GRU with hidden size", + "type": "text" + }, + { + "bbox": [ + 479, + 171, + 486, + 181 + ], + "score": 0.77, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 170, + 506, + 184 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 180, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 506, + 195 + ], + "score": 1.0, + "content": "an initial state of 0. The hidden state at each symbol is taken with a linear transformation to give", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 191, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 107, + 192, + 185, + 204 + ], + "score": 0.92, + "content": "\\phi _ { U } ( s ) = W _ { U } \\Phi ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 191, + 214, + 205 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 215, + 192, + 235, + 204 + ], + "score": 0.92, + "content": "\\Phi ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 191, + 506, + 205 + ], + "score": 1.0, + "content": "is the hidden state of the biGRU. The variable assignment is then", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 204, + 342, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 342, + 216 + ], + "score": 1.0, + "content": "computed as an attention over the according to equation 3.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5.5, + "bbox_fs": [ + 104, + 126, + 506, + 216 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 220, + 505, + 309 + ], + "lines": [ + { + "bbox": [ + 106, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 375, + 232 + ], + "score": 1.0, + "content": "Unification CNN (UCNN) We take symbols embeddings of size", + "type": "text" + }, + { + "bbox": [ + 375, + 221, + 409, + 231 + ], + "score": 0.9, + "content": "d = 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "to obtain an input grid", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 227, + 507, + 245 + ], + "spans": [ + { + "bbox": [ + 107, + 231, + 181, + 243 + ], + "score": 0.92, + "content": "\\phi ( K ) \\in \\mathbb { R } ^ { 3 \\times 3 \\times 3 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 227, + 507, + 245 + ], + "score": 1.0, + "content": ". Similar to UMLP, for each symbol we append the task id as a one-hot vector to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 191, + 254 + ], + "score": 1.0, + "content": "get an input of shape", + "type": "text" + }, + { + "bbox": [ + 192, + 243, + 259, + 254 + ], + "score": 0.92, + "content": "3 \\times 3 \\times ( 3 2 + 4 )", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 242, + 286, + 254 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + }, + { + "bbox": [ + 286, + 243, + 293, + 254 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 242, + 449, + 254 + ], + "score": 1.0, + "content": "consists of 2 convolutional layers with", + "type": "text" + }, + { + "bbox": [ + 449, + 243, + 456, + 252 + ], + "score": 0.8, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "filters each,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "kernel size of 3 and stride 1. We use relu non-linearity in between the layers. We pad the grid with", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 264, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 212, + 277 + ], + "score": 1.0, + "content": "2 columns and 2 rows to a", + "type": "text" + }, + { + "bbox": [ + 212, + 264, + 235, + 275 + ], + "score": 0.89, + "content": "5 \\times 5", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 264, + 505, + 277 + ], + "score": 1.0, + "content": "such that the output of the convolutions yield again a hidden output", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 272, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 160, + 285 + ], + "score": 0.91, + "content": "\\boldsymbol { \\mathsf { H } } \\in \\mathbb { R } ^ { 3 \\times 3 \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 272, + 345, + 289 + ], + "score": 1.0, + "content": "of the same shape. As the final hidden output", + "type": "text" + }, + { + "bbox": [ + 346, + 276, + 353, + 285 + ], + "score": 0.77, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 272, + 495, + 289 + ], + "score": 1.0, + "content": ", we take a global max pool to over", + "type": "text" + }, + { + "bbox": [ + 495, + 275, + 505, + 285 + ], + "score": 0.28, + "content": "\\boldsymbol { \\mathsf { H } }", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 285, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 104, + 285, + 145, + 299 + ], + "score": 1.0, + "content": "to obtain", + "type": "text" + }, + { + "bbox": [ + 145, + 286, + 178, + 296 + ], + "score": 0.9, + "content": "\\pmb { h } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 285, + 268, + 299 + ], + "score": 1.0, + "content": ". Unification function", + "type": "text" + }, + { + "bbox": [ + 268, + 288, + 275, + 298 + ], + "score": 0.8, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 285, + 373, + 299 + ], + "score": 1.0, + "content": "is modelled identical to", + "type": "text" + }, + { + "bbox": [ + 374, + 286, + 381, + 298 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 285, + 506, + 299 + ], + "score": 1.0, + "content": "without the max pooling such", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 296, + 423, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 124, + 310 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 124, + 297, + 190, + 311 + ], + "score": 0.93, + "content": "\\phi _ { U } ( K _ { i j } ) = \\mathsf { H } _ { i j } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 296, + 217, + 310 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 218, + 297, + 229, + 307 + ], + "score": 0.86, + "content": "\\boldsymbol { \\mathsf { H } } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 296, + 423, + 310 + ], + "score": 1.0, + "content": "is the hidden output of the convolutional layers.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13.5, + "bbox_fs": [ + 104, + 220, + 507, + 311 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 321, + 285, + 333 + ], + "lines": [ + { + "bbox": [ + 106, + 320, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 287, + 335 + ], + "score": 1.0, + "content": "A.2 UNIFICATION MEMORY NETWORKS", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 342, + 505, + 399 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 504, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 328, + 355 + ], + "score": 1.0, + "content": "Unlike previous architectures, with UMN we interleave", + "type": "text" + }, + { + "bbox": [ + 328, + 344, + 335, + 354 + ], + "score": 0.8, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 342, + 353, + 355 + ], + "score": 1.0, + "content": "into", + "type": "text" + }, + { + "bbox": [ + 353, + 343, + 360, + 354 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 342, + 474, + 355 + ], + "score": 1.0, + "content": ". We use embedding sizes of", + "type": "text" + }, + { + "bbox": [ + 474, + 343, + 504, + 353 + ], + "score": 0.88, + "content": "d = 3 2", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 152, + 365 + ], + "score": 1.0, + "content": "and model", + "type": "text" + }, + { + "bbox": [ + 152, + 354, + 159, + 365 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "with an iterative memory network. We take the final hidden state of a bi-directional", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 364, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 216, + 378 + ], + "score": 1.0, + "content": "GRU, with initial state 0,", + "type": "text" + }, + { + "bbox": [ + 217, + 365, + 234, + 375 + ], + "score": 0.78, + "content": "\\Phi _ { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 364, + 407, + 378 + ], + "score": 1.0, + "content": "to represent the sentences of the context", + "type": "text" + }, + { + "bbox": [ + 407, + 365, + 418, + 375 + ], + "score": 0.81, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 364, + 464, + 378 + ], + "score": 1.0, + "content": "and query", + "type": "text" + }, + { + "bbox": [ + 465, + 366, + 472, + 376 + ], + "score": 0.76, + "content": "\\pmb q", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 364, + 494, + 378 + ], + "score": 1.0, + "content": "in a", + "type": "text" + }, + { + "bbox": [ + 494, + 365, + 501, + 374 + ], + "score": 0.75, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 364, + 505, + 378 + ], + "score": 1.0, + "content": "-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 374, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 186, + 390 + ], + "score": 1.0, + "content": "dimensional vector", + "type": "text" + }, + { + "bbox": [ + 186, + 376, + 252, + 388 + ], + "score": 0.92, + "content": "M _ { i } = \\Phi _ { M } ( C _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 374, + 311, + 390 + ], + "score": 1.0, + "content": "and the query", + "type": "text" + }, + { + "bbox": [ + 312, + 375, + 371, + 388 + ], + "score": 0.94, + "content": "{ \\pmb m } _ { \\ b q } = \\Phi _ { M } ( { \\pmb q } )", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 374, + 505, + 390 + ], + "score": 1.0, + "content": ". The initial state of the memory", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 385, + 275, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 151, + 402 + ], + "score": 1.0, + "content": "network is", + "type": "text" + }, + { + "bbox": [ + 151, + 387, + 191, + 400 + ], + "score": 0.92, + "content": "h ^ { 0 } = m _ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 385, + 264, + 402 + ], + "score": 1.0, + "content": ". At each iteration", + "type": "text" + }, + { + "bbox": [ + 264, + 388, + 270, + 399 + ], + "score": 0.81, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 385, + 275, + 402 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 342, + 505, + 402 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 404, + 371, + 438 + ], + "lines": [ + { + "bbox": [ + 239, + 404, + 371, + 438 + ], + "spans": [ + { + "bbox": [ + 239, + 404, + 371, + 438 + ], + "score": 0.91, + "content": "\\begin{array} { c } { { A _ { i } ^ { j } = \\operatorname { t a n h } ( W \\rho ( M _ { i } , h ^ { j } ) + b ) } } \\\\ { { \\beta ^ { j } = \\operatorname { s o f t m a x } ( W \\Phi _ { A } ( A ^ { j } ) + b ) } } \\end{array}", + "type": "interline_equation", + "image_path": "b17d2332dda7b22d96bbade43e656c75a1713d2c81331c84317246e400303740.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 239, + 404, + 371, + 421.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 239, + 421.0, + 371, + 438.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 442, + 505, + 476 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 133, + 456 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 443, + 148, + 454 + ], + "score": 0.89, + "content": "\\Phi _ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 441, + 190, + 456 + ], + "score": 1.0, + "content": "is another", + "type": "text" + }, + { + "bbox": [ + 191, + 443, + 197, + 453 + ], + "score": 0.82, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 441, + 346, + 456 + ], + "score": 1.0, + "content": "-dimensional bi-directional GRU and", + "type": "text" + }, + { + "bbox": [ + 347, + 442, + 484, + 455 + ], + "score": 0.91, + "content": "\\rho ( { \\pmb x } , { \\pmb y } ) = [ { \\pmb x } ; { \\pmb y } ; { \\pmb x } \\odot { \\pmb y } ; ( { \\pmb x } - { \\pmb y } ) ^ { 2 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 441, + 505, + 456 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 454, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 107, + 455, + 116, + 464 + ], + "score": 0.82, + "content": "\\odot", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 454, + 266, + 466 + ], + "score": 1.0, + "content": "the element-wise multiplication and", + "type": "text" + }, + { + "bbox": [ + 267, + 454, + 277, + 466 + ], + "score": 0.64, + "content": "[ ; ]", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 454, + 433, + 466 + ], + "score": 1.0, + "content": "the concatenation of vectors. Taking", + "type": "text" + }, + { + "bbox": [ + 433, + 455, + 444, + 466 + ], + "score": 0.87, + "content": "\\beta ^ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 454, + 505, + 466 + ], + "score": 1.0, + "content": "as the context", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 465, + 341, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 341, + 476 + ], + "score": 1.0, + "content": "attention, we obtain the next state of the memory network:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 441, + 505, + 476 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 482, + 388, + 508 + ], + "lines": [ + { + "bbox": [ + 223, + 482, + 388, + 508 + ], + "spans": [ + { + "bbox": [ + 223, + 482, + 388, + 508 + ], + "score": 0.94, + "content": "h ^ { j + 1 } = \\sum _ { i } \\beta _ { i } ^ { j } \\operatorname { t a n h } ( W \\rho ( M _ { i } , h ^ { j } ) + b )", + "type": "interline_equation", + "image_path": "85e2e7ec827f8b1efbb28c02265ab15e959abdf242665360146e66fedab5e494.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 223, + 482, + 388, + 508 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "list", + "bbox": [ + 106, + 514, + 505, + 537 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 150, + 527 + ], + "score": 1.0, + "content": "and iterate", + "type": "text" + }, + { + "bbox": [ + 150, + 515, + 158, + 524 + ], + "score": 0.82, + "content": "J", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 513, + 372, + 527 + ], + "score": 1.0, + "content": "many times in advance. The final prediction becomes", + "type": "text" + }, + { + "bbox": [ + 372, + 514, + 501, + 527 + ], + "score": 0.92, + "content": "f ( C , \\boldsymbol { q } ) = \\mathrm { s o f t m a x } ( W h ^ { J } + b )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 513, + 505, + 527 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 525, + 459, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 187, + 538 + ], + "score": 1.0, + "content": "All weight matrices", + "type": "text" + }, + { + "bbox": [ + 188, + 526, + 201, + 536 + ], + "score": 0.63, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 525, + 269, + 538 + ], + "score": 1.0, + "content": "and bias vectors", + "type": "text" + }, + { + "bbox": [ + 270, + 526, + 276, + 536 + ], + "score": 0.8, + "content": "^ { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 525, + 459, + 538 + ], + "score": 1.0, + "content": "are independent but are tied across iterations.", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 513, + 505, + 538 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 122, + 418, + 395 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 81, + 469, + 114 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 79, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 294, + 96 + ], + "score": 1.0, + "content": "B GENERATED DATASET SAMPLES", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 101, + 469, + 115 + ], + "spans": [ + { + "bbox": [ + 141, + 101, + 469, + 115 + ], + "score": 1.0, + "content": "Table 4: Sample context, query and answer triples from sequences and grid tasks.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 192, + 122, + 418, + 395 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 192, + 122, + 418, + 395 + ], + "spans": [ + { + "bbox": [ + 192, + 122, + 418, + 395 + ], + "score": 0.968, + "html": "
DatasetTaskContextQueryAnswer
Sequencei1488constant2
Sequenceii6157head6
Sequenceii1837tail7
Sequenceiv3563duplicate3
Gridi００８ 022 022box2
Gridii4 -070 0 页 0 8 1head4
Gridii0 673 0 1 2 0 0centre7
Gridiv8 0 0 2 60 41corner2
", + "type": "table", + "image_path": "0614697dfdc3aacbd0a222b3317ffc3cd4f9611857ee0057c4b330f335467084.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 192, + 122, + 418, + 136.3684210526316 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 192, + 136.3684210526316, + 418, + 150.73684210526318 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 192, + 150.73684210526318, + 418, + 165.10526315789477 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 192, + 165.10526315789477, + 418, + 179.47368421052636 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 192, + 179.47368421052636, + 418, + 193.84210526315795 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 192, + 193.84210526315795, + 418, + 208.21052631578954 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 192, + 208.21052631578954, + 418, + 222.57894736842113 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 192, + 222.57894736842113, + 418, + 236.94736842105272 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 192, + 236.94736842105272, + 418, + 251.3157894736843 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 192, + 251.3157894736843, + 418, + 265.6842105263159 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 192, + 265.6842105263159, + 418, + 280.05263157894746 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 192, + 280.05263157894746, + 418, + 294.421052631579 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 192, + 294.421052631579, + 418, + 308.7894736842106 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 192, + 308.7894736842106, + 418, + 323.15789473684214 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 192, + 323.15789473684214, + 418, + 337.5263157894737 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 192, + 337.5263157894737, + 418, + 351.89473684210526 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 192, + 351.89473684210526, + 418, + 366.2631578947368 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 192, + 366.2631578947368, + 418, + 380.6315789473684 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 192, + 380.6315789473684, + 418, + 394.99999999999994 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 397, + 505, + 431 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 396, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 410 + ], + "score": 1.0, + "content": "Table 5: Training sizes for randomly generated fixed length sequences and grid tasks with 8 unique", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 408, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 506, + 420 + ], + "score": 1.0, + "content": "symbols. The reason for Grid task (i) to be smaller is because there are at most 32 combinations of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 420, + 312, + 432 + ], + "spans": [ + { + "bbox": [ + 107, + 420, + 130, + 430 + ], + "score": 0.89, + "content": "2 \\times 2", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 420, + 173, + 432 + ], + "score": 1.0, + "content": "boxes in a", + "type": "text" + }, + { + "bbox": [ + 174, + 420, + 198, + 430 + ], + "score": 0.89, + "content": "3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 420, + 312, + 432 + ], + "score": 1.0, + "content": "grid with 8 unique symbols.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 230, + 440, + 380, + 508 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 230, + 440, + 380, + 508 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 230, + 440, + 380, + 508 + ], + "spans": [ + { + "bbox": [ + 230, + 440, + 380, + 508 + ], + "score": 0.972, + "html": "
Task SequencesGrid
i704.7 ± 12.825.6 ± 1.8
ii709.4 ± 13.8623.7 ± 14.1
ii709.7 ± 14.0768.2 ±12.5
iv624.8 ± 12.4795.2 ± 10.3
", + "type": "table", + "image_path": "164eed252519d30abfd4208cfebc52987bd6e5948e3125927523777305272436.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 440, + 380, + 474.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 230, + 474.0, + 380, + 508.0 + ], + "spans": [], + "index": 25 + } + ] + } + ], + "index": 24.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 122, + 418, + 395 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 81, + 469, + 114 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 79, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 294, + 96 + ], + "score": 1.0, + "content": "B GENERATED DATASET SAMPLES", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 101, + 469, + 115 + ], + "spans": [ + { + "bbox": [ + 141, + 101, + 469, + 115 + ], + "score": 1.0, + "content": "Table 4: Sample context, query and answer triples from sequences and grid tasks.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 192, + 122, + 418, + 395 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 192, + 122, + 418, + 395 + ], + "spans": [ + { + "bbox": [ + 192, + 122, + 418, + 395 + ], + "score": 0.968, + "html": "
DatasetTaskContextQueryAnswer
Sequencei1488constant2
Sequenceii6157head6
Sequenceii1837tail7
Sequenceiv3563duplicate3
Gridi００８ 022 022box2
Gridii4 -070 0 页 0 8 1head4
Gridii0 673 0 1 2 0 0centre7
Gridiv8 0 0 2 60 41corner2
", + "type": "table", + "image_path": "0614697dfdc3aacbd0a222b3317ffc3cd4f9611857ee0057c4b330f335467084.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 192, + 122, + 418, + 136.3684210526316 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 192, + 136.3684210526316, + 418, + 150.73684210526318 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 192, + 150.73684210526318, + 418, + 165.10526315789477 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 192, + 165.10526315789477, + 418, + 179.47368421052636 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 192, + 179.47368421052636, + 418, + 193.84210526315795 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 192, + 193.84210526315795, + 418, + 208.21052631578954 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 192, + 208.21052631578954, + 418, + 222.57894736842113 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 192, + 222.57894736842113, + 418, + 236.94736842105272 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 192, + 236.94736842105272, + 418, + 251.3157894736843 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 192, + 251.3157894736843, + 418, + 265.6842105263159 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 192, + 265.6842105263159, + 418, + 280.05263157894746 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 192, + 280.05263157894746, + 418, + 294.421052631579 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 192, + 294.421052631579, + 418, + 308.7894736842106 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 192, + 308.7894736842106, + 418, + 323.15789473684214 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 192, + 323.15789473684214, + 418, + 337.5263157894737 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 192, + 337.5263157894737, + 418, + 351.89473684210526 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 192, + 351.89473684210526, + 418, + 366.2631578947368 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 192, + 366.2631578947368, + 418, + 380.6315789473684 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 192, + 380.6315789473684, + 418, + 394.99999999999994 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 397, + 505, + 431 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 396, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 410 + ], + "score": 1.0, + "content": "Table 5: Training sizes for randomly generated fixed length sequences and grid tasks with 8 unique", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 408, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 506, + 420 + ], + "score": 1.0, + "content": "symbols. The reason for Grid task (i) to be smaller is because there are at most 32 combinations of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 420, + 312, + 432 + ], + "spans": [ + { + "bbox": [ + 107, + 420, + 130, + 430 + ], + "score": 0.89, + "content": "2 \\times 2", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 420, + 173, + 432 + ], + "score": 1.0, + "content": "boxes in a", + "type": "text" + }, + { + "bbox": [ + 174, + 420, + 198, + 430 + ], + "score": 0.89, + "content": "3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 420, + 312, + 432 + ], + "score": 1.0, + "content": "grid with 8 unique symbols.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 230, + 440, + 380, + 508 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 230, + 440, + 380, + 508 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 230, + 440, + 380, + 508 + ], + "spans": [ + { + "bbox": [ + 230, + 440, + 380, + 508 + ], + "score": 0.972, + "html": "
Task SequencesGrid
i704.7 ± 12.825.6 ± 1.8
ii709.4 ± 13.8623.7 ± 14.1
ii709.7 ± 14.0768.2 ±12.5
iv624.8 ± 12.4795.2 ± 10.3
", + "type": "table", + "image_path": "164eed252519d30abfd4208cfebc52987bd6e5948e3125927523777305272436.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 440, + 380, + 474.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 230, + 474.0, + 380, + 508.0 + ], + "spans": [], + "index": 25 + } + ] + } + ], + "index": 24.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 80, + 228, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 230, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 230, + 97 + ], + "score": 1.0, + "content": "C TRAINING DETAILS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 105, + 250, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 252, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 252, + 119 + ], + "score": 1.0, + "content": "C.1 UNIFICATION MLP & CNN", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 126, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "score": 1.0, + "content": "Both unification models are trained on a 5-fold cross-validation over the generated datasets for 2000", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "iterations with a batch size of 64. We don’t use any weight decay and save the training and test", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 149, + 330, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 330, + 160 + ], + "score": 1.0, + "content": "accuracies every 10 iterations, as presented in Figure 3.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 108, + 173, + 285, + 185 + ], + "lines": [ + { + "bbox": [ + 106, + 172, + 286, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 286, + 186 + ], + "score": 1.0, + "content": "C.2 UNIFICATION MEMORY NETWORKS", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 108, + 193, + 504, + 252 + ], + "lines": [ + { + "bbox": [ + 105, + 193, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 286, + 207 + ], + "score": 1.0, + "content": "We again use a batch size of 64 and pre-train", + "type": "text" + }, + { + "bbox": [ + 286, + 195, + 293, + 205 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 193, + 370, + 207 + ], + "score": 1.0, + "content": "for 40 epochs then", + "type": "text" + }, + { + "bbox": [ + 370, + 195, + 378, + 205 + ], + "score": 0.87, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 193, + 433, + 207 + ], + "score": 1.0, + "content": "together with", + "type": "text" + }, + { + "bbox": [ + 434, + 196, + 440, + 205 + ], + "score": 0.8, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 193, + 505, + 207 + ], + "score": 1.0, + "content": "for 260 epochs.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 204, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 372, + 218 + ], + "score": 1.0, + "content": "We use epochs for UMN since the dataset sizes are fixed. To learn", + "type": "text" + }, + { + "bbox": [ + 373, + 207, + 380, + 217 + ], + "score": 0.79, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 204, + 421, + 218 + ], + "score": 1.0, + "content": "alongside", + "type": "text" + }, + { + "bbox": [ + 421, + 206, + 428, + 217 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 204, + 506, + 218 + ], + "score": 1.0, + "content": ", we combine error", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 216, + 506, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 506, + 229 + ], + "score": 1.0, + "content": "signals from the unification of the invariant and the example. Following equation 4, the objective", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 227, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 349, + 239 + ], + "score": 1.0, + "content": "function not only incorporates the negative log-likelihood", + "type": "text" + }, + { + "bbox": [ + 349, + 227, + 365, + 238 + ], + "score": 0.89, + "content": "{ \\mathcal L } _ { \\mathrm { n l l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 227, + 505, + 239 + ], + "score": 1.0, + "content": "of the answer but also the mean", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 238, + 470, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 274, + 252 + ], + "score": 1.0, + "content": "squared error between intermediate states", + "type": "text" + }, + { + "bbox": [ + 275, + 238, + 287, + 252 + ], + "score": 0.9, + "content": "h _ { I } ^ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 239, + 305, + 252 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 305, + 238, + 321, + 252 + ], + "score": 0.92, + "content": "h _ { K } ^ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 239, + 470, + 252 + ], + "score": 1.0, + "content": "at each iteration as an auxiliary loss:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8 + }, + { + "type": "interline_equation", + "bbox": [ + 133, + 257, + 478, + 297 + ], + "lines": [ + { + "bbox": [ + 133, + 257, + 478, + 297 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 478, + 297 + ], + "score": 0.93, + "content": "J = \\mathcal { L } _ { \\mathrm { n l l } } ( f ( K ) , a ) + \\lambda _ { U } \\left[ \\mathcal { L } _ { \\mathrm { n l l } } ( f \\circ g ( I , K ) , a ) + \\frac { 1 } { J } \\sum _ { j = 1 } ^ { J } ( h _ { I } ^ { j } - h _ { K } ^ { j } ) ^ { 2 } + \\tau \\sum _ { s \\in \\mathcal { S } } \\psi _ { w } ( s ) \\right]", + "type": "interline_equation", + "image_path": "d642031dec0d0c1fb74310b00fa6ec184944a8e1c66f93234e97a793ae2cb952.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 133, + 257, + 478, + 270.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 133, + 270.3333333333333, + 478, + 283.66666666666663 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 133, + 283.66666666666663, + 478, + 296.99999999999994 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 301, + 504, + 346 + ], + "lines": [ + { + "bbox": [ + 106, + 301, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 200, + 314 + ], + "score": 1.0, + "content": "We pre-train by setting", + "type": "text" + }, + { + "bbox": [ + 201, + 302, + 234, + 312 + ], + "score": 0.91, + "content": "\\lambda _ { U } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 301, + 342, + 314 + ], + "score": 1.0, + "content": "for 40 epochs and then set", + "type": "text" + }, + { + "bbox": [ + 343, + 302, + 375, + 312 + ], + "score": 0.91, + "content": "\\lambda _ { U } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 301, + 506, + 314 + ], + "score": 1.0, + "content": ". For strong supervision we also", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 312, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 254, + 325 + ], + "score": 1.0, + "content": "compute the negative log-likelihood", + "type": "text" + }, + { + "bbox": [ + 255, + 313, + 271, + 323 + ], + "score": 0.89, + "content": "{ \\mathcal { L } } _ { \\mathrm { n l l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 312, + 372, + 325 + ], + "score": 1.0, + "content": "for the context attention", + "type": "text" + }, + { + "bbox": [ + 372, + 312, + 384, + 324 + ], + "score": 0.88, + "content": "\\beta ^ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 312, + 506, + 325 + ], + "score": 1.0, + "content": ", described in Appendix A, at", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 324, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 335 + ], + "score": 1.0, + "content": "each iteration using the supporting facts of the tasks. We apply a dropout of 0.1 for all recurrent", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 334, + 485, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 485, + 347 + ], + "score": 1.0, + "content": "neural networks used and only for the bAbI dataset weight decay with 0.001 as the coefficient.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 80, + 228, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 230, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 230, + 97 + ], + "score": 1.0, + "content": "C TRAINING DETAILS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 105, + 250, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 252, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 252, + 119 + ], + "score": 1.0, + "content": "C.1 UNIFICATION MLP & CNN", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 126, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "score": 1.0, + "content": "Both unification models are trained on a 5-fold cross-validation over the generated datasets for 2000", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "iterations with a batch size of 64. We don’t use any weight decay and save the training and test", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 149, + 330, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 330, + 160 + ], + "score": 1.0, + "content": "accuracies every 10 iterations, as presented in Figure 3.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 127, + 505, + 160 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 173, + 285, + 185 + ], + "lines": [ + { + "bbox": [ + 106, + 172, + 286, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 286, + 186 + ], + "score": 1.0, + "content": "C.2 UNIFICATION MEMORY NETWORKS", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 108, + 193, + 504, + 252 + ], + "lines": [ + { + "bbox": [ + 105, + 193, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 286, + 207 + ], + "score": 1.0, + "content": "We again use a batch size of 64 and pre-train", + "type": "text" + }, + { + "bbox": [ + 286, + 195, + 293, + 205 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 193, + 370, + 207 + ], + "score": 1.0, + "content": "for 40 epochs then", + "type": "text" + }, + { + "bbox": [ + 370, + 195, + 378, + 205 + ], + "score": 0.87, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 193, + 433, + 207 + ], + "score": 1.0, + "content": "together with", + "type": "text" + }, + { + "bbox": [ + 434, + 196, + 440, + 205 + ], + "score": 0.8, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 193, + 505, + 207 + ], + "score": 1.0, + "content": "for 260 epochs.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 204, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 372, + 218 + ], + "score": 1.0, + "content": "We use epochs for UMN since the dataset sizes are fixed. To learn", + "type": "text" + }, + { + "bbox": [ + 373, + 207, + 380, + 217 + ], + "score": 0.79, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 204, + 421, + 218 + ], + "score": 1.0, + "content": "alongside", + "type": "text" + }, + { + "bbox": [ + 421, + 206, + 428, + 217 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 204, + 506, + 218 + ], + "score": 1.0, + "content": ", we combine error", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 216, + 506, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 506, + 229 + ], + "score": 1.0, + "content": "signals from the unification of the invariant and the example. Following equation 4, the objective", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 227, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 349, + 239 + ], + "score": 1.0, + "content": "function not only incorporates the negative log-likelihood", + "type": "text" + }, + { + "bbox": [ + 349, + 227, + 365, + 238 + ], + "score": 0.89, + "content": "{ \\mathcal L } _ { \\mathrm { n l l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 227, + 505, + 239 + ], + "score": 1.0, + "content": "of the answer but also the mean", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 238, + 470, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 274, + 252 + ], + "score": 1.0, + "content": "squared error between intermediate states", + "type": "text" + }, + { + "bbox": [ + 275, + 238, + 287, + 252 + ], + "score": 0.9, + "content": "h _ { I } ^ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 239, + 305, + 252 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 305, + 238, + 321, + 252 + ], + "score": 0.92, + "content": "h _ { K } ^ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 239, + 470, + 252 + ], + "score": 1.0, + "content": "at each iteration as an auxiliary loss:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 193, + 506, + 252 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 133, + 257, + 478, + 297 + ], + "lines": [ + { + "bbox": [ + 133, + 257, + 478, + 297 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 478, + 297 + ], + "score": 0.93, + "content": "J = \\mathcal { L } _ { \\mathrm { n l l } } ( f ( K ) , a ) + \\lambda _ { U } \\left[ \\mathcal { L } _ { \\mathrm { n l l } } ( f \\circ g ( I , K ) , a ) + \\frac { 1 } { J } \\sum _ { j = 1 } ^ { J } ( h _ { I } ^ { j } - h _ { K } ^ { j } ) ^ { 2 } + \\tau \\sum _ { s \\in \\mathcal { S } } \\psi _ { w } ( s ) \\right]", + "type": "interline_equation", + "image_path": "d642031dec0d0c1fb74310b00fa6ec184944a8e1c66f93234e97a793ae2cb952.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 133, + 257, + 478, + 270.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 133, + 270.3333333333333, + 478, + 283.66666666666663 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 133, + 283.66666666666663, + 478, + 296.99999999999994 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 301, + 504, + 346 + ], + "lines": [ + { + "bbox": [ + 106, + 301, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 200, + 314 + ], + "score": 1.0, + "content": "We pre-train by setting", + "type": "text" + }, + { + "bbox": [ + 201, + 302, + 234, + 312 + ], + "score": 0.91, + "content": "\\lambda _ { U } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 301, + 342, + 314 + ], + "score": 1.0, + "content": "for 40 epochs and then set", + "type": "text" + }, + { + "bbox": [ + 343, + 302, + 375, + 312 + ], + "score": 0.91, + "content": "\\lambda _ { U } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 301, + 506, + 314 + ], + "score": 1.0, + "content": ". For strong supervision we also", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 312, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 254, + 325 + ], + "score": 1.0, + "content": "compute the negative log-likelihood", + "type": "text" + }, + { + "bbox": [ + 255, + 313, + 271, + 323 + ], + "score": 0.89, + "content": "{ \\mathcal { L } } _ { \\mathrm { n l l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 312, + 372, + 325 + ], + "score": 1.0, + "content": "for the context attention", + "type": "text" + }, + { + "bbox": [ + 372, + 312, + 384, + 324 + ], + "score": 0.88, + "content": "\\beta ^ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 312, + 506, + 325 + ], + "score": 1.0, + "content": ", described in Appendix A, at", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 324, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 335 + ], + "score": 1.0, + "content": "each iteration using the supporting facts of the tasks. We apply a dropout of 0.1 for all recurrent", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 334, + 485, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 485, + 347 + ], + "score": 1.0, + "content": "neural networks used and only for the bAbI dataset weight decay with 0.001 as the coefficient.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 301, + 506, + 347 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 227, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 228, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 228, + 96 + ], + "score": 1.0, + "content": "D FURTHER RESULTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 128, + 102, + 483, + 261 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 102, + 483, + 261 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 102, + 483, + 261 + ], + "spans": [ + { + "bbox": [ + 128, + 102, + 483, + 261 + ], + "score": 0.972, + "type": "image", + "image_path": "c02ba1f113eff3023fe0e72f8db887cb964bac4e0057d50edcd72e8e1f55364f.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 128, + 102, + 483, + 155.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 128, + 155.0, + 483, + 208.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 128, + 208.0, + 483, + 261.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 271, + 506, + 316 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "Figure 6: Results of Unification MLP and CNN on increasing number of invariants. There is no", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 282, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 294 + ], + "score": 1.0, + "content": "impact on performance when more invariants per task are given. Upon closer inspection, we noticed", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 448, + 306 + ], + "score": 1.0, + "content": "the models ignore the extra invariants and only use 1. We speculate the regularisation", + "type": "text" + }, + { + "bbox": [ + 448, + 294, + 457, + 305 + ], + "score": 0.83, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "encourages", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 304, + 258, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 258, + 316 + ], + "score": 1.0, + "content": "the models to use a single 1 invariant.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + } + ], + "index": 3.75 + }, + { + "type": "table", + "bbox": [ + 209, + 340, + 400, + 617 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 135, + 319, + 474, + 331 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 318, + 475, + 332 + ], + "spans": [ + { + "bbox": [ + 135, + 318, + 475, + 332 + ], + "score": 1.0, + "content": "Table 6: Individual task error rates on bAbI tasks for Unification Memory Networks.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "table_body", + "bbox": [ + 209, + 340, + 400, + 617 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 209, + 340, + 400, + 617 + ], + "spans": [ + { + "bbox": [ + 209, + 340, + 400, + 617 + ], + "score": 0.979, + "html": "
SupervisionWeakStrong
#Invs Training Size1 1k3 1k1 1k3 1k3 50
10.00.00.00.01.4
265.663.10.30.730.0
367.162.61.02.439.8
40.00.00.00.037.0
53.44.00.81.126.5
60.20.60.00.018.4
722.022.810.711.322.8
810.38.57.47.624.7
90.125.70.00.033.8
100.12.00.00.332.6
110.00.00.00.011.9
120.00.10.00.021.3
132.13.70.00.15.8
1419.713.50.50.154.8
150.00.70.00.00.0
1655.256.20.00.039.7
1739.249.051.149.348.8
184.48.00.60.510.4
1991.889.653.946.790.2
200.00.00.00.02.7
Mean19.120.56.36.027.6
Std27.927.015.614.321.0
#>5%8104417
", + "type": "table", + "image_path": "7d93fa930a25e39a8a24ed1d60a6f940986c8670a778209ead567218a03c63c2.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 209, + 340, + 400, + 352.59090909090907 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 209, + 352.59090909090907, + 400, + 365.18181818181813 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 209, + 365.18181818181813, + 400, + 377.7727272727272 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 209, + 377.7727272727272, + 400, + 390.36363636363626 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 209, + 390.36363636363626, + 400, + 402.9545454545453 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 209, + 402.9545454545453, + 400, + 415.5454545454544 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 209, + 415.5454545454544, + 400, + 428.13636363636346 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 209, + 428.13636363636346, + 400, + 440.7272727272725 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 209, + 440.7272727272725, + 400, + 453.3181818181816 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 209, + 453.3181818181816, + 400, + 465.90909090909065 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 209, + 465.90909090909065, + 400, + 478.4999999999997 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 209, + 478.4999999999997, + 400, + 491.0909090909088 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 209, + 491.0909090909088, + 400, + 503.68181818181785 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 209, + 503.68181818181785, + 400, + 516.2727272727269 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 209, + 516.2727272727269, + 400, + 528.863636363636 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 209, + 528.863636363636, + 400, + 541.4545454545452 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 209, + 541.4545454545452, + 400, + 554.0454545454543 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 209, + 554.0454545454543, + 400, + 566.6363636363634 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 209, + 566.6363636363634, + 400, + 579.2272727272725 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 209, + 579.2272727272725, + 400, + 591.8181818181816 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 209, + 591.8181818181816, + 400, + 604.4090909090908 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 209, + 604.4090909090908, + 400, + 616.9999999999999 + ], + "spans": [], + "index": 30 + } + ] + } + ], + "index": 13.75 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 227, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 228, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 228, + 96 + ], + "score": 1.0, + "content": "D FURTHER RESULTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 128, + 102, + 483, + 261 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 102, + 483, + 261 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 102, + 483, + 261 + ], + "spans": [ + { + "bbox": [ + 128, + 102, + 483, + 261 + ], + "score": 0.972, + "type": "image", + "image_path": "c02ba1f113eff3023fe0e72f8db887cb964bac4e0057d50edcd72e8e1f55364f.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 128, + 102, + 483, + 155.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 128, + 155.0, + 483, + 208.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 128, + 208.0, + 483, + 261.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 271, + 506, + 316 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "Figure 6: Results of Unification MLP and CNN on increasing number of invariants. There is no", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 282, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 294 + ], + "score": 1.0, + "content": "impact on performance when more invariants per task are given. Upon closer inspection, we noticed", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 448, + 306 + ], + "score": 1.0, + "content": "the models ignore the extra invariants and only use 1. We speculate the regularisation", + "type": "text" + }, + { + "bbox": [ + 448, + 294, + 457, + 305 + ], + "score": 0.83, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "encourages", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 304, + 258, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 258, + 316 + ], + "score": 1.0, + "content": "the models to use a single 1 invariant.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + } + ], + "index": 3.75 + }, + { + "type": "table", + "bbox": [ + 209, + 340, + 400, + 617 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 135, + 319, + 474, + 331 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 318, + 475, + 332 + ], + "spans": [ + { + "bbox": [ + 135, + 318, + 475, + 332 + ], + "score": 1.0, + "content": "Table 6: Individual task error rates on bAbI tasks for Unification Memory Networks.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "table_body", + "bbox": [ + 209, + 340, + 400, + 617 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 209, + 340, + 400, + 617 + ], + "spans": [ + { + "bbox": [ + 209, + 340, + 400, + 617 + ], + "score": 0.979, + "html": "
SupervisionWeakStrong
#Invs Training Size1 1k3 1k1 1k3 1k3 50
10.00.00.00.01.4
265.663.10.30.730.0
367.162.61.02.439.8
40.00.00.00.037.0
53.44.00.81.126.5
60.20.60.00.018.4
722.022.810.711.322.8
810.38.57.47.624.7
90.125.70.00.033.8
100.12.00.00.332.6
110.00.00.00.011.9
120.00.10.00.021.3
132.13.70.00.15.8
1419.713.50.50.154.8
150.00.70.00.00.0
1655.256.20.00.039.7
1739.249.051.149.348.8
184.48.00.60.510.4
1991.889.653.946.790.2
200.00.00.00.02.7
Mean19.120.56.36.027.6
Std27.927.015.614.321.0
#>5%8104417
", + "type": "table", + "image_path": "7d93fa930a25e39a8a24ed1d60a6f940986c8670a778209ead567218a03c63c2.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 209, + 340, + 400, + 352.59090909090907 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 209, + 352.59090909090907, + 400, + 365.18181818181813 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 209, + 365.18181818181813, + 400, + 377.7727272727272 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 209, + 377.7727272727272, + 400, + 390.36363636363626 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 209, + 390.36363636363626, + 400, + 402.9545454545453 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 209, + 402.9545454545453, + 400, + 415.5454545454544 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 209, + 415.5454545454544, + 400, + 428.13636363636346 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 209, + 428.13636363636346, + 400, + 440.7272727272725 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 209, + 440.7272727272725, + 400, + 453.3181818181816 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 209, + 453.3181818181816, + 400, + 465.90909090909065 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 209, + 465.90909090909065, + 400, + 478.4999999999997 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 209, + 478.4999999999997, + 400, + 491.0909090909088 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 209, + 491.0909090909088, + 400, + 503.68181818181785 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 209, + 503.68181818181785, + 400, + 516.2727272727269 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 209, + 516.2727272727269, + 400, + 528.863636363636 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 209, + 528.863636363636, + 400, + 541.4545454545452 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 209, + 541.4545454545452, + 400, + 554.0454545454543 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 209, + 554.0454545454543, + 400, + 566.6363636363634 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 209, + 566.6363636363634, + 400, + 579.2272727272725 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 209, + 579.2272727272725, + 400, + 591.8181818181816 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 209, + 591.8181818181816, + 400, + 604.4090909090908 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 209, + 604.4090909090908, + 400, + 616.9999999999999 + ], + "spans": [], + "index": 30 + } + ] + } + ], + "index": 13.75 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 115, + 165, + 495, + 433 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 100, + 505, + 157 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 101, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 101, + 315, + 113 + ], + "score": 1.0, + "content": "Table 7: Comparison of individual task error rates", + "type": "text" + }, + { + "bbox": [ + 316, + 102, + 331, + 112 + ], + "score": 0.72, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 101, + 505, + 113 + ], + "score": 1.0, + "content": "on the bAbI (Weston et al., 2016) dataset", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "score": 1.0, + "content": "of the best run. We preferred 1k results if a model had experiments published on both 1k and 10k", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 123, + 504, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 504, + 135 + ], + "score": 1.0, + "content": "for data efficiency. References from left to right: (Sukhbaatar et al., 2015) - (Liu & Perez, 2017) -", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 133, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 505, + 147 + ], + "score": 1.0, + "content": "(Henaff et al., 2017) - (Seo et al., 2017) - Ours - (Xiong et al., 2016) - (Graves et al., 2016) - (Weston", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 145, + 273, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 273, + 158 + ], + "score": 1.0, + "content": "et al., 2015) - Ours - (Kumar et al., 2016)", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "table_body", + "bbox": [ + 115, + 165, + 495, + 433 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 165, + 495, + 433 + ], + "spans": [ + { + "bbox": [ + 115, + 165, + 495, + 433 + ], + "score": 0.985, + "html": "
Support Size ModelWeakStrong
GN2N1k EntNetQRNUMN10k DMN+ DNC1k MemNN UMN10k DMN
10.00.00.70.00.00.00.00.00.00.0
28.38.156.40.565.60.30.40.00.71.8
340.338.869.71.267.11.11.80.02.44.8
42.80.41.40.70.00.00.00.00.00.0
513.11.04.61.23.40.50.82.01.10.7
67.68.430.01.20.20.00.00.00.00.0
717.317.822.39.422.02.40.615.011.33.1
810.012.519.23.710.30.00.39.07.63.5
913.210.731.50.00.10.00.20.00.00.0
1015.116.515.60.00.10.00.22.00.32.5
110.90.08.00.00.00.00.00.00.00.1
120.20.00.80.00.00.00.00.00.00.0
130.40.09.00.32.10.00.10.00.10.2
141.71.262.93.819.70.20.41.00.10.0
150.00.057.80.00.00.00.00.00.00.0
161.30.153.253.455.245.355.10.00.00.6
1751.041.746.451.839.24.212.035.049.340.6
1811.19.28.88.84.42.10.85.00.54.7
1982.888.590.490.791.80.03.964.046.765.5
200.00.02.60.30.00.00.00.00.00.0
Mean
#>5%13.9 1112.7 1029.6 1511.3 519.1 82.8 13.8 26.7 46.0 46.4 2
", + "type": "table", + "image_path": "ecc837386c7d436084e3ba2bb9df9b93eb4bf84050d41176fe6f0c107c83ab6a.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 115, + 165, + 495, + 254.33333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 115, + 254.33333333333331, + 495, + 343.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 115, + 343.66666666666663, + 495, + 432.99999999999994 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 4.0 + }, + { + "type": "table", + "bbox": [ + 109, + 501, + 504, + 709 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 161, + 481, + 449, + 493 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 160, + 479, + 450, + 495 + ], + "spans": [ + { + "bbox": [ + 160, + 479, + 302, + 495 + ], + "score": 1.0, + "content": "Table 8: Individual task error rates", + "type": "text" + }, + { + "bbox": [ + 302, + 482, + 318, + 492 + ], + "score": 0.75, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 479, + 450, + 495 + ], + "score": 1.0, + "content": "on the logical reasoning dataset.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "table_body", + "bbox": [ + 109, + 501, + 504, + 709 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 501, + 504, + 709 + ], + "spans": [ + { + "bbox": [ + 109, + 501, + 504, + 709 + ], + "score": 0.984, + "html": "
Size1k10k5020k
SupportWeakStrongWeakWeak
Arity1111Strong 22
# Invs /Model12 312 312 312 132 3DMN2 IMA
Facts1.20.90.00.40.00.00.00.00.033.50.00.0
Unification0.010.30.010.80.00.00.00.00.041.313.010.0
1 Step50.349.84.420.01.227.80.11.35.750.226.02.0
2 Steps47.550.05.735.037.247.80.029.728.749.933.05.0
3 Steps47.649.210.438.739.645.60.026.026.148.323.06.0
AND31.337.410.716.429.829.00.20.41.250.020.05.0
OR25.238.121.035.020.530.24.420.617.447.613.03.0
Transitivity50.026.639.65.06.049.250.050.0
1 Step NBF46.438.73.828.81.121.60.11.18.047.621.02.0
2 Steps NBF48.548.97.739.630.448.20.133.428.750.315.04.0
AND NBF51.050.143.148.629.444.20.11.340.149.516.08.0
OR NBF51.448.450.847.347.647.821.327.630.547.325.014.0
Mean36.439.314.328.921.531.82.412.216.047.121.29.1
Std18.715.916.414.117.116.76.113.213.64.712.313.4
#>5%91171171015912115
", + "type": "table", + "image_path": "0ff92c677511d717c131f4e89e502bc5ae32309f23904fe815fac80a615b0be8.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 109, + 501, + 504, + 570.3333333333334 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 109, + 570.3333333333334, + 504, + 639.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 109, + 639.6666666666667, + 504, + 709.0000000000001 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 9.0 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 115, + 165, + 495, + 433 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 100, + 505, + 157 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 101, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 101, + 315, + 113 + ], + "score": 1.0, + "content": "Table 7: Comparison of individual task error rates", + "type": "text" + }, + { + "bbox": [ + 316, + 102, + 331, + 112 + ], + "score": 0.72, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 101, + 505, + 113 + ], + "score": 1.0, + "content": "on the bAbI (Weston et al., 2016) dataset", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "score": 1.0, + "content": "of the best run. We preferred 1k results if a model had experiments published on both 1k and 10k", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 123, + 504, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 504, + 135 + ], + "score": 1.0, + "content": "for data efficiency. References from left to right: (Sukhbaatar et al., 2015) - (Liu & Perez, 2017) -", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 133, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 505, + 147 + ], + "score": 1.0, + "content": "(Henaff et al., 2017) - (Seo et al., 2017) - Ours - (Xiong et al., 2016) - (Graves et al., 2016) - (Weston", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 145, + 273, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 273, + 158 + ], + "score": 1.0, + "content": "et al., 2015) - Ours - (Kumar et al., 2016)", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "table_body", + "bbox": [ + 115, + 165, + 495, + 433 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 165, + 495, + 433 + ], + "spans": [ + { + "bbox": [ + 115, + 165, + 495, + 433 + ], + "score": 0.985, + "html": "
Support Size ModelWeakStrong
GN2N1k EntNetQRNUMN10k DMN+ DNC1k MemNN UMN10k DMN
10.00.00.70.00.00.00.00.00.00.0
28.38.156.40.565.60.30.40.00.71.8
340.338.869.71.267.11.11.80.02.44.8
42.80.41.40.70.00.00.00.00.00.0
513.11.04.61.23.40.50.82.01.10.7
67.68.430.01.20.20.00.00.00.00.0
717.317.822.39.422.02.40.615.011.33.1
810.012.519.23.710.30.00.39.07.63.5
913.210.731.50.00.10.00.20.00.00.0
1015.116.515.60.00.10.00.22.00.32.5
110.90.08.00.00.00.00.00.00.00.1
120.20.00.80.00.00.00.00.00.00.0
130.40.09.00.32.10.00.10.00.10.2
141.71.262.93.819.70.20.41.00.10.0
150.00.057.80.00.00.00.00.00.00.0
161.30.153.253.455.245.355.10.00.00.6
1751.041.746.451.839.24.212.035.049.340.6
1811.19.28.88.84.42.10.85.00.54.7
1982.888.590.490.791.80.03.964.046.765.5
200.00.02.60.30.00.00.00.00.00.0
Mean
#>5%13.9 1112.7 1029.6 1511.3 519.1 82.8 13.8 26.7 46.0 46.4 2
", + "type": "table", + "image_path": "ecc837386c7d436084e3ba2bb9df9b93eb4bf84050d41176fe6f0c107c83ab6a.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 115, + 165, + 495, + 254.33333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 115, + 254.33333333333331, + 495, + 343.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 115, + 343.66666666666663, + 495, + 432.99999999999994 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 4.0 + }, + { + "type": "table", + "bbox": [ + 109, + 501, + 504, + 709 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 161, + 481, + 449, + 493 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 160, + 479, + 450, + 495 + ], + "spans": [ + { + "bbox": [ + 160, + 479, + 302, + 495 + ], + "score": 1.0, + "content": "Table 8: Individual task error rates", + "type": "text" + }, + { + "bbox": [ + 302, + 482, + 318, + 492 + ], + "score": 0.75, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 479, + 450, + 495 + ], + "score": 1.0, + "content": "on the logical reasoning dataset.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "table_body", + "bbox": [ + 109, + 501, + 504, + 709 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 501, + 504, + 709 + ], + "spans": [ + { + "bbox": [ + 109, + 501, + 504, + 709 + ], + "score": 0.984, + "html": "
Size1k10k5020k
SupportWeakStrongWeakWeak
Arity1111Strong 22
# Invs /Model12 312 312 312 132 3DMN2 IMA
Facts1.20.90.00.40.00.00.00.00.033.50.00.0
Unification0.010.30.010.80.00.00.00.00.041.313.010.0
1 Step50.349.84.420.01.227.80.11.35.750.226.02.0
2 Steps47.550.05.735.037.247.80.029.728.749.933.05.0
3 Steps47.649.210.438.739.645.60.026.026.148.323.06.0
AND31.337.410.716.429.829.00.20.41.250.020.05.0
OR25.238.121.035.020.530.24.420.617.447.613.03.0
Transitivity50.026.639.65.06.049.250.050.0
1 Step NBF46.438.73.828.81.121.60.11.18.047.621.02.0
2 Steps NBF48.548.97.739.630.448.20.133.428.750.315.04.0
AND NBF51.050.143.148.629.444.20.11.340.149.516.08.0
OR NBF51.448.450.847.347.647.821.327.630.547.325.014.0
Mean36.439.314.328.921.531.82.412.216.047.121.29.1
Std18.715.916.414.117.116.76.113.213.64.712.313.4
#>5%91171171015912115
", + "type": "table", + "image_path": "0ff92c677511d717c131f4e89e502bc5ae32309f23904fe815fac80a615b0be8.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 109, + 501, + 504, + 570.3333333333334 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 109, + 570.3333333333334, + 504, + 639.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 109, + 639.6666666666667, + 504, + 709.0000000000001 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 9.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 224, + 82, + 387, + 124 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 224, + 82, + 387, + 124 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 224, + 82, + 387, + 124 + ], + "spans": [ + { + "bbox": [ + 224, + 82, + 387, + 124 + ], + "score": 0.267, + "html": "
Xsandra went back to the Y:bathroom
is X:sandra in the Y:bathroom
yes
", + "type": "table", + "image_path": "d8edacdbc450e3ee1dc957f53d9955390327d1fc6d5530051b7e0a7c2cf7a8b4.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 224, + 82, + 387, + 96.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 224, + 96.0, + 387, + 110.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 224, + 110.0, + 387, + 124.0 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 129, + 133, + 480, + 145 + ], + "lines": [ + { + "bbox": [ + 128, + 133, + 483, + 147 + ], + "spans": [ + { + "bbox": [ + 128, + 133, + 483, + 147 + ], + "score": 1.0, + "content": "Figure 7: bAbI task 6, yes or no questions. The invariant does not variablise the answer.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 149, + 400, + 207 + ], + "lines": [ + { + "bbox": [ + 210, + 149, + 400, + 207 + ], + "spans": [ + { + "bbox": [ + 210, + 149, + 400, + 207 + ], + "score": 0.93, + "content": "\\begin{array} { r l } { X { \\mathrm { : m ~ ( ~ Y : e ~ ) ~ } } } & { { } { \\vdash X { \\mathrm { : m ~ ( ~ Y : e ~ ) ~ } } } } \\\\ { X { \\mathrm { : a ~ ( ~ } Y { \\mathrm { : w ~ , ~ } } { \\mathrm { ~ Z : e ~ ) ~ } } } } & { { } { \\vdash X { \\mathrm { : a ~ ( ~ } Y { \\mathrm { : w ~ , ~ } } { \\mathrm { Z : e ~ ) ~ } } } } } \\\\ { X { \\mathrm { : m ~ ( ~ T ~ ) ~ } } } & { { } { \\vdash X { \\mathrm { : m ~ ( ~ c ~ ) ~ } } } } \\\\ { X { \\mathrm { : x ~ ( ~ \\mathbb { A } ~ ) n o t ~ \\mathbf { Z } { \\mathrm { : q } } ~ ( ~ \\mathbb { A } ~ ) ~ } } } & { { } { \\vdash X { \\mathrm { : x ~ ( ~ } Y { \\mathrm { : z } } ~ ) } } } \\end{array}", + "type": "interline_equation", + "image_path": "53b241b6490c42e9c05a37ee42eac354b639bb17d0da4793cbec9044f1ca384f.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 210, + 149, + 400, + 163.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 210, + 163.5, + 400, + 178.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 210, + 178.0, + 400, + 192.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 210, + 192.5, + 400, + 207.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 215, + 504, + 248 + ], + "lines": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "score": 1.0, + "content": "Figure 8: Invariants learned on tasks 1, 2 and 11 with arity 1 and 2 from the logical reasoning dataset.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "score": 1.0, + "content": "Last invariant on task 11 lifts the example around the negation by failure, denoted as not, capturing", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 238, + 163, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 163, + 249 + ], + "score": 1.0, + "content": "its semantics.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 168, + 252, + 242, + 297 + ], + "lines": [ + { + "bbox": [ + 167, + 251, + 226, + 264 + ], + "spans": [ + { + "bbox": [ + 167, + 251, + 226, + 264 + ], + "score": 1.0, + "content": "3 X 7 Y head", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 167, + 262, + 226, + 275 + ], + "spans": [ + { + "bbox": [ + 167, + 262, + 226, + 275 + ], + "score": 1.0, + "content": "7 4 X Y head", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 167, + 273, + 221, + 286 + ], + "spans": [ + { + "bbox": [ + 167, + 273, + 221, + 286 + ], + "score": 1.0, + "content": "X 3 1 X tail", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 165, + 285, + 243, + 298 + ], + "spans": [ + { + "bbox": [ + 165, + 285, + 243, + 298 + ], + "score": 1.0, + "content": "X X 5 Y duplicate", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "interline_equation", + "bbox": [ + 351, + 250, + 465, + 297 + ], + "lines": [ + { + "bbox": [ + 351, + 250, + 465, + 297 + ], + "spans": [ + { + "bbox": [ + 351, + 250, + 465, + 297 + ], + "score": 0.27, + "content": "\\begin{array} { r l } { X \\mathrm { ~ 0 ~ ~ 0 ~ ~ } } & { \\mathrm { ~ 0 ~ ~ Y ~ 0 ~ ~ 6 ~ 4 ~ } Y } \\\\ { \\mathrm { ~ 0 ~ 1 ~ 0 ~ } } & { \\mathrm { ~ 1 ~ } X \\mathrm { ~ 8 ~ } } \\\\ { \\mathrm { ~ 0 ~ 0 ~ } Y } & { \\mathrm { ~ 0 ~ 2 ~ 0 ~ } } \\end{array}", + "type": "interline_equation", + "image_path": "af9a832b2966a56b80d8d6d927634fd72447a59e0433605421d6b017bdc935e8.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 351, + 250, + 465, + 273.5 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 351, + 273.5, + 465, + 297.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 302, + 493, + 314 + ], + "lines": [ + { + "bbox": [ + 105, + 300, + 494, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 494, + 314 + ], + "score": 1.0, + "content": "(a) Invariants with extra varialbes learned with UMLP. (b) Mismatching invariants learned with UCNN.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 322, + 506, + 366 + ], + "lines": [ + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "score": 1.0, + "content": "Figure 9: Invariants learned that do not match the data generating distribution from UMLP and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 333, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 160, + 345 + ], + "score": 1.0, + "content": "UCNN using", + "type": "text" + }, + { + "bbox": [ + 161, + 333, + 193, + 344 + ], + "score": 0.85, + "content": "\\leq 1 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 333, + 506, + 345 + ], + "score": 1.0, + "content": "examples to train. In these instances the unification still bind to the the correct", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 344, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 506, + 357 + ], + "score": 1.0, + "content": "symbols in order to predict the desired answer; quantitatively we get the same results. Variable", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 354, + 263, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 263, + 367 + ], + "score": 1.0, + "content": "default symbols are omitted for clarity.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 470, + 236, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 469, + 236, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 236, + 480 + ], + "score": 1.0, + "content": "(a) bAbI task 2. When a variable", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 478, + 237, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 237, + 491 + ], + "score": 1.0, + "content": "is unused in the next iteration, e.g.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 488, + 236, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 236, + 500 + ], + "score": 1.0, + "content": "Z:football, it unifies with random", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 498, + 223, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 223, + 510 + ], + "score": 1.0, + "content": "tokens often biased by position.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5 + }, + { + "type": "image", + "bbox": [ + 117, + 374, + 496, + 461 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 374, + 496, + 461 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 374, + 496, + 461 + ], + "spans": [ + { + "bbox": [ + 117, + 374, + 496, + 461 + ], + "score": 0.5, + "type": "image", + "image_path": "9073cfbc8cdd1d7408ab2bf49ed3c03468c3850a67e1c59f65e66ccdfccc89e1.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 117, + 374, + 496, + 403.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 117, + 403.0, + 496, + 432.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 117, + 432.0, + 496, + 461.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 374, + 470, + 504, + 500 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 375, + 469, + 504, + 480 + ], + "spans": [ + { + "bbox": [ + 375, + 469, + 504, + 480 + ], + "score": 1.0, + "content": "(c) Logical reasoning task 3. Arity", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 375, + 478, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 375, + 478, + 505, + 490 + ], + "score": 1.0, + "content": "1 atom forced to bind with arity 2", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 375, + 489, + 493, + 501 + ], + "spans": [ + { + "bbox": [ + 375, + 489, + 493, + 501 + ], + "score": 1.0, + "content": "creates a one-to-many mapping.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + } + ], + "index": 28.0 + }, + { + "type": "text", + "bbox": [ + 240, + 470, + 370, + 500 + ], + "lines": [ + { + "bbox": [ + 241, + 470, + 370, + 480 + ], + "spans": [ + { + "bbox": [ + 241, + 470, + 370, + 480 + ], + "score": 1.0, + "content": "(b) Logical reasoning task 1. A", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 240, + 480, + 370, + 489 + ], + "spans": [ + { + "bbox": [ + 240, + 480, + 370, + 489 + ], + "score": 1.0, + "content": "one-to-one alignment is created be-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 240, + 489, + 356, + 500 + ], + "spans": [ + { + "bbox": [ + 240, + 489, + 356, + 500 + ], + "score": 1.0, + "content": "tween predicates and constants.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 116, + 519, + 493, + 531 + ], + "lines": [ + { + "bbox": [ + 116, + 518, + 494, + 533 + ], + "spans": [ + { + "bbox": [ + 116, + 518, + 494, + 533 + ], + "score": 1.0, + "content": "Figure 10: Further attention maps for equation 3, darker cells indicate higher attention values.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "17", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 224, + 82, + 387, + 124 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 224, + 82, + 387, + 124 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 224, + 82, + 387, + 124 + ], + "spans": [ + { + "bbox": [ + 224, + 82, + 387, + 124 + ], + "score": 0.267, + "html": "
Xsandra went back to the Y:bathroom
is X:sandra in the Y:bathroom
yes
", + "type": "table", + "image_path": "d8edacdbc450e3ee1dc957f53d9955390327d1fc6d5530051b7e0a7c2cf7a8b4.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 224, + 82, + 387, + 96.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 224, + 96.0, + 387, + 110.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 224, + 110.0, + 387, + 124.0 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 129, + 133, + 480, + 145 + ], + "lines": [ + { + "bbox": [ + 128, + 133, + 483, + 147 + ], + "spans": [ + { + "bbox": [ + 128, + 133, + 483, + 147 + ], + "score": 1.0, + "content": "Figure 7: bAbI task 6, yes or no questions. The invariant does not variablise the answer.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3, + "bbox_fs": [ + 128, + 133, + 483, + 147 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 149, + 400, + 207 + ], + "lines": [ + { + "bbox": [ + 210, + 149, + 400, + 207 + ], + "spans": [ + { + "bbox": [ + 210, + 149, + 400, + 207 + ], + "score": 0.93, + "content": "\\begin{array} { r l } { X { \\mathrm { : m ~ ( ~ Y : e ~ ) ~ } } } & { { } { \\vdash X { \\mathrm { : m ~ ( ~ Y : e ~ ) ~ } } } } \\\\ { X { \\mathrm { : a ~ ( ~ } Y { \\mathrm { : w ~ , ~ } } { \\mathrm { ~ Z : e ~ ) ~ } } } } & { { } { \\vdash X { \\mathrm { : a ~ ( ~ } Y { \\mathrm { : w ~ , ~ } } { \\mathrm { Z : e ~ ) ~ } } } } } \\\\ { X { \\mathrm { : m ~ ( ~ T ~ ) ~ } } } & { { } { \\vdash X { \\mathrm { : m ~ ( ~ c ~ ) ~ } } } } \\\\ { X { \\mathrm { : x ~ ( ~ \\mathbb { A } ~ ) n o t ~ \\mathbf { Z } { \\mathrm { : q } } ~ ( ~ \\mathbb { A } ~ ) ~ } } } & { { } { \\vdash X { \\mathrm { : x ~ ( ~ } Y { \\mathrm { : z } } ~ ) } } } \\end{array}", + "type": "interline_equation", + "image_path": "53b241b6490c42e9c05a37ee42eac354b639bb17d0da4793cbec9044f1ca384f.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 210, + 149, + 400, + 163.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 210, + 163.5, + 400, + 178.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 210, + 178.0, + 400, + 192.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 210, + 192.5, + 400, + 207.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 215, + 504, + 248 + ], + "lines": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "score": 1.0, + "content": "Figure 8: Invariants learned on tasks 1, 2 and 11 with arity 1 and 2 from the logical reasoning dataset.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "score": 1.0, + "content": "Last invariant on task 11 lifts the example around the negation by failure, denoted as not, capturing", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 238, + 163, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 163, + 249 + ], + "score": 1.0, + "content": "its semantics.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 214, + 505, + 249 + ] + }, + { + "type": "text", + "bbox": [ + 168, + 252, + 242, + 297 + ], + "lines": [ + { + "bbox": [ + 167, + 251, + 226, + 264 + ], + "spans": [ + { + "bbox": [ + 167, + 251, + 226, + 264 + ], + "score": 1.0, + "content": "3 X 7 Y head", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 167, + 262, + 226, + 275 + ], + "spans": [ + { + "bbox": [ + 167, + 262, + 226, + 275 + ], + "score": 1.0, + "content": "7 4 X Y head", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 167, + 273, + 221, + 286 + ], + "spans": [ + { + "bbox": [ + 167, + 273, + 221, + 286 + ], + "score": 1.0, + "content": "X 3 1 X tail", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 165, + 285, + 243, + 298 + ], + "spans": [ + { + "bbox": [ + 165, + 285, + 243, + 298 + ], + "score": 1.0, + "content": "X X 5 Y duplicate", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 165, + 251, + 243, + 298 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 351, + 250, + 465, + 297 + ], + "lines": [ + { + "bbox": [ + 351, + 250, + 465, + 297 + ], + "spans": [ + { + "bbox": [ + 351, + 250, + 465, + 297 + ], + "score": 0.27, + "content": "\\begin{array} { r l } { X \\mathrm { ~ 0 ~ ~ 0 ~ ~ } } & { \\mathrm { ~ 0 ~ ~ Y ~ 0 ~ ~ 6 ~ 4 ~ } Y } \\\\ { \\mathrm { ~ 0 ~ 1 ~ 0 ~ } } & { \\mathrm { ~ 1 ~ } X \\mathrm { ~ 8 ~ } } \\\\ { \\mathrm { ~ 0 ~ 0 ~ } Y } & { \\mathrm { ~ 0 ~ 2 ~ 0 ~ } } \\end{array}", + "type": "interline_equation", + "image_path": "af9a832b2966a56b80d8d6d927634fd72447a59e0433605421d6b017bdc935e8.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 351, + 250, + 465, + 273.5 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 351, + 273.5, + 465, + 297.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 302, + 493, + 314 + ], + "lines": [ + { + "bbox": [ + 105, + 300, + 494, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 494, + 314 + ], + "score": 1.0, + "content": "(a) Invariants with extra varialbes learned with UMLP. (b) Mismatching invariants learned with UCNN.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 300, + 494, + 314 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 322, + 506, + 366 + ], + "lines": [ + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "score": 1.0, + "content": "Figure 9: Invariants learned that do not match the data generating distribution from UMLP and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 333, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 160, + 345 + ], + "score": 1.0, + "content": "UCNN using", + "type": "text" + }, + { + "bbox": [ + 161, + 333, + 193, + 344 + ], + "score": 0.85, + "content": "\\leq 1 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 333, + 506, + 345 + ], + "score": 1.0, + "content": "examples to train. In these instances the unification still bind to the the correct", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 344, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 506, + 357 + ], + "score": 1.0, + "content": "symbols in order to predict the desired answer; quantitatively we get the same results. Variable", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 354, + 263, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 263, + 367 + ], + "score": 1.0, + "content": "default symbols are omitted for clarity.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 322, + 506, + 367 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 470, + 236, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 469, + 236, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 236, + 480 + ], + "score": 1.0, + "content": "(a) bAbI task 2. When a variable", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 478, + 237, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 237, + 491 + ], + "score": 1.0, + "content": "is unused in the next iteration, e.g.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 488, + 236, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 236, + 500 + ], + "score": 1.0, + "content": "Z:football, it unifies with random", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 498, + 223, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 223, + 510 + ], + "score": 1.0, + "content": "tokens often biased by position.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 469, + 237, + 510 + ] + }, + { + "type": "image", + "bbox": [ + 117, + 374, + 496, + 461 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 374, + 496, + 461 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 374, + 496, + 461 + ], + "spans": [ + { + "bbox": [ + 117, + 374, + 496, + 461 + ], + "score": 0.5, + "type": "image", + "image_path": "9073cfbc8cdd1d7408ab2bf49ed3c03468c3850a67e1c59f65e66ccdfccc89e1.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 117, + 374, + 496, + 403.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 117, + 403.0, + 496, + 432.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 117, + 432.0, + 496, + 461.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 374, + 470, + 504, + 500 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 375, + 469, + 504, + 480 + ], + "spans": [ + { + "bbox": [ + 375, + 469, + 504, + 480 + ], + "score": 1.0, + "content": "(c) Logical reasoning task 3. Arity", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 375, + 478, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 375, + 478, + 505, + 490 + ], + "score": 1.0, + "content": "1 atom forced to bind with arity 2", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 375, + 489, + 493, + 501 + ], + "spans": [ + { + "bbox": [ + 375, + 489, + 493, + 501 + ], + "score": 1.0, + "content": "creates a one-to-many mapping.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + } + ], + "index": 28.0 + }, + { + "type": "text", + "bbox": [ + 240, + 470, + 370, + 500 + ], + "lines": [ + { + "bbox": [ + 241, + 470, + 370, + 480 + ], + "spans": [ + { + "bbox": [ + 241, + 470, + 370, + 480 + ], + "score": 1.0, + "content": "(b) Logical reasoning task 1. A", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 240, + 480, + 370, + 489 + ], + "spans": [ + { + "bbox": [ + 240, + 480, + 370, + 489 + ], + "score": 1.0, + "content": "one-to-one alignment is created be-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 240, + 489, + 356, + 500 + ], + "spans": [ + { + "bbox": [ + 240, + 489, + 356, + 500 + ], + "score": 1.0, + "content": "tween predicates and constants.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 240, + 470, + 370, + 500 + ] + }, + { + "type": "text", + "bbox": [ + 116, + 519, + 493, + 531 + ], + "lines": [ + { + "bbox": [ + 116, + 518, + 494, + 533 + ], + "spans": [ + { + "bbox": [ + 116, + 518, + 494, + 533 + ], + "score": 1.0, + "content": "Figure 10: Further attention maps for equation 3, darker cells indicate higher attention values.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35, + "bbox_fs": [ + 116, + 518, + 494, + 533 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/r1xwA34KDB/r1xwA34KDB_model.json b/parse/train/r1xwA34KDB/r1xwA34KDB_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9403e1b9ee6ac19dbc465c46f43fe92278a0eb70 --- /dev/null +++ b/parse/train/r1xwA34KDB/r1xwA34KDB_model.json @@ -0,0 +1,22940 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1025, + 1403, + 1025, + 1403, + 1362, + 298, + 1362 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 398, + 540, + 1302, + 540, + 1302, + 876, + 398, + 876 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 300, + 1377, + 989, + 1377, + 989, + 1772, + 300, + 1772 + ], + "score": 0.982 + }, + { + "category_id": 0, + "poly": [ + 295, + 218, + 1397, + 218, + 1397, + 269, + 295, + 269 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 299, + 1814, + 1404, + 1814, + 1404, + 2034, + 299, + 2034 + ], + "score": 0.908 + }, + { + "category_id": 1, + "poly": [ + 313, + 322, + 680, + 322, + 680, + 383, + 313, + 383 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 302, + 950, + 573, + 950, + 573, + 985, + 302, + 985 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 773, + 464, + 927, + 464, + 927, + 498, + 773, + 498 + ], + "score": 0.882 + }, + { + "category_id": 6, + "poly": [ + 1011, + 1607, + 1403, + 1607, + 1403, + 1791, + 1011, + 1791 + ], + "score": 0.88 + }, + { + "category_id": 5, + "poly": [ + 1039, + 1396, + 1369, + 1396, + 1369, + 1580, + 1039, + 1580 + ], + "score": 0.78, + "html": "
X:bernhard is a Y:frog Z:lily is a Y:frog
Z:lily is A:green
what colour is X:bernhard
A:green
" + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 857, + 2088, + 857, + 2112, + 842, + 2112 + ], + "score": 0.656 + }, + { + "category_id": 1, + "poly": [ + 300, + 1790, + 990, + 1790, + 990, + 1820, + 300, + 1820 + ], + "score": 0.392 + }, + { + "category_id": 1, + "poly": [ + 1039, + 1396, + 1369, + 1396, + 1369, + 1580, + 1039, + 1580 + ], + "score": 0.103 + }, + { + "category_id": 0, + "poly": [ + 300, + 1790, + 990, + 1790, + 990, + 1820, + 300, + 1820 + ], + "score": 0.098 + }, + { + "category_id": 13, + "poly": [ + 1089, + 1669, + 1113, + 1669, + 1113, + 1697, + 1089, + 1697 + ], + "score": 0.54, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 930, + 1209, + 968, + 1209, + 968, + 1237, + 930, + 1237 + ], + "score": 0.36, + "latex": "^ { 6 6 } \\mathrm { X }" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 218.0, + 1406.0, + 218.0, + 1406.0, + 277.0, + 291.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 947.0, + 579.0, + 947.0, + 579.0, + 994.0, + 294.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 463.0, + 932.0, + 463.0, + 932.0, + 501.0, + 770.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1607.0, + 1404.0, + 1607.0, + 1404.0, + 1639.0, + 1011.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 1636.0, + 1404.0, + 1636.0, + 1404.0, + 1670.0, + 1009.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 1667.0, + 1088.0, + 1667.0, + 1088.0, + 1700.0, + 1009.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 1667.0, + 1404.0, + 1667.0, + 1404.0, + 1700.0, + 1114.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1700.0, + 1403.0, + 1700.0, + 1403.0, + 1731.0, + 1011.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 1730.0, + 1404.0, + 1730.0, + 1404.0, + 1761.0, + 1010.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 1758.0, + 1392.0, + 1758.0, + 1392.0, + 1795.0, + 1009.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 858.0, + 2088.0, + 858.0, + 2116.0, + 841.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1789.0, + 991.0, + 1789.0, + 991.0, + 1824.0, + 295.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1025.0, + 1404.0, + 1025.0, + 1404.0, + 1062.0, + 293.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1056.0, + 1403.0, + 1056.0, + 1403.0, + 1090.0, + 294.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1087.0, + 1405.0, + 1087.0, + 1405.0, + 1121.0, + 294.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1117.0, + 1406.0, + 1117.0, + 1406.0, + 1153.0, + 291.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1146.0, + 1406.0, + 1146.0, + 1406.0, + 1182.0, + 292.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1177.0, + 1408.0, + 1177.0, + 1408.0, + 1216.0, + 292.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1207.0, + 929.0, + 1207.0, + 929.0, + 1246.0, + 293.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 1207.0, + 1407.0, + 1207.0, + 1407.0, + 1246.0, + 969.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1238.0, + 1405.0, + 1238.0, + 1405.0, + 1275.0, + 292.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1267.0, + 1405.0, + 1267.0, + 1405.0, + 1305.0, + 293.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1299.0, + 1405.0, + 1299.0, + 1405.0, + 1334.0, + 293.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1329.0, + 745.0, + 1329.0, + 745.0, + 1367.0, + 292.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 539.0, + 1305.0, + 539.0, + 1305.0, + 577.0, + 393.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 570.0, + 1306.0, + 570.0, + 1306.0, + 607.0, + 393.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 602.0, + 1306.0, + 602.0, + 1306.0, + 638.0, + 393.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 631.0, + 1305.0, + 631.0, + 1305.0, + 666.0, + 394.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 662.0, + 1306.0, + 662.0, + 1306.0, + 698.0, + 393.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 693.0, + 1305.0, + 693.0, + 1305.0, + 727.0, + 393.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 723.0, + 1306.0, + 723.0, + 1306.0, + 758.0, + 393.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 754.0, + 1305.0, + 754.0, + 1305.0, + 787.0, + 394.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 782.0, + 1304.0, + 782.0, + 1304.0, + 818.0, + 394.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 814.0, + 1305.0, + 814.0, + 1305.0, + 850.0, + 395.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 845.0, + 1221.0, + 845.0, + 1221.0, + 879.0, + 393.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1376.0, + 991.0, + 1376.0, + 991.0, + 1411.0, + 296.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1405.0, + 992.0, + 1405.0, + 992.0, + 1446.0, + 293.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1436.0, + 992.0, + 1436.0, + 992.0, + 1473.0, + 294.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1468.0, + 993.0, + 1468.0, + 993.0, + 1503.0, + 294.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1499.0, + 993.0, + 1499.0, + 993.0, + 1534.0, + 293.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1530.0, + 991.0, + 1530.0, + 991.0, + 1562.0, + 295.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1560.0, + 993.0, + 1560.0, + 993.0, + 1595.0, + 294.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1591.0, + 991.0, + 1591.0, + 991.0, + 1623.0, + 294.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1622.0, + 992.0, + 1622.0, + 992.0, + 1653.0, + 294.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1652.0, + 993.0, + 1652.0, + 993.0, + 1687.0, + 294.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1682.0, + 992.0, + 1682.0, + 992.0, + 1714.0, + 295.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1713.0, + 993.0, + 1713.0, + 993.0, + 1748.0, + 294.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1739.0, + 795.0, + 1739.0, + 795.0, + 1778.0, + 296.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1818.0, + 1404.0, + 1818.0, + 1404.0, + 1854.0, + 294.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1849.0, + 1404.0, + 1849.0, + 1404.0, + 1883.0, + 293.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1883.0, + 1406.0, + 1883.0, + 1406.0, + 1914.0, + 295.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1913.0, + 1405.0, + 1913.0, + 1405.0, + 1944.0, + 294.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1976.0, + 293.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2000.0, + 1404.0, + 2000.0, + 1404.0, + 2038.0, + 293.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 322.0, + 560.0, + 322.0, + 560.0, + 355.0, + 315.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 352.0, + 682.0, + 352.0, + 682.0, + 385.0, + 311.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1789.0, + 991.0, + 1789.0, + 991.0, + 1824.0, + 295.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 1395.0, + 1317.0, + 1395.0, + 1317.0, + 1434.0, + 1058.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 1431.0, + 1256.0, + 1431.0, + 1256.0, + 1473.0, + 1056.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 1468.0, + 1252.0, + 1468.0, + 1252.0, + 1508.0, + 1055.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 1510.0, + 1357.0, + 1510.0, + 1357.0, + 1541.0, + 1059.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 1544.0, + 1160.0, + 1544.0, + 1160.0, + 1587.0, + 1052.0, + 1587.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 307, + 1404, + 307, + 1404, + 523, + 298, + 523 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 634, + 1404, + 634, + 1404, + 757, + 298, + 757 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1046, + 1403, + 1046, + 1403, + 1201, + 298, + 1201 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1212, + 1404, + 1212, + 1404, + 1337, + 297, + 1337 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 296, + 767, + 1404, + 767, + 1404, + 892, + 296, + 892 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1582, + 1403, + 1582, + 1403, + 1706, + 297, + 1706 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1479, + 1404, + 1479, + 1404, + 1574, + 298, + 1574 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1863, + 1403, + 1863, + 1403, + 1957, + 298, + 1957 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1401, + 229, + 1401, + 293, + 294, + 293 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 566, + 903, + 1132, + 903, + 1132, + 976, + 566, + 976 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 297, + 1393, + 1398, + 1393, + 1398, + 1458, + 297, + 1458 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 619, + 1346, + 1083, + 1346, + 1083, + 1385, + 619, + 1385 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 297, + 1783, + 1400, + 1783, + 1400, + 1850, + 297, + 1850 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 298, + 1971, + 1399, + 1971, + 1399, + 2035, + 298, + 2035 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 300, + 990, + 1096, + 990, + 1096, + 1023, + 300, + 1023 + ], + "score": 0.936 + }, + { + "category_id": 0, + "poly": [ + 300, + 565, + 625, + 565, + 625, + 602, + 300, + 602 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.915 + }, + { + "category_id": 8, + "poly": [ + 499, + 1717, + 1201, + 1717, + 1201, + 1757, + 499, + 1757 + ], + "score": 0.912 + }, + { + "category_id": 9, + "poly": [ + 1366, + 914, + 1400, + 914, + 1400, + 944, + 1366, + 944 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1351, + 1400, + 1351, + 1400, + 1380, + 1366, + 1380 + ], + "score": 0.879 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1723, + 1400, + 1723, + 1400, + 1752, + 1366, + 1752 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.724 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.198 + }, + { + "category_id": 13, + "poly": [ + 511, + 1426, + 609, + 1426, + 609, + 1459, + 511, + 1459 + ], + "score": 0.93, + "latex": "\\phi _ { V } ( \\pmb { X } ; s )" + }, + { + "category_id": 14, + "poly": [ + 566, + 903, + 1133, + 903, + 1133, + 978, + 566, + 978 + ], + "score": 0.93, + "latex": "\\phi _ { V } ( X { : } s _ { d } ) = \\mathbb { E } _ { \\mathbf { x } \\sim P } [ \\phi ( x ) ] = \\sum _ { x \\in \\mathbb { S } } P ( \\mathbf { x } = x | s _ { d } ) \\phi ( x )" + }, + { + "category_id": 13, + "poly": [ + 372, + 989, + 506, + 989, + 506, + 1023, + 372, + 1023 + ], + "score": 0.93, + "latex": "\\phi : \\mathbb { S } \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 983, + 1783, + 1189, + 1783, + 1189, + 1821, + 983, + 1821 + ], + "score": 0.92, + "latex": "\\phi _ { U } ( K ) \\in \\mathbb { R } ^ { | K | \\times d }" + }, + { + "category_id": 13, + "poly": [ + 864, + 1394, + 918, + 1394, + 918, + 1428, + 864, + 1428 + ], + "score": 0.92, + "latex": "\\phi ( s )" + }, + { + "category_id": 13, + "poly": [ + 1196, + 767, + 1326, + 767, + 1326, + 803, + 1196, + 803 + ], + "score": 0.92, + "latex": "\\pmb { X } \\triangleq \\left( \\mathbf { x } , s _ { d } \\right)" + }, + { + "category_id": 13, + "poly": [ + 803, + 1048, + 916, + 1048, + 916, + 1082, + 803, + 1082 + ], + "score": 0.92, + "latex": "\\phi _ { V } ( { \\pmb X } ; s _ { d } )" + }, + { + "category_id": 13, + "poly": [ + 373, + 1784, + 533, + 1784, + 533, + 1819, + 373, + 1819 + ], + "score": 0.92, + "latex": "\\phi _ { U } : \\mathbb { S } \\to \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 716, + 831, + 828, + 831, + 828, + 864, + 716, + 864 + ], + "score": 0.92, + "latex": "\\phi _ { V } ( { \\pmb X } ; s _ { d } )" + }, + { + "category_id": 13, + "poly": [ + 994, + 1974, + 1092, + 1974, + 1092, + 2005, + 994, + 2005 + ], + "score": 0.91, + "latex": "\\phi , \\psi , \\phi _ { U }" + }, + { + "category_id": 13, + "poly": [ + 372, + 1245, + 448, + 1245, + 448, + 1274, + 372, + 1274 + ], + "score": 0.91, + "latex": "G \\in { \\mathcal { G } }" + }, + { + "category_id": 14, + "poly": [ + 616, + 1346, + 1083, + 1346, + 1083, + 1384, + 616, + 1384 + ], + "score": 0.91, + "latex": "\\phi _ { I } ( s ) = ( 1 - \\psi ( s ) ) \\phi ( s ) + \\psi ( s ) \\phi _ { V } ( X { : } s )" + }, + { + "category_id": 13, + "poly": [ + 1054, + 1582, + 1148, + 1582, + 1148, + 1612, + 1054, + 1612 + ], + "score": 0.9, + "latex": "K \\in \\mathcal G" + }, + { + "category_id": 13, + "poly": [ + 911, + 1613, + 1091, + 1613, + 1091, + 1643, + 911, + 1643 + ], + "score": 0.9, + "latex": "g : \\mathcal { T } \\times \\mathcal { G } \\to \\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 298, + 801, + 376, + 801, + 376, + 831, + 298, + 831 + ], + "score": 0.89, + "latex": "s _ { d } \\in \\mathbb { S }" + }, + { + "category_id": 13, + "poly": [ + 1271, + 1211, + 1400, + 1211, + 1400, + 1246, + 1271, + 1246 + ], + "score": 0.89, + "latex": "I \\triangleq ( G , \\psi )" + }, + { + "category_id": 14, + "poly": [ + 499, + 1715, + 1201, + 1715, + 1201, + 1757, + 499, + 1757 + ], + "score": 0.88, + "latex": "P ( { \\bf x } = x | s _ { d } ) = \\mathrm { s o f t m a x } ( \\phi _ { U } ( s _ { d } ) \\phi _ { U } ( K ) ^ { T } ) , \\ : K = \\mathrm { s u p p o r t \\ o f \\ x }" + }, + { + "category_id": 13, + "poly": [ + 1306, + 1246, + 1404, + 1246, + 1404, + 1275, + 1306, + 1275 + ], + "score": 0.87, + "latex": "\\psi : \\mathbb { S } \\to \\quad" + }, + { + "category_id": 13, + "poly": [ + 298, + 1930, + 335, + 1930, + 335, + 1959, + 298, + 1959 + ], + "score": 0.85, + "latex": "\\phi _ { U }" + }, + { + "category_id": 13, + "poly": [ + 979, + 1675, + 1036, + 1675, + 1036, + 1704, + 979, + 1704 + ], + "score": 0.84, + "latex": "X { : } s _ { d }" + }, + { + "category_id": 13, + "poly": [ + 786, + 866, + 813, + 866, + 813, + 892, + 786, + 892 + ], + "score": 0.84, + "latex": "s _ { d }" + }, + { + "category_id": 13, + "poly": [ + 454, + 1050, + 476, + 1050, + 476, + 1081, + 454, + 1081 + ], + "score": 0.83, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 936, + 1215, + 958, + 1215, + 958, + 1243, + 936, + 1243 + ], + "score": 0.83, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 803, + 1644, + 827, + 1644, + 827, + 1670, + 803, + 1670 + ], + "score": 0.83, + "latex": "P" + }, + { + "category_id": 13, + "poly": [ + 578, + 1822, + 595, + 1822, + 595, + 1850, + 578, + 1850 + ], + "score": 0.82, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 916, + 1141, + 940, + 1141, + 940, + 1167, + 916, + 1167 + ], + "score": 0.82, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 963, + 771, + 987, + 771, + 987, + 797, + 963, + 797 + ], + "score": 0.81, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 547, + 1866, + 584, + 1866, + 584, + 1896, + 547, + 1896 + ], + "score": 0.81, + "latex": "\\phi _ { U }" + }, + { + "category_id": 13, + "poly": [ + 555, + 993, + 572, + 993, + 572, + 1019, + 555, + 1019 + ], + "score": 0.81, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 512, + 1819, + 540, + 1819, + 540, + 1845, + 512, + 1845 + ], + "score": 0.81, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 617, + 1614, + 645, + 1614, + 645, + 1640, + 617, + 1640 + ], + "score": 0.8, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 466, + 1866, + 486, + 1866, + 486, + 1896, + 466, + 1896 + ], + "score": 0.8, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 434, + 1615, + 453, + 1615, + 453, + 1640, + 434, + 1640 + ], + "score": 0.8, + "latex": "I" + }, + { + "category_id": 13, + "poly": [ + 840, + 1305, + 907, + 1305, + 907, + 1338, + 840, + 1338 + ], + "score": 0.77, + "latex": "\\phi _ { I } ( s )" + }, + { + "category_id": 13, + "poly": [ + 707, + 1141, + 730, + 1141, + 730, + 1167, + 707, + 1167 + ], + "score": 0.77, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 837, + 1584, + 856, + 1584, + 856, + 1610, + 837, + 1610 + ], + "score": 0.77, + "latex": "I" + }, + { + "category_id": 13, + "poly": [ + 775, + 1896, + 794, + 1896, + 794, + 1927, + 775, + 1927 + ], + "score": 0.75, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 655, + 1310, + 671, + 1310, + 671, + 1332, + 655, + 1332 + ], + "score": 0.75, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1073, + 999, + 1088, + 999, + 1088, + 1019, + 1073, + 1019 + ], + "score": 0.73, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 826, + 807, + 844, + 807, + 844, + 828, + 826, + 828 + ], + "score": 0.71, + "latex": "\\mathbf { X }" + }, + { + "category_id": 13, + "poly": [ + 484, + 865, + 502, + 865, + 502, + 888, + 484, + 888 + ], + "score": 0.66, + "latex": "\\mathbf { X }" + }, + { + "category_id": 13, + "poly": [ + 471, + 1140, + 561, + 1140, + 561, + 1170, + 471, + 1170 + ], + "score": 0.66, + "latex": "X , Y , A" + }, + { + "category_id": 13, + "poly": [ + 805, + 1306, + 830, + 1306, + 830, + 1333, + 805, + 1333 + ], + "score": 0.65, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 823, + 771, + 841, + 771, + 841, + 797, + 823, + 797 + ], + "score": 0.64, + "latex": "\\mathbb { S }" + }, + { + "category_id": 13, + "poly": [ + 1014, + 1215, + 1034, + 1215, + 1034, + 1241, + 1014, + 1241 + ], + "score": 0.5, + "latex": "\\mathbb { S }" + }, + { + "category_id": 13, + "poly": [ + 323, + 833, + 342, + 833, + 342, + 858, + 323, + 858 + ], + "score": 0.46, + "latex": "\\mathbb { S }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1274, + 355, + 1274, + 355, + 1308, + 298, + 1308 + ], + "score": 0.46, + "latex": "[ 0 , 1 ]" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 561.0, + 629.0, + 561.0, + 629.0, + 608.0, + 292.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 304.0, + 1405.0, + 304.0, + 1405.0, + 347.0, + 292.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 337.0, + 1405.0, + 337.0, + 1405.0, + 374.0, + 293.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 366.0, + 1405.0, + 366.0, + 1405.0, + 408.0, + 292.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 400.0, + 1405.0, + 400.0, + 1405.0, + 435.0, + 296.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 427.0, + 1405.0, + 427.0, + 1405.0, + 468.0, + 292.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 461.0, + 1406.0, + 461.0, + 1406.0, + 495.0, + 294.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 493.0, + 1224.0, + 493.0, + 1224.0, + 524.0, + 301.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 636.0, + 1405.0, + 636.0, + 1405.0, + 669.0, + 294.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 666.0, + 1404.0, + 666.0, + 1404.0, + 699.0, + 294.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 697.0, + 1402.0, + 697.0, + 1402.0, + 729.0, + 296.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 728.0, + 719.0, + 728.0, + 719.0, + 761.0, + 294.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1046.0, + 453.0, + 1046.0, + 453.0, + 1084.0, + 294.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 1046.0, + 802.0, + 1046.0, + 802.0, + 1084.0, + 477.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1046.0, + 1405.0, + 1046.0, + 1405.0, + 1084.0, + 917.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1078.0, + 1405.0, + 1078.0, + 1405.0, + 1112.0, + 296.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1108.0, + 1407.0, + 1108.0, + 1407.0, + 1143.0, + 293.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1138.0, + 470.0, + 1138.0, + 470.0, + 1175.0, + 294.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 1138.0, + 706.0, + 1138.0, + 706.0, + 1175.0, + 562.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 1138.0, + 915.0, + 1138.0, + 915.0, + 1175.0, + 731.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1138.0, + 1405.0, + 1138.0, + 1405.0, + 1175.0, + 941.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1170.0, + 753.0, + 1170.0, + 753.0, + 1203.0, + 296.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1206.0, + 935.0, + 1206.0, + 935.0, + 1251.0, + 291.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 1206.0, + 1013.0, + 1206.0, + 1013.0, + 1251.0, + 959.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 1206.0, + 1270.0, + 1206.0, + 1270.0, + 1251.0, + 1035.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1206.0, + 1405.0, + 1206.0, + 1405.0, + 1251.0, + 1401.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1244.0, + 371.0, + 1244.0, + 371.0, + 1277.0, + 295.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1244.0, + 1305.0, + 1244.0, + 1305.0, + 1277.0, + 449.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1273.0, + 1405.0, + 1273.0, + 1405.0, + 1310.0, + 356.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1304.0, + 654.0, + 1304.0, + 654.0, + 1340.0, + 294.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 1304.0, + 804.0, + 1304.0, + 804.0, + 1340.0, + 672.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 1304.0, + 839.0, + 1304.0, + 839.0, + 1340.0, + 831.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 1304.0, + 946.0, + 1304.0, + 946.0, + 1340.0, + 908.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 767.0, + 822.0, + 767.0, + 822.0, + 804.0, + 294.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 767.0, + 962.0, + 767.0, + 962.0, + 804.0, + 842.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 767.0, + 1195.0, + 767.0, + 1195.0, + 804.0, + 988.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1327.0, + 767.0, + 1404.0, + 767.0, + 1404.0, + 804.0, + 1327.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 799.0, + 297.0, + 799.0, + 297.0, + 835.0, + 293.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 799.0, + 825.0, + 799.0, + 825.0, + 835.0, + 377.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 799.0, + 1405.0, + 799.0, + 1405.0, + 835.0, + 845.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 827.0, + 322.0, + 827.0, + 322.0, + 868.0, + 291.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 827.0, + 715.0, + 827.0, + 715.0, + 868.0, + 343.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 827.0, + 1404.0, + 827.0, + 1404.0, + 868.0, + 829.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 859.0, + 483.0, + 859.0, + 483.0, + 898.0, + 293.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 859.0, + 785.0, + 859.0, + 785.0, + 898.0, + 503.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 859.0, + 826.0, + 859.0, + 826.0, + 898.0, + 814.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1582.0, + 836.0, + 1582.0, + 836.0, + 1615.0, + 295.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 1582.0, + 1053.0, + 1582.0, + 1053.0, + 1615.0, + 857.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 1582.0, + 1403.0, + 1582.0, + 1403.0, + 1615.0, + 1149.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1612.0, + 433.0, + 1612.0, + 433.0, + 1645.0, + 296.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 1612.0, + 616.0, + 1612.0, + 616.0, + 1645.0, + 454.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 1612.0, + 910.0, + 1612.0, + 910.0, + 1645.0, + 646.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 1612.0, + 1405.0, + 1612.0, + 1405.0, + 1645.0, + 1092.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1643.0, + 802.0, + 1643.0, + 802.0, + 1677.0, + 292.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 1643.0, + 1404.0, + 1643.0, + 1404.0, + 1677.0, + 828.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1673.0, + 978.0, + 1673.0, + 978.0, + 1708.0, + 293.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 1673.0, + 1075.0, + 1673.0, + 1075.0, + 1708.0, + 1037.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1479.0, + 1403.0, + 1479.0, + 1403.0, + 1516.0, + 295.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1513.0, + 1404.0, + 1513.0, + 1404.0, + 1544.0, + 297.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1544.0, + 1340.0, + 1544.0, + 1340.0, + 1574.0, + 296.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1863.0, + 465.0, + 1863.0, + 465.0, + 1900.0, + 294.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 1863.0, + 546.0, + 1863.0, + 546.0, + 1900.0, + 487.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 1863.0, + 1404.0, + 1863.0, + 1404.0, + 1900.0, + 585.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1893.0, + 774.0, + 1893.0, + 774.0, + 1929.0, + 293.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 1893.0, + 1403.0, + 1893.0, + 1403.0, + 1929.0, + 795.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1925.0, + 297.0, + 1925.0, + 297.0, + 1962.0, + 294.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 1925.0, + 721.0, + 1925.0, + 721.0, + 1962.0, + 336.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 260.0, + 1205.0, + 260.0, + 1205.0, + 297.0, + 293.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1391.0, + 863.0, + 1391.0, + 863.0, + 1431.0, + 293.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1391.0, + 1405.0, + 1391.0, + 1405.0, + 1431.0, + 919.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1422.0, + 510.0, + 1422.0, + 510.0, + 1462.0, + 293.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1422.0, + 621.0, + 1422.0, + 621.0, + 1462.0, + 610.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1776.0, + 372.0, + 1776.0, + 372.0, + 1827.0, + 292.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 1776.0, + 982.0, + 1776.0, + 982.0, + 1827.0, + 534.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 1776.0, + 1407.0, + 1776.0, + 1407.0, + 1827.0, + 1190.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1815.0, + 511.0, + 1815.0, + 511.0, + 1852.0, + 295.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 1815.0, + 577.0, + 1815.0, + 577.0, + 1852.0, + 541.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1815.0, + 1167.0, + 1815.0, + 1167.0, + 1852.0, + 596.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1970.0, + 993.0, + 1970.0, + 993.0, + 2006.0, + 296.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 1970.0, + 1401.0, + 1970.0, + 1401.0, + 2006.0, + 1093.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2000.0, + 1403.0, + 2000.0, + 1403.0, + 2039.0, + 293.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 984.0, + 371.0, + 984.0, + 371.0, + 1030.0, + 293.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 984.0, + 554.0, + 984.0, + 554.0, + 1030.0, + 507.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 984.0, + 1072.0, + 984.0, + 1072.0, + 1030.0, + 573.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 984.0, + 1102.0, + 984.0, + 1102.0, + 1030.0, + 1089.0, + 1030.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 538, + 1405, + 538, + 1405, + 914, + 297, + 914 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 296, + 926, + 1404, + 926, + 1404, + 1204, + 296, + 1204 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 1649, + 1404, + 1649, + 1404, + 1897, + 297, + 1897 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1910, + 1403, + 1910, + 1403, + 2034, + 299, + 2034 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 378, + 1223, + 1321, + 1223, + 1321, + 1506, + 378, + 1506 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 293, + 228, + 1399, + 228, + 1399, + 293, + 293, + 293 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 293, + 361, + 1401, + 361, + 1401, + 425, + 293, + 425 + ], + "score": 0.951 + }, + { + "category_id": 4, + "poly": [ + 297, + 1542, + 1405, + 1542, + 1405, + 1636, + 297, + 1636 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 706, + 308, + 994, + 308, + 994, + 347, + 706, + 347 + ], + "score": 0.942 + }, + { + "category_id": 2, + "poly": [ + 297, + 74, + 858, + 74, + 858, + 105, + 297, + 105 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 302, + 469, + 714, + 469, + 714, + 505, + 302, + 505 + ], + "score": 0.909 + }, + { + "category_id": 9, + "poly": [ + 1366, + 313, + 1400, + 313, + 1400, + 343, + 1366, + 343 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.626 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.472 + }, + { + "category_id": 13, + "poly": [ + 595, + 1109, + 721, + 1109, + 721, + 1143, + 595, + 1143 + ], + "score": 0.93, + "latex": "\\Phi ( k ) \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 647, + 662, + 834, + 662, + 834, + 698, + 647, + 698 + ], + "score": 0.93, + "latex": "\\phi ( s ) = O [ s ] ^ { T } E" + }, + { + "category_id": 14, + "poly": [ + 704, + 308, + 996, + 308, + 996, + 347, + 704, + 347 + ], + "score": 0.93, + "latex": "f \\circ g ( I , K ) = f ( K ) = a" + }, + { + "category_id": 13, + "poly": [ + 1254, + 1109, + 1401, + 1109, + 1401, + 1141, + 1254, + 1141 + ], + "score": 0.92, + "latex": "\\dot { W _ { U } } \\in \\mathbb { R } ^ { d \\times d }" + }, + { + "category_id": 13, + "poly": [ + 298, + 728, + 477, + 728, + 477, + 762, + 298, + 762 + ], + "score": 0.92, + "latex": "\\psi _ { w } ( s ) = \\sigma ( w _ { s } )" + }, + { + "category_id": 13, + "poly": [ + 1008, + 850, + 1082, + 850, + 1082, + 883, + 1008, + 883 + ], + "score": 0.92, + "latex": "\\psi _ { w } ( s )" + }, + { + "category_id": 13, + "poly": [ + 896, + 661, + 1037, + 661, + 1037, + 693, + 896, + 693 + ], + "score": 0.92, + "latex": "E \\in \\mathbb { R } ^ { | \\mathbb { S } | \\times d }" + }, + { + "category_id": 13, + "poly": [ + 555, + 725, + 660, + 725, + 660, + 757, + 555, + 757 + ], + "score": 0.92, + "latex": "\\boldsymbol { w } \\in \\mathbb { R } ^ { | \\mathbb { S } | }" + }, + { + "category_id": 13, + "poly": [ + 1033, + 571, + 1101, + 571, + 1101, + 603, + 1033, + 603 + ], + "score": 0.92, + "latex": "f \\circ g" + }, + { + "category_id": 13, + "poly": [ + 560, + 1048, + 656, + 1048, + 656, + 1078, + 560, + 1078 + ], + "score": 0.92, + "latex": "\\hat { h } \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1111, + 515, + 1111, + 515, + 1144, + 298, + 1144 + ], + "score": 0.92, + "latex": "\\phi _ { U } ( k ) = W _ { U } \\Phi ( k )" + }, + { + "category_id": 13, + "poly": [ + 1309, + 663, + 1400, + 663, + 1400, + 697, + 1309, + 697 + ], + "score": 0.92, + "latex": "\\mathcal { N } ( 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 960, + 394, + 1019, + 394, + 1019, + 425, + 960, + 425 + ], + "score": 0.92, + "latex": "f \\circ g" + }, + { + "category_id": 13, + "poly": [ + 1162, + 231, + 1248, + 231, + 1248, + 261, + 1162, + 261 + ], + "score": 0.91, + "latex": "K \\in { \\mathcal { G } }" + }, + { + "category_id": 13, + "poly": [ + 547, + 1944, + 608, + 1944, + 608, + 1975, + 547, + 1975 + ], + "score": 0.91, + "latex": "f \\circ g" + }, + { + "category_id": 13, + "poly": [ + 550, + 231, + 681, + 231, + 681, + 262, + 550, + 262 + ], + "score": 0.91, + "latex": "f : \\mathcal G \\to \\mathbb S" + }, + { + "category_id": 13, + "poly": [ + 1238, + 1081, + 1295, + 1081, + 1295, + 1112, + 1238, + 1112 + ], + "score": 0.91, + "latex": "\\phi ( k )" + }, + { + "category_id": 13, + "poly": [ + 832, + 231, + 907, + 231, + 907, + 259, + 832, + 259 + ], + "score": 0.91, + "latex": "I \\in \\mathcal { Z }" + }, + { + "category_id": 13, + "poly": [ + 914, + 1806, + 1038, + 1806, + 1038, + 1833, + 914, + 1833 + ], + "score": 0.91, + "latex": "w \\times h \\times d" + }, + { + "category_id": 13, + "poly": [ + 1252, + 1143, + 1307, + 1143, + 1307, + 1174, + 1252, + 1174 + ], + "score": 0.91, + "latex": "f \\circ g" + }, + { + "category_id": 13, + "poly": [ + 1295, + 1650, + 1403, + 1650, + 1403, + 1685, + 1295, + 1685 + ], + "score": 0.9, + "latex": "\\phi ( K ) \\ \\in" + }, + { + "category_id": 13, + "poly": [ + 347, + 694, + 400, + 694, + 400, + 727, + 347, + 727 + ], + "score": 0.9, + "latex": "O [ s ]" + }, + { + "category_id": 13, + "poly": [ + 297, + 1679, + 397, + 1679, + 397, + 1709, + 297, + 1709 + ], + "score": 0.9, + "latex": "\\mathbb { R } ^ { w \\times h \\times d }" + }, + { + "category_id": 13, + "poly": [ + 838, + 1773, + 1116, + 1773, + 1116, + 1806, + 838, + 1806 + ], + "score": 0.9, + "latex": "\\phi _ { U } ( k ) = c _ { 2 } ( \\mathrm { r e l u } ( c _ { 1 } ( k ) )" + }, + { + "category_id": 13, + "poly": [ + 1233, + 987, + 1393, + 987, + 1393, + 1021, + 1233, + 1021 + ], + "score": 0.89, + "latex": "\\phi ( \\pmb { k } ) \\in R ^ { l \\times d }" + }, + { + "category_id": 13, + "poly": [ + 538, + 1711, + 961, + 1711, + 961, + 1746, + 538, + 1746 + ], + "score": 0.89, + "latex": "f ( \\phi ( K ) ) = \\mathrm { s o f t m a x } ( ( W h + b ) E ^ { T } )" + }, + { + "category_id": 13, + "poly": [ + 570, + 263, + 637, + 263, + 637, + 290, + 570, + 290 + ], + "score": 0.89, + "latex": "a \\in \\mathbb { S }" + }, + { + "category_id": 13, + "poly": [ + 1273, + 1021, + 1403, + 1021, + 1403, + 1053, + 1273, + 1053 + ], + "score": 0.88, + "latex": "f ( \\phi ( \\pmb { k } ) ) =" + }, + { + "category_id": 13, + "poly": [ + 1196, + 1778, + 1261, + 1778, + 1261, + 1805, + 1196, + 1805 + ], + "score": 0.88, + "latex": "c _ { 1 } , c _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 396, + 1048, + 477, + 1048, + 477, + 1082, + 396, + 1082 + ], + "score": 0.87, + "latex": "( h E ^ { T } )" + }, + { + "category_id": 13, + "poly": [ + 391, + 1083, + 428, + 1083, + 428, + 1112, + 391, + 1112 + ], + "score": 0.86, + "latex": "\\phi _ { U }" + }, + { + "category_id": 13, + "poly": [ + 372, + 363, + 392, + 363, + 392, + 394, + 372, + 394 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 727, + 1775, + 748, + 1775, + 748, + 1805, + 727, + 1805 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 475, + 1944, + 495, + 1944, + 495, + 1975, + 475, + 1975 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 803, + 855, + 838, + 855, + 838, + 881, + 803, + 881 + ], + "score": 0.85, + "latex": "w _ { s }" + }, + { + "category_id": 13, + "poly": [ + 495, + 1744, + 560, + 1744, + 560, + 1774, + 495, + 1774 + ], + "score": 0.82, + "latex": "W , b" + }, + { + "category_id": 13, + "poly": [ + 1082, + 960, + 1105, + 960, + 1105, + 989, + 1082, + 989 + ], + "score": 0.81, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 1183, + 1112, + 1202, + 1112, + 1202, + 1138, + 1183, + 1138 + ], + "score": 0.8, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1032, + 1748, + 1050, + 1748, + 1050, + 1774, + 1032, + 1774 + ], + "score": 0.8, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 699, + 1683, + 719, + 1683, + 719, + 1709, + 699, + 1709 + ], + "score": 0.79, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 711, + 732, + 732, + 732, + 732, + 757, + 711, + 757 + ], + "score": 0.78, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 987, + 632, + 1006, + 632, + 1006, + 658, + 987, + 658 + ], + "score": 0.78, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1084, + 367, + 1101, + 367, + 1101, + 394, + 1084, + 394 + ], + "score": 0.77, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 1037, + 1713, + 1060, + 1713, + 1060, + 1741, + 1037, + 1741 + ], + "score": 0.77, + "latex": "^ { h }" + }, + { + "category_id": 13, + "poly": [ + 479, + 1687, + 504, + 1687, + 504, + 1709, + 479, + 1709 + ], + "score": 0.75, + "latex": "w" + }, + { + "category_id": 13, + "poly": [ + 772, + 1916, + 790, + 1916, + 790, + 1944, + 772, + 1944 + ], + "score": 0.74, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 578, + 1085, + 597, + 1085, + 597, + 1112, + 578, + 1112 + ], + "score": 0.72, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 298, + 400, + 315, + 400, + 315, + 425, + 298, + 425 + ], + "score": 0.71, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 750, + 1656, + 769, + 1656, + 769, + 1684, + 750, + 1684 + ], + "score": 0.69, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 744, + 933, + 762, + 933, + 762, + 961, + 744, + 961 + ], + "score": 0.68, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 521, + 991, + 534, + 991, + 534, + 1017, + 521, + 1017 + ], + "score": 0.6, + "latex": "l" + }, + { + "category_id": 13, + "poly": [ + 632, + 930, + 651, + 930, + 651, + 960, + 632, + 960 + ], + "score": 0.6, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 921, + 1921, + 938, + 1921, + 938, + 1944, + 921, + 1944 + ], + "score": 0.59, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 632, + 1652, + 651, + 1652, + 651, + 1684, + 632, + 1684 + ], + "score": 0.58, + "latex": "f" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1234.0, + 518.0, + 1234.0, + 518.0, + 1250.0, + 505.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1228.0, + 631.0, + 1228.0, + 631.0, + 1254.0, + 523.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 1228.0, + 909.0, + 1228.0, + 909.0, + 1263.0, + 784.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1254.0, + 459.0, + 1254.0, + 459.0, + 1279.0, + 392.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 1256.0, + 646.0, + 1256.0, + 646.0, + 1284.0, + 503.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 1273.0, + 461.0, + 1273.0, + 461.0, + 1301.0, + 382.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 1271.0, + 1157.0, + 1271.0, + 1157.0, + 1300.0, + 1018.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 1271.0, + 1283.0, + 1271.0, + 1283.0, + 1297.0, + 1254.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 1304.0, + 448.0, + 1304.0, + 448.0, + 1339.0, + 411.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 1301.0, + 706.0, + 1301.0, + 706.0, + 1340.0, + 503.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 1306.0, + 1152.0, + 1306.0, + 1152.0, + 1337.0, + 1023.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1248.0, + 1305.0, + 1282.0, + 1305.0, + 1282.0, + 1337.0, + 1248.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1336.0, + 449.0, + 1336.0, + 449.0, + 1384.0, + 406.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 1340.0, + 680.0, + 1340.0, + 680.0, + 1379.0, + 531.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 1341.0, + 1159.0, + 1341.0, + 1159.0, + 1376.0, + 1014.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 1333.0, + 1292.0, + 1333.0, + 1292.0, + 1419.0, + 1242.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1368.0, + 450.0, + 1368.0, + 450.0, + 1416.0, + 407.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 1375.0, + 680.0, + 1375.0, + 680.0, + 1416.0, + 532.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 1382.0, + 921.0, + 1382.0, + 921.0, + 1407.0, + 898.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 1381.0, + 1152.0, + 1381.0, + 1152.0, + 1412.0, + 1020.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1415.0, + 448.0, + 1415.0, + 448.0, + 1449.0, + 408.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 1419.0, + 721.0, + 1419.0, + 721.0, + 1446.0, + 488.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 1414.0, + 1175.0, + 1414.0, + 1175.0, + 1451.0, + 999.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 1413.0, + 1285.0, + 1413.0, + 1285.0, + 1448.0, + 1246.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 1460.0, + 542.0, + 1460.0, + 542.0, + 1489.0, + 468.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 1455.0, + 726.0, + 1455.0, + 726.0, + 1491.0, + 640.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1459.0, + 1025.0, + 1459.0, + 1025.0, + 1491.0, + 951.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 1454.0, + 1206.0, + 1454.0, + 1206.0, + 1484.0, + 1146.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 1448.0, + 1311.0, + 1448.0, + 1311.0, + 1473.0, + 1244.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.75, + 1310.0, + 910.75, + 1310.0, + 910.75, + 1331.0, + 904.75, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.25, + 1309.5, + 800.25, + 1309.5, + 800.25, + 1332.5, + 766.25, + 1332.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 1381.5, + 789.0, + 1381.5, + 789.0, + 1407.0, + 778.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 1420.5, + 914.0, + 1420.5, + 914.0, + 1443.5, + 907.0, + 1443.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1542.0, + 1402.0, + 1542.0, + 1402.0, + 1579.0, + 295.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1576.0, + 1402.0, + 1576.0, + 1402.0, + 1606.0, + 296.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1605.0, + 1373.0, + 1605.0, + 1373.0, + 1639.0, + 295.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 465.0, + 719.0, + 465.0, + 719.0, + 510.0, + 291.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 538.0, + 1405.0, + 538.0, + 1405.0, + 576.0, + 294.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 569.0, + 1032.0, + 569.0, + 1032.0, + 608.0, + 292.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 569.0, + 1405.0, + 569.0, + 1405.0, + 608.0, + 1102.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 601.0, + 1403.0, + 601.0, + 1403.0, + 636.0, + 295.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 629.0, + 986.0, + 629.0, + 986.0, + 667.0, + 292.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 629.0, + 1405.0, + 629.0, + 1405.0, + 667.0, + 1007.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 657.0, + 646.0, + 657.0, + 646.0, + 702.0, + 291.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 657.0, + 895.0, + 657.0, + 895.0, + 702.0, + 835.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 657.0, + 1308.0, + 657.0, + 1308.0, + 702.0, + 1038.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 657.0, + 1407.0, + 657.0, + 1407.0, + 702.0, + 1401.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 695.0, + 346.0, + 695.0, + 346.0, + 730.0, + 295.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 695.0, + 1406.0, + 695.0, + 1406.0, + 730.0, + 401.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 722.0, + 297.0, + 722.0, + 297.0, + 768.0, + 290.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 722.0, + 554.0, + 722.0, + 554.0, + 768.0, + 478.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 722.0, + 710.0, + 722.0, + 710.0, + 768.0, + 661.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 722.0, + 1407.0, + 722.0, + 1407.0, + 768.0, + 733.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 759.0, + 1405.0, + 759.0, + 1405.0, + 794.0, + 295.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 789.0, + 1405.0, + 789.0, + 1405.0, + 824.0, + 294.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 819.0, + 1407.0, + 819.0, + 1407.0, + 855.0, + 291.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 851.0, + 802.0, + 851.0, + 802.0, + 886.0, + 294.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 851.0, + 1007.0, + 851.0, + 1007.0, + 886.0, + 839.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 851.0, + 1406.0, + 851.0, + 1406.0, + 886.0, + 1083.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 882.0, + 1187.0, + 882.0, + 1187.0, + 917.0, + 294.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 927.0, + 631.0, + 927.0, + 631.0, + 964.0, + 295.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 927.0, + 743.0, + 927.0, + 743.0, + 964.0, + 652.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 927.0, + 1406.0, + 927.0, + 1406.0, + 964.0, + 763.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 959.0, + 1081.0, + 959.0, + 1081.0, + 993.0, + 293.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 959.0, + 1406.0, + 959.0, + 1406.0, + 993.0, + 1106.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 985.0, + 520.0, + 985.0, + 520.0, + 1024.0, + 290.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 985.0, + 1232.0, + 985.0, + 1232.0, + 1024.0, + 535.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 985.0, + 1407.0, + 985.0, + 1407.0, + 1024.0, + 1394.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1016.0, + 1272.0, + 1016.0, + 1272.0, + 1057.0, + 293.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1043.0, + 395.0, + 1043.0, + 395.0, + 1090.0, + 290.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 1043.0, + 559.0, + 1043.0, + 559.0, + 1090.0, + 478.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 1043.0, + 1406.0, + 1043.0, + 1406.0, + 1090.0, + 657.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1079.0, + 390.0, + 1079.0, + 390.0, + 1116.0, + 294.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 1079.0, + 577.0, + 1079.0, + 577.0, + 1116.0, + 429.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1079.0, + 1237.0, + 1079.0, + 1237.0, + 1116.0, + 598.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 1079.0, + 1405.0, + 1079.0, + 1405.0, + 1116.0, + 1296.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1106.0, + 297.0, + 1106.0, + 297.0, + 1148.0, + 290.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 1106.0, + 594.0, + 1106.0, + 594.0, + 1148.0, + 516.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 1106.0, + 1182.0, + 1106.0, + 1182.0, + 1148.0, + 722.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 1106.0, + 1253.0, + 1106.0, + 1253.0, + 1148.0, + 1203.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1106.0, + 1408.0, + 1106.0, + 1408.0, + 1148.0, + 1402.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1140.0, + 1251.0, + 1140.0, + 1251.0, + 1177.0, + 293.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 1140.0, + 1406.0, + 1140.0, + 1406.0, + 1177.0, + 1308.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1167.0, + 1055.0, + 1167.0, + 1055.0, + 1211.0, + 291.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1650.0, + 631.0, + 1650.0, + 631.0, + 1687.0, + 295.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 1650.0, + 749.0, + 1650.0, + 749.0, + 1687.0, + 652.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 1650.0, + 1294.0, + 1650.0, + 1294.0, + 1687.0, + 770.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1674.0, + 296.0, + 1674.0, + 296.0, + 1719.0, + 291.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1674.0, + 478.0, + 1674.0, + 478.0, + 1719.0, + 398.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1674.0, + 698.0, + 1674.0, + 698.0, + 1719.0, + 505.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 1674.0, + 1409.0, + 1674.0, + 1409.0, + 1719.0, + 720.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1711.0, + 537.0, + 1711.0, + 537.0, + 1746.0, + 292.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1711.0, + 1036.0, + 1711.0, + 1036.0, + 1746.0, + 962.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 1711.0, + 1406.0, + 1711.0, + 1406.0, + 1746.0, + 1061.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1742.0, + 494.0, + 1742.0, + 494.0, + 1779.0, + 294.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 1742.0, + 1031.0, + 1742.0, + 1031.0, + 1779.0, + 561.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 1742.0, + 1406.0, + 1742.0, + 1406.0, + 1779.0, + 1051.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1774.0, + 726.0, + 1774.0, + 726.0, + 1808.0, + 295.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 1774.0, + 837.0, + 1774.0, + 837.0, + 1808.0, + 749.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 1774.0, + 1195.0, + 1774.0, + 1195.0, + 1808.0, + 1117.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1774.0, + 1404.0, + 1774.0, + 1404.0, + 1808.0, + 1262.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1802.0, + 913.0, + 1802.0, + 913.0, + 1836.0, + 296.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 1802.0, + 1405.0, + 1802.0, + 1405.0, + 1836.0, + 1039.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1835.0, + 1404.0, + 1835.0, + 1404.0, + 1868.0, + 295.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1866.0, + 1293.0, + 1866.0, + 1293.0, + 1900.0, + 295.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1910.0, + 771.0, + 1910.0, + 771.0, + 1946.0, + 295.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 1910.0, + 920.0, + 1910.0, + 920.0, + 1946.0, + 791.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 1910.0, + 1407.0, + 1910.0, + 1407.0, + 1946.0, + 939.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1945.0, + 474.0, + 1945.0, + 474.0, + 1974.0, + 297.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 1945.0, + 546.0, + 1945.0, + 546.0, + 1974.0, + 496.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 1945.0, + 1403.0, + 1945.0, + 1403.0, + 1974.0, + 609.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2004.0, + 1405.0, + 2004.0, + 1405.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 549.0, + 227.0, + 549.0, + 267.0, + 295.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 227.0, + 831.0, + 227.0, + 831.0, + 267.0, + 682.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 227.0, + 1161.0, + 227.0, + 1161.0, + 267.0, + 908.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 267.0, + 1249.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 261.0, + 569.0, + 261.0, + 569.0, + 297.0, + 293.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 261.0, + 877.0, + 261.0, + 877.0, + 297.0, + 638.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 357.0, + 371.0, + 357.0, + 371.0, + 398.0, + 293.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 357.0, + 1083.0, + 357.0, + 1083.0, + 398.0, + 393.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 357.0, + 1406.0, + 357.0, + 1406.0, + 398.0, + 1102.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 390.0, + 297.0, + 390.0, + 297.0, + 430.0, + 291.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 390.0, + 959.0, + 390.0, + 959.0, + 430.0, + 316.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 390.0, + 1032.0, + 390.0, + 1032.0, + 430.0, + 1020.0, + 430.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 339, + 311, + 1359, + 311, + 1359, + 602, + 339, + 602 + ], + "score": 0.985, + "html": "
DatasetContextQueryAnswerTraining Size
Sequence8384duplicate8≤1k,≤50
Grid00 3 1 859corner7≤1k,≤50
bAbIMary went to the kitchen. Sandra journeyed to the garden.Where is Mary?kitchen1k,50
Logicp(X)← q(X). q(a).p(a).True1k, 10k,50
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1038, + 1404, + 1038, + 1404, + 1321, + 297, + 1321 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 657, + 1405, + 657, + 1405, + 1024, + 298, + 1024 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1434, + 1404, + 1434, + 1404, + 1650, + 298, + 1650 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1665, + 1402, + 1665, + 1402, + 1788, + 299, + 1788 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1803, + 1402, + 1803, + 1402, + 1958, + 298, + 1958 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1972, + 1400, + 1972, + 1400, + 2034, + 297, + 2034 + ], + "score": 0.94 + }, + { + "category_id": 6, + "poly": [ + 299, + 224, + 1399, + 224, + 1399, + 287, + 299, + 287 + ], + "score": 0.939 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.913 + }, + { + "category_id": 0, + "poly": [ + 299, + 1365, + 498, + 1365, + 498, + 1400, + 299, + 1400 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.757 + }, + { + "category_id": 13, + "poly": [ + 870, + 900, + 1046, + 900, + 1046, + 935, + 870, + 935 + ], + "score": 0.93, + "latex": "\\boldsymbol { \\phi } ( \\pmb { k } ) \\in \\mathbb { R } ^ { m \\times d }" + }, + { + "category_id": 13, + "poly": [ + 995, + 1435, + 1097, + 1435, + 1097, + 1469, + 995, + 1469 + ], + "score": 0.93, + "latex": "( C , \\pmb { q } , a )" + }, + { + "category_id": 13, + "poly": [ + 500, + 839, + 707, + 839, + 707, + 869, + 500, + 869 + ], + "score": 0.92, + "latex": "\\pmb { \\mathsf { U } } \\in \\dot { \\mathbb { R } } ^ { M \\times m \\times N \\times d }" + }, + { + "category_id": 13, + "poly": [ + 880, + 1131, + 1084, + 1131, + 1084, + 1167, + 880, + 1167 + ], + "score": 0.92, + "latex": "\\textstyle \\sum _ { I \\in \\mathbb { I } } f \\circ g ( I , K )" + }, + { + "category_id": 13, + "poly": [ + 758, + 1226, + 802, + 1226, + 802, + 1262, + 758, + 1262 + ], + "score": 0.91, + "latex": "h _ { K } ^ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 483, + 961, + 620, + 961, + 620, + 997, + 483, + 997 + ], + "score": 0.9, + "latex": "( W h _ { I } ^ { J } + b )" + }, + { + "category_id": 13, + "poly": [ + 638, + 1101, + 703, + 1101, + 703, + 1129, + 638, + 1129 + ], + "score": 0.9, + "latex": "I \\in \\mathbb { I }" + }, + { + "category_id": 13, + "poly": [ + 916, + 1222, + 1196, + 1222, + 1196, + 1263, + 916, + 1263 + ], + "score": 0.9, + "latex": "\\eta = \\mathrm { s o f t m a x } ( h _ { I } ^ { 0 } W h _ { K } ^ { 0 ^ { T } } )" + }, + { + "category_id": 13, + "poly": [ + 950, + 1666, + 1014, + 1666, + 1014, + 1694, + 950, + 1694 + ], + "score": 0.89, + "latex": "l = 4" + }, + { + "category_id": 13, + "poly": [ + 671, + 1226, + 707, + 1226, + 707, + 1262, + 671, + 1262 + ], + "score": 0.89, + "latex": "h _ { I } ^ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 348, + 1835, + 415, + 1835, + 415, + 1864, + 348, + 1864 + ], + "score": 0.89, + "latex": "2 \\times 2" + }, + { + "category_id": 13, + "poly": [ + 980, + 1804, + 1036, + 1804, + 1036, + 1833, + 980, + 1833 + ], + "score": 0.89, + "latex": "3 \\times 3" + }, + { + "category_id": 13, + "poly": [ + 702, + 961, + 739, + 961, + 739, + 996, + 702, + 996 + ], + "score": 0.89, + "latex": "h _ { I } ^ { J }" + }, + { + "category_id": 13, + "poly": [ + 1154, + 660, + 1174, + 660, + 1174, + 691, + 1154, + 691 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 788, + 842, + 821, + 842, + 821, + 868, + 788, + 868 + ], + "score": 0.84, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 710, + 996, + 729, + 996, + 729, + 1025, + 710, + 1025 + ], + "score": 0.83, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 596, + 873, + 625, + 873, + 625, + 899, + 596, + 899 + ], + "score": 0.83, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 676, + 812, + 696, + 812, + 696, + 841, + 676, + 841 + ], + "score": 0.83, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 703, + 903, + 731, + 903, + 731, + 930, + 703, + 930 + ], + "score": 0.81, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1198, + 964, + 1219, + 964, + 1219, + 991, + 1198, + 991 + ], + "score": 0.81, + "latex": "J" + }, + { + "category_id": 13, + "poly": [ + 391, + 780, + 658, + 780, + 658, + 813, + 391, + 813 + ], + "score": 0.79, + "latex": "\\phi _ { U } ( \\mathbf { b e r n h a r d } ) = W _ { U } \\Phi (" + }, + { + "category_id": 13, + "poly": [ + 1228, + 1198, + 1248, + 1198, + 1248, + 1225, + 1228, + 1225 + ], + "score": 0.78, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 873, + 845, + 901, + 845, + 901, + 869, + 873, + 869 + ], + "score": 0.78, + "latex": "m" + }, + { + "category_id": 13, + "poly": [ + 612, + 903, + 645, + 903, + 645, + 930, + 612, + 930 + ], + "score": 0.76, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 1242, + 816, + 1261, + 816, + 1261, + 843, + 1242, + 843 + ], + "score": 0.76, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 1178, + 1040, + 1204, + 1040, + 1204, + 1068, + 1178, + 1068 + ], + "score": 0.26, + "latex": "\\mathrm { X }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 224.0, + 1404.0, + 224.0, + 1404.0, + 259.0, + 295.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 254.0, + 1309.0, + 254.0, + 1309.0, + 290.0, + 295.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1362.0, + 504.0, + 1362.0, + 504.0, + 1406.0, + 292.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1037.0, + 1177.0, + 1037.0, + 1177.0, + 1074.0, + 294.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1037.0, + 1407.0, + 1037.0, + 1407.0, + 1074.0, + 1205.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1069.0, + 1405.0, + 1069.0, + 1405.0, + 1104.0, + 294.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1101.0, + 637.0, + 1101.0, + 637.0, + 1138.0, + 295.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1101.0, + 1405.0, + 1101.0, + 1405.0, + 1138.0, + 704.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1128.0, + 879.0, + 1128.0, + 879.0, + 1172.0, + 291.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 1128.0, + 1407.0, + 1128.0, + 1407.0, + 1172.0, + 1085.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1160.0, + 1404.0, + 1160.0, + 1404.0, + 1196.0, + 294.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1191.0, + 1227.0, + 1191.0, + 1227.0, + 1227.0, + 294.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 1191.0, + 1404.0, + 1191.0, + 1404.0, + 1227.0, + 1249.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1218.0, + 670.0, + 1218.0, + 670.0, + 1274.0, + 286.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 1218.0, + 757.0, + 1218.0, + 757.0, + 1274.0, + 708.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 1218.0, + 915.0, + 1218.0, + 915.0, + 1274.0, + 803.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 1218.0, + 1411.0, + 1218.0, + 1411.0, + 1274.0, + 1197.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1258.0, + 1405.0, + 1258.0, + 1405.0, + 1296.0, + 294.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1290.0, + 1333.0, + 1290.0, + 1333.0, + 1323.0, + 295.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 657.0, + 1153.0, + 657.0, + 1153.0, + 695.0, + 293.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 657.0, + 1406.0, + 657.0, + 1406.0, + 695.0, + 1175.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 691.0, + 1403.0, + 691.0, + 1403.0, + 721.0, + 296.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 719.0, + 1405.0, + 719.0, + 1405.0, + 753.0, + 293.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 744.0, + 1407.0, + 744.0, + 1407.0, + 789.0, + 291.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 778.0, + 390.0, + 778.0, + 390.0, + 813.0, + 295.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 778.0, + 1405.0, + 778.0, + 1405.0, + 813.0, + 659.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 811.0, + 675.0, + 811.0, + 675.0, + 845.0, + 295.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 811.0, + 1241.0, + 811.0, + 1241.0, + 845.0, + 697.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 811.0, + 1406.0, + 811.0, + 1406.0, + 845.0, + 1262.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 834.0, + 499.0, + 834.0, + 499.0, + 876.0, + 292.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 834.0, + 787.0, + 834.0, + 787.0, + 876.0, + 708.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 834.0, + 872.0, + 834.0, + 872.0, + 876.0, + 822.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 834.0, + 1407.0, + 834.0, + 1407.0, + 876.0, + 902.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 872.0, + 595.0, + 872.0, + 595.0, + 905.0, + 293.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 872.0, + 1406.0, + 872.0, + 1406.0, + 905.0, + 626.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 896.0, + 611.0, + 896.0, + 611.0, + 938.0, + 292.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 896.0, + 702.0, + 896.0, + 702.0, + 938.0, + 646.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 896.0, + 869.0, + 896.0, + 869.0, + 938.0, + 732.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 896.0, + 1407.0, + 896.0, + 1407.0, + 938.0, + 1047.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 932.0, + 1405.0, + 932.0, + 1405.0, + 966.0, + 295.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 961.0, + 482.0, + 961.0, + 482.0, + 999.0, + 293.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 961.0, + 701.0, + 961.0, + 701.0, + 999.0, + 621.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 961.0, + 1197.0, + 961.0, + 1197.0, + 999.0, + 740.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 961.0, + 1406.0, + 961.0, + 1406.0, + 999.0, + 1220.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 994.0, + 709.0, + 994.0, + 709.0, + 1025.0, + 293.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 994.0, + 1403.0, + 994.0, + 1403.0, + 1025.0, + 730.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1432.0, + 994.0, + 1432.0, + 994.0, + 1474.0, + 293.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 1432.0, + 1409.0, + 1432.0, + 1409.0, + 1474.0, + 1098.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1465.0, + 1406.0, + 1465.0, + 1406.0, + 1501.0, + 292.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1494.0, + 1408.0, + 1494.0, + 1408.0, + 1536.0, + 291.0, + 1536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1526.0, + 1408.0, + 1526.0, + 1408.0, + 1564.0, + 293.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1554.0, + 1406.0, + 1554.0, + 1406.0, + 1594.0, + 293.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1586.0, + 1405.0, + 1586.0, + 1405.0, + 1623.0, + 293.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1617.0, + 980.0, + 1617.0, + 980.0, + 1654.0, + 293.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1663.0, + 949.0, + 1663.0, + 949.0, + 1701.0, + 293.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 1663.0, + 1406.0, + 1663.0, + 1406.0, + 1701.0, + 1015.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1695.0, + 1407.0, + 1695.0, + 1407.0, + 1730.0, + 292.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1728.0, + 1407.0, + 1728.0, + 1407.0, + 1760.0, + 293.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1758.0, + 1337.0, + 1758.0, + 1337.0, + 1790.0, + 294.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1801.0, + 979.0, + 1801.0, + 979.0, + 1841.0, + 294.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 1801.0, + 1406.0, + 1801.0, + 1406.0, + 1841.0, + 1037.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1831.0, + 347.0, + 1831.0, + 347.0, + 1871.0, + 293.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 1831.0, + 1406.0, + 1831.0, + 1406.0, + 1871.0, + 416.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1863.0, + 1406.0, + 1863.0, + 1406.0, + 1900.0, + 294.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1892.0, + 1407.0, + 1892.0, + 1407.0, + 1931.0, + 292.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1923.0, + 1404.0, + 1923.0, + 1404.0, + 1962.0, + 293.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1970.0, + 1402.0, + 1970.0, + 1402.0, + 2007.0, + 293.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1404.0, + 2002.0, + 1404.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 337, + 1405, + 337, + 1405, + 585, + 297, + 585 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 296, + 963, + 1404, + 963, + 1404, + 1214, + 296, + 1214 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 695, + 1403, + 695, + 1403, + 850, + 298, + 850 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1667, + 1404, + 1667, + 1404, + 2034, + 297, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 305, + 1230, + 1392, + 1230, + 1392, + 1527, + 305, + 1527 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 228, + 1403, + 228, + 1403, + 324, + 299, + 324 + ], + "score": 0.965 + }, + { + "category_id": 8, + "poly": [ + 476, + 860, + 1216, + 860, + 1216, + 952, + 476, + 952 + ], + "score": 0.957 + }, + { + "category_id": 4, + "poly": [ + 297, + 1558, + 1404, + 1558, + 1404, + 1653, + 297, + 1653 + ], + "score": 0.948 + }, + { + "category_id": 0, + "poly": [ + 300, + 628, + 558, + 628, + 558, + 663, + 300, + 663 + ], + "score": 0.909 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 104, + 298, + 104 + ], + "score": 0.898 + }, + { + "category_id": 9, + "poly": [ + 1366, + 891, + 1400, + 891, + 1400, + 921, + 1366, + 921 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.74 + }, + { + "category_id": 14, + "poly": [ + 476, + 858, + 1223, + 858, + 1223, + 954, + 476, + 954 + ], + "score": 0.93, + "latex": "J = \\lambda _ { K } \\mathcal { L } _ { \\mathrm { n l l } } ( f ( K ) , a ) + \\lambda _ { U } \\left[ \\mathcal { L } _ { \\mathrm { n l l } } ( f \\circ g ( I , K ) , a ) + \\tau \\sum _ { s \\in \\mathbb { S } } \\psi _ { w } ( s ) \\right]" + }, + { + "category_id": 13, + "poly": [ + 1258, + 1086, + 1400, + 1086, + 1400, + 1125, + 1258, + 1125 + ], + "score": 0.92, + "latex": "( { h _ { I } ^ { j } - h _ { K } ^ { j } } ) ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 488, + 1729, + 551, + 1729, + 551, + 1761, + 488, + 1761 + ], + "score": 0.91, + "latex": "f \\circ g" + }, + { + "category_id": 13, + "poly": [ + 1139, + 1057, + 1233, + 1057, + 1233, + 1086, + 1139, + 1086 + ], + "score": 0.91, + "latex": "\\lambda _ { U } = 1" + }, + { + "category_id": 13, + "poly": [ + 395, + 429, + 697, + 429, + 697, + 463, + 395, + 463 + ], + "score": 0.91, + "latex": "f ( \\bar { C } , \\bar { Q } ) \\stackrel { - } { = } \\mathrm { T r u e } \\bar { C } \\vdash \\bar { Q }" + }, + { + "category_id": 13, + "poly": [ + 998, + 552, + 1052, + 552, + 1052, + 585, + 998, + 585 + ], + "score": 0.91, + "latex": "p ( \\mathrm { x } )" + }, + { + "category_id": 13, + "poly": [ + 346, + 551, + 398, + 551, + 398, + 585, + 346, + 585 + ], + "score": 0.9, + "latex": "p ( p )" + }, + { + "category_id": 13, + "poly": [ + 372, + 966, + 415, + 966, + 415, + 995, + 372, + 995 + ], + "score": 0.9, + "latex": "{ \\mathcal { L } } _ { \\mathrm { n l l } }" + }, + { + "category_id": 13, + "poly": [ + 1156, + 965, + 1246, + 965, + 1246, + 993, + 1156, + 993 + ], + "score": 0.89, + "latex": "\\tau = 0 . 1" + }, + { + "category_id": 13, + "poly": [ + 1324, + 995, + 1403, + 995, + 1403, + 1026, + 1324, + 1026 + ], + "score": 0.89, + "latex": "\\lambda _ { K } ~ =" + }, + { + "category_id": 13, + "poly": [ + 353, + 1790, + 446, + 1790, + 446, + 1820, + 353, + 1820 + ], + "score": 0.87, + "latex": "\\leq 1 0 0 0" + }, + { + "category_id": 13, + "poly": [ + 1337, + 1790, + 1402, + 1790, + 1402, + 1821, + 1337, + 1821 + ], + "score": 0.86, + "latex": "\\leq 5 0" + }, + { + "category_id": 13, + "poly": [ + 992, + 1730, + 1013, + 1730, + 1013, + 1761, + 992, + 1761 + ], + "score": 0.86, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1105, + 966, + 1127, + 966, + 1127, + 996, + 1105, + 996 + ], + "score": 0.86, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 993, + 1790, + 1065, + 1790, + 1065, + 1819, + 993, + 1819 + ], + "score": 0.85, + "latex": "\\approx 0 . 3" + }, + { + "category_id": 13, + "poly": [ + 421, + 1944, + 444, + 1944, + 444, + 1974, + 421, + 1974 + ], + "score": 0.83, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 1186, + 1974, + 1209, + 1974, + 1209, + 2005, + 1186, + 2005 + ], + "score": 0.83, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 484, + 1123, + 501, + 1123, + 501, + 1154, + 484, + 1154 + ], + "score": 0.82, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 328, + 1026, + 427, + 1026, + 427, + 1057, + 328, + 1057 + ], + "score": 0.81, + "latex": "\\lambda _ { U } = 1" + }, + { + "category_id": 13, + "poly": [ + 675, + 1056, + 770, + 1056, + 770, + 1086, + 675, + 1086 + ], + "score": 0.81, + "latex": "\\lambda _ { K } = 1" + }, + { + "category_id": 13, + "poly": [ + 786, + 1056, + 880, + 1056, + 880, + 1087, + 786, + 1087 + ], + "score": 0.77, + "latex": "\\lambda _ { U } = 0" + }, + { + "category_id": 13, + "poly": [ + 675, + 1056, + 881, + 1056, + 881, + 1088, + 675, + 1088 + ], + "score": 0.29, + "latex": "\\lambda _ { K } = 1 , \\lambda _ { U } = 0" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 1228.0, + 600.0, + 1228.0, + 600.0, + 1254.0, + 555.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 1230.0, + 1046.0, + 1230.0, + 1046.0, + 1256.0, + 1001.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1242.0, + 374.0, + 1242.0, + 374.0, + 1272.0, + 323.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1293.0, + 364.0, + 1293.0, + 364.0, + 1319.0, + 324.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 1308.0, + 332.0, + 1308.0, + 332.0, + 1427.0, + 303.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 1312.0, + 1376.0, + 1312.0, + 1376.0, + 1336.0, + 1285.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1344.0, + 361.0, + 1344.0, + 361.0, + 1367.0, + 327.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 1333.0, + 1320.0, + 1333.0, + 1320.0, + 1360.0, + 1281.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 1355.0, + 1325.0, + 1355.0, + 1325.0, + 1381.0, + 1282.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 1375.0, + 1394.0, + 1375.0, + 1394.0, + 1406.0, + 1283.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1391.0, + 363.0, + 1391.0, + 363.0, + 1418.0, + 324.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 1398.0, + 1352.0, + 1398.0, + 1352.0, + 1423.0, + 1280.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 1419.0, + 1334.0, + 1419.0, + 1334.0, + 1448.0, + 1260.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1439.0, + 363.0, + 1439.0, + 363.0, + 1468.0, + 324.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 1484.0, + 393.0, + 1484.0, + 393.0, + 1505.0, + 374.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 1483.0, + 504.0, + 1483.0, + 504.0, + 1508.0, + 411.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1481.0, + 797.0, + 1481.0, + 797.0, + 1509.0, + 505.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 1485.0, + 840.0, + 1485.0, + 840.0, + 1505.0, + 819.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 1481.0, + 1242.0, + 1481.0, + 1242.0, + 1509.0, + 855.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 1501.0, + 616.0, + 1501.0, + 616.0, + 1529.0, + 540.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 1501.0, + 1063.0, + 1501.0, + 1063.0, + 1529.0, + 986.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1559.0, + 1405.0, + 1559.0, + 1405.0, + 1592.0, + 295.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1591.0, + 1405.0, + 1591.0, + 1405.0, + 1625.0, + 295.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1621.0, + 1301.0, + 1621.0, + 1301.0, + 1655.0, + 296.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 623.0, + 561.0, + 623.0, + 561.0, + 670.0, + 291.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 337.0, + 1407.0, + 337.0, + 1407.0, + 374.0, + 294.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 368.0, + 1406.0, + 368.0, + 1406.0, + 405.0, + 291.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 397.0, + 1407.0, + 397.0, + 1407.0, + 438.0, + 292.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 428.0, + 394.0, + 428.0, + 394.0, + 466.0, + 294.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 428.0, + 1406.0, + 428.0, + 1406.0, + 466.0, + 698.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 461.0, + 1405.0, + 461.0, + 1405.0, + 495.0, + 295.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 490.0, + 1407.0, + 490.0, + 1407.0, + 526.0, + 291.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 520.0, + 1406.0, + 520.0, + 1406.0, + 558.0, + 294.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 552.0, + 345.0, + 552.0, + 345.0, + 588.0, + 291.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 552.0, + 997.0, + 552.0, + 997.0, + 588.0, + 399.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 552.0, + 1064.0, + 552.0, + 1064.0, + 588.0, + 1053.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 962.0, + 371.0, + 962.0, + 371.0, + 1002.0, + 294.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 962.0, + 1104.0, + 962.0, + 1104.0, + 1002.0, + 416.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 962.0, + 1155.0, + 962.0, + 1155.0, + 1002.0, + 1128.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 962.0, + 1405.0, + 962.0, + 1405.0, + 1002.0, + 1247.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 995.0, + 1323.0, + 995.0, + 1323.0, + 1029.0, + 295.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1025.0, + 327.0, + 1025.0, + 327.0, + 1059.0, + 293.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 1025.0, + 1405.0, + 1025.0, + 1405.0, + 1059.0, + 428.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1054.0, + 674.0, + 1054.0, + 674.0, + 1091.0, + 291.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 1054.0, + 1138.0, + 1054.0, + 1138.0, + 1091.0, + 882.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1054.0, + 1405.0, + 1054.0, + 1405.0, + 1091.0, + 1234.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1086.0, + 1257.0, + 1086.0, + 1257.0, + 1129.0, + 290.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1086.0, + 1407.0, + 1086.0, + 1407.0, + 1129.0, + 1401.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1120.0, + 483.0, + 1120.0, + 483.0, + 1156.0, + 291.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 1120.0, + 1407.0, + 1120.0, + 1407.0, + 1156.0, + 502.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1151.0, + 1406.0, + 1151.0, + 1406.0, + 1187.0, + 291.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1182.0, + 1133.0, + 1182.0, + 1133.0, + 1216.0, + 294.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 697.0, + 1405.0, + 697.0, + 1405.0, + 730.0, + 297.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 726.0, + 1404.0, + 726.0, + 1404.0, + 758.0, + 294.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 754.0, + 1405.0, + 754.0, + 1405.0, + 792.0, + 293.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 782.0, + 1404.0, + 782.0, + 1404.0, + 827.0, + 293.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 815.0, + 1397.0, + 815.0, + 1397.0, + 852.0, + 294.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1669.0, + 1405.0, + 1669.0, + 1405.0, + 1703.0, + 296.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1700.0, + 1404.0, + 1700.0, + 1404.0, + 1733.0, + 293.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1730.0, + 487.0, + 1730.0, + 487.0, + 1762.0, + 293.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1730.0, + 991.0, + 1730.0, + 991.0, + 1762.0, + 552.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 1730.0, + 1405.0, + 1730.0, + 1405.0, + 1762.0, + 1014.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1759.0, + 1406.0, + 1759.0, + 1406.0, + 1793.0, + 294.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1788.0, + 352.0, + 1788.0, + 352.0, + 1824.0, + 294.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1788.0, + 992.0, + 1788.0, + 992.0, + 1824.0, + 447.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 1788.0, + 1336.0, + 1788.0, + 1336.0, + 1824.0, + 1066.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1820.0, + 1407.0, + 1820.0, + 1407.0, + 1855.0, + 291.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1851.0, + 1407.0, + 1851.0, + 1407.0, + 1887.0, + 292.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1880.0, + 1407.0, + 1880.0, + 1407.0, + 1919.0, + 294.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1910.0, + 1404.0, + 1910.0, + 1404.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1943.0, + 420.0, + 1943.0, + 420.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1943.0, + 1405.0, + 1943.0, + 1405.0, + 1977.0, + 445.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1973.0, + 1185.0, + 1973.0, + 1185.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2007.0, + 1210.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2000.0, + 1083.0, + 2000.0, + 1083.0, + 2039.0, + 292.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 296.0, + 294.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 289.0, + 1289.0, + 289.0, + 1289.0, + 327.0, + 294.0, + 327.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1346, + 1405, + 1346, + 1405, + 1805, + 297, + 1805 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 570, + 1404, + 570, + 1404, + 907, + 297, + 907 + ], + "score": 0.982 + }, + { + "category_id": 5, + "poly": [ + 345, + 341, + 1354, + 341, + 1354, + 518, + 345, + 518 + ], + "score": 0.981, + "html": "
Training Size Supervision #Invs /Model1k501k
Weak 13Strong 3WeakQRNStrong MemNN
31N2N GN2NEntNet
Mean19.120.56.36.027.613.912.729.611.36.7
#>5%810441711101554
" + }, + { + "category_id": 5, + "poly": [ + 307, + 1035, + 1393, + 1035, + 1393, + 1239, + 307, + 1239 + ], + "score": 0.98, + "html": "
Training Size Supervision1k5020k Weak
WeakStrongWeakStrong
121212122222
# Invs /Model1313131133DMNIMA
Mean36.439.314.328.921.531.82.412.216.047.121.29.1
#>5%91171171015912115
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1819, + 1403, + 1819, + 1403, + 2035, + 297, + 2035 + ], + "score": 0.979 + }, + { + "category_id": 6, + "poly": [ + 297, + 223, + 1405, + 223, + 1405, + 317, + 297, + 317 + ], + "score": 0.964 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.914 + }, + { + "category_id": 0, + "poly": [ + 298, + 1278, + 497, + 1278, + 497, + 1312, + 298, + 1312 + ], + "score": 0.912 + }, + { + "category_id": 6, + "poly": [ + 298, + 916, + 1400, + 916, + 1400, + 1010, + 298, + 1010 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 859, + 2089, + 859, + 2112, + 840, + 2112 + ], + "score": 0.783 + }, + { + "category_id": 13, + "poly": [ + 514, + 1773, + 666, + 1773, + 666, + 1807, + 514, + 1807 + ], + "score": 0.93, + "latex": "g ( I , K ) \\approx K" + }, + { + "category_id": 13, + "poly": [ + 1248, + 694, + 1315, + 694, + 1315, + 725, + 1248, + 725 + ], + "score": 0.92, + "latex": "f \\circ g" + }, + { + "category_id": 13, + "poly": [ + 449, + 816, + 510, + 816, + 510, + 847, + 449, + 847 + ], + "score": 0.92, + "latex": "f \\circ g" + }, + { + "category_id": 13, + "poly": [ + 958, + 1743, + 1018, + 1743, + 1018, + 1775, + 958, + 1775 + ], + "score": 0.91, + "latex": "f \\circ g" + }, + { + "category_id": 13, + "poly": [ + 1185, + 1347, + 1402, + 1347, + 1402, + 1381, + 1185, + 1381 + ], + "score": 0.91, + "latex": "\\forall s \\in \\mathbb { S } : \\psi ( s ) > t" + }, + { + "category_id": 13, + "poly": [ + 1047, + 1379, + 1144, + 1379, + 1144, + 1407, + 1047, + 1407 + ], + "score": 0.9, + "latex": "t = 0 . 0" + }, + { + "category_id": 13, + "poly": [ + 503, + 1409, + 600, + 1409, + 600, + 1436, + 503, + 1436 + ], + "score": 0.9, + "latex": "t = 0 . 1" + }, + { + "category_id": 13, + "poly": [ + 931, + 1683, + 969, + 1683, + 969, + 1714, + 931, + 1714 + ], + "score": 0.89, + "latex": "\\phi _ { U }" + }, + { + "category_id": 13, + "poly": [ + 715, + 2004, + 734, + 2004, + 734, + 2035, + 715, + 2035 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 416, + 1470, + 439, + 1470, + 439, + 1500, + 416, + 1500 + ], + "score": 0.85, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 848, + 1714, + 868, + 1714, + 868, + 1745, + 848, + 1745 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1257, + 785, + 1278, + 785, + 1278, + 816, + 1257, + 816 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 607, + 725, + 628, + 725, + 628, + 756, + 607, + 756 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1240, + 1622, + 1262, + 1622, + 1262, + 1653, + 1240, + 1653 + ], + "score": 0.84, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 970, + 1944, + 989, + 1944, + 989, + 1975, + 970, + 1975 + ], + "score": 0.83, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 297, + 1774, + 317, + 1774, + 317, + 1806, + 297, + 1806 + ], + "score": 0.82, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 638, + 226, + 681, + 226, + 681, + 255, + 638, + 255 + ], + "score": 0.72, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 457, + 1444, + 475, + 1444, + 475, + 1467, + 457, + 1467 + ], + "score": 0.71, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 683, + 917, + 725, + 917, + 725, + 947, + 683, + 947 + ], + "score": 0.68, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 985, + 1651, + 1122, + 1651, + 1122, + 1681, + 985, + 1681 + ], + "score": 0.68, + "latex": "^ { * } X { : } 8 \\ 5 \\ 2 \\ 2 ^ { , }" + }, + { + "category_id": 13, + "poly": [ + 296, + 1973, + 318, + 1973, + 318, + 2001, + 296, + 2001 + ], + "score": 0.47, + "latex": "\\mathbf { Z }" + }, + { + "category_id": 13, + "poly": [ + 327, + 2004, + 349, + 2004, + 349, + 2030, + 327, + 2030 + ], + "score": 0.4, + "latex": "\\mathbf { Z }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 221.0, + 637.0, + 221.0, + 637.0, + 260.0, + 292.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 221.0, + 1405.0, + 221.0, + 1405.0, + 260.0, + 682.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 252.0, + 1405.0, + 252.0, + 1405.0, + 291.0, + 292.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 286.0, + 723.0, + 286.0, + 723.0, + 320.0, + 295.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1272.0, + 503.0, + 1272.0, + 503.0, + 1318.0, + 292.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 913.0, + 682.0, + 913.0, + 682.0, + 953.0, + 293.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 913.0, + 1405.0, + 913.0, + 1405.0, + 953.0, + 726.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 946.0, + 1407.0, + 946.0, + 1407.0, + 984.0, + 294.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 978.0, + 1364.0, + 978.0, + 1364.0, + 1012.0, + 296.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 862.0, + 2087.0, + 862.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1346.0, + 1184.0, + 1346.0, + 1184.0, + 1381.0, + 294.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1375.0, + 1046.0, + 1375.0, + 1046.0, + 1415.0, + 291.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 1375.0, + 1407.0, + 1375.0, + 1407.0, + 1415.0, + 1145.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1407.0, + 502.0, + 1407.0, + 502.0, + 1442.0, + 294.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1407.0, + 1406.0, + 1407.0, + 1406.0, + 1442.0, + 601.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1434.0, + 456.0, + 1434.0, + 456.0, + 1478.0, + 290.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 1434.0, + 1407.0, + 1434.0, + 1407.0, + 1478.0, + 476.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1467.0, + 415.0, + 1467.0, + 415.0, + 1504.0, + 292.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 1467.0, + 1405.0, + 1467.0, + 1405.0, + 1504.0, + 440.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1499.0, + 1406.0, + 1499.0, + 1406.0, + 1534.0, + 292.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1531.0, + 1405.0, + 1531.0, + 1405.0, + 1563.0, + 292.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1562.0, + 1403.0, + 1562.0, + 1403.0, + 1593.0, + 296.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1589.0, + 1406.0, + 1589.0, + 1406.0, + 1625.0, + 291.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1620.0, + 1239.0, + 1620.0, + 1239.0, + 1655.0, + 292.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 1620.0, + 1407.0, + 1620.0, + 1407.0, + 1655.0, + 1263.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1649.0, + 984.0, + 1649.0, + 984.0, + 1688.0, + 294.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 1649.0, + 1407.0, + 1649.0, + 1407.0, + 1688.0, + 1123.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1680.0, + 930.0, + 1680.0, + 930.0, + 1719.0, + 294.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 1680.0, + 1407.0, + 1680.0, + 1407.0, + 1719.0, + 970.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1713.0, + 847.0, + 1713.0, + 847.0, + 1748.0, + 295.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1713.0, + 1405.0, + 1713.0, + 1405.0, + 1748.0, + 869.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1741.0, + 957.0, + 1741.0, + 957.0, + 1779.0, + 294.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 1741.0, + 1406.0, + 1741.0, + 1406.0, + 1779.0, + 1019.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1773.0, + 513.0, + 1773.0, + 513.0, + 1808.0, + 318.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 1773.0, + 677.0, + 1773.0, + 677.0, + 1808.0, + 667.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 570.0, + 1404.0, + 570.0, + 1404.0, + 607.0, + 294.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 602.0, + 1404.0, + 602.0, + 1404.0, + 638.0, + 294.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 632.0, + 1406.0, + 632.0, + 1406.0, + 666.0, + 295.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 663.0, + 1404.0, + 663.0, + 1404.0, + 698.0, + 295.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 691.0, + 1247.0, + 691.0, + 1247.0, + 729.0, + 292.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 691.0, + 1405.0, + 691.0, + 1405.0, + 729.0, + 1316.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 722.0, + 606.0, + 722.0, + 606.0, + 760.0, + 294.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 722.0, + 1405.0, + 722.0, + 1405.0, + 760.0, + 629.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 753.0, + 1405.0, + 753.0, + 1405.0, + 788.0, + 295.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 784.0, + 1256.0, + 784.0, + 1256.0, + 819.0, + 295.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 784.0, + 1404.0, + 784.0, + 1404.0, + 819.0, + 1279.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 813.0, + 448.0, + 813.0, + 448.0, + 851.0, + 294.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 813.0, + 1405.0, + 813.0, + 1405.0, + 851.0, + 511.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 847.0, + 1405.0, + 847.0, + 1405.0, + 881.0, + 294.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 875.0, + 1307.0, + 875.0, + 1307.0, + 912.0, + 292.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1818.0, + 1403.0, + 1818.0, + 1403.0, + 1857.0, + 294.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1848.0, + 1405.0, + 1848.0, + 1405.0, + 1888.0, + 294.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1880.0, + 1404.0, + 1880.0, + 1404.0, + 1917.0, + 294.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1943.0, + 969.0, + 1943.0, + 969.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1943.0, + 1405.0, + 1943.0, + 1405.0, + 1978.0, + 990.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1971.0, + 295.0, + 1971.0, + 295.0, + 2007.0, + 292.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2007.0, + 319.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1999.0, + 326.0, + 1999.0, + 326.0, + 2039.0, + 292.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1999.0, + 714.0, + 1999.0, + 714.0, + 2039.0, + 350.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1999.0, + 1405.0, + 1999.0, + 1405.0, + 2039.0, + 735.0, + 2039.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1105, + 1405, + 1105, + 1405, + 1442, + 296, + 1442 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 905, + 1404, + 905, + 1404, + 1090, + 298, + 1090 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 857, + 390, + 1401, + 390, + 1401, + 474, + 857, + 474 + ], + "score": 0.925 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 105, + 297, + 105 + ], + "score": 0.905 + }, + { + "category_id": 1, + "poly": [ + 296, + 390, + 840, + 390, + 840, + 475, + 296, + 475 + ], + "score": 0.896 + }, + { + "category_id": 3, + "poly": [ + 309, + 1509, + 1322, + 1509, + 1322, + 1848, + 309, + 1848 + ], + "score": 0.882 + }, + { + "category_id": 1, + "poly": [ + 298, + 1993, + 1398, + 1993, + 1398, + 2027, + 298, + 2027 + ], + "score": 0.845 + }, + { + "category_id": 1, + "poly": [ + 298, + 779, + 1400, + 779, + 1400, + 843, + 298, + 843 + ], + "score": 0.812 + }, + { + "category_id": 1, + "poly": [ + 297, + 1853, + 656, + 1853, + 656, + 1965, + 297, + 1965 + ], + "score": 0.76 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 858, + 2087, + 858, + 2111, + 841, + 2111 + ], + "score": 0.687 + }, + { + "category_id": 4, + "poly": [ + 1041, + 1853, + 1402, + 1853, + 1402, + 1965, + 1041, + 1965 + ], + "score": 0.494 + }, + { + "category_id": 1, + "poly": [ + 872, + 226, + 1380, + 226, + 1380, + 374, + 872, + 374 + ], + "score": 0.437 + }, + { + "category_id": 1, + "poly": [ + 668, + 1856, + 1028, + 1856, + 1028, + 1967, + 668, + 1967 + ], + "score": 0.385 + }, + { + "category_id": 1, + "poly": [ + 1041, + 1853, + 1402, + 1853, + 1402, + 1965, + 1041, + 1965 + ], + "score": 0.351 + }, + { + "category_id": 5, + "poly": [ + 377, + 226, + 759, + 226, + 759, + 373, + 377, + 373 + ], + "score": 0.343, + "html": "
Y:john Z:left the X:football Y:john travelled to the A:office
where is the X:football
A:office
" + }, + { + "category_id": 1, + "poly": [ + 377, + 226, + 759, + 226, + 759, + 373, + 377, + 373 + ], + "score": 0.213 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.209 + }, + { + "category_id": 4, + "poly": [ + 668, + 1856, + 1028, + 1856, + 1028, + 1967, + 668, + 1967 + ], + "score": 0.168 + }, + { + "category_id": 3, + "poly": [ + 297, + 525, + 1394, + 525, + 1394, + 756, + 297, + 756 + ], + "score": 0.125 + }, + { + "category_id": 4, + "poly": [ + 298, + 779, + 1400, + 779, + 1400, + 843, + 298, + 843 + ], + "score": 0.105 + }, + { + "category_id": 13, + "poly": [ + 402, + 1380, + 478, + 1380, + 478, + 1412, + 402, + 1412 + ], + "score": 0.92, + "latex": "p ( q , q )" + }, + { + "category_id": 13, + "poly": [ + 525, + 725, + 621, + 725, + 621, + 755, + 525, + 755 + ], + "score": 0.92, + "latex": "( C , Q , a )" + }, + { + "category_id": 13, + "poly": [ + 446, + 1289, + 502, + 1289, + 502, + 1321, + 446, + 1321 + ], + "score": 0.92, + "latex": "\\psi ( s )" + }, + { + "category_id": 13, + "poly": [ + 1021, + 1379, + 1189, + 1379, + 1189, + 1413, + 1021, + 1413 + ], + "score": 0.91, + "latex": "\\psi ( w ) = 0 . 0 3 7" + }, + { + "category_id": 13, + "poly": [ + 297, + 1379, + 349, + 1379, + 349, + 1413, + 297, + 1413 + ], + "score": 0.91, + "latex": "p ( q )" + }, + { + "category_id": 13, + "poly": [ + 1241, + 1379, + 1402, + 1379, + 1402, + 1413, + 1241, + 1413 + ], + "score": 0.91, + "latex": "\\psi ( e ) = 0 . 0 4 2" + }, + { + "category_id": 13, + "poly": [ + 1337, + 906, + 1392, + 906, + 1392, + 940, + 1337, + 940 + ], + "score": 0.9, + "latex": "\\psi ( s )" + }, + { + "category_id": 13, + "poly": [ + 394, + 999, + 415, + 999, + 415, + 1030, + 394, + 1030 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 298, + 1031, + 320, + 1031, + 320, + 1061, + 298, + 1061 + ], + "score": 0.83, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 607, + 1034, + 627, + 1034, + 627, + 1056, + 607, + 1056 + ], + "score": 0.75, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 347, + 1198, + 371, + 1198, + 371, + 1225, + 347, + 1225 + ], + "score": 0.71, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 531, + 1354, + 548, + 1354, + 548, + 1382, + 531, + 1382 + ], + "score": 0.7, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 362, + 1910, + 384, + 1910, + 384, + 1934, + 362, + 1934 + ], + "score": 0.69, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1090, + 531, + 1260, + 531, + 1260, + 561, + 1090, + 561 + ], + "score": 0.67, + "latex": "X \\colon \\mathrm { ( ~ T ~ ) } \\gets Z \\colon \\mathrm { l }" + }, + { + "category_id": 13, + "poly": [ + 1003, + 265, + 1027, + 265, + 1027, + 292, + 1003, + 292 + ], + "score": 0.6, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1088, + 592, + 1272, + 592, + 1272, + 622, + 1088, + 622 + ], + "score": 0.57, + "latex": "\\pmb { R } \\mathrm { : x } \\left( \\mathrm { ~ K ~ } \\right) \\gets S \\mathrm { : n }" + }, + { + "category_id": 13, + "poly": [ + 296, + 908, + 319, + 908, + 319, + 934, + 296, + 934 + ], + "score": 0.55, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1206, + 725, + 1242, + 725, + 1242, + 751, + 1206, + 751 + ], + "score": 0.49, + "latex": "X ; \\mathbf { i }" + }, + { + "category_id": 14, + "poly": [ + 1086, + 622, + 1355, + 622, + 1355, + 655, + 1086, + 655 + ], + "score": 0.39, + "latex": "S { : } \\mathrm { n } \\left( Y { : } \\mathrm { o } \\right) \\mathrm { \\vdash } X { : } \\mathrm { i } \\left( Y { : } \\mathrm { o } \\right)" + }, + { + "category_id": 13, + "poly": [ + 1014, + 304, + 1038, + 304, + 1038, + 331, + 1014, + 331 + ], + "score": 0.38, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1200, + 391, + 1221, + 391, + 1221, + 415, + 1200, + 415 + ], + "score": 0.36, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1063, + 230, + 1086, + 230, + 1086, + 257, + 1063, + 257 + ], + "score": 0.33, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1087, + 562, + 1266, + 562, + 1266, + 591, + 1087, + 591 + ], + "score": 0.32, + "latex": "Z : \\mathrm { l } \\left( \\mathrm { ~ U ~ } \\right) \\gets R : \\mathrm { x }" + }, + { + "category_id": 13, + "poly": [ + 741, + 908, + 762, + 908, + 762, + 934, + 741, + 934 + ], + "score": 0.27, + "latex": "\\mathbf { Z }" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 1545.0, + 496.0, + 1545.0, + 496.0, + 1568.0, + 471.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 1547.0, + 552.0, + 1547.0, + 552.0, + 1566.0, + 534.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 1547.0, + 806.0, + 1547.0, + 806.0, + 1566.0, + 785.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 1545.0, + 899.0, + 1545.0, + 899.0, + 1569.0, + 875.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1550.0, + 1268.0, + 1550.0, + 1268.0, + 1566.0, + 1253.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 1600.0, + 756.0, + 1600.0, + 756.0, + 1622.0, + 731.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 1588.0, + 1156.0, + 1588.0, + 1156.0, + 1618.0, + 1111.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 1625.0, + 1167.0, + 1625.0, + 1167.0, + 1659.0, + 1132.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1664.0, + 382.0, + 1664.0, + 382.0, + 1688.0, + 355.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 1661.0, + 1170.0, + 1661.0, + 1170.0, + 1698.0, + 1108.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 1692.0, + 756.0, + 1692.0, + 756.0, + 1715.0, + 731.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1722.0, + 383.0, + 1722.0, + 383.0, + 1741.0, + 361.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 1715.0, + 1158.0, + 1715.0, + 1158.0, + 1733.0, + 1138.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 1741.0, + 1169.0, + 1741.0, + 1169.0, + 1775.0, + 1111.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1534.0, + 456.0, + 1534.0, + 456.0, + 1554.5, + 395.0, + 1554.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1538.0, + 627.0, + 1538.0, + 627.0, + 1558.0, + 573.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 1538.0, + 870.0, + 1538.0, + 870.0, + 1557.0, + 812.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 1541.0, + 955.0, + 1541.0, + 955.0, + 1560.5, + 909.0, + 1560.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 1527.5, + 1017.0, + 1527.5, + 1017.0, + 1549.5, + 939.0, + 1549.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 1619.0, + 406.0, + 1619.0, + 406.0, + 1643.0, + 301.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 1662.5, + 765.0, + 1662.5, + 765.0, + 1681.5, + 695.0, + 1681.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 1742.0, + 766.0, + 1742.0, + 766.0, + 1764.0, + 714.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1771.0, + 410.0, + 1771.0, + 410.0, + 1804.0, + 331.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1800.0, + 779.0, + 1800.0, + 779.0, + 1821.5, + 678.0, + 1821.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 1851.0, + 1404.0, + 1851.0, + 1404.0, + 1881.0, + 1040.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1881.0, + 1404.0, + 1881.0, + 1404.0, + 1909.0, + 1041.0, + 1909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 1907.0, + 1404.0, + 1907.0, + 1404.0, + 1941.0, + 1042.0, + 1941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 1930.0, + 1137.0, + 1930.0, + 1137.0, + 1971.0, + 1040.0, + 1971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 1854.0, + 1031.0, + 1854.0, + 1031.0, + 1885.0, + 667.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 1882.0, + 1032.0, + 1882.0, + 1032.0, + 1914.0, + 667.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1912.0, + 1030.0, + 1912.0, + 1030.0, + 1941.0, + 669.0, + 1941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1940.0, + 859.0, + 1940.0, + 859.0, + 1967.0, + 669.0, + 1967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 522.0, + 612.0, + 522.0, + 612.0, + 567.0, + 393.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 527.0, + 792.0, + 527.0, + 792.0, + 562.0, + 700.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 527.0, + 899.0, + 527.0, + 899.0, + 562.0, + 813.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 527.0, + 1006.0, + 527.0, + 1006.0, + 562.0, + 919.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 525.0, + 1089.0, + 525.0, + 1089.0, + 566.0, + 1085.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 525.0, + 1326.0, + 525.0, + 1326.0, + 566.0, + 1261.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 556.0, + 609.0, + 556.0, + 609.0, + 593.0, + 369.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 556.0, + 791.0, + 556.0, + 791.0, + 591.0, + 698.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 555.0, + 900.0, + 555.0, + 900.0, + 594.0, + 812.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 555.0, + 1009.0, + 555.0, + 1009.0, + 594.0, + 918.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 555.0, + 1331.0, + 555.0, + 1331.0, + 597.0, + 1267.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 585.0, + 611.0, + 585.0, + 611.0, + 626.0, + 367.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 586.0, + 793.0, + 586.0, + 793.0, + 625.0, + 697.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 589.0, + 899.0, + 589.0, + 899.0, + 624.0, + 813.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 589.0, + 1008.0, + 589.0, + 1008.0, + 624.0, + 919.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 589.0, + 1336.0, + 589.0, + 1336.0, + 625.0, + 1273.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 619.0, + 607.0, + 619.0, + 607.0, + 656.0, + 348.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 617.0, + 771.0, + 617.0, + 771.0, + 657.0, + 710.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 624.0, + 894.0, + 624.0, + 894.0, + 653.0, + 813.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 624.0, + 1003.0, + 624.0, + 1003.0, + 653.0, + 919.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 665.0, + 1400.0, + 665.0, + 1400.0, + 701.0, + 296.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 692.0, + 1402.0, + 692.0, + 1402.0, + 730.0, + 294.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 722.0, + 524.0, + 722.0, + 524.0, + 758.0, + 296.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 722.0, + 949.0, + 722.0, + 949.0, + 758.0, + 622.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 723.0, + 1205.0, + 723.0, + 1205.0, + 756.0, + 1039.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 723.0, + 1354.0, + 723.0, + 1354.0, + 756.0, + 1243.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 780.0, + 1402.0, + 780.0, + 1402.0, + 812.0, + 297.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 810.0, + 1351.0, + 810.0, + 1351.0, + 844.0, + 293.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1103.0, + 1403.0, + 1103.0, + 1403.0, + 1142.0, + 294.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1136.0, + 1405.0, + 1136.0, + 1405.0, + 1170.0, + 295.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1163.0, + 1406.0, + 1163.0, + 1406.0, + 1205.0, + 294.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1195.0, + 346.0, + 1195.0, + 346.0, + 1234.0, + 293.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 1195.0, + 1406.0, + 1195.0, + 1406.0, + 1234.0, + 372.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1224.0, + 1406.0, + 1224.0, + 1406.0, + 1265.0, + 291.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1254.0, + 1405.0, + 1254.0, + 1405.0, + 1295.0, + 293.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1288.0, + 445.0, + 1288.0, + 445.0, + 1323.0, + 294.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 1288.0, + 1405.0, + 1288.0, + 1405.0, + 1323.0, + 503.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1314.0, + 1406.0, + 1314.0, + 1406.0, + 1358.0, + 293.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1349.0, + 530.0, + 1349.0, + 530.0, + 1384.0, + 294.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 1349.0, + 1405.0, + 1349.0, + 1405.0, + 1384.0, + 549.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1379.0, + 296.0, + 1379.0, + 296.0, + 1415.0, + 291.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1379.0, + 401.0, + 1379.0, + 401.0, + 1415.0, + 350.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 1379.0, + 1020.0, + 1379.0, + 1020.0, + 1415.0, + 479.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 1379.0, + 1240.0, + 1379.0, + 1240.0, + 1415.0, + 1190.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1379.0, + 1406.0, + 1379.0, + 1406.0, + 1415.0, + 1403.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1410.0, + 1185.0, + 1410.0, + 1185.0, + 1445.0, + 294.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 906.0, + 740.0, + 906.0, + 740.0, + 942.0, + 320.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 906.0, + 1336.0, + 906.0, + 1336.0, + 942.0, + 763.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 906.0, + 1404.0, + 906.0, + 1404.0, + 942.0, + 1393.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 936.0, + 1406.0, + 936.0, + 1406.0, + 972.0, + 294.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 967.0, + 1404.0, + 967.0, + 1404.0, + 1001.0, + 293.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 994.0, + 393.0, + 994.0, + 393.0, + 1036.0, + 292.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 994.0, + 1405.0, + 994.0, + 1405.0, + 1036.0, + 416.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1027.0, + 297.0, + 1027.0, + 297.0, + 1062.0, + 294.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1027.0, + 606.0, + 1027.0, + 606.0, + 1062.0, + 321.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1027.0, + 1404.0, + 1027.0, + 1404.0, + 1062.0, + 628.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1061.0, + 1165.0, + 1061.0, + 1165.0, + 1093.0, + 296.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 389.0, + 1199.0, + 389.0, + 1199.0, + 419.0, + 856.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 389.0, + 1404.0, + 389.0, + 1404.0, + 419.0, + 1222.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 413.0, + 1404.0, + 413.0, + 1404.0, + 451.0, + 852.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 444.0, + 1402.0, + 444.0, + 1402.0, + 474.0, + 855.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 389.0, + 843.0, + 389.0, + 843.0, + 419.0, + 295.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 416.0, + 844.0, + 416.0, + 844.0, + 449.0, + 294.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 444.0, + 678.0, + 444.0, + 678.0, + 476.0, + 292.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1990.0, + 1403.0, + 1990.0, + 1403.0, + 2030.0, + 295.0, + 2030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 780.0, + 1402.0, + 780.0, + 1402.0, + 812.0, + 297.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 810.0, + 1351.0, + 810.0, + 1351.0, + 844.0, + 293.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1850.0, + 659.0, + 1850.0, + 659.0, + 1881.0, + 295.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1882.0, + 657.0, + 1882.0, + 657.0, + 1908.0, + 295.0, + 1908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1908.0, + 361.0, + 1908.0, + 361.0, + 1937.0, + 296.0, + 1937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 1908.0, + 658.0, + 1908.0, + 658.0, + 1937.0, + 385.0, + 1937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1933.0, + 475.0, + 1933.0, + 475.0, + 1968.0, + 294.0, + 1968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 227.0, + 1062.0, + 227.0, + 1062.0, + 260.0, + 891.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 227.0, + 1369.0, + 227.0, + 1369.0, + 260.0, + 1087.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 265.0, + 1002.0, + 265.0, + 1002.0, + 298.0, + 886.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 265.0, + 1344.0, + 265.0, + 1344.0, + 298.0, + 1028.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 303.0, + 1013.0, + 303.0, + 1013.0, + 334.0, + 890.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 303.0, + 1309.0, + 303.0, + 1309.0, + 334.0, + 1039.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 336.0, + 980.0, + 336.0, + 980.0, + 378.0, + 884.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 1854.0, + 1031.0, + 1854.0, + 1031.0, + 1885.0, + 667.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 1882.0, + 1032.0, + 1882.0, + 1032.0, + 1914.0, + 667.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1912.0, + 1030.0, + 1912.0, + 1030.0, + 1941.0, + 669.0, + 1941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1940.0, + 859.0, + 1940.0, + 859.0, + 1967.0, + 669.0, + 1967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 1851.0, + 1404.0, + 1851.0, + 1404.0, + 1881.0, + 1040.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1881.0, + 1404.0, + 1881.0, + 1404.0, + 1909.0, + 1041.0, + 1909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 1907.0, + 1404.0, + 1907.0, + 1404.0, + 1941.0, + 1042.0, + 1941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 1930.0, + 1137.0, + 1930.0, + 1137.0, + 1971.0, + 1040.0, + 1971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 227.0, + 707.0, + 227.0, + 707.0, + 258.0, + 395.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 264.0, + 747.0, + 264.0, + 747.0, + 298.0, + 395.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 302.0, + 655.0, + 302.0, + 655.0, + 333.0, + 395.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 337.0, + 494.0, + 337.0, + 494.0, + 373.0, + 390.0, + 373.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1133, + 1403, + 1133, + 1403, + 1501, + 298, + 1501 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 298, + 309, + 1405, + 309, + 1405, + 736, + 298, + 736 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 752, + 1404, + 752, + 1404, + 1118, + 298, + 1118 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1850, + 1403, + 1850, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1516, + 1403, + 1516, + 1403, + 1699, + 298, + 1699 + ], + "score": 0.979 + }, + { + "category_id": 0, + "poly": [ + 300, + 1767, + 543, + 1767, + 543, + 1802, + 300, + 1802 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 104, + 298, + 104 + ], + "score": 0.889 + }, + { + "category_id": 0, + "poly": [ + 301, + 225, + 586, + 225, + 586, + 261, + 301, + 261 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.778 + }, + { + "category_id": 13, + "poly": [ + 443, + 585, + 462, + 585, + 462, + 616, + 443, + 616 + ], + "score": 0.86, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 548, + 494, + 567, + 494, + 567, + 524, + 548, + 524 + ], + "score": 0.86, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1105, + 1440, + 1126, + 1440, + 1126, + 1471, + 1105, + 1471 + ], + "score": 0.86, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 455, + 1288, + 475, + 1288, + 475, + 1318, + 455, + 1318 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 716, + 1165, + 737, + 1165, + 737, + 1197, + 716, + 1197 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1138, + 1410, + 1160, + 1410, + 1160, + 1440, + 1138, + 1440 + ], + "score": 0.84, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 1325, + 1197, + 1346, + 1197, + 1346, + 1227, + 1325, + 1227 + ], + "score": 0.84, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1381, + 1548, + 1401, + 1548, + 1401, + 1579, + 1381, + 1579 + ], + "score": 0.84, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 577, + 1553, + 596, + 1553, + 596, + 1579, + 577, + 1579 + ], + "score": 0.79, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 1099, + 558, + 1118, + 558, + 1118, + 585, + 1099, + 585 + ], + "score": 0.74, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 454, + 646, + 478, + 646, + 478, + 672, + 454, + 672 + ], + "score": 0.74, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 999, + 904, + 1071, + 904, + 1071, + 933, + 999, + 933 + ], + "score": 0.36, + "latex": "\\delta I L P" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1761.0, + 549.0, + 1761.0, + 549.0, + 1812.0, + 291.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 223.0, + 592.0, + 223.0, + 592.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1131.0, + 1405.0, + 1131.0, + 1405.0, + 1172.0, + 293.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1163.0, + 715.0, + 1163.0, + 715.0, + 1201.0, + 293.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 1163.0, + 1407.0, + 1163.0, + 1407.0, + 1201.0, + 738.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1194.0, + 1324.0, + 1194.0, + 1324.0, + 1232.0, + 293.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1347.0, + 1194.0, + 1405.0, + 1194.0, + 1405.0, + 1232.0, + 1347.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1226.0, + 1404.0, + 1226.0, + 1404.0, + 1260.0, + 294.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1257.0, + 1405.0, + 1257.0, + 1405.0, + 1291.0, + 294.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1288.0, + 454.0, + 1288.0, + 454.0, + 1322.0, + 293.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 1288.0, + 1405.0, + 1288.0, + 1405.0, + 1322.0, + 476.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1318.0, + 1404.0, + 1318.0, + 1404.0, + 1352.0, + 294.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1347.0, + 1405.0, + 1347.0, + 1405.0, + 1382.0, + 293.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1377.0, + 1404.0, + 1377.0, + 1404.0, + 1416.0, + 292.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1406.0, + 1137.0, + 1406.0, + 1137.0, + 1446.0, + 292.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 1406.0, + 1407.0, + 1406.0, + 1407.0, + 1446.0, + 1161.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1436.0, + 1104.0, + 1436.0, + 1104.0, + 1474.0, + 293.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 1436.0, + 1405.0, + 1436.0, + 1405.0, + 1474.0, + 1127.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1468.0, + 1404.0, + 1468.0, + 1404.0, + 1503.0, + 293.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 306.0, + 1406.0, + 306.0, + 1406.0, + 348.0, + 292.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 340.0, + 1406.0, + 340.0, + 1406.0, + 377.0, + 292.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 370.0, + 1406.0, + 370.0, + 1406.0, + 405.0, + 293.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 403.0, + 1403.0, + 403.0, + 1403.0, + 435.0, + 296.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 431.0, + 1406.0, + 431.0, + 1406.0, + 467.0, + 295.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 463.0, + 1407.0, + 463.0, + 1407.0, + 498.0, + 295.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 491.0, + 547.0, + 491.0, + 547.0, + 528.0, + 293.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 491.0, + 1407.0, + 491.0, + 1407.0, + 528.0, + 568.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 521.0, + 1406.0, + 521.0, + 1406.0, + 559.0, + 291.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 550.0, + 1098.0, + 550.0, + 1098.0, + 592.0, + 291.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 550.0, + 1406.0, + 550.0, + 1406.0, + 592.0, + 1119.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 583.0, + 442.0, + 583.0, + 442.0, + 619.0, + 295.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 583.0, + 1405.0, + 583.0, + 1405.0, + 619.0, + 463.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 611.0, + 1406.0, + 611.0, + 1406.0, + 653.0, + 292.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 643.0, + 453.0, + 643.0, + 453.0, + 681.0, + 292.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 643.0, + 1403.0, + 643.0, + 1403.0, + 681.0, + 479.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 673.0, + 1406.0, + 673.0, + 1406.0, + 711.0, + 292.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 705.0, + 1102.0, + 705.0, + 1102.0, + 740.0, + 295.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 752.0, + 1405.0, + 752.0, + 1405.0, + 789.0, + 293.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 782.0, + 1404.0, + 782.0, + 1404.0, + 820.0, + 295.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 815.0, + 1404.0, + 815.0, + 1404.0, + 849.0, + 294.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 842.0, + 1405.0, + 842.0, + 1405.0, + 882.0, + 292.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 876.0, + 1405.0, + 876.0, + 1405.0, + 910.0, + 294.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 905.0, + 998.0, + 905.0, + 998.0, + 939.0, + 293.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 905.0, + 1405.0, + 905.0, + 1405.0, + 939.0, + 1072.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 936.0, + 1405.0, + 936.0, + 1405.0, + 970.0, + 292.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 965.0, + 1404.0, + 965.0, + 1404.0, + 999.0, + 293.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 996.0, + 1404.0, + 996.0, + 1404.0, + 1031.0, + 295.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1024.0, + 1405.0, + 1024.0, + 1405.0, + 1065.0, + 292.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1054.0, + 1404.0, + 1054.0, + 1404.0, + 1096.0, + 293.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1088.0, + 1358.0, + 1088.0, + 1358.0, + 1124.0, + 292.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1848.0, + 1404.0, + 1848.0, + 1404.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1917.0, + 294.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 1403.0, + 1942.0, + 1403.0, + 1977.0, + 293.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2000.0, + 879.0, + 2000.0, + 879.0, + 2040.0, + 293.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1515.0, + 1405.0, + 1515.0, + 1405.0, + 1551.0, + 294.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1546.0, + 576.0, + 1546.0, + 576.0, + 1580.0, + 293.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1546.0, + 1380.0, + 1546.0, + 1380.0, + 1580.0, + 597.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1577.0, + 1405.0, + 1577.0, + 1405.0, + 1613.0, + 294.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1608.0, + 1405.0, + 1608.0, + 1405.0, + 1643.0, + 294.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1638.0, + 1405.0, + 1638.0, + 1405.0, + 1673.0, + 294.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1668.0, + 1349.0, + 1668.0, + 1349.0, + 1704.0, + 296.0, + 1704.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.899 + }, + { + "category_id": 1, + "poly": [ + 293, + 1307, + 1404, + 1307, + 1404, + 1371, + 293, + 1371 + ], + "score": 0.839 + }, + { + "category_id": 1, + "poly": [ + 298, + 279, + 1404, + 279, + 1404, + 370, + 298, + 370 + ], + "score": 0.82 + }, + { + "category_id": 1, + "poly": [ + 294, + 1389, + 1402, + 1389, + 1402, + 1453, + 294, + 1453 + ], + "score": 0.817 + }, + { + "category_id": 1, + "poly": [ + 295, + 1808, + 1398, + 1808, + 1398, + 1871, + 295, + 1871 + ], + "score": 0.811 + }, + { + "category_id": 1, + "poly": [ + 295, + 1471, + 1369, + 1471, + 1369, + 1505, + 295, + 1505 + ], + "score": 0.81 + }, + { + "category_id": 1, + "poly": [ + 292, + 1226, + 1405, + 1226, + 1405, + 1290, + 292, + 1290 + ], + "score": 0.806 + }, + { + "category_id": 0, + "poly": [ + 300, + 228, + 488, + 228, + 488, + 261, + 300, + 261 + ], + "score": 0.801 + }, + { + "category_id": 1, + "poly": [ + 296, + 1726, + 1400, + 1726, + 1400, + 1790, + 296, + 1790 + ], + "score": 0.795 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.779 + }, + { + "category_id": 1, + "poly": [ + 298, + 390, + 1404, + 390, + 1404, + 513, + 298, + 513 + ], + "score": 0.762 + }, + { + "category_id": 1, + "poly": [ + 298, + 1889, + 1354, + 1889, + 1354, + 1923, + 298, + 1923 + ], + "score": 0.75 + }, + { + "category_id": 1, + "poly": [ + 301, + 1942, + 1402, + 1942, + 1402, + 2035, + 301, + 2035 + ], + "score": 0.732 + }, + { + "category_id": 1, + "poly": [ + 297, + 696, + 1405, + 696, + 1405, + 820, + 297, + 820 + ], + "score": 0.722 + }, + { + "category_id": 1, + "poly": [ + 294, + 838, + 1402, + 838, + 1402, + 903, + 294, + 903 + ], + "score": 0.713 + }, + { + "category_id": 1, + "poly": [ + 298, + 1143, + 1399, + 1143, + 1399, + 1208, + 298, + 1208 + ], + "score": 0.706 + }, + { + "category_id": 1, + "poly": [ + 293, + 613, + 1400, + 613, + 1400, + 678, + 293, + 678 + ], + "score": 0.701 + }, + { + "category_id": 1, + "poly": [ + 294, + 532, + 1401, + 532, + 1401, + 597, + 294, + 597 + ], + "score": 0.695 + }, + { + "category_id": 1, + "poly": [ + 298, + 1063, + 1398, + 1063, + 1398, + 1127, + 298, + 1127 + ], + "score": 0.633 + }, + { + "category_id": 1, + "poly": [ + 298, + 1521, + 1404, + 1521, + 1404, + 1708, + 298, + 1708 + ], + "score": 0.632 + }, + { + "category_id": 1, + "poly": [ + 297, + 920, + 1404, + 920, + 1404, + 1044, + 297, + 1044 + ], + "score": 0.596 + }, + { + "category_id": 13, + "poly": [ + 390, + 731, + 422, + 731, + 422, + 755, + 390, + 755 + ], + "score": 0.52, + "latex": "^ { + + }" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 491.0, + 227.0, + 491.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1307.0, + 1405.0, + 1307.0, + 1405.0, + 1343.0, + 295.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1336.0, + 439.0, + 1336.0, + 439.0, + 1374.0, + 321.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 279.0, + 1404.0, + 279.0, + 1404.0, + 313.0, + 294.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 309.0, + 1404.0, + 309.0, + 1404.0, + 342.0, + 322.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 334.0, + 438.0, + 334.0, + 438.0, + 373.0, + 320.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1388.0, + 1403.0, + 1388.0, + 1403.0, + 1424.0, + 296.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1421.0, + 1385.0, + 1421.0, + 1385.0, + 1454.0, + 325.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1806.0, + 1403.0, + 1806.0, + 1403.0, + 1846.0, + 294.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1840.0, + 1177.0, + 1840.0, + 1177.0, + 1871.0, + 321.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1468.0, + 1369.0, + 1468.0, + 1369.0, + 1508.0, + 295.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1225.0, + 1407.0, + 1225.0, + 1407.0, + 1263.0, + 294.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1255.0, + 1215.0, + 1255.0, + 1215.0, + 1291.0, + 322.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1726.0, + 1404.0, + 1726.0, + 1404.0, + 1762.0, + 295.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1757.0, + 878.0, + 1757.0, + 878.0, + 1790.0, + 321.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 389.0, + 1405.0, + 389.0, + 1405.0, + 426.0, + 293.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 421.0, + 1401.0, + 421.0, + 1401.0, + 454.0, + 323.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 450.0, + 1404.0, + 450.0, + 1404.0, + 489.0, + 321.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 483.0, + 435.0, + 483.0, + 435.0, + 513.0, + 322.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1886.0, + 1353.0, + 1886.0, + 1353.0, + 1926.0, + 295.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1939.0, + 1406.0, + 1939.0, + 1406.0, + 1979.0, + 294.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2007.0, + 324.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2004.0, + 1009.0, + 2004.0, + 1009.0, + 2036.0, + 322.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 697.0, + 1403.0, + 697.0, + 1403.0, + 730.0, + 295.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 725.0, + 389.0, + 725.0, + 389.0, + 763.0, + 323.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 725.0, + 1405.0, + 725.0, + 1405.0, + 763.0, + 423.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 757.0, + 1406.0, + 757.0, + 1406.0, + 794.0, + 320.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 790.0, + 906.0, + 790.0, + 906.0, + 819.0, + 324.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 834.0, + 1408.0, + 834.0, + 1408.0, + 876.0, + 293.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 869.0, + 738.0, + 869.0, + 738.0, + 904.0, + 323.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1145.0, + 1403.0, + 1145.0, + 1403.0, + 1181.0, + 296.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1175.0, + 696.0, + 1175.0, + 696.0, + 1208.0, + 327.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 611.0, + 1403.0, + 611.0, + 1403.0, + 652.0, + 293.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 646.0, + 819.0, + 646.0, + 819.0, + 676.0, + 323.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 528.0, + 1405.0, + 528.0, + 1405.0, + 571.0, + 293.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 563.0, + 940.0, + 563.0, + 940.0, + 597.0, + 321.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1059.0, + 1403.0, + 1059.0, + 1403.0, + 1101.0, + 292.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1095.0, + 702.0, + 1095.0, + 702.0, + 1128.0, + 322.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1522.0, + 1402.0, + 1522.0, + 1402.0, + 1558.0, + 296.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1550.0, + 1404.0, + 1550.0, + 1404.0, + 1593.0, + 320.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1582.0, + 1404.0, + 1582.0, + 1404.0, + 1621.0, + 321.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1613.0, + 1404.0, + 1613.0, + 1404.0, + 1650.0, + 320.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1641.0, + 1405.0, + 1641.0, + 1405.0, + 1682.0, + 320.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1675.0, + 904.0, + 1675.0, + 904.0, + 1707.0, + 322.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 920.0, + 1404.0, + 920.0, + 1404.0, + 956.0, + 295.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 951.0, + 1405.0, + 951.0, + 1405.0, + 985.0, + 320.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 982.0, + 1402.0, + 982.0, + 1402.0, + 1017.0, + 324.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1012.0, + 935.0, + 1012.0, + 935.0, + 1046.0, + 322.0, + 1046.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 106, + 298, + 106 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 866, + 2088, + 866, + 2113, + 836, + 2113 + ], + "score": 0.836 + }, + { + "category_id": 1, + "poly": [ + 296, + 738, + 1401, + 738, + 1401, + 803, + 296, + 803 + ], + "score": 0.697 + }, + { + "category_id": 1, + "poly": [ + 296, + 654, + 1400, + 654, + 1400, + 720, + 296, + 720 + ], + "score": 0.668 + }, + { + "category_id": 1, + "poly": [ + 296, + 821, + 1403, + 821, + 1403, + 915, + 296, + 915 + ], + "score": 0.657 + }, + { + "category_id": 1, + "poly": [ + 295, + 229, + 1401, + 229, + 1401, + 293, + 295, + 293 + ], + "score": 0.652 + }, + { + "category_id": 1, + "poly": [ + 290, + 1407, + 1404, + 1407, + 1404, + 1473, + 290, + 1473 + ], + "score": 0.652 + }, + { + "category_id": 1, + "poly": [ + 289, + 1103, + 1306, + 1103, + 1306, + 1139, + 289, + 1139 + ], + "score": 0.644 + }, + { + "category_id": 1, + "poly": [ + 297, + 1020, + 1401, + 1020, + 1401, + 1085, + 297, + 1085 + ], + "score": 0.642 + }, + { + "category_id": 1, + "poly": [ + 298, + 1240, + 1404, + 1240, + 1404, + 1305, + 298, + 1305 + ], + "score": 0.627 + }, + { + "category_id": 1, + "poly": [ + 303, + 395, + 1400, + 395, + 1400, + 491, + 303, + 491 + ], + "score": 0.615 + }, + { + "category_id": 1, + "poly": [ + 297, + 1157, + 1406, + 1157, + 1406, + 1222, + 297, + 1222 + ], + "score": 0.613 + }, + { + "category_id": 1, + "poly": [ + 295, + 1323, + 1404, + 1323, + 1404, + 1389, + 295, + 1389 + ], + "score": 0.611 + }, + { + "category_id": 1, + "poly": [ + 289, + 1492, + 1401, + 1492, + 1401, + 1556, + 289, + 1556 + ], + "score": 0.607 + }, + { + "category_id": 1, + "poly": [ + 297, + 936, + 1399, + 936, + 1399, + 1001, + 297, + 1001 + ], + "score": 0.594 + }, + { + "category_id": 1, + "poly": [ + 294, + 313, + 1402, + 313, + 1402, + 377, + 294, + 377 + ], + "score": 0.582 + }, + { + "category_id": 1, + "poly": [ + 287, + 1719, + 1402, + 1719, + 1402, + 1785, + 287, + 1785 + ], + "score": 0.553 + }, + { + "category_id": 1, + "poly": [ + 297, + 1971, + 1398, + 1971, + 1398, + 2034, + 297, + 2034 + ], + "score": 0.551 + }, + { + "category_id": 1, + "poly": [ + 294, + 1803, + 1401, + 1803, + 1401, + 1869, + 294, + 1869 + ], + "score": 0.549 + }, + { + "category_id": 1, + "poly": [ + 296, + 1887, + 1400, + 1887, + 1400, + 1952, + 296, + 1952 + ], + "score": 0.529 + }, + { + "category_id": 1, + "poly": [ + 296, + 510, + 1405, + 510, + 1405, + 636, + 296, + 636 + ], + "score": 0.507 + }, + { + "category_id": 1, + "poly": [ + 295, + 1574, + 1405, + 1574, + 1405, + 1699, + 295, + 1699 + ], + "score": 0.419 + }, + { + "category_id": 13, + "poly": [ + 1155, + 1408, + 1193, + 1408, + 1193, + 1438, + 1155, + 1438 + ], + "score": 0.73, + "latex": "\\langle 3 ^ { r d }" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 740.0, + 1403.0, + 740.0, + 1403.0, + 776.0, + 296.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 769.0, + 872.0, + 769.0, + 872.0, + 801.0, + 323.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 655.0, + 1404.0, + 655.0, + 1404.0, + 691.0, + 296.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 687.0, + 1098.0, + 687.0, + 1098.0, + 721.0, + 322.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 822.0, + 1404.0, + 822.0, + 1404.0, + 860.0, + 295.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 851.0, + 1406.0, + 851.0, + 1406.0, + 892.0, + 321.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 880.0, + 435.0, + 880.0, + 435.0, + 918.0, + 321.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 225.0, + 1406.0, + 225.0, + 1406.0, + 269.0, + 293.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 261.0, + 546.0, + 261.0, + 546.0, + 294.0, + 321.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1402.0, + 1154.0, + 1402.0, + 1154.0, + 1448.0, + 293.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 1402.0, + 1406.0, + 1402.0, + 1406.0, + 1448.0, + 1194.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1440.0, + 741.0, + 1440.0, + 741.0, + 1472.0, + 323.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1102.0, + 1308.0, + 1102.0, + 1308.0, + 1141.0, + 293.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1019.0, + 1403.0, + 1019.0, + 1403.0, + 1056.0, + 293.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1051.0, + 1115.0, + 1051.0, + 1115.0, + 1087.0, + 324.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1238.0, + 1405.0, + 1238.0, + 1405.0, + 1278.0, + 293.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1270.0, + 397.0, + 1270.0, + 397.0, + 1305.0, + 321.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 394.0, + 1404.0, + 394.0, + 1404.0, + 434.0, + 294.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 426.0, + 1405.0, + 426.0, + 1405.0, + 466.0, + 322.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 458.0, + 1404.0, + 458.0, + 1404.0, + 496.0, + 323.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1157.0, + 1404.0, + 1157.0, + 1404.0, + 1193.0, + 295.0, + 1193.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1188.0, + 668.0, + 1188.0, + 668.0, + 1224.0, + 324.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1324.0, + 1404.0, + 1324.0, + 1404.0, + 1360.0, + 294.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1356.0, + 642.0, + 1356.0, + 642.0, + 1388.0, + 327.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1491.0, + 1402.0, + 1491.0, + 1402.0, + 1527.0, + 295.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1523.0, + 823.0, + 1523.0, + 823.0, + 1555.0, + 322.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 936.0, + 1403.0, + 936.0, + 1403.0, + 972.0, + 295.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 967.0, + 706.0, + 967.0, + 706.0, + 1002.0, + 322.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 313.0, + 1404.0, + 313.0, + 1404.0, + 349.0, + 294.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 344.0, + 866.0, + 344.0, + 866.0, + 376.0, + 323.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1720.0, + 1404.0, + 1720.0, + 1404.0, + 1756.0, + 295.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1752.0, + 700.0, + 1752.0, + 700.0, + 1784.0, + 323.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1968.0, + 1402.0, + 1968.0, + 1402.0, + 2009.0, + 295.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2000.0, + 675.0, + 2000.0, + 675.0, + 2036.0, + 323.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1802.0, + 1403.0, + 1802.0, + 1403.0, + 1842.0, + 294.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1836.0, + 558.0, + 1836.0, + 558.0, + 1868.0, + 322.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1887.0, + 1404.0, + 1887.0, + 1404.0, + 1923.0, + 296.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1920.0, + 729.0, + 1920.0, + 729.0, + 1952.0, + 322.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 513.0, + 1405.0, + 513.0, + 1405.0, + 546.0, + 296.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 541.0, + 1406.0, + 541.0, + 1406.0, + 577.0, + 320.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 571.0, + 1407.0, + 571.0, + 1407.0, + 611.0, + 319.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 604.0, + 480.0, + 604.0, + 480.0, + 634.0, + 323.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1576.0, + 1406.0, + 1576.0, + 1406.0, + 1612.0, + 294.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1605.0, + 1405.0, + 1605.0, + 1405.0, + 1642.0, + 321.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1635.0, + 1406.0, + 1635.0, + 1406.0, + 1674.0, + 319.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1668.0, + 797.0, + 1668.0, + 797.0, + 1701.0, + 323.0, + 1701.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 229, + 1403, + 229, + 1403, + 384, + 297, + 384 + ], + "score": 0.899 + }, + { + "category_id": 1, + "poly": [ + 294, + 403, + 1399, + 403, + 1399, + 467, + 294, + 467 + ], + "score": 0.87 + }, + { + "category_id": 1, + "poly": [ + 294, + 653, + 1401, + 653, + 1401, + 716, + 294, + 716 + ], + "score": 0.869 + }, + { + "category_id": 1, + "poly": [ + 294, + 539, + 1403, + 539, + 1403, + 632, + 294, + 632 + ], + "score": 0.837 + }, + { + "category_id": 1, + "poly": [ + 294, + 486, + 1231, + 486, + 1231, + 519, + 294, + 519 + ], + "score": 0.835 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.831 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.798 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1405.0, + 228.0, + 1405.0, + 267.0, + 294.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 263.0, + 1404.0, + 263.0, + 1404.0, + 296.0, + 322.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 291.0, + 1404.0, + 291.0, + 1404.0, + 326.0, + 322.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 318.0, + 1405.0, + 318.0, + 1405.0, + 360.0, + 321.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 353.0, + 538.0, + 353.0, + 538.0, + 387.0, + 320.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 398.0, + 1405.0, + 398.0, + 1405.0, + 445.0, + 289.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 434.0, + 1021.0, + 434.0, + 1021.0, + 468.0, + 322.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 653.0, + 1404.0, + 653.0, + 1404.0, + 688.0, + 296.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 685.0, + 967.0, + 685.0, + 967.0, + 718.0, + 322.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 540.0, + 1405.0, + 540.0, + 1405.0, + 573.0, + 294.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 566.0, + 1404.0, + 566.0, + 1404.0, + 608.0, + 320.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 600.0, + 541.0, + 600.0, + 541.0, + 632.0, + 321.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 485.0, + 1232.0, + 485.0, + 1232.0, + 524.0, + 293.0, + 524.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 612, + 1405, + 612, + 1405, + 860, + 297, + 860 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 352, + 1405, + 352, + 1405, + 599, + 297, + 599 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 950, + 1403, + 950, + 1403, + 1110, + 297, + 1110 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1230, + 1404, + 1230, + 1404, + 1324, + 298, + 1324 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 295, + 1429, + 1405, + 1429, + 1405, + 1494, + 295, + 1494 + ], + "score": 0.952 + }, + { + "category_id": 0, + "poly": [ + 302, + 894, + 792, + 894, + 792, + 926, + 302, + 926 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1180, + 1400, + 1180, + 1400, + 1210, + 1366, + 1210 + ], + "score": 0.876 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1135, + 1400, + 1135, + 1400, + 1164, + 1366, + 1164 + ], + "score": 0.875 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1352, + 1400, + 1352, + 1400, + 1381, + 1366, + 1381 + ], + "score": 0.873 + }, + { + "category_id": 0, + "poly": [ + 301, + 294, + 699, + 294, + 699, + 328, + 301, + 328 + ], + "score": 0.869 + }, + { + "category_id": 8, + "poly": [ + 669, + 1129, + 1027, + 1129, + 1027, + 1168, + 669, + 1168 + ], + "score": 0.861 + }, + { + "category_id": 8, + "poly": [ + 671, + 1176, + 1033, + 1176, + 1033, + 1212, + 671, + 1212 + ], + "score": 0.848 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.841 + }, + { + "category_id": 8, + "poly": [ + 619, + 1340, + 1080, + 1340, + 1080, + 1410, + 619, + 1410 + ], + "score": 0.803 + }, + { + "category_id": 1, + "poly": [ + 302, + 226, + 601, + 226, + 601, + 262, + 302, + 262 + ], + "score": 0.782 + }, + { + "category_id": 8, + "poly": [ + 620, + 1340, + 1079, + 1340, + 1079, + 1410, + 620, + 1410 + ], + "score": 0.492 + }, + { + "category_id": 8, + "poly": [ + 667, + 1128, + 1031, + 1128, + 1031, + 1213, + 667, + 1213 + ], + "score": 0.166 + }, + { + "category_id": 0, + "poly": [ + 302, + 226, + 601, + 226, + 601, + 262, + 302, + 262 + ], + "score": 0.104 + }, + { + "category_id": 14, + "poly": [ + 620, + 1339, + 1080, + 1339, + 1080, + 1412, + 620, + 1412 + ], + "score": 0.94, + "latex": "h ^ { j + 1 } = \\sum _ { i } \\beta _ { i } ^ { j } \\operatorname { t a n h } ( W \\rho ( M _ { i } , h ^ { j } ) + b )" + }, + { + "category_id": 13, + "poly": [ + 867, + 1044, + 1033, + 1044, + 1033, + 1078, + 867, + 1078 + ], + "score": 0.94, + "latex": "{ \\pmb m } _ { \\ b q } = \\Phi _ { M } ( { \\pmb q } )" + }, + { + "category_id": 13, + "poly": [ + 347, + 825, + 528, + 825, + 528, + 864, + 347, + 864 + ], + "score": 0.93, + "latex": "\\phi _ { U } ( K _ { i j } ) = \\mathsf { H } _ { i j } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 519, + 1045, + 700, + 1045, + 700, + 1078, + 519, + 1078 + ], + "score": 0.92, + "latex": "M _ { i } = \\Phi _ { M } ( C _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 1193, + 382, + 1335, + 382, + 1335, + 416, + 1193, + 416 + ], + "score": 0.92, + "latex": "\\phi ( \\pmb { k } ) \\in \\mathbb { R } ^ { 6 4 }" + }, + { + "category_id": 13, + "poly": [ + 421, + 1076, + 531, + 1076, + 531, + 1112, + 421, + 1112 + ], + "score": 0.92, + "latex": "h ^ { 0 } = m _ { q }" + }, + { + "category_id": 13, + "poly": [ + 298, + 642, + 503, + 642, + 503, + 677, + 298, + 677 + ], + "score": 0.92, + "latex": "\\phi ( K ) \\in \\mathbb { R } ^ { 3 \\times 3 \\times 3 2 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 535, + 515, + 535, + 515, + 569, + 298, + 569 + ], + "score": 0.92, + "latex": "\\phi _ { U } ( s ) = W _ { U } \\Phi ( s )" + }, + { + "category_id": 13, + "poly": [ + 1020, + 445, + 1078, + 445, + 1078, + 477, + 1020, + 477 + ], + "score": 0.92, + "latex": "\\phi ( k )" + }, + { + "category_id": 13, + "poly": [ + 534, + 675, + 720, + 675, + 720, + 707, + 534, + 707 + ], + "score": 0.92, + "latex": "3 \\times 3 \\times ( 3 2 + 4 )" + }, + { + "category_id": 13, + "poly": [ + 1036, + 1428, + 1394, + 1428, + 1394, + 1464, + 1036, + 1464 + ], + "score": 0.92, + "latex": "f ( C , \\boldsymbol { q } ) = \\mathrm { s o f t m a x } ( W h ^ { J } + b )" + }, + { + "category_id": 13, + "poly": [ + 598, + 535, + 655, + 535, + 655, + 569, + 598, + 569 + ], + "score": 0.92, + "latex": "\\Phi ( s )" + }, + { + "category_id": 13, + "poly": [ + 297, + 763, + 445, + 763, + 445, + 794, + 297, + 794 + ], + "score": 0.91, + "latex": "\\boldsymbol { \\mathsf { H } } \\in \\mathbb { R } ^ { 3 \\times 3 \\times d }" + }, + { + "category_id": 14, + "poly": [ + 665, + 1124, + 1033, + 1124, + 1033, + 1219, + 665, + 1219 + ], + "score": 0.91, + "latex": "\\begin{array} { c } { { A _ { i } ^ { j } = \\operatorname { t a n h } ( W \\rho ( M _ { i } , h ^ { j } ) + b ) } } \\\\ { { \\beta ^ { j } = \\operatorname { s o f t m a x } ( W \\Phi _ { A } ( A ^ { j } ) + b ) } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 964, + 1230, + 1345, + 1230, + 1345, + 1265, + 964, + 1265 + ], + "score": 0.91, + "latex": "\\rho ( { \\pmb x } , { \\pmb y } ) = [ { \\pmb x } ; { \\pmb y } ; { \\pmb x } \\odot { \\pmb y } ; ( { \\pmb x } - { \\pmb y } ) ^ { 2 } ]" + }, + { + "category_id": 13, + "poly": [ + 297, + 476, + 455, + 476, + 455, + 505, + 297, + 505 + ], + "score": 0.91, + "latex": "6 4 + 4 = 6 8" + }, + { + "category_id": 13, + "poly": [ + 404, + 795, + 497, + 795, + 497, + 824, + 404, + 824 + ], + "score": 0.9, + "latex": "\\pmb { h } \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 1044, + 614, + 1137, + 614, + 1137, + 642, + 1044, + 642 + ], + "score": 0.9, + "latex": "d = 3 2" + }, + { + "category_id": 13, + "poly": [ + 349, + 384, + 440, + 384, + 440, + 412, + 349, + 412 + ], + "score": 0.9, + "latex": "d = 1 6" + }, + { + "category_id": 13, + "poly": [ + 372, + 1233, + 412, + 1233, + 412, + 1263, + 372, + 1263 + ], + "score": 0.89, + "latex": "\\Phi _ { A }" + }, + { + "category_id": 13, + "poly": [ + 803, + 384, + 874, + 384, + 874, + 412, + 803, + 412 + ], + "score": 0.89, + "latex": "l = 4" + }, + { + "category_id": 13, + "poly": [ + 591, + 736, + 654, + 736, + 654, + 765, + 591, + 765 + ], + "score": 0.89, + "latex": "5 \\times 5" + }, + { + "category_id": 13, + "poly": [ + 748, + 476, + 786, + 476, + 786, + 507, + 748, + 507 + ], + "score": 0.89, + "latex": "\\phi _ { U }" + }, + { + "category_id": 13, + "poly": [ + 1319, + 953, + 1402, + 953, + 1402, + 981, + 1319, + 981 + ], + "score": 0.88, + "latex": "d = 3 2" + }, + { + "category_id": 13, + "poly": [ + 1204, + 1264, + 1234, + 1264, + 1234, + 1295, + 1204, + 1295 + ], + "score": 0.87, + "latex": "\\beta ^ { j }" + }, + { + "category_id": 13, + "poly": [ + 606, + 825, + 637, + 825, + 637, + 854, + 606, + 854 + ], + "score": 0.86, + "latex": "\\boldsymbol { \\mathsf { H } } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 1039, + 797, + 1059, + 797, + 1059, + 829, + 1039, + 829 + ], + "score": 0.86, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 424, + 985, + 444, + 985, + 444, + 1015, + 424, + 1015 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 720, + 355, + 741, + 355, + 741, + 386, + 720, + 386 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 796, + 676, + 816, + 676, + 816, + 707, + 796, + 707 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 983, + 954, + 1002, + 954, + 1002, + 985, + 983, + 985 + ], + "score": 0.84, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 419, + 1433, + 440, + 1433, + 440, + 1458, + 419, + 1458 + ], + "score": 0.82, + "latex": "J" + }, + { + "category_id": 13, + "poly": [ + 531, + 1233, + 549, + 1233, + 549, + 1259, + 531, + 1259 + ], + "score": 0.82, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 298, + 1266, + 323, + 1266, + 323, + 1291, + 298, + 1291 + ], + "score": 0.82, + "latex": "\\odot" + }, + { + "category_id": 13, + "poly": [ + 1133, + 1015, + 1162, + 1015, + 1162, + 1042, + 1133, + 1042 + ], + "score": 0.81, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 736, + 1080, + 751, + 1080, + 751, + 1110, + 736, + 1110 + ], + "score": 0.81, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 1026, + 415, + 1058, + 415, + 1058, + 442, + 1026, + 442 + ], + "score": 0.81, + "latex": "2 d" + }, + { + "category_id": 13, + "poly": [ + 1249, + 676, + 1268, + 676, + 1268, + 702, + 1249, + 702 + ], + "score": 0.8, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 913, + 958, + 931, + 958, + 931, + 985, + 913, + 985 + ], + "score": 0.8, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 746, + 801, + 764, + 801, + 764, + 829, + 746, + 829 + ], + "score": 0.8, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 750, + 1463, + 767, + 1463, + 767, + 1489, + 750, + 1489 + ], + "score": 0.8, + "latex": "^ { b }" + }, + { + "category_id": 13, + "poly": [ + 1109, + 416, + 1128, + 416, + 1128, + 442, + 1109, + 442 + ], + "score": 0.8, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 603, + 1015, + 651, + 1015, + 651, + 1044, + 603, + 1044 + ], + "score": 0.78, + "latex": "\\Phi _ { M }" + }, + { + "category_id": 13, + "poly": [ + 1333, + 477, + 1352, + 477, + 1352, + 503, + 1333, + 503 + ], + "score": 0.77, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 962, + 767, + 981, + 767, + 981, + 793, + 962, + 793 + ], + "score": 0.77, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 1292, + 1019, + 1313, + 1019, + 1313, + 1046, + 1292, + 1046 + ], + "score": 0.76, + "latex": "\\pmb q" + }, + { + "category_id": 13, + "poly": [ + 1374, + 1015, + 1392, + 1015, + 1392, + 1041, + 1374, + 1041 + ], + "score": 0.75, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 742, + 1262, + 772, + 1262, + 772, + 1296, + 742, + 1296 + ], + "score": 0.64, + "latex": "[ ; ]" + }, + { + "category_id": 13, + "poly": [ + 523, + 1462, + 560, + 1462, + 560, + 1489, + 523, + 1489 + ], + "score": 0.63, + "latex": "W" + }, + { + "category_id": 13, + "poly": [ + 1377, + 765, + 1403, + 765, + 1403, + 794, + 1377, + 794 + ], + "score": 0.28, + "latex": "\\boldsymbol { \\mathsf { H } }" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 891.0, + 798.0, + 891.0, + 798.0, + 931.0, + 296.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 293.0, + 702.0, + 293.0, + 702.0, + 331.0, + 294.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 224.0, + 605.0, + 224.0, + 605.0, + 267.0, + 295.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 613.0, + 1043.0, + 613.0, + 1043.0, + 647.0, + 295.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 613.0, + 1405.0, + 613.0, + 1405.0, + 647.0, + 1138.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 633.0, + 297.0, + 633.0, + 297.0, + 683.0, + 291.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 633.0, + 1411.0, + 633.0, + 1411.0, + 683.0, + 504.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 673.0, + 533.0, + 673.0, + 533.0, + 708.0, + 292.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 673.0, + 795.0, + 673.0, + 795.0, + 708.0, + 721.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 673.0, + 1248.0, + 673.0, + 1248.0, + 708.0, + 817.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 673.0, + 1405.0, + 673.0, + 1405.0, + 708.0, + 1269.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 705.0, + 1405.0, + 705.0, + 1405.0, + 738.0, + 294.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 734.0, + 590.0, + 734.0, + 590.0, + 770.0, + 292.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 734.0, + 1405.0, + 734.0, + 1405.0, + 770.0, + 655.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 758.0, + 296.0, + 758.0, + 296.0, + 804.0, + 290.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 758.0, + 961.0, + 758.0, + 961.0, + 804.0, + 446.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 758.0, + 1376.0, + 758.0, + 1376.0, + 804.0, + 982.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 758.0, + 1407.0, + 758.0, + 1407.0, + 804.0, + 1404.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 792.0, + 403.0, + 792.0, + 403.0, + 833.0, + 291.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 792.0, + 745.0, + 792.0, + 745.0, + 833.0, + 498.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 792.0, + 1038.0, + 792.0, + 1038.0, + 833.0, + 765.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 792.0, + 1406.0, + 792.0, + 1406.0, + 833.0, + 1060.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 823.0, + 346.0, + 823.0, + 346.0, + 863.0, + 292.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 823.0, + 605.0, + 823.0, + 605.0, + 863.0, + 529.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 823.0, + 1177.0, + 823.0, + 1177.0, + 863.0, + 638.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 351.0, + 719.0, + 351.0, + 719.0, + 389.0, + 295.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 351.0, + 1408.0, + 351.0, + 1408.0, + 389.0, + 742.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 382.0, + 348.0, + 382.0, + 348.0, + 418.0, + 290.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 382.0, + 802.0, + 382.0, + 802.0, + 418.0, + 441.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 382.0, + 1192.0, + 382.0, + 1192.0, + 418.0, + 875.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1336.0, + 382.0, + 1407.0, + 382.0, + 1407.0, + 418.0, + 1336.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 410.0, + 1025.0, + 410.0, + 1025.0, + 451.0, + 291.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 410.0, + 1108.0, + 410.0, + 1108.0, + 451.0, + 1059.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 410.0, + 1406.0, + 410.0, + 1406.0, + 451.0, + 1129.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 445.0, + 1019.0, + 445.0, + 1019.0, + 479.0, + 295.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 445.0, + 1405.0, + 445.0, + 1405.0, + 479.0, + 1079.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 474.0, + 747.0, + 474.0, + 747.0, + 512.0, + 456.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 474.0, + 1332.0, + 474.0, + 1332.0, + 512.0, + 787.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 474.0, + 1406.0, + 474.0, + 1406.0, + 512.0, + 1353.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 502.0, + 1406.0, + 502.0, + 1406.0, + 543.0, + 292.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 533.0, + 297.0, + 533.0, + 297.0, + 572.0, + 294.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 533.0, + 597.0, + 533.0, + 597.0, + 572.0, + 516.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 533.0, + 1406.0, + 533.0, + 1406.0, + 572.0, + 656.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 567.0, + 950.0, + 567.0, + 950.0, + 601.0, + 295.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 950.0, + 912.0, + 950.0, + 912.0, + 988.0, + 296.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 950.0, + 982.0, + 950.0, + 982.0, + 988.0, + 932.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 950.0, + 1318.0, + 950.0, + 1318.0, + 988.0, + 1003.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 983.0, + 423.0, + 983.0, + 423.0, + 1016.0, + 296.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 983.0, + 1404.0, + 983.0, + 1404.0, + 1016.0, + 445.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1012.0, + 602.0, + 1012.0, + 602.0, + 1050.0, + 296.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 1012.0, + 1132.0, + 1012.0, + 1132.0, + 1050.0, + 652.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 1012.0, + 1291.0, + 1012.0, + 1291.0, + 1050.0, + 1163.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 1012.0, + 1373.0, + 1012.0, + 1373.0, + 1050.0, + 1314.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 1012.0, + 1405.0, + 1012.0, + 1405.0, + 1050.0, + 1393.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1039.0, + 518.0, + 1039.0, + 518.0, + 1084.0, + 292.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 1039.0, + 866.0, + 1039.0, + 866.0, + 1084.0, + 701.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 1039.0, + 1405.0, + 1039.0, + 1405.0, + 1084.0, + 1034.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1072.0, + 420.0, + 1072.0, + 420.0, + 1117.0, + 292.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 1072.0, + 735.0, + 1072.0, + 735.0, + 1117.0, + 532.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 1072.0, + 765.0, + 1072.0, + 765.0, + 1117.0, + 752.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1227.0, + 371.0, + 1227.0, + 371.0, + 1267.0, + 292.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 1227.0, + 530.0, + 1227.0, + 530.0, + 1267.0, + 413.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 1227.0, + 963.0, + 1227.0, + 963.0, + 1267.0, + 550.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1346.0, + 1227.0, + 1405.0, + 1227.0, + 1405.0, + 1267.0, + 1346.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1262.0, + 741.0, + 1262.0, + 741.0, + 1296.0, + 324.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 1262.0, + 1203.0, + 1262.0, + 1203.0, + 1296.0, + 773.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 1262.0, + 1405.0, + 1262.0, + 1405.0, + 1296.0, + 1235.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1294.0, + 949.0, + 1294.0, + 949.0, + 1324.0, + 296.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1425.0, + 418.0, + 1425.0, + 418.0, + 1465.0, + 294.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 1425.0, + 1035.0, + 1425.0, + 1035.0, + 1465.0, + 441.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 1425.0, + 1403.0, + 1425.0, + 1403.0, + 1465.0, + 1395.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1460.0, + 522.0, + 1460.0, + 522.0, + 1496.0, + 295.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 1460.0, + 749.0, + 1460.0, + 749.0, + 1496.0, + 561.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1460.0, + 1277.0, + 1460.0, + 1277.0, + 1496.0, + 768.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 224.0, + 605.0, + 224.0, + 605.0, + 267.0, + 295.0, + 267.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 640, + 1223, + 1058, + 1223, + 1058, + 1412, + 640, + 1412 + ], + "score": 0.972, + "html": "
Task SequencesGrid
i704.7 ± 12.825.6 ± 1.8
ii709.4 ± 13.8623.7 ± 14.1
ii709.7 ± 14.0768.2 ±12.5
iv624.8 ± 12.4795.2 ± 10.3
" + }, + { + "category_id": 5, + "poly": [ + 536, + 340, + 1163, + 340, + 1163, + 1098, + 536, + 1098 + ], + "score": 0.968, + "html": "
DatasetTaskContextQueryAnswer
Sequencei1488constant2
Sequenceii6157head6
Sequenceii1837tail7
Sequenceiv3563duplicate3
Gridi００８ 022 022box2
Gridii4 -070 0 页 0 8 1head4
Gridii0 673 0 1 2 0 0centre7
Gridiv8 0 0 2 60 41corner2
" + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.842 + }, + { + "category_id": 0, + "poly": [ + 299, + 225, + 815, + 225, + 815, + 262, + 299, + 262 + ], + "score": 0.801 + }, + { + "category_id": 6, + "poly": [ + 393, + 283, + 1304, + 283, + 1304, + 319, + 393, + 319 + ], + "score": 0.611 + }, + { + "category_id": 6, + "poly": [ + 296, + 1105, + 1405, + 1105, + 1405, + 1199, + 296, + 1199 + ], + "score": 0.505 + }, + { + "category_id": 6, + "poly": [ + 396, + 226, + 1302, + 226, + 1302, + 319, + 396, + 319 + ], + "score": 0.213 + }, + { + "category_id": 7, + "poly": [ + 296, + 1105, + 1405, + 1105, + 1405, + 1199, + 296, + 1199 + ], + "score": 0.121 + }, + { + "category_id": 13, + "poly": [ + 484, + 1167, + 550, + 1167, + 550, + 1196, + 484, + 1196 + ], + "score": 0.89, + "latex": "3 \\times 3" + }, + { + "category_id": 13, + "poly": [ + 298, + 1167, + 362, + 1167, + 362, + 1195, + 298, + 1195 + ], + "score": 0.89, + "latex": "2 \\times 2" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 221.0, + 819.0, + 221.0, + 819.0, + 267.0, + 291.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 282.0, + 1305.0, + 282.0, + 1305.0, + 321.0, + 392.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1102.0, + 1403.0, + 1102.0, + 1403.0, + 1141.0, + 293.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1136.0, + 1406.0, + 1136.0, + 1406.0, + 1168.0, + 293.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1167.0, + 297.0, + 1167.0, + 297.0, + 1200.0, + 294.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1167.0, + 483.0, + 1167.0, + 483.0, + 1200.0, + 363.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 1167.0, + 868.0, + 1167.0, + 868.0, + 1200.0, + 551.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 228.0, + 818.0, + 228.0, + 818.0, + 263.0, + 390.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 282.0, + 1304.0, + 282.0, + 1304.0, + 321.0, + 392.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1102.0, + 1403.0, + 1102.0, + 1403.0, + 1141.0, + 293.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1136.0, + 1406.0, + 1136.0, + 1406.0, + 1168.0, + 293.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1167.0, + 297.0, + 1167.0, + 297.0, + 1200.0, + 294.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1167.0, + 483.0, + 1167.0, + 483.0, + 1200.0, + 363.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 1167.0, + 868.0, + 1167.0, + 868.0, + 1200.0, + 551.0, + 1200.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 300, + 538, + 1402, + 538, + 1402, + 700, + 300, + 700 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 837, + 1402, + 837, + 1402, + 962, + 299, + 962 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 352, + 1405, + 352, + 1405, + 447, + 298, + 447 + ], + "score": 0.975 + }, + { + "category_id": 8, + "poly": [ + 374, + 716, + 1320, + 716, + 1320, + 823, + 374, + 823 + ], + "score": 0.942 + }, + { + "category_id": 0, + "poly": [ + 302, + 482, + 792, + 482, + 792, + 515, + 302, + 515 + ], + "score": 0.911 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 858, + 74, + 858, + 105, + 298, + 105 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 299, + 224, + 636, + 224, + 636, + 263, + 299, + 263 + ], + "score": 0.904 + }, + { + "category_id": 0, + "poly": [ + 301, + 294, + 697, + 294, + 697, + 328, + 301, + 328 + ], + "score": 0.89 + }, + { + "category_id": 9, + "poly": [ + 1366, + 755, + 1400, + 755, + 1400, + 784, + 1366, + 784 + ], + "score": 0.858 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.845 + }, + { + "category_id": 8, + "poly": [ + 367, + 716, + 1319, + 716, + 1319, + 824, + 367, + 824 + ], + "score": 0.09 + }, + { + "category_id": 14, + "poly": [ + 370, + 714, + 1328, + 714, + 1328, + 826, + 370, + 826 + ], + "score": 0.93, + "latex": "J = \\mathcal { L } _ { \\mathrm { n l l } } ( f ( K ) , a ) + \\lambda _ { U } \\left[ \\mathcal { L } _ { \\mathrm { n l l } } ( f \\circ g ( I , K ) , a ) + \\frac { 1 } { J } \\sum _ { j = 1 } ^ { J } ( h _ { I } ^ { j } - h _ { K } ^ { j } ) ^ { 2 } + \\tau \\sum _ { s \\in \\mathcal { S } } \\psi _ { w } ( s ) \\right]" + }, + { + "category_id": 13, + "poly": [ + 849, + 662, + 892, + 662, + 892, + 701, + 849, + 701 + ], + "score": 0.92, + "latex": "h _ { K } ^ { j }" + }, + { + "category_id": 13, + "poly": [ + 559, + 839, + 650, + 839, + 650, + 869, + 559, + 869 + ], + "score": 0.91, + "latex": "\\lambda _ { U } = 0" + }, + { + "category_id": 13, + "poly": [ + 953, + 839, + 1043, + 839, + 1043, + 869, + 953, + 869 + ], + "score": 0.91, + "latex": "\\lambda _ { U } = 1" + }, + { + "category_id": 13, + "poly": [ + 764, + 662, + 798, + 662, + 798, + 701, + 764, + 701 + ], + "score": 0.9, + "latex": "h _ { I } ^ { j }" + }, + { + "category_id": 13, + "poly": [ + 971, + 633, + 1016, + 633, + 1016, + 663, + 971, + 663 + ], + "score": 0.89, + "latex": "{ \\mathcal L } _ { \\mathrm { n l l } }" + }, + { + "category_id": 13, + "poly": [ + 709, + 870, + 753, + 870, + 753, + 899, + 709, + 899 + ], + "score": 0.89, + "latex": "{ \\mathcal { L } } _ { \\mathrm { n l l } }" + }, + { + "category_id": 13, + "poly": [ + 1036, + 868, + 1067, + 868, + 1067, + 901, + 1036, + 901 + ], + "score": 0.88, + "latex": "\\beta ^ { j }" + }, + { + "category_id": 13, + "poly": [ + 1030, + 542, + 1050, + 542, + 1050, + 572, + 1030, + 572 + ], + "score": 0.87, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 797, + 542, + 816, + 542, + 816, + 572, + 797, + 572 + ], + "score": 0.86, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1172, + 573, + 1191, + 573, + 1191, + 604, + 1172, + 604 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1206, + 546, + 1224, + 546, + 1224, + 572, + 1206, + 572 + ], + "score": 0.8, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 1037, + 577, + 1056, + 577, + 1056, + 604, + 1037, + 604 + ], + "score": 0.79, + "latex": "g" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 480.0, + 797.0, + 480.0, + 797.0, + 519.0, + 295.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 220.0, + 640.0, + 220.0, + 640.0, + 270.0, + 294.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 292.0, + 701.0, + 292.0, + 701.0, + 332.0, + 295.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 538.0, + 796.0, + 538.0, + 796.0, + 576.0, + 294.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 538.0, + 1029.0, + 538.0, + 1029.0, + 576.0, + 817.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 538.0, + 1205.0, + 538.0, + 1205.0, + 576.0, + 1051.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 538.0, + 1404.0, + 538.0, + 1404.0, + 576.0, + 1225.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 569.0, + 1036.0, + 569.0, + 1036.0, + 607.0, + 294.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 569.0, + 1171.0, + 569.0, + 1171.0, + 607.0, + 1057.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 569.0, + 1407.0, + 569.0, + 1407.0, + 607.0, + 1192.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 601.0, + 1406.0, + 601.0, + 1406.0, + 638.0, + 295.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 633.0, + 970.0, + 633.0, + 970.0, + 666.0, + 296.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 633.0, + 1404.0, + 633.0, + 1404.0, + 666.0, + 1017.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 666.0, + 763.0, + 666.0, + 763.0, + 702.0, + 294.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 666.0, + 848.0, + 666.0, + 848.0, + 702.0, + 799.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 666.0, + 1307.0, + 666.0, + 1307.0, + 702.0, + 893.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 837.0, + 558.0, + 837.0, + 558.0, + 874.0, + 295.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 837.0, + 952.0, + 837.0, + 952.0, + 874.0, + 651.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 837.0, + 1406.0, + 837.0, + 1406.0, + 874.0, + 1044.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 868.0, + 708.0, + 868.0, + 708.0, + 904.0, + 294.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 868.0, + 1035.0, + 868.0, + 1035.0, + 904.0, + 754.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 868.0, + 1407.0, + 868.0, + 1407.0, + 904.0, + 1068.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 901.0, + 1406.0, + 901.0, + 1406.0, + 931.0, + 294.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 930.0, + 1348.0, + 930.0, + 1348.0, + 964.0, + 293.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 354.0, + 1403.0, + 354.0, + 1403.0, + 385.0, + 296.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 382.0, + 1405.0, + 382.0, + 1405.0, + 418.0, + 293.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 415.0, + 918.0, + 415.0, + 918.0, + 447.0, + 294.0, + 447.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 583, + 945, + 1113, + 945, + 1113, + 1716, + 583, + 1716 + ], + "score": 0.979, + "html": "
SupervisionWeakStrong
#Invs Training Size1 1k3 1k1 1k3 1k3 50
10.00.00.00.01.4
265.663.10.30.730.0
367.162.61.02.439.8
40.00.00.00.037.0
53.44.00.81.126.5
60.20.60.00.018.4
722.022.810.711.322.8
810.38.57.47.624.7
90.125.70.00.033.8
100.12.00.00.332.6
110.00.00.00.011.9
120.00.10.00.021.3
132.13.70.00.15.8
1419.713.50.50.154.8
150.00.70.00.00.0
1655.256.20.00.039.7
1739.249.051.149.348.8
184.48.00.60.510.4
1991.889.653.946.790.2
200.00.00.00.02.7
Mean19.120.56.36.027.6
Std27.927.015.614.321.0
#>5%8104417
" + }, + { + "category_id": 3, + "poly": [ + 357, + 285, + 1343, + 285, + 1343, + 726, + 357, + 726 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 295, + 753, + 1408, + 753, + 1408, + 878, + 295, + 878 + ], + "score": 0.925 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 858, + 76, + 858, + 104, + 298, + 104 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.847 + }, + { + "category_id": 6, + "poly": [ + 377, + 887, + 1318, + 887, + 1318, + 921, + 377, + 921 + ], + "score": 0.82 + }, + { + "category_id": 0, + "poly": [ + 298, + 225, + 632, + 225, + 632, + 262, + 298, + 262 + ], + "score": 0.756 + }, + { + "category_id": 13, + "poly": [ + 1247, + 818, + 1270, + 818, + 1270, + 849, + 1247, + 849 + ], + "score": 0.83, + "latex": "\\psi" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 295.0, + 628.0, + 295.0, + 628.0, + 310.0, + 597.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 295.0, + 1054.0, + 295.0, + 1054.0, + 309.0, + 1022.0, + 309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 305.0, + 412.0, + 305.0, + 412.0, + 324.0, + 378.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 366.0, + 407.0, + 366.0, + 407.0, + 386.0, + 377.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 428.0, + 407.0, + 428.0, + 407.0, + 448.0, + 377.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 448.0, + 405.0, + 448.0, + 405.0, + 549.0, + 360.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1248.0, + 480.0, + 1339.0, + 480.0, + 1339.0, + 502.0, + 1248.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 498.0, + 1338.0, + 498.0, + 1338.0, + 518.0, + 1281.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 517.0, + 1322.0, + 517.0, + 1322.0, + 537.0, + 1283.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 551.0, + 407.0, + 551.0, + 407.0, + 572.0, + 377.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 613.0, + 407.0, + 613.0, + 407.0, + 633.0, + 377.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 676.0, + 405.0, + 676.0, + 405.0, + 693.0, + 379.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 693.0, + 467.0, + 693.0, + 467.0, + 703.0, + 456.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 689.0, + 569.0, + 689.0, + 569.0, + 705.0, + 557.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 689.0, + 671.0, + 689.0, + 671.0, + 705.0, + 659.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 693.0, + 771.0, + 693.0, + 771.0, + 703.0, + 762.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 690.0, + 994.0, + 690.0, + 994.0, + 705.0, + 980.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 690.0, + 1095.0, + 690.0, + 1095.0, + 704.0, + 1082.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 693.0, + 1195.0, + 693.0, + 1195.0, + 702.0, + 1187.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 706.0, + 649.0, + 706.0, + 649.0, + 724.0, + 579.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 706.0, + 1074.0, + 706.0, + 1074.0, + 724.0, + 1003.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 756.0, + 1405.0, + 756.0, + 1405.0, + 789.0, + 296.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 786.0, + 1405.0, + 786.0, + 1405.0, + 819.0, + 294.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 815.0, + 1246.0, + 815.0, + 1246.0, + 852.0, + 294.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 815.0, + 1405.0, + 815.0, + 1405.0, + 852.0, + 1271.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 847.0, + 719.0, + 847.0, + 719.0, + 880.0, + 296.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 70.0, + 859.0, + 70.0, + 859.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 884.0, + 1321.0, + 884.0, + 1321.0, + 924.0, + 376.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 222.0, + 635.0, + 222.0, + 635.0, + 268.0, + 294.0, + 268.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 322, + 460, + 1376, + 460, + 1376, + 1205, + 322, + 1205 + ], + "score": 0.985, + "html": "
Support Size ModelWeakStrong
GN2N1k EntNetQRNUMN10k DMN+ DNC1k MemNN UMN10k DMN
10.00.00.70.00.00.00.00.00.00.0
28.38.156.40.565.60.30.40.00.71.8
340.338.869.71.267.11.11.80.02.44.8
42.80.41.40.70.00.00.00.00.00.0
513.11.04.61.23.40.50.82.01.10.7
67.68.430.01.20.20.00.00.00.00.0
717.317.822.39.422.02.40.615.011.33.1
810.012.519.23.710.30.00.39.07.63.5
913.210.731.50.00.10.00.20.00.00.0
1015.116.515.60.00.10.00.22.00.32.5
110.90.08.00.00.00.00.00.00.00.1
120.20.00.80.00.00.00.00.00.00.0
130.40.09.00.32.10.00.10.00.10.2
141.71.262.93.819.70.20.41.00.10.0
150.00.057.80.00.00.00.00.00.00.0
161.30.153.253.455.245.355.10.00.00.6
1751.041.746.451.839.24.212.035.049.340.6
1811.19.28.88.84.42.10.85.00.54.7
1982.888.590.490.791.80.03.964.046.765.5
200.00.02.60.30.00.00.00.00.00.0
Mean
#>5%13.9 1112.7 1029.6 1511.3 519.1 82.8 13.8 26.7 46.0 46.4 2
" + }, + { + "category_id": 5, + "poly": [ + 303, + 1393, + 1400, + 1393, + 1400, + 1970, + 303, + 1970 + ], + "score": 0.984, + "html": "
Size1k10k5020k
SupportWeakStrongWeakWeak
Arity1111Strong 22
# Invs /Model12 312 312 312 132 3DMN2 IMA
Facts1.20.90.00.40.00.00.00.00.033.50.00.0
Unification0.010.30.010.80.00.00.00.00.041.313.010.0
1 Step50.349.84.420.01.227.80.11.35.750.226.02.0
2 Steps47.550.05.735.037.247.80.029.728.749.933.05.0
3 Steps47.649.210.438.739.645.60.026.026.148.323.06.0
AND31.337.410.716.429.829.00.20.41.250.020.05.0
OR25.238.121.035.020.530.24.420.617.447.613.03.0
Transitivity50.026.639.65.06.049.250.050.0
1 Step NBF46.438.73.828.81.121.60.11.18.047.621.02.0
2 Steps NBF48.548.97.739.630.448.20.133.428.750.315.04.0
AND NBF51.050.143.148.629.444.20.11.340.149.516.08.0
OR NBF51.448.450.847.347.647.821.327.630.547.325.014.0
Mean36.439.314.328.921.531.82.412.216.047.121.29.1
Std18.715.916.414.117.116.76.113.213.64.712.313.4
#>5%91171171015912115
" + }, + { + "category_id": 6, + "poly": [ + 294, + 280, + 1405, + 280, + 1405, + 437, + 294, + 437 + ], + "score": 0.981 + }, + { + "category_id": 6, + "poly": [ + 449, + 1337, + 1248, + 1337, + 1248, + 1371, + 449, + 1371 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.86 + }, + { + "category_id": 13, + "poly": [ + 841, + 1339, + 884, + 1339, + 884, + 1368, + 841, + 1368 + ], + "score": 0.75, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 878, + 284, + 922, + 284, + 922, + 313, + 878, + 313 + ], + "score": 0.72, + "latex": "( \\% )" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 281.0, + 877.0, + 281.0, + 877.0, + 314.0, + 296.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 281.0, + 1405.0, + 281.0, + 1405.0, + 314.0, + 923.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 313.0, + 1405.0, + 313.0, + 1405.0, + 346.0, + 296.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 344.0, + 1402.0, + 344.0, + 1402.0, + 377.0, + 296.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 372.0, + 1405.0, + 372.0, + 1405.0, + 409.0, + 294.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 403.0, + 761.0, + 403.0, + 761.0, + 440.0, + 293.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1332.0, + 840.0, + 1332.0, + 840.0, + 1375.0, + 447.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 1332.0, + 1250.0, + 1332.0, + 1250.0, + 1375.0, + 885.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 871.0, + 2085.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 895, + 1406, + 895, + 1406, + 1019, + 297, + 1019 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 298, + 599, + 1402, + 599, + 1402, + 691, + 298, + 691 + ], + "score": 0.956 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 105, + 297, + 105 + ], + "score": 0.907 + }, + { + "category_id": 1, + "poly": [ + 361, + 371, + 1336, + 371, + 1336, + 405, + 361, + 405 + ], + "score": 0.895 + }, + { + "category_id": 1, + "poly": [ + 467, + 702, + 674, + 702, + 674, + 827, + 467, + 827 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.848 + }, + { + "category_id": 8, + "poly": [ + 589, + 418, + 1112, + 418, + 1112, + 576, + 589, + 576 + ], + "score": 0.829 + }, + { + "category_id": 4, + "poly": [ + 1040, + 1306, + 1402, + 1306, + 1402, + 1391, + 1040, + 1391 + ], + "score": 0.816 + }, + { + "category_id": 1, + "poly": [ + 324, + 1443, + 1371, + 1443, + 1371, + 1477, + 324, + 1477 + ], + "score": 0.811 + }, + { + "category_id": 1, + "poly": [ + 296, + 1306, + 656, + 1306, + 656, + 1418, + 296, + 1418 + ], + "score": 0.715 + }, + { + "category_id": 1, + "poly": [ + 290, + 839, + 1370, + 839, + 1370, + 873, + 290, + 873 + ], + "score": 0.629 + }, + { + "category_id": 3, + "poly": [ + 327, + 1040, + 1380, + 1040, + 1380, + 1283, + 327, + 1283 + ], + "score": 0.5 + }, + { + "category_id": 8, + "poly": [ + 983, + 700, + 1286, + 700, + 1286, + 827, + 983, + 827 + ], + "score": 0.426 + }, + { + "category_id": 5, + "poly": [ + 623, + 229, + 1077, + 229, + 1077, + 346, + 623, + 346 + ], + "score": 0.267, + "html": "
Xsandra went back to the Y:bathroom
is X:sandra in the Y:bathroom
yes
" + }, + { + "category_id": 1, + "poly": [ + 669, + 1306, + 1028, + 1306, + 1028, + 1390, + 669, + 1390 + ], + "score": 0.191 + }, + { + "category_id": 1, + "poly": [ + 589, + 418, + 1112, + 418, + 1112, + 576, + 589, + 576 + ], + "score": 0.162 + }, + { + "category_id": 6, + "poly": [ + 634, + 230, + 1066, + 230, + 1066, + 261, + 634, + 261 + ], + "score": 0.159 + }, + { + "category_id": 3, + "poly": [ + 1059, + 1058, + 1374, + 1058, + 1374, + 1276, + 1059, + 1276 + ], + "score": 0.131 + }, + { + "category_id": 4, + "poly": [ + 669, + 1306, + 1028, + 1306, + 1028, + 1390, + 669, + 1390 + ], + "score": 0.116 + }, + { + "category_id": 8, + "poly": [ + 623, + 229, + 1077, + 229, + 1077, + 346, + 623, + 346 + ], + "score": 0.11 + }, + { + "category_id": 14, + "poly": [ + 586, + 414, + 1112, + 414, + 1112, + 576, + 586, + 576 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } { X { \\mathrm { : m ~ ( ~ Y : e ~ ) ~ } } } & { { } { \\vdash X { \\mathrm { : m ~ ( ~ Y : e ~ ) ~ } } } } \\\\ { X { \\mathrm { : a ~ ( ~ } Y { \\mathrm { : w ~ , ~ } } { \\mathrm { ~ Z : e ~ ) ~ } } } } & { { } { \\vdash X { \\mathrm { : a ~ ( ~ } Y { \\mathrm { : w ~ , ~ } } { \\mathrm { Z : e ~ ) ~ } } } } } \\\\ { X { \\mathrm { : m ~ ( ~ T ~ ) ~ } } } & { { } { \\vdash X { \\mathrm { : m ~ ( ~ c ~ ) ~ } } } } \\\\ { X { \\mathrm { : x ~ ( ~ \\mathbb { A } ~ ) n o t ~ \\mathbf { Z } { \\mathrm { : q } } ~ ( ~ \\mathbb { A } ~ ) ~ } } } & { { } { \\vdash X { \\mathrm { : x ~ ( ~ } Y { \\mathrm { : z } } ~ ) } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 448, + 927, + 537, + 927, + 537, + 958, + 448, + 958 + ], + "score": 0.85, + "latex": "\\leq 1 0 0 0" + }, + { + "category_id": 14, + "poly": [ + 975, + 697, + 1292, + 697, + 1292, + 826, + 975, + 826 + ], + "score": 0.27, + "latex": "\\begin{array} { r l } { X \\mathrm { ~ 0 ~ ~ 0 ~ ~ } } & { \\mathrm { ~ 0 ~ ~ Y ~ 0 ~ ~ 6 ~ 4 ~ } Y } \\\\ { \\mathrm { ~ 0 ~ 1 ~ 0 ~ } } & { \\mathrm { ~ 1 ~ } X \\mathrm { ~ 8 ~ } } \\\\ { \\mathrm { ~ 0 ~ 0 ~ } Y } & { \\mathrm { ~ 0 ~ 2 ~ 0 ~ } } \\end{array}" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 1305.0, + 1402.0, + 1305.0, + 1402.0, + 1336.0, + 1043.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 1330.0, + 1404.0, + 1330.0, + 1404.0, + 1363.0, + 1042.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 1359.0, + 1370.0, + 1359.0, + 1370.0, + 1394.0, + 1042.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 1081.0, + 460.0, + 1081.0, + 460.0, + 1086.0, + 454.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 1048.0, + 517.0, + 1048.0, + 517.0, + 1078.0, + 484.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 1078.0, + 551.0, + 1078.0, + 551.0, + 1089.0, + 541.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1078.0, + 595.0, + 1078.0, + 595.0, + 1089.0, + 587.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 1071.0, + 893.0, + 1071.0, + 893.0, + 1086.0, + 879.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1065.0, + 1227.0, + 1065.0, + 1227.0, + 1085.0, + 1205.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 1065.0, + 1316.0, + 1065.0, + 1316.0, + 1084.0, + 1293.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1086.0, + 1221.0, + 1086.0, + 1221.0, + 1096.0, + 1211.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1106.0, + 769.0, + 1106.0, + 769.0, + 1133.0, + 735.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 1153.0, + 778.0, + 1153.0, + 778.0, + 1176.0, + 751.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1192.0, + 779.0, + 1192.0, + 779.0, + 1222.0, + 735.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 1191.0, + 1109.0, + 1191.0, + 1109.0, + 1221.0, + 1062.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1045.0, + 461.0, + 1045.0, + 461.0, + 1080.5, + 451.0, + 1080.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.75, + 1049.5, + 436.75, + 1049.5, + 436.75, + 1067.5, + 385.75, + 1067.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 1051.5, + 568.0, + 1051.5, + 568.0, + 1068.5, + 520.0, + 1068.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 1049.0, + 614.0, + 1049.0, + 614.0, + 1066.0, + 563.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 1103.0, + 1099.0, + 1103.0, + 1099.0, + 1135.0, + 1062.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 1102.5, + 399.0, + 1102.5, + 399.0, + 1125.5, + 336.0, + 1125.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.75, + 1137.5, + 400.75, + 1137.5, + 400.75, + 1170.0, + 338.75, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.25, + 1153.5, + 1097.25, + 1153.5, + 1097.25, + 1173.5, + 1086.25, + 1173.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 1192.0, + 387.0, + 1192.0, + 387.0, + 1217.5, + 342.0, + 1217.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1243.0, + 400.0, + 1243.0, + 400.0, + 1266.0, + 322.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 228.0, + 1063.0, + 228.0, + 1063.0, + 265.0, + 635.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 1059.0, + 1325.0, + 1059.0, + 1325.0, + 1101.0, + 1192.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 1099.0, + 1110.0, + 1099.0, + 1110.0, + 1142.0, + 1058.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 1149.0, + 1106.0, + 1149.0, + 1106.0, + 1177.0, + 1081.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 1190.0, + 1112.0, + 1190.0, + 1112.0, + 1223.0, + 1062.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 1306.0, + 1030.0, + 1306.0, + 1030.0, + 1336.0, + 670.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1335.0, + 1029.0, + 1335.0, + 1029.0, + 1361.0, + 669.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1361.0, + 991.0, + 1361.0, + 991.0, + 1391.0, + 669.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 896.0, + 1404.0, + 896.0, + 1404.0, + 929.0, + 295.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 927.0, + 447.0, + 927.0, + 447.0, + 959.0, + 296.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 927.0, + 1406.0, + 927.0, + 1406.0, + 959.0, + 538.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 958.0, + 1406.0, + 958.0, + 1406.0, + 992.0, + 292.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 986.0, + 733.0, + 986.0, + 733.0, + 1022.0, + 294.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 597.0, + 1404.0, + 597.0, + 1404.0, + 635.0, + 293.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 628.0, + 1404.0, + 628.0, + 1404.0, + 667.0, + 293.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 662.0, + 454.0, + 662.0, + 454.0, + 693.0, + 294.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 370.0, + 1342.0, + 370.0, + 1342.0, + 409.0, + 358.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 699.0, + 629.0, + 699.0, + 629.0, + 734.0, + 466.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 730.0, + 629.0, + 730.0, + 629.0, + 766.0, + 465.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 760.0, + 614.0, + 760.0, + 614.0, + 796.0, + 464.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 792.0, + 677.0, + 792.0, + 677.0, + 830.0, + 460.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1440.0, + 1373.0, + 1440.0, + 1373.0, + 1481.0, + 324.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1303.0, + 658.0, + 1303.0, + 658.0, + 1334.0, + 294.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1328.0, + 659.0, + 1328.0, + 659.0, + 1366.0, + 294.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1358.0, + 658.0, + 1358.0, + 658.0, + 1390.0, + 295.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1386.0, + 620.0, + 1386.0, + 620.0, + 1419.0, + 296.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 836.0, + 1374.0, + 836.0, + 1374.0, + 874.0, + 294.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 1306.0, + 1030.0, + 1306.0, + 1030.0, + 1336.0, + 670.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1335.0, + 1029.0, + 1335.0, + 1029.0, + 1361.0, + 669.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1361.0, + 991.0, + 1361.0, + 991.0, + 1391.0, + 669.0, + 1391.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/rkgz2aEKDr/rkgz2aEKDr.md b/parse/train/rkgz2aEKDr/rkgz2aEKDr.md new file mode 100644 index 0000000000000000000000000000000000000000..bdb4f9c73863f32f282e1b8e301ba25ba0ea08a6 --- /dev/null +++ b/parse/train/rkgz2aEKDr/rkgz2aEKDr.md @@ -0,0 +1,412 @@ +# ON THE VARIANCE OF THE ADAPTIVE LEARNING RATE AND BEYOND + +Liyuan Liu ∗ +University of Illinois, Urbana-Champaign +ll2@illinois + +Haoming Jiang Georgia Tech jianghm@gatech.edu + +Pengcheng He, Weizhu Chen Microsoft Dynamics 365 AI {penhe,wzchen}@microsoft.com + +Xiaodong Liu, Jianfeng Gao Microsoft Research {xiaodl,jfgao}@microsoft.com + +Jiawei Han +University of Illinois, Urbana-Champaign +hanj@illinois + +# ABSTRACT + +The learning rate warmup heuristic achieves remarkable success in stabilizing training, accelerating convergence and improving generalization for adaptive stochastic optimization algorithms like RMSprop and Adam. Pursuing the theory behind warmup, we identify a problem of the adaptive learning rate – its variance is problematically large in the early stage, and presume warmup works as a variance reduction technique. We provide both empirical and theoretical evidence to verify our hypothesis. We further propose Rectified Adam (RAdam), a novel variant of Adam, by introducing a term to rectify the variance of the adaptive learning rate. Experimental results on image classification, language modeling,500550 and neural machine translation verify our intuition and demonstrate the efficacy400450 and robustness of RAdam.1 + +# 1 INTRODUCTION + +Fast and stable optimization algorithms are what generations of researchers have been pursuing (Gauss, 1823; Cauchy, 1847). Remarkably, stochastic gradient-based optimization, such as stochastic gradient descent (SGD), has witnessed tremendous success in many fields of science and engineering despite its simplicity. Recently, many efforts have been made to accelerate optimization by applying adaptive learning rate. In particular, Adagrad (Duchi et al., 2010) and its variants, e.g., RMSprop (Hinton et al., 2012), Adam (Kingma & Ba, 2014), Adadelta (Zeiler, 2012) and Nadam (Dozat, 2016), stand out due to their fast convergence, and have been considered as the optimizer of choice in many applications. + +![](images/9e07baec4bf47bd9075fa89a9a4a4cab421e0a129a16eb042e7e1b7d64da16fe.jpg) +Figure 1: Training loss v.s. # of iterations of Transformers on the De-En IWSLT’14 dataset. + +However, it has been observed that these optimization methods may converge to bad/suspicious local optima, and have to resort to a warmup heuristic – using a small learning rate in the first few epochs of training to mitigate such problem (Vaswani et al., 2017; Popel & Bojar, 2018). For example, when training typical Transformers based neural machine translation models on the De-En IWSLT’14 dataset, removing the warmup stage increases the training loss from 3 to around 10, as shown in Figure 1. Similar phenomena are observed in other scenarios like BERT (a bidirectional transformer language model) pre-training (Devlin et al., 2019). + +Duo to the lack of the theoretical underpinnings, there is neither guarantee that warmup would bring consistent improvements for various machine learning settings nor guidance on how we should conduct warmup. Thus, researchers typically use different settings in different applications and have to take a trial-and-error approach, which can be tedious and time-consuming. + +In this paper, we conduct both empirical and theoretical analysis of the convergence issue to identify its origin. We show that its root cause is: the adaptive learning rate has undesirably large variance in the early stage of model training, due to the limited amount of training samples being used. Thus, to reduce such variance, it is better to use smaller learning rates in the first few epochs of training, which justifies the warmup heuristic. + +Inspired by our analysis results, we propose a new variant of Adam, called Rectified Adam (RAdam), which explicitly rectifies the variance of the adaptive learning rate based on derivations. We conduct extensive experiments on language modeling, image classification, and neural machine translation. RAdam brings consistent improvement over the vanilla Adam, which verifies the variance issue generally exists on various tasks across different network architectures. + +In summary, our main contributions are two-fold: + +• We identify the variance issue of the adaptive learning rate and present a theoretical justification for the warmup heuristic. We show that the convergence issue is due to the undesirably large variance of the adaptive learning rate in the early stage of model training. • We propose a new variant of Adam (i.e., RAdam), which not only explicitly rectifies the variance and is theoretically sound, but also compares favorably with the heuristic warmup. + +# 2 PRELIMINARIES AND MOTIVATIONS + +Generic adaptive methods. Algorithm 1 is a generic framework (all operations are element-wise). It describes various popular stochastic gradient descent algorithms (Reddi et al., 2018). Specifically, different optimization algorithms can be specified by different choices of $\phi ( . )$ and $\psi ( . )$ , where $\phi ( . )$ specifies how the momentum at time step $t$ is calculated, and $\psi ( . )$ how the adaptive learning rate at $t$ is calculated. For example, in the Adam algorithm, we have: + +$$ +\phi ( g _ { 1 } , \cdots , g _ { t } ) = { \frac { \dot { ( 1 - \beta _ { 1 } ) } \sum _ { i = 1 } ^ { t } \beta _ { 1 } ^ { t - i } g _ { t } } { 1 - \beta _ { 1 } ^ { t } } } \quad \mathrm { a n d } \quad \psi ( g _ { 1 } , \cdots , g _ { t } ) = { \sqrt { \frac { 1 - \beta _ { 2 } ^ { t } } { ( 1 - \beta _ { 2 } ) \sum _ { i = 1 } ^ { t } \beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } } } . +$$ + +For numerical stability, the function $\psi ( . )$ in Equation 1 is usually calculated as $\widehat { \psi } ( g _ { 1 } , \cdots , g _ { t } ) =$ $\frac { \sqrt { 1 - \beta _ { 2 } ^ { t } } } { \epsilon + \sqrt { \left( 1 - \beta _ { 2 } \right) \sum _ { i = 1 } ^ { t } \beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } }$ , where $\epsilon$ is a relatively small / negligible value (e.g., $1 \times 1 0 ^ { - 8 }$ ). + +Algorithm 1: Generic adaptive optimization method setup. All operations are element-wise. + +Input: $\{ \alpha _ { t } \} _ { t = 1 } ^ { T }$ : step size, $\{ \phi _ { t } , \psi _ { t } \} _ { t = 1 } ^ { T }$ : function to calculate momentum and adaptive rate, $\theta _ { 0 }$ : initial parameter, $f ( \theta )$ : stochastic objective function. + +Output: $\theta _ { T }$ : resulting parameters + +1 while $t = 1$ to $T$ do +2 $g _ { t } \gets \Delta _ { \theta } f _ { t } ( \theta _ { t - 1 } )$ (Calculate gradients w.r.t. stochastic objective at timestep t) +3 $m _ { t } \gets \phi _ { t } ( g _ { 1 } , \cdot \cdot \cdot , g _ { t } )$ (Calculate momentum) +4 $l _ { t } \gets \psi _ { t } ( g _ { 1 } , \cdot \cdot \cdot , g _ { t } )$ (Calculate adaptive learning rate) +5 $\theta _ { t } \gets \theta _ { t - 1 } - \alpha _ { t } m _ { t } l _ { t }$ (Update parameters) + +6 return $\theta _ { T }$ + +Learning rate warmup. Instead of setting the learning rate $\alpha _ { t }$ as a constant or in a decreasing order, a learning rate warmup strategy sets $\alpha _ { t }$ as smaller values in the first few steps, thus not satisfying $\forall t \alpha _ { t + 1 } \leq \alpha _ { t }$ . For example, linear warmup sets $\alpha _ { t } = t \alpha _ { 0 }$ when $t < T _ { w }$ . Warmup has been demonstrated to be beneficial in many deep learning applications. For example, in the NMT experiments in Figure 1, the training loss convergences around 10 when warmup is not applied (Adam-vanilla), and it surprisingly decreases to below 3 after applying warmup (Adam-warmup). + +To further analyze this phenomenon, we visualize the histogram of the absolute value of gradients on a log scale in Figure 2. We observe that, without applying warmup, the gradient distribution is distorted to have a mass center in relatively small values within 10 updates. Such gradient distortion means that the vanilla Adam is trapped in bad/suspicious local optima after the first few updates. Warmup essentially reduces the impact of these problematic updates to avoid the convergence problem. In the following sections, we focus our analysis on learning rate warmup for the Adam algorithm, while it can be applied to other algorithms that use similar adaptive learning rate $( \psi ( . ) )$ designs, e.g., RMSprop (Hinton et al., 2012) and Nadam (Dozat, 2016). + +![](images/6e677d6bab859ece1c2f01a47cda68a7b6ac5a545e1cae5ce58363c6119aa535.jpg) +Figure 2: The absolute gradient histogram of the Transformers on the De-En IWSLT’ 14 dataset during the training (stacked along the y-axis). X-axis is absolute value in the log scale and the height is the frequency. Without warmup, the gradient distribution is distorted in the first 10 steps. + +![](images/dc6f22e695c9b55a2b7cd42c7d714d441222920db3df396076473763529e1315.jpg) +Figure 3: The histogram of the absolute value of gradients (on a log scale) during the training of Transformers on the De-En IWSLT’ 14 dataset. using Adam-2k, RAdam and Adam-eps. + +# 3 VARIANCE OF THE ADAPTIVE LEARNING RATE + +In this section, we first introduce empirical evidence, then analyze the variance of the adaptive learning rate to support our hypothesis – Due to the lack of samples in the early stage, the adaptive learning rate has an undesirably large variance, which leads to suspicious/bad local optima. + +To convey our intuition, we begin with a special case. When $t = 1$ , we have $\psi ( g _ { 1 } ) = \sqrt { 1 / g _ { 1 } ^ { 2 } }$ . We view $\{ g _ { 1 } , \cdots , g _ { t } \}$ as i.i.d. Gaussian random variables following ${ \mathcal { N } } ( 0 , \sigma ^ { 2 } ) ^ { 2 }$ . Therefore, $1 / g _ { 1 } ^ { 2 }$ is subject to the scaled inverse chi-squared distribution, Scale-inv- $\mathcal { X } ^ { 2 } ( 1 , 1 / \sigma ^ { 2 } )$ , and $\mathrm { V a r } [ \sqrt { 1 / g _ { 1 } ^ { 2 } } ]$ is divergent. It means that the adaptive ratio can be undesirably large in the first stage of learning. Meanwhile, setting a small learning rate at the early stage can reduce the variance $\quad ( { \mathrm { V a r } } [ \alpha x ] \ =$ $\alpha ^ { 2 } \mathrm { V a r } [ x ] )$ , thus alleviating this problem. Therefore, we suggest it is the unbounded variance of the adaptive learning rate in the early stage that causes the problematic updates. + +# 3.1 WARMUP AS VARIANCE REDUCTION + +In this section, we design a set of controlled experiments to verify our hypothesis. Particularly, we design two variants of Adam that reducing the variance of the adaptive learning rate: Adam- ${ \it 2 k }$ and Adam-eps. We compare them to vanilla Adam with and without warmup on the IWSLT’14 German to English translation dataset (Cettolo et al., 2014). + +In order to reduce the variance of the adaptive learning rate $( \psi ( . ) )$ , Adam- $2 \mathrm { k }$ only updates $\psi ( . )$ in the first two thousand iterations, while the momentum $( \phi ( . ) )$ and parameters $\mathbf { \eta } ^ { ( \theta ) }$ are fixed3; other than this, it follows the original Adam algorithm. To make comparison with other methods, its iterations are indexed from -1999 instead of 1. In Figure 1, we observe that, after getting these additional two thousand samples for estimating the adaptive learning rate, Adam- $2 \mathrm { k }$ avoids the convergence problem of the vanilla-Adam. Also, comparing Figure 2 and Figure 3, getting large enough samples prevents the gradient distribution from being distorted. These observations verify our hypothesis that the lack of sufficient data samples in the early stage is the root cause of the convergence issue. + +Another straightforward way to reduce the variance is to increase the value of $\epsilon$ in $\widehat { \psi } ( g _ { 1 } , \cdots , g _ { t } ) =$ $\frac { \sqrt { 1 - \beta _ { 2 } ^ { t } } } { \epsilon + \sqrt { \left( 1 - \beta _ { 2 } \right) \sum _ { i = 1 } ^ { t } \beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } }$ Actually, if we assume $\widehat { \psi } ( . )$ is subject to the uniform distribution, its variance equals to $\frac { 1 } { 1 2 \epsilon ^ { 2 } }$ . Therefore, we design Adam-eps, which uses a non-negligibly large $\epsilon = 1 0 ^ { - 4 }$ , while $\epsilon = 1 0 ^ { - 8 }$ for vanilla Adam. Its performance is summarized in Figure 1. We observe that it does not suffer from the serious convergence problem of vanilla-Adam. This further demonstrates that the convergence problem can be alleviated by reducing the variance of the adaptive learning rate, and also explains why tuning $\epsilon$ is important in practice (Liu et al., 2019). Besides, similar to Adam-2k, it prevents the gradient distribution from being distorted (as shown in Figure 3). However, as in Figure 1, it produces a much worse performance comparing to Adam- $2 \mathrm { k }$ and Adam-warmup. We conjecture that this is because large $\epsilon$ induces a large bias into the adaptive learning rate and slows down the optimization process. Thus, we need a more principled and rigorous way to control the variance of the adaptive learning rate. In the next subsection, we will present a theoretical analysis of the variance of the adaptive learning rate. + +# 3.2 ANALYSIS OF ADAPTIVE LEARNING RATE VARIANCE + +As mentioned before, Adam uses the exponential moving average to calculate the adaptive learning rate. For gradients $\{ g _ { 1 } , \cdots , g _ { t } \}$ , their exponential moving average has a larger variance than their simple average. Also, in the early stage $\mathbf { \chi } _ { t }$ is small), the difference of the exponential weights of $\{ g _ { 1 } , \cdot \cdot \cdot , g _ { t } \}$ is relatively small (up to $\breve { 1 } - \beta _ { 2 } ^ { t - 1 } ,$ ). Therefore, for ease of analysis, we approximate the distribution of the exponential moving average as the distribution of the simple average (Nau, 2014), i.e., p(ψ(.)) = p( $\begin{array} { r } { p ( \psi ( . ) ) = p ( \sqrt { \frac { 1 - \beta _ { 2 } ^ { t } } { ( 1 - \beta _ { 2 } ) \sum _ { i = 1 } ^ { t } \beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } } ) \approx p ( \sqrt { \frac { t } { \sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } } ) } \end{array}$ Since $g _ { i } \sim \mathcal { N } ( 0 , \sigma ^ { 2 } )$ , we have ${ \frac { t } { \sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } } \sim \mathrm { S c a l e } { \mathrm { - i n v } } { - \mathcal { X } } ^ { 2 } ( t , { \frac { 1 } { \sigma ^ { 2 } } } )$ . Therefore, we assume $\frac { 1 - \beta _ { 2 } ^ { t } } { ( 1 - \beta _ { 2 } ) \sum _ { i = 1 } ^ { t } \beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } }$ also subjects to a scaled inverse chi-square distribution with $\rho$ degrees of freedom (further analysis on this approximation is conducted in Section 5.3). Based on this assumption, we can calculate $\mathrm { V a r } [ \psi ^ { 2 } ( . ) ]$ and the PDF of $\psi ^ { 2 } ( . )$ . Now, we proceed to the analysis of its square root variance, i.e., $\mathrm { V a r } [ \psi ( . ) ]$ , and show how the variance changes with $\rho$ (which corresponds to number of used training samples). + +Theorem 1. If $\psi ^ { 2 } ( . ) \sim$ Scale-inv- $\scriptstyle { \mathcal { X } } ^ { 2 } ( \rho , { \frac { 1 } { \sigma ^ { 2 } } } )$ , $\mathrm { V a r } [ \psi ( . ) ]$ monotonically decreases as $\rho$ increases. + +Proof. For $\forall \rho > 4$ , we have: + +$$ +\mathrm { V a r } [ \psi ( . ) ] = \mathbb { E } [ \psi ^ { 2 } ( . ) ] - \mathbb { E } [ \psi ( . ) ] ^ { 2 } = \tau ^ { 2 } ( \frac { \rho } { \rho - 2 } - \frac { \rho 2 ^ { 2 \rho - 5 } } { \pi } \mathcal { B } ( \frac { \rho - 1 } { 2 } , \frac { \rho - 1 } { 2 } ) ^ { 2 } ) , +$$ + +where $B ( . )$ is the beta function. By analyzing the derivative of $\mathrm { V a r } [ \psi ( . ) ]$ , we know it monotonically decreases as $\rho$ increases. The detailed derivation is elaborated in the Appendix A. □ + +Theorem 1 gives a qualitative analysis of the variance of the adaptive learning rate. It shows that, due to the lack of used training samples in the early stage, $\mathrm { V a r } [ \bar { \psi } ( . ) ]$ is larger than the late stage (Figure 8). To rigorously constraint the variance, we perform a quantified analysis on $\mathrm { V a r } [ \psi ( . ) ]$ by estimating the degree of freedoms $\rho$ . + +# 4 RECTIFIED ADAPTIVE LEARNING RATE + +In the previous section, Equation 2 gives the analytic form of $\mathrm { V a r } [ \psi ( . ) ]$ , where $\rho$ is the degree of freedoms. Here, we first give an estimation of $\rho$ based on $t$ to conduct a quantified analysis for $\operatorname { V a r } [ \psi ( g _ { 1 } , \cdots , g _ { t } ) ]$ , then we describe the design of the learning rate rectification, and compare it to the heuristic warmup strategies. + +# 4.1 ESTIMATION OF $\rho$ + +The exponential moving average (EMA) can be interpreted as an approximation to the simple moving average (SMA) in real application (Nau, 2014), i.e., + +$$ +p \left( \frac { \left( 1 - \beta _ { 2 } \right) \sum _ { i = 1 } ^ { t } \beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } { 1 - \beta _ { 2 } ^ { t } } \right) \approx p \left( \frac { \sum _ { i = 1 } ^ { f \left( t , \beta _ { 2 } \right) } g _ { t + 1 - i } ^ { 2 } } { f ( t , \beta _ { 2 } ) } \right) . +$$ + +Algorithm 2: Rectified Adam. All operations are element-wise. + +Input: $\{ \alpha _ { t } \} _ { t = 1 } ^ { T }$ : step size, $\{ \beta _ { 1 } , \beta _ { 2 } \}$ : decay rate to calculate moving average and moving 2nd +moment, $\theta _ { 0 }$ : initial parameter, $f _ { t } ( \theta )$ : stochastic objective function. +Output: $\theta _ { t }$ : resulting parameters +1 $m _ { 0 } , v _ { 0 } \gets 0 , 0$ (Initialize moving 1st and 2nd moment) +2 $\rho _ { \infty } 2 / ( 1 - \beta _ { 2 } ) - 1$ (Compute the maximum length of the approximated SMA) +3 while $t = \{ 1 , \cdots , T \}$ do +4 $g _ { t } \gets \Delta _ { \theta } f _ { t } ( \theta _ { t - 1 } )$ (Calculate gradients w.r.t. stochastic objective at timestep t) +5 $v _ { t } 1 / \beta _ { 2 } v _ { t - 1 } + ( 1 - \beta _ { 2 } ) g _ { t } ^ { 2 }$ (Update exponential moving 2nd moment) +6 $m _ { t } \gets \beta _ { 1 } m _ { t - 1 } + ( 1 - \beta _ { 1 } ) g _ { t }$ (Update exponential moving 1st moment) +7 $\widehat { m _ { t } } \gets m _ { t } / ( 1 - \beta _ { 1 } ^ { t } )$ (Compute bias-corrected moving average) +8 $\rho _ { t } \gets \rho _ { \infty } - 2 t \beta _ { 2 } ^ { t } / ( 1 - \beta _ { 2 } ^ { t } )$ (Compute the length of the approximated SMA) +9 if the variance is tractable, i.e., $\rho _ { t } > 4$ then +10 $l _ { t } \sqrt { ( 1 - \beta _ { 2 } ^ { t } ) / v _ { t } }$ (Compute adaptive learning rate) +11 $\begin{array} { r } { r _ { t } \sqrt { \frac { ( \rho _ { t } - 4 ) ( \rho _ { t } - 2 ) \rho _ { \infty } } { ( \rho _ { \infty } - 4 ) ( \rho _ { \infty } - 2 ) \rho _ { t } } } } \end{array}$ (Compute the variance rectification term) +12 $\theta _ { t } \gets \theta _ { t - 1 } - \alpha _ { t } r _ { t } \widehat { m _ { t } } l _ { t }$ (Update parameters with adaptive momentum) +13 else +14 $\begin{array} { r } { \big \lfloor \theta _ { t } \gets \theta _ { t - 1 } - \alpha _ { t } \widehat { m _ { t } } } \end{array}$ (Update parameters with un-adapted momentum) + +15 return $\theta _ { T }$ + +where $f ( t , \beta _ { 2 } )$ is the length of the SMA which allows the SMA to have the same “center of mass” with the EMA. In other words, $f ( t , \beta _ { 2 } )$ satisfies: + +$$ +\frac { ( 1 - \beta _ { 2 } ) \sum _ { i = 1 } ^ { t } \beta _ { 2 } ^ { t - i } \cdot i } { 1 - \beta _ { 2 } ^ { t } } = \frac { \sum _ { i = 1 } ^ { f ( t , \beta _ { 2 } ) } ( t + 1 - i ) } { f ( t , \beta _ { 2 } ) } . +$$ + +By solving Equation 4, we have: $\begin{array} { r l r } { f ( t , \beta _ { 2 } ) } & { { } = } & { { \frac { 2 } { 1 - \beta _ { 2 } } } - 1 - { \frac { 2 t \beta _ { 2 } ^ { t } } { 1 - \beta _ { 2 } ^ { t } } } } \end{array}$ . In the previous section, we assume: $\begin{array} { r } { \frac { 1 - \beta _ { 2 } ^ { t } } { \left( 1 - \beta _ { 2 } \right) \sum _ { i = 1 } ^ { t } \beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } \ \sim \ \mathrm { S c a l e - i n v - } \mathcal { X } ^ { 2 } ( \rho , \frac { 1 } { \sigma ^ { 2 } } ) . } \end{array}$ . Here, since $g _ { i } \sim \mathcal { N } ( 0 , \sigma ^ { 2 } )$ , we have 2) g2t+1−it,β2) ∼ Scale-inv-X 2(f (t, β2), 1σ2 ). Thus, Equation 3 views Scale-inv-X 2(f (t, β2), 1σ2 ) as an approximation to Scale-inv- $\scriptstyle { \mathcal { X } } ^ { 2 } ( \rho , { \frac { 1 } { \sigma ^ { 2 } } } )$ . Therefore, we treat $f ( t , \beta _ { 2 } )$ as an estimation of $\rho$ . For ease of notation, we mark f (t, β2) as ρt. Also, we refer 21−β2 as $\rho _ { \infty }$ (maximum length of the approximated SMA), due to the inequality $\begin{array} { r } { f ( t , \beta _ { 2 } ) \leq \operatorname* { l i m } _ { t \infty } f ( t , \beta _ { 2 } ) = \frac { 2 } { 1 - \beta _ { 2 } } - 1 . } \end{array}$ . + +# 4.2 VARIANCE ESTIMATION AND RECTIFICATION + +Based on previous estimations, we have Var[ψ(.)] = τ 2( ρtρ −2 $\begin{array} { r } { \mathrm { V a r } [ \psi ( . ) ] = \tau ^ { 2 } ( \frac { \rho _ { t } } { \rho _ { t } - 2 } - \frac { \rho _ { t } 2 ^ { 2 \rho _ { t } - 5 } } { \pi } \mathcal { B } ( \frac { \rho _ { t } - 1 } { 2 } , \frac { \rho _ { t } - 1 } { 2 } ) ^ { 2 } ) } \end{array}$ . The value of this function in the early stage is significantly larger than the late stage (as analyzed later, it decays roughly at the speed of $\begin{array} { r } { \dot { O } ( \frac { 1 } { \rho _ { t } } \overline { { ) } } ) } \end{array}$ . For example, the variance at $\rho _ { t } = 5$ is over 100 times larger than the variance at $\rho _ { t } ~ = ~ 5 0 0$ . Additionally, based on Theorem 1, we know $\begin{array} { r } { \operatorname* { m i n } _ { \rho _ { t } } \mathrm { V a r } [ \psi ( . ) ] = } \end{array}$ $\mathrm { V a r } [ \psi ( . ) ] | _ { \rho _ { t } = \rho _ { \infty } }$ and mark this minimal value as $C _ { \mathrm { V a r } }$ . In order to ensure that the adaptive learning rate $( \psi ( . ) )$ has consistent variance, we rectify the variance at the $t$ -th timestamp as below, + +$$ +\mathrm { V a r } [ r _ { t } \psi ( g _ { 1 } , \cdot \cdot \cdot , g _ { t } ) ] = C _ { \mathrm { V a r } } \quad \mathrm { w h e r e } \quad r _ { t } = \sqrt { C _ { \mathrm { V a r } } / \mathrm { V a r } [ \psi ( g _ { 1 } , \cdot \cdot \cdot , g _ { t } ) ] } . +$$ + +Although we have the analytic form of $\mathrm { V a r } [ \psi ( . ) ]$ (i.e., Equation 2), it is not numerically stable. Therefore, we use the first-order approximation to calculate the rectification term. Specifically, by approximating $\sqrt { \psi ^ { 2 } ( . ) }$ to the first order (Wolter, 2007), + +$$ +\sqrt { \psi ^ { 2 } ( . ) } \approx \sqrt { \mathbb { E } [ \psi ^ { 2 } ( . ) ] } + \frac { 1 } { 2 \sqrt { \mathbb { E } [ \psi ^ { 2 } ( . ) ] } } ( \psi ^ { 2 } ( . ) - \mathbb { E } [ \psi ^ { 2 } ( . ) ] ) \quad \mathrm { a n d } \quad \mathrm { V a r } [ \psi ( . ) ] \approx \frac { \mathrm { V a r } [ \psi ^ { 2 } ( . ) ] } { 4 \mathbb { E } [ \psi ^ { 2 } ( . ) ] } . +$$ + +Since $\psi ^ { 2 } ( . ) \sim \mathrm { S c a l e } { \mathrm { - i n v } { - } } \mathcal { X } ^ { 2 } ( \rho _ { t } , \frac { 1 } { \sigma ^ { 2 } } )$ , we have: + +$$ +\mathrm { V a r } [ \psi ( . ) ] \approx \rho _ { t } / [ 2 ( \rho _ { t } - 2 ) ( \rho _ { t } - 4 ) \sigma ^ { 2 } ] . +$$ + +In Section 5.3, we conduct simulation experiments to examine Equation 5 and find that it is a reliable approximation. Based on Equation 5, we know that $\mathrm { V a r } [ \sqrt { \psi ( . ) } ]$ decreases approximately at the + +Table 1: Image Classification + +
MethodAcc.
CITIIIOSGDAdamRAdam91.5190.5491.38
eeeseSGDAdamRAdam69.8666.5467.62
+ +![](images/c8779d131e033f69524303bfd6e6b6c707693f95b1efbdf7a0772e68fcc51b91.jpg) +Figure 4: Language modeling (LSTMs) on the One Billion Word. + +![](images/b8c26508ad736bb85b105e7794abce9ade7ded1d3b0ac44aa62399f7497eb5eb.jpg) +Figure 5: Training of ResNet-18 on the ImageNet and ResNet-20 on the CIFAR10 dataset. + +speed of $O ( \textstyle { \frac { 1 } { \rho _ { t } } } )$ . With this approximation, we can calculate the rectification term as: + +$$ +r _ { t } = \sqrt { \frac { ( \rho _ { t } - 4 ) ( \rho _ { t } - 2 ) \rho _ { \infty } } { ( \rho _ { \infty } - 4 ) ( \rho _ { \infty } - 2 ) \rho _ { t } } } . +$$ + +Applying our rectification term to Adam, we come up with a new variant of Adam, Rectified Adam (RAdam), as summarized in Algorithm 2. Specifically, when the length of the approximated SMA is less or equal than 4, the variance of the adaptive learning rate is intractable and the adaptive learning rate is inactivated. Otherwise, we calculate the variance rectification term and update parameters with the adaptive learning rate. It is worth mentioning that, if $\beta _ { 2 } ~ \leq ~ 0 . 6$ , we have $\rho _ { \infty } \leq 4$ and RAdam is degenerated to SGD with momentum. + +# 4.3 IN COMPARISON WITH WARMUP AND OTHER STABILIZATION TECHNIQUES + +Different from the analysis in this paper, warmup is originally proposed to handle training with very large batches for SGD (Goyal et al., 2017; Gotmare et al., 2019; Bernstein et al., 2018; Xiao et al., 2017). We notice that $r _ { t }$ has a similar form to the heuristic linear warmup, which can be viewed as setting the rectification term as $\frac { m i n ( t , T _ { w } ) } { T _ { w } }$ . It verifies our intuition that warmup works as a variance reduction technique. RAdam deactivates the adaptive learning rate when its variance is divergent, thus avoiding undesired instability in the first few updates. Besides, our method does not require an additional hyperparameter (i.e., $T _ { w . }$ ) and can automatically adapt to different moving average rules. + +Here, we identify and address an underlying issue of adaptive optimization methods independent of (neural) model architectures. Thus, the proposed rectification term is orthogonal to other training stabilization techniques such as gradient clipping (Bengio et al., 2013), smoothing the adaptive learning rate (i.e., increasing $\epsilon$ , applying geometric mean filter (Chen & Gu, 2018), or adding range constraints (Luo et al., 2019)), initialization (Balduzzi et al., 2017; Zhang et al., 2019) and normalization (Ba et al., 2016; Ioffe & Szegedy, 2015). Indeed, these techniques can be combined with the proposed variance rectification method. + +# 5 EXPERIMENTS + +We evaluate RAdam on several benchmarks: One Billion Word for language modeling; Cifar10 and ImageNet for image classification; IWSLT’14 De-En/EN-DE and WMT’16 EN-De for neural machine translation. Following Loshchilov & Hutter (2018), we decouple weight decays in the vanilla Adam, Adam with warmup and RAdam in our experiments. Details are in Appendix B. + +# 5.1 COMPARING TO VANILLA ADAM + +As analyzed before, the adaptive learning rate has undesirably large variance in the early stage of training and leads to suspicious/bad local optima on NMT. One question we are interested in is: whether such an issue widely exits in other similar tasks and applications. Thus, we conduct a set of experiments with two classical tasks of NLP and CV, i.e., language modeling and image classification. RAdam not only results in consistent improvements over the vanilla Adam, but also demonstrates its robustness to the change of learning rates. It verifies that the variance issue exists in various machine learning applications, and has a big impact on the model behavior. + +![](images/6345f7afa7e7a63d9229d9a8dfe723d4ffb9cfb0f3992b2f052dd4d31c0db1fa.jpg) +Figure 6: Performance of RAdam, Adam and SGD with different learning rates on CIFAR10. + +![](images/bb74ac762ae180f94b6c6c55e5c1b8414301b44371f1f8fbfd2e1825dfee7518.jpg) +Figure 7: Performance of RAdam, Adam with warmup on CIFAR10 with different learning rates. + +Performance Comparison. The performances on language modeling (i.e., One Billion Word (Chelba et al., 2013)) and image classification (i.e., CIFAR10 (Krizhevsky et al., 2009) and ImageNet (Deng et al., 2009)) are presented in Figure 4, 5. The results show that RAdam outperforms Adam in all three datasets. As shown in Figure 4, although the rectification term makes RAdam slower than the vanilla Adam in the first few epochs, it allows RAdam to converge faster after that. In other words, by reducing the variance of the adaptive learning rate in the early stage, it gets both faster convergence and better performance, which verifies the impact of the variance issue. We also observe that RAdam obtains consistent improvements over Adam on image classification. It is worth noting that, on both ImageNet and CIFAR10, although RAdam fails to outperform SGD in terms of test accuracy, it results in a better training performance (e.g., the training accuracy of SGD, Adam, and RAdam on ImageNet are 69.57, 69.12 and 70.30 respectively). + +Robustness to Learning Rate Change. Besides performance improvements, RAdam also improves the robustness of model training. We use different initial learning rates, conduct experiments with ResNet-20 on the CIFAR10 datasets, and summarize their performance in Figure 6. For learning rates within a broad range (i.e., $\{ 0 . 1 , 0 . 0 3 , 0 . 0 1 , 0 . 0 0 3 \} ,$ ), RAdam achieves consistent model performances (their test accuracy curves highly overlap with each other), while Adam and SGD are shown to be more sensitive to the learning rate. The observation can be interpreted that by rectifying the variance of the adaptive learning rate, RAdam improves the robustness of model training and can adapt to different learning rates of a broader range. + +# 5.2 COMPARING TO HEURISTIC WARMUP + +To examine the effectiveness of RAdam, we first conduct comparisons on neural machine translation, on which the state-of-the-art employs Adam with the linear warmup. Specifically, we conduct experiments on three datasets, i.e., IWSLT’14 De-En, IWSLT’14 En-De, and WMT’16 En-De. Due to the limited size of the IWSLT’14 dataset, we conduct experiments using 5 different random seeds and report their mean and standard derivation. As discussed before, the vanilla Adam algorithm leads to suspicious/bad local optima (i.e., converges to a training perplexity around 500), and needs a learning rate warmup stage to stabilize the training. + +Table 2: BLEU score on Neural Machine Translation. + +
MethodIWSLT'14 DE-ENIWSLT'14 EN-DEWMT'16 EN-DE
Adam with warmupRAdam34.66± 0.01434.76 ± 0.00328.56± 0.06728.48 ± 0.05427.0327.27
+ +We summarize the performance obtained with the heuristic warmup and our proposed rectification term in Table 2 and visualize the training curve of IWSLT De-En in Figure 1. With a consistent adaptive learning rate variance, our proposed method achieves similar performance to that of previous state-of-the-art warmup heuristics. It verifies our intuition that the problematic updates of Adam are indeed caused by the undesirably large variance in the early stage. + +Moreover, we applied Adam with warmup on the CIFAR10 dataset. Its best accuracy on the test set is 91.29, which is similar to RAdam (91.38). However, we found that RAdam requires less hyperparameter tuning. Specifically, we visualize their learning curves in Figure 7. For some warmup steps, Adam with warmup is relatively more sensitive to the choice of the learning rate. RAdam, at the same time, is not only more robust, but also can automatically control the warmup behavior (i.e., without requiring the length of warmup). For example, when setting the learning rate as 0.1, Adam with 100 steps of warmup fails to get satisfying performance and only results in an accuracy of 90.13; RAdam successfully gets an accuracy of 91.06, with the original setting of the moving average calculation (i.e., $\beta _ { 1 } = 0 . 9 , \beta _ { 2 } = 0 . 9 9 9 )$ . We conjecture the reason is due to the fact that RAdam, which is based on a rigorous variance analysis, explicitly avoids the extreme situation where the variance is divergent, and rectifies the variance to be consistent in other situations. + +# 5.3 SIMULATED VERIFICATION + +In Sections 3 and 4, we approximate $\textstyle \operatorname { V a r } [ \sqrt { t / \sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } ]$ to the first order, and assume $\psi ^ { 2 } ( . ) = $ $\frac { 1 - \beta _ { 2 } ^ { t } } { ( 1 - \beta _ { 2 } ) \sum _ { i = 1 } ^ { t } \beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } }$ subjects to a scaled inverse chi-square distribution (this assumption covers the approximation from EMA to SMA). Here, we examine these two approximations using simulations. + +First Order Approximation of $\textstyle \operatorname { V a r } [ \sqrt { t / \sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } ]$ . To compare Equations 5 and 2, we assume $\tau = 1$ and plot their values and difference for $\nu = \{ 5 , \cdots , 5 0 0 \}$ in Figure 8. The curve of the analytic form and the first-order approximation highly overlap, and their difference is much smaller than their value. This result verifies that our first-order approximation is very accurate. + +Scaled Inverse Chi-Square Distribution Assumption. In this paper, we assume $g _ { i }$ accords to a Normal distribution with a zero mean. We also assume $\psi ^ { 2 } ( . )$ accords to the scaled inverse chi-square distribution to derive the variance of $\mathrm { V a r } [ \psi ( . ) ]$ , based on the similarity between the exponential moving average and simple moving average. Here, we empirically verify this assumption. + +Specifically, since $g _ { i }$ in the optimization problem may not be zero-mean, we assume its expectation is $\mu$ and sample $g _ { i }$ from $\mathcal { N } ( \bar { \boldsymbol { \mu } } , 1 )$ . Then, based on these samples, we calculate the variance of the original adaptive learning rate and the proposed rectified adaptive learning rate, i.e., $\textstyle \operatorname { V a r } [ { \frac { 1 } { \widehat { v _ { t } } } } ]$ and $\textstyle \operatorname { V a r } [ { \frac { r _ { t } } { \widehat { v _ { t } } } } ]$ respectively. We set $\beta _ { 2 }$ bto 0.999, the number of sampled trajectories to 5000, the number bof iterations to 6000, and summarize the simulation results in Figure 9. Across all six settings with different $\mu$ , the adaptive learning rate has a larger variance in the first stage and the rectified adaptive learning rate has relative consistent variance. This verifies the reliability of our assumption. + +# 6 CONCLUSION + +In this paper, we explore the underlying principle of the effectiveness of the warmup heuristic used for adaptive optimization algorithms. Specifically, we identify that, due to the limited amount of samples in the early stage of model training, the adaptive learning rate has an undesirably large variance and can cause the model to converge to suspicious/bad local optima. We provide both empirical and theoretical evidence to support our hypothesis, and further propose a new variant of Adam, whose adaptive learning rate is rectified so as to have a consistent variance. Empirical results demonstrate the effectiveness of our proposed method. In future work, we plan to replace the rectification strategy by sharing the second moment estimation across similar parameters. + +![](images/faa8a54dec9667c124d2828f50a5de48c77fb5bb8cb529cf12a5d90cfc18c0fa.jpg) +Figure 8: The value of Equation 2, Equation 5 and their difference (absolute difference). The $\mathbf { X }$ -axis is $\rho$ and the y-axis is the variance (log scale). + +![](images/2c37e4a5e73292a2c4f76f3284121bfbb32e39ddd248373e6bf36a3c173b8e7f.jpg) +Figure 9: The simulation of $\mathrm { V a r } [ \frac { 1 } { v _ { t } } ]$ and $\textstyle \operatorname { V a r } [ { \frac { c _ { t } } { v _ { t } } } ]$ . The $\mathbf { X }$ -axis is iteration # (from 5), the y-axis is the variance (log scale). + +# ACKNOWLEDGE + +We thank Aeyuan Allen-Zhu for valuable discussions and comments, Microsoft Research Technology Engineering team for setting up GPU machines. Research was sponsored in part by DARPA No. W911NF-17-C-0099 and FA8750-19-2-1004, National Science Foundation IIS 16-18481, IIS 17-04532, and IIS-17-41317, and DTRA HDTRA11810026. + +# REFERENCES + +Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. + +David Balduzzi, Marcus Frean, Lennox Leary, JP Lewis, Kurt Wan-Duo Ma, and Brian McWilliams. The shattered gradients problem: If resnets are the answer, then what is the question? In ICML, 2017. + +Yoshua Bengio, Nicolas Boulanger-Lewandowski, and Razvan Pascanu. Advances in optimizing recurrent networks. In 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 8624–8628. IEEE, 2013. + +Jeremy Bernstein, Yu-Xiang Wang, Kamyar Azizzadenesheli, and Anima Anandkumar. signsgd: Compressed optimisation for non-convex problems. In ICML, 2018. + +Augustin Cauchy. Methode g ´ en´ erale pour la r ´ esolution des systemes d ´ equations simultan ´ ees. ´ Comp. Rend. Sci. Paris, 25(1847):536–538, 1847. + +Mauro Cettolo, Jan Niehues, Sebastian Stuker, Luisa Bentivogli, and Marcello Federico. Report on ¨ the 11th iwslt evaluation campaign, iwslt 2014. In Proceedings of the International Workshop on Spoken Language Translation,, 2014. + +Ciprian Chelba, Tomas Mikolov, Michael Schuster, Qi Ge, Thorsten Brants, Phillipp Koehn, and Tony Robinson. One billion word benchmark for measuring progress in statistical language modeling. In INTERSPEECH, 2013. + +Jinghui Chen and Quanquan Gu. Closing the generalization gap of adaptive gradient methods in training deep neural networks. arXiv preprint arXiv:1806.06763, 2018. + +Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In ICML, 2009. + +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019. + +Timothy Dozat. Incorporating nesterov momentum into adam. 2016. + +John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. In COLT, 2010. + +Carl-Friedrich Gauss. Theoria combinationis observationum erroribus minimis obnoxiae. Commentationes Societatis Regiae Scientiarum Gottingensis Recentiores, 1823. + +Akhilesh Gotmare, Nitish Shirish Keskar, Caiming Xiong, and Richard Socher. A closer look at deep learning heuristics: Learning rate restarts, warmup and distillation. In ICLR, 2019. + +Priya Goyal, Piotr Dollar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, An- ´ drew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. + +Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. + +Geoffrey Hinton, Nitish Srivastava, and Kevin Swersky. Neural networks for machine learning lecture 6a overview of mini-batch gradient descent. Cited on, 2012. + +Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. + +Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. + +Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, Citeseer, 2009. + +Liyuan Liu, Xiang Ren, Jingbo Shang, Jian Peng, and Jiawei Han. Efficient contextualized representation: Language model pruning for sequence labeling. EMNLP, 2018. + +Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. + +Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. In ICLR, 2018. + +Liangchen Luo, Yuanhao Xiong, Yan Liu, and Xu Sun. Adaptive gradient methods with dynamic bound of learning rate. In ICLR, 2019. + +Robert Nau. Forecasting with moving averages. 2014. + +Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. fairseq: A fast, extensible toolkit for sequence modeling. In NAACL, 2019. + +Martin Popel and Ondˇrej Bojar. Training tips for the transformer model. The Prague Bulletin of Mathematical Linguistics, 110(1):43–70, 2018. + +Sashank J Reddi, Satyen Kale, and Sanjiv Kumar. On the convergence of adam and beyond. In ICLR, 2018. + +Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. + +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. + +Kirk M Wolter. Taylor series methods. In Introduction to variance estimation. 2007. + +Lin Xiao, Adams Wei Yu, Qihang Lin, and Weizhu Chen. Dscovr: Randomized primal-dual block coordinate algorithms for asynchronous distributed optimization. J. Mach. Learn. Res., 2017. + +Matthew D Zeiler. Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701, 2012. + +Hongyi Zhang, Yann N Dauphin, and Tengyu Ma. Fixup initialization: Residual learning without normalization. In ICLR, 2019. + +# A PROOF OF THEOREM 1 + +For ease of notation, we refer $\psi ^ { 2 } ( . )$ as $x$ and $\textstyle { \frac { 1 } { \sigma ^ { 2 } } }$ as $\tau ^ { 2 }$ . Thus, $x \sim$ Scale-inv- $\mathcal { X } ^ { 2 } ( \rho , \tau ^ { 2 } )$ and: + +$$ +p ( x ) = \frac { ( \tau ^ { 2 } \rho / 2 ) ^ { \rho / 2 } } { \Gamma ( \rho / 2 ) } \frac { \exp \bigl [ \frac { - \rho \tau ^ { 2 } } { 2 x } \bigr ] } { x ^ { 1 + \rho / 2 } } \quad \mathrm { a n d } \quad \mathbb { E } [ x ] = \frac { \rho } { ( \rho - 2 ) \sigma ^ { 2 } } \left( \forall \rho > 2 \right) +$$ + +where $\Gamma ( . )$ is the gamma function. Therefore, we have: + +$$ +\mathbb { E } [ \sqrt { x } ] = \int _ { 0 } ^ { \infty } \sqrt { x } p ( x ) d x = \frac { \tau \sqrt { \rho } \Gamma ( ( \rho - 1 ) / 2 ) } { \sqrt { 2 } \Gamma ( \rho / 2 ) } ( \forall \rho > 4 ) . +$$ + +Based on Equation 6 and 7, for $\forall \rho > 4$ , we have: + +$$ +\operatorname { V a r } [ \psi ( \cdot ) ] = \operatorname { V a r } [ { \sqrt { x } } ] = \operatorname { \mathbb { E } } [ x ] - \operatorname { \mathbb { E } } [ { \sqrt { x } } ] ^ { 2 } = \tau ^ { 2 } ( { \frac { \rho } { \rho - 2 } } - { \frac { \rho 2 ^ { 2 \rho - 5 } } { \pi } } B ( { \frac { \rho - 1 } { 2 } } , { \frac { \rho - 1 } { 2 } } ) ^ { 2 } ) , +$$ + +where $B ( . )$ is the beta function. To prove the monotonic property of $\mathrm { V a r } [ \psi ( . ) ]$ , we need to show: + +Lemma 1. for $t \geq 4$ $\begin{array} { r } { , \frac { \partial } { \partial t } \big ( \frac { t } { t - 2 } - \frac { t 2 ^ { 2 t - 5 } } { \pi } B ( \frac { t - 1 } { 2 } , \frac { t - 1 } { 2 } ) ^ { 2 } \big ) < 0 } \end{array}$ + +Proof. The target inequality can be re-wrote as + +$$ +\begin{array} { r l } & { \frac { \partial } { \partial t } ( \frac { t } { t - 2 } - \frac { t \dot { 2 } ^ { 2 t - 5 } } { \pi } B ( \frac { t - 1 } { 2 } , \frac { t - 1 } { 2 } ) ^ { 2 } ) } \\ & { = \frac { - 2 } { ( t - 2 ) ^ { 2 } } - \frac { 2 ^ { 2 t - 5 } } { \pi } B ( \frac { t - 1 } { 2 } , \frac { t - 1 } { 2 } ) ^ { 2 } - \frac { t 2 ^ { 2 t - 5 } \ln { 4 } } { \pi } B ( \frac { t - 1 } { 2 } , \frac { t - 1 } { 2 } ) ^ { 2 } } \\ & { - \frac { 2 t 2 ^ { 2 t - 5 } } { \pi } \mathcal { B } ( \frac { t - 1 } { 2 } , \frac { t - 1 } { 2 } ) ^ { 2 } ( \Psi ( \frac { t - 1 } { 2 } ) - \Psi ( t - 1 ) ) , \quad \left( \Psi ( x ) = \frac { \Gamma ^ { \prime } ( x ) } { \Gamma ( x ) } \right) } \\ & { < 0 } \end{array} +$$ + +This inequality is equivalent to: + +$$ +\begin{array} { r l r } & { } & { \frac { 6 4 \pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \mathcal { B } ( \frac { t - 1 } { 2 } , \frac { t - 1 } { 2 } ) ^ { 2 } } + 1 + t \ln 4 + 2 t \Psi ( \frac { t - 1 } { 2 } ) } \\ & { } & { > 2 t \Psi ( t - 1 ) \overset { ( i ) } { = } t [ \Psi ( \frac { t - 1 } { 2 } ) + \Psi ( \frac { t } { 2 } ) + \ln 4 ] , } \end{array} +$$ + +where $( i )$ is derived from Legendre duplication formula. Simplify the above inequality, we get: + +$$ +\frac { 6 4 \pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \mathcal { B } ( \frac { t - 1 } { 2 } , \frac { t - 1 } { 2 } ) ^ { 2 } } + 1 + t \Psi ( \frac { t - \hat { 1 } } { 2 } ) - t \Psi ( \frac { t } { 2 } ) > 0 , +$$ + +We only need to show + +$$ +\begin{array} { r l } & { \frac { 6 4 \pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \mathcal { B } ( \frac { t - 1 } { 2 } , \frac { t - 1 } { 2 } ) ^ { 2 } } + 1 + t \Psi ( \frac { t - 1 } { 2 } ) - t \Psi ( \frac { t } { 2 } ) } \\ & { \geq \frac { 6 4 \pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \mathcal { B } ( \frac { t - 1 } { 2 } , \frac { t - 1 } { 2 } ) ^ { 2 } } + 2 + t ( \ln ( t / 2 ) - 1 / ( t / 2 - 0 . 5 ) ) - t \ln ( t / 2 ) } \\ & { = \frac { 6 4 \pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \mathcal { B } ( \frac { t - 1 } { 2 } , \frac { t - 1 } { 2 } ) ^ { 2 } } - \frac { 2 } { t - 1 } } \\ & { > \frac { 6 4 \pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \mathcal { B } ( \frac { t - 1 } { 2 } , \frac { t - 1 } { 2 } ) ^ { 2 } } - \frac { 2 } { t - 2 } \geq 0 , } \\ & { \mathrm { r s t i n e q u a l i t y i s i f r o m } \ln ( x ) - 1 / ( 2 x ) > \Psi ( x ) > \ln ( x + 0 . 5 ) - 1 / x . } \end{array} +$$ + +where the fi + +Therefore, we only need to show + +$$ +3 2 \pi \geq ( t - 2 ) 4 ^ { t } B ( \frac { t - 1 } { 2 } , \frac { t - 1 } { 2 } ) ^ { 2 } , +$$ + +which is equivalent to + +$$ +\begin{array} { l } { ( t - 2 ) 4 ^ { t } \mathcal { B } ( \displaystyle \frac { t - 1 } { 2 } , \displaystyle \frac { t - 1 } { 2 } ) ^ { 2 } = ( t - 2 ) 4 ^ { t } \frac { \Gamma ( \displaystyle \frac { t - 1 } { 2 } ) ^ { 4 } } { \Gamma ( t - 1 ) ^ { 2 } } } \\ { \overset { ( i ) } { = } ( t - 2 ) 4 ^ { t } \frac { \Gamma ( \displaystyle \frac { t - 1 } { 2 } ) ^ { 2 } } { \Gamma ( t / 2 ) ^ { 2 } } 4 ^ { 2 - t } \pi = 1 6 \pi ( t - 2 ) \frac { \Gamma ( \displaystyle \frac { t - 1 } { 2 } ) ^ { 2 } } { \Gamma ( t / 2 ) ^ { 2 } } \le 3 2 \pi , } \end{array} +$$ + +where (i) is from Legendre duplication formula. + +So we only need to show + +$$ +( t - 2 ) \frac { \Gamma ( \frac { t - 1 } { 2 } ) ^ { 2 } } { \Gamma ( t / 2 ) ^ { 2 } } \leq 2 +$$ + +Using Gautschi’s inequality $( \frac { \Gamma ( x + 1 ) } { \Gamma ( x + s ) } < ( x + 1 ) ^ { 1 - s } )$ , we have + +$$ +( t - 2 ) \frac { \dot { \Gamma } ( \frac { t - 1 } { 2 } ) ^ { 2 } } { \Gamma ( t / 2 ) ^ { 2 } } \leq ( t - 2 ) ( \frac { t - 1 } { 2 } ) ^ { - 1 } = \frac { 2 ( t - 2 ) } { t - 1 } < 2 +$$ + +# B IMPLEMENTATION DETAILS + +# B.1 LANGUAGE MODELING + +Our implementation is based on the previous work (Liu et al., 2018). Specifically, we use two-layer LSTMs with 2048 hidden states with adaptive softmax to conduct experiments on the one billion words dataset. Word embedding (random initialized) of 300 dimensions is used as the input and the adaptive softmax is incorporated with a default setting (cut-offs are set to [4000, 40000, 200000]). Additionally, as pre-processing, we replace all tokens occurring equal or less than 3 times with as UNK, which shrinks the dictionary from 7.9M to 6.4M. Dropout is applied to each layer with a ratio of 0.1, gradients are clipped at 5.0. We use the default hyper-parameters to update moving averages, $i . e . \beta _ { 1 } = 0 . 9$ and $\beta _ { 2 } = 0 . 9 9 9$ . The learning rate is set to start from 0.001, and decayed at the start of 10th epochs. LSTMs are unrolled for 20 steps without resetting the LSTM states and the batch size is set to 128. All models are trained on one NVIDIA Tesla V100 GPU. + +# B.2 IMAGEINE CLASSIFICATION + +We use the default ResNet architectures (He et al., 2016) in a public pytorch re-implementation4. Specifically, we use 20-layer ResNet (9 Basic Blocks) for CIFAR-10 and 18-layer ResNet (8 Basic Blocks) for ImageNet. Batch size is 128 for CIFAR-10 and 256 for ImageNet. The model is trained for 186 epoches and the learning rate decays at the 81-th and the 122-th epoches by 0.1 on CIFAR10, while the model is trained for 90 epoches and the learning rate decays at the 31-th and the 61-th epoch by 0.1 on ImageNet. For Adam and RAdam, we set $\beta _ { 1 } = 0 . 9 , \beta _ { 2 } = 0 . 9 9 9$ . For SGD, we set the momentum factor as 0.9. The weight decay rate is $1 0 ^ { - 4 }$ . Random cropping and random horizontal flipping are applied to training data. + +# B.3 NEURAL MACHINE TRANSLATION + +Our experiments are based on the default Transformers (Vaswani et al., 2017) implementation from the fairseq package (Ott et al., 2019). Specifically, we use word embedding with 512 dimensions and 6-layer encoder / decoder with 4 head and 1024 hidden dimensions on the IWSLT14’ dataset; use word embedding with 512 dimension and 6-layer encoder / decoder with 8 heads and 2048 hidden dimensions. Label smoothed cross entropy is used as the objective function with an uncertainty $=$ 0.1 (Szegedy et al., 2016). We use linear learning rate decay starting from $3 e ^ { - 4 }$ , and the checkpoints of the last 20 epoches are averaged before evaluation. As to the wamrup strategy, we use a linear warmup for Adam in the first 4000 updates, and set $\beta _ { 2 }$ to satisfy $\nu = 4 0 0 0$ $\beta _ { 2 } = 0 . 9 9 9 5 )$ . In the IWSLT’14 dataset, we conduct training on one NVIDIA Tesla V100 GPU, set maximum batch size as 4000, apply dropout with a ratio 0.3, using weight decay of 0.0001 and clip the gradient norm at 25. In the WMT’16 dataset, we conduct training on four NVIDIA Quadro R8000 GPUs and set maximum batch size as 8196. + +# C DOWNGRADING TO SGDM + +As a byproduct determined by math derivations, we degenerated RAdam to SGD with momentum in the first several updates. Although this stage only contains several gradient updates, these updates could be quite damaging (e.g., in our Figure 2, the gradient distribution is distorted within 10 gradient updates). Intuitively, updates with divergent adaptive learning rate variance could be more damaging than the ones with converged variance, as divergent variance implies more instability. As a case study, we performed experiments on the CIFAR10 dataset. Five-run average results are summarized in Table 3. The optimizer fails to get an equally reliably model when changing the first 4 updates to Adam, yet the influence of switching is less deleterious when we change 5-8 updates instead. This result verifies our intuition and is in agreement with our theory the first few updates could be more damaging than later updates. By saying that, we still want to emphasize that this part (downgrading to SGDM) is only a minor part of our algorithm design whereas our main focus is on the mechanism of warmup and the derivation of the rectification term. + +Table 3: Performance on CIFAR10 $( \mathrm { l r } = 0 . 1 $ ). + +
1-4 steps5-8 steps8+ stepstest acctrain losstrain error
RAdamRAdamRAdam91.080.0210.74
Adam (w. divergent var.)RAdamRAdam89.980.0602.12
SGDAdam (w. convergent var.)RAdam90.290.0381.23
\ No newline at end of file diff --git a/parse/train/rkgz2aEKDr/rkgz2aEKDr_content_list.json b/parse/train/rkgz2aEKDr/rkgz2aEKDr_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..52441ca7d4c6d4def14d53a412df507431905ccf --- /dev/null +++ b/parse/train/rkgz2aEKDr/rkgz2aEKDr_content_list.json @@ -0,0 +1,2029 @@ +[ + { + "type": "text", + "text": "ON THE VARIANCE OF THE ADAPTIVE LEARNING RATE AND BEYOND ", + "text_level": 1, + "bbox": [ + 174, + 99, + 823, + 145 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liyuan Liu ∗ \nUniversity of Illinois, Urbana-Champaign \nll2@illinois ", + "bbox": [ + 184, + 170, + 431, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoming Jiang Georgia Tech jianghm@gatech.edu ", + "bbox": [ + 459, + 170, + 583, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pengcheng He, Weizhu Chen Microsoft Dynamics 365 AI {penhe,wzchen}@microsoft.com ", + "bbox": [ + 609, + 170, + 813, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaodong Liu, Jianfeng Gao Microsoft Research {xiaodl,jfgao}@microsoft.com ", + "bbox": [ + 183, + 233, + 382, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiawei Han \nUniversity of Illinois, Urbana-Champaign \nhanj@illinois ", + "bbox": [ + 482, + 234, + 730, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 313, + 544, + 327 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The learning rate warmup heuristic achieves remarkable success in stabilizing training, accelerating convergence and improving generalization for adaptive stochastic optimization algorithms like RMSprop and Adam. Pursuing the theory behind warmup, we identify a problem of the adaptive learning rate – its variance is problematically large in the early stage, and presume warmup works as a variance reduction technique. We provide both empirical and theoretical evidence to verify our hypothesis. We further propose Rectified Adam (RAdam), a novel variant of Adam, by introducing a term to rectify the variance of the adaptive learning rate. Experimental results on image classification, language modeling,500550 and neural machine translation verify our intuition and demonstrate the efficacy400450 and robustness of RAdam.1 ", + "bbox": [ + 233, + 345, + 766, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 531, + 334, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fast and stable optimization algorithms are what generations of researchers have been pursuing (Gauss, 1823; Cauchy, 1847). Remarkably, stochastic gradient-based optimization, such as stochastic gradient descent (SGD), has witnessed tremendous success in many fields of science and engineering despite its simplicity. Recently, many efforts have been made to accelerate optimization by applying adaptive learning rate. In particular, Adagrad (Duchi et al., 2010) and its variants, e.g., RMSprop (Hinton et al., 2012), Adam (Kingma & Ba, 2014), Adadelta (Zeiler, 2012) and Nadam (Dozat, 2016), stand out due to their fast convergence, and have been considered as the optimizer of choice in many applications. ", + "bbox": [ + 176, + 561, + 581, + 728 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9e07baec4bf47bd9075fa89a9a4a4cab421e0a129a16eb042e7e1b7d64da16fe.jpg", + "image_caption": [ + "Figure 1: Training loss v.s. # of iterations of Transformers on the De-En IWSLT’14 dataset. " + ], + "image_footnote": [], + "bbox": [ + 596, + 551, + 821, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, it has been observed that these optimization methods may converge to bad/suspicious local optima, and have to resort to a warmup heuristic – using a small learning rate in the first few epochs of training to mitigate such problem (Vaswani et al., 2017; Popel & Bojar, 2018). For example, when training typical Transformers based neural machine translation models on the De-En IWSLT’14 dataset, removing the warmup stage increases the training loss from 3 to around 10, as shown in Figure 1. Similar phenomena are observed in other scenarios like BERT (a bidirectional transformer language model) pre-training (Devlin et al., 2019). ", + "bbox": [ + 174, + 734, + 825, + 833 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Duo to the lack of the theoretical underpinnings, there is neither guarantee that warmup would bring consistent improvements for various machine learning settings nor guidance on how we should conduct warmup. Thus, researchers typically use different settings in different applications and have to take a trial-and-error approach, which can be tedious and time-consuming. ", + "bbox": [ + 174, + 839, + 823, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we conduct both empirical and theoretical analysis of the convergence issue to identify its origin. We show that its root cause is: the adaptive learning rate has undesirably large variance in the early stage of model training, due to the limited amount of training samples being used. Thus, to reduce such variance, it is better to use smaller learning rates in the first few epochs of training, which justifies the warmup heuristic. ", + "bbox": [ + 174, + 138, + 825, + 208 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by our analysis results, we propose a new variant of Adam, called Rectified Adam (RAdam), which explicitly rectifies the variance of the adaptive learning rate based on derivations. We conduct extensive experiments on language modeling, image classification, and neural machine translation. RAdam brings consistent improvement over the vanilla Adam, which verifies the variance issue generally exists on various tasks across different network architectures. ", + "bbox": [ + 174, + 215, + 825, + 285 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are two-fold: ", + "bbox": [ + 174, + 292, + 498, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "• We identify the variance issue of the adaptive learning rate and present a theoretical justification for the warmup heuristic. We show that the convergence issue is due to the undesirably large variance of the adaptive learning rate in the early stage of model training. • We propose a new variant of Adam (i.e., RAdam), which not only explicitly rectifies the variance and is theoretically sound, but also compares favorably with the heuristic warmup. ", + "bbox": [ + 174, + 309, + 825, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 PRELIMINARIES AND MOTIVATIONS ", + "text_level": 1, + "bbox": [ + 173, + 402, + 504, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generic adaptive methods. Algorithm 1 is a generic framework (all operations are element-wise). It describes various popular stochastic gradient descent algorithms (Reddi et al., 2018). Specifically, different optimization algorithms can be specified by different choices of $\\phi ( . )$ and $\\psi ( . )$ , where $\\phi ( . )$ specifies how the momentum at time step $t$ is calculated, and $\\psi ( . )$ how the adaptive learning rate at $t$ is calculated. For example, in the Adam algorithm, we have: ", + "bbox": [ + 173, + 433, + 825, + 502 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/ecf5e4b7f18515a650387401a902ea879805440618de5f9ded0e1c20c87ef512.jpg", + "text": "$$\n\\phi ( g _ { 1 } , \\cdots , g _ { t } ) = { \\frac { \\dot { ( 1 - \\beta _ { 1 } ) } \\sum _ { i = 1 } ^ { t } \\beta _ { 1 } ^ { t - i } g _ { t } } { 1 - \\beta _ { 1 } ^ { t } } } \\quad \\mathrm { a n d } \\quad \\psi ( g _ { 1 } , \\cdots , g _ { t } ) = { \\sqrt { \\frac { 1 - \\beta _ { 2 } ^ { t } } { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } } } .\n$$", + "text_format": "latex", + "bbox": [ + 184, + 501, + 792, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For numerical stability, the function $\\psi ( . )$ in Equation 1 is usually calculated as $\\widehat { \\psi } ( g _ { 1 } , \\cdots , g _ { t } ) =$ $\\frac { \\sqrt { 1 - \\beta _ { 2 } ^ { t } } } { \\epsilon + \\sqrt { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } }$ , where $\\epsilon$ is a relatively small / negligible value (e.g., $1 \\times 1 0 ^ { - 8 }$ ). ", + "bbox": [ + 176, + 545, + 825, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Algorithm 1: Generic adaptive optimization method setup. All operations are element-wise. ", + "bbox": [ + 173, + 607, + 777, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Input: $\\{ \\alpha _ { t } \\} _ { t = 1 } ^ { T }$ : step size, $\\{ \\phi _ { t } , \\psi _ { t } \\} _ { t = 1 } ^ { T }$ : function to calculate momentum and adaptive rate, $\\theta _ { 0 }$ : initial parameter, $f ( \\theta )$ : stochastic objective function. ", + "bbox": [ + 169, + 626, + 772, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Output: $\\theta _ { T }$ : resulting parameters ", + "bbox": [ + 173, + 656, + 397, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 while $t = 1$ to $T$ do \n2 $g _ { t } \\gets \\Delta _ { \\theta } f _ { t } ( \\theta _ { t - 1 } )$ (Calculate gradients w.r.t. stochastic objective at timestep t) \n3 $m _ { t } \\gets \\phi _ { t } ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } )$ (Calculate momentum) \n4 $l _ { t } \\gets \\psi _ { t } ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } )$ (Calculate adaptive learning rate) \n5 $\\theta _ { t } \\gets \\theta _ { t - 1 } - \\alpha _ { t } m _ { t } l _ { t }$ (Update parameters) ", + "bbox": [ + 161, + 666, + 710, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "6 return $\\theta _ { T }$ ", + "bbox": [ + 160, + 742, + 241, + 756 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning rate warmup. Instead of setting the learning rate $\\alpha _ { t }$ as a constant or in a decreasing order, a learning rate warmup strategy sets $\\alpha _ { t }$ as smaller values in the first few steps, thus not satisfying $\\forall t \\alpha _ { t + 1 } \\leq \\alpha _ { t }$ . For example, linear warmup sets $\\alpha _ { t } = t \\alpha _ { 0 }$ when $t < T _ { w }$ . Warmup has been demonstrated to be beneficial in many deep learning applications. For example, in the NMT experiments in Figure 1, the training loss convergences around 10 when warmup is not applied (Adam-vanilla), and it surprisingly decreases to below 3 after applying warmup (Adam-warmup). ", + "bbox": [ + 173, + 777, + 825, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further analyze this phenomenon, we visualize the histogram of the absolute value of gradients on a log scale in Figure 2. We observe that, without applying warmup, the gradient distribution is distorted to have a mass center in relatively small values within 10 updates. Such gradient distortion means that the vanilla Adam is trapped in bad/suspicious local optima after the first few updates. Warmup essentially reduces the impact of these problematic updates to avoid the convergence problem. In the following sections, we focus our analysis on learning rate warmup for the Adam algorithm, while it can be applied to other algorithms that use similar adaptive learning rate $( \\psi ( . ) )$ designs, e.g., RMSprop (Hinton et al., 2012) and Nadam (Dozat, 2016). ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6e677d6bab859ece1c2f01a47cda68a7b6ac5a545e1cae5ce58363c6119aa535.jpg", + "image_caption": [ + "Figure 2: The absolute gradient histogram of the Transformers on the De-En IWSLT’ 14 dataset during the training (stacked along the y-axis). X-axis is absolute value in the log scale and the height is the frequency. Without warmup, the gradient distribution is distorted in the first 10 steps. " + ], + "image_footnote": [], + "bbox": [ + 178, + 77, + 813, + 191 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dc6f22e695c9b55a2b7cd42c7d714d441222920db3df396076473763529e1315.jpg", + "image_caption": [ + "Figure 3: The histogram of the absolute value of gradients (on a log scale) during the training of Transformers on the De-En IWSLT’ 14 dataset. using Adam-2k, RAdam and Adam-eps. " + ], + "image_footnote": [], + "bbox": [ + 236, + 250, + 753, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 406, + 825, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 VARIANCE OF THE ADAPTIVE LEARNING RATE ", + "text_level": 1, + "bbox": [ + 174, + 479, + 598, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce empirical evidence, then analyze the variance of the adaptive learning rate to support our hypothesis – Due to the lack of samples in the early stage, the adaptive learning rate has an undesirably large variance, which leads to suspicious/bad local optima. ", + "bbox": [ + 174, + 508, + 825, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To convey our intuition, we begin with a special case. When $t = 1$ , we have $\\psi ( g _ { 1 } ) = \\sqrt { 1 / g _ { 1 } ^ { 2 } }$ . We view $\\{ g _ { 1 } , \\cdots , g _ { t } \\}$ as i.i.d. Gaussian random variables following ${ \\mathcal { N } } ( 0 , \\sigma ^ { 2 } ) ^ { 2 }$ . Therefore, $1 / g _ { 1 } ^ { 2 }$ is subject to the scaled inverse chi-squared distribution, Scale-inv- $\\mathcal { X } ^ { 2 } ( 1 , 1 / \\sigma ^ { 2 } )$ , and $\\mathrm { V a r } [ \\sqrt { 1 / g _ { 1 } ^ { 2 } } ]$ is divergent. It means that the adaptive ratio can be undesirably large in the first stage of learning. Meanwhile, setting a small learning rate at the early stage can reduce the variance $\\quad ( { \\mathrm { V a r } } [ \\alpha x ] \\ =$ $\\alpha ^ { 2 } \\mathrm { V a r } [ x ] )$ , thus alleviating this problem. Therefore, we suggest it is the unbounded variance of the adaptive learning rate in the early stage that causes the problematic updates. ", + "bbox": [ + 174, + 559, + 825, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 WARMUP AS VARIANCE REDUCTION ", + "text_level": 1, + "bbox": [ + 176, + 675, + 467, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we design a set of controlled experiments to verify our hypothesis. Particularly, we design two variants of Adam that reducing the variance of the adaptive learning rate: Adam- ${ \\it 2 k }$ and Adam-eps. We compare them to vanilla Adam with and without warmup on the IWSLT’14 German to English translation dataset (Cettolo et al., 2014). ", + "bbox": [ + 174, + 698, + 825, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to reduce the variance of the adaptive learning rate $( \\psi ( . ) )$ , Adam- $2 \\mathrm { k }$ only updates $\\psi ( . )$ in the first two thousand iterations, while the momentum $( \\phi ( . ) )$ and parameters $\\mathbf { \\eta } ^ { ( \\theta ) }$ are fixed3; other than this, it follows the original Adam algorithm. To make comparison with other methods, its iterations are indexed from -1999 instead of 1. In Figure 1, we observe that, after getting these additional two thousand samples for estimating the adaptive learning rate, Adam- $2 \\mathrm { k }$ avoids the convergence problem of the vanilla-Adam. Also, comparing Figure 2 and Figure 3, getting large enough samples prevents the gradient distribution from being distorted. These observations verify our hypothesis that the lack of sufficient data samples in the early stage is the root cause of the convergence issue. ", + "bbox": [ + 173, + 761, + 825, + 872 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another straightforward way to reduce the variance is to increase the value of $\\epsilon$ in $\\widehat { \\psi } ( g _ { 1 } , \\cdots , g _ { t } ) =$ $\\frac { \\sqrt { 1 - \\beta _ { 2 } ^ { t } } } { \\epsilon + \\sqrt { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } }$ Actually, if we assume $\\widehat { \\psi } ( . )$ is subject to the uniform distribution, its variance equals to $\\frac { 1 } { 1 2 \\epsilon ^ { 2 } }$ . Therefore, we design Adam-eps, which uses a non-negligibly large $\\epsilon = 1 0 ^ { - 4 }$ , while $\\epsilon = 1 0 ^ { - 8 }$ for vanilla Adam. Its performance is summarized in Figure 1. We observe that it does not suffer from the serious convergence problem of vanilla-Adam. This further demonstrates that the convergence problem can be alleviated by reducing the variance of the adaptive learning rate, and also explains why tuning $\\epsilon$ is important in practice (Liu et al., 2019). Besides, similar to Adam-2k, it prevents the gradient distribution from being distorted (as shown in Figure 3). However, as in Figure 1, it produces a much worse performance comparing to Adam- $2 \\mathrm { k }$ and Adam-warmup. We conjecture that this is because large $\\epsilon$ induces a large bias into the adaptive learning rate and slows down the optimization process. Thus, we need a more principled and rigorous way to control the variance of the adaptive learning rate. In the next subsection, we will present a theoretical analysis of the variance of the adaptive learning rate. ", + "bbox": [ + 173, + 102, + 825, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 ANALYSIS OF ADAPTIVE LEARNING RATE VARIANCE ", + "text_level": 1, + "bbox": [ + 173, + 318, + 586, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned before, Adam uses the exponential moving average to calculate the adaptive learning rate. For gradients $\\{ g _ { 1 } , \\cdots , g _ { t } \\}$ , their exponential moving average has a larger variance than their simple average. Also, in the early stage $\\mathbf { \\chi } _ { t }$ is small), the difference of the exponential weights of $\\{ g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } \\}$ is relatively small (up to $\\breve { 1 } - \\beta _ { 2 } ^ { t - 1 } ,$ ). Therefore, for ease of analysis, we approximate the distribution of the exponential moving average as the distribution of the simple average (Nau, 2014), i.e., p(ψ(.)) = p( $\\begin{array} { r } { p ( \\psi ( . ) ) = p ( \\sqrt { \\frac { 1 - \\beta _ { 2 } ^ { t } } { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } } ) \\approx p ( \\sqrt { \\frac { t } { \\sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } } ) } \\end{array}$ Since $g _ { i } \\sim \\mathcal { N } ( 0 , \\sigma ^ { 2 } )$ , we have ${ \\frac { t } { \\sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } } \\sim \\mathrm { S c a l e } { \\mathrm { - i n v } } { - \\mathcal { X } } ^ { 2 } ( t , { \\frac { 1 } { \\sigma ^ { 2 } } } )$ . Therefore, we assume $\\frac { 1 - \\beta _ { 2 } ^ { t } } { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } }$ also subjects to a scaled inverse chi-square distribution with $\\rho$ degrees of freedom (further analysis on this approximation is conducted in Section 5.3). Based on this assumption, we can calculate $\\mathrm { V a r } [ \\psi ^ { 2 } ( . ) ]$ and the PDF of $\\psi ^ { 2 } ( . )$ . Now, we proceed to the analysis of its square root variance, i.e., $\\mathrm { V a r } [ \\psi ( . ) ]$ , and show how the variance changes with $\\rho$ (which corresponds to number of used training samples). ", + "bbox": [ + 173, + 340, + 825, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. If $\\psi ^ { 2 } ( . ) \\sim$ Scale-inv- $\\scriptstyle { \\mathcal { X } } ^ { 2 } ( \\rho , { \\frac { 1 } { \\sigma ^ { 2 } } } )$ , $\\mathrm { V a r } [ \\psi ( . ) ]$ monotonically decreases as $\\rho$ increases. ", + "bbox": [ + 171, + 526, + 803, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. For $\\forall \\rho > 4$ , we have: ", + "bbox": [ + 174, + 556, + 367, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/ec5a58e45b0206f51a703fec8e95e0c988017c720c273353d4e10f20a19e4e08.jpg", + "text": "$$\n\\mathrm { V a r } [ \\psi ( . ) ] = \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] - \\mathbb { E } [ \\psi ( . ) ] ^ { 2 } = \\tau ^ { 2 } ( \\frac { \\rho } { \\rho - 2 } - \\frac { \\rho 2 ^ { 2 \\rho - 5 } } { \\pi } \\mathcal { B } ( \\frac { \\rho - 1 } { 2 } , \\frac { \\rho - 1 } { 2 } ) ^ { 2 } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 243, + 574, + 751, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $B ( . )$ is the beta function. By analyzing the derivative of $\\mathrm { V a r } [ \\psi ( . ) ]$ , we know it monotonically decreases as $\\rho$ increases. The detailed derivation is elaborated in the Appendix A. □ ", + "bbox": [ + 173, + 612, + 820, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1 gives a qualitative analysis of the variance of the adaptive learning rate. It shows that, due to the lack of used training samples in the early stage, $\\mathrm { V a r } [ \\bar { \\psi } ( . ) ]$ is larger than the late stage (Figure 8). To rigorously constraint the variance, we perform a quantified analysis on $\\mathrm { V a r } [ \\psi ( . ) ]$ by estimating the degree of freedoms $\\rho$ . ", + "bbox": [ + 173, + 650, + 825, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 RECTIFIED ADAPTIVE LEARNING RATE ", + "text_level": 1, + "bbox": [ + 174, + 724, + 535, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the previous section, Equation 2 gives the analytic form of $\\mathrm { V a r } [ \\psi ( . ) ]$ , where $\\rho$ is the degree of freedoms. Here, we first give an estimation of $\\rho$ based on $t$ to conduct a quantified analysis for $\\operatorname { V a r } [ \\psi ( g _ { 1 } , \\cdots , g _ { t } ) ]$ , then we describe the design of the learning rate rectification, and compare it to the heuristic warmup strategies. ", + "bbox": [ + 173, + 753, + 825, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1 ESTIMATION OF $\\rho$ ", + "text_level": 1, + "bbox": [ + 174, + 825, + 339, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The exponential moving average (EMA) can be interpreted as an approximation to the simple moving average (SMA) in real application (Nau, 2014), i.e., ", + "bbox": [ + 174, + 848, + 823, + 877 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/793d8562d11ed3ea916b4ad47f5781c1153cbebc559f3a479f34f31434576002.jpg", + "text": "$$\np \\left( \\frac { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } { 1 - \\beta _ { 2 } ^ { t } } \\right) \\approx p \\left( \\frac { \\sum _ { i = 1 } ^ { f \\left( t , \\beta _ { 2 } \\right) } g _ { t + 1 - i } ^ { 2 } } { f ( t , \\beta _ { 2 } ) } \\right) .\n$$", + "text_format": "latex", + "bbox": [ + 315, + 880, + 683, + 922 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 2: Rectified Adam. All operations are element-wise. ", + "bbox": [ + 173, + 107, + 594, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: $\\{ \\alpha _ { t } \\} _ { t = 1 } ^ { T }$ : step size, $\\{ \\beta _ { 1 } , \\beta _ { 2 } \\}$ : decay rate to calculate moving average and moving 2nd \nmoment, $\\theta _ { 0 }$ : initial parameter, $f _ { t } ( \\theta )$ : stochastic objective function. \nOutput: $\\theta _ { t }$ : resulting parameters \n1 $m _ { 0 } , v _ { 0 } \\gets 0 , 0$ (Initialize moving 1st and 2nd moment) \n2 $\\rho _ { \\infty } 2 / ( 1 - \\beta _ { 2 } ) - 1$ (Compute the maximum length of the approximated SMA) \n3 while $t = \\{ 1 , \\cdots , T \\}$ do \n4 $g _ { t } \\gets \\Delta _ { \\theta } f _ { t } ( \\theta _ { t - 1 } )$ (Calculate gradients w.r.t. stochastic objective at timestep t) \n5 $v _ { t } 1 / \\beta _ { 2 } v _ { t - 1 } + ( 1 - \\beta _ { 2 } ) g _ { t } ^ { 2 }$ (Update exponential moving 2nd moment) \n6 $m _ { t } \\gets \\beta _ { 1 } m _ { t - 1 } + ( 1 - \\beta _ { 1 } ) g _ { t }$ (Update exponential moving 1st moment) \n7 $\\widehat { m _ { t } } \\gets m _ { t } / ( 1 - \\beta _ { 1 } ^ { t } )$ (Compute bias-corrected moving average) \n8 $\\rho _ { t } \\gets \\rho _ { \\infty } - 2 t \\beta _ { 2 } ^ { t } / ( 1 - \\beta _ { 2 } ^ { t } )$ (Compute the length of the approximated SMA) \n9 if the variance is tractable, i.e., $\\rho _ { t } > 4$ then \n10 $l _ { t } \\sqrt { ( 1 - \\beta _ { 2 } ^ { t } ) / v _ { t } }$ (Compute adaptive learning rate) \n11 $\\begin{array} { r } { r _ { t } \\sqrt { \\frac { ( \\rho _ { t } - 4 ) ( \\rho _ { t } - 2 ) \\rho _ { \\infty } } { ( \\rho _ { \\infty } - 4 ) ( \\rho _ { \\infty } - 2 ) \\rho _ { t } } } } \\end{array}$ (Compute the variance rectification term) \n12 $\\theta _ { t } \\gets \\theta _ { t - 1 } - \\alpha _ { t } r _ { t } \\widehat { m _ { t } } l _ { t }$ (Update parameters with adaptive momentum) \n13 else \n14 $\\begin{array} { r } { \\big \\lfloor \\theta _ { t } \\gets \\theta _ { t - 1 } - \\alpha _ { t } \\widehat { m _ { t } } } \\end{array}$ (Update parameters with un-adapted momentum) ", + "bbox": [ + 158, + 118, + 808, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "15 return $\\theta _ { T }$ ", + "bbox": [ + 155, + 392, + 243, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f ( t , \\beta _ { 2 } )$ is the length of the SMA which allows the SMA to have the same “center of mass” with the EMA. In other words, $f ( t , \\beta _ { 2 } )$ satisfies: ", + "bbox": [ + 176, + 439, + 821, + 468 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/9250d1cc74971050e057c907265dd057a56c53c19498aa7f986c61efd12b2617.jpg", + "text": "$$\n\\frac { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } \\cdot i } { 1 - \\beta _ { 2 } ^ { t } } = \\frac { \\sum _ { i = 1 } ^ { f ( t , \\beta _ { 2 } ) } ( t + 1 - i ) } { f ( t , \\beta _ { 2 } ) } .\n$$", + "text_format": "latex", + "bbox": [ + 339, + 468, + 656, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By solving Equation 4, we have: $\\begin{array} { r l r } { f ( t , \\beta _ { 2 } ) } & { { } = } & { { \\frac { 2 } { 1 - \\beta _ { 2 } } } - 1 - { \\frac { 2 t \\beta _ { 2 } ^ { t } } { 1 - \\beta _ { 2 } ^ { t } } } } \\end{array}$ . In the previous section, we assume: $\\begin{array} { r } { \\frac { 1 - \\beta _ { 2 } ^ { t } } { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } \\ \\sim \\ \\mathrm { S c a l e - i n v - } \\mathcal { X } ^ { 2 } ( \\rho , \\frac { 1 } { \\sigma ^ { 2 } } ) . } \\end{array}$ . Here, since $g _ { i } \\sim \\mathcal { N } ( 0 , \\sigma ^ { 2 } )$ , we have 2) g2t+1−it,β2) ∼ Scale-inv-X 2(f (t, β2), 1σ2 ). Thus, Equation 3 views Scale-inv-X 2(f (t, β2), 1σ2 ) as an approximation to Scale-inv- $\\scriptstyle { \\mathcal { X } } ^ { 2 } ( \\rho , { \\frac { 1 } { \\sigma ^ { 2 } } } )$ . Therefore, we treat $f ( t , \\beta _ { 2 } )$ as an estimation of $\\rho$ . For ease of notation, we mark f (t, β2) as ρt. Also, we refer 21−β2 as $\\rho _ { \\infty }$ (maximum length of the approximated SMA), due to the inequality $\\begin{array} { r } { f ( t , \\beta _ { 2 } ) \\leq \\operatorname* { l i m } _ { t \\infty } f ( t , \\beta _ { 2 } ) = \\frac { 2 } { 1 - \\beta _ { 2 } } - 1 . } \\end{array}$ . ", + "bbox": [ + 173, + 508, + 826, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2 VARIANCE ESTIMATION AND RECTIFICATION ", + "text_level": 1, + "bbox": [ + 173, + 640, + 531, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on previous estimations, we have Var[ψ(.)] = τ 2( ρtρ −2 $\\begin{array} { r } { \\mathrm { V a r } [ \\psi ( . ) ] = \\tau ^ { 2 } ( \\frac { \\rho _ { t } } { \\rho _ { t } - 2 } - \\frac { \\rho _ { t } 2 ^ { 2 \\rho _ { t } - 5 } } { \\pi } \\mathcal { B } ( \\frac { \\rho _ { t } - 1 } { 2 } , \\frac { \\rho _ { t } - 1 } { 2 } ) ^ { 2 } ) } \\end{array}$ . The value of this function in the early stage is significantly larger than the late stage (as analyzed later, it decays roughly at the speed of $\\begin{array} { r } { \\dot { O } ( \\frac { 1 } { \\rho _ { t } } \\overline { { ) } } ) } \\end{array}$ . For example, the variance at $\\rho _ { t } = 5$ is over 100 times larger than the variance at $\\rho _ { t } ~ = ~ 5 0 0$ . Additionally, based on Theorem 1, we know $\\begin{array} { r } { \\operatorname* { m i n } _ { \\rho _ { t } } \\mathrm { V a r } [ \\psi ( . ) ] = } \\end{array}$ $\\mathrm { V a r } [ \\psi ( . ) ] | _ { \\rho _ { t } = \\rho _ { \\infty } }$ and mark this minimal value as $C _ { \\mathrm { V a r } }$ . In order to ensure that the adaptive learning rate $( \\psi ( . ) )$ has consistent variance, we rectify the variance at the $t$ -th timestamp as below, ", + "bbox": [ + 173, + 660, + 826, + 752 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/bfa5cdf77c87e44a8a93bc2010ec5df6421df1855b31d0f5551a951ec91161c1.jpg", + "text": "$$\n\\mathrm { V a r } [ r _ { t } \\psi ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } ) ] = C _ { \\mathrm { V a r } } \\quad \\mathrm { w h e r e } \\quad r _ { t } = \\sqrt { C _ { \\mathrm { V a r } } / \\mathrm { V a r } [ \\psi ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } ) ] } .\n$$", + "text_format": "latex", + "bbox": [ + 248, + 753, + 746, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although we have the analytic form of $\\mathrm { V a r } [ \\psi ( . ) ]$ (i.e., Equation 2), it is not numerically stable. Therefore, we use the first-order approximation to calculate the rectification term. Specifically, by approximating $\\sqrt { \\psi ^ { 2 } ( . ) }$ to the first order (Wolter, 2007), ", + "bbox": [ + 174, + 773, + 823, + 818 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/2bcbe7d69530063b78d06b38316d33f1623fb2c1ec18a55d5780992c1a9a7bd1.jpg", + "text": "$$\n\\sqrt { \\psi ^ { 2 } ( . ) } \\approx \\sqrt { \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] } + \\frac { 1 } { 2 \\sqrt { \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] } } ( \\psi ^ { 2 } ( . ) - \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] ) \\quad \\mathrm { a n d } \\quad \\mathrm { V a r } [ \\psi ( . ) ] \\approx \\frac { \\mathrm { V a r } [ \\psi ^ { 2 } ( . ) ] } { 4 \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] } .\n$$", + "text_format": "latex", + "bbox": [ + 197, + 819, + 799, + 856 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since $\\psi ^ { 2 } ( . ) \\sim \\mathrm { S c a l e } { \\mathrm { - i n v } { - } } \\mathcal { X } ^ { 2 } ( \\rho _ { t } , \\frac { 1 } { \\sigma ^ { 2 } } )$ , we have: ", + "bbox": [ + 174, + 856, + 480, + 873 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/7f29c5400faf9a111206eaa21ee57789c31d688b9384dae42e160d85e180cfb7.jpg", + "text": "$$\n\\mathrm { V a r } [ \\psi ( . ) ] \\approx \\rho _ { t } / [ 2 ( \\rho _ { t } - 2 ) ( \\rho _ { t } - 4 ) \\sigma ^ { 2 } ] .\n$$", + "text_format": "latex", + "bbox": [ + 369, + 875, + 629, + 892 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Section 5.3, we conduct simulation experiments to examine Equation 5 and find that it is a reliable approximation. Based on Equation 5, we know that $\\mathrm { V a r } [ \\sqrt { \\psi ( . ) } ]$ decreases approximately at the ", + "bbox": [ + 173, + 893, + 828, + 925 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2b06ae5fbfd40f478bad176ad4b8bc6635da54c9f4a2f28b5e5c18ea02d67a58.jpg", + "table_caption": [ + "Table 1: Image Classification " + ], + "table_footnote": [], + "table_body": "
MethodAcc.
CITIIIOSGDAdamRAdam91.5190.5491.38
eeeseSGDAdamRAdam69.8666.5467.62
", + "bbox": [ + 642, + 119, + 803, + 243 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8779d131e033f69524303bfd6e6b6c707693f95b1efbdf7a0772e68fcc51b91.jpg", + "image_caption": [ + "Figure 4: Language modeling (LSTMs) on the One Billion Word. " + ], + "image_footnote": [], + "bbox": [ + 174, + 106, + 622, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b8c26508ad736bb85b105e7794abce9ade7ded1d3b0ac44aa62399f7497eb5eb.jpg", + "image_caption": [ + "Figure 5: Training of ResNet-18 on the ImageNet and ResNet-20 on the CIFAR10 dataset. " + ], + "image_footnote": [], + "bbox": [ + 178, + 248, + 802, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "speed of $O ( \\textstyle { \\frac { 1 } { \\rho _ { t } } } )$ . With this approximation, we can calculate the rectification term as: ", + "bbox": [ + 174, + 387, + 722, + 401 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/8719717e482e66abfab474babc6767344179cb7f72e9817c59b381c68cbf46f8.jpg", + "text": "$$\nr _ { t } = \\sqrt { \\frac { ( \\rho _ { t } - 4 ) ( \\rho _ { t } - 2 ) \\rho _ { \\infty } } { ( \\rho _ { \\infty } - 4 ) ( \\rho _ { \\infty } - 2 ) \\rho _ { t } } } .\n$$", + "text_format": "latex", + "bbox": [ + 397, + 404, + 599, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Applying our rectification term to Adam, we come up with a new variant of Adam, Rectified Adam (RAdam), as summarized in Algorithm 2. Specifically, when the length of the approximated SMA is less or equal than 4, the variance of the adaptive learning rate is intractable and the adaptive learning rate is inactivated. Otherwise, we calculate the variance rectification term and update parameters with the adaptive learning rate. It is worth mentioning that, if $\\beta _ { 2 } ~ \\leq ~ 0 . 6$ , we have $\\rho _ { \\infty } \\leq 4$ and RAdam is degenerated to SGD with momentum. ", + "bbox": [ + 174, + 443, + 825, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3 IN COMPARISON WITH WARMUP AND OTHER STABILIZATION TECHNIQUES ", + "text_level": 1, + "bbox": [ + 178, + 529, + 736, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different from the analysis in this paper, warmup is originally proposed to handle training with very large batches for SGD (Goyal et al., 2017; Gotmare et al., 2019; Bernstein et al., 2018; Xiao et al., 2017). We notice that $r _ { t }$ has a similar form to the heuristic linear warmup, which can be viewed as setting the rectification term as $\\frac { m i n ( t , T _ { w } ) } { T _ { w } }$ . It verifies our intuition that warmup works as a variance reduction technique. RAdam deactivates the adaptive learning rate when its variance is divergent, thus avoiding undesired instability in the first few updates. Besides, our method does not require an additional hyperparameter (i.e., $T _ { w . }$ ) and can automatically adapt to different moving average rules. ", + "bbox": [ + 173, + 554, + 825, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, we identify and address an underlying issue of adaptive optimization methods independent of (neural) model architectures. Thus, the proposed rectification term is orthogonal to other training stabilization techniques such as gradient clipping (Bengio et al., 2013), smoothing the adaptive learning rate (i.e., increasing $\\epsilon$ , applying geometric mean filter (Chen & Gu, 2018), or adding range constraints (Luo et al., 2019)), initialization (Balduzzi et al., 2017; Zhang et al., 2019) and normalization (Ba et al., 2016; Ioffe & Szegedy, 2015). Indeed, these techniques can be combined with the proposed variance rectification method. ", + "bbox": [ + 174, + 662, + 825, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 776, + 326, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate RAdam on several benchmarks: One Billion Word for language modeling; Cifar10 and ImageNet for image classification; IWSLT’14 De-En/EN-DE and WMT’16 EN-De for neural machine translation. Following Loshchilov & Hutter (2018), we decouple weight decays in the vanilla Adam, Adam with warmup and RAdam in our experiments. Details are in Appendix B. ", + "bbox": [ + 174, + 804, + 825, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1 COMPARING TO VANILLA ADAM ", + "text_level": 1, + "bbox": [ + 176, + 873, + 441, + 887 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As analyzed before, the adaptive learning rate has undesirably large variance in the early stage of training and leads to suspicious/bad local optima on NMT. One question we are interested in is: whether such an issue widely exits in other similar tasks and applications. Thus, we conduct a set of experiments with two classical tasks of NLP and CV, i.e., language modeling and image classification. RAdam not only results in consistent improvements over the vanilla Adam, but also demonstrates its robustness to the change of learning rates. It verifies that the variance issue exists in various machine learning applications, and has a big impact on the model behavior. ", + "bbox": [ + 174, + 895, + 823, + 924 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6345f7afa7e7a63d9229d9a8dfe723d4ffb9cfb0f3992b2f052dd4d31c0db1fa.jpg", + "image_caption": [ + "Figure 6: Performance of RAdam, Adam and SGD with different learning rates on CIFAR10. " + ], + "image_footnote": [], + "bbox": [ + 194, + 109, + 802, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb74ac762ae180f94b6c6c55e5c1b8414301b44371f1f8fbfd2e1825dfee7518.jpg", + "image_caption": [ + "Figure 7: Performance of RAdam, Adam with warmup on CIFAR10 with different learning rates. " + ], + "image_footnote": [], + "bbox": [ + 174, + 310, + 825, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 496, + 825, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance Comparison. The performances on language modeling (i.e., One Billion Word (Chelba et al., 2013)) and image classification (i.e., CIFAR10 (Krizhevsky et al., 2009) and ImageNet (Deng et al., 2009)) are presented in Figure 4, 5. The results show that RAdam outperforms Adam in all three datasets. As shown in Figure 4, although the rectification term makes RAdam slower than the vanilla Adam in the first few epochs, it allows RAdam to converge faster after that. In other words, by reducing the variance of the adaptive learning rate in the early stage, it gets both faster convergence and better performance, which verifies the impact of the variance issue. We also observe that RAdam obtains consistent improvements over Adam on image classification. It is worth noting that, on both ImageNet and CIFAR10, although RAdam fails to outperform SGD in terms of test accuracy, it results in a better training performance (e.g., the training accuracy of SGD, Adam, and RAdam on ImageNet are 69.57, 69.12 and 70.30 respectively). ", + "bbox": [ + 173, + 573, + 825, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness to Learning Rate Change. Besides performance improvements, RAdam also improves the robustness of model training. We use different initial learning rates, conduct experiments with ResNet-20 on the CIFAR10 datasets, and summarize their performance in Figure 6. For learning rates within a broad range (i.e., $\\{ 0 . 1 , 0 . 0 3 , 0 . 0 1 , 0 . 0 0 3 \\} ,$ ), RAdam achieves consistent model performances (their test accuracy curves highly overlap with each other), while Adam and SGD are shown to be more sensitive to the learning rate. The observation can be interpreted that by rectifying the variance of the adaptive learning rate, RAdam improves the robustness of model training and can adapt to different learning rates of a broader range. ", + "bbox": [ + 174, + 732, + 825, + 843 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2 COMPARING TO HEURISTIC WARMUP ", + "text_level": 1, + "bbox": [ + 176, + 858, + 473, + 872 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To examine the effectiveness of RAdam, we first conduct comparisons on neural machine translation, on which the state-of-the-art employs Adam with the linear warmup. Specifically, we conduct experiments on three datasets, i.e., IWSLT’14 De-En, IWSLT’14 En-De, and WMT’16 En-De. Due to the limited size of the IWSLT’14 dataset, we conduct experiments using 5 different random seeds and report their mean and standard derivation. As discussed before, the vanilla Adam algorithm leads to suspicious/bad local optima (i.e., converges to a training perplexity around 500), and needs a learning rate warmup stage to stabilize the training. ", + "bbox": [ + 176, + 882, + 823, + 922 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a8fe3dfdbb40c1eb36ceeadac1a0d9d2b8896d7e820a4e3e7cf0047e895f9ed8.jpg", + "table_caption": [ + "Table 2: BLEU score on Neural Machine Translation. " + ], + "table_footnote": [], + "table_body": "
MethodIWSLT'14 DE-ENIWSLT'14 EN-DEWMT'16 EN-DE
Adam with warmupRAdam34.66± 0.01434.76 ± 0.00328.56± 0.06728.48 ± 0.05427.0327.27
", + "bbox": [ + 214, + 111, + 779, + 156 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 162, + 825, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We summarize the performance obtained with the heuristic warmup and our proposed rectification term in Table 2 and visualize the training curve of IWSLT De-En in Figure 1. With a consistent adaptive learning rate variance, our proposed method achieves similar performance to that of previous state-of-the-art warmup heuristics. It verifies our intuition that the problematic updates of Adam are indeed caused by the undesirably large variance in the early stage. ", + "bbox": [ + 174, + 224, + 825, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, we applied Adam with warmup on the CIFAR10 dataset. Its best accuracy on the test set is 91.29, which is similar to RAdam (91.38). However, we found that RAdam requires less hyperparameter tuning. Specifically, we visualize their learning curves in Figure 7. For some warmup steps, Adam with warmup is relatively more sensitive to the choice of the learning rate. RAdam, at the same time, is not only more robust, but also can automatically control the warmup behavior (i.e., without requiring the length of warmup). For example, when setting the learning rate as 0.1, Adam with 100 steps of warmup fails to get satisfying performance and only results in an accuracy of 90.13; RAdam successfully gets an accuracy of 91.06, with the original setting of the moving average calculation (i.e., $\\beta _ { 1 } = 0 . 9 , \\beta _ { 2 } = 0 . 9 9 9 )$ . We conjecture the reason is due to the fact that RAdam, which is based on a rigorous variance analysis, explicitly avoids the extreme situation where the variance is divergent, and rectifies the variance to be consistent in other situations. ", + "bbox": [ + 173, + 301, + 825, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3 SIMULATED VERIFICATION ", + "text_level": 1, + "bbox": [ + 176, + 472, + 403, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Sections 3 and 4, we approximate $\\textstyle \\operatorname { V a r } [ \\sqrt { t / \\sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } ]$ to the first order, and assume $\\psi ^ { 2 } ( . ) = $ $\\frac { 1 - \\beta _ { 2 } ^ { t } } { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } }$ subjects to a scaled inverse chi-square distribution (this assumption covers the approximation from EMA to SMA). Here, we examine these two approximations using simulations. ", + "bbox": [ + 173, + 496, + 825, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First Order Approximation of $\\textstyle \\operatorname { V a r } [ \\sqrt { t / \\sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } ]$ . To compare Equations 5 and 2, we assume $\\tau = 1$ and plot their values and difference for $\\nu = \\{ 5 , \\cdots , 5 0 0 \\}$ in Figure 8. The curve of the analytic form and the first-order approximation highly overlap, and their difference is much smaller than their value. This result verifies that our first-order approximation is very accurate. ", + "bbox": [ + 174, + 564, + 825, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scaled Inverse Chi-Square Distribution Assumption. In this paper, we assume $g _ { i }$ accords to a Normal distribution with a zero mean. We also assume $\\psi ^ { 2 } ( . )$ accords to the scaled inverse chi-square distribution to derive the variance of $\\mathrm { V a r } [ \\psi ( . ) ]$ , based on the similarity between the exponential moving average and simple moving average. Here, we empirically verify this assumption. ", + "bbox": [ + 174, + 636, + 825, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, since $g _ { i }$ in the optimization problem may not be zero-mean, we assume its expectation is $\\mu$ and sample $g _ { i }$ from $\\mathcal { N } ( \\bar { \\boldsymbol { \\mu } } , 1 )$ . Then, based on these samples, we calculate the variance of the original adaptive learning rate and the proposed rectified adaptive learning rate, i.e., $\\textstyle \\operatorname { V a r } [ { \\frac { 1 } { \\widehat { v _ { t } } } } ]$ and $\\textstyle \\operatorname { V a r } [ { \\frac { r _ { t } } { \\widehat { v _ { t } } } } ]$ respectively. We set $\\beta _ { 2 }$ bto 0.999, the number of sampled trajectories to 5000, the number bof iterations to 6000, and summarize the simulation results in Figure 9. Across all six settings with different $\\mu$ , the adaptive learning rate has a larger variance in the first stage and the rectified adaptive learning rate has relative consistent variance. This verifies the reliability of our assumption. ", + "bbox": [ + 173, + 699, + 825, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6 CONCLUSION ", + "text_level": 1, + "bbox": [ + 174, + 821, + 318, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we explore the underlying principle of the effectiveness of the warmup heuristic used for adaptive optimization algorithms. Specifically, we identify that, due to the limited amount of samples in the early stage of model training, the adaptive learning rate has an undesirably large variance and can cause the model to converge to suspicious/bad local optima. We provide both empirical and theoretical evidence to support our hypothesis, and further propose a new variant of Adam, whose adaptive learning rate is rectified so as to have a consistent variance. Empirical results demonstrate the effectiveness of our proposed method. In future work, we plan to replace the rectification strategy by sharing the second moment estimation across similar parameters. ", + "bbox": [ + 174, + 853, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/faa8a54dec9667c124d2828f50a5de48c77fb5bb8cb529cf12a5d90cfc18c0fa.jpg", + "image_caption": [ + "Figure 8: The value of Equation 2, Equation 5 and their difference (absolute difference). The $\\mathbf { X }$ -axis is $\\rho$ and the y-axis is the variance (log scale). " + ], + "image_footnote": [], + "bbox": [ + 176, + 122, + 415, + 257 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2c37e4a5e73292a2c4f76f3284121bfbb32e39ddd248373e6bf36a3c173b8e7f.jpg", + "image_caption": [ + "Figure 9: The simulation of $\\mathrm { V a r } [ \\frac { 1 } { v _ { t } } ]$ and $\\textstyle \\operatorname { V a r } [ { \\frac { c _ { t } } { v _ { t } } } ]$ . The $\\mathbf { X }$ -axis is iteration # (from 5), the y-axis is the variance (log scale). " + ], + "image_footnote": [], + "bbox": [ + 441, + 116, + 836, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 335, + 825, + 377 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ACKNOWLEDGE ", + "text_level": 1, + "bbox": [ + 176, + 391, + 310, + 405 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Aeyuan Allen-Zhu for valuable discussions and comments, Microsoft Research Technology Engineering team for setting up GPU machines. Research was sponsored in part by DARPA No. W911NF-17-C-0099 and FA8750-19-2-1004, National Science Foundation IIS 16-18481, IIS 17-04532, and IIS-17-41317, and DTRA HDTRA11810026. ", + "bbox": [ + 176, + 410, + 825, + 465 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 487, + 285, + 502 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. ", + "bbox": [ + 173, + 508, + 823, + 537 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "David Balduzzi, Marcus Frean, Lennox Leary, JP Lewis, Kurt Wan-Duo Ma, and Brian McWilliams. The shattered gradients problem: If resnets are the answer, then what is the question? In ICML, 2017. ", + "bbox": [ + 173, + 545, + 823, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Yoshua Bengio, Nicolas Boulanger-Lewandowski, and Razvan Pascanu. Advances in optimizing recurrent networks. In 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 8624–8628. IEEE, 2013. ", + "bbox": [ + 174, + 597, + 823, + 638 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jeremy Bernstein, Yu-Xiang Wang, Kamyar Azizzadenesheli, and Anima Anandkumar. signsgd: Compressed optimisation for non-convex problems. In ICML, 2018. ", + "bbox": [ + 171, + 647, + 821, + 676 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Augustin Cauchy. Methode g ´ en´ erale pour la r ´ esolution des systemes d ´ equations simultan ´ ees. ´ Comp. Rend. Sci. Paris, 25(1847):536–538, 1847. ", + "bbox": [ + 173, + 684, + 821, + 713 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Mauro Cettolo, Jan Niehues, Sebastian Stuker, Luisa Bentivogli, and Marcello Federico. Report on ¨ the 11th iwslt evaluation campaign, iwslt 2014. In Proceedings of the International Workshop on Spoken Language Translation,, 2014. ", + "bbox": [ + 173, + 719, + 823, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ciprian Chelba, Tomas Mikolov, Michael Schuster, Qi Ge, Thorsten Brants, Phillipp Koehn, and Tony Robinson. One billion word benchmark for measuring progress in statistical language modeling. In INTERSPEECH, 2013. ", + "bbox": [ + 173, + 771, + 823, + 814 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jinghui Chen and Quanquan Gu. Closing the generalization gap of adaptive gradient methods in training deep neural networks. arXiv preprint arXiv:1806.06763, 2018. ", + "bbox": [ + 171, + 821, + 823, + 851 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In ICML, 2009. ", + "bbox": [ + 173, + 858, + 821, + 887 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019. ", + "bbox": [ + 173, + 895, + 820, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Timothy Dozat. Incorporating nesterov momentum into adam. 2016. ", + "bbox": [ + 173, + 103, + 625, + 119 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. In COLT, 2010. ", + "bbox": [ + 173, + 126, + 823, + 155 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Carl-Friedrich Gauss. Theoria combinationis observationum erroribus minimis obnoxiae. Commentationes Societatis Regiae Scientiarum Gottingensis Recentiores, 1823. ", + "bbox": [ + 176, + 161, + 823, + 191 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Akhilesh Gotmare, Nitish Shirish Keskar, Caiming Xiong, and Richard Socher. A closer look at deep learning heuristics: Learning rate restarts, warmup and distillation. In ICLR, 2019. ", + "bbox": [ + 173, + 198, + 823, + 227 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Priya Goyal, Piotr Dollar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, An- ´ drew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. ", + "bbox": [ + 174, + 233, + 823, + 277 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. ", + "bbox": [ + 173, + 284, + 821, + 314 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Geoffrey Hinton, Nitish Srivastava, and Kevin Swersky. Neural networks for machine learning lecture 6a overview of mini-batch gradient descent. Cited on, 2012. ", + "bbox": [ + 173, + 320, + 821, + 351 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. ", + "bbox": [ + 174, + 357, + 820, + 386 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. ", + "bbox": [ + 173, + 393, + 810, + 409 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, Citeseer, 2009. ", + "bbox": [ + 171, + 415, + 821, + 444 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Liyuan Liu, Xiang Ren, Jingbo Shang, Jian Peng, and Jiawei Han. Efficient contextualized representation: Language model pruning for sequence labeling. EMNLP, 2018. ", + "bbox": [ + 171, + 452, + 821, + 481 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. ", + "bbox": [ + 178, + 487, + 821, + 531 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. In ICLR, 2018. ", + "bbox": [ + 173, + 537, + 795, + 554 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Liangchen Luo, Yuanhao Xiong, Yan Liu, and Xu Sun. Adaptive gradient methods with dynamic bound of learning rate. In ICLR, 2019. ", + "bbox": [ + 174, + 560, + 821, + 589 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Robert Nau. Forecasting with moving averages. 2014. ", + "bbox": [ + 173, + 597, + 532, + 612 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. fairseq: A fast, extensible toolkit for sequence modeling. In NAACL, 2019. ", + "bbox": [ + 173, + 619, + 823, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Martin Popel and Ondˇrej Bojar. Training tips for the transformer model. The Prague Bulletin of Mathematical Linguistics, 110(1):43–70, 2018. ", + "bbox": [ + 173, + 655, + 825, + 684 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sashank J Reddi, Satyen Kale, and Sanjiv Kumar. On the convergence of adam and beyond. In ICLR, 2018. ", + "bbox": [ + 173, + 690, + 823, + 720 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. ", + "bbox": [ + 171, + 727, + 823, + 757 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. ", + "bbox": [ + 171, + 763, + 821, + 794 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kirk M Wolter. Taylor series methods. In Introduction to variance estimation. 2007. ", + "bbox": [ + 173, + 800, + 730, + 815 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Lin Xiao, Adams Wei Yu, Qihang Lin, and Weizhu Chen. Dscovr: Randomized primal-dual block coordinate algorithms for asynchronous distributed optimization. J. Mach. Learn. Res., 2017. ", + "bbox": [ + 173, + 823, + 825, + 852 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Matthew D Zeiler. Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701, 2012. ", + "bbox": [ + 173, + 858, + 823, + 888 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Hongyi Zhang, Yann N Dauphin, and Tengyu Ma. Fixup initialization: Residual learning without normalization. In ICLR, 2019. ", + "bbox": [ + 174, + 895, + 821, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A PROOF OF THEOREM 1 ", + "text_level": 1, + "bbox": [ + 176, + 102, + 398, + 118 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For ease of notation, we refer $\\psi ^ { 2 } ( . )$ as $x$ and $\\textstyle { \\frac { 1 } { \\sigma ^ { 2 } } }$ as $\\tau ^ { 2 }$ . Thus, $x \\sim$ Scale-inv- $\\mathcal { X } ^ { 2 } ( \\rho , \\tau ^ { 2 } )$ and: ", + "bbox": [ + 169, + 131, + 772, + 148 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/a68306e481a2e257390d61e1e82b658f0b513631373ef01210937ac2b15d1d5c.jpg", + "text": "$$\np ( x ) = \\frac { ( \\tau ^ { 2 } \\rho / 2 ) ^ { \\rho / 2 } } { \\Gamma ( \\rho / 2 ) } \\frac { \\exp \\bigl [ \\frac { - \\rho \\tau ^ { 2 } } { 2 x } \\bigr ] } { x ^ { 1 + \\rho / 2 } } \\quad \\mathrm { a n d } \\quad \\mathbb { E } [ x ] = \\frac { \\rho } { ( \\rho - 2 ) \\sigma ^ { 2 } } \\left( \\forall \\rho > 2 \\right)\n$$", + "text_format": "latex", + "bbox": [ + 271, + 150, + 727, + 186 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $\\Gamma ( . )$ is the gamma function. Therefore, we have: ", + "bbox": [ + 176, + 185, + 535, + 199 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/f6a97576169a32c6305a6ee627ea0b484b90eb44fab15f8c21938439fed9c40e.jpg", + "text": "$$\n\\mathbb { E } [ \\sqrt { x } ] = \\int _ { 0 } ^ { \\infty } \\sqrt { x } p ( x ) d x = \\frac { \\tau \\sqrt { \\rho } \\Gamma ( ( \\rho - 1 ) / 2 ) } { \\sqrt { 2 } \\Gamma ( \\rho / 2 ) } ( \\forall \\rho > 4 ) .\n$$", + "text_format": "latex", + "bbox": [ + 300, + 198, + 696, + 234 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Based on Equation 6 and 7, for $\\forall \\rho > 4$ , we have: ", + "bbox": [ + 173, + 233, + 500, + 247 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/4731c0000ffbb79e62b7117d5e904f2d8871c0e1860b8538b72d763ea671e1c1.jpg", + "text": "$$\n\\operatorname { V a r } [ \\psi ( \\cdot ) ] = \\operatorname { V a r } [ { \\sqrt { x } } ] = \\operatorname { \\mathbb { E } } [ x ] - \\operatorname { \\mathbb { E } } [ { \\sqrt { x } } ] ^ { 2 } = \\tau ^ { 2 } ( { \\frac { \\rho } { \\rho - 2 } } - { \\frac { \\rho 2 ^ { 2 \\rho - 5 } } { \\pi } } B ( { \\frac { \\rho - 1 } { 2 } } , { \\frac { \\rho - 1 } { 2 } } ) ^ { 2 } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 220, + 244, + 776, + 279 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $B ( . )$ is the beta function. To prove the monotonic property of $\\mathrm { V a r } [ \\psi ( . ) ]$ , we need to show: ", + "bbox": [ + 176, + 279, + 807, + 294 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Lemma 1. for $t \\geq 4$ $\\begin{array} { r } { , \\frac { \\partial } { \\partial t } \\big ( \\frac { t } { t - 2 } - \\frac { t 2 ^ { 2 t - 5 } } { \\pi } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } \\big ) < 0 } \\end{array}$ ", + "bbox": [ + 173, + 299, + 560, + 321 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Proof. The target inequality can be re-wrote as ", + "bbox": [ + 174, + 334, + 483, + 348 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/933b90ff22dd689badad34a6560dc3d45725acedc2e44a4efa8630b010dc48e9.jpg", + "text": "$$\n\\begin{array} { r l } & { \\frac { \\partial } { \\partial t } ( \\frac { t } { t - 2 } - \\frac { t \\dot { 2 } ^ { 2 t - 5 } } { \\pi } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } ) } \\\\ & { = \\frac { - 2 } { ( t - 2 ) ^ { 2 } } - \\frac { 2 ^ { 2 t - 5 } } { \\pi } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } - \\frac { t 2 ^ { 2 t - 5 } \\ln { 4 } } { \\pi } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } \\\\ & { - \\frac { 2 t 2 ^ { 2 t - 5 } } { \\pi } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } ( \\Psi ( \\frac { t - 1 } { 2 } ) - \\Psi ( t - 1 ) ) , \\quad \\left( \\Psi ( x ) = \\frac { \\Gamma ^ { \\prime } ( x ) } { \\Gamma ( x ) } \\right) } \\\\ & { < 0 } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 251, + 347, + 743, + 468 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "This inequality is equivalent to: ", + "bbox": [ + 174, + 486, + 382, + 501 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/a2a70ac3e31572d07fa579594e5e8c99c30ff3fd34366528b34a910690d7a0da.jpg", + "text": "$$\n\\begin{array} { r l r } & { } & { \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } + 1 + t \\ln 4 + 2 t \\Psi ( \\frac { t - 1 } { 2 } ) } \\\\ & { } & { > 2 t \\Psi ( t - 1 ) \\overset { ( i ) } { = } t [ \\Psi ( \\frac { t - 1 } { 2 } ) + \\Psi ( \\frac { t } { 2 } ) + \\ln 4 ] , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 326, + 498, + 669, + 564 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $( i )$ is derived from Legendre duplication formula. Simplify the above inequality, we get: ", + "bbox": [ + 173, + 563, + 794, + 577 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/373a41f5932e3ffef8ce1ac227a41b1b954c78ade3aa31f36219c7646156a52d.jpg", + "text": "$$\n\\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } + 1 + t \\Psi ( \\frac { t - \\hat { 1 } } { 2 } ) - t \\Psi ( \\frac { t } { 2 } ) > 0 ,\n$$", + "text_format": "latex", + "bbox": [ + 308, + 575, + 687, + 609 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We only need to show ", + "bbox": [ + 174, + 616, + 320, + 630 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/daac562cbae9f2bac8ef57763e4772f95a5f145921747210700cb77d99e8242b.jpg", + "text": "$$\n\\begin{array} { r l } & { \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } + 1 + t \\Psi ( \\frac { t - 1 } { 2 } ) - t \\Psi ( \\frac { t } { 2 } ) } \\\\ & { \\geq \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } + 2 + t ( \\ln ( t / 2 ) - 1 / ( t / 2 - 0 . 5 ) ) - t \\ln ( t / 2 ) } \\\\ & { = \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } - \\frac { 2 } { t - 1 } } \\\\ & { > \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } - \\frac { 2 } { t - 2 } \\geq 0 , } \\\\ & { \\mathrm { r s t i n e q u a l i t y i s i f r o m } \\ln ( x ) - 1 / ( 2 x ) > \\Psi ( x ) > \\ln ( x + 0 . 5 ) - 1 / x . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 253, + 627, + 746, + 782 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where the fi ", + "bbox": [ + 173, + 767, + 699, + 784 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Therefore, we only need to show ", + "bbox": [ + 174, + 789, + 392, + 803 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/a8b52e9f9e5400edde8f198fd0d634374bb233b34ebc0185c2fb2cecbd3c17a0.jpg", + "text": "$$\n3 2 \\pi \\geq ( t - 2 ) 4 ^ { t } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } ,\n$$", + "text_format": "latex", + "bbox": [ + 383, + 804, + 611, + 830 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "which is equivalent to ", + "bbox": [ + 174, + 837, + 320, + 851 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/9c064b4557cec0de11192a2e3afddf9644489286614bd06d38e5f5cee88ba410.jpg", + "text": "$$\n\\begin{array} { l } { ( t - 2 ) 4 ^ { t } \\mathcal { B } ( \\displaystyle \\frac { t - 1 } { 2 } , \\displaystyle \\frac { t - 1 } { 2 } ) ^ { 2 } = ( t - 2 ) 4 ^ { t } \\frac { \\Gamma ( \\displaystyle \\frac { t - 1 } { 2 } ) ^ { 4 } } { \\Gamma ( t - 1 ) ^ { 2 } } } \\\\ { \\overset { ( i ) } { = } ( t - 2 ) 4 ^ { t } \\frac { \\Gamma ( \\displaystyle \\frac { t - 1 } { 2 } ) ^ { 2 } } { \\Gamma ( t / 2 ) ^ { 2 } } 4 ^ { 2 - t } \\pi = 1 6 \\pi ( t - 2 ) \\frac { \\Gamma ( \\displaystyle \\frac { t - 1 } { 2 } ) ^ { 2 } } { \\Gamma ( t / 2 ) ^ { 2 } } \\le 3 2 \\pi , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 302, + 852, + 696, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where (i) is from Legendre duplication formula. ", + "bbox": [ + 174, + 103, + 490, + 118 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "So we only need to show ", + "bbox": [ + 174, + 125, + 339, + 140 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/d56e8f3aad70b54c98bf6133479504e8efa59f49a3d2ddac5840de4e4efcfe5b.jpg", + "text": "$$\n( t - 2 ) \\frac { \\Gamma ( \\frac { t - 1 } { 2 } ) ^ { 2 } } { \\Gamma ( t / 2 ) ^ { 2 } } \\leq 2\n$$", + "text_format": "latex", + "bbox": [ + 428, + 136, + 570, + 172 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Using Gautschi’s inequality $( \\frac { \\Gamma ( x + 1 ) } { \\Gamma ( x + s ) } < ( x + 1 ) ^ { 1 - s } )$ , we have ", + "bbox": [ + 174, + 178, + 575, + 198 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/f18994052fffccd4c8bab55617eb816a201046261b1cb00d5677631ff50f31c9.jpg", + "text": "$$\n( t - 2 ) \\frac { \\dot { \\Gamma } ( \\frac { t - 1 } { 2 } ) ^ { 2 } } { \\Gamma ( t / 2 ) ^ { 2 } } \\leq ( t - 2 ) ( \\frac { t - 1 } { 2 } ) ^ { - 1 } = \\frac { 2 ( t - 2 ) } { t - 1 } < 2\n$$", + "text_format": "latex", + "bbox": [ + 321, + 198, + 678, + 233 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "B IMPLEMENTATION DETAILS ", + "text_level": 1, + "bbox": [ + 176, + 263, + 439, + 280 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "B.1 LANGUAGE MODELING ", + "text_level": 1, + "bbox": [ + 174, + 295, + 380, + 309 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our implementation is based on the previous work (Liu et al., 2018). Specifically, we use two-layer LSTMs with 2048 hidden states with adaptive softmax to conduct experiments on the one billion words dataset. Word embedding (random initialized) of 300 dimensions is used as the input and the adaptive softmax is incorporated with a default setting (cut-offs are set to [4000, 40000, 200000]). Additionally, as pre-processing, we replace all tokens occurring equal or less than 3 times with as UNK, which shrinks the dictionary from 7.9M to 6.4M. Dropout is applied to each layer with a ratio of 0.1, gradients are clipped at 5.0. We use the default hyper-parameters to update moving averages, $i . e . \\beta _ { 1 } = 0 . 9$ and $\\beta _ { 2 } = 0 . 9 9 9$ . The learning rate is set to start from 0.001, and decayed at the start of 10th epochs. LSTMs are unrolled for 20 steps without resetting the LSTM states and the batch size is set to 128. All models are trained on one NVIDIA Tesla V100 GPU. ", + "bbox": [ + 173, + 320, + 825, + 459 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "B.2 IMAGEINE CLASSIFICATION ", + "text_level": 1, + "bbox": [ + 176, + 477, + 410, + 491 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We use the default ResNet architectures (He et al., 2016) in a public pytorch re-implementation4. Specifically, we use 20-layer ResNet (9 Basic Blocks) for CIFAR-10 and 18-layer ResNet (8 Basic Blocks) for ImageNet. Batch size is 128 for CIFAR-10 and 256 for ImageNet. The model is trained for 186 epoches and the learning rate decays at the 81-th and the 122-th epoches by 0.1 on CIFAR10, while the model is trained for 90 epoches and the learning rate decays at the 31-th and the 61-th epoch by 0.1 on ImageNet. For Adam and RAdam, we set $\\beta _ { 1 } = 0 . 9 , \\beta _ { 2 } = 0 . 9 9 9$ . For SGD, we set the momentum factor as 0.9. The weight decay rate is $1 0 ^ { - 4 }$ . Random cropping and random horizontal flipping are applied to training data. ", + "bbox": [ + 174, + 502, + 825, + 613 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "B.3 NEURAL MACHINE TRANSLATION ", + "text_level": 1, + "bbox": [ + 176, + 630, + 454, + 643 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our experiments are based on the default Transformers (Vaswani et al., 2017) implementation from the fairseq package (Ott et al., 2019). Specifically, we use word embedding with 512 dimensions and 6-layer encoder / decoder with 4 head and 1024 hidden dimensions on the IWSLT14’ dataset; use word embedding with 512 dimension and 6-layer encoder / decoder with 8 heads and 2048 hidden dimensions. Label smoothed cross entropy is used as the objective function with an uncertainty $=$ 0.1 (Szegedy et al., 2016). We use linear learning rate decay starting from $3 e ^ { - 4 }$ , and the checkpoints of the last 20 epoches are averaged before evaluation. As to the wamrup strategy, we use a linear warmup for Adam in the first 4000 updates, and set $\\beta _ { 2 }$ to satisfy $\\nu = 4 0 0 0$ $\\beta _ { 2 } = 0 . 9 9 9 5 )$ . In the IWSLT’14 dataset, we conduct training on one NVIDIA Tesla V100 GPU, set maximum batch size as 4000, apply dropout with a ratio 0.3, using weight decay of 0.0001 and clip the gradient norm at 25. In the WMT’16 dataset, we conduct training on four NVIDIA Quadro R8000 GPUs and set maximum batch size as 8196. ", + "bbox": [ + 173, + 656, + 825, + 821 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "C DOWNGRADING TO SGDM ", + "text_level": 1, + "bbox": [ + 176, + 842, + 434, + 858 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "As a byproduct determined by math derivations, we degenerated RAdam to SGD with momentum in the first several updates. Although this stage only contains several gradient updates, these updates could be quite damaging (e.g., in our Figure 2, the gradient distribution is distorted within 10 gradient updates). Intuitively, updates with divergent adaptive learning rate variance could be more damaging than the ones with converged variance, as divergent variance implies more instability. As a case study, we performed experiments on the CIFAR10 dataset. Five-run average results are summarized in Table 3. The optimizer fails to get an equally reliably model when changing the first 4 updates to Adam, yet the influence of switching is less deleterious when we change 5-8 updates instead. This result verifies our intuition and is in agreement with our theory the first few updates could be more damaging than later updates. By saying that, we still want to emphasize that this part (downgrading to SGDM) is only a minor part of our algorithm design whereas our main focus is on the mechanism of warmup and the derivation of the rectification term. ", + "bbox": [ + 178, + 873, + 823, + 901 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 825, + 242 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ad11f900d73d2167f23839df7c4d72b8c5485331e01abbf9ed891a61f460d72f.jpg", + "table_caption": [ + "Table 3: Performance on CIFAR10 $( \\mathrm { l r } = 0 . 1 $ ). " + ], + "table_footnote": [], + "table_body": "
1-4 steps5-8 steps8+ stepstest acctrain losstrain error
RAdamRAdamRAdam91.080.0210.74
Adam (w. divergent var.)RAdamRAdam89.980.0602.12
SGDAdam (w. convergent var.)RAdam90.290.0381.23
", + "bbox": [ + 173, + 287, + 825, + 387 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/parse/train/rkgz2aEKDr/rkgz2aEKDr_middle.json b/parse/train/rkgz2aEKDr/rkgz2aEKDr_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..9059ec4db05d61e9c07aa951f3a8954f5d1c6d3a --- /dev/null +++ b/parse/train/rkgz2aEKDr/rkgz2aEKDr_middle.json @@ -0,0 +1,37988 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 79, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 98 + ], + "score": 1.0, + "content": "ON THE VARIANCE OF THE ADAPTIVE LEARNING", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 98, + 255, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 255, + 118 + ], + "score": 1.0, + "content": "RATE AND BEYOND", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 135, + 264, + 168 + ], + "lines": [ + { + "bbox": [ + 111, + 135, + 173, + 146 + ], + "spans": [ + { + "bbox": [ + 111, + 135, + 173, + 146 + ], + "score": 1.0, + "content": "Liyuan Liu ∗", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 264, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 264, + 159 + ], + "score": 1.0, + "content": "University of Illinois, Urbana-Champaign", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 158, + 165, + 168 + ], + "spans": [ + { + "bbox": [ + 112, + 158, + 165, + 168 + ], + "score": 1.0, + "content": "ll2@illinois", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 281, + 135, + 357, + 168 + ], + "lines": [ + { + "bbox": [ + 280, + 135, + 349, + 148 + ], + "spans": [ + { + "bbox": [ + 280, + 135, + 349, + 148 + ], + "score": 1.0, + "content": "Haoming Jiang", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 281, + 147, + 331, + 157 + ], + "spans": [ + { + "bbox": [ + 281, + 147, + 331, + 157 + ], + "score": 1.0, + "content": "Georgia Tech", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 281, + 159, + 358, + 168 + ], + "spans": [ + { + "bbox": [ + 281, + 159, + 358, + 168 + ], + "score": 1.0, + "content": "jianghm@gatech.edu", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 373, + 135, + 498, + 168 + ], + "lines": [ + { + "bbox": [ + 374, + 136, + 498, + 147 + ], + "spans": [ + { + "bbox": [ + 374, + 136, + 498, + 147 + ], + "score": 1.0, + "content": "Pengcheng He, Weizhu Chen", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 373, + 147, + 478, + 158 + ], + "spans": [ + { + "bbox": [ + 373, + 147, + 478, + 158 + ], + "score": 1.0, + "content": "Microsoft Dynamics 365 AI", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 374, + 159, + 493, + 170 + ], + "spans": [ + { + "bbox": [ + 374, + 159, + 493, + 170 + ], + "score": 1.0, + "content": "{penhe,wzchen}@microsoft.com", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 112, + 185, + 234, + 218 + ], + "lines": [ + { + "bbox": [ + 113, + 186, + 235, + 197 + ], + "spans": [ + { + "bbox": [ + 113, + 186, + 235, + 197 + ], + "score": 1.0, + "content": "Xiaodong Liu, Jianfeng Gao", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 112, + 197, + 186, + 207 + ], + "spans": [ + { + "bbox": [ + 112, + 197, + 186, + 207 + ], + "score": 1.0, + "content": "Microsoft Research", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 113, + 209, + 232, + 219 + ], + "spans": [ + { + "bbox": [ + 113, + 209, + 232, + 219 + ], + "score": 1.0, + "content": "{xiaodl,jfgao}@microsoft.com", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 295, + 186, + 447, + 218 + ], + "lines": [ + { + "bbox": [ + 294, + 185, + 346, + 197 + ], + "spans": [ + { + "bbox": [ + 294, + 185, + 346, + 197 + ], + "score": 1.0, + "content": "Jiawei Han", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 195, + 448, + 209 + ], + "spans": [ + { + "bbox": [ + 295, + 195, + 448, + 209 + ], + "score": 1.0, + "content": "University of Illinois, Urbana-Champaign", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 209, + 352, + 218 + ], + "spans": [ + { + "bbox": [ + 295, + 209, + 352, + 218 + ], + "score": 1.0, + "content": "hanj@illinois", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 278, + 248, + 333, + 259 + ], + "lines": [ + { + "bbox": [ + 276, + 246, + 336, + 262 + ], + "spans": [ + { + "bbox": [ + 276, + 246, + 336, + 262 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 143, + 274, + 469, + 395 + ], + "lines": [ + { + "bbox": [ + 141, + 273, + 469, + 288 + ], + "spans": [ + { + "bbox": [ + 141, + 273, + 469, + 288 + ], + "score": 1.0, + "content": "The learning rate warmup heuristic achieves remarkable success in stabilizing", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 285, + 469, + 299 + ], + "spans": [ + { + "bbox": [ + 141, + 285, + 469, + 299 + ], + "score": 1.0, + "content": "training, accelerating convergence and improving generalization for adaptive", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 296, + 469, + 309 + ], + "spans": [ + { + "bbox": [ + 141, + 296, + 469, + 309 + ], + "score": 1.0, + "content": "stochastic optimization algorithms like RMSprop and Adam. Pursuing the theory", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 307, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 141, + 307, + 469, + 320 + ], + "score": 1.0, + "content": "behind warmup, we identify a problem of the adaptive learning rate – its vari-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 318, + 470, + 331 + ], + "spans": [ + { + "bbox": [ + 141, + 318, + 470, + 331 + ], + "score": 1.0, + "content": "ance is problematically large in the early stage, and presume warmup works as a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 142, + 330, + 469, + 341 + ], + "spans": [ + { + "bbox": [ + 142, + 330, + 469, + 341 + ], + "score": 1.0, + "content": "variance reduction technique. We provide both empirical and theoretical evidence", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 340, + 470, + 353 + ], + "spans": [ + { + "bbox": [ + 141, + 340, + 470, + 353 + ], + "score": 1.0, + "content": "to verify our hypothesis. We further propose Rectified Adam (RAdam), a novel", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 142, + 352, + 469, + 363 + ], + "spans": [ + { + "bbox": [ + 142, + 352, + 469, + 363 + ], + "score": 1.0, + "content": "variant of Adam, by introducing a term to rectify the variance of the adaptive", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 361, + 469, + 375 + ], + "spans": [ + { + "bbox": [ + 141, + 361, + 469, + 375 + ], + "score": 1.0, + "content": "learning rate. Experimental results on image classification, language modeling,500550", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 372, + 469, + 386 + ], + "spans": [ + { + "bbox": [ + 141, + 372, + 469, + 386 + ], + "score": 1.0, + "content": "and neural machine translation verify our intuition and demonstrate the efficacy400450", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 142, + 384, + 253, + 395 + ], + "spans": [ + { + "bbox": [ + 142, + 384, + 253, + 395 + ], + "score": 1.0, + "content": "and robustness of RAdam.1", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 108, + 421, + 205, + 434 + ], + "lines": [ + { + "bbox": [ + 105, + 420, + 208, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 208, + 437 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 108, + 445, + 356, + 577 + ], + "lines": [ + { + "bbox": [ + 106, + 444, + 357, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 357, + 457 + ], + "score": 1.0, + "content": "Fast and stable optimization algorithms are what generations", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 455, + 357, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 357, + 469 + ], + "score": 1.0, + "content": "of researchers have been pursuing (Gauss, 1823; Cauchy,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 466, + 356, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 356, + 479 + ], + "score": 1.0, + "content": "1847). Remarkably, stochastic gradient-based optimization,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 478, + 357, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 357, + 490 + ], + "score": 1.0, + "content": "such as stochastic gradient descent (SGD), has witnessed", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 488, + 357, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 357, + 501 + ], + "score": 1.0, + "content": "tremendous success in many fields of science and engineering", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 500, + 357, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 357, + 512 + ], + "score": 1.0, + "content": "despite its simplicity. Recently, many efforts have been made", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 511, + 357, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 357, + 523 + ], + "score": 1.0, + "content": "to accelerate optimization by applying adaptive learning rate.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 520, + 357, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 357, + 535 + ], + "score": 1.0, + "content": "In particular, Adagrad (Duchi et al., 2010) and its variants, e.g.,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 532, + 357, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 357, + 545 + ], + "score": 1.0, + "content": "RMSprop (Hinton et al., 2012), Adam (Kingma & Ba, 2014),", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 543, + 357, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 357, + 555 + ], + "score": 1.0, + "content": "Adadelta (Zeiler, 2012) and Nadam (Dozat, 2016), stand out", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 554, + 357, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 357, + 567 + ], + "score": 1.0, + "content": "due to their fast convergence, and have been considered as the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 566, + 274, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 274, + 578 + ], + "score": 1.0, + "content": "optimizer of choice in many applications.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 35.5 + }, + { + "type": "image", + "bbox": [ + 365, + 437, + 503, + 539 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 365, + 437, + 503, + 539 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 365, + 437, + 503, + 539 + ], + "spans": [ + { + "bbox": [ + 365, + 437, + 503, + 539 + ], + "score": 0.965, + "type": "image", + "image_path": "9e07baec4bf47bd9075fa89a9a4a4cab421e0a129a16eb042e7e1b7d64da16fe.jpg" + } + ] + } + ], + "index": 42.5, + "virtual_lines": [ + { + "bbox": [ + 365, + 437, + 503, + 488.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 365, + 488.0, + 503, + 539.0 + ], + "spans": [], + "index": 43 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 363, + 542, + 505, + 575 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 363, + 542, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 363, + 542, + 506, + 554 + ], + "score": 1.0, + "content": "Figure 1: Training loss v.s. # of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 363, + 553, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 363, + 553, + 505, + 565 + ], + "score": 1.0, + "content": "iterations of Transformers on the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 363, + 563, + 470, + 576 + ], + "spans": [ + { + "bbox": [ + 363, + 563, + 470, + 576 + ], + "score": 1.0, + "content": "De-En IWSLT’14 dataset.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "index": 43.75 + }, + { + "type": "text", + "bbox": [ + 107, + 582, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 582, + 504, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 504, + 595 + ], + "score": 1.0, + "content": "However, it has been observed that these optimization methods may converge to bad/suspicious", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "local optima, and have to resort to a warmup heuristic – using a small learning rate in the first", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "few epochs of training to mitigate such problem (Vaswani et al., 2017; Popel & Bojar, 2018). For", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 616, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 627 + ], + "score": 1.0, + "content": "example, when training typical Transformers based neural machine translation models on the De-En", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "IWSLT’14 dataset, removing the warmup stage increases the training loss from 3 to around 10, as", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 638, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 649 + ], + "score": 1.0, + "content": "shown in Figure 1. Similar phenomena are observed in other scenarios like BERT (a bidirectional", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 649, + 359, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 359, + 661 + ], + "score": 1.0, + "content": "transformer language model) pre-training (Devlin et al., 2019).", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 50 + }, + { + "type": "text", + "bbox": [ + 107, + 665, + 504, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 664, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 679 + ], + "score": 1.0, + "content": "Duo to the lack of the theoretical underpinnings, there is neither guarantee that warmup would bring", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 676, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 688 + ], + "score": 1.0, + "content": "consistent improvements for various machine learning settings nor guidance on how we should", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 54.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 699, + 472, + 732 + ], + "lines": [ + { + "bbox": [ + 117, + 698, + 303, + 711 + ], + "spans": [ + { + "bbox": [ + 117, + 698, + 303, + 711 + ], + "score": 1.0, + "content": "∗Work was done during an internship at Microsoft.", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 708, + 303, + 722 + ], + "spans": [ + { + "bbox": [ + 118, + 708, + 303, + 722 + ], + "score": 1.0, + "content": "†Work was done during an internship at Microsoft.", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 719, + 472, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 472, + 734 + ], + "score": 1.0, + "content": "1All implementations are available at: https://github.com/LiyuanLucasLiu/RAdam.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 79, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 98 + ], + "score": 1.0, + "content": "ON THE VARIANCE OF THE ADAPTIVE LEARNING", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 98, + 255, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 255, + 118 + ], + "score": 1.0, + "content": "RATE AND BEYOND", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "list", + "bbox": [ + 113, + 135, + 264, + 168 + ], + "lines": [ + { + "bbox": [ + 111, + 135, + 173, + 146 + ], + "spans": [ + { + "bbox": [ + 111, + 135, + 173, + 146 + ], + "score": 1.0, + "content": "Liyuan Liu ∗", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 146, + 264, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 264, + 159 + ], + "score": 1.0, + "content": "University of Illinois, Urbana-Champaign", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 158, + 165, + 168 + ], + "spans": [ + { + "bbox": [ + 112, + 158, + 165, + 168 + ], + "score": 1.0, + "content": "ll2@illinois", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + } + ], + "index": 3, + "bbox_fs": [ + 111, + 135, + 264, + 168 + ] + }, + { + "type": "text", + "bbox": [ + 281, + 135, + 357, + 168 + ], + "lines": [ + { + "bbox": [ + 280, + 135, + 349, + 148 + ], + "spans": [ + { + "bbox": [ + 280, + 135, + 349, + 148 + ], + "score": 1.0, + "content": "Haoming Jiang", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 281, + 147, + 331, + 157 + ], + "spans": [ + { + "bbox": [ + 281, + 147, + 331, + 157 + ], + "score": 1.0, + "content": "Georgia Tech", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 281, + 159, + 358, + 168 + ], + "spans": [ + { + "bbox": [ + 281, + 159, + 358, + 168 + ], + "score": 1.0, + "content": "jianghm@gatech.edu", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 280, + 135, + 358, + 168 + ] + }, + { + "type": "text", + "bbox": [ + 373, + 135, + 498, + 168 + ], + "lines": [ + { + "bbox": [ + 374, + 136, + 498, + 147 + ], + "spans": [ + { + "bbox": [ + 374, + 136, + 498, + 147 + ], + "score": 1.0, + "content": "Pengcheng He, Weizhu Chen", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 373, + 147, + 478, + 158 + ], + "spans": [ + { + "bbox": [ + 373, + 147, + 478, + 158 + ], + "score": 1.0, + "content": "Microsoft Dynamics 365 AI", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 374, + 159, + 493, + 170 + ], + "spans": [ + { + "bbox": [ + 374, + 159, + 493, + 170 + ], + "score": 1.0, + "content": "{penhe,wzchen}@microsoft.com", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 373, + 136, + 498, + 170 + ] + }, + { + "type": "text", + "bbox": [ + 112, + 185, + 234, + 218 + ], + "lines": [ + { + "bbox": [ + 113, + 186, + 235, + 197 + ], + "spans": [ + { + "bbox": [ + 113, + 186, + 235, + 197 + ], + "score": 1.0, + "content": "Xiaodong Liu, Jianfeng Gao", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 112, + 197, + 186, + 207 + ], + "spans": [ + { + "bbox": [ + 112, + 197, + 186, + 207 + ], + "score": 1.0, + "content": "Microsoft Research", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 113, + 209, + 232, + 219 + ], + "spans": [ + { + "bbox": [ + 113, + 209, + 232, + 219 + ], + "score": 1.0, + "content": "{xiaodl,jfgao}@microsoft.com", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 112, + 186, + 235, + 219 + ] + }, + { + "type": "list", + "bbox": [ + 295, + 186, + 447, + 218 + ], + "lines": [ + { + "bbox": [ + 294, + 185, + 346, + 197 + ], + "spans": [ + { + "bbox": [ + 294, + 185, + 346, + 197 + ], + "score": 1.0, + "content": "Jiawei Han", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 295, + 195, + 448, + 209 + ], + "spans": [ + { + "bbox": [ + 295, + 195, + 448, + 209 + ], + "score": 1.0, + "content": "University of Illinois, Urbana-Champaign", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 295, + 209, + 352, + 218 + ], + "spans": [ + { + "bbox": [ + 295, + 209, + 352, + 218 + ], + "score": 1.0, + "content": "hanj@illinois", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + } + ], + "index": 15, + "bbox_fs": [ + 294, + 185, + 448, + 218 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 248, + 333, + 259 + ], + "lines": [ + { + "bbox": [ + 276, + 246, + 336, + 262 + ], + "spans": [ + { + "bbox": [ + 276, + 246, + 336, + 262 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 143, + 274, + 469, + 395 + ], + "lines": [ + { + "bbox": [ + 141, + 273, + 469, + 288 + ], + "spans": [ + { + "bbox": [ + 141, + 273, + 469, + 288 + ], + "score": 1.0, + "content": "The learning rate warmup heuristic achieves remarkable success in stabilizing", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 285, + 469, + 299 + ], + "spans": [ + { + "bbox": [ + 141, + 285, + 469, + 299 + ], + "score": 1.0, + "content": "training, accelerating convergence and improving generalization for adaptive", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 296, + 469, + 309 + ], + "spans": [ + { + "bbox": [ + 141, + 296, + 469, + 309 + ], + "score": 1.0, + "content": "stochastic optimization algorithms like RMSprop and Adam. Pursuing the theory", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 307, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 141, + 307, + 469, + 320 + ], + "score": 1.0, + "content": "behind warmup, we identify a problem of the adaptive learning rate – its vari-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 318, + 470, + 331 + ], + "spans": [ + { + "bbox": [ + 141, + 318, + 470, + 331 + ], + "score": 1.0, + "content": "ance is problematically large in the early stage, and presume warmup works as a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 142, + 330, + 469, + 341 + ], + "spans": [ + { + "bbox": [ + 142, + 330, + 469, + 341 + ], + "score": 1.0, + "content": "variance reduction technique. We provide both empirical and theoretical evidence", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 340, + 470, + 353 + ], + "spans": [ + { + "bbox": [ + 141, + 340, + 470, + 353 + ], + "score": 1.0, + "content": "to verify our hypothesis. We further propose Rectified Adam (RAdam), a novel", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 142, + 352, + 469, + 363 + ], + "spans": [ + { + "bbox": [ + 142, + 352, + 469, + 363 + ], + "score": 1.0, + "content": "variant of Adam, by introducing a term to rectify the variance of the adaptive", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 361, + 469, + 375 + ], + "spans": [ + { + "bbox": [ + 141, + 361, + 469, + 375 + ], + "score": 1.0, + "content": "learning rate. Experimental results on image classification, language modeling,500550", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 372, + 469, + 386 + ], + "spans": [ + { + "bbox": [ + 141, + 372, + 469, + 386 + ], + "score": 1.0, + "content": "and neural machine translation verify our intuition and demonstrate the efficacy400450", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 142, + 384, + 253, + 395 + ], + "spans": [ + { + "bbox": [ + 142, + 384, + 253, + 395 + ], + "score": 1.0, + "content": "and robustness of RAdam.1", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23, + "bbox_fs": [ + 141, + 273, + 470, + 395 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 421, + 205, + 434 + ], + "lines": [ + { + "bbox": [ + 105, + 420, + 208, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 208, + 437 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 108, + 445, + 356, + 577 + ], + "lines": [ + { + "bbox": [ + 106, + 444, + 357, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 357, + 457 + ], + "score": 1.0, + "content": "Fast and stable optimization algorithms are what generations", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 455, + 357, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 357, + 469 + ], + "score": 1.0, + "content": "of researchers have been pursuing (Gauss, 1823; Cauchy,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 466, + 356, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 356, + 479 + ], + "score": 1.0, + "content": "1847). Remarkably, stochastic gradient-based optimization,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 478, + 357, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 357, + 490 + ], + "score": 1.0, + "content": "such as stochastic gradient descent (SGD), has witnessed", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 488, + 357, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 357, + 501 + ], + "score": 1.0, + "content": "tremendous success in many fields of science and engineering", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 500, + 357, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 357, + 512 + ], + "score": 1.0, + "content": "despite its simplicity. Recently, many efforts have been made", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 511, + 357, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 357, + 523 + ], + "score": 1.0, + "content": "to accelerate optimization by applying adaptive learning rate.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 520, + 357, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 357, + 535 + ], + "score": 1.0, + "content": "In particular, Adagrad (Duchi et al., 2010) and its variants, e.g.,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 532, + 357, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 357, + 545 + ], + "score": 1.0, + "content": "RMSprop (Hinton et al., 2012), Adam (Kingma & Ba, 2014),", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 543, + 357, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 357, + 555 + ], + "score": 1.0, + "content": "Adadelta (Zeiler, 2012) and Nadam (Dozat, 2016), stand out", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 554, + 357, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 357, + 567 + ], + "score": 1.0, + "content": "due to their fast convergence, and have been considered as the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 566, + 274, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 274, + 578 + ], + "score": 1.0, + "content": "optimizer of choice in many applications.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 444, + 357, + 578 + ] + }, + { + "type": "image", + "bbox": [ + 365, + 437, + 503, + 539 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 365, + 437, + 503, + 539 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 365, + 437, + 503, + 539 + ], + "spans": [ + { + "bbox": [ + 365, + 437, + 503, + 539 + ], + "score": 0.965, + "type": "image", + "image_path": "9e07baec4bf47bd9075fa89a9a4a4cab421e0a129a16eb042e7e1b7d64da16fe.jpg" + } + ] + } + ], + "index": 42.5, + "virtual_lines": [ + { + "bbox": [ + 365, + 437, + 503, + 488.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 365, + 488.0, + 503, + 539.0 + ], + "spans": [], + "index": 43 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 363, + 542, + 505, + 575 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 363, + 542, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 363, + 542, + 506, + 554 + ], + "score": 1.0, + "content": "Figure 1: Training loss v.s. # of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 363, + 553, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 363, + 553, + 505, + 565 + ], + "score": 1.0, + "content": "iterations of Transformers on the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 363, + 563, + 470, + 576 + ], + "spans": [ + { + "bbox": [ + 363, + 563, + 470, + 576 + ], + "score": 1.0, + "content": "De-En IWSLT’14 dataset.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "index": 43.75 + }, + { + "type": "text", + "bbox": [ + 107, + 582, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 582, + 504, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 504, + 595 + ], + "score": 1.0, + "content": "However, it has been observed that these optimization methods may converge to bad/suspicious", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "local optima, and have to resort to a warmup heuristic – using a small learning rate in the first", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "few epochs of training to mitigate such problem (Vaswani et al., 2017; Popel & Bojar, 2018). For", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 616, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 627 + ], + "score": 1.0, + "content": "example, when training typical Transformers based neural machine translation models on the De-En", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "IWSLT’14 dataset, removing the warmup stage increases the training loss from 3 to around 10, as", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 638, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 649 + ], + "score": 1.0, + "content": "shown in Figure 1. Similar phenomena are observed in other scenarios like BERT (a bidirectional", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 649, + 359, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 359, + 661 + ], + "score": 1.0, + "content": "transformer language model) pre-training (Devlin et al., 2019).", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 50, + "bbox_fs": [ + 105, + 582, + 506, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 665, + 504, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 664, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 679 + ], + "score": 1.0, + "content": "Duo to the lack of the theoretical underpinnings, there is neither guarantee that warmup would bring", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 676, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 688 + ], + "score": 1.0, + "content": "consistent improvements for various machine learning settings nor guidance on how we should", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "conduct warmup. Thus, researchers typically use different settings in different applications and", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 437, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 437, + 107 + ], + "score": 1.0, + "content": "have to take a trial-and-error approach, which can be tedious and time-consuming.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 54.5, + "bbox_fs": [ + 105, + 664, + 505, + 688 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "conduct warmup. Thus, researchers typically use different settings in different applications and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 437, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 437, + 107 + ], + "score": 1.0, + "content": "have to take a trial-and-error approach, which can be tedious and time-consuming.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 504, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 504, + 122 + ], + "score": 1.0, + "content": "In this paper, we conduct both empirical and theoretical analysis of the convergence issue to identify", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 505, + 134 + ], + "score": 1.0, + "content": "its origin. We show that its root cause is: the adaptive learning rate has undesirably large variance in", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 506, + 146 + ], + "score": 1.0, + "content": "the early stage of model training, due to the limited amount of training samples being used. Thus,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 141, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 506, + 158 + ], + "score": 1.0, + "content": "to reduce such variance, it is better to use smaller learning rates in the first few epochs of training,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 155, + 255, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 255, + 167 + ], + "score": 1.0, + "content": "which justifies the warmup heuristic.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 171, + 505, + 226 + ], + "lines": [ + { + "bbox": [ + 105, + 171, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 184 + ], + "score": 1.0, + "content": "Inspired by our analysis results, we propose a new variant of Adam, called Rectified Adam (RAdam),", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "which explicitly rectifies the variance of the adaptive learning rate based on derivations. We conduct", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 193, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 505, + 206 + ], + "score": 1.0, + "content": "extensive experiments on language modeling, image classification, and neural machine translation.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 203, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 218 + ], + "score": 1.0, + "content": "RAdam brings consistent improvement over the vanilla Adam, which verifies the variance issue", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 215, + 393, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 393, + 227 + ], + "score": 1.0, + "content": "generally exists on various tasks across different network architectures.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 232, + 305, + 243 + ], + "lines": [ + { + "bbox": [ + 105, + 231, + 306, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 306, + 244 + ], + "score": 1.0, + "content": "In summary, our main contributions are two-fold:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 245, + 505, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 505, + 257 + ], + "score": 1.0, + "content": "• We identify the variance issue of the adaptive learning rate and present a theoretical justification", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 254, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 115, + 254, + 505, + 269 + ], + "score": 1.0, + "content": "for the warmup heuristic. We show that the convergence issue is due to the undesirably large", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 114, + 266, + 411, + 280 + ], + "spans": [ + { + "bbox": [ + 114, + 266, + 411, + 280 + ], + "score": 1.0, + "content": "variance of the adaptive learning rate in the early stage of model training.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 280, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 294 + ], + "score": 1.0, + "content": "• We propose a new variant of Adam (i.e., RAdam), which not only explicitly rectifies the variance", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 292, + 446, + 305 + ], + "spans": [ + { + "bbox": [ + 115, + 292, + 446, + 305 + ], + "score": 1.0, + "content": "and is theoretically sound, but also compares favorably with the heuristic warmup.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 106, + 319, + 309, + 332 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 309, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 309, + 333 + ], + "score": 1.0, + "content": "2 PRELIMINARIES AND MOTIVATIONS", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 343, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "Generic adaptive methods. Algorithm 1 is a generic framework (all operations are element-wise).", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 353, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 368 + ], + "score": 1.0, + "content": "It describes various popular stochastic gradient descent algorithms (Reddi et al., 2018). Specifically,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 363, + 504, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 401, + 380 + ], + "score": 1.0, + "content": "different optimization algorithms can be specified by different choices of", + "type": "text" + }, + { + "bbox": [ + 401, + 365, + 418, + 378 + ], + "score": 0.91, + "content": "\\phi ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 363, + 436, + 380 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 437, + 366, + 455, + 378 + ], + "score": 0.9, + "content": "\\psi ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 363, + 486, + 380 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 486, + 366, + 504, + 378 + ], + "score": 0.89, + "content": "\\phi ( . )", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 376, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 274, + 390 + ], + "score": 1.0, + "content": "specifies how the momentum at time step", + "type": "text" + }, + { + "bbox": [ + 275, + 378, + 280, + 387 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 376, + 352, + 390 + ], + "score": 1.0, + "content": "is calculated, and", + "type": "text" + }, + { + "bbox": [ + 353, + 376, + 371, + 389 + ], + "score": 0.91, + "content": "\\psi ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 376, + 506, + 390 + ], + "score": 1.0, + "content": "how the adaptive learning rate at", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 386, + 356, + 401 + ], + "spans": [ + { + "bbox": [ + 107, + 389, + 111, + 397 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 111, + 386, + 356, + 401 + ], + "score": 1.0, + "content": "is calculated. For example, in the Adam algorithm, we have:", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21 + }, + { + "type": "interline_equation", + "bbox": [ + 113, + 397, + 485, + 431 + ], + "lines": [ + { + "bbox": [ + 113, + 397, + 485, + 431 + ], + "spans": [ + { + "bbox": [ + 113, + 397, + 485, + 431 + ], + "score": 0.92, + "content": "\\phi ( g _ { 1 } , \\cdots , g _ { t } ) = { \\frac { \\dot { ( 1 - \\beta _ { 1 } ) } \\sum _ { i = 1 } ^ { t } \\beta _ { 1 } ^ { t - i } g _ { t } } { 1 - \\beta _ { 1 } ^ { t } } } \\quad \\mathrm { a n d } \\quad \\psi ( g _ { 1 } , \\cdots , g _ { t } ) = { \\sqrt { \\frac { 1 - \\beta _ { 2 } ^ { t } } { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } } } .", + "type": "interline_equation", + "image_path": "ecf5e4b7f18515a650387401a902ea879805440618de5f9ded0e1c20c87ef512.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 113, + 397, + 485, + 408.3333333333333 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 113, + 408.3333333333333, + 485, + 419.66666666666663 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 113, + 419.66666666666663, + 485, + 430.99999999999994 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 432, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 430, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 257, + 446 + ], + "score": 1.0, + "content": "For numerical stability, the function", + "type": "text" + }, + { + "bbox": [ + 257, + 432, + 275, + 444 + ], + "score": 0.9, + "content": "\\psi ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 430, + 434, + 446 + ], + "score": 1.0, + "content": "in Equation 1 is usually calculated as", + "type": "text" + }, + { + "bbox": [ + 434, + 430, + 505, + 445 + ], + "score": 0.89, + "content": "\\widehat { \\psi } ( g _ { 1 } , \\cdots , g _ { t } ) =", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 443, + 456, + 468 + ], + "spans": [ + { + "bbox": [ + 107, + 443, + 200, + 468 + ], + "score": 0.92, + "content": "\\frac { \\sqrt { 1 - \\beta _ { 2 } ^ { t } } } { \\epsilon + \\sqrt { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 448, + 230, + 461 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 230, + 452, + 236, + 459 + ], + "score": 0.77, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 448, + 410, + 461 + ], + "score": 1.0, + "content": "is a relatively small / negligible value (e.g.,", + "type": "text" + }, + { + "bbox": [ + 410, + 448, + 450, + 460 + ], + "score": 0.8, + "content": "1 \\times 1 0 ^ { - 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 448, + 456, + 461 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 481, + 476, + 493 + ], + "lines": [ + { + "bbox": [ + 106, + 480, + 478, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 478, + 495 + ], + "score": 1.0, + "content": "Algorithm 1: Generic adaptive optimization method setup. All operations are element-wise.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 104, + 496, + 473, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 494, + 474, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 137, + 511 + ], + "score": 1.0, + "content": "Input:", + "type": "text" + }, + { + "bbox": [ + 137, + 495, + 171, + 509 + ], + "score": 0.9, + "content": "\\{ \\alpha _ { t } \\} _ { t = 1 } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 494, + 215, + 511 + ], + "score": 1.0, + "content": ": step size,", + "type": "text" + }, + { + "bbox": [ + 216, + 495, + 264, + 509 + ], + "score": 0.9, + "content": "\\{ \\phi _ { t } , \\psi _ { t } \\} _ { t = 1 } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 494, + 474, + 511 + ], + "score": 1.0, + "content": ": function to calculate momentum and adaptive rate,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 137, + 507, + 365, + 520 + ], + "spans": [ + { + "bbox": [ + 137, + 509, + 147, + 519 + ], + "score": 0.86, + "content": "\\theta _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 507, + 222, + 520 + ], + "score": 1.0, + "content": ": initial parameter,", + "type": "text" + }, + { + "bbox": [ + 222, + 509, + 241, + 520 + ], + "score": 0.9, + "content": "f ( \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 507, + 365, + 520 + ], + "score": 1.0, + "content": ": stochastic objective function.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 520, + 243, + 530 + ], + "lines": [ + { + "bbox": [ + 106, + 516, + 245, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 144, + 533 + ], + "score": 1.0, + "content": "Output:", + "type": "text" + }, + { + "bbox": [ + 144, + 519, + 156, + 529 + ], + "score": 0.86, + "content": "\\theta _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 516, + 245, + 533 + ], + "score": 1.0, + "content": ": resulting parameters", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 99, + 528, + 435, + 586 + ], + "lines": [ + { + "bbox": [ + 97, + 528, + 190, + 541 + ], + "spans": [ + { + "bbox": [ + 97, + 528, + 131, + 541 + ], + "score": 1.0, + "content": "1 while", + "type": "text" + }, + { + "bbox": [ + 132, + 530, + 155, + 540 + ], + "score": 0.89, + "content": "t = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 528, + 166, + 541 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 166, + 530, + 175, + 540 + ], + "score": 0.73, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 528, + 190, + 541 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 98, + 540, + 436, + 553 + ], + "spans": [ + { + "bbox": [ + 98, + 543, + 104, + 551 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 122, + 541, + 194, + 552 + ], + "score": 0.86, + "content": "g _ { t } \\gets \\Delta _ { \\theta } f _ { t } ( \\theta _ { t - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 540, + 436, + 553 + ], + "score": 1.0, + "content": "(Calculate gradients w.r.t. stochastic objective at timestep t)", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 98, + 551, + 306, + 564 + ], + "spans": [ + { + "bbox": [ + 98, + 554, + 104, + 562 + ], + "score": 1.0, + "content": "3", + "type": "text" + }, + { + "bbox": [ + 122, + 552, + 208, + 563 + ], + "score": 0.86, + "content": "m _ { t } \\gets \\phi _ { t } ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 551, + 306, + 564 + ], + "score": 1.0, + "content": "(Calculate momentum)", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 98, + 561, + 342, + 576 + ], + "spans": [ + { + "bbox": [ + 98, + 565, + 105, + 572 + ], + "score": 1.0, + "content": "4", + "type": "text" + }, + { + "bbox": [ + 122, + 563, + 204, + 574 + ], + "score": 0.83, + "content": "l _ { t } \\gets \\psi _ { t } ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 561, + 342, + 576 + ], + "score": 1.0, + "content": "(Calculate adaptive learning rate)", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 98, + 572, + 292, + 587 + ], + "spans": [ + { + "bbox": [ + 98, + 576, + 104, + 583 + ], + "score": 1.0, + "content": "5", + "type": "text" + }, + { + "bbox": [ + 122, + 574, + 205, + 586 + ], + "score": 0.82, + "content": "\\theta _ { t } \\gets \\theta _ { t - 1 } - \\alpha _ { t } m _ { t } l _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 572, + 292, + 587 + ], + "score": 1.0, + "content": "(Update parameters)", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 98, + 588, + 148, + 599 + ], + "lines": [ + { + "bbox": [ + 97, + 586, + 149, + 601 + ], + "spans": [ + { + "bbox": [ + 97, + 586, + 136, + 601 + ], + "score": 1.0, + "content": "6 return", + "type": "text" + }, + { + "bbox": [ + 136, + 588, + 149, + 600 + ], + "score": 0.68, + "content": "\\theta _ { T }", + "type": "inline_equation" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 616, + 505, + 683 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 359, + 629 + ], + "score": 1.0, + "content": "Learning rate warmup. Instead of setting the learning rate", + "type": "text" + }, + { + "bbox": [ + 359, + 618, + 370, + 627 + ], + "score": 0.86, + "content": "\\alpha _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "as a constant or in a decreasing", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 288, + 640 + ], + "score": 1.0, + "content": "order, a learning rate warmup strategy sets", + "type": "text" + }, + { + "bbox": [ + 288, + 629, + 300, + 638 + ], + "score": 0.85, + "content": "\\alpha _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "as smaller values in the first few steps, thus not", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 147, + 651 + ], + "score": 1.0, + "content": "satisfying", + "type": "text" + }, + { + "bbox": [ + 148, + 638, + 205, + 650 + ], + "score": 0.92, + "content": "\\forall t \\alpha _ { t + 1 } \\leq \\alpha _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 637, + 345, + 651 + ], + "score": 1.0, + "content": ". For example, linear warmup sets", + "type": "text" + }, + { + "bbox": [ + 346, + 639, + 388, + 649 + ], + "score": 0.92, + "content": "\\alpha _ { t } = t \\alpha _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 637, + 414, + 651 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 414, + 638, + 446, + 649 + ], + "score": 0.91, + "content": "t < T _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 637, + 506, + 651 + ], + "score": 1.0, + "content": ". Warmup has", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "been demonstrated to be beneficial in many deep learning applications. For example, in the NMT", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "experiments in Figure 1, the training loss convergences around 10 when warmup is not applied", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 671, + 497, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 497, + 684 + ], + "score": 1.0, + "content": "(Adam-vanilla), and it surprisingly decreases to below 3 after applying warmup (Adam-warmup).", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "To further analyze this phenomenon, we visualize the histogram of the absolute value of gradients", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "on a log scale in Figure 2. We observe that, without applying warmup, the gradient distribution", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "is distorted to have a mass center in relatively small values within 10 updates. Such gradient dis-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "tortion means that the vanilla Adam is trapped in bad/suspicious local optima after the first few", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 505, + 107 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 504, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 504, + 122 + ], + "score": 1.0, + "content": "In this paper, we conduct both empirical and theoretical analysis of the convergence issue to identify", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 505, + 134 + ], + "score": 1.0, + "content": "its origin. We show that its root cause is: the adaptive learning rate has undesirably large variance in", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 506, + 146 + ], + "score": 1.0, + "content": "the early stage of model training, due to the limited amount of training samples being used. Thus,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 141, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 506, + 158 + ], + "score": 1.0, + "content": "to reduce such variance, it is better to use smaller learning rates in the first few epochs of training,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 155, + 255, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 255, + 167 + ], + "score": 1.0, + "content": "which justifies the warmup heuristic.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 110, + 506, + 167 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 171, + 505, + 226 + ], + "lines": [ + { + "bbox": [ + 105, + 171, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 184 + ], + "score": 1.0, + "content": "Inspired by our analysis results, we propose a new variant of Adam, called Rectified Adam (RAdam),", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "which explicitly rectifies the variance of the adaptive learning rate based on derivations. We conduct", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 193, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 505, + 206 + ], + "score": 1.0, + "content": "extensive experiments on language modeling, image classification, and neural machine translation.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 203, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 218 + ], + "score": 1.0, + "content": "RAdam brings consistent improvement over the vanilla Adam, which verifies the variance issue", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 215, + 393, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 393, + 227 + ], + "score": 1.0, + "content": "generally exists on various tasks across different network architectures.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 171, + 506, + 227 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 232, + 305, + 243 + ], + "lines": [ + { + "bbox": [ + 105, + 231, + 306, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 306, + 244 + ], + "score": 1.0, + "content": "In summary, our main contributions are two-fold:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 231, + 306, + 244 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 245, + 505, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 505, + 257 + ], + "score": 1.0, + "content": "• We identify the variance issue of the adaptive learning rate and present a theoretical justification", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 254, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 115, + 254, + 505, + 269 + ], + "score": 1.0, + "content": "for the warmup heuristic. We show that the convergence issue is due to the undesirably large", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 114, + 266, + 411, + 280 + ], + "spans": [ + { + "bbox": [ + 114, + 266, + 411, + 280 + ], + "score": 1.0, + "content": "variance of the adaptive learning rate in the early stage of model training.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 280, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 294 + ], + "score": 1.0, + "content": "• We propose a new variant of Adam (i.e., RAdam), which not only explicitly rectifies the variance", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 292, + 446, + 305 + ], + "spans": [ + { + "bbox": [ + 115, + 292, + 446, + 305 + ], + "score": 1.0, + "content": "and is theoretically sound, but also compares favorably with the heuristic warmup.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 244, + 505, + 305 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 319, + 309, + 332 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 309, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 309, + 333 + ], + "score": 1.0, + "content": "2 PRELIMINARIES AND MOTIVATIONS", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 343, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "Generic adaptive methods. Algorithm 1 is a generic framework (all operations are element-wise).", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 353, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 368 + ], + "score": 1.0, + "content": "It describes various popular stochastic gradient descent algorithms (Reddi et al., 2018). Specifically,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 363, + 504, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 401, + 380 + ], + "score": 1.0, + "content": "different optimization algorithms can be specified by different choices of", + "type": "text" + }, + { + "bbox": [ + 401, + 365, + 418, + 378 + ], + "score": 0.91, + "content": "\\phi ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 363, + 436, + 380 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 437, + 366, + 455, + 378 + ], + "score": 0.9, + "content": "\\psi ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 363, + 486, + 380 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 486, + 366, + 504, + 378 + ], + "score": 0.89, + "content": "\\phi ( . )", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 376, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 274, + 390 + ], + "score": 1.0, + "content": "specifies how the momentum at time step", + "type": "text" + }, + { + "bbox": [ + 275, + 378, + 280, + 387 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 376, + 352, + 390 + ], + "score": 1.0, + "content": "is calculated, and", + "type": "text" + }, + { + "bbox": [ + 353, + 376, + 371, + 389 + ], + "score": 0.91, + "content": "\\psi ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 376, + 506, + 390 + ], + "score": 1.0, + "content": "how the adaptive learning rate at", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 386, + 356, + 401 + ], + "spans": [ + { + "bbox": [ + 107, + 389, + 111, + 397 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 111, + 386, + 356, + 401 + ], + "score": 1.0, + "content": "is calculated. For example, in the Adam algorithm, we have:", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 342, + 506, + 401 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 113, + 397, + 485, + 431 + ], + "lines": [ + { + "bbox": [ + 113, + 397, + 485, + 431 + ], + "spans": [ + { + "bbox": [ + 113, + 397, + 485, + 431 + ], + "score": 0.92, + "content": "\\phi ( g _ { 1 } , \\cdots , g _ { t } ) = { \\frac { \\dot { ( 1 - \\beta _ { 1 } ) } \\sum _ { i = 1 } ^ { t } \\beta _ { 1 } ^ { t - i } g _ { t } } { 1 - \\beta _ { 1 } ^ { t } } } \\quad \\mathrm { a n d } \\quad \\psi ( g _ { 1 } , \\cdots , g _ { t } ) = { \\sqrt { \\frac { 1 - \\beta _ { 2 } ^ { t } } { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } } } .", + "type": "interline_equation", + "image_path": "ecf5e4b7f18515a650387401a902ea879805440618de5f9ded0e1c20c87ef512.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 113, + 397, + 485, + 408.3333333333333 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 113, + 408.3333333333333, + 485, + 419.66666666666663 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 113, + 419.66666666666663, + 485, + 430.99999999999994 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 432, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 430, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 257, + 446 + ], + "score": 1.0, + "content": "For numerical stability, the function", + "type": "text" + }, + { + "bbox": [ + 257, + 432, + 275, + 444 + ], + "score": 0.9, + "content": "\\psi ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 430, + 434, + 446 + ], + "score": 1.0, + "content": "in Equation 1 is usually calculated as", + "type": "text" + }, + { + "bbox": [ + 434, + 430, + 505, + 445 + ], + "score": 0.89, + "content": "\\widehat { \\psi } ( g _ { 1 } , \\cdots , g _ { t } ) =", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 443, + 456, + 468 + ], + "spans": [ + { + "bbox": [ + 107, + 443, + 200, + 468 + ], + "score": 0.92, + "content": "\\frac { \\sqrt { 1 - \\beta _ { 2 } ^ { t } } } { \\epsilon + \\sqrt { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 448, + 230, + 461 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 230, + 452, + 236, + 459 + ], + "score": 0.77, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 448, + 410, + 461 + ], + "score": 1.0, + "content": "is a relatively small / negligible value (e.g.,", + "type": "text" + }, + { + "bbox": [ + 410, + 448, + 450, + 460 + ], + "score": 0.8, + "content": "1 \\times 1 0 ^ { - 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 448, + 456, + 461 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 430, + 505, + 468 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 481, + 476, + 493 + ], + "lines": [ + { + "bbox": [ + 106, + 480, + 478, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 478, + 495 + ], + "score": 1.0, + "content": "Algorithm 1: Generic adaptive optimization method setup. All operations are element-wise.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29, + "bbox_fs": [ + 106, + 480, + 478, + 495 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 496, + 473, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 494, + 474, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 137, + 511 + ], + "score": 1.0, + "content": "Input:", + "type": "text" + }, + { + "bbox": [ + 137, + 495, + 171, + 509 + ], + "score": 0.9, + "content": "\\{ \\alpha _ { t } \\} _ { t = 1 } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 494, + 215, + 511 + ], + "score": 1.0, + "content": ": step size,", + "type": "text" + }, + { + "bbox": [ + 216, + 495, + 264, + 509 + ], + "score": 0.9, + "content": "\\{ \\phi _ { t } , \\psi _ { t } \\} _ { t = 1 } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 494, + 474, + 511 + ], + "score": 1.0, + "content": ": function to calculate momentum and adaptive rate,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 137, + 507, + 365, + 520 + ], + "spans": [ + { + "bbox": [ + 137, + 509, + 147, + 519 + ], + "score": 0.86, + "content": "\\theta _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 507, + 222, + 520 + ], + "score": 1.0, + "content": ": initial parameter,", + "type": "text" + }, + { + "bbox": [ + 222, + 509, + 241, + 520 + ], + "score": 0.9, + "content": "f ( \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 507, + 365, + 520 + ], + "score": 1.0, + "content": ": stochastic objective function.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 494, + 474, + 520 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 520, + 243, + 530 + ], + "lines": [ + { + "bbox": [ + 106, + 516, + 245, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 144, + 533 + ], + "score": 1.0, + "content": "Output:", + "type": "text" + }, + { + "bbox": [ + 144, + 519, + 156, + 529 + ], + "score": 0.86, + "content": "\\theta _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 516, + 245, + 533 + ], + "score": 1.0, + "content": ": resulting parameters", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 516, + 245, + 533 + ] + }, + { + "type": "index", + "bbox": [ + 99, + 528, + 435, + 586 + ], + "lines": [ + { + "bbox": [ + 97, + 528, + 190, + 541 + ], + "spans": [ + { + "bbox": [ + 97, + 528, + 131, + 541 + ], + "score": 1.0, + "content": "1 while", + "type": "text" + }, + { + "bbox": [ + 132, + 530, + 155, + 540 + ], + "score": 0.89, + "content": "t = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 528, + 166, + 541 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 166, + 530, + 175, + 540 + ], + "score": 0.73, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 528, + 190, + 541 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 98, + 540, + 436, + 553 + ], + "spans": [ + { + "bbox": [ + 98, + 543, + 104, + 551 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 122, + 541, + 194, + 552 + ], + "score": 0.86, + "content": "g _ { t } \\gets \\Delta _ { \\theta } f _ { t } ( \\theta _ { t - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 540, + 436, + 553 + ], + "score": 1.0, + "content": "(Calculate gradients w.r.t. stochastic objective at timestep t)", + "type": "text" + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 98, + 551, + 306, + 564 + ], + "spans": [ + { + "bbox": [ + 98, + 554, + 104, + 562 + ], + "score": 1.0, + "content": "3", + "type": "text" + }, + { + "bbox": [ + 122, + 552, + 208, + 563 + ], + "score": 0.86, + "content": "m _ { t } \\gets \\phi _ { t } ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 551, + 306, + 564 + ], + "score": 1.0, + "content": "(Calculate momentum)", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 98, + 561, + 342, + 576 + ], + "spans": [ + { + "bbox": [ + 98, + 565, + 105, + 572 + ], + "score": 1.0, + "content": "4", + "type": "text" + }, + { + "bbox": [ + 122, + 563, + 204, + 574 + ], + "score": 0.83, + "content": "l _ { t } \\gets \\psi _ { t } ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 561, + 342, + 576 + ], + "score": 1.0, + "content": "(Calculate adaptive learning rate)", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 98, + 572, + 292, + 587 + ], + "spans": [ + { + "bbox": [ + 98, + 576, + 104, + 583 + ], + "score": 1.0, + "content": "5", + "type": "text" + }, + { + "bbox": [ + 122, + 574, + 205, + 586 + ], + "score": 0.82, + "content": "\\theta _ { t } \\gets \\theta _ { t - 1 } - \\alpha _ { t } m _ { t } l _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 572, + 292, + 587 + ], + "score": 1.0, + "content": "(Update parameters)", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + } + ], + "index": 35, + "bbox_fs": [ + 97, + 528, + 436, + 587 + ] + }, + { + "type": "text", + "bbox": [ + 98, + 588, + 148, + 599 + ], + "lines": [ + { + "bbox": [ + 97, + 586, + 149, + 601 + ], + "spans": [ + { + "bbox": [ + 97, + 586, + 136, + 601 + ], + "score": 1.0, + "content": "6 return", + "type": "text" + }, + { + "bbox": [ + 136, + 588, + 149, + 600 + ], + "score": 0.68, + "content": "\\theta _ { T }", + "type": "inline_equation" + } + ], + "index": 38 + } + ], + "index": 38, + "bbox_fs": [ + 97, + 586, + 149, + 601 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 616, + 505, + 683 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 359, + 629 + ], + "score": 1.0, + "content": "Learning rate warmup. Instead of setting the learning rate", + "type": "text" + }, + { + "bbox": [ + 359, + 618, + 370, + 627 + ], + "score": 0.86, + "content": "\\alpha _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "as a constant or in a decreasing", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 288, + 640 + ], + "score": 1.0, + "content": "order, a learning rate warmup strategy sets", + "type": "text" + }, + { + "bbox": [ + 288, + 629, + 300, + 638 + ], + "score": 0.85, + "content": "\\alpha _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "as smaller values in the first few steps, thus not", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 147, + 651 + ], + "score": 1.0, + "content": "satisfying", + "type": "text" + }, + { + "bbox": [ + 148, + 638, + 205, + 650 + ], + "score": 0.92, + "content": "\\forall t \\alpha _ { t + 1 } \\leq \\alpha _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 637, + 345, + 651 + ], + "score": 1.0, + "content": ". For example, linear warmup sets", + "type": "text" + }, + { + "bbox": [ + 346, + 639, + 388, + 649 + ], + "score": 0.92, + "content": "\\alpha _ { t } = t \\alpha _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 637, + 414, + 651 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 414, + 638, + 446, + 649 + ], + "score": 0.91, + "content": "t < T _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 637, + 506, + 651 + ], + "score": 1.0, + "content": ". Warmup has", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "been demonstrated to be beneficial in many deep learning applications. For example, in the NMT", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "experiments in Figure 1, the training loss convergences around 10 when warmup is not applied", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 671, + 497, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 497, + 684 + ], + "score": 1.0, + "content": "(Adam-vanilla), and it surprisingly decreases to below 3 after applying warmup (Adam-warmup).", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 615, + 506, + 684 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "To further analyze this phenomenon, we visualize the histogram of the absolute value of gradients", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "on a log scale in Figure 2. We observe that, without applying warmup, the gradient distribution", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "is distorted to have a mass center in relatively small values within 10 updates. Such gradient dis-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "tortion means that the vanilla Adam is trapped in bad/suspicious local optima after the first few", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "updates. Warmup essentially reduces the impact of these problematic updates to avoid the conver-", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "gence problem. In the following sections, we focus our analysis on learning rate warmup for the", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "Adam algorithm, while it can be applied to other algorithms that use similar adaptive learning rate", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 107, + 355, + 421, + 368 + ], + "spans": [ + { + "bbox": [ + 107, + 355, + 130, + 367 + ], + "score": 0.82, + "content": "( \\psi ( . ) )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 131, + 356, + 421, + 368 + ], + "score": 1.0, + "content": "designs, e.g., RMSprop (Hinton et al., 2012) and Nadam (Dozat, 2016).", + "type": "text", + "cross_page": true + } + ], + "index": 14 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 687, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 61, + 498, + 152 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 61, + 498, + 152 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 61, + 498, + 152 + ], + "spans": [ + { + "bbox": [ + 109, + 61, + 498, + 152 + ], + "score": 0.969, + "type": "image", + "image_path": "6e677d6bab859ece1c2f01a47cda68a7b6ac5a545e1cae5ce58363c6119aa535.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 61, + 498, + 91.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 91.33333333333333, + 498, + 121.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 121.66666666666666, + 498, + 152.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 158, + 505, + 192 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 158, + 506, + 170 + ], + "spans": [ + { + "bbox": [ + 106, + 158, + 506, + 170 + ], + "score": 1.0, + "content": "Figure 2: The absolute gradient histogram of the Transformers on the De-En IWSLT’ 14 dataset", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 169, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 169, + 505, + 182 + ], + "score": 1.0, + "content": "during the training (stacked along the y-axis). X-axis is absolute value in the log scale and the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 180, + 499, + 192 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 499, + 192 + ], + "score": 1.0, + "content": "height is the frequency. Without warmup, the gradient distribution is distorted in the first 10 steps.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 145, + 198, + 461, + 285 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 145, + 198, + 461, + 285 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 145, + 198, + 461, + 285 + ], + "spans": [ + { + "bbox": [ + 145, + 198, + 461, + 285 + ], + "score": 0.968, + "type": "image", + "image_path": "dc6f22e695c9b55a2b7cd42c7d714d441222920db3df396076473763529e1315.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 145, + 198, + 461, + 227.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 145, + 227.0, + 461, + 256.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 145, + 256.0, + 461, + 285.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 289, + 505, + 312 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 289, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 506, + 302 + ], + "score": 1.0, + "content": "Figure 3: The histogram of the absolute value of gradients (on a log scale) during the training of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 298, + 462, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 462, + 313 + ], + "score": 1.0, + "content": "Transformers on the De-En IWSLT’ 14 dataset. using Adam-2k, RAdam and Adam-eps.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + } + ], + "index": 8.25 + }, + { + "type": "text", + "bbox": [ + 107, + 322, + 505, + 367 + ], + "lines": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "updates. Warmup essentially reduces the impact of these problematic updates to avoid the conver-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "gence problem. In the following sections, we focus our analysis on learning rate warmup for the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "Adam algorithm, while it can be applied to other algorithms that use similar adaptive learning rate", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 107, + 355, + 421, + 368 + ], + "spans": [ + { + "bbox": [ + 107, + 355, + 130, + 367 + ], + "score": 0.82, + "content": "( \\psi ( . ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 356, + 421, + 368 + ], + "score": 1.0, + "content": "designs, e.g., RMSprop (Hinton et al., 2012) and Nadam (Dozat, 2016).", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 107, + 380, + 366, + 393 + ], + "lines": [ + { + "bbox": [ + 104, + 379, + 368, + 395 + ], + "spans": [ + { + "bbox": [ + 104, + 379, + 368, + 395 + ], + "score": 1.0, + "content": "3 VARIANCE OF THE ADAPTIVE LEARNING RATE", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 403, + 505, + 436 + ], + "lines": [ + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "score": 1.0, + "content": "In this section, we first introduce empirical evidence, then analyze the variance of the adaptive", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "learning rate to support our hypothesis – Due to the lack of samples in the early stage, the adaptive", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 425, + 479, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 479, + 437 + ], + "score": 1.0, + "content": "learning rate has an undesirably large variance, which leads to suspicious/bad local optima.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 504, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 361, + 456 + ], + "score": 1.0, + "content": "To convey our intuition, we begin with a special case. When", + "type": "text" + }, + { + "bbox": [ + 361, + 444, + 388, + 454 + ], + "score": 0.89, + "content": "t = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 442, + 430, + 456 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + }, + { + "bbox": [ + 430, + 442, + 501, + 456 + ], + "score": 0.93, + "content": "\\psi ( g _ { 1 } ) = \\sqrt { 1 / g _ { 1 } ^ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 442, + 504, + 456 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 453, + 504, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 145, + 470 + ], + "score": 1.0, + "content": "We view", + "type": "text" + }, + { + "bbox": [ + 145, + 455, + 198, + 468 + ], + "score": 0.94, + "content": "\\{ g _ { 1 } , \\cdots , g _ { t } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 453, + 388, + 470 + ], + "score": 1.0, + "content": "as i.i.d. Gaussian random variables following", + "type": "text" + }, + { + "bbox": [ + 389, + 455, + 431, + 468 + ], + "score": 0.94, + "content": "{ \\mathcal { N } } ( 0 , \\sigma ^ { 2 } ) ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 453, + 483, + 470 + ], + "score": 1.0, + "content": ". Therefore,", + "type": "text" + }, + { + "bbox": [ + 483, + 456, + 504, + 468 + ], + "score": 0.88, + "content": "1 / g _ { 1 } ^ { 2 }", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 467, + 504, + 483 + ], + "spans": [ + { + "bbox": [ + 104, + 467, + 378, + 483 + ], + "score": 1.0, + "content": "is subject to the scaled inverse chi-squared distribution, Scale-inv-", + "type": "text" + }, + { + "bbox": [ + 378, + 468, + 429, + 481 + ], + "score": 0.92, + "content": "\\mathcal { X } ^ { 2 } ( 1 , 1 / \\sigma ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 467, + 452, + 483 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 452, + 468, + 504, + 481 + ], + "score": 0.91, + "content": "\\mathrm { V a r } [ \\sqrt { 1 / g _ { 1 } ^ { 2 } } ]", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 478, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 104, + 478, + 505, + 493 + ], + "score": 1.0, + "content": "is divergent. It means that the adaptive ratio can be undesirably large in the first stage of learning.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 489, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 454, + 503 + ], + "score": 1.0, + "content": "Meanwhile, setting a small learning rate at the early stage can reduce the variance", + "type": "text" + }, + { + "bbox": [ + 454, + 491, + 505, + 503 + ], + "score": 0.88, + "content": "\\quad ( { \\mathrm { V a r } } [ \\alpha x ] \\ =", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 500, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 107, + 501, + 149, + 513 + ], + "score": 0.92, + "content": "\\alpha ^ { 2 } \\mathrm { V a r } [ x ] )", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 500, + 506, + 515 + ], + "score": 1.0, + "content": ", thus alleviating this problem. Therefore, we suggest it is the unbounded variance of the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 513, + 412, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 412, + 525 + ], + "score": 1.0, + "content": "adaptive learning rate in the early stage that causes the problematic updates.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 108, + 535, + 286, + 546 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 289, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 289, + 548 + ], + "score": 1.0, + "content": "3.1 WARMUP AS VARIANCE REDUCTION", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 553, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 552, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 566 + ], + "score": 1.0, + "content": "In this section, we design a set of controlled experiments to verify our hypothesis. Particularly, we", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 564, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 475, + 576 + ], + "score": 1.0, + "content": "design two variants of Adam that reducing the variance of the adaptive learning rate: Adam-", + "type": "text" + }, + { + "bbox": [ + 476, + 565, + 487, + 574 + ], + "score": 0.82, + "content": "{ \\it 2 k }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 564, + 505, + 576 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 576, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 587 + ], + "score": 1.0, + "content": "Adam-eps. We compare them to vanilla Adam with and without warmup on the IWSLT’14 German", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 586, + 312, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 312, + 598 + ], + "score": 1.0, + "content": "to English translation dataset (Cettolo et al., 2014).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 603, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 602, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 340, + 615 + ], + "score": 1.0, + "content": "In order to reduce the variance of the adaptive learning rate", + "type": "text" + }, + { + "bbox": [ + 340, + 603, + 365, + 615 + ], + "score": 0.88, + "content": "( \\psi ( . ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 602, + 396, + 615 + ], + "score": 1.0, + "content": ", Adam-", + "type": "text" + }, + { + "bbox": [ + 397, + 604, + 408, + 613 + ], + "score": 0.34, + "content": "2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 602, + 461, + 615 + ], + "score": 1.0, + "content": "only updates", + "type": "text" + }, + { + "bbox": [ + 461, + 603, + 479, + 615 + ], + "score": 0.91, + "content": "\\psi ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 602, + 505, + 615 + ], + "score": 1.0, + "content": "in the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 313, + 626 + ], + "score": 1.0, + "content": "first two thousand iterations, while the momentum", + "type": "text" + }, + { + "bbox": [ + 313, + 614, + 337, + 626 + ], + "score": 0.86, + "content": "( \\phi ( . ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 613, + 403, + 626 + ], + "score": 1.0, + "content": "and parameters", + "type": "text" + }, + { + "bbox": [ + 403, + 614, + 415, + 625 + ], + "score": 0.44, + "content": "\\mathbf { \\eta } ^ { ( \\theta ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "are fixed3; other than", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 625, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 505, + 637 + ], + "score": 1.0, + "content": "this, it follows the original Adam algorithm. To make comparison with other methods, its iterations", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "score": 1.0, + "content": "are indexed from -1999 instead of 1. In Figure 1, we observe that, after getting these additional", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 646, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 104, + 646, + 395, + 660 + ], + "score": 1.0, + "content": "two thousand samples for estimating the adaptive learning rate, Adam-", + "type": "text" + }, + { + "bbox": [ + 396, + 647, + 407, + 657 + ], + "score": 0.37, + "content": "2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 646, + 505, + 660 + ], + "score": 1.0, + "content": "avoids the convergence", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 657, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 671 + ], + "score": 1.0, + "content": "problem of the vanilla-Adam. Also, comparing Figure 2 and Figure 3, getting large enough samples", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "prevents the gradient distribution from being distorted. These observations verify our hypothesis", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 679, + 501, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 501, + 692 + ], + "score": 1.0, + "content": "that the lack of sufficient data samples in the early stage is the root cause of the convergence issue.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 700, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 117, + 698, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 117, + 698, + 506, + 714 + ], + "score": 1.0, + "content": "2The mean zero normal assumption is valid at the beginning of the training, since weights are sampled from", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 711, + 482, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 482, + 721 + ], + "score": 1.0, + "content": "normal distributions with mean zero (Balduzzi et al., 2017), further analysis is conducted in Section 5.3.", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 719, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 506, + 734 + ], + "score": 1.0, + "content": "3Different from Gotmare et al. (2019), all parameters and first moments are frozen in the first 2000 iterations.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 61, + 498, + 152 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 61, + 498, + 152 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 61, + 498, + 152 + ], + "spans": [ + { + "bbox": [ + 109, + 61, + 498, + 152 + ], + "score": 0.969, + "type": "image", + "image_path": "6e677d6bab859ece1c2f01a47cda68a7b6ac5a545e1cae5ce58363c6119aa535.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 61, + 498, + 91.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 91.33333333333333, + 498, + 121.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 121.66666666666666, + 498, + 152.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 158, + 505, + 192 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 158, + 506, + 170 + ], + "spans": [ + { + "bbox": [ + 106, + 158, + 506, + 170 + ], + "score": 1.0, + "content": "Figure 2: The absolute gradient histogram of the Transformers on the De-En IWSLT’ 14 dataset", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 169, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 169, + 505, + 182 + ], + "score": 1.0, + "content": "during the training (stacked along the y-axis). X-axis is absolute value in the log scale and the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 180, + 499, + 192 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 499, + 192 + ], + "score": 1.0, + "content": "height is the frequency. Without warmup, the gradient distribution is distorted in the first 10 steps.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 145, + 198, + 461, + 285 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 145, + 198, + 461, + 285 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 145, + 198, + 461, + 285 + ], + "spans": [ + { + "bbox": [ + 145, + 198, + 461, + 285 + ], + "score": 0.968, + "type": "image", + "image_path": "dc6f22e695c9b55a2b7cd42c7d714d441222920db3df396076473763529e1315.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 145, + 198, + 461, + 227.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 145, + 227.0, + 461, + 256.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 145, + 256.0, + 461, + 285.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 289, + 505, + 312 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 289, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 506, + 302 + ], + "score": 1.0, + "content": "Figure 3: The histogram of the absolute value of gradients (on a log scale) during the training of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 298, + 462, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 462, + 313 + ], + "score": 1.0, + "content": "Transformers on the De-En IWSLT’ 14 dataset. using Adam-2k, RAdam and Adam-eps.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + } + ], + "index": 8.25 + }, + { + "type": "text", + "bbox": [ + 107, + 322, + 505, + 367 + ], + "lines": [], + "index": 12.5, + "bbox_fs": [ + 105, + 322, + 505, + 368 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 380, + 366, + 393 + ], + "lines": [ + { + "bbox": [ + 104, + 379, + 368, + 395 + ], + "spans": [ + { + "bbox": [ + 104, + 379, + 368, + 395 + ], + "score": 1.0, + "content": "3 VARIANCE OF THE ADAPTIVE LEARNING RATE", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 403, + 505, + 436 + ], + "lines": [ + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "score": 1.0, + "content": "In this section, we first introduce empirical evidence, then analyze the variance of the adaptive", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "learning rate to support our hypothesis – Due to the lack of samples in the early stage, the adaptive", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 425, + 479, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 479, + 437 + ], + "score": 1.0, + "content": "learning rate has an undesirably large variance, which leads to suspicious/bad local optima.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 402, + 505, + 437 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 504, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 361, + 456 + ], + "score": 1.0, + "content": "To convey our intuition, we begin with a special case. When", + "type": "text" + }, + { + "bbox": [ + 361, + 444, + 388, + 454 + ], + "score": 0.89, + "content": "t = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 442, + 430, + 456 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + }, + { + "bbox": [ + 430, + 442, + 501, + 456 + ], + "score": 0.93, + "content": "\\psi ( g _ { 1 } ) = \\sqrt { 1 / g _ { 1 } ^ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 442, + 504, + 456 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 453, + 504, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 145, + 470 + ], + "score": 1.0, + "content": "We view", + "type": "text" + }, + { + "bbox": [ + 145, + 455, + 198, + 468 + ], + "score": 0.94, + "content": "\\{ g _ { 1 } , \\cdots , g _ { t } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 453, + 388, + 470 + ], + "score": 1.0, + "content": "as i.i.d. Gaussian random variables following", + "type": "text" + }, + { + "bbox": [ + 389, + 455, + 431, + 468 + ], + "score": 0.94, + "content": "{ \\mathcal { N } } ( 0 , \\sigma ^ { 2 } ) ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 453, + 483, + 470 + ], + "score": 1.0, + "content": ". Therefore,", + "type": "text" + }, + { + "bbox": [ + 483, + 456, + 504, + 468 + ], + "score": 0.88, + "content": "1 / g _ { 1 } ^ { 2 }", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 467, + 504, + 483 + ], + "spans": [ + { + "bbox": [ + 104, + 467, + 378, + 483 + ], + "score": 1.0, + "content": "is subject to the scaled inverse chi-squared distribution, Scale-inv-", + "type": "text" + }, + { + "bbox": [ + 378, + 468, + 429, + 481 + ], + "score": 0.92, + "content": "\\mathcal { X } ^ { 2 } ( 1 , 1 / \\sigma ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 467, + 452, + 483 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 452, + 468, + 504, + 481 + ], + "score": 0.91, + "content": "\\mathrm { V a r } [ \\sqrt { 1 / g _ { 1 } ^ { 2 } } ]", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 478, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 104, + 478, + 505, + 493 + ], + "score": 1.0, + "content": "is divergent. It means that the adaptive ratio can be undesirably large in the first stage of learning.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 489, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 454, + 503 + ], + "score": 1.0, + "content": "Meanwhile, setting a small learning rate at the early stage can reduce the variance", + "type": "text" + }, + { + "bbox": [ + 454, + 491, + 505, + 503 + ], + "score": 0.88, + "content": "\\quad ( { \\mathrm { V a r } } [ \\alpha x ] \\ =", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 500, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 107, + 501, + 149, + 513 + ], + "score": 0.92, + "content": "\\alpha ^ { 2 } \\mathrm { V a r } [ x ] )", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 500, + 506, + 515 + ], + "score": 1.0, + "content": ", thus alleviating this problem. Therefore, we suggest it is the unbounded variance of the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 513, + 412, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 412, + 525 + ], + "score": 1.0, + "content": "adaptive learning rate in the early stage that causes the problematic updates.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22, + "bbox_fs": [ + 104, + 442, + 506, + 525 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 535, + 286, + 546 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 289, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 289, + 548 + ], + "score": 1.0, + "content": "3.1 WARMUP AS VARIANCE REDUCTION", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 553, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 552, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 566 + ], + "score": 1.0, + "content": "In this section, we design a set of controlled experiments to verify our hypothesis. Particularly, we", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 564, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 475, + 576 + ], + "score": 1.0, + "content": "design two variants of Adam that reducing the variance of the adaptive learning rate: Adam-", + "type": "text" + }, + { + "bbox": [ + 476, + 565, + 487, + 574 + ], + "score": 0.82, + "content": "{ \\it 2 k }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 564, + 505, + 576 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 576, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 587 + ], + "score": 1.0, + "content": "Adam-eps. We compare them to vanilla Adam with and without warmup on the IWSLT’14 German", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 586, + 312, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 312, + 598 + ], + "score": 1.0, + "content": "to English translation dataset (Cettolo et al., 2014).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 552, + 505, + 598 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 603, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 602, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 340, + 615 + ], + "score": 1.0, + "content": "In order to reduce the variance of the adaptive learning rate", + "type": "text" + }, + { + "bbox": [ + 340, + 603, + 365, + 615 + ], + "score": 0.88, + "content": "( \\psi ( . ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 602, + 396, + 615 + ], + "score": 1.0, + "content": ", Adam-", + "type": "text" + }, + { + "bbox": [ + 397, + 604, + 408, + 613 + ], + "score": 0.34, + "content": "2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 602, + 461, + 615 + ], + "score": 1.0, + "content": "only updates", + "type": "text" + }, + { + "bbox": [ + 461, + 603, + 479, + 615 + ], + "score": 0.91, + "content": "\\psi ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 602, + 505, + 615 + ], + "score": 1.0, + "content": "in the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 313, + 626 + ], + "score": 1.0, + "content": "first two thousand iterations, while the momentum", + "type": "text" + }, + { + "bbox": [ + 313, + 614, + 337, + 626 + ], + "score": 0.86, + "content": "( \\phi ( . ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 613, + 403, + 626 + ], + "score": 1.0, + "content": "and parameters", + "type": "text" + }, + { + "bbox": [ + 403, + 614, + 415, + 625 + ], + "score": 0.44, + "content": "\\mathbf { \\eta } ^ { ( \\theta ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "are fixed3; other than", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 625, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 505, + 637 + ], + "score": 1.0, + "content": "this, it follows the original Adam algorithm. To make comparison with other methods, its iterations", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "score": 1.0, + "content": "are indexed from -1999 instead of 1. In Figure 1, we observe that, after getting these additional", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 646, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 104, + 646, + 395, + 660 + ], + "score": 1.0, + "content": "two thousand samples for estimating the adaptive learning rate, Adam-", + "type": "text" + }, + { + "bbox": [ + 396, + 647, + 407, + 657 + ], + "score": 0.37, + "content": "2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 646, + 505, + 660 + ], + "score": 1.0, + "content": "avoids the convergence", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 657, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 671 + ], + "score": 1.0, + "content": "problem of the vanilla-Adam. Also, comparing Figure 2 and Figure 3, getting large enough samples", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "prevents the gradient distribution from being distorted. These observations verify our hypothesis", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 679, + 501, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 501, + 692 + ], + "score": 1.0, + "content": "that the lack of sufficient data samples in the early stage is the root cause of the convergence issue.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5, + "bbox_fs": [ + 104, + 602, + 506, + 692 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 240 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 419, + 96 + ], + "score": 1.0, + "content": "Another straightforward way to reduce the variance is to increase the value of", + "type": "text" + }, + { + "bbox": [ + 419, + 85, + 425, + 92 + ], + "score": 0.7, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 81, + 436, + 96 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 436, + 81, + 506, + 95 + ], + "score": 0.91, + "content": "\\widehat { \\psi } ( g _ { 1 } , \\cdots , g _ { t } ) =", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 93, + 504, + 118 + ], + "spans": [ + { + "bbox": [ + 107, + 93, + 200, + 118 + ], + "score": 0.91, + "content": "\\frac { \\sqrt { 1 - \\beta _ { 2 } ^ { t } } } { \\epsilon + \\sqrt { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 99, + 299, + 112 + ], + "score": 1.0, + "content": "Actually, if we assume", + "type": "text" + }, + { + "bbox": [ + 299, + 97, + 318, + 111 + ], + "score": 0.91, + "content": "\\widehat { \\psi } ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 99, + 504, + 112 + ], + "score": 1.0, + "content": "is subject to the uniform distribution, its vari-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 103, + 115, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 103, + 115, + 167, + 133 + ], + "score": 1.0, + "content": "ance equals to", + "type": "text" + }, + { + "bbox": [ + 167, + 117, + 185, + 130 + ], + "score": 0.91, + "content": "\\frac { 1 } { 1 2 \\epsilon ^ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 115, + 461, + 133 + ], + "score": 1.0, + "content": ". Therefore, we design Adam-eps, which uses a non-negligibly large", + "type": "text" + }, + { + "bbox": [ + 461, + 117, + 501, + 128 + ], + "score": 0.9, + "content": "\\epsilon = 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 115, + 506, + 133 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 128, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 131, + 144 + ], + "score": 1.0, + "content": "while", + "type": "text" + }, + { + "bbox": [ + 131, + 129, + 172, + 140 + ], + "score": 0.9, + "content": "\\epsilon = 1 0 ^ { - 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 128, + 506, + 144 + ], + "score": 1.0, + "content": "for vanilla Adam. Its performance is summarized in Figure 1. We observe that it", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 141, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 505, + 153 + ], + "score": 1.0, + "content": "does not suffer from the serious convergence problem of vanilla-Adam. This further demonstrates", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 151, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 505, + 165 + ], + "score": 1.0, + "content": "that the convergence problem can be alleviated by reducing the variance of the adaptive learning", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 162, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 248, + 176 + ], + "score": 1.0, + "content": "rate, and also explains why tuning", + "type": "text" + }, + { + "bbox": [ + 248, + 165, + 254, + 173 + ], + "score": 0.64, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 162, + 506, + 176 + ], + "score": 1.0, + "content": "is important in practice (Liu et al., 2019). Besides, similar to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 172, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 104, + 172, + 506, + 187 + ], + "score": 1.0, + "content": "Adam-2k, it prevents the gradient distribution from being distorted (as shown in Figure 3). However,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 185, + 504, + 197 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 410, + 197 + ], + "score": 1.0, + "content": "as in Figure 1, it produces a much worse performance comparing to Adam-", + "type": "text" + }, + { + "bbox": [ + 410, + 185, + 422, + 195 + ], + "score": 0.31, + "content": "2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 185, + 504, + 197 + ], + "score": 1.0, + "content": "and Adam-warmup.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 195, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 271, + 208 + ], + "score": 1.0, + "content": "We conjecture that this is because large", + "type": "text" + }, + { + "bbox": [ + 271, + 198, + 277, + 205 + ], + "score": 0.57, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 195, + 505, + 208 + ], + "score": 1.0, + "content": "induces a large bias into the adaptive learning rate and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 207, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 505, + 219 + ], + "score": 1.0, + "content": "slows down the optimization process. Thus, we need a more principled and rigorous way to con-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 217, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 505, + 230 + ], + "score": 1.0, + "content": "trol the variance of the adaptive learning rate. In the next subsection, we will present a theoretical", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 228, + 318, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 318, + 241 + ], + "score": 1.0, + "content": "analysis of the variance of the adaptive learning rate.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 6 + }, + { + "type": "title", + "bbox": [ + 106, + 252, + 359, + 263 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 360, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 360, + 264 + ], + "score": 1.0, + "content": "3.2 ANALYSIS OF ADAPTIVE LEARNING RATE VARIANCE", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 270, + 505, + 413 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 284 + ], + "score": 1.0, + "content": "As mentioned before, Adam uses the exponential moving average to calculate the adaptive learning", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 185, + 294 + ], + "score": 1.0, + "content": "rate. For gradients", + "type": "text" + }, + { + "bbox": [ + 185, + 281, + 236, + 293 + ], + "score": 0.96, + "content": "\\{ g _ { 1 } , \\cdots , g _ { t } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 281, + 505, + 294 + ], + "score": 1.0, + "content": ", their exponential moving average has a larger variance than their", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 293, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 276, + 304 + ], + "score": 1.0, + "content": "simple average. Also, in the early stage", + "type": "text" + }, + { + "bbox": [ + 276, + 294, + 281, + 302 + ], + "score": 0.66, + "content": "\\mathbf { \\chi } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 293, + 505, + 304 + ], + "score": 1.0, + "content": "is small), the difference of the exponential weights of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 300, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 107, + 303, + 159, + 315 + ], + "score": 0.92, + "content": "\\{ g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 300, + 262, + 317 + ], + "score": 1.0, + "content": "is relatively small (up to", + "type": "text" + }, + { + "bbox": [ + 262, + 302, + 302, + 316 + ], + "score": 0.9, + "content": "\\breve { 1 } - \\beta _ { 2 } ^ { t - 1 } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 300, + 506, + 317 + ], + "score": 1.0, + "content": "). Therefore, for ease of analysis, we approximate", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "score": 1.0, + "content": "the distribution of the exponential moving average as the distribution of the simple average (Nau,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 325, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 221, + 348 + ], + "score": 1.0, + "content": "2014), i.e., p(ψ(.)) = p(", + "type": "text" + }, + { + "bbox": [ + 154, + 325, + 369, + 352 + ], + "score": 0.94, + "content": "\\begin{array} { r } { p ( \\psi ( . ) ) = p ( \\sqrt { \\frac { 1 - \\beta _ { 2 } ^ { t } } { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } } ) \\approx p ( \\sqrt { \\frac { t } { \\sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 331, + 401, + 345 + ], + "score": 1.0, + "content": "Since", + "type": "text" + }, + { + "bbox": [ + 402, + 331, + 464, + 344 + ], + "score": 0.92, + "content": "g _ { i } \\sim \\mathcal { N } ( 0 , \\sigma ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 331, + 505, + 345 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 347, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 107, + 353, + 234, + 370 + ], + "score": 0.25, + "content": "{ \\frac { t } { \\sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } } \\sim \\mathrm { S c a l e } { \\mathrm { - i n v } } { - \\mathcal { X } } ^ { 2 } ( t , { \\frac { 1 } { \\sigma ^ { 2 } } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 347, + 330, + 373 + ], + "score": 1.0, + "content": ". Therefore, we assume", + "type": "text" + }, + { + "bbox": [ + 330, + 350, + 405, + 370 + ], + "score": 0.94, + "content": "\\frac { 1 - \\beta _ { 2 } ^ { t } } { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 347, + 505, + 373 + ], + "score": 1.0, + "content": "also subjects to a scaled", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 250, + 381 + ], + "score": 1.0, + "content": "inverse chi-square distribution with", + "type": "text" + }, + { + "bbox": [ + 250, + 371, + 257, + 380 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 368, + 505, + 381 + ], + "score": 1.0, + "content": "degrees of freedom (further analysis on this approximation is", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 394, + 392 + ], + "score": 1.0, + "content": "conducted in Section 5.3). Based on this assumption, we can calculate", + "type": "text" + }, + { + "bbox": [ + 395, + 379, + 439, + 391 + ], + "score": 0.92, + "content": "\\mathrm { V a r } [ \\psi ^ { 2 } ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "and the PDF of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 389, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 107, + 390, + 129, + 402 + ], + "score": 0.9, + "content": "\\psi ^ { 2 } ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 389, + 388, + 403 + ], + "score": 1.0, + "content": ". Now, we proceed to the analysis of its square root variance, i.e.,", + "type": "text" + }, + { + "bbox": [ + 388, + 391, + 428, + 402 + ], + "score": 0.91, + "content": "\\mathrm { V a r } [ \\psi ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 389, + 506, + 403 + ], + "score": 1.0, + "content": ", and show how the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 401, + 434, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 197, + 414 + ], + "score": 1.0, + "content": "variance changes with", + "type": "text" + }, + { + "bbox": [ + 197, + 403, + 204, + 413 + ], + "score": 0.79, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 401, + 434, + 414 + ], + "score": 1.0, + "content": "(which corresponds to number of used training samples).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 105, + 417, + 492, + 432 + ], + "lines": [ + { + "bbox": [ + 106, + 417, + 493, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 169, + 432 + ], + "score": 1.0, + "content": "Theorem 1. If", + "type": "text" + }, + { + "bbox": [ + 169, + 418, + 203, + 431 + ], + "score": 0.65, + "content": "\\psi ^ { 2 } ( . ) \\sim", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 417, + 245, + 432 + ], + "score": 1.0, + "content": "Scale-inv-", + "type": "text" + }, + { + "bbox": [ + 245, + 417, + 287, + 432 + ], + "score": 0.84, + "content": "\\scriptstyle { \\mathcal { X } } ^ { 2 } ( \\rho , { \\frac { 1 } { \\sigma ^ { 2 } } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 417, + 291, + 432 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 291, + 418, + 330, + 431 + ], + "score": 0.69, + "content": "\\mathrm { V a r } [ \\psi ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 417, + 443, + 432 + ], + "score": 1.0, + "content": "monotonically decreases as", + "type": "text" + }, + { + "bbox": [ + 443, + 421, + 450, + 430 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 417, + 493, + 432 + ], + "score": 1.0, + "content": "increases.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 225, + 453 + ], + "lines": [ + { + "bbox": [ + 106, + 441, + 226, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 151, + 454 + ], + "score": 1.0, + "content": "Proof. For", + "type": "text" + }, + { + "bbox": [ + 152, + 442, + 184, + 453 + ], + "score": 0.88, + "content": "\\forall \\rho > 4", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 441, + 226, + 454 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "interline_equation", + "bbox": [ + 149, + 455, + 460, + 483 + ], + "lines": [ + { + "bbox": [ + 149, + 455, + 460, + 483 + ], + "spans": [ + { + "bbox": [ + 149, + 455, + 460, + 483 + ], + "score": 0.92, + "content": "\\mathrm { V a r } [ \\psi ( . ) ] = \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] - \\mathbb { E } [ \\psi ( . ) ] ^ { 2 } = \\tau ^ { 2 } ( \\frac { \\rho } { \\rho - 2 } - \\frac { \\rho 2 ^ { 2 \\rho - 5 } } { \\pi } \\mathcal { B } ( \\frac { \\rho - 1 } { 2 } , \\frac { \\rho - 1 } { 2 } ) ^ { 2 } ) ,", + "type": "interline_equation", + "image_path": "ec5a58e45b0206f51a703fec8e95e0c988017c720c273353d4e10f20a19e4e08.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 149, + 455, + 460, + 483 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 485, + 502, + 508 + ], + "lines": [ + { + "bbox": [ + 106, + 484, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 133, + 497 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 485, + 151, + 497 + ], + "score": 0.91, + "content": "B ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 484, + 356, + 497 + ], + "score": 1.0, + "content": "is the beta function. By analyzing the derivative of", + "type": "text" + }, + { + "bbox": [ + 356, + 485, + 395, + 497 + ], + "score": 0.92, + "content": "\\mathrm { V a r } [ \\psi ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 484, + 504, + 497 + ], + "score": 1.0, + "content": ", we know it monotonically", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 496, + 504, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 158, + 508 + ], + "score": 1.0, + "content": "decreases as", + "type": "text" + }, + { + "bbox": [ + 158, + 498, + 164, + 507 + ], + "score": 0.83, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 496, + 434, + 508 + ], + "score": 1.0, + "content": "increases. The detailed derivation is elaborated in the Appendix A.", + "type": "text" + }, + { + "bbox": [ + 495, + 497, + 504, + 507 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 515, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "Theorem 1 gives a qualitative analysis of the variance of the adaptive learning rate. It shows that,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 525, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 350, + 540 + ], + "score": 1.0, + "content": "due to the lack of used training samples in the early stage,", + "type": "text" + }, + { + "bbox": [ + 350, + 527, + 390, + 539 + ], + "score": 0.92, + "content": "\\mathrm { V a r } [ \\bar { \\psi } ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 525, + 505, + 540 + ], + "score": 1.0, + "content": "is larger than the late stage", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 536, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 451, + 551 + ], + "score": 1.0, + "content": "(Figure 8). To rigorously constraint the variance, we perform a quantified analysis on", + "type": "text" + }, + { + "bbox": [ + 452, + 537, + 491, + 550 + ], + "score": 0.91, + "content": "\\mathrm { V a r } [ \\psi ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 536, + 505, + 551 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 548, + 255, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 244, + 563 + ], + "score": 1.0, + "content": "estimating the degree of freedoms", + "type": "text" + }, + { + "bbox": [ + 245, + 551, + 251, + 560 + ], + "score": 0.78, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 548, + 255, + 563 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "title", + "bbox": [ + 107, + 574, + 328, + 587 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 329, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 329, + 588 + ], + "score": 1.0, + "content": "4 RECTIFIED ADAPTIVE LEARNING RATE", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 597, + 505, + 642 + ], + "lines": [ + { + "bbox": [ + 105, + 597, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 360, + 611 + ], + "score": 1.0, + "content": "In the previous section, Equation 2 gives the analytic form of", + "type": "text" + }, + { + "bbox": [ + 360, + 598, + 399, + 610 + ], + "score": 0.92, + "content": "\\mathrm { V a r } [ \\psi ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 597, + 431, + 611 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 431, + 600, + 438, + 609 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 597, + 506, + 611 + ], + "score": 1.0, + "content": "is the degree of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 608, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 302, + 621 + ], + "score": 1.0, + "content": "freedoms. Here, we first give an estimation of", + "type": "text" + }, + { + "bbox": [ + 302, + 610, + 309, + 621 + ], + "score": 0.83, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 608, + 350, + 621 + ], + "score": 1.0, + "content": "based on", + "type": "text" + }, + { + "bbox": [ + 351, + 610, + 356, + 619 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 608, + 506, + 621 + ], + "score": 1.0, + "content": "to conduct a quantified analysis for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 618, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 184, + 632 + ], + "score": 0.93, + "content": "\\operatorname { V a r } [ \\psi ( g _ { 1 } , \\cdots , g _ { t } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 618, + 506, + 633 + ], + "score": 1.0, + "content": ", then we describe the design of the learning rate rectification, and compare it to", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 630, + 235, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 235, + 644 + ], + "score": 1.0, + "content": "the heuristic warmup strategies.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5 + }, + { + "type": "title", + "bbox": [ + 107, + 654, + 208, + 666 + ], + "lines": [ + { + "bbox": [ + 105, + 651, + 207, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 200, + 669 + ], + "score": 1.0, + "content": "4.1 ESTIMATION OF", + "type": "text" + }, + { + "bbox": [ + 200, + 657, + 207, + 666 + ], + "score": 0.69, + "content": "\\rho", + "type": "inline_equation" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 672, + 504, + 695 + ], + "lines": [ + { + "bbox": [ + 106, + 672, + 504, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 672, + 504, + 685 + ], + "score": 1.0, + "content": "The exponential moving average (EMA) can be interpreted as an approximation to the simple mov-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 683, + 330, + 697 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 330, + 697 + ], + "score": 1.0, + "content": "ing average (SMA) in real application (Nau, 2014), i.e.,", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "interline_equation", + "bbox": [ + 193, + 697, + 418, + 731 + ], + "lines": [ + { + "bbox": [ + 193, + 697, + 418, + 731 + ], + "spans": [ + { + "bbox": [ + 193, + 697, + 418, + 731 + ], + "score": 0.93, + "content": "p \\left( \\frac { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } { 1 - \\beta _ { 2 } ^ { t } } \\right) \\approx p \\left( \\frac { \\sum _ { i = 1 } ^ { f \\left( t , \\beta _ { 2 } \\right) } g _ { t + 1 - i } ^ { 2 } } { f ( t , \\beta _ { 2 } ) } \\right) .", + "type": "interline_equation", + "image_path": "793d8562d11ed3ea916b4ad47f5781c1153cbebc559f3a479f34f31434576002.jpg" + } + ] + } + ], + "index": 42.5, + "virtual_lines": [ + { + "bbox": [ + 193, + 697, + 418, + 714.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 193, + 714.0, + 418, + 731.0 + ], + "spans": [], + "index": 43 + } + ] + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 240 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 419, + 96 + ], + "score": 1.0, + "content": "Another straightforward way to reduce the variance is to increase the value of", + "type": "text" + }, + { + "bbox": [ + 419, + 85, + 425, + 92 + ], + "score": 0.7, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 81, + 436, + 96 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 436, + 81, + 506, + 95 + ], + "score": 0.91, + "content": "\\widehat { \\psi } ( g _ { 1 } , \\cdots , g _ { t } ) =", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 93, + 504, + 118 + ], + "spans": [ + { + "bbox": [ + 107, + 93, + 200, + 118 + ], + "score": 0.91, + "content": "\\frac { \\sqrt { 1 - \\beta _ { 2 } ^ { t } } } { \\epsilon + \\sqrt { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 99, + 299, + 112 + ], + "score": 1.0, + "content": "Actually, if we assume", + "type": "text" + }, + { + "bbox": [ + 299, + 97, + 318, + 111 + ], + "score": 0.91, + "content": "\\widehat { \\psi } ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 99, + 504, + 112 + ], + "score": 1.0, + "content": "is subject to the uniform distribution, its vari-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 103, + 115, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 103, + 115, + 167, + 133 + ], + "score": 1.0, + "content": "ance equals to", + "type": "text" + }, + { + "bbox": [ + 167, + 117, + 185, + 130 + ], + "score": 0.91, + "content": "\\frac { 1 } { 1 2 \\epsilon ^ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 115, + 461, + 133 + ], + "score": 1.0, + "content": ". Therefore, we design Adam-eps, which uses a non-negligibly large", + "type": "text" + }, + { + "bbox": [ + 461, + 117, + 501, + 128 + ], + "score": 0.9, + "content": "\\epsilon = 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 115, + 506, + 133 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 128, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 131, + 144 + ], + "score": 1.0, + "content": "while", + "type": "text" + }, + { + "bbox": [ + 131, + 129, + 172, + 140 + ], + "score": 0.9, + "content": "\\epsilon = 1 0 ^ { - 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 128, + 506, + 144 + ], + "score": 1.0, + "content": "for vanilla Adam. Its performance is summarized in Figure 1. We observe that it", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 141, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 505, + 153 + ], + "score": 1.0, + "content": "does not suffer from the serious convergence problem of vanilla-Adam. This further demonstrates", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 151, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 505, + 165 + ], + "score": 1.0, + "content": "that the convergence problem can be alleviated by reducing the variance of the adaptive learning", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 162, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 248, + 176 + ], + "score": 1.0, + "content": "rate, and also explains why tuning", + "type": "text" + }, + { + "bbox": [ + 248, + 165, + 254, + 173 + ], + "score": 0.64, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 162, + 506, + 176 + ], + "score": 1.0, + "content": "is important in practice (Liu et al., 2019). Besides, similar to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 172, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 104, + 172, + 506, + 187 + ], + "score": 1.0, + "content": "Adam-2k, it prevents the gradient distribution from being distorted (as shown in Figure 3). However,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 185, + 504, + 197 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 410, + 197 + ], + "score": 1.0, + "content": "as in Figure 1, it produces a much worse performance comparing to Adam-", + "type": "text" + }, + { + "bbox": [ + 410, + 185, + 422, + 195 + ], + "score": 0.31, + "content": "2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 185, + 504, + 197 + ], + "score": 1.0, + "content": "and Adam-warmup.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 195, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 271, + 208 + ], + "score": 1.0, + "content": "We conjecture that this is because large", + "type": "text" + }, + { + "bbox": [ + 271, + 198, + 277, + 205 + ], + "score": 0.57, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 195, + 505, + 208 + ], + "score": 1.0, + "content": "induces a large bias into the adaptive learning rate and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 207, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 505, + 219 + ], + "score": 1.0, + "content": "slows down the optimization process. Thus, we need a more principled and rigorous way to con-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 217, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 505, + 230 + ], + "score": 1.0, + "content": "trol the variance of the adaptive learning rate. In the next subsection, we will present a theoretical", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 228, + 318, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 318, + 241 + ], + "score": 1.0, + "content": "analysis of the variance of the adaptive learning rate.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 6, + "bbox_fs": [ + 103, + 81, + 506, + 241 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 252, + 359, + 263 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 360, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 360, + 264 + ], + "score": 1.0, + "content": "3.2 ANALYSIS OF ADAPTIVE LEARNING RATE VARIANCE", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 270, + 505, + 413 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 284 + ], + "score": 1.0, + "content": "As mentioned before, Adam uses the exponential moving average to calculate the adaptive learning", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 185, + 294 + ], + "score": 1.0, + "content": "rate. For gradients", + "type": "text" + }, + { + "bbox": [ + 185, + 281, + 236, + 293 + ], + "score": 0.96, + "content": "\\{ g _ { 1 } , \\cdots , g _ { t } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 281, + 505, + 294 + ], + "score": 1.0, + "content": ", their exponential moving average has a larger variance than their", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 293, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 276, + 304 + ], + "score": 1.0, + "content": "simple average. Also, in the early stage", + "type": "text" + }, + { + "bbox": [ + 276, + 294, + 281, + 302 + ], + "score": 0.66, + "content": "\\mathbf { \\chi } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 293, + 505, + 304 + ], + "score": 1.0, + "content": "is small), the difference of the exponential weights of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 300, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 107, + 303, + 159, + 315 + ], + "score": 0.92, + "content": "\\{ g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 300, + 262, + 317 + ], + "score": 1.0, + "content": "is relatively small (up to", + "type": "text" + }, + { + "bbox": [ + 262, + 302, + 302, + 316 + ], + "score": 0.9, + "content": "\\breve { 1 } - \\beta _ { 2 } ^ { t - 1 } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 300, + 506, + 317 + ], + "score": 1.0, + "content": "). Therefore, for ease of analysis, we approximate", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "score": 1.0, + "content": "the distribution of the exponential moving average as the distribution of the simple average (Nau,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 325, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 221, + 348 + ], + "score": 1.0, + "content": "2014), i.e., p(ψ(.)) = p(", + "type": "text" + }, + { + "bbox": [ + 154, + 325, + 369, + 352 + ], + "score": 0.94, + "content": "\\begin{array} { r } { p ( \\psi ( . ) ) = p ( \\sqrt { \\frac { 1 - \\beta _ { 2 } ^ { t } } { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } } ) \\approx p ( \\sqrt { \\frac { t } { \\sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 331, + 401, + 345 + ], + "score": 1.0, + "content": "Since", + "type": "text" + }, + { + "bbox": [ + 402, + 331, + 464, + 344 + ], + "score": 0.92, + "content": "g _ { i } \\sim \\mathcal { N } ( 0 , \\sigma ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 331, + 505, + 345 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 347, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 107, + 353, + 234, + 370 + ], + "score": 0.25, + "content": "{ \\frac { t } { \\sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } } \\sim \\mathrm { S c a l e } { \\mathrm { - i n v } } { - \\mathcal { X } } ^ { 2 } ( t , { \\frac { 1 } { \\sigma ^ { 2 } } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 347, + 330, + 373 + ], + "score": 1.0, + "content": ". Therefore, we assume", + "type": "text" + }, + { + "bbox": [ + 330, + 350, + 405, + 370 + ], + "score": 0.94, + "content": "\\frac { 1 - \\beta _ { 2 } ^ { t } } { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 347, + 505, + 373 + ], + "score": 1.0, + "content": "also subjects to a scaled", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 250, + 381 + ], + "score": 1.0, + "content": "inverse chi-square distribution with", + "type": "text" + }, + { + "bbox": [ + 250, + 371, + 257, + 380 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 368, + 505, + 381 + ], + "score": 1.0, + "content": "degrees of freedom (further analysis on this approximation is", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 394, + 392 + ], + "score": 1.0, + "content": "conducted in Section 5.3). Based on this assumption, we can calculate", + "type": "text" + }, + { + "bbox": [ + 395, + 379, + 439, + 391 + ], + "score": 0.92, + "content": "\\mathrm { V a r } [ \\psi ^ { 2 } ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "and the PDF of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 389, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 107, + 390, + 129, + 402 + ], + "score": 0.9, + "content": "\\psi ^ { 2 } ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 389, + 388, + 403 + ], + "score": 1.0, + "content": ". Now, we proceed to the analysis of its square root variance, i.e.,", + "type": "text" + }, + { + "bbox": [ + 388, + 391, + 428, + 402 + ], + "score": 0.91, + "content": "\\mathrm { V a r } [ \\psi ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 389, + 506, + 403 + ], + "score": 1.0, + "content": ", and show how the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 401, + 434, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 197, + 414 + ], + "score": 1.0, + "content": "variance changes with", + "type": "text" + }, + { + "bbox": [ + 197, + 403, + 204, + 413 + ], + "score": 0.79, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 401, + 434, + 414 + ], + "score": 1.0, + "content": "(which corresponds to number of used training samples).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 268, + 506, + 414 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 417, + 492, + 432 + ], + "lines": [ + { + "bbox": [ + 106, + 417, + 493, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 169, + 432 + ], + "score": 1.0, + "content": "Theorem 1. If", + "type": "text" + }, + { + "bbox": [ + 169, + 418, + 203, + 431 + ], + "score": 0.65, + "content": "\\psi ^ { 2 } ( . ) \\sim", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 417, + 245, + 432 + ], + "score": 1.0, + "content": "Scale-inv-", + "type": "text" + }, + { + "bbox": [ + 245, + 417, + 287, + 432 + ], + "score": 0.84, + "content": "\\scriptstyle { \\mathcal { X } } ^ { 2 } ( \\rho , { \\frac { 1 } { \\sigma ^ { 2 } } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 417, + 291, + 432 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 291, + 418, + 330, + 431 + ], + "score": 0.69, + "content": "\\mathrm { V a r } [ \\psi ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 417, + 443, + 432 + ], + "score": 1.0, + "content": "monotonically decreases as", + "type": "text" + }, + { + "bbox": [ + 443, + 421, + 450, + 430 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 417, + 493, + 432 + ], + "score": 1.0, + "content": "increases.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 417, + 493, + 432 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 225, + 453 + ], + "lines": [ + { + "bbox": [ + 106, + 441, + 226, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 151, + 454 + ], + "score": 1.0, + "content": "Proof. For", + "type": "text" + }, + { + "bbox": [ + 152, + 442, + 184, + 453 + ], + "score": 0.88, + "content": "\\forall \\rho > 4", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 441, + 226, + 454 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 106, + 441, + 226, + 454 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 149, + 455, + 460, + 483 + ], + "lines": [ + { + "bbox": [ + 149, + 455, + 460, + 483 + ], + "spans": [ + { + "bbox": [ + 149, + 455, + 460, + 483 + ], + "score": 0.92, + "content": "\\mathrm { V a r } [ \\psi ( . ) ] = \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] - \\mathbb { E } [ \\psi ( . ) ] ^ { 2 } = \\tau ^ { 2 } ( \\frac { \\rho } { \\rho - 2 } - \\frac { \\rho 2 ^ { 2 \\rho - 5 } } { \\pi } \\mathcal { B } ( \\frac { \\rho - 1 } { 2 } , \\frac { \\rho - 1 } { 2 } ) ^ { 2 } ) ,", + "type": "interline_equation", + "image_path": "ec5a58e45b0206f51a703fec8e95e0c988017c720c273353d4e10f20a19e4e08.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 149, + 455, + 460, + 483 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 485, + 502, + 508 + ], + "lines": [ + { + "bbox": [ + 106, + 484, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 133, + 497 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 485, + 151, + 497 + ], + "score": 0.91, + "content": "B ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 484, + 356, + 497 + ], + "score": 1.0, + "content": "is the beta function. By analyzing the derivative of", + "type": "text" + }, + { + "bbox": [ + 356, + 485, + 395, + 497 + ], + "score": 0.92, + "content": "\\mathrm { V a r } [ \\psi ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 484, + 504, + 497 + ], + "score": 1.0, + "content": ", we know it monotonically", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 496, + 504, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 158, + 508 + ], + "score": 1.0, + "content": "decreases as", + "type": "text" + }, + { + "bbox": [ + 158, + 498, + 164, + 507 + ], + "score": 0.83, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 496, + 434, + 508 + ], + "score": 1.0, + "content": "increases. The detailed derivation is elaborated in the Appendix A.", + "type": "text" + }, + { + "bbox": [ + 495, + 497, + 504, + 507 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 484, + 504, + 508 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 515, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "Theorem 1 gives a qualitative analysis of the variance of the adaptive learning rate. It shows that,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 525, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 350, + 540 + ], + "score": 1.0, + "content": "due to the lack of used training samples in the early stage,", + "type": "text" + }, + { + "bbox": [ + 350, + 527, + 390, + 539 + ], + "score": 0.92, + "content": "\\mathrm { V a r } [ \\bar { \\psi } ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 525, + 505, + 540 + ], + "score": 1.0, + "content": "is larger than the late stage", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 536, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 451, + 551 + ], + "score": 1.0, + "content": "(Figure 8). To rigorously constraint the variance, we perform a quantified analysis on", + "type": "text" + }, + { + "bbox": [ + 452, + 537, + 491, + 550 + ], + "score": 0.91, + "content": "\\mathrm { V a r } [ \\psi ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 536, + 505, + 551 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 548, + 255, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 244, + 563 + ], + "score": 1.0, + "content": "estimating the degree of freedoms", + "type": "text" + }, + { + "bbox": [ + 245, + 551, + 251, + 560 + ], + "score": 0.78, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 548, + 255, + 563 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 515, + 505, + 563 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 574, + 328, + 587 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 329, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 329, + 588 + ], + "score": 1.0, + "content": "4 RECTIFIED ADAPTIVE LEARNING RATE", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 597, + 505, + 642 + ], + "lines": [ + { + "bbox": [ + 105, + 597, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 360, + 611 + ], + "score": 1.0, + "content": "In the previous section, Equation 2 gives the analytic form of", + "type": "text" + }, + { + "bbox": [ + 360, + 598, + 399, + 610 + ], + "score": 0.92, + "content": "\\mathrm { V a r } [ \\psi ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 597, + 431, + 611 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 431, + 600, + 438, + 609 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 597, + 506, + 611 + ], + "score": 1.0, + "content": "is the degree of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 608, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 302, + 621 + ], + "score": 1.0, + "content": "freedoms. Here, we first give an estimation of", + "type": "text" + }, + { + "bbox": [ + 302, + 610, + 309, + 621 + ], + "score": 0.83, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 608, + 350, + 621 + ], + "score": 1.0, + "content": "based on", + "type": "text" + }, + { + "bbox": [ + 351, + 610, + 356, + 619 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 608, + 506, + 621 + ], + "score": 1.0, + "content": "to conduct a quantified analysis for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 618, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 184, + 632 + ], + "score": 0.93, + "content": "\\operatorname { V a r } [ \\psi ( g _ { 1 } , \\cdots , g _ { t } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 618, + 506, + 633 + ], + "score": 1.0, + "content": ", then we describe the design of the learning rate rectification, and compare it to", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 630, + 235, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 235, + 644 + ], + "score": 1.0, + "content": "the heuristic warmup strategies.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 597, + 506, + 644 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 654, + 208, + 666 + ], + "lines": [ + { + "bbox": [ + 105, + 651, + 207, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 200, + 669 + ], + "score": 1.0, + "content": "4.1 ESTIMATION OF", + "type": "text" + }, + { + "bbox": [ + 200, + 657, + 207, + 666 + ], + "score": 0.69, + "content": "\\rho", + "type": "inline_equation" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 672, + 504, + 695 + ], + "lines": [ + { + "bbox": [ + 106, + 672, + 504, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 672, + 504, + 685 + ], + "score": 1.0, + "content": "The exponential moving average (EMA) can be interpreted as an approximation to the simple mov-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 683, + 330, + 697 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 330, + 697 + ], + "score": 1.0, + "content": "ing average (SMA) in real application (Nau, 2014), i.e.,", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 672, + 504, + 697 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 193, + 697, + 418, + 731 + ], + "lines": [ + { + "bbox": [ + 193, + 697, + 418, + 731 + ], + "spans": [ + { + "bbox": [ + 193, + 697, + 418, + 731 + ], + "score": 0.93, + "content": "p \\left( \\frac { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } { 1 - \\beta _ { 2 } ^ { t } } \\right) \\approx p \\left( \\frac { \\sum _ { i = 1 } ^ { f \\left( t , \\beta _ { 2 } \\right) } g _ { t + 1 - i } ^ { 2 } } { f ( t , \\beta _ { 2 } ) } \\right) .", + "type": "interline_equation", + "image_path": "793d8562d11ed3ea916b4ad47f5781c1153cbebc559f3a479f34f31434576002.jpg" + } + ] + } + ], + "index": 42.5, + "virtual_lines": [ + { + "bbox": [ + 193, + 697, + 418, + 714.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 193, + 714.0, + 418, + 731.0 + ], + "spans": [], + "index": 43 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 85, + 364, + 97 + ], + "lines": [ + { + "bbox": [ + 106, + 84, + 365, + 99 + ], + "spans": [ + { + "bbox": [ + 106, + 84, + 365, + 99 + ], + "score": 1.0, + "content": "Algorithm 2: Rectified Adam. All operations are element-wise.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 97, + 94, + 495, + 308 + ], + "lines": [ + { + "bbox": [ + 105, + 98, + 481, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 137, + 115 + ], + "score": 1.0, + "content": "Input:", + "type": "text" + }, + { + "bbox": [ + 137, + 100, + 171, + 113 + ], + "score": 0.91, + "content": "\\{ \\alpha _ { t } \\} _ { t = 1 } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 98, + 215, + 115 + ], + "score": 1.0, + "content": ": step size,", + "type": "text" + }, + { + "bbox": [ + 216, + 100, + 251, + 113 + ], + "score": 0.92, + "content": "\\{ \\beta _ { 1 } , \\beta _ { 2 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 98, + 481, + 115 + ], + "score": 1.0, + "content": ": decay rate to calculate moving average and moving 2nd", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 136, + 111, + 405, + 124 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 174, + 124 + ], + "score": 1.0, + "content": "moment,", + "type": "text" + }, + { + "bbox": [ + 174, + 112, + 185, + 123 + ], + "score": 0.87, + "content": "\\theta _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 111, + 259, + 124 + ], + "score": 1.0, + "content": ": initial parameter,", + "type": "text" + }, + { + "bbox": [ + 260, + 111, + 281, + 124 + ], + "score": 0.93, + "content": "f _ { t } ( \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 111, + 405, + 124 + ], + "score": 1.0, + "content": ": stochastic objective function.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 243, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 143, + 135 + ], + "score": 1.0, + "content": "Output:", + "type": "text" + }, + { + "bbox": [ + 144, + 123, + 153, + 133 + ], + "score": 0.85, + "content": "\\theta _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 121, + 243, + 135 + ], + "score": 1.0, + "content": ": resulting parameters", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 133, + 327, + 145 + ], + "spans": [ + { + "bbox": [ + 97, + 133, + 106, + 145 + ], + "score": 1.0, + "content": "1", + "type": "text" + }, + { + "bbox": [ + 107, + 134, + 164, + 145 + ], + "score": 0.87, + "content": "m _ { 0 } , v _ { 0 } \\gets 0 , 0", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 133, + 327, + 145 + ], + "score": 1.0, + "content": "(Initialize moving 1st and 2nd moment)", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 144, + 438, + 158 + ], + "spans": [ + { + "bbox": [ + 96, + 144, + 106, + 158 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 106, + 145, + 198, + 156 + ], + "score": 0.85, + "content": "\\rho _ { \\infty } 2 / ( 1 - \\beta _ { 2 } ) - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 144, + 438, + 158 + ], + "score": 1.0, + "content": "(Compute the maximum length of the approximated SMA)", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 155, + 210, + 167 + ], + "spans": [ + { + "bbox": [ + 97, + 155, + 132, + 167 + ], + "score": 1.0, + "content": "3 while", + "type": "text" + }, + { + "bbox": [ + 132, + 156, + 195, + 167 + ], + "score": 0.88, + "content": "t = \\{ 1 , \\cdots , T \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 155, + 210, + 167 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 165, + 436, + 179 + ], + "spans": [ + { + "bbox": [ + 99, + 170, + 104, + 176 + ], + "score": 1.0, + "content": "4", + "type": "text" + }, + { + "bbox": [ + 122, + 167, + 194, + 178 + ], + "score": 0.87, + "content": "g _ { t } \\gets \\Delta _ { \\theta } f _ { t } ( \\theta _ { t - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 165, + 436, + 179 + ], + "score": 1.0, + "content": "(Calculate gradients w.r.t. stochastic objective at timestep t)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 98, + 178, + 416, + 191 + ], + "spans": [ + { + "bbox": [ + 98, + 182, + 104, + 189 + ], + "score": 1.0, + "content": "5", + "type": "text" + }, + { + "bbox": [ + 122, + 178, + 241, + 190 + ], + "score": 0.87, + "content": "v _ { t } 1 / \\beta _ { 2 } v _ { t - 1 } + ( 1 - \\beta _ { 2 } ) g _ { t } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 178, + 416, + 191 + ], + "score": 1.0, + "content": "(Update exponential moving 2nd moment)", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 97, + 190, + 408, + 202 + ], + "spans": [ + { + "bbox": [ + 97, + 193, + 104, + 201 + ], + "score": 1.0, + "content": "6", + "type": "text" + }, + { + "bbox": [ + 122, + 191, + 239, + 202 + ], + "score": 0.86, + "content": "m _ { t } \\gets \\beta _ { 1 } m _ { t - 1 } + ( 1 - \\beta _ { 1 } ) g _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 190, + 408, + 202 + ], + "score": 1.0, + "content": "(Update exponential moving 1st moment)", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 97, + 201, + 375, + 215 + ], + "spans": [ + { + "bbox": [ + 97, + 204, + 104, + 212 + ], + "score": 1.0, + "content": "7", + "type": "text" + }, + { + "bbox": [ + 122, + 202, + 203, + 214 + ], + "score": 0.89, + "content": "\\widehat { m _ { t } } \\gets m _ { t } / ( 1 - \\beta _ { 1 } ^ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 201, + 375, + 215 + ], + "score": 1.0, + "content": "(Compute bias-corrected moving average)", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 97, + 213, + 425, + 227 + ], + "spans": [ + { + "bbox": [ + 97, + 217, + 104, + 225 + ], + "score": 1.0, + "content": "8", + "type": "text" + }, + { + "bbox": [ + 122, + 214, + 231, + 226 + ], + "score": 0.81, + "content": "\\rho _ { t } \\gets \\rho _ { \\infty } - 2 t \\beta _ { 2 } ^ { t } / ( 1 - \\beta _ { 2 } ^ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 213, + 425, + 227 + ], + "score": 1.0, + "content": "(Compute the length of the approximated SMA)", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 99, + 224, + 300, + 238 + ], + "spans": [ + { + "bbox": [ + 99, + 228, + 104, + 235 + ], + "score": 1.0, + "content": "9", + "type": "text" + }, + { + "bbox": [ + 120, + 224, + 249, + 238 + ], + "score": 1.0, + "content": "if the variance is tractable, i.e.,", + "type": "text" + }, + { + "bbox": [ + 249, + 226, + 277, + 237 + ], + "score": 0.9, + "content": "\\rho _ { t } > 4", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 224, + 300, + 238 + ], + "score": 1.0, + "content": "then", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 96, + 236, + 354, + 251 + ], + "spans": [ + { + "bbox": [ + 96, + 240, + 104, + 248 + ], + "score": 1.0, + "content": "10", + "type": "text" + }, + { + "bbox": [ + 137, + 236, + 218, + 250 + ], + "score": 0.92, + "content": "l _ { t } \\sqrt { ( 1 - \\beta _ { 2 } ^ { t } ) / v _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 236, + 354, + 251 + ], + "score": 1.0, + "content": "(Compute adaptive learning rate)", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 96, + 248, + 406, + 270 + ], + "spans": [ + { + "bbox": [ + 96, + 257, + 104, + 266 + ], + "score": 1.0, + "content": "11", + "type": "text" + }, + { + "bbox": [ + 137, + 250, + 237, + 269 + ], + "score": 0.91, + "content": "\\begin{array} { r } { r _ { t } \\sqrt { \\frac { ( \\rho _ { t } - 4 ) ( \\rho _ { t } - 2 ) \\rho _ { \\infty } } { ( \\rho _ { \\infty } - 4 ) ( \\rho _ { \\infty } - 2 ) \\rho _ { t } } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 248, + 406, + 270 + ], + "score": 1.0, + "content": "(Compute the variance rectification term)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 95, + 267, + 419, + 283 + ], + "spans": [ + { + "bbox": [ + 95, + 271, + 104, + 281 + ], + "score": 1.0, + "content": "12", + "type": "text" + }, + { + "bbox": [ + 137, + 270, + 228, + 281 + ], + "score": 0.88, + "content": "\\theta _ { t } \\gets \\theta _ { t - 1 } - \\alpha _ { t } r _ { t } \\widehat { m _ { t } } l _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 267, + 419, + 283 + ], + "score": 1.0, + "content": "(Update parameters with adaptive momentum)", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 95, + 281, + 141, + 293 + ], + "spans": [ + { + "bbox": [ + 95, + 283, + 105, + 293 + ], + "score": 1.0, + "content": "13", + "type": "text" + }, + { + "bbox": [ + 120, + 281, + 141, + 292 + ], + "score": 1.0, + "content": "else", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 95, + 290, + 416, + 305 + ], + "spans": [ + { + "bbox": [ + 95, + 294, + 105, + 303 + ], + "score": 1.0, + "content": "14", + "type": "text" + }, + { + "bbox": [ + 125, + 291, + 214, + 305 + ], + "score": 0.48, + "content": "\\begin{array} { r } { \\big \\lfloor \\theta _ { t } \\gets \\theta _ { t - 1 } - \\alpha _ { t } \\widehat { m _ { t } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 290, + 416, + 305 + ], + "score": 1.0, + "content": "(Update parameters with un-adapted momentum)", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 95, + 311, + 149, + 322 + ], + "lines": [ + { + "bbox": [ + 94, + 310, + 149, + 323 + ], + "spans": [ + { + "bbox": [ + 94, + 310, + 136, + 323 + ], + "score": 1.0, + "content": "15 return", + "type": "text" + }, + { + "bbox": [ + 136, + 311, + 149, + 322 + ], + "score": 0.37, + "content": "\\theta _ { T }", + "type": "inline_equation" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 108, + 348, + 503, + 371 + ], + "lines": [ + { + "bbox": [ + 106, + 348, + 504, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 133, + 361 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 349, + 167, + 361 + ], + "score": 0.92, + "content": "f ( t , \\beta _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 348, + 504, + 361 + ], + "score": 1.0, + "content": "is the length of the SMA which allows the SMA to have the same “center of mass”", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 359, + 303, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 231, + 372 + ], + "score": 1.0, + "content": "with the EMA. In other words,", + "type": "text" + }, + { + "bbox": [ + 232, + 360, + 264, + 372 + ], + "score": 0.92, + "content": "f ( t , \\beta _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 359, + 303, + 372 + ], + "score": 1.0, + "content": "satisfies:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 371, + 402, + 402 + ], + "lines": [ + { + "bbox": [ + 208, + 371, + 402, + 402 + ], + "spans": [ + { + "bbox": [ + 208, + 371, + 402, + 402 + ], + "score": 0.93, + "content": "\\frac { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } \\cdot i } { 1 - \\beta _ { 2 } ^ { t } } = \\frac { \\sum _ { i = 1 } ^ { f ( t , \\beta _ { 2 } ) } ( t + 1 - i ) } { f ( t , \\beta _ { 2 } ) } .", + "type": "interline_equation", + "image_path": "9250d1cc74971050e057c907265dd057a56c53c19498aa7f986c61efd12b2617.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 371, + 402, + 386.5 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 208, + 386.5, + 402, + 402.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 403, + 506, + 501 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 257, + 423 + ], + "score": 1.0, + "content": "By solving Equation 4, we have:", + "type": "text" + }, + { + "bbox": [ + 257, + 403, + 390, + 421 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } { f ( t , \\beta _ { 2 } ) } & { { } = } & { { \\frac { 2 } { 1 - \\beta _ { 2 } } } - 1 - { \\frac { 2 t \\beta _ { 2 } ^ { t } } { 1 - \\beta _ { 2 } ^ { t } } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 401, + 397, + 423 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 399, + 405, + 505, + 419 + ], + "score": 1.0, + "content": "In the previous section,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 418, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 158, + 437 + ], + "score": 1.0, + "content": "we assume:", + "type": "text" + }, + { + "bbox": [ + 161, + 420, + 338, + 441 + ], + "score": 0.49, + "content": "\\begin{array} { r } { \\frac { 1 - \\beta _ { 2 } ^ { t } } { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } \\ \\sim \\ \\mathrm { S c a l e - i n v - } \\mathcal { X } ^ { 2 } ( \\rho , \\frac { 1 } { \\sigma ^ { 2 } } ) . } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 418, + 397, + 443 + ], + "score": 1.0, + "content": ". Here, since", + "type": "text" + }, + { + "bbox": [ + 397, + 423, + 462, + 436 + ], + "score": 0.93, + "content": "g _ { i } \\sim \\mathcal { N } ( 0 , \\sigma ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 418, + 506, + 443 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 134, + 437, + 508, + 463 + ], + "spans": [ + { + "bbox": [ + 134, + 437, + 508, + 463 + ], + "score": 1.0, + "content": "2) g2t+1−it,β2) ∼ Scale-inv-X 2(f (t, β2), 1σ2 ). Thus, Equation 3 views Scale-inv-X 2(f (t, β2), 1σ2 )", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 458, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 241, + 474 + ], + "score": 1.0, + "content": "as an approximation to Scale-inv-", + "type": "text" + }, + { + "bbox": [ + 241, + 459, + 283, + 473 + ], + "score": 0.91, + "content": "\\scriptstyle { \\mathcal { X } } ^ { 2 } ( \\rho , { \\frac { 1 } { \\sigma ^ { 2 } } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 458, + 366, + 474 + ], + "score": 1.0, + "content": ". Therefore, we treat", + "type": "text" + }, + { + "bbox": [ + 367, + 460, + 399, + 472 + ], + "score": 0.93, + "content": "f ( t , \\beta _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 458, + 478, + 474 + ], + "score": 1.0, + "content": "as an estimation of", + "type": "text" + }, + { + "bbox": [ + 478, + 462, + 484, + 472 + ], + "score": 0.78, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 458, + 506, + 474 + ], + "score": 1.0, + "content": ". For", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 469, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 104, + 469, + 363, + 489 + ], + "score": 1.0, + "content": "ease of notation, we mark f (t, β2) as ρt. Also, we refer 21−β2", + "type": "text" + }, + { + "bbox": [ + 376, + 472, + 388, + 486 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 388, + 474, + 403, + 485 + ], + "score": 0.88, + "content": "\\rho _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 472, + 505, + 486 + ], + "score": 1.0, + "content": "(maximum length of the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 102, + 483, + 451, + 506 + ], + "spans": [ + { + "bbox": [ + 102, + 483, + 278, + 506 + ], + "score": 1.0, + "content": "approximated SMA), due to the inequality", + "type": "text" + }, + { + "bbox": [ + 279, + 486, + 443, + 502 + ], + "score": 0.86, + "content": "\\begin{array} { r } { f ( t , \\beta _ { 2 } ) \\leq \\operatorname* { l i m } _ { t \\infty } f ( t , \\beta _ { 2 } ) = \\frac { 2 } { 1 - \\beta _ { 2 } } - 1 . } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 483, + 451, + 506 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25.5 + }, + { + "type": "title", + "bbox": [ + 106, + 507, + 325, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 507, + 325, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 325, + 520 + ], + "score": 1.0, + "content": "4.2 VARIANCE ESTIMATION AND RECTIFICATION", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 523, + 506, + 596 + ], + "lines": [ + { + "bbox": [ + 103, + 521, + 507, + 541 + ], + "spans": [ + { + "bbox": [ + 103, + 522, + 363, + 541 + ], + "score": 1.0, + "content": "Based on previous estimations, we have Var[ψ(.)] = τ 2( ρtρ −2", + "type": "text" + }, + { + "bbox": [ + 274, + 522, + 479, + 540 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\mathrm { V a r } [ \\psi ( . ) ] = \\tau ^ { 2 } ( \\frac { \\rho _ { t } } { \\rho _ { t } - 2 } - \\frac { \\rho _ { t } 2 ^ { 2 \\rho _ { t } - 5 } } { \\pi } \\mathcal { B } ( \\frac { \\rho _ { t } - 1 } { 2 } , \\frac { \\rho _ { t } - 1 } { 2 } ) ^ { 2 } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 521, + 507, + 539 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 537, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 506, + 551 + ], + "score": 1.0, + "content": "value of this function in the early stage is significantly larger than the late stage (as analyzed later, it", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 548, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 230, + 563 + ], + "score": 1.0, + "content": "decays roughly at the speed of", + "type": "text" + }, + { + "bbox": [ + 230, + 548, + 259, + 563 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\dot { O } ( \\frac { 1 } { \\rho _ { t } } \\overline { { ) } } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 548, + 379, + 563 + ], + "score": 1.0, + "content": ". For example, the variance at", + "type": "text" + }, + { + "bbox": [ + 380, + 550, + 408, + 561 + ], + "score": 0.91, + "content": "\\rho _ { t } = 5", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 548, + 506, + 563 + ], + "score": 1.0, + "content": "is over 100 times larger", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 561, + 504, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 190, + 576 + ], + "score": 1.0, + "content": "than the variance at", + "type": "text" + }, + { + "bbox": [ + 190, + 563, + 232, + 574 + ], + "score": 0.89, + "content": "\\rho _ { t } ~ = ~ 5 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 561, + 425, + 576 + ], + "score": 1.0, + "content": ". Additionally, based on Theorem 1, we know", + "type": "text" + }, + { + "bbox": [ + 426, + 561, + 504, + 575 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\operatorname* { m i n } _ { \\rho _ { t } } \\mathrm { V a r } [ \\psi ( . ) ] = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 33 + }, + { + "bbox": [ + 107, + 572, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 107, + 573, + 174, + 586 + ], + "score": 0.91, + "content": "\\mathrm { V a r } [ \\psi ( . ) ] | _ { \\rho _ { t } = \\rho _ { \\infty } }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 572, + 303, + 588 + ], + "score": 1.0, + "content": "and mark this minimal value as", + "type": "text" + }, + { + "bbox": [ + 303, + 574, + 325, + 585 + ], + "score": 0.82, + "content": "C _ { \\mathrm { V a r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 572, + 506, + 588 + ], + "score": 1.0, + "content": ". In order to ensure that the adaptive learning", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 584, + 466, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 124, + 597 + ], + "score": 1.0, + "content": "rate", + "type": "text" + }, + { + "bbox": [ + 124, + 586, + 149, + 596 + ], + "score": 0.88, + "content": "( \\psi ( . ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 584, + 364, + 597 + ], + "score": 1.0, + "content": "has consistent variance, we rectify the variance at the", + "type": "text" + }, + { + "bbox": [ + 365, + 586, + 370, + 594 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 584, + 466, + 597 + ], + "score": 1.0, + "content": "-th timestamp as below,", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5 + }, + { + "type": "interline_equation", + "bbox": [ + 152, + 597, + 457, + 612 + ], + "lines": [ + { + "bbox": [ + 152, + 597, + 457, + 612 + ], + "spans": [ + { + "bbox": [ + 152, + 597, + 457, + 612 + ], + "score": 0.35, + "content": "\\mathrm { V a r } [ r _ { t } \\psi ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } ) ] = C _ { \\mathrm { V a r } } \\quad \\mathrm { w h e r e } \\quad r _ { t } = \\sqrt { C _ { \\mathrm { V a r } } / \\mathrm { V a r } [ \\psi ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } ) ] } .", + "type": "interline_equation", + "image_path": "bfa5cdf77c87e44a8a93bc2010ec5df6421df1855b31d0f5551a951ec91161c1.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 152, + 597, + 457, + 612 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 613, + 504, + 648 + ], + "lines": [ + { + "bbox": [ + 105, + 612, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 271, + 626 + ], + "score": 1.0, + "content": "Although we have the analytic form of", + "type": "text" + }, + { + "bbox": [ + 271, + 613, + 311, + 625 + ], + "score": 0.91, + "content": "\\mathrm { V a r } [ \\psi ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 612, + 505, + 626 + ], + "score": 1.0, + "content": "(i.e., Equation 2), it is not numerically stable.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 622, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 637 + ], + "score": 1.0, + "content": "Therefore, we use the first-order approximation to calculate the rectification term. Specifically, by", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 635, + 331, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 167, + 649 + ], + "score": 1.0, + "content": "approximating", + "type": "text" + }, + { + "bbox": [ + 168, + 635, + 200, + 649 + ], + "score": 0.93, + "content": "\\sqrt { \\psi ^ { 2 } ( . ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 636, + 331, + 649 + ], + "score": 1.0, + "content": "to the first order (Wolter, 2007),", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "interline_equation", + "bbox": [ + 121, + 649, + 489, + 678 + ], + "lines": [ + { + "bbox": [ + 121, + 649, + 489, + 678 + ], + "spans": [ + { + "bbox": [ + 121, + 649, + 489, + 678 + ], + "score": 0.92, + "content": "\\sqrt { \\psi ^ { 2 } ( . ) } \\approx \\sqrt { \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] } + \\frac { 1 } { 2 \\sqrt { \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] } } ( \\psi ^ { 2 } ( . ) - \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] ) \\quad \\mathrm { a n d } \\quad \\mathrm { V a r } [ \\psi ( . ) ] \\approx \\frac { \\mathrm { V a r } [ \\psi ^ { 2 } ( . ) ] } { 4 \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] } .", + "type": "interline_equation", + "image_path": "2bcbe7d69530063b78d06b38316d33f1623fb2c1ec18a55d5780992c1a9a7bd1.jpg" + } + ] + } + ], + "index": 41, + "virtual_lines": [ + { + "bbox": [ + 121, + 649, + 489, + 658.6666666666666 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 121, + 658.6666666666666, + 489, + 668.3333333333333 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 121, + 668.3333333333333, + 489, + 677.9999999999999 + ], + "spans": [], + "index": 42 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 294, + 692 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 131, + 694 + ], + "score": 1.0, + "content": "Since", + "type": "text" + }, + { + "bbox": [ + 131, + 679, + 253, + 693 + ], + "score": 0.5, + "content": "\\psi ^ { 2 } ( . ) \\sim \\mathrm { S c a l e } { \\mathrm { - i n v } { - } } \\mathcal { X } ^ { 2 } ( \\rho _ { t } , \\frac { 1 } { \\sigma ^ { 2 } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 676, + 295, + 694 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "interline_equation", + "bbox": [ + 226, + 693, + 385, + 707 + ], + "lines": [ + { + "bbox": [ + 226, + 693, + 385, + 707 + ], + "spans": [ + { + "bbox": [ + 226, + 693, + 385, + 707 + ], + "score": 0.93, + "content": "\\mathrm { V a r } [ \\psi ( . ) ] \\approx \\rho _ { t } / [ 2 ( \\rho _ { t } - 2 ) ( \\rho _ { t } - 4 ) \\sigma ^ { 2 } ] .", + "type": "interline_equation", + "image_path": "7f29c5400faf9a111206eaa21ee57789c31d688b9384dae42e160d85e180cfb7.jpg" + } + ] + } + ], + "index": 44, + "virtual_lines": [ + { + "bbox": [ + 226, + 693, + 385, + 707 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 708, + 507, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 707, + 505, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 505, + 720 + ], + "score": 1.0, + "content": "In Section 5.3, we conduct simulation experiments to examine Equation 5 and find that it is a reliable", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 719, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 325, + 734 + ], + "score": 1.0, + "content": "approximation. Based on Equation 5, we know that", + "type": "text" + }, + { + "bbox": [ + 325, + 719, + 374, + 733 + ], + "score": 0.93, + "content": "\\mathrm { V a r } [ \\sqrt { \\psi ( . ) } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "decreases approximately at the", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 85, + 364, + 97 + ], + "lines": [ + { + "bbox": [ + 106, + 84, + 365, + 99 + ], + "spans": [ + { + "bbox": [ + 106, + 84, + 365, + 99 + ], + "score": 1.0, + "content": "Algorithm 2: Rectified Adam. All operations are element-wise.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 106, + 84, + 365, + 99 + ] + }, + { + "type": "index", + "bbox": [ + 97, + 94, + 495, + 308 + ], + "lines": [ + { + "bbox": [ + 105, + 98, + 481, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 137, + 115 + ], + "score": 1.0, + "content": "Input:", + "type": "text" + }, + { + "bbox": [ + 137, + 100, + 171, + 113 + ], + "score": 0.91, + "content": "\\{ \\alpha _ { t } \\} _ { t = 1 } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 98, + 215, + 115 + ], + "score": 1.0, + "content": ": step size,", + "type": "text" + }, + { + "bbox": [ + 216, + 100, + 251, + 113 + ], + "score": 0.92, + "content": "\\{ \\beta _ { 1 } , \\beta _ { 2 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 98, + 481, + 115 + ], + "score": 1.0, + "content": ": decay rate to calculate moving average and moving 2nd", + "type": "text" + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 136, + 111, + 405, + 124 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 174, + 124 + ], + "score": 1.0, + "content": "moment,", + "type": "text" + }, + { + "bbox": [ + 174, + 112, + 185, + 123 + ], + "score": 0.87, + "content": "\\theta _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 111, + 259, + 124 + ], + "score": 1.0, + "content": ": initial parameter,", + "type": "text" + }, + { + "bbox": [ + 260, + 111, + 281, + 124 + ], + "score": 0.93, + "content": "f _ { t } ( \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 111, + 405, + 124 + ], + "score": 1.0, + "content": ": stochastic objective function.", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 121, + 243, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 143, + 135 + ], + "score": 1.0, + "content": "Output:", + "type": "text" + }, + { + "bbox": [ + 144, + 123, + 153, + 133 + ], + "score": 0.85, + "content": "\\theta _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 121, + 243, + 135 + ], + "score": 1.0, + "content": ": resulting parameters", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 97, + 133, + 327, + 145 + ], + "spans": [ + { + "bbox": [ + 97, + 133, + 106, + 145 + ], + "score": 1.0, + "content": "1", + "type": "text" + }, + { + "bbox": [ + 107, + 134, + 164, + 145 + ], + "score": 0.87, + "content": "m _ { 0 } , v _ { 0 } \\gets 0 , 0", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 133, + 327, + 145 + ], + "score": 1.0, + "content": "(Initialize moving 1st and 2nd moment)", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 96, + 144, + 438, + 158 + ], + "spans": [ + { + "bbox": [ + 96, + 144, + 106, + 158 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 106, + 145, + 198, + 156 + ], + "score": 0.85, + "content": "\\rho _ { \\infty } 2 / ( 1 - \\beta _ { 2 } ) - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 144, + 438, + 158 + ], + "score": 1.0, + "content": "(Compute the maximum length of the approximated SMA)", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 97, + 155, + 210, + 167 + ], + "spans": [ + { + "bbox": [ + 97, + 155, + 132, + 167 + ], + "score": 1.0, + "content": "3 while", + "type": "text" + }, + { + "bbox": [ + 132, + 156, + 195, + 167 + ], + "score": 0.88, + "content": "t = \\{ 1 , \\cdots , T \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 155, + 210, + 167 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 99, + 165, + 436, + 179 + ], + "spans": [ + { + "bbox": [ + 99, + 170, + 104, + 176 + ], + "score": 1.0, + "content": "4", + "type": "text" + }, + { + "bbox": [ + 122, + 167, + 194, + 178 + ], + "score": 0.87, + "content": "g _ { t } \\gets \\Delta _ { \\theta } f _ { t } ( \\theta _ { t - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 165, + 436, + 179 + ], + "score": 1.0, + "content": "(Calculate gradients w.r.t. stochastic objective at timestep t)", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 98, + 178, + 416, + 191 + ], + "spans": [ + { + "bbox": [ + 98, + 182, + 104, + 189 + ], + "score": 1.0, + "content": "5", + "type": "text" + }, + { + "bbox": [ + 122, + 178, + 241, + 190 + ], + "score": 0.87, + "content": "v _ { t } 1 / \\beta _ { 2 } v _ { t - 1 } + ( 1 - \\beta _ { 2 } ) g _ { t } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 178, + 416, + 191 + ], + "score": 1.0, + "content": "(Update exponential moving 2nd moment)", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 97, + 190, + 408, + 202 + ], + "spans": [ + { + "bbox": [ + 97, + 193, + 104, + 201 + ], + "score": 1.0, + "content": "6", + "type": "text" + }, + { + "bbox": [ + 122, + 191, + 239, + 202 + ], + "score": 0.86, + "content": "m _ { t } \\gets \\beta _ { 1 } m _ { t - 1 } + ( 1 - \\beta _ { 1 } ) g _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 190, + 408, + 202 + ], + "score": 1.0, + "content": "(Update exponential moving 1st moment)", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 97, + 201, + 375, + 215 + ], + "spans": [ + { + "bbox": [ + 97, + 204, + 104, + 212 + ], + "score": 1.0, + "content": "7", + "type": "text" + }, + { + "bbox": [ + 122, + 202, + 203, + 214 + ], + "score": 0.89, + "content": "\\widehat { m _ { t } } \\gets m _ { t } / ( 1 - \\beta _ { 1 } ^ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 201, + 375, + 215 + ], + "score": 1.0, + "content": "(Compute bias-corrected moving average)", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 97, + 213, + 425, + 227 + ], + "spans": [ + { + "bbox": [ + 97, + 217, + 104, + 225 + ], + "score": 1.0, + "content": "8", + "type": "text" + }, + { + "bbox": [ + 122, + 214, + 231, + 226 + ], + "score": 0.81, + "content": "\\rho _ { t } \\gets \\rho _ { \\infty } - 2 t \\beta _ { 2 } ^ { t } / ( 1 - \\beta _ { 2 } ^ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 213, + 425, + 227 + ], + "score": 1.0, + "content": "(Compute the length of the approximated SMA)", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 99, + 224, + 300, + 238 + ], + "spans": [ + { + "bbox": [ + 99, + 228, + 104, + 235 + ], + "score": 1.0, + "content": "9", + "type": "text" + }, + { + "bbox": [ + 120, + 224, + 249, + 238 + ], + "score": 1.0, + "content": "if the variance is tractable, i.e.,", + "type": "text" + }, + { + "bbox": [ + 249, + 226, + 277, + 237 + ], + "score": 0.9, + "content": "\\rho _ { t } > 4", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 224, + 300, + 238 + ], + "score": 1.0, + "content": "then", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 96, + 236, + 354, + 251 + ], + "spans": [ + { + "bbox": [ + 96, + 240, + 104, + 248 + ], + "score": 1.0, + "content": "10", + "type": "text" + }, + { + "bbox": [ + 137, + 236, + 218, + 250 + ], + "score": 0.92, + "content": "l _ { t } \\sqrt { ( 1 - \\beta _ { 2 } ^ { t } ) / v _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 236, + 354, + 251 + ], + "score": 1.0, + "content": "(Compute adaptive learning rate)", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 96, + 248, + 406, + 270 + ], + "spans": [ + { + "bbox": [ + 96, + 257, + 104, + 266 + ], + "score": 1.0, + "content": "11", + "type": "text" + }, + { + "bbox": [ + 137, + 250, + 237, + 269 + ], + "score": 0.91, + "content": "\\begin{array} { r } { r _ { t } \\sqrt { \\frac { ( \\rho _ { t } - 4 ) ( \\rho _ { t } - 2 ) \\rho _ { \\infty } } { ( \\rho _ { \\infty } - 4 ) ( \\rho _ { \\infty } - 2 ) \\rho _ { t } } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 248, + 406, + 270 + ], + "score": 1.0, + "content": "(Compute the variance rectification term)", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 267, + 419, + 283 + ], + "spans": [ + { + "bbox": [ + 95, + 271, + 104, + 281 + ], + "score": 1.0, + "content": "12", + "type": "text" + }, + { + "bbox": [ + 137, + 270, + 228, + 281 + ], + "score": 0.88, + "content": "\\theta _ { t } \\gets \\theta _ { t - 1 } - \\alpha _ { t } r _ { t } \\widehat { m _ { t } } l _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 267, + 419, + 283 + ], + "score": 1.0, + "content": "(Update parameters with adaptive momentum)", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 281, + 141, + 293 + ], + "spans": [ + { + "bbox": [ + 95, + 283, + 105, + 293 + ], + "score": 1.0, + "content": "13", + "type": "text" + }, + { + "bbox": [ + 120, + 281, + 141, + 292 + ], + "score": 1.0, + "content": "else", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 290, + 416, + 305 + ], + "spans": [ + { + "bbox": [ + 95, + 294, + 105, + 303 + ], + "score": 1.0, + "content": "14", + "type": "text" + }, + { + "bbox": [ + 125, + 291, + 214, + 305 + ], + "score": 0.48, + "content": "\\begin{array} { r } { \\big \\lfloor \\theta _ { t } \\gets \\theta _ { t - 1 } - \\alpha _ { t } \\widehat { m _ { t } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 290, + 416, + 305 + ], + "score": 1.0, + "content": "(Update parameters with un-adapted momentum)", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + } + ], + "index": 9, + "bbox_fs": [ + 95, + 98, + 481, + 305 + ] + }, + { + "type": "text", + "bbox": [ + 95, + 311, + 149, + 322 + ], + "lines": [ + { + "bbox": [ + 94, + 310, + 149, + 323 + ], + "spans": [ + { + "bbox": [ + 94, + 310, + 136, + 323 + ], + "score": 1.0, + "content": "15 return", + "type": "text" + }, + { + "bbox": [ + 136, + 311, + 149, + 322 + ], + "score": 0.37, + "content": "\\theta _ { T }", + "type": "inline_equation" + } + ], + "index": 18 + } + ], + "index": 18, + "bbox_fs": [ + 94, + 310, + 149, + 323 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 348, + 503, + 371 + ], + "lines": [ + { + "bbox": [ + 106, + 348, + 504, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 133, + 361 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 349, + 167, + 361 + ], + "score": 0.92, + "content": "f ( t , \\beta _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 348, + 504, + 361 + ], + "score": 1.0, + "content": "is the length of the SMA which allows the SMA to have the same “center of mass”", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 359, + 303, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 231, + 372 + ], + "score": 1.0, + "content": "with the EMA. In other words,", + "type": "text" + }, + { + "bbox": [ + 232, + 360, + 264, + 372 + ], + "score": 0.92, + "content": "f ( t , \\beta _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 359, + 303, + 372 + ], + "score": 1.0, + "content": "satisfies:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 348, + 504, + 372 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 371, + 402, + 402 + ], + "lines": [ + { + "bbox": [ + 208, + 371, + 402, + 402 + ], + "spans": [ + { + "bbox": [ + 208, + 371, + 402, + 402 + ], + "score": 0.93, + "content": "\\frac { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } \\cdot i } { 1 - \\beta _ { 2 } ^ { t } } = \\frac { \\sum _ { i = 1 } ^ { f ( t , \\beta _ { 2 } ) } ( t + 1 - i ) } { f ( t , \\beta _ { 2 } ) } .", + "type": "interline_equation", + "image_path": "9250d1cc74971050e057c907265dd057a56c53c19498aa7f986c61efd12b2617.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 371, + 402, + 386.5 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 208, + 386.5, + 402, + 402.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 403, + 506, + 501 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 257, + 423 + ], + "score": 1.0, + "content": "By solving Equation 4, we have:", + "type": "text" + }, + { + "bbox": [ + 257, + 403, + 390, + 421 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } { f ( t , \\beta _ { 2 } ) } & { { } = } & { { \\frac { 2 } { 1 - \\beta _ { 2 } } } - 1 - { \\frac { 2 t \\beta _ { 2 } ^ { t } } { 1 - \\beta _ { 2 } ^ { t } } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 401, + 397, + 423 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 399, + 405, + 505, + 419 + ], + "score": 1.0, + "content": "In the previous section,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 418, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 158, + 437 + ], + "score": 1.0, + "content": "we assume:", + "type": "text" + }, + { + "bbox": [ + 161, + 420, + 338, + 441 + ], + "score": 0.49, + "content": "\\begin{array} { r } { \\frac { 1 - \\beta _ { 2 } ^ { t } } { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } \\ \\sim \\ \\mathrm { S c a l e - i n v - } \\mathcal { X } ^ { 2 } ( \\rho , \\frac { 1 } { \\sigma ^ { 2 } } ) . } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 418, + 397, + 443 + ], + "score": 1.0, + "content": ". Here, since", + "type": "text" + }, + { + "bbox": [ + 397, + 423, + 462, + 436 + ], + "score": 0.93, + "content": "g _ { i } \\sim \\mathcal { N } ( 0 , \\sigma ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 418, + 506, + 443 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 134, + 437, + 508, + 463 + ], + "spans": [ + { + "bbox": [ + 134, + 437, + 508, + 463 + ], + "score": 1.0, + "content": "2) g2t+1−it,β2) ∼ Scale-inv-X 2(f (t, β2), 1σ2 ). Thus, Equation 3 views Scale-inv-X 2(f (t, β2), 1σ2 )", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 458, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 241, + 474 + ], + "score": 1.0, + "content": "as an approximation to Scale-inv-", + "type": "text" + }, + { + "bbox": [ + 241, + 459, + 283, + 473 + ], + "score": 0.91, + "content": "\\scriptstyle { \\mathcal { X } } ^ { 2 } ( \\rho , { \\frac { 1 } { \\sigma ^ { 2 } } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 458, + 366, + 474 + ], + "score": 1.0, + "content": ". Therefore, we treat", + "type": "text" + }, + { + "bbox": [ + 367, + 460, + 399, + 472 + ], + "score": 0.93, + "content": "f ( t , \\beta _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 458, + 478, + 474 + ], + "score": 1.0, + "content": "as an estimation of", + "type": "text" + }, + { + "bbox": [ + 478, + 462, + 484, + 472 + ], + "score": 0.78, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 458, + 506, + 474 + ], + "score": 1.0, + "content": ". For", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 469, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 104, + 469, + 363, + 489 + ], + "score": 1.0, + "content": "ease of notation, we mark f (t, β2) as ρt. Also, we refer 21−β2", + "type": "text" + }, + { + "bbox": [ + 376, + 472, + 388, + 486 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 388, + 474, + 403, + 485 + ], + "score": 0.88, + "content": "\\rho _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 472, + 505, + 486 + ], + "score": 1.0, + "content": "(maximum length of the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 102, + 483, + 451, + 506 + ], + "spans": [ + { + "bbox": [ + 102, + 483, + 278, + 506 + ], + "score": 1.0, + "content": "approximated SMA), due to the inequality", + "type": "text" + }, + { + "bbox": [ + 279, + 486, + 443, + 502 + ], + "score": 0.86, + "content": "\\begin{array} { r } { f ( t , \\beta _ { 2 } ) \\leq \\operatorname* { l i m } _ { t \\infty } f ( t , \\beta _ { 2 } ) = \\frac { 2 } { 1 - \\beta _ { 2 } } - 1 . } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 483, + 451, + 506 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25.5, + "bbox_fs": [ + 102, + 401, + 508, + 506 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 507, + 325, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 507, + 325, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 325, + 520 + ], + "score": 1.0, + "content": "4.2 VARIANCE ESTIMATION AND RECTIFICATION", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 523, + 506, + 596 + ], + "lines": [ + { + "bbox": [ + 103, + 521, + 507, + 541 + ], + "spans": [ + { + "bbox": [ + 103, + 522, + 363, + 541 + ], + "score": 1.0, + "content": "Based on previous estimations, we have Var[ψ(.)] = τ 2( ρtρ −2", + "type": "text" + }, + { + "bbox": [ + 274, + 522, + 479, + 540 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\mathrm { V a r } [ \\psi ( . ) ] = \\tau ^ { 2 } ( \\frac { \\rho _ { t } } { \\rho _ { t } - 2 } - \\frac { \\rho _ { t } 2 ^ { 2 \\rho _ { t } - 5 } } { \\pi } \\mathcal { B } ( \\frac { \\rho _ { t } - 1 } { 2 } , \\frac { \\rho _ { t } - 1 } { 2 } ) ^ { 2 } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 521, + 507, + 539 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 537, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 506, + 551 + ], + "score": 1.0, + "content": "value of this function in the early stage is significantly larger than the late stage (as analyzed later, it", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 548, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 230, + 563 + ], + "score": 1.0, + "content": "decays roughly at the speed of", + "type": "text" + }, + { + "bbox": [ + 230, + 548, + 259, + 563 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\dot { O } ( \\frac { 1 } { \\rho _ { t } } \\overline { { ) } } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 548, + 379, + 563 + ], + "score": 1.0, + "content": ". For example, the variance at", + "type": "text" + }, + { + "bbox": [ + 380, + 550, + 408, + 561 + ], + "score": 0.91, + "content": "\\rho _ { t } = 5", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 548, + 506, + 563 + ], + "score": 1.0, + "content": "is over 100 times larger", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 561, + 504, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 190, + 576 + ], + "score": 1.0, + "content": "than the variance at", + "type": "text" + }, + { + "bbox": [ + 190, + 563, + 232, + 574 + ], + "score": 0.89, + "content": "\\rho _ { t } ~ = ~ 5 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 561, + 425, + 576 + ], + "score": 1.0, + "content": ". Additionally, based on Theorem 1, we know", + "type": "text" + }, + { + "bbox": [ + 426, + 561, + 504, + 575 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\operatorname* { m i n } _ { \\rho _ { t } } \\mathrm { V a r } [ \\psi ( . ) ] = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 33 + }, + { + "bbox": [ + 107, + 572, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 107, + 573, + 174, + 586 + ], + "score": 0.91, + "content": "\\mathrm { V a r } [ \\psi ( . ) ] | _ { \\rho _ { t } = \\rho _ { \\infty } }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 572, + 303, + 588 + ], + "score": 1.0, + "content": "and mark this minimal value as", + "type": "text" + }, + { + "bbox": [ + 303, + 574, + 325, + 585 + ], + "score": 0.82, + "content": "C _ { \\mathrm { V a r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 572, + 506, + 588 + ], + "score": 1.0, + "content": ". In order to ensure that the adaptive learning", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 584, + 466, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 124, + 597 + ], + "score": 1.0, + "content": "rate", + "type": "text" + }, + { + "bbox": [ + 124, + 586, + 149, + 596 + ], + "score": 0.88, + "content": "( \\psi ( . ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 584, + 364, + 597 + ], + "score": 1.0, + "content": "has consistent variance, we rectify the variance at the", + "type": "text" + }, + { + "bbox": [ + 365, + 586, + 370, + 594 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 584, + 466, + 597 + ], + "score": 1.0, + "content": "-th timestamp as below,", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5, + "bbox_fs": [ + 103, + 521, + 507, + 597 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 152, + 597, + 457, + 612 + ], + "lines": [ + { + "bbox": [ + 152, + 597, + 457, + 612 + ], + "spans": [ + { + "bbox": [ + 152, + 597, + 457, + 612 + ], + "score": 0.35, + "content": "\\mathrm { V a r } [ r _ { t } \\psi ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } ) ] = C _ { \\mathrm { V a r } } \\quad \\mathrm { w h e r e } \\quad r _ { t } = \\sqrt { C _ { \\mathrm { V a r } } / \\mathrm { V a r } [ \\psi ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } ) ] } .", + "type": "interline_equation", + "image_path": "bfa5cdf77c87e44a8a93bc2010ec5df6421df1855b31d0f5551a951ec91161c1.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 152, + 597, + 457, + 612 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 613, + 504, + 648 + ], + "lines": [ + { + "bbox": [ + 105, + 612, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 271, + 626 + ], + "score": 1.0, + "content": "Although we have the analytic form of", + "type": "text" + }, + { + "bbox": [ + 271, + 613, + 311, + 625 + ], + "score": 0.91, + "content": "\\mathrm { V a r } [ \\psi ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 612, + 505, + 626 + ], + "score": 1.0, + "content": "(i.e., Equation 2), it is not numerically stable.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 622, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 637 + ], + "score": 1.0, + "content": "Therefore, we use the first-order approximation to calculate the rectification term. Specifically, by", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 635, + 331, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 167, + 649 + ], + "score": 1.0, + "content": "approximating", + "type": "text" + }, + { + "bbox": [ + 168, + 635, + 200, + 649 + ], + "score": 0.93, + "content": "\\sqrt { \\psi ^ { 2 } ( . ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 636, + 331, + 649 + ], + "score": 1.0, + "content": "to the first order (Wolter, 2007),", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 612, + 505, + 649 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 121, + 649, + 489, + 678 + ], + "lines": [ + { + "bbox": [ + 121, + 649, + 489, + 678 + ], + "spans": [ + { + "bbox": [ + 121, + 649, + 489, + 678 + ], + "score": 0.92, + "content": "\\sqrt { \\psi ^ { 2 } ( . ) } \\approx \\sqrt { \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] } + \\frac { 1 } { 2 \\sqrt { \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] } } ( \\psi ^ { 2 } ( . ) - \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] ) \\quad \\mathrm { a n d } \\quad \\mathrm { V a r } [ \\psi ( . ) ] \\approx \\frac { \\mathrm { V a r } [ \\psi ^ { 2 } ( . ) ] } { 4 \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] } .", + "type": "interline_equation", + "image_path": "2bcbe7d69530063b78d06b38316d33f1623fb2c1ec18a55d5780992c1a9a7bd1.jpg" + } + ] + } + ], + "index": 41, + "virtual_lines": [ + { + "bbox": [ + 121, + 649, + 489, + 658.6666666666666 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 121, + 658.6666666666666, + 489, + 668.3333333333333 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 121, + 668.3333333333333, + 489, + 677.9999999999999 + ], + "spans": [], + "index": 42 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 294, + 692 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 131, + 694 + ], + "score": 1.0, + "content": "Since", + "type": "text" + }, + { + "bbox": [ + 131, + 679, + 253, + 693 + ], + "score": 0.5, + "content": "\\psi ^ { 2 } ( . ) \\sim \\mathrm { S c a l e } { \\mathrm { - i n v } { - } } \\mathcal { X } ^ { 2 } ( \\rho _ { t } , \\frac { 1 } { \\sigma ^ { 2 } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 676, + 295, + 694 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 676, + 295, + 694 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 226, + 693, + 385, + 707 + ], + "lines": [ + { + "bbox": [ + 226, + 693, + 385, + 707 + ], + "spans": [ + { + "bbox": [ + 226, + 693, + 385, + 707 + ], + "score": 0.93, + "content": "\\mathrm { V a r } [ \\psi ( . ) ] \\approx \\rho _ { t } / [ 2 ( \\rho _ { t } - 2 ) ( \\rho _ { t } - 4 ) \\sigma ^ { 2 } ] .", + "type": "interline_equation", + "image_path": "7f29c5400faf9a111206eaa21ee57789c31d688b9384dae42e160d85e180cfb7.jpg" + } + ] + } + ], + "index": 44, + "virtual_lines": [ + { + "bbox": [ + 226, + 693, + 385, + 707 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 708, + 507, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 707, + 505, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 505, + 720 + ], + "score": 1.0, + "content": "In Section 5.3, we conduct simulation experiments to examine Equation 5 and find that it is a reliable", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 719, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 325, + 734 + ], + "score": 1.0, + "content": "approximation. Based on Equation 5, we know that", + "type": "text" + }, + { + "bbox": [ + 325, + 719, + 374, + 733 + ], + "score": 0.93, + "content": "\\mathrm { V a r } [ \\sqrt { \\psi ( . ) } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "decreases approximately at the", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 707, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 393, + 95, + 492, + 193 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 384, + 81, + 502, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 383, + 80, + 502, + 93 + ], + "spans": [ + { + "bbox": [ + 383, + 80, + 502, + 93 + ], + "score": 1.0, + "content": "Table 1: Image Classification", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 393, + 95, + 492, + 193 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 393, + 95, + 492, + 193 + ], + "spans": [ + { + "bbox": [ + 393, + 95, + 492, + 193 + ], + "score": 0.903, + "html": "
MethodAcc.
CITIIIOSGDAdamRAdam91.5190.5491.38
eeeseSGDAdamRAdam69.8666.5467.62
", + "type": "table", + "image_path": "2b06ae5fbfd40f478bad176ad4b8bc6635da54c9f4a2f28b5e5c18ea02d67a58.jpg" + } + ] + } + ], + "index": 4.0, + "virtual_lines": [ + { + "bbox": [ + 393, + 95, + 492, + 144.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 393, + 144.0, + 492, + 193.0 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 107, + 84, + 381, + 175 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 84, + 381, + 175 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 84, + 381, + 175 + ], + "spans": [ + { + "bbox": [ + 107, + 84, + 381, + 175 + ], + "score": 0.957, + "type": "image", + "image_path": "c8779d131e033f69524303bfd6e6b6c707693f95b1efbdf7a0772e68fcc51b91.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 107, + 84, + 381, + 114.33333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 114.33333333333333, + 381, + 144.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 107, + 144.66666666666666, + 381, + 175.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 111, + 177, + 376, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 176, + 376, + 190 + ], + "spans": [ + { + "bbox": [ + 112, + 176, + 376, + 190 + ], + "score": 1.0, + "content": "Figure 4: Language modeling (LSTMs) on the One Billion Word.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + } + ], + "index": 4.0 + }, + { + "type": "image", + "bbox": [ + 109, + 197, + 491, + 281 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 197, + 491, + 281 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 197, + 491, + 281 + ], + "spans": [ + { + "bbox": [ + 109, + 197, + 491, + 281 + ], + "score": 0.954, + "type": "image", + "image_path": "b8c26508ad736bb85b105e7794abce9ade7ded1d3b0ac44aa62399f7497eb5eb.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 109, + 197, + 491, + 225.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 109, + 225.0, + 491, + 253.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 109, + 253.0, + 491, + 281.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 122, + 285, + 486, + 297 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 122, + 284, + 488, + 298 + ], + "spans": [ + { + "bbox": [ + 122, + 284, + 488, + 298 + ], + "score": 1.0, + "content": "Figure 5: Training of ResNet-18 on the ImageNet and ResNet-20 on the CIFAR10 dataset.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + } + ], + "index": 9.0 + }, + { + "type": "text", + "bbox": [ + 107, + 307, + 442, + 318 + ], + "lines": [ + { + "bbox": [ + 105, + 305, + 444, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 142, + 321 + ], + "score": 1.0, + "content": "speed of", + "type": "text" + }, + { + "bbox": [ + 143, + 306, + 169, + 322 + ], + "score": 0.92, + "content": "O ( \\textstyle { \\frac { 1 } { \\rho _ { t } } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 305, + 444, + 321 + ], + "score": 1.0, + "content": ". With this approximation, we can calculate the rectification term as:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "interline_equation", + "bbox": [ + 243, + 320, + 367, + 353 + ], + "lines": [ + { + "bbox": [ + 243, + 320, + 367, + 353 + ], + "spans": [ + { + "bbox": [ + 243, + 320, + 367, + 353 + ], + "score": 0.95, + "content": "r _ { t } = \\sqrt { \\frac { ( \\rho _ { t } - 4 ) ( \\rho _ { t } - 2 ) \\rho _ { \\infty } } { ( \\rho _ { \\infty } - 4 ) ( \\rho _ { \\infty } - 2 ) \\rho _ { t } } } .", + "type": "interline_equation", + "image_path": "8719717e482e66abfab474babc6767344179cb7f72e9817c59b381c68cbf46f8.jpg" + } + ] + } + ], + "index": 12.5, + "virtual_lines": [ + { + "bbox": [ + 243, + 320, + 367, + 336.5 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 243, + 336.5, + 367, + 353.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 351, + 505, + 417 + ], + "lines": [ + { + "bbox": [ + 106, + 351, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 506, + 363 + ], + "score": 1.0, + "content": "Applying our rectification term to Adam, we come up with a new variant of Adam, Rectified Adam", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "(RAdam), as summarized in Algorithm 2. Specifically, when the length of the approximated SMA is", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 371, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 506, + 387 + ], + "score": 1.0, + "content": "less or equal than 4, the variance of the adaptive learning rate is intractable and the adaptive learning", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "rate is inactivated. Otherwise, we calculate the variance rectification term and update parameters", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 366, + 408 + ], + "score": 1.0, + "content": "with the adaptive learning rate. It is worth mentioning that, if", + "type": "text" + }, + { + "bbox": [ + 366, + 395, + 407, + 407 + ], + "score": 0.91, + "content": "\\beta _ { 2 } ~ \\leq ~ 0 . 6", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 394, + 448, + 408 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + }, + { + "bbox": [ + 449, + 395, + 486, + 407 + ], + "score": 0.9, + "content": "\\rho _ { \\infty } \\leq 4", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 406, + 302, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 302, + 417 + ], + "score": 1.0, + "content": "RAdam is degenerated to SGD with momentum.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 109, + 419, + 451, + 430 + ], + "lines": [ + { + "bbox": [ + 106, + 418, + 452, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 452, + 432 + ], + "score": 1.0, + "content": "4.3 IN COMPARISON WITH WARMUP AND OTHER STABILIZATION TECHNIQUES", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 439, + 505, + 521 + ], + "lines": [ + { + "bbox": [ + 105, + 438, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 505, + 453 + ], + "score": 1.0, + "content": "Different from the analysis in this paper, warmup is originally proposed to handle training with very", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 450, + 504, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 504, + 462 + ], + "score": 1.0, + "content": "large batches for SGD (Goyal et al., 2017; Gotmare et al., 2019; Bernstein et al., 2018; Xiao et al.,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 460, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 196, + 474 + ], + "score": 1.0, + "content": "2017). We notice that", + "type": "text" + }, + { + "bbox": [ + 196, + 463, + 205, + 473 + ], + "score": 0.85, + "content": "r _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 460, + 505, + 474 + ], + "score": 1.0, + "content": "has a similar form to the heuristic linear warmup, which can be viewed as", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 103, + 469, + 507, + 490 + ], + "spans": [ + { + "bbox": [ + 103, + 469, + 233, + 490 + ], + "score": 1.0, + "content": "setting the rectification term as", + "type": "text" + }, + { + "bbox": [ + 234, + 473, + 273, + 489 + ], + "score": 0.93, + "content": "\\frac { m i n ( t , T _ { w } ) } { T _ { w } }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 469, + 507, + 490 + ], + "score": 1.0, + "content": ". It verifies our intuition that warmup works as a variance", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 485, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 104, + 485, + 505, + 500 + ], + "score": 1.0, + "content": "reduction technique. RAdam deactivates the adaptive learning rate when its variance is divergent,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "thus avoiding undesired instability in the first few updates. Besides, our method does not require an", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 509, + 504, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 234, + 522 + ], + "score": 1.0, + "content": "additional hyperparameter (i.e.,", + "type": "text" + }, + { + "bbox": [ + 234, + 509, + 248, + 520 + ], + "score": 0.73, + "content": "T _ { w . }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 509, + 504, + 522 + ], + "score": 1.0, + "content": ") and can automatically adapt to different moving average rules.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 505, + 603 + ], + "lines": [ + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "Here, we identify and address an underlying issue of adaptive optimization methods independent", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "of (neural) model architectures. Thus, the proposed rectification term is orthogonal to other train-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 548, + 504, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 504, + 560 + ], + "score": 1.0, + "content": "ing stabilization techniques such as gradient clipping (Bengio et al., 2013), smoothing the adaptive", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 223, + 572 + ], + "score": 1.0, + "content": "learning rate (i.e., increasing", + "type": "text" + }, + { + "bbox": [ + 223, + 561, + 228, + 569 + ], + "score": 0.3, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 559, + 505, + 572 + ], + "score": 1.0, + "content": ", applying geometric mean filter (Chen & Gu, 2018), or adding range", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "constraints (Luo et al., 2019)), initialization (Balduzzi et al., 2017; Zhang et al., 2019) and normal-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 581, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 506, + 594 + ], + "score": 1.0, + "content": "ization (Ba et al., 2016; Ioffe & Szegedy, 2015). Indeed, these techniques can be combined with the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 592, + 267, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 267, + 604 + ], + "score": 1.0, + "content": "proposed variance rectification method.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31 + }, + { + "type": "title", + "bbox": [ + 108, + 615, + 200, + 628 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 201, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 201, + 630 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "score": 1.0, + "content": "We evaluate RAdam on several benchmarks: One Billion Word for language modeling; Cifar10", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "and ImageNet for image classification; IWSLT’14 De-En/EN-DE and WMT’16 EN-De for neural", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "machine translation. Following Loshchilov & Hutter (2018), we decouple weight decays in the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 671, + 487, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 487, + 682 + ], + "score": 1.0, + "content": "vanilla Adam, Adam with warmup and RAdam in our experiments. Details are in Appendix B.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + }, + { + "type": "title", + "bbox": [ + 108, + 692, + 270, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 691, + 272, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 272, + 705 + ], + "score": 1.0, + "content": "5.1 COMPARING TO VANILLA ADAM", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "As analyzed before, the adaptive learning rate has undesirably large variance in the early stage", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "of training and leads to suspicious/bad local optima on NMT. One question we are interested in", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 393, + 95, + 492, + 193 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 384, + 81, + 502, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 383, + 80, + 502, + 93 + ], + "spans": [ + { + "bbox": [ + 383, + 80, + 502, + 93 + ], + "score": 1.0, + "content": "Table 1: Image Classification", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 393, + 95, + 492, + 193 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 393, + 95, + 492, + 193 + ], + "spans": [ + { + "bbox": [ + 393, + 95, + 492, + 193 + ], + "score": 0.903, + "html": "
MethodAcc.
CITIIIOSGDAdamRAdam91.5190.5491.38
eeeseSGDAdamRAdam69.8666.5467.62
", + "type": "table", + "image_path": "2b06ae5fbfd40f478bad176ad4b8bc6635da54c9f4a2f28b5e5c18ea02d67a58.jpg" + } + ] + } + ], + "index": 4.0, + "virtual_lines": [ + { + "bbox": [ + 393, + 95, + 492, + 144.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 393, + 144.0, + 492, + 193.0 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 107, + 84, + 381, + 175 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 84, + 381, + 175 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 84, + 381, + 175 + ], + "spans": [ + { + "bbox": [ + 107, + 84, + 381, + 175 + ], + "score": 0.957, + "type": "image", + "image_path": "c8779d131e033f69524303bfd6e6b6c707693f95b1efbdf7a0772e68fcc51b91.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 107, + 84, + 381, + 114.33333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 114.33333333333333, + 381, + 144.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 107, + 144.66666666666666, + 381, + 175.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 111, + 177, + 376, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 176, + 376, + 190 + ], + "spans": [ + { + "bbox": [ + 112, + 176, + 376, + 190 + ], + "score": 1.0, + "content": "Figure 4: Language modeling (LSTMs) on the One Billion Word.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + } + ], + "index": 4.0 + }, + { + "type": "image", + "bbox": [ + 109, + 197, + 491, + 281 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 197, + 491, + 281 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 197, + 491, + 281 + ], + "spans": [ + { + "bbox": [ + 109, + 197, + 491, + 281 + ], + "score": 0.954, + "type": "image", + "image_path": "b8c26508ad736bb85b105e7794abce9ade7ded1d3b0ac44aa62399f7497eb5eb.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 109, + 197, + 491, + 225.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 109, + 225.0, + 491, + 253.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 109, + 253.0, + 491, + 281.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 122, + 285, + 486, + 297 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 122, + 284, + 488, + 298 + ], + "spans": [ + { + "bbox": [ + 122, + 284, + 488, + 298 + ], + "score": 1.0, + "content": "Figure 5: Training of ResNet-18 on the ImageNet and ResNet-20 on the CIFAR10 dataset.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + } + ], + "index": 9.0 + }, + { + "type": "text", + "bbox": [ + 107, + 307, + 442, + 318 + ], + "lines": [ + { + "bbox": [ + 105, + 305, + 444, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 142, + 321 + ], + "score": 1.0, + "content": "speed of", + "type": "text" + }, + { + "bbox": [ + 143, + 306, + 169, + 322 + ], + "score": 0.92, + "content": "O ( \\textstyle { \\frac { 1 } { \\rho _ { t } } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 305, + 444, + 321 + ], + "score": 1.0, + "content": ". With this approximation, we can calculate the rectification term as:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 305, + 444, + 322 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 243, + 320, + 367, + 353 + ], + "lines": [ + { + "bbox": [ + 243, + 320, + 367, + 353 + ], + "spans": [ + { + "bbox": [ + 243, + 320, + 367, + 353 + ], + "score": 0.95, + "content": "r _ { t } = \\sqrt { \\frac { ( \\rho _ { t } - 4 ) ( \\rho _ { t } - 2 ) \\rho _ { \\infty } } { ( \\rho _ { \\infty } - 4 ) ( \\rho _ { \\infty } - 2 ) \\rho _ { t } } } .", + "type": "interline_equation", + "image_path": "8719717e482e66abfab474babc6767344179cb7f72e9817c59b381c68cbf46f8.jpg" + } + ] + } + ], + "index": 12.5, + "virtual_lines": [ + { + "bbox": [ + 243, + 320, + 367, + 336.5 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 243, + 336.5, + 367, + 353.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 351, + 505, + 417 + ], + "lines": [ + { + "bbox": [ + 106, + 351, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 506, + 363 + ], + "score": 1.0, + "content": "Applying our rectification term to Adam, we come up with a new variant of Adam, Rectified Adam", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "(RAdam), as summarized in Algorithm 2. Specifically, when the length of the approximated SMA is", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 371, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 506, + 387 + ], + "score": 1.0, + "content": "less or equal than 4, the variance of the adaptive learning rate is intractable and the adaptive learning", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "rate is inactivated. Otherwise, we calculate the variance rectification term and update parameters", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 366, + 408 + ], + "score": 1.0, + "content": "with the adaptive learning rate. It is worth mentioning that, if", + "type": "text" + }, + { + "bbox": [ + 366, + 395, + 407, + 407 + ], + "score": 0.91, + "content": "\\beta _ { 2 } ~ \\leq ~ 0 . 6", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 394, + 448, + 408 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + }, + { + "bbox": [ + 449, + 395, + 486, + 407 + ], + "score": 0.9, + "content": "\\rho _ { \\infty } \\leq 4", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 406, + 302, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 302, + 417 + ], + "score": 1.0, + "content": "RAdam is degenerated to SGD with momentum.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 351, + 506, + 417 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 419, + 451, + 430 + ], + "lines": [ + { + "bbox": [ + 106, + 418, + 452, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 452, + 432 + ], + "score": 1.0, + "content": "4.3 IN COMPARISON WITH WARMUP AND OTHER STABILIZATION TECHNIQUES", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 439, + 505, + 521 + ], + "lines": [ + { + "bbox": [ + 105, + 438, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 505, + 453 + ], + "score": 1.0, + "content": "Different from the analysis in this paper, warmup is originally proposed to handle training with very", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 450, + 504, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 504, + 462 + ], + "score": 1.0, + "content": "large batches for SGD (Goyal et al., 2017; Gotmare et al., 2019; Bernstein et al., 2018; Xiao et al.,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 460, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 196, + 474 + ], + "score": 1.0, + "content": "2017). We notice that", + "type": "text" + }, + { + "bbox": [ + 196, + 463, + 205, + 473 + ], + "score": 0.85, + "content": "r _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 460, + 505, + 474 + ], + "score": 1.0, + "content": "has a similar form to the heuristic linear warmup, which can be viewed as", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 103, + 469, + 507, + 490 + ], + "spans": [ + { + "bbox": [ + 103, + 469, + 233, + 490 + ], + "score": 1.0, + "content": "setting the rectification term as", + "type": "text" + }, + { + "bbox": [ + 234, + 473, + 273, + 489 + ], + "score": 0.93, + "content": "\\frac { m i n ( t , T _ { w } ) } { T _ { w } }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 469, + 507, + 490 + ], + "score": 1.0, + "content": ". It verifies our intuition that warmup works as a variance", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 485, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 104, + 485, + 505, + 500 + ], + "score": 1.0, + "content": "reduction technique. RAdam deactivates the adaptive learning rate when its variance is divergent,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "thus avoiding undesired instability in the first few updates. Besides, our method does not require an", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 509, + 504, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 234, + 522 + ], + "score": 1.0, + "content": "additional hyperparameter (i.e.,", + "type": "text" + }, + { + "bbox": [ + 234, + 509, + 248, + 520 + ], + "score": 0.73, + "content": "T _ { w . }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 509, + 504, + 522 + ], + "score": 1.0, + "content": ") and can automatically adapt to different moving average rules.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24, + "bbox_fs": [ + 103, + 438, + 507, + 522 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 505, + 603 + ], + "lines": [ + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "Here, we identify and address an underlying issue of adaptive optimization methods independent", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "of (neural) model architectures. Thus, the proposed rectification term is orthogonal to other train-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 548, + 504, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 504, + 560 + ], + "score": 1.0, + "content": "ing stabilization techniques such as gradient clipping (Bengio et al., 2013), smoothing the adaptive", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 223, + 572 + ], + "score": 1.0, + "content": "learning rate (i.e., increasing", + "type": "text" + }, + { + "bbox": [ + 223, + 561, + 228, + 569 + ], + "score": 0.3, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 559, + 505, + 572 + ], + "score": 1.0, + "content": ", applying geometric mean filter (Chen & Gu, 2018), or adding range", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "constraints (Luo et al., 2019)), initialization (Balduzzi et al., 2017; Zhang et al., 2019) and normal-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 581, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 506, + 594 + ], + "score": 1.0, + "content": "ization (Ba et al., 2016; Ioffe & Szegedy, 2015). Indeed, these techniques can be combined with the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 592, + 267, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 267, + 604 + ], + "score": 1.0, + "content": "proposed variance rectification method.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 525, + 506, + 604 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 615, + 200, + 628 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 201, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 201, + 630 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "score": 1.0, + "content": "We evaluate RAdam on several benchmarks: One Billion Word for language modeling; Cifar10", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "and ImageNet for image classification; IWSLT’14 De-En/EN-DE and WMT’16 EN-De for neural", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "machine translation. Following Loshchilov & Hutter (2018), we decouple weight decays in the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 671, + 487, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 487, + 682 + ], + "score": 1.0, + "content": "vanilla Adam, Adam with warmup and RAdam in our experiments. Details are in Appendix B.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5, + "bbox_fs": [ + 106, + 637, + 505, + 682 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 692, + 270, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 691, + 272, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 272, + 705 + ], + "score": 1.0, + "content": "5.1 COMPARING TO VANILLA ADAM", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "As analyzed before, the adaptive learning rate has undesirably large variance in the early stage", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "of training and leads to suspicious/bad local optima on NMT. One question we are interested in", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "is: whether such an issue widely exits in other similar tasks and applications. Thus, we conduct", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 403, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 417 + ], + "score": 1.0, + "content": "a set of experiments with two classical tasks of NLP and CV, i.e., language modeling and image", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 415, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 506, + 428 + ], + "score": 1.0, + "content": "classification. RAdam not only results in consistent improvements over the vanilla Adam, but also", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 427, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 437 + ], + "score": 1.0, + "content": "demonstrates its robustness to the change of learning rates. It verifies that the variance issue exists", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 437, + 450, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 450, + 449 + ], + "score": 1.0, + "content": "in various machine learning applications, and has a big impact on the model behavior.", + "type": "text", + "cross_page": true + } + ], + "index": 12 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 708, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 119, + 87, + 491, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 87, + 491, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 119, + 87, + 491, + 225 + ], + "spans": [ + { + "bbox": [ + 119, + 87, + 491, + 225 + ], + "score": 0.971, + "type": "image", + "image_path": "6345f7afa7e7a63d9229d9a8dfe723d4ffb9cfb0f3992b2f052dd4d31c0db1fa.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 119, + 87, + 491, + 133.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 119, + 133.0, + 491, + 179.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 179.0, + 491, + 225.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 119, + 226, + 492, + 237 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 224, + 493, + 239 + ], + "spans": [ + { + "bbox": [ + 117, + 224, + 493, + 239 + ], + "score": 1.0, + "content": "Figure 6: Performance of RAdam, Adam and SGD with different learning rates on CIFAR10.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 107, + 246, + 505, + 372 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 246, + 505, + 372 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 107, + 246, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 107, + 246, + 505, + 372 + ], + "score": 0.957, + "type": "image", + "image_path": "bb74ac762ae180f94b6c6c55e5c1b8414301b44371f1f8fbfd2e1825dfee7518.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 107, + 246, + 505, + 288.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 107, + 288.0, + 505, + 330.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 330.0, + 505, + 372.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 108, + 374, + 500, + 384 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 371, + 501, + 386 + ], + "spans": [ + { + "bbox": [ + 109, + 371, + 501, + 386 + ], + "score": 1.0, + "content": "Figure 7: Performance of RAdam, Adam with warmup on CIFAR10 with different learning rates.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 107, + 393, + 505, + 448 + ], + "lines": [ + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "is: whether such an issue widely exits in other similar tasks and applications. Thus, we conduct", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 403, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 417 + ], + "score": 1.0, + "content": "a set of experiments with two classical tasks of NLP and CV, i.e., language modeling and image", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 415, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 506, + 428 + ], + "score": 1.0, + "content": "classification. RAdam not only results in consistent improvements over the vanilla Adam, but also", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 427, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 437 + ], + "score": 1.0, + "content": "demonstrates its robustness to the change of learning rates. It verifies that the variance issue exists", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 437, + 450, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 450, + 449 + ], + "score": 1.0, + "content": "in various machine learning applications, and has a big impact on the model behavior.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 454, + 505, + 575 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "Performance Comparison. The performances on language modeling (i.e., One Billion", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "Word (Chelba et al., 2013)) and image classification (i.e., CIFAR10 (Krizhevsky et al., 2009) and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 476, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 488 + ], + "score": 1.0, + "content": "ImageNet (Deng et al., 2009)) are presented in Figure 4, 5. The results show that RAdam out-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 486, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 104, + 486, + 506, + 499 + ], + "score": 1.0, + "content": "performs Adam in all three datasets. As shown in Figure 4, although the rectification term makes", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 498, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 506, + 510 + ], + "score": 1.0, + "content": "RAdam slower than the vanilla Adam in the first few epochs, it allows RAdam to converge faster", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 509, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 506, + 522 + ], + "score": 1.0, + "content": "after that. In other words, by reducing the variance of the adaptive learning rate in the early stage, it", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 519, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 533 + ], + "score": 1.0, + "content": "gets both faster convergence and better performance, which verifies the impact of the variance issue.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 531, + 504, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 504, + 542 + ], + "score": 1.0, + "content": "We also observe that RAdam obtains consistent improvements over Adam on image classification.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 542, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 554 + ], + "score": 1.0, + "content": "It is worth noting that, on both ImageNet and CIFAR10, although RAdam fails to outperform SGD", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 552, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 506, + 565 + ], + "score": 1.0, + "content": "in terms of test accuracy, it results in a better training performance (e.g., the training accuracy of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 563, + 431, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 431, + 576 + ], + "score": 1.0, + "content": "SGD, Adam, and RAdam on ImageNet are 69.57, 69.12 and 70.30 respectively).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 580, + 505, + 668 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "score": 1.0, + "content": "Robustness to Learning Rate Change. Besides performance improvements, RAdam also improves", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "the robustness of model training. We use different initial learning rates, conduct experiments with", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 602, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 505, + 615 + ], + "score": 1.0, + "content": "ResNet-20 on the CIFAR10 datasets, and summarize their performance in Figure 6. For learning", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 233, + 626 + ], + "score": 1.0, + "content": "rates within a broad range (i.e.,", + "type": "text" + }, + { + "bbox": [ + 233, + 613, + 329, + 626 + ], + "score": 0.86, + "content": "\\{ 0 . 1 , 0 . 0 3 , 0 . 0 1 , 0 . 0 0 3 \\} ,", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 613, + 505, + 626 + ], + "score": 1.0, + "content": "), RAdam achieves consistent model perfor-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "score": 1.0, + "content": "mances (their test accuracy curves highly overlap with each other), while Adam and SGD are shown", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 634, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 506, + 648 + ], + "score": 1.0, + "content": "to be more sensitive to the learning rate. The observation can be interpreted that by rectifying the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 647, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 647, + 505, + 659 + ], + "score": 1.0, + "content": "variance of the adaptive learning rate, RAdam improves the robustness of model training and can", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 657, + 311, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 311, + 671 + ], + "score": 1.0, + "content": "adapt to different learning rates of a broader range.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27.5 + }, + { + "type": "title", + "bbox": [ + 108, + 680, + 290, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 679, + 292, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 292, + 693 + ], + "score": 1.0, + "content": "5.2 COMPARING TO HEURISTIC WARMUP", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "To examine the effectiveness of RAdam, we first conduct comparisons on neural machine transla-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "tion, on which the state-of-the-art employs Adam with the linear warmup. Specifically, we conduct", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "experiments on three datasets, i.e., IWSLT’14 De-En, IWSLT’14 En-De, and WMT’16 En-De. Due", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 119, + 87, + 491, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 87, + 491, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 119, + 87, + 491, + 225 + ], + "spans": [ + { + "bbox": [ + 119, + 87, + 491, + 225 + ], + "score": 0.971, + "type": "image", + "image_path": "6345f7afa7e7a63d9229d9a8dfe723d4ffb9cfb0f3992b2f052dd4d31c0db1fa.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 119, + 87, + 491, + 133.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 119, + 133.0, + 491, + 179.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 179.0, + 491, + 225.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 119, + 226, + 492, + 237 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 224, + 493, + 239 + ], + "spans": [ + { + "bbox": [ + 117, + 224, + 493, + 239 + ], + "score": 1.0, + "content": "Figure 6: Performance of RAdam, Adam and SGD with different learning rates on CIFAR10.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 107, + 246, + 505, + 372 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 246, + 505, + 372 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 107, + 246, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 107, + 246, + 505, + 372 + ], + "score": 0.957, + "type": "image", + "image_path": "bb74ac762ae180f94b6c6c55e5c1b8414301b44371f1f8fbfd2e1825dfee7518.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 107, + 246, + 505, + 288.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 107, + 288.0, + 505, + 330.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 330.0, + 505, + 372.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 108, + 374, + 500, + 384 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 371, + 501, + 386 + ], + "spans": [ + { + "bbox": [ + 109, + 371, + 501, + 386 + ], + "score": 1.0, + "content": "Figure 7: Performance of RAdam, Adam with warmup on CIFAR10 with different learning rates.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 107, + 393, + 505, + 448 + ], + "lines": [], + "index": 10, + "bbox_fs": [ + 105, + 393, + 506, + 449 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 454, + 505, + 575 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "Performance Comparison. The performances on language modeling (i.e., One Billion", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "Word (Chelba et al., 2013)) and image classification (i.e., CIFAR10 (Krizhevsky et al., 2009) and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 476, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 488 + ], + "score": 1.0, + "content": "ImageNet (Deng et al., 2009)) are presented in Figure 4, 5. The results show that RAdam out-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 486, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 104, + 486, + 506, + 499 + ], + "score": 1.0, + "content": "performs Adam in all three datasets. As shown in Figure 4, although the rectification term makes", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 498, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 506, + 510 + ], + "score": 1.0, + "content": "RAdam slower than the vanilla Adam in the first few epochs, it allows RAdam to converge faster", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 509, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 506, + 522 + ], + "score": 1.0, + "content": "after that. In other words, by reducing the variance of the adaptive learning rate in the early stage, it", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 519, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 533 + ], + "score": 1.0, + "content": "gets both faster convergence and better performance, which verifies the impact of the variance issue.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 531, + 504, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 504, + 542 + ], + "score": 1.0, + "content": "We also observe that RAdam obtains consistent improvements over Adam on image classification.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 542, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 554 + ], + "score": 1.0, + "content": "It is worth noting that, on both ImageNet and CIFAR10, although RAdam fails to outperform SGD", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 552, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 506, + 565 + ], + "score": 1.0, + "content": "in terms of test accuracy, it results in a better training performance (e.g., the training accuracy of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 563, + 431, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 431, + 576 + ], + "score": 1.0, + "content": "SGD, Adam, and RAdam on ImageNet are 69.57, 69.12 and 70.30 respectively).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18, + "bbox_fs": [ + 104, + 453, + 506, + 576 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 580, + 505, + 668 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "score": 1.0, + "content": "Robustness to Learning Rate Change. Besides performance improvements, RAdam also improves", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "the robustness of model training. We use different initial learning rates, conduct experiments with", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 602, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 505, + 615 + ], + "score": 1.0, + "content": "ResNet-20 on the CIFAR10 datasets, and summarize their performance in Figure 6. For learning", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 233, + 626 + ], + "score": 1.0, + "content": "rates within a broad range (i.e.,", + "type": "text" + }, + { + "bbox": [ + 233, + 613, + 329, + 626 + ], + "score": 0.86, + "content": "\\{ 0 . 1 , 0 . 0 3 , 0 . 0 1 , 0 . 0 0 3 \\} ,", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 613, + 505, + 626 + ], + "score": 1.0, + "content": "), RAdam achieves consistent model perfor-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "score": 1.0, + "content": "mances (their test accuracy curves highly overlap with each other), while Adam and SGD are shown", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 634, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 506, + 648 + ], + "score": 1.0, + "content": "to be more sensitive to the learning rate. The observation can be interpreted that by rectifying the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 647, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 647, + 505, + 659 + ], + "score": 1.0, + "content": "variance of the adaptive learning rate, RAdam improves the robustness of model training and can", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 657, + 311, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 311, + 671 + ], + "score": 1.0, + "content": "adapt to different learning rates of a broader range.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 579, + 506, + 671 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 680, + 290, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 679, + 292, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 292, + 693 + ], + "score": 1.0, + "content": "5.2 COMPARING TO HEURISTIC WARMUP", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "To examine the effectiveness of RAdam, we first conduct comparisons on neural machine transla-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "tion, on which the state-of-the-art employs Adam with the linear warmup. Specifically, we conduct", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "experiments on three datasets, i.e., IWSLT’14 De-En, IWSLT’14 En-De, and WMT’16 En-De. Due", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 129, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 505, + 140 + ], + "score": 1.0, + "content": "to the limited size of the IWSLT’14 dataset, we conduct experiments using 5 different random seeds", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 140, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 505, + 151 + ], + "score": 1.0, + "content": "and report their mean and standard derivation. As discussed before, the vanilla Adam algorithm", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 150, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 505, + 164 + ], + "score": 1.0, + "content": "leads to suspicious/bad local optima (i.e., converges to a training perplexity around 500), and needs", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 161, + 321, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 321, + 175 + ], + "score": 1.0, + "content": "a learning rate warmup stage to stabilize the training.", + "type": "text", + "cross_page": true + } + ], + "index": 7 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 698, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 131, + 88, + 477, + 124 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 197, + 74, + 414, + 86 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 196, + 73, + 414, + 87 + ], + "spans": [ + { + "bbox": [ + 196, + 73, + 414, + 87 + ], + "score": 1.0, + "content": "Table 2: BLEU score on Neural Machine Translation.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 131, + 88, + 477, + 124 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 131, + 88, + 477, + 124 + ], + "spans": [ + { + "bbox": [ + 131, + 88, + 477, + 124 + ], + "score": 0.967, + "html": "
MethodIWSLT'14 DE-ENIWSLT'14 EN-DEWMT'16 EN-DE
Adam with warmupRAdam34.66± 0.01434.76 ± 0.00328.56± 0.06728.48 ± 0.05427.0327.27
", + "type": "table", + "image_path": "a8fe3dfdbb40c1eb36ceeadac1a0d9d2b8896d7e820a4e3e7cf0047e895f9ed8.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 131, + 88, + 477, + 100.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 131, + 100.0, + 477, + 112.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 131, + 112.0, + 477, + 124.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 108, + 129, + 505, + 173 + ], + "lines": [ + { + "bbox": [ + 106, + 129, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 505, + 140 + ], + "score": 1.0, + "content": "to the limited size of the IWSLT’14 dataset, we conduct experiments using 5 different random seeds", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 140, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 505, + 151 + ], + "score": 1.0, + "content": "and report their mean and standard derivation. As discussed before, the vanilla Adam algorithm", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 150, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 505, + 164 + ], + "score": 1.0, + "content": "leads to suspicious/bad local optima (i.e., converges to a training perplexity around 500), and needs", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 161, + 321, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 321, + 175 + ], + "score": 1.0, + "content": "a learning rate warmup stage to stabilize the training.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 178, + 505, + 234 + ], + "lines": [ + { + "bbox": [ + 106, + 178, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 505, + 190 + ], + "score": 1.0, + "content": "We summarize the performance obtained with the heuristic warmup and our proposed rectification", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 189, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 505, + 202 + ], + "score": 1.0, + "content": "term in Table 2 and visualize the training curve of IWSLT De-En in Figure 1. With a consistent", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "adaptive learning rate variance, our proposed method achieves similar performance to that of previ-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 211, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 505, + 224 + ], + "score": 1.0, + "content": "ous state-of-the-art warmup heuristics. It verifies our intuition that the problematic updates of Adam", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 221, + 387, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 387, + 236 + ], + "score": 1.0, + "content": "are indeed caused by the undesirably large variance in the early stage.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 506, + 252 + ], + "score": 1.0, + "content": "Moreover, we applied Adam with warmup on the CIFAR10 dataset. Its best accuracy on the test", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 250, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 263 + ], + "score": 1.0, + "content": "set is 91.29, which is similar to RAdam (91.38). However, we found that RAdam requires less hy-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 261, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 506, + 275 + ], + "score": 1.0, + "content": "perparameter tuning. Specifically, we visualize their learning curves in Figure 7. For some warmup", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 272, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 505, + 285 + ], + "score": 1.0, + "content": "steps, Adam with warmup is relatively more sensitive to the choice of the learning rate. RAdam,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 283, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 505, + 296 + ], + "score": 1.0, + "content": "at the same time, is not only more robust, but also can automatically control the warmup behav-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "ior (i.e., without requiring the length of warmup). For example, when setting the learning rate as", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 305, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 305, + 505, + 318 + ], + "score": 1.0, + "content": "0.1, Adam with 100 steps of warmup fails to get satisfying performance and only results in an ac-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 316, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 506, + 329 + ], + "score": 1.0, + "content": "curacy of 90.13; RAdam successfully gets an accuracy of 91.06, with the original setting of the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 327, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 239, + 339 + ], + "score": 1.0, + "content": "moving average calculation (i.e.,", + "type": "text" + }, + { + "bbox": [ + 239, + 327, + 329, + 339 + ], + "score": 0.89, + "content": "\\beta _ { 1 } = 0 . 9 , \\beta _ { 2 } = 0 . 9 9 9 )", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 327, + 505, + 339 + ], + "score": 1.0, + "content": ". We conjecture the reason is due to the fact", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "score": 1.0, + "content": "that RAdam, which is based on a rigorous variance analysis, explicitly avoids the extreme situation", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 349, + 477, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 477, + 361 + ], + "score": 1.0, + "content": "where the variance is divergent, and rectifies the variance to be consistent in other situations.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 108, + 374, + 247, + 385 + ], + "lines": [ + { + "bbox": [ + 106, + 373, + 248, + 386 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 248, + 386 + ], + "score": 1.0, + "content": "5.3 SIMULATED VERIFICATION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 393, + 505, + 442 + ], + "lines": [ + { + "bbox": [ + 104, + 393, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 104, + 394, + 263, + 414 + ], + "score": 1.0, + "content": "In Sections 3 and 4, we approximate", + "type": "text" + }, + { + "bbox": [ + 263, + 393, + 340, + 413 + ], + "score": 0.93, + "content": "\\textstyle \\operatorname { V a r } [ \\sqrt { t / \\sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 394, + 468, + 414 + ], + "score": 1.0, + "content": "to the first order, and assume", + "type": "text" + }, + { + "bbox": [ + 468, + 397, + 505, + 410 + ], + "score": 0.89, + "content": "\\psi ^ { 2 } ( . ) = ", + "type": "inline_equation" + } + ], + "index": 25 + }, + { + "bbox": [ + 108, + 410, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 108, + 411, + 182, + 432 + ], + "score": 0.88, + "content": "\\frac { 1 - \\beta _ { 2 } ^ { t } } { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 410, + 505, + 435 + ], + "score": 1.0, + "content": "subjects to a scaled inverse chi-square distribution (this assumption covers the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "score": 1.0, + "content": "approximation from EMA to SMA). Here, we examine these two approximations using simulations.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 447, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 102, + 446, + 509, + 470 + ], + "spans": [ + { + "bbox": [ + 102, + 446, + 240, + 470 + ], + "score": 1.0, + "content": "First Order Approximation of", + "type": "text" + }, + { + "bbox": [ + 241, + 447, + 318, + 467 + ], + "score": 0.93, + "content": "\\textstyle \\operatorname { V a r } [ \\sqrt { t / \\sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 446, + 509, + 470 + ], + "score": 1.0, + "content": ". To compare Equations 5 and 2, we assume", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 465, + 504, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 135, + 476 + ], + "score": 0.87, + "content": "\\tau = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 465, + 301, + 478 + ], + "score": 1.0, + "content": "and plot their values and difference for", + "type": "text" + }, + { + "bbox": [ + 301, + 465, + 378, + 478 + ], + "score": 0.92, + "content": "\\nu = \\{ 5 , \\cdots , 5 0 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 465, + 504, + 478 + ], + "score": 1.0, + "content": "in Figure 8. The curve of the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 477, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 506, + 489 + ], + "score": 1.0, + "content": "analytic form and the first-order approximation highly overlap, and their difference is much smaller", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 487, + 453, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 453, + 500 + ], + "score": 1.0, + "content": "than their value. This result verifies that our first-order approximation is very accurate.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 504, + 505, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 504, + 507, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 442, + 518 + ], + "score": 1.0, + "content": "Scaled Inverse Chi-Square Distribution Assumption. In this paper, we assume", + "type": "text" + }, + { + "bbox": [ + 442, + 507, + 451, + 516 + ], + "score": 0.83, + "content": "g _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 504, + 507, + 518 + ], + "score": 1.0, + "content": "accords to a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 515, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 324, + 529 + ], + "score": 1.0, + "content": "Normal distribution with a zero mean. We also assume", + "type": "text" + }, + { + "bbox": [ + 324, + 515, + 346, + 528 + ], + "score": 0.91, + "content": "\\psi ^ { 2 } ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 515, + 506, + 529 + ], + "score": 1.0, + "content": "accords to the scaled inverse chi-square", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 260, + 540 + ], + "score": 1.0, + "content": "distribution to derive the variance of", + "type": "text" + }, + { + "bbox": [ + 261, + 527, + 300, + 539 + ], + "score": 0.91, + "content": "\\mathrm { V a r } [ \\psi ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 527, + 506, + 540 + ], + "score": 1.0, + "content": ", based on the similarity between the exponential", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 538, + 467, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 467, + 551 + ], + "score": 1.0, + "content": "moving average and simple moving average. Here, we empirically verify this assumption.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 554, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 180, + 567 + ], + "score": 1.0, + "content": "Specifically, since", + "type": "text" + }, + { + "bbox": [ + 181, + 556, + 190, + 566 + ], + "score": 0.85, + "content": "g _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "in the optimization problem may not be zero-mean, we assume its expectation", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 116, + 578 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 116, + 567, + 124, + 577 + ], + "score": 0.79, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 565, + 173, + 578 + ], + "score": 1.0, + "content": "and sample", + "type": "text" + }, + { + "bbox": [ + 174, + 567, + 183, + 577 + ], + "score": 0.85, + "content": "g _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 565, + 207, + 578 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 208, + 565, + 241, + 578 + ], + "score": 0.93, + "content": "\\mathcal { N } ( \\bar { \\boldsymbol { \\mu } } , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 565, + 505, + 578 + ], + "score": 1.0, + "content": ". Then, based on these samples, we calculate the variance of the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 576, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 453, + 591 + ], + "score": 1.0, + "content": "original adaptive learning rate and the proposed rectified adaptive learning rate, i.e.,", + "type": "text" + }, + { + "bbox": [ + 454, + 576, + 486, + 591 + ], + "score": 0.91, + "content": "\\textstyle \\operatorname { V a r } [ { \\frac { 1 } { \\widehat { v _ { t } } } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 576, + 506, + 591 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 588, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 138, + 603 + ], + "score": 0.9, + "content": "\\textstyle \\operatorname { V a r } [ { \\frac { r _ { t } } { \\widehat { v _ { t } } } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 588, + 225, + 604 + ], + "score": 1.0, + "content": "respectively. We set", + "type": "text" + }, + { + "bbox": [ + 225, + 590, + 236, + 601 + ], + "score": 0.87, + "content": "\\beta _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 588, + 506, + 604 + ], + "score": 1.0, + "content": "bto 0.999, the number of sampled trajectories to 5000, the number", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 600, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 614 + ], + "score": 1.0, + "content": "bof iterations to 6000, and summarize the simulation results in Figure 9. Across all six settings with", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 612, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 142, + 626 + ], + "score": 1.0, + "content": "different", + "type": "text" + }, + { + "bbox": [ + 143, + 614, + 150, + 624 + ], + "score": 0.78, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 612, + 505, + 626 + ], + "score": 1.0, + "content": ", the adaptive learning rate has a larger variance in the first stage and the rectified adaptive", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 623, + 472, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 472, + 636 + ], + "score": 1.0, + "content": "learning rate has relative consistent variance. This verifies the reliability of our assumption.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39 + }, + { + "type": "title", + "bbox": [ + 107, + 651, + 195, + 664 + ], + "lines": [ + { + "bbox": [ + 105, + 649, + 198, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 198, + 667 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "In this paper, we explore the underlying principle of the effectiveness of the warmup heuristic used", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "for adaptive optimization algorithms. Specifically, we identify that, due to the limited amount of", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 698, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 104, + 698, + 506, + 713 + ], + "score": 1.0, + "content": "samples in the early stage of model training, the adaptive learning rate has an undesirably large", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "variance and can cause the model to converge to suspicious/bad local optima. We provide both", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 721, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 734 + ], + "score": 1.0, + "content": "empirical and theoretical evidence to support our hypothesis, and further propose a new variant", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 131, + 88, + 477, + 124 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 197, + 74, + 414, + 86 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 196, + 73, + 414, + 87 + ], + "spans": [ + { + "bbox": [ + 196, + 73, + 414, + 87 + ], + "score": 1.0, + "content": "Table 2: BLEU score on Neural Machine Translation.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 131, + 88, + 477, + 124 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 131, + 88, + 477, + 124 + ], + "spans": [ + { + "bbox": [ + 131, + 88, + 477, + 124 + ], + "score": 0.967, + "html": "
MethodIWSLT'14 DE-ENIWSLT'14 EN-DEWMT'16 EN-DE
Adam with warmupRAdam34.66± 0.01434.76 ± 0.00328.56± 0.06728.48 ± 0.05427.0327.27
", + "type": "table", + "image_path": "a8fe3dfdbb40c1eb36ceeadac1a0d9d2b8896d7e820a4e3e7cf0047e895f9ed8.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 131, + 88, + 477, + 100.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 131, + 100.0, + 477, + 112.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 131, + 112.0, + 477, + 124.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 108, + 129, + 505, + 173 + ], + "lines": [], + "index": 5.5, + "bbox_fs": [ + 105, + 129, + 505, + 175 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 178, + 505, + 234 + ], + "lines": [ + { + "bbox": [ + 106, + 178, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 505, + 190 + ], + "score": 1.0, + "content": "We summarize the performance obtained with the heuristic warmup and our proposed rectification", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 189, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 505, + 202 + ], + "score": 1.0, + "content": "term in Table 2 and visualize the training curve of IWSLT De-En in Figure 1. With a consistent", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "adaptive learning rate variance, our proposed method achieves similar performance to that of previ-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 211, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 505, + 224 + ], + "score": 1.0, + "content": "ous state-of-the-art warmup heuristics. It verifies our intuition that the problematic updates of Adam", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 221, + 387, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 387, + 236 + ], + "score": 1.0, + "content": "are indeed caused by the undesirably large variance in the early stage.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 178, + 505, + 236 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 506, + 252 + ], + "score": 1.0, + "content": "Moreover, we applied Adam with warmup on the CIFAR10 dataset. Its best accuracy on the test", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 250, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 263 + ], + "score": 1.0, + "content": "set is 91.29, which is similar to RAdam (91.38). However, we found that RAdam requires less hy-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 261, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 506, + 275 + ], + "score": 1.0, + "content": "perparameter tuning. Specifically, we visualize their learning curves in Figure 7. For some warmup", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 272, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 505, + 285 + ], + "score": 1.0, + "content": "steps, Adam with warmup is relatively more sensitive to the choice of the learning rate. RAdam,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 283, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 505, + 296 + ], + "score": 1.0, + "content": "at the same time, is not only more robust, but also can automatically control the warmup behav-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "ior (i.e., without requiring the length of warmup). For example, when setting the learning rate as", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 305, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 305, + 505, + 318 + ], + "score": 1.0, + "content": "0.1, Adam with 100 steps of warmup fails to get satisfying performance and only results in an ac-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 316, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 506, + 329 + ], + "score": 1.0, + "content": "curacy of 90.13; RAdam successfully gets an accuracy of 91.06, with the original setting of the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 327, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 239, + 339 + ], + "score": 1.0, + "content": "moving average calculation (i.e.,", + "type": "text" + }, + { + "bbox": [ + 239, + 327, + 329, + 339 + ], + "score": 0.89, + "content": "\\beta _ { 1 } = 0 . 9 , \\beta _ { 2 } = 0 . 9 9 9 )", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 327, + 505, + 339 + ], + "score": 1.0, + "content": ". We conjecture the reason is due to the fact", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "score": 1.0, + "content": "that RAdam, which is based on a rigorous variance analysis, explicitly avoids the extreme situation", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 349, + 477, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 477, + 361 + ], + "score": 1.0, + "content": "where the variance is divergent, and rectifies the variance to be consistent in other situations.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 239, + 506, + 361 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 374, + 247, + 385 + ], + "lines": [ + { + "bbox": [ + 106, + 373, + 248, + 386 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 248, + 386 + ], + "score": 1.0, + "content": "5.3 SIMULATED VERIFICATION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 393, + 505, + 442 + ], + "lines": [ + { + "bbox": [ + 104, + 393, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 104, + 394, + 263, + 414 + ], + "score": 1.0, + "content": "In Sections 3 and 4, we approximate", + "type": "text" + }, + { + "bbox": [ + 263, + 393, + 340, + 413 + ], + "score": 0.93, + "content": "\\textstyle \\operatorname { V a r } [ \\sqrt { t / \\sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 394, + 468, + 414 + ], + "score": 1.0, + "content": "to the first order, and assume", + "type": "text" + }, + { + "bbox": [ + 468, + 397, + 505, + 410 + ], + "score": 0.89, + "content": "\\psi ^ { 2 } ( . ) = ", + "type": "inline_equation" + } + ], + "index": 25 + }, + { + "bbox": [ + 108, + 410, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 108, + 411, + 182, + 432 + ], + "score": 0.88, + "content": "\\frac { 1 - \\beta _ { 2 } ^ { t } } { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 410, + 505, + 435 + ], + "score": 1.0, + "content": "subjects to a scaled inverse chi-square distribution (this assumption covers the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "score": 1.0, + "content": "approximation from EMA to SMA). Here, we examine these two approximations using simulations.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 104, + 393, + 505, + 443 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 447, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 102, + 446, + 509, + 470 + ], + "spans": [ + { + "bbox": [ + 102, + 446, + 240, + 470 + ], + "score": 1.0, + "content": "First Order Approximation of", + "type": "text" + }, + { + "bbox": [ + 241, + 447, + 318, + 467 + ], + "score": 0.93, + "content": "\\textstyle \\operatorname { V a r } [ \\sqrt { t / \\sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 446, + 509, + 470 + ], + "score": 1.0, + "content": ". To compare Equations 5 and 2, we assume", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 465, + 504, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 135, + 476 + ], + "score": 0.87, + "content": "\\tau = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 465, + 301, + 478 + ], + "score": 1.0, + "content": "and plot their values and difference for", + "type": "text" + }, + { + "bbox": [ + 301, + 465, + 378, + 478 + ], + "score": 0.92, + "content": "\\nu = \\{ 5 , \\cdots , 5 0 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 465, + 504, + 478 + ], + "score": 1.0, + "content": "in Figure 8. The curve of the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 477, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 506, + 489 + ], + "score": 1.0, + "content": "analytic form and the first-order approximation highly overlap, and their difference is much smaller", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 487, + 453, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 453, + 500 + ], + "score": 1.0, + "content": "than their value. This result verifies that our first-order approximation is very accurate.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 102, + 446, + 509, + 500 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 504, + 505, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 504, + 507, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 442, + 518 + ], + "score": 1.0, + "content": "Scaled Inverse Chi-Square Distribution Assumption. In this paper, we assume", + "type": "text" + }, + { + "bbox": [ + 442, + 507, + 451, + 516 + ], + "score": 0.83, + "content": "g _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 504, + 507, + 518 + ], + "score": 1.0, + "content": "accords to a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 515, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 324, + 529 + ], + "score": 1.0, + "content": "Normal distribution with a zero mean. We also assume", + "type": "text" + }, + { + "bbox": [ + 324, + 515, + 346, + 528 + ], + "score": 0.91, + "content": "\\psi ^ { 2 } ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 515, + 506, + 529 + ], + "score": 1.0, + "content": "accords to the scaled inverse chi-square", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 260, + 540 + ], + "score": 1.0, + "content": "distribution to derive the variance of", + "type": "text" + }, + { + "bbox": [ + 261, + 527, + 300, + 539 + ], + "score": 0.91, + "content": "\\mathrm { V a r } [ \\psi ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 527, + 506, + 540 + ], + "score": 1.0, + "content": ", based on the similarity between the exponential", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 538, + 467, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 467, + 551 + ], + "score": 1.0, + "content": "moving average and simple moving average. Here, we empirically verify this assumption.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 504, + 507, + 551 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 554, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 180, + 567 + ], + "score": 1.0, + "content": "Specifically, since", + "type": "text" + }, + { + "bbox": [ + 181, + 556, + 190, + 566 + ], + "score": 0.85, + "content": "g _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "in the optimization problem may not be zero-mean, we assume its expectation", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 116, + 578 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 116, + 567, + 124, + 577 + ], + "score": 0.79, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 565, + 173, + 578 + ], + "score": 1.0, + "content": "and sample", + "type": "text" + }, + { + "bbox": [ + 174, + 567, + 183, + 577 + ], + "score": 0.85, + "content": "g _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 565, + 207, + 578 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 208, + 565, + 241, + 578 + ], + "score": 0.93, + "content": "\\mathcal { N } ( \\bar { \\boldsymbol { \\mu } } , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 565, + 505, + 578 + ], + "score": 1.0, + "content": ". Then, based on these samples, we calculate the variance of the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 576, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 453, + 591 + ], + "score": 1.0, + "content": "original adaptive learning rate and the proposed rectified adaptive learning rate, i.e.,", + "type": "text" + }, + { + "bbox": [ + 454, + 576, + 486, + 591 + ], + "score": 0.91, + "content": "\\textstyle \\operatorname { V a r } [ { \\frac { 1 } { \\widehat { v _ { t } } } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 576, + 506, + 591 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 588, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 138, + 603 + ], + "score": 0.9, + "content": "\\textstyle \\operatorname { V a r } [ { \\frac { r _ { t } } { \\widehat { v _ { t } } } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 588, + 225, + 604 + ], + "score": 1.0, + "content": "respectively. We set", + "type": "text" + }, + { + "bbox": [ + 225, + 590, + 236, + 601 + ], + "score": 0.87, + "content": "\\beta _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 588, + 506, + 604 + ], + "score": 1.0, + "content": "bto 0.999, the number of sampled trajectories to 5000, the number", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 600, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 614 + ], + "score": 1.0, + "content": "bof iterations to 6000, and summarize the simulation results in Figure 9. Across all six settings with", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 612, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 142, + 626 + ], + "score": 1.0, + "content": "different", + "type": "text" + }, + { + "bbox": [ + 143, + 614, + 150, + 624 + ], + "score": 0.78, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 612, + 505, + 626 + ], + "score": 1.0, + "content": ", the adaptive learning rate has a larger variance in the first stage and the rectified adaptive", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 623, + 472, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 472, + 636 + ], + "score": 1.0, + "content": "learning rate has relative consistent variance. This verifies the reliability of our assumption.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 555, + 506, + 636 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 651, + 195, + 664 + ], + "lines": [ + { + "bbox": [ + 105, + 649, + 198, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 198, + 667 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "In this paper, we explore the underlying principle of the effectiveness of the warmup heuristic used", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "for adaptive optimization algorithms. Specifically, we identify that, due to the limited amount of", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 698, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 104, + 698, + 506, + 713 + ], + "score": 1.0, + "content": "samples in the early stage of model training, the adaptive learning rate has an undesirably large", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "variance and can cause the model to converge to suspicious/bad local optima. We provide both", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 721, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 734 + ], + "score": 1.0, + "content": "empirical and theoretical evidence to support our hypothesis, and further propose a new variant", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 266, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 505, + 278 + ], + "score": 1.0, + "content": "of Adam, whose adaptive learning rate is rectified so as to have a consistent variance. Empirical", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 277, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 505, + 289 + ], + "score": 1.0, + "content": "results demonstrate the effectiveness of our proposed method. In future work, we plan to replace the", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 288, + 465, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 465, + 300 + ], + "score": 1.0, + "content": "rectification strategy by sharing the second moment estimation across similar parameters.", + "type": "text", + "cross_page": true + } + ], + "index": 20 + } + ], + "index": 46, + "bbox_fs": [ + 104, + 677, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 97, + 254, + 204 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 97, + 254, + 204 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 97, + 254, + 204 + ], + "spans": [ + { + "bbox": [ + 108, + 97, + 254, + 204 + ], + "score": 0.964, + "type": "image", + "image_path": "faa8a54dec9667c124d2828f50a5de48c77fb5bb8cb529cf12a5d90cfc18c0fa.jpg" + } + ] + } + ], + "index": 0.5, + "virtual_lines": [ + { + "bbox": [ + 108, + 97, + 254, + 150.5 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 150.5, + 254, + 204.0 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 209, + 259, + 254 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 208, + 259, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 259, + 221 + ], + "score": 1.0, + "content": "Figure 8: The value of Equation 2,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 220, + 259, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 259, + 231 + ], + "score": 1.0, + "content": "Equation 5 and their difference (abso-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 231, + 260, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 195, + 243 + ], + "score": 1.0, + "content": "lute difference). The", + "type": "text" + }, + { + "bbox": [ + 196, + 233, + 202, + 241 + ], + "score": 0.56, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 231, + 234, + 243 + ], + "score": 1.0, + "content": "-axis is", + "type": "text" + }, + { + "bbox": [ + 234, + 233, + 241, + 243 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 231, + 260, + 243 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 241, + 255, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 255, + 255 + ], + "score": 1.0, + "content": "the y-axis is the variance (log scale).", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 270, + 92, + 512, + 228 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 270, + 92, + 512, + 228 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 270, + 92, + 512, + 227 + ], + "spans": [ + { + "bbox": [ + 270, + 92, + 512, + 227 + ], + "score": 0.972, + "type": "image", + "image_path": "2c37e4a5e73292a2c4f76f3284121bfbb32e39ddd248373e6bf36a3c173b8e7f.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 270, + 92, + 512, + 105.6 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 270, + 105.6, + 512, + 119.19999999999999 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 270, + 119.19999999999999, + 512, + 132.79999999999998 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 270, + 132.79999999999998, + 512, + 146.39999999999998 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 270, + 146.39999999999998, + 512, + 159.99999999999997 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 270, + 159.99999999999997, + 512, + 173.59999999999997 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 270, + 173.59999999999997, + 512, + 187.19999999999996 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 270, + 187.19999999999996, + 512, + 200.79999999999995 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 270, + 200.79999999999995, + 512, + 214.39999999999995 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 270, + 214.39999999999995, + 512, + 227.99999999999994 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 263, + 231, + 507, + 256 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 262, + 231, + 508, + 246 + ], + "spans": [ + { + "bbox": [ + 262, + 231, + 376, + 246 + ], + "score": 1.0, + "content": "Figure 9: The simulation of", + "type": "text" + }, + { + "bbox": [ + 376, + 231, + 408, + 246 + ], + "score": 0.92, + "content": "\\mathrm { V a r } [ \\frac { 1 } { v _ { t } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 231, + 426, + 246 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 426, + 232, + 458, + 246 + ], + "score": 0.91, + "content": "\\textstyle \\operatorname { V a r } [ { \\frac { c _ { t } } { v _ { t } } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 231, + 480, + 246 + ], + "score": 1.0, + "content": ". The", + "type": "text" + }, + { + "bbox": [ + 480, + 234, + 487, + 243 + ], + "score": 0.62, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 231, + 508, + 246 + ], + "score": 1.0, + "content": "-axis", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 262, + 243, + 502, + 258 + ], + "spans": [ + { + "bbox": [ + 262, + 243, + 502, + 258 + ], + "score": 1.0, + "content": "is iteration # (from 5), the y-axis is the variance (log scale).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 266, + 505, + 299 + ], + "lines": [ + { + "bbox": [ + 106, + 266, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 505, + 278 + ], + "score": 1.0, + "content": "of Adam, whose adaptive learning rate is rectified so as to have a consistent variance. Empirical", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 277, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 505, + 289 + ], + "score": 1.0, + "content": "results demonstrate the effectiveness of our proposed method. In future work, we plan to replace the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 288, + 465, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 465, + 300 + ], + "score": 1.0, + "content": "rectification strategy by sharing the second moment estimation across similar parameters.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 108, + 310, + 190, + 321 + ], + "lines": [ + { + "bbox": [ + 106, + 309, + 191, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 191, + 324 + ], + "score": 1.0, + "content": "ACKNOWLEDGE", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 108, + 325, + 505, + 369 + ], + "lines": [ + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "We thank Aeyuan Allen-Zhu for valuable discussions and comments, Microsoft Research Technol-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "score": 1.0, + "content": "ogy Engineering team for setting up GPU machines. Research was sponsored in part by DARPA", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 347, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 505, + 359 + ], + "score": 1.0, + "content": "No. W911NF-17-C-0099 and FA8750-19-2-1004, National Science Foundation IIS 16-18481, IIS", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 359, + 350, + 369 + ], + "spans": [ + { + "bbox": [ + 107, + 359, + 350, + 369 + ], + "score": 1.0, + "content": "17-04532, and IIS-17-41317, and DTRA HDTRA11810026.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "title", + "bbox": [ + 107, + 386, + 175, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 386, + 176, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 176, + 399 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 403, + 504, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 402, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 417 + ], + "score": 1.0, + "content": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 414, + 220, + 426 + ], + "spans": [ + { + "bbox": [ + 115, + 414, + 220, + 426 + ], + "score": 1.0, + "content": "arXiv:1607.06450, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 432, + 504, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 431, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 505, + 446 + ], + "score": 1.0, + "content": "David Balduzzi, Marcus Frean, Lennox Leary, JP Lewis, Kurt Wan-Duo Ma, and Brian McWilliams.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 115, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "The shattered gradients problem: If resnets are the answer, then what is the question? In ICML,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 453, + 144, + 468 + ], + "spans": [ + { + "bbox": [ + 115, + 453, + 144, + 468 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 473, + 504, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 487 + ], + "score": 1.0, + "content": "Yoshua Bengio, Nicolas Boulanger-Lewandowski, and Razvan Pascanu. Advances in optimizing", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 115, + 483, + 505, + 497 + ], + "score": 1.0, + "content": "recurrent networks. In 2013 IEEE International Conference on Acoustics, Speech and Signal", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 495, + 281, + 507 + ], + "spans": [ + { + "bbox": [ + 116, + 495, + 281, + 507 + ], + "score": 1.0, + "content": "Processing, pp. 8624–8628. IEEE, 2013.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 105, + 513, + 503, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 511, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 526 + ], + "score": 1.0, + "content": "Jeremy Bernstein, Yu-Xiang Wang, Kamyar Azizzadenesheli, and Anima Anandkumar. signsgd:", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 524, + 391, + 537 + ], + "spans": [ + { + "bbox": [ + 116, + 524, + 391, + 537 + ], + "score": 1.0, + "content": "Compressed optimisation for non-convex problems. In ICML, 2018.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 106, + 542, + 503, + 565 + ], + "lines": [ + { + "bbox": [ + 106, + 540, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 556 + ], + "score": 1.0, + "content": "Augustin Cauchy. Methode g ´ en´ erale pour la r ´ esolution des systemes d ´ equations simultan ´ ees. ´ Comp.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 117, + 553, + 289, + 565 + ], + "spans": [ + { + "bbox": [ + 117, + 553, + 289, + 565 + ], + "score": 1.0, + "content": "Rend. Sci. Paris, 25(1847):536–538, 1847.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 106, + 570, + 504, + 605 + ], + "lines": [ + { + "bbox": [ + 105, + 570, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 584 + ], + "score": 1.0, + "content": "Mauro Cettolo, Jan Niehues, Sebastian Stuker, Luisa Bentivogli, and Marcello Federico. Report on ¨", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 115, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "the 11th iwslt evaluation campaign, iwslt 2014. In Proceedings of the International Workshop on", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 593, + 268, + 605 + ], + "spans": [ + { + "bbox": [ + 116, + 593, + 268, + 605 + ], + "score": 1.0, + "content": "Spoken Language Translation,, 2014.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 106, + 611, + 504, + 645 + ], + "lines": [ + { + "bbox": [ + 106, + 611, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 506, + 624 + ], + "score": 1.0, + "content": "Ciprian Chelba, Tomas Mikolov, Michael Schuster, Qi Ge, Thorsten Brants, Phillipp Koehn, and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 116, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "Tony Robinson. One billion word benchmark for measuring progress in statistical language mod-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 633, + 248, + 646 + ], + "spans": [ + { + "bbox": [ + 116, + 633, + 248, + 646 + ], + "score": 1.0, + "content": "eling. In INTERSPEECH, 2013.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 105, + 651, + 504, + 674 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 506, + 665 + ], + "score": 1.0, + "content": "Jinghui Chen and Quanquan Gu. Closing the generalization gap of adaptive gradient methods in", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 662, + 404, + 675 + ], + "spans": [ + { + "bbox": [ + 115, + 662, + 404, + 675 + ], + "score": 1.0, + "content": "training deep neural networks. arXiv preprint arXiv:1806.06763, 2018.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + }, + { + "type": "text", + "bbox": [ + 106, + 680, + 503, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 504, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 504, + 693 + ], + "score": 1.0, + "content": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 691, + 299, + 704 + ], + "spans": [ + { + "bbox": [ + 115, + 691, + 299, + 704 + ], + "score": 1.0, + "content": "hierarchical image database. In ICML, 2009.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 115, + 720, + 430, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 430, + 732 + ], + "score": 1.0, + "content": "bidirectional transformers for language understanding. In NAACL-HLT, 2019.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 97, + 254, + 204 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 97, + 254, + 204 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 97, + 254, + 204 + ], + "spans": [ + { + "bbox": [ + 108, + 97, + 254, + 204 + ], + "score": 0.964, + "type": "image", + "image_path": "faa8a54dec9667c124d2828f50a5de48c77fb5bb8cb529cf12a5d90cfc18c0fa.jpg" + } + ] + } + ], + "index": 0.5, + "virtual_lines": [ + { + "bbox": [ + 108, + 97, + 254, + 150.5 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 150.5, + 254, + 204.0 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 209, + 259, + 254 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 208, + 259, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 259, + 221 + ], + "score": 1.0, + "content": "Figure 8: The value of Equation 2,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 220, + 259, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 259, + 231 + ], + "score": 1.0, + "content": "Equation 5 and their difference (abso-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 231, + 260, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 195, + 243 + ], + "score": 1.0, + "content": "lute difference). The", + "type": "text" + }, + { + "bbox": [ + 196, + 233, + 202, + 241 + ], + "score": 0.56, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 231, + 234, + 243 + ], + "score": 1.0, + "content": "-axis is", + "type": "text" + }, + { + "bbox": [ + 234, + 233, + 241, + 243 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 231, + 260, + 243 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 241, + 255, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 255, + 255 + ], + "score": 1.0, + "content": "the y-axis is the variance (log scale).", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 270, + 92, + 512, + 228 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 270, + 92, + 512, + 228 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 270, + 92, + 512, + 227 + ], + "spans": [ + { + "bbox": [ + 270, + 92, + 512, + 227 + ], + "score": 0.972, + "type": "image", + "image_path": "2c37e4a5e73292a2c4f76f3284121bfbb32e39ddd248373e6bf36a3c173b8e7f.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 270, + 92, + 512, + 105.6 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 270, + 105.6, + 512, + 119.19999999999999 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 270, + 119.19999999999999, + 512, + 132.79999999999998 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 270, + 132.79999999999998, + 512, + 146.39999999999998 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 270, + 146.39999999999998, + 512, + 159.99999999999997 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 270, + 159.99999999999997, + 512, + 173.59999999999997 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 270, + 173.59999999999997, + 512, + 187.19999999999996 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 270, + 187.19999999999996, + 512, + 200.79999999999995 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 270, + 200.79999999999995, + 512, + 214.39999999999995 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 270, + 214.39999999999995, + 512, + 227.99999999999994 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 263, + 231, + 507, + 256 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 262, + 231, + 508, + 246 + ], + "spans": [ + { + "bbox": [ + 262, + 231, + 376, + 246 + ], + "score": 1.0, + "content": "Figure 9: The simulation of", + "type": "text" + }, + { + "bbox": [ + 376, + 231, + 408, + 246 + ], + "score": 0.92, + "content": "\\mathrm { V a r } [ \\frac { 1 } { v _ { t } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 231, + 426, + 246 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 426, + 232, + 458, + 246 + ], + "score": 0.91, + "content": "\\textstyle \\operatorname { V a r } [ { \\frac { c _ { t } } { v _ { t } } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 231, + 480, + 246 + ], + "score": 1.0, + "content": ". The", + "type": "text" + }, + { + "bbox": [ + 480, + 234, + 487, + 243 + ], + "score": 0.62, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 231, + 508, + 246 + ], + "score": 1.0, + "content": "-axis", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 262, + 243, + 502, + 258 + ], + "spans": [ + { + "bbox": [ + 262, + 243, + 502, + 258 + ], + "score": 1.0, + "content": "is iteration # (from 5), the y-axis is the variance (log scale).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 266, + 505, + 299 + ], + "lines": [], + "index": 19, + "bbox_fs": [ + 106, + 266, + 505, + 300 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 310, + 190, + 321 + ], + "lines": [ + { + "bbox": [ + 106, + 309, + 191, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 191, + 324 + ], + "score": 1.0, + "content": "ACKNOWLEDGE", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 108, + 325, + 505, + 369 + ], + "lines": [ + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "We thank Aeyuan Allen-Zhu for valuable discussions and comments, Microsoft Research Technol-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "score": 1.0, + "content": "ogy Engineering team for setting up GPU machines. Research was sponsored in part by DARPA", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 347, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 505, + 359 + ], + "score": 1.0, + "content": "No. W911NF-17-C-0099 and FA8750-19-2-1004, National Science Foundation IIS 16-18481, IIS", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 359, + 350, + 369 + ], + "spans": [ + { + "bbox": [ + 107, + 359, + 350, + 369 + ], + "score": 1.0, + "content": "17-04532, and IIS-17-41317, and DTRA HDTRA11810026.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 325, + 505, + 369 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 386, + 175, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 386, + 176, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 176, + 399 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 403, + 504, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 402, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 417 + ], + "score": 1.0, + "content": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 414, + 220, + 426 + ], + "spans": [ + { + "bbox": [ + 115, + 414, + 220, + 426 + ], + "score": 1.0, + "content": "arXiv:1607.06450, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 402, + 506, + 426 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 432, + 504, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 431, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 505, + 446 + ], + "score": 1.0, + "content": "David Balduzzi, Marcus Frean, Lennox Leary, JP Lewis, Kurt Wan-Duo Ma, and Brian McWilliams.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 115, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "The shattered gradients problem: If resnets are the answer, then what is the question? In ICML,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 453, + 144, + 468 + ], + "spans": [ + { + "bbox": [ + 115, + 453, + 144, + 468 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 431, + 505, + 468 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 473, + 504, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 487 + ], + "score": 1.0, + "content": "Yoshua Bengio, Nicolas Boulanger-Lewandowski, and Razvan Pascanu. Advances in optimizing", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 115, + 483, + 505, + 497 + ], + "score": 1.0, + "content": "recurrent networks. In 2013 IEEE International Conference on Acoustics, Speech and Signal", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 495, + 281, + 507 + ], + "spans": [ + { + "bbox": [ + 116, + 495, + 281, + 507 + ], + "score": 1.0, + "content": "Processing, pp. 8624–8628. IEEE, 2013.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 471, + 505, + 507 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 513, + 503, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 511, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 526 + ], + "score": 1.0, + "content": "Jeremy Bernstein, Yu-Xiang Wang, Kamyar Azizzadenesheli, and Anima Anandkumar. signsgd:", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 524, + 391, + 537 + ], + "spans": [ + { + "bbox": [ + 116, + 524, + 391, + 537 + ], + "score": 1.0, + "content": "Compressed optimisation for non-convex problems. In ICML, 2018.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 511, + 505, + 537 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 542, + 503, + 565 + ], + "lines": [ + { + "bbox": [ + 106, + 540, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 556 + ], + "score": 1.0, + "content": "Augustin Cauchy. Methode g ´ en´ erale pour la r ´ esolution des systemes d ´ equations simultan ´ ees. ´ Comp.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 117, + 553, + 289, + 565 + ], + "spans": [ + { + "bbox": [ + 117, + 553, + 289, + 565 + ], + "score": 1.0, + "content": "Rend. Sci. Paris, 25(1847):536–538, 1847.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 106, + 540, + 505, + 565 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 570, + 504, + 605 + ], + "lines": [ + { + "bbox": [ + 105, + 570, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 584 + ], + "score": 1.0, + "content": "Mauro Cettolo, Jan Niehues, Sebastian Stuker, Luisa Bentivogli, and Marcello Federico. Report on ¨", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 115, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "the 11th iwslt evaluation campaign, iwslt 2014. In Proceedings of the International Workshop on", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 593, + 268, + 605 + ], + "spans": [ + { + "bbox": [ + 116, + 593, + 268, + 605 + ], + "score": 1.0, + "content": "Spoken Language Translation,, 2014.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 570, + 505, + 605 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 611, + 504, + 645 + ], + "lines": [ + { + "bbox": [ + 106, + 611, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 506, + 624 + ], + "score": 1.0, + "content": "Ciprian Chelba, Tomas Mikolov, Michael Schuster, Qi Ge, Thorsten Brants, Phillipp Koehn, and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 116, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "Tony Robinson. One billion word benchmark for measuring progress in statistical language mod-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 633, + 248, + 646 + ], + "spans": [ + { + "bbox": [ + 116, + 633, + 248, + 646 + ], + "score": 1.0, + "content": "eling. In INTERSPEECH, 2013.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 106, + 611, + 506, + 646 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 651, + 504, + 674 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 506, + 665 + ], + "score": 1.0, + "content": "Jinghui Chen and Quanquan Gu. Closing the generalization gap of adaptive gradient methods in", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 662, + 404, + 675 + ], + "spans": [ + { + "bbox": [ + 115, + 662, + 404, + 675 + ], + "score": 1.0, + "content": "training deep neural networks. arXiv preprint arXiv:1806.06763, 2018.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 650, + 506, + 675 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 680, + 503, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 504, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 504, + 693 + ], + "score": 1.0, + "content": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 691, + 299, + 704 + ], + "spans": [ + { + "bbox": [ + 115, + 691, + 299, + 704 + ], + "score": 1.0, + "content": "hierarchical image database. In ICML, 2009.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5, + "bbox_fs": [ + 106, + 680, + 504, + 704 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 115, + 720, + 430, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 430, + 732 + ], + "score": 1.0, + "content": "bidirectional transformers for language understanding. In NAACL-HLT, 2019.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49.5, + "bbox_fs": [ + 105, + 708, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 383, + 95 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 384, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 384, + 95 + ], + "score": 1.0, + "content": "Timothy Dozat. Incorporating nesterov momentum into adam. 2016.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 504, + 123 + ], + "lines": [ + { + "bbox": [ + 106, + 100, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 505, + 113 + ], + "score": 1.0, + "content": "John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 112, + 280, + 122 + ], + "spans": [ + { + "bbox": [ + 116, + 112, + 280, + 122 + ], + "score": 1.0, + "content": "stochastic optimization. In COLT, 2010.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 108, + 128, + 504, + 152 + ], + "lines": [ + { + "bbox": [ + 106, + 128, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 505, + 141 + ], + "score": 1.0, + "content": "Carl-Friedrich Gauss. Theoria combinationis observationum erroribus minimis obnoxiae. Commen-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 140, + 401, + 151 + ], + "spans": [ + { + "bbox": [ + 115, + 140, + 401, + 151 + ], + "score": 1.0, + "content": "tationes Societatis Regiae Scientiarum Gottingensis Recentiores, 1823.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 157, + 504, + 180 + ], + "lines": [ + { + "bbox": [ + 106, + 156, + 506, + 170 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 506, + 170 + ], + "score": 1.0, + "content": "Akhilesh Gotmare, Nitish Shirish Keskar, Caiming Xiong, and Richard Socher. A closer look at", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 168, + 470, + 182 + ], + "spans": [ + { + "bbox": [ + 115, + 168, + 470, + 182 + ], + "score": 1.0, + "content": "deep learning heuristics: Learning rate restarts, warmup and distillation. In ICLR, 2019.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 185, + 504, + 220 + ], + "lines": [ + { + "bbox": [ + 105, + 185, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 505, + 199 + ], + "score": 1.0, + "content": "Priya Goyal, Piotr Dollar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, An- ´", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 197, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 116, + 197, + 505, + 210 + ], + "score": 1.0, + "content": "drew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 208, + 320, + 221 + ], + "spans": [ + { + "bbox": [ + 116, + 208, + 320, + 221 + ], + "score": 1.0, + "content": "in 1 hour. arXiv preprint arXiv:1706.02677, 2017.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 225, + 503, + 249 + ], + "lines": [ + { + "bbox": [ + 105, + 223, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 241 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 236, + 212, + 249 + ], + "spans": [ + { + "bbox": [ + 116, + 236, + 212, + 249 + ], + "score": 1.0, + "content": "nition. In CVPR, 2016.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 254, + 503, + 278 + ], + "lines": [ + { + "bbox": [ + 106, + 252, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 505, + 268 + ], + "score": 1.0, + "content": "Geoffrey Hinton, Nitish Srivastava, and Kevin Swersky. Neural networks for machine learning", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 117, + 266, + 389, + 277 + ], + "spans": [ + { + "bbox": [ + 117, + 266, + 389, + 277 + ], + "score": 1.0, + "content": "lecture 6a overview of mini-batch gradient descent. Cited on, 2012.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 283, + 502, + 306 + ], + "lines": [ + { + "bbox": [ + 105, + 281, + 504, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 504, + 297 + ], + "score": 1.0, + "content": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 294, + 313, + 306 + ], + "spans": [ + { + "bbox": [ + 115, + 294, + 313, + 306 + ], + "score": 1.0, + "content": "reducing internal covariate shift. In ICML, 2015.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 312, + 496, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 497, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 497, + 325 + ], + "score": 1.0, + "content": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 105, + 329, + 503, + 352 + ], + "lines": [ + { + "bbox": [ + 105, + 327, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 505, + 344 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 117, + 341, + 249, + 352 + ], + "spans": [ + { + "bbox": [ + 117, + 341, + 249, + 352 + ], + "score": 1.0, + "content": "Technical report, Citeseer, 2009.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 105, + 358, + 503, + 381 + ], + "lines": [ + { + "bbox": [ + 105, + 357, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 505, + 372 + ], + "score": 1.0, + "content": "Liyuan Liu, Xiang Ren, Jingbo Shang, Jian Peng, and Jiawei Han. Efficient contextualized repre-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 369, + 416, + 382 + ], + "spans": [ + { + "bbox": [ + 115, + 369, + 416, + 382 + ], + "score": 1.0, + "content": "sentation: Language model pruning for sequence labeling. EMNLP, 2018.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 109, + 386, + 503, + 421 + ], + "lines": [ + { + "bbox": [ + 106, + 386, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 505, + 400 + ], + "score": 1.0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 397, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 115, + 397, + 505, + 412 + ], + "score": 1.0, + "content": "Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 410, + 321, + 421 + ], + "spans": [ + { + "bbox": [ + 116, + 410, + 321, + 421 + ], + "score": 1.0, + "content": "approach. arXiv preprint arXiv:1907.11692, 2019.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 426, + 487, + 439 + ], + "lines": [ + { + "bbox": [ + 105, + 426, + 487, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 487, + 439 + ], + "score": 1.0, + "content": "Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. In ICLR, 2018.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 444, + 503, + 467 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 505, + 458 + ], + "score": 1.0, + "content": "Liangchen Luo, Yuanhao Xiong, Yan Liu, and Xu Sun. Adaptive gradient methods with dynamic", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 455, + 273, + 468 + ], + "spans": [ + { + "bbox": [ + 116, + 455, + 273, + 468 + ], + "score": 1.0, + "content": "bound of learning rate. In ICLR, 2019.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 473, + 326, + 485 + ], + "lines": [ + { + "bbox": [ + 105, + 472, + 326, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 326, + 486 + ], + "score": 1.0, + "content": "Robert Nau. Forecasting with moving averages. 2014.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 491, + 504, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 505 + ], + "score": 1.0, + "content": "Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 501, + 494, + 515 + ], + "spans": [ + { + "bbox": [ + 115, + 501, + 494, + 515 + ], + "score": 1.0, + "content": "and Michael Auli. fairseq: A fast, extensible toolkit for sequence modeling. In NAACL, 2019.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 519, + 505, + 542 + ], + "lines": [ + { + "bbox": [ + 105, + 518, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 533 + ], + "score": 1.0, + "content": "Martin Popel and Ondˇrej Bojar. Training tips for the transformer model. The Prague Bulletin of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 530, + 307, + 543 + ], + "spans": [ + { + "bbox": [ + 116, + 530, + 307, + 543 + ], + "score": 1.0, + "content": "Mathematical Linguistics, 110(1):43–70, 2018.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 547, + 504, + 571 + ], + "lines": [ + { + "bbox": [ + 105, + 547, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 506, + 561 + ], + "score": 1.0, + "content": "Sashank J Reddi, Satyen Kale, and Sanjiv Kumar. On the convergence of adam and beyond. In", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 559, + 168, + 571 + ], + "spans": [ + { + "bbox": [ + 115, + 559, + 168, + 571 + ], + "score": 1.0, + "content": "ICLR, 2018.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 105, + 576, + 504, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 575, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 591 + ], + "score": 1.0, + "content": "Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 587, + 371, + 600 + ], + "spans": [ + { + "bbox": [ + 115, + 587, + 371, + 600 + ], + "score": 1.0, + "content": "the inception architecture for computer vision. In CVPR, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 105, + 605, + 503, + 629 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 619 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 616, + 431, + 629 + ], + "spans": [ + { + "bbox": [ + 115, + 616, + 431, + 629 + ], + "score": 1.0, + "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 106, + 634, + 447, + 646 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 446, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 446, + 647 + ], + "score": 1.0, + "content": "Kirk M Wolter. Taylor series methods. In Introduction to variance estimation. 2007.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 652, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 106, + 652, + 505, + 665 + ], + "spans": [ + { + "bbox": [ + 106, + 652, + 505, + 665 + ], + "score": 1.0, + "content": "Lin Xiao, Adams Wei Yu, Qihang Lin, and Weizhu Chen. Dscovr: Randomized primal-dual block", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 663, + 491, + 675 + ], + "spans": [ + { + "bbox": [ + 115, + 663, + 491, + 675 + ], + "score": 1.0, + "content": "coordinate algorithms for asynchronous distributed optimization. J. Mach. Learn. Res., 2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 106, + 680, + 504, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 506, + 694 + ], + "score": 1.0, + "content": "Matthew D Zeiler. Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 114, + 690, + 144, + 704 + ], + "spans": [ + { + "bbox": [ + 114, + 690, + 144, + 704 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Hongyi Zhang, Yann N Dauphin, and Tengyu Ma. Fixup initialization: Residual learning without", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 721, + 241, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 241, + 732 + ], + "score": 1.0, + "content": "normalization. In ICLR, 2019.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 383, + 95 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 384, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 384, + 95 + ], + "score": 1.0, + "content": "Timothy Dozat. Incorporating nesterov momentum into adam. 2016.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 106, + 82, + 384, + 95 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 504, + 123 + ], + "lines": [ + { + "bbox": [ + 106, + 100, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 505, + 113 + ], + "score": 1.0, + "content": "John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 112, + 280, + 122 + ], + "spans": [ + { + "bbox": [ + 116, + 112, + 280, + 122 + ], + "score": 1.0, + "content": "stochastic optimization. In COLT, 2010.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 106, + 100, + 505, + 122 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 128, + 504, + 152 + ], + "lines": [ + { + "bbox": [ + 106, + 128, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 505, + 141 + ], + "score": 1.0, + "content": "Carl-Friedrich Gauss. Theoria combinationis observationum erroribus minimis obnoxiae. Commen-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 140, + 401, + 151 + ], + "spans": [ + { + "bbox": [ + 115, + 140, + 401, + 151 + ], + "score": 1.0, + "content": "tationes Societatis Regiae Scientiarum Gottingensis Recentiores, 1823.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 106, + 128, + 505, + 151 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 157, + 504, + 180 + ], + "lines": [ + { + "bbox": [ + 106, + 156, + 506, + 170 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 506, + 170 + ], + "score": 1.0, + "content": "Akhilesh Gotmare, Nitish Shirish Keskar, Caiming Xiong, and Richard Socher. A closer look at", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 168, + 470, + 182 + ], + "spans": [ + { + "bbox": [ + 115, + 168, + 470, + 182 + ], + "score": 1.0, + "content": "deep learning heuristics: Learning rate restarts, warmup and distillation. In ICLR, 2019.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 106, + 156, + 506, + 182 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 185, + 504, + 220 + ], + "lines": [ + { + "bbox": [ + 105, + 185, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 505, + 199 + ], + "score": 1.0, + "content": "Priya Goyal, Piotr Dollar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, An- ´", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 197, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 116, + 197, + 505, + 210 + ], + "score": 1.0, + "content": "drew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 208, + 320, + 221 + ], + "spans": [ + { + "bbox": [ + 116, + 208, + 320, + 221 + ], + "score": 1.0, + "content": "in 1 hour. arXiv preprint arXiv:1706.02677, 2017.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 185, + 505, + 221 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 225, + 503, + 249 + ], + "lines": [ + { + "bbox": [ + 105, + 223, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 241 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 236, + 212, + 249 + ], + "spans": [ + { + "bbox": [ + 116, + 236, + 212, + 249 + ], + "score": 1.0, + "content": "nition. In CVPR, 2016.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 223, + 505, + 249 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 254, + 503, + 278 + ], + "lines": [ + { + "bbox": [ + 106, + 252, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 505, + 268 + ], + "score": 1.0, + "content": "Geoffrey Hinton, Nitish Srivastava, and Kevin Swersky. Neural networks for machine learning", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 117, + 266, + 389, + 277 + ], + "spans": [ + { + "bbox": [ + 117, + 266, + 389, + 277 + ], + "score": 1.0, + "content": "lecture 6a overview of mini-batch gradient descent. Cited on, 2012.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 106, + 252, + 505, + 277 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 283, + 502, + 306 + ], + "lines": [ + { + "bbox": [ + 105, + 281, + 504, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 504, + 297 + ], + "score": 1.0, + "content": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 294, + 313, + 306 + ], + "spans": [ + { + "bbox": [ + 115, + 294, + 313, + 306 + ], + "score": 1.0, + "content": "reducing internal covariate shift. In ICML, 2015.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 281, + 504, + 306 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 312, + 496, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 497, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 497, + 325 + ], + "score": 1.0, + "content": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16, + "bbox_fs": [ + 106, + 311, + 497, + 325 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 329, + 503, + 352 + ], + "lines": [ + { + "bbox": [ + 105, + 327, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 505, + 344 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 117, + 341, + 249, + 352 + ], + "spans": [ + { + "bbox": [ + 117, + 341, + 249, + 352 + ], + "score": 1.0, + "content": "Technical report, Citeseer, 2009.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 327, + 505, + 352 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 358, + 503, + 381 + ], + "lines": [ + { + "bbox": [ + 105, + 357, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 505, + 372 + ], + "score": 1.0, + "content": "Liyuan Liu, Xiang Ren, Jingbo Shang, Jian Peng, and Jiawei Han. Efficient contextualized repre-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 369, + 416, + 382 + ], + "spans": [ + { + "bbox": [ + 115, + 369, + 416, + 382 + ], + "score": 1.0, + "content": "sentation: Language model pruning for sequence labeling. EMNLP, 2018.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 357, + 505, + 382 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 386, + 503, + 421 + ], + "lines": [ + { + "bbox": [ + 106, + 386, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 505, + 400 + ], + "score": 1.0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 397, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 115, + 397, + 505, + 412 + ], + "score": 1.0, + "content": "Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 410, + 321, + 421 + ], + "spans": [ + { + "bbox": [ + 116, + 410, + 321, + 421 + ], + "score": 1.0, + "content": "approach. arXiv preprint arXiv:1907.11692, 2019.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 386, + 505, + 421 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 426, + 487, + 439 + ], + "lines": [ + { + "bbox": [ + 105, + 426, + 487, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 487, + 439 + ], + "score": 1.0, + "content": "Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. In ICLR, 2018.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 426, + 487, + 439 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 444, + 503, + 467 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 505, + 458 + ], + "score": 1.0, + "content": "Liangchen Luo, Yuanhao Xiong, Yan Liu, and Xu Sun. Adaptive gradient methods with dynamic", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 455, + 273, + 468 + ], + "spans": [ + { + "bbox": [ + 116, + 455, + 273, + 468 + ], + "score": 1.0, + "content": "bound of learning rate. In ICLR, 2019.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 443, + 505, + 468 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 473, + 326, + 485 + ], + "lines": [ + { + "bbox": [ + 105, + 472, + 326, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 326, + 486 + ], + "score": 1.0, + "content": "Robert Nau. Forecasting with moving averages. 2014.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 472, + 326, + 486 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 491, + 504, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 505 + ], + "score": 1.0, + "content": "Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 501, + 494, + 515 + ], + "spans": [ + { + "bbox": [ + 115, + 501, + 494, + 515 + ], + "score": 1.0, + "content": "and Michael Auli. fairseq: A fast, extensible toolkit for sequence modeling. In NAACL, 2019.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 489, + 506, + 515 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 519, + 505, + 542 + ], + "lines": [ + { + "bbox": [ + 105, + 518, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 533 + ], + "score": 1.0, + "content": "Martin Popel and Ondˇrej Bojar. Training tips for the transformer model. The Prague Bulletin of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 530, + 307, + 543 + ], + "spans": [ + { + "bbox": [ + 116, + 530, + 307, + 543 + ], + "score": 1.0, + "content": "Mathematical Linguistics, 110(1):43–70, 2018.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 518, + 506, + 543 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 547, + 504, + 571 + ], + "lines": [ + { + "bbox": [ + 105, + 547, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 506, + 561 + ], + "score": 1.0, + "content": "Sashank J Reddi, Satyen Kale, and Sanjiv Kumar. On the convergence of adam and beyond. In", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 559, + 168, + 571 + ], + "spans": [ + { + "bbox": [ + 115, + 559, + 168, + 571 + ], + "score": 1.0, + "content": "ICLR, 2018.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 547, + 506, + 571 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 576, + 504, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 575, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 591 + ], + "score": 1.0, + "content": "Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 587, + 371, + 600 + ], + "spans": [ + { + "bbox": [ + 115, + 587, + 371, + 600 + ], + "score": 1.0, + "content": "the inception architecture for computer vision. In CVPR, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 575, + 506, + 600 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 605, + 503, + 629 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 619 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 616, + 431, + 629 + ], + "spans": [ + { + "bbox": [ + 115, + 616, + 431, + 629 + ], + "score": 1.0, + "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 605, + 505, + 629 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 634, + 447, + 646 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 446, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 446, + 647 + ], + "score": 1.0, + "content": "Kirk M Wolter. Taylor series methods. In Introduction to variance estimation. 2007.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 633, + 446, + 647 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 652, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 106, + 652, + 505, + 665 + ], + "spans": [ + { + "bbox": [ + 106, + 652, + 505, + 665 + ], + "score": 1.0, + "content": "Lin Xiao, Adams Wei Yu, Qihang Lin, and Weizhu Chen. Dscovr: Randomized primal-dual block", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 663, + 491, + 675 + ], + "spans": [ + { + "bbox": [ + 115, + 663, + 491, + 675 + ], + "score": 1.0, + "content": "coordinate algorithms for asynchronous distributed optimization. J. Mach. Learn. Res., 2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 106, + 652, + 505, + 675 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 680, + 504, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 506, + 694 + ], + "score": 1.0, + "content": "Matthew D Zeiler. Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 114, + 690, + 144, + 704 + ], + "spans": [ + { + "bbox": [ + 114, + 690, + 144, + 704 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 678, + 506, + 704 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Hongyi Zhang, Yann N Dauphin, and Tengyu Ma. Fixup initialization: Residual learning without", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 721, + 241, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 241, + 732 + ], + "score": 1.0, + "content": "normalization. In ICLR, 2019.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 244, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 245, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 245, + 96 + ], + "score": 1.0, + "content": "A PROOF OF THEOREM 1", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 104, + 104, + 473, + 118 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 474, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 226, + 119 + ], + "score": 1.0, + "content": "For ease of notation, we refer", + "type": "text" + }, + { + "bbox": [ + 226, + 105, + 249, + 118 + ], + "score": 0.92, + "content": "\\psi ^ { 2 } ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 104, + 261, + 119 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 261, + 108, + 268, + 116 + ], + "score": 0.74, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 104, + 286, + 119 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 287, + 105, + 299, + 118 + ], + "score": 0.89, + "content": "\\textstyle { \\frac { 1 } { \\sigma ^ { 2 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 104, + 311, + 119 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 311, + 105, + 322, + 116 + ], + "score": 0.86, + "content": "\\tau ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 104, + 351, + 119 + ], + "score": 1.0, + "content": ". Thus,", + "type": "text" + }, + { + "bbox": [ + 351, + 107, + 369, + 116 + ], + "score": 0.37, + "content": "x \\sim", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 104, + 411, + 119 + ], + "score": 1.0, + "content": "Scale-inv-", + "type": "text" + }, + { + "bbox": [ + 411, + 105, + 452, + 118 + ], + "score": 0.93, + "content": "\\mathcal { X } ^ { 2 } ( \\rho , \\tau ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 104, + 474, + 119 + ], + "score": 1.0, + "content": "and:", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "interline_equation", + "bbox": [ + 166, + 119, + 445, + 148 + ], + "lines": [ + { + "bbox": [ + 166, + 119, + 445, + 148 + ], + "spans": [ + { + "bbox": [ + 166, + 119, + 445, + 148 + ], + "score": 0.89, + "content": "p ( x ) = \\frac { ( \\tau ^ { 2 } \\rho / 2 ) ^ { \\rho / 2 } } { \\Gamma ( \\rho / 2 ) } \\frac { \\exp \\bigl [ \\frac { - \\rho \\tau ^ { 2 } } { 2 x } \\bigr ] } { x ^ { 1 + \\rho / 2 } } \\quad \\mathrm { a n d } \\quad \\mathbb { E } [ x ] = \\frac { \\rho } { ( \\rho - 2 ) \\sigma ^ { 2 } } \\left( \\forall \\rho > 2 \\right)", + "type": "interline_equation", + "image_path": "a68306e481a2e257390d61e1e82b658f0b513631373ef01210937ac2b15d1d5c.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 166, + 119, + 445, + 148 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 147, + 328, + 158 + ], + "lines": [ + { + "bbox": [ + 106, + 146, + 329, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 146, + 133, + 160 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 147, + 151, + 159 + ], + "score": 0.9, + "content": "\\Gamma ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 146, + 329, + 160 + ], + "score": 1.0, + "content": "is the gamma function. Therefore, we have:", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "interline_equation", + "bbox": [ + 184, + 157, + 426, + 186 + ], + "lines": [ + { + "bbox": [ + 184, + 157, + 426, + 186 + ], + "spans": [ + { + "bbox": [ + 184, + 157, + 426, + 186 + ], + "score": 0.92, + "content": "\\mathbb { E } [ \\sqrt { x } ] = \\int _ { 0 } ^ { \\infty } \\sqrt { x } p ( x ) d x = \\frac { \\tau \\sqrt { \\rho } \\Gamma ( ( \\rho - 1 ) / 2 ) } { \\sqrt { 2 } \\Gamma ( \\rho / 2 ) } ( \\forall \\rho > 4 ) .", + "type": "interline_equation", + "image_path": "f6a97576169a32c6305a6ee627ea0b484b90eb44fab15f8c21938439fed9c40e.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 184, + 157, + 426, + 186 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 185, + 306, + 196 + ], + "lines": [ + { + "bbox": [ + 106, + 184, + 306, + 197 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 232, + 197 + ], + "score": 1.0, + "content": "Based on Equation 6 and 7, for", + "type": "text" + }, + { + "bbox": [ + 232, + 185, + 264, + 196 + ], + "score": 0.91, + "content": "\\forall \\rho > 4", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 184, + 306, + 197 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "interline_equation", + "bbox": [ + 135, + 194, + 475, + 221 + ], + "lines": [ + { + "bbox": [ + 135, + 194, + 475, + 221 + ], + "spans": [ + { + "bbox": [ + 135, + 194, + 475, + 221 + ], + "score": 0.92, + "content": "\\operatorname { V a r } [ \\psi ( \\cdot ) ] = \\operatorname { V a r } [ { \\sqrt { x } } ] = \\operatorname { \\mathbb { E } } [ x ] - \\operatorname { \\mathbb { E } } [ { \\sqrt { x } } ] ^ { 2 } = \\tau ^ { 2 } ( { \\frac { \\rho } { \\rho - 2 } } - { \\frac { \\rho 2 ^ { 2 \\rho - 5 } } { \\pi } } B ( { \\frac { \\rho - 1 } { 2 } } , { \\frac { \\rho - 1 } { 2 } } ) ^ { 2 } ) ,", + "type": "interline_equation", + "image_path": "4731c0000ffbb79e62b7117d5e904f2d8871c0e1860b8538b72d763ea671e1c1.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 135, + 194, + 475, + 221 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 221, + 494, + 233 + ], + "lines": [ + { + "bbox": [ + 105, + 218, + 495, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 133, + 234 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 221, + 151, + 233 + ], + "score": 0.91, + "content": "B ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 218, + 379, + 234 + ], + "score": 1.0, + "content": "is the beta function. To prove the monotonic property of", + "type": "text" + }, + { + "bbox": [ + 380, + 221, + 419, + 233 + ], + "score": 0.92, + "content": "\\mathrm { V a r } [ \\psi ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 218, + 495, + 234 + ], + "score": 1.0, + "content": ", we need to show:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 343, + 255 + ], + "lines": [ + { + "bbox": [ + 101, + 235, + 342, + 257 + ], + "spans": [ + { + "bbox": [ + 101, + 235, + 168, + 257 + ], + "score": 1.0, + "content": "Lemma 1. for", + "type": "text" + }, + { + "bbox": [ + 168, + 239, + 191, + 252 + ], + "score": 0.66, + "content": "t \\geq 4", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 237, + 342, + 255 + ], + "score": 0.71, + "content": "\\begin{array} { r } { , \\frac { \\partial } { \\partial t } \\big ( \\frac { t } { t - 2 } - \\frac { t 2 ^ { 2 t - 5 } } { \\pi } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } \\big ) < 0 } \\end{array}", + "type": "inline_equation" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 265, + 296, + 276 + ], + "lines": [ + { + "bbox": [ + 105, + 263, + 298, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 298, + 279 + ], + "score": 1.0, + "content": "Proof. The target inequality can be re-wrote as", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 154, + 275, + 455, + 371 + ], + "lines": [ + { + "bbox": [ + 154, + 275, + 455, + 371 + ], + "spans": [ + { + "bbox": [ + 154, + 275, + 455, + 371 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\frac { \\partial } { \\partial t } ( \\frac { t } { t - 2 } - \\frac { t \\dot { 2 } ^ { 2 t - 5 } } { \\pi } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } ) } \\\\ & { = \\frac { - 2 } { ( t - 2 ) ^ { 2 } } - \\frac { 2 ^ { 2 t - 5 } } { \\pi } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } - \\frac { t 2 ^ { 2 t - 5 } \\ln { 4 } } { \\pi } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } \\\\ & { - \\frac { 2 t 2 ^ { 2 t - 5 } } { \\pi } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } ( \\Psi ( \\frac { t - 1 } { 2 } ) - \\Psi ( t - 1 ) ) , \\quad \\left( \\Psi ( x ) = \\frac { \\Gamma ^ { \\prime } ( x ) } { \\Gamma ( x ) } \\right) } \\\\ & { < 0 } \\end{array}", + "type": "interline_equation", + "image_path": "933b90ff22dd689badad34a6560dc3d45725acedc2e44a4efa8630b010dc48e9.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 154, + 275, + 455, + 307.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 154, + 307.0, + 455, + 339.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 154, + 339.0, + 455, + 371.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 234, + 397 + ], + "lines": [ + { + "bbox": [ + 106, + 383, + 235, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 235, + 399 + ], + "score": 1.0, + "content": "This inequality is equivalent to:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "interline_equation", + "bbox": [ + 200, + 395, + 410, + 447 + ], + "lines": [ + { + "bbox": [ + 200, + 395, + 410, + 447 + ], + "spans": [ + { + "bbox": [ + 200, + 395, + 410, + 447 + ], + "score": 0.88, + "content": "\\begin{array} { r l r } & { } & { \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } + 1 + t \\ln 4 + 2 t \\Psi ( \\frac { t - 1 } { 2 } ) } \\\\ & { } & { > 2 t \\Psi ( t - 1 ) \\overset { ( i ) } { = } t [ \\Psi ( \\frac { t - 1 } { 2 } ) + \\Psi ( \\frac { t } { 2 } ) + \\ln 4 ] , } \\end{array}", + "type": "interline_equation", + "image_path": "a2a70ac3e31572d07fa579594e5e8c99c30ff3fd34366528b34a910690d7a0da.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 200, + 395, + 410, + 412.3333333333333 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 200, + 412.3333333333333, + 410, + 429.66666666666663 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 200, + 429.66666666666663, + 410, + 446.99999999999994 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 446, + 486, + 457 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 488, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 133, + 460 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 446, + 145, + 458 + ], + "score": 0.84, + "content": "( i )", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 443, + 488, + 460 + ], + "score": 1.0, + "content": "is derived from Legendre duplication formula. Simplify the above inequality, we get:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 189, + 456, + 421, + 483 + ], + "lines": [ + { + "bbox": [ + 189, + 456, + 421, + 483 + ], + "spans": [ + { + "bbox": [ + 189, + 456, + 421, + 483 + ], + "score": 0.88, + "content": "\\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } + 1 + t \\Psi ( \\frac { t - \\hat { 1 } } { 2 } ) - t \\Psi ( \\frac { t } { 2 } ) > 0 ,", + "type": "interline_equation", + "image_path": "373a41f5932e3ffef8ce1ac227a41b1b954c78ade3aa31f36219c7646156a52d.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 189, + 456, + 421, + 483 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 488, + 196, + 499 + ], + "lines": [ + { + "bbox": [ + 106, + 486, + 198, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 198, + 501 + ], + "score": 1.0, + "content": "We only need to show", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "interline_equation", + "bbox": [ + 155, + 497, + 457, + 620 + ], + "lines": [ + { + "bbox": [ + 155, + 497, + 457, + 620 + ], + "spans": [ + { + "bbox": [ + 155, + 497, + 457, + 620 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } + 1 + t \\Psi ( \\frac { t - 1 } { 2 } ) - t \\Psi ( \\frac { t } { 2 } ) } \\\\ & { \\geq \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } + 2 + t ( \\ln ( t / 2 ) - 1 / ( t / 2 - 0 . 5 ) ) - t \\ln ( t / 2 ) } \\\\ & { = \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } - \\frac { 2 } { t - 1 } } \\\\ & { > \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } - \\frac { 2 } { t - 2 } \\geq 0 , } \\\\ & { \\mathrm { r s t i n e q u a l i t y i s i f r o m } \\ln ( x ) - 1 / ( 2 x ) > \\Psi ( x ) > \\ln ( x + 0 . 5 ) - 1 / x . } \\end{array}", + "type": "interline_equation", + "image_path": "daac562cbae9f2bac8ef57763e4772f95a5f145921747210700cb77d99e8242b.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 155, + 497, + 457, + 538.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 155, + 538.0, + 457, + 579.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 155, + 579.0, + 457, + 620.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 608, + 428, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 155, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 155, + 622 + ], + "score": 1.0, + "content": "where the fi", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 625, + 240, + 636 + ], + "lines": [ + { + "bbox": [ + 106, + 624, + 239, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 239, + 637 + ], + "score": 1.0, + "content": "Therefore, we only need to show", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 637, + 374, + 658 + ], + "lines": [ + { + "bbox": [ + 235, + 637, + 374, + 658 + ], + "spans": [ + { + "bbox": [ + 235, + 637, + 374, + 658 + ], + "score": 0.92, + "content": "3 2 \\pi \\geq ( t - 2 ) 4 ^ { t } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } ,", + "type": "interline_equation", + "image_path": "a8b52e9f9e5400edde8f198fd0d634374bb233b34ebc0185c2fb2cecbd3c17a0.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 235, + 637, + 374, + 658 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 663, + 196, + 674 + ], + "lines": [ + { + "bbox": [ + 106, + 662, + 197, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 197, + 676 + ], + "score": 1.0, + "content": "which is equivalent to", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 675, + 426, + 732 + ], + "lines": [ + { + "bbox": [ + 185, + 675, + 426, + 732 + ], + "spans": [ + { + "bbox": [ + 185, + 675, + 426, + 732 + ], + "score": 0.91, + "content": "\\begin{array} { l } { ( t - 2 ) 4 ^ { t } \\mathcal { B } ( \\displaystyle \\frac { t - 1 } { 2 } , \\displaystyle \\frac { t - 1 } { 2 } ) ^ { 2 } = ( t - 2 ) 4 ^ { t } \\frac { \\Gamma ( \\displaystyle \\frac { t - 1 } { 2 } ) ^ { 4 } } { \\Gamma ( t - 1 ) ^ { 2 } } } \\\\ { \\overset { ( i ) } { = } ( t - 2 ) 4 ^ { t } \\frac { \\Gamma ( \\displaystyle \\frac { t - 1 } { 2 } ) ^ { 2 } } { \\Gamma ( t / 2 ) ^ { 2 } } 4 ^ { 2 - t } \\pi = 1 6 \\pi ( t - 2 ) \\frac { \\Gamma ( \\displaystyle \\frac { t - 1 } { 2 } ) ^ { 2 } } { \\Gamma ( t / 2 ) ^ { 2 } } \\le 3 2 \\pi , } \\end{array}", + "type": "interline_equation", + "image_path": "9c064b4557cec0de11192a2e3afddf9644489286614bd06d38e5f5cee88ba410.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 185, + 675, + 426, + 694.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 185, + 694.0, + 426, + 713.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 185, + 713.0, + 426, + 732.0 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 244, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 245, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 245, + 96 + ], + "score": 1.0, + "content": "A PROOF OF THEOREM 1", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 104, + 104, + 473, + 118 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 474, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 226, + 119 + ], + "score": 1.0, + "content": "For ease of notation, we refer", + "type": "text" + }, + { + "bbox": [ + 226, + 105, + 249, + 118 + ], + "score": 0.92, + "content": "\\psi ^ { 2 } ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 104, + 261, + 119 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 261, + 108, + 268, + 116 + ], + "score": 0.74, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 104, + 286, + 119 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 287, + 105, + 299, + 118 + ], + "score": 0.89, + "content": "\\textstyle { \\frac { 1 } { \\sigma ^ { 2 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 104, + 311, + 119 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 311, + 105, + 322, + 116 + ], + "score": 0.86, + "content": "\\tau ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 104, + 351, + 119 + ], + "score": 1.0, + "content": ". Thus,", + "type": "text" + }, + { + "bbox": [ + 351, + 107, + 369, + 116 + ], + "score": 0.37, + "content": "x \\sim", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 104, + 411, + 119 + ], + "score": 1.0, + "content": "Scale-inv-", + "type": "text" + }, + { + "bbox": [ + 411, + 105, + 452, + 118 + ], + "score": 0.93, + "content": "\\mathcal { X } ^ { 2 } ( \\rho , \\tau ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 104, + 474, + 119 + ], + "score": 1.0, + "content": "and:", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 104, + 474, + 119 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 166, + 119, + 445, + 148 + ], + "lines": [ + { + "bbox": [ + 166, + 119, + 445, + 148 + ], + "spans": [ + { + "bbox": [ + 166, + 119, + 445, + 148 + ], + "score": 0.89, + "content": "p ( x ) = \\frac { ( \\tau ^ { 2 } \\rho / 2 ) ^ { \\rho / 2 } } { \\Gamma ( \\rho / 2 ) } \\frac { \\exp \\bigl [ \\frac { - \\rho \\tau ^ { 2 } } { 2 x } \\bigr ] } { x ^ { 1 + \\rho / 2 } } \\quad \\mathrm { a n d } \\quad \\mathbb { E } [ x ] = \\frac { \\rho } { ( \\rho - 2 ) \\sigma ^ { 2 } } \\left( \\forall \\rho > 2 \\right)", + "type": "interline_equation", + "image_path": "a68306e481a2e257390d61e1e82b658f0b513631373ef01210937ac2b15d1d5c.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 166, + 119, + 445, + 148 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 147, + 328, + 158 + ], + "lines": [ + { + "bbox": [ + 106, + 146, + 329, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 146, + 133, + 160 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 147, + 151, + 159 + ], + "score": 0.9, + "content": "\\Gamma ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 146, + 329, + 160 + ], + "score": 1.0, + "content": "is the gamma function. Therefore, we have:", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3, + "bbox_fs": [ + 106, + 146, + 329, + 160 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 184, + 157, + 426, + 186 + ], + "lines": [ + { + "bbox": [ + 184, + 157, + 426, + 186 + ], + "spans": [ + { + "bbox": [ + 184, + 157, + 426, + 186 + ], + "score": 0.92, + "content": "\\mathbb { E } [ \\sqrt { x } ] = \\int _ { 0 } ^ { \\infty } \\sqrt { x } p ( x ) d x = \\frac { \\tau \\sqrt { \\rho } \\Gamma ( ( \\rho - 1 ) / 2 ) } { \\sqrt { 2 } \\Gamma ( \\rho / 2 ) } ( \\forall \\rho > 4 ) .", + "type": "interline_equation", + "image_path": "f6a97576169a32c6305a6ee627ea0b484b90eb44fab15f8c21938439fed9c40e.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 184, + 157, + 426, + 186 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 185, + 306, + 196 + ], + "lines": [ + { + "bbox": [ + 106, + 184, + 306, + 197 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 232, + 197 + ], + "score": 1.0, + "content": "Based on Equation 6 and 7, for", + "type": "text" + }, + { + "bbox": [ + 232, + 185, + 264, + 196 + ], + "score": 0.91, + "content": "\\forall \\rho > 4", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 184, + 306, + 197 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5, + "bbox_fs": [ + 106, + 184, + 306, + 197 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 135, + 194, + 475, + 221 + ], + "lines": [ + { + "bbox": [ + 135, + 194, + 475, + 221 + ], + "spans": [ + { + "bbox": [ + 135, + 194, + 475, + 221 + ], + "score": 0.92, + "content": "\\operatorname { V a r } [ \\psi ( \\cdot ) ] = \\operatorname { V a r } [ { \\sqrt { x } } ] = \\operatorname { \\mathbb { E } } [ x ] - \\operatorname { \\mathbb { E } } [ { \\sqrt { x } } ] ^ { 2 } = \\tau ^ { 2 } ( { \\frac { \\rho } { \\rho - 2 } } - { \\frac { \\rho 2 ^ { 2 \\rho - 5 } } { \\pi } } B ( { \\frac { \\rho - 1 } { 2 } } , { \\frac { \\rho - 1 } { 2 } } ) ^ { 2 } ) ,", + "type": "interline_equation", + "image_path": "4731c0000ffbb79e62b7117d5e904f2d8871c0e1860b8538b72d763ea671e1c1.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 135, + 194, + 475, + 221 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 221, + 494, + 233 + ], + "lines": [ + { + "bbox": [ + 105, + 218, + 495, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 133, + 234 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 221, + 151, + 233 + ], + "score": 0.91, + "content": "B ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 218, + 379, + 234 + ], + "score": 1.0, + "content": "is the beta function. To prove the monotonic property of", + "type": "text" + }, + { + "bbox": [ + 380, + 221, + 419, + 233 + ], + "score": 0.92, + "content": "\\mathrm { V a r } [ \\psi ( . ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 218, + 495, + 234 + ], + "score": 1.0, + "content": ", we need to show:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 218, + 495, + 234 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 343, + 255 + ], + "lines": [ + { + "bbox": [ + 101, + 235, + 342, + 257 + ], + "spans": [ + { + "bbox": [ + 101, + 235, + 168, + 257 + ], + "score": 1.0, + "content": "Lemma 1. for", + "type": "text" + }, + { + "bbox": [ + 168, + 239, + 191, + 252 + ], + "score": 0.66, + "content": "t \\geq 4", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 237, + 342, + 255 + ], + "score": 0.71, + "content": "\\begin{array} { r } { , \\frac { \\partial } { \\partial t } \\big ( \\frac { t } { t - 2 } - \\frac { t 2 ^ { 2 t - 5 } } { \\pi } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } \\big ) < 0 } \\end{array}", + "type": "inline_equation" + } + ], + "index": 8 + } + ], + "index": 8, + "bbox_fs": [ + 101, + 235, + 342, + 257 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 265, + 296, + 276 + ], + "lines": [ + { + "bbox": [ + 105, + 263, + 298, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 298, + 279 + ], + "score": 1.0, + "content": "Proof. The target inequality can be re-wrote as", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 263, + 298, + 279 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 154, + 275, + 455, + 371 + ], + "lines": [ + { + "bbox": [ + 154, + 275, + 455, + 371 + ], + "spans": [ + { + "bbox": [ + 154, + 275, + 455, + 371 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\frac { \\partial } { \\partial t } ( \\frac { t } { t - 2 } - \\frac { t \\dot { 2 } ^ { 2 t - 5 } } { \\pi } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } ) } \\\\ & { = \\frac { - 2 } { ( t - 2 ) ^ { 2 } } - \\frac { 2 ^ { 2 t - 5 } } { \\pi } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } - \\frac { t 2 ^ { 2 t - 5 } \\ln { 4 } } { \\pi } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } \\\\ & { - \\frac { 2 t 2 ^ { 2 t - 5 } } { \\pi } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } ( \\Psi ( \\frac { t - 1 } { 2 } ) - \\Psi ( t - 1 ) ) , \\quad \\left( \\Psi ( x ) = \\frac { \\Gamma ^ { \\prime } ( x ) } { \\Gamma ( x ) } \\right) } \\\\ & { < 0 } \\end{array}", + "type": "interline_equation", + "image_path": "933b90ff22dd689badad34a6560dc3d45725acedc2e44a4efa8630b010dc48e9.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 154, + 275, + 455, + 307.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 154, + 307.0, + 455, + 339.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 154, + 339.0, + 455, + 371.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 234, + 397 + ], + "lines": [ + { + "bbox": [ + 106, + 383, + 235, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 235, + 399 + ], + "score": 1.0, + "content": "This inequality is equivalent to:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 383, + 235, + 399 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 200, + 395, + 410, + 447 + ], + "lines": [ + { + "bbox": [ + 200, + 395, + 410, + 447 + ], + "spans": [ + { + "bbox": [ + 200, + 395, + 410, + 447 + ], + "score": 0.88, + "content": "\\begin{array} { r l r } & { } & { \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } + 1 + t \\ln 4 + 2 t \\Psi ( \\frac { t - 1 } { 2 } ) } \\\\ & { } & { > 2 t \\Psi ( t - 1 ) \\overset { ( i ) } { = } t [ \\Psi ( \\frac { t - 1 } { 2 } ) + \\Psi ( \\frac { t } { 2 } ) + \\ln 4 ] , } \\end{array}", + "type": "interline_equation", + "image_path": "a2a70ac3e31572d07fa579594e5e8c99c30ff3fd34366528b34a910690d7a0da.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 200, + 395, + 410, + 412.3333333333333 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 200, + 412.3333333333333, + 410, + 429.66666666666663 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 200, + 429.66666666666663, + 410, + 446.99999999999994 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 446, + 486, + 457 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 488, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 133, + 460 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 446, + 145, + 458 + ], + "score": 0.84, + "content": "( i )", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 443, + 488, + 460 + ], + "score": 1.0, + "content": "is derived from Legendre duplication formula. Simplify the above inequality, we get:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 443, + 488, + 460 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 189, + 456, + 421, + 483 + ], + "lines": [ + { + "bbox": [ + 189, + 456, + 421, + 483 + ], + "spans": [ + { + "bbox": [ + 189, + 456, + 421, + 483 + ], + "score": 0.88, + "content": "\\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } + 1 + t \\Psi ( \\frac { t - \\hat { 1 } } { 2 } ) - t \\Psi ( \\frac { t } { 2 } ) > 0 ,", + "type": "interline_equation", + "image_path": "373a41f5932e3ffef8ce1ac227a41b1b954c78ade3aa31f36219c7646156a52d.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 189, + 456, + 421, + 483 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 488, + 196, + 499 + ], + "lines": [ + { + "bbox": [ + 106, + 486, + 198, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 198, + 501 + ], + "score": 1.0, + "content": "We only need to show", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19, + "bbox_fs": [ + 106, + 486, + 198, + 501 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 155, + 497, + 457, + 620 + ], + "lines": [ + { + "bbox": [ + 155, + 497, + 457, + 620 + ], + "spans": [ + { + "bbox": [ + 155, + 497, + 457, + 620 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } + 1 + t \\Psi ( \\frac { t - 1 } { 2 } ) - t \\Psi ( \\frac { t } { 2 } ) } \\\\ & { \\geq \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } + 2 + t ( \\ln ( t / 2 ) - 1 / ( t / 2 - 0 . 5 ) ) - t \\ln ( t / 2 ) } \\\\ & { = \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } - \\frac { 2 } { t - 1 } } \\\\ & { > \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } - \\frac { 2 } { t - 2 } \\geq 0 , } \\\\ & { \\mathrm { r s t i n e q u a l i t y i s i f r o m } \\ln ( x ) - 1 / ( 2 x ) > \\Psi ( x ) > \\ln ( x + 0 . 5 ) - 1 / x . } \\end{array}", + "type": "interline_equation", + "image_path": "daac562cbae9f2bac8ef57763e4772f95a5f145921747210700cb77d99e8242b.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 155, + 497, + 457, + 538.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 155, + 538.0, + 457, + 579.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 155, + 579.0, + 457, + 620.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 608, + 428, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 155, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 155, + 622 + ], + "score": 1.0, + "content": "where the fi", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 606, + 155, + 622 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 625, + 240, + 636 + ], + "lines": [ + { + "bbox": [ + 106, + 624, + 239, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 239, + 637 + ], + "score": 1.0, + "content": "Therefore, we only need to show", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 106, + 624, + 239, + 637 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 637, + 374, + 658 + ], + "lines": [ + { + "bbox": [ + 235, + 637, + 374, + 658 + ], + "spans": [ + { + "bbox": [ + 235, + 637, + 374, + 658 + ], + "score": 0.92, + "content": "3 2 \\pi \\geq ( t - 2 ) 4 ^ { t } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } ,", + "type": "interline_equation", + "image_path": "a8b52e9f9e5400edde8f198fd0d634374bb233b34ebc0185c2fb2cecbd3c17a0.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 235, + 637, + 374, + 658 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 663, + 196, + 674 + ], + "lines": [ + { + "bbox": [ + 106, + 662, + 197, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 197, + 676 + ], + "score": 1.0, + "content": "which is equivalent to", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 106, + 662, + 197, + 676 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 675, + 426, + 732 + ], + "lines": [ + { + "bbox": [ + 185, + 675, + 426, + 732 + ], + "spans": [ + { + "bbox": [ + 185, + 675, + 426, + 732 + ], + "score": 0.91, + "content": "\\begin{array} { l } { ( t - 2 ) 4 ^ { t } \\mathcal { B } ( \\displaystyle \\frac { t - 1 } { 2 } , \\displaystyle \\frac { t - 1 } { 2 } ) ^ { 2 } = ( t - 2 ) 4 ^ { t } \\frac { \\Gamma ( \\displaystyle \\frac { t - 1 } { 2 } ) ^ { 4 } } { \\Gamma ( t - 1 ) ^ { 2 } } } \\\\ { \\overset { ( i ) } { = } ( t - 2 ) 4 ^ { t } \\frac { \\Gamma ( \\displaystyle \\frac { t - 1 } { 2 } ) ^ { 2 } } { \\Gamma ( t / 2 ) ^ { 2 } } 4 ^ { 2 - t } \\pi = 1 6 \\pi ( t - 2 ) \\frac { \\Gamma ( \\displaystyle \\frac { t - 1 } { 2 } ) ^ { 2 } } { \\Gamma ( t / 2 ) ^ { 2 } } \\le 3 2 \\pi , } \\end{array}", + "type": "interline_equation", + "image_path": "9c064b4557cec0de11192a2e3afddf9644489286614bd06d38e5f5cee88ba410.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 185, + 675, + 426, + 694.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 185, + 694.0, + 426, + 713.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 185, + 713.0, + 426, + 732.0 + ], + "spans": [], + "index": 29 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 300, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 301, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 301, + 96 + ], + "score": 1.0, + "content": "where (i) is from Legendre duplication formula.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 99, + 208, + 111 + ], + "lines": [ + { + "bbox": [ + 106, + 98, + 208, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 98, + 208, + 113 + ], + "score": 1.0, + "content": "So we only need to show", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "interline_equation", + "bbox": [ + 262, + 108, + 349, + 137 + ], + "lines": [ + { + "bbox": [ + 262, + 108, + 349, + 137 + ], + "spans": [ + { + "bbox": [ + 262, + 108, + 349, + 137 + ], + "score": 0.94, + "content": "( t - 2 ) \\frac { \\Gamma ( \\frac { t - 1 } { 2 } ) ^ { 2 } } { \\Gamma ( t / 2 ) ^ { 2 } } \\leq 2", + "type": "interline_equation", + "image_path": "d56e8f3aad70b54c98bf6133479504e8efa59f49a3d2ddac5840de4e4efcfe5b.jpg" + } + ] + } + ], + "index": 2.5, + "virtual_lines": [ + { + "bbox": [ + 262, + 108, + 349, + 122.5 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 262, + 122.5, + 349, + 137.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 141, + 352, + 157 + ], + "lines": [ + { + "bbox": [ + 102, + 136, + 353, + 164 + ], + "spans": [ + { + "bbox": [ + 102, + 136, + 220, + 164 + ], + "score": 1.0, + "content": "Using Gautschi’s inequality", + "type": "text" + }, + { + "bbox": [ + 221, + 141, + 313, + 158 + ], + "score": 0.91, + "content": "( \\frac { \\Gamma ( x + 1 ) } { \\Gamma ( x + s ) } < ( x + 1 ) ^ { 1 - s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 136, + 353, + 164 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "interline_equation", + "bbox": [ + 197, + 157, + 415, + 185 + ], + "lines": [ + { + "bbox": [ + 197, + 157, + 415, + 185 + ], + "spans": [ + { + "bbox": [ + 197, + 157, + 415, + 185 + ], + "score": 0.89, + "content": "( t - 2 ) \\frac { \\dot { \\Gamma } ( \\frac { t - 1 } { 2 } ) ^ { 2 } } { \\Gamma ( t / 2 ) ^ { 2 } } \\leq ( t - 2 ) ( \\frac { t - 1 } { 2 } ) ^ { - 1 } = \\frac { 2 ( t - 2 ) } { t - 1 } < 2", + "type": "interline_equation", + "image_path": "f18994052fffccd4c8bab55617eb816a201046261b1cb00d5677631ff50f31c9.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 197, + 157, + 415, + 171.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 197, + 171.0, + 415, + 185.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "title", + "bbox": [ + 108, + 209, + 269, + 222 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 270, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 270, + 225 + ], + "score": 1.0, + "content": "B IMPLEMENTATION DETAILS", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "title", + "bbox": [ + 107, + 234, + 233, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 232, + 234, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 234, + 246 + ], + "score": 1.0, + "content": "B.1 LANGUAGE MODELING", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 254, + 505, + 364 + ], + "lines": [ + { + "bbox": [ + 106, + 254, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 267 + ], + "score": 1.0, + "content": "Our implementation is based on the previous work (Liu et al., 2018). Specifically, we use two-layer", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 265, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 505, + 278 + ], + "score": 1.0, + "content": "LSTMs with 2048 hidden states with adaptive softmax to conduct experiments on the one billion", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "words dataset. Word embedding (random initialized) of 300 dimensions is used as the input and the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 287, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 505, + 300 + ], + "score": 1.0, + "content": "adaptive softmax is incorporated with a default setting (cut-offs are set to [4000, 40000, 200000]).", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 298, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 312 + ], + "score": 1.0, + "content": "Additionally, as pre-processing, we replace all tokens occurring equal or less than 3 times with as", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "UNK, which shrinks the dictionary from 7.9M to 6.4M. Dropout is applied to each layer with a ratio", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 319, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 506, + 334 + ], + "score": 1.0, + "content": "of 0.1, gradients are clipped at 5.0. We use the default hyper-parameters to update moving averages,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 107, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 107, + 331, + 156, + 343 + ], + "score": 0.89, + "content": "i . e . \\beta _ { 1 } = 0 . 9", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 331, + 174, + 344 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 174, + 331, + 221, + 343 + ], + "score": 0.91, + "content": "\\beta _ { 2 } = 0 . 9 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 331, + 506, + 344 + ], + "score": 1.0, + "content": ". The learning rate is set to start from 0.001, and decayed at the start of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 355 + ], + "score": 1.0, + "content": "10th epochs. LSTMs are unrolled for 20 steps without resetting the LSTM states and the batch size", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 353, + 392, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 392, + 365 + ], + "score": 1.0, + "content": "is set to 128. All models are trained on one NVIDIA Tesla V100 GPU.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 13.5 + }, + { + "type": "title", + "bbox": [ + 108, + 378, + 251, + 389 + ], + "lines": [ + { + "bbox": [ + 106, + 377, + 252, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 252, + 390 + ], + "score": 1.0, + "content": "B.2 IMAGEINE CLASSIFICATION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 398, + 505, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "We use the default ResNet architectures (He et al., 2016) in a public pytorch re-implementation4.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 409, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 506, + 421 + ], + "score": 1.0, + "content": "Specifically, we use 20-layer ResNet (9 Basic Blocks) for CIFAR-10 and 18-layer ResNet (8 Basic", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 420, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 505, + 432 + ], + "score": 1.0, + "content": "Blocks) for ImageNet. Batch size is 128 for CIFAR-10 and 256 for ImageNet. The model is trained", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 431, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 505, + 443 + ], + "score": 1.0, + "content": "for 186 epoches and the learning rate decays at the 81-th and the 122-th epoches by 0.1 on CIFAR-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "10, while the model is trained for 90 epoches and the learning rate decays at the 31-th and the 61-th", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 349, + 466 + ], + "score": 1.0, + "content": "epoch by 0.1 on ImageNet. For Adam and RAdam, we set", + "type": "text" + }, + { + "bbox": [ + 349, + 453, + 442, + 464 + ], + "score": 0.9, + "content": "\\beta _ { 1 } = 0 . 9 , \\beta _ { 2 } = 0 . 9 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 452, + 505, + 466 + ], + "score": 1.0, + "content": ". For SGD, we", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 464, + 504, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 349, + 476 + ], + "score": 1.0, + "content": "set the momentum factor as 0.9. The weight decay rate is", + "type": "text" + }, + { + "bbox": [ + 349, + 464, + 371, + 474 + ], + "score": 0.89, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 464, + 504, + 476 + ], + "score": 1.0, + "content": ". Random cropping and random", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 475, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 295, + 487 + ], + "score": 1.0, + "content": "horizontal flipping are applied to training data.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + }, + { + "type": "title", + "bbox": [ + 108, + 499, + 278, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 280, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 280, + 511 + ], + "score": 1.0, + "content": "B.3 NEURAL MACHINE TRANSLATION", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 520, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 520, + 504, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 504, + 532 + ], + "score": 1.0, + "content": "Our experiments are based on the default Transformers (Vaswani et al., 2017) implementation from", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "the fairseq package (Ott et al., 2019). Specifically, we use word embedding with 512 dimensions and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "score": 1.0, + "content": "6-layer encoder / decoder with 4 head and 1024 hidden dimensions on the IWSLT14’ dataset; use", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 553, + 504, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 504, + 564 + ], + "score": 1.0, + "content": "word embedding with 512 dimension and 6-layer encoder / decoder with 8 heads and 2048 hidden", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 493, + 577 + ], + "score": 1.0, + "content": "dimensions. Label smoothed cross entropy is used as the objective function with an uncertainty", + "type": "text" + }, + { + "bbox": [ + 494, + 565, + 505, + 574 + ], + "score": 0.77, + "content": "=", + "type": "inline_equation" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 399, + 586 + ], + "score": 1.0, + "content": "0.1 (Szegedy et al., 2016). We use linear learning rate decay starting from", + "type": "text" + }, + { + "bbox": [ + 399, + 574, + 420, + 585 + ], + "score": 0.9, + "content": "3 e ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 574, + 505, + 586 + ], + "score": 1.0, + "content": ", and the checkpoints", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "of the last 20 epoches are averaged before evaluation. As to the wamrup strategy, we use a linear", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 317, + 609 + ], + "score": 1.0, + "content": "warmup for Adam in the first 4000 updates, and set", + "type": "text" + }, + { + "bbox": [ + 318, + 597, + 329, + 608 + ], + "score": 0.88, + "content": "\\beta _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 596, + 370, + 609 + ], + "score": 1.0, + "content": "to satisfy", + "type": "text" + }, + { + "bbox": [ + 370, + 597, + 412, + 607 + ], + "score": 0.87, + "content": "\\nu = 4 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 596, + 416, + 609 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 416, + 597, + 472, + 608 + ], + "score": 0.86, + "content": "\\beta _ { 2 } = 0 . 9 9 9 5 )", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 596, + 505, + 609 + ], + "score": 1.0, + "content": ". In the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "IWSLT’14 dataset, we conduct training on one NVIDIA Tesla V100 GPU, set maximum batch size", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 619, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 631 + ], + "score": 1.0, + "content": "as 4000, apply dropout with a ratio 0.3, using weight decay of 0.0001 and clip the gradient norm", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 629, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 506, + 641 + ], + "score": 1.0, + "content": "at 25. In the WMT’16 dataset, we conduct training on four NVIDIA Quadro R8000 GPUs and set", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 641, + 227, + 652 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 227, + 652 + ], + "score": 1.0, + "content": "maximum batch size as 8196.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34.5 + }, + { + "type": "title", + "bbox": [ + 108, + 667, + 266, + 680 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 268, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 268, + 682 + ], + "score": 1.0, + "content": "C DOWNGRADING TO SGDM", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 109, + 692, + 504, + 714 + ], + "lines": [ + { + "bbox": [ + 107, + 691, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 107, + 691, + 505, + 704 + ], + "score": 1.0, + "content": "As a byproduct determined by math derivations, we degenerated RAdam to SGD with momentum", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 701, + 504, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 701, + 504, + 717 + ], + "score": 1.0, + "content": "in the first several updates. Although this stage only contains several gradient updates, these up-", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 117, + 722, + 389, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 720, + 390, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 720, + 390, + 734 + ], + "score": 1.0, + "content": "4https://github.com/bearpaw/pytorch-classification", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 183, + 504, + 193 + ], + "lines": [ + { + "bbox": [ + 496, + 185, + 504, + 192 + ], + "spans": [ + { + "bbox": [ + 496, + 185, + 504, + 192 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 300, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 301, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 301, + 96 + ], + "score": 1.0, + "content": "where (i) is from Legendre duplication formula.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 106, + 81, + 301, + 96 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 99, + 208, + 111 + ], + "lines": [ + { + "bbox": [ + 106, + 98, + 208, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 98, + 208, + 113 + ], + "score": 1.0, + "content": "So we only need to show", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 98, + 208, + 113 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 262, + 108, + 349, + 137 + ], + "lines": [ + { + "bbox": [ + 262, + 108, + 349, + 137 + ], + "spans": [ + { + "bbox": [ + 262, + 108, + 349, + 137 + ], + "score": 0.94, + "content": "( t - 2 ) \\frac { \\Gamma ( \\frac { t - 1 } { 2 } ) ^ { 2 } } { \\Gamma ( t / 2 ) ^ { 2 } } \\leq 2", + "type": "interline_equation", + "image_path": "d56e8f3aad70b54c98bf6133479504e8efa59f49a3d2ddac5840de4e4efcfe5b.jpg" + } + ] + } + ], + "index": 2.5, + "virtual_lines": [ + { + "bbox": [ + 262, + 108, + 349, + 122.5 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 262, + 122.5, + 349, + 137.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 141, + 352, + 157 + ], + "lines": [ + { + "bbox": [ + 102, + 136, + 353, + 164 + ], + "spans": [ + { + "bbox": [ + 102, + 136, + 220, + 164 + ], + "score": 1.0, + "content": "Using Gautschi’s inequality", + "type": "text" + }, + { + "bbox": [ + 221, + 141, + 313, + 158 + ], + "score": 0.91, + "content": "( \\frac { \\Gamma ( x + 1 ) } { \\Gamma ( x + s ) } < ( x + 1 ) ^ { 1 - s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 136, + 353, + 164 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 102, + 136, + 353, + 164 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 197, + 157, + 415, + 185 + ], + "lines": [ + { + "bbox": [ + 197, + 157, + 415, + 185 + ], + "spans": [ + { + "bbox": [ + 197, + 157, + 415, + 185 + ], + "score": 0.89, + "content": "( t - 2 ) \\frac { \\dot { \\Gamma } ( \\frac { t - 1 } { 2 } ) ^ { 2 } } { \\Gamma ( t / 2 ) ^ { 2 } } \\leq ( t - 2 ) ( \\frac { t - 1 } { 2 } ) ^ { - 1 } = \\frac { 2 ( t - 2 ) } { t - 1 } < 2", + "type": "interline_equation", + "image_path": "f18994052fffccd4c8bab55617eb816a201046261b1cb00d5677631ff50f31c9.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 197, + 157, + 415, + 171.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 197, + 171.0, + 415, + 185.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "title", + "bbox": [ + 108, + 209, + 269, + 222 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 270, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 270, + 225 + ], + "score": 1.0, + "content": "B IMPLEMENTATION DETAILS", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "title", + "bbox": [ + 107, + 234, + 233, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 232, + 234, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 234, + 246 + ], + "score": 1.0, + "content": "B.1 LANGUAGE MODELING", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 254, + 505, + 364 + ], + "lines": [ + { + "bbox": [ + 106, + 254, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 267 + ], + "score": 1.0, + "content": "Our implementation is based on the previous work (Liu et al., 2018). Specifically, we use two-layer", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 265, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 505, + 278 + ], + "score": 1.0, + "content": "LSTMs with 2048 hidden states with adaptive softmax to conduct experiments on the one billion", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "words dataset. Word embedding (random initialized) of 300 dimensions is used as the input and the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 287, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 505, + 300 + ], + "score": 1.0, + "content": "adaptive softmax is incorporated with a default setting (cut-offs are set to [4000, 40000, 200000]).", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 298, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 312 + ], + "score": 1.0, + "content": "Additionally, as pre-processing, we replace all tokens occurring equal or less than 3 times with as", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "UNK, which shrinks the dictionary from 7.9M to 6.4M. Dropout is applied to each layer with a ratio", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 319, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 506, + 334 + ], + "score": 1.0, + "content": "of 0.1, gradients are clipped at 5.0. We use the default hyper-parameters to update moving averages,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 107, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 107, + 331, + 156, + 343 + ], + "score": 0.89, + "content": "i . e . \\beta _ { 1 } = 0 . 9", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 331, + 174, + 344 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 174, + 331, + 221, + 343 + ], + "score": 0.91, + "content": "\\beta _ { 2 } = 0 . 9 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 331, + 506, + 344 + ], + "score": 1.0, + "content": ". The learning rate is set to start from 0.001, and decayed at the start of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 355 + ], + "score": 1.0, + "content": "10th epochs. LSTMs are unrolled for 20 steps without resetting the LSTM states and the batch size", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 353, + 392, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 392, + 365 + ], + "score": 1.0, + "content": "is set to 128. All models are trained on one NVIDIA Tesla V100 GPU.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 254, + 506, + 365 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 378, + 251, + 389 + ], + "lines": [ + { + "bbox": [ + 106, + 377, + 252, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 252, + 390 + ], + "score": 1.0, + "content": "B.2 IMAGEINE CLASSIFICATION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 398, + 505, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "We use the default ResNet architectures (He et al., 2016) in a public pytorch re-implementation4.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 409, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 506, + 421 + ], + "score": 1.0, + "content": "Specifically, we use 20-layer ResNet (9 Basic Blocks) for CIFAR-10 and 18-layer ResNet (8 Basic", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 420, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 505, + 432 + ], + "score": 1.0, + "content": "Blocks) for ImageNet. Batch size is 128 for CIFAR-10 and 256 for ImageNet. The model is trained", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 431, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 505, + 443 + ], + "score": 1.0, + "content": "for 186 epoches and the learning rate decays at the 81-th and the 122-th epoches by 0.1 on CIFAR-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "10, while the model is trained for 90 epoches and the learning rate decays at the 31-th and the 61-th", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 349, + 466 + ], + "score": 1.0, + "content": "epoch by 0.1 on ImageNet. For Adam and RAdam, we set", + "type": "text" + }, + { + "bbox": [ + 349, + 453, + 442, + 464 + ], + "score": 0.9, + "content": "\\beta _ { 1 } = 0 . 9 , \\beta _ { 2 } = 0 . 9 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 452, + 505, + 466 + ], + "score": 1.0, + "content": ". For SGD, we", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 464, + 504, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 349, + 476 + ], + "score": 1.0, + "content": "set the momentum factor as 0.9. The weight decay rate is", + "type": "text" + }, + { + "bbox": [ + 349, + 464, + 371, + 474 + ], + "score": 0.89, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 464, + 504, + 476 + ], + "score": 1.0, + "content": ". Random cropping and random", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 475, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 295, + 487 + ], + "score": 1.0, + "content": "horizontal flipping are applied to training data.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 397, + 506, + 487 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 499, + 278, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 280, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 280, + 511 + ], + "score": 1.0, + "content": "B.3 NEURAL MACHINE TRANSLATION", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 520, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 520, + 504, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 504, + 532 + ], + "score": 1.0, + "content": "Our experiments are based on the default Transformers (Vaswani et al., 2017) implementation from", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "the fairseq package (Ott et al., 2019). Specifically, we use word embedding with 512 dimensions and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "score": 1.0, + "content": "6-layer encoder / decoder with 4 head and 1024 hidden dimensions on the IWSLT14’ dataset; use", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 553, + 504, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 504, + 564 + ], + "score": 1.0, + "content": "word embedding with 512 dimension and 6-layer encoder / decoder with 8 heads and 2048 hidden", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 493, + 577 + ], + "score": 1.0, + "content": "dimensions. Label smoothed cross entropy is used as the objective function with an uncertainty", + "type": "text" + }, + { + "bbox": [ + 494, + 565, + 505, + 574 + ], + "score": 0.77, + "content": "=", + "type": "inline_equation" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 399, + 586 + ], + "score": 1.0, + "content": "0.1 (Szegedy et al., 2016). We use linear learning rate decay starting from", + "type": "text" + }, + { + "bbox": [ + 399, + 574, + 420, + 585 + ], + "score": 0.9, + "content": "3 e ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 574, + 505, + 586 + ], + "score": 1.0, + "content": ", and the checkpoints", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "of the last 20 epoches are averaged before evaluation. As to the wamrup strategy, we use a linear", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 317, + 609 + ], + "score": 1.0, + "content": "warmup for Adam in the first 4000 updates, and set", + "type": "text" + }, + { + "bbox": [ + 318, + 597, + 329, + 608 + ], + "score": 0.88, + "content": "\\beta _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 596, + 370, + 609 + ], + "score": 1.0, + "content": "to satisfy", + "type": "text" + }, + { + "bbox": [ + 370, + 597, + 412, + 607 + ], + "score": 0.87, + "content": "\\nu = 4 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 596, + 416, + 609 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 416, + 597, + 472, + 608 + ], + "score": 0.86, + "content": "\\beta _ { 2 } = 0 . 9 9 9 5 )", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 596, + 505, + 609 + ], + "score": 1.0, + "content": ". In the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "IWSLT’14 dataset, we conduct training on one NVIDIA Tesla V100 GPU, set maximum batch size", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 619, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 631 + ], + "score": 1.0, + "content": "as 4000, apply dropout with a ratio 0.3, using weight decay of 0.0001 and clip the gradient norm", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 629, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 506, + 641 + ], + "score": 1.0, + "content": "at 25. In the WMT’16 dataset, we conduct training on four NVIDIA Quadro R8000 GPUs and set", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 641, + 227, + 652 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 227, + 652 + ], + "score": 1.0, + "content": "maximum batch size as 8196.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 520, + 506, + 652 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 667, + 266, + 680 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 268, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 268, + 682 + ], + "score": 1.0, + "content": "C DOWNGRADING TO SGDM", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 109, + 692, + 504, + 714 + ], + "lines": [ + { + "bbox": [ + 107, + 691, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 107, + 691, + 505, + 704 + ], + "score": 1.0, + "content": "As a byproduct determined by math derivations, we degenerated RAdam to SGD with momentum", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 701, + 504, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 701, + 504, + 717 + ], + "score": 1.0, + "content": "in the first several updates. Although this stage only contains several gradient updates, these up-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "dates could be quite damaging (e.g., in our Figure 2, the gradient distribution is distorted within 10", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "gradient updates). Intuitively, updates with divergent adaptive learning rate variance could be more", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "damaging than the ones with converged variance, as divergent variance implies more instability. As", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "a case study, we performed experiments on the CIFAR10 dataset. Five-run average results are sum-", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 104, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "marized in Table 3. The optimizer fails to get an equally reliably model when changing the first", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "4 updates to Adam, yet the influence of switching is less deleterious when we change 5-8 updates", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "instead. This result verifies our intuition and is in agreement with our theory the first few updates", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 506, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 506, + 172 + ], + "score": 1.0, + "content": "could be more damaging than later updates. By saying that, we still want to emphasize that this part", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 170, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 505, + 182 + ], + "score": 1.0, + "content": "(downgrading to SGDM) is only a minor part of our algorithm design whereas our main focus is on", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 180, + 388, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 388, + 194 + ], + "score": 1.0, + "content": "the mechanism of warmup and the derivation of the rectification term.", + "type": "text", + "cross_page": true + } + ], + "index": 9 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 691, + 505, + 717 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "dates could be quite damaging (e.g., in our Figure 2, the gradient distribution is distorted within 10", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "gradient updates). Intuitively, updates with divergent adaptive learning rate variance could be more", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "damaging than the ones with converged variance, as divergent variance implies more instability. As", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "a case study, we performed experiments on the CIFAR10 dataset. Five-run average results are sum-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 104, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "marized in Table 3. The optimizer fails to get an equally reliably model when changing the first", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "4 updates to Adam, yet the influence of switching is less deleterious when we change 5-8 updates", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "instead. This result verifies our intuition and is in agreement with our theory the first few updates", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 506, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 506, + 172 + ], + "score": 1.0, + "content": "could be more damaging than later updates. By saying that, we still want to emphasize that this part", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 170, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 505, + 182 + ], + "score": 1.0, + "content": "(downgrading to SGDM) is only a minor part of our algorithm design whereas our main focus is on", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 180, + 388, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 388, + 194 + ], + "score": 1.0, + "content": "the mechanism of warmup and the derivation of the rectification term.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5 + }, + { + "type": "table", + "bbox": [ + 106, + 228, + 505, + 307 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 214, + 203, + 396, + 215 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 213, + 201, + 397, + 217 + ], + "spans": [ + { + "bbox": [ + 213, + 201, + 358, + 217 + ], + "score": 1.0, + "content": "Table 3: Performance on CIFAR10", + "type": "text" + }, + { + "bbox": [ + 358, + 203, + 391, + 214 + ], + "score": 0.75, + "content": "( \\mathrm { l r } = 0 . 1 ", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 201, + 397, + 217 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "table_body", + "bbox": [ + 106, + 228, + 505, + 307 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 228, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 505, + 307 + ], + "score": 0.982, + "html": "
1-4 steps5-8 steps8+ stepstest acctrain losstrain error
RAdamRAdamRAdam91.080.0210.74
Adam (w. divergent var.)RAdamRAdam89.980.0602.12
SGDAdam (w. convergent var.)RAdam90.290.0381.23
", + "type": "table", + "image_path": "ad11f900d73d2167f23839df7c4d72b8c5485331e01abbf9ed891a61f460d72f.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 106, + 228, + 505, + 254.33333333333334 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 106, + 254.33333333333334, + 505, + 280.6666666666667 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 106, + 280.6666666666667, + 505, + 307.0 + ], + "spans": [], + "index": 13 + } + ] + } + ], + "index": 11.0 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 192 + ], + "lines": [], + "index": 4.5, + "bbox_fs": [ + 104, + 82, + 506, + 194 + ], + "lines_deleted": true + }, + { + "type": "table", + "bbox": [ + 106, + 228, + 505, + 307 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 214, + 203, + 396, + 215 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 213, + 201, + 397, + 217 + ], + "spans": [ + { + "bbox": [ + 213, + 201, + 358, + 217 + ], + "score": 1.0, + "content": "Table 3: Performance on CIFAR10", + "type": "text" + }, + { + "bbox": [ + 358, + 203, + 391, + 214 + ], + "score": 0.75, + "content": "( \\mathrm { l r } = 0 . 1 ", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 201, + 397, + 217 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "table_body", + "bbox": [ + 106, + 228, + 505, + 307 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 228, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 505, + 307 + ], + "score": 0.982, + "html": "
1-4 steps5-8 steps8+ stepstest acctrain losstrain error
RAdamRAdamRAdam91.080.0210.74
Adam (w. divergent var.)RAdamRAdam89.980.0602.12
SGDAdam (w. convergent var.)RAdam90.290.0381.23
", + "type": "table", + "image_path": "ad11f900d73d2167f23839df7c4d72b8c5485331e01abbf9ed891a61f460d72f.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 106, + 228, + 505, + 254.33333333333334 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 106, + 254.33333333333334, + 505, + 280.6666666666667 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 106, + 280.6666666666667, + 505, + 307.0 + ], + "spans": [], + "index": 13 + } + ] + } + ], + "index": 11.0 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/rkgz2aEKDr/rkgz2aEKDr_model.json b/parse/train/rkgz2aEKDr/rkgz2aEKDr_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4403df3a190d48770cd3c706acb75308a30aa2d8 --- /dev/null +++ b/parse/train/rkgz2aEKDr/rkgz2aEKDr_model.json @@ -0,0 +1,24231 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 300, + 1237, + 989, + 1237, + 989, + 1603, + 300, + 1603 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 398, + 762, + 1303, + 762, + 1303, + 1099, + 398, + 1099 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1619, + 1404, + 1619, + 1404, + 1834, + 298, + 1834 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 1014, + 1215, + 1399, + 1215, + 1399, + 1499, + 1014, + 1499 + ], + "score": 0.965 + }, + { + "category_id": 0, + "poly": [ + 299, + 221, + 1402, + 221, + 1402, + 322, + 299, + 322 + ], + "score": 0.959 + }, + { + "category_id": 4, + "poly": [ + 1010, + 1508, + 1403, + 1508, + 1403, + 1599, + 1010, + 1599 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 314, + 377, + 736, + 377, + 736, + 467, + 314, + 467 + ], + "score": 0.943 + }, + { + "category_id": 2, + "poly": [ + 331, + 1944, + 1312, + 1944, + 1312, + 2034, + 331, + 2034 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 1038, + 377, + 1385, + 377, + 1385, + 469, + 1038, + 469 + ], + "score": 0.937 + }, + { + "category_id": 1, + "poly": [ + 298, + 1849, + 1400, + 1849, + 1400, + 1911, + 298, + 1911 + ], + "score": 0.937 + }, + { + "category_id": 1, + "poly": [ + 781, + 376, + 993, + 376, + 993, + 468, + 781, + 468 + ], + "score": 0.937 + }, + { + "category_id": 0, + "poly": [ + 302, + 1171, + 572, + 1171, + 572, + 1206, + 302, + 1206 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 773, + 689, + 927, + 689, + 927, + 722, + 773, + 722 + ], + "score": 0.893 + }, + { + "category_id": 1, + "poly": [ + 313, + 516, + 652, + 516, + 652, + 608, + 313, + 608 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 857, + 2088, + 857, + 2112, + 841, + 2112 + ], + "score": 0.729 + }, + { + "category_id": 1, + "poly": [ + 822, + 517, + 1244, + 517, + 1244, + 608, + 822, + 608 + ], + "score": 0.716 + }, + { + "category_id": 0, + "poly": [ + 823, + 518, + 959, + 518, + 959, + 546, + 823, + 546 + ], + "score": 0.12 + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 1235.0, + 1059.0, + 1235.0, + 1059.0, + 1427.0, + 1036.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 1285.0, + 1036.0, + 1285.0, + 1036.0, + 1359.0, + 1016.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 1343.0, + 1131.0, + 1343.0, + 1131.0, + 1353.0, + 1122.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 1358.0, + 1147.0, + 1358.0, + 1147.0, + 1383.0, + 1077.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 1418.0, + 1069.0, + 1418.0, + 1069.0, + 1436.0, + 1053.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 1414.0, + 1402.0, + 1414.0, + 1402.0, + 1439.0, + 1089.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 1450.0, + 1402.0, + 1450.0, + 1402.0, + 1475.0, + 1017.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 1474.0, + 1186.0, + 1474.0, + 1186.0, + 1502.0, + 1087.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 1474.0, + 1367.0, + 1474.0, + 1367.0, + 1504.0, + 1209.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 215.0, + 1404.0, + 215.0, + 1404.0, + 273.0, + 294.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 273.0, + 709.0, + 273.0, + 709.0, + 328.0, + 294.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 1506.0, + 1408.0, + 1506.0, + 1408.0, + 1541.0, + 1009.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 1538.0, + 1404.0, + 1538.0, + 1404.0, + 1570.0, + 1009.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 1566.0, + 1308.0, + 1566.0, + 1308.0, + 1600.0, + 1009.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1939.0, + 842.0, + 1939.0, + 842.0, + 1976.0, + 327.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1968.0, + 844.0, + 1968.0, + 844.0, + 2008.0, + 328.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1998.0, + 1312.0, + 1998.0, + 1312.0, + 2039.0, + 328.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1168.0, + 579.0, + 1168.0, + 579.0, + 1215.0, + 293.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 685.0, + 934.0, + 685.0, + 934.0, + 728.0, + 768.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 840.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 514.0, + 963.0, + 514.0, + 963.0, + 548.0, + 817.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1236.0, + 992.0, + 1236.0, + 992.0, + 1271.0, + 295.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1264.0, + 992.0, + 1264.0, + 992.0, + 1304.0, + 294.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1297.0, + 991.0, + 1297.0, + 991.0, + 1333.0, + 296.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1329.0, + 992.0, + 1329.0, + 992.0, + 1362.0, + 296.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1358.0, + 992.0, + 1358.0, + 992.0, + 1394.0, + 295.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1390.0, + 992.0, + 1390.0, + 992.0, + 1423.0, + 295.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1420.0, + 992.0, + 1420.0, + 992.0, + 1455.0, + 293.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1447.0, + 993.0, + 1447.0, + 993.0, + 1487.0, + 293.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1480.0, + 992.0, + 1480.0, + 992.0, + 1514.0, + 294.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1509.0, + 992.0, + 1509.0, + 992.0, + 1544.0, + 294.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1541.0, + 992.0, + 1541.0, + 992.0, + 1575.0, + 295.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1573.0, + 763.0, + 1573.0, + 763.0, + 1606.0, + 296.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 761.0, + 1305.0, + 761.0, + 1305.0, + 800.0, + 394.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 793.0, + 1305.0, + 793.0, + 1305.0, + 832.0, + 393.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 823.0, + 1304.0, + 823.0, + 1304.0, + 861.0, + 392.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 855.0, + 1305.0, + 855.0, + 1305.0, + 890.0, + 393.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 885.0, + 1308.0, + 885.0, + 1308.0, + 921.0, + 394.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 919.0, + 1305.0, + 919.0, + 1305.0, + 948.0, + 396.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 946.0, + 1307.0, + 946.0, + 1307.0, + 982.0, + 393.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 979.0, + 1304.0, + 979.0, + 1304.0, + 1011.0, + 395.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1005.0, + 1305.0, + 1005.0, + 1305.0, + 1044.0, + 393.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1034.0, + 1304.0, + 1034.0, + 1304.0, + 1074.0, + 393.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1069.0, + 703.0, + 1069.0, + 703.0, + 1099.0, + 395.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1619.0, + 1402.0, + 1619.0, + 1402.0, + 1654.0, + 294.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1650.0, + 1406.0, + 1650.0, + 1406.0, + 1685.0, + 294.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1680.0, + 1405.0, + 1680.0, + 1405.0, + 1714.0, + 293.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1712.0, + 1404.0, + 1712.0, + 1404.0, + 1742.0, + 294.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1740.0, + 1405.0, + 1740.0, + 1405.0, + 1777.0, + 293.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1773.0, + 1404.0, + 1773.0, + 1404.0, + 1804.0, + 296.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1803.0, + 999.0, + 1803.0, + 999.0, + 1837.0, + 294.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 377.0, + 481.0, + 377.0, + 481.0, + 407.0, + 309.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 406.0, + 736.0, + 406.0, + 736.0, + 442.0, + 312.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 441.0, + 461.0, + 441.0, + 461.0, + 469.0, + 312.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 379.0, + 1386.0, + 379.0, + 1386.0, + 411.0, + 1039.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 410.0, + 1330.0, + 410.0, + 1330.0, + 441.0, + 1038.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 442.0, + 1372.0, + 442.0, + 1372.0, + 473.0, + 1041.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1845.0, + 1404.0, + 1845.0, + 1404.0, + 1888.0, + 294.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 1404.0, + 1880.0, + 1404.0, + 1913.0, + 293.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 375.0, + 970.0, + 375.0, + 970.0, + 412.0, + 778.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 409.0, + 921.0, + 409.0, + 921.0, + 438.0, + 781.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 443.0, + 995.0, + 443.0, + 995.0, + 469.0, + 781.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 517.0, + 653.0, + 517.0, + 653.0, + 549.0, + 314.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 549.0, + 517.0, + 549.0, + 517.0, + 577.0, + 313.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 583.0, + 646.0, + 583.0, + 646.0, + 611.0, + 315.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 516.0, + 963.0, + 516.0, + 963.0, + 548.0, + 821.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 544.0, + 1245.0, + 544.0, + 1245.0, + 581.0, + 820.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 583.0, + 980.0, + 583.0, + 980.0, + 608.0, + 822.0, + 608.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 475, + 1404, + 475, + 1404, + 630, + 298, + 630 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 307, + 1403, + 307, + 1403, + 460, + 298, + 460 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 297, + 1712, + 1403, + 1712, + 1403, + 1898, + 297, + 1898 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 297, + 953, + 1403, + 953, + 1403, + 1108, + 297, + 1108 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 298, + 1911, + 1401, + 1911, + 1401, + 2034, + 298, + 2034 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 292, + 229, + 1402, + 229, + 1402, + 293, + 292, + 293 + ], + "score": 0.95 + }, + { + "category_id": 0, + "poly": [ + 297, + 887, + 860, + 887, + 860, + 923, + 297, + 923 + ], + "score": 0.918 + }, + { + "category_id": 1, + "poly": [ + 299, + 646, + 848, + 646, + 848, + 677, + 299, + 677 + ], + "score": 0.916 + }, + { + "category_id": 1, + "poly": [ + 300, + 1201, + 1403, + 1201, + 1403, + 1300, + 300, + 1300 + ], + "score": 0.915 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 105, + 300, + 105 + ], + "score": 0.901 + }, + { + "category_id": 1, + "poly": [ + 273, + 1636, + 413, + 1636, + 413, + 1666, + 273, + 1666 + ], + "score": 0.847 + }, + { + "category_id": 1, + "poly": [ + 298, + 681, + 1405, + 681, + 1405, + 846, + 298, + 846 + ], + "score": 0.817 + }, + { + "category_id": 9, + "poly": [ + 1367, + 1138, + 1400, + 1138, + 1400, + 1168, + 1367, + 1168 + ], + "score": 0.782 + }, + { + "category_id": 1, + "poly": [ + 296, + 1337, + 1324, + 1337, + 1324, + 1371, + 296, + 1371 + ], + "score": 0.766 + }, + { + "category_id": 8, + "poly": [ + 311, + 1111, + 1342, + 1111, + 1342, + 1196, + 311, + 1196 + ], + "score": 0.762 + }, + { + "category_id": 1, + "poly": [ + 290, + 1379, + 1314, + 1379, + 1314, + 1443, + 290, + 1443 + ], + "score": 0.744 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.706 + }, + { + "category_id": 1, + "poly": [ + 275, + 1468, + 1210, + 1468, + 1210, + 1629, + 275, + 1629 + ], + "score": 0.438 + }, + { + "category_id": 1, + "poly": [ + 295, + 1445, + 677, + 1445, + 677, + 1473, + 295, + 1473 + ], + "score": 0.359 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.173 + }, + { + "category_id": 1, + "poly": [ + 311, + 1111, + 1342, + 1111, + 1342, + 1196, + 311, + 1196 + ], + "score": 0.093 + }, + { + "category_id": 13, + "poly": [ + 299, + 1233, + 557, + 1233, + 557, + 1300, + 299, + 1300 + ], + "score": 0.92, + "latex": "\\frac { \\sqrt { 1 - \\beta _ { 2 } ^ { t } } } { \\epsilon + \\sqrt { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } }" + }, + { + "category_id": 14, + "poly": [ + 316, + 1105, + 1348, + 1105, + 1348, + 1198, + 316, + 1198 + ], + "score": 0.92, + "latex": "\\phi ( g _ { 1 } , \\cdots , g _ { t } ) = { \\frac { \\dot { ( 1 - \\beta _ { 1 } ) } \\sum _ { i = 1 } ^ { t } \\beta _ { 1 } ^ { t - i } g _ { t } } { 1 - \\beta _ { 1 } ^ { t } } } \\quad \\mathrm { a n d } \\quad \\psi ( g _ { 1 } , \\cdots , g _ { t } ) = { \\sqrt { \\frac { 1 - \\beta _ { 2 } ^ { t } } { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } } } ." + }, + { + "category_id": 13, + "poly": [ + 962, + 1775, + 1079, + 1775, + 1079, + 1804, + 962, + 1804 + ], + "score": 0.92, + "latex": "\\alpha _ { t } = t \\alpha _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 412, + 1774, + 571, + 1774, + 571, + 1806, + 412, + 1806 + ], + "score": 0.92, + "latex": "\\forall t \\alpha _ { t + 1 } \\leq \\alpha _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1151, + 1774, + 1241, + 1774, + 1241, + 1804, + 1151, + 1804 + ], + "score": 0.91, + "latex": "t < T _ { w }" + }, + { + "category_id": 13, + "poly": [ + 1115, + 1016, + 1163, + 1016, + 1163, + 1050, + 1115, + 1050 + ], + "score": 0.91, + "latex": "\\phi ( . )" + }, + { + "category_id": 13, + "poly": [ + 981, + 1047, + 1031, + 1047, + 1031, + 1081, + 981, + 1081 + ], + "score": 0.91, + "latex": "\\psi ( . )" + }, + { + "category_id": 13, + "poly": [ + 1214, + 1017, + 1265, + 1017, + 1265, + 1050, + 1214, + 1050 + ], + "score": 0.9, + "latex": "\\psi ( . )" + }, + { + "category_id": 13, + "poly": [ + 618, + 1414, + 672, + 1414, + 672, + 1445, + 618, + 1445 + ], + "score": 0.9, + "latex": "f ( \\theta )" + }, + { + "category_id": 13, + "poly": [ + 715, + 1202, + 766, + 1202, + 766, + 1236, + 715, + 1236 + ], + "score": 0.9, + "latex": "\\psi ( . )" + }, + { + "category_id": 13, + "poly": [ + 600, + 1377, + 734, + 1377, + 734, + 1414, + 600, + 1414 + ], + "score": 0.9, + "latex": "\\{ \\phi _ { t } , \\psi _ { t } \\} _ { t = 1 } ^ { T }" + }, + { + "category_id": 13, + "poly": [ + 382, + 1377, + 477, + 1377, + 477, + 1414, + 382, + 1414 + ], + "score": 0.9, + "latex": "\\{ \\alpha _ { t } \\} _ { t = 1 } ^ { T }" + }, + { + "category_id": 13, + "poly": [ + 1352, + 1017, + 1400, + 1017, + 1400, + 1050, + 1352, + 1050 + ], + "score": 0.89, + "latex": "\\phi ( . )" + }, + { + "category_id": 13, + "poly": [ + 367, + 1474, + 432, + 1474, + 432, + 1500, + 367, + 1500 + ], + "score": 0.89, + "latex": "t = 1" + }, + { + "category_id": 13, + "poly": [ + 1208, + 1197, + 1404, + 1197, + 1404, + 1237, + 1208, + 1237 + ], + "score": 0.89, + "latex": "\\widehat { \\psi } ( g _ { 1 } , \\cdots , g _ { t } ) =" + }, + { + "category_id": 13, + "poly": [ + 339, + 1503, + 539, + 1503, + 539, + 1535, + 339, + 1535 + ], + "score": 0.86, + "latex": "g _ { t } \\gets \\Delta _ { \\theta } f _ { t } ( \\theta _ { t - 1 } )" + }, + { + "category_id": 13, + "poly": [ + 999, + 1719, + 1030, + 1719, + 1030, + 1743, + 999, + 1743 + ], + "score": 0.86, + "latex": "\\alpha _ { t }" + }, + { + "category_id": 13, + "poly": [ + 340, + 1535, + 580, + 1535, + 580, + 1565, + 340, + 1565 + ], + "score": 0.86, + "latex": "m _ { t } \\gets \\phi _ { t } ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 401, + 1443, + 434, + 1443, + 434, + 1472, + 401, + 1472 + ], + "score": 0.86, + "latex": "\\theta _ { T }" + }, + { + "category_id": 13, + "poly": [ + 381, + 1414, + 409, + 1414, + 409, + 1442, + 381, + 1442 + ], + "score": 0.86, + "latex": "\\theta _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 802, + 1748, + 834, + 1748, + 834, + 1773, + 802, + 1773 + ], + "score": 0.85, + "latex": "\\alpha _ { t }" + }, + { + "category_id": 13, + "poly": [ + 339, + 1566, + 568, + 1566, + 568, + 1596, + 339, + 1596 + ], + "score": 0.83, + "latex": "l _ { t } \\gets \\psi _ { t } ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 339, + 1596, + 571, + 1596, + 571, + 1628, + 339, + 1628 + ], + "score": 0.82, + "latex": "\\theta _ { t } \\gets \\theta _ { t - 1 } - \\alpha _ { t } m _ { t } l _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1140, + 1247, + 1251, + 1247, + 1251, + 1280, + 1140, + 1280 + ], + "score": 0.8, + "latex": "1 \\times 1 0 ^ { - 8 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1081, + 310, + 1081, + 310, + 1105, + 298, + 1105 + ], + "score": 0.78, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 641, + 1256, + 657, + 1256, + 657, + 1277, + 641, + 1277 + ], + "score": 0.77, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 764, + 1051, + 778, + 1051, + 778, + 1075, + 764, + 1075 + ], + "score": 0.74, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 463, + 1474, + 487, + 1474, + 487, + 1500, + 463, + 1500 + ], + "score": 0.73, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 380, + 1635, + 414, + 1635, + 414, + 1667, + 380, + 1667 + ], + "score": 0.68, + "latex": "\\theta _ { T }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 885.0, + 861.0, + 885.0, + 861.0, + 927.0, + 292.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 475.0, + 1404.0, + 475.0, + 1404.0, + 513.0, + 294.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 507.0, + 1405.0, + 507.0, + 1405.0, + 541.0, + 297.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 537.0, + 1405.0, + 537.0, + 1405.0, + 574.0, + 292.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 566.0, + 1406.0, + 566.0, + 1406.0, + 606.0, + 292.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 599.0, + 1092.0, + 599.0, + 1092.0, + 633.0, + 293.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 308.0, + 1402.0, + 308.0, + 1402.0, + 341.0, + 296.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 336.0, + 1405.0, + 336.0, + 1405.0, + 374.0, + 292.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 365.0, + 1407.0, + 365.0, + 1407.0, + 406.0, + 293.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 394.0, + 1407.0, + 394.0, + 1407.0, + 440.0, + 292.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 431.0, + 711.0, + 431.0, + 711.0, + 464.0, + 296.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1710.0, + 998.0, + 1710.0, + 998.0, + 1749.0, + 294.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 1710.0, + 1405.0, + 1710.0, + 1405.0, + 1749.0, + 1031.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1742.0, + 801.0, + 1742.0, + 801.0, + 1778.0, + 295.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 1742.0, + 1407.0, + 1742.0, + 1407.0, + 1778.0, + 835.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1771.0, + 411.0, + 1771.0, + 411.0, + 1810.0, + 293.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 1771.0, + 961.0, + 1771.0, + 961.0, + 1810.0, + 572.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 1771.0, + 1150.0, + 1771.0, + 1150.0, + 1810.0, + 1080.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 1771.0, + 1407.0, + 1771.0, + 1407.0, + 1810.0, + 1242.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1801.0, + 1407.0, + 1801.0, + 1407.0, + 1841.0, + 293.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1834.0, + 1405.0, + 1834.0, + 1405.0, + 1870.0, + 295.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1864.0, + 1383.0, + 1864.0, + 1383.0, + 1901.0, + 294.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 952.0, + 1404.0, + 952.0, + 1404.0, + 991.0, + 295.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 983.0, + 1405.0, + 983.0, + 1405.0, + 1023.0, + 292.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1011.0, + 1114.0, + 1011.0, + 1114.0, + 1056.0, + 292.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 1011.0, + 1213.0, + 1011.0, + 1213.0, + 1056.0, + 1164.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 1011.0, + 1351.0, + 1011.0, + 1351.0, + 1056.0, + 1266.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1011.0, + 1405.0, + 1011.0, + 1405.0, + 1056.0, + 1401.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1045.0, + 763.0, + 1045.0, + 763.0, + 1085.0, + 292.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 1045.0, + 980.0, + 1045.0, + 980.0, + 1085.0, + 779.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1045.0, + 1408.0, + 1045.0, + 1408.0, + 1085.0, + 1032.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1074.0, + 297.0, + 1074.0, + 297.0, + 1114.0, + 291.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 1074.0, + 989.0, + 1074.0, + 989.0, + 1114.0, + 311.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1945.0, + 294.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1943.0, + 1405.0, + 1943.0, + 1405.0, + 1976.0, + 294.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1403.0, + 1971.0, + 1403.0, + 2006.0, + 293.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2004.0, + 1405.0, + 2004.0, + 1405.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 258.0, + 1214.0, + 258.0, + 1214.0, + 298.0, + 294.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 643.0, + 851.0, + 643.0, + 851.0, + 680.0, + 294.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1197.0, + 714.0, + 1197.0, + 714.0, + 1239.0, + 293.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1197.0, + 1207.0, + 1197.0, + 1207.0, + 1239.0, + 767.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 1247.0, + 640.0, + 1247.0, + 640.0, + 1283.0, + 559.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 658.0, + 1247.0, + 1139.0, + 1247.0, + 1139.0, + 1283.0, + 658.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 1247.0, + 1269.0, + 1247.0, + 1269.0, + 1283.0, + 1252.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 1263.0, + 561.0, + 1263.0, + 561.0, + 1306.0, + 558.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 270.0, + 1629.0, + 379.0, + 1629.0, + 379.0, + 1672.0, + 270.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 1629.0, + 418.0, + 1629.0, + 418.0, + 1672.0, + 415.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 678.0, + 1405.0, + 678.0, + 1405.0, + 716.0, + 295.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 708.0, + 1405.0, + 708.0, + 1405.0, + 749.0, + 320.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 740.0, + 1142.0, + 740.0, + 1142.0, + 779.0, + 318.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 780.0, + 1405.0, + 780.0, + 1405.0, + 819.0, + 293.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 812.0, + 1241.0, + 812.0, + 1241.0, + 849.0, + 320.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1335.0, + 1329.0, + 1335.0, + 1329.0, + 1375.0, + 295.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1373.0, + 381.0, + 1373.0, + 381.0, + 1420.0, + 294.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 1373.0, + 599.0, + 1373.0, + 599.0, + 1420.0, + 478.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1373.0, + 1318.0, + 1373.0, + 1318.0, + 1420.0, + 735.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 1411.0, + 617.0, + 1411.0, + 617.0, + 1447.0, + 410.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 1411.0, + 1014.0, + 1411.0, + 1014.0, + 1447.0, + 673.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 272.0, + 1468.0, + 366.0, + 1468.0, + 366.0, + 1504.0, + 272.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 1468.0, + 462.0, + 1468.0, + 462.0, + 1504.0, + 433.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 1468.0, + 528.0, + 1468.0, + 528.0, + 1504.0, + 488.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 274.0, + 1511.0, + 291.0, + 1511.0, + 291.0, + 1532.0, + 274.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 1501.0, + 338.0, + 1501.0, + 338.0, + 1538.0, + 335.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 1501.0, + 1213.0, + 1501.0, + 1213.0, + 1538.0, + 540.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 274.0, + 1540.0, + 291.0, + 1540.0, + 291.0, + 1562.0, + 274.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 1532.0, + 339.0, + 1532.0, + 339.0, + 1568.0, + 336.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 1532.0, + 850.0, + 1532.0, + 850.0, + 1568.0, + 581.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 274.0, + 1572.0, + 292.0, + 1572.0, + 292.0, + 1591.0, + 274.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 1560.0, + 338.0, + 1560.0, + 338.0, + 1601.0, + 333.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 1560.0, + 950.0, + 1560.0, + 950.0, + 1601.0, + 569.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 274.0, + 1602.0, + 291.0, + 1602.0, + 291.0, + 1622.0, + 274.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1590.0, + 338.0, + 1590.0, + 338.0, + 1631.0, + 330.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 1590.0, + 812.0, + 1590.0, + 812.0, + 1631.0, + 572.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1436.0, + 400.0, + 1436.0, + 400.0, + 1481.0, + 296.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1436.0, + 682.0, + 1436.0, + 682.0, + 1481.0, + 435.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 1100.5, + 782.0, + 1100.5, + 782.0, + 1179.0, + 303.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 1129.5, + 1346.0, + 1129.5, + 1346.0, + 1198.5, + 776.0, + 1198.5 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1232, + 1404, + 1232, + 1404, + 1457, + 298, + 1457 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 896, + 1403, + 896, + 1403, + 1020, + 299, + 1020 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1537, + 1404, + 1537, + 1404, + 1660, + 298, + 1660 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 1675, + 1404, + 1675, + 1404, + 1921, + 297, + 1921 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1120, + 1404, + 1120, + 1404, + 1213, + 299, + 1213 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 304, + 172, + 1384, + 172, + 1384, + 424, + 304, + 424 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 405, + 551, + 1283, + 551, + 1283, + 794, + 405, + 794 + ], + "score": 0.968 + }, + { + "category_id": 2, + "poly": [ + 298, + 1947, + 1402, + 1947, + 1402, + 2034, + 298, + 2034 + ], + "score": 0.956 + }, + { + "category_id": 4, + "poly": [ + 297, + 804, + 1403, + 804, + 1403, + 867, + 297, + 867 + ], + "score": 0.938 + }, + { + "category_id": 0, + "poly": [ + 299, + 1058, + 1019, + 1058, + 1019, + 1093, + 299, + 1093 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 301, + 1487, + 797, + 1487, + 797, + 1518, + 301, + 1518 + ], + "score": 0.904 + }, + { + "category_id": 4, + "poly": [ + 296, + 439, + 1405, + 439, + 1405, + 534, + 296, + 534 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 815, + 76, + 815, + 104, + 298, + 104 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.672 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.286 + }, + { + "category_id": 13, + "poly": [ + 1081, + 1266, + 1198, + 1266, + 1198, + 1301, + 1081, + 1301 + ], + "score": 0.94, + "latex": "{ \\mathcal { N } } ( 0 , \\sigma ^ { 2 } ) ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 405, + 1266, + 551, + 1266, + 551, + 1302, + 405, + 1302 + ], + "score": 0.94, + "latex": "\\{ g _ { 1 } , \\cdots , g _ { t } \\}" + }, + { + "category_id": 13, + "poly": [ + 1197, + 1228, + 1393, + 1228, + 1393, + 1268, + 1197, + 1268 + ], + "score": 0.93, + "latex": "\\psi ( g _ { 1 } ) = \\sqrt { 1 / g _ { 1 } ^ { 2 } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1393, + 416, + 1393, + 416, + 1427, + 298, + 1427 + ], + "score": 0.92, + "latex": "\\alpha ^ { 2 } \\mathrm { V a r } [ x ] )" + }, + { + "category_id": 13, + "poly": [ + 1051, + 1302, + 1192, + 1302, + 1192, + 1337, + 1051, + 1337 + ], + "score": 0.92, + "latex": "\\mathcal { X } ^ { 2 } ( 1 , 1 / \\sigma ^ { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 1257, + 1300, + 1400, + 1300, + 1400, + 1338, + 1257, + 1338 + ], + "score": 0.91, + "latex": "\\mathrm { V a r } [ \\sqrt { 1 / g _ { 1 } ^ { 2 } } ]" + }, + { + "category_id": 13, + "poly": [ + 1282, + 1676, + 1333, + 1676, + 1333, + 1709, + 1282, + 1709 + ], + "score": 0.91, + "latex": "\\psi ( . )" + }, + { + "category_id": 13, + "poly": [ + 1005, + 1235, + 1078, + 1235, + 1078, + 1262, + 1005, + 1262 + ], + "score": 0.89, + "latex": "t = 1" + }, + { + "category_id": 13, + "poly": [ + 1343, + 1267, + 1400, + 1267, + 1400, + 1300, + 1343, + 1300 + ], + "score": 0.88, + "latex": "1 / g _ { 1 } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1263, + 1365, + 1403, + 1365, + 1403, + 1398, + 1263, + 1398 + ], + "score": 0.88, + "latex": "\\quad ( { \\mathrm { V a r } } [ \\alpha x ] \\ =" + }, + { + "category_id": 13, + "poly": [ + 947, + 1676, + 1016, + 1676, + 1016, + 1709, + 947, + 1709 + ], + "score": 0.88, + "latex": "( \\psi ( . ) )" + }, + { + "category_id": 13, + "poly": [ + 871, + 1707, + 937, + 1707, + 937, + 1740, + 871, + 1740 + ], + "score": 0.86, + "latex": "( \\phi ( . ) )" + }, + { + "category_id": 13, + "poly": [ + 298, + 988, + 363, + 988, + 363, + 1022, + 298, + 1022 + ], + "score": 0.82, + "latex": "( \\psi ( . ) )" + }, + { + "category_id": 13, + "poly": [ + 1323, + 1570, + 1355, + 1570, + 1355, + 1597, + 1323, + 1597 + ], + "score": 0.82, + "latex": "{ \\it 2 k }" + }, + { + "category_id": 13, + "poly": [ + 1121, + 1708, + 1155, + 1708, + 1155, + 1738, + 1121, + 1738 + ], + "score": 0.44, + "latex": "\\mathbf { \\eta } ^ { ( \\theta ) }" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1799, + 1132, + 1799, + 1132, + 1827, + 1100, + 1827 + ], + "score": 0.37, + "latex": "2 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 1103, + 1678, + 1135, + 1678, + 1135, + 1705, + 1103, + 1705 + ], + "score": 0.34, + "latex": "2 \\mathrm { k }" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 191.0, + 469.0, + 191.0, + 469.0, + 206.0, + 420.0, + 206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 183.0, + 871.0, + 183.0, + 871.0, + 208.0, + 542.0, + 208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 188.0, + 976.0, + 188.0, + 976.0, + 204.0, + 929.0, + 204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 190.0, + 1274.0, + 190.0, + 1274.0, + 206.0, + 1229.0, + 206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 203.0, + 720.0, + 203.0, + 720.0, + 219.0, + 672.0, + 219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 274.0, + 325.0, + 274.0, + 325.0, + 333.0, + 306.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 255.0, + 607.0, + 255.0, + 607.0, + 343.0, + 560.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 264.0, + 855.0, + 264.0, + 855.0, + 325.0, + 834.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 253.0, + 1155.0, + 253.0, + 1155.0, + 293.0, + 1117.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 287.0, + 1133.0, + 287.0, + 1133.0, + 303.0, + 1118.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 308.0, + 1151.0, + 308.0, + 1151.0, + 321.0, + 1135.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 311.0, + 1130.0, + 311.0, + 1130.0, + 320.0, + 1119.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 319.0, + 341.0, + 319.0, + 341.0, + 333.0, + 323.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 319.0, + 871.0, + 319.0, + 871.0, + 333.0, + 851.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 346.0, + 596.0, + 346.0, + 596.0, + 357.0, + 585.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 364.0, + 1149.0, + 364.0, + 1149.0, + 374.0, + 1133.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 371.0, + 344.0, + 371.0, + 344.0, + 385.0, + 324.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 370.0, + 602.0, + 370.0, + 602.0, + 387.0, + 574.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 381.0, + 748.0, + 381.0, + 748.0, + 390.0, + 733.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 374.0, + 871.0, + 374.0, + 871.0, + 387.0, + 849.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 386.0, + 733.0, + 386.0, + 733.0, + 395.0, + 724.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 396.0, + 675.0, + 396.0, + 675.0, + 428.0, + 460.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 396.0, + 1215.0, + 396.0, + 1215.0, + 428.0, + 1030.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 375.5, + 391.0, + 375.5, + 391.0, + 396.5, + 343.0, + 396.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 377.0, + 427.0, + 377.0, + 427.0, + 394.0, + 393.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 376.0, + 477.0, + 376.0, + 477.0, + 394.5, + 440.0, + 394.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 486.0, + 376.5, + 523.0, + 376.5, + 523.0, + 394.5, + 486.0, + 394.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 376.0, + 564.0, + 376.0, + 564.0, + 394.0, + 531.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 377.0, + 649.0, + 377.0, + 649.0, + 397.5, + 601.0, + 397.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 377.5, + 699.0, + 377.5, + 699.0, + 394.5, + 663.0, + 394.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 378.0, + 927.0, + 378.0, + 927.0, + 395.5, + 884.0, + 395.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.75, + 378.0, + 977.75, + 378.0, + 977.75, + 395.5, + 939.75, + 395.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 377.5, + 1032.0, + 377.5, + 1032.0, + 394.5, + 997.0, + 394.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 381.0, + 1081.0, + 381.0, + 1081.0, + 395.0, + 1058.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 378.5, + 1205.0, + 378.5, + 1205.0, + 396.0, + 1164.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 378.0, + 1257.0, + 378.0, + 1257.0, + 396.5, + 1220.0, + 396.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1275.0, + 378.5, + 1311.0, + 378.5, + 1311.0, + 395.5, + 1275.0, + 395.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 562.0, + 564.0, + 562.0, + 564.0, + 577.0, + 507.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 561.0, + 859.0, + 561.0, + 859.0, + 576.0, + 801.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 570.0, + 1127.0, + 570.0, + 1127.0, + 587.0, + 1071.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 637.0, + 685.0, + 637.0, + 685.0, + 645.0, + 671.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 633.0, + 741.0, + 633.0, + 741.0, + 641.0, + 733.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 619.0, + 950.0, + 619.0, + 950.0, + 634.0, + 930.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 626.0, + 978.0, + 626.0, + 978.0, + 645.0, + 963.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 633.0, + 1038.0, + 633.0, + 1038.0, + 642.0, + 1027.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 668.0, + 447.0, + 668.0, + 447.0, + 718.0, + 409.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 669.0, + 745.0, + 669.0, + 745.0, + 718.0, + 706.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 651.0, + 983.0, + 651.0, + 983.0, + 711.0, + 965.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 668.0, + 1044.0, + 668.0, + 1044.0, + 718.0, + 1004.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 661.0, + 1244.0, + 661.0, + 1244.0, + 732.0, + 1183.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 750.0, + 447.0, + 750.0, + 447.0, + 761.0, + 425.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 750.0, + 743.0, + 750.0, + 743.0, + 761.0, + 723.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 750.0, + 1043.0, + 750.0, + 1043.0, + 761.0, + 1021.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 761.0, + 672.0, + 761.0, + 672.0, + 769.0, + 663.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 767.0, + 659.0, + 767.0, + 659.0, + 772.0, + 653.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1227.0, + 765.0, + 1234.0, + 765.0, + 1234.0, + 773.0, + 1227.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 771.0, + 571.0, + 771.0, + 571.0, + 789.0, + 517.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 770.0, + 881.0, + 770.0, + 881.0, + 790.0, + 835.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 756.0, + 560.0, + 756.0, + 560.0, + 774.5, + 521.0, + 774.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 757.0, + 502.0, + 757.0, + 502.0, + 776.0, + 454.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 756.5, + 618.0, + 756.5, + 618.0, + 774.5, + 580.0, + 774.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.75, + 757.0, + 797.75, + 757.0, + 797.75, + 775.0, + 748.75, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 756.5, + 855.0, + 756.5, + 855.0, + 774.5, + 814.0, + 774.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.75, + 755.0, + 915.75, + 755.0, + 915.75, + 775.0, + 875.75, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 762.0, + 969.0, + 762.0, + 969.0, + 774.0, + 942.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.75, + 755.5, + 1092.75, + 755.5, + 1092.75, + 777.0, + 1042.75, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.75, + 756.5, + 1143.75, + 756.5, + 1143.75, + 775.5, + 1103.75, + 775.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.75, + 755.5, + 1199.75, + 755.5, + 1199.75, + 775.5, + 1160.75, + 775.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 772.0, + 1170.0, + 772.0, + 1170.0, + 790.0, + 1113.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1940.0, + 1407.0, + 1940.0, + 1407.0, + 1985.0, + 327.0, + 1985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1975.0, + 1340.0, + 1975.0, + 1340.0, + 2005.0, + 294.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1999.0, + 1406.0, + 1999.0, + 1406.0, + 2040.0, + 328.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 804.0, + 1407.0, + 804.0, + 1407.0, + 839.0, + 296.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 829.0, + 1285.0, + 829.0, + 1285.0, + 872.0, + 294.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1054.0, + 1024.0, + 1054.0, + 1024.0, + 1099.0, + 291.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1484.0, + 803.0, + 1484.0, + 803.0, + 1523.0, + 294.0, + 1523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 440.0, + 1406.0, + 440.0, + 1406.0, + 474.0, + 296.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 472.0, + 1405.0, + 472.0, + 1405.0, + 506.0, + 295.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 502.0, + 1388.0, + 502.0, + 1388.0, + 536.0, + 295.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1229.0, + 1004.0, + 1229.0, + 1004.0, + 1268.0, + 293.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1229.0, + 1196.0, + 1229.0, + 1196.0, + 1268.0, + 1079.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1229.0, + 1402.0, + 1229.0, + 1402.0, + 1268.0, + 1394.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1261.0, + 404.0, + 1261.0, + 404.0, + 1306.0, + 292.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1261.0, + 1080.0, + 1261.0, + 1080.0, + 1306.0, + 552.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 1261.0, + 1342.0, + 1261.0, + 1342.0, + 1306.0, + 1199.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1261.0, + 1404.0, + 1261.0, + 1404.0, + 1306.0, + 1401.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1299.0, + 1050.0, + 1299.0, + 1050.0, + 1342.0, + 291.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 1299.0, + 1256.0, + 1299.0, + 1256.0, + 1342.0, + 1193.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1329.0, + 1404.0, + 1329.0, + 1404.0, + 1371.0, + 291.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1361.0, + 1262.0, + 1361.0, + 1262.0, + 1399.0, + 292.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1389.0, + 297.0, + 1389.0, + 297.0, + 1432.0, + 292.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 1389.0, + 1406.0, + 1389.0, + 1406.0, + 1432.0, + 417.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1426.0, + 1145.0, + 1426.0, + 1145.0, + 1461.0, + 294.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 896.0, + 1403.0, + 896.0, + 1403.0, + 932.0, + 294.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 928.0, + 1405.0, + 928.0, + 1405.0, + 963.0, + 292.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 956.0, + 1404.0, + 956.0, + 1404.0, + 994.0, + 292.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 990.0, + 297.0, + 990.0, + 297.0, + 1023.0, + 293.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 990.0, + 1171.0, + 990.0, + 1171.0, + 1023.0, + 364.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1535.0, + 1405.0, + 1535.0, + 1405.0, + 1573.0, + 292.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1569.0, + 1322.0, + 1569.0, + 1322.0, + 1602.0, + 294.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 1569.0, + 1405.0, + 1569.0, + 1405.0, + 1602.0, + 1356.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1600.0, + 1404.0, + 1600.0, + 1404.0, + 1632.0, + 293.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1630.0, + 869.0, + 1630.0, + 869.0, + 1662.0, + 293.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1674.0, + 946.0, + 1674.0, + 946.0, + 1711.0, + 294.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 1674.0, + 1102.0, + 1674.0, + 1102.0, + 1711.0, + 1017.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1674.0, + 1281.0, + 1674.0, + 1281.0, + 1711.0, + 1136.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1334.0, + 1674.0, + 1405.0, + 1674.0, + 1405.0, + 1711.0, + 1334.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1704.0, + 870.0, + 1704.0, + 870.0, + 1740.0, + 292.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 1704.0, + 1120.0, + 1704.0, + 1120.0, + 1740.0, + 938.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1704.0, + 1406.0, + 1704.0, + 1406.0, + 1740.0, + 1156.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1738.0, + 1404.0, + 1738.0, + 1404.0, + 1772.0, + 295.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1768.0, + 1405.0, + 1768.0, + 1405.0, + 1802.0, + 295.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1796.0, + 1099.0, + 1796.0, + 1099.0, + 1835.0, + 291.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1796.0, + 1405.0, + 1796.0, + 1405.0, + 1835.0, + 1133.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1827.0, + 1405.0, + 1827.0, + 1405.0, + 1865.0, + 292.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1858.0, + 1405.0, + 1858.0, + 1405.0, + 1895.0, + 292.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1888.0, + 1394.0, + 1888.0, + 1394.0, + 1924.0, + 292.0, + 1924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1118.0, + 1405.0, + 1118.0, + 1405.0, + 1157.0, + 292.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1152.0, + 1404.0, + 1152.0, + 1404.0, + 1186.0, + 295.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1182.0, + 1331.0, + 1182.0, + 1331.0, + 1216.0, + 295.0, + 1216.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 226, + 1405, + 226, + 1405, + 668, + 297, + 668 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 295, + 750, + 1405, + 750, + 1405, + 1149, + 295, + 1149 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1660, + 1405, + 1660, + 1405, + 1786, + 297, + 1786 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 296, + 1433, + 1404, + 1433, + 1404, + 1558, + 296, + 1558 + ], + "score": 0.976 + }, + { + "category_id": 8, + "poly": [ + 534, + 1938, + 1159, + 1938, + 1159, + 2031, + 534, + 2031 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 299, + 1868, + 1400, + 1868, + 1400, + 1932, + 299, + 1932 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 296, + 1348, + 1396, + 1348, + 1396, + 1412, + 296, + 1412 + ], + "score": 0.948 + }, + { + "category_id": 0, + "poly": [ + 299, + 1597, + 912, + 1597, + 912, + 1633, + 299, + 1633 + ], + "score": 0.932 + }, + { + "category_id": 1, + "poly": [ + 292, + 1160, + 1368, + 1160, + 1368, + 1200, + 292, + 1200 + ], + "score": 0.916 + }, + { + "category_id": 0, + "poly": [ + 297, + 700, + 998, + 700, + 998, + 733, + 297, + 733 + ], + "score": 0.916 + }, + { + "category_id": 0, + "poly": [ + 298, + 1817, + 579, + 1817, + 579, + 1850, + 298, + 1850 + ], + "score": 0.909 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.903 + }, + { + "category_id": 1, + "poly": [ + 299, + 1227, + 626, + 1227, + 626, + 1260, + 299, + 1260 + ], + "score": 0.901 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1290, + 1400, + 1290, + 1400, + 1320, + 1366, + 1320 + ], + "score": 0.872 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1969, + 1400, + 1969, + 1400, + 2000, + 1365, + 2000 + ], + "score": 0.854 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.768 + }, + { + "category_id": 8, + "poly": [ + 421, + 1268, + 1281, + 1268, + 1281, + 1341, + 421, + 1341 + ], + "score": 0.651 + }, + { + "category_id": 8, + "poly": [ + 418, + 1267, + 1280, + 1267, + 1280, + 1341, + 418, + 1341 + ], + "score": 0.474 + }, + { + "category_id": 13, + "poly": [ + 515, + 782, + 658, + 782, + 658, + 816, + 515, + 816 + ], + "score": 0.96, + "latex": "\\{ g _ { 1 } , \\cdots , g _ { t } \\}" + }, + { + "category_id": 13, + "poly": [ + 428, + 905, + 1027, + 905, + 1027, + 978, + 428, + 978 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { p ( \\psi ( . ) ) = p ( \\sqrt { \\frac { 1 - \\beta _ { 2 } ^ { t } } { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } } ) \\approx p ( \\sqrt { \\frac { t } { \\sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 919, + 973, + 1125, + 973, + 1125, + 1030, + 919, + 1030 + ], + "score": 0.94, + "latex": "\\frac { 1 - \\beta _ { 2 } ^ { t } } { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1722, + 513, + 1722, + 513, + 1756, + 297, + 1756 + ], + "score": 0.93, + "latex": "\\operatorname { V a r } [ \\psi ( g _ { 1 } , \\cdots , g _ { t } ) ]" + }, + { + "category_id": 14, + "poly": [ + 537, + 1937, + 1162, + 1937, + 1162, + 2031, + 537, + 2031 + ], + "score": 0.93, + "latex": "p \\left( \\frac { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } { 1 - \\beta _ { 2 } ^ { t } } \\right) \\approx p \\left( \\frac { \\sum _ { i = 1 } ^ { f \\left( t , \\beta _ { 2 } \\right) } g _ { t + 1 - i } ^ { 2 } } { f ( t , \\beta _ { 2 } ) } \\right) ." + }, + { + "category_id": 13, + "poly": [ + 1117, + 921, + 1290, + 921, + 1290, + 957, + 1117, + 957 + ], + "score": 0.92, + "latex": "g _ { i } \\sim \\mathcal { N } ( 0 , \\sigma ^ { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 974, + 1464, + 1084, + 1464, + 1084, + 1498, + 974, + 1498 + ], + "score": 0.92, + "latex": "\\mathrm { V a r } [ \\bar { \\psi } ( . ) ]" + }, + { + "category_id": 13, + "poly": [ + 990, + 1348, + 1099, + 1348, + 1099, + 1382, + 990, + 1382 + ], + "score": 0.92, + "latex": "\\mathrm { V a r } [ \\psi ( . ) ]" + }, + { + "category_id": 13, + "poly": [ + 298, + 843, + 443, + 843, + 443, + 877, + 298, + 877 + ], + "score": 0.92, + "latex": "\\{ g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } \\}" + }, + { + "category_id": 13, + "poly": [ + 1001, + 1662, + 1110, + 1662, + 1110, + 1696, + 1001, + 1696 + ], + "score": 0.92, + "latex": "\\mathrm { V a r } [ \\psi ( . ) ]" + }, + { + "category_id": 14, + "poly": [ + 414, + 1265, + 1280, + 1265, + 1280, + 1342, + 414, + 1342 + ], + "score": 0.92, + "latex": "\\mathrm { V a r } [ \\psi ( . ) ] = \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] - \\mathbb { E } [ \\psi ( . ) ] ^ { 2 } = \\tau ^ { 2 } ( \\frac { \\rho } { \\rho - 2 } - \\frac { \\rho 2 ^ { 2 \\rho - 5 } } { \\pi } \\mathcal { B } ( \\frac { \\rho - 1 } { 2 } , \\frac { \\rho - 1 } { 2 } ) ^ { 2 } ) ," + }, + { + "category_id": 13, + "poly": [ + 1098, + 1054, + 1220, + 1054, + 1220, + 1088, + 1098, + 1088 + ], + "score": 0.92, + "latex": "\\mathrm { V a r } [ \\psi ^ { 2 } ( . ) ]" + }, + { + "category_id": 13, + "poly": [ + 465, + 326, + 514, + 326, + 514, + 362, + 465, + 362 + ], + "score": 0.91, + "latex": "\\frac { 1 } { 1 2 \\epsilon ^ { 2 } }" + }, + { + "category_id": 13, + "poly": [ + 1080, + 1087, + 1189, + 1087, + 1189, + 1119, + 1080, + 1119 + ], + "score": 0.91, + "latex": "\\mathrm { V a r } [ \\psi ( . ) ]" + }, + { + "category_id": 13, + "poly": [ + 372, + 1349, + 422, + 1349, + 422, + 1382, + 372, + 1382 + ], + "score": 0.91, + "latex": "B ( . )" + }, + { + "category_id": 13, + "poly": [ + 833, + 270, + 885, + 270, + 885, + 311, + 833, + 311 + ], + "score": 0.91, + "latex": "\\widehat { \\psi } ( . )" + }, + { + "category_id": 13, + "poly": [ + 298, + 261, + 557, + 261, + 557, + 328, + 298, + 328 + ], + "score": 0.91, + "latex": "\\frac { \\sqrt { 1 - \\beta _ { 2 } ^ { t } } } { \\epsilon + \\sqrt { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } }" + }, + { + "category_id": 13, + "poly": [ + 1256, + 1494, + 1365, + 1494, + 1365, + 1528, + 1256, + 1528 + ], + "score": 0.91, + "latex": "\\mathrm { V a r } [ \\psi ( . ) ]" + }, + { + "category_id": 13, + "poly": [ + 1213, + 225, + 1406, + 225, + 1406, + 265, + 1213, + 265 + ], + "score": 0.91, + "latex": "\\widehat { \\psi } ( g _ { 1 } , \\cdots , g _ { t } ) =" + }, + { + "category_id": 13, + "poly": [ + 366, + 361, + 480, + 361, + 480, + 391, + 366, + 391 + ], + "score": 0.9, + "latex": "\\epsilon = 1 0 ^ { - 8 }" + }, + { + "category_id": 13, + "poly": [ + 730, + 841, + 839, + 841, + 839, + 878, + 730, + 878 + ], + "score": 0.9, + "latex": "\\breve { 1 } - \\beta _ { 2 } ^ { t - 1 } ," + }, + { + "category_id": 13, + "poly": [ + 1282, + 325, + 1393, + 325, + 1393, + 357, + 1282, + 357 + ], + "score": 0.9, + "latex": "\\epsilon = 1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1084, + 361, + 1084, + 361, + 1119, + 298, + 1119 + ], + "score": 0.9, + "latex": "\\psi ^ { 2 } ( . )" + }, + { + "category_id": 13, + "poly": [ + 423, + 1230, + 513, + 1230, + 513, + 1261, + 423, + 1261 + ], + "score": 0.88, + "latex": "\\forall \\rho > 4" + }, + { + "category_id": 13, + "poly": [ + 538, + 982, + 651, + 982, + 651, + 1020, + 538, + 1020 + ], + "score": 0.85, + "latex": "\\textstyle \\mathcal { X } ^ { 2 } ( t , { \\frac { 1 } { \\sigma ^ { 2 } } } )" + }, + { + "category_id": 13, + "poly": [ + 683, + 1161, + 799, + 1161, + 799, + 1201, + 683, + 1201 + ], + "score": 0.84, + "latex": "\\scriptstyle { \\mathcal { X } } ^ { 2 } ( \\rho , { \\frac { 1 } { \\sigma ^ { 2 } } } )" + }, + { + "category_id": 13, + "poly": [ + 840, + 1697, + 860, + 1697, + 860, + 1725, + 840, + 1725 + ], + "score": 0.83, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 440, + 1386, + 458, + 1386, + 458, + 1411, + 440, + 1411 + ], + "score": 0.83, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 1233, + 1171, + 1252, + 1171, + 1252, + 1197, + 1233, + 1197 + ], + "score": 0.81, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 1199, + 1668, + 1218, + 1668, + 1218, + 1694, + 1199, + 1694 + ], + "score": 0.81, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 696, + 1031, + 715, + 1031, + 715, + 1058, + 696, + 1058 + ], + "score": 0.81, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 549, + 1121, + 568, + 1121, + 568, + 1149, + 549, + 1149 + ], + "score": 0.79, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 681, + 1532, + 698, + 1532, + 698, + 1558, + 681, + 1558 + ], + "score": 0.78, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 975, + 1696, + 989, + 1696, + 989, + 1720, + 975, + 1720 + ], + "score": 0.78, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1165, + 237, + 1181, + 237, + 1181, + 258, + 1165, + 258 + ], + "score": 0.7, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 558, + 1825, + 577, + 1825, + 577, + 1851, + 558, + 1851 + ], + "score": 0.69, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 810, + 1162, + 919, + 1162, + 919, + 1198, + 810, + 1198 + ], + "score": 0.69, + "latex": "\\mathrm { V a r } [ \\psi ( . ) ]" + }, + { + "category_id": 13, + "poly": [ + 769, + 817, + 783, + 817, + 783, + 840, + 769, + 840 + ], + "score": 0.66, + "latex": "\\mathbf { \\chi } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 472, + 1162, + 566, + 1162, + 566, + 1198, + 472, + 1198 + ], + "score": 0.65, + "latex": "\\psi ^ { 2 } ( . ) \\sim" + }, + { + "category_id": 13, + "poly": [ + 690, + 460, + 706, + 460, + 706, + 481, + 690, + 481 + ], + "score": 0.64, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 755, + 551, + 772, + 551, + 772, + 572, + 755, + 572 + ], + "score": 0.57, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 299, + 981, + 422, + 981, + 422, + 1029, + 299, + 1029 + ], + "score": 0.52, + "latex": "\\frac { t } { \\sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } \\sim" + }, + { + "category_id": 13, + "poly": [ + 1140, + 515, + 1173, + 515, + 1173, + 544, + 1140, + 544 + ], + "score": 0.31, + "latex": "2 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 298, + 981, + 651, + 981, + 651, + 1029, + 298, + 1029 + ], + "score": 0.25, + "latex": "{ \\frac { t } { \\sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } } \\sim \\mathrm { S c a l e } { \\mathrm { - i n v } } { - \\mathcal { X } } ^ { 2 } ( t , { \\frac { 1 } { \\sigma ^ { 2 } } } )" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1595.0, + 914.0, + 1595.0, + 914.0, + 1636.0, + 293.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 700.0, + 1000.0, + 700.0, + 1000.0, + 736.0, + 294.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1810.0, + 557.0, + 1810.0, + 557.0, + 1859.0, + 292.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 1810.0, + 584.0, + 1810.0, + 584.0, + 1859.0, + 578.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 863.0, + 2085.0, + 863.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 227.0, + 1164.0, + 227.0, + 1164.0, + 267.0, + 296.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 227.0, + 1212.0, + 227.0, + 1212.0, + 267.0, + 1182.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 276.0, + 832.0, + 276.0, + 832.0, + 313.0, + 565.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 276.0, + 1402.0, + 276.0, + 1402.0, + 313.0, + 886.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 321.0, + 464.0, + 321.0, + 464.0, + 371.0, + 287.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 321.0, + 1281.0, + 321.0, + 1281.0, + 371.0, + 515.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 321.0, + 1407.0, + 321.0, + 1407.0, + 371.0, + 1394.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 356.0, + 365.0, + 356.0, + 365.0, + 400.0, + 292.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 356.0, + 1408.0, + 356.0, + 1408.0, + 400.0, + 481.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 392.0, + 1403.0, + 392.0, + 1403.0, + 425.0, + 295.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 421.0, + 1405.0, + 421.0, + 1405.0, + 460.0, + 292.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 452.0, + 689.0, + 452.0, + 689.0, + 489.0, + 292.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 452.0, + 1406.0, + 452.0, + 1406.0, + 489.0, + 707.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 478.0, + 1406.0, + 478.0, + 1406.0, + 521.0, + 291.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 516.0, + 1139.0, + 516.0, + 1139.0, + 549.0, + 295.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 516.0, + 1402.0, + 516.0, + 1402.0, + 549.0, + 1174.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 542.0, + 754.0, + 542.0, + 754.0, + 580.0, + 294.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 542.0, + 1405.0, + 542.0, + 1405.0, + 580.0, + 773.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 575.0, + 1403.0, + 575.0, + 1403.0, + 610.0, + 294.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 605.0, + 1405.0, + 605.0, + 1405.0, + 641.0, + 294.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 636.0, + 886.0, + 636.0, + 886.0, + 672.0, + 295.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 747.0, + 1406.0, + 747.0, + 1406.0, + 791.0, + 293.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 781.0, + 514.0, + 781.0, + 514.0, + 817.0, + 293.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 781.0, + 1405.0, + 781.0, + 1405.0, + 817.0, + 659.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 814.0, + 768.0, + 814.0, + 768.0, + 847.0, + 295.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 814.0, + 1405.0, + 814.0, + 1405.0, + 847.0, + 784.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 834.0, + 297.0, + 834.0, + 297.0, + 882.0, + 290.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 834.0, + 729.0, + 834.0, + 729.0, + 882.0, + 444.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 834.0, + 1408.0, + 834.0, + 1408.0, + 882.0, + 840.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 872.0, + 1405.0, + 872.0, + 1405.0, + 911.0, + 293.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 922.0, + 1116.0, + 922.0, + 1116.0, + 959.0, + 1038.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 922.0, + 1403.0, + 922.0, + 1403.0, + 959.0, + 1291.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 966.0, + 918.0, + 966.0, + 918.0, + 1038.0, + 652.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 966.0, + 1404.0, + 966.0, + 1404.0, + 1038.0, + 1126.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1024.0, + 695.0, + 1024.0, + 695.0, + 1060.0, + 294.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1024.0, + 1405.0, + 1024.0, + 1405.0, + 1060.0, + 716.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1053.0, + 1097.0, + 1053.0, + 1097.0, + 1090.0, + 294.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 1053.0, + 1407.0, + 1053.0, + 1407.0, + 1090.0, + 1221.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1082.0, + 1079.0, + 1082.0, + 1079.0, + 1122.0, + 362.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 1082.0, + 1407.0, + 1082.0, + 1407.0, + 1122.0, + 1190.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1116.0, + 548.0, + 1116.0, + 548.0, + 1152.0, + 295.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 1116.0, + 1207.0, + 1116.0, + 1207.0, + 1152.0, + 569.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 916.0, + 614.0, + 916.0, + 614.0, + 969.0, + 292.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.25, + 918.5, + 927.25, + 918.5, + 927.25, + 966.5, + 827.25, + 966.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1659.0, + 1000.0, + 1659.0, + 1000.0, + 1698.0, + 292.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 1659.0, + 1198.0, + 1659.0, + 1198.0, + 1698.0, + 1111.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 1659.0, + 1407.0, + 1659.0, + 1407.0, + 1698.0, + 1219.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1691.0, + 839.0, + 1691.0, + 839.0, + 1727.0, + 294.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1691.0, + 974.0, + 1691.0, + 974.0, + 1727.0, + 861.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1691.0, + 1406.0, + 1691.0, + 1406.0, + 1727.0, + 990.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 1718.0, + 1407.0, + 1718.0, + 1407.0, + 1761.0, + 514.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1752.0, + 655.0, + 1752.0, + 655.0, + 1789.0, + 294.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1432.0, + 1405.0, + 1432.0, + 1405.0, + 1470.0, + 294.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1461.0, + 973.0, + 1461.0, + 973.0, + 1501.0, + 294.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 1461.0, + 1405.0, + 1461.0, + 1405.0, + 1501.0, + 1085.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1491.0, + 1255.0, + 1491.0, + 1255.0, + 1533.0, + 293.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1366.0, + 1491.0, + 1405.0, + 1491.0, + 1405.0, + 1533.0, + 1366.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1524.0, + 680.0, + 1524.0, + 680.0, + 1564.0, + 293.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1524.0, + 710.0, + 1524.0, + 710.0, + 1564.0, + 699.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1867.0, + 1402.0, + 1867.0, + 1402.0, + 1905.0, + 295.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1899.0, + 918.0, + 1899.0, + 918.0, + 1937.0, + 294.0, + 1937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1347.0, + 371.0, + 1347.0, + 371.0, + 1383.0, + 296.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1347.0, + 989.0, + 1347.0, + 989.0, + 1383.0, + 423.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 1347.0, + 1400.0, + 1347.0, + 1400.0, + 1383.0, + 1100.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1379.0, + 439.0, + 1379.0, + 439.0, + 1413.0, + 295.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 1379.0, + 1208.0, + 1379.0, + 1208.0, + 1413.0, + 459.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1375.0, + 1381.0, + 1401.0, + 1381.0, + 1401.0, + 1411.0, + 1375.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1160.0, + 471.0, + 1160.0, + 471.0, + 1202.0, + 295.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 1160.0, + 682.0, + 1160.0, + 682.0, + 1202.0, + 567.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 1160.0, + 809.0, + 1160.0, + 809.0, + 1202.0, + 800.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1160.0, + 1232.0, + 1160.0, + 1232.0, + 1202.0, + 920.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1160.0, + 1371.0, + 1160.0, + 1371.0, + 1202.0, + 1253.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1227.0, + 422.0, + 1227.0, + 422.0, + 1263.0, + 296.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 1227.0, + 629.0, + 1227.0, + 629.0, + 1263.0, + 514.0, + 1263.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1453, + 1406, + 1453, + 1406, + 1657, + 296, + 1657 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 295, + 1122, + 1406, + 1122, + 1406, + 1393, + 295, + 1393 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1704, + 1401, + 1704, + 1401, + 1801, + 298, + 1801 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 581, + 1035, + 1119, + 1035, + 1119, + 1117, + 581, + 1117 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 300, + 968, + 1398, + 968, + 1398, + 1032, + 300, + 1032 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 297, + 1969, + 1410, + 1969, + 1410, + 2037, + 297, + 2037 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 345, + 1809, + 1355, + 1809, + 1355, + 1883, + 345, + 1883 + ], + "score": 0.922 + }, + { + "category_id": 0, + "poly": [ + 297, + 1410, + 904, + 1410, + 904, + 1444, + 297, + 1444 + ], + "score": 0.917 + }, + { + "category_id": 8, + "poly": [ + 631, + 1928, + 1066, + 1928, + 1066, + 1966, + 631, + 1966 + ], + "score": 0.915 + }, + { + "category_id": 1, + "poly": [ + 298, + 1886, + 817, + 1886, + 817, + 1924, + 298, + 1924 + ], + "score": 0.915 + }, + { + "category_id": 8, + "poly": [ + 431, + 1662, + 1269, + 1662, + 1269, + 1701, + 431, + 1701 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.9 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1933, + 1401, + 1933, + 1401, + 1961, + 1365, + 1961 + ], + "score": 0.867 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1064, + 1401, + 1064, + 1401, + 1093, + 1366, + 1093 + ], + "score": 0.858 + }, + { + "category_id": 1, + "poly": [ + 272, + 262, + 1377, + 262, + 1377, + 857, + 272, + 857 + ], + "score": 0.774 + }, + { + "category_id": 2, + "poly": [ + 840, + 2087, + 859, + 2087, + 859, + 2112, + 840, + 2112 + ], + "score": 0.749 + }, + { + "category_id": 1, + "poly": [ + 266, + 865, + 414, + 865, + 414, + 896, + 266, + 896 + ], + "score": 0.578 + }, + { + "category_id": 1, + "poly": [ + 297, + 237, + 1012, + 237, + 1012, + 270, + 297, + 270 + ], + "score": 0.416 + }, + { + "category_id": 0, + "poly": [ + 266, + 865, + 414, + 865, + 414, + 896, + 266, + 896 + ], + "score": 0.303 + }, + { + "category_id": 0, + "poly": [ + 297, + 237, + 1012, + 237, + 1012, + 270, + 297, + 270 + ], + "score": 0.103 + }, + { + "category_id": 13, + "poly": [ + 935, + 1312, + 1045, + 1312, + 1045, + 1354, + 935, + 1354 + ], + "score": 0.93, + "latex": "\\frac { 2 } { 1 - \\beta _ { 2 } } - 1" + }, + { + "category_id": 13, + "poly": [ + 904, + 1999, + 1041, + 1999, + 1041, + 2038, + 904, + 2038 + ], + "score": 0.93, + "latex": "\\mathrm { V a r } [ \\sqrt { \\psi ( . ) } ]" + }, + { + "category_id": 14, + "poly": [ + 628, + 1926, + 1070, + 1926, + 1070, + 1966, + 628, + 1966 + ], + "score": 0.93, + "latex": "\\mathrm { V a r } [ \\psi ( . ) ] \\approx \\rho _ { t } / [ 2 ( \\rho _ { t } - 2 ) ( \\rho _ { t } - 4 ) \\sigma ^ { 2 } ] ." + }, + { + "category_id": 13, + "poly": [ + 723, + 311, + 783, + 311, + 783, + 345, + 723, + 345 + ], + "score": 0.93, + "latex": "f _ { t } ( \\theta )" + }, + { + "category_id": 13, + "poly": [ + 716, + 1122, + 1086, + 1122, + 1086, + 1171, + 716, + 1171 + ], + "score": 0.93, + "latex": "\\begin{array} { r l r } { f ( t , \\beta _ { 2 } ) } & { { } = } & { { \\frac { 2 } { 1 - \\beta _ { 2 } } } - 1 - { \\frac { 2 t \\beta _ { 2 } ^ { t } } { 1 - \\beta _ { 2 } ^ { t } } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 596, + 1315, + 686, + 1315, + 686, + 1349, + 596, + 1349 + ], + "score": 0.93, + "latex": "f ( t , \\beta _ { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 1020, + 1278, + 1110, + 1278, + 1110, + 1312, + 1020, + 1312 + ], + "score": 0.93, + "latex": "f ( t , \\beta _ { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 467, + 1764, + 557, + 1764, + 557, + 1803, + 467, + 1803 + ], + "score": 0.93, + "latex": "\\sqrt { \\psi ^ { 2 } ( . ) }" + }, + { + "category_id": 13, + "poly": [ + 1104, + 1177, + 1285, + 1177, + 1285, + 1213, + 1104, + 1213 + ], + "score": 0.93, + "latex": "g _ { i } \\sim \\mathcal { N } ( 0 , \\sigma ^ { 2 } )" + }, + { + "category_id": 14, + "poly": [ + 580, + 1033, + 1119, + 1033, + 1119, + 1117, + 580, + 1117 + ], + "score": 0.93, + "latex": "\\frac { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } \\cdot i } { 1 - \\beta _ { 2 } ^ { t } } = \\frac { \\sum _ { i = 1 } ^ { f ( t , \\beta _ { 2 } ) } ( t + 1 - i ) } { f ( t , \\beta _ { 2 } ) } ." + }, + { + "category_id": 13, + "poly": [ + 645, + 1000, + 736, + 1000, + 736, + 1034, + 645, + 1034 + ], + "score": 0.92, + "latex": "f ( t , \\beta _ { 2 } )" + }, + { + "category_id": 14, + "poly": [ + 338, + 1804, + 1359, + 1804, + 1359, + 1885, + 338, + 1885 + ], + "score": 0.92, + "latex": "\\sqrt { \\psi ^ { 2 } ( . ) } \\approx \\sqrt { \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] } + \\frac { 1 } { 2 \\sqrt { \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] } } ( \\psi ^ { 2 } ( . ) - \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] ) \\quad \\mathrm { a n d } \\quad \\mathrm { V a r } [ \\psi ( . ) ] \\approx \\frac { \\mathrm { V a r } [ \\psi ^ { 2 } ( . ) ] } { 4 \\mathbb { E } [ \\psi ^ { 2 } ( . ) ] } ." + }, + { + "category_id": 13, + "poly": [ + 381, + 658, + 608, + 658, + 608, + 696, + 381, + 696 + ], + "score": 0.92, + "latex": "l _ { t } \\sqrt { ( 1 - \\beta _ { 2 } ^ { t } ) / v _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 373, + 970, + 464, + 970, + 464, + 1003, + 373, + 1003 + ], + "score": 0.92, + "latex": "f ( t , \\beta _ { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 1208, + 1235, + 1400, + 1235, + 1400, + 1275, + 1208, + 1275 + ], + "score": 0.92, + "latex": "\\cdot \\mathcal { X } ^ { 2 } ( f ( t , \\beta _ { 2 } ) , \\frac { 1 } { \\sigma ^ { 2 } } )" + }, + { + "category_id": 13, + "poly": [ + 600, + 280, + 698, + 280, + 698, + 314, + 600, + 314 + ], + "score": 0.92, + "latex": "\\{ \\beta _ { 1 } , \\beta _ { 2 } \\}" + }, + { + "category_id": 13, + "poly": [ + 671, + 1277, + 787, + 1277, + 787, + 1315, + 671, + 1315 + ], + "score": 0.91, + "latex": "\\scriptstyle { \\mathcal { X } } ^ { 2 } ( \\rho , { \\frac { 1 } { \\sigma ^ { 2 } } } )" + }, + { + "category_id": 13, + "poly": [ + 381, + 697, + 660, + 697, + 660, + 749, + 381, + 749 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { r _ { t } \\sqrt { \\frac { ( \\rho _ { t } - 4 ) ( \\rho _ { t } - 2 ) \\rho _ { \\infty } } { ( \\rho _ { \\infty } - 4 ) ( \\rho _ { \\infty } - 2 ) \\rho _ { t } } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 755, + 1703, + 865, + 1703, + 865, + 1737, + 755, + 1737 + ], + "score": 0.91, + "latex": "\\mathrm { V a r } [ \\psi ( . ) ]" + }, + { + "category_id": 13, + "poly": [ + 1056, + 1528, + 1134, + 1528, + 1134, + 1559, + 1056, + 1559 + ], + "score": 0.91, + "latex": "\\rho _ { t } = 5" + }, + { + "category_id": 13, + "poly": [ + 641, + 1524, + 721, + 1524, + 721, + 1566, + 641, + 1566 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\dot { O } ( \\frac { 1 } { \\rho _ { t } } \\overline { { ) } } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 382, + 278, + 477, + 278, + 477, + 315, + 382, + 315 + ], + "score": 0.91, + "latex": "\\{ \\alpha _ { t } \\} _ { t = 1 } ^ { T }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1594, + 484, + 1594, + 484, + 1628, + 298, + 1628 + ], + "score": 0.91, + "latex": "\\mathrm { V a r } [ \\psi ( . ) ] | _ { \\rho _ { t } = \\rho _ { \\infty } }" + }, + { + "category_id": 13, + "poly": [ + 694, + 629, + 772, + 629, + 772, + 659, + 694, + 659 + ], + "score": 0.9, + "latex": "\\rho _ { t } > 4" + }, + { + "category_id": 13, + "poly": [ + 1184, + 1561, + 1402, + 1561, + 1402, + 1598, + 1184, + 1598 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { \\operatorname* { m i n } _ { \\rho _ { t } } \\mathrm { V a r } [ \\psi ( . ) ] = } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 530, + 1564, + 645, + 1564, + 645, + 1595, + 530, + 1595 + ], + "score": 0.89, + "latex": "\\rho _ { t } ~ = ~ 5 0 0" + }, + { + "category_id": 13, + "poly": [ + 340, + 562, + 564, + 562, + 564, + 595, + 340, + 595 + ], + "score": 0.89, + "latex": "\\widehat { m _ { t } } \\gets m _ { t } / ( 1 - \\beta _ { 1 } ^ { t } )" + }, + { + "category_id": 13, + "poly": [ + 762, + 1452, + 1333, + 1452, + 1333, + 1500, + 762, + 1500 + ], + "score": 0.88, + "latex": "\\begin{array} { r } { \\mathrm { V a r } [ \\psi ( . ) ] = \\tau ^ { 2 } ( \\frac { \\rho _ { t } } { \\rho _ { t } - 2 } - \\frac { \\rho _ { t } 2 ^ { 2 \\rho _ { t } - 5 } } { \\pi } \\mathcal { B } ( \\frac { \\rho _ { t } - 1 } { 2 } , \\frac { \\rho _ { t } - 1 } { 2 } ) ^ { 2 } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 382, + 750, + 636, + 750, + 636, + 781, + 382, + 781 + ], + "score": 0.88, + "latex": "\\theta _ { t } \\gets \\theta _ { t - 1 } - \\alpha _ { t } r _ { t } \\widehat { m _ { t } } l _ { t }" + }, + { + "category_id": 13, + "poly": [ + 347, + 1628, + 415, + 1628, + 415, + 1657, + 347, + 1657 + ], + "score": 0.88, + "latex": "( \\psi ( . ) )" + }, + { + "category_id": 13, + "poly": [ + 1079, + 1319, + 1121, + 1319, + 1121, + 1349, + 1079, + 1349 + ], + "score": 0.88, + "latex": "\\rho _ { \\infty }" + }, + { + "category_id": 13, + "poly": [ + 368, + 435, + 543, + 435, + 543, + 465, + 368, + 465 + ], + "score": 0.88, + "latex": "t = \\{ 1 , \\cdots , T \\}" + }, + { + "category_id": 13, + "poly": [ + 298, + 374, + 458, + 374, + 458, + 403, + 298, + 403 + ], + "score": 0.87, + "latex": "m _ { 0 } , v _ { 0 } \\gets 0 , 0" + }, + { + "category_id": 13, + "poly": [ + 485, + 313, + 514, + 313, + 514, + 342, + 485, + 342 + ], + "score": 0.87, + "latex": "\\theta _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 615, + 1235, + 806, + 1235, + 806, + 1274, + 615, + 1274 + ], + "score": 0.87, + "latex": "\\mathcal { X } ^ { 2 } ( f ( t , \\beta _ { 2 } ) , \\frac { 1 } { \\sigma ^ { 2 } } )" + }, + { + "category_id": 13, + "poly": [ + 339, + 495, + 672, + 495, + 672, + 530, + 339, + 530 + ], + "score": 0.87, + "latex": "v _ { t } 1 / \\beta _ { 2 } v _ { t - 1 } + ( 1 - \\beta _ { 2 } ) g _ { t } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 339, + 465, + 540, + 465, + 540, + 497, + 339, + 497 + ], + "score": 0.87, + "latex": "g _ { t } \\gets \\Delta _ { \\theta } f _ { t } ( \\theta _ { t - 1 } )" + }, + { + "category_id": 13, + "poly": [ + 339, + 531, + 665, + 531, + 665, + 562, + 339, + 562 + ], + "score": 0.86, + "latex": "m _ { t } \\gets \\beta _ { 1 } m _ { t - 1 } + ( 1 - \\beta _ { 1 } ) g _ { t }" + }, + { + "category_id": 13, + "poly": [ + 775, + 1352, + 1232, + 1352, + 1232, + 1395, + 775, + 1395 + ], + "score": 0.86, + "latex": "\\begin{array} { r } { f ( t , \\beta _ { 2 } ) \\leq \\operatorname* { l i m } _ { t \\infty } f ( t , \\beta _ { 2 } ) = \\frac { 2 } { 1 - \\beta _ { 2 } } - 1 . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 297, + 404, + 552, + 404, + 552, + 435, + 297, + 435 + ], + "score": 0.85, + "latex": "\\rho _ { \\infty } 2 / ( 1 - \\beta _ { 2 } ) - 1" + }, + { + "category_id": 13, + "poly": [ + 400, + 343, + 426, + 343, + 426, + 372, + 400, + 372 + ], + "score": 0.85, + "latex": "\\theta _ { t }" + }, + { + "category_id": 13, + "poly": [ + 722, + 1321, + 749, + 1321, + 749, + 1348, + 722, + 1348 + ], + "score": 0.84, + "latex": "\\rho _ { t }" + }, + { + "category_id": 13, + "poly": [ + 884, + 1659, + 1261, + 1659, + 1261, + 1702, + 884, + 1702 + ], + "score": 0.84, + "latex": "r _ { t } = \\sqrt { C _ { \\mathrm { V a r } } / \\mathrm { V a r } [ \\psi ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } ) ] }" + }, + { + "category_id": 13, + "poly": [ + 820, + 1176, + 937, + 1176, + 937, + 1215, + 820, + 1215 + ], + "score": 0.83, + "latex": "\\scriptstyle { \\mathcal { X } } ^ { 2 } ( \\rho , { \\frac { 1 } { \\sigma ^ { 2 } } } )" + }, + { + "category_id": 13, + "poly": [ + 843, + 1595, + 903, + 1595, + 903, + 1625, + 843, + 1625 + ], + "score": 0.82, + "latex": "C _ { \\mathrm { V a r } }" + }, + { + "category_id": 13, + "poly": [ + 340, + 596, + 642, + 596, + 642, + 629, + 340, + 629 + ], + "score": 0.81, + "latex": "\\rho _ { t } \\gets \\rho _ { \\infty } - 2 t \\beta _ { 2 } ^ { t } / ( 1 - \\beta _ { 2 } ^ { t } )" + }, + { + "category_id": 13, + "poly": [ + 577, + 1888, + 703, + 1888, + 703, + 1926, + 577, + 1926 + ], + "score": 0.8, + "latex": "\\scriptstyle \\mathcal { X } ^ { 2 } ( \\rho _ { t } , \\frac { 1 } { \\sigma ^ { 2 } } )" + }, + { + "category_id": 13, + "poly": [ + 1014, + 1628, + 1028, + 1628, + 1028, + 1652, + 1014, + 1652 + ], + "score": 0.78, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1330, + 1285, + 1347, + 1285, + 1347, + 1312, + 1330, + 1312 + ], + "score": 0.78, + "latex": "\\rho" + }, + { + "category_id": 14, + "poly": [ + 428, + 1663, + 772, + 1663, + 772, + 1701, + 428, + 1701 + ], + "score": 0.78, + "latex": "\\mathrm { V a r } [ r _ { t } \\psi ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } ) ] = C _ { \\mathrm { V a r } }" + }, + { + "category_id": 13, + "poly": [ + 448, + 1168, + 664, + 1168, + 664, + 1226, + 448, + 1226 + ], + "score": 0.68, + "latex": "\\frac { 1 - \\beta _ { 2 } ^ { t } } { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1225, + 806, + 1225, + 806, + 1280, + 298, + 1280 + ], + "score": 0.66, + "latex": "{ \\frac { \\sum _ { i = 1 } ^ { f ( t , \\beta _ { 2 } ) } g _ { t + 1 - i } ^ { 2 } } { f ( t , \\beta _ { 2 } ) } } \\sim \\mathrm { S c a l e } { - \\mathrm { i n v } - \\chi ^ { 2 } ( f ( t , \\beta _ { 2 } ) , { \\frac { 1 } { \\sigma ^ { 2 } } } ) }" + }, + { + "category_id": 13, + "poly": [ + 366, + 1888, + 704, + 1888, + 704, + 1925, + 366, + 1925 + ], + "score": 0.5, + "latex": "\\psi ^ { 2 } ( . ) \\sim \\mathrm { S c a l e } { \\mathrm { - i n v } { - } } \\mathcal { X } ^ { 2 } ( \\rho _ { t } , \\frac { 1 } { \\sigma ^ { 2 } } )" + }, + { + "category_id": 14, + "poly": [ + 884, + 1659, + 1261, + 1659, + 1261, + 1701, + 884, + 1701 + ], + "score": 0.49, + "latex": "r _ { t } = \\sqrt { C _ { \\mathrm { V a r } } / \\mathrm { V a r } [ \\psi ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } ) ] }" + }, + { + "category_id": 13, + "poly": [ + 448, + 1168, + 940, + 1168, + 940, + 1226, + 448, + 1226 + ], + "score": 0.49, + "latex": "\\begin{array} { r } { \\frac { 1 - \\beta _ { 2 } ^ { t } } { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } } \\ \\sim \\ \\mathrm { S c a l e - i n v - } \\mathcal { X } ^ { 2 } ( \\rho , \\frac { 1 } { \\sigma ^ { 2 } } ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 348, + 810, + 597, + 810, + 597, + 849, + 348, + 849 + ], + "score": 0.48, + "latex": "\\begin{array} { r } { \\big \\lfloor \\theta _ { t } \\gets \\theta _ { t - 1 } - \\alpha _ { t } \\widehat { m _ { t } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 380, + 864, + 414, + 864, + 414, + 895, + 380, + 895 + ], + "score": 0.37, + "latex": "\\theta _ { T }" + }, + { + "category_id": 14, + "poly": [ + 424, + 1660, + 1270, + 1660, + 1270, + 1702, + 424, + 1702 + ], + "score": 0.35, + "latex": "\\mathrm { V a r } [ r _ { t } \\psi ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } ) ] = C _ { \\mathrm { V a r } } \\quad \\mathrm { w h e r e } \\quad r _ { t } = \\sqrt { C _ { \\mathrm { V a r } } / \\mathrm { V a r } [ \\psi ( g _ { 1 } , \\cdot \\cdot \\cdot , g _ { t } ) ] } ." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1410.0, + 904.0, + 1410.0, + 904.0, + 1447.0, + 294.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 861.0, + 2086.0, + 861.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 263.0, + 862.0, + 379.0, + 862.0, + 379.0, + 899.0, + 263.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 236.0, + 1014.0, + 236.0, + 1014.0, + 275.0, + 296.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1334.0, + 1448.0, + 1410.0, + 1448.0, + 1410.0, + 1499.0, + 1334.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1494.0, + 1407.0, + 1494.0, + 1407.0, + 1531.0, + 295.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1524.0, + 640.0, + 1524.0, + 640.0, + 1565.0, + 294.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 1524.0, + 1055.0, + 1524.0, + 1055.0, + 1565.0, + 722.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 1524.0, + 1407.0, + 1524.0, + 1407.0, + 1565.0, + 1135.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1561.0, + 529.0, + 1561.0, + 529.0, + 1601.0, + 293.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 1561.0, + 1183.0, + 1561.0, + 1183.0, + 1601.0, + 646.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1561.0, + 1407.0, + 1561.0, + 1407.0, + 1601.0, + 1403.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1591.0, + 297.0, + 1591.0, + 297.0, + 1635.0, + 294.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 1591.0, + 842.0, + 1591.0, + 842.0, + 1635.0, + 485.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1591.0, + 1407.0, + 1591.0, + 1407.0, + 1635.0, + 904.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1623.0, + 346.0, + 1623.0, + 346.0, + 1659.0, + 294.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 1623.0, + 1013.0, + 1623.0, + 1013.0, + 1659.0, + 416.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1623.0, + 1295.0, + 1623.0, + 1295.0, + 1659.0, + 1029.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.75, + 1451.0, + 1008.75, + 1451.0, + 1008.75, + 1503.5, + 288.75, + 1503.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1446.5, + 1151.0, + 1446.5, + 1151.0, + 1482.0, + 1041.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1115.0, + 715.0, + 1115.0, + 715.0, + 1176.0, + 294.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 1115.0, + 1103.0, + 1115.0, + 1103.0, + 1176.0, + 1087.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 1125.0, + 1403.0, + 1125.0, + 1403.0, + 1165.0, + 1111.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1180.0, + 439.0, + 1180.0, + 439.0, + 1215.0, + 295.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 1162.0, + 447.0, + 1162.0, + 447.0, + 1233.0, + 443.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1162.0, + 1103.0, + 1162.0, + 1103.0, + 1233.0, + 941.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 1162.0, + 1407.0, + 1162.0, + 1407.0, + 1233.0, + 1286.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 1219.0, + 395.0, + 1219.0, + 395.0, + 1248.0, + 302.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1274.0, + 670.0, + 1274.0, + 670.0, + 1318.0, + 292.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 1274.0, + 1019.0, + 1274.0, + 1019.0, + 1318.0, + 788.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 1274.0, + 1329.0, + 1274.0, + 1329.0, + 1318.0, + 1111.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1348.0, + 1274.0, + 1407.0, + 1274.0, + 1407.0, + 1318.0, + 1348.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 1313.0, + 1078.0, + 1313.0, + 1078.0, + 1350.0, + 1046.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 1313.0, + 1405.0, + 1313.0, + 1405.0, + 1350.0, + 1122.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1343.0, + 774.0, + 1343.0, + 774.0, + 1408.0, + 285.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 1343.0, + 1253.0, + 1343.0, + 1253.0, + 1408.0, + 1233.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.25, + 1215.0, + 1412.25, + 1215.0, + 1412.25, + 1288.5, + 373.25, + 1288.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1305.5, + 1009.0, + 1305.5, + 1009.0, + 1359.5, + 290.0, + 1359.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1702.0, + 754.0, + 1702.0, + 754.0, + 1740.0, + 294.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 1702.0, + 1405.0, + 1702.0, + 1405.0, + 1740.0, + 866.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1730.0, + 1405.0, + 1730.0, + 1405.0, + 1771.0, + 293.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1767.0, + 466.0, + 1767.0, + 466.0, + 1804.0, + 293.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 1767.0, + 921.0, + 1767.0, + 921.0, + 1804.0, + 558.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 967.0, + 372.0, + 967.0, + 372.0, + 1005.0, + 295.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 967.0, + 1402.0, + 967.0, + 1402.0, + 1005.0, + 465.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 999.0, + 644.0, + 999.0, + 644.0, + 1035.0, + 295.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 999.0, + 843.0, + 999.0, + 843.0, + 1035.0, + 737.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1966.0, + 1404.0, + 1966.0, + 1404.0, + 2002.0, + 294.0, + 2002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2000.0, + 903.0, + 2000.0, + 903.0, + 2040.0, + 292.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 2000.0, + 1406.0, + 2000.0, + 1406.0, + 2040.0, + 1042.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1879.0, + 365.0, + 1879.0, + 365.0, + 1930.0, + 294.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 1879.0, + 821.0, + 1879.0, + 821.0, + 1930.0, + 705.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 273.0, + 381.0, + 273.0, + 381.0, + 322.0, + 292.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 273.0, + 599.0, + 273.0, + 599.0, + 322.0, + 478.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 273.0, + 1338.0, + 273.0, + 1338.0, + 322.0, + 699.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 310.0, + 484.0, + 310.0, + 484.0, + 346.0, + 378.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 310.0, + 722.0, + 310.0, + 722.0, + 346.0, + 515.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 310.0, + 1127.0, + 310.0, + 1127.0, + 346.0, + 784.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 338.0, + 399.0, + 338.0, + 399.0, + 377.0, + 294.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 338.0, + 675.0, + 338.0, + 675.0, + 377.0, + 427.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 272.0, + 370.0, + 297.0, + 370.0, + 297.0, + 405.0, + 272.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 370.0, + 909.0, + 370.0, + 909.0, + 405.0, + 459.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 268.0, + 400.0, + 296.0, + 400.0, + 296.0, + 439.0, + 268.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 400.0, + 1217.0, + 400.0, + 1217.0, + 439.0, + 553.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 272.0, + 433.0, + 367.0, + 433.0, + 367.0, + 465.0, + 272.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 433.0, + 586.0, + 433.0, + 586.0, + 465.0, + 544.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 275.0, + 474.0, + 290.0, + 474.0, + 290.0, + 491.0, + 275.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 461.0, + 338.0, + 461.0, + 338.0, + 499.0, + 332.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 461.0, + 1212.0, + 461.0, + 1212.0, + 499.0, + 541.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 273.0, + 506.0, + 289.0, + 506.0, + 289.0, + 526.0, + 273.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 497.0, + 1156.0, + 497.0, + 1156.0, + 533.0, + 673.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 272.0, + 537.0, + 291.0, + 537.0, + 291.0, + 560.0, + 272.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 528.0, + 1136.0, + 528.0, + 1136.0, + 563.0, + 666.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 272.0, + 568.0, + 291.0, + 568.0, + 291.0, + 590.0, + 272.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 560.0, + 339.0, + 560.0, + 339.0, + 598.0, + 334.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 560.0, + 1044.0, + 560.0, + 1044.0, + 598.0, + 565.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 272.0, + 603.0, + 291.0, + 603.0, + 291.0, + 626.0, + 272.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 593.0, + 339.0, + 593.0, + 339.0, + 631.0, + 334.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 593.0, + 1183.0, + 593.0, + 1183.0, + 631.0, + 643.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 275.0, + 635.0, + 289.0, + 635.0, + 289.0, + 655.0, + 275.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 623.0, + 693.0, + 623.0, + 693.0, + 662.0, + 335.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 623.0, + 836.0, + 623.0, + 836.0, + 662.0, + 773.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 267.0, + 669.0, + 290.0, + 669.0, + 290.0, + 691.0, + 267.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 657.0, + 380.0, + 657.0, + 380.0, + 699.0, + 374.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 657.0, + 985.0, + 657.0, + 985.0, + 699.0, + 609.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 267.0, + 715.0, + 291.0, + 715.0, + 291.0, + 740.0, + 267.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 691.0, + 1130.0, + 691.0, + 1130.0, + 752.0, + 661.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 266.0, + 754.0, + 291.0, + 754.0, + 291.0, + 781.0, + 266.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 743.0, + 381.0, + 743.0, + 381.0, + 787.0, + 376.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 743.0, + 1165.0, + 743.0, + 1165.0, + 787.0, + 637.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 781.0, + 394.0, + 781.0, + 394.0, + 812.0, + 336.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 265.0, + 787.0, + 292.0, + 787.0, + 292.0, + 814.0, + 265.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 265.0, + 818.0, + 292.0, + 818.0, + 292.0, + 843.0, + 265.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 807.0, + 347.0, + 807.0, + 347.0, + 849.0, + 344.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 807.0, + 1156.0, + 807.0, + 1156.0, + 849.0, + 598.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 708.0, + 485.0, + 708.0, + 485.0, + 747.0, + 377.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 263.0, + 862.0, + 379.0, + 862.0, + 379.0, + 899.0, + 263.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 236.0, + 1014.0, + 236.0, + 1014.0, + 275.0, + 296.0, + 275.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1460, + 1404, + 1460, + 1404, + 1676, + 298, + 1676 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 977, + 1405, + 977, + 1405, + 1159, + 298, + 1159 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1771, + 1404, + 1771, + 1404, + 1895, + 299, + 1895 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1220, + 1405, + 1220, + 1405, + 1448, + 297, + 1448 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 299, + 235, + 1059, + 235, + 1059, + 487, + 299, + 487 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 678, + 896, + 1020, + 896, + 1020, + 976, + 678, + 976 + ], + "score": 0.954 + }, + { + "category_id": 3, + "poly": [ + 303, + 548, + 1366, + 548, + 1366, + 782, + 303, + 782 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 299, + 1972, + 1402, + 1972, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 299, + 853, + 1230, + 853, + 1230, + 886, + 299, + 886 + ], + "score": 0.935 + }, + { + "category_id": 4, + "poly": [ + 341, + 792, + 1350, + 792, + 1350, + 825, + 341, + 825 + ], + "score": 0.923 + }, + { + "category_id": 4, + "poly": [ + 310, + 492, + 1046, + 492, + 1046, + 524, + 310, + 524 + ], + "score": 0.922 + }, + { + "category_id": 5, + "poly": [ + 1094, + 265, + 1368, + 265, + 1368, + 537, + 1094, + 537 + ], + "score": 0.903, + "html": "
MethodAcc.
CITIIIOSGDAdamRAdam91.5190.5491.38
eeeseSGDAdamRAdam69.8666.5467.62
" + }, + { + "category_id": 0, + "poly": [ + 300, + 1711, + 558, + 1711, + 558, + 1746, + 300, + 1746 + ], + "score": 0.903 + }, + { + "category_id": 6, + "poly": [ + 1067, + 225, + 1395, + 225, + 1395, + 256, + 1067, + 256 + ], + "score": 0.901 + }, + { + "category_id": 0, + "poly": [ + 300, + 1923, + 752, + 1923, + 752, + 1954, + 300, + 1954 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.779 + }, + { + "category_id": 0, + "poly": [ + 304, + 1164, + 1253, + 1164, + 1253, + 1197, + 304, + 1197 + ], + "score": 0.775 + }, + { + "category_id": 1, + "poly": [ + 304, + 1164, + 1253, + 1164, + 1253, + 1197, + 304, + 1197 + ], + "score": 0.115 + }, + { + "category_id": 14, + "poly": [ + 676, + 889, + 1021, + 889, + 1021, + 981, + 676, + 981 + ], + "score": 0.95, + "latex": "r _ { t } = \\sqrt { \\frac { ( \\rho _ { t } - 4 ) ( \\rho _ { t } - 2 ) \\rho _ { \\infty } } { ( \\rho _ { \\infty } - 4 ) ( \\rho _ { \\infty } - 2 ) \\rho _ { t } } } ." + }, + { + "category_id": 13, + "poly": [ + 650, + 1315, + 759, + 1315, + 759, + 1359, + 650, + 1359 + ], + "score": 0.93, + "latex": "\\frac { m i n ( t , T _ { w } ) } { T _ { w } }" + }, + { + "category_id": 13, + "poly": [ + 398, + 851, + 470, + 851, + 470, + 895, + 398, + 895 + ], + "score": 0.92, + "latex": "O ( \\textstyle { \\frac { 1 } { \\rho _ { t } } } )" + }, + { + "category_id": 13, + "poly": [ + 1018, + 1098, + 1132, + 1098, + 1132, + 1131, + 1018, + 1131 + ], + "score": 0.91, + "latex": "\\beta _ { 2 } ~ \\leq ~ 0 . 6" + }, + { + "category_id": 13, + "poly": [ + 1248, + 1099, + 1352, + 1099, + 1352, + 1131, + 1248, + 1131 + ], + "score": 0.9, + "latex": "\\rho _ { \\infty } \\leq 4" + }, + { + "category_id": 13, + "poly": [ + 546, + 1287, + 572, + 1287, + 572, + 1314, + 546, + 1314 + ], + "score": 0.85, + "latex": "r _ { t }" + }, + { + "category_id": 13, + "poly": [ + 652, + 1416, + 690, + 1416, + 690, + 1445, + 652, + 1445 + ], + "score": 0.73, + "latex": "T _ { w . }" + }, + { + "category_id": 13, + "poly": [ + 621, + 1561, + 636, + 1561, + 636, + 1581, + 621, + 1581 + ], + "score": 0.3, + "latex": "\\epsilon" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 240.0, + 606.0, + 240.0, + 606.0, + 266.0, + 495.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 239.0, + 976.0, + 239.0, + 976.0, + 265.0, + 893.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 278.0, + 331.0, + 278.0, + 331.0, + 300.0, + 300.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 270.0, + 580.0, + 270.0, + 580.0, + 298.0, + 556.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 266.0, + 834.0, + 266.0, + 834.0, + 365.0, + 811.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 296.0, + 330.0, + 296.0, + 330.0, + 318.0, + 300.0, + 318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 293.0, + 573.0, + 293.0, + 573.0, + 323.0, + 557.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 314.0, + 331.0, + 314.0, + 331.0, + 336.0, + 300.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 317.0, + 573.0, + 317.0, + 573.0, + 334.0, + 557.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 332.0, + 331.0, + 332.0, + 331.0, + 355.0, + 300.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 330.0, + 573.0, + 330.0, + 573.0, + 358.0, + 557.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 350.0, + 331.0, + 350.0, + 331.0, + 391.0, + 299.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 367.0, + 574.0, + 367.0, + 574.0, + 396.0, + 557.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 354.0, + 835.0, + 354.0, + 835.0, + 405.0, + 811.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 363.0, + 975.0, + 363.0, + 975.0, + 400.0, + 853.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 386.0, + 539.0, + 386.0, + 539.0, + 421.0, + 299.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 392.0, + 796.0, + 392.0, + 796.0, + 423.0, + 556.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 397.0, + 853.0, + 397.0, + 853.0, + 425.0, + 814.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 407.0, + 1051.0, + 407.0, + 1051.0, + 427.0, + 854.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 426.0, + 619.0, + 426.0, + 619.0, + 453.0, + 465.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 425.0, + 1056.0, + 425.0, + 1056.0, + 450.0, + 829.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 459.0, + 601.0, + 459.0, + 601.0, + 491.0, + 527.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 459.0, + 891.0, + 459.0, + 891.0, + 492.0, + 829.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.25, + 297.5, + 398.25, + 297.5, + 398.25, + 315.5, + 346.25, + 315.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.75, + 299.0, + 617.75, + 299.0, + 617.75, + 317.0, + 596.75, + 317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 306.5, + 901.0, + 306.5, + 901.0, + 335.0, + 849.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 331.5, + 727.0, + 331.5, + 727.0, + 368.5, + 619.0, + 368.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 353.5, + 571.0, + 353.5, + 571.0, + 369.5, + 558.0, + 369.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 547.0, + 622.0, + 547.0, + 622.0, + 570.0, + 570.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 549.0, + 879.0, + 549.0, + 879.0, + 569.0, + 843.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 551.0, + 1159.0, + 551.0, + 1159.0, + 568.0, + 1120.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 563.0, + 471.0, + 563.0, + 471.0, + 589.0, + 390.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 564.0, + 740.0, + 564.0, + 740.0, + 586.0, + 654.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 565.0, + 1003.0, + 565.0, + 1003.0, + 587.0, + 922.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 565.0, + 1272.0, + 565.0, + 1272.0, + 586.0, + 1186.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 584.0, + 329.0, + 584.0, + 329.0, + 626.0, + 301.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 585.0, + 588.0, + 585.0, + 588.0, + 736.0, + 560.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 593.0, + 867.0, + 593.0, + 867.0, + 610.0, + 838.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 597.0, + 1113.0, + 597.0, + 1113.0, + 610.0, + 1094.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 605.0, + 866.0, + 605.0, + 866.0, + 622.0, + 842.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 616.0, + 326.0, + 616.0, + 326.0, + 634.0, + 304.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 615.0, + 867.0, + 615.0, + 867.0, + 633.0, + 838.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 613.0, + 1114.0, + 613.0, + 1114.0, + 630.0, + 1093.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 628.0, + 327.0, + 628.0, + 327.0, + 653.0, + 306.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 627.0, + 867.0, + 627.0, + 867.0, + 656.0, + 838.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 630.0, + 1113.0, + 630.0, + 1113.0, + 649.0, + 1093.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 646.0, + 328.0, + 646.0, + 328.0, + 703.0, + 306.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 650.0, + 867.0, + 650.0, + 867.0, + 701.0, + 838.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 646.0, + 1115.0, + 646.0, + 1115.0, + 701.0, + 1092.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 697.0, + 327.0, + 697.0, + 327.0, + 724.0, + 306.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 710.0, + 651.0, + 710.0, + 651.0, + 727.0, + 604.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 695.0, + 867.0, + 695.0, + 867.0, + 723.0, + 838.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 703.0, + 1113.0, + 703.0, + 1113.0, + 717.0, + 1094.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 728.0, + 340.0, + 728.0, + 340.0, + 738.0, + 329.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 725.0, + 517.0, + 725.0, + 517.0, + 740.0, + 349.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 726.0, + 541.0, + 726.0, + 541.0, + 739.0, + 520.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 726.0, + 564.0, + 726.0, + 564.0, + 739.0, + 546.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 728.0, + 593.0, + 728.0, + 593.0, + 738.0, + 583.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 724.0, + 675.0, + 724.0, + 675.0, + 743.0, + 602.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 726.0, + 696.0, + 726.0, + 696.0, + 740.0, + 677.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 726.0, + 721.0, + 726.0, + 721.0, + 739.0, + 702.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 725.0, + 774.0, + 725.0, + 774.0, + 740.0, + 727.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 726.0, + 797.0, + 726.0, + 797.0, + 739.0, + 776.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 726.0, + 821.0, + 726.0, + 821.0, + 740.0, + 801.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 726.0, + 905.0, + 726.0, + 905.0, + 739.0, + 884.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 725.0, + 1091.0, + 725.0, + 1091.0, + 740.0, + 908.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 730.0, + 1122.0, + 730.0, + 1122.0, + 745.0, + 1106.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 731.0, + 1152.0, + 731.0, + 1152.0, + 745.0, + 1134.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 731.0, + 1183.0, + 731.0, + 1183.0, + 744.0, + 1163.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 731.0, + 1216.0, + 731.0, + 1216.0, + 744.0, + 1192.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 730.0, + 1363.0, + 730.0, + 1363.0, + 746.0, + 1220.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 743.0, + 505.0, + 743.0, + 505.0, + 758.0, + 354.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 739.0, + 775.0, + 739.0, + 775.0, + 759.0, + 621.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 740.0, + 1043.0, + 740.0, + 1043.0, + 758.0, + 889.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 739.0, + 1309.0, + 739.0, + 1309.0, + 759.0, + 1154.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 758.0, + 604.0, + 758.0, + 604.0, + 781.0, + 526.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 757.0, + 1135.0, + 757.0, + 1135.0, + 780.0, + 1063.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 655.5, + 917.0, + 655.5, + 917.0, + 672.5, + 899.0, + 672.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 722.0, + 868.0, + 722.0, + 868.0, + 736.0, + 859.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 789.0, + 1357.0, + 789.0, + 1357.0, + 829.0, + 341.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 489.0, + 1045.0, + 489.0, + 1045.0, + 529.0, + 312.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1711.0, + 560.0, + 1711.0, + 560.0, + 1751.0, + 293.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 223.0, + 1397.0, + 223.0, + 1397.0, + 259.0, + 1064.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1920.0, + 756.0, + 1920.0, + 756.0, + 1959.0, + 294.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1163.0, + 1256.0, + 1163.0, + 1256.0, + 1200.0, + 297.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1459.0, + 1404.0, + 1459.0, + 1404.0, + 1496.0, + 293.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1493.0, + 1404.0, + 1493.0, + 1404.0, + 1527.0, + 294.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1524.0, + 1402.0, + 1524.0, + 1402.0, + 1558.0, + 294.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1554.0, + 620.0, + 1554.0, + 620.0, + 1589.0, + 294.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1554.0, + 1404.0, + 1554.0, + 1404.0, + 1589.0, + 637.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1584.0, + 1404.0, + 1584.0, + 1404.0, + 1619.0, + 294.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1614.0, + 1406.0, + 1614.0, + 1406.0, + 1652.0, + 293.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1646.0, + 743.0, + 1646.0, + 743.0, + 1678.0, + 293.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 976.0, + 1406.0, + 976.0, + 1406.0, + 1011.0, + 296.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1006.0, + 1405.0, + 1006.0, + 1405.0, + 1042.0, + 295.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1033.0, + 1406.0, + 1033.0, + 1406.0, + 1077.0, + 292.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1069.0, + 1403.0, + 1069.0, + 1403.0, + 1101.0, + 295.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1097.0, + 1017.0, + 1097.0, + 1017.0, + 1136.0, + 293.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1097.0, + 1247.0, + 1097.0, + 1247.0, + 1136.0, + 1133.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 1097.0, + 1407.0, + 1097.0, + 1407.0, + 1136.0, + 1353.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1130.0, + 841.0, + 1130.0, + 841.0, + 1161.0, + 296.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1771.0, + 1405.0, + 1771.0, + 1405.0, + 1804.0, + 297.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1801.0, + 1404.0, + 1801.0, + 1404.0, + 1834.0, + 295.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1834.0, + 1404.0, + 1834.0, + 1404.0, + 1867.0, + 295.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1865.0, + 1354.0, + 1865.0, + 1354.0, + 1897.0, + 297.0, + 1897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1217.0, + 1405.0, + 1217.0, + 1405.0, + 1261.0, + 292.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1252.0, + 1402.0, + 1252.0, + 1402.0, + 1284.0, + 296.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1280.0, + 545.0, + 1280.0, + 545.0, + 1319.0, + 294.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1280.0, + 1405.0, + 1280.0, + 1405.0, + 1319.0, + 573.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 1305.0, + 649.0, + 1305.0, + 649.0, + 1363.0, + 287.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 1305.0, + 1411.0, + 1305.0, + 1411.0, + 1363.0, + 760.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1349.0, + 1405.0, + 1349.0, + 1405.0, + 1390.0, + 291.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1384.0, + 1405.0, + 1384.0, + 1405.0, + 1420.0, + 295.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1415.0, + 651.0, + 1415.0, + 651.0, + 1451.0, + 295.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 1415.0, + 1400.0, + 1415.0, + 1400.0, + 1451.0, + 691.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1969.0, + 1406.0, + 1969.0, + 1406.0, + 2011.0, + 294.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 2004.0, + 1403.0, + 2004.0, + 1403.0, + 2036.0, + 297.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 849.0, + 397.0, + 849.0, + 397.0, + 892.0, + 292.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 849.0, + 1236.0, + 849.0, + 1236.0, + 892.0, + 471.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1163.0, + 1256.0, + 1163.0, + 1256.0, + 1200.0, + 297.0, + 1200.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1613, + 1403, + 1613, + 1403, + 1858, + 298, + 1858 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 1262, + 1404, + 1262, + 1404, + 1598, + 297, + 1598 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1092, + 1403, + 1092, + 1403, + 1247, + 298, + 1247 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 300, + 1943, + 1402, + 1943, + 1402, + 2033, + 300, + 2033 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 332, + 244, + 1365, + 244, + 1365, + 625, + 332, + 625 + ], + "score": 0.971 + }, + { + "category_id": 3, + "poly": [ + 298, + 684, + 1405, + 684, + 1405, + 1035, + 298, + 1035 + ], + "score": 0.957 + }, + { + "category_id": 4, + "poly": [ + 302, + 1039, + 1391, + 1039, + 1391, + 1069, + 302, + 1069 + ], + "score": 0.923 + }, + { + "category_id": 0, + "poly": [ + 301, + 1891, + 808, + 1891, + 808, + 1922, + 301, + 1922 + ], + "score": 0.908 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.887 + }, + { + "category_id": 4, + "poly": [ + 333, + 629, + 1367, + 629, + 1367, + 660, + 333, + 660 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2111, + 842, + 2111 + ], + "score": 0.637 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2111, + 842, + 2111 + ], + "score": 0.292 + }, + { + "category_id": 4, + "poly": [ + 382, + 684, + 1322, + 684, + 1322, + 709, + 382, + 709 + ], + "score": 0.141 + }, + { + "category_id": 13, + "poly": [ + 649, + 1705, + 916, + 1705, + 916, + 1739, + 649, + 1739 + ], + "score": 0.86, + "latex": "\\{ 0 . 1 , 0 . 0 3 , 0 . 0 1 , 0 . 0 0 3 \\} ," + }, + { + "category_id": 13, + "poly": [ + 1049, + 1014, + 1123, + 1014, + 1123, + 1035, + 1049, + 1035 + ], + "score": 0.51, + "latex": "\\vert \\boldsymbol { \\mathsf { r } } = 0 . 0 0 3" + }, + { + "category_id": 13, + "poly": [ + 1285, + 415, + 1360, + 415, + 1360, + 436, + 1285, + 436 + ], + "score": 0.35, + "latex": "\\vert \\boldsymbol { \\mathsf { r } } = 0 . 0 0 3" + }, + { + "category_id": 13, + "poly": [ + 429, + 1014, + 484, + 1014, + 484, + 1036, + 429, + 1036 + ], + "score": 0.32, + "latex": "\\vert \\boldsymbol { \\mathsf { r } } = 0 . 1" + }, + { + "category_id": 13, + "poly": [ + 630, + 1014, + 695, + 1014, + 695, + 1036, + 630, + 1036 + ], + "score": 0.31, + "latex": "{ \\| r = 0 . 0 3 }" + }, + { + "category_id": 13, + "poly": [ + 850, + 1014, + 914, + 1014, + 914, + 1035, + 850, + 1035 + ], + "score": 0.25, + "latex": "\\vert \\mathsf { r } = 0 . 0 1" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 246.0, + 540.0, + 246.0, + 540.0, + 269.0, + 476.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 243.0, + 827.0, + 243.0, + 827.0, + 271.0, + 770.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 251.0, + 1037.0, + 251.0, + 1037.0, + 269.0, + 874.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 246.0, + 1101.0, + 246.0, + 1101.0, + 269.0, + 1055.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 266.0, + 1027.0, + 266.0, + 1027.0, + 289.0, + 882.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 308.0, + 361.0, + 308.0, + 361.0, + 426.0, + 334.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 288.0, + 398.0, + 288.0, + 398.0, + 344.0, + 376.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 288.0, + 671.0, + 288.0, + 671.0, + 342.0, + 650.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 289.0, + 968.0, + 289.0, + 968.0, + 341.0, + 944.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 291.0, + 1355.0, + 291.0, + 1355.0, + 318.0, + 1290.0, + 318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 345.0, + 396.0, + 345.0, + 396.0, + 363.0, + 376.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 343.0, + 671.0, + 343.0, + 671.0, + 362.0, + 650.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 347.0, + 962.0, + 347.0, + 962.0, + 359.0, + 946.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 333.0, + 1360.0, + 333.0, + 1360.0, + 356.0, + 1286.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 365.0, + 398.0, + 365.0, + 398.0, + 382.0, + 376.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 354.0, + 642.0, + 354.0, + 642.0, + 382.0, + 513.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 362.0, + 671.0, + 362.0, + 671.0, + 379.0, + 651.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 364.0, + 962.0, + 364.0, + 962.0, + 378.0, + 946.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 384.0, + 396.0, + 384.0, + 396.0, + 401.0, + 376.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 377.0, + 650.0, + 377.0, + 650.0, + 395.0, + 506.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 382.0, + 670.0, + 382.0, + 670.0, + 396.0, + 651.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 384.0, + 963.0, + 384.0, + 963.0, + 396.0, + 946.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 374.0, + 1357.0, + 374.0, + 1357.0, + 398.0, + 1286.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 402.0, + 396.0, + 402.0, + 396.0, + 419.0, + 376.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 396.0, + 627.0, + 396.0, + 627.0, + 415.0, + 529.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 400.0, + 673.0, + 400.0, + 673.0, + 418.0, + 651.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 402.0, + 965.0, + 402.0, + 965.0, + 416.0, + 946.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 422.0, + 399.0, + 422.0, + 399.0, + 439.0, + 376.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 419.0, + 673.0, + 419.0, + 673.0, + 433.0, + 653.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 417.0, + 970.0, + 417.0, + 970.0, + 434.0, + 945.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 414.0, + 1284.0, + 414.0, + 1284.0, + 437.0, + 1280.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 436.0, + 406.0, + 436.0, + 406.0, + 450.0, + 392.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 433.0, + 437.0, + 433.0, + 437.0, + 451.0, + 413.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 434.0, + 490.0, + 434.0, + 490.0, + 450.0, + 439.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 434.0, + 644.0, + 434.0, + 644.0, + 450.0, + 492.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 431.0, + 682.0, + 431.0, + 682.0, + 446.0, + 668.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 430.0, + 716.0, + 430.0, + 716.0, + 447.0, + 695.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 430.0, + 748.0, + 430.0, + 748.0, + 447.0, + 726.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 430.0, + 779.0, + 430.0, + 779.0, + 447.0, + 756.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 430.0, + 811.0, + 430.0, + 811.0, + 447.0, + 787.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 431.0, + 843.0, + 431.0, + 843.0, + 447.0, + 814.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 431.0, + 933.0, + 431.0, + 933.0, + 447.0, + 847.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 433.0, + 970.0, + 433.0, + 970.0, + 448.0, + 954.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 432.0, + 1004.0, + 432.0, + 1004.0, + 450.0, + 982.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 432.0, + 1035.0, + 432.0, + 1035.0, + 450.0, + 1013.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 432.0, + 1066.0, + 432.0, + 1066.0, + 450.0, + 1043.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 432.0, + 1096.0, + 432.0, + 1096.0, + 450.0, + 1075.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 432.0, + 1131.0, + 432.0, + 1131.0, + 448.0, + 1103.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 431.0, + 1221.0, + 431.0, + 1221.0, + 450.0, + 1133.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 463.0, + 405.0, + 463.0, + 405.0, + 481.0, + 376.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 462.0, + 682.0, + 462.0, + 682.0, + 482.0, + 649.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 461.0, + 974.0, + 461.0, + 974.0, + 482.0, + 942.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 480.0, + 363.0, + 480.0, + 363.0, + 591.0, + 331.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 475.0, + 405.0, + 475.0, + 405.0, + 503.0, + 373.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 473.0, + 682.0, + 473.0, + 682.0, + 507.0, + 649.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 475.0, + 972.0, + 475.0, + 972.0, + 506.0, + 941.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 499.0, + 405.0, + 499.0, + 405.0, + 517.0, + 380.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 500.0, + 681.0, + 500.0, + 681.0, + 518.0, + 655.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 498.0, + 974.0, + 498.0, + 974.0, + 519.0, + 944.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 510.0, + 404.0, + 510.0, + 404.0, + 527.0, + 373.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 511.0, + 682.0, + 511.0, + 682.0, + 530.0, + 649.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 511.0, + 974.0, + 511.0, + 974.0, + 530.0, + 941.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 524.0, + 405.0, + 524.0, + 405.0, + 541.0, + 380.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 525.0, + 682.0, + 525.0, + 682.0, + 542.0, + 655.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 525.0, + 972.0, + 525.0, + 972.0, + 542.0, + 946.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 535.0, + 405.0, + 535.0, + 405.0, + 552.0, + 376.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 536.0, + 680.0, + 536.0, + 680.0, + 553.0, + 648.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 537.0, + 972.0, + 537.0, + 972.0, + 553.0, + 939.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 547.0, + 406.0, + 547.0, + 406.0, + 564.0, + 381.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 548.0, + 683.0, + 548.0, + 683.0, + 569.0, + 654.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 549.0, + 974.0, + 549.0, + 974.0, + 566.0, + 946.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 559.0, + 405.0, + 559.0, + 405.0, + 588.0, + 376.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 561.0, + 682.0, + 561.0, + 682.0, + 592.0, + 649.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 561.0, + 974.0, + 561.0, + 974.0, + 592.0, + 941.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 567.0, + 1355.0, + 567.0, + 1355.0, + 594.0, + 1232.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 584.0, + 406.0, + 584.0, + 406.0, + 611.0, + 376.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 586.0, + 682.0, + 586.0, + 682.0, + 616.0, + 649.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 585.0, + 974.0, + 585.0, + 974.0, + 616.0, + 941.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 591.0, + 1323.0, + 591.0, + 1323.0, + 618.0, + 1230.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 610.0, + 415.0, + 610.0, + 415.0, + 622.0, + 405.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 608.0, + 450.0, + 608.0, + 450.0, + 624.0, + 426.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 609.0, + 643.0, + 609.0, + 643.0, + 624.0, + 454.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 609.0, + 692.0, + 609.0, + 692.0, + 625.0, + 677.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 609.0, + 726.0, + 609.0, + 726.0, + 625.0, + 701.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 609.0, + 755.0, + 609.0, + 755.0, + 625.0, + 730.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 609.0, + 785.0, + 609.0, + 785.0, + 625.0, + 760.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 608.0, + 930.0, + 608.0, + 930.0, + 626.0, + 787.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 608.0, + 982.0, + 608.0, + 982.0, + 623.0, + 968.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 609.0, + 1015.0, + 609.0, + 1015.0, + 623.0, + 995.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 607.0, + 1220.0, + 607.0, + 1220.0, + 625.0, + 1021.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 375.0, + 790.0, + 375.0, + 790.0, + 395.0, + 743.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.75, + 331.0, + 531.75, + 331.0, + 531.75, + 385.0, + 362.75, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 341.5, + 1120.0, + 341.5, + 1120.0, + 406.5, + 937.0, + 406.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 681.0, + 1318.0, + 681.0, + 1318.0, + 708.0, + 400.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 713.0, + 477.0, + 713.0, + 477.0, + 743.0, + 411.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 728.0, + 714.0, + 728.0, + 714.0, + 754.0, + 601.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 726.0, + 909.0, + 726.0, + 909.0, + 754.0, + 869.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 714.0, + 1063.0, + 714.0, + 1063.0, + 743.0, + 912.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 728.0, + 1119.0, + 728.0, + 1119.0, + 754.0, + 1079.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 728.0, + 1329.0, + 728.0, + 1329.0, + 753.0, + 1281.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 757.0, + 330.0, + 757.0, + 330.0, + 866.0, + 301.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 761.0, + 357.0, + 761.0, + 357.0, + 777.0, + 340.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 761.0, + 502.0, + 761.0, + 502.0, + 770.0, + 473.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 761.0, + 589.0, + 761.0, + 589.0, + 776.0, + 571.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 761.0, + 800.0, + 761.0, + 800.0, + 776.0, + 782.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 761.0, + 929.0, + 761.0, + 929.0, + 770.0, + 914.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 761.0, + 1009.0, + 761.0, + 1009.0, + 776.0, + 992.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 763.0, + 1217.0, + 763.0, + 1217.0, + 775.0, + 1205.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 771.0, + 359.0, + 771.0, + 359.0, + 801.0, + 335.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 770.0, + 503.0, + 770.0, + 503.0, + 779.0, + 483.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 771.0, + 590.0, + 771.0, + 590.0, + 801.0, + 565.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 780.0, + 764.0, + 780.0, + 764.0, + 793.0, + 715.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 771.0, + 800.0, + 771.0, + 800.0, + 801.0, + 776.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 771.0, + 1011.0, + 771.0, + 1011.0, + 801.0, + 986.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 774.0, + 1220.0, + 774.0, + 1220.0, + 801.0, + 1197.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 794.0, + 359.0, + 794.0, + 359.0, + 813.0, + 335.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 794.0, + 590.0, + 794.0, + 590.0, + 813.0, + 565.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 795.0, + 799.0, + 795.0, + 799.0, + 811.0, + 777.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 794.0, + 1011.0, + 794.0, + 1011.0, + 812.0, + 986.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 794.0, + 1221.0, + 794.0, + 1221.0, + 812.0, + 1196.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 808.0, + 357.0, + 808.0, + 357.0, + 824.0, + 340.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 808.0, + 589.0, + 808.0, + 589.0, + 824.0, + 571.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 808.0, + 799.0, + 808.0, + 799.0, + 825.0, + 782.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 808.0, + 1010.0, + 808.0, + 1010.0, + 824.0, + 991.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 810.0, + 1217.0, + 810.0, + 1217.0, + 821.0, + 1205.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 818.0, + 359.0, + 818.0, + 359.0, + 847.0, + 333.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 818.0, + 590.0, + 818.0, + 590.0, + 847.0, + 565.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 820.0, + 799.0, + 820.0, + 799.0, + 847.0, + 777.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 818.0, + 1011.0, + 818.0, + 1011.0, + 847.0, + 986.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 818.0, + 1221.0, + 818.0, + 1221.0, + 847.0, + 1196.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1275.0, + 825.0, + 1295.0, + 825.0, + 1295.0, + 845.0, + 1275.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 840.0, + 359.0, + 840.0, + 359.0, + 861.0, + 333.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 842.0, + 590.0, + 842.0, + 590.0, + 860.0, + 565.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 842.0, + 800.0, + 842.0, + 800.0, + 860.0, + 775.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 843.0, + 1009.0, + 843.0, + 1009.0, + 858.0, + 987.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 842.0, + 1221.0, + 842.0, + 1221.0, + 860.0, + 1196.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 854.0, + 365.0, + 854.0, + 365.0, + 875.0, + 341.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 865.0, + 388.0, + 865.0, + 388.0, + 875.0, + 374.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 865.0, + 452.0, + 865.0, + 452.0, + 874.0, + 440.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 860.0, + 541.0, + 860.0, + 541.0, + 880.0, + 453.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 854.0, + 595.0, + 854.0, + 595.0, + 875.0, + 571.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 863.0, + 618.0, + 863.0, + 618.0, + 875.0, + 605.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 862.0, + 750.0, + 862.0, + 750.0, + 875.0, + 687.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 862.0, + 770.0, + 862.0, + 770.0, + 876.0, + 751.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 856.0, + 800.0, + 856.0, + 800.0, + 870.0, + 782.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 862.0, + 830.0, + 862.0, + 830.0, + 877.0, + 814.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 865.0, + 893.0, + 865.0, + 893.0, + 874.0, + 880.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 861.0, + 982.0, + 861.0, + 982.0, + 877.0, + 899.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 856.0, + 1016.0, + 856.0, + 1016.0, + 875.0, + 992.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 863.0, + 1041.0, + 863.0, + 1041.0, + 877.0, + 1025.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 865.0, + 1104.0, + 865.0, + 1104.0, + 874.0, + 1090.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 862.0, + 1191.0, + 862.0, + 1191.0, + 877.0, + 1109.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 856.0, + 1226.0, + 856.0, + 1226.0, + 874.0, + 1202.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 865.0, + 1249.0, + 865.0, + 1249.0, + 875.0, + 1236.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1300.0, + 865.0, + 1312.0, + 865.0, + 1312.0, + 875.0, + 1300.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 860.0, + 1402.0, + 860.0, + 1402.0, + 876.0, + 1318.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 877.0, + 361.0, + 877.0, + 361.0, + 990.0, + 301.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 877.0, + 591.0, + 877.0, + 591.0, + 916.0, + 565.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 876.0, + 804.0, + 876.0, + 804.0, + 908.0, + 772.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 880.0, + 1010.0, + 880.0, + 1010.0, + 903.0, + 989.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 879.0, + 1221.0, + 879.0, + 1221.0, + 916.0, + 1193.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 895.0, + 802.0, + 895.0, + 802.0, + 927.0, + 772.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 897.0, + 1012.0, + 897.0, + 1012.0, + 915.0, + 986.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 907.0, + 360.0, + 907.0, + 360.0, + 944.0, + 332.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 908.0, + 594.0, + 908.0, + 594.0, + 994.0, + 560.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 918.0, + 799.0, + 918.0, + 799.0, + 933.0, + 776.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 908.0, + 1015.0, + 908.0, + 1015.0, + 945.0, + 985.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 908.0, + 1219.0, + 908.0, + 1219.0, + 933.0, + 1196.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 930.0, + 364.0, + 930.0, + 364.0, + 995.0, + 328.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 925.0, + 804.0, + 925.0, + 804.0, + 994.0, + 771.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 936.0, + 1012.0, + 936.0, + 1012.0, + 979.0, + 986.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 927.0, + 1221.0, + 927.0, + 1221.0, + 982.0, + 1192.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 974.0, + 1010.0, + 974.0, + 1010.0, + 988.0, + 989.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 975.0, + 1217.0, + 975.0, + 1217.0, + 989.0, + 1196.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 985.0, + 365.0, + 985.0, + 365.0, + 1004.0, + 346.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 991.0, + 390.0, + 991.0, + 390.0, + 1006.0, + 372.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 993.0, + 410.0, + 993.0, + 410.0, + 1003.0, + 396.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 989.0, + 540.0, + 989.0, + 540.0, + 1006.0, + 415.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 985.0, + 589.0, + 985.0, + 589.0, + 995.0, + 579.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 993.0, + 620.0, + 993.0, + 620.0, + 1003.0, + 605.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 993.0, + 640.0, + 993.0, + 640.0, + 1003.0, + 625.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 991.0, + 663.0, + 991.0, + 663.0, + 1006.0, + 647.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 991.0, + 684.0, + 991.0, + 684.0, + 1006.0, + 668.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 989.0, + 771.0, + 989.0, + 771.0, + 1006.0, + 686.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 993.0, + 830.0, + 993.0, + 830.0, + 1003.0, + 815.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 993.0, + 852.0, + 993.0, + 852.0, + 1003.0, + 836.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 993.0, + 873.0, + 993.0, + 873.0, + 1003.0, + 858.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 991.0, + 896.0, + 991.0, + 896.0, + 1006.0, + 878.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 989.0, + 982.0, + 989.0, + 982.0, + 1006.0, + 897.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 985.0, + 1017.0, + 985.0, + 1017.0, + 1003.0, + 999.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 991.0, + 1041.0, + 991.0, + 1041.0, + 1006.0, + 1024.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 993.0, + 1060.0, + 993.0, + 1060.0, + 1003.0, + 1048.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 993.0, + 1083.0, + 993.0, + 1083.0, + 1003.0, + 1068.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 989.0, + 1192.0, + 989.0, + 1192.0, + 1006.0, + 1085.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 985.0, + 1226.0, + 985.0, + 1226.0, + 1004.0, + 1207.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 991.0, + 1251.0, + 991.0, + 1251.0, + 1006.0, + 1234.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 994.0, + 1271.0, + 994.0, + 1271.0, + 1003.0, + 1256.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 991.0, + 1294.0, + 991.0, + 1294.0, + 1006.0, + 1276.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 988.0, + 1403.0, + 988.0, + 1403.0, + 1008.0, + 1295.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1015.0, + 849.0, + 1015.0, + 849.0, + 1036.0, + 844.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1015.0, + 1048.0, + 1015.0, + 1048.0, + 1035.0, + 1045.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 1011.0, + 1401.0, + 1011.0, + 1401.0, + 1035.0, + 1212.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.75, + 863.0, + 1273.75, + 863.0, + 1273.75, + 874.0, + 1252.75, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 907.5, + 433.0, + 907.5, + 433.0, + 928.5, + 420.0, + 928.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.75, + 899.0, + 1090.75, + 899.0, + 1090.75, + 927.0, + 1061.75, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 906.5, + 1296.0, + 906.5, + 1296.0, + 924.5, + 1281.0, + 924.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 926.5, + 880.0, + 926.5, + 880.0, + 937.5, + 873.0, + 937.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 988.0, + 801.0, + 988.0, + 801.0, + 1001.0, + 791.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1032.0, + 1393.0, + 1032.0, + 1393.0, + 1074.0, + 304.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1888.0, + 812.0, + 1888.0, + 812.0, + 1927.0, + 293.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 624.0, + 1370.0, + 624.0, + 1370.0, + 665.0, + 326.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 683.0, + 1320.0, + 683.0, + 1320.0, + 710.0, + 401.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1611.0, + 1405.0, + 1611.0, + 1405.0, + 1649.0, + 293.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1643.0, + 1404.0, + 1643.0, + 1404.0, + 1679.0, + 293.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1673.0, + 1404.0, + 1673.0, + 1404.0, + 1710.0, + 293.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1704.0, + 648.0, + 1704.0, + 648.0, + 1740.0, + 293.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1704.0, + 1405.0, + 1704.0, + 1405.0, + 1740.0, + 917.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1736.0, + 1405.0, + 1736.0, + 1405.0, + 1769.0, + 293.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1763.0, + 1407.0, + 1763.0, + 1407.0, + 1801.0, + 292.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1798.0, + 1405.0, + 1798.0, + 1405.0, + 1831.0, + 296.0, + 1831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1825.0, + 866.0, + 1825.0, + 866.0, + 1864.0, + 295.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1260.0, + 1404.0, + 1260.0, + 1404.0, + 1297.0, + 294.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1293.0, + 1404.0, + 1293.0, + 1404.0, + 1327.0, + 295.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1323.0, + 1405.0, + 1323.0, + 1405.0, + 1358.0, + 295.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1352.0, + 1406.0, + 1352.0, + 1406.0, + 1388.0, + 291.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1384.0, + 1406.0, + 1384.0, + 1406.0, + 1419.0, + 294.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1415.0, + 1406.0, + 1415.0, + 1406.0, + 1450.0, + 294.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1444.0, + 1405.0, + 1444.0, + 1405.0, + 1482.0, + 294.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1477.0, + 1402.0, + 1477.0, + 1402.0, + 1508.0, + 297.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1506.0, + 1405.0, + 1506.0, + 1405.0, + 1540.0, + 294.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1535.0, + 1406.0, + 1535.0, + 1406.0, + 1572.0, + 292.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1565.0, + 1198.0, + 1565.0, + 1198.0, + 1602.0, + 294.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1094.0, + 1404.0, + 1094.0, + 1404.0, + 1127.0, + 294.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1122.0, + 1405.0, + 1122.0, + 1405.0, + 1160.0, + 292.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1154.0, + 1407.0, + 1154.0, + 1407.0, + 1191.0, + 294.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1187.0, + 1403.0, + 1187.0, + 1403.0, + 1216.0, + 297.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1216.0, + 1252.0, + 1216.0, + 1252.0, + 1249.0, + 294.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1940.0, + 1403.0, + 1940.0, + 1403.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1974.0, + 1404.0, + 1974.0, + 1404.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1406.0, + 2002.0, + 1406.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 664, + 1405, + 664, + 1405, + 1002, + 297, + 1002 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1541, + 1404, + 1541, + 1404, + 1765, + 297, + 1765 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 495, + 1403, + 495, + 1403, + 650, + 298, + 650 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1880, + 1403, + 1880, + 1403, + 2035, + 298, + 2035 + ], + "score": 0.976 + }, + { + "category_id": 5, + "poly": [ + 364, + 246, + 1325, + 246, + 1325, + 347, + 364, + 347 + ], + "score": 0.967, + "html": "
MethodIWSLT'14 DE-ENIWSLT'14 EN-DEWMT'16 EN-DE
Adam with warmupRAdam34.66± 0.01434.76 ± 0.00328.56± 0.06728.48 ± 0.05427.0327.27
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1402, + 1404, + 1402, + 1404, + 1528, + 298, + 1528 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 300, + 359, + 1403, + 359, + 1403, + 481, + 300, + 481 + ], + "score": 0.966 + }, + { + "category_id": 6, + "poly": [ + 548, + 208, + 1151, + 208, + 1151, + 240, + 548, + 240 + ], + "score": 0.924 + }, + { + "category_id": 0, + "poly": [ + 300, + 1039, + 687, + 1039, + 687, + 1072, + 300, + 1072 + ], + "score": 0.903 + }, + { + "category_id": 0, + "poly": [ + 299, + 1810, + 544, + 1810, + 544, + 1846, + 299, + 1846 + ], + "score": 0.901 + }, + { + "category_id": 1, + "poly": [ + 298, + 1244, + 1404, + 1244, + 1404, + 1389, + 298, + 1389 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 817, + 75, + 817, + 105, + 298, + 105 + ], + "score": 0.894 + }, + { + "category_id": 1, + "poly": [ + 297, + 1092, + 1405, + 1092, + 1405, + 1230, + 297, + 1230 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.792 + }, + { + "category_id": 13, + "poly": [ + 578, + 1572, + 672, + 1572, + 672, + 1606, + 578, + 1606 + ], + "score": 0.93, + "latex": "\\mathcal { N } ( \\bar { \\boldsymbol { \\mu } } , 1 )" + }, + { + "category_id": 13, + "poly": [ + 732, + 1092, + 947, + 1092, + 947, + 1149, + 732, + 1149 + ], + "score": 0.93, + "latex": "\\textstyle \\operatorname { V a r } [ \\sqrt { t / \\sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } ]" + }, + { + "category_id": 13, + "poly": [ + 670, + 1242, + 884, + 1242, + 884, + 1299, + 670, + 1299 + ], + "score": 0.93, + "latex": "\\textstyle \\operatorname { V a r } [ \\sqrt { t / \\sum _ { i = 1 } ^ { t } g _ { i } ^ { 2 } } ]" + }, + { + "category_id": 13, + "poly": [ + 838, + 1294, + 1051, + 1294, + 1051, + 1329, + 838, + 1329 + ], + "score": 0.92, + "latex": "\\nu = \\{ 5 , \\cdots , 5 0 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 901, + 1433, + 963, + 1433, + 963, + 1467, + 901, + 1467 + ], + "score": 0.91, + "latex": "\\psi ^ { 2 } ( . )" + }, + { + "category_id": 13, + "poly": [ + 725, + 1464, + 835, + 1464, + 835, + 1498, + 725, + 1498 + ], + "score": 0.91, + "latex": "\\mathrm { V a r } [ \\psi ( . ) ]" + }, + { + "category_id": 13, + "poly": [ + 1262, + 1601, + 1351, + 1601, + 1351, + 1642, + 1262, + 1642 + ], + "score": 0.91, + "latex": "\\textstyle \\operatorname { V a r } [ { \\frac { 1 } { \\widehat { v _ { t } } } } ]" + }, + { + "category_id": 13, + "poly": [ + 296, + 1637, + 386, + 1637, + 386, + 1677, + 296, + 1677 + ], + "score": 0.9, + "latex": "\\textstyle \\operatorname { V a r } [ { \\frac { r _ { t } } { \\widehat { v _ { t } } } } ]" + }, + { + "category_id": 13, + "poly": [ + 665, + 909, + 915, + 909, + 915, + 942, + 665, + 942 + ], + "score": 0.89, + "latex": "\\beta _ { 1 } = 0 . 9 , \\beta _ { 2 } = 0 . 9 9 9 )" + }, + { + "category_id": 13, + "poly": [ + 1301, + 1104, + 1403, + 1104, + 1403, + 1141, + 1301, + 1141 + ], + "score": 0.89, + "latex": "\\psi ^ { 2 } ( . ) = " + }, + { + "category_id": 13, + "poly": [ + 300, + 1144, + 507, + 1144, + 507, + 1202, + 300, + 1202 + ], + "score": 0.88, + "latex": "\\frac { 1 - \\beta _ { 2 } ^ { t } } { ( 1 - \\beta _ { 2 } ) \\sum _ { i = 1 } ^ { t } \\beta _ { 2 } ^ { t - i } g _ { i } ^ { 2 } }" + }, + { + "category_id": 13, + "poly": [ + 626, + 1639, + 657, + 1639, + 657, + 1671, + 626, + 1671 + ], + "score": 0.87, + "latex": "\\beta _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1295, + 377, + 1295, + 377, + 1324, + 297, + 1324 + ], + "score": 0.87, + "latex": "\\tau = 1" + }, + { + "category_id": 13, + "poly": [ + 484, + 1577, + 511, + 1577, + 511, + 1605, + 484, + 1605 + ], + "score": 0.85, + "latex": "g _ { i }" + }, + { + "category_id": 13, + "poly": [ + 503, + 1547, + 529, + 1547, + 529, + 1574, + 503, + 1574 + ], + "score": 0.85, + "latex": "g _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1229, + 1410, + 1255, + 1410, + 1255, + 1436, + 1229, + 1436 + ], + "score": 0.83, + "latex": "g _ { i }" + }, + { + "category_id": 13, + "poly": [ + 324, + 1577, + 345, + 1577, + 345, + 1605, + 324, + 1605 + ], + "score": 0.79, + "latex": "\\mu" + }, + { + "category_id": 13, + "poly": [ + 398, + 1708, + 418, + 1708, + 418, + 1735, + 398, + 1735 + ], + "score": 0.78, + "latex": "\\mu" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 204.0, + 1151.0, + 204.0, + 1151.0, + 244.0, + 547.0, + 244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1038.0, + 690.0, + 1038.0, + 690.0, + 1074.0, + 295.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1803.0, + 550.0, + 1803.0, + 550.0, + 1855.0, + 292.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 665.0, + 1406.0, + 665.0, + 1406.0, + 700.0, + 295.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 695.0, + 1403.0, + 695.0, + 1403.0, + 732.0, + 292.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 725.0, + 1406.0, + 725.0, + 1406.0, + 766.0, + 292.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 758.0, + 1405.0, + 758.0, + 1405.0, + 793.0, + 295.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 787.0, + 1405.0, + 787.0, + 1405.0, + 823.0, + 292.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 817.0, + 1405.0, + 817.0, + 1405.0, + 853.0, + 292.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 849.0, + 1405.0, + 849.0, + 1405.0, + 884.0, + 295.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 878.0, + 1406.0, + 878.0, + 1406.0, + 915.0, + 292.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 909.0, + 664.0, + 909.0, + 664.0, + 944.0, + 294.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 909.0, + 1405.0, + 909.0, + 1405.0, + 944.0, + 916.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 939.0, + 1406.0, + 939.0, + 1406.0, + 975.0, + 294.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 972.0, + 1327.0, + 972.0, + 1327.0, + 1003.0, + 297.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1542.0, + 502.0, + 1542.0, + 502.0, + 1577.0, + 295.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 1542.0, + 1405.0, + 1542.0, + 1405.0, + 1577.0, + 530.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1572.0, + 323.0, + 1572.0, + 323.0, + 1606.0, + 292.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1572.0, + 483.0, + 1572.0, + 483.0, + 1606.0, + 346.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 1572.0, + 577.0, + 1572.0, + 577.0, + 1606.0, + 512.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 1572.0, + 1405.0, + 1572.0, + 1405.0, + 1606.0, + 673.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1600.0, + 1261.0, + 1600.0, + 1261.0, + 1643.0, + 292.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 1600.0, + 1406.0, + 1600.0, + 1406.0, + 1643.0, + 1352.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1635.0, + 625.0, + 1635.0, + 625.0, + 1678.0, + 387.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 658.0, + 1635.0, + 1407.0, + 1635.0, + 1407.0, + 1678.0, + 658.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1669.0, + 1404.0, + 1669.0, + 1404.0, + 1707.0, + 294.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1700.0, + 397.0, + 1700.0, + 397.0, + 1740.0, + 294.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 1700.0, + 1405.0, + 1700.0, + 1405.0, + 1740.0, + 419.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1731.0, + 1313.0, + 1731.0, + 1313.0, + 1769.0, + 294.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 497.0, + 1404.0, + 497.0, + 1404.0, + 530.0, + 297.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 525.0, + 1405.0, + 525.0, + 1405.0, + 563.0, + 293.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 559.0, + 1404.0, + 559.0, + 1404.0, + 592.0, + 296.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 588.0, + 1405.0, + 588.0, + 1405.0, + 624.0, + 293.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 616.0, + 1075.0, + 616.0, + 1075.0, + 658.0, + 293.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 1404.0, + 1882.0, + 1404.0, + 1915.0, + 294.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1910.0, + 1407.0, + 1910.0, + 1407.0, + 1946.0, + 293.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1939.0, + 1408.0, + 1939.0, + 1408.0, + 1981.0, + 291.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2009.0, + 296.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1407.0, + 2003.0, + 1407.0, + 2040.0, + 294.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1400.0, + 1228.0, + 1400.0, + 1228.0, + 1439.0, + 293.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1400.0, + 1409.0, + 1400.0, + 1409.0, + 1439.0, + 1256.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1433.0, + 900.0, + 1433.0, + 900.0, + 1470.0, + 293.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 1433.0, + 1408.0, + 1433.0, + 1408.0, + 1470.0, + 964.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1464.0, + 724.0, + 1464.0, + 724.0, + 1500.0, + 294.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 1464.0, + 1406.0, + 1464.0, + 1406.0, + 1500.0, + 836.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1496.0, + 1298.0, + 1496.0, + 1298.0, + 1531.0, + 292.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 359.0, + 1404.0, + 359.0, + 1404.0, + 391.0, + 295.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 389.0, + 1404.0, + 389.0, + 1404.0, + 421.0, + 295.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 418.0, + 1405.0, + 418.0, + 1405.0, + 457.0, + 293.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 449.0, + 892.0, + 449.0, + 892.0, + 487.0, + 293.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1239.0, + 669.0, + 1239.0, + 669.0, + 1306.0, + 286.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 1239.0, + 1414.0, + 1239.0, + 1414.0, + 1306.0, + 885.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 1292.0, + 837.0, + 1292.0, + 837.0, + 1328.0, + 378.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 1292.0, + 1402.0, + 1292.0, + 1402.0, + 1328.0, + 1052.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1325.0, + 1406.0, + 1325.0, + 1406.0, + 1361.0, + 294.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1355.0, + 1261.0, + 1355.0, + 1261.0, + 1391.0, + 295.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1095.0, + 731.0, + 1095.0, + 731.0, + 1150.0, + 289.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 1095.0, + 1300.0, + 1095.0, + 1300.0, + 1150.0, + 948.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 1095.0, + 1410.0, + 1095.0, + 1410.0, + 1150.0, + 1404.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1139.0, + 299.0, + 1139.0, + 299.0, + 1211.0, + 294.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 1139.0, + 1405.0, + 1139.0, + 1405.0, + 1211.0, + 508.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1194.0, + 1405.0, + 1194.0, + 1405.0, + 1233.0, + 292.0, + 1233.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 751, + 258, + 1424, + 258, + 1424, + 632, + 751, + 632 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 300, + 905, + 1404, + 905, + 1404, + 1027, + 300, + 1027 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 298, + 739, + 1405, + 739, + 1405, + 832, + 298, + 832 + ], + "score": 0.968 + }, + { + "category_id": 3, + "poly": [ + 301, + 271, + 708, + 271, + 708, + 567, + 301, + 567 + ], + "score": 0.964 + }, + { + "category_id": 4, + "poly": [ + 296, + 582, + 721, + 582, + 721, + 706, + 296, + 706 + ], + "score": 0.964 + }, + { + "category_id": 4, + "poly": [ + 731, + 644, + 1409, + 644, + 1409, + 713, + 731, + 713 + ], + "score": 0.947 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 816, + 76, + 816, + 104, + 299, + 104 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 300, + 862, + 529, + 862, + 529, + 894, + 300, + 894 + ], + "score": 0.887 + }, + { + "category_id": 0, + "poly": [ + 299, + 1073, + 488, + 1073, + 488, + 1106, + 299, + 1106 + ], + "score": 0.8 + }, + { + "category_id": 1, + "poly": [ + 295, + 1122, + 1401, + 1122, + 1401, + 1185, + 295, + 1185 + ], + "score": 0.769 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.765 + }, + { + "category_id": 1, + "poly": [ + 297, + 1202, + 1400, + 1202, + 1400, + 1296, + 297, + 1296 + ], + "score": 0.661 + }, + { + "category_id": 1, + "poly": [ + 297, + 1506, + 1399, + 1506, + 1399, + 1570, + 297, + 1570 + ], + "score": 0.637 + }, + { + "category_id": 1, + "poly": [ + 297, + 1972, + 1397, + 1972, + 1397, + 2035, + 297, + 2035 + ], + "score": 0.623 + }, + { + "category_id": 1, + "poly": [ + 293, + 1425, + 1399, + 1425, + 1399, + 1489, + 293, + 1489 + ], + "score": 0.595 + }, + { + "category_id": 1, + "poly": [ + 298, + 1314, + 1400, + 1314, + 1400, + 1408, + 298, + 1408 + ], + "score": 0.59 + }, + { + "category_id": 1, + "poly": [ + 296, + 1586, + 1402, + 1586, + 1402, + 1682, + 296, + 1682 + ], + "score": 0.583 + }, + { + "category_id": 1, + "poly": [ + 297, + 1698, + 1401, + 1698, + 1401, + 1792, + 297, + 1792 + ], + "score": 0.575 + }, + { + "category_id": 1, + "poly": [ + 294, + 1809, + 1401, + 1809, + 1401, + 1874, + 294, + 1874 + ], + "score": 0.563 + }, + { + "category_id": 1, + "poly": [ + 296, + 1890, + 1398, + 1890, + 1398, + 1955, + 296, + 1955 + ], + "score": 0.541 + }, + { + "category_id": 13, + "poly": [ + 1047, + 643, + 1135, + 643, + 1135, + 684, + 1047, + 684 + ], + "score": 0.92, + "latex": "\\mathrm { V a r } [ \\frac { 1 } { v _ { t } } ]" + }, + { + "category_id": 13, + "poly": [ + 1186, + 645, + 1274, + 645, + 1274, + 684, + 1186, + 684 + ], + "score": 0.91, + "latex": "\\textstyle \\operatorname { V a r } [ { \\frac { c _ { t } } { v _ { t } } } ]" + }, + { + "category_id": 13, + "poly": [ + 651, + 649, + 670, + 649, + 670, + 676, + 651, + 676 + ], + "score": 0.81, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 989, + 309, + 1022, + 309, + 1022, + 324, + 989, + 324 + ], + "score": 0.66, + "latex": "1 0 ^ { - 2 } \\nmid" + }, + { + "category_id": 14, + "poly": [ + 1278, + 599, + 1367, + 599, + 1367, + 633, + 1278, + 633 + ], + "score": 0.63, + "latex": "\\mu = 1 0" + }, + { + "category_id": 13, + "poly": [ + 1336, + 651, + 1354, + 651, + 1354, + 675, + 1336, + 675 + ], + "score": 0.62, + "latex": "\\mathbf { X }" + }, + { + "category_id": 14, + "poly": [ + 805, + 599, + 901, + 599, + 901, + 633, + 805, + 633 + ], + "score": 0.61, + "latex": "\\mu = 0 . 1" + }, + { + "category_id": 13, + "poly": [ + 1270, + 420, + 1375, + 420, + 1375, + 446, + 1270, + 446 + ], + "score": 0.59, + "latex": "\\mu = 0 . 0 1" + }, + { + "category_id": 13, + "poly": [ + 1028, + 421, + 1146, + 421, + 1146, + 446, + 1028, + 446 + ], + "score": 0.56, + "latex": "\\mu = 0 . 0 0 1" + }, + { + "category_id": 13, + "poly": [ + 545, + 649, + 563, + 649, + 563, + 671, + 545, + 671 + ], + "score": 0.56, + "latex": "\\mathbf { X }" + }, + { + "category_id": 13, + "poly": [ + 306, + 367, + 347, + 367, + 347, + 387, + 306, + 387 + ], + "score": 0.54, + "latex": "1 0 ^ { - 2 }" + }, + { + "category_id": 13, + "poly": [ + 305, + 412, + 347, + 412, + 347, + 432, + 305, + 432 + ], + "score": 0.52, + "latex": "1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 1366, + 275, + 1407, + 275, + 1407, + 296, + 1366, + 296 + ], + "score": 0.51, + "latex": "\\begin{array} { r } { \\mathrm { V a r } [ \\frac { 1 } { v _ { t } } ] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 989, + 271, + 1019, + 271, + 1019, + 287, + 989, + 287 + ], + "score": 0.48, + "latex": "1 0 ^ { - 1 } \\cdot" + }, + { + "category_id": 13, + "poly": [ + 1366, + 303, + 1406, + 303, + 1406, + 322, + 1366, + 322 + ], + "score": 0.47, + "latex": "\\mathrm { V a r } [ \\textstyle { \\frac { c _ { i } } { v _ { t } } } ]" + }, + { + "category_id": 13, + "poly": [ + 988, + 347, + 1019, + 347, + 1019, + 362, + 988, + 362 + ], + "score": 0.42, + "latex": "1 0 ^ { - 3 } \\cdot" + }, + { + "category_id": 13, + "poly": [ + 1224, + 310, + 1257, + 310, + 1257, + 325, + 1224, + 325 + ], + "score": 0.39, + "latex": "1 0 ^ { - 2 } \\sharp" + }, + { + "category_id": 13, + "poly": [ + 306, + 321, + 348, + 321, + 348, + 341, + 306, + 341 + ], + "score": 0.38, + "latex": "1 0 ^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 754, + 346, + 786, + 346, + 786, + 361, + 754, + 361 + ], + "score": 0.37, + "latex": "1 0 ^ { - 3 } \\sharp" + }, + { + "category_id": 14, + "poly": [ + 1050, + 599, + 1125, + 599, + 1125, + 634, + 1050, + 634 + ], + "score": 0.32, + "latex": "\\mu = 1" + }, + { + "category_id": 13, + "poly": [ + 1224, + 274, + 1252, + 274, + 1252, + 287, + 1224, + 287 + ], + "score": 0.31, + "latex": "1 0 ^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1224, + 493, + 1257, + 493, + 1257, + 509, + 1224, + 509 + ], + "score": 0.29, + "latex": "1 0 ^ { - 6 } \\ddagger" + }, + { + "category_id": 13, + "poly": [ + 1257, + 391, + 1266, + 391, + 1266, + 403, + 1257, + 403 + ], + "score": 0.26, + "latex": "0" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 265.0, + 789.0, + 265.0, + 789.0, + 291.0, + 750.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 267.0, + 988.0, + 267.0, + 988.0, + 289.0, + 982.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 266.0, + 1223.0, + 266.0, + 1223.0, + 292.0, + 1218.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 266.0, + 1258.0, + 266.0, + 1258.0, + 292.0, + 1253.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 265.0, + 1365.0, + 265.0, + 1365.0, + 305.0, + 1356.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1408.0, + 265.0, + 1416.0, + 265.0, + 1416.0, + 305.0, + 1408.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 301.0, + 789.0, + 301.0, + 789.0, + 328.0, + 747.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 302.0, + 988.0, + 302.0, + 988.0, + 329.0, + 981.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 303.0, + 1223.0, + 303.0, + 1223.0, + 332.0, + 1217.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1357.0, + 290.0, + 1365.0, + 290.0, + 1365.0, + 329.0, + 1357.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1407.0, + 290.0, + 1415.0, + 290.0, + 1415.0, + 329.0, + 1407.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 338.0, + 753.0, + 338.0, + 753.0, + 367.0, + 748.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 338.0, + 790.0, + 338.0, + 790.0, + 367.0, + 787.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 341.0, + 1258.0, + 341.0, + 1258.0, + 366.0, + 1216.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 387.0, + 802.0, + 387.0, + 802.0, + 405.0, + 785.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 382.0, + 877.0, + 382.0, + 877.0, + 409.0, + 833.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 381.0, + 937.0, + 381.0, + 937.0, + 409.0, + 897.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 388.0, + 1035.0, + 388.0, + 1035.0, + 405.0, + 1019.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 384.0, + 1111.0, + 384.0, + 1111.0, + 408.0, + 1068.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 381.0, + 1173.0, + 381.0, + 1173.0, + 409.0, + 1131.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 388.0, + 1256.0, + 388.0, + 1256.0, + 406.0, + 1253.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 388.0, + 1272.0, + 388.0, + 1272.0, + 406.0, + 1267.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 383.0, + 1344.0, + 383.0, + 1344.0, + 408.0, + 1303.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 384.0, + 1407.0, + 384.0, + 1407.0, + 408.0, + 1364.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 413.0, + 896.0, + 413.0, + 896.0, + 446.0, + 807.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 413.0, + 1027.0, + 413.0, + 1027.0, + 445.0, + 1019.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 413.0, + 1153.0, + 413.0, + 1153.0, + 445.0, + 1147.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 412.0, + 1269.0, + 412.0, + 1269.0, + 445.0, + 1261.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 412.0, + 1379.0, + 412.0, + 1379.0, + 445.0, + 1376.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 451.0, + 790.0, + 451.0, + 790.0, + 480.0, + 748.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 451.0, + 1260.0, + 451.0, + 1260.0, + 477.0, + 1218.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 472.0, + 1024.0, + 472.0, + 1024.0, + 496.0, + 980.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 487.0, + 791.0, + 487.0, + 791.0, + 517.0, + 748.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 486.0, + 1223.0, + 486.0, + 1223.0, + 516.0, + 1217.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 486.0, + 1261.0, + 486.0, + 1261.0, + 516.0, + 1258.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 509.0, + 1025.0, + 509.0, + 1025.0, + 535.0, + 981.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 526.0, + 790.0, + 526.0, + 790.0, + 555.0, + 748.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1214.0, + 527.0, + 1256.0, + 527.0, + 1256.0, + 551.0, + 1214.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 547.0, + 1022.0, + 547.0, + 1022.0, + 572.0, + 984.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 573.0, + 802.0, + 573.0, + 802.0, + 591.0, + 785.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 569.0, + 876.0, + 569.0, + 876.0, + 594.0, + 834.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 570.0, + 937.0, + 570.0, + 937.0, + 593.0, + 897.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 573.0, + 1035.0, + 573.0, + 1035.0, + 590.0, + 1019.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 569.0, + 1112.0, + 569.0, + 1112.0, + 594.0, + 1068.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 570.0, + 1171.0, + 570.0, + 1171.0, + 593.0, + 1132.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 573.0, + 1271.0, + 573.0, + 1271.0, + 590.0, + 1254.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 569.0, + 1344.0, + 569.0, + 1344.0, + 594.0, + 1303.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1365.0, + 571.0, + 1404.0, + 571.0, + 1404.0, + 592.0, + 1365.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 600.0, + 804.0, + 600.0, + 804.0, + 631.0, + 799.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 601.0, + 1277.0, + 601.0, + 1277.0, + 631.0, + 1272.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 601.0, + 1371.0, + 601.0, + 1371.0, + 631.0, + 1368.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.25, + 338.5, + 1029.25, + 338.5, + 1029.25, + 367.5, + 979.25, + 367.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.25, + 598.0, + 1132.25, + 598.0, + 1132.25, + 635.0, + 1042.25, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 268.0, + 355.0, + 268.0, + 355.0, + 303.0, + 312.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 284.0, + 380.0, + 284.0, + 380.0, + 386.0, + 349.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 286.0, + 606.0, + 286.0, + 606.0, + 320.0, + 503.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 312.0, + 305.0, + 312.0, + 305.0, + 346.0, + 295.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 312.0, + 354.0, + 312.0, + 354.0, + 346.0, + 349.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 328.0, + 480.0, + 328.0, + 480.0, + 340.0, + 466.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 317.0, + 595.0, + 317.0, + 595.0, + 354.0, + 503.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 347.0, + 692.0, + 347.0, + 692.0, + 384.0, + 504.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 405.0, + 304.0, + 405.0, + 304.0, + 438.0, + 295.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 405.0, + 356.0, + 405.0, + 356.0, + 438.0, + 348.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 379.0, + 442.0, + 379.0, + 442.0, + 401.0, + 420.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 386.0, + 475.0, + 386.0, + 475.0, + 409.0, + 453.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 393.0, + 505.0, + 393.0, + 505.0, + 412.0, + 488.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 397.0, + 540.0, + 397.0, + 540.0, + 419.0, + 517.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 401.0, + 572.0, + 401.0, + 572.0, + 422.0, + 551.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 404.0, + 605.0, + 404.0, + 605.0, + 426.0, + 583.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 408.0, + 636.0, + 408.0, + 636.0, + 428.0, + 615.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 411.0, + 669.0, + 411.0, + 669.0, + 430.0, + 650.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 412.0, + 703.0, + 412.0, + 703.0, + 433.0, + 681.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 496.0, + 356.0, + 496.0, + 356.0, + 529.0, + 295.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 541.0, + 378.0, + 541.0, + 378.0, + 569.0, + 355.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 538.0, + 519.0, + 538.0, + 519.0, + 573.0, + 473.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 536.0, + 650.0, + 536.0, + 650.0, + 574.0, + 602.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 355.0, + 361.0, + 355.0, + 361.0, + 395.5, + 294.0, + 395.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 446.0, + 361.0, + 446.0, + 361.0, + 487.5, + 295.0, + 487.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 579.0, + 722.0, + 579.0, + 722.0, + 616.0, + 294.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 612.0, + 722.0, + 612.0, + 722.0, + 644.0, + 296.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 643.0, + 544.0, + 643.0, + 544.0, + 675.0, + 296.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 643.0, + 650.0, + 643.0, + 650.0, + 675.0, + 564.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 643.0, + 723.0, + 643.0, + 723.0, + 675.0, + 671.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 672.0, + 710.0, + 672.0, + 710.0, + 709.0, + 294.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 642.0, + 1046.0, + 642.0, + 1046.0, + 685.0, + 729.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 642.0, + 1185.0, + 642.0, + 1185.0, + 685.0, + 1136.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1275.0, + 642.0, + 1335.0, + 642.0, + 1335.0, + 685.0, + 1275.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 642.0, + 1412.0, + 642.0, + 1412.0, + 685.0, + 1355.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 676.0, + 1396.0, + 676.0, + 1396.0, + 717.0, + 730.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 859.0, + 533.0, + 859.0, + 533.0, + 902.0, + 296.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1073.0, + 490.0, + 1073.0, + 490.0, + 1110.0, + 297.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 905.0, + 1403.0, + 905.0, + 1403.0, + 937.0, + 296.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 936.0, + 1405.0, + 936.0, + 1405.0, + 972.0, + 294.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 966.0, + 1404.0, + 966.0, + 1404.0, + 998.0, + 294.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 998.0, + 974.0, + 998.0, + 974.0, + 1027.0, + 298.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 740.0, + 1403.0, + 740.0, + 1403.0, + 773.0, + 295.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 770.0, + 1403.0, + 770.0, + 1403.0, + 804.0, + 295.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 801.0, + 1293.0, + 801.0, + 1293.0, + 835.0, + 295.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1119.0, + 1406.0, + 1119.0, + 1406.0, + 1160.0, + 293.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1152.0, + 612.0, + 1152.0, + 612.0, + 1185.0, + 322.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1199.0, + 1404.0, + 1199.0, + 1404.0, + 1241.0, + 293.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1231.0, + 1405.0, + 1231.0, + 1405.0, + 1271.0, + 321.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1259.0, + 400.0, + 1259.0, + 400.0, + 1300.0, + 320.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1502.0, + 1403.0, + 1502.0, + 1403.0, + 1546.0, + 295.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1538.0, + 804.0, + 1538.0, + 804.0, + 1570.0, + 325.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1967.0, + 1403.0, + 1967.0, + 1403.0, + 2011.0, + 292.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2001.0, + 1196.0, + 2001.0, + 1196.0, + 2035.0, + 322.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1422.0, + 1404.0, + 1422.0, + 1404.0, + 1463.0, + 292.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1456.0, + 1088.0, + 1456.0, + 1088.0, + 1492.0, + 323.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1311.0, + 1405.0, + 1311.0, + 1405.0, + 1353.0, + 294.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1344.0, + 1405.0, + 1344.0, + 1405.0, + 1382.0, + 321.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1377.0, + 781.0, + 1377.0, + 781.0, + 1411.0, + 324.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1585.0, + 1404.0, + 1585.0, + 1404.0, + 1623.0, + 294.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1617.0, + 1404.0, + 1617.0, + 1404.0, + 1654.0, + 321.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1649.0, + 746.0, + 1649.0, + 746.0, + 1683.0, + 323.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1698.0, + 1406.0, + 1698.0, + 1406.0, + 1734.0, + 295.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1730.0, + 1405.0, + 1730.0, + 1405.0, + 1764.0, + 324.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1761.0, + 691.0, + 1761.0, + 691.0, + 1795.0, + 324.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1808.0, + 1406.0, + 1808.0, + 1406.0, + 1848.0, + 294.0, + 1848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1840.0, + 1123.0, + 1840.0, + 1123.0, + 1875.0, + 322.0, + 1875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1890.0, + 1402.0, + 1890.0, + 1402.0, + 1926.0, + 295.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1921.0, + 831.0, + 1921.0, + 831.0, + 1957.0, + 322.0, + 1957.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 866, + 2088, + 866, + 2113, + 836, + 2113 + ], + "score": 0.833 + }, + { + "category_id": 1, + "poly": [ + 295, + 228, + 1066, + 228, + 1066, + 264, + 295, + 264 + ], + "score": 0.811 + }, + { + "category_id": 1, + "poly": [ + 295, + 867, + 1379, + 867, + 1379, + 902, + 295, + 902 + ], + "score": 0.753 + }, + { + "category_id": 1, + "poly": [ + 296, + 1314, + 906, + 1314, + 906, + 1348, + 296, + 1348 + ], + "score": 0.751 + }, + { + "category_id": 1, + "poly": [ + 293, + 916, + 1399, + 916, + 1399, + 980, + 293, + 980 + ], + "score": 0.728 + }, + { + "category_id": 1, + "poly": [ + 293, + 995, + 1398, + 995, + 1398, + 1061, + 293, + 1061 + ], + "score": 0.704 + }, + { + "category_id": 1, + "poly": [ + 298, + 787, + 1396, + 787, + 1396, + 852, + 298, + 852 + ], + "score": 0.687 + }, + { + "category_id": 1, + "poly": [ + 294, + 1683, + 1399, + 1683, + 1399, + 1748, + 294, + 1748 + ], + "score": 0.687 + }, + { + "category_id": 1, + "poly": [ + 296, + 1185, + 1353, + 1185, + 1353, + 1221, + 296, + 1221 + ], + "score": 0.679 + }, + { + "category_id": 1, + "poly": [ + 297, + 627, + 1399, + 627, + 1399, + 692, + 297, + 692 + ], + "score": 0.675 + }, + { + "category_id": 1, + "poly": [ + 295, + 1762, + 1244, + 1762, + 1244, + 1796, + 295, + 1796 + ], + "score": 0.671 + }, + { + "category_id": 1, + "poly": [ + 297, + 1522, + 1402, + 1522, + 1402, + 1587, + 297, + 1587 + ], + "score": 0.662 + }, + { + "category_id": 1, + "poly": [ + 299, + 1234, + 1398, + 1234, + 1398, + 1299, + 299, + 1299 + ], + "score": 0.656 + }, + { + "category_id": 1, + "poly": [ + 293, + 1602, + 1401, + 1602, + 1401, + 1667, + 293, + 1667 + ], + "score": 0.655 + }, + { + "category_id": 1, + "poly": [ + 296, + 278, + 1401, + 278, + 1401, + 343, + 296, + 343 + ], + "score": 0.642 + }, + { + "category_id": 1, + "poly": [ + 296, + 1812, + 1403, + 1812, + 1403, + 1876, + 296, + 1876 + ], + "score": 0.637 + }, + { + "category_id": 1, + "poly": [ + 303, + 1074, + 1399, + 1074, + 1399, + 1170, + 303, + 1170 + ], + "score": 0.636 + }, + { + "category_id": 1, + "poly": [ + 297, + 437, + 1400, + 437, + 1400, + 502, + 297, + 502 + ], + "score": 0.633 + }, + { + "category_id": 1, + "poly": [ + 297, + 706, + 1399, + 706, + 1399, + 773, + 297, + 773 + ], + "score": 0.631 + }, + { + "category_id": 1, + "poly": [ + 299, + 516, + 1402, + 516, + 1402, + 612, + 299, + 612 + ], + "score": 0.631 + }, + { + "category_id": 1, + "poly": [ + 297, + 1364, + 1401, + 1364, + 1401, + 1429, + 297, + 1429 + ], + "score": 0.621 + }, + { + "category_id": 1, + "poly": [ + 295, + 1443, + 1403, + 1443, + 1403, + 1508, + 295, + 1508 + ], + "score": 0.602 + }, + { + "category_id": 1, + "poly": [ + 295, + 1891, + 1402, + 1891, + 1402, + 1957, + 295, + 1957 + ], + "score": 0.583 + }, + { + "category_id": 1, + "poly": [ + 298, + 1971, + 1398, + 1971, + 1398, + 2035, + 298, + 2035 + ], + "score": 0.566 + }, + { + "category_id": 1, + "poly": [ + 300, + 357, + 1400, + 357, + 1400, + 423, + 300, + 423 + ], + "score": 0.565 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 229.0, + 1067.0, + 229.0, + 1067.0, + 266.0, + 297.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 865.0, + 1382.0, + 865.0, + 1382.0, + 904.0, + 296.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1312.0, + 908.0, + 1312.0, + 908.0, + 1352.0, + 294.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 911.0, + 1404.0, + 911.0, + 1404.0, + 956.0, + 293.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 948.0, + 693.0, + 948.0, + 693.0, + 980.0, + 325.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 993.0, + 1403.0, + 993.0, + 1403.0, + 1035.0, + 293.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1027.0, + 1156.0, + 1027.0, + 1156.0, + 1062.0, + 322.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 783.0, + 1401.0, + 783.0, + 1401.0, + 827.0, + 294.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 818.0, + 872.0, + 818.0, + 872.0, + 850.0, + 322.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1681.0, + 1404.0, + 1681.0, + 1404.0, + 1721.0, + 294.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1713.0, + 1198.0, + 1713.0, + 1198.0, + 1749.0, + 322.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1186.0, + 1355.0, + 1186.0, + 1355.0, + 1221.0, + 294.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 621.0, + 1404.0, + 621.0, + 1404.0, + 670.0, + 292.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 657.0, + 590.0, + 657.0, + 590.0, + 693.0, + 323.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1760.0, + 1240.0, + 1760.0, + 1240.0, + 1798.0, + 294.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1521.0, + 1406.0, + 1521.0, + 1406.0, + 1561.0, + 293.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1554.0, + 468.0, + 1554.0, + 468.0, + 1588.0, + 322.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1233.0, + 1404.0, + 1233.0, + 1404.0, + 1273.0, + 294.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1265.0, + 761.0, + 1265.0, + 761.0, + 1300.0, + 324.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1598.0, + 1406.0, + 1598.0, + 1406.0, + 1644.0, + 293.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1633.0, + 1031.0, + 1633.0, + 1031.0, + 1668.0, + 322.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 278.0, + 1405.0, + 278.0, + 1405.0, + 314.0, + 295.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 312.0, + 779.0, + 312.0, + 779.0, + 341.0, + 324.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1812.0, + 1405.0, + 1812.0, + 1405.0, + 1848.0, + 295.0, + 1848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1844.0, + 1366.0, + 1844.0, + 1366.0, + 1877.0, + 321.0, + 1877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1074.0, + 1404.0, + 1074.0, + 1404.0, + 1112.0, + 297.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1105.0, + 1404.0, + 1105.0, + 1404.0, + 1145.0, + 320.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1139.0, + 894.0, + 1139.0, + 894.0, + 1171.0, + 323.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 436.0, + 1406.0, + 436.0, + 1406.0, + 473.0, + 295.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 467.0, + 1308.0, + 467.0, + 1308.0, + 507.0, + 322.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 702.0, + 1404.0, + 702.0, + 1404.0, + 747.0, + 295.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 740.0, + 1081.0, + 740.0, + 1081.0, + 772.0, + 325.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 515.0, + 1403.0, + 515.0, + 1403.0, + 554.0, + 293.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 548.0, + 1404.0, + 548.0, + 1404.0, + 585.0, + 323.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 580.0, + 891.0, + 580.0, + 891.0, + 614.0, + 323.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1360.0, + 1406.0, + 1360.0, + 1406.0, + 1404.0, + 292.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1393.0, + 1374.0, + 1393.0, + 1374.0, + 1431.0, + 321.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1439.0, + 1408.0, + 1439.0, + 1408.0, + 1482.0, + 292.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1474.0, + 854.0, + 1474.0, + 854.0, + 1509.0, + 323.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1886.0, + 1407.0, + 1886.0, + 1407.0, + 1928.0, + 292.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1919.0, + 400.0, + 1919.0, + 400.0, + 1958.0, + 318.0, + 1958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1403.0, + 1971.0, + 1403.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2003.0, + 670.0, + 2003.0, + 670.0, + 2034.0, + 323.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 356.0, + 1403.0, + 356.0, + 1403.0, + 394.0, + 295.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 391.0, + 1115.0, + 391.0, + 1115.0, + 421.0, + 322.0, + 421.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 8, + "poly": [ + 560, + 1102, + 1144, + 1102, + 1144, + 1238, + 560, + 1238 + ], + "score": 0.969 + }, + { + "category_id": 8, + "poly": [ + 430, + 770, + 1270, + 770, + 1270, + 1033, + 430, + 1033 + ], + "score": 0.968 + }, + { + "category_id": 8, + "poly": [ + 436, + 1385, + 1268, + 1385, + 1268, + 1689, + 436, + 1689 + ], + "score": 0.967 + }, + { + "category_id": 8, + "poly": [ + 513, + 1875, + 1185, + 1875, + 1185, + 2031, + 513, + 2031 + ], + "score": 0.966 + }, + { + "category_id": 8, + "poly": [ + 653, + 1767, + 1044, + 1767, + 1044, + 1832, + 653, + 1832 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 526, + 1273, + 1171, + 1273, + 1171, + 1345, + 526, + 1345 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 298, + 1844, + 545, + 1844, + 545, + 1874, + 298, + 1874 + ], + "score": 0.933 + }, + { + "category_id": 8, + "poly": [ + 510, + 441, + 1186, + 441, + 1186, + 515, + 510, + 515 + ], + "score": 0.932 + }, + { + "category_id": 1, + "poly": [ + 295, + 660, + 953, + 660, + 953, + 709, + 295, + 709 + ], + "score": 0.929 + }, + { + "category_id": 1, + "poly": [ + 298, + 1071, + 652, + 1071, + 652, + 1103, + 298, + 1103 + ], + "score": 0.929 + }, + { + "category_id": 1, + "poly": [ + 298, + 1357, + 546, + 1357, + 546, + 1387, + 298, + 1387 + ], + "score": 0.918 + }, + { + "category_id": 1, + "poly": [ + 295, + 1240, + 1350, + 1240, + 1350, + 1272, + 295, + 1272 + ], + "score": 0.917 + }, + { + "category_id": 8, + "poly": [ + 458, + 332, + 1238, + 332, + 1238, + 410, + 458, + 410 + ], + "score": 0.916 + }, + { + "category_id": 1, + "poly": [ + 298, + 738, + 824, + 738, + 824, + 769, + 298, + 769 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 818, + 73, + 818, + 106, + 297, + 106 + ], + "score": 0.914 + }, + { + "category_id": 0, + "poly": [ + 300, + 225, + 680, + 225, + 680, + 263, + 300, + 263 + ], + "score": 0.91 + }, + { + "category_id": 1, + "poly": [ + 296, + 514, + 850, + 514, + 850, + 547, + 296, + 547 + ], + "score": 0.908 + }, + { + "category_id": 1, + "poly": [ + 299, + 1737, + 667, + 1737, + 667, + 1768, + 299, + 1768 + ], + "score": 0.907 + }, + { + "category_id": 1, + "poly": [ + 289, + 291, + 1314, + 291, + 1314, + 330, + 289, + 330 + ], + "score": 0.903 + }, + { + "category_id": 1, + "poly": [ + 301, + 410, + 913, + 410, + 913, + 440, + 301, + 440 + ], + "score": 0.901 + }, + { + "category_id": 1, + "poly": [ + 297, + 1690, + 1190, + 1690, + 1190, + 1726, + 297, + 1726 + ], + "score": 0.9 + }, + { + "category_id": 1, + "poly": [ + 300, + 615, + 1374, + 615, + 1374, + 648, + 300, + 648 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1365, + 461, + 1401, + 461, + 1401, + 492, + 1365, + 492 + ], + "score": 0.862 + }, + { + "category_id": 9, + "poly": [ + 1365, + 360, + 1401, + 360, + 1401, + 389, + 1365, + 389 + ], + "score": 0.861 + }, + { + "category_id": 8, + "poly": [ + 368, + 548, + 1322, + 548, + 1322, + 612, + 368, + 612 + ], + "score": 0.86 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.855 + }, + { + "category_id": 9, + "poly": [ + 1365, + 566, + 1401, + 566, + 1401, + 595, + 1365, + 595 + ], + "score": 0.84 + }, + { + "category_id": 14, + "poly": [ + 428, + 766, + 1266, + 766, + 1266, + 1033, + 428, + 1033 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\frac { \\partial } { \\partial t } ( \\frac { t } { t - 2 } - \\frac { t \\dot { 2 } ^ { 2 t - 5 } } { \\pi } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } ) } \\\\ & { = \\frac { - 2 } { ( t - 2 ) ^ { 2 } } - \\frac { 2 ^ { 2 t - 5 } } { \\pi } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } - \\frac { t 2 ^ { 2 t - 5 } \\ln { 4 } } { \\pi } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } \\\\ & { - \\frac { 2 t 2 ^ { 2 t - 5 } } { \\pi } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } ( \\Psi ( \\frac { t - 1 } { 2 } ) - \\Psi ( t - 1 ) ) , \\quad \\left( \\Psi ( x ) = \\frac { \\Gamma ^ { \\prime } ( x ) } { \\Gamma ( x ) } \\right) } \\\\ & { < 0 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1144, + 294, + 1257, + 294, + 1257, + 329, + 1144, + 329 + ], + "score": 0.93, + "latex": "\\mathcal { X } ^ { 2 } ( \\rho , \\tau ^ { 2 } )" + }, + { + "category_id": 14, + "poly": [ + 432, + 1382, + 1270, + 1382, + 1270, + 1723, + 432, + 1723 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } + 1 + t \\Psi ( \\frac { t - 1 } { 2 } ) - t \\Psi ( \\frac { t } { 2 } ) } \\\\ & { \\geq \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } + 2 + t ( \\ln ( t / 2 ) - 1 / ( t / 2 - 0 . 5 ) ) - t \\ln ( t / 2 ) } \\\\ & { = \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } - \\frac { 2 } { t - 1 } } \\\\ & { > \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } - \\frac { 2 } { t - 2 } \\geq 0 , } \\\\ & { \\mathrm { r s t i n e q u a l i t y i s i f r o m } \\ln ( x ) - 1 / ( 2 x ) > \\Psi ( x ) > \\ln ( x + 0 . 5 ) - 1 / x . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 653, + 1772, + 1041, + 1772, + 1041, + 1829, + 653, + 1829 + ], + "score": 0.92, + "latex": "3 2 \\pi \\geq ( t - 2 ) 4 ^ { t } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } ," + }, + { + "category_id": 14, + "poly": [ + 376, + 540, + 1321, + 540, + 1321, + 616, + 376, + 616 + ], + "score": 0.92, + "latex": "\\operatorname { V a r } [ \\psi ( \\cdot ) ] = \\operatorname { V a r } [ { \\sqrt { x } } ] = \\operatorname { \\mathbb { E } } [ x ] - \\operatorname { \\mathbb { E } } [ { \\sqrt { x } } ] ^ { 2 } = \\tau ^ { 2 } ( { \\frac { \\rho } { \\rho - 2 } } - { \\frac { \\rho 2 ^ { 2 \\rho - 5 } } { \\pi } } B ( { \\frac { \\rho - 1 } { 2 } } , { \\frac { \\rho - 1 } { 2 } } ) ^ { 2 } ) ," + }, + { + "category_id": 14, + "poly": [ + 512, + 437, + 1186, + 437, + 1186, + 517, + 512, + 517 + ], + "score": 0.92, + "latex": "\\mathbb { E } [ \\sqrt { x } ] = \\int _ { 0 } ^ { \\infty } \\sqrt { x } p ( x ) d x = \\frac { \\tau \\sqrt { \\rho } \\Gamma ( ( \\rho - 1 ) / 2 ) } { \\sqrt { 2 } \\Gamma ( \\rho / 2 ) } ( \\forall \\rho > 4 ) ." + }, + { + "category_id": 13, + "poly": [ + 1056, + 615, + 1164, + 615, + 1164, + 648, + 1056, + 648 + ], + "score": 0.92, + "latex": "\\mathrm { V a r } [ \\psi ( . ) ]" + }, + { + "category_id": 13, + "poly": [ + 630, + 294, + 693, + 294, + 693, + 329, + 630, + 329 + ], + "score": 0.92, + "latex": "\\psi ^ { 2 } ( . )" + }, + { + "category_id": 14, + "poly": [ + 514, + 1875, + 1184, + 1875, + 1184, + 2036, + 514, + 2036 + ], + "score": 0.91, + "latex": "\\begin{array} { l } { ( t - 2 ) 4 ^ { t } \\mathcal { B } ( \\displaystyle \\frac { t - 1 } { 2 } , \\displaystyle \\frac { t - 1 } { 2 } ) ^ { 2 } = ( t - 2 ) 4 ^ { t } \\frac { \\Gamma ( \\displaystyle \\frac { t - 1 } { 2 } ) ^ { 4 } } { \\Gamma ( t - 1 ) ^ { 2 } } } \\\\ { \\overset { ( i ) } { = } ( t - 2 ) 4 ^ { t } \\frac { \\Gamma ( \\displaystyle \\frac { t - 1 } { 2 } ) ^ { 2 } } { \\Gamma ( t / 2 ) ^ { 2 } } 4 ^ { 2 - t } \\pi = 1 6 \\pi ( t - 2 ) \\frac { \\Gamma ( \\displaystyle \\frac { t - 1 } { 2 } ) ^ { 2 } } { \\Gamma ( t / 2 ) ^ { 2 } } \\le 3 2 \\pi , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 372, + 614, + 422, + 614, + 422, + 648, + 372, + 648 + ], + "score": 0.91, + "latex": "B ( . )" + }, + { + "category_id": 13, + "poly": [ + 647, + 516, + 736, + 516, + 736, + 547, + 647, + 547 + ], + "score": 0.91, + "latex": "\\forall \\rho > 4" + }, + { + "category_id": 13, + "poly": [ + 372, + 409, + 421, + 409, + 421, + 443, + 372, + 443 + ], + "score": 0.9, + "latex": "\\Gamma ( . )" + }, + { + "category_id": 14, + "poly": [ + 462, + 331, + 1238, + 331, + 1238, + 412, + 462, + 412 + ], + "score": 0.89, + "latex": "p ( x ) = \\frac { ( \\tau ^ { 2 } \\rho / 2 ) ^ { \\rho / 2 } } { \\Gamma ( \\rho / 2 ) } \\frac { \\exp \\bigl [ \\frac { - \\rho \\tau ^ { 2 } } { 2 x } \\bigr ] } { x ^ { 1 + \\rho / 2 } } \\quad \\mathrm { a n d } \\quad \\mathbb { E } [ x ] = \\frac { \\rho } { ( \\rho - 2 ) \\sigma ^ { 2 } } \\left( \\forall \\rho > 2 \\right)" + }, + { + "category_id": 13, + "poly": [ + 798, + 293, + 831, + 293, + 831, + 330, + 798, + 330 + ], + "score": 0.89, + "latex": "\\textstyle { \\frac { 1 } { \\sigma ^ { 2 } } }" + }, + { + "category_id": 14, + "poly": [ + 526, + 1268, + 1170, + 1268, + 1170, + 1344, + 526, + 1344 + ], + "score": 0.88, + "latex": "\\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } + 1 + t \\Psi ( \\frac { t - \\hat { 1 } } { 2 } ) - t \\Psi ( \\frac { t } { 2 } ) > 0 ," + }, + { + "category_id": 14, + "poly": [ + 557, + 1099, + 1141, + 1099, + 1141, + 1242, + 557, + 1242 + ], + "score": 0.88, + "latex": "\\begin{array} { r l r } & { } & { \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } + 1 + t \\ln 4 + 2 t \\Psi ( \\frac { t - 1 } { 2 } ) } \\\\ & { } & { > 2 t \\Psi ( t - 1 ) \\overset { ( i ) } { = } t [ \\Psi ( \\frac { t - 1 } { 2 } ) + \\Psi ( \\frac { t } { 2 } ) + \\ln 4 ] , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 865, + 294, + 895, + 294, + 895, + 323, + 865, + 323 + ], + "score": 0.86, + "latex": "\\tau ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 373, + 1240, + 405, + 1240, + 405, + 1273, + 373, + 1273 + ], + "score": 0.84, + "latex": "( i )" + }, + { + "category_id": 13, + "poly": [ + 727, + 302, + 746, + 302, + 746, + 323, + 727, + 323 + ], + "score": 0.74, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 533, + 660, + 951, + 660, + 951, + 709, + 533, + 709 + ], + "score": 0.71, + "latex": "\\begin{array} { r } { , \\frac { \\partial } { \\partial t } \\big ( \\frac { t } { t - 2 } - \\frac { t 2 ^ { 2 t - 5 } } { \\pi } B ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } \\big ) < 0 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 469, + 666, + 533, + 666, + 533, + 702, + 469, + 702 + ], + "score": 0.66, + "latex": "t \\geq 4" + }, + { + "category_id": 13, + "poly": [ + 666, + 1694, + 1187, + 1694, + 1187, + 1724, + 666, + 1724 + ], + "score": 0.44, + "latex": "\\ln ( x ) - 1 / ( 2 x ) > \\Psi ( x ) > \\ln ( x + 0 . 5 ) - 1 / x ." + }, + { + "category_id": 13, + "poly": [ + 977, + 298, + 1027, + 298, + 1027, + 324, + 977, + 324 + ], + "score": 0.37, + "latex": "x \\sim" + }, + { + "category_id": 13, + "poly": [ + 442, + 1619, + 899, + 1619, + 899, + 1692, + 442, + 1692 + ], + "score": 0.34, + "latex": "> \\frac { 6 4 \\pi } { ( t - 2 ) ^ { 2 } 4 ^ { t } \\mathcal { B } ( \\frac { t - 1 } { 2 } , \\frac { t - 1 } { 2 } ) ^ { 2 } } - \\frac { 2 } { t - 2 } \\ge 0 ," + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 224.0, + 682.0, + 224.0, + 682.0, + 268.0, + 295.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1839.0, + 548.0, + 1839.0, + 548.0, + 1879.0, + 296.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 282.0, + 655.0, + 468.0, + 655.0, + 468.0, + 716.0, + 282.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 655.0, + 960.0, + 655.0, + 960.0, + 716.0, + 952.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1066.0, + 654.0, + 1066.0, + 654.0, + 1109.0, + 295.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1352.0, + 550.0, + 1352.0, + 550.0, + 1392.0, + 295.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1233.0, + 372.0, + 1233.0, + 372.0, + 1279.0, + 293.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1233.0, + 1356.0, + 1233.0, + 1356.0, + 1279.0, + 406.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 733.0, + 829.0, + 733.0, + 829.0, + 777.0, + 293.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 513.0, + 646.0, + 513.0, + 646.0, + 549.0, + 296.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 513.0, + 851.0, + 513.0, + 851.0, + 549.0, + 737.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1736.0, + 666.0, + 1736.0, + 666.0, + 1772.0, + 296.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 289.0, + 629.0, + 289.0, + 629.0, + 333.0, + 293.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 289.0, + 726.0, + 289.0, + 726.0, + 333.0, + 694.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 289.0, + 797.0, + 289.0, + 797.0, + 333.0, + 747.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 289.0, + 864.0, + 289.0, + 864.0, + 333.0, + 832.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 289.0, + 976.0, + 289.0, + 976.0, + 333.0, + 896.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 289.0, + 1143.0, + 289.0, + 1143.0, + 333.0, + 1028.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 289.0, + 1317.0, + 289.0, + 1317.0, + 333.0, + 1258.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 407.0, + 371.0, + 407.0, + 371.0, + 445.0, + 297.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 407.0, + 915.0, + 407.0, + 915.0, + 445.0, + 422.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1685.0, + 431.0, + 1685.0, + 431.0, + 1729.0, + 294.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 608.0, + 371.0, + 608.0, + 371.0, + 652.0, + 294.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 608.0, + 1055.0, + 608.0, + 1055.0, + 652.0, + 423.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 608.0, + 1375.0, + 608.0, + 1375.0, + 652.0, + 1165.0, + 652.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1445, + 1404, + 1445, + 1404, + 1811, + 297, + 1811 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 707, + 1405, + 707, + 1405, + 1013, + 297, + 1013 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1106, + 1404, + 1106, + 1404, + 1352, + 298, + 1352 + ], + "score": 0.982 + }, + { + "category_id": 8, + "poly": [ + 731, + 303, + 970, + 303, + 970, + 381, + 731, + 381 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 549, + 439, + 1154, + 439, + 1154, + 514, + 549, + 514 + ], + "score": 0.942 + }, + { + "category_id": 1, + "poly": [ + 303, + 1923, + 1402, + 1923, + 1402, + 1985, + 303, + 1985 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 298, + 394, + 980, + 394, + 980, + 438, + 298, + 438 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 834, + 229, + 834, + 263, + 299, + 263 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 300, + 582, + 748, + 582, + 748, + 618, + 300, + 618 + ], + "score": 0.908 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 817, + 74, + 817, + 106, + 299, + 106 + ], + "score": 0.907 + }, + { + "category_id": 1, + "poly": [ + 298, + 277, + 579, + 277, + 579, + 309, + 298, + 309 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 327, + 2006, + 1082, + 2006, + 1082, + 2034, + 327, + 2034 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 301, + 1854, + 741, + 1854, + 741, + 1890, + 301, + 1890 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 299, + 650, + 648, + 650, + 648, + 682, + 299, + 682 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 302, + 1388, + 774, + 1388, + 774, + 1419, + 302, + 1419 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 301, + 1050, + 699, + 1050, + 699, + 1081, + 301, + 1081 + ], + "score": 0.889 + }, + { + "category_id": 9, + "poly": [ + 1365, + 329, + 1401, + 329, + 1401, + 359, + 1365, + 359 + ], + "score": 0.889 + }, + { + "category_id": 9, + "poly": [ + 1352, + 461, + 1400, + 461, + 1400, + 493, + 1352, + 493 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.854 + }, + { + "category_id": 2, + "poly": [ + 1374, + 509, + 1402, + 509, + 1402, + 537, + 1374, + 537 + ], + "score": 0.748 + }, + { + "category_id": 14, + "poly": [ + 729, + 301, + 972, + 301, + 972, + 382, + 729, + 382 + ], + "score": 0.94, + "latex": "( t - 2 ) \\frac { \\Gamma ( \\frac { t - 1 } { 2 } ) ^ { 2 } } { \\Gamma ( t / 2 ) ^ { 2 } } \\leq 2" + }, + { + "category_id": 13, + "poly": [ + 485, + 922, + 616, + 922, + 616, + 953, + 485, + 953 + ], + "score": 0.91, + "latex": "\\beta _ { 2 } = 0 . 9 9 9" + }, + { + "category_id": 13, + "poly": [ + 614, + 394, + 872, + 394, + 872, + 440, + 614, + 440 + ], + "score": 0.91, + "latex": "( \\frac { \\Gamma ( x + 1 ) } { \\Gamma ( x + s ) } < ( x + 1 ) ^ { 1 - s } )" + }, + { + "category_id": 13, + "poly": [ + 972, + 1260, + 1230, + 1260, + 1230, + 1291, + 972, + 1291 + ], + "score": 0.9, + "latex": "\\beta _ { 1 } = 0 . 9 , \\beta _ { 2 } = 0 . 9 9 9" + }, + { + "category_id": 13, + "poly": [ + 1110, + 1596, + 1169, + 1596, + 1169, + 1627, + 1110, + 1627 + ], + "score": 0.9, + "latex": "3 e ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 972, + 1290, + 1031, + 1290, + 1031, + 1319, + 972, + 1319 + ], + "score": 0.89, + "latex": "1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 922, + 435, + 922, + 435, + 953, + 298, + 953 + ], + "score": 0.89, + "latex": "i . e . \\beta _ { 1 } = 0 . 9" + }, + { + "category_id": 14, + "poly": [ + 548, + 438, + 1154, + 438, + 1154, + 516, + 548, + 516 + ], + "score": 0.89, + "latex": "( t - 2 ) \\frac { \\dot { \\Gamma } ( \\frac { t - 1 } { 2 } ) ^ { 2 } } { \\Gamma ( t / 2 ) ^ { 2 } } \\leq ( t - 2 ) ( \\frac { t - 1 } { 2 } ) ^ { - 1 } = \\frac { 2 ( t - 2 ) } { t - 1 } < 2" + }, + { + "category_id": 13, + "poly": [ + 884, + 1660, + 915, + 1660, + 915, + 1690, + 884, + 1690 + ], + "score": 0.88, + "latex": "\\beta _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1030, + 1660, + 1146, + 1660, + 1146, + 1688, + 1030, + 1688 + ], + "score": 0.87, + "latex": "\\nu = 4 0 0 0" + }, + { + "category_id": 13, + "poly": [ + 1157, + 1659, + 1313, + 1659, + 1313, + 1690, + 1157, + 1690 + ], + "score": 0.86, + "latex": "\\beta _ { 2 } = 0 . 9 9 9 5 )" + }, + { + "category_id": 13, + "poly": [ + 1373, + 1571, + 1404, + 1571, + 1404, + 1597, + 1373, + 1597 + ], + "score": 0.77, + "latex": "=" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 581.0, + 751.0, + 581.0, + 751.0, + 625.0, + 292.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 109.0, + 295.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 2000.0, + 1084.0, + 2000.0, + 1084.0, + 2040.0, + 330.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1850.0, + 746.0, + 1850.0, + 746.0, + 1896.0, + 295.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 647.0, + 651.0, + 647.0, + 651.0, + 686.0, + 294.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1386.0, + 778.0, + 1386.0, + 778.0, + 1422.0, + 296.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1049.0, + 702.0, + 1049.0, + 702.0, + 1085.0, + 295.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1379.0, + 514.0, + 1401.0, + 514.0, + 1401.0, + 536.0, + 1379.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1447.0, + 1401.0, + 1447.0, + 1401.0, + 1478.0, + 297.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1475.0, + 1405.0, + 1475.0, + 1405.0, + 1509.0, + 294.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1505.0, + 1406.0, + 1505.0, + 1406.0, + 1541.0, + 294.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1538.0, + 1402.0, + 1538.0, + 1402.0, + 1569.0, + 296.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1565.0, + 1372.0, + 1565.0, + 1372.0, + 1605.0, + 294.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1596.0, + 1109.0, + 1596.0, + 1109.0, + 1630.0, + 294.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 1596.0, + 1404.0, + 1596.0, + 1404.0, + 1630.0, + 1170.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1628.0, + 1405.0, + 1628.0, + 1405.0, + 1663.0, + 294.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1658.0, + 883.0, + 1658.0, + 883.0, + 1692.0, + 295.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1658.0, + 1029.0, + 1658.0, + 1029.0, + 1692.0, + 916.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1658.0, + 1156.0, + 1658.0, + 1156.0, + 1692.0, + 1147.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 1658.0, + 1405.0, + 1658.0, + 1405.0, + 1692.0, + 1314.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1689.0, + 1405.0, + 1689.0, + 1405.0, + 1723.0, + 295.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1720.0, + 1404.0, + 1720.0, + 1404.0, + 1754.0, + 294.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1749.0, + 1406.0, + 1749.0, + 1406.0, + 1783.0, + 294.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1782.0, + 632.0, + 1782.0, + 632.0, + 1812.0, + 295.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 708.0, + 1405.0, + 708.0, + 1405.0, + 742.0, + 295.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 738.0, + 1405.0, + 738.0, + 1405.0, + 774.0, + 294.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 769.0, + 1405.0, + 769.0, + 1405.0, + 804.0, + 295.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 799.0, + 1403.0, + 799.0, + 1403.0, + 835.0, + 295.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 828.0, + 1406.0, + 828.0, + 1406.0, + 868.0, + 294.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 859.0, + 1406.0, + 859.0, + 1406.0, + 896.0, + 294.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 888.0, + 1406.0, + 888.0, + 1406.0, + 930.0, + 294.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 921.0, + 297.0, + 921.0, + 297.0, + 956.0, + 294.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 921.0, + 484.0, + 921.0, + 484.0, + 956.0, + 436.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 921.0, + 1406.0, + 921.0, + 1406.0, + 956.0, + 617.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 950.0, + 1405.0, + 950.0, + 1405.0, + 987.0, + 295.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 983.0, + 1089.0, + 983.0, + 1089.0, + 1015.0, + 294.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1104.0, + 1404.0, + 1104.0, + 1404.0, + 1141.0, + 293.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1138.0, + 1406.0, + 1138.0, + 1406.0, + 1170.0, + 293.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1168.0, + 1405.0, + 1168.0, + 1405.0, + 1201.0, + 294.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1199.0, + 1404.0, + 1199.0, + 1404.0, + 1233.0, + 294.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1228.0, + 1404.0, + 1228.0, + 1404.0, + 1264.0, + 295.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1258.0, + 971.0, + 1258.0, + 971.0, + 1295.0, + 292.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 1258.0, + 1405.0, + 1258.0, + 1405.0, + 1295.0, + 1231.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1290.0, + 971.0, + 1290.0, + 971.0, + 1324.0, + 294.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1290.0, + 1402.0, + 1290.0, + 1402.0, + 1324.0, + 1032.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1322.0, + 820.0, + 1322.0, + 820.0, + 1355.0, + 294.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1921.0, + 1403.0, + 1921.0, + 1403.0, + 1957.0, + 298.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1949.0, + 1402.0, + 1949.0, + 1402.0, + 1993.0, + 293.0, + 1993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 379.0, + 613.0, + 379.0, + 613.0, + 458.0, + 286.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 379.0, + 983.0, + 379.0, + 983.0, + 458.0, + 873.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 227.0, + 838.0, + 227.0, + 838.0, + 267.0, + 296.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 274.0, + 580.0, + 274.0, + 580.0, + 314.0, + 296.0, + 314.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 229, + 1404, + 229, + 1404, + 536, + 297, + 536 + ], + "score": 0.983 + }, + { + "category_id": 5, + "poly": [ + 295, + 635, + 1405, + 635, + 1405, + 854, + 295, + 854 + ], + "score": 0.982, + "html": "
1-4 steps5-8 steps8+ stepstest acctrain losstrain error
RAdamRAdamRAdam91.080.0210.74
Adam (w. divergent var.)RAdamRAdam89.980.0602.12
SGDAdam (w. convergent var.)RAdam90.290.0381.23
" + }, + { + "category_id": 6, + "poly": [ + 597, + 564, + 1100, + 564, + 1100, + 598, + 597, + 598 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.844 + }, + { + "category_id": 13, + "poly": [ + 997, + 566, + 1087, + 566, + 1087, + 596, + 997, + 596 + ], + "score": 0.75, + "latex": "( \\mathrm { l r } = 0 . 1 " + }, + { + "category_id": 13, + "poly": [ + 951, + 647, + 986, + 647, + 986, + 676, + 951, + 676 + ], + "score": 0.28, + "latex": "^ { 8 + }" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 559.0, + 996.0, + 559.0, + 996.0, + 603.0, + 593.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 559.0, + 1103.0, + 559.0, + 1103.0, + 603.0, + 1088.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 1406.0, + 230.0, + 1406.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 260.0, + 1407.0, + 260.0, + 1407.0, + 297.0, + 292.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 293.0, + 1405.0, + 293.0, + 1405.0, + 325.0, + 296.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 321.0, + 1405.0, + 321.0, + 1405.0, + 356.0, + 292.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 352.0, + 1406.0, + 352.0, + 1406.0, + 388.0, + 291.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 383.0, + 1405.0, + 383.0, + 1405.0, + 418.0, + 294.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 413.0, + 1405.0, + 413.0, + 1405.0, + 449.0, + 294.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 444.0, + 1406.0, + 444.0, + 1406.0, + 479.0, + 295.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 474.0, + 1404.0, + 474.0, + 1404.0, + 508.0, + 295.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 502.0, + 1078.0, + 502.0, + 1078.0, + 539.0, + 294.0, + 539.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/ryxdEkHtPS/ryxdEkHtPS.md b/parse/train/ryxdEkHtPS/ryxdEkHtPS.md new file mode 100644 index 0000000000000000000000000000000000000000..aabcd462d1a9d20ed0995776800d050e5814654f --- /dev/null +++ b/parse/train/ryxdEkHtPS/ryxdEkHtPS.md @@ -0,0 +1,382 @@ +# A CLOSER LOOK AT DEEP POLICY GRADIENTS + +Andrew Ilyas∗1, Logan Engstrom∗1, Shibani Santurkar1, Dimitris Tsipras1, Firdaus Janoos2, Larry Rudolph1,2, and Aleksander M ˛adry1 + +1MIT 2Two Sigma {ailyas,engstrom,shibani,tsipras,madry}@mit.edu rudolph@csail.mit.edu, firdaus.janoos@twosigma.com + +# ABSTRACT + +We study how the behavior of deep policy gradient algorithms reflects the conceptual framework motivating their development. To this end, we propose a finegrained analysis of state-of-the-art methods based on key elements of this framework: gradient estimation, value prediction, and optimization landscapes. Our results show that the behavior of deep policy gradient algorithms often deviates from what their motivating framework would predict: the surrogate objective does not match the true reward landscape, learned value estimators fail to fit the true value function, and gradient estimates poorly correlate with the “true” gradient. The mismatch between predicted and empirical behavior we uncover highlights our poor understanding of current methods, and indicates the need to move beyond current benchmark-centric evaluation methods. + +# 1 INTRODUCTION + +Deep reinforcement learning (RL) is behind some of the most publicized achievements of modern machine learning (Silver et al., 2017; OpenAI, 2018; Dayarathna et al., 2016; OpenAI et al., 2018). In fact, to many, this framework embodies the promise of the real-world impact of machine learning. However, the deep RL toolkit has not yet attained the same level of engineering stability as, for example, the current deep (supervised) learning framework. Indeed, recent studies demonstrate that state-of-the-art deep RL algorithms suffer from oversensitivity to hyperparameter choices, lack of consistency, and poor reproducibility (Henderson et al., 2017). + +This state of affairs suggests that it might be necessary to re-examine the conceptual underpinnings of deep RL methodology. More precisely, the overarching question that motivates this work is: + +To what degree does current practice in deep RL reflect the principles informing its development? + +Our specific focus is on deep policy gradient methods, a widely used class of deep RL algorithms. Our goal is to explore the extent to which state-of-the-art implementations of these methods succeed at realizing the key primitives of the general policy gradient framework. + +Our contributions. We take a broader look at policy gradient algorithms and their relation to their underlying framework. With this perspective in mind, we perform a fine-grained examination of key RL primitives as they manifest in practice. Concretely, we study: + +• Gradient Estimation: we find that even when agents improve in reward, their gradient estimates used in parameter updates poorly correlate with the “true” gradient. We additionally show that gradient estimate quality decays with training progress and task complexity. Finally, we demonstrate that varying the sample regime yields training dynamics that are unexplained by the motivating framework and run contrary to supervised learning intuition. + +• Value Prediction: our experiments indicate that value networks successfully solve the supervised learning task they are trained on, but do not fit the true value function. Additionally, employing a value network as a baseline function only marginally decreases the variance of gradient estimates compared to using true value as a baseline (but still dramatically increases agent’s performance compared to using no baseline at all). + +• Optimization Landscapes: we show that the optimization landscape induced by modern policy gradient algorithms is often not reflective of the underlying true reward landscape, and that the latter is frequently poorly behaved in the relevant sample regime. + +Overall, our results demonstrate that the motivating theoretical framework for deep RL algorithms is often unpredictive of phenomena arising in practice. This suggests that building reliable deep RL algorithms requires moving past benchmark-centric evaluations to a multi-faceted understanding of their often unintuitive behavior. We conclude (in Section 3) by discussing several areas where such understanding is most critically needed. + +# 2 EXAMINING THE PRIMITIVES OF DEEP POLICY GRADIENT ALGORITHMS + +In this section, we investigate the degree to which our theoretical understanding of RL applies to modern methods. We consider key primitives of policy gradient algorithms: gradient estimation, value prediction and reward fitting. In what follows, we perform a fine-grained analysis of state-ofthe-art policy gradient algorithms (PPO and TRPO) through the lens of these primitives—detailed preliminaries, background, and notation can be found in Appendix A.1. + +# 2.1 GRADIENT ESTIMATE QUALITY + +A central premise of policy gradient methods is that stochastic gradient ascent on a suitable objective function yields a good policy. These algorithms use as a primitive the gradient of that objective function: + +$$ +\hat { g } = \nabla _ { \theta } \mathbb { E } _ { ( s _ { t } , a _ { t } ) \sim \pi _ { 0 } } \left[ \frac { \pi _ { \theta } ( a _ { t } | s _ { t } ) } { \pi _ { 0 } ( a _ { t } | s _ { t } ) } \widehat { A } _ { \pi _ { 0 } } ( s _ { t } , a _ { t } ) \right] = \mathbb { E } _ { ( s _ { t } , a _ { t } ) \sim \pi _ { 0 } } \left[ \frac { \nabla _ { \theta } \pi _ { \theta } ( a _ { t } | s _ { t } ) } { \pi _ { 0 } ( a _ { t } | s _ { t } ) } \widehat { A } _ { \pi _ { 0 } } ( s _ { t } , a _ { t } ) \right] , +$$ + +where in the above we use standard RL notation (see Appendix A.1 for more details). An underlying assumption behind these methods is that we have access to a reasonable estimate of this quantity. This assumption effectively translates into an assumption that we can accurately estimate the expectation above using an empirical mean of finite (typically $\sim 1 0 ^ { 3 }$ ) samples. Evidently (since the agent attains a high reward) these estimates are sufficient to consistently improve reward—we are thus interested in the relative quality of these gradient estimates in practice, and the effect of gradient quality on optimization. + +![](images/564328a99ce532599345890f1e098eb6f8665f3c9643714f6b4dbc1bb4fb90ac.jpg) +Figure 1: Empirical variance of the estimated gradient (c.f. (1)) as a function of the number of stateaction pairs used in estimation in the MuJoCo Humanoid task. We measure the average pairwise cosine similarity between ten repeated gradient measurements taken from the same policy, with the $9 5 \%$ confidence intervals (shaded). For each algorithm, we perform multiple trials with the same hyperparameter configurations but different random seeds, shown as repeated lines in the figure. The vertical line (at $x = 2 \mathsf { K }$ ) indicates the sample regime used for gradient estimation in standard implementations of policy gradient methods. In general, it seems that obtaining tightly concentrated gradient estimates would require significantly more samples than are used in practice, particularly after the first few timesteps. For other tasks – such as Walker2d-v2 and Hopper-v2 – the plots (seen in Appendix Figure 9) have similar trends, except that gradient variance is slightly lower. Confidence intervals calculated with 500 sample bootstrapping. + +![](images/0a11f724707726202a961f210b6493d422e426616dc7c49610d275b33adf05b7.jpg) +Figure 2: Convergence of gradient estimates (c.f. (1)) to the “true” expected gradient in the MuJoCo Humanoid task. We measure the mean cosine similarity between the “true” gradient approximated using ten million state-action pairs, and ten gradient estimates which use increasing numbers of state-action pairs (with $9 5 \%$ confidence intervals). For each algorithm, we perform multiple trials with the same hyperparameter configurations but different random seeds. The vertical line (at $x =$ 2K) indicates the sample regime used for gradient estimation in standard implementations of policy gradient methods. Observe that although it is possible to empirically estimate the true gradient, this requires several-fold more samples than are used commonly in practical applications of these algorithms. See additionally that the estimation task becomes more difficult further into training. For other tasks – such as Walker2d-v2 and Hopper-v2 – the plots (seen in Appendix Figure 10) have similar trends, except that gradient estimation is slightly better. Confidence intervals calculated with 500 sample bootstrapping. + +How accurate are the gradient estimates we compute? To answer this question, we examine two of the most natural measures of estimate quality: the empirical variance and the convergence to the “true” gradient. To evaluate the former, we measure the average pairwise cosine similarity between estimates of the gradient computed from the same policy with independent rollouts (Figure 1). We evaluate the latter by first forming an estimate of the true gradient with a large number of state-action pairs. We then examine the convergence of gradient estimates to this “true” gradient (which we once again measure using cosine similarity) as we increase the number of samples (Figure 2). + +We observe that deep policy gradient methods operate with relatively poor estimates of the gradient, especially as task complexity increases and as training progresses (contrast Humanoid-v2, a “hard” task, to other tasks and contrast successive checkpoints in Figures 1 and 2). This is in spite of the fact that our agents continually improve throughout training, and attain nowhere near the maximum reward possible on each task. In fact, we sometimes observe a zero or even negative correlation in the relevant sample regime1. + +While these results might be reminiscent of the well-studied “noisy gradients” problem in supervised learning (Robbins & Monro, 1951; d’Aspremont, 2008; Kawaguchi, 2016; Safran & Shamir, 2018; Livni et al., 2014; Keskar et al., 2016; Hochreiter & Schmidhuber, 1997), we have very little understanding of how gradient quality affects optimization in the substantially different reinforcement learning setting. For example: + +• The sample regime in which RL algorithms operate seems to have a profound impact on the robustness and stability of agent training—in particular, many of the sensitivity issues reported by Henderson et al. (2017) are claimed to disappear (Sutskever, 2018) in highersample regimes. Understanding the implications of working in this sample regime, and more generally the impact of sample complexity on training stability remains to be precisely understood. +• Agent policy networks are trained concurrently with value networks (discussed more in the following section) meant to reduce the variance of gradient estimates. Under our conceptual framework, we might expect these networks to help gradient estimates more as training progresses, contrary to what we observe in Figure 1. The value network also makes the now two-player optimization landscape and training dynamics even more difficult to grasp, as such interactions are poorly understood. + +• The relevant measure of sample complexity for many settings (number of state-action pairs) can differ drastically from the number of independent samples used at each training iteration (the number of complete trajectories). The latter quantity (a) tends to be much lower than the number of state-action pairs, and (b) decreases across iterations during training. + +All the above factors make it unclear to what degree our intuition from classical settings transfer to the deep RL regime. And the policy gradient framework, as of now, provides little predictive power regarding the variance of gradient estimates and its impact on reward optimization. + +Our results indicate that despite having a rigorous theoretical framework for RL, we lack a precise understanding of the structure of the reward landscape and optimization process. + +# 2.2 VALUE PREDICTION + +Our findings from the previous section motivate a deeper look into gradient estimation. After all, the policy gradient in its original formulation (Sutton et al., 1999) is known to be hard to estimate, and thus algorithms employ a variety of variance reduction methods. The most popular of these techniques is a baseline function. Concretely, an equivalent form of the policy gradient is given by: + +$$ +\widehat { g } _ { \theta } = \mathbb { E } _ { \tau \sim \pi _ { \theta } } \left[ \sum _ { ( s _ { t } , a _ { t } ) \in \tau } \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } | s _ { t } ) \cdot ( Q _ { \pi _ { \theta } } ( s _ { t } , a _ { t } ) - b ( s _ { t } ) ) \right] +$$ + +where $b ( s _ { t } )$ is some fixed function of the state $s _ { t }$ . A canonical choice of baseline function is the value function $V _ { \pi } ( s )$ , the expected return from a given state (more details and motivation in A.1): + +$$ +V _ { \pi _ { \theta } } ( s _ { t } ) = \mathbb { E } _ { \pi _ { \theta } } [ R _ { t } | s _ { t } ] \ . +$$ + +Indeed, fitting a value-estimating function (Schulman et al., 2015c; Sutton & Barto, 2018) (a neural network, in the deep RL setting) and using it as a baseline function is precisely the approach taken by most deep policy gradient methods. Concretely, one trains a value network $V _ { \theta _ { t } } ^ { \pi }$ such that: + +$$ +\theta _ { t } = \operatorname* { m i n } _ { \theta } \mathbb { E } \left[ \left( V _ { \theta } ^ { \pi } ( s _ { t } ) - ( V _ { \theta _ { t - 1 } } ^ { \pi } ( s _ { t } ) + A _ { t } ) \right) ^ { 2 } \right] +$$ + +where $V _ { \theta _ { t - 1 } } ^ { \pi } ( s _ { t } )$ are estimates given by the last value function, and $A _ { t }$ is the advantage of the policy, i.e. the returns minus the estimated values. (Typically, $A _ { t }$ is estimated using generalized advantage estimation, as described in (Schulman et al., 2015c).) Our findings in the previous section prompt us to take a closer look at the value network and its impact on the variance of gradient estimates. + +![](images/3400f4111aa816678fb290803993e451edc4053c39f8141eab8c837fd6cfb8ce.jpg) +Figure 3: Quality of value prediction in terms of mean relative error (MRE) on heldout state-action pairs for agents trained to solve the MuJoCo Walker2d-v2 task. We observe in (left) that the agents do indeed succeed at solving the supervised learning task they are trained for—the MRE on the GAE-based value loss $( V _ { o l d } \bar { + } A _ { G A E } ) ^ { 2 }$ (c.f. (4)) is small. On the other hand, in (right) we see that the returns MRE is still quite high—the learned value function is off by about $5 0 \%$ with respect to the underlying true value function. Similar plots for other MuJoCo tasks are in Appendix A.5. + +![](images/3623c047652f57efd9b89a08dd419cc64f3d9383a3fca27b9f80f90d631a8c19.jpg) +Figure 4: Efficacy of the value network as a variance reducing baseline for Walker2d-v2 (top) and Hopper-v2 (bottom) agents. We measure the empirical variance of the gradient (c.f. (1)) as a function of the number of state-action pairs used in estimation, for different choices of baseline functions: the value network (used by the agent in training), the “true” value function (fit to the returns using $5 \cdot 1 0 ^ { 6 }$ state-action pairs sampled from the current policy) and the “zero” value function (i.e. replacing advantages with returns). We observe that using the true value function leads to a significantly lower-variance estimate of the gradient compared to the value network. In turn, employing the value network yields a noticeable variance reduction compared to the zero baseline function, even though this difference may appear rather small in the small-sample regime (2K). Confidence intervals calculated with 10 sample bootstrapping. + +Value prediction as a supervised learning problem. We first analyze the value network through the lens of the supervised learning problem it solves. After all, (4) describes an empirical risk minimization, where a loss is minimized over a set of sampled $\left( { { s _ { t } } , { a _ { t } } } \right)$ . So, how does $V _ { \theta } ^ { \pi }$ perform as a solution to (4)? And in turn, how does (4) perform as a proxy for learning the true value function? + +Our results (Figure 3a) show that the value network does succeed at both fitting the given loss function and generalizing to unseen data, showing low and stable mean relative error (MRE). However, the significant drop in performance as shown in Figure 3 indicates that the supervised learning problem induced by (4) does not lead to $V _ { \theta } ^ { \pi }$ learning the underlying true value function. + +Does the value network lead to a reduction in variance? Though evaluating the $V _ { \theta } ^ { \pi }$ baseline function as a value predictor as we did above is informative, in the end the sole purpose of the value function is to reduce variance. So: how does using our value function actually impact the variance of our gradient estimates? To answer this question, we compare the variance reduction that results from employing our value network against both a “true” value function and a trivial “zero” baseline function (i.e. simply replacing advantages with returns). Our results, captured in Figure 4, show that the “true” value function yields a much lower-variance estimate of the gradient. This is especially true in the sample regime in which we operate. We note, however, that despite not effectively predicting the true value function or inducing the same degree of variance reduction, the value network does help to some degree (compared to the “zero” baseline). Additionally, the seemingly marginal increase in gradient correlation provided by the value network (compared to the “true” baseline function) turns out to result in a significant improvement in agent performance. (Indeed, agents trained without a baseline reach almost an order of magnitude worse reward.) + +Our findings suggest that we still need a better understanding of the role of the value network in agent training, and raise several questions that we discuss in Section 3. + +# 2.3 EXPLORING THE OPTIMIZATION LANDSCAPE + +Another key assumption of policy gradient algorithms is that first-order updates (w.r.t. policy parameters) actually yield better policies. It is thus natural to examine how valid this assumption is. + +The true rewards landscape. We begin by examining the landscape of agent reward with respect to the policy parameters. Indeed, even if deep policy gradient methods do not optimize for the true reward directly (e.g. if they use a surrogate objective), the ultimate goal of any policy gradient algorithm is to navigate this landscape. First, Figure 5 shows that while estimating the true reward landscape with a high number of samples yields a relatively smooth reward landscape (perhaps suggesting viability of direct reward optimization), estimating the true reward landscape in the typical, low sample regime results in a landscape that appears jagged and poorly-behaved. The low-sample regime thus gives rise to a certain kind of barrier to direct reward optimization. Indeed, applying our algorithms in this regime makes it impossible to distinguish between good and bad points in the landscape, even though the true underlying landscape is fairly well-behaved. + +The surrogate objective landscape. The untamed nature of the rewards landscape has led to the development of alternate approaches to reward maximization. Recall that an important element of many modern policy gradient methods is the maximization of a surrogate objective function in place of the true rewards (the exact mechanism behind the surrogate objective is detailed in Appendix A.1, and particularly in (14)). The surrogate objective, based on relaxing the policy improvement theorem of Kakade and Langford (Kakade & Langford, 2002), can be viewed as a simplification of the reward maximization objective. + +As a purported approximation of the true returns, one would expect that the surrogate objective landscape approximates the true reward landscape fairly well. That is, parameters corresponding to good surrogate objective will also correspond to good true reward. + +Figure 6 shows that in the early stages of training, the optimization landscapes of the true reward and surrogate objective are indeed approximately aligned. However, as training progresses, the surrogate objective becomes much less predictive of the true reward in the relevant sample regime. In particular, we often observe that directions that increase the surrogate objective lead to a decrease of the true reward (see Figures 6, 7). In a higher-sample regime (using several orders of magnitude more samples), we find that PPO and TRPO turn out to behave rather differently. In the case of TRPO, the update direction following the surrogate objective matches the true reward much more closely. However, for PPO we consistently observe landscapes where the step direction leads to lower true reward, even in the high-sample regime. This suggests that even when estimated accurately enough, the surrogate objective might not be an accurate proxy for the true reward. (Recall from Section 2.1 that this is a sample regime where we are able to estimate the true gradient of the reward fairly well.) + +![](images/988f3e20351aeff96237e9e3741cc998c67968391db52269442dd22142615d7c.jpg) +2,000 state-action pairs + +![](images/01017d8520cd26b620c161d401d7fa8673dacd4c65563407b6eb21bc583fe7d4.jpg) +20,000 state-action pairs + +![](images/9133c34da248b4810278f657205cb43af0e0b9d2b1c5c14e17b97c87979a954b.jpg) +100,000 state-action pairs +Figure 5: True reward landscape concentration for TRPO on Humanoid-v2. We visualize the landscape at a training iteration 150 while varying the number of trajectories used in reward estimation (each subplot), both in the direction of the step taken and a random direction. Moving one unit along the “step direction” axis corresponds to moving one full step in parameter space. In the random direction one unit corresponds to moving along a random norm 2 Gaussian vector in the parameter space. In practice, the norm of the step is typically an order of magnitude lower than the random direction. While the landscape is very noisy in the low-sample regime, large numbers of samples reveal a well-behaved underlying landscape. See Figures 20, 19 of the Appendix for additional plots. + +![](images/1cb608251899f8943400cff475cebaf32b427d895045caac4cdea6cd45719592.jpg) +Figure 6: True reward and surrogate objective landscapes for TRPO on the Humanoid-v2 MuJoCo task. We visualize the landscapes in the direction of the update step and a random direction (as in Figure 5). The surrogate objective corresponds to the actual function optimized by the algorithm at each step. We estimate true reward with $1 { \dot { 0 } } ^ { 6 }$ state-action pairs per point. We compare the landscapes at different points in training and with varying numbers of state-action pairs used in the update step. Early in training the true and surrogate landscapes align fairly well in both sample regimes, but later become misaligned in the low-sample regime. More landscapes in Appendix Figures 13-18. + +# 3 TOWARDS STRONGER FOUNDATIONS FOR DEEP RL + +Deep reinforcement learning (RL) algorithms have shown great practical promise, and are rooted in a well-grounded theoretical framework. However, our results indicate that this framework often fails to provide insight into the practical performance of these algorithms. This disconnect impedes our understanding of why these algorithms succeed (or fail), and is a major barrier to addressing key challenges facing deep RL such as brittleness and poor reproducibility. + +To close this gap, we need to either develop methods that adhere more closely to theory, or build theory that can capture what makes existing policy gradient methods successful. In both cases, the first step is to precisely pinpoint where theory and practice diverge. To this end, we analyze and consolidate our findings from the previous section. + +Gradient estimation. Our analysis in Section 2.1 shows that the quality of gradient estimates that deep policy gradient algorithms use is rather poor. Indeed, even when agents improve, such gradient estimates often poorly correlate with the true gradient (c.f. Figure 2). We also note that gradient correlation decreases as training progresses and task complexity increases. While this certainly does not preclude the estimates from conveying useful signal, the exact underpinnings of this phenomenon in deep RL still elude us. In particular, in Section 2.1 we outline a few keys ways in which the deep RL setting is quite unique and difficult to understand from an optimization perspective, both theoretically and in practice Overall, understanding the impact of gradient estimate quality on deep RL algorithms is challenging and largely unexplored. + +Value prediction. The findings presented in Section 2.2 identify two key issues. First, while the value network successfully solves the supervised learning task it is trained on, it does not accurately model the “true” value function. Second, employing the value network as a baseline does decrease the gradient variance (compared to the trivial (“zero”) baseline). However, this decrease is rather marginal compared to the variance reduction offered by the “true” value function. + +It is natural to wonder whether this failure in modeling the value function is inevitable. For example, how does the loss function used to train the value network impact value prediction and variance reduction? More broadly, we lack an understanding of the precise role of the value network in training. Can we empirically quantify the relationship between variance reduction and performance? And does the value network play a broader role than just variance reduction? + +![](images/0f929acd579bd6910d6fbd92bfdf56ec0bf9d21185294ff124330ed9cb583b77.jpg) +Figure 7: True reward and surrogate objective landscapes for PPO on the Humanoid-v2 MuJoCo task. See Figure 6 for a description. We observe that early in training the true and surrogate landscapes align well. However, later increasing the surrogate objective leads to lower true reward. + +Optimization landscape. We have also seen, in Section 2.3, that the optimization landscape induced by modern policy gradient algorithms, the surrogate objective, is often not reflective of the underlying true reward landscape. We thus need a deeper understanding of why current methods succeed despite these issues, and, more broadly, how to better navigate the true reward landscape. + +# 4 RELATED WORK + +The idea of using gradient estimates to update neural network-based RL agents dates back at least to the REINFORCE (Williams, 1992) algorithm. Later, Sutton (Sutton et al., 1999) established a unifying framework casting these algorithms as instances of the policy gradient class of algorithms. + +Our work focuses on proximal policy optimization (PPO) (Schulman et al., 2017) and trust region policy optimization (TRPO) (Schulman et al., 2015a), which are two of the most prominent policy gradient algorithms used in deep RL, drawing inspiration from works on related algorithms, such as (Peters et al., 2010) and Kakade (2001). + +Many recent works document the brittleness of deep RL algorithms (Henderson et al., 2018; 2017; Islam et al., 2017). (Rajeswaran et al., 2017) and (Mania et al., 2018) demonstrate that on many benchmark tasks, state-of-the-art performance can be attained by augmented randomized search approaches. McCandlish et al. (2018) investigates gradient noise in large-batch settings, and Ahmed et al. (2018) investigates the role of entropy regularization (which we do not study) on optimization. + +# 5 CONCLUSION + +In this work, we analyze the degree to which key primitives of deep policy gradient algorithms follow their conceptual underpinnings. Our experiments show that these primitives often do not conform to the expected behavior: gradient estimates poorly correlate with the true gradient, better gradient estimates can require lower learning rates and can induce degenerate agent behavior, value networks reduce gradient estimation variance to a significantly smaller extent than the true value, and the underlying optimization landscape can be misleading. + +This demonstrates that there is a significant gap between the theory inspiring current algorithms and the actual mechanisms driving their performance. Overall, our findings suggest that developing a deep RL toolkit that is truly robust and reliable will require moving beyond the current benchmarkdriven evaluation model to a more fine-grained understanding of deep RL algorithms. + +# REFERENCES + +Zafarali Ahmed, Nicolas Le Roux, Mohammad Norouzi, and Dale Schuurmans. Understanding the impact of entropy on policy optimization, 2018. + +Alexandre d’Aspremont. Smooth optimization with approximate gradient. SIAM Journal on Optimization, 19:1171–1183, 2008. + +Miyuru Dayarathna, Yonggang Wen, and Rui Fan. Data center energy consumption modeling: A survey. IEEE Communications Surveys & Tutorials, 18(1):732–794, 2016. + +Peter Henderson, Riashat Islam, Philip Bachman, Joelle Pineau, Doina Precup, and David Meger. Deep reinforcement learning that matters. arXiv preprint arXiv:1709.06560, 2017. + +Peter Henderson, Joshua Romoff, and Joelle Pineau. Where did my optimum go?: An empirical analysis of gradient descent optimization in policy gradient methods, 2018. + +Sepp Hochreiter and Jürgen Schmidhuber. Flat minima. Neural Computation, 9:1–42, 1997. + +Riashat Islam, Peter Henderson, Maziar Gomrokchi, and Doina Precup. Reproducibility of benchmarked deep reinforcement learning tasks for continuous control. In ICML Reproducibility in Machine Learning Workshop, 2017. + +Sham M. Kakade. A natural policy gradient. In NIPS, 2001. + +Sham M. Kakade and John Langford. Approximately optimal approximate reinforcement learning. In ICML, 2002. + +Kenji Kawaguchi. Deep learning without poor local minima. In NIPS, 2016. + +Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. CoRR, abs/1609.04836, 2016. + +Roi Livni, Shai Shalev-Shwartz, and Ohad Shamir. On the computational efficiency of training neural networks. In NIPS, 2014. + +Horia Mania, Aurelia Guy, and Benjamin Recht. Simple random search provides a competitive approach to reinforcement learning. CoRR, abs/1803.07055, 2018. + +Sam McCandlish, Jared Kaplan, Dario Amodei, and OpenAI Dota Team. An empirical model of large-batch training, 2018. + +OpenAI. Openai five. https://blog.openai.com/openai-five/, 2018. + +OpenAI, :, Marcin Andrychowicz, Bowen Baker, Maciek Chociej, Rafal Jozefowicz, Bob McGrew, Jakub Pachocki, Arthur Petron, Matthias Plappert, Glenn Powell, Alex Ray, Jonas Schneider, Szymon Sidor, Josh Tobin, Peter Welinder, Lilian Weng, and Wojciech Zaremba. Learning dexterous in-hand manipulation, 2018. + +Jan Peters, Katharina Mülling, and Yasemin Altun. Relative entropy policy search. In AAAI, 2010. + +Aravind Rajeswaran, Kendall Lowrey, Emanuel Todorov, and Sham M. Kakade. Towards generalization and simplicity in continuous control. In NIPS, 2017. + +Herbert Robbins and Sutton Monro. A stochastic approximation method. Ann. Math. Statist., 22(3): 400–407, 09 1951. doi: 10.1214/aoms/1177729586. URL https://doi.org/10.1214/ aoms/1177729586. + +Itay Safran and Ohad Shamir. Spurious local minima are common in two-layer relu neural networks. In ICML, 2018. + +John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In International Conference on Machine Learning, pp. 1889–1897, 2015a. + +John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. Highdimensional continuous control using generalized advantage estimation. arXiv preprint arXiv:1506.02438, 2015b. + +John Schulman, Philipp Moritz, Sergey Levine, Michael I. Jordan, and Pieter Abbeel. Highdimensional continuous control using generalized advantage estimation. CoRR, abs/1506.02438, 2015c. + +John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. + +David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go without human knowledge. Nature, 550(7676):354, 2017. + +Ilya Sutskever. Keynote talk. NVIDIA NTECH, 2018. URL https://www.youtube.com/ watch?v $=$ w3ues-NayAs&t $= 4 6 7 s$ . + +Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction. MIT press, 2018. + +Richard S. Sutton, David A. McAllester, Satinder P. Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. In NIPS, 1999. + +Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8:229–256, 1992. + +# A APPENDIX + +# A.1 BACKGROUND + +In the reinforcement learning (RL) setting, an agent interacts with a stateful environment with the goal of maximizing cumulative reward. Formally, we model the environment as a (possibly randomized) function mapping its current state $s$ and an action $a$ supplied by the agent to a new state $s ^ { \prime }$ and a resulting reward $r$ . The choice of actions of the agent is governed by the its policy $\pi$ . This policy is a function mapping environment states to a distribution over the actions to take. The objective of an RL algorithm is to find a policy $\pi$ which maximizes the expected cumulative reward, where the expectation is taken over both environment randomness and the (randomized) action choices. + +Preliminaries and notation. For a given policy $\pi$ , we denote by $\pi ( a | s )$ the probability that this policy assigns to taking action $a$ when the environment is in the state $s$ . We use $r ( s , a )$ to denote the reward that the agent earns for playing action $a$ in response to the state $s$ . A trajectory $\tau =$ $\{ ( a _ { t } , s _ { t } ) : t \in \{ 1 \ldots \overline { { T } } \} \}$ is a sequence of state-action pairs that constitutes a valid transcript of interactions of the agent with the environment. (Here, $a _ { t }$ (resp. $s _ { t }$ ) corresponds to the action taken by the agent (resp. state of the environment) in the $t$ -th round of interaction.) We then define $\pi ( \tau )$ to be the probability that the trajectory $\tau$ is executed if the agent follows policy $\pi$ (provided the initial state of the environment is $s _ { 1 }$ ). Similarly, $\begin{array} { r } { r ( \tau ) = \sum _ { t } r ( \bar { s _ { t } } , a _ { t } ) } \end{array}$ denotes the cumulative reward earned by the agent when following this trajectory, where $s _ { t }$ (resp. $a _ { t }$ ) denote the $t { \cdot }$ -th state (resp. action) in the trajectory $\tau$ . In the RL setting, however, we often choose to maximize the discounted cumulative reward of a policy $R : = R _ { 1 }$ , where $R _ { t }$ is defined as + +$$ +R _ { t } ( \tau ) = \sum _ { t ^ { \prime } = t } ^ { \infty } \gamma ^ { ( t ^ { \prime } - t ) } r _ { t ^ { \prime } } \ . +$$ + +and $0 < \gamma < 1$ is a “discount factor”. The discount factor ensures that the cumulative reward of a policy is well-defined even for an infinite time horizon, and it also incentivizes achieving reward earlier. + +Policy gradient methods. A widely used class of RL algorithms that will be the focus of our analysis is the class of so-called policy gradient methods. The central idea behind these algorithms is to first parameterize the policy $\pi _ { \theta }$ using a parameter vector $\theta$ . (In the deep RL context, $\pi _ { \theta }$ is expressed by a neural network with weights $\theta$ .) Then, we perform stochastic gradient ascent on the cumulative reward with respect to $\theta$ . In other words, we want to apply the stochastic ascent approach to our problem: + +$$ +\operatorname* { m a x } _ { \theta } \operatorname { \mathbb { E } } _ { \tau \sim \pi _ { \theta } } [ r ( \tau ) ] ~ , +$$ + +where $\tau \sim \pi _ { \theta }$ represents trajectories (rollouts) sampled from the distribution induced by the policy $\pi _ { \theta }$ . This approach relies on the key observation (Sutton et al., 1999) that under mild conditions, the gradient of our objective can be written as: + +$$ +\begin{array} { r } { \nabla _ { \theta } \mathbb { E } _ { \tau \sim \pi _ { \theta } } [ r ( \tau ) ] = \mathbb { E } _ { \tau \sim \pi _ { \theta } } [ \nabla _ { \theta } \log ( \pi _ { \theta } ( \tau ) ) r ( \tau ) ] , } \end{array} +$$ + +and the latter quantity can be estimated directly by sampling trajectories according to the policy $\pi _ { \theta }$ + +When we use the discounted variant of the cumulative reward and note that the action of the policy at time $t$ cannot affect its performance at earlier times, we can express our gradient estimate as: + +$$ +\widehat { g _ { \theta } } = \mathbb { E } _ { \tau \sim \pi _ { \theta } } \left[ \sum _ { \left( s _ { t } , a _ { t } \right) \in \tau } \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } | s _ { t } ) \cdot Q _ { \pi _ { \theta } } ( s _ { t } , a _ { t } ) \right] \ , +$$ + +where $Q _ { \pi _ { \theta } } ( s _ { t } , a _ { t } )$ represents the expected returns after taking action $a _ { t }$ from state $s _ { t }$ + +$$ +Q _ { \pi _ { \theta } } ( s _ { t } , a _ { t } ) = \mathbb { E } _ { \pi _ { \theta } } [ R _ { t } | a _ { t } , s _ { t } ] ~ . +$$ + +Value estimation and advantage. Unfortunately, the variance of the expectation in (7) can be (and often is) very large, which makes getting an accurate estimate of this expectation quite challenging. To alleviate this issue, a number of variance reduction techniques have been developed. One of the most popular such techniques is the use of a so-called baseline function, wherein a state-dependent value is subtracted from $Q _ { \pi _ { \theta } }$ . Thus, instead of estimating (7) directly, we use: + +$$ +\widehat { g _ { \theta } } = \mathbb { E } _ { \tau \sim \pi _ { \theta } } \left[ \sum _ { ( s _ { t } , a _ { t } ) \in \tau } \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } | s _ { t } ) \cdot ( Q _ { \pi _ { \theta } } ( s _ { t } , a _ { t } ) - b ( s _ { t } ) ) \right] , +$$ + +where $b ( \cdot )$ is a baseline function of our choice. + +A natural choice of the baseline function is the value function, i.e. + +$$ +V _ { \pi _ { \theta } } ( s _ { t } ) = \mathbb { E } _ { \pi _ { \theta } } [ R _ { t } | s _ { t } ] . +$$ + +When we use the value function as our baseline, the resulting gradient estimation problem becomes: + +$$ +\widehat { g } _ { \theta } = \mathbb { E } _ { \tau \sim \pi _ { \theta } } \left[ \sum _ { \left( s _ { t } , a _ { t } \right) \in \tau } \nabla _ { \theta } \log \pi _ { \theta } ( a _ { t } | s _ { t } ) \cdot A _ { \pi _ { \theta } } ( s _ { t } , a _ { t } ) \right] , +$$ + +where + +$$ +A _ { \pi _ { \theta } } ( s _ { t } , a _ { t } ) = Q _ { \pi _ { \theta } } ( s _ { t } , a _ { t } ) - V _ { \pi _ { \theta } } ( s _ { t } ) +$$ + +is referred to as the advantage of performing action $a _ { t }$ . Different methods of estimating $V _ { \pi _ { \theta } }$ have been proposed, with techniques ranging from moving averages to the use of neural network predictors Schulman et al. (2015b). + +Surrogate Objective. So far, our focus has been on extracting a good estimate of the gradient with respect to the policy parameters $\theta$ . However, it turns out that directly optimizing the cumulative rewards can be challenging. Thus, a modification used by modern policy gradient algorithms is to optimize a “surrogate objective” instead. We will focus on maximizing the following local approximation of the true reward Schulman et al. (2015a): + +$$ +\operatorname* { m a x } _ { \theta } \mathbb { E } _ { ( s _ { t } , a _ { t } ) \sim \pi } \left[ \frac { \pi _ { \theta } ( a _ { t } | s _ { t } ) } { \pi ( a _ { t } | s _ { t } ) } A _ { \pi } ( s _ { t } , a _ { t } ) \right] \qquad \Big ( = \mathbb { E } _ { \pi _ { \theta } } \left[ A _ { \pi } \right] \Big ) , +$$ + +or the normalized advantage variant proposed to reduce variance Schulman et al. (2017): + +$$ +\operatorname* { m a x } _ { \theta } \mathbb { E } _ { ( s _ { t } , a _ { t } ) \sim \pi } \left[ \frac { \pi _ { \theta } ( a _ { t } | s _ { t } ) } { \pi ( a _ { t } | s _ { t } ) } \widehat { A } _ { \pi } ( s _ { t } , a _ { t } ) \right] +$$ + +where + +$$ +\widehat { A } _ { \pi } = \frac { A _ { \pi } - \mu ( A _ { \pi } ) } { \sigma ( A _ { \pi } ) } +$$ + +and $\pi$ is the current policy. + +Trust region methods. The surrogate objective function, although easier to optimize, comes at a cost: the gradient of the surrogate objective is only predictive of the policy gradient locally (at the current policy). Thus, to ensure that our update steps we derive based on the surrogate objective are predictive, they need to be confined to a “trust region” around the current policy. The resulting trust region methods (Kakade, 2001; Schulman et al., 2015a; 2017) try to constrain the local variation of the parameters in policy-space by restricting the distributional distance between successive policies. + +A popular method in this class is trust region policy optimization (TRPO) Schulman et al. (2015a), which constrains the KL divergence between successive policies on the optimization trajectory, leading to the following problem: + +$$ +\begin{array} { r l } { \displaystyle \operatorname* { m a x } _ { \theta } } & { \mathbb { E } _ { ( s _ { t } , a _ { t } ) \sim \pi } \left[ \frac { \pi _ { \theta } ( a _ { t } | s _ { t } ) } { \pi ( a _ { t } | s _ { t } ) } \widehat { A } _ { \pi } ( s _ { t } , a _ { t } ) \right] } \\ { \mathrm { s . t . } } & { D _ { K L } ( \pi _ { \theta } ( \cdot \mid s ) | | \pi ( \cdot \mid s ) ) \leq \delta , \quad \forall s ~ . } \end{array} +$$ + +In practice, this objective is maximized using a second-order approximation of the KL divergence and natural gradient descent, while replacing the worst-case KL constraints over all possible states with an approximation of the mean KL based on the states observed in the current trajectory. + +Proximal policy optimization. In practice, the TRPO algorithm can be computationally costly— the step direction is estimated with nonlinear conjugate gradients, which requires the computation of multiple Hessian-vector products. To address this issue, Schulman et al. Schulman et al. (2017) propose proximal policy optimization (PPO), which utilizes a different objective and does not compute a projection. Concretely, PPO proposes replacing the KL-constrained objective (16) of TRPO by clipping the objective function directly as: + +$$ +\operatorname* { m a x } _ { \theta } \mathbb { E } _ { ( s _ { t } , a _ { t } ) \sim \pi } \left[ \operatorname* { m i n } \left( \mathrm { c l i p } \left( \rho _ { t } , 1 - \varepsilon , 1 + \varepsilon \right) \widehat A _ { \pi } ( s _ { t } , a _ { t } ) , \rho _ { t } \widehat A _ { \pi } ( s _ { t } , a _ { t } ) \right) \right] +$$ + +where + +$$ +\rho _ { t } = \frac { \pi _ { \theta } ( a _ { t } | s _ { t } ) } { \pi ( a _ { t } | s _ { t } ) } +$$ + +In addition to being simpler, PPO is intended to be faster and more sample-efficient than TRPO (Schulman et al., 2017). + +# A.2 EXPERIMENTAL SETUP + +We use the following parameters for PPO and TRPO based on a hyperparameter grid search: + +Table 1: Hyperparameters for PPO and TRPO algorithms. + +
Humanoid-v2Walker2d-v2Hopper-v2
PPOTRPOPPOTRPOPPOTRPO
Timesteps per iteration204820482048204820482048
Discount factor ()0.990.990.990.990.990.99
GAE discount (λ)0.950.950.950.950.950.95
Value network LR0.00010.00030.00030.00030.00020.0002
Value net num. epochs101010101010
Policy net hidden layers[64, 64][64, 64][64,64][64, 64][64, 64][64,64]
Value net hidden layers[64, 64][64, 64][64, 64][64,64][64,64][64,64]
KL constraint (δ)N/A0.07N/A0.04N/A0.13
Fisher est. fractionN/A0.1N/A0.1N/A0.1
Conjugate grad. stepsN/A10N/A10N/A10
CG dampingN/A0.1N/A0.1N/A0.1
Backtracking stepsN/A10N/A10N/A10
Policy LR (Adam)0.00025N/A0.0004N/A0.00045N/A
Policy epochs10N/A10N/A10N/A
PPO Clipping ε0.2N/A0.2N/A0.2N/A
Entropy coeff.0.00.00.00.00.00.0
Reward clipping[-10,10]一[-10,10]一[-10, 10]1
Reward normalizationOnOffOnOffOnOff
State clipping[-10,10]1[-10,10]1[-10, 10]1
+ +All error bars we plot are $9 5 \%$ confidence intervals, obtained via bootstrapped sampling. + +![](images/b09d1c68040ce8eb485111bb7b2e2abe53cfbd6471cca62b69680e86279cd880.jpg) +A.3 STANDARD REWARD PLOTS +Figure 8: Mean reward for the studied policy gradient algorithms on standard MuJoCo benchmark tasks. For each algorithm, we perform 24 random trials using the best performing hyperparameter configuration, with 10 of the random agents shown here. + +![](images/db6c9fe908d6aea5f69b186d211735147297b28e55664e1b76de44ed708f234e.jpg) +Figure 9: Empirical variance of the gradient (c.f. (1)) as a function of the number of state-action pairs used in estimation for policy gradient methods. We obtain multiple gradient estimates using a given number of state-action pairs from the policy at a particular iteration. We then measure the average pairwise cosine similarity between these repeated gradient measurements, along with the $9 5 \%$ confidence intervals (shaded). Each of the colored lines (for a specific algorithm) represents a particular trained agent (we perform multiple trials with the same hyperparameter configurations but different random seeds). The dotted vertical black line (at 2K) indicates the sample regime used for gradient estimation in standard practical implementations of policy gradient methods. + +![](images/eaf9745d78b6e2ca858d1ecaaa601f2cabfcaa1e6ed6129a80fb84f04fa5c503.jpg) +Figure 10: Convergence of gradient estimates to the “true” expected gradient (c.f. (1)). We measure the cosine similarity between the true gradient (approximated using around 1M samples) and gradient estimates, as a function of number of state-action pairs used to obtain the later. For a particular policy and state-action pair count, we obtain multiple estimates of this cosine similarity and then report the average, along with the $9 5 \%$ confidence intervals (shaded). Each of the colored lines (for a specific algorithm) represents a particular trained agent (we perform multiple trials with the same hyperparameter configurations but different random seeds). The dotted vertical black line (at 2K) indicates the sample regime used for gradient estimation in standard practical implementations of policy gradient methods. + +![](images/e16e2ef2341635befb3f55136fc77ddb35d3971e8920c0eb52be8368c4792f55.jpg) +Figure 11: Quality of value prediction in terms of mean relative error (MRE) on train state-action pairs for agents trained to solve the MuJoCo tasks. We see in that the agents do indeed succeed at solving the supervised learning task they are trained for – the train MRE on the GAE-based value loss $( \bar { V } _ { o l d } + \bar { A _ { G A E } } ) ^ { 2 }$ (c.f. (4)) is small (left column). We observe that the returns MRE is quite small as well (right column). + +![](images/9e3757b24093255f914c2704a40dc876cbbcfa06c61e80505054ebedbb2a1d41.jpg) +Figure 12: Quality of value prediction in terms of mean relative error (MRE) on heldout state-action pairs for agents trained to solve MuJoCo tasks. We see in that the agents do indeed succeed at solving the supervised learning task they are trained for – the validation MRE on the GAE-based value loss $( \bar { V _ { o l d } } + A _ { G A E } ) ^ { 2 }$ (c.f. (4)) is small (left column). On the other hand, we see that the returns MRE is still quite high – the learned value function is off by about $5 0 \%$ with respect to the underlying true value function (right column). + +![](images/492ad53b0b623ea1c8de6f1b55e166845454e05d8dda57581999fbe4a8caaa6b.jpg) +Figure 13: Humanoid-v2 – PPO reward landscapes. + +![](images/e71942f5b97f65b0a5ab9ef5c63749592e3e90f18e0d9113c71bc5245529686f.jpg) +Figure 14: Humanoid-v2 – TRPO reward landscapes. + +![](images/92360b41ac24e124107a6a3bf5b314f4ede8a0c180a149bf671750f3eabe09d8.jpg) +Figure 15: Walker2d-v2 – PPO reward landscapes. + +![](images/23b96f1084ce71237a61e54cba14cf0f7f00b45b8386ebcb9faee6b98b81ec8d.jpg) +Figure 16: Walker2d-v2 – TRPO reward landscapes. + +![](images/e1181885263802be8a0612e7e76a1b0d23fe109de1c0065c63a1b6753c7b9d7a.jpg) +Figure 17: Hopper-v2 – PPO reward landscapes. + +![](images/46b2f4466bc46a834c09c86d56fe0c9566d040a40989424a57e720c34dae5420.jpg) +Figure 18: Hopper-v2 – TRPO reward landscapes. + +![](images/32b3359e0631a1a531f1b03fe88fc25bb6a0e14ccbec9ce8e5281584831351b2.jpg) +Figure 19: Humanoid-v2 TRPO landscape concentration (see Figure 5 for a description). + +# + +![](images/c672ad624699fce7bdb3cf34e092df2e245eb748620fa547f07a8e91bbfe6af0.jpg) + + +Figure 20: Humanoid-v2 PPO landscape concentration (see Figure 5 for a description). \ No newline at end of file diff --git a/parse/train/ryxdEkHtPS/ryxdEkHtPS_content_list.json b/parse/train/ryxdEkHtPS/ryxdEkHtPS_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..83167390b484822e534d65110150977633a55daa --- /dev/null +++ b/parse/train/ryxdEkHtPS/ryxdEkHtPS_content_list.json @@ -0,0 +1,1914 @@ +[ + { + "type": "text", + "text": "A CLOSER LOOK AT DEEP POLICY GRADIENTS ", + "text_level": 1, + "bbox": [ + 178, + 98, + 741, + 122 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Andrew Ilyas∗1, Logan Engstrom∗1, Shibani Santurkar1, Dimitris Tsipras1, Firdaus Janoos2, Larry Rudolph1,2, and Aleksander M ˛adry1 ", + "bbox": [ + 259, + 150, + 738, + 179 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1MIT 2Two Sigma {ailyas,engstrom,shibani,tsipras,madry}@mit.edu rudolph@csail.mit.edu, firdaus.janoos@twosigma.com ", + "bbox": [ + 256, + 189, + 745, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 273, + 544, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We study how the behavior of deep policy gradient algorithms reflects the conceptual framework motivating their development. To this end, we propose a finegrained analysis of state-of-the-art methods based on key elements of this framework: gradient estimation, value prediction, and optimization landscapes. Our results show that the behavior of deep policy gradient algorithms often deviates from what their motivating framework would predict: the surrogate objective does not match the true reward landscape, learned value estimators fail to fit the true value function, and gradient estimates poorly correlate with the “true” gradient. The mismatch between predicted and empirical behavior we uncover highlights our poor understanding of current methods, and indicates the need to move beyond current benchmark-centric evaluation methods. ", + "bbox": [ + 233, + 304, + 764, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 479, + 336, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep reinforcement learning (RL) is behind some of the most publicized achievements of modern machine learning (Silver et al., 2017; OpenAI, 2018; Dayarathna et al., 2016; OpenAI et al., 2018). In fact, to many, this framework embodies the promise of the real-world impact of machine learning. However, the deep RL toolkit has not yet attained the same level of engineering stability as, for example, the current deep (supervised) learning framework. Indeed, recent studies demonstrate that state-of-the-art deep RL algorithms suffer from oversensitivity to hyperparameter choices, lack of consistency, and poor reproducibility (Henderson et al., 2017). ", + "bbox": [ + 174, + 511, + 825, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This state of affairs suggests that it might be necessary to re-examine the conceptual underpinnings of deep RL methodology. More precisely, the overarching question that motivates this work is: ", + "bbox": [ + 174, + 616, + 821, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To what degree does current practice in deep RL reflect the principles informing its development? ", + "bbox": [ + 179, + 654, + 818, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our specific focus is on deep policy gradient methods, a widely used class of deep RL algorithms. Our goal is to explore the extent to which state-of-the-art implementations of these methods succeed at realizing the key primitives of the general policy gradient framework. ", + "bbox": [ + 176, + 678, + 821, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our contributions. We take a broader look at policy gradient algorithms and their relation to their underlying framework. With this perspective in mind, we perform a fine-grained examination of key RL primitives as they manifest in practice. Concretely, we study: ", + "bbox": [ + 176, + 734, + 823, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "• Gradient Estimation: we find that even when agents improve in reward, their gradient estimates used in parameter updates poorly correlate with the “true” gradient. We additionally show that gradient estimate quality decays with training progress and task complexity. Finally, we demonstrate that varying the sample regime yields training dynamics that are unexplained by the motivating framework and run contrary to supervised learning intuition. ", + "bbox": [ + 215, + 787, + 821, + 856 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "• Value Prediction: our experiments indicate that value networks successfully solve the supervised learning task they are trained on, but do not fit the true value function. Additionally, employing a value network as a baseline function only marginally decreases the variance of gradient estimates compared to using true value as a baseline (but still dramatically increases agent’s performance compared to using no baseline at all). ", + "bbox": [ + 217, + 861, + 821, + 904 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 227, + 103, + 821, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "• Optimization Landscapes: we show that the optimization landscape induced by modern policy gradient algorithms is often not reflective of the underlying true reward landscape, and that the latter is frequently poorly behaved in the relevant sample regime. ", + "bbox": [ + 218, + 138, + 821, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, our results demonstrate that the motivating theoretical framework for deep RL algorithms is often unpredictive of phenomena arising in practice. This suggests that building reliable deep RL algorithms requires moving past benchmark-centric evaluations to a multi-faceted understanding of their often unintuitive behavior. We conclude (in Section 3) by discussing several areas where such understanding is most critically needed. ", + "bbox": [ + 174, + 193, + 825, + 262 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 EXAMINING THE PRIMITIVES OF DEEP POLICY GRADIENT ALGORITHMS ", + "text_level": 1, + "bbox": [ + 173, + 284, + 810, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we investigate the degree to which our theoretical understanding of RL applies to modern methods. We consider key primitives of policy gradient algorithms: gradient estimation, value prediction and reward fitting. In what follows, we perform a fine-grained analysis of state-ofthe-art policy gradient algorithms (PPO and TRPO) through the lens of these primitives—detailed preliminaries, background, and notation can be found in Appendix A.1. ", + "bbox": [ + 174, + 316, + 825, + 387 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 GRADIENT ESTIMATE QUALITY ", + "text_level": 1, + "bbox": [ + 176, + 406, + 429, + 420 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A central premise of policy gradient methods is that stochastic gradient ascent on a suitable objective function yields a good policy. These algorithms use as a primitive the gradient of that objective function: ", + "bbox": [ + 174, + 433, + 825, + 474 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/b4c1f9d57651640aeb4c5861406f489ab4db657183174ab069a0805ac2964331.jpg", + "text": "$$\n\\hat { g } = \\nabla _ { \\theta } \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi _ { 0 } } \\left[ \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi _ { 0 } ( a _ { t } | s _ { t } ) } \\widehat { A } _ { \\pi _ { 0 } } ( s _ { t } , a _ { t } ) \\right] = \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi _ { 0 } } \\left[ \\frac { \\nabla _ { \\theta } \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi _ { 0 } ( a _ { t } | s _ { t } ) } \\widehat { A } _ { \\pi _ { 0 } } ( s _ { t } , a _ { t } ) \\right] ,\n$$", + "text_format": "latex", + "bbox": [ + 212, + 481, + 784, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where in the above we use standard RL notation (see Appendix A.1 for more details). An underlying assumption behind these methods is that we have access to a reasonable estimate of this quantity. This assumption effectively translates into an assumption that we can accurately estimate the expectation above using an empirical mean of finite (typically $\\sim 1 0 ^ { 3 }$ ) samples. Evidently (since the agent attains a high reward) these estimates are sufficient to consistently improve reward—we are thus interested in the relative quality of these gradient estimates in practice, and the effect of gradient quality on optimization. ", + "bbox": [ + 173, + 522, + 825, + 621 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/564328a99ce532599345890f1e098eb6f8665f3c9643714f6b4dbc1bb4fb90ac.jpg", + "image_caption": [ + "Figure 1: Empirical variance of the estimated gradient (c.f. (1)) as a function of the number of stateaction pairs used in estimation in the MuJoCo Humanoid task. We measure the average pairwise cosine similarity between ten repeated gradient measurements taken from the same policy, with the $9 5 \\%$ confidence intervals (shaded). For each algorithm, we perform multiple trials with the same hyperparameter configurations but different random seeds, shown as repeated lines in the figure. The vertical line (at $x = 2 \\mathsf { K }$ ) indicates the sample regime used for gradient estimation in standard implementations of policy gradient methods. In general, it seems that obtaining tightly concentrated gradient estimates would require significantly more samples than are used in practice, particularly after the first few timesteps. For other tasks – such as Walker2d-v2 and Hopper-v2 – the plots (seen in Appendix Figure 9) have similar trends, except that gradient variance is slightly lower. Confidence intervals calculated with 500 sample bootstrapping. " + ], + "image_footnote": [], + "bbox": [ + 176, + 636, + 821, + 746 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0a11f724707726202a961f210b6493d422e426616dc7c49610d275b33adf05b7.jpg", + "image_caption": [ + "Figure 2: Convergence of gradient estimates (c.f. (1)) to the “true” expected gradient in the MuJoCo Humanoid task. We measure the mean cosine similarity between the “true” gradient approximated using ten million state-action pairs, and ten gradient estimates which use increasing numbers of state-action pairs (with $9 5 \\%$ confidence intervals). For each algorithm, we perform multiple trials with the same hyperparameter configurations but different random seeds. The vertical line (at $x =$ 2K) indicates the sample regime used for gradient estimation in standard implementations of policy gradient methods. Observe that although it is possible to empirically estimate the true gradient, this requires several-fold more samples than are used commonly in practical applications of these algorithms. See additionally that the estimation task becomes more difficult further into training. For other tasks – such as Walker2d-v2 and Hopper-v2 – the plots (seen in Appendix Figure 10) have similar trends, except that gradient estimation is slightly better. Confidence intervals calculated with 500 sample bootstrapping. " + ], + "image_footnote": [], + "bbox": [ + 176, + 101, + 821, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "How accurate are the gradient estimates we compute? To answer this question, we examine two of the most natural measures of estimate quality: the empirical variance and the convergence to the “true” gradient. To evaluate the former, we measure the average pairwise cosine similarity between estimates of the gradient computed from the same policy with independent rollouts (Figure 1). We evaluate the latter by first forming an estimate of the true gradient with a large number of state-action pairs. We then examine the convergence of gradient estimates to this “true” gradient (which we once again measure using cosine similarity) as we increase the number of samples (Figure 2). ", + "bbox": [ + 173, + 422, + 825, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We observe that deep policy gradient methods operate with relatively poor estimates of the gradient, especially as task complexity increases and as training progresses (contrast Humanoid-v2, a “hard” task, to other tasks and contrast successive checkpoints in Figures 1 and 2). This is in spite of the fact that our agents continually improve throughout training, and attain nowhere near the maximum reward possible on each task. In fact, we sometimes observe a zero or even negative correlation in the relevant sample regime1. ", + "bbox": [ + 174, + 526, + 825, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While these results might be reminiscent of the well-studied “noisy gradients” problem in supervised learning (Robbins & Monro, 1951; d’Aspremont, 2008; Kawaguchi, 2016; Safran & Shamir, 2018; Livni et al., 2014; Keskar et al., 2016; Hochreiter & Schmidhuber, 1997), we have very little understanding of how gradient quality affects optimization in the substantially different reinforcement learning setting. For example: ", + "bbox": [ + 176, + 617, + 825, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "• The sample regime in which RL algorithms operate seems to have a profound impact on the robustness and stability of agent training—in particular, many of the sensitivity issues reported by Henderson et al. (2017) are claimed to disappear (Sutskever, 2018) in highersample regimes. Understanding the implications of working in this sample regime, and more generally the impact of sample complexity on training stability remains to be precisely understood. \n• Agent policy networks are trained concurrently with value networks (discussed more in the following section) meant to reduce the variance of gradient estimates. Under our conceptual framework, we might expect these networks to help gradient estimates more as training progresses, contrary to what we observe in Figure 1. The value network also makes the now two-player optimization landscape and training dynamics even more difficult to grasp, as such interactions are poorly understood. ", + "bbox": [ + 215, + 700, + 825, + 877 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "• The relevant measure of sample complexity for many settings (number of state-action pairs) can differ drastically from the number of independent samples used at each training iteration (the number of complete trajectories). The latter quantity (a) tends to be much lower than the number of state-action pairs, and (b) decreases across iterations during training. ", + "bbox": [ + 215, + 103, + 823, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All the above factors make it unclear to what degree our intuition from classical settings transfer to the deep RL regime. And the policy gradient framework, as of now, provides little predictive power regarding the variance of gradient estimates and its impact on reward optimization. ", + "bbox": [ + 178, + 172, + 825, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our results indicate that despite having a rigorous theoretical framework for RL, we lack a precise understanding of the structure of the reward landscape and optimization process. ", + "bbox": [ + 173, + 220, + 823, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 VALUE PREDICTION ", + "text_level": 1, + "bbox": [ + 174, + 268, + 352, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our findings from the previous section motivate a deeper look into gradient estimation. After all, the policy gradient in its original formulation (Sutton et al., 1999) is known to be hard to estimate, and thus algorithms employ a variety of variance reduction methods. The most popular of these techniques is a baseline function. Concretely, an equivalent form of the policy gradient is given by: ", + "bbox": [ + 174, + 295, + 825, + 352 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/f97b21943b4d8c6911ca51665274bf2a059c625f68d0c9ae94e1d287c9aca258.jpg", + "text": "$$\n\\widehat { g } _ { \\theta } = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } \\left[ \\sum _ { ( s _ { t } , a _ { t } ) \\in \\tau } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\cdot ( Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) - b ( s _ { t } ) ) \\right]\n$$", + "text_format": "latex", + "bbox": [ + 284, + 359, + 712, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $b ( s _ { t } )$ is some fixed function of the state $s _ { t }$ . A canonical choice of baseline function is the value function $V _ { \\pi } ( s )$ , the expected return from a given state (more details and motivation in A.1): ", + "bbox": [ + 173, + 417, + 823, + 448 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/95f468dd3a3fa419fc819d3265a3cbc8bd54e69d2c6317d9b9b4dd2659dd1adb.jpg", + "text": "$$\nV _ { \\pi _ { \\theta } } ( s _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ R _ { t } | s _ { t } ] \\ .\n$$", + "text_format": "latex", + "bbox": [ + 421, + 455, + 576, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Indeed, fitting a value-estimating function (Schulman et al., 2015c; Sutton & Barto, 2018) (a neural network, in the deep RL setting) and using it as a baseline function is precisely the approach taken by most deep policy gradient methods. Concretely, one trains a value network $V _ { \\theta _ { t } } ^ { \\pi }$ such that: ", + "bbox": [ + 173, + 481, + 825, + 525 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/f64770b6b2bf7c80a932982ac714a8638b2449b39f25297a583d64774361b58e.jpg", + "text": "$$\n\\theta _ { t } = \\operatorname* { m i n } _ { \\theta } \\mathbb { E } \\left[ \\left( V _ { \\theta } ^ { \\pi } ( s _ { t } ) - ( V _ { \\theta _ { t - 1 } } ^ { \\pi } ( s _ { t } ) + A _ { t } ) \\right) ^ { 2 } \\right]\n$$", + "text_format": "latex", + "bbox": [ + 343, + 534, + 653, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $V _ { \\theta _ { t - 1 } } ^ { \\pi } ( s _ { t } )$ are estimates given by the last value function, and $A _ { t }$ is the advantage of the policy, i.e. the returns minus the estimated values. (Typically, $A _ { t }$ is estimated using generalized advantage estimation, as described in (Schulman et al., 2015c).) Our findings in the previous section prompt us to take a closer look at the value network and its impact on the variance of gradient estimates. ", + "bbox": [ + 173, + 577, + 825, + 636 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3400f4111aa816678fb290803993e451edc4053c39f8141eab8c837fd6cfb8ce.jpg", + "image_caption": [ + "Figure 3: Quality of value prediction in terms of mean relative error (MRE) on heldout state-action pairs for agents trained to solve the MuJoCo Walker2d-v2 task. We observe in (left) that the agents do indeed succeed at solving the supervised learning task they are trained for—the MRE on the GAE-based value loss $( V _ { o l d } \\bar { + } A _ { G A E } ) ^ { 2 }$ (c.f. (4)) is small. On the other hand, in (right) we see that the returns MRE is still quite high—the learned value function is off by about $5 0 \\%$ with respect to the underlying true value function. Similar plots for other MuJoCo tasks are in Appendix A.5. " + ], + "image_footnote": [], + "bbox": [ + 176, + 655, + 818, + 814 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3623c047652f57efd9b89a08dd419cc64f3d9383a3fca27b9f80f90d631a8c19.jpg", + "image_caption": [ + "Figure 4: Efficacy of the value network as a variance reducing baseline for Walker2d-v2 (top) and Hopper-v2 (bottom) agents. We measure the empirical variance of the gradient (c.f. (1)) as a function of the number of state-action pairs used in estimation, for different choices of baseline functions: the value network (used by the agent in training), the “true” value function (fit to the returns using $5 \\cdot 1 0 ^ { 6 }$ state-action pairs sampled from the current policy) and the “zero” value function (i.e. replacing advantages with returns). We observe that using the true value function leads to a significantly lower-variance estimate of the gradient compared to the value network. In turn, employing the value network yields a noticeable variance reduction compared to the zero baseline function, even though this difference may appear rather small in the small-sample regime (2K). Confidence intervals calculated with 10 sample bootstrapping. " + ], + "image_footnote": [], + "bbox": [ + 176, + 101, + 821, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Value prediction as a supervised learning problem. We first analyze the value network through the lens of the supervised learning problem it solves. After all, (4) describes an empirical risk minimization, where a loss is minimized over a set of sampled $\\left( { { s _ { t } } , { a _ { t } } } \\right)$ . So, how does $V _ { \\theta } ^ { \\pi }$ perform as a solution to (4)? And in turn, how does (4) perform as a proxy for learning the true value function? ", + "bbox": [ + 174, + 556, + 823, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our results (Figure 3a) show that the value network does succeed at both fitting the given loss function and generalizing to unseen data, showing low and stable mean relative error (MRE). However, the significant drop in performance as shown in Figure 3 indicates that the supervised learning problem induced by (4) does not lead to $V _ { \\theta } ^ { \\pi }$ learning the underlying true value function. ", + "bbox": [ + 174, + 619, + 823, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Does the value network lead to a reduction in variance? Though evaluating the $V _ { \\theta } ^ { \\pi }$ baseline function as a value predictor as we did above is informative, in the end the sole purpose of the value function is to reduce variance. So: how does using our value function actually impact the variance of our gradient estimates? To answer this question, we compare the variance reduction that results from employing our value network against both a “true” value function and a trivial “zero” baseline function (i.e. simply replacing advantages with returns). Our results, captured in Figure 4, show that the “true” value function yields a much lower-variance estimate of the gradient. This is especially true in the sample regime in which we operate. We note, however, that despite not effectively predicting the true value function or inducing the same degree of variance reduction, the value network does help to some degree (compared to the “zero” baseline). Additionally, the seemingly marginal increase in gradient correlation provided by the value network (compared to the “true” baseline function) turns out to result in a significant improvement in agent performance. (Indeed, agents trained without a baseline reach almost an order of magnitude worse reward.) ", + "bbox": [ + 174, + 708, + 825, + 888 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our findings suggest that we still need a better understanding of the role of the value network in agent training, and raise several questions that we discuss in Section 3. ", + "bbox": [ + 174, + 895, + 821, + 922 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 EXPLORING THE OPTIMIZATION LANDSCAPE ", + "text_level": 1, + "bbox": [ + 178, + 104, + 522, + 117 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Another key assumption of policy gradient algorithms is that first-order updates (w.r.t. policy parameters) actually yield better policies. It is thus natural to examine how valid this assumption is. ", + "bbox": [ + 173, + 131, + 823, + 159 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The true rewards landscape. We begin by examining the landscape of agent reward with respect to the policy parameters. Indeed, even if deep policy gradient methods do not optimize for the true reward directly (e.g. if they use a surrogate objective), the ultimate goal of any policy gradient algorithm is to navigate this landscape. First, Figure 5 shows that while estimating the true reward landscape with a high number of samples yields a relatively smooth reward landscape (perhaps suggesting viability of direct reward optimization), estimating the true reward landscape in the typical, low sample regime results in a landscape that appears jagged and poorly-behaved. The low-sample regime thus gives rise to a certain kind of barrier to direct reward optimization. Indeed, applying our algorithms in this regime makes it impossible to distinguish between good and bad points in the landscape, even though the true underlying landscape is fairly well-behaved. ", + "bbox": [ + 174, + 178, + 825, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The surrogate objective landscape. The untamed nature of the rewards landscape has led to the development of alternate approaches to reward maximization. Recall that an important element of many modern policy gradient methods is the maximization of a surrogate objective function in place of the true rewards (the exact mechanism behind the surrogate objective is detailed in Appendix A.1, and particularly in (14)). The surrogate objective, based on relaxing the policy improvement theorem of Kakade and Langford (Kakade & Langford, 2002), can be viewed as a simplification of the reward maximization objective. ", + "bbox": [ + 174, + 335, + 825, + 433 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As a purported approximation of the true returns, one would expect that the surrogate objective landscape approximates the true reward landscape fairly well. That is, parameters corresponding to good surrogate objective will also correspond to good true reward. ", + "bbox": [ + 178, + 440, + 821, + 482 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 6 shows that in the early stages of training, the optimization landscapes of the true reward and surrogate objective are indeed approximately aligned. However, as training progresses, the surrogate objective becomes much less predictive of the true reward in the relevant sample regime. In particular, we often observe that directions that increase the surrogate objective lead to a decrease of the true reward (see Figures 6, 7). In a higher-sample regime (using several orders of magnitude more samples), we find that PPO and TRPO turn out to behave rather differently. In the case of TRPO, the update direction following the surrogate objective matches the true reward much more closely. However, for PPO we consistently observe landscapes where the step direction leads to lower true reward, even in the high-sample regime. This suggests that even when estimated accurately enough, the surrogate objective might not be an accurate proxy for the true reward. (Recall from Section 2.1 that this is a sample regime where we are able to estimate the true gradient of the reward fairly well.) ", + "bbox": [ + 173, + 489, + 825, + 614 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/988f3e20351aeff96237e9e3741cc998c67968391db52269442dd22142615d7c.jpg", + "image_caption": [ + "2,000 state-action pairs " + ], + "image_footnote": [], + "bbox": [ + 212, + 664, + 372, + 791 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/01017d8520cd26b620c161d401d7fa8673dacd4c65563407b6eb21bc583fe7d4.jpg", + "image_caption": [ + "20,000 state-action pairs " + ], + "image_footnote": [], + "bbox": [ + 416, + 664, + 576, + 790 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9133c34da248b4810278f657205cb43af0e0b9d2b1c5c14e17b97c87979a954b.jpg", + "image_caption": [ + "100,000 state-action pairs ", + "Figure 5: True reward landscape concentration for TRPO on Humanoid-v2. We visualize the landscape at a training iteration 150 while varying the number of trajectories used in reward estimation (each subplot), both in the direction of the step taken and a random direction. Moving one unit along the “step direction” axis corresponds to moving one full step in parameter space. In the random direction one unit corresponds to moving along a random norm 2 Gaussian vector in the parameter space. In practice, the norm of the step is typically an order of magnitude lower than the random direction. While the landscape is very noisy in the low-sample regime, large numbers of samples reveal a well-behaved underlying landscape. See Figures 20, 19 of the Appendix for additional plots. " + ], + "image_footnote": [], + "bbox": [ + 624, + 665, + 781, + 791 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1cb608251899f8943400cff475cebaf32b427d895045caac4cdea6cd45719592.jpg", + "image_caption": [ + "Figure 6: True reward and surrogate objective landscapes for TRPO on the Humanoid-v2 MuJoCo task. We visualize the landscapes in the direction of the update step and a random direction (as in Figure 5). The surrogate objective corresponds to the actual function optimized by the algorithm at each step. We estimate true reward with $1 { \\dot { 0 } } ^ { 6 }$ state-action pairs per point. We compare the landscapes at different points in training and with varying numbers of state-action pairs used in the update step. Early in training the true and surrogate landscapes align fairly well in both sample regimes, but later become misaligned in the low-sample regime. More landscapes in Appendix Figures 13-18. " + ], + "image_footnote": [], + "bbox": [ + 207, + 99, + 769, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 444, + 823, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3 TOWARDS STRONGER FOUNDATIONS FOR DEEP RL ", + "text_level": 1, + "bbox": [ + 174, + 498, + 632, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Deep reinforcement learning (RL) algorithms have shown great practical promise, and are rooted in a well-grounded theoretical framework. However, our results indicate that this framework often fails to provide insight into the practical performance of these algorithms. This disconnect impedes our understanding of why these algorithms succeed (or fail), and is a major barrier to addressing key challenges facing deep RL such as brittleness and poor reproducibility. ", + "bbox": [ + 174, + 532, + 825, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To close this gap, we need to either develop methods that adhere more closely to theory, or build theory that can capture what makes existing policy gradient methods successful. In both cases, the first step is to precisely pinpoint where theory and practice diverge. To this end, we analyze and consolidate our findings from the previous section. ", + "bbox": [ + 174, + 609, + 825, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Gradient estimation. Our analysis in Section 2.1 shows that the quality of gradient estimates that deep policy gradient algorithms use is rather poor. Indeed, even when agents improve, such gradient estimates often poorly correlate with the true gradient (c.f. Figure 2). We also note that gradient correlation decreases as training progresses and task complexity increases. While this certainly does not preclude the estimates from conveying useful signal, the exact underpinnings of this phenomenon in deep RL still elude us. In particular, in Section 2.1 we outline a few keys ways in which the deep RL setting is quite unique and difficult to understand from an optimization perspective, both theoretically and in practice Overall, understanding the impact of gradient estimate quality on deep RL algorithms is challenging and largely unexplored. ", + "bbox": [ + 174, + 672, + 825, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Value prediction. The findings presented in Section 2.2 identify two key issues. First, while the value network successfully solves the supervised learning task it is trained on, it does not accurately model the “true” value function. Second, employing the value network as a baseline does decrease the gradient variance (compared to the trivial (“zero”) baseline). However, this decrease is rather marginal compared to the variance reduction offered by the “true” value function. ", + "bbox": [ + 176, + 805, + 825, + 875 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is natural to wonder whether this failure in modeling the value function is inevitable. For example, how does the loss function used to train the value network impact value prediction and variance reduction? More broadly, we lack an understanding of the precise role of the value network in training. Can we empirically quantify the relationship between variance reduction and performance? And does the value network play a broader role than just variance reduction? ", + "bbox": [ + 176, + 882, + 823, + 924 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0f929acd579bd6910d6fbd92bfdf56ec0bf9d21185294ff124330ed9cb583b77.jpg", + "image_caption": [ + "Figure 7: True reward and surrogate objective landscapes for PPO on the Humanoid-v2 MuJoCo task. See Figure 6 for a description. We observe that early in training the true and surrogate landscapes align well. However, later increasing the surrogate objective leads to lower true reward. " + ], + "image_footnote": [], + "bbox": [ + 207, + 99, + 763, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 376, + 821, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Optimization landscape. We have also seen, in Section 2.3, that the optimization landscape induced by modern policy gradient algorithms, the surrogate objective, is often not reflective of the underlying true reward landscape. We thus need a deeper understanding of why current methods succeed despite these issues, and, more broadly, how to better navigate the true reward landscape. ", + "bbox": [ + 174, + 410, + 825, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 487, + 344, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The idea of using gradient estimates to update neural network-based RL agents dates back at least to the REINFORCE (Williams, 1992) algorithm. Later, Sutton (Sutton et al., 1999) established a unifying framework casting these algorithms as instances of the policy gradient class of algorithms. ", + "bbox": [ + 174, + 518, + 825, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our work focuses on proximal policy optimization (PPO) (Schulman et al., 2017) and trust region policy optimization (TRPO) (Schulman et al., 2015a), which are two of the most prominent policy gradient algorithms used in deep RL, drawing inspiration from works on related algorithms, such as (Peters et al., 2010) and Kakade (2001). ", + "bbox": [ + 174, + 568, + 825, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Many recent works document the brittleness of deep RL algorithms (Henderson et al., 2018; 2017; Islam et al., 2017). (Rajeswaran et al., 2017) and (Mania et al., 2018) demonstrate that on many benchmark tasks, state-of-the-art performance can be attained by augmented randomized search approaches. McCandlish et al. (2018) investigates gradient noise in large-batch settings, and Ahmed et al. (2018) investigates the role of entropy regularization (which we do not study) on optimization. ", + "bbox": [ + 174, + 631, + 825, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 720, + 318, + 737 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we analyze the degree to which key primitives of deep policy gradient algorithms follow their conceptual underpinnings. Our experiments show that these primitives often do not conform to the expected behavior: gradient estimates poorly correlate with the true gradient, better gradient estimates can require lower learning rates and can induce degenerate agent behavior, value networks reduce gradient estimation variance to a significantly smaller extent than the true value, and the underlying optimization landscape can be misleading. ", + "bbox": [ + 174, + 752, + 825, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This demonstrates that there is a significant gap between the theory inspiring current algorithms and the actual mechanisms driving their performance. Overall, our findings suggest that developing a deep RL toolkit that is truly robust and reliable will require moving beyond the current benchmarkdriven evaluation model to a more fine-grained understanding of deep RL algorithms. ", + "bbox": [ + 176, + 843, + 823, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 102, + 287, + 118 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Zafarali Ahmed, Nicolas Le Roux, Mohammad Norouzi, and Dale Schuurmans. Understanding the impact of entropy on policy optimization, 2018. ", + "bbox": [ + 173, + 126, + 825, + 155 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Alexandre d’Aspremont. Smooth optimization with approximate gradient. SIAM Journal on Optimization, 19:1171–1183, 2008. ", + "bbox": [ + 173, + 164, + 823, + 193 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Miyuru Dayarathna, Yonggang Wen, and Rui Fan. Data center energy consumption modeling: A survey. IEEE Communications Surveys & Tutorials, 18(1):732–794, 2016. ", + "bbox": [ + 173, + 203, + 823, + 232 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Peter Henderson, Riashat Islam, Philip Bachman, Joelle Pineau, Doina Precup, and David Meger. Deep reinforcement learning that matters. arXiv preprint arXiv:1709.06560, 2017. ", + "bbox": [ + 174, + 241, + 823, + 271 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Peter Henderson, Joshua Romoff, and Joelle Pineau. Where did my optimum go?: An empirical analysis of gradient descent optimization in policy gradient methods, 2018. ", + "bbox": [ + 173, + 280, + 823, + 309 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sepp Hochreiter and Jürgen Schmidhuber. Flat minima. Neural Computation, 9:1–42, 1997. ", + "bbox": [ + 174, + 318, + 779, + 333 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Riashat Islam, Peter Henderson, Maziar Gomrokchi, and Doina Precup. Reproducibility of benchmarked deep reinforcement learning tasks for continuous control. In ICML Reproducibility in Machine Learning Workshop, 2017. ", + "bbox": [ + 174, + 343, + 825, + 386 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sham M. Kakade. A natural policy gradient. In NIPS, 2001. ", + "bbox": [ + 173, + 395, + 570, + 411 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sham M. Kakade and John Langford. Approximately optimal approximate reinforcement learning. In ICML, 2002. ", + "bbox": [ + 173, + 420, + 825, + 449 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Kenji Kawaguchi. Deep learning without poor local minima. In NIPS, 2016. ", + "bbox": [ + 174, + 458, + 678, + 473 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. CoRR, abs/1609.04836, 2016. ", + "bbox": [ + 174, + 483, + 826, + 525 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Roi Livni, Shai Shalev-Shwartz, and Ohad Shamir. On the computational efficiency of training neural networks. In NIPS, 2014. ", + "bbox": [ + 173, + 535, + 823, + 564 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Horia Mania, Aurelia Guy, and Benjamin Recht. Simple random search provides a competitive approach to reinforcement learning. CoRR, abs/1803.07055, 2018. ", + "bbox": [ + 173, + 574, + 823, + 603 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sam McCandlish, Jared Kaplan, Dario Amodei, and OpenAI Dota Team. An empirical model of large-batch training, 2018. ", + "bbox": [ + 171, + 612, + 825, + 641 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "OpenAI. Openai five. https://blog.openai.com/openai-five/, 2018. ", + "bbox": [ + 173, + 650, + 722, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "OpenAI, :, Marcin Andrychowicz, Bowen Baker, Maciek Chociej, Rafal Jozefowicz, Bob McGrew, Jakub Pachocki, Arthur Petron, Matthias Plappert, Glenn Powell, Alex Ray, Jonas Schneider, Szymon Sidor, Josh Tobin, Peter Welinder, Lilian Weng, and Wojciech Zaremba. Learning dexterous in-hand manipulation, 2018. ", + "bbox": [ + 173, + 674, + 825, + 731 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jan Peters, Katharina Mülling, and Yasemin Altun. Relative entropy policy search. In AAAI, 2010. ", + "bbox": [ + 173, + 741, + 820, + 757 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Aravind Rajeswaran, Kendall Lowrey, Emanuel Todorov, and Sham M. Kakade. Towards generalization and simplicity in continuous control. In NIPS, 2017. ", + "bbox": [ + 171, + 765, + 821, + 795 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Herbert Robbins and Sutton Monro. A stochastic approximation method. Ann. Math. Statist., 22(3): 400–407, 09 1951. doi: 10.1214/aoms/1177729586. URL https://doi.org/10.1214/ aoms/1177729586. ", + "bbox": [ + 173, + 804, + 821, + 847 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Itay Safran and Ohad Shamir. Spurious local minima are common in two-layer relu neural networks. In ICML, 2018. ", + "bbox": [ + 173, + 856, + 821, + 886 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In International Conference on Machine Learning, pp. 1889–1897, 2015a. ", + "bbox": [ + 173, + 895, + 821, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. Highdimensional continuous control using generalized advantage estimation. arXiv preprint arXiv:1506.02438, 2015b. ", + "bbox": [ + 174, + 103, + 823, + 146 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "John Schulman, Philipp Moritz, Sergey Levine, Michael I. Jordan, and Pieter Abbeel. Highdimensional continuous control using generalized advantage estimation. CoRR, abs/1506.02438, 2015c. ", + "bbox": [ + 174, + 155, + 821, + 196 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. ", + "bbox": [ + 171, + 205, + 823, + 236 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go without human knowledge. Nature, 550(7676):354, 2017. ", + "bbox": [ + 174, + 244, + 821, + 287 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ilya Sutskever. Keynote talk. NVIDIA NTECH, 2018. URL https://www.youtube.com/ watch?v $=$ w3ues-NayAs&t $= 4 6 7 s$ . ", + "bbox": [ + 173, + 295, + 820, + 325 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction. MIT press, 2018. ", + "bbox": [ + 173, + 333, + 821, + 349 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Richard S. Sutton, David A. McAllester, Satinder P. Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. In NIPS, 1999. ", + "bbox": [ + 173, + 357, + 821, + 387 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8:229–256, 1992. ", + "bbox": [ + 171, + 395, + 823, + 424 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A APPENDIX ", + "text_level": 1, + "bbox": [ + 176, + 102, + 297, + 117 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A.1 BACKGROUND ", + "text_level": 1, + "bbox": [ + 176, + 136, + 318, + 150 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In the reinforcement learning (RL) setting, an agent interacts with a stateful environment with the goal of maximizing cumulative reward. Formally, we model the environment as a (possibly randomized) function mapping its current state $s$ and an action $a$ supplied by the agent to a new state $s ^ { \\prime }$ and a resulting reward $r$ . The choice of actions of the agent is governed by the its policy $\\pi$ . This policy is a function mapping environment states to a distribution over the actions to take. The objective of an RL algorithm is to find a policy $\\pi$ which maximizes the expected cumulative reward, where the expectation is taken over both environment randomness and the (randomized) action choices. ", + "bbox": [ + 173, + 162, + 825, + 261 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Preliminaries and notation. For a given policy $\\pi$ , we denote by $\\pi ( a | s )$ the probability that this policy assigns to taking action $a$ when the environment is in the state $s$ . We use $r ( s , a )$ to denote the reward that the agent earns for playing action $a$ in response to the state $s$ . A trajectory $\\tau =$ $\\{ ( a _ { t } , s _ { t } ) : t \\in \\{ 1 \\ldots \\overline { { T } } \\} \\}$ is a sequence of state-action pairs that constitutes a valid transcript of interactions of the agent with the environment. (Here, $a _ { t }$ (resp. $s _ { t }$ ) corresponds to the action taken by the agent (resp. state of the environment) in the $t$ -th round of interaction.) We then define $\\pi ( \\tau )$ to be the probability that the trajectory $\\tau$ is executed if the agent follows policy $\\pi$ (provided the initial state of the environment is $s _ { 1 }$ ). Similarly, $\\begin{array} { r } { r ( \\tau ) = \\sum _ { t } r ( \\bar { s _ { t } } , a _ { t } ) } \\end{array}$ denotes the cumulative reward earned by the agent when following this trajectory, where $s _ { t }$ (resp. $a _ { t }$ ) denote the $t { \\cdot }$ -th state (resp. action) in the trajectory $\\tau$ . In the RL setting, however, we often choose to maximize the discounted cumulative reward of a policy $R : = R _ { 1 }$ , where $R _ { t }$ is defined as ", + "bbox": [ + 173, + 280, + 825, + 434 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/7177816dfa109e29770a8a4ba0c8349b35b39718e020cdaaa78900f2e4b057b4.jpg", + "text": "$$\nR _ { t } ( \\tau ) = \\sum _ { t ^ { \\prime } = t } ^ { \\infty } \\gamma ^ { ( t ^ { \\prime } - t ) } r _ { t ^ { \\prime } } \\ .\n$$", + "text_format": "latex", + "bbox": [ + 416, + 443, + 581, + 486 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "and $0 < \\gamma < 1$ is a “discount factor”. The discount factor ensures that the cumulative reward of a policy is well-defined even for an infinite time horizon, and it also incentivizes achieving reward earlier. ", + "bbox": [ + 174, + 494, + 825, + 536 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Policy gradient methods. A widely used class of RL algorithms that will be the focus of our analysis is the class of so-called policy gradient methods. The central idea behind these algorithms is to first parameterize the policy $\\pi _ { \\theta }$ using a parameter vector $\\theta$ . (In the deep RL context, $\\pi _ { \\theta }$ is expressed by a neural network with weights $\\theta$ .) Then, we perform stochastic gradient ascent on the cumulative reward with respect to $\\theta$ . In other words, we want to apply the stochastic ascent approach to our problem: ", + "bbox": [ + 173, + 555, + 825, + 640 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/52f92bb8b77400b38dd49bb2027623ec9202623fb151fa7b124e747141c5ed2b.jpg", + "text": "$$\n\\operatorname* { m a x } _ { \\theta } \\operatorname { \\mathbb { E } } _ { \\tau \\sim \\pi _ { \\theta } } [ r ( \\tau ) ] ~ ,\n$$", + "text_format": "latex", + "bbox": [ + 434, + 650, + 562, + 672 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $\\tau \\sim \\pi _ { \\theta }$ represents trajectories (rollouts) sampled from the distribution induced by the policy $\\pi _ { \\theta }$ . This approach relies on the key observation (Sutton et al., 1999) that under mild conditions, the gradient of our objective can be written as: ", + "bbox": [ + 174, + 683, + 825, + 724 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/cb967e99733cf851321178d8db2dc9741c07633c81aacb7351b7f5c82c2729ae.jpg", + "text": "$$\n\\begin{array} { r } { \\nabla _ { \\theta } \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } [ r ( \\tau ) ] = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } [ \\nabla _ { \\theta } \\log ( \\pi _ { \\theta } ( \\tau ) ) r ( \\tau ) ] , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 341, + 736, + 655, + 753 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "and the latter quantity can be estimated directly by sampling trajectories according to the policy $\\pi _ { \\theta }$ ", + "bbox": [ + 171, + 762, + 821, + 779 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "When we use the discounted variant of the cumulative reward and note that the action of the policy at time $t$ cannot affect its performance at earlier times, we can express our gradient estimate as: ", + "bbox": [ + 173, + 784, + 823, + 813 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/fa60162cee4cb315f9cd5be57fa54136309fa25e9a534683aaca7db116bd4e24.jpg", + "text": "$$\n\\widehat { g _ { \\theta } } = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } \\left[ \\sum _ { \\left( s _ { t } , a _ { t } \\right) \\in \\tau } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\cdot Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) \\right] \\ ,\n$$", + "text_format": "latex", + "bbox": [ + 312, + 821, + 684, + 873 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } )$ represents the expected returns after taking action $a _ { t }$ from state $s _ { t }$ ", + "bbox": [ + 173, + 882, + 733, + 898 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/1e400ac0480533e756b11b13673171d64954b7d6d3e744df8f772414a66f7a18.jpg", + "text": "$$\nQ _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ R _ { t } | a _ { t } , s _ { t } ] ~ .\n$$", + "text_format": "latex", + "bbox": [ + 398, + 909, + 599, + 925 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Value estimation and advantage. Unfortunately, the variance of the expectation in (7) can be (and often is) very large, which makes getting an accurate estimate of this expectation quite challenging. To alleviate this issue, a number of variance reduction techniques have been developed. One of the most popular such techniques is the use of a so-called baseline function, wherein a state-dependent value is subtracted from $Q _ { \\pi _ { \\theta } }$ . Thus, instead of estimating (7) directly, we use: ", + "bbox": [ + 174, + 102, + 825, + 174 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/714c95f4101a60583a3bb3e9d4f137696057e71dc390133e51fed1fe55633dba.jpg", + "text": "$$\n\\widehat { g _ { \\theta } } = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } \\left[ \\sum _ { ( s _ { t } , a _ { t } ) \\in \\tau } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\cdot ( Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) - b ( s _ { t } ) ) \\right] ,\n$$", + "text_format": "latex", + "bbox": [ + 281, + 176, + 715, + 227 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "where $b ( \\cdot )$ is a baseline function of our choice. ", + "bbox": [ + 174, + 229, + 478, + 244 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A natural choice of the baseline function is the value function, i.e. ", + "bbox": [ + 173, + 250, + 606, + 265 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/d536625ed85c40e5b51a28cd421cbbbaedf06c7b52fbd587f8037c122929a8c0.jpg", + "text": "$$\nV _ { \\pi _ { \\theta } } ( s _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ R _ { t } | s _ { t } ] .\n$$", + "text_format": "latex", + "bbox": [ + 419, + 267, + 576, + 285 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "When we use the value function as our baseline, the resulting gradient estimation problem becomes: ", + "bbox": [ + 173, + 287, + 820, + 303 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/22c9b56669575dfaf8f6f350ffa5bf3fd46bb9d7d2bbb38748407885d7357c67.jpg", + "text": "$$\n\\widehat { g } _ { \\theta } = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } \\left[ \\sum _ { \\left( s _ { t } , a _ { t } \\right) \\in \\tau } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\cdot A _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) \\right] ,\n$$", + "text_format": "latex", + "bbox": [ + 313, + 304, + 681, + 354 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "where ", + "bbox": [ + 174, + 356, + 217, + 369 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/97e07d10f77d32fc619464effb2b00b133ed31357ab03cac5ba7f0861f0a87d4.jpg", + "text": "$$\nA _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) = Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) - V _ { \\pi _ { \\theta } } ( s _ { t } )\n$$", + "text_format": "latex", + "bbox": [ + 375, + 373, + 622, + 391 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "is referred to as the advantage of performing action $a _ { t }$ . Different methods of estimating $V _ { \\pi _ { \\theta } }$ have been proposed, with techniques ranging from moving averages to the use of neural network predictors Schulman et al. (2015b). ", + "bbox": [ + 174, + 392, + 825, + 435 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Surrogate Objective. So far, our focus has been on extracting a good estimate of the gradient with respect to the policy parameters $\\theta$ . However, it turns out that directly optimizing the cumulative rewards can be challenging. Thus, a modification used by modern policy gradient algorithms is to optimize a “surrogate objective” instead. We will focus on maximizing the following local approximation of the true reward Schulman et al. (2015a): ", + "bbox": [ + 173, + 449, + 825, + 520 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/c33cc875141510bc73c214796c859a14f7047b6273794d97c0874caa7cec7038.jpg", + "text": "$$\n\\operatorname* { m a x } _ { \\theta } \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi } \\left[ \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi ( a _ { t } | s _ { t } ) } A _ { \\pi } ( s _ { t } , a _ { t } ) \\right] \\qquad \\Big ( = \\mathbb { E } _ { \\pi _ { \\theta } } \\left[ A _ { \\pi } \\right] \\Big ) ,\n$$", + "text_format": "latex", + "bbox": [ + 297, + 522, + 700, + 556 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "or the normalized advantage variant proposed to reduce variance Schulman et al. (2017): ", + "bbox": [ + 169, + 559, + 751, + 573 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/6c93024d7e9186175c5a0326f39c803d8c4ea4e6ba71fd228107b27c706b95a6.jpg", + "text": "$$\n\\operatorname* { m a x } _ { \\theta } \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi } \\left[ \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi ( a _ { t } | s _ { t } ) } \\widehat { A } _ { \\pi } ( s _ { t } , a _ { t } ) \\right]\n$$", + "text_format": "latex", + "bbox": [ + 366, + 575, + 630, + 611 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "where ", + "bbox": [ + 174, + 613, + 217, + 626 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/4398683e8c15fdbb5407bd0c1c00ef8578d14f18944f7f1ed863715cf8cb7974.jpg", + "text": "$$\n\\widehat { A } _ { \\pi } = \\frac { A _ { \\pi } - \\mu ( A _ { \\pi } ) } { \\sigma ( A _ { \\pi } ) }\n$$", + "text_format": "latex", + "bbox": [ + 431, + 626, + 566, + 660 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "and $\\pi$ is the current policy. ", + "bbox": [ + 174, + 662, + 349, + 678 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Trust region methods. The surrogate objective function, although easier to optimize, comes at a cost: the gradient of the surrogate objective is only predictive of the policy gradient locally (at the current policy). Thus, to ensure that our update steps we derive based on the surrogate objective are predictive, they need to be confined to a “trust region” around the current policy. The resulting trust region methods (Kakade, 2001; Schulman et al., 2015a; 2017) try to constrain the local variation of the parameters in policy-space by restricting the distributional distance between successive policies. ", + "bbox": [ + 173, + 690, + 825, + 775 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A popular method in this class is trust region policy optimization (TRPO) Schulman et al. (2015a), which constrains the KL divergence between successive policies on the optimization trajectory, leading to the following problem: ", + "bbox": [ + 174, + 781, + 825, + 824 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/1c76c841578edf86898b179efb59aca8cc71ebb10567e69380e287ef7135954b.jpg", + "text": "$$\n\\begin{array} { r l } { \\displaystyle \\operatorname* { m a x } _ { \\theta } } & { \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi } \\left[ \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi ( a _ { t } | s _ { t } ) } \\widehat { A } _ { \\pi } ( s _ { t } , a _ { t } ) \\right] } \\\\ { \\mathrm { s . t . } } & { D _ { K L } ( \\pi _ { \\theta } ( \\cdot \\mid s ) | | \\pi ( \\cdot \\mid s ) ) \\leq \\delta , \\quad \\forall s ~ . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 354, + 827, + 643, + 881 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In practice, this objective is maximized using a second-order approximation of the KL divergence and natural gradient descent, while replacing the worst-case KL constraints over all possible states with an approximation of the mean KL based on the states observed in the current trajectory. ", + "bbox": [ + 174, + 882, + 825, + 924 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Proximal policy optimization. In practice, the TRPO algorithm can be computationally costly— the step direction is estimated with nonlinear conjugate gradients, which requires the computation of multiple Hessian-vector products. To address this issue, Schulman et al. Schulman et al. (2017) propose proximal policy optimization (PPO), which utilizes a different objective and does not compute a projection. Concretely, PPO proposes replacing the KL-constrained objective (16) of TRPO by clipping the objective function directly as: ", + "bbox": [ + 173, + 102, + 825, + 188 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/1e19169b4d9e120dfdabc5f1d991706c1453aa0fabc2b94420b16107ba0848be.jpg", + "text": "$$\n\\operatorname* { m a x } _ { \\theta } \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi } \\left[ \\operatorname* { m i n } \\left( \\mathrm { c l i p } \\left( \\rho _ { t } , 1 - \\varepsilon , 1 + \\varepsilon \\right) \\widehat A _ { \\pi } ( s _ { t } , a _ { t } ) , \\rho _ { t } \\widehat A _ { \\pi } ( s _ { t } , a _ { t } ) \\right) \\right]\n$$", + "text_format": "latex", + "bbox": [ + 258, + 194, + 741, + 222 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "where ", + "bbox": [ + 173, + 228, + 217, + 242 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/1c272ca739a1a0925bbbb93b732b445452eed9d91d04e517aa57b22c7891cb53.jpg", + "text": "$$\n\\rho _ { t } = \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi ( a _ { t } | s _ { t } ) }\n$$", + "text_format": "latex", + "bbox": [ + 446, + 246, + 550, + 281 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In addition to being simpler, PPO is intended to be faster and more sample-efficient than TRPO (Schulman et al., 2017). ", + "bbox": [ + 171, + 286, + 825, + 315 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A.2 EXPERIMENTAL SETUP ", + "text_level": 1, + "bbox": [ + 176, + 103, + 379, + 118 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We use the following parameters for PPO and TRPO based on a hyperparameter grid search: ", + "bbox": [ + 171, + 128, + 777, + 145 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/ce554d2ca3dd0022c9b04c16ce0c50ead26d6bb99ca0925eaebce576f01b932d.jpg", + "table_caption": [ + "Table 1: Hyperparameters for PPO and TRPO algorithms. " + ], + "table_footnote": [ + "All error bars we plot are $9 5 \\%$ confidence intervals, obtained via bootstrapped sampling. " + ], + "table_body": "
Humanoid-v2Walker2d-v2Hopper-v2
PPOTRPOPPOTRPOPPOTRPO
Timesteps per iteration204820482048204820482048
Discount factor ()0.990.990.990.990.990.99
GAE discount (λ)0.950.950.950.950.950.95
Value network LR0.00010.00030.00030.00030.00020.0002
Value net num. epochs101010101010
Policy net hidden layers[64, 64][64, 64][64,64][64, 64][64, 64][64,64]
Value net hidden layers[64, 64][64, 64][64, 64][64,64][64,64][64,64]
KL constraint (δ)N/A0.07N/A0.04N/A0.13
Fisher est. fractionN/A0.1N/A0.1N/A0.1
Conjugate grad. stepsN/A10N/A10N/A10
CG dampingN/A0.1N/A0.1N/A0.1
Backtracking stepsN/A10N/A10N/A10
Policy LR (Adam)0.00025N/A0.0004N/A0.00045N/A
Policy epochs10N/A10N/A10N/A
PPO Clipping ε0.2N/A0.2N/A0.2N/A
Entropy coeff.0.00.00.00.00.00.0
Reward clipping[-10,10]一[-10,10]一[-10, 10]1
Reward normalizationOnOffOnOffOnOff
State clipping[-10,10]1[-10,10]1[-10, 10]1
", + "bbox": [ + 196, + 183, + 802, + 503 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b09d1c68040ce8eb485111bb7b2e2abe53cfbd6471cca62b69680e86279cd880.jpg", + "image_caption": [ + "A.3 STANDARD REWARD PLOTS ", + "Figure 8: Mean reward for the studied policy gradient algorithms on standard MuJoCo benchmark tasks. For each algorithm, we perform 24 random trials using the best performing hyperparameter configuration, with 10 of the random agents shown here. " + ], + "image_footnote": [], + "bbox": [ + 307, + 128, + 691, + 751 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/db6c9fe908d6aea5f69b186d211735147297b28e55664e1b76de44ed708f234e.jpg", + "image_caption": [ + "Figure 9: Empirical variance of the gradient (c.f. (1)) as a function of the number of state-action pairs used in estimation for policy gradient methods. We obtain multiple gradient estimates using a given number of state-action pairs from the policy at a particular iteration. We then measure the average pairwise cosine similarity between these repeated gradient measurements, along with the $9 5 \\%$ confidence intervals (shaded). Each of the colored lines (for a specific algorithm) represents a particular trained agent (we perform multiple trials with the same hyperparameter configurations but different random seeds). The dotted vertical black line (at 2K) indicates the sample regime used for gradient estimation in standard practical implementations of policy gradient methods. " + ], + "image_footnote": [], + "bbox": [ + 174, + 135, + 821, + 398 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/eaf9745d78b6e2ca858d1ecaaa601f2cabfcaa1e6ed6129a80fb84f04fa5c503.jpg", + "image_caption": [ + "Figure 10: Convergence of gradient estimates to the “true” expected gradient (c.f. (1)). We measure the cosine similarity between the true gradient (approximated using around 1M samples) and gradient estimates, as a function of number of state-action pairs used to obtain the later. For a particular policy and state-action pair count, we obtain multiple estimates of this cosine similarity and then report the average, along with the $9 5 \\%$ confidence intervals (shaded). Each of the colored lines (for a specific algorithm) represents a particular trained agent (we perform multiple trials with the same hyperparameter configurations but different random seeds). The dotted vertical black line (at 2K) indicates the sample regime used for gradient estimation in standard practical implementations of policy gradient methods. " + ], + "image_footnote": [], + "bbox": [ + 174, + 310, + 821, + 573 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/e16e2ef2341635befb3f55136fc77ddb35d3971e8920c0eb52be8368c4792f55.jpg", + "image_caption": [ + "Figure 11: Quality of value prediction in terms of mean relative error (MRE) on train state-action pairs for agents trained to solve the MuJoCo tasks. We see in that the agents do indeed succeed at solving the supervised learning task they are trained for – the train MRE on the GAE-based value loss $( \\bar { V } _ { o l d } + \\bar { A _ { G A E } } ) ^ { 2 }$ (c.f. (4)) is small (left column). We observe that the returns MRE is quite small as well (right column). " + ], + "image_footnote": [], + "bbox": [ + 207, + 137, + 789, + 584 + ], + "page_idx": 17 + }, + { + "type": "image", + "img_path": "images/9e3757b24093255f914c2704a40dc876cbbcfa06c61e80505054ebedbb2a1d41.jpg", + "image_caption": [ + "Figure 12: Quality of value prediction in terms of mean relative error (MRE) on heldout state-action pairs for agents trained to solve MuJoCo tasks. We see in that the agents do indeed succeed at solving the supervised learning task they are trained for – the validation MRE on the GAE-based value loss $( \\bar { V _ { o l d } } + A _ { G A E } ) ^ { 2 }$ (c.f. (4)) is small (left column). On the other hand, we see that the returns MRE is still quite high – the learned value function is off by about $5 0 \\%$ with respect to the underlying true value function (right column). " + ], + "image_footnote": [], + "bbox": [ + 174, + 301, + 820, + 627 + ], + "page_idx": 18 + }, + { + "type": "image", + "img_path": "images/492ad53b0b623ea1c8de6f1b55e166845454e05d8dda57581999fbe4a8caaa6b.jpg", + "image_caption": [ + "Figure 13: Humanoid-v2 – PPO reward landscapes. " + ], + "image_footnote": [], + "bbox": [ + 176, + 133, + 805, + 550 + ], + "page_idx": 19 + }, + { + "type": "image", + "img_path": "images/e71942f5b97f65b0a5ab9ef5c63749592e3e90f18e0d9113c71bc5245529686f.jpg", + "image_caption": [ + "Figure 14: Humanoid-v2 – TRPO reward landscapes. " + ], + "image_footnote": [], + "bbox": [ + 176, + 290, + 805, + 707 + ], + "page_idx": 20 + }, + { + "type": "image", + "img_path": "images/92360b41ac24e124107a6a3bf5b314f4ede8a0c180a149bf671750f3eabe09d8.jpg", + "image_caption": [ + "Figure 15: Walker2d-v2 – PPO reward landscapes. " + ], + "image_footnote": [], + "bbox": [ + 176, + 290, + 805, + 707 + ], + "page_idx": 21 + }, + { + "type": "image", + "img_path": "images/23b96f1084ce71237a61e54cba14cf0f7f00b45b8386ebcb9faee6b98b81ec8d.jpg", + "image_caption": [ + "Figure 16: Walker2d-v2 – TRPO reward landscapes. " + ], + "image_footnote": [], + "bbox": [ + 176, + 290, + 805, + 707 + ], + "page_idx": 22 + }, + { + "type": "image", + "img_path": "images/e1181885263802be8a0612e7e76a1b0d23fe109de1c0065c63a1b6753c7b9d7a.jpg", + "image_caption": [ + "Figure 17: Hopper-v2 – PPO reward landscapes. " + ], + "image_footnote": [], + "bbox": [ + 176, + 290, + 805, + 707 + ], + "page_idx": 23 + }, + { + "type": "image", + "img_path": "images/46b2f4466bc46a834c09c86d56fe0c9566d040a40989424a57e720c34dae5420.jpg", + "image_caption": [ + "Figure 18: Hopper-v2 – TRPO reward landscapes. " + ], + "image_footnote": [], + "bbox": [ + 176, + 290, + 805, + 707 + ], + "page_idx": 24 + }, + { + "type": "image", + "img_path": "images/32b3359e0631a1a531f1b03fe88fc25bb6a0e14ccbec9ce8e5281584831351b2.jpg", + "image_caption": [ + "Figure 19: Humanoid-v2 TRPO landscape concentration (see Figure 5 for a description). " + ], + "image_footnote": [], + "bbox": [ + 196, + 208, + 843, + 789 + ], + "page_idx": 25 + }, + { + "type": "text", + "text": "", + "text_level": 1, + "bbox": [ + 250, + 208, + 403, + 222 + ], + "page_idx": 26 + }, + { + "type": "image", + "img_path": "images/c672ad624699fce7bdb3cf34e092df2e245eb748620fa547f07a8e91bbfe6af0.jpg", + "image_caption": [ + "", + "", + "Figure 20: Humanoid-v2 PPO landscape concentration (see Figure 5 for a description). " + ], + "image_footnote": [], + "bbox": [ + 197, + 212, + 839, + 789 + ], + "page_idx": 26 + } +] \ No newline at end of file diff --git a/parse/train/ryxdEkHtPS/ryxdEkHtPS_middle.json b/parse/train/ryxdEkHtPS/ryxdEkHtPS_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..a82e62472a763629cc261a45cb560ea0aa0a3356 --- /dev/null +++ b/parse/train/ryxdEkHtPS/ryxdEkHtPS_middle.json @@ -0,0 +1,37142 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 109, + 78, + 454, + 97 + ], + "lines": [ + { + "bbox": [ + 106, + 77, + 456, + 99 + ], + "spans": [ + { + "bbox": [ + 106, + 77, + 456, + 99 + ], + "score": 1.0, + "content": "A CLOSER LOOK AT DEEP POLICY GRADIENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 159, + 119, + 452, + 142 + ], + "lines": [ + { + "bbox": [ + 159, + 118, + 452, + 133 + ], + "spans": [ + { + "bbox": [ + 159, + 118, + 452, + 133 + ], + "score": 1.0, + "content": "Andrew Ilyas∗1, Logan Engstrom∗1, Shibani Santurkar1, Dimitris Tsipras1,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 184, + 129, + 425, + 144 + ], + "spans": [ + { + "bbox": [ + 184, + 129, + 425, + 144 + ], + "score": 1.0, + "content": "Firdaus Janoos2, Larry Rudolph1,2, and Aleksander M ˛adry1", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 157, + 150, + 456, + 185 + ], + "lines": [ + { + "bbox": [ + 264, + 149, + 350, + 165 + ], + "spans": [ + { + "bbox": [ + 264, + 149, + 350, + 165 + ], + "score": 1.0, + "content": "1MIT 2Two Sigma", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 167, + 164, + 448, + 175 + ], + "spans": [ + { + "bbox": [ + 167, + 164, + 448, + 175 + ], + "score": 1.0, + "content": "{ailyas,engstrom,shibani,tsipras,madry}@mit.edu", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 156, + 174, + 458, + 186 + ], + "spans": [ + { + "bbox": [ + 156, + 174, + 458, + 186 + ], + "score": 1.0, + "content": "rudolph@csail.mit.edu, firdaus.janoos@twosigma.com", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 278, + 217, + 333, + 230 + ], + "lines": [ + { + "bbox": [ + 277, + 217, + 335, + 231 + ], + "spans": [ + { + "bbox": [ + 277, + 217, + 335, + 231 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 143, + 241, + 468, + 361 + ], + "lines": [ + { + "bbox": [ + 142, + 241, + 469, + 254 + ], + "spans": [ + { + "bbox": [ + 142, + 241, + 469, + 254 + ], + "score": 1.0, + "content": "We study how the behavior of deep policy gradient algorithms reflects the con-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 252, + 469, + 264 + ], + "spans": [ + { + "bbox": [ + 141, + 252, + 469, + 264 + ], + "score": 1.0, + "content": "ceptual framework motivating their development. To this end, we propose a fine-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 263, + 469, + 276 + ], + "spans": [ + { + "bbox": [ + 141, + 263, + 469, + 276 + ], + "score": 1.0, + "content": "grained analysis of state-of-the-art methods based on key elements of this frame-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 274, + 469, + 286 + ], + "spans": [ + { + "bbox": [ + 142, + 274, + 469, + 286 + ], + "score": 1.0, + "content": "work: gradient estimation, value prediction, and optimization landscapes. Our re-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 285, + 470, + 298 + ], + "spans": [ + { + "bbox": [ + 141, + 285, + 470, + 298 + ], + "score": 1.0, + "content": "sults show that the behavior of deep policy gradient algorithms often deviates from", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 296, + 470, + 308 + ], + "spans": [ + { + "bbox": [ + 142, + 296, + 470, + 308 + ], + "score": 1.0, + "content": "what their motivating framework would predict: the surrogate objective does not", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 307, + 469, + 319 + ], + "spans": [ + { + "bbox": [ + 141, + 307, + 469, + 319 + ], + "score": 1.0, + "content": "match the true reward landscape, learned value estimators fail to fit the true value", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 318, + 469, + 329 + ], + "spans": [ + { + "bbox": [ + 141, + 318, + 469, + 329 + ], + "score": 1.0, + "content": "function, and gradient estimates poorly correlate with the “true” gradient. The", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 329, + 469, + 340 + ], + "spans": [ + { + "bbox": [ + 142, + 329, + 469, + 340 + ], + "score": 1.0, + "content": "mismatch between predicted and empirical behavior we uncover highlights our", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 339, + 470, + 352 + ], + "spans": [ + { + "bbox": [ + 141, + 339, + 470, + 352 + ], + "score": 1.0, + "content": "poor understanding of current methods, and indicates the need to move beyond", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 351, + 332, + 362 + ], + "spans": [ + { + "bbox": [ + 141, + 351, + 332, + 362 + ], + "score": 1.0, + "content": "current benchmark-centric evaluation methods.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 12 + }, + { + "type": "title", + "bbox": [ + 108, + 380, + 206, + 393 + ], + "lines": [ + { + "bbox": [ + 105, + 379, + 208, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 208, + 396 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 405, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 405, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 505, + 417 + ], + "score": 1.0, + "content": "Deep reinforcement learning (RL) is behind some of the most publicized achievements of modern", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 416, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 428 + ], + "score": 1.0, + "content": "machine learning (Silver et al., 2017; OpenAI, 2018; Dayarathna et al., 2016; OpenAI et al., 2018).", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 426, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 441 + ], + "score": 1.0, + "content": "In fact, to many, this framework embodies the promise of the real-world impact of machine learning.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 438, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 505, + 450 + ], + "score": 1.0, + "content": "However, the deep RL toolkit has not yet attained the same level of engineering stability as, for", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "score": 1.0, + "content": "example, the current deep (supervised) learning framework. Indeed, recent studies demonstrate that", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 459, + 507, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 507, + 473 + ], + "score": 1.0, + "content": "state-of-the-art deep RL algorithms suffer from oversensitivity to hyperparameter choices, lack of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 471, + 358, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 358, + 484 + ], + "score": 1.0, + "content": "consistency, and poor reproducibility (Henderson et al., 2017).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 488, + 503, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 505, + 501 + ], + "score": 1.0, + "content": "This state of affairs suggests that it might be necessary to re-examine the conceptual underpinnings", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 498, + 487, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 487, + 511 + ], + "score": 1.0, + "content": "of deep RL methodology. More precisely, the overarching question that motivates this work is:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 110, + 518, + 501, + 529 + ], + "lines": [ + { + "bbox": [ + 109, + 516, + 501, + 532 + ], + "spans": [ + { + "bbox": [ + 109, + 516, + 501, + 532 + ], + "score": 1.0, + "content": "To what degree does current practice in deep RL reflect the principles informing its development?", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 108, + 537, + 503, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 504, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 504, + 551 + ], + "score": 1.0, + "content": "Our specific focus is on deep policy gradient methods, a widely used class of deep RL algorithms.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 550, + 504, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 504, + 560 + ], + "score": 1.0, + "content": "Our goal is to explore the extent to which state-of-the-art implementations of these methods succeed", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 560, + 395, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 395, + 573 + ], + "score": 1.0, + "content": "at realizing the key primitives of the general policy gradient framework.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 108, + 582, + 504, + 615 + ], + "lines": [ + { + "bbox": [ + 106, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 506, + 595 + ], + "score": 1.0, + "content": "Our contributions. We take a broader look at policy gradient algorithms and their relation to their", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 592, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 505, + 607 + ], + "score": 1.0, + "content": "underlying framework. With this perspective in mind, we perform a fine-grained examination of key", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 604, + 367, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 367, + 617 + ], + "score": 1.0, + "content": "RL primitives as they manifest in practice. Concretely, we study:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 132, + 624, + 503, + 678 + ], + "lines": [ + { + "bbox": [ + 132, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "• Gradient Estimation: we find that even when agents improve in reward, their gradient", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 141, + 635, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 141, + 635, + 505, + 648 + ], + "score": 1.0, + "content": "estimates used in parameter updates poorly correlate with the “true” gradient. We addition-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 141, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 141, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "ally show that gradient estimate quality decays with training progress and task complexity.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 141, + 657, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 141, + 657, + 505, + 669 + ], + "score": 1.0, + "content": "Finally, we demonstrate that varying the sample regime yields training dynamics that are", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 141, + 668, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 141, + 668, + 505, + 680 + ], + "score": 1.0, + "content": "unexplained by the motivating framework and run contrary to supervised learning intuition.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 133, + 682, + 503, + 716 + ], + "lines": [ + { + "bbox": [ + 132, + 681, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 132, + 681, + 506, + 696 + ], + "score": 1.0, + "content": "• Value Prediction: our experiments indicate that value networks successfully solve the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 142, + 693, + 504, + 705 + ], + "spans": [ + { + "bbox": [ + 142, + 693, + 504, + 705 + ], + "score": 1.0, + "content": "supervised learning task they are trained on, but do not fit the true value function. Addi-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 142, + 705, + 505, + 717 + ], + "spans": [ + { + "bbox": [ + 142, + 705, + 505, + 717 + ], + "score": 1.0, + "content": "tionally, employing a value network as a baseline function only marginally decreases the", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 721, + 358, + 731 + ], + "lines": [ + { + "bbox": [ + 119, + 721, + 358, + 733 + ], + "spans": [ + { + "bbox": [ + 119, + 721, + 358, + 733 + ], + "score": 1.0, + "content": "∗Equal contribution. Work done in part as an intern at Two Sigma.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 109, + 78, + 454, + 97 + ], + "lines": [ + { + "bbox": [ + 106, + 77, + 456, + 99 + ], + "spans": [ + { + "bbox": [ + 106, + 77, + 456, + 99 + ], + "score": 1.0, + "content": "A CLOSER LOOK AT DEEP POLICY GRADIENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 159, + 119, + 452, + 142 + ], + "lines": [ + { + "bbox": [ + 159, + 118, + 452, + 133 + ], + "spans": [ + { + "bbox": [ + 159, + 118, + 452, + 133 + ], + "score": 1.0, + "content": "Andrew Ilyas∗1, Logan Engstrom∗1, Shibani Santurkar1, Dimitris Tsipras1,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 184, + 129, + 425, + 144 + ], + "spans": [ + { + "bbox": [ + 184, + 129, + 425, + 144 + ], + "score": 1.0, + "content": "Firdaus Janoos2, Larry Rudolph1,2, and Aleksander M ˛adry1", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 159, + 118, + 452, + 144 + ] + }, + { + "type": "text", + "bbox": [ + 157, + 150, + 456, + 185 + ], + "lines": [ + { + "bbox": [ + 264, + 149, + 350, + 165 + ], + "spans": [ + { + "bbox": [ + 264, + 149, + 350, + 165 + ], + "score": 1.0, + "content": "1MIT 2Two Sigma", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 167, + 164, + 448, + 175 + ], + "spans": [ + { + "bbox": [ + 167, + 164, + 448, + 175 + ], + "score": 1.0, + "content": "{ailyas,engstrom,shibani,tsipras,madry}@mit.edu", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 156, + 174, + 458, + 186 + ], + "spans": [ + { + "bbox": [ + 156, + 174, + 458, + 186 + ], + "score": 1.0, + "content": "rudolph@csail.mit.edu, firdaus.janoos@twosigma.com", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 156, + 149, + 458, + 186 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 217, + 333, + 230 + ], + "lines": [ + { + "bbox": [ + 277, + 217, + 335, + 231 + ], + "spans": [ + { + "bbox": [ + 277, + 217, + 335, + 231 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 143, + 241, + 468, + 361 + ], + "lines": [ + { + "bbox": [ + 142, + 241, + 469, + 254 + ], + "spans": [ + { + "bbox": [ + 142, + 241, + 469, + 254 + ], + "score": 1.0, + "content": "We study how the behavior of deep policy gradient algorithms reflects the con-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 252, + 469, + 264 + ], + "spans": [ + { + "bbox": [ + 141, + 252, + 469, + 264 + ], + "score": 1.0, + "content": "ceptual framework motivating their development. To this end, we propose a fine-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 263, + 469, + 276 + ], + "spans": [ + { + "bbox": [ + 141, + 263, + 469, + 276 + ], + "score": 1.0, + "content": "grained analysis of state-of-the-art methods based on key elements of this frame-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 274, + 469, + 286 + ], + "spans": [ + { + "bbox": [ + 142, + 274, + 469, + 286 + ], + "score": 1.0, + "content": "work: gradient estimation, value prediction, and optimization landscapes. Our re-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 285, + 470, + 298 + ], + "spans": [ + { + "bbox": [ + 141, + 285, + 470, + 298 + ], + "score": 1.0, + "content": "sults show that the behavior of deep policy gradient algorithms often deviates from", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 296, + 470, + 308 + ], + "spans": [ + { + "bbox": [ + 142, + 296, + 470, + 308 + ], + "score": 1.0, + "content": "what their motivating framework would predict: the surrogate objective does not", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 307, + 469, + 319 + ], + "spans": [ + { + "bbox": [ + 141, + 307, + 469, + 319 + ], + "score": 1.0, + "content": "match the true reward landscape, learned value estimators fail to fit the true value", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 318, + 469, + 329 + ], + "spans": [ + { + "bbox": [ + 141, + 318, + 469, + 329 + ], + "score": 1.0, + "content": "function, and gradient estimates poorly correlate with the “true” gradient. The", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 329, + 469, + 340 + ], + "spans": [ + { + "bbox": [ + 142, + 329, + 469, + 340 + ], + "score": 1.0, + "content": "mismatch between predicted and empirical behavior we uncover highlights our", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 339, + 470, + 352 + ], + "spans": [ + { + "bbox": [ + 141, + 339, + 470, + 352 + ], + "score": 1.0, + "content": "poor understanding of current methods, and indicates the need to move beyond", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 351, + 332, + 362 + ], + "spans": [ + { + "bbox": [ + 141, + 351, + 332, + 362 + ], + "score": 1.0, + "content": "current benchmark-centric evaluation methods.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 12, + "bbox_fs": [ + 141, + 241, + 470, + 362 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 380, + 206, + 393 + ], + "lines": [ + { + "bbox": [ + 105, + 379, + 208, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 208, + 396 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 405, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 405, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 505, + 417 + ], + "score": 1.0, + "content": "Deep reinforcement learning (RL) is behind some of the most publicized achievements of modern", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 416, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 428 + ], + "score": 1.0, + "content": "machine learning (Silver et al., 2017; OpenAI, 2018; Dayarathna et al., 2016; OpenAI et al., 2018).", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 426, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 441 + ], + "score": 1.0, + "content": "In fact, to many, this framework embodies the promise of the real-world impact of machine learning.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 438, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 505, + 450 + ], + "score": 1.0, + "content": "However, the deep RL toolkit has not yet attained the same level of engineering stability as, for", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "score": 1.0, + "content": "example, the current deep (supervised) learning framework. Indeed, recent studies demonstrate that", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 459, + 507, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 507, + 473 + ], + "score": 1.0, + "content": "state-of-the-art deep RL algorithms suffer from oversensitivity to hyperparameter choices, lack of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 471, + 358, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 358, + 484 + ], + "score": 1.0, + "content": "consistency, and poor reproducibility (Henderson et al., 2017).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 405, + 507, + 484 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 488, + 503, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 505, + 501 + ], + "score": 1.0, + "content": "This state of affairs suggests that it might be necessary to re-examine the conceptual underpinnings", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 498, + 487, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 487, + 511 + ], + "score": 1.0, + "content": "of deep RL methodology. More precisely, the overarching question that motivates this work is:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 106, + 487, + 505, + 511 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 518, + 501, + 529 + ], + "lines": [ + { + "bbox": [ + 109, + 516, + 501, + 532 + ], + "spans": [ + { + "bbox": [ + 109, + 516, + 501, + 532 + ], + "score": 1.0, + "content": "To what degree does current practice in deep RL reflect the principles informing its development?", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28, + "bbox_fs": [ + 109, + 516, + 501, + 532 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 537, + 503, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 504, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 504, + 551 + ], + "score": 1.0, + "content": "Our specific focus is on deep policy gradient methods, a widely used class of deep RL algorithms.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 550, + 504, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 504, + 560 + ], + "score": 1.0, + "content": "Our goal is to explore the extent to which state-of-the-art implementations of these methods succeed", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 560, + 395, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 395, + 573 + ], + "score": 1.0, + "content": "at realizing the key primitives of the general policy gradient framework.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 537, + 504, + 573 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 582, + 504, + 615 + ], + "lines": [ + { + "bbox": [ + 106, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 506, + 595 + ], + "score": 1.0, + "content": "Our contributions. We take a broader look at policy gradient algorithms and their relation to their", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 592, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 505, + 607 + ], + "score": 1.0, + "content": "underlying framework. With this perspective in mind, we perform a fine-grained examination of key", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 604, + 367, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 367, + 617 + ], + "score": 1.0, + "content": "RL primitives as they manifest in practice. Concretely, we study:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 106, + 582, + 506, + 617 + ] + }, + { + "type": "text", + "bbox": [ + 132, + 624, + 503, + 678 + ], + "lines": [ + { + "bbox": [ + 132, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "• Gradient Estimation: we find that even when agents improve in reward, their gradient", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 141, + 635, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 141, + 635, + 505, + 648 + ], + "score": 1.0, + "content": "estimates used in parameter updates poorly correlate with the “true” gradient. We addition-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 141, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 141, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "ally show that gradient estimate quality decays with training progress and task complexity.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 141, + 657, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 141, + 657, + 505, + 669 + ], + "score": 1.0, + "content": "Finally, we demonstrate that varying the sample regime yields training dynamics that are", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 141, + 668, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 141, + 668, + 505, + 680 + ], + "score": 1.0, + "content": "unexplained by the motivating framework and run contrary to supervised learning intuition.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37, + "bbox_fs": [ + 132, + 624, + 505, + 680 + ] + }, + { + "type": "text", + "bbox": [ + 133, + 682, + 503, + 716 + ], + "lines": [ + { + "bbox": [ + 132, + 681, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 132, + 681, + 506, + 696 + ], + "score": 1.0, + "content": "• Value Prediction: our experiments indicate that value networks successfully solve the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 142, + 693, + 504, + 705 + ], + "spans": [ + { + "bbox": [ + 142, + 693, + 504, + 705 + ], + "score": 1.0, + "content": "supervised learning task they are trained on, but do not fit the true value function. Addi-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 142, + 705, + 505, + 717 + ], + "spans": [ + { + "bbox": [ + 142, + 705, + 505, + 717 + ], + "score": 1.0, + "content": "tionally, employing a value network as a baseline function only marginally decreases the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 142, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 142, + 82, + 504, + 95 + ], + "score": 1.0, + "content": "variance of gradient estimates compared to using true value as a baseline (but still dramat-", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 142, + 94, + 441, + 106 + ], + "spans": [ + { + "bbox": [ + 142, + 94, + 441, + 106 + ], + "score": 1.0, + "content": "ically increases agent’s performance compared to using no baseline at all).", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 41, + "bbox_fs": [ + 132, + 681, + 506, + 717 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 139, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 142, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 142, + 82, + 504, + 95 + ], + "score": 1.0, + "content": "variance of gradient estimates compared to using true value as a baseline (but still dramat-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 142, + 94, + 441, + 106 + ], + "spans": [ + { + "bbox": [ + 142, + 94, + 441, + 106 + ], + "score": 1.0, + "content": "ically increases agent’s performance compared to using no baseline at all).", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 134, + 110, + 503, + 143 + ], + "lines": [ + { + "bbox": [ + 133, + 109, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 133, + 109, + 505, + 123 + ], + "score": 1.0, + "content": "• Optimization Landscapes: we show that the optimization landscape induced by modern", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 120, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 142, + 120, + 506, + 134 + ], + "score": 1.0, + "content": "policy gradient algorithms is often not reflective of the underlying true reward landscape,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 131, + 453, + 145 + ], + "spans": [ + { + "bbox": [ + 141, + 131, + 453, + 145 + ], + "score": 1.0, + "content": "and that the latter is frequently poorly behaved in the relevant sample regime.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 153, + 505, + 208 + ], + "lines": [ + { + "bbox": [ + 106, + 153, + 504, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 504, + 165 + ], + "score": 1.0, + "content": "Overall, our results demonstrate that the motivating theoretical framework for deep RL algorithms", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 163, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 505, + 177 + ], + "score": 1.0, + "content": "is often unpredictive of phenomena arising in practice. This suggests that building reliable deep RL", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 174, + 506, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 506, + 188 + ], + "score": 1.0, + "content": "algorithms requires moving past benchmark-centric evaluations to a multi-faceted understanding of", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 186, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 186, + 505, + 198 + ], + "score": 1.0, + "content": "their often unintuitive behavior. We conclude (in Section 3) by discussing several areas where such", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 197, + 267, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 197, + 267, + 209 + ], + "score": 1.0, + "content": "understanding is most critically needed.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + }, + { + "type": "title", + "bbox": [ + 106, + 225, + 496, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 498, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 498, + 240 + ], + "score": 1.0, + "content": "2 EXAMINING THE PRIMITIVES OF DEEP POLICY GRADIENT ALGORITHMS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 251, + 505, + 307 + ], + "lines": [ + { + "bbox": [ + 106, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "In this section, we investigate the degree to which our theoretical understanding of RL applies to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 262, + 504, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 504, + 276 + ], + "score": 1.0, + "content": "modern methods. We consider key primitives of policy gradient algorithms: gradient estimation,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "score": 1.0, + "content": "value prediction and reward fitting. In what follows, we perform a fine-grained analysis of state-of-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 285, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 298 + ], + "score": 1.0, + "content": "the-art policy gradient algorithms (PPO and TRPO) through the lens of these primitives—detailed", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 295, + 394, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 394, + 309 + ], + "score": 1.0, + "content": "preliminaries, background, and notation can be found in Appendix A.1.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 108, + 322, + 263, + 333 + ], + "lines": [ + { + "bbox": [ + 105, + 321, + 265, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 265, + 335 + ], + "score": 1.0, + "content": "2.1 GRADIENT ESTIMATE QUALITY", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 505, + 376 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "A central premise of policy gradient methods is that stochastic gradient ascent on a suitable objective", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "function yields a good policy. These algorithms use as a primitive the gradient of that objective", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 365, + 146, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 146, + 377 + ], + "score": 1.0, + "content": "function:", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "interline_equation", + "bbox": [ + 130, + 381, + 480, + 408 + ], + "lines": [ + { + "bbox": [ + 130, + 381, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 480, + 408 + ], + "score": 0.95, + "content": "\\hat { g } = \\nabla _ { \\theta } \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi _ { 0 } } \\left[ \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi _ { 0 } ( a _ { t } | s _ { t } ) } \\widehat { A } _ { \\pi _ { 0 } } ( s _ { t } , a _ { t } ) \\right] = \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi _ { 0 } } \\left[ \\frac { \\nabla _ { \\theta } \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi _ { 0 } ( a _ { t } | s _ { t } ) } \\widehat { A } _ { \\pi _ { 0 } } ( s _ { t } , a _ { t } ) \\right] ,", + "type": "interline_equation", + "image_path": "b4c1f9d57651640aeb4c5861406f489ab4db657183174ab069a0805ac2964331.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 130, + 381, + 480, + 390.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 130, + 390.0, + 480, + 399.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 130, + 399.0, + 480, + 408.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 414, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 505, + 428 + ], + "score": 1.0, + "content": "where in the above we use standard RL notation (see Appendix A.1 for more details). An underlying", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 425, + 504, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 504, + 438 + ], + "score": 1.0, + "content": "assumption behind these methods is that we have access to a reasonable estimate of this quantity.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "score": 1.0, + "content": "This assumption effectively translates into an assumption that we can accurately estimate the expec-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 446, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 330, + 460 + ], + "score": 1.0, + "content": "tation above using an empirical mean of finite (typically", + "type": "text" + }, + { + "bbox": [ + 331, + 447, + 357, + 458 + ], + "score": 0.88, + "content": "\\sim 1 0 ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 446, + 506, + 460 + ], + "score": 1.0, + "content": ") samples. Evidently (since the agent", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 459, + 504, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 504, + 470 + ], + "score": 1.0, + "content": "attains a high reward) these estimates are sufficient to consistently improve reward—we are thus", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 469, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 483 + ], + "score": 1.0, + "content": "interested in the relative quality of these gradient estimates in practice, and the effect of gradient", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 481, + 204, + 494 + ], + "spans": [ + { + "bbox": [ + 106, + 481, + 204, + 494 + ], + "score": 1.0, + "content": "quality on optimization.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 108, + 504, + 503, + 591 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 504, + 503, + 591 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 504, + 503, + 591 + ], + "spans": [ + { + "bbox": [ + 108, + 504, + 503, + 591 + ], + "score": 0.965, + "type": "image", + "image_path": "564328a99ce532599345890f1e098eb6f8665f3c9643714f6b4dbc1bb4fb90ac.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 108, + 504, + 503, + 533.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 108, + 533.0, + 503, + 562.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 108, + 562.0, + 503, + 591.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 600, + 505, + 722 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 601, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 505, + 613 + ], + "score": 1.0, + "content": "Figure 1: Empirical variance of the estimated gradient (c.f. (1)) as a function of the number of state-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 612, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 624 + ], + "score": 1.0, + "content": "action pairs used in estimation in the MuJoCo Humanoid task. We measure the average pairwise", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 623, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 505, + 635 + ], + "score": 1.0, + "content": "cosine similarity between ten repeated gradient measurements taken from the same policy, with the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 127, + 644 + ], + "score": 0.87, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "confidence intervals (shaded). For each algorithm, we perform multiple trials with the same", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 645, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 657 + ], + "score": 1.0, + "content": "hyperparameter configurations but different random seeds, shown as repeated lines in the figure.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 655, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 189, + 669 + ], + "score": 1.0, + "content": "The vertical line (at", + "type": "text" + }, + { + "bbox": [ + 189, + 655, + 223, + 666 + ], + "score": 0.87, + "content": "x = 2 \\mathsf { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 655, + 506, + 669 + ], + "score": 1.0, + "content": ") indicates the sample regime used for gradient estimation in standard", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 666, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 679 + ], + "score": 1.0, + "content": "implementations of policy gradient methods. In general, it seems that obtaining tightly concentrated", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "gradient estimates would require significantly more samples than are used in practice, particularly", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "after the first few timesteps. For other tasks – such as Walker2d-v2 and Hopper-v2 – the plots (seen", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "in Appendix Figure 9) have similar trends, except that gradient variance is slightly lower. Confidence", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 709, + 314, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 314, + 723 + ], + "score": 1.0, + "content": "intervals calculated with 500 sample bootstrapping.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 38 + } + ], + "index": 34.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 139, + 82, + 503, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 142, + 82, + 504, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 134, + 110, + 503, + 143 + ], + "lines": [ + { + "bbox": [ + 133, + 109, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 133, + 109, + 505, + 123 + ], + "score": 1.0, + "content": "• Optimization Landscapes: we show that the optimization landscape induced by modern", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 120, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 142, + 120, + 506, + 134 + ], + "score": 1.0, + "content": "policy gradient algorithms is often not reflective of the underlying true reward landscape,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 131, + 453, + 145 + ], + "spans": [ + { + "bbox": [ + 141, + 131, + 453, + 145 + ], + "score": 1.0, + "content": "and that the latter is frequently poorly behaved in the relevant sample regime.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 133, + 109, + 506, + 145 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 153, + 505, + 208 + ], + "lines": [ + { + "bbox": [ + 106, + 153, + 504, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 504, + 165 + ], + "score": 1.0, + "content": "Overall, our results demonstrate that the motivating theoretical framework for deep RL algorithms", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 163, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 505, + 177 + ], + "score": 1.0, + "content": "is often unpredictive of phenomena arising in practice. This suggests that building reliable deep RL", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 174, + 506, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 506, + 188 + ], + "score": 1.0, + "content": "algorithms requires moving past benchmark-centric evaluations to a multi-faceted understanding of", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 186, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 186, + 505, + 198 + ], + "score": 1.0, + "content": "their often unintuitive behavior. We conclude (in Section 3) by discussing several areas where such", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 197, + 267, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 197, + 267, + 209 + ], + "score": 1.0, + "content": "understanding is most critically needed.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 153, + 506, + 209 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 225, + 496, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 498, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 498, + 240 + ], + "score": 1.0, + "content": "2 EXAMINING THE PRIMITIVES OF DEEP POLICY GRADIENT ALGORITHMS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 251, + 505, + 307 + ], + "lines": [ + { + "bbox": [ + 106, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "In this section, we investigate the degree to which our theoretical understanding of RL applies to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 262, + 504, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 504, + 276 + ], + "score": 1.0, + "content": "modern methods. We consider key primitives of policy gradient algorithms: gradient estimation,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "score": 1.0, + "content": "value prediction and reward fitting. In what follows, we perform a fine-grained analysis of state-of-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 285, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 298 + ], + "score": 1.0, + "content": "the-art policy gradient algorithms (PPO and TRPO) through the lens of these primitives—detailed", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 295, + 394, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 394, + 309 + ], + "score": 1.0, + "content": "preliminaries, background, and notation can be found in Appendix A.1.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 252, + 506, + 309 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 322, + 263, + 333 + ], + "lines": [ + { + "bbox": [ + 105, + 321, + 265, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 265, + 335 + ], + "score": 1.0, + "content": "2.1 GRADIENT ESTIMATE QUALITY", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 505, + 376 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "A central premise of policy gradient methods is that stochastic gradient ascent on a suitable objective", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "function yields a good policy. These algorithms use as a primitive the gradient of that objective", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 365, + 146, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 146, + 377 + ], + "score": 1.0, + "content": "function:", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 342, + 505, + 377 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 130, + 381, + 480, + 408 + ], + "lines": [ + { + "bbox": [ + 130, + 381, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 480, + 408 + ], + "score": 0.95, + "content": "\\hat { g } = \\nabla _ { \\theta } \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi _ { 0 } } \\left[ \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi _ { 0 } ( a _ { t } | s _ { t } ) } \\widehat { A } _ { \\pi _ { 0 } } ( s _ { t } , a _ { t } ) \\right] = \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi _ { 0 } } \\left[ \\frac { \\nabla _ { \\theta } \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi _ { 0 } ( a _ { t } | s _ { t } ) } \\widehat { A } _ { \\pi _ { 0 } } ( s _ { t } , a _ { t } ) \\right] ,", + "type": "interline_equation", + "image_path": "b4c1f9d57651640aeb4c5861406f489ab4db657183174ab069a0805ac2964331.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 130, + 381, + 480, + 390.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 130, + 390.0, + 480, + 399.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 130, + 399.0, + 480, + 408.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 414, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 505, + 428 + ], + "score": 1.0, + "content": "where in the above we use standard RL notation (see Appendix A.1 for more details). An underlying", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 425, + 504, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 504, + 438 + ], + "score": 1.0, + "content": "assumption behind these methods is that we have access to a reasonable estimate of this quantity.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "score": 1.0, + "content": "This assumption effectively translates into an assumption that we can accurately estimate the expec-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 446, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 330, + 460 + ], + "score": 1.0, + "content": "tation above using an empirical mean of finite (typically", + "type": "text" + }, + { + "bbox": [ + 331, + 447, + 357, + 458 + ], + "score": 0.88, + "content": "\\sim 1 0 ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 446, + 506, + 460 + ], + "score": 1.0, + "content": ") samples. Evidently (since the agent", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 459, + 504, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 504, + 470 + ], + "score": 1.0, + "content": "attains a high reward) these estimates are sufficient to consistently improve reward—we are thus", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 469, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 483 + ], + "score": 1.0, + "content": "interested in the relative quality of these gradient estimates in practice, and the effect of gradient", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 481, + 204, + 494 + ], + "spans": [ + { + "bbox": [ + 106, + 481, + 204, + 494 + ], + "score": 1.0, + "content": "quality on optimization.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 413, + 506, + 494 + ] + }, + { + "type": "image", + "bbox": [ + 108, + 504, + 503, + 591 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 504, + 503, + 591 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 504, + 503, + 591 + ], + "spans": [ + { + "bbox": [ + 108, + 504, + 503, + 591 + ], + "score": 0.965, + "type": "image", + "image_path": "564328a99ce532599345890f1e098eb6f8665f3c9643714f6b4dbc1bb4fb90ac.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 108, + 504, + 503, + 533.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 108, + 533.0, + 503, + 562.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 108, + 562.0, + 503, + 591.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 600, + 505, + 722 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 601, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 505, + 613 + ], + "score": 1.0, + "content": "Figure 1: Empirical variance of the estimated gradient (c.f. (1)) as a function of the number of state-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 612, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 624 + ], + "score": 1.0, + "content": "action pairs used in estimation in the MuJoCo Humanoid task. We measure the average pairwise", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 623, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 505, + 635 + ], + "score": 1.0, + "content": "cosine similarity between ten repeated gradient measurements taken from the same policy, with the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 127, + 644 + ], + "score": 0.87, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "confidence intervals (shaded). For each algorithm, we perform multiple trials with the same", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 645, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 657 + ], + "score": 1.0, + "content": "hyperparameter configurations but different random seeds, shown as repeated lines in the figure.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 655, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 189, + 669 + ], + "score": 1.0, + "content": "The vertical line (at", + "type": "text" + }, + { + "bbox": [ + 189, + 655, + 223, + 666 + ], + "score": 0.87, + "content": "x = 2 \\mathsf { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 655, + 506, + 669 + ], + "score": 1.0, + "content": ") indicates the sample regime used for gradient estimation in standard", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 666, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 679 + ], + "score": 1.0, + "content": "implementations of policy gradient methods. In general, it seems that obtaining tightly concentrated", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "gradient estimates would require significantly more samples than are used in practice, particularly", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "after the first few timesteps. For other tasks – such as Walker2d-v2 and Hopper-v2 – the plots (seen", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "in Appendix Figure 9) have similar trends, except that gradient variance is slightly lower. Confidence", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 709, + 314, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 314, + 723 + ], + "score": 1.0, + "content": "intervals calculated with 500 sample bootstrapping.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 38 + } + ], + "index": 34.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 80, + 503, + 167 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 80, + 503, + 167 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 80, + 503, + 167 + ], + "spans": [ + { + "bbox": [ + 108, + 80, + 503, + 167 + ], + "score": 0.964, + "type": "image", + "image_path": "0a11f724707726202a961f210b6493d422e426616dc7c49610d275b33adf05b7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 80, + 503, + 109.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 109.0, + 503, + 138.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 138.0, + 503, + 167.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 177, + 505, + 309 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 177, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 505, + 191 + ], + "score": 1.0, + "content": "Figure 2: Convergence of gradient estimates (c.f. (1)) to the “true” expected gradient in the MuJoCo", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 189, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 505, + 201 + ], + "score": 1.0, + "content": "Humanoid task. We measure the mean cosine similarity between the “true” gradient approximated", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 200, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 506, + 212 + ], + "score": 1.0, + "content": "using ten million state-action pairs, and ten gradient estimates which use increasing numbers of", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 210, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 201, + 223 + ], + "score": 1.0, + "content": "state-action pairs (with", + "type": "text" + }, + { + "bbox": [ + 202, + 210, + 222, + 221 + ], + "score": 0.86, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 210, + 505, + 223 + ], + "score": 1.0, + "content": "confidence intervals). For each algorithm, we perform multiple trials", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 221, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 486, + 235 + ], + "score": 1.0, + "content": "with the same hyperparameter configurations but different random seeds. The vertical line (at", + "type": "text" + }, + { + "bbox": [ + 486, + 222, + 505, + 232 + ], + "score": 0.83, + "content": "x =", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 232, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 505, + 245 + ], + "score": 1.0, + "content": "2K) indicates the sample regime used for gradient estimation in standard implementations of policy", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 243, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 256 + ], + "score": 1.0, + "content": "gradient methods. Observe that although it is possible to empirically estimate the true gradient,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 254, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 267 + ], + "score": 1.0, + "content": "this requires several-fold more samples than are used commonly in practical applications of these", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 264, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 279 + ], + "score": 1.0, + "content": "algorithms. See additionally that the estimation task becomes more difficult further into training.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "For other tasks – such as Walker2d-v2 and Hopper-v2 – the plots (seen in Appendix Figure 10) have", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 287, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 505, + 299 + ], + "score": 1.0, + "content": "similar trends, except that gradient estimation is slightly better. Confidence intervals calculated with", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 297, + 214, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 214, + 312 + ], + "score": 1.0, + "content": "500 sample bootstrapping.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 8.5 + } + ], + "index": 4.75 + }, + { + "type": "text", + "bbox": [ + 106, + 335, + 505, + 412 + ], + "lines": [ + { + "bbox": [ + 106, + 335, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 506, + 347 + ], + "score": 1.0, + "content": "How accurate are the gradient estimates we compute? To answer this question, we examine two", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 345, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 359 + ], + "score": 1.0, + "content": "of the most natural measures of estimate quality: the empirical variance and the convergence to the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "“true” gradient. To evaluate the former, we measure the average pairwise cosine similarity between", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 369, + 504, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 504, + 380 + ], + "score": 1.0, + "content": "estimates of the gradient computed from the same policy with independent rollouts (Figure 1). We", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 379, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 505, + 391 + ], + "score": 1.0, + "content": "evaluate the latter by first forming an estimate of the true gradient with a large number of state-action", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 390, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 506, + 402 + ], + "score": 1.0, + "content": "pairs. We then examine the convergence of gradient estimates to this “true” gradient (which we once", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 401, + 460, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 460, + 414 + ], + "score": 1.0, + "content": "again measure using cosine similarity) as we increase the number of samples (Figure 2).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 417, + 505, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "We observe that deep policy gradient methods operate with relatively poor estimates of the gradient,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "especially as task complexity increases and as training progresses (contrast Humanoid-v2, a “hard”", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 440, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 505, + 452 + ], + "score": 1.0, + "content": "task, to other tasks and contrast successive checkpoints in Figures 1 and 2). This is in spite of the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "score": 1.0, + "content": "fact that our agents continually improve throughout training, and attain nowhere near the maximum", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "reward possible on each task. In fact, we sometimes observe a zero or even negative correlation in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 471, + 222, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 222, + 485 + ], + "score": 1.0, + "content": "the relevant sample regime1.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 108, + 489, + 505, + 545 + ], + "lines": [ + { + "bbox": [ + 106, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "While these results might be reminiscent of the well-studied “noisy gradients” problem in supervised", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 499, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 513 + ], + "score": 1.0, + "content": "learning (Robbins & Monro, 1951; d’Aspremont, 2008; Kawaguchi, 2016; Safran & Shamir, 2018;", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "Livni et al., 2014; Keskar et al., 2016; Hochreiter & Schmidhuber, 1997), we have very little un-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 522, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 505, + 535 + ], + "score": 1.0, + "content": "derstanding of how gradient quality affects optimization in the substantially different reinforcement", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 533, + 228, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 228, + 546 + ], + "score": 1.0, + "content": "learning setting. For example:", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 132, + 555, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 132, + 555, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 506, + 569 + ], + "score": 1.0, + "content": "• The sample regime in which RL algorithms operate seems to have a profound impact on", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 141, + 567, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 141, + 567, + 506, + 580 + ], + "score": 1.0, + "content": "the robustness and stability of agent training—in particular, many of the sensitivity issues", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 141, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 141, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "reported by Henderson et al. (2017) are claimed to disappear (Sutskever, 2018) in higher-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 141, + 589, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 141, + 589, + 506, + 602 + ], + "score": 1.0, + "content": "sample regimes. Understanding the implications of working in this sample regime, and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 141, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 141, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "more generally the impact of sample complexity on training stability remains to be pre-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 142, + 612, + 217, + 623 + ], + "spans": [ + { + "bbox": [ + 142, + 612, + 217, + 623 + ], + "score": 1.0, + "content": "cisely understood.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 135, + 630, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 135, + 630, + 506, + 641 + ], + "score": 1.0, + "content": "• Agent policy networks are trained concurrently with value networks (discussed more in the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 140, + 640, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 140, + 640, + 506, + 654 + ], + "score": 1.0, + "content": "following section) meant to reduce the variance of gradient estimates. Under our conceptual", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 142, + 652, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 142, + 652, + 505, + 664 + ], + "score": 1.0, + "content": "framework, we might expect these networks to help gradient estimates more as training", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 141, + 663, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 141, + 663, + 506, + 674 + ], + "score": 1.0, + "content": "progresses, contrary to what we observe in Figure 1. The value network also makes the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 141, + 673, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 141, + 673, + 506, + 687 + ], + "score": 1.0, + "content": "now two-player optimization landscape and training dynamics even more difficult to grasp,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 142, + 685, + 315, + 696 + ], + "spans": [ + { + "bbox": [ + 142, + 685, + 315, + 696 + ], + "score": 1.0, + "content": "as such interactions are poorly understood.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 38.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 711, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 119, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "1Deep policy gradient algorithms use gradients indirectly to compute steps—in Appendix A.4 we show that", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 721, + 291, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 291, + 733 + ], + "score": 1.0, + "content": "our results also hold true for these computed steps.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 80, + 503, + 167 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 80, + 503, + 167 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 80, + 503, + 167 + ], + "spans": [ + { + "bbox": [ + 108, + 80, + 503, + 167 + ], + "score": 0.964, + "type": "image", + "image_path": "0a11f724707726202a961f210b6493d422e426616dc7c49610d275b33adf05b7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 80, + 503, + 109.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 109.0, + 503, + 138.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 138.0, + 503, + 167.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 177, + 505, + 309 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 177, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 505, + 191 + ], + "score": 1.0, + "content": "Figure 2: Convergence of gradient estimates (c.f. (1)) to the “true” expected gradient in the MuJoCo", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 189, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 505, + 201 + ], + "score": 1.0, + "content": "Humanoid task. We measure the mean cosine similarity between the “true” gradient approximated", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 200, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 506, + 212 + ], + "score": 1.0, + "content": "using ten million state-action pairs, and ten gradient estimates which use increasing numbers of", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 210, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 201, + 223 + ], + "score": 1.0, + "content": "state-action pairs (with", + "type": "text" + }, + { + "bbox": [ + 202, + 210, + 222, + 221 + ], + "score": 0.86, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 210, + 505, + 223 + ], + "score": 1.0, + "content": "confidence intervals). For each algorithm, we perform multiple trials", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 221, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 486, + 235 + ], + "score": 1.0, + "content": "with the same hyperparameter configurations but different random seeds. The vertical line (at", + "type": "text" + }, + { + "bbox": [ + 486, + 222, + 505, + 232 + ], + "score": 0.83, + "content": "x =", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 232, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 505, + 245 + ], + "score": 1.0, + "content": "2K) indicates the sample regime used for gradient estimation in standard implementations of policy", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 243, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 256 + ], + "score": 1.0, + "content": "gradient methods. Observe that although it is possible to empirically estimate the true gradient,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 254, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 267 + ], + "score": 1.0, + "content": "this requires several-fold more samples than are used commonly in practical applications of these", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 264, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 279 + ], + "score": 1.0, + "content": "algorithms. See additionally that the estimation task becomes more difficult further into training.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "For other tasks – such as Walker2d-v2 and Hopper-v2 – the plots (seen in Appendix Figure 10) have", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 287, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 505, + 299 + ], + "score": 1.0, + "content": "similar trends, except that gradient estimation is slightly better. Confidence intervals calculated with", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 297, + 214, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 214, + 312 + ], + "score": 1.0, + "content": "500 sample bootstrapping.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 8.5 + } + ], + "index": 4.75 + }, + { + "type": "text", + "bbox": [ + 106, + 335, + 505, + 412 + ], + "lines": [ + { + "bbox": [ + 106, + 335, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 506, + 347 + ], + "score": 1.0, + "content": "How accurate are the gradient estimates we compute? To answer this question, we examine two", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 345, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 359 + ], + "score": 1.0, + "content": "of the most natural measures of estimate quality: the empirical variance and the convergence to the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "“true” gradient. To evaluate the former, we measure the average pairwise cosine similarity between", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 369, + 504, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 504, + 380 + ], + "score": 1.0, + "content": "estimates of the gradient computed from the same policy with independent rollouts (Figure 1). We", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 379, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 505, + 391 + ], + "score": 1.0, + "content": "evaluate the latter by first forming an estimate of the true gradient with a large number of state-action", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 390, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 506, + 402 + ], + "score": 1.0, + "content": "pairs. We then examine the convergence of gradient estimates to this “true” gradient (which we once", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 401, + 460, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 460, + 414 + ], + "score": 1.0, + "content": "again measure using cosine similarity) as we increase the number of samples (Figure 2).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 335, + 506, + 414 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 417, + 505, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "We observe that deep policy gradient methods operate with relatively poor estimates of the gradient,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "especially as task complexity increases and as training progresses (contrast Humanoid-v2, a “hard”", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 440, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 505, + 452 + ], + "score": 1.0, + "content": "task, to other tasks and contrast successive checkpoints in Figures 1 and 2). This is in spite of the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "score": 1.0, + "content": "fact that our agents continually improve throughout training, and attain nowhere near the maximum", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "reward possible on each task. In fact, we sometimes observe a zero or even negative correlation in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 471, + 222, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 222, + 485 + ], + "score": 1.0, + "content": "the relevant sample regime1.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 417, + 505, + 485 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 489, + 505, + 545 + ], + "lines": [ + { + "bbox": [ + 106, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "While these results might be reminiscent of the well-studied “noisy gradients” problem in supervised", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 499, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 513 + ], + "score": 1.0, + "content": "learning (Robbins & Monro, 1951; d’Aspremont, 2008; Kawaguchi, 2016; Safran & Shamir, 2018;", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "Livni et al., 2014; Keskar et al., 2016; Hochreiter & Schmidhuber, 1997), we have very little un-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 522, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 505, + 535 + ], + "score": 1.0, + "content": "derstanding of how gradient quality affects optimization in the substantially different reinforcement", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 533, + 228, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 228, + 546 + ], + "score": 1.0, + "content": "learning setting. For example:", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 489, + 505, + 546 + ] + }, + { + "type": "list", + "bbox": [ + 132, + 555, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 132, + 555, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 506, + 569 + ], + "score": 1.0, + "content": "• The sample regime in which RL algorithms operate seems to have a profound impact on", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 567, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 141, + 567, + 506, + 580 + ], + "score": 1.0, + "content": "the robustness and stability of agent training—in particular, many of the sensitivity issues", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 141, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 141, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "reported by Henderson et al. (2017) are claimed to disappear (Sutskever, 2018) in higher-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 141, + 589, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 141, + 589, + 506, + 602 + ], + "score": 1.0, + "content": "sample regimes. Understanding the implications of working in this sample regime, and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 141, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 141, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "more generally the impact of sample complexity on training stability remains to be pre-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 142, + 612, + 217, + 623 + ], + "spans": [ + { + "bbox": [ + 142, + 612, + 217, + 623 + ], + "score": 1.0, + "content": "cisely understood.", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 135, + 630, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 135, + 630, + 506, + 641 + ], + "score": 1.0, + "content": "• Agent policy networks are trained concurrently with value networks (discussed more in the", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 140, + 640, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 140, + 640, + 506, + 654 + ], + "score": 1.0, + "content": "following section) meant to reduce the variance of gradient estimates. Under our conceptual", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 142, + 652, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 142, + 652, + 505, + 664 + ], + "score": 1.0, + "content": "framework, we might expect these networks to help gradient estimates more as training", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 141, + 663, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 141, + 663, + 506, + 674 + ], + "score": 1.0, + "content": "progresses, contrary to what we observe in Figure 1. The value network also makes the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 141, + 673, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 141, + 673, + 506, + 687 + ], + "score": 1.0, + "content": "now two-player optimization landscape and training dynamics even more difficult to grasp,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 142, + 685, + 315, + 696 + ], + "spans": [ + { + "bbox": [ + 142, + 685, + 315, + 696 + ], + "score": 1.0, + "content": "as such interactions are poorly understood.", + "type": "text" + } + ], + "index": 44, + "is_list_end_line": true + } + ], + "index": 38.5, + "bbox_fs": [ + 132, + 555, + 506, + 696 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 132, + 82, + 504, + 127 + ], + "lines": [ + { + "bbox": [ + 132, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 132, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "• The relevant measure of sample complexity for many settings (number of state-action pairs)", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 142, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 142, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "can differ drastically from the number of independent samples used at each training itera-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 142, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 142, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "tion (the number of complete trajectories). The latter quantity (a) tends to be much lower", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 115, + 496, + 128 + ], + "spans": [ + { + "bbox": [ + 142, + 115, + 496, + 128 + ], + "score": 1.0, + "content": "than the number of state-action pairs, and (b) decreases across iterations during training.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 109, + 137, + 505, + 170 + ], + "lines": [ + { + "bbox": [ + 108, + 137, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 108, + 137, + 505, + 149 + ], + "score": 1.0, + "content": "All the above factors make it unclear to what degree our intuition from classical settings transfer to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 149, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 108, + 149, + 505, + 160 + ], + "score": 1.0, + "content": "the deep RL regime. And the policy gradient framework, as of now, provides little predictive power", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 159, + 439, + 171 + ], + "spans": [ + { + "bbox": [ + 107, + 159, + 439, + 171 + ], + "score": 1.0, + "content": "regarding the variance of gradient estimates and its impact on reward optimization.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 175, + 504, + 199 + ], + "lines": [ + { + "bbox": [ + 106, + 174, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 505, + 189 + ], + "score": 1.0, + "content": "Our results indicate that despite having a rigorous theoretical framework for RL, we lack a precise", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 186, + 430, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 186, + 430, + 200 + ], + "score": 1.0, + "content": "understanding of the structure of the reward landscape and optimization process.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "title", + "bbox": [ + 107, + 213, + 216, + 225 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 218, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 218, + 226 + ], + "score": 1.0, + "content": "2.2 VALUE PREDICTION", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 234, + 505, + 279 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 505, + 248 + ], + "score": 1.0, + "content": "Our findings from the previous section motivate a deeper look into gradient estimation. After all,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "the policy gradient in its original formulation (Sutton et al., 1999) is known to be hard to estimate,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "and thus algorithms employ a variety of variance reduction methods. The most popular of these", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 268, + 504, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 504, + 281 + ], + "score": 1.0, + "content": "techniques is a baseline function. Concretely, an equivalent form of the policy gradient is given by:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "interline_equation", + "bbox": [ + 174, + 285, + 436, + 325 + ], + "lines": [ + { + "bbox": [ + 174, + 285, + 436, + 325 + ], + "spans": [ + { + "bbox": [ + 174, + 285, + 436, + 325 + ], + "score": 0.95, + "content": "\\widehat { g } _ { \\theta } = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } \\left[ \\sum _ { ( s _ { t } , a _ { t } ) \\in \\tau } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\cdot ( Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) - b ( s _ { t } ) ) \\right]", + "type": "interline_equation", + "image_path": "f97b21943b4d8c6911ca51665274bf2a059c625f68d0c9ae94e1d287c9aca258.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 174, + 285, + 436, + 298.3333333333333 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 174, + 298.3333333333333, + 436, + 311.66666666666663 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 174, + 311.66666666666663, + 436, + 324.99999999999994 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 331, + 504, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 332, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 134, + 344 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 332, + 155, + 344 + ], + "score": 0.92, + "content": "b ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 332, + 299, + 344 + ], + "score": 1.0, + "content": "is some fixed function of the state", + "type": "text" + }, + { + "bbox": [ + 300, + 334, + 309, + 343 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 332, + 505, + 344 + ], + "score": 1.0, + "content": ". A canonical choice of baseline function is the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 343, + 498, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 166, + 356 + ], + "score": 1.0, + "content": "value function", + "type": "text" + }, + { + "bbox": [ + 166, + 343, + 190, + 355 + ], + "score": 0.92, + "content": "V _ { \\pi } ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 343, + 498, + 356 + ], + "score": 1.0, + "content": ", the expected return from a given state (more details and motivation in A.1):", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "interline_equation", + "bbox": [ + 258, + 361, + 353, + 375 + ], + "lines": [ + { + "bbox": [ + 258, + 361, + 353, + 375 + ], + "spans": [ + { + "bbox": [ + 258, + 361, + 353, + 375 + ], + "score": 0.93, + "content": "V _ { \\pi _ { \\theta } } ( s _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ R _ { t } | s _ { t } ] \\ .", + "type": "interline_equation", + "image_path": "95f468dd3a3fa419fc819d3265a3cbc8bd54e69d2c6317d9b9b4dd2659dd1adb.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 258, + 361, + 353, + 375 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 381, + 505, + 416 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "Indeed, fitting a value-estimating function (Schulman et al., 2015c; Sutton & Barto, 2018) (a neural", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 393, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 406 + ], + "score": 1.0, + "content": "network, in the deep RL setting) and using it as a baseline function is precisely the approach taken", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 402, + 476, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 419, + 417 + ], + "score": 1.0, + "content": "by most deep policy gradient methods. Concretely, one trains a value network", + "type": "text" + }, + { + "bbox": [ + 419, + 404, + 434, + 417 + ], + "score": 0.9, + "content": "V _ { \\theta _ { t } } ^ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 402, + 476, + 417 + ], + "score": 1.0, + "content": "such that:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 423, + 400, + 451 + ], + "lines": [ + { + "bbox": [ + 210, + 423, + 400, + 451 + ], + "spans": [ + { + "bbox": [ + 210, + 423, + 400, + 451 + ], + "score": 0.94, + "content": "\\theta _ { t } = \\operatorname* { m i n } _ { \\theta } \\mathbb { E } \\left[ \\left( V _ { \\theta } ^ { \\pi } ( s _ { t } ) - ( V _ { \\theta _ { t - 1 } } ^ { \\pi } ( s _ { t } ) + A _ { t } ) \\right) ^ { 2 } \\right]", + "type": "interline_equation", + "image_path": "f64770b6b2bf7c80a932982ac714a8638b2449b39f25297a583d64774361b58e.jpg" + } + ] + } + ], + "index": 23.5, + "virtual_lines": [ + { + "bbox": [ + 210, + 423, + 400, + 437.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 210, + 437.0, + 400, + 451.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 457, + 505, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 457, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 133, + 472 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 458, + 172, + 471 + ], + "score": 0.93, + "content": "V _ { \\theta _ { t - 1 } } ^ { \\pi } ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 457, + 371, + 472 + ], + "score": 1.0, + "content": "are estimates given by the last value function, and", + "type": "text" + }, + { + "bbox": [ + 371, + 459, + 383, + 469 + ], + "score": 0.89, + "content": "A _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 457, + 506, + 472 + ], + "score": 1.0, + "content": "is the advantage of the policy,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 326, + 483 + ], + "score": 1.0, + "content": "i.e. the returns minus the estimated values. (Typically,", + "type": "text" + }, + { + "bbox": [ + 326, + 470, + 338, + 481 + ], + "score": 0.88, + "content": "A _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "is estimated using generalized advantage", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 481, + 506, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 506, + 494 + ], + "score": 1.0, + "content": "estimation, as described in (Schulman et al., 2015c).) Our findings in the previous section prompt", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 492, + 493, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 493, + 505 + ], + "score": 1.0, + "content": "us to take a closer look at the value network and its impact on the variance of gradient estimates.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5 + }, + { + "type": "image", + "bbox": [ + 108, + 519, + 501, + 645 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 519, + 501, + 645 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 519, + 501, + 645 + ], + "spans": [ + { + "bbox": [ + 108, + 519, + 501, + 645 + ], + "score": 0.971, + "type": "image", + "image_path": "3400f4111aa816678fb290803993e451edc4053c39f8141eab8c837fd6cfb8ce.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 108, + 519, + 501, + 561.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 108, + 561.0, + 501, + 603.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 108, + 603.0, + 501, + 645.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 654, + 505, + 722 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "score": 1.0, + "content": "Figure 3: Quality of value prediction in terms of mean relative error (MRE) on heldout state-action", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 104, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "pairs for agents trained to solve the MuJoCo Walker2d-v2 task. We observe in (left) that the agents", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "do indeed succeed at solving the supervised learning task they are trained for—the MRE on the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 198, + 701 + ], + "score": 1.0, + "content": "GAE-based value loss", + "type": "text" + }, + { + "bbox": [ + 199, + 687, + 267, + 700 + ], + "score": 0.91, + "content": "( V _ { o l d } \\bar { + } A _ { G A E } ) ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 688, + 506, + 701 + ], + "score": 1.0, + "content": "(c.f. (4)) is small. On the other hand, in (right) we see that", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 422, + 711 + ], + "score": 1.0, + "content": "the returns MRE is still quite high—the learned value function is off by about", + "type": "text" + }, + { + "bbox": [ + 422, + 699, + 442, + 709 + ], + "score": 0.88, + "content": "5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "with respect to", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 709, + 484, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 484, + 722 + ], + "score": 1.0, + "content": "the underlying true value function. Similar plots for other MuJoCo tasks are in Appendix A.5.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5 + } + ], + "index": 32.25 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 8 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 132, + 82, + 504, + 127 + ], + "lines": [ + { + "bbox": [ + 132, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 132, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "• The relevant measure of sample complexity for many settings (number of state-action pairs)", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 142, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 142, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "can differ drastically from the number of independent samples used at each training itera-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 142, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 142, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "tion (the number of complete trajectories). The latter quantity (a) tends to be much lower", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 115, + 496, + 128 + ], + "spans": [ + { + "bbox": [ + 142, + 115, + 496, + 128 + ], + "score": 1.0, + "content": "than the number of state-action pairs, and (b) decreases across iterations during training.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 132, + 82, + 505, + 128 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 137, + 505, + 170 + ], + "lines": [ + { + "bbox": [ + 108, + 137, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 108, + 137, + 505, + 149 + ], + "score": 1.0, + "content": "All the above factors make it unclear to what degree our intuition from classical settings transfer to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 149, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 108, + 149, + 505, + 160 + ], + "score": 1.0, + "content": "the deep RL regime. And the policy gradient framework, as of now, provides little predictive power", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 159, + 439, + 171 + ], + "spans": [ + { + "bbox": [ + 107, + 159, + 439, + 171 + ], + "score": 1.0, + "content": "regarding the variance of gradient estimates and its impact on reward optimization.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 107, + 137, + 505, + 171 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 175, + 504, + 199 + ], + "lines": [ + { + "bbox": [ + 106, + 174, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 505, + 189 + ], + "score": 1.0, + "content": "Our results indicate that despite having a rigorous theoretical framework for RL, we lack a precise", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 186, + 430, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 186, + 430, + 200 + ], + "score": 1.0, + "content": "understanding of the structure of the reward landscape and optimization process.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 174, + 505, + 200 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 213, + 216, + 225 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 218, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 218, + 226 + ], + "score": 1.0, + "content": "2.2 VALUE PREDICTION", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 234, + 505, + 279 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 505, + 248 + ], + "score": 1.0, + "content": "Our findings from the previous section motivate a deeper look into gradient estimation. After all,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "the policy gradient in its original formulation (Sutton et al., 1999) is known to be hard to estimate,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "and thus algorithms employ a variety of variance reduction methods. The most popular of these", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 268, + 504, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 504, + 281 + ], + "score": 1.0, + "content": "techniques is a baseline function. Concretely, an equivalent form of the policy gradient is given by:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 234, + 505, + 281 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 174, + 285, + 436, + 325 + ], + "lines": [ + { + "bbox": [ + 174, + 285, + 436, + 325 + ], + "spans": [ + { + "bbox": [ + 174, + 285, + 436, + 325 + ], + "score": 0.95, + "content": "\\widehat { g } _ { \\theta } = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } \\left[ \\sum _ { ( s _ { t } , a _ { t } ) \\in \\tau } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\cdot ( Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) - b ( s _ { t } ) ) \\right]", + "type": "interline_equation", + "image_path": "f97b21943b4d8c6911ca51665274bf2a059c625f68d0c9ae94e1d287c9aca258.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 174, + 285, + 436, + 298.3333333333333 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 174, + 298.3333333333333, + 436, + 311.66666666666663 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 174, + 311.66666666666663, + 436, + 324.99999999999994 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 331, + 504, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 332, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 134, + 344 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 332, + 155, + 344 + ], + "score": 0.92, + "content": "b ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 332, + 299, + 344 + ], + "score": 1.0, + "content": "is some fixed function of the state", + "type": "text" + }, + { + "bbox": [ + 300, + 334, + 309, + 343 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 332, + 505, + 344 + ], + "score": 1.0, + "content": ". A canonical choice of baseline function is the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 343, + 498, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 166, + 356 + ], + "score": 1.0, + "content": "value function", + "type": "text" + }, + { + "bbox": [ + 166, + 343, + 190, + 355 + ], + "score": 0.92, + "content": "V _ { \\pi } ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 343, + 498, + 356 + ], + "score": 1.0, + "content": ", the expected return from a given state (more details and motivation in A.1):", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 106, + 332, + 505, + 356 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 258, + 361, + 353, + 375 + ], + "lines": [ + { + "bbox": [ + 258, + 361, + 353, + 375 + ], + "spans": [ + { + "bbox": [ + 258, + 361, + 353, + 375 + ], + "score": 0.93, + "content": "V _ { \\pi _ { \\theta } } ( s _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ R _ { t } | s _ { t } ] \\ .", + "type": "interline_equation", + "image_path": "95f468dd3a3fa419fc819d3265a3cbc8bd54e69d2c6317d9b9b4dd2659dd1adb.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 258, + 361, + 353, + 375 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 381, + 505, + 416 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "Indeed, fitting a value-estimating function (Schulman et al., 2015c; Sutton & Barto, 2018) (a neural", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 393, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 406 + ], + "score": 1.0, + "content": "network, in the deep RL setting) and using it as a baseline function is precisely the approach taken", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 402, + 476, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 419, + 417 + ], + "score": 1.0, + "content": "by most deep policy gradient methods. Concretely, one trains a value network", + "type": "text" + }, + { + "bbox": [ + 419, + 404, + 434, + 417 + ], + "score": 0.9, + "content": "V _ { \\theta _ { t } } ^ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 402, + 476, + 417 + ], + "score": 1.0, + "content": "such that:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 381, + 505, + 417 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 423, + 400, + 451 + ], + "lines": [ + { + "bbox": [ + 210, + 423, + 400, + 451 + ], + "spans": [ + { + "bbox": [ + 210, + 423, + 400, + 451 + ], + "score": 0.94, + "content": "\\theta _ { t } = \\operatorname* { m i n } _ { \\theta } \\mathbb { E } \\left[ \\left( V _ { \\theta } ^ { \\pi } ( s _ { t } ) - ( V _ { \\theta _ { t - 1 } } ^ { \\pi } ( s _ { t } ) + A _ { t } ) \\right) ^ { 2 } \\right]", + "type": "interline_equation", + "image_path": "f64770b6b2bf7c80a932982ac714a8638b2449b39f25297a583d64774361b58e.jpg" + } + ] + } + ], + "index": 23.5, + "virtual_lines": [ + { + "bbox": [ + 210, + 423, + 400, + 437.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 210, + 437.0, + 400, + 451.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 457, + 505, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 457, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 133, + 472 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 458, + 172, + 471 + ], + "score": 0.93, + "content": "V _ { \\theta _ { t - 1 } } ^ { \\pi } ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 457, + 371, + 472 + ], + "score": 1.0, + "content": "are estimates given by the last value function, and", + "type": "text" + }, + { + "bbox": [ + 371, + 459, + 383, + 469 + ], + "score": 0.89, + "content": "A _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 457, + 506, + 472 + ], + "score": 1.0, + "content": "is the advantage of the policy,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 326, + 483 + ], + "score": 1.0, + "content": "i.e. the returns minus the estimated values. (Typically,", + "type": "text" + }, + { + "bbox": [ + 326, + 470, + 338, + 481 + ], + "score": 0.88, + "content": "A _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "is estimated using generalized advantage", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 481, + 506, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 506, + 494 + ], + "score": 1.0, + "content": "estimation, as described in (Schulman et al., 2015c).) Our findings in the previous section prompt", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 492, + 493, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 493, + 505 + ], + "score": 1.0, + "content": "us to take a closer look at the value network and its impact on the variance of gradient estimates.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 457, + 506, + 505 + ] + }, + { + "type": "image", + "bbox": [ + 108, + 519, + 501, + 645 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 519, + 501, + 645 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 519, + 501, + 645 + ], + "spans": [ + { + "bbox": [ + 108, + 519, + 501, + 645 + ], + "score": 0.971, + "type": "image", + "image_path": "3400f4111aa816678fb290803993e451edc4053c39f8141eab8c837fd6cfb8ce.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 108, + 519, + 501, + 561.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 108, + 561.0, + 501, + 603.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 108, + 603.0, + 501, + 645.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 654, + 505, + 722 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "score": 1.0, + "content": "Figure 3: Quality of value prediction in terms of mean relative error (MRE) on heldout state-action", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 104, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "pairs for agents trained to solve the MuJoCo Walker2d-v2 task. We observe in (left) that the agents", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "do indeed succeed at solving the supervised learning task they are trained for—the MRE on the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 198, + 701 + ], + "score": 1.0, + "content": "GAE-based value loss", + "type": "text" + }, + { + "bbox": [ + 199, + 687, + 267, + 700 + ], + "score": 0.91, + "content": "( V _ { o l d } \\bar { + } A _ { G A E } ) ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 688, + 506, + 701 + ], + "score": 1.0, + "content": "(c.f. (4)) is small. On the other hand, in (right) we see that", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 422, + 711 + ], + "score": 1.0, + "content": "the returns MRE is still quite high—the learned value function is off by about", + "type": "text" + }, + { + "bbox": [ + 422, + 699, + 442, + 709 + ], + "score": 0.88, + "content": "5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "with respect to", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 709, + 484, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 484, + 722 + ], + "score": 1.0, + "content": "the underlying true value function. Similar plots for other MuJoCo tasks are in Appendix A.5.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5 + } + ], + "index": 32.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 80, + 503, + 287 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 80, + 503, + 287 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 80, + 503, + 287 + ], + "spans": [ + { + "bbox": [ + 108, + 80, + 503, + 287 + ], + "score": 0.978, + "type": "image", + "image_path": "3623c047652f57efd9b89a08dd419cc64f3d9383a3fca27b9f80f90d631a8c19.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 80, + 503, + 149.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 149.0, + 503, + 218.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 218.0, + 503, + 287.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 297, + 505, + 407 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "Figure 4: Efficacy of the value network as a variance reducing baseline for Walker2d-v2 (top) and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "Hopper-v2 (bottom) agents. We measure the empirical variance of the gradient (c.f. (1)) as a function", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 318, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 332 + ], + "score": 1.0, + "content": "of the number of state-action pairs used in estimation, for different choices of baseline functions:", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 329, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 343 + ], + "score": 1.0, + "content": "the value network (used by the agent in training), the “true” value function (fit to the returns us-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 104, + 340, + 122, + 354 + ], + "score": 1.0, + "content": "ing", + "type": "text" + }, + { + "bbox": [ + 122, + 340, + 152, + 352 + ], + "score": 0.88, + "content": "5 \\cdot 1 0 ^ { 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 340, + 505, + 354 + ], + "score": 1.0, + "content": "state-action pairs sampled from the current policy) and the “zero” value function (i.e.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "replacing advantages with returns). We observe that using the true value function leads to a signif-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 361, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 506, + 377 + ], + "score": 1.0, + "content": "icantly lower-variance estimate of the gradient compared to the value network. In turn, employing", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "score": 1.0, + "content": "the value network yields a noticeable variance reduction compared to the zero baseline function,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 384, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 398 + ], + "score": 1.0, + "content": "even though this difference may appear rather small in the small-sample regime (2K). Confidence", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 394, + 309, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 309, + 410 + ], + "score": 1.0, + "content": "intervals calculated with 10 sample bootstrapping.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 7.5 + } + ], + "index": 4.25 + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 504, + 485 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "score": 1.0, + "content": "Value prediction as a supervised learning problem. We first analyze the value network through", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "the lens of the supervised learning problem it solves. After all, (4) describes an empirical risk", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 463, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 355, + 476 + ], + "score": 1.0, + "content": "minimization, where a loss is minimized over a set of sampled", + "type": "text" + }, + { + "bbox": [ + 356, + 463, + 385, + 475 + ], + "score": 0.92, + "content": "\\left( { { s _ { t } } , { a _ { t } } } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 463, + 444, + 476 + ], + "score": 1.0, + "content": ". So, how does", + "type": "text" + }, + { + "bbox": [ + 444, + 463, + 459, + 475 + ], + "score": 0.89, + "content": "V _ { \\theta } ^ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 463, + 506, + 476 + ], + "score": 1.0, + "content": "perform as", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "score": 1.0, + "content": "a solution to (4)? And in turn, how does (4) perform as a proxy for learning the true value function?", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 491, + 504, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 490, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 506, + 504 + ], + "score": 1.0, + "content": "Our results (Figure 3a) show that the value network does succeed at both fitting the given loss func-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 501, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 515 + ], + "score": 1.0, + "content": "tion and generalizing to unseen data, showing low and stable mean relative error (MRE). However,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "the significant drop in performance as shown in Figure 3 indicates that the supervised learning prob-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 524, + 442, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 250, + 537 + ], + "score": 1.0, + "content": "lem induced by (4) does not lead to", + "type": "text" + }, + { + "bbox": [ + 250, + 524, + 264, + 536 + ], + "score": 0.89, + "content": "V _ { \\theta } ^ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 524, + 442, + 537 + ], + "score": 1.0, + "content": "learning the underlying true value function.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 453, + 574 + ], + "score": 1.0, + "content": "Does the value network lead to a reduction in variance? Though evaluating the", + "type": "text" + }, + { + "bbox": [ + 453, + 562, + 468, + 573 + ], + "score": 0.89, + "content": "V _ { \\theta } ^ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "baseline", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "score": 1.0, + "content": "function as a value predictor as we did above is informative, in the end the sole purpose of the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "value function is to reduce variance. So: how does using our value function actually impact the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "variance of our gradient estimates? To answer this question, we compare the variance reduction", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 604, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 619 + ], + "score": 1.0, + "content": "that results from employing our value network against both a “true” value function and a trivial", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 614, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 506, + 629 + ], + "score": 1.0, + "content": "“zero” baseline function (i.e. simply replacing advantages with returns). Our results, captured in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "Figure 4, show that the “true” value function yields a much lower-variance estimate of the gradient.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "This is especially true in the sample regime in which we operate. We note, however, that despite", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "not effectively predicting the true value function or inducing the same degree of variance reduction,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "the value network does help to some degree (compared to the “zero” baseline). Additionally, the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "score": 1.0, + "content": "seemingly marginal increase in gradient correlation provided by the value network (compared to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 681, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 695 + ], + "score": 1.0, + "content": "the “true” baseline function) turns out to result in a significant improvement in agent performance.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 692, + 480, + 706 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 480, + 706 + ], + "score": 1.0, + "content": "(Indeed, agents trained without a baseline reach almost an order of magnitude worse reward.)", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 721 + ], + "score": 1.0, + "content": "Our findings suggest that we still need a better understanding of the role of the value network in", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 721, + 391, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 391, + 732 + ], + "score": 1.0, + "content": "agent training, and raise several questions that we discuss in Section 3.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 80, + 503, + 287 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 80, + 503, + 287 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 80, + 503, + 287 + ], + "spans": [ + { + "bbox": [ + 108, + 80, + 503, + 287 + ], + "score": 0.978, + "type": "image", + "image_path": "3623c047652f57efd9b89a08dd419cc64f3d9383a3fca27b9f80f90d631a8c19.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 80, + 503, + 149.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 149.0, + 503, + 218.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 218.0, + 503, + 287.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 297, + 505, + 407 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "Figure 4: Efficacy of the value network as a variance reducing baseline for Walker2d-v2 (top) and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "Hopper-v2 (bottom) agents. We measure the empirical variance of the gradient (c.f. (1)) as a function", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 318, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 332 + ], + "score": 1.0, + "content": "of the number of state-action pairs used in estimation, for different choices of baseline functions:", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 329, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 343 + ], + "score": 1.0, + "content": "the value network (used by the agent in training), the “true” value function (fit to the returns us-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 104, + 340, + 122, + 354 + ], + "score": 1.0, + "content": "ing", + "type": "text" + }, + { + "bbox": [ + 122, + 340, + 152, + 352 + ], + "score": 0.88, + "content": "5 \\cdot 1 0 ^ { 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 340, + 505, + 354 + ], + "score": 1.0, + "content": "state-action pairs sampled from the current policy) and the “zero” value function (i.e.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "replacing advantages with returns). We observe that using the true value function leads to a signif-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 361, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 506, + 377 + ], + "score": 1.0, + "content": "icantly lower-variance estimate of the gradient compared to the value network. In turn, employing", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "score": 1.0, + "content": "the value network yields a noticeable variance reduction compared to the zero baseline function,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 384, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 398 + ], + "score": 1.0, + "content": "even though this difference may appear rather small in the small-sample regime (2K). Confidence", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 394, + 309, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 309, + 410 + ], + "score": 1.0, + "content": "intervals calculated with 10 sample bootstrapping.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 7.5 + } + ], + "index": 4.25 + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 504, + 485 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "score": 1.0, + "content": "Value prediction as a supervised learning problem. We first analyze the value network through", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "the lens of the supervised learning problem it solves. After all, (4) describes an empirical risk", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 463, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 355, + 476 + ], + "score": 1.0, + "content": "minimization, where a loss is minimized over a set of sampled", + "type": "text" + }, + { + "bbox": [ + 356, + 463, + 385, + 475 + ], + "score": 0.92, + "content": "\\left( { { s _ { t } } , { a _ { t } } } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 463, + 444, + 476 + ], + "score": 1.0, + "content": ". So, how does", + "type": "text" + }, + { + "bbox": [ + 444, + 463, + 459, + 475 + ], + "score": 0.89, + "content": "V _ { \\theta } ^ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 463, + 506, + 476 + ], + "score": 1.0, + "content": "perform as", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "score": 1.0, + "content": "a solution to (4)? And in turn, how does (4) perform as a proxy for learning the true value function?", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 440, + 506, + 487 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 491, + 504, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 490, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 506, + 504 + ], + "score": 1.0, + "content": "Our results (Figure 3a) show that the value network does succeed at both fitting the given loss func-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 501, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 515 + ], + "score": 1.0, + "content": "tion and generalizing to unseen data, showing low and stable mean relative error (MRE). However,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "the significant drop in performance as shown in Figure 3 indicates that the supervised learning prob-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 524, + 442, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 250, + 537 + ], + "score": 1.0, + "content": "lem induced by (4) does not lead to", + "type": "text" + }, + { + "bbox": [ + 250, + 524, + 264, + 536 + ], + "score": 0.89, + "content": "V _ { \\theta } ^ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 524, + 442, + 537 + ], + "score": 1.0, + "content": "learning the underlying true value function.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 490, + 506, + 537 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 453, + 574 + ], + "score": 1.0, + "content": "Does the value network lead to a reduction in variance? Though evaluating the", + "type": "text" + }, + { + "bbox": [ + 453, + 562, + 468, + 573 + ], + "score": 0.89, + "content": "V _ { \\theta } ^ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "baseline", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "score": 1.0, + "content": "function as a value predictor as we did above is informative, in the end the sole purpose of the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "value function is to reduce variance. So: how does using our value function actually impact the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "variance of our gradient estimates? To answer this question, we compare the variance reduction", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 604, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 619 + ], + "score": 1.0, + "content": "that results from employing our value network against both a “true” value function and a trivial", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 614, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 506, + 629 + ], + "score": 1.0, + "content": "“zero” baseline function (i.e. simply replacing advantages with returns). Our results, captured in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "Figure 4, show that the “true” value function yields a much lower-variance estimate of the gradient.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "This is especially true in the sample regime in which we operate. We note, however, that despite", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "not effectively predicting the true value function or inducing the same degree of variance reduction,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "the value network does help to some degree (compared to the “zero” baseline). Additionally, the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "score": 1.0, + "content": "seemingly marginal increase in gradient correlation provided by the value network (compared to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 681, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 695 + ], + "score": 1.0, + "content": "the “true” baseline function) turns out to result in a significant improvement in agent performance.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 692, + 480, + 706 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 480, + 706 + ], + "score": 1.0, + "content": "(Indeed, agents trained without a baseline reach almost an order of magnitude worse reward.)", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 560, + 506, + 706 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 721 + ], + "score": 1.0, + "content": "Our findings suggest that we still need a better understanding of the role of the value network in", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 721, + 391, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 391, + 732 + ], + "score": 1.0, + "content": "agent training, and raise several questions that we discuss in Section 3.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 109, + 83, + 320, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 322, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 322, + 95 + ], + "score": 1.0, + "content": "2.3 EXPLORING THE OPTIMIZATION LANDSCAPE", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 104, + 504, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 505, + 118 + ], + "score": 1.0, + "content": "Another key assumption of policy gradient algorithms is that first-order updates (w.r.t. policy pa-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 115, + 496, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 496, + 127 + ], + "score": 1.0, + "content": "rameters) actually yield better policies. It is thus natural to examine how valid this assumption is.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 107, + 141, + 505, + 252 + ], + "lines": [ + { + "bbox": [ + 106, + 141, + 506, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 506, + 154 + ], + "score": 1.0, + "content": "The true rewards landscape. We begin by examining the landscape of agent reward with respect", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 153, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 164 + ], + "score": 1.0, + "content": "to the policy parameters. Indeed, even if deep policy gradient methods do not optimize for the true", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 163, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 505, + 176 + ], + "score": 1.0, + "content": "reward directly (e.g. if they use a surrogate objective), the ultimate goal of any policy gradient al-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 174, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 505, + 187 + ], + "score": 1.0, + "content": "gorithm is to navigate this landscape. First, Figure 5 shows that while estimating the true reward", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 199 + ], + "score": 1.0, + "content": "landscape with a high number of samples yields a relatively smooth reward landscape (perhaps sug-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "score": 1.0, + "content": "gesting viability of direct reward optimization), estimating the true reward landscape in the typical,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 207, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 505, + 220 + ], + "score": 1.0, + "content": "low sample regime results in a landscape that appears jagged and poorly-behaved. The low-sample", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 217, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 505, + 232 + ], + "score": 1.0, + "content": "regime thus gives rise to a certain kind of barrier to direct reward optimization. Indeed, applying", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 230, + 504, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 504, + 241 + ], + "score": 1.0, + "content": "our algorithms in this regime makes it impossible to distinguish between good and bad points in the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 240, + 414, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 414, + 253 + ], + "score": 1.0, + "content": "landscape, even though the true underlying landscape is fairly well-behaved.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 266, + 505, + 343 + ], + "lines": [ + { + "bbox": [ + 105, + 266, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 506, + 280 + ], + "score": 1.0, + "content": "The surrogate objective landscape. The untamed nature of the rewards landscape has led to the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 277, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 506, + 290 + ], + "score": 1.0, + "content": "development of alternate approaches to reward maximization. Recall that an important element of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 289, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 505, + 301 + ], + "score": 1.0, + "content": "many modern policy gradient methods is the maximization of a surrogate objective function in place", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 299, + 504, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 504, + 312 + ], + "score": 1.0, + "content": "of the true rewards (the exact mechanism behind the surrogate objective is detailed in Appendix A.1,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 309, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 323 + ], + "score": 1.0, + "content": "and particularly in (14)). The surrogate objective, based on relaxing the policy improvement theorem", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 321, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 506, + 335 + ], + "score": 1.0, + "content": "of Kakade and Langford (Kakade & Langford, 2002), can be viewed as a simplification of the reward", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 333, + 205, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 205, + 344 + ], + "score": 1.0, + "content": "maximization objective.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 109, + 349, + 503, + 382 + ], + "lines": [ + { + "bbox": [ + 106, + 348, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 505, + 362 + ], + "score": 1.0, + "content": "As a purported approximation of the true returns, one would expect that the surrogate objective", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 360, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 373 + ], + "score": 1.0, + "content": "landscape approximates the true reward landscape fairly well. That is, parameters corresponding to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 370, + 374, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 374, + 384 + ], + "score": 1.0, + "content": "good surrogate objective will also correspond to good true reward.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 388, + 505, + 487 + ], + "lines": [ + { + "bbox": [ + 106, + 387, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 506, + 402 + ], + "score": 1.0, + "content": "Figure 6 shows that in the early stages of training, the optimization landscapes of the true reward", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 412 + ], + "score": 1.0, + "content": "and surrogate objective are indeed approximately aligned. However, as training progresses, the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 409, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 506, + 423 + ], + "score": 1.0, + "content": "surrogate objective becomes much less predictive of the true reward in the relevant sample regime. In", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 420, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 506, + 434 + ], + "score": 1.0, + "content": "particular, we often observe that directions that increase the surrogate objective lead to a decrease of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 432, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 506, + 444 + ], + "score": 1.0, + "content": "the true reward (see Figures 6, 7). In a higher-sample regime (using several orders of magnitude more", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "samples), we find that PPO and TRPO turn out to behave rather differently. In the case of TRPO,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 454, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 504, + 466 + ], + "score": 1.0, + "content": "the update direction following the surrogate objective matches the true reward much more closely.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "However, for PPO we consistently observe landscapes where the step direction leads to lower true", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "reward, even in the high-sample regime. This suggests that even when estimated accurately enough,", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 130, + 526, + 228, + 627 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 127, + 508, + 221, + 519 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 125, + 506, + 222, + 521 + ], + "spans": [ + { + "bbox": [ + 125, + 506, + 222, + 521 + ], + "score": 1.0, + "content": "2,000 state-action pairs", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "image_body", + "bbox": [ + 130, + 526, + 228, + 627 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 228, + 627 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 228, + 627 + ], + "score": 0.54, + "type": "image", + "image_path": "988f3e20351aeff96237e9e3741cc998c67968391db52269442dd22142615d7c.jpg" + } + ] + } + ], + "index": 36.5, + "virtual_lines": [ + { + "bbox": [ + 130, + 526, + 228, + 576.5 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 130, + 576.5, + 228, + 627.0 + ], + "spans": [], + "index": 38 + } + ] + } + ], + "index": 34.25 + }, + { + "type": "image", + "bbox": [ + 255, + 526, + 353, + 626 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 256, + 508, + 355, + 519 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 255, + 506, + 356, + 521 + ], + "spans": [ + { + "bbox": [ + 255, + 506, + 356, + 521 + ], + "score": 1.0, + "content": "20,000 state-action pairs", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "image_body", + "bbox": [ + 255, + 526, + 353, + 626 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 255, + 526, + 353, + 626 + ], + "spans": [ + { + "bbox": [ + 255, + 526, + 353, + 626 + ], + "score": 0.521, + "type": "image", + "image_path": "01017d8520cd26b620c161d401d7fa8673dacd4c65563407b6eb21bc583fe7d4.jpg" + } + ] + } + ], + "index": 37.5, + "virtual_lines": [ + { + "bbox": [ + 255, + 526, + 353, + 576.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 255, + 576.0, + 353, + 626.0 + ], + "spans": [], + "index": 39 + } + ] + } + ], + "index": 35.25 + }, + { + "type": "image", + "bbox": [ + 382, + 527, + 478, + 627 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 386, + 509, + 489, + 519 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 385, + 507, + 490, + 520 + ], + "spans": [ + { + "bbox": [ + 385, + 507, + 490, + 520 + ], + "score": 1.0, + "content": "100,000 state-action pairs", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "image_body", + "bbox": [ + 382, + 527, + 478, + 627 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 382, + 527, + 478, + 627 + ], + "spans": [ + { + "bbox": [ + 382, + 527, + 478, + 627 + ], + "score": 0.841, + "type": "image", + "image_path": "9133c34da248b4810278f657205cb43af0e0b9d2b1c5c14e17b97c87979a954b.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 382, + 527, + 478, + 577.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 382, + 577.0, + 478, + 627.0 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 640, + 505, + 729 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "score": 1.0, + "content": "Figure 5: True reward landscape concentration for TRPO on Humanoid-v2. We visualize the land-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 652, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 506, + 664 + ], + "score": 1.0, + "content": "scape at a training iteration 150 while varying the number of trajectories used in reward estimation", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 662, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 505, + 676 + ], + "score": 1.0, + "content": "(each subplot), both in the direction of the step taken and a random direction. Moving one unit along", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "score": 1.0, + "content": "the “step direction” axis corresponds to moving one full step in parameter space. In the random di-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 684, + 505, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 505, + 696 + ], + "score": 1.0, + "content": "rection one unit corresponds to moving along a random norm 2 Gaussian vector in the parameter", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 695, + 505, + 708 + ], + "spans": [ + { + "bbox": [ + 104, + 695, + 505, + 708 + ], + "score": 1.0, + "content": "space. In practice, the norm of the step is typically an order of magnitude lower than the random", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 706, + 505, + 719 + ], + "spans": [ + { + "bbox": [ + 105, + 706, + 505, + 719 + ], + "score": 1.0, + "content": "direction. While the landscape is very noisy in the low-sample regime, large numbers of samples", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 718, + 504, + 730 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 504, + 730 + ], + "score": 1.0, + "content": "reveal a well-behaved underlying landscape. See Figures 20, 19 of the Appendix for additional plots.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 44.5 + } + ], + "index": 38.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 109, + 83, + 320, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 322, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 322, + 95 + ], + "score": 1.0, + "content": "2.3 EXPLORING THE OPTIMIZATION LANDSCAPE", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 104, + 504, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 505, + 118 + ], + "score": 1.0, + "content": "Another key assumption of policy gradient algorithms is that first-order updates (w.r.t. policy pa-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 115, + 496, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 496, + 127 + ], + "score": 1.0, + "content": "rameters) actually yield better policies. It is thus natural to examine how valid this assumption is.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 102, + 505, + 127 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 141, + 505, + 252 + ], + "lines": [ + { + "bbox": [ + 106, + 141, + 506, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 506, + 154 + ], + "score": 1.0, + "content": "The true rewards landscape. We begin by examining the landscape of agent reward with respect", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 153, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 164 + ], + "score": 1.0, + "content": "to the policy parameters. Indeed, even if deep policy gradient methods do not optimize for the true", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 163, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 505, + 176 + ], + "score": 1.0, + "content": "reward directly (e.g. if they use a surrogate objective), the ultimate goal of any policy gradient al-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 174, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 505, + 187 + ], + "score": 1.0, + "content": "gorithm is to navigate this landscape. First, Figure 5 shows that while estimating the true reward", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 199 + ], + "score": 1.0, + "content": "landscape with a high number of samples yields a relatively smooth reward landscape (perhaps sug-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "score": 1.0, + "content": "gesting viability of direct reward optimization), estimating the true reward landscape in the typical,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 207, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 505, + 220 + ], + "score": 1.0, + "content": "low sample regime results in a landscape that appears jagged and poorly-behaved. The low-sample", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 217, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 505, + 232 + ], + "score": 1.0, + "content": "regime thus gives rise to a certain kind of barrier to direct reward optimization. Indeed, applying", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 230, + 504, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 504, + 241 + ], + "score": 1.0, + "content": "our algorithms in this regime makes it impossible to distinguish between good and bad points in the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 240, + 414, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 414, + 253 + ], + "score": 1.0, + "content": "landscape, even though the true underlying landscape is fairly well-behaved.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 141, + 506, + 253 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 266, + 505, + 343 + ], + "lines": [ + { + "bbox": [ + 105, + 266, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 506, + 280 + ], + "score": 1.0, + "content": "The surrogate objective landscape. The untamed nature of the rewards landscape has led to the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 277, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 506, + 290 + ], + "score": 1.0, + "content": "development of alternate approaches to reward maximization. Recall that an important element of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 289, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 505, + 301 + ], + "score": 1.0, + "content": "many modern policy gradient methods is the maximization of a surrogate objective function in place", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 299, + 504, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 504, + 312 + ], + "score": 1.0, + "content": "of the true rewards (the exact mechanism behind the surrogate objective is detailed in Appendix A.1,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 309, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 323 + ], + "score": 1.0, + "content": "and particularly in (14)). The surrogate objective, based on relaxing the policy improvement theorem", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 321, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 506, + 335 + ], + "score": 1.0, + "content": "of Kakade and Langford (Kakade & Langford, 2002), can be viewed as a simplification of the reward", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 333, + 205, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 205, + 344 + ], + "score": 1.0, + "content": "maximization objective.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 266, + 506, + 344 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 349, + 503, + 382 + ], + "lines": [ + { + "bbox": [ + 106, + 348, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 505, + 362 + ], + "score": 1.0, + "content": "As a purported approximation of the true returns, one would expect that the surrogate objective", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 360, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 373 + ], + "score": 1.0, + "content": "landscape approximates the true reward landscape fairly well. That is, parameters corresponding to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 370, + 374, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 374, + 384 + ], + "score": 1.0, + "content": "good surrogate objective will also correspond to good true reward.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 348, + 505, + 384 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 388, + 505, + 487 + ], + "lines": [ + { + "bbox": [ + 106, + 387, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 506, + 402 + ], + "score": 1.0, + "content": "Figure 6 shows that in the early stages of training, the optimization landscapes of the true reward", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 412 + ], + "score": 1.0, + "content": "and surrogate objective are indeed approximately aligned. However, as training progresses, the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 409, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 506, + 423 + ], + "score": 1.0, + "content": "surrogate objective becomes much less predictive of the true reward in the relevant sample regime. In", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 420, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 506, + 434 + ], + "score": 1.0, + "content": "particular, we often observe that directions that increase the surrogate objective lead to a decrease of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 432, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 506, + 444 + ], + "score": 1.0, + "content": "the true reward (see Figures 6, 7). In a higher-sample regime (using several orders of magnitude more", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "samples), we find that PPO and TRPO turn out to behave rather differently. In the case of TRPO,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 454, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 504, + 466 + ], + "score": 1.0, + "content": "the update direction following the surrogate objective matches the true reward much more closely.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "However, for PPO we consistently observe landscapes where the step direction leads to lower true", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "reward, even in the high-sample regime. This suggests that even when estimated accurately enough,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 352, + 504, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 504, + 364 + ], + "score": 1.0, + "content": "the surrogate objective might not be an accurate proxy for the true reward. (Recall from Section 2.1", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 364, + 504, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 504, + 375 + ], + "score": 1.0, + "content": "that this is a sample regime where we are able to estimate the true gradient of the reward fairly well.)", + "type": "text", + "cross_page": true + } + ], + "index": 11 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 387, + 506, + 489 + ] + }, + { + "type": "image", + "bbox": [ + 130, + 526, + 228, + 627 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 127, + 508, + 221, + 519 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 125, + 506, + 222, + 521 + ], + "spans": [ + { + "bbox": [ + 125, + 506, + 222, + 521 + ], + "score": 1.0, + "content": "2,000 state-action pairs", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "image_body", + "bbox": [ + 130, + 526, + 228, + 627 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 228, + 627 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 228, + 627 + ], + "score": 0.54, + "type": "image", + "image_path": "988f3e20351aeff96237e9e3741cc998c67968391db52269442dd22142615d7c.jpg" + } + ] + } + ], + "index": 36.5, + "virtual_lines": [ + { + "bbox": [ + 130, + 526, + 228, + 576.5 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 130, + 576.5, + 228, + 627.0 + ], + "spans": [], + "index": 38 + } + ] + } + ], + "index": 34.25 + }, + { + "type": "image", + "bbox": [ + 255, + 526, + 353, + 626 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 256, + 508, + 355, + 519 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 255, + 506, + 356, + 521 + ], + "spans": [ + { + "bbox": [ + 255, + 506, + 356, + 521 + ], + "score": 1.0, + "content": "20,000 state-action pairs", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "image_body", + "bbox": [ + 255, + 526, + 353, + 626 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 255, + 526, + 353, + 626 + ], + "spans": [ + { + "bbox": [ + 255, + 526, + 353, + 626 + ], + "score": 0.521, + "type": "image", + "image_path": "01017d8520cd26b620c161d401d7fa8673dacd4c65563407b6eb21bc583fe7d4.jpg" + } + ] + } + ], + "index": 37.5, + "virtual_lines": [ + { + "bbox": [ + 255, + 526, + 353, + 576.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 255, + 576.0, + 353, + 626.0 + ], + "spans": [], + "index": 39 + } + ] + } + ], + "index": 35.25 + }, + { + "type": "image", + "bbox": [ + 382, + 527, + 478, + 627 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 386, + 509, + 489, + 519 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 385, + 507, + 490, + 520 + ], + "spans": [ + { + "bbox": [ + 385, + 507, + 490, + 520 + ], + "score": 1.0, + "content": "100,000 state-action pairs", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "image_body", + "bbox": [ + 382, + 527, + 478, + 627 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 382, + 527, + 478, + 627 + ], + "spans": [ + { + "bbox": [ + 382, + 527, + 478, + 627 + ], + "score": 0.841, + "type": "image", + "image_path": "9133c34da248b4810278f657205cb43af0e0b9d2b1c5c14e17b97c87979a954b.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 382, + 527, + 478, + 577.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 382, + 577.0, + 478, + 627.0 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 640, + 505, + 729 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "score": 1.0, + "content": "Figure 5: True reward landscape concentration for TRPO on Humanoid-v2. We visualize the land-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 652, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 506, + 664 + ], + "score": 1.0, + "content": "scape at a training iteration 150 while varying the number of trajectories used in reward estimation", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 662, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 505, + 676 + ], + "score": 1.0, + "content": "(each subplot), both in the direction of the step taken and a random direction. Moving one unit along", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "score": 1.0, + "content": "the “step direction” axis corresponds to moving one full step in parameter space. In the random di-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 684, + 505, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 505, + 696 + ], + "score": 1.0, + "content": "rection one unit corresponds to moving along a random norm 2 Gaussian vector in the parameter", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 695, + 505, + 708 + ], + "spans": [ + { + "bbox": [ + 104, + 695, + 505, + 708 + ], + "score": 1.0, + "content": "space. In practice, the norm of the step is typically an order of magnitude lower than the random", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 706, + 505, + 719 + ], + "spans": [ + { + "bbox": [ + 105, + 706, + 505, + 719 + ], + "score": 1.0, + "content": "direction. While the landscape is very noisy in the low-sample regime, large numbers of samples", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 718, + 504, + 730 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 504, + 730 + ], + "score": 1.0, + "content": "reveal a well-behaved underlying landscape. See Figures 20, 19 of the Appendix for additional plots.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 44.5 + } + ], + "index": 38.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 127, + 79, + 471, + 240 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 127, + 79, + 471, + 240 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 79, + 471, + 240 + ], + "spans": [ + { + "bbox": [ + 127, + 79, + 471, + 240 + ], + "score": 0.976, + "type": "image", + "image_path": "1cb608251899f8943400cff475cebaf32b427d895045caac4cdea6cd45719592.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 127, + 79, + 471, + 132.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 127, + 132.66666666666666, + 471, + 186.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 127, + 186.33333333333331, + 471, + 239.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 250, + 505, + 327 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "Figure 6: True reward and surrogate objective landscapes for TRPO on the Humanoid-v2 MuJoCo", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "task. We visualize the landscapes in the direction of the update step and a random direction (as in", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 273, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 505, + 284 + ], + "score": 1.0, + "content": "Figure 5). The surrogate objective corresponds to the actual function optimized by the algorithm at", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 281, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 104, + 281, + 266, + 296 + ], + "score": 1.0, + "content": "each step. We estimate true reward with", + "type": "text" + }, + { + "bbox": [ + 266, + 282, + 282, + 293 + ], + "score": 0.87, + "content": "1 { \\dot { 0 } } ^ { 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 281, + 506, + 296 + ], + "score": 1.0, + "content": "state-action pairs per point. We compare the landscapes", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 292, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 104, + 292, + 505, + 308 + ], + "score": 1.0, + "content": "at different points in training and with varying numbers of state-action pairs used in the update step.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 304, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 506, + 317 + ], + "score": 1.0, + "content": "Early in training the true and surrogate landscapes align fairly well in both sample regimes, but later", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 315, + 475, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 475, + 329 + ], + "score": 1.0, + "content": "become misaligned in the low-sample regime. More landscapes in Appendix Figures 13-18.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 352, + 504, + 374 + ], + "lines": [ + { + "bbox": [ + 106, + 352, + 504, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 504, + 364 + ], + "score": 1.0, + "content": "the surrogate objective might not be an accurate proxy for the true reward. (Recall from Section 2.1", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 364, + 504, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 504, + 375 + ], + "score": 1.0, + "content": "that this is a sample regime where we are able to estimate the true gradient of the reward fairly well.)", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 107, + 395, + 387, + 408 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 389, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 389, + 409 + ], + "score": 1.0, + "content": "3 TOWARDS STRONGER FOUNDATIONS FOR DEEP RL", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 422, + 505, + 478 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "Deep reinforcement learning (RL) algorithms have shown great practical promise, and are rooted", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "in a well-grounded theoretical framework. However, our results indicate that this framework often", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 445, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 505, + 457 + ], + "score": 1.0, + "content": "fails to provide insight into the practical performance of these algorithms. This disconnect impedes", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 454, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 505, + 470 + ], + "score": 1.0, + "content": "our understanding of why these algorithms succeed (or fail), and is a major barrier to addressing key", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 466, + 391, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 391, + 480 + ], + "score": 1.0, + "content": "challenges facing deep RL such as brittleness and poor reproducibility.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 483, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "To close this gap, we need to either develop methods that adhere more closely to theory, or build", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "theory that can capture what makes existing policy gradient methods successful. In both cases, the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "first step is to precisely pinpoint where theory and practice diverge. To this end, we analyze and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 516, + 311, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 311, + 529 + ], + "score": 1.0, + "content": "consolidate our findings from the previous section.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 532, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 506, + 546 + ], + "score": 1.0, + "content": "Gradient estimation. Our analysis in Section 2.1 shows that the quality of gradient estimates that", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 544, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 505, + 558 + ], + "score": 1.0, + "content": "deep policy gradient algorithms use is rather poor. Indeed, even when agents improve, such gradient", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "estimates often poorly correlate with the true gradient (c.f. Figure 2). We also note that gradient cor-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "score": 1.0, + "content": "relation decreases as training progresses and task complexity increases. While this certainly does not", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "preclude the estimates from conveying useful signal, the exact underpinnings of this phenomenon", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "in deep RL still elude us. In particular, in Section 2.1 we outline a few keys ways in which the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "deep RL setting is quite unique and difficult to understand from an optimization perspective, both", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 608, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 506, + 624 + ], + "score": 1.0, + "content": "theoretically and in practice Overall, understanding the impact of gradient estimate quality on deep", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 621, + 321, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 321, + 634 + ], + "score": 1.0, + "content": "RL algorithms is challenging and largely unexplored.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 108, + 638, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 107, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 107, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "Value prediction. The findings presented in Section 2.2 identify two key issues. First, while the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "value network successfully solves the supervised learning task it is trained on, it does not accurately", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "score": 1.0, + "content": "model the “true” value function. Second, employing the value network as a baseline does decrease", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 670, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 683 + ], + "score": 1.0, + "content": "the gradient variance (compared to the trivial (“zero”) baseline). However, this decrease is rather", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 681, + 434, + 695 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 434, + 695 + ], + "score": 1.0, + "content": "marginal compared to the variance reduction offered by the “true” value function.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "It is natural to wonder whether this failure in modeling the value function is inevitable. For example,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "how does the loss function used to train the value network impact value prediction and variance", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "reduction? More broadly, we lack an understanding of the precise role of the value network in", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 127, + 79, + 471, + 240 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 127, + 79, + 471, + 240 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 79, + 471, + 240 + ], + "spans": [ + { + "bbox": [ + 127, + 79, + 471, + 240 + ], + "score": 0.976, + "type": "image", + "image_path": "1cb608251899f8943400cff475cebaf32b427d895045caac4cdea6cd45719592.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 127, + 79, + 471, + 132.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 127, + 132.66666666666666, + 471, + 186.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 127, + 186.33333333333331, + 471, + 239.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 250, + 505, + 327 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "Figure 6: True reward and surrogate objective landscapes for TRPO on the Humanoid-v2 MuJoCo", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "task. We visualize the landscapes in the direction of the update step and a random direction (as in", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 273, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 505, + 284 + ], + "score": 1.0, + "content": "Figure 5). The surrogate objective corresponds to the actual function optimized by the algorithm at", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 281, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 104, + 281, + 266, + 296 + ], + "score": 1.0, + "content": "each step. We estimate true reward with", + "type": "text" + }, + { + "bbox": [ + 266, + 282, + 282, + 293 + ], + "score": 0.87, + "content": "1 { \\dot { 0 } } ^ { 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 281, + 506, + 296 + ], + "score": 1.0, + "content": "state-action pairs per point. We compare the landscapes", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 292, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 104, + 292, + 505, + 308 + ], + "score": 1.0, + "content": "at different points in training and with varying numbers of state-action pairs used in the update step.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 304, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 506, + 317 + ], + "score": 1.0, + "content": "Early in training the true and surrogate landscapes align fairly well in both sample regimes, but later", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 315, + 475, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 475, + 329 + ], + "score": 1.0, + "content": "become misaligned in the low-sample regime. More landscapes in Appendix Figures 13-18.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 352, + 504, + 374 + ], + "lines": [], + "index": 10.5, + "bbox_fs": [ + 106, + 352, + 504, + 375 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 395, + 387, + 408 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 389, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 389, + 409 + ], + "score": 1.0, + "content": "3 TOWARDS STRONGER FOUNDATIONS FOR DEEP RL", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 422, + 505, + 478 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "Deep reinforcement learning (RL) algorithms have shown great practical promise, and are rooted", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "in a well-grounded theoretical framework. However, our results indicate that this framework often", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 445, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 505, + 457 + ], + "score": 1.0, + "content": "fails to provide insight into the practical performance of these algorithms. This disconnect impedes", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 454, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 505, + 470 + ], + "score": 1.0, + "content": "our understanding of why these algorithms succeed (or fail), and is a major barrier to addressing key", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 466, + 391, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 391, + 480 + ], + "score": 1.0, + "content": "challenges facing deep RL such as brittleness and poor reproducibility.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 423, + 505, + 480 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 483, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "To close this gap, we need to either develop methods that adhere more closely to theory, or build", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "theory that can capture what makes existing policy gradient methods successful. In both cases, the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "first step is to precisely pinpoint where theory and practice diverge. To this end, we analyze and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 516, + 311, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 311, + 529 + ], + "score": 1.0, + "content": "consolidate our findings from the previous section.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 483, + 505, + 529 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 532, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 506, + 546 + ], + "score": 1.0, + "content": "Gradient estimation. Our analysis in Section 2.1 shows that the quality of gradient estimates that", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 544, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 505, + 558 + ], + "score": 1.0, + "content": "deep policy gradient algorithms use is rather poor. Indeed, even when agents improve, such gradient", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "estimates often poorly correlate with the true gradient (c.f. Figure 2). We also note that gradient cor-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "score": 1.0, + "content": "relation decreases as training progresses and task complexity increases. While this certainly does not", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "preclude the estimates from conveying useful signal, the exact underpinnings of this phenomenon", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "in deep RL still elude us. In particular, in Section 2.1 we outline a few keys ways in which the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "deep RL setting is quite unique and difficult to understand from an optimization perspective, both", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 608, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 506, + 624 + ], + "score": 1.0, + "content": "theoretically and in practice Overall, understanding the impact of gradient estimate quality on deep", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 621, + 321, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 321, + 634 + ], + "score": 1.0, + "content": "RL algorithms is challenging and largely unexplored.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 532, + 506, + 634 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 638, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 107, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 107, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "Value prediction. The findings presented in Section 2.2 identify two key issues. First, while the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "value network successfully solves the supervised learning task it is trained on, it does not accurately", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "score": 1.0, + "content": "model the “true” value function. Second, employing the value network as a baseline does decrease", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 670, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 683 + ], + "score": 1.0, + "content": "the gradient variance (compared to the trivial (“zero”) baseline). However, this decrease is rather", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 681, + 434, + 695 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 434, + 695 + ], + "score": 1.0, + "content": "marginal compared to the variance reduction offered by the “true” value function.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 638, + 506, + 695 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "It is natural to wonder whether this failure in modeling the value function is inevitable. For example,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "how does the loss function used to train the value network impact value prediction and variance", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "reduction? More broadly, we lack an understanding of the precise role of the value network in", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 299, + 504, + 310 + ], + "spans": [ + { + "bbox": [ + 107, + 299, + 504, + 310 + ], + "score": 1.0, + "content": "training. Can we empirically quantify the relationship between variance reduction and performance?", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 309, + 414, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 414, + 321 + ], + "score": 1.0, + "content": "And does the value network play a broader role than just variance reduction?", + "type": "text", + "cross_page": true + } + ], + "index": 7 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 698, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 127, + 79, + 467, + 234 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 127, + 79, + 467, + 234 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 79, + 467, + 234 + ], + "spans": [ + { + "bbox": [ + 127, + 79, + 467, + 234 + ], + "score": 0.976, + "type": "image", + "image_path": "0f929acd579bd6910d6fbd92bfdf56ec0bf9d21185294ff124330ed9cb583b77.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 127, + 79, + 467, + 130.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 127, + 130.66666666666666, + 467, + 182.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 127, + 182.33333333333331, + 467, + 233.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 243, + 504, + 276 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 243, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 255 + ], + "score": 1.0, + "content": "Figure 7: True reward and surrogate objective landscapes for PPO on the Humanoid-v2 MuJoCo", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "task. See Figure 6 for a description. We observe that early in training the true and surrogate land-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 265, + 485, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 485, + 277 + ], + "score": 1.0, + "content": "scapes align well. However, later increasing the surrogate objective leads to lower true reward.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 503, + 320 + ], + "lines": [ + { + "bbox": [ + 107, + 299, + 504, + 310 + ], + "spans": [ + { + "bbox": [ + 107, + 299, + 504, + 310 + ], + "score": 1.0, + "content": "training. Can we empirically quantify the relationship between variance reduction and performance?", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 309, + 414, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 414, + 321 + ], + "score": 1.0, + "content": "And does the value network play a broader role than just variance reduction?", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 325, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 106, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "Optimization landscape. We have also seen, in Section 2.3, that the optimization landscape in-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 337, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 505, + 350 + ], + "score": 1.0, + "content": "duced by modern policy gradient algorithms, the surrogate objective, is often not reflective of the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 348, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 505, + 360 + ], + "score": 1.0, + "content": "underlying true reward landscape. We thus need a deeper understanding of why current methods", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 359, + 497, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 497, + 372 + ], + "score": 1.0, + "content": "succeed despite these issues, and, more broadly, how to better navigate the true reward landscape.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 108, + 386, + 211, + 399 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 213, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 213, + 401 + ], + "score": 1.0, + "content": "4 RELATED WORK", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 411, + 505, + 444 + ], + "lines": [ + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "The idea of using gradient estimates to update neural network-based RL agents dates back at least", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 422, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 506, + 435 + ], + "score": 1.0, + "content": "to the REINFORCE (Williams, 1992) algorithm. Later, Sutton (Sutton et al., 1999) established a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "unifying framework casting these algorithms as instances of the policy gradient class of algorithms.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 505, + 494 + ], + "lines": [ + { + "bbox": [ + 105, + 449, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 505, + 462 + ], + "score": 1.0, + "content": "Our work focuses on proximal policy optimization (PPO) (Schulman et al., 2017) and trust region", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 460, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 474 + ], + "score": 1.0, + "content": "policy optimization (TRPO) (Schulman et al., 2015a), which are two of the most prominent policy", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 472, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 485 + ], + "score": 1.0, + "content": "gradient algorithms used in deep RL, drawing inspiration from works on related algorithms, such", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 484, + 279, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 279, + 495 + ], + "score": 1.0, + "content": "as (Peters et al., 2010) and Kakade (2001).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 505, + 512 + ], + "score": 1.0, + "content": "Many recent works document the brittleness of deep RL algorithms (Henderson et al., 2018; 2017;", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "Islam et al., 2017). (Rajeswaran et al., 2017) and (Mania et al., 2018) demonstrate that on many", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "score": 1.0, + "content": "benchmark tasks, state-of-the-art performance can be attained by augmented randomized search", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 532, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 506, + 546 + ], + "score": 1.0, + "content": "approaches. McCandlish et al. (2018) investigates gradient noise in large-batch settings, and Ahmed", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 543, + 502, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 502, + 557 + ], + "score": 1.0, + "content": "et al. (2018) investigates the role of entropy regularization (which we do not study) on optimization.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 108, + 571, + 195, + 584 + ], + "lines": [ + { + "bbox": [ + 104, + 569, + 198, + 587 + ], + "spans": [ + { + "bbox": [ + 104, + 569, + 198, + 587 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "score": 1.0, + "content": "In this work, we analyze the degree to which key primitives of deep policy gradient algorithms", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 606, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 620 + ], + "score": 1.0, + "content": "follow their conceptual underpinnings. Our experiments show that these primitives often do not", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 618, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 631 + ], + "score": 1.0, + "content": "conform to the expected behavior: gradient estimates poorly correlate with the true gradient, better", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "gradient estimates can require lower learning rates and can induce degenerate agent behavior, value", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 640, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 506, + 653 + ], + "score": 1.0, + "content": "networks reduce gradient estimation variance to a significantly smaller extent than the true value,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 651, + 355, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 355, + 664 + ], + "score": 1.0, + "content": "and the underlying optimization landscape can be misleading.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 108, + 668, + 504, + 712 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "score": 1.0, + "content": "This demonstrates that there is a significant gap between the theory inspiring current algorithms and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 679, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 506, + 691 + ], + "score": 1.0, + "content": "the actual mechanisms driving their performance. Overall, our findings suggest that developing a", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 690, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 506, + 703 + ], + "score": 1.0, + "content": "deep RL toolkit that is truly robust and reliable will require moving beyond the current benchmark-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 701, + 449, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 701, + 449, + 713 + ], + "score": 1.0, + "content": "driven evaluation model to a more fine-grained understanding of deep RL algorithms.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 127, + 79, + 467, + 234 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 127, + 79, + 467, + 234 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 79, + 467, + 234 + ], + "spans": [ + { + "bbox": [ + 127, + 79, + 467, + 234 + ], + "score": 0.976, + "type": "image", + "image_path": "0f929acd579bd6910d6fbd92bfdf56ec0bf9d21185294ff124330ed9cb583b77.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 127, + 79, + 467, + 130.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 127, + 130.66666666666666, + 467, + 182.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 127, + 182.33333333333331, + 467, + 233.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 243, + 504, + 276 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 243, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 255 + ], + "score": 1.0, + "content": "Figure 7: True reward and surrogate objective landscapes for PPO on the Humanoid-v2 MuJoCo", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "task. See Figure 6 for a description. We observe that early in training the true and surrogate land-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 265, + 485, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 485, + 277 + ], + "score": 1.0, + "content": "scapes align well. However, later increasing the surrogate objective leads to lower true reward.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 503, + 320 + ], + "lines": [], + "index": 6.5, + "bbox_fs": [ + 106, + 299, + 504, + 321 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 325, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 106, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "Optimization landscape. We have also seen, in Section 2.3, that the optimization landscape in-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 337, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 505, + 350 + ], + "score": 1.0, + "content": "duced by modern policy gradient algorithms, the surrogate objective, is often not reflective of the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 348, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 505, + 360 + ], + "score": 1.0, + "content": "underlying true reward landscape. We thus need a deeper understanding of why current methods", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 359, + 497, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 497, + 372 + ], + "score": 1.0, + "content": "succeed despite these issues, and, more broadly, how to better navigate the true reward landscape.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 325, + 505, + 372 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 386, + 211, + 399 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 213, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 213, + 401 + ], + "score": 1.0, + "content": "4 RELATED WORK", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 411, + 505, + 444 + ], + "lines": [ + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "The idea of using gradient estimates to update neural network-based RL agents dates back at least", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 422, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 506, + 435 + ], + "score": 1.0, + "content": "to the REINFORCE (Williams, 1992) algorithm. Later, Sutton (Sutton et al., 1999) established a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "unifying framework casting these algorithms as instances of the policy gradient class of algorithms.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 411, + 506, + 447 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 505, + 494 + ], + "lines": [ + { + "bbox": [ + 105, + 449, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 505, + 462 + ], + "score": 1.0, + "content": "Our work focuses on proximal policy optimization (PPO) (Schulman et al., 2017) and trust region", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 460, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 474 + ], + "score": 1.0, + "content": "policy optimization (TRPO) (Schulman et al., 2015a), which are two of the most prominent policy", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 472, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 485 + ], + "score": 1.0, + "content": "gradient algorithms used in deep RL, drawing inspiration from works on related algorithms, such", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 484, + 279, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 279, + 495 + ], + "score": 1.0, + "content": "as (Peters et al., 2010) and Kakade (2001).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 449, + 506, + 495 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 505, + 512 + ], + "score": 1.0, + "content": "Many recent works document the brittleness of deep RL algorithms (Henderson et al., 2018; 2017;", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "Islam et al., 2017). (Rajeswaran et al., 2017) and (Mania et al., 2018) demonstrate that on many", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "score": 1.0, + "content": "benchmark tasks, state-of-the-art performance can be attained by augmented randomized search", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 532, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 506, + 546 + ], + "score": 1.0, + "content": "approaches. McCandlish et al. (2018) investigates gradient noise in large-batch settings, and Ahmed", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 543, + 502, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 502, + 557 + ], + "score": 1.0, + "content": "et al. (2018) investigates the role of entropy regularization (which we do not study) on optimization.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 500, + 506, + 557 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 571, + 195, + 584 + ], + "lines": [ + { + "bbox": [ + 104, + 569, + 198, + 587 + ], + "spans": [ + { + "bbox": [ + 104, + 569, + 198, + 587 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "score": 1.0, + "content": "In this work, we analyze the degree to which key primitives of deep policy gradient algorithms", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 606, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 620 + ], + "score": 1.0, + "content": "follow their conceptual underpinnings. Our experiments show that these primitives often do not", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 618, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 631 + ], + "score": 1.0, + "content": "conform to the expected behavior: gradient estimates poorly correlate with the true gradient, better", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "gradient estimates can require lower learning rates and can induce degenerate agent behavior, value", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 640, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 506, + 653 + ], + "score": 1.0, + "content": "networks reduce gradient estimation variance to a significantly smaller extent than the true value,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 651, + 355, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 355, + 664 + ], + "score": 1.0, + "content": "and the underlying optimization landscape can be misleading.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 596, + 506, + 664 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 668, + 504, + 712 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "score": 1.0, + "content": "This demonstrates that there is a significant gap between the theory inspiring current algorithms and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 679, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 506, + 691 + ], + "score": 1.0, + "content": "the actual mechanisms driving their performance. Overall, our findings suggest that developing a", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 690, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 506, + 703 + ], + "score": 1.0, + "content": "deep RL toolkit that is truly robust and reliable will require moving beyond the current benchmark-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 701, + 449, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 701, + 449, + 713 + ], + "score": 1.0, + "content": "driven evaluation model to a more fine-grained understanding of deep RL algorithms.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 667, + 506, + 713 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 176, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 505, + 123 + ], + "lines": [ + { + "bbox": [ + 106, + 100, + 504, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 504, + 113 + ], + "score": 1.0, + "content": "Zafarali Ahmed, Nicolas Le Roux, Mohammad Norouzi, and Dale Schuurmans. Understanding the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 310, + 123 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 310, + 123 + ], + "score": 1.0, + "content": "impact of entropy on policy optimization, 2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 106, + 130, + 504, + 153 + ], + "lines": [ + { + "bbox": [ + 106, + 129, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 505, + 144 + ], + "score": 1.0, + "content": "Alexandre d’Aspremont. Smooth optimization with approximate gradient. SIAM Journal on Opti-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 142, + 243, + 153 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 243, + 153 + ], + "score": 1.0, + "content": "mization, 19:1171–1183, 2008.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 161, + 504, + 184 + ], + "lines": [ + { + "bbox": [ + 105, + 160, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 505, + 174 + ], + "score": 1.0, + "content": "Miyuru Dayarathna, Yonggang Wen, and Rui Fan. Data center energy consumption modeling: A", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 171, + 416, + 185 + ], + "spans": [ + { + "bbox": [ + 115, + 171, + 416, + 185 + ], + "score": 1.0, + "content": "survey. IEEE Communications Surveys & Tutorials, 18(1):732–794, 2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 191, + 504, + 215 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 504, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 504, + 205 + ], + "score": 1.0, + "content": "Peter Henderson, Riashat Islam, Philip Bachman, Joelle Pineau, Doina Precup, and David Meger.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 202, + 449, + 215 + ], + "spans": [ + { + "bbox": [ + 116, + 202, + 449, + 215 + ], + "score": 1.0, + "content": "Deep reinforcement learning that matters. arXiv preprint arXiv:1709.06560, 2017.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 504, + 245 + ], + "lines": [ + { + "bbox": [ + 106, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "Peter Henderson, Joshua Romoff, and Joelle Pineau. Where did my optimum go?: An empirical", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 234, + 419, + 245 + ], + "spans": [ + { + "bbox": [ + 115, + 234, + 419, + 245 + ], + "score": 1.0, + "content": "analysis of gradient descent optimization in policy gradient methods, 2018.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 252, + 477, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 478, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 478, + 266 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jürgen Schmidhuber. Flat minima. Neural Computation, 9:1–42, 1997.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 272, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 105, + 271, + 504, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 504, + 284 + ], + "score": 1.0, + "content": "Riashat Islam, Peter Henderson, Maziar Gomrokchi, and Doina Precup. Reproducibility of bench-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 282, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 115, + 282, + 506, + 296 + ], + "score": 1.0, + "content": "marked deep reinforcement learning tasks for continuous control. In ICML Reproducibility in", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 294, + 262, + 306 + ], + "spans": [ + { + "bbox": [ + 116, + 294, + 262, + 306 + ], + "score": 1.0, + "content": "Machine Learning Workshop, 2017.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 313, + 349, + 326 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 349, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 349, + 326 + ], + "score": 1.0, + "content": "Sham M. Kakade. A natural policy gradient. In NIPS, 2001.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 333, + 505, + 356 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 347 + ], + "score": 1.0, + "content": "Sham M. Kakade and John Langford. Approximately optimal approximate reinforcement learning.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 344, + 181, + 356 + ], + "spans": [ + { + "bbox": [ + 116, + 344, + 181, + 356 + ], + "score": 1.0, + "content": "In ICML, 2002.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 363, + 415, + 375 + ], + "lines": [ + { + "bbox": [ + 105, + 362, + 414, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 414, + 377 + ], + "score": 1.0, + "content": "Kenji Kawaguchi. Deep learning without poor local minima. In NIPS, 2016.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 383, + 506, + 416 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 396 + ], + "score": 1.0, + "content": "Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Pe-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 394, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 116, + 394, + 505, + 406 + ], + "score": 1.0, + "content": "ter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. CoRR,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 404, + 209, + 417 + ], + "spans": [ + { + "bbox": [ + 116, + 404, + 209, + 417 + ], + "score": 1.0, + "content": "abs/1609.04836, 2016.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 424, + 504, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 506, + 438 + ], + "score": 1.0, + "content": "Roi Livni, Shai Shalev-Shwartz, and Ohad Shamir. On the computational efficiency of training", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 435, + 248, + 447 + ], + "spans": [ + { + "bbox": [ + 115, + 435, + 248, + 447 + ], + "score": 1.0, + "content": "neural networks. In NIPS, 2014.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 455, + 504, + 478 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "Horia Mania, Aurelia Guy, and Benjamin Recht. Simple random search provides a competitive", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 466, + 386, + 478 + ], + "spans": [ + { + "bbox": [ + 116, + 466, + 386, + 478 + ], + "score": 1.0, + "content": "approach to reinforcement learning. CoRR, abs/1803.07055, 2018.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 105, + 485, + 505, + 508 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "score": 1.0, + "content": "Sam McCandlish, Jared Kaplan, Dario Amodei, and OpenAI Dota Team. An empirical model of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 496, + 224, + 509 + ], + "spans": [ + { + "bbox": [ + 115, + 496, + 224, + 509 + ], + "score": 1.0, + "content": "large-batch training, 2018.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 515, + 442, + 528 + ], + "lines": [ + { + "bbox": [ + 105, + 514, + 443, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 443, + 530 + ], + "score": 1.0, + "content": "OpenAI. Openai five. https://blog.openai.com/openai-five/, 2018.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 505, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 506, + 548 + ], + "score": 1.0, + "content": "OpenAI, :, Marcin Andrychowicz, Bowen Baker, Maciek Chociej, Rafal Jozefowicz, Bob McGrew,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 114, + 545, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 114, + 545, + 505, + 559 + ], + "score": 1.0, + "content": "Jakub Pachocki, Arthur Petron, Matthias Plappert, Glenn Powell, Alex Ray, Jonas Schneider, Szy-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 557, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 115, + 557, + 505, + 570 + ], + "score": 1.0, + "content": "mon Sidor, Josh Tobin, Peter Welinder, Lilian Weng, and Wojciech Zaremba. Learning dexterous", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 568, + 232, + 580 + ], + "spans": [ + { + "bbox": [ + 116, + 568, + 232, + 580 + ], + "score": 1.0, + "content": "in-hand manipulation, 2018.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 587, + 502, + 600 + ], + "lines": [ + { + "bbox": [ + 104, + 586, + 502, + 601 + ], + "spans": [ + { + "bbox": [ + 104, + 586, + 502, + 601 + ], + "score": 1.0, + "content": "Jan Peters, Katharina Mülling, and Yasemin Altun. Relative entropy policy search. In AAAI, 2010.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 105, + 606, + 503, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 608, + 504, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 504, + 619 + ], + "score": 1.0, + "content": "Aravind Rajeswaran, Kendall Lowrey, Emanuel Todorov, and Sham M. Kakade. Towards general-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 618, + 358, + 630 + ], + "spans": [ + { + "bbox": [ + 115, + 618, + 358, + 630 + ], + "score": 1.0, + "content": "ization and simplicity in continuous control. In NIPS, 2017.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 637, + 503, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "Herbert Robbins and Sutton Monro. A stochastic approximation method. Ann. Math. Statist., 22(3):", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 648, + 504, + 661 + ], + "spans": [ + { + "bbox": [ + 115, + 648, + 504, + 661 + ], + "score": 1.0, + "content": "400–407, 09 1951. doi: 10.1214/aoms/1177729586. URL https://doi.org/10.1214/", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 660, + 212, + 671 + ], + "spans": [ + { + "bbox": [ + 116, + 660, + 212, + 671 + ], + "score": 1.0, + "content": "aoms/1177729586.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 678, + 503, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 504, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 504, + 691 + ], + "score": 1.0, + "content": "Itay Safran and Ohad Shamir. Spurious local minima are common in two-layer relu neural networks.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 690, + 181, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 181, + 702 + ], + "score": 1.0, + "content": "In ICML, 2018.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 721, + 500, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 500, + 732 + ], + "score": 1.0, + "content": "policy optimization. In International Conference on Machine Learning, pp. 1889–1897, 2015a.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 176, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 505, + 123 + ], + "lines": [ + { + "bbox": [ + 106, + 100, + 504, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 504, + 113 + ], + "score": 1.0, + "content": "Zafarali Ahmed, Nicolas Le Roux, Mohammad Norouzi, and Dale Schuurmans. Understanding the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 310, + 123 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 310, + 123 + ], + "score": 1.0, + "content": "impact of entropy on policy optimization, 2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 106, + 100, + 504, + 123 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 130, + 504, + 153 + ], + "lines": [ + { + "bbox": [ + 106, + 129, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 505, + 144 + ], + "score": 1.0, + "content": "Alexandre d’Aspremont. Smooth optimization with approximate gradient. SIAM Journal on Opti-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 142, + 243, + 153 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 243, + 153 + ], + "score": 1.0, + "content": "mization, 19:1171–1183, 2008.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 106, + 129, + 505, + 153 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 161, + 504, + 184 + ], + "lines": [ + { + "bbox": [ + 105, + 160, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 505, + 174 + ], + "score": 1.0, + "content": "Miyuru Dayarathna, Yonggang Wen, and Rui Fan. Data center energy consumption modeling: A", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 171, + 416, + 185 + ], + "spans": [ + { + "bbox": [ + 115, + 171, + 416, + 185 + ], + "score": 1.0, + "content": "survey. IEEE Communications Surveys & Tutorials, 18(1):732–794, 2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 160, + 505, + 185 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 191, + 504, + 215 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 504, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 504, + 205 + ], + "score": 1.0, + "content": "Peter Henderson, Riashat Islam, Philip Bachman, Joelle Pineau, Doina Precup, and David Meger.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 202, + 449, + 215 + ], + "spans": [ + { + "bbox": [ + 116, + 202, + 449, + 215 + ], + "score": 1.0, + "content": "Deep reinforcement learning that matters. arXiv preprint arXiv:1709.06560, 2017.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 191, + 504, + 215 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 504, + 245 + ], + "lines": [ + { + "bbox": [ + 106, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "Peter Henderson, Joshua Romoff, and Joelle Pineau. Where did my optimum go?: An empirical", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 234, + 419, + 245 + ], + "spans": [ + { + "bbox": [ + 115, + 234, + 419, + 245 + ], + "score": 1.0, + "content": "analysis of gradient descent optimization in policy gradient methods, 2018.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 222, + 505, + 245 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 252, + 477, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 478, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 478, + 266 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jürgen Schmidhuber. Flat minima. Neural Computation, 9:1–42, 1997.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 252, + 478, + 266 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 272, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 105, + 271, + 504, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 504, + 284 + ], + "score": 1.0, + "content": "Riashat Islam, Peter Henderson, Maziar Gomrokchi, and Doina Precup. Reproducibility of bench-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 282, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 115, + 282, + 506, + 296 + ], + "score": 1.0, + "content": "marked deep reinforcement learning tasks for continuous control. In ICML Reproducibility in", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 294, + 262, + 306 + ], + "spans": [ + { + "bbox": [ + 116, + 294, + 262, + 306 + ], + "score": 1.0, + "content": "Machine Learning Workshop, 2017.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 271, + 506, + 306 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 313, + 349, + 326 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 349, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 349, + 326 + ], + "score": 1.0, + "content": "Sham M. Kakade. A natural policy gradient. In NIPS, 2001.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 106, + 313, + 349, + 326 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 333, + 505, + 356 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 347 + ], + "score": 1.0, + "content": "Sham M. Kakade and John Langford. Approximately optimal approximate reinforcement learning.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 344, + 181, + 356 + ], + "spans": [ + { + "bbox": [ + 116, + 344, + 181, + 356 + ], + "score": 1.0, + "content": "In ICML, 2002.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 331, + 505, + 356 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 363, + 415, + 375 + ], + "lines": [ + { + "bbox": [ + 105, + 362, + 414, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 414, + 377 + ], + "score": 1.0, + "content": "Kenji Kawaguchi. Deep learning without poor local minima. In NIPS, 2016.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 362, + 414, + 377 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 383, + 506, + 416 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 396 + ], + "score": 1.0, + "content": "Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Pe-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 394, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 116, + 394, + 505, + 406 + ], + "score": 1.0, + "content": "ter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. CoRR,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 404, + 209, + 417 + ], + "spans": [ + { + "bbox": [ + 116, + 404, + 209, + 417 + ], + "score": 1.0, + "content": "abs/1609.04836, 2016.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 381, + 505, + 417 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 424, + 504, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 506, + 438 + ], + "score": 1.0, + "content": "Roi Livni, Shai Shalev-Shwartz, and Ohad Shamir. On the computational efficiency of training", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 435, + 248, + 447 + ], + "spans": [ + { + "bbox": [ + 115, + 435, + 248, + 447 + ], + "score": 1.0, + "content": "neural networks. In NIPS, 2014.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 422, + 506, + 447 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 455, + 504, + 478 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "Horia Mania, Aurelia Guy, and Benjamin Recht. Simple random search provides a competitive", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 466, + 386, + 478 + ], + "spans": [ + { + "bbox": [ + 116, + 466, + 386, + 478 + ], + "score": 1.0, + "content": "approach to reinforcement learning. CoRR, abs/1803.07055, 2018.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 106, + 455, + 505, + 478 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 485, + 505, + 508 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "score": 1.0, + "content": "Sam McCandlish, Jared Kaplan, Dario Amodei, and OpenAI Dota Team. An empirical model of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 496, + 224, + 509 + ], + "spans": [ + { + "bbox": [ + 115, + 496, + 224, + 509 + ], + "score": 1.0, + "content": "large-batch training, 2018.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 484, + 506, + 509 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 515, + 442, + 528 + ], + "lines": [ + { + "bbox": [ + 105, + 514, + 443, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 443, + 530 + ], + "score": 1.0, + "content": "OpenAI. Openai five. https://blog.openai.com/openai-five/, 2018.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 514, + 443, + 530 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 505, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 506, + 548 + ], + "score": 1.0, + "content": "OpenAI, :, Marcin Andrychowicz, Bowen Baker, Maciek Chociej, Rafal Jozefowicz, Bob McGrew,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 114, + 545, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 114, + 545, + 505, + 559 + ], + "score": 1.0, + "content": "Jakub Pachocki, Arthur Petron, Matthias Plappert, Glenn Powell, Alex Ray, Jonas Schneider, Szy-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 557, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 115, + 557, + 505, + 570 + ], + "score": 1.0, + "content": "mon Sidor, Josh Tobin, Peter Welinder, Lilian Weng, and Wojciech Zaremba. Learning dexterous", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 568, + 232, + 580 + ], + "spans": [ + { + "bbox": [ + 116, + 568, + 232, + 580 + ], + "score": 1.0, + "content": "in-hand manipulation, 2018.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 534, + 506, + 580 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 587, + 502, + 600 + ], + "lines": [ + { + "bbox": [ + 104, + 586, + 502, + 601 + ], + "spans": [ + { + "bbox": [ + 104, + 586, + 502, + 601 + ], + "score": 1.0, + "content": "Jan Peters, Katharina Mülling, and Yasemin Altun. Relative entropy policy search. In AAAI, 2010.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33, + "bbox_fs": [ + 104, + 586, + 502, + 601 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 606, + 503, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 608, + 504, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 504, + 619 + ], + "score": 1.0, + "content": "Aravind Rajeswaran, Kendall Lowrey, Emanuel Todorov, and Sham M. Kakade. Towards general-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 618, + 358, + 630 + ], + "spans": [ + { + "bbox": [ + 115, + 618, + 358, + 630 + ], + "score": 1.0, + "content": "ization and simplicity in continuous control. In NIPS, 2017.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 106, + 608, + 504, + 630 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 637, + 503, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "Herbert Robbins and Sutton Monro. A stochastic approximation method. Ann. Math. Statist., 22(3):", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 648, + 504, + 661 + ], + "spans": [ + { + "bbox": [ + 115, + 648, + 504, + 661 + ], + "score": 1.0, + "content": "400–407, 09 1951. doi: 10.1214/aoms/1177729586. URL https://doi.org/10.1214/", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 660, + 212, + 671 + ], + "spans": [ + { + "bbox": [ + 116, + 660, + 212, + 671 + ], + "score": 1.0, + "content": "aoms/1177729586.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 106, + 637, + 505, + 671 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 678, + 503, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 504, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 504, + 691 + ], + "score": 1.0, + "content": "Itay Safran and Ohad Shamir. Spurious local minima are common in two-layer relu neural networks.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 690, + 181, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 181, + 702 + ], + "score": 1.0, + "content": "In ICML, 2018.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 678, + 504, + 702 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 721, + 500, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 500, + 732 + ], + "score": 1.0, + "content": "policy optimization. In International Conference on Machine Learning, pp. 1889–1897, 2015a.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 708, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "dimensional continuous control using generalized advantage estimation. arXiv preprint", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 223, + 115 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 223, + 115 + ], + "score": 1.0, + "content": "arXiv:1506.02438, 2015b.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 503, + 156 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 135 + ], + "score": 1.0, + "content": "John Schulman, Philipp Moritz, Sergey Levine, Michael I. Jordan, and Pieter Abbeel. High-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 134, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 116, + 134, + 505, + 147 + ], + "score": 1.0, + "content": "dimensional continuous control using generalized advantage estimation. CoRR, abs/1506.02438,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 144, + 147, + 157 + ], + "spans": [ + { + "bbox": [ + 115, + 144, + 147, + 157 + ], + "score": 1.0, + "content": "2015c.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 105, + 163, + 504, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 163, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 505, + 178 + ], + "score": 1.0, + "content": "John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 175, + 380, + 187 + ], + "spans": [ + { + "bbox": [ + 116, + 175, + 380, + 187 + ], + "score": 1.0, + "content": "optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 194, + 503, + 228 + ], + "lines": [ + { + "bbox": [ + 105, + 192, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 208 + ], + "score": 1.0, + "content": "David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 203, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 115, + 203, + 505, + 219 + ], + "score": 1.0, + "content": "Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 216, + 351, + 228 + ], + "spans": [ + { + "bbox": [ + 116, + 216, + 351, + 228 + ], + "score": 1.0, + "content": "without human knowledge. Nature, 550(7676):354, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 234, + 502, + 258 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 504, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 504, + 248 + ], + "score": 1.0, + "content": "Ilya Sutskever. Keynote talk. NVIDIA NTECH, 2018. URL https://www.youtube.com/", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 246, + 278, + 259 + ], + "spans": [ + { + "bbox": [ + 115, + 246, + 158, + 259 + ], + "score": 1.0, + "content": "watch?v", + "type": "text" + }, + { + "bbox": [ + 158, + 248, + 165, + 255 + ], + "score": 0.36, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 246, + 242, + 259 + ], + "score": 1.0, + "content": "w3ues-NayAs&t", + "type": "text" + }, + { + "bbox": [ + 243, + 247, + 273, + 257 + ], + "score": 0.62, + "content": "= 4 6 7 s", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 246, + 278, + 259 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 264, + 503, + 277 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 505, + 278 + ], + "score": 1.0, + "content": "Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction. MIT press, 2018.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 503, + 307 + ], + "lines": [ + { + "bbox": [ + 106, + 282, + 504, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 504, + 296 + ], + "score": 1.0, + "content": "Richard S. Sutton, David A. McAllester, Satinder P. Singh, and Yishay Mansour. Policy gradient", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 294, + 442, + 307 + ], + "spans": [ + { + "bbox": [ + 115, + 294, + 442, + 307 + ], + "score": 1.0, + "content": "methods for reinforcement learning with function approximation. In NIPS, 1999.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 105, + 313, + 504, + 336 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 324, + 306, + 337 + ], + "spans": [ + { + "bbox": [ + 116, + 324, + 306, + 337 + ], + "score": 1.0, + "content": "learning. Machine Learning, 8:229–256, 1992.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "dimensional continuous control using generalized advantage estimation. arXiv preprint", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 223, + 115 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 223, + 115 + ], + "score": 1.0, + "content": "arXiv:1506.02438, 2015b.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 82, + 505, + 115 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 503, + 156 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 135 + ], + "score": 1.0, + "content": "John Schulman, Philipp Moritz, Sergey Levine, Michael I. Jordan, and Pieter Abbeel. High-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 134, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 116, + 134, + 505, + 147 + ], + "score": 1.0, + "content": "dimensional continuous control using generalized advantage estimation. CoRR, abs/1506.02438,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 144, + 147, + 157 + ], + "spans": [ + { + "bbox": [ + 115, + 144, + 147, + 157 + ], + "score": 1.0, + "content": "2015c.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 123, + 505, + 157 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 163, + 504, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 163, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 505, + 178 + ], + "score": 1.0, + "content": "John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 175, + 380, + 187 + ], + "spans": [ + { + "bbox": [ + 116, + 175, + 380, + 187 + ], + "score": 1.0, + "content": "optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 163, + 505, + 187 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 194, + 503, + 228 + ], + "lines": [ + { + "bbox": [ + 105, + 192, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 208 + ], + "score": 1.0, + "content": "David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 203, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 115, + 203, + 505, + 219 + ], + "score": 1.0, + "content": "Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 216, + 351, + 228 + ], + "spans": [ + { + "bbox": [ + 116, + 216, + 351, + 228 + ], + "score": 1.0, + "content": "without human knowledge. Nature, 550(7676):354, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 192, + 505, + 228 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 234, + 502, + 258 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 504, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 504, + 248 + ], + "score": 1.0, + "content": "Ilya Sutskever. Keynote talk. NVIDIA NTECH, 2018. URL https://www.youtube.com/", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 246, + 278, + 259 + ], + "spans": [ + { + "bbox": [ + 115, + 246, + 158, + 259 + ], + "score": 1.0, + "content": "watch?v", + "type": "text" + }, + { + "bbox": [ + 158, + 248, + 165, + 255 + ], + "score": 0.36, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 246, + 242, + 259 + ], + "score": 1.0, + "content": "w3ues-NayAs&t", + "type": "text" + }, + { + "bbox": [ + 243, + 247, + 273, + 257 + ], + "score": 0.62, + "content": "= 4 6 7 s", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 246, + 278, + 259 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 234, + 504, + 259 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 264, + 503, + 277 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 505, + 278 + ], + "score": 1.0, + "content": "Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction. MIT press, 2018.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 264, + 505, + 278 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 503, + 307 + ], + "lines": [ + { + "bbox": [ + 106, + 282, + 504, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 504, + 296 + ], + "score": 1.0, + "content": "Richard S. Sutton, David A. McAllester, Satinder P. Singh, and Yishay Mansour. Policy gradient", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 294, + 442, + 307 + ], + "spans": [ + { + "bbox": [ + 115, + 294, + 442, + 307 + ], + "score": 1.0, + "content": "methods for reinforcement learning with function approximation. In NIPS, 1999.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 282, + 504, + 307 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 313, + 504, + 336 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 324, + 306, + 337 + ], + "spans": [ + { + "bbox": [ + 116, + 324, + 306, + 337 + ], + "score": 1.0, + "content": "learning. Machine Learning, 8:229–256, 1992.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 313, + 505, + 337 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 182, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 108, + 195, + 119 + ], + "lines": [ + { + "bbox": [ + 106, + 107, + 196, + 120 + ], + "spans": [ + { + "bbox": [ + 106, + 107, + 196, + 120 + ], + "score": 1.0, + "content": "A.1 BACKGROUND", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 129, + 505, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 129, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 505, + 144 + ], + "score": 1.0, + "content": "In the reinforcement learning (RL) setting, an agent interacts with a stateful environment with the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 141, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 505, + 154 + ], + "score": 1.0, + "content": "goal of maximizing cumulative reward. Formally, we model the environment as a (possibly random-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 152, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 264, + 164 + ], + "score": 1.0, + "content": "ized) function mapping its current state", + "type": "text" + }, + { + "bbox": [ + 265, + 154, + 271, + 162 + ], + "score": 0.56, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 152, + 327, + 164 + ], + "score": 1.0, + "content": "and an action", + "type": "text" + }, + { + "bbox": [ + 327, + 154, + 334, + 162 + ], + "score": 0.71, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 152, + 478, + 164 + ], + "score": 1.0, + "content": "supplied by the agent to a new state", + "type": "text" + }, + { + "bbox": [ + 478, + 152, + 487, + 162 + ], + "score": 0.85, + "content": "s ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 152, + 505, + 164 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 162, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 104, + 162, + 181, + 176 + ], + "score": 1.0, + "content": "a resulting reward", + "type": "text" + }, + { + "bbox": [ + 181, + 165, + 187, + 173 + ], + "score": 0.7, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 162, + 445, + 176 + ], + "score": 1.0, + "content": ". The choice of actions of the agent is governed by the its policy", + "type": "text" + }, + { + "bbox": [ + 446, + 164, + 453, + 173 + ], + "score": 0.73, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 162, + 506, + 176 + ], + "score": 1.0, + "content": ". This policy", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 173, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 506, + 186 + ], + "score": 1.0, + "content": "is a function mapping environment states to a distribution over the actions to take. The objective of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 185, + 504, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 248, + 196 + ], + "score": 1.0, + "content": "an RL algorithm is to find a policy", + "type": "text" + }, + { + "bbox": [ + 249, + 187, + 257, + 195 + ], + "score": 0.75, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 185, + 504, + 196 + ], + "score": 1.0, + "content": "which maximizes the expected cumulative reward, where the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 195, + 481, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 481, + 208 + ], + "score": 1.0, + "content": "expectation is taken over both environment randomness and the (randomized) action choices.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 505, + 344 + ], + "lines": [ + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 310, + 235 + ], + "score": 1.0, + "content": "Preliminaries and notation. For a given policy", + "type": "text" + }, + { + "bbox": [ + 310, + 225, + 317, + 233 + ], + "score": 0.74, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 222, + 379, + 235 + ], + "score": 1.0, + "content": ", we denote by", + "type": "text" + }, + { + "bbox": [ + 379, + 222, + 406, + 235 + ], + "score": 0.93, + "content": "\\pi ( a | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "the probability that this", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 233, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 232, + 246 + ], + "score": 1.0, + "content": "policy assigns to taking action", + "type": "text" + }, + { + "bbox": [ + 232, + 236, + 239, + 243 + ], + "score": 0.62, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 233, + 390, + 246 + ], + "score": 1.0, + "content": "when the environment is in the state", + "type": "text" + }, + { + "bbox": [ + 391, + 236, + 397, + 243 + ], + "score": 0.69, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 233, + 435, + 246 + ], + "score": 1.0, + "content": ". We use", + "type": "text" + }, + { + "bbox": [ + 435, + 234, + 463, + 245 + ], + "score": 0.93, + "content": "r ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 233, + 505, + 246 + ], + "score": 1.0, + "content": "to denote", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 312, + 257 + ], + "score": 1.0, + "content": "the reward that the agent earns for playing action", + "type": "text" + }, + { + "bbox": [ + 312, + 247, + 319, + 254 + ], + "score": 0.71, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 244, + 419, + 257 + ], + "score": 1.0, + "content": "in response to the state", + "type": "text" + }, + { + "bbox": [ + 419, + 247, + 425, + 254 + ], + "score": 0.65, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 244, + 484, + 257 + ], + "score": 1.0, + "content": ". A trajectory", + "type": "text" + }, + { + "bbox": [ + 485, + 245, + 505, + 255 + ], + "score": 0.83, + "content": "\\tau =", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 107, + 255, + 214, + 267 + ], + "score": 0.91, + "content": "\\{ ( a _ { t } , s _ { t } ) : t \\in \\{ 1 \\ldots \\overline { { T } } \\} \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "is a sequence of state-action pairs that constitutes a valid transcript of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 267, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 325, + 279 + ], + "score": 1.0, + "content": "interactions of the agent with the environment. (Here,", + "type": "text" + }, + { + "bbox": [ + 326, + 268, + 336, + 277 + ], + "score": 0.84, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 267, + 363, + 279 + ], + "score": 1.0, + "content": "(resp.", + "type": "text" + }, + { + "bbox": [ + 363, + 268, + 373, + 277 + ], + "score": 0.77, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 267, + 505, + 279 + ], + "score": 1.0, + "content": ") corresponds to the action taken", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 277, + 504, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 313, + 290 + ], + "score": 1.0, + "content": "by the agent (resp. state of the environment) in the", + "type": "text" + }, + { + "bbox": [ + 313, + 279, + 318, + 287 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 277, + 483, + 290 + ], + "score": 1.0, + "content": "-th round of interaction.) We then define", + "type": "text" + }, + { + "bbox": [ + 483, + 277, + 504, + 289 + ], + "score": 0.91, + "content": "\\pi ( \\tau )", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 289, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 269, + 301 + ], + "score": 1.0, + "content": "to be the probability that the trajectory", + "type": "text" + }, + { + "bbox": [ + 269, + 290, + 276, + 298 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 289, + 438, + 301 + ], + "score": 1.0, + "content": "is executed if the agent follows policy", + "type": "text" + }, + { + "bbox": [ + 438, + 290, + 446, + 299 + ], + "score": 0.72, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 289, + 505, + 301 + ], + "score": 1.0, + "content": "(provided the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 299, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 239, + 313 + ], + "score": 1.0, + "content": "initial state of the environment is", + "type": "text" + }, + { + "bbox": [ + 239, + 301, + 249, + 310 + ], + "score": 0.83, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 299, + 299, + 313 + ], + "score": 1.0, + "content": "). Similarly,", + "type": "text" + }, + { + "bbox": [ + 299, + 299, + 380, + 312 + ], + "score": 0.92, + "content": "\\begin{array} { r } { r ( \\tau ) = \\sum _ { t } r ( \\bar { s _ { t } } , a _ { t } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 299, + 506, + 313 + ], + "score": 1.0, + "content": "denotes the cumulative reward", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 309, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 343, + 324 + ], + "score": 1.0, + "content": "earned by the agent when following this trajectory, where", + "type": "text" + }, + { + "bbox": [ + 343, + 312, + 352, + 321 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 309, + 381, + 324 + ], + "score": 1.0, + "content": "(resp.", + "type": "text" + }, + { + "bbox": [ + 381, + 312, + 392, + 321 + ], + "score": 0.79, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 309, + 441, + 324 + ], + "score": 1.0, + "content": ") denote the", + "type": "text" + }, + { + "bbox": [ + 441, + 311, + 446, + 320 + ], + "score": 0.75, + "content": "t { \\cdot }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 309, + 506, + 324 + ], + "score": 1.0, + "content": "-th state (resp.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 321, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 202, + 333 + ], + "score": 1.0, + "content": "action) in the trajectory", + "type": "text" + }, + { + "bbox": [ + 203, + 323, + 209, + 331 + ], + "score": 0.67, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 321, + 506, + 333 + ], + "score": 1.0, + "content": ". In the RL setting, however, we often choose to maximize the discounted", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 332, + 361, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 227, + 344 + ], + "score": 1.0, + "content": "cumulative reward of a policy", + "type": "text" + }, + { + "bbox": [ + 228, + 332, + 264, + 343 + ], + "score": 0.92, + "content": "R : = R _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 333, + 295, + 344 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 295, + 333, + 307, + 343 + ], + "score": 0.89, + "content": "R _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 333, + 361, + 344 + ], + "score": 1.0, + "content": "is defined as", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14 + }, + { + "type": "interline_equation", + "bbox": [ + 255, + 351, + 356, + 385 + ], + "lines": [ + { + "bbox": [ + 255, + 351, + 356, + 385 + ], + "spans": [ + { + "bbox": [ + 255, + 351, + 356, + 385 + ], + "score": 0.94, + "content": "R _ { t } ( \\tau ) = \\sum _ { t ^ { \\prime } = t } ^ { \\infty } \\gamma ^ { ( t ^ { \\prime } - t ) } r _ { t ^ { \\prime } } \\ .", + "type": "interline_equation", + "image_path": "7177816dfa109e29770a8a4ba0c8349b35b39718e020cdaaa78900f2e4b057b4.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 255, + 351, + 356, + 368.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 255, + 368.0, + 356, + 385.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 124, + 403 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 393, + 173, + 403 + ], + "score": 0.92, + "content": "0 < \\gamma < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 392, + 506, + 403 + ], + "score": 1.0, + "content": "is a “discount factor”. The discount factor ensures that the cumulative reward of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "a policy is well-defined even for an infinite time horizon, and it also incentivizes achieving reward", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 413, + 138, + 426 + ], + "spans": [ + { + "bbox": [ + 104, + 413, + 138, + 426 + ], + "score": 1.0, + "content": "earlier.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 440, + 505, + 507 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "score": 1.0, + "content": "Policy gradient methods. A widely used class of RL algorithms that will be the focus of our", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "analysis is the class of so-called policy gradient methods. The central idea behind these algorithms", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 245, + 475 + ], + "score": 1.0, + "content": "is to first parameterize the policy", + "type": "text" + }, + { + "bbox": [ + 246, + 465, + 257, + 474 + ], + "score": 0.84, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 463, + 364, + 475 + ], + "score": 1.0, + "content": "using a parameter vector", + "type": "text" + }, + { + "bbox": [ + 364, + 463, + 370, + 473 + ], + "score": 0.75, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 463, + 482, + 475 + ], + "score": 1.0, + "content": ". (In the deep RL context,", + "type": "text" + }, + { + "bbox": [ + 482, + 464, + 493, + 474 + ], + "score": 0.82, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 473, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 284, + 486 + ], + "score": 1.0, + "content": "expressed by a neural network with weights", + "type": "text" + }, + { + "bbox": [ + 285, + 474, + 291, + 483 + ], + "score": 0.74, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 473, + 506, + 486 + ], + "score": 1.0, + "content": ".) Then, we perform stochastic gradient ascent on the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 242, + 497 + ], + "score": 1.0, + "content": "cumulative reward with respect to", + "type": "text" + }, + { + "bbox": [ + 242, + 485, + 248, + 495 + ], + "score": 0.8, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 484, + 505, + 497 + ], + "score": 1.0, + "content": ". In other words, we want to apply the stochastic ascent approach", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 496, + 171, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 171, + 507 + ], + "score": 1.0, + "content": "to our problem:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5 + }, + { + "type": "interline_equation", + "bbox": [ + 266, + 515, + 344, + 533 + ], + "lines": [ + { + "bbox": [ + 266, + 515, + 344, + 533 + ], + "spans": [ + { + "bbox": [ + 266, + 515, + 344, + 533 + ], + "score": 0.91, + "content": "\\operatorname* { m a x } _ { \\theta } \\operatorname { \\mathbb { E } } _ { \\tau \\sim \\pi _ { \\theta } } [ r ( \\tau ) ] ~ ,", + "type": "interline_equation", + "image_path": "52f92bb8b77400b38dd49bb2027623ec9202623fb151fa7b124e747141c5ed2b.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 266, + 515, + 344, + 533 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 541, + 505, + 574 + ], + "lines": [ + { + "bbox": [ + 105, + 540, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 133, + 555 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 544, + 164, + 552 + ], + "score": 0.88, + "content": "\\tau \\sim \\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 540, + 505, + 555 + ], + "score": 1.0, + "content": "represents trajectories (rollouts) sampled from the distribution induced by the policy", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 552, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 118, + 563 + ], + "score": 0.81, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 552, + 505, + 565 + ], + "score": 1.0, + "content": ". This approach relies on the key observation (Sutton et al., 1999) that under mild conditions, the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 563, + 280, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 280, + 576 + ], + "score": 1.0, + "content": "gradient of our objective can be written as:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 583, + 401, + 597 + ], + "lines": [ + { + "bbox": [ + 209, + 583, + 401, + 597 + ], + "spans": [ + { + "bbox": [ + 209, + 583, + 401, + 597 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\nabla _ { \\theta } \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } [ r ( \\tau ) ] = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } [ \\nabla _ { \\theta } \\log ( \\pi _ { \\theta } ( \\tau ) ) r ( \\tau ) ] , } \\end{array}", + "type": "interline_equation", + "image_path": "cb967e99733cf851321178d8db2dc9741c07633c81aacb7351b7f5c82c2729ae.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 209, + 583, + 401, + 597 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 604, + 503, + 617 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 501, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 490, + 619 + ], + "score": 1.0, + "content": "and the latter quantity can be estimated directly by sampling trajectories according to the policy", + "type": "text" + }, + { + "bbox": [ + 490, + 607, + 501, + 616 + ], + "score": 0.76, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 621, + 504, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 620, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 505, + 635 + ], + "score": 1.0, + "content": "When we use the discounted variant of the cumulative reward and note that the action of the policy", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 632, + 488, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 136, + 646 + ], + "score": 1.0, + "content": "at time", + "type": "text" + }, + { + "bbox": [ + 136, + 634, + 141, + 642 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 632, + 488, + 646 + ], + "score": 1.0, + "content": "cannot affect its performance at earlier times, we can express our gradient estimate as:", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 651, + 419, + 692 + ], + "lines": [ + { + "bbox": [ + 191, + 651, + 419, + 692 + ], + "spans": [ + { + "bbox": [ + 191, + 651, + 419, + 692 + ], + "score": 0.94, + "content": "\\widehat { g _ { \\theta } } = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } \\left[ \\sum _ { \\left( s _ { t } , a _ { t } \\right) \\in \\tau } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\cdot Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) \\right] \\ ,", + "type": "interline_equation", + "image_path": "fa60162cee4cb315f9cd5be57fa54136309fa25e9a534683aaca7db116bd4e24.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 191, + 651, + 419, + 664.6666666666666 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 191, + 664.6666666666666, + 419, + 678.3333333333333 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 191, + 678.3333333333333, + 419, + 691.9999999999999 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 699, + 449, + 712 + ], + "lines": [ + { + "bbox": [ + 105, + 697, + 446, + 715 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 133, + 715 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 700, + 181, + 712 + ], + "score": 0.93, + "content": "Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 697, + 383, + 715 + ], + "score": 1.0, + "content": "represents the expected returns after taking action", + "type": "text" + }, + { + "bbox": [ + 383, + 704, + 392, + 711 + ], + "score": 0.82, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 697, + 436, + 715 + ], + "score": 1.0, + "content": "from state", + "type": "text" + }, + { + "bbox": [ + 437, + 702, + 446, + 711 + ], + "score": 0.83, + "content": "s _ { t }", + "type": "inline_equation" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "interline_equation", + "bbox": [ + 244, + 720, + 367, + 733 + ], + "lines": [ + { + "bbox": [ + 244, + 720, + 367, + 733 + ], + "spans": [ + { + "bbox": [ + 244, + 720, + 367, + 733 + ], + "score": 0.92, + "content": "Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ R _ { t } | a _ { t } , s _ { t } ] ~ .", + "type": "interline_equation", + "image_path": "1e400ac0480533e756b11b13673171d64954b7d6d3e744df8f772414a66f7a18.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 244, + 720, + 367, + 733 + ], + "spans": [], + "index": 43 + } + ] + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 182, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 108, + 195, + 119 + ], + "lines": [ + { + "bbox": [ + 106, + 107, + 196, + 120 + ], + "spans": [ + { + "bbox": [ + 106, + 107, + 196, + 120 + ], + "score": 1.0, + "content": "A.1 BACKGROUND", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 129, + 505, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 129, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 505, + 144 + ], + "score": 1.0, + "content": "In the reinforcement learning (RL) setting, an agent interacts with a stateful environment with the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 141, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 505, + 154 + ], + "score": 1.0, + "content": "goal of maximizing cumulative reward. Formally, we model the environment as a (possibly random-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 152, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 264, + 164 + ], + "score": 1.0, + "content": "ized) function mapping its current state", + "type": "text" + }, + { + "bbox": [ + 265, + 154, + 271, + 162 + ], + "score": 0.56, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 152, + 327, + 164 + ], + "score": 1.0, + "content": "and an action", + "type": "text" + }, + { + "bbox": [ + 327, + 154, + 334, + 162 + ], + "score": 0.71, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 152, + 478, + 164 + ], + "score": 1.0, + "content": "supplied by the agent to a new state", + "type": "text" + }, + { + "bbox": [ + 478, + 152, + 487, + 162 + ], + "score": 0.85, + "content": "s ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 152, + 505, + 164 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 162, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 104, + 162, + 181, + 176 + ], + "score": 1.0, + "content": "a resulting reward", + "type": "text" + }, + { + "bbox": [ + 181, + 165, + 187, + 173 + ], + "score": 0.7, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 162, + 445, + 176 + ], + "score": 1.0, + "content": ". The choice of actions of the agent is governed by the its policy", + "type": "text" + }, + { + "bbox": [ + 446, + 164, + 453, + 173 + ], + "score": 0.73, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 162, + 506, + 176 + ], + "score": 1.0, + "content": ". This policy", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 173, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 506, + 186 + ], + "score": 1.0, + "content": "is a function mapping environment states to a distribution over the actions to take. The objective of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 185, + 504, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 248, + 196 + ], + "score": 1.0, + "content": "an RL algorithm is to find a policy", + "type": "text" + }, + { + "bbox": [ + 249, + 187, + 257, + 195 + ], + "score": 0.75, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 185, + 504, + 196 + ], + "score": 1.0, + "content": "which maximizes the expected cumulative reward, where the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 195, + 481, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 481, + 208 + ], + "score": 1.0, + "content": "expectation is taken over both environment randomness and the (randomized) action choices.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5, + "bbox_fs": [ + 104, + 129, + 506, + 208 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 505, + 344 + ], + "lines": [ + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 310, + 235 + ], + "score": 1.0, + "content": "Preliminaries and notation. For a given policy", + "type": "text" + }, + { + "bbox": [ + 310, + 225, + 317, + 233 + ], + "score": 0.74, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 222, + 379, + 235 + ], + "score": 1.0, + "content": ", we denote by", + "type": "text" + }, + { + "bbox": [ + 379, + 222, + 406, + 235 + ], + "score": 0.93, + "content": "\\pi ( a | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "the probability that this", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 233, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 232, + 246 + ], + "score": 1.0, + "content": "policy assigns to taking action", + "type": "text" + }, + { + "bbox": [ + 232, + 236, + 239, + 243 + ], + "score": 0.62, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 233, + 390, + 246 + ], + "score": 1.0, + "content": "when the environment is in the state", + "type": "text" + }, + { + "bbox": [ + 391, + 236, + 397, + 243 + ], + "score": 0.69, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 233, + 435, + 246 + ], + "score": 1.0, + "content": ". We use", + "type": "text" + }, + { + "bbox": [ + 435, + 234, + 463, + 245 + ], + "score": 0.93, + "content": "r ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 233, + 505, + 246 + ], + "score": 1.0, + "content": "to denote", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 312, + 257 + ], + "score": 1.0, + "content": "the reward that the agent earns for playing action", + "type": "text" + }, + { + "bbox": [ + 312, + 247, + 319, + 254 + ], + "score": 0.71, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 244, + 419, + 257 + ], + "score": 1.0, + "content": "in response to the state", + "type": "text" + }, + { + "bbox": [ + 419, + 247, + 425, + 254 + ], + "score": 0.65, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 244, + 484, + 257 + ], + "score": 1.0, + "content": ". A trajectory", + "type": "text" + }, + { + "bbox": [ + 485, + 245, + 505, + 255 + ], + "score": 0.83, + "content": "\\tau =", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 107, + 255, + 214, + 267 + ], + "score": 0.91, + "content": "\\{ ( a _ { t } , s _ { t } ) : t \\in \\{ 1 \\ldots \\overline { { T } } \\} \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "is a sequence of state-action pairs that constitutes a valid transcript of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 267, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 325, + 279 + ], + "score": 1.0, + "content": "interactions of the agent with the environment. (Here,", + "type": "text" + }, + { + "bbox": [ + 326, + 268, + 336, + 277 + ], + "score": 0.84, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 267, + 363, + 279 + ], + "score": 1.0, + "content": "(resp.", + "type": "text" + }, + { + "bbox": [ + 363, + 268, + 373, + 277 + ], + "score": 0.77, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 267, + 505, + 279 + ], + "score": 1.0, + "content": ") corresponds to the action taken", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 277, + 504, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 313, + 290 + ], + "score": 1.0, + "content": "by the agent (resp. state of the environment) in the", + "type": "text" + }, + { + "bbox": [ + 313, + 279, + 318, + 287 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 277, + 483, + 290 + ], + "score": 1.0, + "content": "-th round of interaction.) We then define", + "type": "text" + }, + { + "bbox": [ + 483, + 277, + 504, + 289 + ], + "score": 0.91, + "content": "\\pi ( \\tau )", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 289, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 269, + 301 + ], + "score": 1.0, + "content": "to be the probability that the trajectory", + "type": "text" + }, + { + "bbox": [ + 269, + 290, + 276, + 298 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 289, + 438, + 301 + ], + "score": 1.0, + "content": "is executed if the agent follows policy", + "type": "text" + }, + { + "bbox": [ + 438, + 290, + 446, + 299 + ], + "score": 0.72, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 289, + 505, + 301 + ], + "score": 1.0, + "content": "(provided the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 299, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 239, + 313 + ], + "score": 1.0, + "content": "initial state of the environment is", + "type": "text" + }, + { + "bbox": [ + 239, + 301, + 249, + 310 + ], + "score": 0.83, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 299, + 299, + 313 + ], + "score": 1.0, + "content": "). Similarly,", + "type": "text" + }, + { + "bbox": [ + 299, + 299, + 380, + 312 + ], + "score": 0.92, + "content": "\\begin{array} { r } { r ( \\tau ) = \\sum _ { t } r ( \\bar { s _ { t } } , a _ { t } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 299, + 506, + 313 + ], + "score": 1.0, + "content": "denotes the cumulative reward", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 309, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 343, + 324 + ], + "score": 1.0, + "content": "earned by the agent when following this trajectory, where", + "type": "text" + }, + { + "bbox": [ + 343, + 312, + 352, + 321 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 309, + 381, + 324 + ], + "score": 1.0, + "content": "(resp.", + "type": "text" + }, + { + "bbox": [ + 381, + 312, + 392, + 321 + ], + "score": 0.79, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 309, + 441, + 324 + ], + "score": 1.0, + "content": ") denote the", + "type": "text" + }, + { + "bbox": [ + 441, + 311, + 446, + 320 + ], + "score": 0.75, + "content": "t { \\cdot }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 309, + 506, + 324 + ], + "score": 1.0, + "content": "-th state (resp.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 321, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 202, + 333 + ], + "score": 1.0, + "content": "action) in the trajectory", + "type": "text" + }, + { + "bbox": [ + 203, + 323, + 209, + 331 + ], + "score": 0.67, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 321, + 506, + 333 + ], + "score": 1.0, + "content": ". In the RL setting, however, we often choose to maximize the discounted", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 332, + 361, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 227, + 344 + ], + "score": 1.0, + "content": "cumulative reward of a policy", + "type": "text" + }, + { + "bbox": [ + 228, + 332, + 264, + 343 + ], + "score": 0.92, + "content": "R : = R _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 333, + 295, + 344 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 295, + 333, + 307, + 343 + ], + "score": 0.89, + "content": "R _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 333, + 361, + 344 + ], + "score": 1.0, + "content": "is defined as", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 222, + 506, + 344 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 255, + 351, + 356, + 385 + ], + "lines": [ + { + "bbox": [ + 255, + 351, + 356, + 385 + ], + "spans": [ + { + "bbox": [ + 255, + 351, + 356, + 385 + ], + "score": 0.94, + "content": "R _ { t } ( \\tau ) = \\sum _ { t ^ { \\prime } = t } ^ { \\infty } \\gamma ^ { ( t ^ { \\prime } - t ) } r _ { t ^ { \\prime } } \\ .", + "type": "interline_equation", + "image_path": "7177816dfa109e29770a8a4ba0c8349b35b39718e020cdaaa78900f2e4b057b4.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 255, + 351, + 356, + 368.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 255, + 368.0, + 356, + 385.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 124, + 403 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 393, + 173, + 403 + ], + "score": 0.92, + "content": "0 < \\gamma < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 392, + 506, + 403 + ], + "score": 1.0, + "content": "is a “discount factor”. The discount factor ensures that the cumulative reward of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "a policy is well-defined even for an infinite time horizon, and it also incentivizes achieving reward", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 413, + 138, + 426 + ], + "spans": [ + { + "bbox": [ + 104, + 413, + 138, + 426 + ], + "score": 1.0, + "content": "earlier.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 104, + 392, + 506, + 426 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 440, + 505, + 507 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "score": 1.0, + "content": "Policy gradient methods. A widely used class of RL algorithms that will be the focus of our", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "analysis is the class of so-called policy gradient methods. The central idea behind these algorithms", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 245, + 475 + ], + "score": 1.0, + "content": "is to first parameterize the policy", + "type": "text" + }, + { + "bbox": [ + 246, + 465, + 257, + 474 + ], + "score": 0.84, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 463, + 364, + 475 + ], + "score": 1.0, + "content": "using a parameter vector", + "type": "text" + }, + { + "bbox": [ + 364, + 463, + 370, + 473 + ], + "score": 0.75, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 463, + 482, + 475 + ], + "score": 1.0, + "content": ". (In the deep RL context,", + "type": "text" + }, + { + "bbox": [ + 482, + 464, + 493, + 474 + ], + "score": 0.82, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 473, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 284, + 486 + ], + "score": 1.0, + "content": "expressed by a neural network with weights", + "type": "text" + }, + { + "bbox": [ + 285, + 474, + 291, + 483 + ], + "score": 0.74, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 473, + 506, + 486 + ], + "score": 1.0, + "content": ".) Then, we perform stochastic gradient ascent on the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 242, + 497 + ], + "score": 1.0, + "content": "cumulative reward with respect to", + "type": "text" + }, + { + "bbox": [ + 242, + 485, + 248, + 495 + ], + "score": 0.8, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 484, + 505, + 497 + ], + "score": 1.0, + "content": ". In other words, we want to apply the stochastic ascent approach", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 496, + 171, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 171, + 507 + ], + "score": 1.0, + "content": "to our problem:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 440, + 506, + 507 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 266, + 515, + 344, + 533 + ], + "lines": [ + { + "bbox": [ + 266, + 515, + 344, + 533 + ], + "spans": [ + { + "bbox": [ + 266, + 515, + 344, + 533 + ], + "score": 0.91, + "content": "\\operatorname* { m a x } _ { \\theta } \\operatorname { \\mathbb { E } } _ { \\tau \\sim \\pi _ { \\theta } } [ r ( \\tau ) ] ~ ,", + "type": "interline_equation", + "image_path": "52f92bb8b77400b38dd49bb2027623ec9202623fb151fa7b124e747141c5ed2b.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 266, + 515, + 344, + 533 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 541, + 505, + 574 + ], + "lines": [ + { + "bbox": [ + 105, + 540, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 133, + 555 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 544, + 164, + 552 + ], + "score": 0.88, + "content": "\\tau \\sim \\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 540, + 505, + 555 + ], + "score": 1.0, + "content": "represents trajectories (rollouts) sampled from the distribution induced by the policy", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 552, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 118, + 563 + ], + "score": 0.81, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 552, + 505, + 565 + ], + "score": 1.0, + "content": ". This approach relies on the key observation (Sutton et al., 1999) that under mild conditions, the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 563, + 280, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 280, + 576 + ], + "score": 1.0, + "content": "gradient of our objective can be written as:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 540, + 505, + 576 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 583, + 401, + 597 + ], + "lines": [ + { + "bbox": [ + 209, + 583, + 401, + 597 + ], + "spans": [ + { + "bbox": [ + 209, + 583, + 401, + 597 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\nabla _ { \\theta } \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } [ r ( \\tau ) ] = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } [ \\nabla _ { \\theta } \\log ( \\pi _ { \\theta } ( \\tau ) ) r ( \\tau ) ] , } \\end{array}", + "type": "interline_equation", + "image_path": "cb967e99733cf851321178d8db2dc9741c07633c81aacb7351b7f5c82c2729ae.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 209, + 583, + 401, + 597 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 604, + 503, + 617 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 501, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 490, + 619 + ], + "score": 1.0, + "content": "and the latter quantity can be estimated directly by sampling trajectories according to the policy", + "type": "text" + }, + { + "bbox": [ + 490, + 607, + 501, + 616 + ], + "score": 0.76, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + } + ], + "index": 36 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 603, + 501, + 619 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 621, + 504, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 620, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 505, + 635 + ], + "score": 1.0, + "content": "When we use the discounted variant of the cumulative reward and note that the action of the policy", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 632, + 488, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 136, + 646 + ], + "score": 1.0, + "content": "at time", + "type": "text" + }, + { + "bbox": [ + 136, + 634, + 141, + 642 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 632, + 488, + 646 + ], + "score": 1.0, + "content": "cannot affect its performance at earlier times, we can express our gradient estimate as:", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 620, + 505, + 646 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 651, + 419, + 692 + ], + "lines": [ + { + "bbox": [ + 191, + 651, + 419, + 692 + ], + "spans": [ + { + "bbox": [ + 191, + 651, + 419, + 692 + ], + "score": 0.94, + "content": "\\widehat { g _ { \\theta } } = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } \\left[ \\sum _ { \\left( s _ { t } , a _ { t } \\right) \\in \\tau } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\cdot Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) \\right] \\ ,", + "type": "interline_equation", + "image_path": "fa60162cee4cb315f9cd5be57fa54136309fa25e9a534683aaca7db116bd4e24.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 191, + 651, + 419, + 664.6666666666666 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 191, + 664.6666666666666, + 419, + 678.3333333333333 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 191, + 678.3333333333333, + 419, + 691.9999999999999 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 699, + 449, + 712 + ], + "lines": [ + { + "bbox": [ + 105, + 697, + 446, + 715 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 133, + 715 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 700, + 181, + 712 + ], + "score": 0.93, + "content": "Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 697, + 383, + 715 + ], + "score": 1.0, + "content": "represents the expected returns after taking action", + "type": "text" + }, + { + "bbox": [ + 383, + 704, + 392, + 711 + ], + "score": 0.82, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 697, + 436, + 715 + ], + "score": 1.0, + "content": "from state", + "type": "text" + }, + { + "bbox": [ + 437, + 702, + 446, + 711 + ], + "score": 0.83, + "content": "s _ { t }", + "type": "inline_equation" + } + ], + "index": 42 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 697, + 446, + 715 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 244, + 720, + 367, + 733 + ], + "lines": [ + { + "bbox": [ + 244, + 720, + 367, + 733 + ], + "spans": [ + { + "bbox": [ + 244, + 720, + 367, + 733 + ], + "score": 0.92, + "content": "Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ R _ { t } | a _ { t } , s _ { t } ] ~ .", + "type": "interline_equation", + "image_path": "1e400ac0480533e756b11b13673171d64954b7d6d3e744df8f772414a66f7a18.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 244, + 720, + 367, + 733 + ], + "spans": [], + "index": 43 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 81, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Value estimation and advantage. Unfortunately, the variance of the expectation in (7) can be (and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 505, + 107 + ], + "score": 1.0, + "content": "often is) very large, which makes getting an accurate estimate of this expectation quite challenging.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "To alleviate this issue, a number of variance reduction techniques have been developed. One of the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "most popular such techniques is the use of a so-called baseline function, wherein a state-dependent", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 419, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 205, + 140 + ], + "score": 1.0, + "content": "value is subtracted from", + "type": "text" + }, + { + "bbox": [ + 205, + 127, + 222, + 138 + ], + "score": 0.9, + "content": "Q _ { \\pi _ { \\theta } }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 126, + 419, + 140 + ], + "score": 1.0, + "content": ". Thus, instead of estimating (7) directly, we use:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "interline_equation", + "bbox": [ + 172, + 140, + 438, + 180 + ], + "lines": [ + { + "bbox": [ + 172, + 140, + 438, + 180 + ], + "spans": [ + { + "bbox": [ + 172, + 140, + 438, + 180 + ], + "score": 0.94, + "content": "\\widehat { g _ { \\theta } } = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } \\left[ \\sum _ { ( s _ { t } , a _ { t } ) \\in \\tau } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\cdot ( Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) - b ( s _ { t } ) ) \\right] ,", + "type": "interline_equation", + "image_path": "714c95f4101a60583a3bb3e9d4f137696057e71dc390133e51fed1fe55633dba.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 172, + 140, + 438, + 153.33333333333334 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 172, + 153.33333333333334, + 438, + 166.66666666666669 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 172, + 166.66666666666669, + 438, + 180.00000000000003 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 182, + 293, + 194 + ], + "lines": [ + { + "bbox": [ + 106, + 181, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 133, + 194 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 182, + 149, + 194 + ], + "score": 0.9, + "content": "b ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 181, + 295, + 194 + ], + "score": 1.0, + "content": "is a baseline function of our choice.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 371, + 210 + ], + "lines": [ + { + "bbox": [ + 105, + 197, + 372, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 372, + 212 + ], + "score": 1.0, + "content": "A natural choice of the baseline function is the value function, i.e.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 257, + 212, + 353, + 226 + ], + "lines": [ + { + "bbox": [ + 257, + 212, + 353, + 226 + ], + "spans": [ + { + "bbox": [ + 257, + 212, + 353, + 226 + ], + "score": 0.93, + "content": "V _ { \\pi _ { \\theta } } ( s _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ R _ { t } | s _ { t } ] .", + "type": "interline_equation", + "image_path": "d536625ed85c40e5b51a28cd421cbbbaedf06c7b52fbd587f8037c122929a8c0.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 257, + 212, + 353, + 226 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 228, + 502, + 240 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 242 + ], + "score": 1.0, + "content": "When we use the value function as our baseline, the resulting gradient estimation problem becomes:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "interline_equation", + "bbox": [ + 192, + 241, + 417, + 281 + ], + "lines": [ + { + "bbox": [ + 192, + 241, + 417, + 281 + ], + "spans": [ + { + "bbox": [ + 192, + 241, + 417, + 281 + ], + "score": 0.93, + "content": "\\widehat { g } _ { \\theta } = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } \\left[ \\sum _ { \\left( s _ { t } , a _ { t } \\right) \\in \\tau } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\cdot A _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) \\right] ,", + "type": "interline_equation", + "image_path": "22c9b56669575dfaf8f6f350ffa5bf3fd46bb9d7d2bbb38748407885d7357c67.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 192, + 241, + 417, + 254.33333333333334 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 192, + 254.33333333333334, + 417, + 267.6666666666667 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 192, + 267.6666666666667, + 417, + 281.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 282, + 133, + 293 + ], + "lines": [ + { + "bbox": [ + 105, + 281, + 135, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 135, + 295 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 296, + 381, + 310 + ], + "lines": [ + { + "bbox": [ + 230, + 296, + 381, + 310 + ], + "spans": [ + { + "bbox": [ + 230, + 296, + 381, + 310 + ], + "score": 0.92, + "content": "A _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) = Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) - V _ { \\pi _ { \\theta } } ( s _ { t } )", + "type": "interline_equation", + "image_path": "97e07d10f77d32fc619464effb2b00b133ed31357ab03cac5ba7f0861f0a87d4.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 230, + 296, + 381, + 310 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 311, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 310, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 318, + 325 + ], + "score": 1.0, + "content": "is referred to as the advantage of performing action", + "type": "text" + }, + { + "bbox": [ + 318, + 314, + 328, + 323 + ], + "score": 0.86, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 310, + 466, + 325 + ], + "score": 1.0, + "content": ". Different methods of estimating", + "type": "text" + }, + { + "bbox": [ + 466, + 312, + 482, + 324 + ], + "score": 0.9, + "content": "V _ { \\pi _ { \\theta } }", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 310, + 505, + 325 + ], + "score": 1.0, + "content": "have", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "been proposed, with techniques ranging from moving averages to the use of neural network predic-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 333, + 225, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 225, + 345 + ], + "score": 1.0, + "content": "tors Schulman et al. (2015b).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 356, + 505, + 412 + ], + "lines": [ + { + "bbox": [ + 106, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "Surrogate Objective. So far, our focus has been on extracting a good estimate of the gradient", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 367, + 504, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 258, + 379 + ], + "score": 1.0, + "content": "with respect to the policy parameters", + "type": "text" + }, + { + "bbox": [ + 258, + 368, + 264, + 378 + ], + "score": 0.77, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 367, + 504, + 379 + ], + "score": 1.0, + "content": ". However, it turns out that directly optimizing the cumula-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 379, + 504, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 504, + 391 + ], + "score": 1.0, + "content": "tive rewards can be challenging. Thus, a modification used by modern policy gradient algorithms", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 390, + 504, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 504, + 402 + ], + "score": 1.0, + "content": "is to optimize a “surrogate objective” instead. We will focus on maximizing the following local", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 401, + 341, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 341, + 414 + ], + "score": 1.0, + "content": "approximation of the true reward Schulman et al. (2015a):", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + }, + { + "type": "interline_equation", + "bbox": [ + 182, + 414, + 429, + 441 + ], + "lines": [ + { + "bbox": [ + 182, + 414, + 429, + 441 + ], + "spans": [ + { + "bbox": [ + 182, + 414, + 429, + 441 + ], + "score": 0.93, + "content": "\\operatorname* { m a x } _ { \\theta } \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi } \\left[ \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi ( a _ { t } | s _ { t } ) } A _ { \\pi } ( s _ { t } , a _ { t } ) \\right] \\qquad \\Big ( = \\mathbb { E } _ { \\pi _ { \\theta } } \\left[ A _ { \\pi } \\right] \\Big ) ,", + "type": "interline_equation", + "image_path": "c33cc875141510bc73c214796c859a14f7047b6273794d97c0874caa7cec7038.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 182, + 414, + 429, + 423.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 182, + 423.0, + 429, + 432.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 182, + 432.0, + 429, + 441.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 443, + 460, + 454 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 462, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 462, + 456 + ], + "score": 1.0, + "content": "or the normalized advantage variant proposed to reduce variance Schulman et al. (2017):", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "interline_equation", + "bbox": [ + 224, + 456, + 386, + 484 + ], + "lines": [ + { + "bbox": [ + 224, + 456, + 386, + 484 + ], + "spans": [ + { + "bbox": [ + 224, + 456, + 386, + 484 + ], + "score": 0.94, + "content": "\\operatorname* { m a x } _ { \\theta } \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi } \\left[ \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi ( a _ { t } | s _ { t } ) } \\widehat { A } _ { \\pi } ( s _ { t } , a _ { t } ) \\right]", + "type": "interline_equation", + "image_path": "6c93024d7e9186175c5a0326f39c803d8c4ea4e6ba71fd228107b27c706b95a6.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 224, + 456, + 386, + 470.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 224, + 470.0, + 386, + 484.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 133, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 135, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 135, + 497 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "interline_equation", + "bbox": [ + 264, + 496, + 347, + 523 + ], + "lines": [ + { + "bbox": [ + 264, + 496, + 347, + 523 + ], + "spans": [ + { + "bbox": [ + 264, + 496, + 347, + 523 + ], + "score": 0.96, + "content": "\\widehat { A } _ { \\pi } = \\frac { A _ { \\pi } - \\mu ( A _ { \\pi } ) } { \\sigma ( A _ { \\pi } ) }", + "type": "interline_equation", + "image_path": "4398683e8c15fdbb5407bd0c1c00ef8578d14f18944f7f1ed863715cf8cb7974.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 264, + 496, + 347, + 509.5 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 264, + 509.5, + 347, + 523.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 214, + 537 + ], + "lines": [ + { + "bbox": [ + 105, + 523, + 216, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 123, + 538 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 527, + 131, + 535 + ], + "score": 0.74, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 523, + 216, + 538 + ], + "score": 1.0, + "content": "is the current policy.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 547, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 106, + 547, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 506, + 561 + ], + "score": 1.0, + "content": "Trust region methods. The surrogate objective function, although easier to optimize, comes at a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "cost: the gradient of the surrogate objective is only predictive of the policy gradient locally (at the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 570, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 506, + 583 + ], + "score": 1.0, + "content": "current policy). Thus, to ensure that our update steps we derive based on the surrogate objective are", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 581, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 506, + 594 + ], + "score": 1.0, + "content": "predictive, they need to be confined to a “trust region” around the current policy. The resulting trust", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 591, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 104, + 591, + 506, + 604 + ], + "score": 1.0, + "content": "region methods (Kakade, 2001; Schulman et al., 2015a; 2017) try to constrain the local variation of", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 603, + 504, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 504, + 614 + ], + "score": 1.0, + "content": "the parameters in policy-space by restricting the distributional distance between successive policies.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 107, + 619, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 632 + ], + "score": 1.0, + "content": "A popular method in this class is trust region policy optimization (TRPO) Schulman et al. (2015a),", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "which constrains the KL divergence between successive policies on the optimization trajectory, lead-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 642, + 226, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 226, + 654 + ], + "score": 1.0, + "content": "ing to the following problem:", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 655, + 394, + 698 + ], + "lines": [ + { + "bbox": [ + 217, + 655, + 394, + 698 + ], + "spans": [ + { + "bbox": [ + 217, + 655, + 394, + 698 + ], + "score": 0.9, + "content": "\\begin{array} { r l } { \\displaystyle \\operatorname* { m a x } _ { \\theta } } & { \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi } \\left[ \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi ( a _ { t } | s _ { t } ) } \\widehat { A } _ { \\pi } ( s _ { t } , a _ { t } ) \\right] } \\\\ { \\mathrm { s . t . } } & { D _ { K L } ( \\pi _ { \\theta } ( \\cdot \\mid s ) | | \\pi ( \\cdot \\mid s ) ) \\leq \\delta , \\quad \\forall s ~ . } \\end{array}", + "type": "interline_equation", + "image_path": "1c76c841578edf86898b179efb59aca8cc71ebb10567e69380e287ef7135954b.jpg" + } + ] + } + ], + "index": 45, + "virtual_lines": [ + { + "bbox": [ + 217, + 655, + 394, + 669.3333333333334 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 217, + 669.3333333333334, + 394, + 683.6666666666667 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 217, + 683.6666666666667, + 394, + 698.0000000000001 + ], + "spans": [], + "index": 46 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "In practice, this objective is maximized using a second-order approximation of the KL divergence", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "and natural gradient descent, while replacing the worst-case KL constraints over all possible states", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 719, + 478, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 478, + 734 + ], + "score": 1.0, + "content": "with an approximation of the mean KL based on the states observed in the current trajectory.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 81, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Value estimation and advantage. Unfortunately, the variance of the expectation in (7) can be (and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 505, + 107 + ], + "score": 1.0, + "content": "often is) very large, which makes getting an accurate estimate of this expectation quite challenging.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "To alleviate this issue, a number of variance reduction techniques have been developed. One of the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "most popular such techniques is the use of a so-called baseline function, wherein a state-dependent", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 419, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 205, + 140 + ], + "score": 1.0, + "content": "value is subtracted from", + "type": "text" + }, + { + "bbox": [ + 205, + 127, + 222, + 138 + ], + "score": 0.9, + "content": "Q _ { \\pi _ { \\theta } }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 126, + 419, + 140 + ], + "score": 1.0, + "content": ". Thus, instead of estimating (7) directly, we use:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 82, + 506, + 140 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 172, + 140, + 438, + 180 + ], + "lines": [ + { + "bbox": [ + 172, + 140, + 438, + 180 + ], + "spans": [ + { + "bbox": [ + 172, + 140, + 438, + 180 + ], + "score": 0.94, + "content": "\\widehat { g _ { \\theta } } = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } \\left[ \\sum _ { ( s _ { t } , a _ { t } ) \\in \\tau } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\cdot ( Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) - b ( s _ { t } ) ) \\right] ,", + "type": "interline_equation", + "image_path": "714c95f4101a60583a3bb3e9d4f137696057e71dc390133e51fed1fe55633dba.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 172, + 140, + 438, + 153.33333333333334 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 172, + 153.33333333333334, + 438, + 166.66666666666669 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 172, + 166.66666666666669, + 438, + 180.00000000000003 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 182, + 293, + 194 + ], + "lines": [ + { + "bbox": [ + 106, + 181, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 133, + 194 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 182, + 149, + 194 + ], + "score": 0.9, + "content": "b ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 181, + 295, + 194 + ], + "score": 1.0, + "content": "is a baseline function of our choice.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8, + "bbox_fs": [ + 106, + 181, + 295, + 194 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 371, + 210 + ], + "lines": [ + { + "bbox": [ + 105, + 197, + 372, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 372, + 212 + ], + "score": 1.0, + "content": "A natural choice of the baseline function is the value function, i.e.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 197, + 372, + 212 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 257, + 212, + 353, + 226 + ], + "lines": [ + { + "bbox": [ + 257, + 212, + 353, + 226 + ], + "spans": [ + { + "bbox": [ + 257, + 212, + 353, + 226 + ], + "score": 0.93, + "content": "V _ { \\pi _ { \\theta } } ( s _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ R _ { t } | s _ { t } ] .", + "type": "interline_equation", + "image_path": "d536625ed85c40e5b51a28cd421cbbbaedf06c7b52fbd587f8037c122929a8c0.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 257, + 212, + 353, + 226 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 228, + 502, + 240 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 242 + ], + "score": 1.0, + "content": "When we use the value function as our baseline, the resulting gradient estimation problem becomes:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 225, + 505, + 242 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 192, + 241, + 417, + 281 + ], + "lines": [ + { + "bbox": [ + 192, + 241, + 417, + 281 + ], + "spans": [ + { + "bbox": [ + 192, + 241, + 417, + 281 + ], + "score": 0.93, + "content": "\\widehat { g } _ { \\theta } = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } \\left[ \\sum _ { \\left( s _ { t } , a _ { t } \\right) \\in \\tau } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\cdot A _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) \\right] ,", + "type": "interline_equation", + "image_path": "22c9b56669575dfaf8f6f350ffa5bf3fd46bb9d7d2bbb38748407885d7357c67.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 192, + 241, + 417, + 254.33333333333334 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 192, + 254.33333333333334, + 417, + 267.6666666666667 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 192, + 267.6666666666667, + 417, + 281.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 282, + 133, + 293 + ], + "lines": [ + { + "bbox": [ + 105, + 281, + 135, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 135, + 295 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 281, + 135, + 295 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 296, + 381, + 310 + ], + "lines": [ + { + "bbox": [ + 230, + 296, + 381, + 310 + ], + "spans": [ + { + "bbox": [ + 230, + 296, + 381, + 310 + ], + "score": 0.92, + "content": "A _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) = Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) - V _ { \\pi _ { \\theta } } ( s _ { t } )", + "type": "interline_equation", + "image_path": "97e07d10f77d32fc619464effb2b00b133ed31357ab03cac5ba7f0861f0a87d4.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 230, + 296, + 381, + 310 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 311, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 310, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 318, + 325 + ], + "score": 1.0, + "content": "is referred to as the advantage of performing action", + "type": "text" + }, + { + "bbox": [ + 318, + 314, + 328, + 323 + ], + "score": 0.86, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 310, + 466, + 325 + ], + "score": 1.0, + "content": ". Different methods of estimating", + "type": "text" + }, + { + "bbox": [ + 466, + 312, + 482, + 324 + ], + "score": 0.9, + "content": "V _ { \\pi _ { \\theta } }", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 310, + 505, + 325 + ], + "score": 1.0, + "content": "have", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "been proposed, with techniques ranging from moving averages to the use of neural network predic-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 333, + 225, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 225, + 345 + ], + "score": 1.0, + "content": "tors Schulman et al. (2015b).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 310, + 505, + 345 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 356, + 505, + 412 + ], + "lines": [ + { + "bbox": [ + 106, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "Surrogate Objective. So far, our focus has been on extracting a good estimate of the gradient", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 367, + 504, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 258, + 379 + ], + "score": 1.0, + "content": "with respect to the policy parameters", + "type": "text" + }, + { + "bbox": [ + 258, + 368, + 264, + 378 + ], + "score": 0.77, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 367, + 504, + 379 + ], + "score": 1.0, + "content": ". However, it turns out that directly optimizing the cumula-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 379, + 504, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 504, + 391 + ], + "score": 1.0, + "content": "tive rewards can be challenging. Thus, a modification used by modern policy gradient algorithms", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 390, + 504, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 504, + 402 + ], + "score": 1.0, + "content": "is to optimize a “surrogate objective” instead. We will focus on maximizing the following local", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 401, + 341, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 341, + 414 + ], + "score": 1.0, + "content": "approximation of the true reward Schulman et al. (2015a):", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 356, + 505, + 414 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 182, + 414, + 429, + 441 + ], + "lines": [ + { + "bbox": [ + 182, + 414, + 429, + 441 + ], + "spans": [ + { + "bbox": [ + 182, + 414, + 429, + 441 + ], + "score": 0.93, + "content": "\\operatorname* { m a x } _ { \\theta } \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi } \\left[ \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi ( a _ { t } | s _ { t } ) } A _ { \\pi } ( s _ { t } , a _ { t } ) \\right] \\qquad \\Big ( = \\mathbb { E } _ { \\pi _ { \\theta } } \\left[ A _ { \\pi } \\right] \\Big ) ,", + "type": "interline_equation", + "image_path": "c33cc875141510bc73c214796c859a14f7047b6273794d97c0874caa7cec7038.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 182, + 414, + 429, + 423.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 182, + 423.0, + 429, + 432.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 182, + 432.0, + 429, + 441.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 443, + 460, + 454 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 462, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 462, + 456 + ], + "score": 1.0, + "content": "or the normalized advantage variant proposed to reduce variance Schulman et al. (2017):", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 442, + 462, + 456 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 224, + 456, + 386, + 484 + ], + "lines": [ + { + "bbox": [ + 224, + 456, + 386, + 484 + ], + "spans": [ + { + "bbox": [ + 224, + 456, + 386, + 484 + ], + "score": 0.94, + "content": "\\operatorname* { m a x } _ { \\theta } \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi } \\left[ \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi ( a _ { t } | s _ { t } ) } \\widehat { A } _ { \\pi } ( s _ { t } , a _ { t } ) \\right]", + "type": "interline_equation", + "image_path": "6c93024d7e9186175c5a0326f39c803d8c4ea4e6ba71fd228107b27c706b95a6.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 224, + 456, + 386, + 470.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 224, + 470.0, + 386, + 484.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 133, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 135, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 135, + 497 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 484, + 135, + 497 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 264, + 496, + 347, + 523 + ], + "lines": [ + { + "bbox": [ + 264, + 496, + 347, + 523 + ], + "spans": [ + { + "bbox": [ + 264, + 496, + 347, + 523 + ], + "score": 0.96, + "content": "\\widehat { A } _ { \\pi } = \\frac { A _ { \\pi } - \\mu ( A _ { \\pi } ) } { \\sigma ( A _ { \\pi } ) }", + "type": "interline_equation", + "image_path": "4398683e8c15fdbb5407bd0c1c00ef8578d14f18944f7f1ed863715cf8cb7974.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 264, + 496, + 347, + 509.5 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 264, + 509.5, + 347, + 523.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 214, + 537 + ], + "lines": [ + { + "bbox": [ + 105, + 523, + 216, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 123, + 538 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 527, + 131, + 535 + ], + "score": 0.74, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 523, + 216, + 538 + ], + "score": 1.0, + "content": "is the current policy.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 523, + 216, + 538 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 547, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 106, + 547, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 506, + 561 + ], + "score": 1.0, + "content": "Trust region methods. The surrogate objective function, although easier to optimize, comes at a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "cost: the gradient of the surrogate objective is only predictive of the policy gradient locally (at the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 570, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 506, + 583 + ], + "score": 1.0, + "content": "current policy). Thus, to ensure that our update steps we derive based on the surrogate objective are", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 581, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 506, + 594 + ], + "score": 1.0, + "content": "predictive, they need to be confined to a “trust region” around the current policy. The resulting trust", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 591, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 104, + 591, + 506, + 604 + ], + "score": 1.0, + "content": "region methods (Kakade, 2001; Schulman et al., 2015a; 2017) try to constrain the local variation of", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 603, + 504, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 504, + 614 + ], + "score": 1.0, + "content": "the parameters in policy-space by restricting the distributional distance between successive policies.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5, + "bbox_fs": [ + 104, + 547, + 506, + 614 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 619, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 632 + ], + "score": 1.0, + "content": "A popular method in this class is trust region policy optimization (TRPO) Schulman et al. (2015a),", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "which constrains the KL divergence between successive policies on the optimization trajectory, lead-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 642, + 226, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 226, + 654 + ], + "score": 1.0, + "content": "ing to the following problem:", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 618, + 505, + 654 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 655, + 394, + 698 + ], + "lines": [ + { + "bbox": [ + 217, + 655, + 394, + 698 + ], + "spans": [ + { + "bbox": [ + 217, + 655, + 394, + 698 + ], + "score": 0.9, + "content": "\\begin{array} { r l } { \\displaystyle \\operatorname* { m a x } _ { \\theta } } & { \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi } \\left[ \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi ( a _ { t } | s _ { t } ) } \\widehat { A } _ { \\pi } ( s _ { t } , a _ { t } ) \\right] } \\\\ { \\mathrm { s . t . } } & { D _ { K L } ( \\pi _ { \\theta } ( \\cdot \\mid s ) | | \\pi ( \\cdot \\mid s ) ) \\leq \\delta , \\quad \\forall s ~ . } \\end{array}", + "type": "interline_equation", + "image_path": "1c76c841578edf86898b179efb59aca8cc71ebb10567e69380e287ef7135954b.jpg" + } + ] + } + ], + "index": 45, + "virtual_lines": [ + { + "bbox": [ + 217, + 655, + 394, + 669.3333333333334 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 217, + 669.3333333333334, + 394, + 683.6666666666667 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 217, + 683.6666666666667, + 394, + 698.0000000000001 + ], + "spans": [], + "index": 46 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "In practice, this objective is maximized using a second-order approximation of the KL divergence", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "and natural gradient descent, while replacing the worst-case KL constraints over all possible states", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 719, + 478, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 478, + 734 + ], + "score": 1.0, + "content": "with an approximation of the mean KL based on the states observed in the current trajectory.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48, + "bbox_fs": [ + 105, + 699, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Proximal policy optimization. In practice, the TRPO algorithm can be computationally costly—", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 107 + ], + "score": 1.0, + "content": "the step direction is estimated with nonlinear conjugate gradients, which requires the computation", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "of multiple Hessian-vector products. To address this issue, Schulman et al. Schulman et al. (2017)", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 104, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "propose proximal policy optimization (PPO), which utilizes a different objective and does not com-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "pute a projection. Concretely, PPO proposes replacing the KL-constrained objective (16) of TRPO", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 289, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 289, + 150 + ], + "score": 1.0, + "content": "by clipping the objective function directly as:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "interline_equation", + "bbox": [ + 158, + 154, + 454, + 176 + ], + "lines": [ + { + "bbox": [ + 158, + 154, + 454, + 176 + ], + "spans": [ + { + "bbox": [ + 158, + 154, + 454, + 176 + ], + "score": 0.92, + "content": "\\operatorname* { m a x } _ { \\theta } \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi } \\left[ \\operatorname* { m i n } \\left( \\mathrm { c l i p } \\left( \\rho _ { t } , 1 - \\varepsilon , 1 + \\varepsilon \\right) \\widehat A _ { \\pi } ( s _ { t } , a _ { t } ) , \\rho _ { t } \\widehat A _ { \\pi } ( s _ { t } , a _ { t } ) \\right) \\right]", + "type": "interline_equation", + "image_path": "1e19169b4d9e120dfdabc5f1d991706c1453aa0fabc2b94420b16107ba0848be.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 158, + 154, + 454, + 176 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 181, + 133, + 192 + ], + "lines": [ + { + "bbox": [ + 105, + 180, + 135, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 135, + 194 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "interline_equation", + "bbox": [ + 273, + 195, + 337, + 223 + ], + "lines": [ + { + "bbox": [ + 273, + 195, + 337, + 223 + ], + "spans": [ + { + "bbox": [ + 273, + 195, + 337, + 223 + ], + "score": 0.95, + "content": "\\rho _ { t } = \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi ( a _ { t } | s _ { t } ) }", + "type": "interline_equation", + "image_path": "1c272ca739a1a0925bbbb93b732b445452eed9d91d04e517aa57b22c7891cb53.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 273, + 195, + 337, + 209.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 273, + 209.0, + 337, + 223.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 227, + 505, + 250 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "score": 1.0, + "content": "In addition to being simpler, PPO is intended to be faster and more sample-efficient than", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 237, + 233, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 233, + 250 + ], + "score": 1.0, + "content": "TRPO (Schulman et al., 2017).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Proximal policy optimization. In practice, the TRPO algorithm can be computationally costly—", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 107 + ], + "score": 1.0, + "content": "the step direction is estimated with nonlinear conjugate gradients, which requires the computation", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "of multiple Hessian-vector products. To address this issue, Schulman et al. Schulman et al. (2017)", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 104, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "propose proximal policy optimization (PPO), which utilizes a different objective and does not com-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "pute a projection. Concretely, PPO proposes replacing the KL-constrained objective (16) of TRPO", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 289, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 289, + 150 + ], + "score": 1.0, + "content": "by clipping the objective function directly as:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 104, + 82, + 505, + 150 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 158, + 154, + 454, + 176 + ], + "lines": [ + { + "bbox": [ + 158, + 154, + 454, + 176 + ], + "spans": [ + { + "bbox": [ + 158, + 154, + 454, + 176 + ], + "score": 0.92, + "content": "\\operatorname* { m a x } _ { \\theta } \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi } \\left[ \\operatorname* { m i n } \\left( \\mathrm { c l i p } \\left( \\rho _ { t } , 1 - \\varepsilon , 1 + \\varepsilon \\right) \\widehat A _ { \\pi } ( s _ { t } , a _ { t } ) , \\rho _ { t } \\widehat A _ { \\pi } ( s _ { t } , a _ { t } ) \\right) \\right]", + "type": "interline_equation", + "image_path": "1e19169b4d9e120dfdabc5f1d991706c1453aa0fabc2b94420b16107ba0848be.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 158, + 154, + 454, + 176 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 181, + 133, + 192 + ], + "lines": [ + { + "bbox": [ + 105, + 180, + 135, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 135, + 194 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 180, + 135, + 194 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 273, + 195, + 337, + 223 + ], + "lines": [ + { + "bbox": [ + 273, + 195, + 337, + 223 + ], + "spans": [ + { + "bbox": [ + 273, + 195, + 337, + 223 + ], + "score": 0.95, + "content": "\\rho _ { t } = \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi ( a _ { t } | s _ { t } ) }", + "type": "interline_equation", + "image_path": "1c272ca739a1a0925bbbb93b732b445452eed9d91d04e517aa57b22c7891cb53.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 273, + 195, + 337, + 209.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 273, + 209.0, + 337, + 223.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 227, + 505, + 250 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "score": 1.0, + "content": "In addition to being simpler, PPO is intended to be faster and more sample-efficient than", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 237, + 233, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 233, + 250 + ], + "score": 1.0, + "content": "TRPO (Schulman et al., 2017).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 226, + 505, + 250 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 232, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 233, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 233, + 95 + ], + "score": 1.0, + "content": "A.2 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 105, + 102, + 476, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 101, + 479, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 479, + 118 + ], + "score": 1.0, + "content": "We use the following parameters for PPO and TRPO based on a hyperparameter grid search:", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 120, + 145, + 491, + 399 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 188, + 125, + 422, + 137 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 188, + 124, + 422, + 139 + ], + "spans": [ + { + "bbox": [ + 188, + 124, + 422, + 139 + ], + "score": 1.0, + "content": "Table 1: Hyperparameters for PPO and TRPO algorithms.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "table_body", + "bbox": [ + 120, + 145, + 491, + 399 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 145, + 491, + 399 + ], + "spans": [ + { + "bbox": [ + 120, + 145, + 491, + 399 + ], + "score": 0.987, + "html": "
Humanoid-v2Walker2d-v2Hopper-v2
PPOTRPOPPOTRPOPPOTRPO
Timesteps per iteration204820482048204820482048
Discount factor ()0.990.990.990.990.990.99
GAE discount (λ)0.950.950.950.950.950.95
Value network LR0.00010.00030.00030.00030.00020.0002
Value net num. epochs101010101010
Policy net hidden layers[64, 64][64, 64][64,64][64, 64][64, 64][64,64]
Value net hidden layers[64, 64][64, 64][64, 64][64,64][64,64][64,64]
KL constraint (δ)N/A0.07N/A0.04N/A0.13
Fisher est. fractionN/A0.1N/A0.1N/A0.1
Conjugate grad. stepsN/A10N/A10N/A10
CG dampingN/A0.1N/A0.1N/A0.1
Backtracking stepsN/A10N/A10N/A10
Policy LR (Adam)0.00025N/A0.0004N/A0.00045N/A
Policy epochs10N/A10N/A10N/A
PPO Clipping ε0.2N/A0.2N/A0.2N/A
Entropy coeff.0.00.00.00.00.00.0
Reward clipping[-10,10]一[-10,10]一[-10, 10]1
Reward normalizationOnOffOnOffOnOff
State clipping[-10,10]1[-10,10]1[-10, 10]1
", + "type": "table", + "image_path": "ce554d2ca3dd0022c9b04c16ce0c50ead26d6bb99ca0925eaebce576f01b932d.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 120, + 145, + 491, + 229.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 120, + 229.66666666666669, + 491, + 314.33333333333337 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 120, + 314.33333333333337, + 491, + 399.00000000000006 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 410, + 461, + 422 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 407, + 463, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 209, + 425 + ], + "score": 1.0, + "content": "All error bars we plot are", + "type": "text" + }, + { + "bbox": [ + 209, + 410, + 229, + 421 + ], + "score": 0.86, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 407, + 463, + 425 + ], + "score": 1.0, + "content": "confidence intervals, obtained via bootstrapped sampling.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + } + ], + "index": 4 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 232, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 233, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 233, + 95 + ], + "score": 1.0, + "content": "A.2 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 105, + 102, + 476, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 101, + 479, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 479, + 118 + ], + "score": 1.0, + "content": "We use the following parameters for PPO and TRPO based on a hyperparameter grid search:", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 101, + 479, + 118 + ] + }, + { + "type": "table", + "bbox": [ + 120, + 145, + 491, + 399 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 188, + 125, + 422, + 137 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 188, + 124, + 422, + 139 + ], + "spans": [ + { + "bbox": [ + 188, + 124, + 422, + 139 + ], + "score": 1.0, + "content": "Table 1: Hyperparameters for PPO and TRPO algorithms.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "table_body", + "bbox": [ + 120, + 145, + 491, + 399 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 145, + 491, + 399 + ], + "spans": [ + { + "bbox": [ + 120, + 145, + 491, + 399 + ], + "score": 0.987, + "html": "
Humanoid-v2Walker2d-v2Hopper-v2
PPOTRPOPPOTRPOPPOTRPO
Timesteps per iteration204820482048204820482048
Discount factor ()0.990.990.990.990.990.99
GAE discount (λ)0.950.950.950.950.950.95
Value network LR0.00010.00030.00030.00030.00020.0002
Value net num. epochs101010101010
Policy net hidden layers[64, 64][64, 64][64,64][64, 64][64, 64][64,64]
Value net hidden layers[64, 64][64, 64][64, 64][64,64][64,64][64,64]
KL constraint (δ)N/A0.07N/A0.04N/A0.13
Fisher est. fractionN/A0.1N/A0.1N/A0.1
Conjugate grad. stepsN/A10N/A10N/A10
CG dampingN/A0.1N/A0.1N/A0.1
Backtracking stepsN/A10N/A10N/A10
Policy LR (Adam)0.00025N/A0.0004N/A0.00045N/A
Policy epochs10N/A10N/A10N/A
PPO Clipping ε0.2N/A0.2N/A0.2N/A
Entropy coeff.0.00.00.00.00.00.0
Reward clipping[-10,10]一[-10,10]一[-10, 10]1
Reward normalizationOnOffOnOffOnOff
State clipping[-10,10]1[-10,10]1[-10, 10]1
", + "type": "table", + "image_path": "ce554d2ca3dd0022c9b04c16ce0c50ead26d6bb99ca0925eaebce576f01b932d.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 120, + 145, + 491, + 229.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 120, + 229.66666666666669, + 491, + 314.33333333333337 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 120, + 314.33333333333337, + 491, + 399.00000000000006 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 410, + 461, + 422 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 407, + 463, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 209, + 425 + ], + "score": 1.0, + "content": "All error bars we plot are", + "type": "text" + }, + { + "bbox": [ + 209, + 410, + 229, + 421 + ], + "score": 0.86, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 407, + 463, + 425 + ], + "score": 1.0, + "content": "confidence intervals, obtained via bootstrapped sampling.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + } + ], + "index": 4 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 188, + 102, + 423, + 595 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 108, + 82, + 252, + 93 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 81, + 253, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 253, + 95 + ], + "score": 1.0, + "content": "A.3 STANDARD REWARD PLOTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image_body", + "bbox": [ + 188, + 102, + 423, + 595 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 188, + 102, + 423, + 595 + ], + "spans": [ + { + "bbox": [ + 188, + 102, + 423, + 595 + ], + "score": 0.957, + "type": "image", + "image_path": "b09d1c68040ce8eb485111bb7b2e2abe53cfbd6471cca62b69680e86279cd880.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 188, + 102, + 423, + 116.08571428571429 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 188, + 116.08571428571429, + 423, + 130.17142857142858 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 188, + 130.17142857142858, + 423, + 144.25714285714287 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 188, + 144.25714285714287, + 423, + 158.34285714285716 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 188, + 158.34285714285716, + 423, + 172.42857142857144 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 188, + 172.42857142857144, + 423, + 186.51428571428573 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 188, + 186.51428571428573, + 423, + 200.60000000000002 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 188, + 200.60000000000002, + 423, + 214.6857142857143 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 188, + 214.6857142857143, + 423, + 228.7714285714286 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 188, + 228.7714285714286, + 423, + 242.8571428571429 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 188, + 242.8571428571429, + 423, + 256.9428571428572 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 188, + 256.9428571428572, + 423, + 271.02857142857147 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 188, + 271.02857142857147, + 423, + 285.11428571428576 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 188, + 285.11428571428576, + 423, + 299.20000000000005 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 188, + 299.20000000000005, + 423, + 313.28571428571433 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 188, + 313.28571428571433, + 423, + 327.3714285714286 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 188, + 327.3714285714286, + 423, + 341.4571428571429 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 188, + 341.4571428571429, + 423, + 355.5428571428572 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 188, + 355.5428571428572, + 423, + 369.6285714285715 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 188, + 369.6285714285715, + 423, + 383.7142857142858 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 188, + 383.7142857142858, + 423, + 397.80000000000007 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 188, + 397.80000000000007, + 423, + 411.88571428571436 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 188, + 411.88571428571436, + 423, + 425.97142857142865 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 188, + 425.97142857142865, + 423, + 440.05714285714294 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 188, + 440.05714285714294, + 423, + 454.1428571428572 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 188, + 454.1428571428572, + 423, + 468.2285714285715 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 188, + 468.2285714285715, + 423, + 482.3142857142858 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 188, + 482.3142857142858, + 423, + 496.4000000000001 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 188, + 496.4000000000001, + 423, + 510.4857142857144 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 188, + 510.4857142857144, + 423, + 524.5714285714287 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 188, + 524.5714285714287, + 423, + 538.657142857143 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 188, + 538.657142857143, + 423, + 552.7428571428572 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 188, + 552.7428571428572, + 423, + 566.8285714285715 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 188, + 566.8285714285715, + 423, + 580.9142857142858 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 188, + 580.9142857142858, + 423, + 595.0000000000001 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 603, + 506, + 638 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 603, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 616 + ], + "score": 1.0, + "content": "Figure 8: Mean reward for the studied policy gradient algorithms on standard MuJoCo benchmark", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 614, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 505, + 628 + ], + "score": 1.0, + "content": "tasks. For each algorithm, we perform 24 random trials using the best performing hyperparameter", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 625, + 334, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 334, + 639 + ], + "score": 1.0, + "content": "configuration, with 10 of the random agents shown here.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + } + ], + "index": 18 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 188, + 102, + 423, + 595 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 108, + 82, + 252, + 93 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 81, + 253, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 253, + 95 + ], + "score": 1.0, + "content": "A.3 STANDARD REWARD PLOTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image_body", + "bbox": [ + 188, + 102, + 423, + 595 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 188, + 102, + 423, + 595 + ], + "spans": [ + { + "bbox": [ + 188, + 102, + 423, + 595 + ], + "score": 0.957, + "type": "image", + "image_path": "b09d1c68040ce8eb485111bb7b2e2abe53cfbd6471cca62b69680e86279cd880.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 188, + 102, + 423, + 116.08571428571429 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 188, + 116.08571428571429, + 423, + 130.17142857142858 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 188, + 130.17142857142858, + 423, + 144.25714285714287 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 188, + 144.25714285714287, + 423, + 158.34285714285716 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 188, + 158.34285714285716, + 423, + 172.42857142857144 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 188, + 172.42857142857144, + 423, + 186.51428571428573 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 188, + 186.51428571428573, + 423, + 200.60000000000002 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 188, + 200.60000000000002, + 423, + 214.6857142857143 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 188, + 214.6857142857143, + 423, + 228.7714285714286 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 188, + 228.7714285714286, + 423, + 242.8571428571429 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 188, + 242.8571428571429, + 423, + 256.9428571428572 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 188, + 256.9428571428572, + 423, + 271.02857142857147 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 188, + 271.02857142857147, + 423, + 285.11428571428576 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 188, + 285.11428571428576, + 423, + 299.20000000000005 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 188, + 299.20000000000005, + 423, + 313.28571428571433 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 188, + 313.28571428571433, + 423, + 327.3714285714286 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 188, + 327.3714285714286, + 423, + 341.4571428571429 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 188, + 341.4571428571429, + 423, + 355.5428571428572 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 188, + 355.5428571428572, + 423, + 369.6285714285715 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 188, + 369.6285714285715, + 423, + 383.7142857142858 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 188, + 383.7142857142858, + 423, + 397.80000000000007 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 188, + 397.80000000000007, + 423, + 411.88571428571436 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 188, + 411.88571428571436, + 423, + 425.97142857142865 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 188, + 425.97142857142865, + 423, + 440.05714285714294 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 188, + 440.05714285714294, + 423, + 454.1428571428572 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 188, + 454.1428571428572, + 423, + 468.2285714285715 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 188, + 468.2285714285715, + 423, + 482.3142857142858 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 188, + 482.3142857142858, + 423, + 496.4000000000001 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 188, + 496.4000000000001, + 423, + 510.4857142857144 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 188, + 510.4857142857144, + 423, + 524.5714285714287 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 188, + 524.5714285714287, + 423, + 538.657142857143 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 188, + 538.657142857143, + 423, + 552.7428571428572 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 188, + 552.7428571428572, + 423, + 566.8285714285715 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 188, + 566.8285714285715, + 423, + 580.9142857142858 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 188, + 580.9142857142858, + 423, + 595.0000000000001 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 603, + 506, + 638 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 603, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 616 + ], + "score": 1.0, + "content": "Figure 8: Mean reward for the studied policy gradient algorithms on standard MuJoCo benchmark", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 614, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 505, + 628 + ], + "score": 1.0, + "content": "tasks. For each algorithm, we perform 24 random trials using the best performing hyperparameter", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 625, + 334, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 334, + 639 + ], + "score": 1.0, + "content": "configuration, with 10 of the random agents shown here.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + } + ], + "index": 18 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 107, + 503, + 316 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 107, + 503, + 316 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 107, + 503, + 316 + ], + "spans": [ + { + "bbox": [ + 107, + 107, + 503, + 316 + ], + "score": 0.975, + "type": "image", + "image_path": "db6c9fe908d6aea5f69b186d211735147297b28e55664e1b76de44ed708f234e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 107, + 503, + 176.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 176.66666666666669, + 503, + 246.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 246.33333333333337, + 503, + 316.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 324, + 505, + 413 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "Figure 9: Empirical variance of the gradient (c.f. (1)) as a function of the number of state-action", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 336, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 349 + ], + "score": 1.0, + "content": "pairs used in estimation for policy gradient methods. We obtain multiple gradient estimates using", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 347, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 505, + 359 + ], + "score": 1.0, + "content": "a given number of state-action pairs from the policy at a particular iteration. We then measure the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "score": 1.0, + "content": "average pairwise cosine similarity between these repeated gradient measurements, along with the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 127, + 380 + ], + "score": 0.87, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 368, + 505, + 381 + ], + "score": 1.0, + "content": "confidence intervals (shaded). Each of the colored lines (for a specific algorithm) represents", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 380, + 504, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 504, + 392 + ], + "score": 1.0, + "content": "a particular trained agent (we perform multiple trials with the same hyperparameter configurations", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "but different random seeds). The dotted vertical black line (at 2K) indicates the sample regime used", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 401, + 464, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 464, + 414 + ], + "score": 1.0, + "content": "for gradient estimation in standard practical implementations of policy gradient methods.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5 + } + ], + "index": 3.75 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 752, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 83, + 293, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 294, + 95 + ], + "score": 1.0, + "content": "A.4 QUALITY OF GRADIENT ESTIMATION", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 107, + 503, + 316 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 107, + 503, + 316 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 107, + 503, + 316 + ], + "spans": [ + { + "bbox": [ + 107, + 107, + 503, + 316 + ], + "score": 0.975, + "type": "image", + "image_path": "db6c9fe908d6aea5f69b186d211735147297b28e55664e1b76de44ed708f234e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 107, + 503, + 176.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 176.66666666666669, + 503, + 246.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 246.33333333333337, + 503, + 316.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 324, + 505, + 413 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "Figure 9: Empirical variance of the gradient (c.f. (1)) as a function of the number of state-action", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 336, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 349 + ], + "score": 1.0, + "content": "pairs used in estimation for policy gradient methods. We obtain multiple gradient estimates using", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 347, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 505, + 359 + ], + "score": 1.0, + "content": "a given number of state-action pairs from the policy at a particular iteration. We then measure the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "score": 1.0, + "content": "average pairwise cosine similarity between these repeated gradient measurements, along with the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 127, + 380 + ], + "score": 0.87, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 368, + 505, + 381 + ], + "score": 1.0, + "content": "confidence intervals (shaded). Each of the colored lines (for a specific algorithm) represents", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 380, + 504, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 504, + 392 + ], + "score": 1.0, + "content": "a particular trained agent (we perform multiple trials with the same hyperparameter configurations", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "but different random seeds). The dotted vertical black line (at 2K) indicates the sample regime used", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 401, + 464, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 464, + 414 + ], + "score": 1.0, + "content": "for gradient estimation in standard practical implementations of policy gradient methods.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5 + } + ], + "index": 3.75 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 246, + 503, + 454 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 246, + 503, + 454 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 246, + 503, + 454 + ], + "spans": [ + { + "bbox": [ + 107, + 246, + 503, + 454 + ], + "score": 0.974, + "type": "image", + "image_path": "eaf9745d78b6e2ca858d1ecaaa601f2cabfcaa1e6ed6129a80fb84f04fa5c503.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 246, + 503, + 315.3333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 315.3333333333333, + 503, + 384.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 384.66666666666663, + 503, + 453.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 463, + 505, + 563 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 463, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 506, + 477 + ], + "score": 1.0, + "content": "Figure 10: Convergence of gradient estimates to the “true” expected gradient (c.f. (1)). We measure", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "the cosine similarity between the true gradient (approximated using around 1M samples) and gradi-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 486, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 506, + 499 + ], + "score": 1.0, + "content": "ent estimates, as a function of number of state-action pairs used to obtain the later. For a particular", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "policy and state-action pair count, we obtain multiple estimates of this cosine similarity and then", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 242, + 520 + ], + "score": 1.0, + "content": "report the average, along with the", + "type": "text" + }, + { + "bbox": [ + 243, + 507, + 263, + 518 + ], + "score": 0.89, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "confidence intervals (shaded). Each of the colored lines (for", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 518, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 104, + 518, + 506, + 531 + ], + "score": 1.0, + "content": "a specific algorithm) represents a particular trained agent (we perform multiple trials with the same", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "hyperparameter configurations but different random seeds). The dotted vertical black line (at 2K)", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "score": 1.0, + "content": "indicates the sample regime used for gradient estimation in standard practical implementations of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 551, + 208, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 208, + 564 + ], + "score": 1.0, + "content": "policy gradient methods.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7 + } + ], + "index": 4.0 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "17", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 246, + 503, + 454 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 246, + 503, + 454 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 246, + 503, + 454 + ], + "spans": [ + { + "bbox": [ + 107, + 246, + 503, + 454 + ], + "score": 0.974, + "type": "image", + "image_path": "eaf9745d78b6e2ca858d1ecaaa601f2cabfcaa1e6ed6129a80fb84f04fa5c503.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 246, + 503, + 315.3333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 315.3333333333333, + 503, + 384.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 384.66666666666663, + 503, + 453.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 463, + 505, + 563 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 463, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 506, + 477 + ], + "score": 1.0, + "content": "Figure 10: Convergence of gradient estimates to the “true” expected gradient (c.f. (1)). We measure", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "the cosine similarity between the true gradient (approximated using around 1M samples) and gradi-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 486, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 506, + 499 + ], + "score": 1.0, + "content": "ent estimates, as a function of number of state-action pairs used to obtain the later. For a particular", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "policy and state-action pair count, we obtain multiple estimates of this cosine similarity and then", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 242, + 520 + ], + "score": 1.0, + "content": "report the average, along with the", + "type": "text" + }, + { + "bbox": [ + 243, + 507, + 263, + 518 + ], + "score": 0.89, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "confidence intervals (shaded). Each of the colored lines (for", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 518, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 104, + 518, + 506, + 531 + ], + "score": 1.0, + "content": "a specific algorithm) represents a particular trained agent (we perform multiple trials with the same", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "hyperparameter configurations but different random seeds). The dotted vertical black line (at 2K)", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "score": 1.0, + "content": "indicates the sample regime used for gradient estimation in standard practical implementations of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 551, + 208, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 208, + 564 + ], + "score": 1.0, + "content": "policy gradient methods.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7 + } + ], + "index": 4.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 127, + 109, + 483, + 463 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 127, + 109, + 483, + 463 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 109, + 483, + 463 + ], + "spans": [ + { + "bbox": [ + 127, + 109, + 483, + 463 + ], + "score": 0.978, + "type": "image", + "image_path": "e16e2ef2341635befb3f55136fc77ddb35d3971e8920c0eb52be8368c4792f55.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 127, + 109, + 483, + 227.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 127, + 227.0, + 483, + 345.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 127, + 345.0, + 483, + 463.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 472, + 506, + 528 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 473, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 505, + 484 + ], + "score": 1.0, + "content": "Figure 11: Quality of value prediction in terms of mean relative error (MRE) on train state-action", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "score": 1.0, + "content": "pairs for agents trained to solve the MuJoCo tasks. We see in that the agents do indeed succeed at", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 495, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 506 + ], + "score": 1.0, + "content": "solving the supervised learning task they are trained for – the train MRE on the GAE-based value", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 124, + 518 + ], + "score": 1.0, + "content": "loss", + "type": "text" + }, + { + "bbox": [ + 125, + 505, + 190, + 518 + ], + "score": 0.92, + "content": "( \\bar { V } _ { o l d } + \\bar { A _ { G A E } } ) ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "(c.f. (4)) is small (left column). We observe that the returns MRE is quite small", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 516, + 200, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 200, + 529 + ], + "score": 1.0, + "content": "as well (right column).", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 82, + 220, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 221, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 221, + 95 + ], + "score": 1.0, + "content": "A.5 VALUE PREDICTION", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 127, + 109, + 483, + 463 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 127, + 109, + 483, + 463 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 109, + 483, + 463 + ], + "spans": [ + { + "bbox": [ + 127, + 109, + 483, + 463 + ], + "score": 0.978, + "type": "image", + "image_path": "e16e2ef2341635befb3f55136fc77ddb35d3971e8920c0eb52be8368c4792f55.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 127, + 109, + 483, + 227.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 127, + 227.0, + 483, + 345.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 127, + 345.0, + 483, + 463.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 472, + 506, + 528 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 473, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 505, + 484 + ], + "score": 1.0, + "content": "Figure 11: Quality of value prediction in terms of mean relative error (MRE) on train state-action", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "score": 1.0, + "content": "pairs for agents trained to solve the MuJoCo tasks. We see in that the agents do indeed succeed at", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 495, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 506 + ], + "score": 1.0, + "content": "solving the supervised learning task they are trained for – the train MRE on the GAE-based value", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 124, + 518 + ], + "score": 1.0, + "content": "loss", + "type": "text" + }, + { + "bbox": [ + 125, + 505, + 190, + 518 + ], + "score": 0.92, + "content": "( \\bar { V } _ { o l d } + \\bar { A _ { G A E } } ) ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "(c.f. (4)) is small (left column). We observe that the returns MRE is quite small", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 516, + 200, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 200, + 529 + ], + "score": 1.0, + "content": "as well (right column).", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 239, + 502, + 497 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 239, + 502, + 497 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 239, + 502, + 497 + ], + "spans": [ + { + "bbox": [ + 107, + 239, + 502, + 497 + ], + "score": 0.966, + "type": "image", + "image_path": "9e3757b24093255f914c2704a40dc876cbbcfa06c61e80505054ebedbb2a1d41.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 239, + 502, + 325.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 325.0, + 502, + 411.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 411.0, + 502, + 497.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 506, + 506, + 573 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 507, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 505, + 518 + ], + "score": 1.0, + "content": "Figure 12: Quality of value prediction in terms of mean relative error (MRE) on heldout state-action", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 517, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 506, + 529 + ], + "score": 1.0, + "content": "pairs for agents trained to solve MuJoCo tasks. We see in that the agents do indeed succeed at", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "solving the supervised learning task they are trained for – the validation MRE on the GAE-based", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 151, + 552 + ], + "score": 1.0, + "content": "value loss", + "type": "text" + }, + { + "bbox": [ + 151, + 539, + 221, + 551 + ], + "score": 0.92, + "content": "( \\bar { V _ { o l d } } + A _ { G A E } ) ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "(c.f. (4)) is small (left column). On the other hand, we see that the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 550, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 407, + 562 + ], + "score": 1.0, + "content": "returns MRE is still quite high – the learned value function is off by about", + "type": "text" + }, + { + "bbox": [ + 407, + 550, + 427, + 561 + ], + "score": 0.88, + "content": "5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 551, + 505, + 562 + ], + "score": 1.0, + "content": "with respect to the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 561, + 292, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 292, + 574 + ], + "score": 1.0, + "content": "underlying true value function (right column).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 239, + 502, + 497 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 239, + 502, + 497 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 239, + 502, + 497 + ], + "spans": [ + { + "bbox": [ + 107, + 239, + 502, + 497 + ], + "score": 0.966, + "type": "image", + "image_path": "9e3757b24093255f914c2704a40dc876cbbcfa06c61e80505054ebedbb2a1d41.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 239, + 502, + 325.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 325.0, + 502, + 411.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 411.0, + 502, + 497.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 506, + 506, + 573 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 507, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 505, + 518 + ], + "score": 1.0, + "content": "Figure 12: Quality of value prediction in terms of mean relative error (MRE) on heldout state-action", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 517, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 506, + 529 + ], + "score": 1.0, + "content": "pairs for agents trained to solve MuJoCo tasks. We see in that the agents do indeed succeed at", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "solving the supervised learning task they are trained for – the validation MRE on the GAE-based", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 151, + 552 + ], + "score": 1.0, + "content": "value loss", + "type": "text" + }, + { + "bbox": [ + 151, + 539, + 221, + 551 + ], + "score": 0.92, + "content": "( \\bar { V _ { o l d } } + A _ { G A E } ) ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "(c.f. (4)) is small (left column). On the other hand, we see that the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 550, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 407, + 562 + ], + "score": 1.0, + "content": "returns MRE is still quite high – the learned value function is off by about", + "type": "text" + }, + { + "bbox": [ + 407, + 550, + 427, + 561 + ], + "score": 0.88, + "content": "5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 551, + 505, + 562 + ], + "score": 1.0, + "content": "with respect to the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 561, + 292, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 292, + 574 + ], + "score": 1.0, + "content": "underlying true value function (right column).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 106, + 493, + 436 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 106, + 493, + 436 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 106, + 493, + 436 + ], + "spans": [ + { + "bbox": [ + 108, + 106, + 493, + 436 + ], + "score": 0.974, + "type": "image", + "image_path": "492ad53b0b623ea1c8de6f1b55e166845454e05d8dda57581999fbe4a8caaa6b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 106, + 493, + 216.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 216.0, + 493, + 326.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 326.0, + 493, + 436.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 201, + 445, + 409, + 457 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 200, + 443, + 410, + 459 + ], + "spans": [ + { + "bbox": [ + 200, + 443, + 410, + 459 + ], + "score": 1.0, + "content": "Figure 13: Humanoid-v2 – PPO reward landscapes.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 83, + 253, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 254, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 254, + 95 + ], + "score": 1.0, + "content": "A.6 OPTIMIZATION LANDSCAPE", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 110, + 132, + 124, + 417 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 106, + 493, + 436 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 106, + 493, + 436 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 106, + 493, + 436 + ], + "spans": [ + { + "bbox": [ + 108, + 106, + 493, + 436 + ], + "score": 0.974, + "type": "image", + "image_path": "492ad53b0b623ea1c8de6f1b55e166845454e05d8dda57581999fbe4a8caaa6b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 106, + 493, + 216.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 216.0, + 493, + 326.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 326.0, + 493, + 436.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 201, + 445, + 409, + 457 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 200, + 443, + 410, + 459 + ], + "spans": [ + { + "bbox": [ + 200, + 443, + 410, + 459 + ], + "score": 1.0, + "content": "Figure 13: Humanoid-v2 – PPO reward landscapes.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "spans": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "score": 0.972, + "type": "image", + "image_path": "e71942f5b97f65b0a5ab9ef5c63749592e3e90f18e0d9113c71bc5245529686f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 230, + 493, + 340.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 340.0, + 493, + 450.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 450.0, + 493, + 560.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 197, + 568, + 413, + 580 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 197, + 567, + 414, + 582 + ], + "spans": [ + { + "bbox": [ + 197, + 567, + 414, + 582 + ], + "score": 1.0, + "content": "Figure 14: Humanoid-v2 – TRPO reward landscapes.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ], + "page_idx": 20, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 109, + 260, + 124, + 536 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "spans": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "score": 0.972, + "type": "image", + "image_path": "e71942f5b97f65b0a5ab9ef5c63749592e3e90f18e0d9113c71bc5245529686f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 230, + 493, + 340.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 340.0, + 493, + 450.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 450.0, + 493, + 560.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 197, + 568, + 413, + 580 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 197, + 567, + 414, + 582 + ], + "spans": [ + { + "bbox": [ + 197, + 567, + 414, + 582 + ], + "score": 1.0, + "content": "Figure 14: Humanoid-v2 – TRPO reward landscapes.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "spans": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "score": 0.97, + "type": "image", + "image_path": "92360b41ac24e124107a6a3bf5b314f4ede8a0c180a149bf671750f3eabe09d8.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 230, + 493, + 340.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 340.0, + 493, + 450.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 450.0, + 493, + 560.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 204, + 568, + 406, + 580 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 202, + 567, + 408, + 582 + ], + "spans": [ + { + "bbox": [ + 202, + 567, + 408, + 582 + ], + "score": 1.0, + "content": "Figure 15: Walker2d-v2 – PPO reward landscapes.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ], + "page_idx": 21, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "spans": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "score": 0.97, + "type": "image", + "image_path": "92360b41ac24e124107a6a3bf5b314f4ede8a0c180a149bf671750f3eabe09d8.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 230, + 493, + 340.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 340.0, + 493, + 450.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 450.0, + 493, + 560.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 204, + 568, + 406, + 580 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 202, + 567, + 408, + 582 + ], + "spans": [ + { + "bbox": [ + 202, + 567, + 408, + 582 + ], + "score": 1.0, + "content": "Figure 15: Walker2d-v2 – PPO reward landscapes.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "spans": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "score": 0.972, + "type": "image", + "image_path": "23b96f1084ce71237a61e54cba14cf0f7f00b45b8386ebcb9faee6b98b81ec8d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 230, + 493, + 340.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 340.0, + 493, + 450.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 450.0, + 493, + 560.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 201, + 568, + 411, + 580 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 198, + 567, + 412, + 582 + ], + "spans": [ + { + "bbox": [ + 198, + 567, + 412, + 582 + ], + "score": 1.0, + "content": "Figure 16: Walker2d-v2 – TRPO reward landscapes.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ], + "page_idx": 22, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "spans": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "score": 0.972, + "type": "image", + "image_path": "23b96f1084ce71237a61e54cba14cf0f7f00b45b8386ebcb9faee6b98b81ec8d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 230, + 493, + 340.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 340.0, + 493, + 450.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 450.0, + 493, + 560.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 201, + 568, + 411, + 580 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 198, + 567, + 412, + 582 + ], + "spans": [ + { + "bbox": [ + 198, + 567, + 412, + 582 + ], + "score": 1.0, + "content": "Figure 16: Walker2d-v2 – TRPO reward landscapes.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "spans": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "score": 0.972, + "type": "image", + "image_path": "e1181885263802be8a0612e7e76a1b0d23fe109de1c0065c63a1b6753c7b9d7a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 230, + 493, + 340.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 340.0, + 493, + 450.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 450.0, + 493, + 560.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 209, + 568, + 403, + 580 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 207, + 567, + 403, + 582 + ], + "spans": [ + { + "bbox": [ + 207, + 567, + 403, + 582 + ], + "score": 1.0, + "content": "Figure 17: Hopper-v2 – PPO reward landscapes.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ], + "page_idx": 23, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "spans": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "score": 0.972, + "type": "image", + "image_path": "e1181885263802be8a0612e7e76a1b0d23fe109de1c0065c63a1b6753c7b9d7a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 230, + 493, + 340.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 340.0, + 493, + 450.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 450.0, + 493, + 560.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 209, + 568, + 403, + 580 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 207, + 567, + 403, + 582 + ], + "spans": [ + { + "bbox": [ + 207, + 567, + 403, + 582 + ], + "score": 1.0, + "content": "Figure 17: Hopper-v2 – PPO reward landscapes.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "spans": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "score": 0.968, + "type": "image", + "image_path": "46b2f4466bc46a834c09c86d56fe0c9566d040a40989424a57e720c34dae5420.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 230, + 493, + 340.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 340.0, + 493, + 450.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 450.0, + 493, + 560.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 204, + 568, + 406, + 580 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 203, + 567, + 407, + 582 + ], + "spans": [ + { + "bbox": [ + 203, + 567, + 407, + 582 + ], + "score": 1.0, + "content": "Figure 18: Hopper-v2 – TRPO reward landscapes.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ], + "page_idx": 24, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 230, + 493, + 560 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "spans": [ + { + "bbox": [ + 108, + 230, + 493, + 560 + ], + "score": 0.968, + "type": "image", + "image_path": "46b2f4466bc46a834c09c86d56fe0c9566d040a40989424a57e720c34dae5420.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 230, + 493, + 340.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 340.0, + 493, + 450.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 450.0, + 493, + 560.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 204, + 568, + 406, + 580 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 203, + 567, + 407, + 582 + ], + "spans": [ + { + "bbox": [ + 203, + 567, + 407, + 582 + ], + "score": 1.0, + "content": "Figure 18: Hopper-v2 – TRPO reward landscapes.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 120, + 165, + 516, + 625 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 120, + 165, + 516, + 625 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 165, + 516, + 625 + ], + "spans": [ + { + "bbox": [ + 120, + 165, + 516, + 625 + ], + "score": 0.965, + "type": "image", + "image_path": "32b3359e0631a1a531f1b03fe88fc25bb6a0e14ccbec9ce8e5281584831351b2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 120, + 165, + 516, + 318.33333333333337 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 120, + 318.33333333333337, + 516, + 471.66666666666674 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 120, + 471.66666666666674, + 516, + 625.0000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 127, + 632, + 483, + 645 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 631, + 484, + 647 + ], + "spans": [ + { + "bbox": [ + 126, + 631, + 484, + 647 + ], + "score": 1.0, + "content": "Figure 19: Humanoid-v2 TRPO landscape concentration (see Figure 5 for a description).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ], + "page_idx": 25, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "26", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 111, + 213, + 123, + 240 + ], + "lines": [ + { + "bbox": [ + 110, + 210, + 125, + 241 + ], + "spans": [ + { + "bbox": [ + 110, + 210, + 125, + 241 + ], + "score": 0.0, + "content": "", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 120, + 165, + 516, + 625 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 120, + 165, + 516, + 625 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 165, + 516, + 625 + ], + "spans": [ + { + "bbox": [ + 120, + 165, + 516, + 625 + ], + "score": 0.965, + "type": "image", + "image_path": "32b3359e0631a1a531f1b03fe88fc25bb6a0e14ccbec9ce8e5281584831351b2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 120, + 165, + 516, + 318.33333333333337 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 120, + 318.33333333333337, + 516, + 471.66666666666674 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 120, + 471.66666666666674, + 516, + 625.0000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 127, + 632, + 483, + 645 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 631, + 484, + 647 + ], + "spans": [ + { + "bbox": [ + 126, + 631, + 484, + 647 + ], + "score": 1.0, + "content": "Figure 19: Humanoid-v2 TRPO landscape concentration (see Figure 5 for a description).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 153, + 165, + 247, + 176 + ], + "lines": [ + { + "bbox": [ + 153, + 165, + 247, + 176 + ], + "spans": [], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 121, + 168, + 514, + 625 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 282, + 165, + 382, + 176 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 282, + 165, + 382, + 176 + ], + "spans": [], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "image_caption", + "bbox": [ + 413, + 165, + 516, + 176 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 413, + 165, + 516, + 176 + ], + "spans": [], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "image_body", + "bbox": [ + 121, + 168, + 514, + 625 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 121, + 168, + 514, + 625 + ], + "spans": [ + { + "bbox": [ + 121, + 168, + 514, + 625 + ], + "score": 0.966, + "type": "image", + "image_path": "c672ad624699fce7bdb3cf34e092df2e245eb748620fa547f07a8e91bbfe6af0.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 121, + 168, + 514, + 320.33333333333337 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 121, + 320.33333333333337, + 514, + 472.66666666666674 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 121, + 472.66666666666674, + 514, + 625.0000000000001 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 130, + 632, + 480, + 645 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 631, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 129, + 631, + 481, + 647 + ], + "score": 1.0, + "content": "Figure 20: Humanoid-v2 PPO landscape concentration (see Figure 5 for a description).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + } + ], + "index": 3.0 + } + ], + "page_idx": 26, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 111, + 213, + 123, + 240 + ], + "lines": [ + { + "bbox": [ + 110, + 210, + 125, + 241 + ], + "spans": [ + { + "bbox": [ + 110, + 210, + 125, + 241 + ], + "score": 0.0, + "content": "", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 153, + 165, + 247, + 176 + ], + "lines": [ + { + "bbox": [ + 153, + 165, + 247, + 176 + ], + "spans": [], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 121, + 168, + 514, + 625 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 282, + 165, + 382, + 176 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 282, + 165, + 382, + 176 + ], + "spans": [], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "image_caption", + "bbox": [ + 413, + 165, + 516, + 176 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 413, + 165, + 516, + 176 + ], + "spans": [], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "image_body", + "bbox": [ + 121, + 168, + 514, + 625 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 121, + 168, + 514, + 625 + ], + "spans": [ + { + "bbox": [ + 121, + 168, + 514, + 625 + ], + "score": 0.966, + "type": "image", + "image_path": "c672ad624699fce7bdb3cf34e092df2e245eb748620fa547f07a8e91bbfe6af0.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 121, + 168, + 514, + 320.33333333333337 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 121, + 320.33333333333337, + 514, + 472.66666666666674 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 121, + 472.66666666666674, + 514, + 625.0000000000001 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 130, + 632, + 480, + 645 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 631, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 129, + 631, + 481, + 647 + ], + "score": 1.0, + "content": "Figure 20: Humanoid-v2 PPO landscape concentration (see Figure 5 for a description).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + } + ], + "index": 3.0 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/ryxdEkHtPS/ryxdEkHtPS_model.json b/parse/train/ryxdEkHtPS/ryxdEkHtPS_model.json new file mode 100644 index 0000000000000000000000000000000000000000..01dae1c97fde00fbed9d2f55e80c4bb9385d113d --- /dev/null +++ b/parse/train/ryxdEkHtPS/ryxdEkHtPS_model.json @@ -0,0 +1,38410 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 670, + 1301, + 670, + 1301, + 1004, + 398, + 1004 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1125, + 1404, + 1125, + 1404, + 1341, + 299, + 1341 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 302, + 1494, + 1398, + 1494, + 1398, + 1588, + 302, + 1588 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 302, + 1619, + 1402, + 1619, + 1402, + 1711, + 302, + 1711 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 299, + 1357, + 1399, + 1357, + 1399, + 1418, + 299, + 1418 + ], + "score": 0.937 + }, + { + "category_id": 1, + "poly": [ + 443, + 332, + 1257, + 332, + 1257, + 397, + 443, + 397 + ], + "score": 0.93 + }, + { + "category_id": 0, + "poly": [ + 303, + 217, + 1262, + 217, + 1262, + 270, + 303, + 270 + ], + "score": 0.926 + }, + { + "category_id": 0, + "poly": [ + 302, + 1058, + 573, + 1058, + 573, + 1093, + 302, + 1093 + ], + "score": 0.893 + }, + { + "category_id": 1, + "poly": [ + 437, + 419, + 1268, + 419, + 1268, + 515, + 437, + 515 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 332, + 2005, + 997, + 2005, + 997, + 2033, + 332, + 2033 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 773, + 605, + 927, + 605, + 927, + 639, + 773, + 639 + ], + "score": 0.882 + }, + { + "category_id": 1, + "poly": [ + 308, + 1439, + 1393, + 1439, + 1393, + 1472, + 308, + 1472 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 104, + 299, + 104 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 842, + 2089, + 857, + 2089, + 857, + 2112, + 842, + 2112 + ], + "score": 0.666 + }, + { + "category_id": 1, + "poly": [ + 367, + 1735, + 1399, + 1735, + 1399, + 1885, + 367, + 1885 + ], + "score": 0.59 + }, + { + "category_id": 1, + "poly": [ + 370, + 1897, + 1399, + 1897, + 1399, + 1989, + 370, + 1989 + ], + "score": 0.574 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 216.0, + 1269.0, + 216.0, + 1269.0, + 277.0, + 295.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1055.0, + 579.0, + 1055.0, + 579.0, + 1102.0, + 294.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 2003.0, + 996.0, + 2003.0, + 996.0, + 2038.0, + 333.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 604.0, + 932.0, + 604.0, + 932.0, + 642.0, + 770.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 860.0, + 2088.0, + 860.0, + 2118.0, + 841.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 670.0, + 1305.0, + 670.0, + 1305.0, + 706.0, + 396.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 700.0, + 1304.0, + 700.0, + 1304.0, + 736.0, + 394.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 732.0, + 1305.0, + 732.0, + 1305.0, + 767.0, + 393.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 763.0, + 1304.0, + 763.0, + 1304.0, + 795.0, + 395.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 792.0, + 1306.0, + 792.0, + 1306.0, + 828.0, + 394.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 823.0, + 1306.0, + 823.0, + 1306.0, + 856.0, + 395.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 855.0, + 1305.0, + 855.0, + 1305.0, + 887.0, + 394.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 884.0, + 1305.0, + 884.0, + 1305.0, + 916.0, + 394.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 916.0, + 1304.0, + 916.0, + 1304.0, + 946.0, + 396.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 943.0, + 1306.0, + 943.0, + 1306.0, + 980.0, + 392.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 975.0, + 924.0, + 975.0, + 924.0, + 1006.0, + 393.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1127.0, + 1404.0, + 1127.0, + 1404.0, + 1161.0, + 295.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1156.0, + 1403.0, + 1156.0, + 1403.0, + 1191.0, + 294.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1184.0, + 1404.0, + 1184.0, + 1404.0, + 1226.0, + 292.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1218.0, + 1404.0, + 1218.0, + 1404.0, + 1252.0, + 294.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1249.0, + 1406.0, + 1249.0, + 1406.0, + 1284.0, + 294.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1277.0, + 1409.0, + 1277.0, + 1409.0, + 1316.0, + 294.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1311.0, + 996.0, + 1311.0, + 996.0, + 1345.0, + 295.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1494.0, + 1402.0, + 1494.0, + 1402.0, + 1531.0, + 297.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1528.0, + 1402.0, + 1528.0, + 1402.0, + 1558.0, + 297.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1556.0, + 1098.0, + 1556.0, + 1098.0, + 1593.0, + 294.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1617.0, + 1406.0, + 1617.0, + 1406.0, + 1655.0, + 295.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1647.0, + 1405.0, + 1647.0, + 1405.0, + 1688.0, + 295.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1679.0, + 1021.0, + 1679.0, + 1021.0, + 1715.0, + 296.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1354.0, + 1403.0, + 1354.0, + 1403.0, + 1393.0, + 295.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1386.0, + 1353.0, + 1386.0, + 1353.0, + 1421.0, + 295.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 330.0, + 1258.0, + 330.0, + 1258.0, + 370.0, + 442.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 360.0, + 1181.0, + 360.0, + 1181.0, + 402.0, + 513.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 416.0, + 974.0, + 416.0, + 974.0, + 459.0, + 734.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 456.0, + 1246.0, + 456.0, + 1246.0, + 487.0, + 464.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 484.0, + 1273.0, + 484.0, + 1273.0, + 519.0, + 435.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 1436.0, + 1392.0, + 1436.0, + 1392.0, + 1478.0, + 305.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 1734.0, + 1405.0, + 1734.0, + 1405.0, + 1770.0, + 367.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1765.0, + 1405.0, + 1765.0, + 1405.0, + 1800.0, + 394.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1795.0, + 1403.0, + 1795.0, + 1403.0, + 1832.0, + 394.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1826.0, + 1405.0, + 1826.0, + 1405.0, + 1861.0, + 392.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1857.0, + 1405.0, + 1857.0, + 1405.0, + 1890.0, + 394.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1894.0, + 1407.0, + 1894.0, + 1407.0, + 1935.0, + 368.0, + 1935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1927.0, + 1402.0, + 1927.0, + 1402.0, + 1960.0, + 395.0, + 1960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1959.0, + 1404.0, + 1959.0, + 1404.0, + 1992.0, + 396.0, + 1992.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 699, + 1404, + 699, + 1404, + 855, + 298, + 855 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 296, + 1151, + 1404, + 1151, + 1404, + 1367, + 296, + 1367 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 426, + 1404, + 426, + 1404, + 580, + 298, + 580 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 953, + 1404, + 953, + 1404, + 1046, + 299, + 1046 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 300, + 1401, + 1398, + 1401, + 1398, + 1642, + 300, + 1642 + ], + "score": 0.965 + }, + { + "category_id": 4, + "poly": [ + 296, + 1669, + 1405, + 1669, + 1405, + 2006, + 296, + 2006 + ], + "score": 0.963 + }, + { + "category_id": 8, + "poly": [ + 358, + 1059, + 1332, + 1059, + 1332, + 1137, + 358, + 1137 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 387, + 230, + 1398, + 230, + 1398, + 293, + 387, + 293 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 373, + 306, + 1399, + 306, + 1399, + 398, + 373, + 398 + ], + "score": 0.928 + }, + { + "category_id": 0, + "poly": [ + 297, + 626, + 1380, + 626, + 1380, + 664, + 297, + 664 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 300, + 896, + 732, + 896, + 732, + 927, + 300, + 927 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 105, + 300, + 105 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1367, + 1083, + 1399, + 1083, + 1399, + 1113, + 1367, + 1113 + ], + "score": 0.84 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.719 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.103 + }, + { + "category_id": 14, + "poly": [ + 362, + 1059, + 1335, + 1059, + 1335, + 1136, + 362, + 1136 + ], + "score": 0.95, + "latex": "\\hat { g } = \\nabla _ { \\theta } \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi _ { 0 } } \\left[ \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi _ { 0 } ( a _ { t } | s _ { t } ) } \\widehat { A } _ { \\pi _ { 0 } } ( s _ { t } , a _ { t } ) \\right] = \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi _ { 0 } } \\left[ \\frac { \\nabla _ { \\theta } \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi _ { 0 } ( a _ { t } | s _ { t } ) } \\widehat { A } _ { \\pi _ { 0 } } ( s _ { t } , a _ { t } ) \\right] ," + }, + { + "category_id": 13, + "poly": [ + 920, + 1242, + 993, + 1242, + 993, + 1273, + 920, + 1273 + ], + "score": 0.88, + "latex": "\\sim 1 0 ^ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1760, + 353, + 1760, + 353, + 1791, + 297, + 1791 + ], + "score": 0.87, + "latex": "9 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 526, + 1822, + 621, + 1822, + 621, + 1850, + 526, + 1850 + ], + "score": 0.87, + "latex": "x = 2 \\mathsf { K }" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 1401.0, + 513.0, + 1401.0, + 513.0, + 1427.0, + 400.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1401.0, + 800.0, + 1401.0, + 800.0, + 1427.0, + 669.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 1401.0, + 1079.0, + 1401.0, + 1079.0, + 1427.0, + 948.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 1401.0, + 1357.0, + 1401.0, + 1357.0, + 1427.0, + 1225.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1426.0, + 355.0, + 1426.0, + 355.0, + 1602.0, + 299.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1428.0, + 447.0, + 1428.0, + 447.0, + 1456.0, + 392.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 1426.0, + 633.0, + 1426.0, + 633.0, + 1600.0, + 577.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1428.0, + 724.0, + 1428.0, + 724.0, + 1456.0, + 669.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 1426.0, + 909.0, + 1426.0, + 909.0, + 1602.0, + 855.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 1427.0, + 1002.0, + 1427.0, + 1002.0, + 1456.0, + 948.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 1426.0, + 1188.0, + 1426.0, + 1188.0, + 1602.0, + 1132.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1227.0, + 1428.0, + 1281.0, + 1428.0, + 1281.0, + 1456.0, + 1227.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1446.0, + 438.0, + 1446.0, + 438.0, + 1478.0, + 392.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1446.0, + 715.0, + 1446.0, + 715.0, + 1478.0, + 669.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 1443.0, + 995.0, + 1443.0, + 995.0, + 1477.0, + 948.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 1446.0, + 1272.0, + 1446.0, + 1272.0, + 1477.0, + 1226.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1482.0, + 352.0, + 1482.0, + 352.0, + 1501.0, + 327.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 1480.0, + 633.0, + 1480.0, + 633.0, + 1503.0, + 603.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 1480.0, + 910.0, + 1480.0, + 910.0, + 1503.0, + 881.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 1480.0, + 1187.0, + 1480.0, + 1187.0, + 1503.0, + 1160.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1538.0, + 354.0, + 1538.0, + 354.0, + 1556.0, + 327.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 1537.0, + 633.0, + 1537.0, + 633.0, + 1558.0, + 603.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 1538.0, + 909.0, + 1538.0, + 909.0, + 1557.0, + 883.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 1538.0, + 1187.0, + 1538.0, + 1187.0, + 1556.0, + 1161.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1591.0, + 354.0, + 1591.0, + 354.0, + 1613.0, + 321.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1594.0, + 631.0, + 1594.0, + 631.0, + 1611.0, + 601.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 1594.0, + 908.0, + 1594.0, + 908.0, + 1611.0, + 878.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1594.0, + 1186.0, + 1594.0, + 1186.0, + 1611.0, + 1156.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1603.0, + 376.0, + 1603.0, + 376.0, + 1621.0, + 351.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1603.0, + 425.0, + 1603.0, + 425.0, + 1621.0, + 396.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 1603.0, + 471.0, + 1603.0, + 471.0, + 1621.0, + 443.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 490.0, + 1603.0, + 517.0, + 1603.0, + 517.0, + 1621.0, + 490.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 1602.0, + 564.0, + 1602.0, + 564.0, + 1623.0, + 535.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 1603.0, + 654.0, + 1603.0, + 654.0, + 1622.0, + 629.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 1603.0, + 702.0, + 1603.0, + 702.0, + 1621.0, + 674.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 1603.0, + 749.0, + 1603.0, + 749.0, + 1621.0, + 722.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1603.0, + 795.0, + 1603.0, + 795.0, + 1621.0, + 768.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 1602.0, + 843.0, + 1602.0, + 843.0, + 1623.0, + 812.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1603.0, + 933.0, + 1603.0, + 933.0, + 1622.0, + 906.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1603.0, + 979.0, + 1603.0, + 979.0, + 1621.0, + 953.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 1603.0, + 1027.0, + 1603.0, + 1027.0, + 1621.0, + 999.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1603.0, + 1075.0, + 1603.0, + 1075.0, + 1621.0, + 1045.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 1603.0, + 1119.0, + 1603.0, + 1119.0, + 1622.0, + 1092.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1603.0, + 1211.0, + 1603.0, + 1211.0, + 1621.0, + 1183.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 1603.0, + 1257.0, + 1603.0, + 1257.0, + 1621.0, + 1230.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 1603.0, + 1303.0, + 1603.0, + 1303.0, + 1621.0, + 1277.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1323.0, + 1603.0, + 1352.0, + 1603.0, + 1352.0, + 1621.0, + 1323.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1371.0, + 1603.0, + 1396.0, + 1603.0, + 1396.0, + 1621.0, + 1371.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 1616.0, + 538.0, + 1616.0, + 538.0, + 1642.0, + 375.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 1616.0, + 817.0, + 1616.0, + 817.0, + 1642.0, + 653.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 1616.0, + 1094.0, + 1616.0, + 1094.0, + 1642.0, + 931.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 1614.0, + 1372.0, + 1614.0, + 1372.0, + 1642.0, + 1207.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1670.0, + 1403.0, + 1670.0, + 1403.0, + 1705.0, + 295.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1700.0, + 1405.0, + 1700.0, + 1405.0, + 1734.0, + 295.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1731.0, + 1405.0, + 1731.0, + 1405.0, + 1765.0, + 294.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1761.0, + 1405.0, + 1761.0, + 1405.0, + 1795.0, + 354.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1792.0, + 1403.0, + 1792.0, + 1403.0, + 1826.0, + 294.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1820.0, + 525.0, + 1820.0, + 525.0, + 1859.0, + 295.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 1820.0, + 1406.0, + 1820.0, + 1406.0, + 1859.0, + 622.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 1405.0, + 1850.0, + 1405.0, + 1887.0, + 294.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1883.0, + 1403.0, + 1883.0, + 1403.0, + 1918.0, + 295.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1911.0, + 1405.0, + 1911.0, + 1405.0, + 1948.0, + 293.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1944.0, + 1405.0, + 1944.0, + 1405.0, + 1979.0, + 294.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1970.0, + 873.0, + 1970.0, + 873.0, + 2011.0, + 293.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 625.0, + 1386.0, + 625.0, + 1386.0, + 669.0, + 293.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 892.0, + 737.0, + 892.0, + 737.0, + 933.0, + 293.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 71.0, + 817.0, + 71.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 701.0, + 1404.0, + 701.0, + 1404.0, + 734.0, + 296.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 730.0, + 1402.0, + 730.0, + 1402.0, + 768.0, + 294.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 761.0, + 1404.0, + 761.0, + 1404.0, + 798.0, + 294.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 792.0, + 1406.0, + 792.0, + 1406.0, + 829.0, + 294.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 822.0, + 1097.0, + 822.0, + 1097.0, + 860.0, + 294.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1148.0, + 1405.0, + 1148.0, + 1405.0, + 1190.0, + 293.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1182.0, + 1402.0, + 1182.0, + 1402.0, + 1218.0, + 293.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1212.0, + 1404.0, + 1212.0, + 1404.0, + 1249.0, + 294.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1240.0, + 919.0, + 1240.0, + 919.0, + 1280.0, + 293.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 1240.0, + 1407.0, + 1240.0, + 1407.0, + 1280.0, + 994.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1276.0, + 1402.0, + 1276.0, + 1402.0, + 1307.0, + 295.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1303.0, + 1406.0, + 1303.0, + 1406.0, + 1342.0, + 293.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1337.0, + 569.0, + 1337.0, + 569.0, + 1373.0, + 295.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 426.0, + 1402.0, + 426.0, + 1402.0, + 459.0, + 296.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 454.0, + 1405.0, + 454.0, + 1405.0, + 492.0, + 292.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 486.0, + 1406.0, + 486.0, + 1406.0, + 523.0, + 292.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 519.0, + 1404.0, + 519.0, + 1404.0, + 552.0, + 296.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 549.0, + 743.0, + 549.0, + 743.0, + 582.0, + 296.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 952.0, + 1405.0, + 952.0, + 1405.0, + 990.0, + 294.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 983.0, + 1405.0, + 983.0, + 1405.0, + 1021.0, + 293.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1014.0, + 406.0, + 1014.0, + 406.0, + 1049.0, + 294.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 230.0, + 1402.0, + 230.0, + 1402.0, + 265.0, + 397.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 263.0, + 1225.0, + 263.0, + 1225.0, + 295.0, + 396.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 305.0, + 1403.0, + 305.0, + 1403.0, + 342.0, + 372.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 336.0, + 1406.0, + 336.0, + 1406.0, + 374.0, + 395.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 364.0, + 1260.0, + 364.0, + 1260.0, + 405.0, + 393.0, + 405.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 931, + 1404, + 931, + 1404, + 1146, + 297, + 1146 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1161, + 1403, + 1161, + 1403, + 1345, + 298, + 1345 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 300, + 1360, + 1403, + 1360, + 1403, + 1514, + 300, + 1514 + ], + "score": 0.971 + }, + { + "category_id": 3, + "poly": [ + 300, + 224, + 1398, + 224, + 1398, + 466, + 300, + 466 + ], + "score": 0.964 + }, + { + "category_id": 4, + "poly": [ + 296, + 493, + 1405, + 493, + 1405, + 861, + 296, + 861 + ], + "score": 0.955 + }, + { + "category_id": 2, + "poly": [ + 297, + 1977, + 1404, + 1977, + 1404, + 2034, + 297, + 2034 + ], + "score": 0.927 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 815, + 76, + 815, + 104, + 298, + 104 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.693 + }, + { + "category_id": 1, + "poly": [ + 368, + 1543, + 1403, + 1543, + 1403, + 1933, + 368, + 1933 + ], + "score": 0.594 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.199 + }, + { + "category_id": 13, + "poly": [ + 562, + 586, + 617, + 586, + 617, + 615, + 562, + 615 + ], + "score": 0.86, + "latex": "9 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1351, + 619, + 1404, + 619, + 1404, + 647, + 1351, + 647 + ], + "score": 0.83, + "latex": "x =" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 354.0, + 230.0, + 354.0, + 447.0, + 297.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 226.0, + 513.0, + 226.0, + 513.0, + 282.0, + 394.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 230.0, + 632.0, + 230.0, + 632.0, + 447.0, + 576.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 226.0, + 800.0, + 226.0, + 800.0, + 282.0, + 669.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 229.0, + 909.0, + 229.0, + 909.0, + 451.0, + 854.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 226.0, + 1077.0, + 226.0, + 1077.0, + 282.0, + 948.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 227.0, + 1186.0, + 227.0, + 1186.0, + 450.0, + 1132.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 226.0, + 1357.0, + 226.0, + 1357.0, + 282.0, + 1225.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 270.0, + 437.0, + 270.0, + 437.0, + 302.0, + 392.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 271.0, + 715.0, + 271.0, + 715.0, + 301.0, + 671.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 273.0, + 991.0, + 273.0, + 991.0, + 300.0, + 950.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 271.0, + 1271.0, + 271.0, + 1271.0, + 301.0, + 1226.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 307.0, + 352.0, + 307.0, + 352.0, + 326.0, + 327.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 308.0, + 632.0, + 308.0, + 632.0, + 326.0, + 604.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 308.0, + 909.0, + 308.0, + 909.0, + 326.0, + 884.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 307.0, + 1186.0, + 307.0, + 1186.0, + 326.0, + 1161.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 363.0, + 354.0, + 363.0, + 354.0, + 381.0, + 327.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 363.0, + 632.0, + 363.0, + 632.0, + 381.0, + 604.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 363.0, + 908.0, + 363.0, + 908.0, + 382.0, + 883.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 363.0, + 1186.0, + 363.0, + 1186.0, + 381.0, + 1161.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 419.0, + 352.0, + 419.0, + 352.0, + 436.0, + 325.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 419.0, + 629.0, + 419.0, + 629.0, + 436.0, + 603.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 419.0, + 908.0, + 419.0, + 908.0, + 436.0, + 879.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 416.0, + 1187.0, + 416.0, + 1187.0, + 438.0, + 1157.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 428.0, + 376.0, + 428.0, + 376.0, + 447.0, + 351.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 428.0, + 425.0, + 428.0, + 425.0, + 446.0, + 396.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 428.0, + 470.0, + 428.0, + 470.0, + 446.0, + 443.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 427.0, + 518.0, + 427.0, + 518.0, + 448.0, + 488.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 428.0, + 563.0, + 428.0, + 563.0, + 447.0, + 537.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 428.0, + 654.0, + 428.0, + 654.0, + 447.0, + 628.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 428.0, + 702.0, + 428.0, + 702.0, + 446.0, + 674.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 428.0, + 749.0, + 428.0, + 749.0, + 446.0, + 720.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 428.0, + 794.0, + 428.0, + 794.0, + 447.0, + 768.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 427.0, + 843.0, + 427.0, + 843.0, + 448.0, + 813.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 428.0, + 933.0, + 428.0, + 933.0, + 447.0, + 906.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 428.0, + 979.0, + 428.0, + 979.0, + 446.0, + 953.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 428.0, + 1026.0, + 428.0, + 1026.0, + 446.0, + 999.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 428.0, + 1074.0, + 428.0, + 1074.0, + 446.0, + 1044.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 428.0, + 1119.0, + 428.0, + 1119.0, + 447.0, + 1094.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 427.0, + 1212.0, + 427.0, + 1212.0, + 448.0, + 1182.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 428.0, + 1258.0, + 428.0, + 1258.0, + 446.0, + 1231.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 428.0, + 1303.0, + 428.0, + 1303.0, + 446.0, + 1277.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1322.0, + 427.0, + 1352.0, + 427.0, + 1352.0, + 448.0, + 1322.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1371.0, + 428.0, + 1397.0, + 428.0, + 1397.0, + 447.0, + 1371.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 440.0, + 537.0, + 440.0, + 537.0, + 466.0, + 375.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 439.0, + 817.0, + 439.0, + 817.0, + 467.0, + 652.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 440.0, + 1094.0, + 440.0, + 1094.0, + 466.0, + 932.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 439.0, + 1372.0, + 439.0, + 1372.0, + 467.0, + 1207.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 493.0, + 1405.0, + 493.0, + 1405.0, + 531.0, + 295.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 525.0, + 1405.0, + 525.0, + 1405.0, + 559.0, + 295.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 556.0, + 1406.0, + 556.0, + 1406.0, + 589.0, + 295.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 586.0, + 561.0, + 586.0, + 561.0, + 620.0, + 294.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 586.0, + 1405.0, + 586.0, + 1405.0, + 620.0, + 618.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 615.0, + 1350.0, + 615.0, + 1350.0, + 653.0, + 294.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 645.0, + 1403.0, + 645.0, + 1403.0, + 682.0, + 294.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 676.0, + 1405.0, + 676.0, + 1405.0, + 712.0, + 294.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 707.0, + 1405.0, + 707.0, + 1405.0, + 742.0, + 295.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 734.0, + 1405.0, + 734.0, + 1405.0, + 776.0, + 293.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 768.0, + 1405.0, + 768.0, + 1405.0, + 803.0, + 295.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 799.0, + 1405.0, + 799.0, + 1405.0, + 833.0, + 294.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 827.0, + 595.0, + 827.0, + 595.0, + 867.0, + 294.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 1974.0, + 1405.0, + 1974.0, + 1405.0, + 2010.0, + 333.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 809.0, + 2003.0, + 809.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 931.0, + 1406.0, + 931.0, + 1406.0, + 966.0, + 295.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 961.0, + 1405.0, + 961.0, + 1405.0, + 998.0, + 294.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 990.0, + 1405.0, + 990.0, + 1405.0, + 1027.0, + 294.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1025.0, + 1402.0, + 1025.0, + 1402.0, + 1056.0, + 296.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1053.0, + 1404.0, + 1053.0, + 1404.0, + 1088.0, + 295.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1084.0, + 1406.0, + 1084.0, + 1406.0, + 1118.0, + 295.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1114.0, + 1280.0, + 1114.0, + 1280.0, + 1152.0, + 294.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1159.0, + 1405.0, + 1159.0, + 1405.0, + 1196.0, + 294.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1192.0, + 1405.0, + 1192.0, + 1405.0, + 1225.0, + 293.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1224.0, + 1404.0, + 1224.0, + 1404.0, + 1256.0, + 296.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1250.0, + 1405.0, + 1250.0, + 1405.0, + 1287.0, + 293.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1283.0, + 1405.0, + 1283.0, + 1405.0, + 1319.0, + 294.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1311.0, + 618.0, + 1311.0, + 618.0, + 1348.0, + 294.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1359.0, + 1404.0, + 1359.0, + 1404.0, + 1395.0, + 296.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1388.0, + 1405.0, + 1388.0, + 1405.0, + 1426.0, + 293.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1418.0, + 1405.0, + 1418.0, + 1405.0, + 1457.0, + 293.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1450.0, + 1405.0, + 1450.0, + 1405.0, + 1487.0, + 295.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1482.0, + 636.0, + 1482.0, + 636.0, + 1519.0, + 295.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 1544.0, + 1406.0, + 1544.0, + 1406.0, + 1583.0, + 367.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1577.0, + 1406.0, + 1577.0, + 1406.0, + 1613.0, + 394.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1606.0, + 1404.0, + 1606.0, + 1404.0, + 1643.0, + 392.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1637.0, + 1406.0, + 1637.0, + 1406.0, + 1673.0, + 392.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1667.0, + 1405.0, + 1667.0, + 1405.0, + 1705.0, + 393.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1700.0, + 604.0, + 1700.0, + 604.0, + 1731.0, + 396.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 1750.0, + 1406.0, + 1750.0, + 1406.0, + 1783.0, + 375.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1778.0, + 1406.0, + 1778.0, + 1406.0, + 1817.0, + 391.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1812.0, + 1405.0, + 1812.0, + 1405.0, + 1845.0, + 395.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1843.0, + 1406.0, + 1843.0, + 1406.0, + 1873.0, + 392.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1871.0, + 1406.0, + 1871.0, + 1406.0, + 1909.0, + 393.0, + 1909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1903.0, + 876.0, + 1903.0, + 876.0, + 1936.0, + 395.0, + 1936.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1271, + 1405, + 1271, + 1405, + 1401, + 296, + 1401 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1060, + 1404, + 1060, + 1404, + 1156, + 297, + 1156 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 652, + 1403, + 652, + 1403, + 777, + 299, + 777 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 305, + 382, + 1405, + 382, + 1405, + 474, + 305, + 474 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 302, + 1443, + 1393, + 1443, + 1393, + 1794, + 302, + 1794 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 369, + 229, + 1402, + 229, + 1402, + 353, + 369, + 353 + ], + "score": 0.971 + }, + { + "category_id": 4, + "poly": [ + 294, + 1819, + 1405, + 1819, + 1405, + 2007, + 294, + 2007 + ], + "score": 0.962 + }, + { + "category_id": 8, + "poly": [ + 582, + 1177, + 1112, + 1177, + 1112, + 1253, + 582, + 1253 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 295, + 922, + 1402, + 922, + 1402, + 987, + 295, + 987 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 297, + 488, + 1401, + 488, + 1401, + 553, + 297, + 553 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 717, + 1002, + 980, + 1002, + 980, + 1045, + 717, + 1045 + ], + "score": 0.939 + }, + { + "category_id": 0, + "poly": [ + 299, + 593, + 602, + 593, + 602, + 625, + 299, + 625 + ], + "score": 0.909 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 105, + 300, + 105 + ], + "score": 0.895 + }, + { + "category_id": 9, + "poly": [ + 1366, + 834, + 1400, + 834, + 1400, + 864, + 1366, + 864 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1008, + 1400, + 1008, + 1400, + 1038, + 1366, + 1038 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1200, + 1400, + 1200, + 1400, + 1230, + 1366, + 1230 + ], + "score": 0.88 + }, + { + "category_id": 8, + "poly": [ + 481, + 794, + 1207, + 794, + 1207, + 904, + 481, + 904 + ], + "score": 0.798 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.773 + }, + { + "category_id": 8, + "poly": [ + 481, + 795, + 1206, + 795, + 1206, + 905, + 481, + 905 + ], + "score": 0.336 + }, + { + "category_id": 14, + "poly": [ + 484, + 793, + 1212, + 793, + 1212, + 905, + 484, + 905 + ], + "score": 0.95, + "latex": "\\widehat { g } _ { \\theta } = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } \\left[ \\sum _ { ( s _ { t } , a _ { t } ) \\in \\tau } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\cdot ( Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) - b ( s _ { t } ) ) \\right]" + }, + { + "category_id": 14, + "poly": [ + 584, + 1177, + 1113, + 1177, + 1113, + 1254, + 584, + 1254 + ], + "score": 0.94, + "latex": "\\theta _ { t } = \\operatorname* { m i n } _ { \\theta } \\mathbb { E } \\left[ \\left( V _ { \\theta } ^ { \\pi } ( s _ { t } ) - ( V _ { \\theta _ { t - 1 } } ^ { \\pi } ( s _ { t } ) + A _ { t } ) \\right) ^ { 2 } \\right]" + }, + { + "category_id": 14, + "poly": [ + 717, + 1004, + 983, + 1004, + 983, + 1044, + 717, + 1044 + ], + "score": 0.93, + "latex": "V _ { \\pi _ { \\theta } } ( s _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ R _ { t } | s _ { t } ] \\ ." + }, + { + "category_id": 13, + "poly": [ + 371, + 1273, + 479, + 1273, + 479, + 1311, + 371, + 1311 + ], + "score": 0.93, + "latex": "V _ { \\theta _ { t - 1 } } ^ { \\pi } ( s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 463, + 954, + 530, + 954, + 530, + 988, + 463, + 988 + ], + "score": 0.92, + "latex": "V _ { \\pi } ( s )" + }, + { + "category_id": 13, + "poly": [ + 374, + 923, + 432, + 923, + 432, + 957, + 374, + 957 + ], + "score": 0.92, + "latex": "b ( s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 553, + 1911, + 743, + 1911, + 743, + 1946, + 553, + 1946 + ], + "score": 0.91, + "latex": "( V _ { o l d } \\bar { + } A _ { G A E } ) ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1166, + 1123, + 1206, + 1123, + 1206, + 1160, + 1166, + 1160 + ], + "score": 0.9, + "latex": "V _ { \\theta _ { t } } ^ { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 1033, + 1275, + 1066, + 1275, + 1066, + 1304, + 1033, + 1304 + ], + "score": 0.89, + "latex": "A _ { t }" + }, + { + "category_id": 13, + "poly": [ + 907, + 1308, + 941, + 1308, + 941, + 1338, + 907, + 1338 + ], + "score": 0.88, + "latex": "A _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1174, + 1943, + 1229, + 1943, + 1229, + 1972, + 1174, + 1972 + ], + "score": 0.88, + "latex": "5 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 834, + 929, + 859, + 929, + 859, + 954, + 834, + 954 + ], + "score": 0.84, + "latex": "s _ { t }" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1449.0, + 377.0, + 1449.0, + 377.0, + 1490.0, + 330.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 1456.0, + 815.0, + 1456.0, + 815.0, + 1492.0, + 745.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 1453.0, + 936.0, + 1453.0, + 936.0, + 1593.0, + 872.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1469.0, + 967.0, + 1469.0, + 967.0, + 1494.0, + 953.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 1501.0, + 380.0, + 1501.0, + 380.0, + 1619.0, + 302.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 1484.0, + 802.0, + 1484.0, + 802.0, + 1521.0, + 745.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 1488.0, + 923.0, + 1488.0, + 923.0, + 1636.0, + 850.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 1602.0, + 1362.0, + 1602.0, + 1362.0, + 1636.0, + 1294.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 1634.0, + 1347.0, + 1634.0, + 1347.0, + 1666.0, + 1294.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1692.0, + 416.0, + 1692.0, + 416.0, + 1715.0, + 396.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 1686.0, + 515.0, + 1686.0, + 515.0, + 1718.0, + 464.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 1686.0, + 598.0, + 1686.0, + 598.0, + 1718.0, + 547.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 1686.0, + 681.0, + 1686.0, + 681.0, + 1718.0, + 630.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 1686.0, + 766.0, + 1686.0, + 766.0, + 1718.0, + 714.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 1686.0, + 846.0, + 1686.0, + 846.0, + 1718.0, + 795.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1692.0, + 965.0, + 1692.0, + 965.0, + 1715.0, + 944.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 1688.0, + 1063.0, + 1688.0, + 1063.0, + 1718.0, + 1013.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 1686.0, + 1146.0, + 1686.0, + 1146.0, + 1718.0, + 1094.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1686.0, + 1230.0, + 1686.0, + 1230.0, + 1718.0, + 1179.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1686.0, + 1313.0, + 1686.0, + 1313.0, + 1718.0, + 1262.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 1685.0, + 1395.0, + 1685.0, + 1395.0, + 1718.0, + 1345.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 1709.0, + 673.0, + 1709.0, + 673.0, + 1746.0, + 539.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 1709.0, + 1222.0, + 1709.0, + 1222.0, + 1746.0, + 1086.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 1763.0, + 871.0, + 1763.0, + 871.0, + 1799.0, + 829.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1536.0, + 366.0, + 1536.0, + 366.0, + 1611.0, + 346.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1616.5, + 380.0, + 1616.5, + 380.0, + 1649.5, + 326.0, + 1649.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.25, + 1629.0, + 926.25, + 1629.0, + 926.25, + 1656.0, + 878.25, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1819.0, + 1406.0, + 1819.0, + 1406.0, + 1855.0, + 294.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1850.0, + 1405.0, + 1850.0, + 1405.0, + 1886.0, + 291.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1881.0, + 1406.0, + 1881.0, + 1406.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 552.0, + 1912.0, + 552.0, + 1948.0, + 294.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 1912.0, + 1406.0, + 1912.0, + 1406.0, + 1948.0, + 744.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1173.0, + 1941.0, + 1173.0, + 1976.0, + 294.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1976.0, + 1230.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 1345.0, + 1972.0, + 1345.0, + 2007.0, + 293.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 590.0, + 607.0, + 590.0, + 607.0, + 630.0, + 294.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1271.0, + 370.0, + 1271.0, + 370.0, + 1312.0, + 294.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 1271.0, + 1032.0, + 1271.0, + 1032.0, + 1312.0, + 480.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 1271.0, + 1406.0, + 1271.0, + 1406.0, + 1312.0, + 1067.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1307.0, + 906.0, + 1307.0, + 906.0, + 1344.0, + 293.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 1307.0, + 1405.0, + 1307.0, + 1405.0, + 1344.0, + 942.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1337.0, + 1406.0, + 1337.0, + 1406.0, + 1374.0, + 294.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1368.0, + 1371.0, + 1368.0, + 1371.0, + 1403.0, + 293.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1059.0, + 1405.0, + 1059.0, + 1405.0, + 1096.0, + 292.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1093.0, + 1405.0, + 1093.0, + 1405.0, + 1128.0, + 294.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1119.0, + 1165.0, + 1119.0, + 1165.0, + 1161.0, + 292.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 1119.0, + 1324.0, + 1119.0, + 1324.0, + 1161.0, + 1207.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 652.0, + 1405.0, + 652.0, + 1405.0, + 689.0, + 294.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 682.0, + 1405.0, + 682.0, + 1405.0, + 719.0, + 294.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 714.0, + 1404.0, + 714.0, + 1404.0, + 751.0, + 294.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 745.0, + 1400.0, + 745.0, + 1400.0, + 781.0, + 294.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 382.0, + 1404.0, + 382.0, + 1404.0, + 416.0, + 300.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 414.0, + 1405.0, + 414.0, + 1405.0, + 447.0, + 300.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 442.0, + 1221.0, + 442.0, + 1221.0, + 477.0, + 299.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 228.0, + 1405.0, + 228.0, + 1405.0, + 265.0, + 367.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 261.0, + 1404.0, + 261.0, + 1404.0, + 296.0, + 395.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 293.0, + 1405.0, + 293.0, + 1405.0, + 326.0, + 396.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 320.0, + 1379.0, + 320.0, + 1379.0, + 358.0, + 395.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 924.0, + 373.0, + 924.0, + 373.0, + 956.0, + 297.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 924.0, + 833.0, + 924.0, + 833.0, + 956.0, + 433.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 924.0, + 1404.0, + 924.0, + 1404.0, + 956.0, + 860.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 953.0, + 462.0, + 953.0, + 462.0, + 989.0, + 297.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 953.0, + 1384.0, + 953.0, + 1384.0, + 989.0, + 531.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 484.0, + 1403.0, + 484.0, + 1403.0, + 527.0, + 296.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 519.0, + 1195.0, + 519.0, + 1195.0, + 557.0, + 295.0, + 557.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1560, + 1404, + 1560, + 1404, + 1957, + 298, + 1957 + ], + "score": 0.983 + }, + { + "category_id": 3, + "poly": [ + 300, + 224, + 1399, + 224, + 1399, + 799, + 300, + 799 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1226, + 1402, + 1226, + 1402, + 1349, + 299, + 1349 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1365, + 1402, + 1365, + 1402, + 1489, + 298, + 1489 + ], + "score": 0.971 + }, + { + "category_id": 4, + "poly": [ + 296, + 825, + 1404, + 825, + 1404, + 1132, + 296, + 1132 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 299, + 1972, + 1399, + 1972, + 1399, + 2033, + 299, + 2033 + ], + "score": 0.941 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.705 + }, + { + "category_id": 13, + "poly": [ + 989, + 1288, + 1070, + 1288, + 1070, + 1321, + 989, + 1321 + ], + "score": 0.92, + "latex": "\\left( { { s _ { t } } , { a _ { t } } } \\right)" + }, + { + "category_id": 13, + "poly": [ + 696, + 1457, + 736, + 1457, + 736, + 1491, + 696, + 1491 + ], + "score": 0.89, + "latex": "V _ { \\theta } ^ { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 1260, + 1562, + 1300, + 1562, + 1300, + 1594, + 1260, + 1594 + ], + "score": 0.89, + "latex": "V _ { \\theta } ^ { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 1236, + 1288, + 1275, + 1288, + 1275, + 1322, + 1236, + 1322 + ], + "score": 0.89, + "latex": "V _ { \\theta } ^ { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 341, + 947, + 423, + 947, + 423, + 978, + 341, + 978 + ], + "score": 0.88, + "latex": "5 \\cdot 1 0 ^ { 6 }" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 241.0, + 351.0, + 241.0, + 351.0, + 427.0, + 299.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 225.0, + 558.0, + 225.0, + 558.0, + 282.0, + 356.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 227.0, + 936.0, + 227.0, + 936.0, + 444.0, + 662.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 227.0, + 1315.0, + 227.0, + 1315.0, + 340.0, + 1050.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 275.0, + 351.0, + 275.0, + 351.0, + 297.0, + 319.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 273.0, + 439.0, + 273.0, + 439.0, + 301.0, + 392.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 273.0, + 818.0, + 273.0, + 818.0, + 299.0, + 772.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 273.0, + 1112.0, + 273.0, + 1112.0, + 301.0, + 1076.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 273.0, + 1198.0, + 273.0, + 1198.0, + 301.0, + 1152.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 291.0, + 425.0, + 291.0, + 425.0, + 342.0, + 392.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 296.0, + 731.0, + 296.0, + 731.0, + 318.0, + 696.0, + 318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 289.0, + 804.0, + 289.0, + 804.0, + 341.0, + 768.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 293.0, + 1129.0, + 293.0, + 1129.0, + 442.0, + 1042.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 291.0, + 1183.0, + 291.0, + 1183.0, + 342.0, + 1150.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 309.0, + 350.0, + 309.0, + 350.0, + 328.0, + 320.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 317.0, + 1111.0, + 317.0, + 1111.0, + 340.0, + 1077.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 340.0, + 352.0, + 340.0, + 352.0, + 363.0, + 319.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 329.0, + 732.0, + 329.0, + 732.0, + 353.0, + 698.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 337.0, + 1111.0, + 337.0, + 1111.0, + 360.0, + 1076.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 375.0, + 350.0, + 375.0, + 350.0, + 394.0, + 320.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 404.0, + 352.0, + 404.0, + 352.0, + 429.0, + 319.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 430.0, + 370.0, + 430.0, + 370.0, + 446.0, + 355.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 427.0, + 434.0, + 427.0, + 434.0, + 448.0, + 396.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 428.0, + 484.0, + 428.0, + 484.0, + 446.0, + 449.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 427.0, + 535.0, + 427.0, + 535.0, + 448.0, + 497.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 427.0, + 586.0, + 427.0, + 586.0, + 448.0, + 550.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 427.0, + 641.0, + 427.0, + 641.0, + 448.0, + 598.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 430.0, + 750.0, + 430.0, + 750.0, + 446.0, + 733.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 427.0, + 813.0, + 427.0, + 813.0, + 448.0, + 776.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 428.0, + 862.0, + 428.0, + 862.0, + 446.0, + 828.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 428.0, + 913.0, + 428.0, + 913.0, + 446.0, + 878.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 427.0, + 966.0, + 427.0, + 966.0, + 448.0, + 927.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 427.0, + 1021.0, + 427.0, + 1021.0, + 448.0, + 978.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 430.0, + 1129.0, + 430.0, + 1129.0, + 446.0, + 1112.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 427.0, + 1192.0, + 427.0, + 1192.0, + 448.0, + 1154.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 428.0, + 1242.0, + 428.0, + 1242.0, + 446.0, + 1207.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 428.0, + 1293.0, + 428.0, + 1293.0, + 446.0, + 1258.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 427.0, + 1344.0, + 427.0, + 1344.0, + 448.0, + 1308.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1358.0, + 427.0, + 1399.0, + 427.0, + 1399.0, + 448.0, + 1358.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 438.0, + 574.0, + 438.0, + 574.0, + 466.0, + 411.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 438.0, + 953.0, + 438.0, + 953.0, + 466.0, + 790.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 440.0, + 1333.0, + 440.0, + 1333.0, + 467.0, + 1169.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 480.0, + 939.0, + 480.0, + 939.0, + 511.0, + 761.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 514.0, + 553.0, + 514.0, + 553.0, + 729.0, + 288.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 514.0, + 936.0, + 514.0, + 936.0, + 736.0, + 656.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 539.0, + 1108.0, + 539.0, + 1108.0, + 613.0, + 1057.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 513.0, + 1314.0, + 513.0, + 1314.0, + 570.0, + 1112.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 564.0, + 346.0, + 564.0, + 346.0, + 587.0, + 319.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 561.0, + 434.0, + 561.0, + 434.0, + 588.0, + 387.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 562.0, + 816.0, + 562.0, + 816.0, + 588.0, + 768.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 565.0, + 1109.0, + 565.0, + 1109.0, + 588.0, + 1076.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 561.0, + 1195.0, + 561.0, + 1195.0, + 588.0, + 1150.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 602.0, + 345.0, + 602.0, + 345.0, + 622.0, + 320.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 581.0, + 419.0, + 581.0, + 419.0, + 627.0, + 387.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 585.0, + 727.0, + 585.0, + 727.0, + 608.0, + 693.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 581.0, + 798.0, + 581.0, + 798.0, + 613.0, + 770.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 583.0, + 1118.0, + 583.0, + 1118.0, + 688.0, + 1038.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 579.0, + 1182.0, + 579.0, + 1182.0, + 629.0, + 1150.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 636.0, + 345.0, + 636.0, + 345.0, + 660.0, + 319.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 622.0, + 728.0, + 622.0, + 728.0, + 644.0, + 693.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 629.0, + 1109.0, + 629.0, + 1109.0, + 653.0, + 1076.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 663.0, + 1118.0, + 663.0, + 1118.0, + 724.0, + 1049.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 721.0, + 362.0, + 721.0, + 362.0, + 734.0, + 351.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 716.0, + 427.0, + 716.0, + 427.0, + 737.0, + 391.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 717.0, + 479.0, + 717.0, + 479.0, + 735.0, + 444.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 716.0, + 531.0, + 716.0, + 531.0, + 737.0, + 494.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 716.0, + 582.0, + 716.0, + 582.0, + 737.0, + 545.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 716.0, + 637.0, + 716.0, + 637.0, + 737.0, + 596.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 720.0, + 746.0, + 720.0, + 746.0, + 735.0, + 731.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 716.0, + 810.0, + 716.0, + 810.0, + 737.0, + 772.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 717.0, + 859.0, + 717.0, + 859.0, + 735.0, + 825.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 717.0, + 911.0, + 717.0, + 911.0, + 735.0, + 877.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 716.0, + 963.0, + 716.0, + 963.0, + 737.0, + 927.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 716.0, + 1018.0, + 716.0, + 1018.0, + 737.0, + 977.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 719.0, + 1127.0, + 719.0, + 1127.0, + 736.0, + 1111.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 716.0, + 1190.0, + 716.0, + 1190.0, + 737.0, + 1153.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 717.0, + 1239.0, + 717.0, + 1239.0, + 735.0, + 1205.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 717.0, + 1293.0, + 717.0, + 1293.0, + 735.0, + 1258.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 716.0, + 1344.0, + 716.0, + 1344.0, + 737.0, + 1308.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1358.0, + 716.0, + 1399.0, + 716.0, + 1399.0, + 737.0, + 1358.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 731.0, + 570.0, + 731.0, + 570.0, + 755.0, + 406.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 729.0, + 952.0, + 729.0, + 952.0, + 756.0, + 786.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 731.0, + 1332.0, + 731.0, + 1332.0, + 755.0, + 1168.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 767.0, + 928.0, + 767.0, + 928.0, + 804.0, + 771.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 604.0, + 800.0, + 604.0, + 800.0, + 625.0, + 769.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 826.0, + 1405.0, + 826.0, + 1405.0, + 861.0, + 294.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 858.0, + 1405.0, + 858.0, + 1405.0, + 893.0, + 294.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 886.0, + 1404.0, + 886.0, + 1404.0, + 923.0, + 293.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 916.0, + 1405.0, + 916.0, + 1405.0, + 954.0, + 293.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 947.0, + 340.0, + 947.0, + 340.0, + 984.0, + 291.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 947.0, + 1405.0, + 947.0, + 1405.0, + 984.0, + 424.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 979.0, + 1404.0, + 979.0, + 1404.0, + 1013.0, + 293.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1005.0, + 1406.0, + 1005.0, + 1406.0, + 1048.0, + 293.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1037.0, + 1406.0, + 1037.0, + 1406.0, + 1075.0, + 293.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1069.0, + 1405.0, + 1069.0, + 1405.0, + 1106.0, + 293.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1097.0, + 859.0, + 1097.0, + 859.0, + 1139.0, + 293.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1557.0, + 1259.0, + 1557.0, + 1259.0, + 1597.0, + 293.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1301.0, + 1557.0, + 1405.0, + 1557.0, + 1405.0, + 1597.0, + 1301.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1588.0, + 1405.0, + 1588.0, + 1405.0, + 1627.0, + 293.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1620.0, + 1405.0, + 1620.0, + 1405.0, + 1655.0, + 296.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1651.0, + 1405.0, + 1651.0, + 1405.0, + 1687.0, + 294.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1680.0, + 1408.0, + 1680.0, + 1408.0, + 1720.0, + 293.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1708.0, + 1406.0, + 1708.0, + 1406.0, + 1749.0, + 292.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1742.0, + 1404.0, + 1742.0, + 1404.0, + 1778.0, + 294.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1771.0, + 1404.0, + 1771.0, + 1404.0, + 1810.0, + 293.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1804.0, + 1404.0, + 1804.0, + 1404.0, + 1840.0, + 293.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1834.0, + 1404.0, + 1834.0, + 1404.0, + 1870.0, + 294.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1865.0, + 1404.0, + 1865.0, + 1404.0, + 1901.0, + 293.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1892.0, + 1404.0, + 1892.0, + 1404.0, + 1931.0, + 293.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1924.0, + 1336.0, + 1924.0, + 1336.0, + 1962.0, + 295.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1224.0, + 1406.0, + 1224.0, + 1406.0, + 1261.0, + 294.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1257.0, + 1404.0, + 1257.0, + 1404.0, + 1290.0, + 295.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1288.0, + 988.0, + 1288.0, + 988.0, + 1324.0, + 293.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 1288.0, + 1235.0, + 1288.0, + 1235.0, + 1324.0, + 1071.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 1288.0, + 1407.0, + 1288.0, + 1407.0, + 1324.0, + 1276.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1318.0, + 1407.0, + 1318.0, + 1407.0, + 1354.0, + 293.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1363.0, + 1406.0, + 1363.0, + 1406.0, + 1401.0, + 293.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1394.0, + 1406.0, + 1394.0, + 1406.0, + 1432.0, + 292.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1427.0, + 1404.0, + 1427.0, + 1404.0, + 1460.0, + 294.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1456.0, + 695.0, + 1456.0, + 695.0, + 1492.0, + 293.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 1456.0, + 1228.0, + 1456.0, + 1228.0, + 1492.0, + 737.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1970.0, + 1403.0, + 1970.0, + 1403.0, + 2005.0, + 295.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1088.0, + 2003.0, + 1088.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 394, + 1404, + 394, + 1404, + 700, + 298, + 700 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 741, + 1404, + 741, + 1404, + 954, + 298, + 954 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1078, + 1405, + 1078, + 1405, + 1355, + 297, + 1355 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 303, + 970, + 1399, + 970, + 1399, + 1062, + 303, + 1062 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 295, + 290, + 1400, + 290, + 1400, + 352, + 295, + 352 + ], + "score": 0.95 + }, + { + "category_id": 4, + "poly": [ + 296, + 1780, + 1404, + 1780, + 1404, + 2025, + 296, + 2025 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.881 + }, + { + "category_id": 3, + "poly": [ + 1062, + 1464, + 1328, + 1464, + 1328, + 1742, + 1062, + 1742 + ], + "score": 0.841 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.776 + }, + { + "category_id": 4, + "poly": [ + 1074, + 1414, + 1361, + 1414, + 1361, + 1442, + 1074, + 1442 + ], + "score": 0.762 + }, + { + "category_id": 4, + "poly": [ + 712, + 1413, + 988, + 1413, + 988, + 1442, + 712, + 1442 + ], + "score": 0.753 + }, + { + "category_id": 0, + "poly": [ + 304, + 232, + 890, + 232, + 890, + 260, + 304, + 260 + ], + "score": 0.61 + }, + { + "category_id": 4, + "poly": [ + 353, + 1413, + 614, + 1413, + 614, + 1442, + 353, + 1442 + ], + "score": 0.607 + }, + { + "category_id": 3, + "poly": [ + 363, + 1463, + 634, + 1463, + 634, + 1743, + 363, + 1743 + ], + "score": 0.54 + }, + { + "category_id": 3, + "poly": [ + 711, + 1463, + 983, + 1463, + 983, + 1741, + 711, + 1741 + ], + "score": 0.521 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1780.0, + 1404.0, + 1780.0, + 1404.0, + 1814.0, + 295.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1813.0, + 1406.0, + 1813.0, + 1406.0, + 1845.0, + 293.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1839.0, + 1405.0, + 1839.0, + 1405.0, + 1879.0, + 293.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1873.0, + 1404.0, + 1873.0, + 1404.0, + 1906.0, + 295.0, + 1906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1902.0, + 1404.0, + 1902.0, + 1404.0, + 1936.0, + 294.0, + 1936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1932.0, + 1405.0, + 1932.0, + 1405.0, + 1967.0, + 291.0, + 1967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1962.0, + 1404.0, + 1962.0, + 1404.0, + 1998.0, + 293.0, + 1998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1995.0, + 1402.0, + 1995.0, + 1402.0, + 2028.0, + 294.0, + 2028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 1469.0, + 1266.0, + 1469.0, + 1266.0, + 1485.0, + 1129.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 1520.0, + 1085.0, + 1520.0, + 1085.0, + 1534.0, + 1062.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 1543.0, + 1086.0, + 1543.0, + 1086.0, + 1554.0, + 1065.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 1562.0, + 1087.0, + 1562.0, + 1087.0, + 1575.0, + 1066.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 1584.0, + 1088.0, + 1584.0, + 1088.0, + 1594.0, + 1069.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 1616.0, + 1096.0, + 1616.0, + 1096.0, + 1631.0, + 1075.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 1632.0, + 1101.0, + 1632.0, + 1101.0, + 1646.0, + 1081.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 1646.0, + 1108.0, + 1646.0, + 1108.0, + 1662.0, + 1087.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 1717.0, + 1131.0, + 1717.0, + 1131.0, + 1728.0, + 1113.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 1726.0, + 1162.0, + 1726.0, + 1162.0, + 1741.0, + 1139.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 1722.0, + 1187.0, + 1722.0, + 1187.0, + 1736.0, + 1166.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 1704.0, + 1292.0, + 1704.0, + 1292.0, + 1731.0, + 1192.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 1698.0, + 1319.0, + 1698.0, + 1319.0, + 1712.0, + 1295.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 1528.0, + 1357.0, + 1528.0, + 1357.0, + 1664.0, + 1055.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 1662.5, + 1137.0, + 1662.5, + 1137.0, + 1720.5, + 1072.0, + 1720.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.75, + 1721.0, + 1274.75, + 1721.0, + 1274.75, + 1738.0, + 1188.75, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 1411.0, + 1363.0, + 1411.0, + 1363.0, + 1446.0, + 1072.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 1406.0, + 990.0, + 1406.0, + 990.0, + 1448.0, + 709.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 896.0, + 230.0, + 896.0, + 265.0, + 297.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1407.0, + 617.0, + 1407.0, + 617.0, + 1448.0, + 348.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 1466.0, + 565.0, + 1466.0, + 565.0, + 1488.0, + 441.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1500.0, + 395.0, + 1500.0, + 395.0, + 1517.0, + 364.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1516.0, + 389.0, + 1516.0, + 389.0, + 1531.0, + 366.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1532.0, + 391.0, + 1532.0, + 391.0, + 1546.0, + 368.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1546.0, + 392.0, + 1546.0, + 392.0, + 1559.0, + 369.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 1559.0, + 393.0, + 1559.0, + 393.0, + 1574.0, + 371.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 1573.0, + 395.0, + 1573.0, + 395.0, + 1588.0, + 372.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 1588.0, + 396.0, + 1588.0, + 396.0, + 1602.0, + 374.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 1616.0, + 401.0, + 1616.0, + 401.0, + 1631.0, + 381.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1629.0, + 409.0, + 1629.0, + 409.0, + 1716.0, + 360.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1662.0, + 419.0, + 1662.0, + 419.0, + 1680.0, + 396.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 1700.0, + 623.0, + 1700.0, + 623.0, + 1711.0, + 603.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 1704.0, + 602.0, + 1704.0, + 602.0, + 1715.0, + 571.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 1716.0, + 437.0, + 1716.0, + 437.0, + 1730.0, + 417.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 1728.0, + 465.0, + 1728.0, + 465.0, + 1740.0, + 443.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1723.0, + 493.0, + 1723.0, + 493.0, + 1737.0, + 470.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1718.0, + 521.0, + 1718.0, + 521.0, + 1732.0, + 497.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.75, + 1708.5, + 577.75, + 1708.5, + 577.75, + 1725.0, + 517.75, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.75, + 1722.5, + 579.75, + 1722.5, + 579.75, + 1738.0, + 495.75, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1468.0, + 914.0, + 1468.0, + 914.0, + 1487.0, + 786.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 1502.0, + 742.0, + 1502.0, + 742.0, + 1518.0, + 710.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 1517.0, + 738.0, + 1517.0, + 738.0, + 1535.0, + 712.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1535.0, + 738.0, + 1535.0, + 738.0, + 1550.0, + 715.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1551.0, + 742.0, + 1551.0, + 742.0, + 1568.0, + 715.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 1566.0, + 743.0, + 1566.0, + 743.0, + 1583.0, + 717.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1575.0, + 893.0, + 1575.0, + 893.0, + 1585.0, + 871.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 1582.0, + 742.0, + 1582.0, + 742.0, + 1597.0, + 719.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 1615.0, + 748.0, + 1615.0, + 748.0, + 1631.0, + 728.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 1620.0, + 761.0, + 1620.0, + 761.0, + 1686.0, + 722.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 1680.0, + 773.0, + 1680.0, + 773.0, + 1697.0, + 750.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1697.0, + 971.0, + 1697.0, + 971.0, + 1712.0, + 949.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 1702.0, + 945.0, + 1702.0, + 945.0, + 1717.0, + 924.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 1707.0, + 920.0, + 1707.0, + 920.0, + 1722.0, + 896.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 1715.0, + 784.0, + 1715.0, + 784.0, + 1730.0, + 764.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1716.0, + 867.0, + 1716.0, + 867.0, + 1731.0, + 845.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1722.0, + 839.0, + 1722.0, + 839.0, + 1736.0, + 817.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1727.0, + 814.0, + 1727.0, + 814.0, + 1741.0, + 792.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1645.0, + 759.0, + 1645.0, + 759.0, + 1663.5, + 740.0, + 1663.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 1682.0, + 765.0, + 1682.0, + 765.0, + 1710.0, + 743.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.25, + 1663.0, + 762.25, + 1663.0, + 762.25, + 1679.0, + 745.25, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1710.5, + 895.0, + 1710.5, + 895.0, + 1727.0, + 868.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1721.5, + 927.0, + 1721.5, + 927.0, + 1737.5, + 843.0, + 1737.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 393.0, + 1406.0, + 393.0, + 1406.0, + 429.0, + 296.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 426.0, + 1404.0, + 426.0, + 1404.0, + 458.0, + 296.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 454.0, + 1404.0, + 454.0, + 1404.0, + 490.0, + 293.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 486.0, + 1405.0, + 486.0, + 1405.0, + 520.0, + 293.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 511.0, + 1405.0, + 511.0, + 1405.0, + 555.0, + 292.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 546.0, + 1405.0, + 546.0, + 1405.0, + 584.0, + 292.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 575.0, + 1405.0, + 575.0, + 1405.0, + 613.0, + 292.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 604.0, + 1405.0, + 604.0, + 1405.0, + 646.0, + 292.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 640.0, + 1402.0, + 640.0, + 1402.0, + 671.0, + 296.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 669.0, + 1150.0, + 669.0, + 1150.0, + 704.0, + 293.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 740.0, + 1406.0, + 740.0, + 1406.0, + 778.0, + 294.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 772.0, + 1406.0, + 772.0, + 1406.0, + 806.0, + 296.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 803.0, + 1404.0, + 803.0, + 1404.0, + 838.0, + 294.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 833.0, + 1402.0, + 833.0, + 1402.0, + 867.0, + 294.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 861.0, + 1404.0, + 861.0, + 1404.0, + 898.0, + 293.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 893.0, + 1406.0, + 893.0, + 1406.0, + 931.0, + 293.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 927.0, + 571.0, + 927.0, + 571.0, + 958.0, + 294.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1076.0, + 1406.0, + 1076.0, + 1406.0, + 1117.0, + 295.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1109.0, + 1405.0, + 1109.0, + 1405.0, + 1146.0, + 295.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1138.0, + 1407.0, + 1138.0, + 1407.0, + 1176.0, + 292.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1169.0, + 1407.0, + 1169.0, + 1407.0, + 1206.0, + 294.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1202.0, + 1406.0, + 1202.0, + 1406.0, + 1235.0, + 296.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1230.0, + 1403.0, + 1230.0, + 1403.0, + 1266.0, + 292.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1262.0, + 1402.0, + 1262.0, + 1402.0, + 1295.0, + 295.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1292.0, + 1403.0, + 1292.0, + 1403.0, + 1325.0, + 296.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1323.0, + 1403.0, + 1323.0, + 1403.0, + 1359.0, + 294.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 969.0, + 1404.0, + 969.0, + 1404.0, + 1006.0, + 297.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1000.0, + 1405.0, + 1000.0, + 1405.0, + 1038.0, + 295.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1030.0, + 1039.0, + 1030.0, + 1039.0, + 1068.0, + 295.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 284.0, + 1404.0, + 284.0, + 1404.0, + 329.0, + 294.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 320.0, + 1380.0, + 320.0, + 1380.0, + 355.0, + 293.0, + 355.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1481, + 1403, + 1481, + 1403, + 1758, + 298, + 1758 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 300, + 1773, + 1403, + 1773, + 1403, + 1926, + 300, + 1926 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 353, + 221, + 1310, + 221, + 1310, + 668, + 353, + 668 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1174, + 1403, + 1174, + 1403, + 1329, + 299, + 1329 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1344, + 1403, + 1344, + 1403, + 1466, + 299, + 1466 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 300, + 1942, + 1401, + 1942, + 1401, + 2034, + 300, + 2034 + ], + "score": 0.97 + }, + { + "category_id": 4, + "poly": [ + 296, + 695, + 1405, + 695, + 1405, + 911, + 296, + 911 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 297, + 980, + 1400, + 980, + 1400, + 1041, + 297, + 1041 + ], + "score": 0.938 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.88 + }, + { + "category_id": 0, + "poly": [ + 299, + 1098, + 1077, + 1098, + 1077, + 1134, + 299, + 1134 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2111, + 842, + 2111 + ], + "score": 0.633 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2111, + 842, + 2111 + ], + "score": 0.256 + }, + { + "category_id": 13, + "poly": [ + 741, + 785, + 784, + 785, + 784, + 815, + 741, + 815 + ], + "score": 0.87, + "latex": "1 { \\dot { 0 } } ^ { 6 }" + }, + { + "category_id": 13, + "poly": [ + 1218, + 224, + 1279, + 224, + 1279, + 257, + 1218, + 257 + ], + "score": 0.76, + "latex": "( 1 0 ^ { 6 } )" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 224.0, + 812.0, + 224.0, + 812.0, + 258.0, + 473.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 222.0, + 1217.0, + 222.0, + 1217.0, + 257.0, + 947.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 222.0, + 1284.0, + 222.0, + 1284.0, + 257.0, + 1280.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 253.0, + 592.0, + 253.0, + 592.0, + 292.0, + 475.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 256.0, + 822.0, + 256.0, + 822.0, + 284.0, + 681.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 253.0, + 1066.0, + 253.0, + 1066.0, + 291.0, + 948.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 255.0, + 1297.0, + 255.0, + 1297.0, + 286.0, + 1155.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 301.0, + 398.0, + 301.0, + 398.0, + 390.0, + 356.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 410.0, + 479.0, + 410.0, + 479.0, + 453.0, + 455.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 403.0, + 951.0, + 403.0, + 951.0, + 452.0, + 927.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 479.0, + 394.0, + 479.0, + 394.0, + 592.0, + 353.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 536.0, + 678.0, + 536.0, + 678.0, + 545.0, + 670.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 546.0, + 680.0, + 546.0, + 680.0, + 555.0, + 671.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 553.0, + 682.0, + 553.0, + 682.0, + 573.0, + 667.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 581.0, + 684.0, + 581.0, + 684.0, + 587.0, + 678.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 603.0, + 479.0, + 603.0, + 479.0, + 626.0, + 463.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 650.0, + 728.0, + 650.0, + 728.0, + 658.0, + 717.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 647.0, + 746.0, + 647.0, + 746.0, + 656.0, + 732.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 647.0, + 1221.0, + 647.0, + 1221.0, + 656.0, + 1208.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 342.0, + 630.0, + 342.0, + 630.0, + 417.0, + 442.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 342.5, + 848.0, + 342.5, + 848.0, + 417.0, + 666.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 339.5, + 1106.0, + 339.5, + 1106.0, + 414.5, + 916.0, + 414.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 343.5, + 1329.0, + 343.5, + 1329.0, + 418.0, + 1136.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.25, + 440.0, + 703.25, + 440.0, + 703.25, + 449.0, + 693.25, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 532.0, + 634.0, + 532.0, + 634.0, + 608.5, + 444.0, + 608.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.75, + 531.0, + 1105.75, + 531.0, + 1105.75, + 607.0, + 914.75, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 527.5, + 1306.0, + 527.5, + 1306.0, + 602.0, + 1136.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.75, + 639.5, + 796.75, + 639.5, + 796.75, + 657.5, + 762.75, + 657.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 696.0, + 1405.0, + 696.0, + 1405.0, + 730.0, + 294.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 724.0, + 1405.0, + 724.0, + 1405.0, + 760.0, + 293.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 759.0, + 1403.0, + 759.0, + 1403.0, + 789.0, + 296.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 783.0, + 740.0, + 783.0, + 740.0, + 824.0, + 291.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 783.0, + 1406.0, + 783.0, + 1406.0, + 824.0, + 785.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 813.0, + 1405.0, + 813.0, + 1405.0, + 856.0, + 291.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 847.0, + 1406.0, + 847.0, + 1406.0, + 882.0, + 294.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 877.0, + 1322.0, + 877.0, + 1322.0, + 915.0, + 294.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1097.0, + 1082.0, + 1097.0, + 1082.0, + 1137.0, + 294.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1479.0, + 1407.0, + 1479.0, + 1407.0, + 1519.0, + 294.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1513.0, + 1405.0, + 1513.0, + 1405.0, + 1550.0, + 296.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1544.0, + 1405.0, + 1544.0, + 1405.0, + 1580.0, + 294.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1574.0, + 1405.0, + 1574.0, + 1405.0, + 1610.0, + 293.0, + 1610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1605.0, + 1407.0, + 1605.0, + 1407.0, + 1642.0, + 293.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1635.0, + 1404.0, + 1635.0, + 1404.0, + 1668.0, + 296.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1665.0, + 1405.0, + 1665.0, + 1405.0, + 1702.0, + 294.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1690.0, + 1407.0, + 1690.0, + 1407.0, + 1736.0, + 292.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1726.0, + 894.0, + 1726.0, + 894.0, + 1762.0, + 294.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1773.0, + 1403.0, + 1773.0, + 1403.0, + 1806.0, + 298.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1801.0, + 1403.0, + 1801.0, + 1403.0, + 1839.0, + 295.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1832.0, + 1407.0, + 1832.0, + 1407.0, + 1869.0, + 294.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1863.0, + 1407.0, + 1863.0, + 1407.0, + 1898.0, + 294.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1894.0, + 1206.0, + 1894.0, + 1206.0, + 1931.0, + 296.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1176.0, + 1404.0, + 1176.0, + 1404.0, + 1209.0, + 297.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1204.0, + 1405.0, + 1204.0, + 1405.0, + 1240.0, + 293.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1237.0, + 1404.0, + 1237.0, + 1404.0, + 1270.0, + 295.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1263.0, + 1404.0, + 1263.0, + 1404.0, + 1306.0, + 292.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1295.0, + 1087.0, + 1295.0, + 1087.0, + 1335.0, + 294.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1344.0, + 1405.0, + 1344.0, + 1405.0, + 1380.0, + 294.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1374.0, + 1405.0, + 1374.0, + 1405.0, + 1410.0, + 294.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1406.0, + 1405.0, + 1406.0, + 1405.0, + 1442.0, + 294.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1436.0, + 865.0, + 1436.0, + 865.0, + 1471.0, + 293.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1940.0, + 1403.0, + 1940.0, + 1403.0, + 1976.0, + 294.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 980.0, + 1402.0, + 980.0, + 1402.0, + 1013.0, + 295.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1012.0, + 1402.0, + 1012.0, + 1402.0, + 1043.0, + 296.0, + 1043.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1657, + 1403, + 1657, + 1403, + 1843, + 298, + 1843 + ], + "score": 0.976 + }, + { + "category_id": 3, + "poly": [ + 353, + 221, + 1299, + 221, + 1299, + 652, + 353, + 652 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1389, + 1404, + 1389, + 1404, + 1544, + 299, + 1544 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1250, + 1403, + 1250, + 1403, + 1374, + 298, + 1374 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 905, + 1404, + 905, + 1404, + 1029, + 299, + 1029 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1144, + 1403, + 1144, + 1403, + 1236, + 299, + 1236 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 300, + 1857, + 1402, + 1857, + 1402, + 1980, + 300, + 1980 + ], + "score": 0.973 + }, + { + "category_id": 4, + "poly": [ + 298, + 676, + 1402, + 676, + 1402, + 769, + 298, + 769 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 299, + 829, + 1398, + 829, + 1398, + 890, + 299, + 890 + ], + "score": 0.944 + }, + { + "category_id": 0, + "poly": [ + 301, + 1074, + 588, + 1074, + 588, + 1109, + 301, + 1109 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 300, + 1588, + 544, + 1588, + 544, + 1624, + 300, + 1624 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.787 + }, + { + "category_id": 13, + "poly": [ + 1218, + 224, + 1279, + 224, + 1279, + 257, + 1218, + 257 + ], + "score": 0.77, + "latex": "( 1 0 ^ { 6 } )" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 223.0, + 812.0, + 223.0, + 812.0, + 258.0, + 473.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 223.0, + 1217.0, + 223.0, + 1217.0, + 258.0, + 948.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 223.0, + 1284.0, + 223.0, + 1284.0, + 258.0, + 1280.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 252.0, + 591.0, + 252.0, + 591.0, + 292.0, + 475.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 255.0, + 823.0, + 255.0, + 823.0, + 287.0, + 680.0, + 287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 253.0, + 1066.0, + 253.0, + 1066.0, + 292.0, + 950.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 252.0, + 1298.0, + 252.0, + 1298.0, + 290.0, + 1153.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 290.0, + 398.0, + 290.0, + 398.0, + 380.0, + 357.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 317.0, + 782.0, + 317.0, + 782.0, + 323.0, + 777.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 406.0, + 963.0, + 406.0, + 963.0, + 426.0, + 946.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 459.0, + 396.0, + 459.0, + 396.0, + 574.0, + 352.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 531.0, + 679.0, + 531.0, + 679.0, + 561.0, + 667.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 578.0, + 961.0, + 578.0, + 961.0, + 613.0, + 943.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 625.0, + 772.0, + 625.0, + 772.0, + 632.0, + 758.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 629.0, + 738.0, + 629.0, + 738.0, + 637.0, + 725.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 629.0, + 1228.0, + 629.0, + 1228.0, + 638.0, + 1200.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.75, + 324.0, + 842.75, + 324.0, + 842.75, + 429.0, + 657.75, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 340.0, + 629.0, + 340.0, + 629.0, + 411.0, + 461.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 339.5, + 1105.0, + 339.5, + 1105.0, + 412.5, + 936.0, + 412.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 340.5, + 1313.0, + 340.5, + 1313.0, + 419.5, + 1143.0, + 419.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 518.0, + 633.0, + 518.0, + 633.0, + 591.0, + 456.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.75, + 517.0, + 1108.75, + 517.0, + 1108.75, + 590.0, + 928.75, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 675.0, + 1404.0, + 675.0, + 1404.0, + 709.0, + 294.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 707.0, + 1404.0, + 707.0, + 1404.0, + 741.0, + 296.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 738.0, + 1349.0, + 738.0, + 1349.0, + 771.0, + 294.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1071.0, + 592.0, + 1071.0, + 592.0, + 1115.0, + 292.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1582.0, + 550.0, + 1582.0, + 550.0, + 1633.0, + 290.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1658.0, + 1404.0, + 1658.0, + 1404.0, + 1694.0, + 294.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1686.0, + 1407.0, + 1686.0, + 1407.0, + 1723.0, + 292.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1718.0, + 1407.0, + 1718.0, + 1407.0, + 1754.0, + 294.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1750.0, + 1405.0, + 1750.0, + 1405.0, + 1784.0, + 293.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1779.0, + 1407.0, + 1779.0, + 1407.0, + 1815.0, + 294.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1809.0, + 987.0, + 1809.0, + 987.0, + 1847.0, + 293.0, + 1847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1391.0, + 1405.0, + 1391.0, + 1405.0, + 1424.0, + 295.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1417.0, + 1404.0, + 1417.0, + 1404.0, + 1457.0, + 293.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1451.0, + 1404.0, + 1451.0, + 1404.0, + 1486.0, + 294.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1478.0, + 1406.0, + 1478.0, + 1406.0, + 1518.0, + 293.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1510.0, + 1396.0, + 1510.0, + 1396.0, + 1549.0, + 293.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1249.0, + 1404.0, + 1249.0, + 1404.0, + 1286.0, + 293.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1280.0, + 1403.0, + 1280.0, + 1403.0, + 1317.0, + 293.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1313.0, + 1407.0, + 1313.0, + 1407.0, + 1349.0, + 294.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1345.0, + 775.0, + 1345.0, + 775.0, + 1377.0, + 296.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 905.0, + 1404.0, + 905.0, + 1404.0, + 941.0, + 295.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 937.0, + 1405.0, + 937.0, + 1405.0, + 973.0, + 294.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 969.0, + 1404.0, + 969.0, + 1404.0, + 1002.0, + 297.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 998.0, + 1382.0, + 998.0, + 1382.0, + 1034.0, + 294.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1144.0, + 1405.0, + 1144.0, + 1405.0, + 1178.0, + 295.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1173.0, + 1408.0, + 1173.0, + 1408.0, + 1209.0, + 293.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1204.0, + 1404.0, + 1204.0, + 1404.0, + 1242.0, + 295.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1855.0, + 1406.0, + 1855.0, + 1406.0, + 1893.0, + 294.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1887.0, + 1406.0, + 1887.0, + 1406.0, + 1922.0, + 294.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1918.0, + 1406.0, + 1918.0, + 1406.0, + 1955.0, + 295.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1948.0, + 1249.0, + 1948.0, + 1249.0, + 1982.0, + 294.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 831.0, + 1402.0, + 831.0, + 1402.0, + 862.0, + 298.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 861.0, + 1152.0, + 861.0, + 1152.0, + 892.0, + 296.0, + 892.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1486, + 1404, + 1486, + 1404, + 1611, + 296, + 1611 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 298, + 1064, + 1406, + 1064, + 1406, + 1158, + 298, + 1158 + ], + "score": 0.916 + }, + { + "category_id": 1, + "poly": [ + 297, + 1771, + 1399, + 1771, + 1399, + 1865, + 297, + 1865 + ], + "score": 0.915 + }, + { + "category_id": 1, + "poly": [ + 299, + 756, + 1405, + 756, + 1405, + 851, + 299, + 851 + ], + "score": 0.909 + }, + { + "category_id": 1, + "poly": [ + 297, + 925, + 1404, + 925, + 1404, + 989, + 297, + 989 + ], + "score": 0.907 + }, + { + "category_id": 1, + "poly": [ + 297, + 617, + 1401, + 617, + 1401, + 682, + 297, + 682 + ], + "score": 0.905 + }, + { + "category_id": 1, + "poly": [ + 296, + 1264, + 1402, + 1264, + 1402, + 1328, + 296, + 1328 + ], + "score": 0.901 + }, + { + "category_id": 1, + "poly": [ + 294, + 1348, + 1403, + 1348, + 1403, + 1413, + 294, + 1413 + ], + "score": 0.901 + }, + { + "category_id": 1, + "poly": [ + 298, + 532, + 1400, + 532, + 1400, + 598, + 298, + 598 + ], + "score": 0.901 + }, + { + "category_id": 1, + "poly": [ + 297, + 1179, + 1401, + 1179, + 1401, + 1243, + 297, + 1243 + ], + "score": 0.898 + }, + { + "category_id": 1, + "poly": [ + 297, + 1971, + 1398, + 1971, + 1398, + 2035, + 297, + 2035 + ], + "score": 0.898 + }, + { + "category_id": 1, + "poly": [ + 298, + 1010, + 1154, + 1010, + 1154, + 1044, + 298, + 1044 + ], + "score": 0.893 + }, + { + "category_id": 1, + "poly": [ + 296, + 871, + 970, + 871, + 970, + 906, + 296, + 906 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.887 + }, + { + "category_id": 1, + "poly": [ + 299, + 701, + 1326, + 701, + 1326, + 736, + 299, + 736 + ], + "score": 0.887 + }, + { + "category_id": 1, + "poly": [ + 296, + 1886, + 1399, + 1886, + 1399, + 1950, + 296, + 1950 + ], + "score": 0.886 + }, + { + "category_id": 1, + "poly": [ + 294, + 1686, + 1399, + 1686, + 1399, + 1751, + 294, + 1751 + ], + "score": 0.884 + }, + { + "category_id": 1, + "poly": [ + 296, + 448, + 1401, + 448, + 1401, + 512, + 296, + 512 + ], + "score": 0.873 + }, + { + "category_id": 1, + "poly": [ + 297, + 1433, + 1229, + 1433, + 1229, + 1468, + 297, + 1468 + ], + "score": 0.873 + }, + { + "category_id": 1, + "poly": [ + 297, + 279, + 1404, + 279, + 1404, + 343, + 297, + 343 + ], + "score": 0.87 + }, + { + "category_id": 1, + "poly": [ + 295, + 1632, + 1396, + 1632, + 1396, + 1668, + 295, + 1668 + ], + "score": 0.866 + }, + { + "category_id": 1, + "poly": [ + 296, + 362, + 1401, + 362, + 1401, + 427, + 296, + 427 + ], + "score": 0.84 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.781 + }, + { + "category_id": 0, + "poly": [ + 299, + 227, + 489, + 227, + 489, + 262, + 299, + 262 + ], + "score": 0.749 + }, + { + "category_id": 1, + "poly": [ + 299, + 227, + 489, + 227, + 489, + 262, + 299, + 262 + ], + "score": 0.121 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 490.0, + 230.0, + 490.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1486.0, + 1406.0, + 1486.0, + 1406.0, + 1524.0, + 294.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1516.0, + 1404.0, + 1516.0, + 1404.0, + 1554.0, + 318.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1548.0, + 1404.0, + 1548.0, + 1404.0, + 1585.0, + 322.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1580.0, + 646.0, + 1580.0, + 646.0, + 1613.0, + 323.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1061.0, + 1403.0, + 1061.0, + 1403.0, + 1101.0, + 292.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1096.0, + 1404.0, + 1096.0, + 1404.0, + 1130.0, + 324.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1123.0, + 583.0, + 1123.0, + 583.0, + 1160.0, + 324.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1772.0, + 1404.0, + 1772.0, + 1404.0, + 1806.0, + 295.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1801.0, + 1401.0, + 1801.0, + 1401.0, + 1837.0, + 322.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1835.0, + 590.0, + 1835.0, + 590.0, + 1864.0, + 323.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 755.0, + 1402.0, + 755.0, + 1402.0, + 791.0, + 294.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 785.0, + 1406.0, + 785.0, + 1406.0, + 824.0, + 321.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 818.0, + 730.0, + 818.0, + 730.0, + 852.0, + 323.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 920.0, + 1405.0, + 920.0, + 1405.0, + 965.0, + 292.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 957.0, + 503.0, + 957.0, + 503.0, + 989.0, + 324.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 617.0, + 1403.0, + 617.0, + 1403.0, + 653.0, + 295.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 650.0, + 1164.0, + 650.0, + 1164.0, + 683.0, + 322.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1264.0, + 1404.0, + 1264.0, + 1404.0, + 1300.0, + 295.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1296.0, + 1073.0, + 1296.0, + 1073.0, + 1329.0, + 323.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1346.0, + 1408.0, + 1346.0, + 1408.0, + 1386.0, + 294.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1379.0, + 624.0, + 1379.0, + 624.0, + 1414.0, + 322.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 531.0, + 1402.0, + 531.0, + 1402.0, + 570.0, + 293.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 562.0, + 1248.0, + 562.0, + 1248.0, + 599.0, + 323.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1174.0, + 1406.0, + 1174.0, + 1406.0, + 1219.0, + 292.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1211.0, + 691.0, + 1211.0, + 691.0, + 1242.0, + 322.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1968.0, + 1403.0, + 1968.0, + 1403.0, + 2009.0, + 293.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2003.0, + 1390.0, + 2003.0, + 1390.0, + 2036.0, + 322.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1008.0, + 1152.0, + 1008.0, + 1152.0, + 1048.0, + 294.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 871.0, + 972.0, + 871.0, + 972.0, + 907.0, + 297.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 700.0, + 1330.0, + 700.0, + 1330.0, + 741.0, + 292.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1886.0, + 1401.0, + 1886.0, + 1401.0, + 1922.0, + 294.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1918.0, + 503.0, + 1918.0, + 503.0, + 1950.0, + 322.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1689.0, + 1401.0, + 1689.0, + 1401.0, + 1721.0, + 297.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1717.0, + 996.0, + 1717.0, + 996.0, + 1752.0, + 322.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 447.0, + 1403.0, + 447.0, + 1403.0, + 485.0, + 294.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 477.0, + 1156.0, + 477.0, + 1156.0, + 515.0, + 321.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1429.0, + 1233.0, + 1429.0, + 1233.0, + 1473.0, + 293.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 279.0, + 1402.0, + 279.0, + 1402.0, + 315.0, + 295.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 310.0, + 863.0, + 310.0, + 863.0, + 344.0, + 322.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1629.0, + 1396.0, + 1629.0, + 1396.0, + 1671.0, + 291.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 361.0, + 1405.0, + 361.0, + 1405.0, + 401.0, + 295.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 396.0, + 675.0, + 396.0, + 675.0, + 427.0, + 322.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 490.0, + 230.0, + 490.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.863 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 866, + 2088, + 866, + 2113, + 836, + 2113 + ], + "score": 0.835 + }, + { + "category_id": 1, + "poly": [ + 293, + 872, + 1402, + 872, + 1402, + 935, + 293, + 935 + ], + "score": 0.659 + }, + { + "category_id": 1, + "poly": [ + 296, + 736, + 1398, + 736, + 1398, + 770, + 296, + 770 + ], + "score": 0.624 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1401, + 229, + 1401, + 323, + 298, + 323 + ], + "score": 0.621 + }, + { + "category_id": 1, + "poly": [ + 297, + 651, + 1396, + 651, + 1396, + 717, + 297, + 717 + ], + "score": 0.603 + }, + { + "category_id": 1, + "poly": [ + 296, + 787, + 1398, + 787, + 1398, + 853, + 296, + 853 + ], + "score": 0.599 + }, + { + "category_id": 1, + "poly": [ + 299, + 343, + 1399, + 343, + 1399, + 436, + 299, + 436 + ], + "score": 0.533 + }, + { + "category_id": 1, + "poly": [ + 298, + 539, + 1399, + 539, + 1399, + 634, + 298, + 634 + ], + "score": 0.493 + }, + { + "category_id": 1, + "poly": [ + 293, + 455, + 1401, + 455, + 1401, + 520, + 293, + 520 + ], + "score": 0.485 + }, + { + "category_id": 13, + "poly": [ + 675, + 687, + 760, + 687, + 760, + 714, + 675, + 714 + ], + "score": 0.62, + "latex": "= 4 6 7 s" + }, + { + "category_id": 13, + "poly": [ + 441, + 690, + 460, + 690, + 460, + 710, + 441, + 710 + ], + "score": 0.36, + "latex": "=" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 872.0, + 1404.0, + 872.0, + 1404.0, + 907.0, + 295.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 901.0, + 850.0, + 901.0, + 850.0, + 937.0, + 324.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 735.0, + 1403.0, + 735.0, + 1403.0, + 774.0, + 295.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 260.0, + 1405.0, + 260.0, + 1405.0, + 297.0, + 323.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 292.0, + 622.0, + 292.0, + 622.0, + 321.0, + 323.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 650.0, + 1401.0, + 650.0, + 1401.0, + 690.0, + 294.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 684.0, + 440.0, + 684.0, + 440.0, + 720.0, + 322.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 684.0, + 674.0, + 684.0, + 674.0, + 720.0, + 461.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 684.0, + 774.0, + 684.0, + 774.0, + 720.0, + 761.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 786.0, + 1402.0, + 786.0, + 1402.0, + 823.0, + 295.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 819.0, + 1230.0, + 819.0, + 1230.0, + 854.0, + 321.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 344.0, + 1403.0, + 344.0, + 1403.0, + 377.0, + 294.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 373.0, + 1405.0, + 373.0, + 1405.0, + 410.0, + 323.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 402.0, + 409.0, + 402.0, + 409.0, + 438.0, + 320.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 535.0, + 1404.0, + 535.0, + 1404.0, + 578.0, + 292.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 566.0, + 1405.0, + 566.0, + 1405.0, + 610.0, + 322.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 602.0, + 975.0, + 602.0, + 975.0, + 636.0, + 324.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 453.0, + 1404.0, + 453.0, + 1404.0, + 495.0, + 292.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 488.0, + 1057.0, + 488.0, + 1057.0, + 521.0, + 323.0, + 521.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 617, + 1405, + 617, + 1405, + 956, + 297, + 956 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 361, + 1404, + 361, + 1404, + 576, + 297, + 576 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1223, + 1403, + 1223, + 1403, + 1410, + 297, + 1410 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1504, + 1403, + 1504, + 1403, + 1597, + 298, + 1597 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1089, + 1405, + 1089, + 1405, + 1181, + 298, + 1181 + ], + "score": 0.975 + }, + { + "category_id": 8, + "poly": [ + 712, + 981, + 985, + 981, + 985, + 1068, + 712, + 1068 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 529, + 1811, + 1164, + 1811, + 1164, + 1923, + 529, + 1923 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 740, + 1432, + 957, + 1432, + 957, + 1480, + 740, + 1480 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 678, + 1998, + 1020, + 1998, + 1020, + 2040, + 678, + 2040 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 296, + 1727, + 1402, + 1727, + 1402, + 1791, + 296, + 1791 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 584, + 1619, + 1117, + 1619, + 1117, + 1661, + 584, + 1661 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 293, + 1680, + 1398, + 1680, + 1398, + 1714, + 293, + 1714 + ], + "score": 0.927 + }, + { + "category_id": 1, + "poly": [ + 297, + 1944, + 1248, + 1944, + 1248, + 1979, + 297, + 1979 + ], + "score": 0.926 + }, + { + "category_id": 0, + "poly": [ + 300, + 300, + 544, + 300, + 544, + 333, + 300, + 333 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.891 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1852, + 1401, + 1852, + 1401, + 1883, + 1366, + 1883 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 301, + 227, + 508, + 227, + 508, + 261, + 301, + 261 + ], + "score": 0.886 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1625, + 1400, + 1625, + 1400, + 1654, + 1366, + 1654 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1436, + 1400, + 1436, + 1400, + 1466, + 1366, + 1466 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1365, + 2004, + 1401, + 2004, + 1401, + 2034, + 1365, + 2034 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.871 + }, + { + "category_id": 14, + "poly": [ + 531, + 1811, + 1164, + 1811, + 1164, + 1923, + 531, + 1923 + ], + "score": 0.94, + "latex": "\\widehat { g _ { \\theta } } = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } \\left[ \\sum _ { \\left( s _ { t } , a _ { t } \\right) \\in \\tau } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\cdot Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) \\right] \\ ," + }, + { + "category_id": 14, + "poly": [ + 710, + 977, + 990, + 977, + 990, + 1070, + 710, + 1070 + ], + "score": 0.94, + "latex": "R _ { t } ( \\tau ) = \\sum _ { t ^ { \\prime } = t } ^ { \\infty } \\gamma ^ { ( t ^ { \\prime } - t ) } r _ { t ^ { \\prime } } \\ ." + }, + { + "category_id": 13, + "poly": [ + 372, + 1945, + 503, + 1945, + 503, + 1979, + 372, + 1979 + ], + "score": 0.93, + "latex": "Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 1054, + 619, + 1130, + 619, + 1130, + 653, + 1054, + 653 + ], + "score": 0.93, + "latex": "\\pi ( a | s )" + }, + { + "category_id": 13, + "poly": [ + 1210, + 650, + 1288, + 650, + 1288, + 683, + 1210, + 683 + ], + "score": 0.93, + "latex": "r ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 634, + 924, + 735, + 924, + 735, + 954, + 634, + 954 + ], + "score": 0.92, + "latex": "R : = R _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 346, + 1092, + 481, + 1092, + 481, + 1122, + 346, + 1122 + ], + "score": 0.92, + "latex": "0 < \\gamma < 1" + }, + { + "category_id": 13, + "poly": [ + 832, + 831, + 1057, + 831, + 1057, + 867, + 832, + 867 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { r ( \\tau ) = \\sum _ { t } r ( \\bar { s _ { t } } , a _ { t } ) } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 678, + 2000, + 1022, + 2000, + 1022, + 2037, + 678, + 2037 + ], + "score": 0.92, + "latex": "Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ R _ { t } | a _ { t } , s _ { t } ] ~ ." + }, + { + "category_id": 13, + "poly": [ + 298, + 710, + 597, + 710, + 597, + 744, + 298, + 744 + ], + "score": 0.91, + "latex": "\\{ ( a _ { t } , s _ { t } ) : t \\in \\{ 1 \\ldots \\overline { { T } } \\} \\}" + }, + { + "category_id": 13, + "poly": [ + 1344, + 771, + 1400, + 771, + 1400, + 804, + 1344, + 804 + ], + "score": 0.91, + "latex": "\\pi ( \\tau )" + }, + { + "category_id": 14, + "poly": [ + 582, + 1620, + 1115, + 1620, + 1115, + 1659, + 582, + 1659 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\nabla _ { \\theta } \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } [ r ( \\tau ) ] = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } [ \\nabla _ { \\theta } \\log ( \\pi _ { \\theta } ( \\tau ) ) r ( \\tau ) ] , } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 739, + 1432, + 956, + 1432, + 956, + 1482, + 739, + 1482 + ], + "score": 0.91, + "latex": "\\operatorname* { m a x } _ { \\theta } \\operatorname { \\mathbb { E } } _ { \\tau \\sim \\pi _ { \\theta } } [ r ( \\tau ) ] ~ ," + }, + { + "category_id": 13, + "poly": [ + 822, + 925, + 855, + 925, + 855, + 954, + 822, + 954 + ], + "score": 0.89, + "latex": "R _ { t }" + }, + { + "category_id": 13, + "poly": [ + 372, + 1512, + 456, + 1512, + 456, + 1536, + 372, + 1536 + ], + "score": 0.88, + "latex": "\\tau \\sim \\pi _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1330, + 424, + 1354, + 424, + 1354, + 451, + 1330, + 451 + ], + "score": 0.85, + "latex": "s ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 684, + 1292, + 715, + 1292, + 715, + 1317, + 684, + 1317 + ], + "score": 0.84, + "latex": "\\pi _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 954, + 869, + 980, + 869, + 980, + 894, + 954, + 894 + ], + "score": 0.84, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 906, + 745, + 934, + 745, + 934, + 772, + 906, + 772 + ], + "score": 0.84, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 666, + 837, + 694, + 837, + 694, + 863, + 666, + 863 + ], + "score": 0.83, + "latex": "s _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1348, + 683, + 1403, + 683, + 1403, + 711, + 1348, + 711 + ], + "score": 0.83, + "latex": "\\tau =" + }, + { + "category_id": 13, + "poly": [ + 1214, + 1952, + 1239, + 1952, + 1239, + 1976, + 1214, + 1976 + ], + "score": 0.83, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1341, + 1291, + 1372, + 1291, + 1372, + 1317, + 1341, + 1317 + ], + "score": 0.82, + "latex": "\\pi _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1065, + 1958, + 1090, + 1958, + 1090, + 1976, + 1065, + 1976 + ], + "score": 0.82, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1540, + 328, + 1540, + 328, + 1566, + 297, + 1566 + ], + "score": 0.81, + "latex": "\\pi _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 674, + 1349, + 690, + 1349, + 690, + 1375, + 674, + 1375 + ], + "score": 0.8, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 1060, + 867, + 1089, + 867, + 1089, + 894, + 1060, + 894 + ], + "score": 0.79, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 380, + 1762, + 394, + 1762, + 394, + 1786, + 380, + 1786 + ], + "score": 0.78, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1010, + 746, + 1038, + 746, + 1038, + 772, + 1010, + 772 + ], + "score": 0.77, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 872, + 775, + 886, + 775, + 886, + 799, + 872, + 799 + ], + "score": 0.76, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1363, + 1688, + 1392, + 1688, + 1392, + 1712, + 1363, + 1712 + ], + "score": 0.76, + "latex": "\\pi _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1226, + 866, + 1241, + 866, + 1241, + 890, + 1226, + 890 + ], + "score": 0.75, + "latex": "t { \\cdot }" + }, + { + "category_id": 13, + "poly": [ + 749, + 808, + 769, + 808, + 769, + 830, + 749, + 830 + ], + "score": 0.75, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 692, + 520, + 714, + 520, + 714, + 543, + 692, + 543 + ], + "score": 0.75, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1013, + 1288, + 1029, + 1288, + 1029, + 1314, + 1013, + 1314 + ], + "score": 0.75, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 792, + 1318, + 809, + 1318, + 809, + 1344, + 792, + 1344 + ], + "score": 0.74, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 863, + 626, + 883, + 626, + 883, + 648, + 863, + 648 + ], + "score": 0.74, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1239, + 458, + 1259, + 458, + 1259, + 481, + 1239, + 481 + ], + "score": 0.73, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1219, + 808, + 1241, + 808, + 1241, + 831, + 1219, + 831 + ], + "score": 0.72, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 869, + 687, + 887, + 687, + 887, + 708, + 869, + 708 + ], + "score": 0.71, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 910, + 429, + 929, + 429, + 929, + 451, + 910, + 451 + ], + "score": 0.71, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 504, + 459, + 521, + 459, + 521, + 481, + 504, + 481 + ], + "score": 0.7, + "latex": "r" + }, + { + "category_id": 13, + "poly": [ + 1087, + 656, + 1103, + 656, + 1103, + 677, + 1087, + 677 + ], + "score": 0.69, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 564, + 899, + 583, + 899, + 583, + 921, + 564, + 921 + ], + "score": 0.67, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1165, + 687, + 1181, + 687, + 1181, + 708, + 1165, + 708 + ], + "score": 0.65, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 647, + 656, + 665, + 656, + 665, + 677, + 647, + 677 + ], + "score": 0.62, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 737, + 429, + 754, + 429, + 754, + 452, + 737, + 452 + ], + "score": 0.56, + "latex": "s" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 298.0, + 547.0, + 298.0, + 547.0, + 336.0, + 295.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 221.0, + 514.0, + 221.0, + 514.0, + 270.0, + 294.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 619.0, + 862.0, + 619.0, + 862.0, + 654.0, + 294.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 619.0, + 1053.0, + 619.0, + 1053.0, + 654.0, + 884.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 619.0, + 1405.0, + 619.0, + 1405.0, + 654.0, + 1131.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 649.0, + 646.0, + 649.0, + 646.0, + 684.0, + 294.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 649.0, + 1086.0, + 649.0, + 1086.0, + 684.0, + 666.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 649.0, + 1209.0, + 649.0, + 1209.0, + 684.0, + 1104.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1289.0, + 649.0, + 1405.0, + 649.0, + 1405.0, + 684.0, + 1289.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 680.0, + 868.0, + 680.0, + 868.0, + 715.0, + 295.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 680.0, + 1164.0, + 680.0, + 1164.0, + 715.0, + 888.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 680.0, + 1347.0, + 680.0, + 1347.0, + 715.0, + 1182.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 709.0, + 1407.0, + 709.0, + 1407.0, + 748.0, + 598.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 742.0, + 905.0, + 742.0, + 905.0, + 777.0, + 294.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 742.0, + 1009.0, + 742.0, + 1009.0, + 777.0, + 935.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 742.0, + 1405.0, + 742.0, + 1405.0, + 777.0, + 1039.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 772.0, + 871.0, + 772.0, + 871.0, + 806.0, + 295.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 772.0, + 1343.0, + 772.0, + 1343.0, + 806.0, + 887.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 803.0, + 748.0, + 803.0, + 748.0, + 838.0, + 294.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 803.0, + 1218.0, + 803.0, + 1218.0, + 838.0, + 770.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 803.0, + 1405.0, + 803.0, + 1405.0, + 838.0, + 1242.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 831.0, + 665.0, + 831.0, + 665.0, + 870.0, + 294.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 831.0, + 831.0, + 831.0, + 831.0, + 870.0, + 695.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 831.0, + 1408.0, + 831.0, + 1408.0, + 870.0, + 1058.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 860.0, + 953.0, + 860.0, + 953.0, + 900.0, + 292.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 860.0, + 1059.0, + 860.0, + 1059.0, + 900.0, + 981.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 860.0, + 1225.0, + 860.0, + 1225.0, + 900.0, + 1090.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 860.0, + 1406.0, + 860.0, + 1406.0, + 900.0, + 1242.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 894.0, + 563.0, + 894.0, + 563.0, + 927.0, + 294.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 894.0, + 1407.0, + 894.0, + 1407.0, + 927.0, + 584.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 925.0, + 633.0, + 925.0, + 633.0, + 956.0, + 295.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 925.0, + 821.0, + 925.0, + 821.0, + 956.0, + 736.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 925.0, + 1003.0, + 925.0, + 1003.0, + 956.0, + 856.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 361.0, + 1405.0, + 361.0, + 1405.0, + 400.0, + 294.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 392.0, + 1405.0, + 392.0, + 1405.0, + 428.0, + 292.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 423.0, + 736.0, + 423.0, + 736.0, + 457.0, + 294.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 423.0, + 909.0, + 423.0, + 909.0, + 457.0, + 755.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 423.0, + 1329.0, + 423.0, + 1329.0, + 457.0, + 930.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 423.0, + 1405.0, + 423.0, + 1405.0, + 457.0, + 1355.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 451.0, + 503.0, + 451.0, + 503.0, + 491.0, + 290.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 451.0, + 1238.0, + 451.0, + 1238.0, + 491.0, + 522.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 451.0, + 1406.0, + 451.0, + 1406.0, + 491.0, + 1260.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 483.0, + 1406.0, + 483.0, + 1406.0, + 518.0, + 294.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 515.0, + 691.0, + 515.0, + 691.0, + 546.0, + 295.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 515.0, + 1402.0, + 515.0, + 1402.0, + 546.0, + 715.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 544.0, + 1337.0, + 544.0, + 1337.0, + 580.0, + 292.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1224.0, + 1405.0, + 1224.0, + 1405.0, + 1260.0, + 293.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1256.0, + 1405.0, + 1256.0, + 1405.0, + 1291.0, + 293.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1288.0, + 683.0, + 1288.0, + 683.0, + 1320.0, + 295.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1288.0, + 1012.0, + 1288.0, + 1012.0, + 1320.0, + 716.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 1288.0, + 1340.0, + 1288.0, + 1340.0, + 1320.0, + 1030.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 1288.0, + 1404.0, + 1288.0, + 1404.0, + 1320.0, + 1373.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1315.0, + 791.0, + 1315.0, + 791.0, + 1352.0, + 292.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 1315.0, + 1406.0, + 1315.0, + 1406.0, + 1352.0, + 810.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1347.0, + 673.0, + 1347.0, + 673.0, + 1383.0, + 293.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 1347.0, + 1405.0, + 1347.0, + 1405.0, + 1383.0, + 691.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1378.0, + 476.0, + 1378.0, + 476.0, + 1411.0, + 295.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1502.0, + 371.0, + 1502.0, + 371.0, + 1542.0, + 293.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 1502.0, + 1404.0, + 1502.0, + 1404.0, + 1542.0, + 457.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1535.0, + 296.0, + 1535.0, + 296.0, + 1570.0, + 292.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1535.0, + 1405.0, + 1535.0, + 1405.0, + 1570.0, + 329.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1566.0, + 779.0, + 1566.0, + 779.0, + 1600.0, + 294.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1089.0, + 345.0, + 1089.0, + 345.0, + 1122.0, + 293.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1089.0, + 1406.0, + 1089.0, + 1406.0, + 1122.0, + 482.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1121.0, + 1405.0, + 1121.0, + 1405.0, + 1154.0, + 293.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1148.0, + 384.0, + 1148.0, + 384.0, + 1184.0, + 291.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1724.0, + 1403.0, + 1724.0, + 1403.0, + 1764.0, + 295.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1758.0, + 379.0, + 1758.0, + 379.0, + 1795.0, + 293.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1758.0, + 1357.0, + 1758.0, + 1357.0, + 1795.0, + 395.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1675.0, + 1362.0, + 1675.0, + 1362.0, + 1722.0, + 292.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 1675.0, + 1404.0, + 1675.0, + 1404.0, + 1722.0, + 1393.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1938.0, + 371.0, + 1938.0, + 371.0, + 1987.0, + 293.0, + 1987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 1938.0, + 1064.0, + 1938.0, + 1064.0, + 1987.0, + 504.0, + 1987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1938.0, + 1213.0, + 1938.0, + 1213.0, + 1987.0, + 1091.0, + 1987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 1938.0, + 1254.0, + 1938.0, + 1254.0, + 1987.0, + 1240.0, + 1987.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 989, + 1404, + 989, + 1404, + 1145, + 297, + 1145 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1521, + 1405, + 1521, + 1405, + 1708, + 297, + 1708 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 227, + 1404, + 227, + 1404, + 385, + 298, + 385 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1721, + 1403, + 1721, + 1403, + 1815, + 299, + 1815 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 866, + 1403, + 866, + 1403, + 959, + 298, + 959 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1942, + 1403, + 1942, + 1403, + 2035, + 299, + 2035 + ], + "score": 0.974 + }, + { + "category_id": 8, + "poly": [ + 604, + 1821, + 1091, + 1821, + 1091, + 1937, + 604, + 1937 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 477, + 392, + 1216, + 392, + 1216, + 499, + 477, + 499 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 735, + 1380, + 963, + 1380, + 963, + 1455, + 735, + 1455 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 532, + 674, + 1160, + 674, + 1160, + 782, + 532, + 782 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 642, + 823, + 1059, + 823, + 1059, + 862, + 642, + 862 + ], + "score": 0.947 + }, + { + "category_id": 8, + "poly": [ + 507, + 1151, + 1194, + 1151, + 1194, + 1226, + 507, + 1226 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 624, + 1270, + 1072, + 1270, + 1072, + 1345, + 624, + 1345 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 299, + 1459, + 597, + 1459, + 597, + 1492, + 299, + 1492 + ], + "score": 0.937 + }, + { + "category_id": 8, + "poly": [ + 718, + 590, + 979, + 590, + 979, + 629, + 718, + 629 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 298, + 1350, + 370, + 1350, + 370, + 1379, + 298, + 1379 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 298, + 786, + 370, + 786, + 370, + 816, + 298, + 816 + ], + "score": 0.929 + }, + { + "category_id": 1, + "poly": [ + 297, + 552, + 1032, + 552, + 1032, + 585, + 297, + 585 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 299, + 506, + 816, + 506, + 816, + 539, + 299, + 539 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 297, + 634, + 1396, + 634, + 1396, + 667, + 297, + 667 + ], + "score": 0.919 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 817, + 74, + 817, + 105, + 299, + 105 + ], + "score": 0.914 + }, + { + "category_id": 1, + "poly": [ + 291, + 1232, + 1280, + 1232, + 1280, + 1263, + 291, + 1263 + ], + "score": 0.914 + }, + { + "category_id": 9, + "poly": [ + 1352, + 826, + 1400, + 826, + 1400, + 857, + 1352, + 857 + ], + "score": 0.901 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1401, + 1400, + 1401, + 1400, + 1433, + 1352, + 1433 + ], + "score": 0.901 + }, + { + "category_id": 9, + "poly": [ + 1352, + 711, + 1400, + 711, + 1400, + 742, + 1352, + 742 + ], + "score": 0.9 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1291, + 1400, + 1291, + 1400, + 1321, + 1352, + 1321 + ], + "score": 0.898 + }, + { + "category_id": 9, + "poly": [ + 1352, + 593, + 1400, + 593, + 1400, + 624, + 1352, + 624 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1902, + 1400, + 1902, + 1400, + 1932, + 1352, + 1932 + ], + "score": 0.892 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1172, + 1400, + 1172, + 1400, + 1203, + 1351, + 1203 + ], + "score": 0.879 + }, + { + "category_id": 9, + "poly": [ + 1365, + 430, + 1401, + 430, + 1401, + 460, + 1365, + 460 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.872 + }, + { + "category_id": 14, + "poly": [ + 734, + 1379, + 964, + 1379, + 964, + 1455, + 734, + 1455 + ], + "score": 0.96, + "latex": "\\widehat { A } _ { \\pi } = \\frac { A _ { \\pi } - \\mu ( A _ { \\pi } ) } { \\sigma ( A _ { \\pi } ) }" + }, + { + "category_id": 14, + "poly": [ + 479, + 391, + 1218, + 391, + 1218, + 501, + 479, + 501 + ], + "score": 0.94, + "latex": "\\widehat { g _ { \\theta } } = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } \\left[ \\sum _ { ( s _ { t } , a _ { t } ) \\in \\tau } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\cdot ( Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) - b ( s _ { t } ) ) \\right] ," + }, + { + "category_id": 14, + "poly": [ + 624, + 1268, + 1073, + 1268, + 1073, + 1345, + 624, + 1345 + ], + "score": 0.94, + "latex": "\\operatorname* { m a x } _ { \\theta } \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi } \\left[ \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi ( a _ { t } | s _ { t } ) } \\widehat { A } _ { \\pi } ( s _ { t } , a _ { t } ) \\right]" + }, + { + "category_id": 14, + "poly": [ + 534, + 670, + 1159, + 670, + 1159, + 782, + 534, + 782 + ], + "score": 0.93, + "latex": "\\widehat { g } _ { \\theta } = \\mathbb { E } _ { \\tau \\sim \\pi _ { \\theta } } \\left[ \\sum _ { \\left( s _ { t } , a _ { t } \\right) \\in \\tau } \\nabla _ { \\theta } \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) \\cdot A _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) \\right] ," + }, + { + "category_id": 14, + "poly": [ + 716, + 590, + 982, + 590, + 982, + 629, + 716, + 629 + ], + "score": 0.93, + "latex": "V _ { \\pi _ { \\theta } } ( s _ { t } ) = \\mathbb { E } _ { \\pi _ { \\theta } } [ R _ { t } | s _ { t } ] ." + }, + { + "category_id": 14, + "poly": [ + 507, + 1150, + 1192, + 1150, + 1192, + 1227, + 507, + 1227 + ], + "score": 0.93, + "latex": "\\operatorname* { m a x } _ { \\theta } \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi } \\left[ \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi ( a _ { t } | s _ { t } ) } A _ { \\pi } ( s _ { t } , a _ { t } ) \\right] \\qquad \\Big ( = \\mathbb { E } _ { \\pi _ { \\theta } } \\left[ A _ { \\pi } \\right] \\Big ) ," + }, + { + "category_id": 14, + "poly": [ + 640, + 823, + 1060, + 823, + 1060, + 862, + 640, + 862 + ], + "score": 0.92, + "latex": "A _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) = Q _ { \\pi _ { \\theta } } ( s _ { t } , a _ { t } ) - V _ { \\pi _ { \\theta } } ( s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 372, + 506, + 415, + 506, + 415, + 540, + 372, + 540 + ], + "score": 0.9, + "latex": "b ( \\cdot )" + }, + { + "category_id": 14, + "poly": [ + 603, + 1820, + 1096, + 1820, + 1096, + 1940, + 603, + 1940 + ], + "score": 0.9, + "latex": "\\begin{array} { r l } { \\displaystyle \\operatorname* { m a x } _ { \\theta } } & { \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi } \\left[ \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi ( a _ { t } | s _ { t } ) } \\widehat { A } _ { \\pi } ( s _ { t } , a _ { t } ) \\right] } \\\\ { \\mathrm { s . t . } } & { D _ { K L } ( \\pi _ { \\theta } ( \\cdot \\mid s ) | | \\pi ( \\cdot \\mid s ) ) \\leq \\delta , \\quad \\forall s ~ . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 571, + 353, + 619, + 353, + 619, + 386, + 571, + 386 + ], + "score": 0.9, + "latex": "Q _ { \\pi _ { \\theta } }" + }, + { + "category_id": 13, + "poly": [ + 1296, + 867, + 1339, + 867, + 1339, + 900, + 1296, + 900 + ], + "score": 0.9, + "latex": "V _ { \\pi _ { \\theta } }" + }, + { + "category_id": 13, + "poly": [ + 886, + 873, + 913, + 873, + 913, + 898, + 886, + 898 + ], + "score": 0.86, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 719, + 1024, + 735, + 1024, + 735, + 1050, + 719, + 1050 + ], + "score": 0.77, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 344, + 1465, + 365, + 1465, + 365, + 1487, + 344, + 1487 + ], + "score": 0.74, + "latex": "\\pi" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 990.0, + 1405.0, + 990.0, + 1405.0, + 1027.0, + 295.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1022.0, + 718.0, + 1022.0, + 718.0, + 1055.0, + 296.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1022.0, + 1401.0, + 1022.0, + 1401.0, + 1055.0, + 736.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1054.0, + 1402.0, + 1054.0, + 1402.0, + 1087.0, + 295.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1084.0, + 1402.0, + 1084.0, + 1402.0, + 1118.0, + 295.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1114.0, + 948.0, + 1114.0, + 948.0, + 1150.0, + 293.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1521.0, + 1407.0, + 1521.0, + 1407.0, + 1559.0, + 295.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1553.0, + 1405.0, + 1553.0, + 1405.0, + 1589.0, + 295.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1585.0, + 1406.0, + 1585.0, + 1406.0, + 1620.0, + 294.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1615.0, + 1406.0, + 1615.0, + 1406.0, + 1650.0, + 294.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1643.0, + 1407.0, + 1643.0, + 1407.0, + 1680.0, + 291.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1676.0, + 1402.0, + 1676.0, + 1402.0, + 1708.0, + 296.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 267.0, + 296.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 258.0, + 1404.0, + 258.0, + 1404.0, + 299.0, + 293.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 290.0, + 1406.0, + 290.0, + 1406.0, + 327.0, + 294.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 321.0, + 1406.0, + 321.0, + 1406.0, + 357.0, + 292.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 352.0, + 570.0, + 352.0, + 570.0, + 389.0, + 294.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 352.0, + 1166.0, + 352.0, + 1166.0, + 389.0, + 620.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1719.0, + 1404.0, + 1719.0, + 1404.0, + 1758.0, + 293.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1751.0, + 1403.0, + 1751.0, + 1403.0, + 1786.0, + 294.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1784.0, + 629.0, + 1784.0, + 629.0, + 1818.0, + 295.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 863.0, + 885.0, + 863.0, + 885.0, + 903.0, + 292.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 863.0, + 1295.0, + 863.0, + 1295.0, + 903.0, + 914.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1340.0, + 863.0, + 1405.0, + 863.0, + 1405.0, + 903.0, + 1340.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 897.0, + 1404.0, + 897.0, + 1404.0, + 931.0, + 294.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 927.0, + 626.0, + 927.0, + 626.0, + 961.0, + 294.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1999.0, + 1329.0, + 1999.0, + 1329.0, + 2040.0, + 294.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1455.0, + 343.0, + 1455.0, + 343.0, + 1497.0, + 294.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1455.0, + 601.0, + 1455.0, + 601.0, + 1497.0, + 366.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1345.0, + 376.0, + 1345.0, + 376.0, + 1383.0, + 293.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 781.0, + 375.0, + 781.0, + 375.0, + 820.0, + 294.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 549.0, + 1034.0, + 549.0, + 1034.0, + 589.0, + 293.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 505.0, + 371.0, + 505.0, + 371.0, + 540.0, + 296.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 505.0, + 820.0, + 505.0, + 820.0, + 540.0, + 416.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 627.0, + 1404.0, + 627.0, + 1404.0, + 673.0, + 293.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1229.0, + 1286.0, + 1229.0, + 1286.0, + 1267.0, + 294.0, + 1267.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 227, + 1404, + 227, + 1404, + 415, + 297, + 415 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 294, + 632, + 1404, + 632, + 1404, + 696, + 294, + 696 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 759, + 546, + 936, + 546, + 936, + 620, + 759, + 620 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 440, + 430, + 1259, + 430, + 1259, + 491, + 440, + 491 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 297, + 505, + 370, + 505, + 370, + 535, + 297, + 535 + ], + "score": 0.921 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 816, + 74, + 816, + 105, + 299, + 105 + ], + "score": 0.903 + }, + { + "category_id": 9, + "poly": [ + 1352, + 566, + 1400, + 566, + 1400, + 597, + 1352, + 597 + ], + "score": 0.891 + }, + { + "category_id": 9, + "poly": [ + 1351, + 442, + 1400, + 442, + 1400, + 474, + 1351, + 474 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.84 + }, + { + "category_id": 14, + "poly": [ + 760, + 544, + 938, + 544, + 938, + 621, + 760, + 621 + ], + "score": 0.95, + "latex": "\\rho _ { t } = \\frac { \\pi _ { \\theta } ( a _ { t } | s _ { t } ) } { \\pi ( a _ { t } | s _ { t } ) }" + }, + { + "category_id": 14, + "poly": [ + 440, + 428, + 1262, + 428, + 1262, + 489, + 440, + 489 + ], + "score": 0.92, + "latex": "\\operatorname* { m a x } _ { \\theta } \\mathbb { E } _ { ( s _ { t } , a _ { t } ) \\sim \\pi } \\left[ \\operatorname* { m i n } \\left( \\mathrm { c l i p } \\left( \\rho _ { t } , 1 - \\varepsilon , 1 + \\varepsilon \\right) \\widehat A _ { \\pi } ( s _ { t } , a _ { t } ) , \\rho _ { t } \\widehat A _ { \\pi } ( s _ { t } , a _ { t } ) \\right) \\right]" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 268.0, + 294.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 298.0, + 295.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 293.0, + 1404.0, + 293.0, + 1404.0, + 325.0, + 296.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 320.0, + 1405.0, + 320.0, + 1405.0, + 358.0, + 291.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 352.0, + 1404.0, + 352.0, + 1404.0, + 388.0, + 292.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 383.0, + 804.0, + 383.0, + 804.0, + 419.0, + 295.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 628.0, + 1404.0, + 628.0, + 1404.0, + 669.0, + 293.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 660.0, + 649.0, + 660.0, + 649.0, + 696.0, + 296.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 500.0, + 375.0, + 500.0, + 375.0, + 539.0, + 294.0, + 539.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 335, + 404, + 1364, + 404, + 1364, + 1110, + 335, + 1110 + ], + "score": 0.987, + "html": "
Humanoid-v2Walker2d-v2Hopper-v2
PPOTRPOPPOTRPOPPOTRPO
Timesteps per iteration204820482048204820482048
Discount factor ()0.990.990.990.990.990.99
GAE discount (λ)0.950.950.950.950.950.95
Value network LR0.00010.00030.00030.00030.00020.0002
Value net num. epochs101010101010
Policy net hidden layers[64, 64][64, 64][64,64][64, 64][64, 64][64,64]
Value net hidden layers[64, 64][64, 64][64, 64][64,64][64,64][64,64]
KL constraint (δ)N/A0.07N/A0.04N/A0.13
Fisher est. fractionN/A0.1N/A0.1N/A0.1
Conjugate grad. stepsN/A10N/A10N/A10
CG dampingN/A0.1N/A0.1N/A0.1
Backtracking stepsN/A10N/A10N/A10
Policy LR (Adam)0.00025N/A0.0004N/A0.00045N/A
Policy epochs10N/A10N/A10N/A
PPO Clipping ε0.2N/A0.2N/A0.2N/A
Entropy coeff.0.00.00.00.00.00.0
Reward clipping[-10,10]一[-10,10]一[-10, 10]1
Reward normalizationOnOffOnOffOnOff
State clipping[-10,10]1[-10,10]1[-10, 10]1
" + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 105, + 300, + 105 + ], + "score": 0.896 + }, + { + "category_id": 6, + "poly": [ + 523, + 348, + 1173, + 348, + 1173, + 382, + 523, + 382 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 300, + 229, + 647, + 229, + 647, + 262, + 300, + 262 + ], + "score": 0.889 + }, + { + "category_id": 1, + "poly": [ + 293, + 285, + 1324, + 285, + 1324, + 320, + 293, + 320 + ], + "score": 0.86 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.843 + }, + { + "category_id": 7, + "poly": [ + 296, + 1140, + 1283, + 1140, + 1283, + 1173, + 296, + 1173 + ], + "score": 0.368 + }, + { + "category_id": 1, + "poly": [ + 296, + 1140, + 1283, + 1140, + 1283, + 1173, + 296, + 1173 + ], + "score": 0.325 + }, + { + "category_id": 13, + "poly": [ + 582, + 1141, + 638, + 1141, + 638, + 1170, + 582, + 1170 + ], + "score": 0.86, + "latex": "9 5 \\%" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 345.0, + 1174.0, + 345.0, + 1174.0, + 388.0, + 523.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 228.0, + 649.0, + 228.0, + 649.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1133.0, + 581.0, + 1133.0, + 581.0, + 1183.0, + 293.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 1133.0, + 1287.0, + 1133.0, + 1287.0, + 1183.0, + 639.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 282.0, + 1332.0, + 282.0, + 1332.0, + 328.0, + 294.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1133.0, + 581.0, + 1133.0, + 581.0, + 1183.0, + 293.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 1133.0, + 1287.0, + 1133.0, + 1287.0, + 1183.0, + 639.0, + 1183.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 524, + 286, + 1176, + 286, + 1176, + 1653, + 524, + 1653 + ], + "score": 0.957 + }, + { + "category_id": 4, + "poly": [ + 295, + 1677, + 1407, + 1677, + 1407, + 1773, + 295, + 1773 + ], + "score": 0.948 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.854 + }, + { + "category_id": 4, + "poly": [ + 300, + 230, + 702, + 230, + 702, + 261, + 300, + 261 + ], + "score": 0.271 + }, + { + "category_id": 15, + "poly": [ + 560.0, + 314.0, + 632.0, + 314.0, + 632.0, + 351.0, + 560.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 351.0, + 573.0, + 351.0, + 573.0, + 488.0, + 521.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 358.0, + 632.0, + 358.0, + 632.0, + 396.0, + 560.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 404.0, + 632.0, + 404.0, + 632.0, + 441.0, + 560.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 448.0, + 632.0, + 448.0, + 632.0, + 487.0, + 559.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 468.0, + 634.0, + 468.0, + 634.0, + 574.0, + 523.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 503.0, + 1141.0, + 503.0, + 1141.0, + 549.0, + 1035.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 539.0, + 632.0, + 539.0, + 632.0, + 577.0, + 573.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 546.0, + 1123.0, + 546.0, + 1123.0, + 594.0, + 1038.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 587.0, + 629.0, + 587.0, + 629.0, + 617.0, + 602.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 626.0, + 676.0, + 626.0, + 676.0, + 655.0, + 649.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 621.0, + 790.0, + 621.0, + 790.0, + 659.0, + 729.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 620.0, + 888.0, + 620.0, + 888.0, + 659.0, + 825.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 620.0, + 983.0, + 620.0, + 983.0, + 659.0, + 922.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 620.0, + 1080.0, + 620.0, + 1080.0, + 659.0, + 1021.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 621.0, + 1177.0, + 621.0, + 1177.0, + 659.0, + 1116.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 649.0, + 989.0, + 649.0, + 989.0, + 693.0, + 809.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 709.0, + 931.0, + 709.0, + 931.0, + 749.0, + 768.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 775.0, + 839.0, + 775.0, + 839.0, + 823.0, + 736.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 801.0, + 574.0, + 801.0, + 574.0, + 1028.0, + 519.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 812.0, + 632.0, + 812.0, + 632.0, + 850.0, + 560.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 820.0, + 820.0, + 820.0, + 820.0, + 867.0, + 736.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 885.0, + 632.0, + 885.0, + 632.0, + 922.0, + 560.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 959.0, + 632.0, + 959.0, + 632.0, + 997.0, + 560.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1037.0, + 629.0, + 1037.0, + 629.0, + 1066.0, + 602.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1075.0, + 678.0, + 1075.0, + 678.0, + 1109.0, + 648.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 1072.0, + 790.0, + 1072.0, + 790.0, + 1110.0, + 731.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 1072.0, + 886.0, + 1072.0, + 886.0, + 1110.0, + 826.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 1072.0, + 983.0, + 1072.0, + 983.0, + 1110.0, + 924.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 1072.0, + 1080.0, + 1072.0, + 1080.0, + 1110.0, + 1021.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 1072.0, + 1176.0, + 1072.0, + 1176.0, + 1110.0, + 1116.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 1099.0, + 989.0, + 1099.0, + 989.0, + 1144.0, + 808.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1161.0, + 940.0, + 1161.0, + 940.0, + 1197.0, + 757.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 1202.0, + 632.0, + 1202.0, + 632.0, + 1238.0, + 560.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1232.0, + 839.0, + 1232.0, + 839.0, + 1280.0, + 734.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 1257.0, + 571.0, + 1257.0, + 571.0, + 1481.0, + 521.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1266.0, + 632.0, + 1266.0, + 632.0, + 1304.0, + 573.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1275.0, + 820.0, + 1275.0, + 820.0, + 1324.0, + 736.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1330.0, + 632.0, + 1330.0, + 632.0, + 1368.0, + 573.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1396.0, + 631.0, + 1396.0, + 631.0, + 1429.0, + 573.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1457.0, + 632.0, + 1457.0, + 632.0, + 1497.0, + 573.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 1533.0, + 676.0, + 1533.0, + 676.0, + 1562.0, + 649.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 1527.0, + 789.0, + 1527.0, + 789.0, + 1567.0, + 729.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 1527.0, + 888.0, + 1527.0, + 888.0, + 1567.0, + 825.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1527.0, + 983.0, + 1527.0, + 983.0, + 1567.0, + 922.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 1527.0, + 1080.0, + 1527.0, + 1080.0, + 1565.0, + 1021.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 1530.0, + 1174.0, + 1530.0, + 1174.0, + 1564.0, + 1119.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 1553.0, + 990.0, + 1553.0, + 990.0, + 1602.0, + 808.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 1619.0, + 943.0, + 1619.0, + 943.0, + 1655.0, + 754.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1677.0, + 1404.0, + 1677.0, + 1404.0, + 1712.0, + 294.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1708.0, + 1404.0, + 1708.0, + 1404.0, + 1745.0, + 294.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1738.0, + 929.0, + 1738.0, + 929.0, + 1775.0, + 294.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 703.0, + 227.0, + 703.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 298, + 298, + 1399, + 298, + 1399, + 878, + 298, + 878 + ], + "score": 0.975 + }, + { + "category_id": 4, + "poly": [ + 295, + 902, + 1405, + 902, + 1405, + 1149, + 295, + 1149 + ], + "score": 0.935 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 836, + 2089, + 865, + 2089, + 865, + 2112, + 836, + 2112 + ], + "score": 0.843 + }, + { + "category_id": 0, + "poly": [ + 299, + 231, + 814, + 231, + 814, + 261, + 299, + 261 + ], + "score": 0.487 + }, + { + "category_id": 2, + "poly": [ + 299, + 231, + 814, + 231, + 814, + 261, + 299, + 261 + ], + "score": 0.414 + }, + { + "category_id": 13, + "poly": [ + 296, + 1024, + 353, + 1024, + 353, + 1056, + 296, + 1056 + ], + "score": 0.87, + "latex": "9 5 \\%" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 300.0, + 514.0, + 300.0, + 514.0, + 328.0, + 401.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 300.0, + 801.0, + 300.0, + 801.0, + 328.0, + 670.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 300.0, + 1077.0, + 300.0, + 1077.0, + 328.0, + 947.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 300.0, + 1355.0, + 300.0, + 1355.0, + 328.0, + 1225.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 324.0, + 354.0, + 324.0, + 354.0, + 504.0, + 299.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 325.0, + 446.0, + 325.0, + 446.0, + 358.0, + 392.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 324.0, + 633.0, + 324.0, + 633.0, + 503.0, + 575.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 327.0, + 725.0, + 327.0, + 725.0, + 358.0, + 670.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 324.0, + 910.0, + 324.0, + 910.0, + 503.0, + 855.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 327.0, + 1002.0, + 327.0, + 1002.0, + 358.0, + 949.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 324.0, + 1187.0, + 324.0, + 1187.0, + 503.0, + 1131.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 327.0, + 1280.0, + 327.0, + 1280.0, + 358.0, + 1226.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 346.0, + 439.0, + 346.0, + 439.0, + 378.0, + 393.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 347.0, + 716.0, + 347.0, + 716.0, + 378.0, + 670.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 347.0, + 994.0, + 347.0, + 994.0, + 378.0, + 949.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 347.0, + 1273.0, + 347.0, + 1273.0, + 378.0, + 1225.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 383.0, + 353.0, + 383.0, + 353.0, + 403.0, + 328.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 383.0, + 631.0, + 383.0, + 631.0, + 402.0, + 606.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 383.0, + 909.0, + 383.0, + 909.0, + 403.0, + 884.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 383.0, + 1186.0, + 383.0, + 1186.0, + 403.0, + 1161.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 438.0, + 353.0, + 438.0, + 353.0, + 458.0, + 328.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 438.0, + 631.0, + 438.0, + 631.0, + 457.0, + 606.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 438.0, + 909.0, + 438.0, + 909.0, + 457.0, + 884.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 438.0, + 1186.0, + 438.0, + 1186.0, + 457.0, + 1161.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 492.0, + 377.0, + 492.0, + 377.0, + 523.0, + 322.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 502.0, + 424.0, + 502.0, + 424.0, + 524.0, + 394.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 502.0, + 472.0, + 502.0, + 472.0, + 524.0, + 442.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 502.0, + 518.0, + 502.0, + 518.0, + 524.0, + 488.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 500.0, + 566.0, + 500.0, + 566.0, + 524.0, + 534.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 490.0, + 655.0, + 490.0, + 655.0, + 523.0, + 600.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 503.0, + 701.0, + 503.0, + 701.0, + 523.0, + 675.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 503.0, + 748.0, + 503.0, + 748.0, + 521.0, + 716.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 502.0, + 796.0, + 502.0, + 796.0, + 524.0, + 766.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 502.0, + 843.0, + 502.0, + 843.0, + 524.0, + 812.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 490.0, + 932.0, + 490.0, + 932.0, + 523.0, + 876.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 502.0, + 980.0, + 502.0, + 980.0, + 524.0, + 950.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 502.0, + 1027.0, + 502.0, + 1027.0, + 524.0, + 996.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 502.0, + 1075.0, + 502.0, + 1075.0, + 524.0, + 1044.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 500.0, + 1120.0, + 500.0, + 1120.0, + 524.0, + 1090.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 492.0, + 1210.0, + 492.0, + 1210.0, + 523.0, + 1155.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 502.0, + 1259.0, + 502.0, + 1259.0, + 524.0, + 1229.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 504.0, + 1302.0, + 504.0, + 1302.0, + 520.0, + 1274.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1321.0, + 502.0, + 1351.0, + 502.0, + 1351.0, + 524.0, + 1321.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 502.0, + 1399.0, + 502.0, + 1399.0, + 524.0, + 1368.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 513.0, + 538.0, + 513.0, + 538.0, + 544.0, + 375.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 514.0, + 817.0, + 514.0, + 817.0, + 544.0, + 652.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 513.0, + 1095.0, + 513.0, + 1095.0, + 544.0, + 930.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 513.0, + 1373.0, + 513.0, + 1373.0, + 544.0, + 1208.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 556.0, + 939.0, + 556.0, + 939.0, + 589.0, + 759.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 591.0, + 514.0, + 591.0, + 514.0, + 618.0, + 401.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 591.0, + 801.0, + 591.0, + 801.0, + 618.0, + 670.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 591.0, + 1079.0, + 591.0, + 1079.0, + 618.0, + 947.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 591.0, + 1356.0, + 591.0, + 1356.0, + 618.0, + 1225.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 613.0, + 353.0, + 613.0, + 353.0, + 793.0, + 299.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 613.0, + 448.0, + 613.0, + 448.0, + 649.0, + 392.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 614.0, + 631.0, + 614.0, + 631.0, + 790.0, + 576.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 614.0, + 726.0, + 614.0, + 726.0, + 649.0, + 668.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 613.0, + 910.0, + 613.0, + 910.0, + 792.0, + 855.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 614.0, + 1002.0, + 614.0, + 1002.0, + 649.0, + 947.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 615.0, + 1189.0, + 615.0, + 1189.0, + 791.0, + 1132.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 614.0, + 1281.0, + 614.0, + 1281.0, + 649.0, + 1225.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 633.0, + 443.0, + 633.0, + 443.0, + 669.0, + 392.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 636.0, + 716.0, + 636.0, + 716.0, + 668.0, + 670.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 636.0, + 994.0, + 636.0, + 994.0, + 667.0, + 949.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 636.0, + 1271.0, + 636.0, + 1271.0, + 667.0, + 1226.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 673.0, + 353.0, + 673.0, + 353.0, + 693.0, + 328.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 673.0, + 631.0, + 673.0, + 631.0, + 692.0, + 606.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 671.0, + 908.0, + 671.0, + 908.0, + 691.0, + 880.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 672.0, + 1186.0, + 672.0, + 1186.0, + 692.0, + 1161.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 727.0, + 353.0, + 727.0, + 353.0, + 747.0, + 328.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 728.0, + 631.0, + 728.0, + 631.0, + 747.0, + 606.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 728.0, + 909.0, + 728.0, + 909.0, + 747.0, + 884.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 728.0, + 1186.0, + 728.0, + 1186.0, + 747.0, + 1161.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 782.0, + 377.0, + 782.0, + 377.0, + 812.0, + 322.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 793.0, + 423.0, + 793.0, + 423.0, + 812.0, + 396.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 793.0, + 471.0, + 793.0, + 471.0, + 812.0, + 443.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 793.0, + 518.0, + 793.0, + 518.0, + 812.0, + 489.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 790.0, + 565.0, + 790.0, + 565.0, + 813.0, + 534.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 781.0, + 655.0, + 781.0, + 655.0, + 812.0, + 603.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 793.0, + 701.0, + 793.0, + 701.0, + 812.0, + 673.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 793.0, + 748.0, + 793.0, + 748.0, + 812.0, + 720.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 793.0, + 797.0, + 793.0, + 797.0, + 812.0, + 767.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 793.0, + 841.0, + 793.0, + 841.0, + 812.0, + 813.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 782.0, + 932.0, + 782.0, + 932.0, + 812.0, + 877.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 793.0, + 979.0, + 793.0, + 979.0, + 812.0, + 951.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 793.0, + 1026.0, + 793.0, + 1026.0, + 812.0, + 999.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 793.0, + 1074.0, + 793.0, + 1074.0, + 812.0, + 1045.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 793.0, + 1119.0, + 793.0, + 1119.0, + 812.0, + 1091.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 779.0, + 1210.0, + 779.0, + 1210.0, + 812.0, + 1157.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 793.0, + 1256.0, + 793.0, + 1256.0, + 812.0, + 1230.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 793.0, + 1304.0, + 793.0, + 1304.0, + 812.0, + 1276.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1324.0, + 793.0, + 1351.0, + 793.0, + 1351.0, + 812.0, + 1324.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1370.0, + 793.0, + 1396.0, + 793.0, + 1396.0, + 812.0, + 1370.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 805.0, + 537.0, + 805.0, + 537.0, + 832.0, + 376.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 805.0, + 816.0, + 805.0, + 816.0, + 833.0, + 654.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 804.0, + 1094.0, + 804.0, + 1094.0, + 834.0, + 930.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 804.0, + 1372.0, + 804.0, + 1372.0, + 832.0, + 1209.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 844.0, + 928.0, + 844.0, + 928.0, + 882.0, + 768.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 904.0, + 1405.0, + 904.0, + 1405.0, + 938.0, + 295.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 934.0, + 1406.0, + 934.0, + 1406.0, + 971.0, + 293.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 966.0, + 1405.0, + 966.0, + 1405.0, + 998.0, + 292.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 995.0, + 1405.0, + 995.0, + 1405.0, + 1028.0, + 292.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1024.0, + 1403.0, + 1024.0, + 1403.0, + 1060.0, + 354.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1056.0, + 1402.0, + 1056.0, + 1402.0, + 1089.0, + 293.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1087.0, + 1405.0, + 1087.0, + 1405.0, + 1121.0, + 294.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1116.0, + 1289.0, + 1116.0, + 1289.0, + 1151.0, + 293.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 871.0, + 2085.0, + 871.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 817.0, + 230.0, + 817.0, + 265.0, + 297.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 817.0, + 230.0, + 817.0, + 265.0, + 297.0, + 265.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 298, + 684, + 1399, + 684, + 1399, + 1263, + 298, + 1263 + ], + "score": 0.974 + }, + { + "category_id": 4, + "poly": [ + 296, + 1287, + 1405, + 1287, + 1405, + 1565, + 296, + 1565 + ], + "score": 0.961 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.84 + }, + { + "category_id": 13, + "poly": [ + 675, + 1410, + 731, + 1410, + 731, + 1440, + 675, + 1440 + ], + "score": 0.89, + "latex": "9 5 \\%" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 689.0, + 353.0, + 689.0, + 353.0, + 911.0, + 298.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 686.0, + 514.0, + 686.0, + 514.0, + 745.0, + 391.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 689.0, + 631.0, + 689.0, + 631.0, + 911.0, + 576.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 686.0, + 801.0, + 686.0, + 801.0, + 745.0, + 668.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 688.0, + 909.0, + 688.0, + 909.0, + 912.0, + 852.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 686.0, + 1079.0, + 686.0, + 1079.0, + 745.0, + 947.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 688.0, + 1186.0, + 688.0, + 1186.0, + 912.0, + 1133.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 686.0, + 1355.0, + 686.0, + 1355.0, + 743.0, + 1225.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 731.0, + 441.0, + 731.0, + 441.0, + 764.0, + 393.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 731.0, + 716.0, + 731.0, + 716.0, + 764.0, + 670.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 731.0, + 994.0, + 731.0, + 994.0, + 763.0, + 949.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 731.0, + 1271.0, + 731.0, + 1271.0, + 763.0, + 1225.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 769.0, + 353.0, + 769.0, + 353.0, + 789.0, + 328.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 769.0, + 631.0, + 769.0, + 631.0, + 788.0, + 606.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 769.0, + 909.0, + 769.0, + 909.0, + 788.0, + 884.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 769.0, + 1186.0, + 769.0, + 1186.0, + 788.0, + 1161.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 824.0, + 353.0, + 824.0, + 353.0, + 844.0, + 328.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 824.0, + 631.0, + 824.0, + 631.0, + 842.0, + 605.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 824.0, + 909.0, + 824.0, + 909.0, + 844.0, + 884.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 824.0, + 1186.0, + 824.0, + 1186.0, + 842.0, + 1161.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 877.0, + 376.0, + 877.0, + 376.0, + 909.0, + 322.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 889.0, + 423.0, + 889.0, + 423.0, + 909.0, + 397.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 887.0, + 472.0, + 887.0, + 472.0, + 910.0, + 442.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 887.0, + 518.0, + 887.0, + 518.0, + 910.0, + 488.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 887.0, + 565.0, + 887.0, + 565.0, + 910.0, + 534.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 880.0, + 653.0, + 880.0, + 653.0, + 909.0, + 602.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 889.0, + 747.0, + 889.0, + 747.0, + 905.0, + 719.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 889.0, + 795.0, + 889.0, + 795.0, + 909.0, + 768.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 887.0, + 842.0, + 887.0, + 842.0, + 910.0, + 812.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 877.0, + 932.0, + 877.0, + 932.0, + 909.0, + 877.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 889.0, + 979.0, + 889.0, + 979.0, + 909.0, + 952.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 887.0, + 1027.0, + 887.0, + 1027.0, + 910.0, + 996.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 889.0, + 1119.0, + 889.0, + 1119.0, + 909.0, + 1092.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 877.0, + 1210.0, + 877.0, + 1210.0, + 909.0, + 1156.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 889.0, + 1256.0, + 889.0, + 1256.0, + 909.0, + 1230.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 889.0, + 1302.0, + 889.0, + 1302.0, + 905.0, + 1274.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1321.0, + 887.0, + 1351.0, + 887.0, + 1351.0, + 910.0, + 1321.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1370.0, + 889.0, + 1396.0, + 889.0, + 1396.0, + 909.0, + 1370.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 898.0, + 538.0, + 898.0, + 538.0, + 930.0, + 375.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 898.0, + 817.0, + 898.0, + 817.0, + 930.0, + 652.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 898.0, + 1095.0, + 898.0, + 1095.0, + 930.0, + 930.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 898.0, + 1373.0, + 898.0, + 1373.0, + 930.0, + 1208.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 941.0, + 938.0, + 941.0, + 938.0, + 973.0, + 759.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 977.0, + 353.0, + 977.0, + 353.0, + 1201.0, + 298.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 976.0, + 514.0, + 976.0, + 514.0, + 1036.0, + 391.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 978.0, + 631.0, + 978.0, + 631.0, + 1201.0, + 577.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 976.0, + 801.0, + 976.0, + 801.0, + 1035.0, + 670.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 975.0, + 910.0, + 975.0, + 910.0, + 1202.0, + 852.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 976.0, + 1079.0, + 976.0, + 1079.0, + 1035.0, + 947.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 977.0, + 1188.0, + 977.0, + 1188.0, + 1202.0, + 1132.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 976.0, + 1356.0, + 976.0, + 1356.0, + 1033.0, + 1225.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1021.0, + 439.0, + 1021.0, + 439.0, + 1053.0, + 392.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 1021.0, + 716.0, + 1021.0, + 716.0, + 1053.0, + 670.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 1023.0, + 991.0, + 1023.0, + 991.0, + 1051.0, + 950.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 1021.0, + 1271.0, + 1021.0, + 1271.0, + 1052.0, + 1226.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1058.0, + 353.0, + 1058.0, + 353.0, + 1078.0, + 328.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1058.0, + 631.0, + 1058.0, + 631.0, + 1077.0, + 606.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 1058.0, + 909.0, + 1058.0, + 909.0, + 1078.0, + 882.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 1058.0, + 1186.0, + 1058.0, + 1186.0, + 1077.0, + 1161.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1113.0, + 353.0, + 1113.0, + 353.0, + 1132.0, + 328.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 1113.0, + 631.0, + 1113.0, + 631.0, + 1132.0, + 605.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 1113.0, + 909.0, + 1113.0, + 909.0, + 1132.0, + 882.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 1113.0, + 1186.0, + 1113.0, + 1186.0, + 1132.0, + 1161.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1167.0, + 353.0, + 1167.0, + 353.0, + 1189.0, + 322.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1169.0, + 630.0, + 1169.0, + 630.0, + 1188.0, + 602.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1167.0, + 909.0, + 1167.0, + 909.0, + 1189.0, + 877.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1167.0, + 1188.0, + 1167.0, + 1188.0, + 1189.0, + 1156.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1178.0, + 377.0, + 1178.0, + 377.0, + 1198.0, + 351.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1178.0, + 421.0, + 1178.0, + 421.0, + 1195.0, + 392.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 1177.0, + 473.0, + 1177.0, + 473.0, + 1199.0, + 442.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 1178.0, + 518.0, + 1178.0, + 518.0, + 1197.0, + 489.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 1179.0, + 561.0, + 1179.0, + 561.0, + 1195.0, + 534.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1178.0, + 653.0, + 1178.0, + 653.0, + 1198.0, + 628.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 1180.0, + 699.0, + 1180.0, + 699.0, + 1196.0, + 671.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 1179.0, + 747.0, + 1179.0, + 747.0, + 1195.0, + 719.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1178.0, + 797.0, + 1178.0, + 797.0, + 1197.0, + 767.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 1177.0, + 842.0, + 1177.0, + 842.0, + 1199.0, + 812.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1178.0, + 932.0, + 1178.0, + 932.0, + 1198.0, + 906.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1178.0, + 979.0, + 1178.0, + 979.0, + 1197.0, + 951.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 1177.0, + 1027.0, + 1177.0, + 1027.0, + 1199.0, + 996.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1178.0, + 1074.0, + 1178.0, + 1074.0, + 1197.0, + 1045.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1178.0, + 1117.0, + 1178.0, + 1117.0, + 1198.0, + 1091.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1178.0, + 1210.0, + 1178.0, + 1210.0, + 1198.0, + 1185.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1227.0, + 1180.0, + 1254.0, + 1180.0, + 1254.0, + 1196.0, + 1227.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 1179.0, + 1302.0, + 1179.0, + 1302.0, + 1195.0, + 1274.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1323.0, + 1178.0, + 1354.0, + 1178.0, + 1354.0, + 1197.0, + 1323.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 1177.0, + 1398.0, + 1177.0, + 1398.0, + 1199.0, + 1368.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 1191.0, + 538.0, + 1191.0, + 538.0, + 1218.0, + 376.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 1188.0, + 817.0, + 1188.0, + 817.0, + 1220.0, + 652.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 1191.0, + 1094.0, + 1191.0, + 1094.0, + 1218.0, + 931.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 1191.0, + 1373.0, + 1191.0, + 1373.0, + 1218.0, + 1209.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1227.0, + 928.0, + 1227.0, + 928.0, + 1267.0, + 768.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 889.0, + 702.0, + 889.0, + 702.0, + 907.0, + 672.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.25, + 887.0, + 1076.25, + 887.0, + 1076.25, + 909.0, + 1040.25, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1288.0, + 1406.0, + 1288.0, + 1406.0, + 1325.0, + 294.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1321.0, + 1403.0, + 1321.0, + 1403.0, + 1354.0, + 295.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1350.0, + 1406.0, + 1350.0, + 1406.0, + 1387.0, + 294.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1380.0, + 1406.0, + 1380.0, + 1406.0, + 1415.0, + 294.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1410.0, + 674.0, + 1410.0, + 674.0, + 1446.0, + 293.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1410.0, + 1406.0, + 1410.0, + 1406.0, + 1446.0, + 732.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1440.0, + 1406.0, + 1440.0, + 1406.0, + 1477.0, + 291.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1472.0, + 1405.0, + 1472.0, + 1405.0, + 1506.0, + 293.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1499.0, + 1407.0, + 1499.0, + 1407.0, + 1538.0, + 293.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1532.0, + 578.0, + 1532.0, + 578.0, + 1569.0, + 295.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 353, + 303, + 1343, + 303, + 1343, + 1288, + 353, + 1288 + ], + "score": 0.978 + }, + { + "category_id": 4, + "poly": [ + 295, + 1312, + 1407, + 1312, + 1407, + 1468, + 295, + 1468 + ], + "score": 0.961 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 816, + 76, + 816, + 105, + 299, + 105 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.848 + }, + { + "category_id": 0, + "poly": [ + 299, + 230, + 612, + 230, + 612, + 261, + 299, + 261 + ], + "score": 0.639 + }, + { + "category_id": 2, + "poly": [ + 299, + 230, + 612, + 230, + 612, + 261, + 299, + 261 + ], + "score": 0.197 + }, + { + "category_id": 13, + "poly": [ + 348, + 1404, + 530, + 1404, + 530, + 1439, + 348, + 1439 + ], + "score": 0.92, + "latex": "( \\bar { V } _ { o l d } + \\bar { A _ { G A E } } ) ^ { 2 }" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 321.0, + 431.0, + 321.0, + 431.0, + 345.0, + 387.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 318.0, + 694.0, + 318.0, + 694.0, + 346.0, + 628.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 319.0, + 922.0, + 319.0, + 922.0, + 351.0, + 889.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 343.0, + 834.0, + 343.0, + 834.0, + 507.0, + 447.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 343.0, + 679.0, + 343.0, + 679.0, + 372.0, + 631.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 345.0, + 889.0, + 345.0, + 889.0, + 481.0, + 853.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 355.0, + 921.0, + 355.0, + 921.0, + 386.0, + 877.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 362.0, + 389.0, + 362.0, + 389.0, + 462.0, + 356.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 414.0, + 431.0, + 414.0, + 431.0, + 439.0, + 387.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 436.0, + 921.0, + 436.0, + 921.0, + 463.0, + 878.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 449.0, + 1311.0, + 449.0, + 1311.0, + 479.0, + 1249.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 472.0, + 1301.0, + 472.0, + 1301.0, + 508.0, + 1249.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 529.0, + 467.0, + 529.0, + 467.0, + 548.0, + 449.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 523.0, + 555.0, + 523.0, + 555.0, + 553.0, + 510.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 523.0, + 629.0, + 523.0, + 629.0, + 553.0, + 582.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 523.0, + 703.0, + 523.0, + 703.0, + 553.0, + 657.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 523.0, + 779.0, + 523.0, + 779.0, + 553.0, + 732.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 523.0, + 850.0, + 523.0, + 850.0, + 553.0, + 805.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 528.0, + 957.0, + 528.0, + 957.0, + 549.0, + 939.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 524.0, + 1044.0, + 524.0, + 1044.0, + 553.0, + 999.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 524.0, + 1118.0, + 524.0, + 1118.0, + 553.0, + 1073.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 524.0, + 1192.0, + 524.0, + 1192.0, + 553.0, + 1147.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 524.0, + 1267.0, + 524.0, + 1267.0, + 553.0, + 1222.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 523.0, + 1340.0, + 523.0, + 1340.0, + 553.0, + 1294.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 544.0, + 697.0, + 544.0, + 697.0, + 576.0, + 575.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 546.0, + 1187.0, + 546.0, + 1187.0, + 578.0, + 1065.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 591.0, + 926.0, + 591.0, + 926.0, + 629.0, + 772.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 651.0, + 432.0, + 651.0, + 432.0, + 678.0, + 386.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 647.0, + 694.0, + 647.0, + 694.0, + 679.0, + 629.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 649.0, + 923.0, + 649.0, + 923.0, + 681.0, + 888.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 694.0, + 432.0, + 694.0, + 432.0, + 793.0, + 356.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 676.0, + 832.0, + 676.0, + 832.0, + 836.0, + 443.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 678.0, + 888.0, + 678.0, + 888.0, + 809.0, + 855.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 710.0, + 1040.0, + 710.0, + 1040.0, + 773.0, + 963.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 704.0, + 1301.0, + 704.0, + 1301.0, + 756.0, + 1102.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 779.0, + 1311.0, + 779.0, + 1311.0, + 810.0, + 1250.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 803.0, + 1301.0, + 803.0, + 1301.0, + 839.0, + 1249.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 859.0, + 467.0, + 859.0, + 467.0, + 880.0, + 449.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 855.0, + 555.0, + 855.0, + 555.0, + 884.0, + 510.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 855.0, + 629.0, + 855.0, + 629.0, + 883.0, + 582.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 855.0, + 704.0, + 855.0, + 704.0, + 883.0, + 657.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 855.0, + 780.0, + 855.0, + 780.0, + 883.0, + 732.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 855.0, + 850.0, + 855.0, + 850.0, + 884.0, + 803.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 859.0, + 957.0, + 859.0, + 957.0, + 880.0, + 939.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 855.0, + 1044.0, + 855.0, + 1044.0, + 884.0, + 999.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 855.0, + 1119.0, + 855.0, + 1119.0, + 883.0, + 1073.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 855.0, + 1193.0, + 855.0, + 1193.0, + 883.0, + 1147.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 855.0, + 1268.0, + 855.0, + 1268.0, + 883.0, + 1220.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 854.0, + 1340.0, + 854.0, + 1340.0, + 884.0, + 1294.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 877.0, + 696.0, + 877.0, + 696.0, + 909.0, + 575.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 877.0, + 1187.0, + 877.0, + 1187.0, + 909.0, + 1065.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 925.0, + 937.0, + 925.0, + 937.0, + 957.0, + 760.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 962.0, + 917.0, + 962.0, + 917.0, + 993.0, + 883.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 982.0, + 818.0, + 982.0, + 818.0, + 1012.0, + 755.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 982.0, + 1311.0, + 982.0, + 1311.0, + 1012.0, + 1249.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1026.0, + 421.0, + 1026.0, + 421.0, + 1127.0, + 356.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 1007.0, + 806.0, + 1007.0, + 806.0, + 1040.0, + 754.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 1010.0, + 883.0, + 1010.0, + 883.0, + 1141.0, + 849.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1248.0, + 1008.0, + 1300.0, + 1008.0, + 1300.0, + 1041.0, + 1248.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 1108.0, + 916.0, + 1108.0, + 916.0, + 1132.0, + 873.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 1134.0, + 424.0, + 1134.0, + 424.0, + 1163.0, + 378.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 1127.0, + 561.0, + 1127.0, + 561.0, + 1156.0, + 518.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 1190.0, + 461.0, + 1190.0, + 461.0, + 1211.0, + 442.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 1185.0, + 548.0, + 1185.0, + 548.0, + 1215.0, + 503.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 1185.0, + 623.0, + 1185.0, + 623.0, + 1215.0, + 578.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 1185.0, + 698.0, + 1185.0, + 698.0, + 1215.0, + 653.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 1185.0, + 774.0, + 1185.0, + 774.0, + 1215.0, + 728.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 1185.0, + 847.0, + 1185.0, + 847.0, + 1215.0, + 801.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 1190.0, + 954.0, + 1190.0, + 954.0, + 1211.0, + 934.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 1185.0, + 1042.0, + 1185.0, + 1042.0, + 1215.0, + 997.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 1185.0, + 1117.0, + 1185.0, + 1117.0, + 1215.0, + 1071.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 1185.0, + 1192.0, + 1185.0, + 1192.0, + 1215.0, + 1145.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 1186.0, + 1265.0, + 1186.0, + 1265.0, + 1212.0, + 1219.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 1185.0, + 1340.0, + 1185.0, + 1340.0, + 1215.0, + 1294.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 1207.0, + 692.0, + 1207.0, + 692.0, + 1239.0, + 570.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 1207.0, + 1184.0, + 1207.0, + 1184.0, + 1239.0, + 1063.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 1256.0, + 941.0, + 1256.0, + 941.0, + 1289.0, + 758.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.25, + 362.5, + 435.25, + 362.5, + 435.25, + 394.5, + 386.25, + 394.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 362.0, + 1306.0, + 362.0, + 1306.0, + 440.0, + 1070.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 395.5, + 925.0, + 395.5, + 925.0, + 424.5, + 877.0, + 424.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 460.0, + 432.0, + 460.0, + 432.0, + 484.0, + 381.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.75, + 471.0, + 924.75, + 471.0, + 924.75, + 498.0, + 874.75, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.25, + 695.0, + 436.25, + 695.0, + 436.25, + 724.5, + 385.25, + 724.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 688.0, + 925.0, + 688.0, + 925.0, + 716.5, + 877.0, + 716.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 728.5, + 925.0, + 728.5, + 925.0, + 753.5, + 878.0, + 753.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.75, + 766.0, + 924.75, + 766.0, + 924.75, + 791.0, + 876.75, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.25, + 788.5, + 435.25, + 788.5, + 435.25, + 821.0, + 377.25, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.75, + 805.0, + 922.75, + 805.0, + 922.75, + 829.0, + 877.75, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 970.5, + 428.0, + 970.5, + 428.0, + 1004.5, + 377.0, + 1004.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 1024.5, + 429.0, + 1024.5, + 429.0, + 1057.0, + 376.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.25, + 1035.5, + 917.25, + 1035.5, + 917.25, + 1059.5, + 874.25, + 1059.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1060.0, + 868.0, + 1060.0, + 868.0, + 1219.5, + 517.0, + 1219.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.25, + 1062.5, + 1322.25, + 1062.5, + 1322.25, + 1164.5, + 947.25, + 1164.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.25, + 1114.5, + 490.25, + 1114.5, + 490.25, + 1126.0, + 482.25, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1314.0, + 1404.0, + 1314.0, + 1404.0, + 1347.0, + 296.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1343.0, + 1406.0, + 1343.0, + 1406.0, + 1379.0, + 292.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1375.0, + 1405.0, + 1375.0, + 1405.0, + 1406.0, + 294.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1403.0, + 347.0, + 1403.0, + 347.0, + 1441.0, + 294.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 1403.0, + 1405.0, + 1403.0, + 1405.0, + 1441.0, + 531.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1434.0, + 556.0, + 1434.0, + 556.0, + 1471.0, + 293.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 71.0, + 817.0, + 71.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 225.0, + 615.0, + 225.0, + 615.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 225.0, + 615.0, + 225.0, + 615.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 299, + 665, + 1397, + 665, + 1397, + 1382, + 299, + 1382 + ], + "score": 0.966 + }, + { + "category_id": 4, + "poly": [ + 294, + 1407, + 1406, + 1407, + 1406, + 1593, + 294, + 1593 + ], + "score": 0.959 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.843 + }, + { + "category_id": 13, + "poly": [ + 422, + 1498, + 615, + 1498, + 615, + 1532, + 422, + 1532 + ], + "score": 0.92, + "latex": "( \\bar { V _ { o l d } } + A _ { G A E } ) ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1133, + 1529, + 1188, + 1529, + 1188, + 1559, + 1133, + 1559 + ], + "score": 0.88, + "latex": "5 0 \\%" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 681.0, + 383.0, + 681.0, + 383.0, + 708.0, + 336.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 680.0, + 675.0, + 680.0, + 675.0, + 715.0, + 603.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 683.0, + 929.0, + 683.0, + 929.0, + 720.0, + 894.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 707.0, + 832.0, + 707.0, + 832.0, + 880.0, + 401.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 712.0, + 893.0, + 712.0, + 893.0, + 863.0, + 853.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 715.0, + 657.0, + 715.0, + 657.0, + 740.0, + 607.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 732.0, + 336.0, + 732.0, + 336.0, + 842.0, + 304.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 788.0, + 387.0, + 788.0, + 387.0, + 827.0, + 334.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 827.0, + 1362.0, + 827.0, + 1362.0, + 861.0, + 1292.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 861.0, + 925.0, + 861.0, + 925.0, + 887.0, + 882.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 856.0, + 1351.0, + 856.0, + 1351.0, + 892.0, + 1292.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 916.0, + 424.0, + 916.0, + 424.0, + 939.0, + 404.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 911.0, + 522.0, + 911.0, + 522.0, + 943.0, + 472.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 911.0, + 605.0, + 911.0, + 605.0, + 943.0, + 553.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 911.0, + 687.0, + 911.0, + 687.0, + 943.0, + 636.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 911.0, + 771.0, + 911.0, + 771.0, + 943.0, + 718.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 910.0, + 850.0, + 910.0, + 850.0, + 943.0, + 799.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 914.0, + 970.0, + 914.0, + 970.0, + 942.0, + 947.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 911.0, + 1066.0, + 911.0, + 1066.0, + 943.0, + 1016.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 911.0, + 1149.0, + 911.0, + 1149.0, + 943.0, + 1098.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 911.0, + 1230.0, + 911.0, + 1230.0, + 943.0, + 1181.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 911.0, + 1313.0, + 911.0, + 1313.0, + 943.0, + 1264.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 911.0, + 1395.0, + 911.0, + 1395.0, + 943.0, + 1345.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 934.0, + 680.0, + 934.0, + 680.0, + 970.0, + 545.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 934.0, + 1224.0, + 934.0, + 1224.0, + 970.0, + 1089.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 984.0, + 928.0, + 984.0, + 928.0, + 1025.0, + 772.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 1055.0, + 379.0, + 1055.0, + 379.0, + 1207.0, + 303.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 1043.0, + 817.0, + 1043.0, + 817.0, + 1110.0, + 742.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 1035.0, + 927.0, + 1035.0, + 927.0, + 1225.0, + 848.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 1044.0, + 1363.0, + 1044.0, + 1363.0, + 1108.0, + 1294.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1138.0, + 379.0, + 1138.0, + 379.0, + 1183.0, + 329.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1184.0, + 376.0, + 1184.0, + 376.0, + 1219.0, + 328.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 1172.0, + 482.0, + 1172.0, + 482.0, + 1206.0, + 442.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 1184.0, + 1330.0, + 1184.0, + 1330.0, + 1195.0, + 1317.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1230.0, + 373.0, + 1230.0, + 373.0, + 1258.0, + 330.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1276.0, + 419.0, + 1276.0, + 419.0, + 1304.0, + 395.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1274.0, + 515.0, + 1274.0, + 515.0, + 1306.0, + 465.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 1274.0, + 597.0, + 1274.0, + 597.0, + 1306.0, + 547.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 1274.0, + 681.0, + 1274.0, + 681.0, + 1306.0, + 631.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1276.0, + 764.0, + 1276.0, + 764.0, + 1306.0, + 716.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1274.0, + 846.0, + 1274.0, + 846.0, + 1306.0, + 796.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1279.0, + 965.0, + 1279.0, + 965.0, + 1303.0, + 944.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 1274.0, + 1063.0, + 1274.0, + 1063.0, + 1306.0, + 1013.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 1275.0, + 1146.0, + 1275.0, + 1146.0, + 1306.0, + 1095.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1272.0, + 1229.0, + 1272.0, + 1229.0, + 1308.0, + 1179.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1274.0, + 1313.0, + 1274.0, + 1313.0, + 1306.0, + 1262.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1343.0, + 1274.0, + 1395.0, + 1274.0, + 1395.0, + 1306.0, + 1343.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 1296.0, + 674.0, + 1296.0, + 674.0, + 1333.0, + 539.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 1296.0, + 1221.0, + 1296.0, + 1221.0, + 1333.0, + 1087.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 1348.0, + 943.0, + 1348.0, + 943.0, + 1385.0, + 758.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 728.5, + 394.0, + 728.5, + 394.0, + 773.0, + 331.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 725.0, + 933.0, + 725.0, + 933.0, + 756.5, + 878.0, + 756.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.75, + 770.5, + 931.75, + 770.5, + 931.75, + 799.0, + 880.75, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 814.0, + 933.0, + 814.0, + 933.0, + 845.0, + 879.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 844.5, + 386.0, + 844.5, + 386.0, + 877.0, + 327.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.75, + 1129.0, + 926.75, + 1129.0, + 926.75, + 1156.0, + 877.75, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1174.5, + 927.0, + 1174.5, + 927.0, + 1203.0, + 876.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1220.0, + 928.0, + 1220.0, + 928.0, + 1248.5, + 876.0, + 1248.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1409.0, + 1403.0, + 1409.0, + 1403.0, + 1441.0, + 296.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1437.0, + 1407.0, + 1437.0, + 1407.0, + 1472.0, + 293.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1468.0, + 1405.0, + 1468.0, + 1405.0, + 1504.0, + 295.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1499.0, + 421.0, + 1499.0, + 421.0, + 1534.0, + 295.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 1499.0, + 1403.0, + 1499.0, + 1403.0, + 1534.0, + 616.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1531.0, + 1132.0, + 1531.0, + 1132.0, + 1563.0, + 295.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 1531.0, + 1404.0, + 1531.0, + 1404.0, + 1563.0, + 1189.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1560.0, + 813.0, + 1560.0, + 813.0, + 1596.0, + 296.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2086.0, + 870.0, + 2086.0, + 870.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 18, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 302, + 295, + 1370, + 295, + 1370, + 1212, + 302, + 1212 + ], + "score": 0.974 + }, + { + "category_id": 4, + "poly": [ + 559, + 1237, + 1137, + 1237, + 1137, + 1270, + 559, + 1270 + ], + "score": 0.923 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 816, + 76, + 816, + 104, + 299, + 104 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2112, + 835, + 2112 + ], + "score": 0.845 + }, + { + "category_id": 0, + "poly": [ + 301, + 232, + 703, + 232, + 703, + 260, + 301, + 260 + ], + "score": 0.52 + }, + { + "category_id": 2, + "poly": [ + 301, + 232, + 703, + 232, + 703, + 260, + 301, + 260 + ], + "score": 0.368 + }, + { + "category_id": 2, + "poly": [ + 307, + 368, + 345, + 368, + 345, + 1160, + 307, + 1160 + ], + "score": 0.094 + }, + { + "category_id": 13, + "poly": [ + 1253, + 295, + 1314, + 295, + 1314, + 328, + 1253, + 328 + ], + "score": 0.79, + "latex": "( 1 0 ^ { 6 } )" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 294.0, + 792.0, + 294.0, + 792.0, + 329.0, + 453.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 291.0, + 1252.0, + 291.0, + 1252.0, + 329.0, + 982.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 291.0, + 1320.0, + 291.0, + 1320.0, + 329.0, + 1315.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 323.0, + 560.0, + 323.0, + 560.0, + 363.0, + 440.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 324.0, + 819.0, + 324.0, + 819.0, + 360.0, + 672.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 322.0, + 1091.0, + 322.0, + 1091.0, + 365.0, + 967.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 324.0, + 1348.0, + 324.0, + 1348.0, + 360.0, + 1200.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 396.0, + 351.0, + 396.0, + 351.0, + 486.0, + 308.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 401.0, + 1197.0, + 401.0, + 1197.0, + 450.0, + 1177.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 595.0, + 349.0, + 595.0, + 349.0, + 711.0, + 304.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 622.0, + 417.0, + 622.0, + 417.0, + 678.0, + 401.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 651.0, + 667.0, + 651.0, + 667.0, + 662.0, + 655.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 606.0, + 1191.0, + 606.0, + 1191.0, + 679.0, + 1171.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 668.0, + 668.0, + 668.0, + 668.0, + 678.0, + 656.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 685.0, + 673.0, + 685.0, + 673.0, + 697.0, + 662.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 694.0, + 683.0, + 694.0, + 683.0, + 746.0, + 654.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 763.0, + 1267.0, + 763.0, + 1267.0, + 772.0, + 1253.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 810.0, + 351.0, + 810.0, + 351.0, + 925.0, + 307.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 884.0, + 420.0, + 884.0, + 420.0, + 893.0, + 410.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 824.0, + 673.0, + 824.0, + 673.0, + 897.0, + 647.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 884.0, + 950.0, + 884.0, + 950.0, + 893.0, + 938.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 830.0, + 1199.0, + 830.0, + 1199.0, + 947.0, + 1165.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 893.0, + 683.0, + 893.0, + 683.0, + 932.0, + 657.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 976.0, + 736.0, + 976.0, + 736.0, + 985.0, + 726.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 977.0, + 1264.0, + 977.0, + 1264.0, + 983.0, + 1258.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1021.0, + 349.0, + 1021.0, + 349.0, + 1135.0, + 304.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 1075.0, + 669.0, + 1075.0, + 669.0, + 1102.0, + 652.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 1053.0, + 1195.0, + 1053.0, + 1195.0, + 1105.0, + 1176.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1112.0, + 672.0, + 1112.0, + 672.0, + 1123.0, + 661.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1105.0, + 1207.0, + 1105.0, + 1207.0, + 1160.0, + 1178.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 1130.0, + 971.0, + 1130.0, + 971.0, + 1173.0, + 946.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 1157.0, + 1225.0, + 1157.0, + 1225.0, + 1184.0, + 1202.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1325.0, + 1174.0, + 1340.0, + 1174.0, + 1340.0, + 1184.0, + 1325.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1191.0, + 717.0, + 1191.0, + 717.0, + 1200.0, + 707.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 1187.0, + 736.0, + 1187.0, + 736.0, + 1197.0, + 722.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 1191.0, + 1247.0, + 1191.0, + 1247.0, + 1200.0, + 1238.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.75, + 406.5, + 867.75, + 406.5, + 867.75, + 522.0, + 638.75, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 424.5, + 1136.0, + 424.5, + 1136.0, + 509.5, + 935.0, + 509.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 425.5, + 1386.0, + 425.5, + 1386.0, + 516.5, + 1186.0, + 516.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 423.0, + 604.0, + 423.0, + 604.0, + 502.0, + 403.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 632.5, + 605.0, + 632.5, + 605.0, + 714.0, + 402.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 630.0, + 1142.0, + 630.0, + 1142.0, + 717.0, + 922.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.75, + 741.0, + 691.75, + 741.0, + 691.75, + 755.5, + 674.75, + 755.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 848.0, + 701.0, + 848.0, + 701.0, + 863.5, + 683.0, + 863.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 860.5, + 1228.0, + 860.5, + 1228.0, + 869.0, + 1218.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.75, + 842.5, + 610.75, + 842.5, + 610.75, + 928.5, + 395.75, + 928.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.75, + 841.5, + 1143.75, + 841.5, + 1143.75, + 927.5, + 922.75, + 927.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 937.0, + 1229.0, + 937.0, + 1229.0, + 972.5, + 1192.0, + 972.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 1062.5, + 704.0, + 1062.5, + 704.0, + 1080.5, + 688.0, + 1080.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 1053.0, + 611.0, + 1053.0, + 611.0, + 1140.0, + 399.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 1053.0, + 1142.0, + 1053.0, + 1142.0, + 1144.0, + 923.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.75, + 1119.5, + 692.75, + 1119.5, + 692.75, + 1161.5, + 651.75, + 1161.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.75, + 1166.0, + 685.75, + 1166.0, + 685.75, + 1180.0, + 679.75, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 1232.0, + 1141.0, + 1232.0, + 1141.0, + 1276.0, + 557.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2123.0, + 829.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 226.0, + 707.0, + 226.0, + 707.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 226.0, + 707.0, + 226.0, + 707.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 394.0, + 352.0, + 394.0, + 352.0, + 486.0, + 308.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 595.0, + 350.0, + 595.0, + 350.0, + 710.0, + 307.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 808.0, + 348.0, + 808.0, + 348.0, + 923.0, + 305.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 1021.0, + 348.0, + 1021.0, + 348.0, + 1135.0, + 305.0, + 1135.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 19, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 301, + 641, + 1370, + 641, + 1370, + 1558, + 301, + 1558 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 549, + 1580, + 1148, + 1580, + 1148, + 1613, + 549, + 1613 + ], + "score": 0.926 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 863, + 2088, + 863, + 2113, + 834, + 2113 + ], + "score": 0.824 + }, + { + "category_id": 2, + "poly": [ + 304, + 723, + 347, + 723, + 347, + 1491, + 304, + 1491 + ], + "score": 0.171 + }, + { + "category_id": 13, + "poly": [ + 1253, + 636, + 1314, + 636, + 1314, + 670, + 1253, + 670 + ], + "score": 0.79, + "latex": "( 1 0 ^ { 6 } )" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 636.0, + 792.0, + 636.0, + 792.0, + 670.0, + 452.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 634.0, + 1252.0, + 634.0, + 1252.0, + 669.0, + 982.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 634.0, + 1318.0, + 634.0, + 1318.0, + 669.0, + 1315.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 664.0, + 561.0, + 664.0, + 561.0, + 709.0, + 437.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 669.0, + 818.0, + 669.0, + 818.0, + 701.0, + 674.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 666.0, + 1089.0, + 666.0, + 1089.0, + 705.0, + 968.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 666.0, + 1348.0, + 666.0, + 1348.0, + 701.0, + 1201.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 739.0, + 350.0, + 739.0, + 350.0, + 826.0, + 308.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 754.0, + 661.0, + 754.0, + 661.0, + 760.0, + 655.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 743.0, + 1197.0, + 743.0, + 1197.0, + 791.0, + 1176.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 760.0, + 662.0, + 760.0, + 662.0, + 789.0, + 650.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 867.0, + 442.0, + 867.0, + 442.0, + 877.0, + 434.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 937.0, + 349.0, + 937.0, + 349.0, + 1053.0, + 304.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 960.0, + 672.0, + 960.0, + 672.0, + 1025.0, + 644.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1029.0, + 671.0, + 1029.0, + 671.0, + 1038.0, + 661.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1065.0, + 445.0, + 1065.0, + 445.0, + 1094.0, + 422.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 1047.0, + 968.0, + 1047.0, + 968.0, + 1096.0, + 942.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1150.0, + 349.0, + 1150.0, + 349.0, + 1265.0, + 304.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 1201.0, + 672.0, + 1201.0, + 672.0, + 1238.0, + 652.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1240.0, + 674.0, + 1240.0, + 674.0, + 1251.0, + 661.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 1241.0, + 1202.0, + 1241.0, + 1202.0, + 1251.0, + 1191.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 1247.0, + 679.0, + 1247.0, + 679.0, + 1313.0, + 647.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 1280.0, + 1226.0, + 1280.0, + 1226.0, + 1316.0, + 1200.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1363.0, + 349.0, + 1363.0, + 349.0, + 1479.0, + 304.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 1382.0, + 417.0, + 1382.0, + 417.0, + 1441.0, + 401.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 1394.0, + 667.0, + 1394.0, + 667.0, + 1446.0, + 646.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1395.0, + 1197.0, + 1395.0, + 1197.0, + 1441.0, + 1178.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1454.0, + 672.0, + 1454.0, + 672.0, + 1465.0, + 661.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 1454.0, + 1201.0, + 1454.0, + 1201.0, + 1464.0, + 1191.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 1496.0, + 442.0, + 1496.0, + 442.0, + 1520.0, + 424.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 1477.0, + 968.0, + 1477.0, + 968.0, + 1520.0, + 943.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 1521.0, + 1320.0, + 1521.0, + 1320.0, + 1530.0, + 1308.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1529.0, + 742.0, + 1529.0, + 742.0, + 1542.0, + 706.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 1529.0, + 777.0, + 1529.0, + 777.0, + 1544.0, + 743.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 1529.0, + 1268.0, + 1529.0, + 1268.0, + 1542.0, + 1231.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 761.0, + 608.0, + 761.0, + 608.0, + 846.5, + 397.0, + 846.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 761.0, + 855.0, + 761.0, + 855.0, + 847.5, + 650.0, + 847.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 758.5, + 1140.0, + 758.5, + 1140.0, + 845.5, + 924.0, + 845.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.75, + 761.5, + 1389.75, + 761.5, + 1389.75, + 847.5, + 1174.75, + 847.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.25, + 828.0, + 689.25, + 828.0, + 689.25, + 863.0, + 655.25, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 967.0, + 1385.0, + 967.0, + 1385.0, + 1063.5, + 1176.0, + 1063.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 973.0, + 611.0, + 973.0, + 611.0, + 1062.0, + 397.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 971.0, + 1138.0, + 971.0, + 1138.0, + 1057.5, + 927.0, + 1057.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1186.5, + 1349.0, + 1186.5, + 1349.0, + 1266.5, + 1179.0, + 1266.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.75, + 1185.0, + 613.75, + 1185.0, + 613.75, + 1276.0, + 395.75, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.75, + 1184.5, + 1140.75, + 1184.5, + 1140.75, + 1272.0, + 923.75, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 1277.0, + 453.0, + 1277.0, + 453.0, + 1310.0, + 419.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.25, + 1399.5, + 612.25, + 1399.5, + 612.25, + 1486.5, + 401.25, + 1486.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.75, + 1398.0, + 1137.75, + 1398.0, + 1137.75, + 1485.0, + 927.75, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.75, + 1464.0, + 685.75, + 1464.0, + 685.75, + 1491.0, + 652.75, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 1576.0, + 1150.0, + 1576.0, + 1150.0, + 1619.0, + 548.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 830.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 736.0, + 351.0, + 736.0, + 351.0, + 827.0, + 307.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 937.0, + 349.0, + 937.0, + 349.0, + 1052.0, + 304.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 1149.0, + 350.0, + 1149.0, + 350.0, + 1266.0, + 303.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 1362.0, + 351.0, + 1362.0, + 351.0, + 1479.0, + 303.0, + 1479.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 20, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 301, + 641, + 1371, + 641, + 1371, + 1558, + 301, + 1558 + ], + "score": 0.97 + }, + { + "category_id": 4, + "poly": [ + 567, + 1580, + 1130, + 1580, + 1130, + 1613, + 567, + 1613 + ], + "score": 0.909 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2113, + 834, + 2113 + ], + "score": 0.838 + }, + { + "category_id": 3, + "poly": [ + 301, + 641, + 1371, + 641, + 1371, + 1558, + 301, + 1558 + ], + "score": 0.196 + }, + { + "category_id": 13, + "poly": [ + 1253, + 636, + 1314, + 636, + 1314, + 670, + 1253, + 670 + ], + "score": 0.8, + "latex": "( 1 0 ^ { 6 } )" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 636.0, + 792.0, + 636.0, + 792.0, + 672.0, + 452.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 634.0, + 1252.0, + 634.0, + 1252.0, + 669.0, + 981.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 664.0, + 561.0, + 664.0, + 561.0, + 709.0, + 437.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 668.0, + 818.0, + 668.0, + 818.0, + 700.0, + 671.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 664.0, + 1091.0, + 664.0, + 1091.0, + 709.0, + 967.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 669.0, + 1348.0, + 669.0, + 1348.0, + 701.0, + 1202.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 739.0, + 350.0, + 739.0, + 350.0, + 826.0, + 308.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 737.0, + 1203.0, + 737.0, + 1203.0, + 824.0, + 1176.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 849.0, + 684.0, + 849.0, + 684.0, + 859.0, + 670.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 846.0, + 1223.0, + 846.0, + 1223.0, + 885.0, + 1196.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 868.0, + 441.0, + 868.0, + 441.0, + 875.0, + 435.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 937.0, + 349.0, + 937.0, + 349.0, + 1053.0, + 304.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 976.0, + 663.0, + 976.0, + 663.0, + 994.0, + 654.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 986.0, + 1194.0, + 986.0, + 1194.0, + 996.0, + 1182.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 991.0, + 668.0, + 991.0, + 668.0, + 1019.0, + 654.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1009.0, + 1194.0, + 1009.0, + 1194.0, + 1019.0, + 1185.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 1059.0, + 685.0, + 1059.0, + 685.0, + 1074.0, + 668.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 1067.0, + 692.0, + 1067.0, + 692.0, + 1098.0, + 668.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 1074.0, + 1227.0, + 1074.0, + 1227.0, + 1101.0, + 1203.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 1104.0, + 737.0, + 1104.0, + 737.0, + 1114.0, + 723.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 1108.0, + 1248.0, + 1108.0, + 1248.0, + 1117.0, + 1238.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1104.0, + 1267.0, + 1104.0, + 1267.0, + 1114.0, + 1253.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1150.0, + 349.0, + 1150.0, + 349.0, + 1265.0, + 304.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1180.0, + 414.0, + 1180.0, + 414.0, + 1190.0, + 406.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1205.0, + 419.0, + 1205.0, + 419.0, + 1214.0, + 408.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1210.0, + 1200.0, + 1210.0, + 1200.0, + 1236.0, + 1183.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 1238.0, + 1204.0, + 1238.0, + 1204.0, + 1252.0, + 1189.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 1252.0, + 676.0, + 1252.0, + 676.0, + 1263.0, + 664.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1281.0, + 444.0, + 1281.0, + 444.0, + 1307.0, + 423.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1282.0, + 692.0, + 1282.0, + 692.0, + 1312.0, + 669.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 1305.0, + 814.0, + 1305.0, + 814.0, + 1313.0, + 799.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 1301.0, + 830.0, + 1301.0, + 830.0, + 1310.0, + 819.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 1282.0, + 1215.0, + 1282.0, + 1215.0, + 1312.0, + 1193.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1312.0, + 1226.0, + 1312.0, + 1226.0, + 1322.0, + 1216.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1318.0, + 1267.0, + 1318.0, + 1267.0, + 1327.0, + 1253.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 1321.0, + 1247.0, + 1321.0, + 1247.0, + 1329.0, + 1236.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1363.0, + 349.0, + 1363.0, + 349.0, + 1479.0, + 304.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1449.0, + 697.0, + 1449.0, + 697.0, + 1463.0, + 687.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 431.0, + 1508.0, + 444.0, + 1508.0, + 444.0, + 1519.0, + 431.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 1516.0, + 814.0, + 1516.0, + 814.0, + 1525.0, + 798.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 1516.0, + 1341.0, + 1516.0, + 1341.0, + 1526.0, + 1326.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 1531.0, + 492.0, + 1531.0, + 492.0, + 1541.0, + 478.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 1530.0, + 736.0, + 1530.0, + 736.0, + 1540.0, + 723.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 1531.0, + 1019.0, + 1531.0, + 1019.0, + 1540.0, + 1009.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 1533.0, + 1248.0, + 1533.0, + 1248.0, + 1542.0, + 1238.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 1529.0, + 1269.0, + 1529.0, + 1269.0, + 1542.0, + 1252.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 762.5, + 612.0, + 762.5, + 612.0, + 858.5, + 399.0, + 858.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.75, + 766.0, + 859.75, + 766.0, + 859.75, + 859.0, + 650.75, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 760.0, + 1144.0, + 760.0, + 1144.0, + 855.5, + 927.0, + 855.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 775.0, + 1383.0, + 775.0, + 1383.0, + 857.0, + 1195.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 960.5, + 1372.0, + 960.5, + 1372.0, + 1055.0, + 1164.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 973.5, + 603.0, + 973.5, + 603.0, + 1054.0, + 405.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 972.5, + 1136.0, + 972.5, + 1136.0, + 1055.5, + 932.0, + 1055.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1177.5, + 612.0, + 1177.5, + 612.0, + 1270.5, + 389.0, + 1270.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.75, + 1182.5, + 1143.75, + 1182.5, + 1143.75, + 1274.0, + 922.75, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.75, + 1275.5, + 678.75, + 1275.5, + 678.75, + 1284.5, + 674.75, + 1284.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 1409.0, + 1363.0, + 1409.0, + 1363.0, + 1504.0, + 1206.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 1397.0, + 608.0, + 1397.0, + 608.0, + 1480.0, + 403.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 1398.0, + 1137.0, + 1398.0, + 1137.0, + 1483.0, + 927.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1516.5, + 1334.0, + 1516.5, + 1334.0, + 1545.5, + 1262.0, + 1545.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.75, + 1531.0, + 1054.75, + 1531.0, + 1054.75, + 1543.0, + 1025.75, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 1575.0, + 1136.0, + 1575.0, + 1136.0, + 1619.0, + 562.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 829.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 636.0, + 792.0, + 636.0, + 792.0, + 672.0, + 452.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 634.0, + 1252.0, + 634.0, + 1252.0, + 669.0, + 981.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 664.0, + 561.0, + 664.0, + 561.0, + 709.0, + 437.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 668.0, + 818.0, + 668.0, + 818.0, + 700.0, + 671.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 664.0, + 1091.0, + 664.0, + 1091.0, + 709.0, + 967.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 669.0, + 1348.0, + 669.0, + 1348.0, + 701.0, + 1202.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 739.0, + 350.0, + 739.0, + 350.0, + 826.0, + 308.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 737.0, + 1203.0, + 737.0, + 1203.0, + 824.0, + 1176.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 849.0, + 684.0, + 849.0, + 684.0, + 859.0, + 670.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 846.0, + 1223.0, + 846.0, + 1223.0, + 885.0, + 1196.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 868.0, + 441.0, + 868.0, + 441.0, + 875.0, + 435.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 937.0, + 349.0, + 937.0, + 349.0, + 1053.0, + 304.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 976.0, + 663.0, + 976.0, + 663.0, + 994.0, + 654.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 986.0, + 1194.0, + 986.0, + 1194.0, + 996.0, + 1182.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 991.0, + 668.0, + 991.0, + 668.0, + 1019.0, + 654.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1009.0, + 1194.0, + 1009.0, + 1194.0, + 1019.0, + 1185.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 1059.0, + 685.0, + 1059.0, + 685.0, + 1074.0, + 668.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 1067.0, + 692.0, + 1067.0, + 692.0, + 1098.0, + 668.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 1074.0, + 1227.0, + 1074.0, + 1227.0, + 1101.0, + 1203.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 1104.0, + 737.0, + 1104.0, + 737.0, + 1114.0, + 723.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 1108.0, + 1248.0, + 1108.0, + 1248.0, + 1117.0, + 1238.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1104.0, + 1267.0, + 1104.0, + 1267.0, + 1114.0, + 1253.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1150.0, + 349.0, + 1150.0, + 349.0, + 1265.0, + 304.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1180.0, + 414.0, + 1180.0, + 414.0, + 1190.0, + 406.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1205.0, + 419.0, + 1205.0, + 419.0, + 1214.0, + 408.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1210.0, + 1200.0, + 1210.0, + 1200.0, + 1236.0, + 1183.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 1238.0, + 1204.0, + 1238.0, + 1204.0, + 1252.0, + 1189.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 1252.0, + 676.0, + 1252.0, + 676.0, + 1263.0, + 664.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1281.0, + 444.0, + 1281.0, + 444.0, + 1307.0, + 423.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1282.0, + 692.0, + 1282.0, + 692.0, + 1312.0, + 669.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 1305.0, + 814.0, + 1305.0, + 814.0, + 1313.0, + 799.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 1301.0, + 830.0, + 1301.0, + 830.0, + 1310.0, + 819.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 1282.0, + 1215.0, + 1282.0, + 1215.0, + 1312.0, + 1193.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1312.0, + 1226.0, + 1312.0, + 1226.0, + 1322.0, + 1216.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1318.0, + 1267.0, + 1318.0, + 1267.0, + 1327.0, + 1253.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 1321.0, + 1247.0, + 1321.0, + 1247.0, + 1329.0, + 1236.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1363.0, + 349.0, + 1363.0, + 349.0, + 1479.0, + 304.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1449.0, + 697.0, + 1449.0, + 697.0, + 1463.0, + 687.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 431.0, + 1508.0, + 444.0, + 1508.0, + 444.0, + 1519.0, + 431.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 1516.0, + 814.0, + 1516.0, + 814.0, + 1525.0, + 798.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 1516.0, + 1341.0, + 1516.0, + 1341.0, + 1526.0, + 1326.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 1531.0, + 492.0, + 1531.0, + 492.0, + 1541.0, + 478.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 1530.0, + 736.0, + 1530.0, + 736.0, + 1540.0, + 723.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 1531.0, + 1019.0, + 1531.0, + 1019.0, + 1540.0, + 1009.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 1533.0, + 1248.0, + 1533.0, + 1248.0, + 1542.0, + 1238.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 1529.0, + 1269.0, + 1529.0, + 1269.0, + 1542.0, + 1252.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 762.5, + 612.0, + 762.5, + 612.0, + 858.5, + 399.0, + 858.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.75, + 766.0, + 859.75, + 766.0, + 859.75, + 859.0, + 650.75, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 760.0, + 1144.0, + 760.0, + 1144.0, + 855.5, + 927.0, + 855.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 775.0, + 1383.0, + 775.0, + 1383.0, + 857.0, + 1195.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 960.5, + 1372.0, + 960.5, + 1372.0, + 1055.0, + 1164.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 973.5, + 603.0, + 973.5, + 603.0, + 1054.0, + 405.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 972.5, + 1136.0, + 972.5, + 1136.0, + 1055.5, + 932.0, + 1055.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1177.5, + 612.0, + 1177.5, + 612.0, + 1270.5, + 389.0, + 1270.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.75, + 1182.5, + 1143.75, + 1182.5, + 1143.75, + 1274.0, + 922.75, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.75, + 1275.5, + 678.75, + 1275.5, + 678.75, + 1284.5, + 674.75, + 1284.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 1409.0, + 1363.0, + 1409.0, + 1363.0, + 1504.0, + 1206.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 1397.0, + 608.0, + 1397.0, + 608.0, + 1480.0, + 403.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 1398.0, + 1137.0, + 1398.0, + 1137.0, + 1483.0, + 927.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1516.5, + 1334.0, + 1516.5, + 1334.0, + 1545.5, + 1262.0, + 1545.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.75, + 1531.0, + 1054.75, + 1531.0, + 1054.75, + 1543.0, + 1025.75, + 1543.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 21, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 301, + 641, + 1371, + 641, + 1371, + 1557, + 301, + 1557 + ], + "score": 0.972 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2113, + 834, + 2113 + ], + "score": 0.842 + }, + { + "category_id": 4, + "poly": [ + 559, + 1580, + 1142, + 1580, + 1142, + 1613, + 559, + 1613 + ], + "score": 0.728 + }, + { + "category_id": 4, + "poly": [ + 555, + 1580, + 1142, + 1580, + 1142, + 1613, + 555, + 1613 + ], + "score": 0.507 + }, + { + "category_id": 13, + "poly": [ + 1253, + 636, + 1314, + 636, + 1314, + 670, + 1253, + 670 + ], + "score": 0.81, + "latex": "( 1 0 ^ { 6 } )" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 636.0, + 792.0, + 636.0, + 792.0, + 672.0, + 452.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 634.0, + 1252.0, + 634.0, + 1252.0, + 672.0, + 981.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 634.0, + 1320.0, + 634.0, + 1320.0, + 672.0, + 1315.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 664.0, + 561.0, + 664.0, + 561.0, + 709.0, + 437.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 668.0, + 820.0, + 668.0, + 820.0, + 703.0, + 671.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 664.0, + 1091.0, + 664.0, + 1091.0, + 709.0, + 967.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 668.0, + 1348.0, + 668.0, + 1348.0, + 703.0, + 1200.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 739.0, + 350.0, + 739.0, + 350.0, + 827.0, + 308.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 733.0, + 798.0, + 733.0, + 798.0, + 762.0, + 780.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 757.0, + 1192.0, + 757.0, + 1192.0, + 766.0, + 1182.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 766.0, + 1192.0, + 766.0, + 1192.0, + 777.0, + 1183.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 777.0, + 1193.0, + 777.0, + 1193.0, + 789.0, + 1185.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 791.0, + 1193.0, + 791.0, + 1193.0, + 797.0, + 1187.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 845.0, + 692.0, + 845.0, + 692.0, + 862.0, + 667.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 893.0, + 492.0, + 893.0, + 492.0, + 901.0, + 478.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 890.0, + 741.0, + 890.0, + 741.0, + 902.0, + 720.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 888.0, + 796.0, + 888.0, + 796.0, + 900.0, + 775.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 895.0, + 1248.0, + 895.0, + 1248.0, + 904.0, + 1237.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 936.0, + 349.0, + 936.0, + 349.0, + 1053.0, + 304.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 1031.0, + 670.0, + 1031.0, + 670.0, + 1037.0, + 664.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 1029.0, + 1203.0, + 1029.0, + 1203.0, + 1038.0, + 1192.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1066.0, + 441.0, + 1066.0, + 441.0, + 1093.0, + 423.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 1058.0, + 968.0, + 1058.0, + 968.0, + 1085.0, + 950.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 1087.0, + 1222.0, + 1087.0, + 1222.0, + 1095.0, + 1209.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 1092.0, + 1341.0, + 1092.0, + 1341.0, + 1100.0, + 1331.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 1105.0, + 737.0, + 1105.0, + 737.0, + 1114.0, + 723.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 1106.0, + 1292.0, + 1106.0, + 1292.0, + 1119.0, + 1271.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1151.0, + 349.0, + 1151.0, + 349.0, + 1265.0, + 304.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 1171.0, + 1193.0, + 1171.0, + 1193.0, + 1215.0, + 1176.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 1253.0, + 675.0, + 1253.0, + 675.0, + 1263.0, + 664.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 1277.0, + 452.0, + 1277.0, + 452.0, + 1313.0, + 424.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1362.0, + 349.0, + 1362.0, + 349.0, + 1479.0, + 304.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1412.0, + 1194.0, + 1412.0, + 1194.0, + 1421.0, + 1183.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1420.0, + 1196.0, + 1420.0, + 1196.0, + 1432.0, + 1185.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 1517.0, + 813.0, + 1517.0, + 813.0, + 1525.0, + 799.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 1511.0, + 1224.0, + 1511.0, + 1224.0, + 1524.0, + 1204.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 1520.0, + 1321.0, + 1520.0, + 1321.0, + 1530.0, + 1306.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 1517.0, + 1341.0, + 1517.0, + 1341.0, + 1525.0, + 1326.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 1529.0, + 740.0, + 1529.0, + 740.0, + 1541.0, + 720.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 1531.0, + 1019.0, + 1531.0, + 1019.0, + 1540.0, + 1007.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 1533.0, + 1249.0, + 1533.0, + 1249.0, + 1542.0, + 1233.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 1529.0, + 1269.0, + 1529.0, + 1269.0, + 1542.0, + 1250.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 755.0, + 613.0, + 755.0, + 613.0, + 846.5, + 395.0, + 846.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 769.5, + 845.0, + 769.5, + 845.0, + 854.5, + 686.0, + 854.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.25, + 758.5, + 1139.25, + 758.5, + 1139.25, + 844.5, + 926.25, + 844.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 775.5, + 1382.0, + 775.5, + 1382.0, + 859.5, + 1193.0, + 859.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.25, + 819.0, + 1242.25, + 819.0, + 1242.25, + 892.0, + 1173.25, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 963.0, + 610.0, + 963.0, + 610.0, + 1055.0, + 397.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 964.0, + 1377.0, + 964.0, + 1377.0, + 1056.5, + 1171.0, + 1056.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.75, + 967.5, + 1139.75, + 967.5, + 1139.75, + 1059.5, + 925.75, + 1059.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1182.0, + 610.0, + 1182.0, + 610.0, + 1268.5, + 396.0, + 1268.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 1184.5, + 1136.0, + 1184.5, + 1136.0, + 1267.5, + 932.0, + 1267.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 1273.5, + 680.0, + 1273.5, + 680.0, + 1286.5, + 671.0, + 1286.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 1296.0, + 693.0, + 1296.0, + 693.0, + 1310.0, + 676.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.75, + 1298.0, + 1221.75, + 1298.0, + 1221.75, + 1310.0, + 1205.75, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.25, + 1392.5, + 1133.25, + 1392.5, + 1133.25, + 1480.5, + 931.25, + 1480.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1392.5, + 611.0, + 1392.5, + 611.0, + 1483.5, + 392.0, + 1483.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 1530.5, + 1060.0, + 1530.5, + 1060.0, + 1542.5, + 1025.0, + 1542.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 830.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1576.0, + 1145.0, + 1576.0, + 1145.0, + 1619.0, + 552.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1576.0, + 1145.0, + 1576.0, + 1145.0, + 1619.0, + 552.0, + 1619.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 22, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 301, + 641, + 1371, + 641, + 1371, + 1558, + 301, + 1558 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 582, + 1580, + 1120, + 1580, + 1120, + 1613, + 582, + 1613 + ], + "score": 0.914 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2112, + 834, + 2112 + ], + "score": 0.845 + }, + { + "category_id": 13, + "poly": [ + 1253, + 636, + 1314, + 636, + 1314, + 670, + 1253, + 670 + ], + "score": 0.8, + "latex": "( 1 0 ^ { 6 } )" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 636.0, + 792.0, + 636.0, + 792.0, + 672.0, + 452.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 635.0, + 1252.0, + 635.0, + 1252.0, + 670.0, + 982.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 635.0, + 1319.0, + 635.0, + 1319.0, + 670.0, + 1315.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 664.0, + 561.0, + 664.0, + 561.0, + 709.0, + 437.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 668.0, + 820.0, + 668.0, + 820.0, + 703.0, + 671.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 664.0, + 1091.0, + 664.0, + 1091.0, + 709.0, + 967.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 668.0, + 1349.0, + 668.0, + 1349.0, + 703.0, + 1200.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 739.0, + 350.0, + 739.0, + 350.0, + 827.0, + 308.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 799.0, + 951.0, + 799.0, + 951.0, + 809.0, + 936.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 747.0, + 1192.0, + 747.0, + 1192.0, + 810.0, + 1177.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 828.0, + 677.0, + 828.0, + 677.0, + 860.0, + 656.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 813.0, + 1207.0, + 813.0, + 1207.0, + 880.0, + 1169.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 841.0, + 966.0, + 841.0, + 966.0, + 877.0, + 942.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 892.0, + 736.0, + 892.0, + 736.0, + 900.0, + 726.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 937.0, + 349.0, + 937.0, + 349.0, + 1053.0, + 304.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 999.0, + 667.0, + 999.0, + 667.0, + 1009.0, + 654.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 1071.0, + 443.0, + 1071.0, + 443.0, + 1094.0, + 424.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 1080.0, + 1225.0, + 1080.0, + 1225.0, + 1099.0, + 1203.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1327.0, + 1091.0, + 1341.0, + 1091.0, + 1341.0, + 1101.0, + 1327.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 1103.0, + 740.0, + 1103.0, + 740.0, + 1115.0, + 720.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 1108.0, + 1248.0, + 1108.0, + 1248.0, + 1117.0, + 1238.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1104.0, + 1267.0, + 1104.0, + 1267.0, + 1114.0, + 1253.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 1108.0, + 1289.0, + 1108.0, + 1289.0, + 1117.0, + 1272.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1150.0, + 349.0, + 1150.0, + 349.0, + 1266.0, + 304.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 1204.0, + 668.0, + 1204.0, + 668.0, + 1234.0, + 650.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 1160.0, + 1194.0, + 1160.0, + 1194.0, + 1233.0, + 1168.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1241.0, + 675.0, + 1241.0, + 675.0, + 1251.0, + 662.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 1243.0, + 1202.0, + 1243.0, + 1202.0, + 1251.0, + 1192.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1251.0, + 680.0, + 1251.0, + 680.0, + 1265.0, + 663.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 1276.0, + 441.0, + 1276.0, + 441.0, + 1309.0, + 418.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 1283.0, + 970.0, + 1283.0, + 970.0, + 1293.0, + 957.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 1273.0, + 1204.0, + 1273.0, + 1204.0, + 1283.0, + 1194.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 1299.0, + 692.0, + 1299.0, + 692.0, + 1309.0, + 676.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1363.0, + 349.0, + 1363.0, + 349.0, + 1479.0, + 304.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1422.0, + 667.0, + 1422.0, + 667.0, + 1432.0, + 654.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1376.0, + 1197.0, + 1376.0, + 1197.0, + 1454.0, + 1169.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1454.0, + 671.0, + 1454.0, + 671.0, + 1464.0, + 662.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 1447.0, + 1205.0, + 1447.0, + 1205.0, + 1501.0, + 1173.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 1497.0, + 1222.0, + 1497.0, + 1222.0, + 1524.0, + 1197.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 1516.0, + 1341.0, + 1516.0, + 1341.0, + 1526.0, + 1326.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 1530.0, + 492.0, + 1530.0, + 492.0, + 1540.0, + 478.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 1530.0, + 737.0, + 1530.0, + 737.0, + 1540.0, + 721.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 1531.0, + 1019.0, + 1531.0, + 1019.0, + 1540.0, + 1008.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 1532.0, + 1248.0, + 1532.0, + 1248.0, + 1542.0, + 1235.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 1529.0, + 1269.0, + 1529.0, + 1269.0, + 1542.0, + 1250.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 1521.0, + 1324.0, + 1521.0, + 1324.0, + 1546.0, + 1270.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 766.0, + 601.0, + 766.0, + 601.0, + 843.0, + 411.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 773.0, + 847.0, + 773.0, + 847.0, + 846.5, + 666.0, + 846.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.75, + 761.0, + 1137.75, + 761.0, + 1137.75, + 852.5, + 927.75, + 852.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 974.5, + 606.0, + 974.5, + 606.0, + 1060.5, + 403.0, + 1060.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 970.5, + 1141.0, + 970.5, + 1141.0, + 1058.5, + 925.0, + 1058.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 1188.0, + 668.0, + 1188.0, + 668.0, + 1207.0, + 646.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 1185.0, + 1141.0, + 1185.0, + 1141.0, + 1271.5, + 924.0, + 1271.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1173.0, + 1378.0, + 1173.0, + 1378.0, + 1262.0, + 1157.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 1396.0, + 611.0, + 1396.0, + 611.0, + 1482.5, + 399.0, + 1482.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 1399.0, + 1135.0, + 1399.0, + 1135.0, + 1484.0, + 930.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 1496.0, + 978.0, + 1496.0, + 978.0, + 1519.5, + 950.0, + 1519.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 1506.5, + 445.0, + 1506.5, + 445.0, + 1519.5, + 429.0, + 1519.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.75, + 1531.0, + 1054.75, + 1531.0, + 1054.75, + 1543.0, + 1025.75, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 1577.0, + 1122.0, + 1577.0, + 1122.0, + 1619.0, + 576.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 872.0, + 2084.0, + 872.0, + 2124.0, + 829.0, + 2124.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 23, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 301, + 640, + 1370, + 640, + 1370, + 1558, + 301, + 1558 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 568, + 1580, + 1128, + 1580, + 1128, + 1613, + 568, + 1613 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2113, + 834, + 2113 + ], + "score": 0.843 + }, + { + "category_id": 13, + "poly": [ + 1253, + 636, + 1314, + 636, + 1314, + 670, + 1253, + 670 + ], + "score": 0.79, + "latex": "( 1 0 ^ { 6 } )" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 636.0, + 792.0, + 636.0, + 792.0, + 672.0, + 452.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 634.0, + 1252.0, + 634.0, + 1252.0, + 673.0, + 980.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 634.0, + 1320.0, + 634.0, + 1320.0, + 673.0, + 1315.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 664.0, + 561.0, + 664.0, + 561.0, + 709.0, + 438.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 667.0, + 820.0, + 667.0, + 820.0, + 703.0, + 671.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 666.0, + 1090.0, + 666.0, + 1090.0, + 709.0, + 966.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 667.0, + 1348.0, + 667.0, + 1348.0, + 703.0, + 1200.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 739.0, + 350.0, + 739.0, + 350.0, + 827.0, + 308.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 735.0, + 1194.0, + 735.0, + 1194.0, + 811.0, + 1174.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 774.0, + 1274.0, + 774.0, + 1274.0, + 862.0, + 1225.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 855.0, + 442.0, + 855.0, + 442.0, + 882.0, + 424.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 892.0, + 492.0, + 892.0, + 492.0, + 901.0, + 477.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 892.0, + 1267.0, + 892.0, + 1267.0, + 901.0, + 1252.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 937.0, + 349.0, + 937.0, + 349.0, + 1054.0, + 304.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 1007.0, + 670.0, + 1007.0, + 670.0, + 1024.0, + 655.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 961.0, + 1199.0, + 961.0, + 1199.0, + 1027.0, + 1170.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1028.0, + 674.0, + 1028.0, + 674.0, + 1039.0, + 661.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 1024.0, + 1203.0, + 1024.0, + 1203.0, + 1040.0, + 1189.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 1046.0, + 1203.0, + 1046.0, + 1203.0, + 1057.0, + 1191.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1065.0, + 442.0, + 1065.0, + 442.0, + 1096.0, + 421.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 1055.0, + 967.0, + 1055.0, + 967.0, + 1092.0, + 946.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 1104.0, + 736.0, + 1104.0, + 736.0, + 1114.0, + 726.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 1109.0, + 1247.0, + 1109.0, + 1247.0, + 1117.0, + 1236.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 1105.0, + 1267.0, + 1105.0, + 1267.0, + 1115.0, + 1252.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1151.0, + 349.0, + 1151.0, + 349.0, + 1266.0, + 304.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 1165.0, + 669.0, + 1165.0, + 669.0, + 1238.0, + 641.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1182.0, + 1192.0, + 1182.0, + 1192.0, + 1192.0, + 1179.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1281.0, + 441.0, + 1281.0, + 441.0, + 1309.0, + 420.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1236.0, + 677.0, + 1236.0, + 677.0, + 1314.0, + 642.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1272.0, + 966.0, + 1272.0, + 966.0, + 1309.0, + 944.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 1318.0, + 1267.0, + 1318.0, + 1267.0, + 1327.0, + 1252.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 1320.0, + 1247.0, + 1320.0, + 1247.0, + 1329.0, + 1236.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1364.0, + 349.0, + 1364.0, + 349.0, + 1479.0, + 304.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 1393.0, + 1191.0, + 1393.0, + 1191.0, + 1405.0, + 1180.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 1404.0, + 669.0, + 1404.0, + 669.0, + 1417.0, + 650.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 1412.0, + 672.0, + 1412.0, + 672.0, + 1433.0, + 650.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 1428.0, + 671.0, + 1428.0, + 671.0, + 1443.0, + 653.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1455.0, + 672.0, + 1455.0, + 672.0, + 1464.0, + 661.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 1454.0, + 1202.0, + 1454.0, + 1202.0, + 1465.0, + 1191.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 1531.0, + 492.0, + 1531.0, + 492.0, + 1541.0, + 477.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1533.0, + 719.0, + 1533.0, + 719.0, + 1542.0, + 706.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 1531.0, + 741.0, + 1531.0, + 741.0, + 1539.0, + 721.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 1531.0, + 1018.0, + 1531.0, + 1018.0, + 1541.0, + 1008.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1533.0, + 1247.0, + 1533.0, + 1247.0, + 1542.0, + 1237.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 1531.0, + 1267.0, + 1531.0, + 1267.0, + 1539.0, + 1252.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 748.5, + 610.0, + 748.5, + 610.0, + 845.0, + 385.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.75, + 760.0, + 1138.75, + 760.0, + 1138.75, + 846.0, + 921.75, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.25, + 765.5, + 827.25, + 765.5, + 827.25, + 835.5, + 625.25, + 835.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 844.0, + 979.0, + 844.0, + 979.0, + 878.0, + 945.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 887.5, + 756.0, + 887.5, + 756.0, + 905.5, + 719.0, + 905.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 991.0, + 702.0, + 991.0, + 702.0, + 1013.0, + 689.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 978.5, + 1263.0, + 978.5, + 1263.0, + 1048.5, + 1197.0, + 1048.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 955.5, + 600.0, + 955.5, + 600.0, + 1045.5, + 395.0, + 1045.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.25, + 969.0, + 1135.25, + 969.0, + 1135.25, + 1051.5, + 924.25, + 1051.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1035.0, + 694.0, + 1035.0, + 694.0, + 1079.5, + 648.0, + 1079.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 1176.5, + 614.0, + 1176.5, + 614.0, + 1273.0, + 386.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.25, + 1183.5, + 1134.25, + 1183.5, + 1134.25, + 1266.5, + 931.25, + 1266.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.75, + 1414.5, + 466.75, + 1414.5, + 466.75, + 1447.0, + 455.75, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 1414.0, + 537.0, + 1414.0, + 537.0, + 1468.5, + 529.0, + 1468.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.25, + 1418.5, + 702.25, + 1418.5, + 702.25, + 1429.5, + 698.25, + 1429.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1388.5, + 1140.0, + 1388.5, + 1140.0, + 1487.0, + 915.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1464.5, + 685.0, + 1464.5, + 685.0, + 1491.0, + 654.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 1577.0, + 1133.0, + 1577.0, + 1133.0, + 1619.0, + 565.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2083.0, + 871.0, + 2083.0, + 871.0, + 2124.0, + 829.0, + 2124.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 24, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 336, + 461, + 1436, + 461, + 1436, + 1737, + 336, + 1737 + ], + "score": 0.965 + }, + { + "category_id": 4, + "poly": [ + 353, + 1757, + 1344, + 1757, + 1344, + 1792, + 353, + 1792 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 816, + 76, + 816, + 105, + 299, + 105 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2113, + 834, + 2113 + ], + "score": 0.853 + }, + { + "category_id": 2, + "poly": [ + 311, + 592, + 344, + 592, + 344, + 669, + 311, + 669 + ], + "score": 0.158 + }, + { + "category_id": 15, + "poly": [ + 425.0, + 457.0, + 691.0, + 457.0, + 691.0, + 491.0, + 425.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 457.0, + 1065.0, + 457.0, + 1065.0, + 491.0, + 784.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 457.0, + 1440.0, + 457.0, + 1440.0, + 491.0, + 1146.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 501.0, + 625.0, + 501.0, + 625.0, + 524.0, + 494.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 498.0, + 994.0, + 498.0, + 994.0, + 526.0, + 855.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 501.0, + 1365.0, + 501.0, + 1365.0, + 524.0, + 1218.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 543.0, + 442.0, + 543.0, + 442.0, + 563.0, + 413.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 541.0, + 807.0, + 541.0, + 807.0, + 561.0, + 778.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 541.0, + 1170.0, + 541.0, + 1170.0, + 559.0, + 1146.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 556.0, + 440.0, + 556.0, + 440.0, + 577.0, + 414.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 557.0, + 804.0, + 557.0, + 804.0, + 574.0, + 782.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 556.0, + 1170.0, + 556.0, + 1170.0, + 573.0, + 1147.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 570.0, + 442.0, + 570.0, + 442.0, + 592.0, + 416.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 570.0, + 808.0, + 570.0, + 808.0, + 592.0, + 781.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 572.0, + 1172.0, + 572.0, + 1172.0, + 589.0, + 1148.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 586.0, + 440.0, + 586.0, + 440.0, + 603.0, + 420.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 586.0, + 807.0, + 586.0, + 807.0, + 603.0, + 784.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 586.0, + 1173.0, + 586.0, + 1173.0, + 603.0, + 1151.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 596.0, + 445.0, + 596.0, + 445.0, + 617.0, + 419.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 599.0, + 810.0, + 599.0, + 810.0, + 616.0, + 786.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 599.0, + 1174.0, + 599.0, + 1174.0, + 616.0, + 1153.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 612.0, + 811.0, + 612.0, + 811.0, + 629.0, + 788.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 613.0, + 1176.0, + 613.0, + 1176.0, + 630.0, + 1153.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 625.0, + 446.0, + 625.0, + 446.0, + 642.0, + 423.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 629.0, + 808.0, + 629.0, + 808.0, + 640.0, + 791.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 627.0, + 1177.0, + 627.0, + 1177.0, + 643.0, + 1153.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 658.0, + 452.0, + 658.0, + 452.0, + 673.0, + 432.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 658.0, + 818.0, + 658.0, + 818.0, + 673.0, + 798.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 660.0, + 1180.0, + 660.0, + 1180.0, + 672.0, + 1164.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 675.0, + 456.0, + 675.0, + 456.0, + 691.0, + 437.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 678.0, + 821.0, + 678.0, + 821.0, + 689.0, + 808.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 695.0, + 461.0, + 695.0, + 461.0, + 706.0, + 445.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 695.0, + 828.0, + 695.0, + 828.0, + 705.0, + 814.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 695.0, + 1192.0, + 695.0, + 1192.0, + 706.0, + 1177.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 712.0, + 468.0, + 712.0, + 468.0, + 723.0, + 450.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 712.0, + 834.0, + 712.0, + 834.0, + 723.0, + 820.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 723.0, + 453.0, + 723.0, + 453.0, + 731.0, + 446.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 728.0, + 478.0, + 728.0, + 478.0, + 742.0, + 455.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 729.0, + 840.0, + 729.0, + 840.0, + 741.0, + 825.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 729.0, + 1206.0, + 729.0, + 1206.0, + 741.0, + 1189.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 765.0, + 488.0, + 765.0, + 488.0, + 777.0, + 472.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 775.0, + 520.0, + 775.0, + 520.0, + 791.0, + 498.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 769.0, + 548.0, + 769.0, + 548.0, + 785.0, + 524.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 759.0, + 605.0, + 759.0, + 605.0, + 775.0, + 582.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 755.0, + 632.0, + 755.0, + 632.0, + 771.0, + 608.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 752.0, + 657.0, + 752.0, + 657.0, + 764.0, + 641.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 745.0, + 686.0, + 745.0, + 686.0, + 761.0, + 662.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 736.0, + 851.0, + 736.0, + 851.0, + 777.0, + 814.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 775.0, + 886.0, + 775.0, + 886.0, + 791.0, + 864.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 769.0, + 913.0, + 769.0, + 913.0, + 785.0, + 892.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 766.0, + 941.0, + 766.0, + 941.0, + 778.0, + 922.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 762.0, + 968.0, + 762.0, + 968.0, + 774.0, + 949.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 755.0, + 998.0, + 755.0, + 998.0, + 771.0, + 974.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 752.0, + 1023.0, + 752.0, + 1023.0, + 764.0, + 1006.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 748.0, + 1050.0, + 748.0, + 1050.0, + 759.0, + 1033.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 765.0, + 1219.0, + 765.0, + 1219.0, + 777.0, + 1203.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 775.0, + 1252.0, + 775.0, + 1252.0, + 791.0, + 1229.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 769.0, + 1280.0, + 769.0, + 1280.0, + 785.0, + 1257.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 765.0, + 1307.0, + 765.0, + 1307.0, + 781.0, + 1286.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 759.0, + 1335.0, + 759.0, + 1335.0, + 775.0, + 1313.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1339.0, + 756.0, + 1362.0, + 756.0, + 1362.0, + 771.0, + 1339.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1371.0, + 752.0, + 1388.0, + 752.0, + 1388.0, + 764.0, + 1371.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 748.0, + 1415.0, + 748.0, + 1415.0, + 759.0, + 1398.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 814.0, + 625.0, + 814.0, + 625.0, + 832.0, + 495.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 812.0, + 993.0, + 812.0, + 993.0, + 835.0, + 856.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 812.0, + 1365.0, + 812.0, + 1365.0, + 835.0, + 1218.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 848.0, + 443.0, + 848.0, + 443.0, + 864.0, + 414.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 850.0, + 811.0, + 850.0, + 811.0, + 865.0, + 781.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 863.0, + 439.0, + 863.0, + 439.0, + 880.0, + 414.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 865.0, + 804.0, + 865.0, + 804.0, + 883.0, + 782.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 865.0, + 1000.0, + 865.0, + 1000.0, + 885.0, + 983.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 868.0, + 1170.0, + 868.0, + 1170.0, + 883.0, + 1147.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 878.0, + 440.0, + 878.0, + 440.0, + 911.0, + 417.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 883.0, + 808.0, + 883.0, + 808.0, + 917.0, + 784.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 890.0, + 1172.0, + 890.0, + 1172.0, + 905.0, + 1150.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 908.0, + 445.0, + 908.0, + 445.0, + 940.0, + 420.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 914.0, + 812.0, + 914.0, + 812.0, + 936.0, + 785.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 911.0, + 1173.0, + 911.0, + 1173.0, + 927.0, + 1153.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 937.0, + 445.0, + 937.0, + 445.0, + 954.0, + 423.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 931.0, + 812.0, + 931.0, + 812.0, + 949.0, + 788.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 936.0, + 1174.0, + 936.0, + 1174.0, + 947.0, + 1157.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 967.0, + 452.0, + 967.0, + 452.0, + 984.0, + 430.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 967.0, + 820.0, + 967.0, + 820.0, + 984.0, + 797.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 967.0, + 1183.0, + 967.0, + 1183.0, + 984.0, + 1163.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 984.0, + 831.0, + 984.0, + 831.0, + 1074.0, + 776.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 982.0, + 1200.0, + 982.0, + 1200.0, + 1036.0, + 1145.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 1072.0, + 489.0, + 1072.0, + 489.0, + 1089.0, + 469.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 1085.0, + 520.0, + 1085.0, + 520.0, + 1102.0, + 498.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1080.0, + 548.0, + 1080.0, + 548.0, + 1096.0, + 524.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 1076.0, + 577.0, + 1076.0, + 577.0, + 1090.0, + 554.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 1069.0, + 603.0, + 1069.0, + 603.0, + 1085.0, + 583.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1063.0, + 635.0, + 1063.0, + 635.0, + 1083.0, + 606.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 1060.0, + 660.0, + 1060.0, + 660.0, + 1076.0, + 636.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1056.0, + 686.0, + 1056.0, + 686.0, + 1072.0, + 662.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 1073.0, + 856.0, + 1073.0, + 856.0, + 1089.0, + 836.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1085.0, + 885.0, + 1085.0, + 885.0, + 1100.0, + 864.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1080.0, + 913.0, + 1080.0, + 913.0, + 1096.0, + 890.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1075.0, + 942.0, + 1075.0, + 942.0, + 1090.0, + 921.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 1069.0, + 971.0, + 1069.0, + 971.0, + 1085.0, + 948.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 1056.0, + 1052.0, + 1056.0, + 1052.0, + 1083.0, + 973.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 1073.0, + 1221.0, + 1073.0, + 1221.0, + 1089.0, + 1200.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 1085.0, + 1254.0, + 1085.0, + 1254.0, + 1100.0, + 1229.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 1080.0, + 1280.0, + 1080.0, + 1280.0, + 1096.0, + 1257.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 1075.0, + 1307.0, + 1075.0, + 1307.0, + 1090.0, + 1287.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 1069.0, + 1335.0, + 1069.0, + 1335.0, + 1085.0, + 1313.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1336.0, + 1063.0, + 1365.0, + 1063.0, + 1365.0, + 1083.0, + 1336.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 1056.0, + 1418.0, + 1056.0, + 1418.0, + 1076.0, + 1368.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1122.0, + 625.0, + 1122.0, + 625.0, + 1145.0, + 495.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 1122.0, + 994.0, + 1122.0, + 994.0, + 1145.0, + 856.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 1122.0, + 1365.0, + 1122.0, + 1365.0, + 1145.0, + 1218.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 1162.0, + 443.0, + 1162.0, + 443.0, + 1195.0, + 413.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 1161.0, + 810.0, + 1161.0, + 810.0, + 1274.0, + 771.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1175.0, + 1170.0, + 1175.0, + 1170.0, + 1191.0, + 1147.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 1194.0, + 440.0, + 1194.0, + 440.0, + 1211.0, + 417.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 1194.0, + 1173.0, + 1194.0, + 1173.0, + 1211.0, + 1148.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 1207.0, + 443.0, + 1207.0, + 443.0, + 1228.0, + 417.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 1212.0, + 1173.0, + 1212.0, + 1173.0, + 1229.0, + 1151.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1224.0, + 446.0, + 1224.0, + 446.0, + 1255.0, + 422.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1231.0, + 1176.0, + 1231.0, + 1176.0, + 1247.0, + 1153.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 1251.0, + 445.0, + 1251.0, + 445.0, + 1268.0, + 424.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1249.0, + 1176.0, + 1249.0, + 1176.0, + 1265.0, + 1156.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1278.0, + 452.0, + 1278.0, + 452.0, + 1294.0, + 432.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 1278.0, + 818.0, + 1278.0, + 818.0, + 1294.0, + 797.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 1277.0, + 1182.0, + 1277.0, + 1182.0, + 1294.0, + 1163.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 1294.0, + 1193.0, + 1294.0, + 1193.0, + 1385.0, + 1140.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1334.0, + 1190.0, + 1334.0, + 1190.0, + 1348.0, + 1177.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 1383.0, + 489.0, + 1383.0, + 489.0, + 1400.0, + 469.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 1394.0, + 521.0, + 1394.0, + 521.0, + 1411.0, + 498.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1389.0, + 548.0, + 1389.0, + 548.0, + 1406.0, + 524.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1386.0, + 579.0, + 1386.0, + 579.0, + 1401.0, + 553.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 1373.0, + 635.0, + 1373.0, + 635.0, + 1396.0, + 582.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1370.0, + 658.0, + 1370.0, + 658.0, + 1386.0, + 638.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1366.0, + 686.0, + 1366.0, + 686.0, + 1381.0, + 662.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 1384.0, + 854.0, + 1384.0, + 854.0, + 1400.0, + 836.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1396.0, + 885.0, + 1396.0, + 885.0, + 1411.0, + 864.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 1390.0, + 913.0, + 1390.0, + 913.0, + 1406.0, + 892.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1386.0, + 942.0, + 1386.0, + 942.0, + 1401.0, + 919.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 1366.0, + 1052.0, + 1366.0, + 1052.0, + 1396.0, + 948.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 1383.0, + 1219.0, + 1383.0, + 1219.0, + 1400.0, + 1200.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 1394.0, + 1254.0, + 1394.0, + 1254.0, + 1411.0, + 1229.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 1390.0, + 1280.0, + 1390.0, + 1280.0, + 1406.0, + 1257.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 1380.0, + 1335.0, + 1380.0, + 1335.0, + 1401.0, + 1286.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 1370.0, + 1392.0, + 1370.0, + 1392.0, + 1386.0, + 1368.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1367.0, + 1415.0, + 1367.0, + 1415.0, + 1379.0, + 1397.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1432.0, + 625.0, + 1432.0, + 625.0, + 1454.0, + 495.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 1432.0, + 994.0, + 1432.0, + 994.0, + 1454.0, + 856.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 1432.0, + 1361.0, + 1432.0, + 1361.0, + 1454.0, + 1222.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 1479.0, + 436.0, + 1479.0, + 436.0, + 1490.0, + 417.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 1470.0, + 810.0, + 1470.0, + 810.0, + 1486.0, + 779.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 1466.0, + 1170.0, + 1466.0, + 1170.0, + 1496.0, + 1146.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 1489.0, + 804.0, + 1489.0, + 804.0, + 1505.0, + 782.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1492.0, + 1170.0, + 1492.0, + 1170.0, + 1509.0, + 1147.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1503.0, + 439.0, + 1503.0, + 439.0, + 1513.0, + 420.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 1503.0, + 808.0, + 1503.0, + 808.0, + 1525.0, + 782.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 1505.0, + 1172.0, + 1505.0, + 1172.0, + 1522.0, + 1148.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1526.0, + 440.0, + 1526.0, + 440.0, + 1536.0, + 422.0, + 1536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 1522.0, + 807.0, + 1522.0, + 807.0, + 1539.0, + 785.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 1512.0, + 1179.0, + 1512.0, + 1179.0, + 1552.0, + 1146.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1545.0, + 446.0, + 1545.0, + 446.0, + 1561.0, + 422.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1539.0, + 811.0, + 1539.0, + 811.0, + 1556.0, + 786.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1542.0, + 1176.0, + 1542.0, + 1176.0, + 1559.0, + 1153.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1556.0, + 811.0, + 1556.0, + 811.0, + 1573.0, + 789.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 1555.0, + 1176.0, + 1555.0, + 1176.0, + 1571.0, + 1154.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 1566.0, + 446.0, + 1566.0, + 446.0, + 1582.0, + 424.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1588.0, + 452.0, + 1588.0, + 452.0, + 1605.0, + 432.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 1588.0, + 818.0, + 1588.0, + 818.0, + 1605.0, + 797.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 1588.0, + 1183.0, + 1588.0, + 1183.0, + 1606.0, + 1163.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 1622.0, + 462.0, + 1622.0, + 462.0, + 1638.0, + 443.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 1603.0, + 1191.0, + 1603.0, + 1191.0, + 1660.0, + 1146.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1639.0, + 469.0, + 1639.0, + 469.0, + 1655.0, + 449.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 1649.0, + 488.0, + 1649.0, + 488.0, + 1698.0, + 442.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1652.0, + 456.0, + 1652.0, + 456.0, + 1667.0, + 445.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 1685.0, + 635.0, + 1685.0, + 635.0, + 1701.0, + 608.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 1682.0, + 657.0, + 1682.0, + 657.0, + 1694.0, + 639.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1675.0, + 686.0, + 1675.0, + 686.0, + 1691.0, + 662.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1685.0, + 998.0, + 1685.0, + 998.0, + 1701.0, + 974.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 1681.0, + 1026.0, + 1681.0, + 1026.0, + 1697.0, + 1003.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1675.0, + 1052.0, + 1675.0, + 1052.0, + 1691.0, + 1029.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 1681.0, + 1391.0, + 1681.0, + 1391.0, + 1695.0, + 1368.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1678.0, + 1415.0, + 1678.0, + 1415.0, + 1690.0, + 1397.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 1694.0, + 489.0, + 1694.0, + 489.0, + 1710.0, + 469.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 1705.0, + 521.0, + 1705.0, + 521.0, + 1721.0, + 498.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1701.0, + 547.0, + 1701.0, + 547.0, + 1717.0, + 524.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 1695.0, + 579.0, + 1695.0, + 579.0, + 1711.0, + 554.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1697.0, + 853.0, + 1697.0, + 853.0, + 1708.0, + 837.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 1707.0, + 885.0, + 1707.0, + 885.0, + 1718.0, + 866.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1701.0, + 913.0, + 1701.0, + 913.0, + 1717.0, + 890.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1697.0, + 941.0, + 1697.0, + 941.0, + 1708.0, + 922.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 1692.0, + 971.0, + 1692.0, + 971.0, + 1707.0, + 948.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 1694.0, + 1221.0, + 1694.0, + 1221.0, + 1710.0, + 1202.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 1704.0, + 1254.0, + 1704.0, + 1254.0, + 1720.0, + 1229.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 1701.0, + 1280.0, + 1701.0, + 1280.0, + 1717.0, + 1257.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 1695.0, + 1311.0, + 1695.0, + 1311.0, + 1711.0, + 1287.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 608.0, + 450.0, + 608.0, + 450.0, + 629.5, + 416.0, + 629.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.25, + 584.5, + 720.25, + 584.5, + 720.25, + 711.5, + 411.25, + 711.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.25, + 581.0, + 1092.25, + 581.0, + 1092.25, + 714.0, + 773.25, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 582.5, + 1456.0, + 582.5, + 1456.0, + 711.0, + 1141.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 726.5, + 819.0, + 726.5, + 819.0, + 732.5, + 816.0, + 732.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 737.0, + 1215.0, + 737.0, + 1215.0, + 763.0, + 1179.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.75, + 763.5, + 579.75, + 763.5, + 579.75, + 779.5, + 550.75, + 779.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.25, + 768.0, + 639.25, + 768.0, + 639.25, + 788.0, + 550.25, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 768.5, + 1004.0, + 768.5, + 1004.0, + 787.5, + 918.0, + 787.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 769.5, + 1369.0, + 769.5, + 1369.0, + 788.0, + 1283.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.75, + 890.5, + 1077.75, + 890.5, + 1077.75, + 1012.0, + 790.75, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.25, + 880.5, + 1451.25, + 880.5, + 1451.25, + 1015.0, + 1150.25, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.25, + 974.0, + 505.25, + 974.0, + 505.25, + 1084.5, + 405.25, + 1084.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 1019.5, + 1227.0, + 1019.5, + 1227.0, + 1079.5, + 1162.0, + 1079.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1079.0, + 639.0, + 1079.0, + 639.0, + 1097.5, + 552.0, + 1097.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.25, + 1077.5, + 1005.25, + 1077.5, + 1005.25, + 1099.0, + 915.25, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 1077.0, + 1371.0, + 1077.0, + 1371.0, + 1100.0, + 1280.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 1170.0, + 720.0, + 1170.0, + 720.0, + 1352.0, + 415.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 1287.0, + 479.0, + 1287.0, + 479.0, + 1348.0, + 415.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 1286.0, + 850.0, + 1286.0, + 850.0, + 1355.0, + 779.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1326.0, + 464.0, + 1326.0, + 464.0, + 1346.0, + 451.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.75, + 1330.0, + 494.75, + 1330.0, + 494.75, + 1390.0, + 430.75, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 1328.5, + 863.0, + 1328.5, + 863.0, + 1391.5, + 795.0, + 1391.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1336.75, + 1374.0, + 1363.75, + 1374.0, + 1363.75, + 1391.0, + 1336.75, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 1386.5, + 641.0, + 1386.5, + 641.0, + 1409.5, + 549.0, + 1409.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1388.0, + 1006.0, + 1388.0, + 1006.0, + 1409.5, + 915.0, + 1409.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 1389.5, + 1370.0, + 1389.5, + 1370.0, + 1407.5, + 1282.0, + 1407.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.25, + 1597.0, + 850.25, + 1597.0, + 850.25, + 1664.5, + 779.25, + 1664.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 1641.0, + 860.0, + 1641.0, + 860.0, + 1701.0, + 797.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 1639.5, + 1226.0, + 1639.5, + 1226.0, + 1699.0, + 1161.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1336.75, + 1684.0, + 1363.75, + 1684.0, + 1363.75, + 1699.5, + 1336.75, + 1699.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 1688.0, + 609.0, + 1688.0, + 609.0, + 1705.5, + 580.0, + 1705.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 1690.0, + 1338.0, + 1690.0, + 1338.0, + 1706.0, + 1307.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 1698.5, + 639.0, + 1698.5, + 639.0, + 1719.0, + 550.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1697.5, + 1006.0, + 1697.5, + 1006.0, + 1720.5, + 915.0, + 1720.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.25, + 1698.5, + 1370.25, + 1698.5, + 1370.25, + 1718.5, + 1281.25, + 1718.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1753.0, + 1347.0, + 1753.0, + 1347.0, + 1798.0, + 350.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 71.0, + 817.0, + 71.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2121.0, + 830.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 585.0, + 349.0, + 585.0, + 349.0, + 672.0, + 306.0, + 672.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 25, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 337, + 467, + 1429, + 467, + 1429, + 1737, + 337, + 1737 + ], + "score": 0.966 + }, + { + "category_id": 4, + "poly": [ + 363, + 1757, + 1334, + 1757, + 1334, + 1792, + 363, + 1792 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 816, + 76, + 816, + 105, + 299, + 105 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 864, + 2088, + 864, + 2112, + 834, + 2112 + ], + "score": 0.855 + }, + { + "category_id": 4, + "poly": [ + 1148, + 461, + 1436, + 461, + 1436, + 490, + 1148, + 490 + ], + "score": 0.499 + }, + { + "category_id": 0, + "poly": [ + 426, + 461, + 688, + 461, + 688, + 490, + 426, + 490 + ], + "score": 0.252 + }, + { + "category_id": 4, + "poly": [ + 786, + 461, + 1062, + 461, + 1062, + 490, + 786, + 490 + ], + "score": 0.225 + }, + { + "category_id": 2, + "poly": [ + 311, + 592, + 344, + 592, + 344, + 669, + 311, + 669 + ], + "score": 0.204 + }, + { + "category_id": 4, + "poly": [ + 426, + 461, + 688, + 461, + 688, + 490, + 426, + 490 + ], + "score": 0.128 + }, + { + "category_id": 15, + "poly": [ + 425.0, + 460.0, + 693.0, + 460.0, + 693.0, + 493.0, + 425.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 458.0, + 1067.0, + 458.0, + 1067.0, + 493.0, + 784.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 458.0, + 1436.0, + 458.0, + 1436.0, + 493.0, + 1145.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 498.0, + 622.0, + 498.0, + 622.0, + 526.0, + 494.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 498.0, + 995.0, + 498.0, + 995.0, + 526.0, + 856.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 498.0, + 1363.0, + 498.0, + 1363.0, + 526.0, + 1216.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 538.0, + 437.0, + 538.0, + 437.0, + 553.0, + 415.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 548.0, + 437.0, + 548.0, + 437.0, + 564.0, + 417.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 551.0, + 801.0, + 551.0, + 801.0, + 563.0, + 783.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 544.0, + 1170.0, + 544.0, + 1170.0, + 560.0, + 1147.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 560.0, + 444.0, + 560.0, + 444.0, + 647.0, + 410.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 565.0, + 806.0, + 565.0, + 806.0, + 604.0, + 786.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 573.0, + 1175.0, + 573.0, + 1175.0, + 605.0, + 1149.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 607.0, + 807.0, + 607.0, + 807.0, + 618.0, + 788.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 598.0, + 1175.0, + 598.0, + 1175.0, + 620.0, + 1151.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 618.0, + 811.0, + 618.0, + 811.0, + 635.0, + 788.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 614.0, + 1177.0, + 614.0, + 1177.0, + 631.0, + 1154.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 632.0, + 811.0, + 632.0, + 811.0, + 648.0, + 790.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 625.0, + 1180.0, + 625.0, + 1180.0, + 647.0, + 1152.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 658.0, + 453.0, + 658.0, + 453.0, + 674.0, + 432.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 658.0, + 817.0, + 658.0, + 817.0, + 674.0, + 798.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 658.0, + 1184.0, + 658.0, + 1184.0, + 674.0, + 1162.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 665.0, + 463.0, + 665.0, + 463.0, + 730.0, + 427.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 694.0, + 827.0, + 694.0, + 827.0, + 705.0, + 811.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 674.0, + 1194.0, + 674.0, + 1194.0, + 764.0, + 1141.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 710.0, + 470.0, + 710.0, + 470.0, + 727.0, + 449.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 710.0, + 837.0, + 710.0, + 837.0, + 725.0, + 816.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 717.0, + 847.0, + 717.0, + 847.0, + 751.0, + 806.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 767.0, + 485.0, + 767.0, + 485.0, + 778.0, + 472.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 775.0, + 521.0, + 775.0, + 521.0, + 791.0, + 500.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 771.0, + 548.0, + 771.0, + 548.0, + 785.0, + 523.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 768.0, + 576.0, + 768.0, + 576.0, + 778.0, + 556.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 761.0, + 605.0, + 761.0, + 605.0, + 775.0, + 581.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 758.0, + 628.0, + 758.0, + 628.0, + 768.0, + 611.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 752.0, + 657.0, + 752.0, + 657.0, + 764.0, + 639.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 748.0, + 684.0, + 748.0, + 684.0, + 758.0, + 668.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 765.0, + 851.0, + 765.0, + 851.0, + 778.0, + 837.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 775.0, + 887.0, + 775.0, + 887.0, + 791.0, + 864.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 771.0, + 913.0, + 771.0, + 913.0, + 785.0, + 890.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 768.0, + 937.0, + 768.0, + 937.0, + 778.0, + 922.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 759.0, + 972.0, + 759.0, + 972.0, + 775.0, + 947.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 755.0, + 996.0, + 755.0, + 996.0, + 771.0, + 975.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 751.0, + 1025.0, + 751.0, + 1025.0, + 765.0, + 1002.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 745.0, + 1051.0, + 745.0, + 1051.0, + 761.0, + 1031.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 764.0, + 1220.0, + 764.0, + 1220.0, + 779.0, + 1200.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 774.0, + 1254.0, + 774.0, + 1254.0, + 791.0, + 1230.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 771.0, + 1280.0, + 771.0, + 1280.0, + 787.0, + 1257.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 762.0, + 1333.0, + 762.0, + 1333.0, + 774.0, + 1316.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1340.0, + 755.0, + 1363.0, + 755.0, + 1363.0, + 771.0, + 1340.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1369.0, + 751.0, + 1389.0, + 751.0, + 1389.0, + 767.0, + 1369.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 745.0, + 1417.0, + 745.0, + 1417.0, + 761.0, + 1394.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 812.0, + 625.0, + 812.0, + 625.0, + 835.0, + 499.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 812.0, + 990.0, + 812.0, + 990.0, + 835.0, + 857.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 814.0, + 1360.0, + 814.0, + 1360.0, + 832.0, + 1220.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 847.0, + 440.0, + 847.0, + 440.0, + 862.0, + 415.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 847.0, + 804.0, + 847.0, + 804.0, + 862.0, + 781.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 849.0, + 1174.0, + 849.0, + 1174.0, + 865.0, + 1142.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 862.0, + 443.0, + 862.0, + 443.0, + 925.0, + 415.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 859.0, + 812.0, + 859.0, + 812.0, + 967.0, + 774.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 859.0, + 1183.0, + 859.0, + 1183.0, + 950.0, + 1138.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 922.0, + 445.0, + 922.0, + 445.0, + 939.0, + 422.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 935.0, + 447.0, + 935.0, + 447.0, + 956.0, + 422.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 941.0, + 1178.0, + 941.0, + 1178.0, + 955.0, + 1151.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 968.0, + 453.0, + 968.0, + 453.0, + 985.0, + 430.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 966.0, + 817.0, + 966.0, + 817.0, + 985.0, + 797.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 965.0, + 1185.0, + 965.0, + 1185.0, + 986.0, + 1161.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1019.0, + 834.0, + 1019.0, + 834.0, + 1035.0, + 813.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 1018.0, + 1200.0, + 1018.0, + 1200.0, + 1036.0, + 1180.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 1073.0, + 490.0, + 1073.0, + 490.0, + 1089.0, + 469.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 1085.0, + 521.0, + 1085.0, + 521.0, + 1101.0, + 498.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1081.0, + 549.0, + 1081.0, + 549.0, + 1096.0, + 523.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1076.0, + 579.0, + 1076.0, + 579.0, + 1091.0, + 553.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1061.0, + 659.0, + 1061.0, + 659.0, + 1076.0, + 637.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 1058.0, + 685.0, + 1058.0, + 685.0, + 1069.0, + 667.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 1076.0, + 854.0, + 1076.0, + 854.0, + 1088.0, + 836.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 1085.0, + 887.0, + 1085.0, + 887.0, + 1101.0, + 863.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1081.0, + 915.0, + 1081.0, + 915.0, + 1096.0, + 890.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1076.0, + 945.0, + 1076.0, + 945.0, + 1091.0, + 919.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 1066.0, + 999.0, + 1066.0, + 999.0, + 1085.0, + 947.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 1061.0, + 1025.0, + 1061.0, + 1025.0, + 1075.0, + 1002.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1058.0, + 1049.0, + 1058.0, + 1049.0, + 1069.0, + 1032.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 1073.0, + 1223.0, + 1073.0, + 1223.0, + 1089.0, + 1200.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1228.0, + 1085.0, + 1253.0, + 1085.0, + 1253.0, + 1101.0, + 1228.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 1081.0, + 1281.0, + 1081.0, + 1281.0, + 1096.0, + 1257.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 1066.0, + 1367.0, + 1066.0, + 1367.0, + 1091.0, + 1286.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1369.0, + 1061.0, + 1392.0, + 1061.0, + 1392.0, + 1075.0, + 1369.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1058.0, + 1416.0, + 1058.0, + 1416.0, + 1069.0, + 1397.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1122.0, + 627.0, + 1122.0, + 627.0, + 1145.0, + 495.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 1122.0, + 995.0, + 1122.0, + 995.0, + 1145.0, + 856.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 1122.0, + 1359.0, + 1122.0, + 1359.0, + 1145.0, + 1221.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 1156.0, + 440.0, + 1156.0, + 440.0, + 1186.0, + 415.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 1156.0, + 806.0, + 1156.0, + 806.0, + 1222.0, + 771.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1159.0, + 1170.0, + 1159.0, + 1170.0, + 1189.0, + 1147.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 1182.0, + 437.0, + 1182.0, + 437.0, + 1200.0, + 417.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1182.0, + 1171.0, + 1182.0, + 1171.0, + 1203.0, + 1147.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 1191.0, + 445.0, + 1191.0, + 445.0, + 1266.0, + 411.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 1194.0, + 1172.0, + 1194.0, + 1172.0, + 1227.0, + 1144.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 1208.0, + 813.0, + 1208.0, + 813.0, + 1266.0, + 779.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 1216.0, + 1177.0, + 1216.0, + 1177.0, + 1276.0, + 1142.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1278.0, + 452.0, + 1278.0, + 452.0, + 1296.0, + 432.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1276.0, + 820.0, + 1276.0, + 820.0, + 1298.0, + 796.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 1275.0, + 1184.0, + 1275.0, + 1184.0, + 1298.0, + 1161.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 1293.0, + 826.0, + 1293.0, + 826.0, + 1350.0, + 779.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 1383.0, + 489.0, + 1383.0, + 489.0, + 1399.0, + 469.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 1396.0, + 521.0, + 1396.0, + 521.0, + 1412.0, + 499.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 1390.0, + 549.0, + 1390.0, + 549.0, + 1406.0, + 525.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1386.0, + 578.0, + 1386.0, + 578.0, + 1400.0, + 553.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1370.0, + 658.0, + 1370.0, + 658.0, + 1386.0, + 638.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 1369.0, + 685.0, + 1369.0, + 685.0, + 1379.0, + 665.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 1383.0, + 854.0, + 1383.0, + 854.0, + 1399.0, + 836.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1396.0, + 886.0, + 1396.0, + 886.0, + 1412.0, + 864.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 1390.0, + 915.0, + 1390.0, + 915.0, + 1406.0, + 892.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1387.0, + 940.0, + 1387.0, + 940.0, + 1399.0, + 922.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 1373.0, + 999.0, + 1373.0, + 999.0, + 1396.0, + 947.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 1373.0, + 1023.0, + 1373.0, + 1023.0, + 1383.0, + 1005.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1367.0, + 1049.0, + 1367.0, + 1049.0, + 1379.0, + 1032.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 1383.0, + 1221.0, + 1383.0, + 1221.0, + 1400.0, + 1200.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 1396.0, + 1253.0, + 1396.0, + 1253.0, + 1412.0, + 1230.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 1390.0, + 1283.0, + 1390.0, + 1283.0, + 1406.0, + 1257.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 1386.0, + 1309.0, + 1386.0, + 1309.0, + 1400.0, + 1286.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 1376.0, + 1363.0, + 1376.0, + 1363.0, + 1395.0, + 1314.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1373.0, + 1389.0, + 1373.0, + 1389.0, + 1383.0, + 1372.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1369.0, + 1416.0, + 1369.0, + 1416.0, + 1379.0, + 1399.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 1429.0, + 627.0, + 1429.0, + 627.0, + 1459.0, + 494.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 1433.0, + 993.0, + 1433.0, + 993.0, + 1456.0, + 856.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 1433.0, + 1359.0, + 1433.0, + 1359.0, + 1456.0, + 1220.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 1467.0, + 1171.0, + 1467.0, + 1171.0, + 1483.0, + 1145.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 1477.0, + 445.0, + 1477.0, + 445.0, + 1580.0, + 411.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 1494.0, + 642.0, + 1494.0, + 642.0, + 1653.0, + 457.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 1487.0, + 807.0, + 1487.0, + 807.0, + 1524.0, + 781.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 1482.0, + 1174.0, + 1482.0, + 1174.0, + 1540.0, + 1141.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 1530.0, + 810.0, + 1530.0, + 810.0, + 1546.0, + 787.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1534.0, + 1174.0, + 1534.0, + 1174.0, + 1552.0, + 1152.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 1552.0, + 811.0, + 1552.0, + 811.0, + 1567.0, + 787.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1543.0, + 1178.0, + 1543.0, + 1178.0, + 1579.0, + 1147.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1589.0, + 452.0, + 1589.0, + 452.0, + 1606.0, + 432.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 1589.0, + 817.0, + 1589.0, + 817.0, + 1606.0, + 798.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 1586.0, + 1185.0, + 1586.0, + 1185.0, + 1607.0, + 1161.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1604.0, + 462.0, + 1604.0, + 462.0, + 1694.0, + 407.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 1693.0, + 490.0, + 1693.0, + 490.0, + 1710.0, + 469.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 1706.0, + 521.0, + 1706.0, + 521.0, + 1721.0, + 499.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1700.0, + 548.0, + 1700.0, + 548.0, + 1717.0, + 523.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1696.0, + 578.0, + 1696.0, + 578.0, + 1710.0, + 553.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 1691.0, + 606.0, + 1691.0, + 606.0, + 1706.0, + 582.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 1686.0, + 635.0, + 1686.0, + 635.0, + 1700.0, + 608.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 1683.0, + 657.0, + 1683.0, + 657.0, + 1694.0, + 639.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 1679.0, + 685.0, + 1679.0, + 685.0, + 1689.0, + 664.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 1603.0, + 826.0, + 1603.0, + 826.0, + 1659.0, + 779.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1696.0, + 854.0, + 1696.0, + 854.0, + 1707.0, + 837.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 1706.0, + 887.0, + 1706.0, + 887.0, + 1721.0, + 863.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1700.0, + 915.0, + 1700.0, + 915.0, + 1716.0, + 890.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 1686.0, + 999.0, + 1686.0, + 999.0, + 1706.0, + 947.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 1681.0, + 1025.0, + 1681.0, + 1025.0, + 1696.0, + 1002.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1676.0, + 1051.0, + 1676.0, + 1051.0, + 1691.0, + 1029.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 1696.0, + 1220.0, + 1696.0, + 1220.0, + 1707.0, + 1201.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1228.0, + 1706.0, + 1253.0, + 1706.0, + 1253.0, + 1720.0, + 1228.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 1701.0, + 1280.0, + 1701.0, + 1280.0, + 1716.0, + 1257.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 1693.0, + 1333.0, + 1693.0, + 1333.0, + 1704.0, + 1316.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1337.0, + 1686.0, + 1363.0, + 1686.0, + 1363.0, + 1700.0, + 1337.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1367.0, + 1681.0, + 1392.0, + 1681.0, + 1392.0, + 1696.0, + 1367.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1679.0, + 1415.0, + 1679.0, + 1415.0, + 1689.0, + 1397.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 555.5, + 1174.0, + 555.5, + 1174.0, + 575.5, + 1144.0, + 575.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.75, + 594.5, + 1451.75, + 594.5, + 1451.75, + 716.5, + 1154.75, + 716.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.75, + 601.0, + 709.75, + 601.0, + 709.75, + 694.0, + 449.75, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.25, + 594.0, + 1080.25, + 594.0, + 1080.25, + 705.0, + 786.25, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.25, + 677.0, + 827.25, + 677.0, + 827.25, + 693.5, + 788.25, + 693.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 706.0, + 495.0, + 706.0, + 495.0, + 769.0, + 429.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 768.5, + 639.0, + 768.5, + 639.0, + 788.5, + 551.0, + 788.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 768.5, + 1005.0, + 768.5, + 1005.0, + 788.5, + 916.0, + 788.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 768.5, + 1371.0, + 768.5, + 1371.0, + 788.5, + 1282.0, + 788.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 765.5, + 1308.0, + 765.5, + 1308.0, + 779.5, + 1283.0, + 779.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.25, + 978.0, + 486.25, + 978.0, + 486.25, + 1045.0, + 410.25, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 978.0, + 846.0, + 978.0, + 846.0, + 1036.0, + 778.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.75, + 977.0, + 1212.75, + 977.0, + 1212.75, + 1040.0, + 1143.75, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 1016.0, + 498.0, + 1016.0, + 498.0, + 1082.5, + 427.0, + 1082.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1017.0, + 861.0, + 1017.0, + 861.0, + 1080.5, + 794.0, + 1080.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1017.0, + 1227.0, + 1017.0, + 1227.0, + 1081.5, + 1159.0, + 1081.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.75, + 1065.0, + 642.75, + 1065.0, + 642.75, + 1083.5, + 572.75, + 1083.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 1076.5, + 642.0, + 1076.5, + 642.0, + 1099.5, + 550.0, + 1099.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.25, + 1079.5, + 1006.25, + 1079.5, + 1006.25, + 1098.5, + 916.25, + 1098.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.25, + 1078.5, + 1371.25, + 1078.5, + 1371.25, + 1098.5, + 1281.25, + 1098.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 1198.0, + 860.0, + 1198.0, + 860.0, + 1230.0, + 854.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.25, + 1216.5, + 1223.25, + 1216.5, + 1223.25, + 1235.5, + 1203.25, + 1235.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 1286.0, + 486.0, + 1286.0, + 486.0, + 1356.5, + 410.0, + 1356.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 1287.5, + 1217.0, + 1287.5, + 1217.0, + 1356.5, + 1141.0, + 1356.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 1328.0, + 498.0, + 1328.0, + 498.0, + 1391.0, + 427.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1328.0, + 865.0, + 1328.0, + 865.0, + 1391.5, + 792.0, + 1391.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1327.0, + 1229.0, + 1327.0, + 1229.0, + 1392.0, + 1159.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 1372.5, + 639.0, + 1372.5, + 639.0, + 1395.0, + 576.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.75, + 1389.5, + 640.75, + 1389.5, + 640.75, + 1408.5, + 550.75, + 1408.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.75, + 1388.5, + 1004.75, + 1388.5, + 1004.75, + 1409.0, + 917.75, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 1389.5, + 1371.0, + 1389.5, + 1371.0, + 1408.5, + 1282.0, + 1408.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 1595.0, + 1216.0, + 1595.0, + 1216.0, + 1665.0, + 1143.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.75, + 1641.0, + 859.75, + 1641.0, + 859.75, + 1700.0, + 797.75, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.75, + 1639.5, + 1227.75, + 1639.5, + 1227.75, + 1699.5, + 1157.75, + 1699.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1694.5, + 944.0, + 1694.5, + 944.0, + 1710.5, + 916.0, + 1710.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 1695.0, + 1313.0, + 1695.0, + 1313.0, + 1709.5, + 1283.0, + 1709.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 1697.5, + 641.0, + 1697.5, + 641.0, + 1720.5, + 549.0, + 1720.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.25, + 1699.5, + 1004.25, + 1699.5, + 1004.25, + 1717.5, + 916.25, + 1717.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 1699.5, + 1371.0, + 1699.5, + 1371.0, + 1717.5, + 1282.0, + 1717.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1753.0, + 1337.0, + 1753.0, + 1337.0, + 1798.0, + 360.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 71.0, + 817.0, + 71.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2125.0, + 830.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 456.0, + 1438.0, + 456.0, + 1438.0, + 494.0, + 1146.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 457.0, + 691.0, + 457.0, + 691.0, + 496.0, + 423.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 455.0, + 1064.0, + 455.0, + 1064.0, + 496.0, + 784.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 585.0, + 349.0, + 585.0, + 349.0, + 672.0, + 306.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 457.0, + 691.0, + 457.0, + 691.0, + 496.0, + 423.0, + 496.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 26, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/xGZG2kS5bFk/xGZG2kS5bFk.md b/parse/train/xGZG2kS5bFk/xGZG2kS5bFk.md new file mode 100644 index 0000000000000000000000000000000000000000..597e9d0703a6e8d425d5c89c1fe90f8560012667 --- /dev/null +++ b/parse/train/xGZG2kS5bFk/xGZG2kS5bFk.md @@ -0,0 +1,279 @@ +# DANCE REVOLUTION: LONG-TERM DANCE GENERA-TION WITH MUSIC VIA CURRICULUM LEARNING + +Ruozi Huang1∗ † Huang $\mathbf { H } \mathbf { u } ^ { 2 * }$ Wei ${ \bf W } { \bf u } ^ { 3 }$ Kei Sawada4 Mi Zhang1 Daxin Jiang2‡ Fudan University1 Microsoft STCA2 Meituan3 Rinna AI4 +{rzhuang17,mi_zhang}@fudan.edu.cn, {huahu,djiang}@microsoft.com, wuwei19850318@gmail.com, keisawada@rinna.co.jp + +# ABSTRACT + +Dancing to music is one of human’s innate abilities since ancient times. In machine learning research, however, synthesizing dance movements from music is a challenging problem. Recently, researchers synthesize human motion sequences through autoregressive models like recurrent neural network (RNN). Such an approach often generates short sequences due to an accumulation of prediction errors that are fed back into the neural network. This problem becomes even more severe in the long motion sequence generation. Besides, the consistency between dance and music in terms of style, rhythm and beat is yet to be taken into account during modeling. In this paper, we formalize the music-conditioned dance generation as a sequence-to-sequence learning problem and devise a novel seq2seq architecture to efficiently process long sequences of music features and capture the fine-grained correspondence between music and dance. Furthermore, we propose a novel curriculum learning strategy to alleviate error accumulation of autoregressive models in long motion sequence generation, which gently changes the training process from a fully guided teacher-forcing scheme using the previous ground-truth movements, towards a less guided autoregressive scheme mostly using the generated movements instead. Extensive experiments show that our approach significantly outperforms the existing state-of-the-arts on automatic metrics and human evaluation. We also make a demo video to demonstrate the superior performance of our proposed approach at https://www.youtube.com/watch?v $=$ lmE20MEheZ8. + +# 1 INTRODUCTION + +Arguably, dancing to music is one of human’s innate abilities, as we can spontaneously sway along with the tempo of music we hear. The research in neuropsychology indicates that our brain is hardwired to make us move and synchronize with music regardless of our intention (Chen et al., 2008). Another study of archaeology also suggests that dance is a social communication skill among early humans connected to the ability of survival long time ago (Adshead-Lansdale & Layson, 2006). Nowadays, dance (to music) has become a means to the cultural promotion, a method of emotional expression, a tool for socialization and an art form to bring aesthetic enjoyment. The neurological mechanism behind dancing behavior and the unique value of dance to society motivate us to explore a computational approach to dance creation from a piece of music in artificial intelligence research. Such work is potentially beneficial to a wide range of applications such as dance creation assistant in art and sports, character motion generation for audio games and research on cross-modal behavior. + +In literature, the music-conditioned dance generation is a relatively new task that attracts increasing research interests recently. Early works (Fan et al., 2011; Lee et al., 2013) synthesize dance sequences from music by retrieval-based methods, which show the limited creativity in practice. Recently, Lee et al. (2019) formulate the task from the generative perspective, and further propose a decompositionto-composition framework. Their model first generates basic dance units from the music clips and then composes them by using the last pose of current unit to initialize the first pose of the next unit. Although this approach shows better performance than the retrieval-based methods, several challenges still remain. + +First, existing generative methods synthesize new human motion sequences through autoregressive models like RNN, which tend to result in short sequences. In other words, the generated sequences often quickly freeze within a few seconds due to an accumulation of prediction errors that are fed back into the neural network. This problem becomes even more severe in long motion sequence generation. For instance, composing a dance for a 1-minute music clip under 15 Frame Per Second (FPS) means generating 900 poses at one time. In practice, we need novel methods which can effectively generate long motion sequences. Besides, how to enhance the harmony between the synthesized dance movements and the given music is a largely unexplored challenge. Inutitively, the movements need to be consistent with the music in terms of style, rhythm and beat. However, to achieve this goal is non-trivial, which requires the generation model to have the capability to capture the fine-grained correspondence between music and dance. + +In this paper, we formalize music-conditioned dance generation as a sequence-to-sequence learning problem where the fine-grained correspondence between music and dance is represented through sequence modeling and their alignment is established via mapping from a sequence of acoustic features of the music to a sequence of movements of the dance. The model consists of a music encoder and a dance decoder. The encoder transforms low-level acoustic features of an input music clip into high-level latent representations via self-attention with a receptive field restricted within $k$ -nearest neighbors of an element. Thus, the encoder can efficiently process long sequences of music features, e.g., a sequence with more than 1000 elements, and model local characteristics of the music such as chord and rhythm patterns. The decoder exploits a recurrent structure to predict the dance movement frame by frame conditioned on the corresponding element in the latent representations of music feature sequence. Furthermore, we propose a curriculum learning (Bengio et al., 2009) strategy to alleviate error accumulation (Li et al., 2017) of autoregressive models in long motion sequence generation. Specifically, it gently changes the training process from a fully guided teacher-forcing scheme using the previous ground-truth movements, towards a less guided autoregressive scheme which mostly utilizes the generated movements instead. This strategy bridges the gap between training and inference of autoregressive models, and thus alleviates error accumulation at inference. + +The length of each video clip in the dataset released by Lee et al. (2019) is only about 6 seconds, which cannot be used by other methds to generate long-term dance except for their specially designed decomposition-to-composition framework. To facilitate the task of long-term dance generation with music, we collect a high-quality dataset consisting of 1-minute video clips, totaling about 12 hours. And there are three representative styles in our dataset: “Ballet”, “Hiphop” and “Japanese Pop”. + +Our contributions in this work are four-fold: (1) We formalize music-conditioned dance generation as a sequence-to-sequence learning problem and devise a novel seq2seq architecture for the long-term dance generation with music. (2) We propose a novel curriculum learning strategy to alleviate error accumulation of autoregressive models in long motion sequence generation. (3) To facilitate long-term dance generation with music, we collect a high-quality dataset that is available with our code1. (4) The extensive experiments show that our approach significantly outperforms the existing state-of-the-arts on both automatic metrics and human judgements. The demo video in the supplementary material also exhibits our approach can generate diverse minute-length dances that are smooth, natural-looking, style-consistent and beat-matching with the musics from test set. + +# 2 RELATED WORK + +Cross-Modal Learning. Most existing works focus on the modeling between vision and text, such as image captioning (Lu et al., 2017; Xu et al., 2015) and text-to-image generation (Reed et al., 2016; Zhang et al., 2017). There are some other works to study the translation between audio and text like Automatic Speech Recognition (ASR) (Hinton et al., 2012) and Text-To-Speech (TTS) (Oord et al., 2016). While the modeling between audio and vision is largely unexplored and the music-conditioned dance generation is a typical cross-modal learning problem from audio to vision. + +Human Motion Prediction. Prediction of human motion dynamics has been a challenging problem in computer vision, which suffers from the high spatial-temporal complexity. Existing works (Chan et al., 2019; Wang et al., 2018) represent the human pose as 2D or 3D body keyjoints (Cao et al., 2017) and address the problem via sequence modeling. Early methods, such as hidden markov models (Lehrmann et al., 2014), Gaussian processes (Wang et al., 2006) and restricted boltzmann machines (Taylor et al., 2007), have to balance the model capacity and inference complexity due to complicated training procedures. Recently, neural networks dominate the human motion modeling. For instance, Fragkiadaki et al. (2015) present LSTM-3LR and Encoder-Recurrent-Decoder (ERD) as two recurrent architectures for the task; Jain et al. (2016) propose a structural-RNN to model human-object interactions in a spatio-temporal graph; and Ghosh et al. (2017) equip LSTM-3LR with a dropout autoencoder to enhance the long-term prediction. Besides, convolutional neural networks (CNNs) have also been utilized to model the human motion prediction (Li et al., 2018). + +Audio-Conditioned Dance Generation. In the research of audio-conditioned dance generation, most existing works study 2D dance motion generation with music since the training data for paired 2D pose and music can be extracted from the huge amount of dance videos available online. Various methods have been proposed to handle this task, such as adversarial learning based methods (Lee et al., 2019; Sun et al., 2020; Ferreira et al., 2021), autoencoder methods (Tang et al., 2018) and sequence-to-sequence methods (Lee et al., 2018; Ren et al., 2019; Yalta et al., 2019; Ye et al., 2020). While these works mainly focus on exploring the different neural architectures and overlook the freezing motion issue in dance motion synthesis. In this work, we first propose a novel seq2seq architecture to model the fine-grained correspondence between music and dance, and then introduce a novel curriculum learning strategy to address the freezing motion issue caused by error accumulation (Li et al., 2017) in long-term dance motion generation. + +# 3 APPROACH + +In this section, we present our approach to music-conditioned dance generation. After formalization of the problem in question, we elaborate the model architecture and the dynamic auto-condition learning approach that facilitates long-term dance generation according to the given music. + +![](images/f2bd9934103d7f435c8ce5dfd137026280301ab5097a7eec93c1dca3017d0a32.jpg) +Figure 1: The overview of our method. The transformer based encoder is stacked of $N$ identical layers with the local self-attention mechanism, which aims to efficiently process long features sequences extracted from music in a GPU-memory-economic way. The decoder exploits a recurrent structure to predict dance movement $y _ { i }$ frame by frame conditioned on hidden state $h _ { i }$ of decoder and latent vector $z _ { i }$ . Note that $y _ { 0 }$ is a predefined Begin Of Pose (BOP). During the training phase, the proposed curriculum learning strategy gently increases the number of steps of autoregressive scheme according to the number of training epochs. This strategy bridges the gap between training and inference of autoregressive models, and thus alleviates error accumulation at inference. + +# 3.1 PROBLEM FORMALIZATION + +Suppose that there is a dataset $\mathcal { D } = \{ ( X _ { i } , Y _ { i } ) \} _ { i = 1 } ^ { N }$ , where $X = \{ x _ { t } \} _ { t = 1 } ^ { n }$ is a music clip with $x _ { t }$ being a vector of acoustic features at time-step $t$ , and $\bar { Y = } \{ y _ { t } \} _ { t = 1 } ^ { n }$ is a sequence of dance movements with $y _ { t }$ aligned to $x _ { t }$ . The goal is to estimate a generation model $g ( \cdot )$ from $\mathcal { D }$ , and thus given a new music input $X$ , the model can synthesize dance $Y$ to music $X$ based on $g ( X )$ . We first present our seq2seq architectures chosen for music-conditioned dance generation in the following section. Later in the experiments, we empirically justify the choice by comparing the architectures with other alternatives. + +# 3.2 MODEL ARCHITECTURE + +In the architecture of $g ( \cdot )$ , a music encoder first transforms $X = ( x _ { 1 } , . . . , x _ { n } )$ $( x _ { i } \in \mathbb { R } ^ { d _ { x } } )$ into a hidden sequence $Z = ( z _ { 1 } , . . . , z _ { n } )$ $( z _ { i } \in \mathbb { R } ^ { d _ { z } } )$ using a local self-attention mechanism to reduce the memory requirement for long sequence modeling, and then a dance decoder exploits a recurrent structure to autoregressively predicts movements $Y = ( y _ { 1 } , . . . , y _ { n } )$ conditioned on $Z$ . + +Music Encoder. Encouraged by the compelling performance on music generation (Huang et al., 2018), we define the music encoder with a transformer encoder structure. While the self-attention mechanism (Vaswani et al., 2017) in transformer can effectively represent the multi-scale structure of music, the quadratic memory complexity $\scriptstyle { \mathcal { O } } ( n ^ { 2 } )$ about sequence length $n$ impedes its application to long sequence modeling due to huge of GPU memory consumption (Child et al., 2019). To keep the effectiveness of representation and control the cost, we introduce a local self-attention mechanism that modifies the receptive field of self-attention by restricting the element connections within $k$ -nearest neighbors. Thus, the memory complexity is reduced to $\mathcal { O } ( n k )$ . $k$ could be small in our scenario since we only pursue an effective representation for a given music clip. Therefore, the local patterns of music are encoded in $z _ { t }$ that is sufficient for the generation of movement $y _ { t }$ at time-step $t$ , which is aligned with the common sense that the dance movement at certain time-step is highly influenced by the nearby clips of music. Yet we can handle long sequences of acoustic features, e.g., more than 1000 elements, in an efficient and memory-economic way. + +Specifically, we first embed $X = ( x _ { 1 } , . . . , x _ { n } )$ into $U = ( u _ { 1 } , . . . , u _ { n } )$ with a linear layer parameterized by $W ^ { E } \in \mathbf { \bar { \mathbb { R } } } ^ { d _ { x } \times d _ { z } }$ . Then $\forall i \in \{ 1 , . . . , n \}$ , $z _ { i }$ can be formulated as: + +$$ +z _ { i } = \mathcal { F } ( a _ { i } ) , \quad a _ { i } = \sum _ { j = i - \lfloor k / 2 \rfloor } ^ { i + \lfloor k / 2 \rfloor } \alpha _ { i j } ( u _ { j } W _ { l } ^ { V } ) , \quad U = X W ^ { E } , +$$ + +where $\mathcal { F } ( \cdot ) : \mathbb { R } ^ { d _ { v } } \mathbb { R } ^ { d _ { z } }$ is a feed forward neural network. Each $u _ { j }$ is only allowed to attend its $k$ -nearest neighbors including itself, where $k$ is a hyper-parameter referring to the sliding window size of the local self-attention. Then attention weight $\alpha _ { i j }$ is calculated using a softmax function as: + +$$ +\alpha _ { i j } = \frac { \exp { e _ { i j } } } { \sum _ { t = j - \lfloor k / 2 \rfloor } ^ { j + \lfloor k / 2 \rfloor } \exp { e _ { i t } } } , e _ { i j } = \frac { ( u _ { i } W _ { l } ^ { Q } ) ( u _ { j } W _ { l } ^ { K } ) ^ { \top } } { \sqrt { d _ { k } } } , +$$ + +where for the the $l$ -th head, $W _ { l } ^ { Q }$ , $W _ { l } ^ { K } \in \mathbb { R } ^ { d _ { z } \times d _ { k } }$ and $W _ { l } ^ { V } \in \mathbb { R } ^ { d _ { z } \times d _ { v } }$ are parameters that transform $U$ into a query, a key, and a value respectively. $d _ { z }$ is the dimension of hidden state $z _ { i }$ while $d _ { k }$ is the dimension of query, key and $d _ { v }$ is the dimension of value. + +Dance Decoder. We choose a recurrent neural network as the dance decoder in consideration of two factors: (1) the chain structure can well capture the spatial-temporal dependency among human movement dynamics, which has proven to be highly effective in the state-of-the-art methods for human motion prediction (Li et al., 2018; Mao et al., 2019); (2) our proposed learning strategy is tailored for the autoregressive models like RNN, as will be described in the next section. Specifically, with $Z = ( z _ { 1 } , . . . , z _ { n } )$ , the dance movements $\boldsymbol { Y } = ( \hat { y } _ { 1 } , . . . , \hat { y } _ { n } )$ are synthesized by: + +$$ +\begin{array} { l } { \hat { y } _ { i } = [ h _ { i } ; z _ { i } ] W ^ { S } + b , } \\ { h _ { i } = \mathrm { R N N } ( h _ { i - 1 } , \hat { y } _ { i - 1 } ) , } \end{array} +$$ + +where $h _ { i }$ is the $i$ -th hidden state of the decoder and $h _ { 0 }$ is initialized by sampling from the standard normal distribution to enhance variation of the generated sequences. $[ \cdot ; \cdot ]$ denotes the concatenation operation. $W ^ { S } \in \mathbb { R } ^ { ( d _ { s } + d _ { z } ) \times d _ { y } }$ and $b \in \mathbb { R } ^ { d _ { y } }$ are parameters where $d _ { s }$ and $d _ { y }$ are the dimensions of $h _ { i }$ and $\hat { y } _ { i }$ , respectively. At the $i$ -th time-step, the decoder predicts the movement $\hat { y } _ { i }$ conditioned on $h _ { i }$ as well as the latent feature representation $z _ { i }$ , and thus can capture the fine-grained correspondence between music feature sequence and dance movement sequence. + +# 3.3 DYNAMIC AUTO-CONDITION LEARNING APPROACH + +Exposure bias (Ranzato et al., 2015; Lamb et al., 2016; Zhang et al., 2019) is a notorious issue in natural language generation (NLG), which refers to the train-test discrepancy of autoregressive generative models. This kind of models are never exposed to their own prediction errors at the training phase due to the conventional teacher-forcing training scheme. However, compared to NLG, motion sequence generation suffers from the more severe exposure bias problem. In NLG, the bias of predicted probability distribution over vocabulary can be corrected by sampling strategy. For instance, we can still generate the target token with index 2 by sampling on probability distribution [0.3, 0.3, 0.4] whose groundtruth is [0, 0, 1]. While any small biases of the predicted motion at each time-step will be accumulated and propagated to the future, since each generated motion is a real-valued vector (rather than a discrete token) representing 2D or 3D body keyjoints in the continuous space. As a result, the generated motion sequences tend to quickly freeze within a few seconds. This problem becomes even worse in the generation of long motion sequences, e.g., more than 1000 movements. + +Scheduled sampling (Bengio et al., 2015) is proposed to alleviate exposure bias in NLG, which does not work in long motion sequence generation. Since its sampling-based strategy would feed long predicted sub-sequences of high bias into the model at early stage, which causes gradient vanishing at training due to error accumulation of these high-biased sub-sequences. Motivated by this, we propose a dynamic auto-condition training method as a novel curriculum learning strategy to alleviate the error accumulation of autoregressive models in long motion sequence generation. Specifically, our learning strategy gently changes the training process from a fully guided teacher-forcing scheme using the previous ground-truth movements, towards a less guided autoregressive scheme mostly using the generated movements instead, which bridges the gap between training and inference. + +As shown in Figure 1, The decoder predicts $Y _ { t g t } = \{ y _ { 1 } , . . . , y _ { p } , y _ { p + 1 } , . . . , y _ { p + q } , y _ { p + q + 1 } , . . . \}$ from $Y _ { i n } = \{ y _ { 0 } , \hat { y _ { 1 } } , . . . , \hat { y _ { p } } , y _ { p + 1 } , . . . , y _ { p + q } , \hat { y } _ { p + q + 1 } , . . . \}$ which alternates the predicted sub-sequences with length $p$ (e.g., $\hat { y } _ { 1 } , . . . , \hat { y } _ { p } )$ and the ground-truth sub-sequences with length $q$ (e.g., $y _ { p + 1 } , . . . , y _ { { p + q } } )$ . $y _ { 0 }$ is a predefined begin of pose (BOP). During the training phase, we fix $q$ and gradually increase $p$ according to a growth function $f ( t ) \in \{ c o n s t , \lfloor \lambda t \rfloor , \lfloor \lambda t ^ { 2 } \rfloor , \lfloor \lambda e ^ { t } \rfloor \}$ where $t$ is the number of training epochs and $\lambda < 1$ controls how frequently $p$ is updated. We choose $f ( t ) = \lfloor \lambda t \rfloor$ in practice through the empirical comparison. Note that our learning strategy degenerates to the method proposed in Li et al. (2017) when $f ( t ) = c o n s t$ . While the advantage of a dynamic strategy over a static strategy lies in that the former introduces the curriculum learning spirit to dynamically increase the difficulty of curriculum (e.g., larger $p$ ) as the model capacity improves during training, which can further improve the model capacity and alleviate the error accumulation of autoregressive predictions. Finally, we estimate parameters of $g ( \cdot )$ by minimizing the $\ell _ { 1 }$ loss on $\mathcal { D }$ : + +$$ +\ell _ { 1 } = \frac { 1 } { N } \sum _ { i = 1 } ^ { N } \left| \left| g ( X _ { i } ) - Y _ { t g t } ^ { ( i ) } \right| \right| _ { 1 } +$$ + +# 4 EXPERIMENTAL SETUP + +# 4.1 DATASET COLLECTION AND PREPROCESSING + +Although the dataset (Lee et al., 2019) contains about 70-hour video clips, the average length of each clip is only about 6 seconds. Moreover, their dataset cannot be used by other methods to generate longterm dance except for their specially designed decomposition-to-composition framework. To facilitate the task of long-term dance generation with music, we collect a high-quality dataset consisting of 790 one-minute clips of pose data with 15 FPS, totaling about 12 hours. And there are three styles in the dataset including “Ballet”, “Hiphop” and “Japanese Pop”. Table 1 shows the statistics of our dataset. In the experiment, we randomly split the dataset into the $90 \%$ training set and $10 \%$ test set. + +Pose Preprocess. For the human pose estimation, we leverage OpenPose (Cao et al., 2017) to extract 2D body keyjoints from videos. Each pose consists of 25 keyjoints2 and is represented by a 50-dimension vector in the continuous space. In practice, we develop a linear interpolation algorithm to find the missing keyjoints from nearby frames to reduce the noise in extracted pose data. + +Table 1: Statistics of the dataset. + +
CategoryNum of ClipsClip Length (min)FPSResolution
Ballet136115720P
Hiphop298115720P
Japanese Pop356115720P
+ +Audio Preprocess. Librosa (McFee et al., 2015) is a well-known audio and music analysis library in the music information retrieval, which provides flexible ways to extract the spectral and rhythm features of audio data. Specifically, we extract the following features: mel frequency cepstral coefficients (MFCC), MFCC delta, constant- $Q$ chromagram, tempogram and onset strength. To better capture the beat information of music, we convert onset strength into a one-hot vector as an additional feature to explicitly represent the beat sequence of music, i.e., beat one-hot. Finally, we concatenate these audio features as the representation of a music frame, as shown in Table 4 in Section 6.2. During extracting audio features, the sampling rate is $1 5 { , } 4 0 0 \mathrm { H z }$ and hop size is 1024. Hence, we have 15 audio samples per second that is aligned with the 15 FPS of pose data. + +# 4.2 IMPLEMENTATION DETAILS + +The music encoder consists of a stack of $N = 2$ identical layers. Each layer has two sublayers: a local self-attention sublayer with $l = 8$ heads and a position-wise fully connected feed-forward sublayer with 1024 hidden units. Each head contains a scaled dot-product attention layer with $d _ { k } = d _ { v } = 6 4$ and its receptive yield is restricted by setting $k = 1 0 0$ . Then we set the sequence length $n = 9 0 0$ , dimension of acoustic feature vector $d _ { x } = 4 3 8$ , dimension of hidden vector $d _ { z } = 2 5 6$ , respectively. The dance decoder is a 3-layer LSTM with $d _ { s } = 1 0 2 4$ and the dimension of pose vector $d _ { y } = 5 0$ . We set $\lambda = 0 . 0 1$ and $q = 1 0$ for the proposed learning approach and train the model using the Adam optimizer with the learning rate 1e-4 on 2 NVIDIA V100 GPUs. + +# 4.3 BASELINES + +The music-conditioned generation is an emerging task and there are few methods proposed to solve this problem. In our experiment, we compare our proposed approach to the following baselines: (1) Dancing2Music. We use Dancing2Music (Lee et al., 2019) as the primary baseline that is the previous state-of-the-art; (2) Aud-MoCoGAN. Aud-MoCoGAN is an auxiliary baseline used in Dancing2Music, which is original from MoCoGAN (Tulyakov et al., 2018); (3) LSTM. Shlizerman et al. (2018) propose a LSTM network to predict body dynamics from the audio signal. We modify their open-source code to take audio features of music as inputs and produce human pose sequences. + +# 5 EXPERIMENTAL RESULTS + +In this section, we conduct extensive experiments to evaluate our approach and compare it to the aforementioned baselines. Note that we do not include comparison experiment on the dataset collect by Lee et al. (2019) due to the above issue in Section 4.1. + +# 5.1 AUTOMATIC METRICS AND HUMAN EVALUATION + +We evaluate the different methods by automatic metrics and conduct a human evaluation on motion realism and smoothness of generated dance movements, as well as style consistency. Specifically, we randomly sample 60 music clips from the test set to generate dances, then divide the dances generated by three methods and the real dances into four pairs: (LSTM, ours), (Dancing2Music, ours), (real, ours) and (real, real). We invite 10 amateur dancers as the annotators to answer 3 questions for each pair that is blind for them (use (A, B) pair instead): (1) Which dance is more realistic regardless of music? (2) Which dance is more smooth regardless of music? (3) Which dance matches the music better with respect to style? Note that, we do not include Aud-MoCoGAN in the human evaluation, since both it and Dancing2Music are GAN based methods while the latter is the state-of-the-art. + +![](images/160f9ed77699ffb31249fa04fe6af19e09624d27fa86094970947262a6f5633e.jpg) +Figure 2: Human evaluation results on motion realism, style consistency and smoothness. We conduct a human evaluation to ask annotators to select the dances that are more realistic and more smooth regardless of music, more style-consistent with music through pairwise comparison. + +Motion Realism, Style Consistency and Smoothness. We evaluate the visual quality and realism of generated dances by Fréchet Inception Distance (FID) (Heusel et al., 2017), which is used to measure how close the distribution of generated dances is to the real. Similar to Lee et al. (2019), we train a style classifier on dance movement sequences of three styles and use it to extract features for the given dances. Then we calculate FID between the synthesized dances and the real. As shown in Table 2, our FID score is significantly lower than those of baselines and much closer to that of the real, which means our generated dances are more motion-realistic and more likely to the real dances. Besides, we use the same classifier to measure the style accuracy of generated dance to the music, and our approach achieves $7 7 . 6 \%$ accuracy and significantly outperforms Dancing2Music by $1 7 . 2 \%$ . + +Human evaluation in Figure 2 consistently shows the superior performance of our approach, compared to baselines on motion realism, style consistency and smoothness. We observe the dances generated by LSTM have lots of floating movements and tend to quickly freeze within several seconds due to the error accumulation, which results in lower preferences. While Dancing2Music can generate the smooth dance units, the synthesized dances still have significant jumps where dance units are connected due to the inherent gap between them, which harm its overall performance in practice. This is also reflected in the preference comparisons to our approach, in which only $3 5 \%$ annotators prefer Dancing2Music on motion realism and $2 1 . 7 \%$ prefer Dancing2Music on smoothness. Besides, the result on style consistency shows our approach can generate more style-consistent dances with musics. Since our approach models the fine-grained correspondence between music and dance at the methodological level and also introduces more specific music features. However, Dancing2Music only considers a global style feature extracted by a pre-trained classifier in the dance generation. + +In the comparison to real dances, $4 1 . 2 \%$ annotators prefer our method on motion realism while $3 0 . 3 \%$ prefer on style consistency. Additionally, we found that $5 7 . 9 \%$ annotators prefer our approach on smoothness compared to real dances. Since the imperfect OpenPose (Cao et al., 2017) introduces the minor noise on pose data extracted from dance videos. On the other hand, in the preprocessing, we develop an interpolation algorithm to find missing keyjoints and reduce the noise of training pose data. + +Besides, this also indicates the chain structure based decoder can well capture the spatial-temporal dependency among human motion dynamics and produce the smooth movements. + +Beat Coverage and Hit Rate. We also evaluate the beat coverage and hit rate introduced in Lee et al. (2019), which defines the beat coverage as $B _ { k } / B _ { m }$ and beat hit rate as $B _ { a } / B _ { k }$ , where $B _ { k }$ is the number of kinematic beats, $B _ { m }$ is the number of musical beats and $B _ { a }$ is the number of kinematic beats that are aligned with the musical beats. An early study (Ho et al., 2013) reveals that kinematic beat frames occur when the direction of the movement changes. Therefore, similar to Yalta et al. (2019), we use the standard deviation (SD) of movement to detect the kinematic beats in our experiment. The onset strength (Ellis, 2007) is a common way to detect musical beats. Figure 3 shows two short clips of motion SD curves and aligned musical beats. We observed that the kinematic beats occur where the musical beats occur, which is consistent with the common sense that dancers would step on musical beats during dancing but do not step on every musical beat. As we can see in Table 2, LSTM and Aud-MoCoGAN generate dances with few kinematic beats and most of them do not match the musical beats. Our method outperforms Dancing2Music by $6 . 1 \%$ on beat coverage and $2 . 7 \%$ on hit rate, which indicates that introducing the features about musical beat into model is beneficial to the better beat alignment of generated dance and input music. + +![](images/167107ef7ff1cf1ce81b167922e4cc91ebfc7d4110451112c0a7a53b8c70d63b.jpg) +Figure 3: Beat tracking curves for music and dance by onset strength and motion standard deviation respectively. The left is a short clip of ballet while the right is a short clip of hip-hop. The red circles are kinematic beats and dash lines denote the musical beats. + +Diversity and Multimodality. Lee et al. (2019) introduce the diversity metric and the multimodality metric. The diversity metric evaluates the variations among generated dances corresponding to various music, which reflects the generalization ability of the model and its dependency on the input music. Similarly, we use the average feature distance as the measurement and these features are extracted by the same classifier used in measuring FID. We randomly sample 60 music clips from test set to generate dances and randomly pick 500 combinations to compute the average feature distance. Results in Table 2 show the performance of our approach is superior to Dancing2Music on diversity. The reason is that Dancing2Music only considers a global music style feature in generation while different music of the same style have the almost same style feature. Our approach models the fine-grained correspondence between music and dance at the methodological level, simultaneously takes more specific music features into account. In other words, our approach is more dependent on input music and thus can generate different samples for difference music clips of the same style. + +Multimodality metric evaluates the variations of generated dances for the same music clip. Specifically, we generate 5 dances for each of randomly sampled 60 music clips and compute the average feature distance. As we can see, our approach slightly underperforms Dancing2Music and AudMoCoGAN since it has more dependency on music inputs as aforementioned, which make the generated dances have the limited variation patterns given the same music clip. + +# 5.2 LONG-TERM DANCE GENERATION + +To further investigate the performance of different methods on long-term dance generation, we evaluate FID of dances generated by different methods over time. Specifically, we first split a generated dance with 1 minute into 15 four-second clips and measure FID of each clip. Figure 4 shows FID scores of LSTM and Aud-MoCoGAN grow rapidly in the early stage due to error accumulation and converge to the high FID, since the generated dances quickly become frozen. Dancing2Music maintains the relatively stable FID scores all the time, which benefits from its decomposition-tocomposition method. While its curve still has subtle fluctuations since it synthesizes whole dances by composing the generated dance units. Compared to baselines, FID of our method are much lower and close to real dances, which validates the good performance of our method on long-term generation. + +![](images/b131c94c16a3461b63bf6187b0f782cc2dc65245cdd3adee1cfccc5add824bad.jpg) +Figure 4: FID curves of different methods over time. + +
Encoder Structure LSTMFIDACC (%)
7.945
ConvS2S encoder13.354
Global self-attention3.680
Local self-attention4.378.5
+ +Table 3: Left: Automatic metrics of our model with different encoder structures. Right: Automatic metrics of our model training by different learning strategies. + +
Learning ApproachFIDACC (%)
Teacher-forcing61.25.3
Auto-condition (con st)15.735
Ours ([λet J)9.869
Ours ([λt²])6.473
Ours ([λt])5.177.6
+ +# 5.3 ABLATION STUDY + +Comparison on Encoder Structures. We compare the different encoder structures in our model with the same LSTM decoder and the same curriculum learning strategy, as shown in Table 3. The transformer encoders outperform LSTM and the encoder in ConvS2S (Gehring et al., 2017) on both FID and style consistency, due to its superior performance on sequence modeling. Although the transformer encoder with our proposed local self-attention slightly underperforms the global self-attention by 0.7 higher in FID and $1 . 5 \%$ lower in ACC, it has much fewer parameters in our settings. This confirms the effectiveness of the local self-attention on long sequence modeling. + +Comparison on Learning Strategies. To investigate the performance of our proposed learning approach, we conduct an ablation study to compare the performances of our model using different learning strategies. As we can see the right of in Table 3, all curriculum learning strategies significantly outperform the static auto-condition strategy (Li et al., 2017). Among them, the linear growth function $f ( t ) = \lfloor \lambda t \rfloor$ performs best, which might be that increasing the difficulty of curriculum too fast would result in the model degradation. Besides, the original teacher-forcing strategy has the highest FID and the lowest style accuracy due to severe error accumulation problem, which indicates that using our proposed curriculum learning strategy to train the model can effectively alleviate this issue. + +# 6 CONCLUSION + +In this work, we propose a novel seq2seq architecture for long-term dance generation with music, e.g., about one-minute length. To efficiently process long sequences of music features, we introduce a local self-attention mechanism in the transformer encoder structure to reduce the quadratic memory requirements to the linear in the sequence length. Besides, we also propose a dynamic auto-condition training method as a novel curriculum learning strategy to alleviate the error accumulation of autoregressive models in long motion sequence generation, thus facilitate the long-term dance generation. Extensive experiments have demonstrated the superior performance of our proposed approach on automatic metrics and human evaluation. Future works include the exploration on end-to-end methods that could work with raw audio data instead of preprocessed audio features. + +# REFERENCES + +Janet Adshead-Lansdale and June Layson. Dance history: An introduction. Routledge, 2006. + +Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. In Advances in Neural Information Processing Systems, pp. 1171–1179, 2015. + +Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pp. 41–48, 2009. + +Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 7291–7299, 2017. + +Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A Efros. Everybody dance now. In Proceedings of the IEEE International Conference on Computer Vision, pp. 5933–5942, 2019. + +Joyce L Chen, Virginia B Penhune, and Robert J Zatorre. Listening to musical rhythms recruits motor regions of the brain. Cerebral cortex, 18(12):2844–2854, 2008. + +Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. + +Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pp. 2262–2271, 2019. + +Daniel PW Ellis. Beat tracking by dynamic programming. Journal of New Music Research, 36(1): 51–60, 2007. + +Rukun Fan, Songhua Xu, and Weidong Geng. Example-based automatic music-driven conventional dance motion synthesis. IEEE transactions on visualization and computer graphics, 18(3):501–515, 2011. + +Joao P Ferreira, Thiago M Coutinho, Thiago L Gomes, José F Neto, Rafael Azevedo, Renato Martins, and Erickson R Nascimento. Learning to dance: A graph convolutional adversarial network to generate realistic dance motions from audio. Computers & Graphics, 94:11–21, 2021. + +Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models for human dynamics. In Proceedings of the IEEE International Conference on Computer Vision, pp. 4346–4354, 2015. + +Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional sequence to sequence learning. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pp. 1243–1252. JMLR. org, 2017. + +Partha Ghosh, Jie Song, Emre Aksan, and Otmar Hilliges. Learning human motion models for long-term predictions. In 2017 International Conference on 3D Vision (3DV), pp. 458–466. IEEE, 2017. + +Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in neural information processing systems, pp. 6626–6637, 2017. + +Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. IEEE Signal processing magazine, 29(6):82–97, 2012. + +Chieh Ho, Wei-Tze Tsai, Keng-Sheng Lin, and Homer H Chen. Extraction and alignment evaluation of motion beats for street dance. In 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 2429–2433. IEEE, 2013. + +Cheng-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit, Noam Shazeer, Ian Simon, Curtis Hawthorne, Andrew M Dai, Matthew D Hoffman, Monica Dinculescu, and Douglas Eck. Music transformer. arXiv preprint arXiv:1809.04281, 2018. + +Ashesh Jain, Amir R Zamir, Silvio Savarese, and Ashutosh Saxena. Structural-rnn: Deep learning on spatio-temporal graphs. In Proceedings of the ieee conference on computer vision and pattern recognition, pp. 5308–5317, 2016. + +Alex M Lamb, Anirudh Goyal Alias Parth Goyal, Ying Zhang, Saizheng Zhang, Aaron C Courville, and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. In Advances In Neural Information Processing Systems, pp. 4601–4609, 2016. + +Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. In Advances in Neural Information Processing Systems, pp. 3581–3591, 2019. + +Juheon Lee, Seohyun Kim, and Kyogu Lee. Listen to dance: Music-driven choreography generation using autoregressive encoder-decoder network. arXiv preprint arXiv:1811.00818, 2018. + +Minho Lee, Kyogu Lee, and Jaeheung Park. Music similarity-based approach to generating dance motion sequence. Multimedia tools and applications, 62(3):895–912, 2013. + +Andreas M Lehrmann, Peter V Gehler, and Sebastian Nowozin. Efficient nonlinear markov models for human motion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1314–1321, 2014. + +Chen Li, Zhen Zhang, Wee Sun Lee, and Gim Hee Lee. Convolutional sequence to sequence model for human dynamics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 5226–5234, 2018. + +Zimo Li, Yi Zhou, Shuangjiu Xiao, Chong He, Zeng Huang, and Hao Li. Auto-conditioned recurrent networks for extended complex human motion synthesis. arXiv preprint arXiv:1707.05363, 2017. + +Jiasen Lu, Caiming Xiong, Devi Parikh, and Richard Socher. Knowing when to look: Adaptive attention via a visual sentinel for image captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 375–383, 2017. + +Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. In Proceedings of the IEEE International Conference on Computer Vision, pp. 9489–9497, 2019. + +Brian McFee, Colin Raffel, Dawen Liang, Daniel PW Ellis, Matt McVicar, Eric Battenberg, and Oriol Nieto. librosa: Audio and music signal analysis in python. In Proceedings of the 14th python in science conference, volume 8, 2015. + +Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499, 2016. + +Marc’Aurelio Ranzato, Sumit Chopra, Michael Auli, and Wojciech Zaremba. Sequence level training with recurrent neural networks. arXiv preprint arXiv:1511.06732, 2015. + +Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. arXiv preprint arXiv:1605.05396, 2016. + +Xuanchi Ren, Haoran Li, Zijian Huang, and Qifeng Chen. Music-oriented dance video synthesis with pose perceptual loss. arXiv preprint arXiv:1912.06606, 2019. + +Eli Shlizerman, Lucio Dery, Hayden Schoen, and Ira Kemelmacher-Shlizerman. Audio to body dynamics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 7574–7583, 2018. + +Guofei Sun, Yongkang Wong, Zhiyong Cheng, Mohan S Kankanhalli, Weidong Geng, and Xiangdong Li. Deepdance: music-to-dance motion choreography with adversarial learning. IEEE Transactions on Multimedia, 23:497–509, 2020. + +Taoran Tang, Jia Jia, and Hanyang Mao. Dance with melody: An lstm-autoencoder approach to music-oriented dance synthesis. In Proceedings of the 26th ACM international conference on Multimedia, pp. 1598–1606, 2018. + +Graham W Taylor, Geoffrey E Hinton, and Sam T Roweis. Modeling human motion using binary latent variables. In Advances in neural information processing systems, pp. 1345–1352, 2007. + +Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1526–1535, 2018. + +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pp. 5998–6008, 2017. + +Jack Wang, Aaron Hertzmann, and David J Fleet. Gaussian process dynamical models. In Advances in neural information processing systems, pp. 1441–1448, 2006. + +Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. arXiv preprint arXiv:1808.06601, 2018. + +Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In International conference on machine learning, pp. 2048–2057, 2015. + +Nelson Yalta, Shinji Watanabe, Kazuhiro Nakadai, and Tetsuya Ogata. Weakly-supervised deep recurrent neural networks for basic dance step generation. In 2019 International Joint Conference on Neural Networks (IJCNN), pp. 1–8. IEEE, 2019. + +Zijie Ye, Haozhe Wu, Jia Jia, Yaohua Bu, Wei Chen, Fanbo Meng, and Yanfeng Wang. Choreonet: Towards music to dance synthesis with choreographic action unit. In Proceedings of the 28th ACM International Conference on Multimedia, pp. 744–752, 2020. + +Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pp. 5907–5915, 2017. + +Wen Zhang, Yang Feng, Fandong Meng, Di You, and Qun Liu. Bridging the gap between training and inference for neural machine translation. arXiv preprint arXiv:1906.02448, 2019. + +# APPENDIX + +# 6.1 MORE DETAILS ABOUT DATA COLLECTION + +The dance videos are collected from YouTube by crowd workers, which are all solo dance videos with 30FPS. We trim the beginning and ending several seconds for each of collected videos to remove the silence parts, and split them into one-minute video clips. Then, we extract 2D pose data from these video clips by OpenPose (Cao et al., 2017) with 15FPS setting and collect 790 one-minute clips of pose data with 15FPS, totaling about 13 hours. Finally, we extract the corresponding audio data from video clips by FFmpeg3. The code is implemented based on PyTorch framework. MIT License will be used for the released code and data. + +# 6.2 EXTRACTED MUSIC FEATURES + +In this section, we detail the extracted features of music that are feeded into our model. Specifically, we leverage the public Librosa (McFee et al., 2015) to extract the music features including: 20- dim MFCC, 20-dim MFCC delta, 12-dim chroma, 384-dim tempogram, 1-dim onset strength (i.e., envelope) and 1-dim one-hot beat, as shown in Table 4. + +Table 4: Acoustic features of music. + +
FeatureCharacteristicDimension
MFCCPitch20
MFCC deltaPitch20
Constant-QchromagramPitch12
TempogramStrength384
Onset strengthStrength1
Beat one-hotBeat1
+ +# 6.3 DOWNSTREAM APPLICATIONS + +The downstream applications of our proposed audio-conditioned dance generation method include: (1) Our model can be used to help professional people choreograph new dances for a given song and teach human how to dance with regard to this song; (2) With the help of 3D human pose estimation (Ci et al., 2019) and 3D animation driving techniques, our model can be used to drive the various 3D character models, such as the 3D model of Hatsune Miku (very popular virtual character in Japan). This technique has the great potential for the virtual advertisement video generation that could be used for the promotion events on social medias like TikTok, Twitter and etc. + +# 6.4 MUSICAL BEAT DETECTION UNDER LOW SAMPLING RATES + +In this section, we conduct an additional experiment to evaluate the performance of Librosa beat tracker under $1 5 { , } 4 0 0 \mathrm { H z }$ sampling rate. Specifically, we first utilize it to extract the onset beats from audio data with $1 5 { , } 4 0 0 \mathrm { H z }$ and $2 2 { , } 0 5 0 \mathrm { H z }$ respectively, then compare two groups of onset beats under the same time scale. Since $2 2 , 0 5 0 \mathrm { H z }$ is a common sampling rate used in music information retrieval, we define the beat alignment ratio as $B _ { 2 } / B _ { 1 }$ , where $B _ { 1 }$ is the number of beats under $1 5 { , } 4 0 0 \mathrm { H z }$ and $B _ { 2 }$ is the number of beats under 15,400Hz that are aligned with the beats under $2 2 { , } 0 5 0 \mathrm { H z }$ . One beat is counted when $\mid t _ { 1 } - t _ { 2 } \mid \leq \Delta t . \ t _ { 1 }$ denotes the timestamp of a certain beat under $1 5 { , } 4 0 0 \mathrm { H z }$ while $t _ { 2 }$ refers to the timestamp of a certain beat under $2 2 { , } 0 5 0 \mathrm { H z }$ , $\Delta t$ is the time offset threshold. + +In the experiment, we set $\Delta t = 1 / 1 5 s$ (FPS is 15) and calculate the beat alignment ratio for audio data of three styles respectively. We randomly sample 10 audio clips from each style to calculate the beat alignment ratio and do the sampling for 10 times to take the average. As shown in Table 5, most of beats extracted with $1 5 { , } 4 0 0 \mathrm { H z }$ are aligned with those extracted with $2 2 { , } 0 5 0 \mathrm { H z }$ . Besides, we randomly sample an audio clip and visualize beat tracking curves of its first 20 seconds under the sampling rates of $1 5 { , } 4 0 0 \mathrm { H z }$ and $2 2 { , } 0 5 0 \mathrm { H z }$ . As we can see in Figure 5, most of the beats from $1 5 { , } 4 0 0 \mathrm { H z }$ and $2 2 { , } 0 5 0 \mathrm { H z }$ are aligned within the time offset threshold $\Delta t = 1 / 1 5 s$ . + +Table 5: The beat alignment ratio $B _ { 2 } / B _ { 1 }$ . + +
CategoryBeat Alignment Ratio (%)
Ballet88.7
Hiphop93.2
Japanese Pop91.6
+ +![](images/960e4e9bd2cd632b5f494cde13e98e1b266fa0bc90a5990a3085ee12ef8de7cc.jpg) +Figure 5: Beat tracking curves for the first 20 seconds of a music clip randomly sampled from audio data under the sampling rates of $1 5 { , } 4 0 0 \mathrm { H z }$ and $2 2 { , } 0 5 0 \mathrm { H z }$ . Most of the beats are aligned within the time offset threshold $\Delta t = 1 / 1 5 s$ . \ No newline at end of file diff --git a/parse/train/xGZG2kS5bFk/xGZG2kS5bFk_content_list.json b/parse/train/xGZG2kS5bFk/xGZG2kS5bFk_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7b9eba5652bac34d87166d4a8caac5aa4748a742 --- /dev/null +++ b/parse/train/xGZG2kS5bFk/xGZG2kS5bFk_content_list.json @@ -0,0 +1,1528 @@ +[ + { + "type": "text", + "text": "DANCE REVOLUTION: LONG-TERM DANCE GENERA-TION WITH MUSIC VIA CURRICULUM LEARNING", + "text_level": 1, + "bbox": [ + 176, + 98, + 826, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruozi Huang1∗ † Huang $\\mathbf { H } \\mathbf { u } ^ { 2 * }$ Wei ${ \\bf W } { \\bf u } ^ { 3 }$ Kei Sawada4 Mi Zhang1 Daxin Jiang2‡ Fudan University1 Microsoft STCA2 Meituan3 Rinna AI4 \n{rzhuang17,mi_zhang}@fudan.edu.cn, {huahu,djiang}@microsoft.com, wuwei19850318@gmail.com, keisawada@rinna.co.jp ", + "bbox": [ + 183, + 167, + 795, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 263, + 544, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dancing to music is one of human’s innate abilities since ancient times. In machine learning research, however, synthesizing dance movements from music is a challenging problem. Recently, researchers synthesize human motion sequences through autoregressive models like recurrent neural network (RNN). Such an approach often generates short sequences due to an accumulation of prediction errors that are fed back into the neural network. This problem becomes even more severe in the long motion sequence generation. Besides, the consistency between dance and music in terms of style, rhythm and beat is yet to be taken into account during modeling. In this paper, we formalize the music-conditioned dance generation as a sequence-to-sequence learning problem and devise a novel seq2seq architecture to efficiently process long sequences of music features and capture the fine-grained correspondence between music and dance. Furthermore, we propose a novel curriculum learning strategy to alleviate error accumulation of autoregressive models in long motion sequence generation, which gently changes the training process from a fully guided teacher-forcing scheme using the previous ground-truth movements, towards a less guided autoregressive scheme mostly using the generated movements instead. Extensive experiments show that our approach significantly outperforms the existing state-of-the-arts on automatic metrics and human evaluation. We also make a demo video to demonstrate the superior performance of our proposed approach at https://www.youtube.com/watch?v $=$ lmE20MEheZ8. ", + "bbox": [ + 233, + 294, + 766, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 598, + 336, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Arguably, dancing to music is one of human’s innate abilities, as we can spontaneously sway along with the tempo of music we hear. The research in neuropsychology indicates that our brain is hardwired to make us move and synchronize with music regardless of our intention (Chen et al., 2008). Another study of archaeology also suggests that dance is a social communication skill among early humans connected to the ability of survival long time ago (Adshead-Lansdale & Layson, 2006). Nowadays, dance (to music) has become a means to the cultural promotion, a method of emotional expression, a tool for socialization and an art form to bring aesthetic enjoyment. The neurological mechanism behind dancing behavior and the unique value of dance to society motivate us to explore a computational approach to dance creation from a piece of music in artificial intelligence research. Such work is potentially beneficial to a wide range of applications such as dance creation assistant in art and sports, character motion generation for audio games and research on cross-modal behavior. ", + "bbox": [ + 174, + 631, + 825, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In literature, the music-conditioned dance generation is a relatively new task that attracts increasing research interests recently. Early works (Fan et al., 2011; Lee et al., 2013) synthesize dance sequences from music by retrieval-based methods, which show the limited creativity in practice. Recently, Lee et al. (2019) formulate the task from the generative perspective, and further propose a decompositionto-composition framework. Their model first generates basic dance units from the music clips and then composes them by using the last pose of current unit to initialize the first pose of the next unit. Although this approach shows better performance than the retrieval-based methods, several challenges still remain. ", + "bbox": [ + 174, + 790, + 825, + 873 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 823, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First, existing generative methods synthesize new human motion sequences through autoregressive models like RNN, which tend to result in short sequences. In other words, the generated sequences often quickly freeze within a few seconds due to an accumulation of prediction errors that are fed back into the neural network. This problem becomes even more severe in long motion sequence generation. For instance, composing a dance for a 1-minute music clip under 15 Frame Per Second (FPS) means generating 900 poses at one time. In practice, we need novel methods which can effectively generate long motion sequences. Besides, how to enhance the harmony between the synthesized dance movements and the given music is a largely unexplored challenge. Inutitively, the movements need to be consistent with the music in terms of style, rhythm and beat. However, to achieve this goal is non-trivial, which requires the generation model to have the capability to capture the fine-grained correspondence between music and dance. ", + "bbox": [ + 174, + 138, + 825, + 291 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we formalize music-conditioned dance generation as a sequence-to-sequence learning problem where the fine-grained correspondence between music and dance is represented through sequence modeling and their alignment is established via mapping from a sequence of acoustic features of the music to a sequence of movements of the dance. The model consists of a music encoder and a dance decoder. The encoder transforms low-level acoustic features of an input music clip into high-level latent representations via self-attention with a receptive field restricted within $k$ -nearest neighbors of an element. Thus, the encoder can efficiently process long sequences of music features, e.g., a sequence with more than 1000 elements, and model local characteristics of the music such as chord and rhythm patterns. The decoder exploits a recurrent structure to predict the dance movement frame by frame conditioned on the corresponding element in the latent representations of music feature sequence. Furthermore, we propose a curriculum learning (Bengio et al., 2009) strategy to alleviate error accumulation (Li et al., 2017) of autoregressive models in long motion sequence generation. Specifically, it gently changes the training process from a fully guided teacher-forcing scheme using the previous ground-truth movements, towards a less guided autoregressive scheme which mostly utilizes the generated movements instead. This strategy bridges the gap between training and inference of autoregressive models, and thus alleviates error accumulation at inference. ", + "bbox": [ + 174, + 299, + 825, + 520 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The length of each video clip in the dataset released by Lee et al. (2019) is only about 6 seconds, which cannot be used by other methds to generate long-term dance except for their specially designed decomposition-to-composition framework. To facilitate the task of long-term dance generation with music, we collect a high-quality dataset consisting of 1-minute video clips, totaling about 12 hours. And there are three representative styles in our dataset: “Ballet”, “Hiphop” and “Japanese Pop”. ", + "bbox": [ + 174, + 527, + 825, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions in this work are four-fold: (1) We formalize music-conditioned dance generation as a sequence-to-sequence learning problem and devise a novel seq2seq architecture for the long-term dance generation with music. (2) We propose a novel curriculum learning strategy to alleviate error accumulation of autoregressive models in long motion sequence generation. (3) To facilitate long-term dance generation with music, we collect a high-quality dataset that is available with our code1. (4) The extensive experiments show that our approach significantly outperforms the existing state-of-the-arts on both automatic metrics and human judgements. The demo video in the supplementary material also exhibits our approach can generate diverse minute-length dances that are smooth, natural-looking, style-consistent and beat-matching with the musics from test set. ", + "bbox": [ + 174, + 604, + 825, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 760, + 344, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Cross-Modal Learning. Most existing works focus on the modeling between vision and text, such as image captioning (Lu et al., 2017; Xu et al., 2015) and text-to-image generation (Reed et al., 2016; Zhang et al., 2017). There are some other works to study the translation between audio and text like Automatic Speech Recognition (ASR) (Hinton et al., 2012) and Text-To-Speech (TTS) (Oord et al., 2016). While the modeling between audio and vision is largely unexplored and the music-conditioned dance generation is a typical cross-modal learning problem from audio to vision. ", + "bbox": [ + 174, + 797, + 825, + 881 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Motion Prediction. Prediction of human motion dynamics has been a challenging problem in computer vision, which suffers from the high spatial-temporal complexity. Existing works (Chan et al., 2019; Wang et al., 2018) represent the human pose as 2D or 3D body keyjoints (Cao et al., 2017) and address the problem via sequence modeling. Early methods, such as hidden markov models (Lehrmann et al., 2014), Gaussian processes (Wang et al., 2006) and restricted boltzmann machines (Taylor et al., 2007), have to balance the model capacity and inference complexity due to complicated training procedures. Recently, neural networks dominate the human motion modeling. For instance, Fragkiadaki et al. (2015) present LSTM-3LR and Encoder-Recurrent-Decoder (ERD) as two recurrent architectures for the task; Jain et al. (2016) propose a structural-RNN to model human-object interactions in a spatio-temporal graph; and Ghosh et al. (2017) equip LSTM-3LR with a dropout autoencoder to enhance the long-term prediction. Besides, convolutional neural networks (CNNs) have also been utilized to model the human motion prediction (Li et al., 2018). ", + "bbox": [ + 173, + 103, + 825, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Audio-Conditioned Dance Generation. In the research of audio-conditioned dance generation, most existing works study 2D dance motion generation with music since the training data for paired 2D pose and music can be extracted from the huge amount of dance videos available online. Various methods have been proposed to handle this task, such as adversarial learning based methods (Lee et al., 2019; Sun et al., 2020; Ferreira et al., 2021), autoencoder methods (Tang et al., 2018) and sequence-to-sequence methods (Lee et al., 2018; Ren et al., 2019; Yalta et al., 2019; Ye et al., 2020). While these works mainly focus on exploring the different neural architectures and overlook the freezing motion issue in dance motion synthesis. In this work, we first propose a novel seq2seq architecture to model the fine-grained correspondence between music and dance, and then introduce a novel curriculum learning strategy to address the freezing motion issue caused by error accumulation (Li et al., 2017) in long-term dance motion generation. ", + "bbox": [ + 174, + 286, + 825, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 APPROACH", + "text_level": 1, + "bbox": [ + 174, + 460, + 297, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present our approach to music-conditioned dance generation. After formalization of the problem in question, we elaborate the model architecture and the dynamic auto-condition learning approach that facilitates long-term dance generation according to the given music. ", + "bbox": [ + 174, + 493, + 825, + 535 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f2bd9934103d7f435c8ce5dfd137026280301ab5097a7eec93c1dca3017d0a32.jpg", + "image_caption": [ + "Figure 1: The overview of our method. The transformer based encoder is stacked of $N$ identical layers with the local self-attention mechanism, which aims to efficiently process long features sequences extracted from music in a GPU-memory-economic way. The decoder exploits a recurrent structure to predict dance movement $y _ { i }$ frame by frame conditioned on hidden state $h _ { i }$ of decoder and latent vector $z _ { i }$ . Note that $y _ { 0 }$ is a predefined Begin Of Pose (BOP). During the training phase, the proposed curriculum learning strategy gently increases the number of steps of autoregressive scheme according to the number of training epochs. This strategy bridges the gap between training and inference of autoregressive models, and thus alleviates error accumulation at inference. " + ], + "image_footnote": [], + "bbox": [ + 202, + 550, + 789, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 PROBLEM FORMALIZATION ", + "text_level": 1, + "bbox": [ + 176, + 869, + 405, + 882 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Suppose that there is a dataset $\\mathcal { D } = \\{ ( X _ { i } , Y _ { i } ) \\} _ { i = 1 } ^ { N }$ , where $X = \\{ x _ { t } \\} _ { t = 1 } ^ { n }$ is a music clip with $x _ { t }$ being a vector of acoustic features at time-step $t$ , and $\\bar { Y = } \\{ y _ { t } \\} _ { t = 1 } ^ { n }$ is a sequence of dance movements with $y _ { t }$ aligned to $x _ { t }$ . The goal is to estimate a generation model $g ( \\cdot )$ from $\\mathcal { D }$ , and thus given a new music input $X$ , the model can synthesize dance $Y$ to music $X$ based on $g ( X )$ . We first present our seq2seq architectures chosen for music-conditioned dance generation in the following section. Later in the experiments, we empirically justify the choice by comparing the architectures with other alternatives. ", + "bbox": [ + 174, + 893, + 823, + 925 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 MODEL ARCHITECTURE ", + "text_level": 1, + "bbox": [ + 174, + 176, + 383, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the architecture of $g ( \\cdot )$ , a music encoder first transforms $X = ( x _ { 1 } , . . . , x _ { n } )$ $( x _ { i } \\in \\mathbb { R } ^ { d _ { x } } )$ into a hidden sequence $Z = ( z _ { 1 } , . . . , z _ { n } )$ $( z _ { i } \\in \\mathbb { R } ^ { d _ { z } } )$ using a local self-attention mechanism to reduce the memory requirement for long sequence modeling, and then a dance decoder exploits a recurrent structure to autoregressively predicts movements $Y = ( y _ { 1 } , . . . , y _ { n } )$ conditioned on $Z$ . ", + "bbox": [ + 173, + 202, + 825, + 260 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Music Encoder. Encouraged by the compelling performance on music generation (Huang et al., 2018), we define the music encoder with a transformer encoder structure. While the self-attention mechanism (Vaswani et al., 2017) in transformer can effectively represent the multi-scale structure of music, the quadratic memory complexity $\\scriptstyle { \\mathcal { O } } ( n ^ { 2 } )$ about sequence length $n$ impedes its application to long sequence modeling due to huge of GPU memory consumption (Child et al., 2019). To keep the effectiveness of representation and control the cost, we introduce a local self-attention mechanism that modifies the receptive field of self-attention by restricting the element connections within $k$ -nearest neighbors. Thus, the memory complexity is reduced to $\\mathcal { O } ( n k )$ . $k$ could be small in our scenario since we only pursue an effective representation for a given music clip. Therefore, the local patterns of music are encoded in $z _ { t }$ that is sufficient for the generation of movement $y _ { t }$ at time-step $t$ , which is aligned with the common sense that the dance movement at certain time-step is highly influenced by the nearby clips of music. Yet we can handle long sequences of acoustic features, e.g., more than 1000 elements, in an efficient and memory-economic way. ", + "bbox": [ + 173, + 272, + 826, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we first embed $X = ( x _ { 1 } , . . . , x _ { n } )$ into $U = ( u _ { 1 } , . . . , u _ { n } )$ with a linear layer parameterized by $W ^ { E } \\in \\mathbf { \\bar { \\mathbb { R } } } ^ { d _ { x } \\times d _ { z } }$ . Then $\\forall i \\in \\{ 1 , . . . , n \\}$ , $z _ { i }$ can be formulated as: ", + "bbox": [ + 174, + 460, + 821, + 489 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/5dbdfb230aa9b6ecb8673ef1df3a2d5a1901bf5046f8c3547666140729add6dc.jpg", + "text": "$$\nz _ { i } = \\mathcal { F } ( a _ { i } ) , \\quad a _ { i } = \\sum _ { j = i - \\lfloor k / 2 \\rfloor } ^ { i + \\lfloor k / 2 \\rfloor } \\alpha _ { i j } ( u _ { j } W _ { l } ^ { V } ) , \\quad U = X W ^ { E } ,\n$$", + "text_format": "latex", + "bbox": [ + 303, + 496, + 692, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal { F } ( \\cdot ) : \\mathbb { R } ^ { d _ { v } } \\mathbb { R } ^ { d _ { z } }$ is a feed forward neural network. Each $u _ { j }$ is only allowed to attend its $k$ -nearest neighbors including itself, where $k$ is a hyper-parameter referring to the sliding window size of the local self-attention. Then attention weight $\\alpha _ { i j }$ is calculated using a softmax function as: ", + "bbox": [ + 176, + 549, + 825, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/a0a63acb0667d3d33495087c124f59b9e1835abeaea8585f230fab1324b86adb.jpg", + "text": "$$\n\\alpha _ { i j } = \\frac { \\exp { e _ { i j } } } { \\sum _ { t = j - \\lfloor k / 2 \\rfloor } ^ { j + \\lfloor k / 2 \\rfloor } \\exp { e _ { i t } } } , e _ { i j } = \\frac { ( u _ { i } W _ { l } ^ { Q } ) ( u _ { j } W _ { l } ^ { K } ) ^ { \\top } } { \\sqrt { d _ { k } } } ,\n$$", + "text_format": "latex", + "bbox": [ + 307, + 598, + 689, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where for the the $l$ -th head, $W _ { l } ^ { Q }$ , $W _ { l } ^ { K } \\in \\mathbb { R } ^ { d _ { z } \\times d _ { k } }$ and $W _ { l } ^ { V } \\in \\mathbb { R } ^ { d _ { z } \\times d _ { v } }$ are parameters that transform $U$ into a query, a key, and a value respectively. $d _ { z }$ is the dimension of hidden state $z _ { i }$ while $d _ { k }$ is the dimension of query, key and $d _ { v }$ is the dimension of value. ", + "bbox": [ + 174, + 647, + 825, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dance Decoder. We choose a recurrent neural network as the dance decoder in consideration of two factors: (1) the chain structure can well capture the spatial-temporal dependency among human movement dynamics, which has proven to be highly effective in the state-of-the-art methods for human motion prediction (Li et al., 2018; Mao et al., 2019); (2) our proposed learning strategy is tailored for the autoregressive models like RNN, as will be described in the next section. Specifically, with $Z = ( z _ { 1 } , . . . , z _ { n } )$ , the dance movements $\\boldsymbol { Y } = ( \\hat { y } _ { 1 } , . . . , \\hat { y } _ { n } )$ are synthesized by: ", + "bbox": [ + 173, + 707, + 825, + 791 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/3f8ddabcf4f07733079df1da51a1bdcc825cb9c03f3726d9786db20821f1d134.jpg", + "text": "$$\n\\begin{array} { l } { \\hat { y } _ { i } = [ h _ { i } ; z _ { i } ] W ^ { S } + b , } \\\\ { h _ { i } = \\mathrm { R N N } ( h _ { i - 1 } , \\hat { y } _ { i - 1 } ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 416, + 796, + 578, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h _ { i }$ is the $i$ -th hidden state of the decoder and $h _ { 0 }$ is initialized by sampling from the standard normal distribution to enhance variation of the generated sequences. $[ \\cdot ; \\cdot ]$ denotes the concatenation operation. $W ^ { S } \\in \\mathbb { R } ^ { ( d _ { s } + d _ { z } ) \\times d _ { y } }$ and $b \\in \\mathbb { R } ^ { d _ { y } }$ are parameters where $d _ { s }$ and $d _ { y }$ are the dimensions of $h _ { i }$ and $\\hat { y } _ { i }$ , respectively. At the $i$ -th time-step, the decoder predicts the movement $\\hat { y } _ { i }$ conditioned on $h _ { i }$ as well as the latent feature representation $z _ { i }$ , and thus can capture the fine-grained correspondence between music feature sequence and dance movement sequence. ", + "bbox": [ + 173, + 837, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 DYNAMIC AUTO-CONDITION LEARNING APPROACH", + "text_level": 1, + "bbox": [ + 173, + 103, + 576, + 117 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Exposure bias (Ranzato et al., 2015; Lamb et al., 2016; Zhang et al., 2019) is a notorious issue in natural language generation (NLG), which refers to the train-test discrepancy of autoregressive generative models. This kind of models are never exposed to their own prediction errors at the training phase due to the conventional teacher-forcing training scheme. However, compared to NLG, motion sequence generation suffers from the more severe exposure bias problem. In NLG, the bias of predicted probability distribution over vocabulary can be corrected by sampling strategy. For instance, we can still generate the target token with index 2 by sampling on probability distribution [0.3, 0.3, 0.4] whose groundtruth is [0, 0, 1]. While any small biases of the predicted motion at each time-step will be accumulated and propagated to the future, since each generated motion is a real-valued vector (rather than a discrete token) representing 2D or 3D body keyjoints in the continuous space. As a result, the generated motion sequences tend to quickly freeze within a few seconds. This problem becomes even worse in the generation of long motion sequences, e.g., more than 1000 movements. ", + "bbox": [ + 173, + 128, + 826, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scheduled sampling (Bengio et al., 2015) is proposed to alleviate exposure bias in NLG, which does not work in long motion sequence generation. Since its sampling-based strategy would feed long predicted sub-sequences of high bias into the model at early stage, which causes gradient vanishing at training due to error accumulation of these high-biased sub-sequences. Motivated by this, we propose a dynamic auto-condition training method as a novel curriculum learning strategy to alleviate the error accumulation of autoregressive models in long motion sequence generation. Specifically, our learning strategy gently changes the training process from a fully guided teacher-forcing scheme using the previous ground-truth movements, towards a less guided autoregressive scheme mostly using the generated movements instead, which bridges the gap between training and inference. ", + "bbox": [ + 174, + 303, + 825, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 1, The decoder predicts $Y _ { t g t } = \\{ y _ { 1 } , . . . , y _ { p } , y _ { p + 1 } , . . . , y _ { p + q } , y _ { p + q + 1 } , . . . \\}$ from $Y _ { i n } = \\{ y _ { 0 } , \\hat { y _ { 1 } } , . . . , \\hat { y _ { p } } , y _ { p + 1 } , . . . , y _ { p + q } , \\hat { y } _ { p + q + 1 } , . . . \\}$ which alternates the predicted sub-sequences with length $p$ (e.g., $\\hat { y } _ { 1 } , . . . , \\hat { y } _ { p } )$ and the ground-truth sub-sequences with length $q$ (e.g., $y _ { p + 1 } , . . . , y _ { { p + q } } )$ . $y _ { 0 }$ is a predefined begin of pose (BOP). During the training phase, we fix $q$ and gradually increase $p$ according to a growth function $f ( t ) \\in \\{ c o n s t , \\lfloor \\lambda t \\rfloor , \\lfloor \\lambda t ^ { 2 } \\rfloor , \\lfloor \\lambda e ^ { t } \\rfloor \\}$ where $t$ is the number of training epochs and $\\lambda < 1$ controls how frequently $p$ is updated. We choose $f ( t ) = \\lfloor \\lambda t \\rfloor$ in practice through the empirical comparison. Note that our learning strategy degenerates to the method proposed in Li et al. (2017) when $f ( t ) = c o n s t$ . While the advantage of a dynamic strategy over a static strategy lies in that the former introduces the curriculum learning spirit to dynamically increase the difficulty of curriculum (e.g., larger $p$ ) as the model capacity improves during training, which can further improve the model capacity and alleviate the error accumulation of autoregressive predictions. Finally, we estimate parameters of $g ( \\cdot )$ by minimizing the $\\ell _ { 1 }$ loss on $\\mathcal { D }$ : ", + "bbox": [ + 173, + 434, + 825, + 602 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/3096d0a2855285191c1b6add72da2f174905654e997f8cf923920c0b29a39611.jpg", + "text": "$$\n\\ell _ { 1 } = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\left| \\left| g ( X _ { i } ) - Y _ { t g t } ^ { ( i ) } \\right| \\right| _ { 1 }\n$$", + "text_format": "latex", + "bbox": [ + 397, + 609, + 599, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 EXPERIMENTAL SETUP ", + "text_level": 1, + "bbox": [ + 176, + 667, + 398, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 DATASET COLLECTION AND PREPROCESSING ", + "text_level": 1, + "bbox": [ + 174, + 699, + 529, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although the dataset (Lee et al., 2019) contains about 70-hour video clips, the average length of each clip is only about 6 seconds. Moreover, their dataset cannot be used by other methods to generate longterm dance except for their specially designed decomposition-to-composition framework. To facilitate the task of long-term dance generation with music, we collect a high-quality dataset consisting of 790 one-minute clips of pose data with 15 FPS, totaling about 12 hours. And there are three styles in the dataset including “Ballet”, “Hiphop” and “Japanese Pop”. Table 1 shows the statistics of our dataset. In the experiment, we randomly split the dataset into the $90 \\%$ training set and $10 \\%$ test set. ", + "bbox": [ + 173, + 726, + 825, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pose Preprocess. For the human pose estimation, we leverage OpenPose (Cao et al., 2017) to extract 2D body keyjoints from videos. Each pose consists of 25 keyjoints2 and is represented by a 50-dimension vector in the continuous space. In practice, we develop a linear interpolation algorithm to find the missing keyjoints from nearby frames to reduce the noise in extracted pose data. ", + "bbox": [ + 174, + 830, + 825, + 887 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2f2496b0f3b777149324c651c6b22b6b1a131d08728d71168ae9012dcb9e265e.jpg", + "table_caption": [ + "Table 1: Statistics of the dataset. " + ], + "table_footnote": [], + "table_body": "
CategoryNum of ClipsClip Length (min)FPSResolution
Ballet136115720P
Hiphop298115720P
Japanese Pop356115720P
", + "bbox": [ + 254, + 130, + 743, + 202 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Audio Preprocess. Librosa (McFee et al., 2015) is a well-known audio and music analysis library in the music information retrieval, which provides flexible ways to extract the spectral and rhythm features of audio data. Specifically, we extract the following features: mel frequency cepstral coefficients (MFCC), MFCC delta, constant- $Q$ chromagram, tempogram and onset strength. To better capture the beat information of music, we convert onset strength into a one-hot vector as an additional feature to explicitly represent the beat sequence of music, i.e., beat one-hot. Finally, we concatenate these audio features as the representation of a music frame, as shown in Table 4 in Section 6.2. During extracting audio features, the sampling rate is $1 5 { , } 4 0 0 \\mathrm { H z }$ and hop size is 1024. Hence, we have 15 audio samples per second that is aligned with the 15 FPS of pose data. ", + "bbox": [ + 174, + 228, + 825, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 IMPLEMENTATION DETAILS ", + "text_level": 1, + "bbox": [ + 176, + 373, + 405, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The music encoder consists of a stack of $N = 2$ identical layers. Each layer has two sublayers: a local self-attention sublayer with $l = 8$ heads and a position-wise fully connected feed-forward sublayer with 1024 hidden units. Each head contains a scaled dot-product attention layer with $d _ { k } = d _ { v } = 6 4$ and its receptive yield is restricted by setting $k = 1 0 0$ . Then we set the sequence length $n = 9 0 0$ , dimension of acoustic feature vector $d _ { x } = 4 3 8$ , dimension of hidden vector $d _ { z } = 2 5 6$ , respectively. The dance decoder is a 3-layer LSTM with $d _ { s } = 1 0 2 4$ and the dimension of pose vector $d _ { y } = 5 0$ . We set $\\lambda = 0 . 0 1$ and $q = 1 0$ for the proposed learning approach and train the model using the Adam optimizer with the learning rate 1e-4 on 2 NVIDIA V100 GPUs. ", + "bbox": [ + 174, + 400, + 826, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3 BASELINES", + "text_level": 1, + "bbox": [ + 174, + 531, + 294, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The music-conditioned generation is an emerging task and there are few methods proposed to solve this problem. In our experiment, we compare our proposed approach to the following baselines: (1) Dancing2Music. We use Dancing2Music (Lee et al., 2019) as the primary baseline that is the previous state-of-the-art; (2) Aud-MoCoGAN. Aud-MoCoGAN is an auxiliary baseline used in Dancing2Music, which is original from MoCoGAN (Tulyakov et al., 2018); (3) LSTM. Shlizerman et al. (2018) propose a LSTM network to predict body dynamics from the audio signal. We modify their open-source code to take audio features of music as inputs and produce human pose sequences. ", + "bbox": [ + 174, + 558, + 825, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 EXPERIMENTAL RESULTS ", + "text_level": 1, + "bbox": [ + 176, + 678, + 418, + 693 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct extensive experiments to evaluate our approach and compare it to the aforementioned baselines. Note that we do not include comparison experiment on the dataset collect by Lee et al. (2019) due to the above issue in Section 4.1. ", + "bbox": [ + 174, + 710, + 825, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1 AUTOMATIC METRICS AND HUMAN EVALUATION ", + "text_level": 1, + "bbox": [ + 174, + 771, + 558, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the different methods by automatic metrics and conduct a human evaluation on motion realism and smoothness of generated dance movements, as well as style consistency. Specifically, we randomly sample 60 music clips from the test set to generate dances, then divide the dances generated by three methods and the real dances into four pairs: (LSTM, ours), (Dancing2Music, ours), (real, ours) and (real, real). We invite 10 amateur dancers as the annotators to answer 3 questions for each pair that is blind for them (use (A, B) pair instead): (1) Which dance is more realistic regardless of music? (2) Which dance is more smooth regardless of music? (3) Which dance matches the music better with respect to style? Note that, we do not include Aud-MoCoGAN in the human evaluation, since both it and Dancing2Music are GAN based methods while the latter is the state-of-the-art. ", + "bbox": [ + 173, + 797, + 826, + 924 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/160f9ed77699ffb31249fa04fe6af19e09624d27fa86094970947262a6f5633e.jpg", + "image_caption": [ + "Figure 2: Human evaluation results on motion realism, style consistency and smoothness. We conduct a human evaluation to ask annotators to select the dances that are more realistic and more smooth regardless of music, more style-consistent with music through pairwise comparison. " + ], + "image_footnote": [], + "bbox": [ + 171, + 220, + 825, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motion Realism, Style Consistency and Smoothness. We evaluate the visual quality and realism of generated dances by Fréchet Inception Distance (FID) (Heusel et al., 2017), which is used to measure how close the distribution of generated dances is to the real. Similar to Lee et al. (2019), we train a style classifier on dance movement sequences of three styles and use it to extract features for the given dances. Then we calculate FID between the synthesized dances and the real. As shown in Table 2, our FID score is significantly lower than those of baselines and much closer to that of the real, which means our generated dances are more motion-realistic and more likely to the real dances. Besides, we use the same classifier to measure the style accuracy of generated dance to the music, and our approach achieves $7 7 . 6 \\%$ accuracy and significantly outperforms Dancing2Music by $1 7 . 2 \\%$ . ", + "bbox": [ + 174, + 547, + 825, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human evaluation in Figure 2 consistently shows the superior performance of our approach, compared to baselines on motion realism, style consistency and smoothness. We observe the dances generated by LSTM have lots of floating movements and tend to quickly freeze within several seconds due to the error accumulation, which results in lower preferences. While Dancing2Music can generate the smooth dance units, the synthesized dances still have significant jumps where dance units are connected due to the inherent gap between them, which harm its overall performance in practice. This is also reflected in the preference comparisons to our approach, in which only $3 5 \\%$ annotators prefer Dancing2Music on motion realism and $2 1 . 7 \\%$ prefer Dancing2Music on smoothness. Besides, the result on style consistency shows our approach can generate more style-consistent dances with musics. Since our approach models the fine-grained correspondence between music and dance at the methodological level and also introduces more specific music features. However, Dancing2Music only considers a global style feature extracted by a pre-trained classifier in the dance generation. ", + "bbox": [ + 173, + 680, + 825, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the comparison to real dances, $4 1 . 2 \\%$ annotators prefer our method on motion realism while $3 0 . 3 \\%$ prefer on style consistency. Additionally, we found that $5 7 . 9 \\%$ annotators prefer our approach on smoothness compared to real dances. Since the imperfect OpenPose (Cao et al., 2017) introduces the minor noise on pose data extracted from dance videos. On the other hand, in the preprocessing, we develop an interpolation algorithm to find missing keyjoints and reduce the noise of training pose data. ", + "bbox": [ + 174, + 854, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides, this also indicates the chain structure based decoder can well capture the spatial-temporal dependency among human motion dynamics and produce the smooth movements. ", + "bbox": [ + 173, + 103, + 823, + 132 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Beat Coverage and Hit Rate. We also evaluate the beat coverage and hit rate introduced in Lee et al. (2019), which defines the beat coverage as $B _ { k } / B _ { m }$ and beat hit rate as $B _ { a } / B _ { k }$ , where $B _ { k }$ is the number of kinematic beats, $B _ { m }$ is the number of musical beats and $B _ { a }$ is the number of kinematic beats that are aligned with the musical beats. An early study (Ho et al., 2013) reveals that kinematic beat frames occur when the direction of the movement changes. Therefore, similar to Yalta et al. (2019), we use the standard deviation (SD) of movement to detect the kinematic beats in our experiment. The onset strength (Ellis, 2007) is a common way to detect musical beats. Figure 3 shows two short clips of motion SD curves and aligned musical beats. We observed that the kinematic beats occur where the musical beats occur, which is consistent with the common sense that dancers would step on musical beats during dancing but do not step on every musical beat. As we can see in Table 2, LSTM and Aud-MoCoGAN generate dances with few kinematic beats and most of them do not match the musical beats. Our method outperforms Dancing2Music by $6 . 1 \\%$ on beat coverage and $2 . 7 \\%$ on hit rate, which indicates that introducing the features about musical beat into model is beneficial to the better beat alignment of generated dance and input music. ", + "bbox": [ + 173, + 138, + 825, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/167107ef7ff1cf1ce81b167922e4cc91ebfc7d4110451112c0a7a53b8c70d63b.jpg", + "image_caption": [ + "Figure 3: Beat tracking curves for music and dance by onset strength and motion standard deviation respectively. The left is a short clip of ballet while the right is a short clip of hip-hop. The red circles are kinematic beats and dash lines denote the musical beats. " + ], + "image_footnote": [], + "bbox": [ + 202, + 347, + 789, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Diversity and Multimodality. Lee et al. (2019) introduce the diversity metric and the multimodality metric. The diversity metric evaluates the variations among generated dances corresponding to various music, which reflects the generalization ability of the model and its dependency on the input music. Similarly, we use the average feature distance as the measurement and these features are extracted by the same classifier used in measuring FID. We randomly sample 60 music clips from test set to generate dances and randomly pick 500 combinations to compute the average feature distance. Results in Table 2 show the performance of our approach is superior to Dancing2Music on diversity. The reason is that Dancing2Music only considers a global music style feature in generation while different music of the same style have the almost same style feature. Our approach models the fine-grained correspondence between music and dance at the methodological level, simultaneously takes more specific music features into account. In other words, our approach is more dependent on input music and thus can generate different samples for difference music clips of the same style. ", + "bbox": [ + 174, + 540, + 825, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multimodality metric evaluates the variations of generated dances for the same music clip. Specifically, we generate 5 dances for each of randomly sampled 60 music clips and compute the average feature distance. As we can see, our approach slightly underperforms Dancing2Music and AudMoCoGAN since it has more dependency on music inputs as aforementioned, which make the generated dances have the limited variation patterns given the same music clip. ", + "bbox": [ + 174, + 713, + 825, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2 LONG-TERM DANCE GENERATION", + "text_level": 1, + "bbox": [ + 176, + 800, + 455, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further investigate the performance of different methods on long-term dance generation, we evaluate FID of dances generated by different methods over time. Specifically, we first split a generated dance with 1 minute into 15 four-second clips and measure FID of each clip. Figure 4 shows FID scores of LSTM and Aud-MoCoGAN grow rapidly in the early stage due to error accumulation and converge to the high FID, since the generated dances quickly become frozen. Dancing2Music maintains the relatively stable FID scores all the time, which benefits from its decomposition-tocomposition method. While its curve still has subtle fluctuations since it synthesizes whole dances by composing the generated dance units. Compared to baselines, FID of our method are much lower and close to real dances, which validates the good performance of our method on long-term generation. ", + "bbox": [ + 174, + 825, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b131c94c16a3461b63bf6187b0f782cc2dc65245cdd3adee1cfccc5add824bad.jpg", + "image_caption": [ + "Figure 4: FID curves of different methods over time. " + ], + "image_footnote": [], + "bbox": [ + 338, + 103, + 653, + 261 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/ec6b92b6e4c9db67e824443b66fac70b206a85a4a2c1c09c95deafb57a1a89d9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Encoder Structure LSTMFIDACC (%)
7.945
ConvS2S encoder13.354
Global self-attention3.680
Local self-attention4.378.5
", + "bbox": [ + 202, + 339, + 465, + 420 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/93832bc96a22cc8bcbf625fa826d04271ee181ac45bbcbe9616453589df5ecae.jpg", + "table_caption": [ + "Table 3: Left: Automatic metrics of our model with different encoder structures. Right: Automatic metrics of our model training by different learning strategies. " + ], + "table_footnote": [], + "table_body": "
Learning ApproachFIDACC (%)
Teacher-forcing61.25.3
Auto-condition (con st)15.735
Ours ([λet J)9.869
Ours ([λt²])6.473
Ours ([λt])5.177.6
", + "bbox": [ + 501, + 333, + 776, + 425 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 452, + 823, + 479 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.3 ABLATION STUDY ", + "text_level": 1, + "bbox": [ + 176, + 496, + 341, + 511 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparison on Encoder Structures. We compare the different encoder structures in our model with the same LSTM decoder and the same curriculum learning strategy, as shown in Table 3. The transformer encoders outperform LSTM and the encoder in ConvS2S (Gehring et al., 2017) on both FID and style consistency, due to its superior performance on sequence modeling. Although the transformer encoder with our proposed local self-attention slightly underperforms the global self-attention by 0.7 higher in FID and $1 . 5 \\%$ lower in ACC, it has much fewer parameters in our settings. This confirms the effectiveness of the local self-attention on long sequence modeling. ", + "bbox": [ + 174, + 522, + 825, + 619 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparison on Learning Strategies. To investigate the performance of our proposed learning approach, we conduct an ablation study to compare the performances of our model using different learning strategies. As we can see the right of in Table 3, all curriculum learning strategies significantly outperform the static auto-condition strategy (Li et al., 2017). Among them, the linear growth function $f ( t ) = \\lfloor \\lambda t \\rfloor$ performs best, which might be that increasing the difficulty of curriculum too fast would result in the model degradation. Besides, the original teacher-forcing strategy has the highest FID and the lowest style accuracy due to severe error accumulation problem, which indicates that using our proposed curriculum learning strategy to train the model can effectively alleviate this issue. ", + "bbox": [ + 174, + 626, + 825, + 738 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 CONCLUSION ", + "text_level": 1, + "bbox": [ + 174, + 758, + 318, + 775 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we propose a novel seq2seq architecture for long-term dance generation with music, e.g., about one-minute length. To efficiently process long sequences of music features, we introduce a local self-attention mechanism in the transformer encoder structure to reduce the quadratic memory requirements to the linear in the sequence length. Besides, we also propose a dynamic auto-condition training method as a novel curriculum learning strategy to alleviate the error accumulation of autoregressive models in long motion sequence generation, thus facilitate the long-term dance generation. Extensive experiments have demonstrated the superior performance of our proposed approach on automatic metrics and human evaluation. Future works include the exploration on end-to-end methods that could work with raw audio data instead of preprocessed audio features. ", + "bbox": [ + 173, + 790, + 825, + 915 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 102, + 287, + 118 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Janet Adshead-Lansdale and June Layson. Dance history: An introduction. Routledge, 2006. ", + "bbox": [ + 169, + 126, + 784, + 141 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. In Advances in Neural Information Processing Systems, pp. 1171–1179, 2015. ", + "bbox": [ + 174, + 150, + 825, + 193 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pp. 41–48, 2009. ", + "bbox": [ + 169, + 202, + 825, + 233 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 7291–7299, 2017. ", + "bbox": [ + 174, + 241, + 823, + 285 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A Efros. Everybody dance now. In Proceedings of the IEEE International Conference on Computer Vision, pp. 5933–5942, 2019. ", + "bbox": [ + 173, + 294, + 825, + 323 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Joyce L Chen, Virginia B Penhune, and Robert J Zatorre. Listening to musical rhythms recruits motor regions of the brain. Cerebral cortex, 18(12):2844–2854, 2008. ", + "bbox": [ + 173, + 332, + 821, + 362 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. ", + "bbox": [ + 173, + 371, + 821, + 400 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pp. 2262–2271, 2019. ", + "bbox": [ + 173, + 409, + 826, + 452 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Daniel PW Ellis. Beat tracking by dynamic programming. Journal of New Music Research, 36(1): 51–60, 2007. ", + "bbox": [ + 173, + 462, + 823, + 491 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Rukun Fan, Songhua Xu, and Weidong Geng. Example-based automatic music-driven conventional dance motion synthesis. IEEE transactions on visualization and computer graphics, 18(3):501–515, 2011. ", + "bbox": [ + 173, + 501, + 826, + 542 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Joao P Ferreira, Thiago M Coutinho, Thiago L Gomes, José F Neto, Rafael Azevedo, Renato Martins, and Erickson R Nascimento. Learning to dance: A graph convolutional adversarial network to generate realistic dance motions from audio. Computers & Graphics, 94:11–21, 2021. ", + "bbox": [ + 173, + 553, + 825, + 597 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models for human dynamics. In Proceedings of the IEEE International Conference on Computer Vision, pp. 4346–4354, 2015. ", + "bbox": [ + 173, + 604, + 825, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional sequence to sequence learning. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pp. 1243–1252. JMLR. org, 2017. ", + "bbox": [ + 173, + 657, + 826, + 702 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Partha Ghosh, Jie Song, Emre Aksan, and Otmar Hilliges. Learning human motion models for long-term predictions. In 2017 International Conference on 3D Vision (3DV), pp. 458–466. IEEE, 2017. ", + "bbox": [ + 173, + 710, + 826, + 752 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in neural information processing systems, pp. 6626–6637, 2017. ", + "bbox": [ + 173, + 762, + 825, + 806 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. IEEE Signal processing magazine, 29(6):82–97, 2012. ", + "bbox": [ + 174, + 814, + 825, + 872 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Chieh Ho, Wei-Tze Tsai, Keng-Sheng Lin, and Homer H Chen. Extraction and alignment evaluation of motion beats for street dance. In 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 2429–2433. IEEE, 2013. ", + "bbox": [ + 174, + 882, + 823, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Cheng-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit, Noam Shazeer, Ian Simon, Curtis Hawthorne, Andrew M Dai, Matthew D Hoffman, Monica Dinculescu, and Douglas Eck. Music transformer. arXiv preprint arXiv:1809.04281, 2018. ", + "bbox": [ + 176, + 103, + 823, + 146 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ashesh Jain, Amir R Zamir, Silvio Savarese, and Ashutosh Saxena. Structural-rnn: Deep learning on spatio-temporal graphs. In Proceedings of the ieee conference on computer vision and pattern recognition, pp. 5308–5317, 2016. ", + "bbox": [ + 176, + 154, + 823, + 196 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alex M Lamb, Anirudh Goyal Alias Parth Goyal, Ying Zhang, Saizheng Zhang, Aaron C Courville, and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. In Advances In Neural Information Processing Systems, pp. 4601–4609, 2016. ", + "bbox": [ + 174, + 204, + 823, + 247 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. In Advances in Neural Information Processing Systems, pp. 3581–3591, 2019. ", + "bbox": [ + 174, + 255, + 825, + 297 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Juheon Lee, Seohyun Kim, and Kyogu Lee. Listen to dance: Music-driven choreography generation using autoregressive encoder-decoder network. arXiv preprint arXiv:1811.00818, 2018. ", + "bbox": [ + 173, + 305, + 823, + 335 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Minho Lee, Kyogu Lee, and Jaeheung Park. Music similarity-based approach to generating dance motion sequence. Multimedia tools and applications, 62(3):895–912, 2013. ", + "bbox": [ + 171, + 342, + 823, + 372 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Andreas M Lehrmann, Peter V Gehler, and Sebastian Nowozin. Efficient nonlinear markov models for human motion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1314–1321, 2014. ", + "bbox": [ + 176, + 378, + 823, + 422 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Chen Li, Zhen Zhang, Wee Sun Lee, and Gim Hee Lee. Convolutional sequence to sequence model for human dynamics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 5226–5234, 2018. ", + "bbox": [ + 176, + 429, + 823, + 473 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Zimo Li, Yi Zhou, Shuangjiu Xiao, Chong He, Zeng Huang, and Hao Li. Auto-conditioned recurrent networks for extended complex human motion synthesis. arXiv preprint arXiv:1707.05363, 2017. ", + "bbox": [ + 169, + 481, + 825, + 510 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jiasen Lu, Caiming Xiong, Devi Parikh, and Richard Socher. Knowing when to look: Adaptive attention via a visual sentinel for image captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 375–383, 2017. ", + "bbox": [ + 173, + 517, + 823, + 560 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. In Proceedings of the IEEE International Conference on Computer Vision, pp. 9489–9497, 2019. ", + "bbox": [ + 174, + 568, + 823, + 611 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Brian McFee, Colin Raffel, Dawen Liang, Daniel PW Ellis, Matt McVicar, Eric Battenberg, and Oriol Nieto. librosa: Audio and music signal analysis in python. In Proceedings of the 14th python in science conference, volume 8, 2015. ", + "bbox": [ + 173, + 618, + 825, + 661 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499, 2016. ", + "bbox": [ + 171, + 669, + 825, + 713 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Marc’Aurelio Ranzato, Sumit Chopra, Michael Auli, and Wojciech Zaremba. Sequence level training with recurrent neural networks. arXiv preprint arXiv:1511.06732, 2015. ", + "bbox": [ + 173, + 719, + 825, + 750 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. arXiv preprint arXiv:1605.05396, 2016. ", + "bbox": [ + 174, + 757, + 823, + 786 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Xuanchi Ren, Haoran Li, Zijian Huang, and Qifeng Chen. Music-oriented dance video synthesis with pose perceptual loss. arXiv preprint arXiv:1912.06606, 2019. ", + "bbox": [ + 174, + 792, + 823, + 823 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Eli Shlizerman, Lucio Dery, Hayden Schoen, and Ira Kemelmacher-Shlizerman. Audio to body dynamics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 7574–7583, 2018. ", + "bbox": [ + 173, + 830, + 825, + 873 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Guofei Sun, Yongkang Wong, Zhiyong Cheng, Mohan S Kankanhalli, Weidong Geng, and Xiangdong Li. Deepdance: music-to-dance motion choreography with adversarial learning. IEEE Transactions on Multimedia, 23:497–509, 2020. ", + "bbox": [ + 174, + 881, + 825, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Taoran Tang, Jia Jia, and Hanyang Mao. Dance with melody: An lstm-autoencoder approach to music-oriented dance synthesis. In Proceedings of the 26th ACM international conference on Multimedia, pp. 1598–1606, 2018. ", + "bbox": [ + 176, + 103, + 823, + 146 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Graham W Taylor, Geoffrey E Hinton, and Sam T Roweis. Modeling human motion using binary latent variables. In Advances in neural information processing systems, pp. 1345–1352, 2007. ", + "bbox": [ + 171, + 155, + 823, + 184 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1526–1535, 2018. ", + "bbox": [ + 176, + 193, + 823, + 236 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pp. 5998–6008, 2017. ", + "bbox": [ + 173, + 244, + 823, + 287 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Jack Wang, Aaron Hertzmann, and David J Fleet. Gaussian process dynamical models. In Advances in neural information processing systems, pp. 1441–1448, 2006. ", + "bbox": [ + 171, + 295, + 823, + 325 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. arXiv preprint arXiv:1808.06601, 2018. ", + "bbox": [ + 173, + 333, + 821, + 363 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In International conference on machine learning, pp. 2048–2057, 2015. ", + "bbox": [ + 174, + 371, + 823, + 414 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Nelson Yalta, Shinji Watanabe, Kazuhiro Nakadai, and Tetsuya Ogata. Weakly-supervised deep recurrent neural networks for basic dance step generation. In 2019 International Joint Conference on Neural Networks (IJCNN), pp. 1–8. IEEE, 2019. ", + "bbox": [ + 173, + 422, + 823, + 465 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Zijie Ye, Haozhe Wu, Jia Jia, Yaohua Bu, Wei Chen, Fanbo Meng, and Yanfeng Wang. Choreonet: Towards music to dance synthesis with choreographic action unit. In Proceedings of the 28th ACM International Conference on Multimedia, pp. 744–752, 2020. ", + "bbox": [ + 174, + 473, + 825, + 517 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pp. 5907–5915, 2017. ", + "bbox": [ + 173, + 526, + 825, + 582 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Wen Zhang, Yang Feng, Fandong Meng, Di You, and Qun Liu. Bridging the gap between training and inference for neural machine translation. arXiv preprint arXiv:1906.02448, 2019. ", + "bbox": [ + 173, + 592, + 825, + 621 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "APPENDIX ", + "text_level": 1, + "bbox": [ + 176, + 103, + 263, + 117 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.1 MORE DETAILS ABOUT DATA COLLECTION ", + "text_level": 1, + "bbox": [ + 174, + 132, + 516, + 147 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The dance videos are collected from YouTube by crowd workers, which are all solo dance videos with 30FPS. We trim the beginning and ending several seconds for each of collected videos to remove the silence parts, and split them into one-minute video clips. Then, we extract 2D pose data from these video clips by OpenPose (Cao et al., 2017) with 15FPS setting and collect 790 one-minute clips of pose data with 15FPS, totaling about 13 hours. Finally, we extract the corresponding audio data from video clips by FFmpeg3. The code is implemented based on PyTorch framework. MIT License will be used for the released code and data. ", + "bbox": [ + 173, + 159, + 825, + 256 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.2 EXTRACTED MUSIC FEATURES ", + "text_level": 1, + "bbox": [ + 176, + 272, + 429, + 286 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we detail the extracted features of music that are feeded into our model. Specifically, we leverage the public Librosa (McFee et al., 2015) to extract the music features including: 20- dim MFCC, 20-dim MFCC delta, 12-dim chroma, 384-dim tempogram, 1-dim onset strength (i.e., envelope) and 1-dim one-hot beat, as shown in Table 4. ", + "bbox": [ + 174, + 299, + 826, + 354 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b358b082044b77105e16eb9a4b95f99a56e408023a5b5130b4909fa107e10dbf.jpg", + "table_caption": [ + "Table 4: Acoustic features of music. " + ], + "table_footnote": [], + "table_body": "
FeatureCharacteristicDimension
MFCCPitch20
MFCC deltaPitch20
Constant-QchromagramPitch12
TempogramStrength384
Onset strengthStrength1
Beat one-hotBeat1
", + "bbox": [ + 320, + 395, + 676, + 501 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.3 DOWNSTREAM APPLICATIONS ", + "text_level": 1, + "bbox": [ + 176, + 522, + 424, + 537 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The downstream applications of our proposed audio-conditioned dance generation method include: (1) Our model can be used to help professional people choreograph new dances for a given song and teach human how to dance with regard to this song; (2) With the help of 3D human pose estimation (Ci et al., 2019) and 3D animation driving techniques, our model can be used to drive the various 3D character models, such as the 3D model of Hatsune Miku (very popular virtual character in Japan). This technique has the great potential for the virtual advertisement video generation that could be used for the promotion events on social medias like TikTok, Twitter and etc. ", + "bbox": [ + 173, + 547, + 826, + 646 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.4 MUSICAL BEAT DETECTION UNDER LOW SAMPLING RATES ", + "text_level": 1, + "bbox": [ + 174, + 662, + 630, + 676 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we conduct an additional experiment to evaluate the performance of Librosa beat tracker under $1 5 { , } 4 0 0 \\mathrm { H z }$ sampling rate. Specifically, we first utilize it to extract the onset beats from audio data with $1 5 { , } 4 0 0 \\mathrm { H z }$ and $2 2 { , } 0 5 0 \\mathrm { H z }$ respectively, then compare two groups of onset beats under the same time scale. Since $2 2 , 0 5 0 \\mathrm { H z }$ is a common sampling rate used in music information retrieval, we define the beat alignment ratio as $B _ { 2 } / B _ { 1 }$ , where $B _ { 1 }$ is the number of beats under $1 5 { , } 4 0 0 \\mathrm { H z }$ and $B _ { 2 }$ is the number of beats under 15,400Hz that are aligned with the beats under $2 2 { , } 0 5 0 \\mathrm { H z }$ . One beat is counted when $\\mid t _ { 1 } - t _ { 2 } \\mid \\leq \\Delta t . \\ t _ { 1 }$ denotes the timestamp of a certain beat under $1 5 { , } 4 0 0 \\mathrm { H z }$ while $t _ { 2 }$ refers to the timestamp of a certain beat under $2 2 { , } 0 5 0 \\mathrm { H z }$ , $\\Delta t$ is the time offset threshold. ", + "bbox": [ + 174, + 688, + 825, + 800 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In the experiment, we set $\\Delta t = 1 / 1 5 s$ (FPS is 15) and calculate the beat alignment ratio for audio data of three styles respectively. We randomly sample 10 audio clips from each style to calculate the beat alignment ratio and do the sampling for 10 times to take the average. As shown in Table 5, most of beats extracted with $1 5 { , } 4 0 0 \\mathrm { H z }$ are aligned with those extracted with $2 2 { , } 0 5 0 \\mathrm { H z }$ . Besides, we randomly sample an audio clip and visualize beat tracking curves of its first 20 seconds under the sampling rates of $1 5 { , } 4 0 0 \\mathrm { H z }$ and $2 2 { , } 0 5 0 \\mathrm { H z }$ . As we can see in Figure 5, most of the beats from $1 5 { , } 4 0 0 \\mathrm { H z }$ and $2 2 { , } 0 5 0 \\mathrm { H z }$ are aligned within the time offset threshold $\\Delta t = 1 / 1 5 s$ . ", + "bbox": [ + 173, + 806, + 825, + 904 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/c25230cbddf59843dade0c1962c09ff1911fd40378025f0a07c73d06bf2336f2.jpg", + "table_caption": [ + "Table 5: The beat alignment ratio $B _ { 2 } / B _ { 1 }$ . " + ], + "table_footnote": [], + "table_body": "
CategoryBeat Alignment Ratio (%)
Ballet88.7
Hiphop93.2
Japanese Pop91.6
", + "bbox": [ + 352, + 133, + 647, + 207 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/960e4e9bd2cd632b5f494cde13e98e1b266fa0bc90a5990a3085ee12ef8de7cc.jpg", + "image_caption": [ + "Figure 5: Beat tracking curves for the first 20 seconds of a music clip randomly sampled from audio data under the sampling rates of $1 5 { , } 4 0 0 \\mathrm { H z }$ and $2 2 { , } 0 5 0 \\mathrm { H z }$ . Most of the beats are aligned within the time offset threshold $\\Delta t = 1 / 1 5 s$ . " + ], + "image_footnote": [], + "bbox": [ + 174, + 232, + 820, + 479 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/parse/train/xGZG2kS5bFk/xGZG2kS5bFk_middle.json b/parse/train/xGZG2kS5bFk/xGZG2kS5bFk_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..0f487a7d0ea327ea8a65683c6c9202b53a806b2a --- /dev/null +++ b/parse/train/xGZG2kS5bFk/xGZG2kS5bFk_middle.json @@ -0,0 +1,36833 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 506, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 507, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 507, + 97 + ], + "score": 1.0, + "content": "DANCE REVOLUTION: LONG-TERM DANCE GENERA-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 466, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 466, + 118 + ], + "score": 1.0, + "content": "TION WITH MUSIC VIA CURRICULUM LEARNING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 133, + 487, + 180 + ], + "lines": [ + { + "bbox": [ + 110, + 132, + 489, + 148 + ], + "spans": [ + { + "bbox": [ + 110, + 132, + 221, + 148 + ], + "score": 1.0, + "content": "Ruozi Huang1∗ † Huang", + "type": "text" + }, + { + "bbox": [ + 222, + 134, + 246, + 146 + ], + "score": 0.85, + "content": "\\mathbf { H } \\mathbf { u } ^ { 2 * }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 132, + 273, + 148 + ], + "score": 1.0, + "content": "Wei", + "type": "text" + }, + { + "bbox": [ + 273, + 134, + 295, + 146 + ], + "score": 0.66, + "content": "{ \\bf W } { \\bf u } ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 132, + 489, + 148 + ], + "score": 1.0, + "content": "Kei Sawada4 Mi Zhang1 Daxin Jiang2‡", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 145, + 369, + 158 + ], + "spans": [ + { + "bbox": [ + 110, + 145, + 369, + 158 + ], + "score": 1.0, + "content": "Fudan University1 Microsoft STCA2 Meituan3 Rinna AI4", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 156, + 484, + 171 + ], + "spans": [ + { + "bbox": [ + 111, + 156, + 484, + 171 + ], + "score": 1.0, + "content": "{rzhuang17,mi_zhang}@fudan.edu.cn, {huahu,djiang}@microsoft.com,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 168, + 383, + 181 + ], + "spans": [ + { + "bbox": [ + 111, + 168, + 383, + 181 + ], + "score": 1.0, + "content": "wuwei19850318@gmail.com, keisawada@rinna.co.jp", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 278, + 209, + 333, + 221 + ], + "lines": [ + { + "bbox": [ + 277, + 208, + 335, + 222 + ], + "spans": [ + { + "bbox": [ + 277, + 208, + 335, + 222 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 143, + 233, + 469, + 453 + ], + "lines": [ + { + "bbox": [ + 141, + 234, + 470, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 470, + 246 + ], + "score": 1.0, + "content": "Dancing to music is one of human’s innate abilities since ancient times. In ma-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 245, + 470, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 470, + 258 + ], + "score": 1.0, + "content": "chine learning research, however, synthesizing dance movements from music is a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 255, + 470, + 269 + ], + "spans": [ + { + "bbox": [ + 141, + 255, + 470, + 269 + ], + "score": 1.0, + "content": "challenging problem. Recently, researchers synthesize human motion sequences", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 266, + 470, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 470, + 280 + ], + "score": 1.0, + "content": "through autoregressive models like recurrent neural network (RNN). Such an ap-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 278, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 470, + 290 + ], + "score": 1.0, + "content": "proach often generates short sequences due to an accumulation of prediction errors", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 469, + 300 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 300 + ], + "score": 1.0, + "content": "that are fed back into the neural network. This problem becomes even more severe", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 299, + 469, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 469, + 312 + ], + "score": 1.0, + "content": "in the long motion sequence generation. Besides, the consistency between dance", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 310, + 469, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 310, + 469, + 323 + ], + "score": 1.0, + "content": "and music in terms of style, rhythm and beat is yet to be taken into account during", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 322, + 470, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 470, + 334 + ], + "score": 1.0, + "content": "modeling. In this paper, we formalize the music-conditioned dance generation as a", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 333, + 469, + 344 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 469, + 344 + ], + "score": 1.0, + "content": "sequence-to-sequence learning problem and devise a novel seq2seq architecture to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 343, + 470, + 356 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 470, + 356 + ], + "score": 1.0, + "content": "efficiently process long sequences of music features and capture the fine-grained", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 354, + 470, + 367 + ], + "spans": [ + { + "bbox": [ + 141, + 354, + 470, + 367 + ], + "score": 1.0, + "content": "correspondence between music and dance. Furthermore, we propose a novel cur-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 366, + 470, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 470, + 379 + ], + "score": 1.0, + "content": "riculum learning strategy to alleviate error accumulation of autoregressive models", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 376, + 470, + 389 + ], + "spans": [ + { + "bbox": [ + 141, + 376, + 470, + 389 + ], + "score": 1.0, + "content": "in long motion sequence generation, which gently changes the training process from", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 387, + 471, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 387, + 471, + 401 + ], + "score": 1.0, + "content": "a fully guided teacher-forcing scheme using the previous ground-truth movements,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 399, + 470, + 411 + ], + "spans": [ + { + "bbox": [ + 141, + 399, + 470, + 411 + ], + "score": 1.0, + "content": "towards a less guided autoregressive scheme mostly using the generated movements", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 410, + 469, + 421 + ], + "spans": [ + { + "bbox": [ + 141, + 410, + 469, + 421 + ], + "score": 1.0, + "content": "instead. Extensive experiments show that our approach significantly outperforms", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 142, + 421, + 469, + 432 + ], + "spans": [ + { + "bbox": [ + 142, + 421, + 469, + 432 + ], + "score": 1.0, + "content": "the existing state-of-the-arts on automatic metrics and human evaluation. We also", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 431, + 470, + 444 + ], + "spans": [ + { + "bbox": [ + 141, + 431, + 470, + 444 + ], + "score": 1.0, + "content": "make a demo video to demonstrate the superior performance of our proposed", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 443, + 454, + 454 + ], + "spans": [ + { + "bbox": [ + 141, + 443, + 376, + 454 + ], + "score": 1.0, + "content": "approach at https://www.youtube.com/watch?v", + "type": "text" + }, + { + "bbox": [ + 377, + 444, + 384, + 451 + ], + "score": 0.43, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 443, + 454, + 454 + ], + "score": 1.0, + "content": "lmE20MEheZ8.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 108, + 474, + 206, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 208, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 208, + 490 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 499, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 506, + 513 + ], + "score": 1.0, + "content": "Arguably, dancing to music is one of human’s innate abilities, as we can spontaneously sway along", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 511, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 506, + 523 + ], + "score": 1.0, + "content": "with the tempo of music we hear. The research in neuropsychology indicates that our brain is", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 521, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 506, + 534 + ], + "score": 1.0, + "content": "hardwired to make us move and synchronize with music regardless of our intention (Chen et al.,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 531, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 506, + 546 + ], + "score": 1.0, + "content": "2008). Another study of archaeology also suggests that dance is a social communication skill among", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 543, + 506, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 506, + 555 + ], + "score": 1.0, + "content": "early humans connected to the ability of survival long time ago (Adshead-Lansdale & Layson, 2006).", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 554, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 506, + 567 + ], + "score": 1.0, + "content": "Nowadays, dance (to music) has become a means to the cultural promotion, a method of emotional", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 565, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 506, + 578 + ], + "score": 1.0, + "content": "expression, a tool for socialization and an art form to bring aesthetic enjoyment. The neurological", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "mechanism behind dancing behavior and the unique value of dance to society motivate us to explore", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 587, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 600 + ], + "score": 1.0, + "content": "a computational approach to dance creation from a piece of music in artificial intelligence research.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 597, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 505, + 612 + ], + "score": 1.0, + "content": "Such work is potentially beneficial to a wide range of applications such as dance creation assistant in", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 609, + 501, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 501, + 622 + ], + "score": 1.0, + "content": "art and sports, character motion generation for audio games and research on cross-modal behavior.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 626, + 505, + 692 + ], + "lines": [ + { + "bbox": [ + 105, + 625, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 505, + 640 + ], + "score": 1.0, + "content": "In literature, the music-conditioned dance generation is a relatively new task that attracts increasing", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 636, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 506, + 651 + ], + "score": 1.0, + "content": "research interests recently. Early works (Fan et al., 2011; Lee et al., 2013) synthesize dance sequences", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "from music by retrieval-based methods, which show the limited creativity in practice. Recently, Lee", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "score": 1.0, + "content": "et al. (2019) formulate the task from the generative perspective, and further propose a decomposition-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 671, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 505, + 682 + ], + "score": 1.0, + "content": "to-composition framework. Their model first generates basic dance units from the music clips and", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 681, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 505, + 693 + ], + "score": 1.0, + "content": "then composes them by using the last pose of current unit to initialize the first pose of the next", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 700, + 313, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 699, + 194, + 710 + ], + "spans": [ + { + "bbox": [ + 118, + 699, + 194, + 710 + ], + "score": 1.0, + "content": "∗Equal contribution.", + "type": "text" + } + ] + }, + { + "bbox": [ + 117, + 708, + 315, + 723 + ], + "spans": [ + { + "bbox": [ + 117, + 708, + 315, + 723 + ], + "score": 1.0, + "content": "†Work done during the internship at Microsoft STCA.", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 719, + 205, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 205, + 734 + ], + "score": 1.0, + "content": "‡Corresponding author.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 506, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 507, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 507, + 97 + ], + "score": 1.0, + "content": "DANCE REVOLUTION: LONG-TERM DANCE GENERA-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 466, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 466, + 118 + ], + "score": 1.0, + "content": "TION WITH MUSIC VIA CURRICULUM LEARNING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "list", + "bbox": [ + 112, + 133, + 487, + 180 + ], + "lines": [ + { + "bbox": [ + 110, + 132, + 489, + 148 + ], + "spans": [ + { + "bbox": [ + 110, + 132, + 221, + 148 + ], + "score": 1.0, + "content": "Ruozi Huang1∗ † Huang", + "type": "text" + }, + { + "bbox": [ + 222, + 134, + 246, + 146 + ], + "score": 0.85, + "content": "\\mathbf { H } \\mathbf { u } ^ { 2 * }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 132, + 273, + 148 + ], + "score": 1.0, + "content": "Wei", + "type": "text" + }, + { + "bbox": [ + 273, + 134, + 295, + 146 + ], + "score": 0.66, + "content": "{ \\bf W } { \\bf u } ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 132, + 489, + 148 + ], + "score": 1.0, + "content": "Kei Sawada4 Mi Zhang1 Daxin Jiang2‡", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 145, + 369, + 158 + ], + "spans": [ + { + "bbox": [ + 110, + 145, + 369, + 158 + ], + "score": 1.0, + "content": "Fudan University1 Microsoft STCA2 Meituan3 Rinna AI4", + "type": "text" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 156, + 484, + 171 + ], + "spans": [ + { + "bbox": [ + 111, + 156, + 484, + 171 + ], + "score": 1.0, + "content": "{rzhuang17,mi_zhang}@fudan.edu.cn, {huahu,djiang}@microsoft.com,", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 168, + 383, + 181 + ], + "spans": [ + { + "bbox": [ + 111, + 168, + 383, + 181 + ], + "score": 1.0, + "content": "wuwei19850318@gmail.com, keisawada@rinna.co.jp", + "type": "text" + } + ], + "index": 5, + "is_list_end_line": true + } + ], + "index": 3.5, + "bbox_fs": [ + 110, + 132, + 489, + 181 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 209, + 333, + 221 + ], + "lines": [ + { + "bbox": [ + 277, + 208, + 335, + 222 + ], + "spans": [ + { + "bbox": [ + 277, + 208, + 335, + 222 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 143, + 233, + 469, + 453 + ], + "lines": [ + { + "bbox": [ + 141, + 234, + 470, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 470, + 246 + ], + "score": 1.0, + "content": "Dancing to music is one of human’s innate abilities since ancient times. In ma-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 245, + 470, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 470, + 258 + ], + "score": 1.0, + "content": "chine learning research, however, synthesizing dance movements from music is a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 255, + 470, + 269 + ], + "spans": [ + { + "bbox": [ + 141, + 255, + 470, + 269 + ], + "score": 1.0, + "content": "challenging problem. Recently, researchers synthesize human motion sequences", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 266, + 470, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 470, + 280 + ], + "score": 1.0, + "content": "through autoregressive models like recurrent neural network (RNN). Such an ap-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 278, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 470, + 290 + ], + "score": 1.0, + "content": "proach often generates short sequences due to an accumulation of prediction errors", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 469, + 300 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 300 + ], + "score": 1.0, + "content": "that are fed back into the neural network. This problem becomes even more severe", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 299, + 469, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 469, + 312 + ], + "score": 1.0, + "content": "in the long motion sequence generation. Besides, the consistency between dance", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 310, + 469, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 310, + 469, + 323 + ], + "score": 1.0, + "content": "and music in terms of style, rhythm and beat is yet to be taken into account during", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 322, + 470, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 470, + 334 + ], + "score": 1.0, + "content": "modeling. In this paper, we formalize the music-conditioned dance generation as a", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 333, + 469, + 344 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 469, + 344 + ], + "score": 1.0, + "content": "sequence-to-sequence learning problem and devise a novel seq2seq architecture to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 343, + 470, + 356 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 470, + 356 + ], + "score": 1.0, + "content": "efficiently process long sequences of music features and capture the fine-grained", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 354, + 470, + 367 + ], + "spans": [ + { + "bbox": [ + 141, + 354, + 470, + 367 + ], + "score": 1.0, + "content": "correspondence between music and dance. Furthermore, we propose a novel cur-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 366, + 470, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 470, + 379 + ], + "score": 1.0, + "content": "riculum learning strategy to alleviate error accumulation of autoregressive models", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 376, + 470, + 389 + ], + "spans": [ + { + "bbox": [ + 141, + 376, + 470, + 389 + ], + "score": 1.0, + "content": "in long motion sequence generation, which gently changes the training process from", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 387, + 471, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 387, + 471, + 401 + ], + "score": 1.0, + "content": "a fully guided teacher-forcing scheme using the previous ground-truth movements,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 399, + 470, + 411 + ], + "spans": [ + { + "bbox": [ + 141, + 399, + 470, + 411 + ], + "score": 1.0, + "content": "towards a less guided autoregressive scheme mostly using the generated movements", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 410, + 469, + 421 + ], + "spans": [ + { + "bbox": [ + 141, + 410, + 469, + 421 + ], + "score": 1.0, + "content": "instead. Extensive experiments show that our approach significantly outperforms", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 142, + 421, + 469, + 432 + ], + "spans": [ + { + "bbox": [ + 142, + 421, + 469, + 432 + ], + "score": 1.0, + "content": "the existing state-of-the-arts on automatic metrics and human evaluation. We also", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 431, + 470, + 444 + ], + "spans": [ + { + "bbox": [ + 141, + 431, + 470, + 444 + ], + "score": 1.0, + "content": "make a demo video to demonstrate the superior performance of our proposed", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 443, + 454, + 454 + ], + "spans": [ + { + "bbox": [ + 141, + 443, + 376, + 454 + ], + "score": 1.0, + "content": "approach at https://www.youtube.com/watch?v", + "type": "text" + }, + { + "bbox": [ + 377, + 444, + 384, + 451 + ], + "score": 0.43, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 443, + 454, + 454 + ], + "score": 1.0, + "content": "lmE20MEheZ8.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 16.5, + "bbox_fs": [ + 141, + 234, + 471, + 454 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 474, + 206, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 208, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 208, + 490 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 499, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 506, + 513 + ], + "score": 1.0, + "content": "Arguably, dancing to music is one of human’s innate abilities, as we can spontaneously sway along", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 511, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 506, + 523 + ], + "score": 1.0, + "content": "with the tempo of music we hear. The research in neuropsychology indicates that our brain is", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 521, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 506, + 534 + ], + "score": 1.0, + "content": "hardwired to make us move and synchronize with music regardless of our intention (Chen et al.,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 531, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 506, + 546 + ], + "score": 1.0, + "content": "2008). Another study of archaeology also suggests that dance is a social communication skill among", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 543, + 506, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 506, + 555 + ], + "score": 1.0, + "content": "early humans connected to the ability of survival long time ago (Adshead-Lansdale & Layson, 2006).", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 554, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 506, + 567 + ], + "score": 1.0, + "content": "Nowadays, dance (to music) has become a means to the cultural promotion, a method of emotional", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 565, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 506, + 578 + ], + "score": 1.0, + "content": "expression, a tool for socialization and an art form to bring aesthetic enjoyment. The neurological", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "mechanism behind dancing behavior and the unique value of dance to society motivate us to explore", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 587, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 600 + ], + "score": 1.0, + "content": "a computational approach to dance creation from a piece of music in artificial intelligence research.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 597, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 505, + 612 + ], + "score": 1.0, + "content": "Such work is potentially beneficial to a wide range of applications such as dance creation assistant in", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 609, + 501, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 501, + 622 + ], + "score": 1.0, + "content": "art and sports, character motion generation for audio games and research on cross-modal behavior.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 499, + 506, + 622 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 626, + 505, + 692 + ], + "lines": [ + { + "bbox": [ + 105, + 625, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 505, + 640 + ], + "score": 1.0, + "content": "In literature, the music-conditioned dance generation is a relatively new task that attracts increasing", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 636, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 506, + 651 + ], + "score": 1.0, + "content": "research interests recently. Early works (Fan et al., 2011; Lee et al., 2013) synthesize dance sequences", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "from music by retrieval-based methods, which show the limited creativity in practice. Recently, Lee", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "score": 1.0, + "content": "et al. (2019) formulate the task from the generative perspective, and further propose a decomposition-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 671, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 505, + 682 + ], + "score": 1.0, + "content": "to-composition framework. Their model first generates basic dance units from the music clips and", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 681, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 505, + 693 + ], + "score": 1.0, + "content": "then composes them by using the last pose of current unit to initialize the first pose of the next", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "unit. Although this approach shows better performance than the retrieval-based methods, several", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 201, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 201, + 106 + ], + "score": 1.0, + "content": "challenges still remain.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 625, + 506, + 693 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "unit. Although this approach shows better performance than the retrieval-based methods, several", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 201, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 201, + 106 + ], + "score": 1.0, + "content": "challenges still remain.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "First, existing generative methods synthesize new human motion sequences through autoregressive", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "models like RNN, which tend to result in short sequences. In other words, the generated sequences", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "often quickly freeze within a few seconds due to an accumulation of prediction errors that are fed", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "back into the neural network. This problem becomes even more severe in long motion sequence", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 155, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 166 + ], + "score": 1.0, + "content": "generation. For instance, composing a dance for a 1-minute music clip under 15 Frame Per Second", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "score": 1.0, + "content": "(FPS) means generating 900 poses at one time. In practice, we need novel methods which can", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "effectively generate long motion sequences. Besides, how to enhance the harmony between the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "synthesized dance movements and the given music is a largely unexplored challenge. Inutitively, the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "movements need to be consistent with the music in terms of style, rhythm and beat. However, to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "achieve this goal is non-trivial, which requires the generation model to have the capability to capture", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 344, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 344, + 232 + ], + "score": 1.0, + "content": "the fine-grained correspondence between music and dance.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 237, + 505, + 412 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 251 + ], + "score": 1.0, + "content": "In this paper, we formalize music-conditioned dance generation as a sequence-to-sequence learning", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "problem where the fine-grained correspondence between music and dance is represented through", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "sequence modeling and their alignment is established via mapping from a sequence of acoustic", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "score": 1.0, + "content": "features of the music to a sequence of movements of the dance. The model consists of a music", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "encoder and a dance decoder. The encoder transforms low-level acoustic features of an input music", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 292, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 505, + 304 + ], + "score": 1.0, + "content": "clip into high-level latent representations via self-attention with a receptive field restricted within", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 303, + 504, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 113, + 313 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 304, + 504, + 315 + ], + "score": 1.0, + "content": "-nearest neighbors of an element. Thus, the encoder can efficiently process long sequences of music", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 314, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 505, + 326 + ], + "score": 1.0, + "content": "features, e.g., a sequence with more than 1000 elements, and model local characteristics of the music", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "such as chord and rhythm patterns. The decoder exploits a recurrent structure to predict the dance", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "movement frame by frame conditioned on the corresponding element in the latent representations of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 346, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 505, + 360 + ], + "score": 1.0, + "content": "music feature sequence. Furthermore, we propose a curriculum learning (Bengio et al., 2009) strategy", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 356, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 371 + ], + "score": 1.0, + "content": "to alleviate error accumulation (Li et al., 2017) of autoregressive models in long motion sequence", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 367, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 383 + ], + "score": 1.0, + "content": "generation. Specifically, it gently changes the training process from a fully guided teacher-forcing", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 379, + 504, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 504, + 392 + ], + "score": 1.0, + "content": "scheme using the previous ground-truth movements, towards a less guided autoregressive scheme", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "which mostly utilizes the generated movements instead. This strategy bridges the gap between", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 402, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 414 + ], + "score": 1.0, + "content": "training and inference of autoregressive models, and thus alleviates error accumulation at inference.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 418, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 106, + 418, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 506, + 430 + ], + "score": 1.0, + "content": "The length of each video clip in the dataset released by Lee et al. (2019) is only about 6 seconds,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "which cannot be used by other methds to generate long-term dance except for their specially designed", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 441, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 452 + ], + "score": 1.0, + "content": "decomposition-to-composition framework. To facilitate the task of long-term dance generation with", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 451, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 506, + 464 + ], + "score": 1.0, + "content": "music, we collect a high-quality dataset consisting of 1-minute video clips, totaling about 12 hours.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 461, + 490, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 490, + 475 + ], + "score": 1.0, + "content": "And there are three representative styles in our dataset: “Ballet”, “Hiphop” and “Japanese Pop”.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "Our contributions in this work are four-fold: (1) We formalize music-conditioned dance generation as", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "a sequence-to-sequence learning problem and devise a novel seq2seq architecture for the long-term", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "score": 1.0, + "content": "dance generation with music. (2) We propose a novel curriculum learning strategy to alleviate", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 513, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 524 + ], + "score": 1.0, + "content": "error accumulation of autoregressive models in long motion sequence generation. (3) To facilitate", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 522, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 536 + ], + "score": 1.0, + "content": "long-term dance generation with music, we collect a high-quality dataset that is available with", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "our code1. (4) The extensive experiments show that our approach significantly outperforms the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 545, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 557 + ], + "score": 1.0, + "content": "existing state-of-the-arts on both automatic metrics and human judgements. The demo video in the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "supplementary material also exhibits our approach can generate diverse minute-length dances that are", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 567, + 467, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 467, + 579 + ], + "score": 1.0, + "content": "smooth, natural-looking, style-consistent and beat-matching with the musics from test set.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38 + }, + { + "type": "title", + "bbox": [ + 108, + 602, + 211, + 615 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 213, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 213, + 617 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 632, + 505, + 698 + ], + "lines": [ + { + "bbox": [ + 107, + 632, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 107, + 632, + 505, + 644 + ], + "score": 1.0, + "content": "Cross-Modal Learning. Most existing works focus on the modeling between vision and text, such", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 644, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 655 + ], + "score": 1.0, + "content": "as image captioning (Lu et al., 2017; Xu et al., 2015) and text-to-image generation (Reed et al., 2016;", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "score": 1.0, + "content": "Zhang et al., 2017). There are some other works to study the translation between audio and text like", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "Automatic Speech Recognition (ASR) (Hinton et al., 2012) and Text-To-Speech (TTS) (Oord et al.,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 676, + 506, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 506, + 688 + ], + "score": 1.0, + "content": "2016). While the modeling between audio and vision is largely unexplored and the music-conditioned", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 687, + 430, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 430, + 700 + ], + "score": 1.0, + "content": "dance generation is a typical cross-modal learning problem from audio to vision.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 722, + 350, + 731 + ], + "lines": [ + { + "bbox": [ + 119, + 721, + 351, + 733 + ], + "spans": [ + { + "bbox": [ + 119, + 721, + 351, + 733 + ], + "score": 1.0, + "content": "1https://github.com/stonyhu/DanceRevolution", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 13, + "width": 8 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "First, existing generative methods synthesize new human motion sequences through autoregressive", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "models like RNN, which tend to result in short sequences. In other words, the generated sequences", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "often quickly freeze within a few seconds due to an accumulation of prediction errors that are fed", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "back into the neural network. This problem becomes even more severe in long motion sequence", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 155, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 166 + ], + "score": 1.0, + "content": "generation. For instance, composing a dance for a 1-minute music clip under 15 Frame Per Second", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "score": 1.0, + "content": "(FPS) means generating 900 poses at one time. In practice, we need novel methods which can", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "effectively generate long motion sequences. Besides, how to enhance the harmony between the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "synthesized dance movements and the given music is a largely unexplored challenge. Inutitively, the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "movements need to be consistent with the music in terms of style, rhythm and beat. However, to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "achieve this goal is non-trivial, which requires the generation model to have the capability to capture", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 344, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 344, + 232 + ], + "score": 1.0, + "content": "the fine-grained correspondence between music and dance.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 111, + 506, + 232 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 237, + 505, + 412 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 251 + ], + "score": 1.0, + "content": "In this paper, we formalize music-conditioned dance generation as a sequence-to-sequence learning", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "problem where the fine-grained correspondence between music and dance is represented through", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "sequence modeling and their alignment is established via mapping from a sequence of acoustic", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "score": 1.0, + "content": "features of the music to a sequence of movements of the dance. The model consists of a music", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "encoder and a dance decoder. The encoder transforms low-level acoustic features of an input music", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 292, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 505, + 304 + ], + "score": 1.0, + "content": "clip into high-level latent representations via self-attention with a receptive field restricted within", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 303, + 504, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 113, + 313 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 304, + 504, + 315 + ], + "score": 1.0, + "content": "-nearest neighbors of an element. Thus, the encoder can efficiently process long sequences of music", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 314, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 505, + 326 + ], + "score": 1.0, + "content": "features, e.g., a sequence with more than 1000 elements, and model local characteristics of the music", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "such as chord and rhythm patterns. The decoder exploits a recurrent structure to predict the dance", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "movement frame by frame conditioned on the corresponding element in the latent representations of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 346, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 505, + 360 + ], + "score": 1.0, + "content": "music feature sequence. Furthermore, we propose a curriculum learning (Bengio et al., 2009) strategy", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 356, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 371 + ], + "score": 1.0, + "content": "to alleviate error accumulation (Li et al., 2017) of autoregressive models in long motion sequence", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 367, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 383 + ], + "score": 1.0, + "content": "generation. Specifically, it gently changes the training process from a fully guided teacher-forcing", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 379, + 504, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 504, + 392 + ], + "score": 1.0, + "content": "scheme using the previous ground-truth movements, towards a less guided autoregressive scheme", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "which mostly utilizes the generated movements instead. This strategy bridges the gap between", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 402, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 414 + ], + "score": 1.0, + "content": "training and inference of autoregressive models, and thus alleviates error accumulation at inference.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 236, + 506, + 414 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 418, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 106, + 418, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 506, + 430 + ], + "score": 1.0, + "content": "The length of each video clip in the dataset released by Lee et al. (2019) is only about 6 seconds,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "which cannot be used by other methds to generate long-term dance except for their specially designed", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 441, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 452 + ], + "score": 1.0, + "content": "decomposition-to-composition framework. To facilitate the task of long-term dance generation with", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 451, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 506, + 464 + ], + "score": 1.0, + "content": "music, we collect a high-quality dataset consisting of 1-minute video clips, totaling about 12 hours.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 461, + 490, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 490, + 475 + ], + "score": 1.0, + "content": "And there are three representative styles in our dataset: “Ballet”, “Hiphop” and “Japanese Pop”.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 418, + 506, + 475 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "Our contributions in this work are four-fold: (1) We formalize music-conditioned dance generation as", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "a sequence-to-sequence learning problem and devise a novel seq2seq architecture for the long-term", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 514 + ], + "score": 1.0, + "content": "dance generation with music. (2) We propose a novel curriculum learning strategy to alleviate", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 513, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 524 + ], + "score": 1.0, + "content": "error accumulation of autoregressive models in long motion sequence generation. (3) To facilitate", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 522, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 536 + ], + "score": 1.0, + "content": "long-term dance generation with music, we collect a high-quality dataset that is available with", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "our code1. (4) The extensive experiments show that our approach significantly outperforms the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 545, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 557 + ], + "score": 1.0, + "content": "existing state-of-the-arts on both automatic metrics and human judgements. The demo video in the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "supplementary material also exhibits our approach can generate diverse minute-length dances that are", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 567, + 467, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 467, + 579 + ], + "score": 1.0, + "content": "smooth, natural-looking, style-consistent and beat-matching with the musics from test set.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 479, + 506, + 579 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 602, + 211, + 615 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 213, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 213, + 617 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 632, + 505, + 698 + ], + "lines": [ + { + "bbox": [ + 107, + 632, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 107, + 632, + 505, + 644 + ], + "score": 1.0, + "content": "Cross-Modal Learning. Most existing works focus on the modeling between vision and text, such", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 644, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 655 + ], + "score": 1.0, + "content": "as image captioning (Lu et al., 2017; Xu et al., 2015) and text-to-image generation (Reed et al., 2016;", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "score": 1.0, + "content": "Zhang et al., 2017). There are some other works to study the translation between audio and text like", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "Automatic Speech Recognition (ASR) (Hinton et al., 2012) and Text-To-Speech (TTS) (Oord et al.,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 676, + 506, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 506, + 688 + ], + "score": 1.0, + "content": "2016). While the modeling between audio and vision is largely unexplored and the music-conditioned", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 687, + 430, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 430, + 700 + ], + "score": 1.0, + "content": "dance generation is a typical cross-modal learning problem from audio to vision.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 632, + 506, + 700 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Human Motion Prediction. Prediction of human motion dynamics has been a challenging problem", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 104, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "in computer vision, which suffers from the high spatial-temporal complexity. Existing works (Chan", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 506, + 117 + ], + "score": 1.0, + "content": "et al., 2019; Wang et al., 2018) represent the human pose as 2D or 3D body keyjoints (Cao et al.,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "2017) and address the problem via sequence modeling. Early methods, such as hidden markov", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 139 + ], + "score": 1.0, + "content": "models (Lehrmann et al., 2014), Gaussian processes (Wang et al., 2006) and restricted boltzmann", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "machines (Taylor et al., 2007), have to balance the model capacity and inference complexity due to", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 507, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 507, + 162 + ], + "score": 1.0, + "content": "complicated training procedures. Recently, neural networks dominate the human motion modeling.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 171 + ], + "score": 1.0, + "content": "For instance, Fragkiadaki et al. (2015) present LSTM-3LR and Encoder-Recurrent-Decoder (ERD)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "as two recurrent architectures for the task; Jain et al. (2016) propose a structural-RNN to model", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "human-object interactions in a spatio-temporal graph; and Ghosh et al. (2017) equip LSTM-3LR with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 204 + ], + "score": 1.0, + "content": "a dropout autoencoder to enhance the long-term prediction. Besides, convolutional neural networks", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 203, + 455, + 215 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 455, + 215 + ], + "score": 1.0, + "content": "(CNNs) have also been utilized to model the human motion prediction (Li et al., 2018).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 227, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 227, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 506, + 240 + ], + "score": 1.0, + "content": "Audio-Conditioned Dance Generation. In the research of audio-conditioned dance generation,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 239, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 506, + 251 + ], + "score": 1.0, + "content": "most existing works study 2D dance motion generation with music since the training data for paired", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 249, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 505, + 262 + ], + "score": 1.0, + "content": "2D pose and music can be extracted from the huge amount of dance videos available online. Various", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "methods have been proposed to handle this task, such as adversarial learning based methods (Lee", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 271, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 506, + 284 + ], + "score": 1.0, + "content": "et al., 2019; Sun et al., 2020; Ferreira et al., 2021), autoencoder methods (Tang et al., 2018) and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 282, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 506, + 295 + ], + "score": 1.0, + "content": "sequence-to-sequence methods (Lee et al., 2018; Ren et al., 2019; Yalta et al., 2019; Ye et al., 2020).", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "While these works mainly focus on exploring the different neural architectures and overlook the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 303, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 318 + ], + "score": 1.0, + "content": "freezing motion issue in dance motion synthesis. In this work, we first propose a novel seq2seq", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 316, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 506, + 328 + ], + "score": 1.0, + "content": "architecture to model the fine-grained correspondence between music and dance, and then introduce a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 327, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 505, + 339 + ], + "score": 1.0, + "content": "novel curriculum learning strategy to address the freezing motion issue caused by error accumulation", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 337, + 326, + 350 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 326, + 350 + ], + "score": 1.0, + "content": "(Li et al., 2017) in long-term dance motion generation.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 107, + 365, + 182, + 378 + ], + "lines": [ + { + "bbox": [ + 104, + 363, + 185, + 381 + ], + "spans": [ + { + "bbox": [ + 104, + 363, + 185, + 381 + ], + "score": 1.0, + "content": "3 APPROACH", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 391, + 505, + 424 + ], + "lines": [ + { + "bbox": [ + 105, + 390, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 404 + ], + "score": 1.0, + "content": "In this section, we present our approach to music-conditioned dance generation. After formalization", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 402, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 415 + ], + "score": 1.0, + "content": "of the problem in question, we elaborate the model architecture and the dynamic auto-condition", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 413, + 470, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 470, + 426 + ], + "score": 1.0, + "content": "learning approach that facilitates long-term dance generation according to the given music.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 124, + 436, + 483, + 562 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 436, + 483, + 562 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 436, + 483, + 562 + ], + "spans": [ + { + "bbox": [ + 124, + 436, + 483, + 562 + ], + "score": 0.973, + "type": "image", + "image_path": "f2bd9934103d7f435c8ce5dfd137026280301ab5097a7eec93c1dca3017d0a32.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 124, + 436, + 483, + 478.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 124, + 478.0, + 483, + 520.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 124, + 520.0, + 483, + 562.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 573, + 505, + 662 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 432, + 586 + ], + "score": 1.0, + "content": "Figure 1: The overview of our method. The transformer based encoder is stacked of", + "type": "text" + }, + { + "bbox": [ + 432, + 574, + 443, + 583 + ], + "score": 0.77, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "identical layers", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 585, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 596 + ], + "score": 1.0, + "content": "with the local self-attention mechanism, which aims to efficiently process long features sequences", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 595, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 607 + ], + "score": 1.0, + "content": "extracted from music in a GPU-memory-economic way. The decoder exploits a recurrent structure", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 217, + 619 + ], + "score": 1.0, + "content": "to predict dance movement", + "type": "text" + }, + { + "bbox": [ + 218, + 608, + 227, + 618 + ], + "score": 0.84, + "content": "y _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 606, + 406, + 619 + ], + "score": 1.0, + "content": "frame by frame conditioned on hidden state", + "type": "text" + }, + { + "bbox": [ + 406, + 606, + 417, + 617 + ], + "score": 0.88, + "content": "h _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "of decoder and latent", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 133, + 630 + ], + "score": 1.0, + "content": "vector", + "type": "text" + }, + { + "bbox": [ + 133, + 619, + 142, + 628 + ], + "score": 0.84, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 617, + 185, + 630 + ], + "score": 1.0, + "content": ". Note that", + "type": "text" + }, + { + "bbox": [ + 185, + 619, + 196, + 629 + ], + "score": 0.85, + "content": "y _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "is a predefined Begin Of Pose (BOP). During the training phase, the proposed", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 627, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 104, + 627, + 505, + 642 + ], + "score": 1.0, + "content": "curriculum learning strategy gently increases the number of steps of autoregressive scheme according", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 639, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 506, + 652 + ], + "score": 1.0, + "content": "to the number of training epochs. This strategy bridges the gap between training and inference of", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 651, + 405, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 405, + 662 + ], + "score": 1.0, + "content": "autoregressive models, and thus alleviates error accumulation at inference.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5 + } + ], + "index": 30.75 + }, + { + "type": "title", + "bbox": [ + 108, + 689, + 248, + 699 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 250, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 250, + 701 + ], + "score": 1.0, + "content": "3.1 PROBLEM FORMALIZATION", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 708, + 504, + 733 + ], + "lines": [ + { + "bbox": [ + 104, + 706, + 507, + 725 + ], + "spans": [ + { + "bbox": [ + 104, + 706, + 226, + 725 + ], + "score": 1.0, + "content": "Suppose that there is a dataset", + "type": "text" + }, + { + "bbox": [ + 226, + 709, + 304, + 722 + ], + "score": 0.91, + "content": "\\mathcal { D } = \\{ ( X _ { i } , Y _ { i } ) \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 706, + 334, + 725 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 334, + 710, + 390, + 722 + ], + "score": 0.93, + "content": "X = \\{ x _ { t } \\} _ { t = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 706, + 469, + 725 + ], + "score": 1.0, + "content": "is a music clip with", + "type": "text" + }, + { + "bbox": [ + 469, + 711, + 480, + 721 + ], + "score": 0.84, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 706, + 507, + 725 + ], + "score": 1.0, + "content": "being", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 718, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 104, + 718, + 267, + 734 + ], + "score": 1.0, + "content": "a vector of acoustic features at time-step", + "type": "text" + }, + { + "bbox": [ + 267, + 722, + 272, + 730 + ], + "score": 0.77, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 718, + 293, + 734 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 293, + 721, + 347, + 732 + ], + "score": 0.93, + "content": "\\bar { Y = } \\{ y _ { t } \\} _ { t = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 718, + 506, + 734 + ], + "score": 1.0, + "content": "is a sequence of dance movements with", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Human Motion Prediction. Prediction of human motion dynamics has been a challenging problem", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 104, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "in computer vision, which suffers from the high spatial-temporal complexity. Existing works (Chan", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 506, + 117 + ], + "score": 1.0, + "content": "et al., 2019; Wang et al., 2018) represent the human pose as 2D or 3D body keyjoints (Cao et al.,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "2017) and address the problem via sequence modeling. Early methods, such as hidden markov", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 139 + ], + "score": 1.0, + "content": "models (Lehrmann et al., 2014), Gaussian processes (Wang et al., 2006) and restricted boltzmann", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "machines (Taylor et al., 2007), have to balance the model capacity and inference complexity due to", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 507, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 507, + 162 + ], + "score": 1.0, + "content": "complicated training procedures. Recently, neural networks dominate the human motion modeling.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 171 + ], + "score": 1.0, + "content": "For instance, Fragkiadaki et al. (2015) present LSTM-3LR and Encoder-Recurrent-Decoder (ERD)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "as two recurrent architectures for the task; Jain et al. (2016) propose a structural-RNN to model", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "human-object interactions in a spatio-temporal graph; and Ghosh et al. (2017) equip LSTM-3LR with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 204 + ], + "score": 1.0, + "content": "a dropout autoencoder to enhance the long-term prediction. Besides, convolutional neural networks", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 203, + 455, + 215 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 455, + 215 + ], + "score": 1.0, + "content": "(CNNs) have also been utilized to model the human motion prediction (Li et al., 2018).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 5.5, + "bbox_fs": [ + 104, + 81, + 507, + 215 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 227, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 227, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 506, + 240 + ], + "score": 1.0, + "content": "Audio-Conditioned Dance Generation. In the research of audio-conditioned dance generation,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 239, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 506, + 251 + ], + "score": 1.0, + "content": "most existing works study 2D dance motion generation with music since the training data for paired", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 249, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 505, + 262 + ], + "score": 1.0, + "content": "2D pose and music can be extracted from the huge amount of dance videos available online. Various", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "methods have been proposed to handle this task, such as adversarial learning based methods (Lee", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 271, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 506, + 284 + ], + "score": 1.0, + "content": "et al., 2019; Sun et al., 2020; Ferreira et al., 2021), autoencoder methods (Tang et al., 2018) and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 282, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 506, + 295 + ], + "score": 1.0, + "content": "sequence-to-sequence methods (Lee et al., 2018; Ren et al., 2019; Yalta et al., 2019; Ye et al., 2020).", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "While these works mainly focus on exploring the different neural architectures and overlook the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 303, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 318 + ], + "score": 1.0, + "content": "freezing motion issue in dance motion synthesis. In this work, we first propose a novel seq2seq", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 316, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 506, + 328 + ], + "score": 1.0, + "content": "architecture to model the fine-grained correspondence between music and dance, and then introduce a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 327, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 505, + 339 + ], + "score": 1.0, + "content": "novel curriculum learning strategy to address the freezing motion issue caused by error accumulation", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 337, + 326, + 350 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 326, + 350 + ], + "score": 1.0, + "content": "(Li et al., 2017) in long-term dance motion generation.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 227, + 506, + 350 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 365, + 182, + 378 + ], + "lines": [ + { + "bbox": [ + 104, + 363, + 185, + 381 + ], + "spans": [ + { + "bbox": [ + 104, + 363, + 185, + 381 + ], + "score": 1.0, + "content": "3 APPROACH", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 391, + 505, + 424 + ], + "lines": [ + { + "bbox": [ + 105, + 390, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 404 + ], + "score": 1.0, + "content": "In this section, we present our approach to music-conditioned dance generation. After formalization", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 402, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 415 + ], + "score": 1.0, + "content": "of the problem in question, we elaborate the model architecture and the dynamic auto-condition", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 413, + 470, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 470, + 426 + ], + "score": 1.0, + "content": "learning approach that facilitates long-term dance generation according to the given music.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 390, + 505, + 426 + ] + }, + { + "type": "image", + "bbox": [ + 124, + 436, + 483, + 562 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 436, + 483, + 562 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 436, + 483, + 562 + ], + "spans": [ + { + "bbox": [ + 124, + 436, + 483, + 562 + ], + "score": 0.973, + "type": "image", + "image_path": "f2bd9934103d7f435c8ce5dfd137026280301ab5097a7eec93c1dca3017d0a32.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 124, + 436, + 483, + 478.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 124, + 478.0, + 483, + 520.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 124, + 520.0, + 483, + 562.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 573, + 505, + 662 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 432, + 586 + ], + "score": 1.0, + "content": "Figure 1: The overview of our method. The transformer based encoder is stacked of", + "type": "text" + }, + { + "bbox": [ + 432, + 574, + 443, + 583 + ], + "score": 0.77, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "identical layers", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 585, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 596 + ], + "score": 1.0, + "content": "with the local self-attention mechanism, which aims to efficiently process long features sequences", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 595, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 607 + ], + "score": 1.0, + "content": "extracted from music in a GPU-memory-economic way. The decoder exploits a recurrent structure", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 217, + 619 + ], + "score": 1.0, + "content": "to predict dance movement", + "type": "text" + }, + { + "bbox": [ + 218, + 608, + 227, + 618 + ], + "score": 0.84, + "content": "y _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 606, + 406, + 619 + ], + "score": 1.0, + "content": "frame by frame conditioned on hidden state", + "type": "text" + }, + { + "bbox": [ + 406, + 606, + 417, + 617 + ], + "score": 0.88, + "content": "h _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "of decoder and latent", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 133, + 630 + ], + "score": 1.0, + "content": "vector", + "type": "text" + }, + { + "bbox": [ + 133, + 619, + 142, + 628 + ], + "score": 0.84, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 617, + 185, + 630 + ], + "score": 1.0, + "content": ". Note that", + "type": "text" + }, + { + "bbox": [ + 185, + 619, + 196, + 629 + ], + "score": 0.85, + "content": "y _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "is a predefined Begin Of Pose (BOP). During the training phase, the proposed", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 627, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 104, + 627, + 505, + 642 + ], + "score": 1.0, + "content": "curriculum learning strategy gently increases the number of steps of autoregressive scheme according", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 639, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 506, + 652 + ], + "score": 1.0, + "content": "to the number of training epochs. This strategy bridges the gap between training and inference of", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 651, + 405, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 405, + 662 + ], + "score": 1.0, + "content": "autoregressive models, and thus alleviates error accumulation at inference.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5 + } + ], + "index": 30.75 + }, + { + "type": "title", + "bbox": [ + 108, + 689, + 248, + 699 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 250, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 250, + 701 + ], + "score": 1.0, + "content": "3.1 PROBLEM FORMALIZATION", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 708, + 504, + 733 + ], + "lines": [ + { + "bbox": [ + 104, + 706, + 507, + 725 + ], + "spans": [ + { + "bbox": [ + 104, + 706, + 226, + 725 + ], + "score": 1.0, + "content": "Suppose that there is a dataset", + "type": "text" + }, + { + "bbox": [ + 226, + 709, + 304, + 722 + ], + "score": 0.91, + "content": "\\mathcal { D } = \\{ ( X _ { i } , Y _ { i } ) \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 706, + 334, + 725 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 334, + 710, + 390, + 722 + ], + "score": 0.93, + "content": "X = \\{ x _ { t } \\} _ { t = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 706, + 469, + 725 + ], + "score": 1.0, + "content": "is a music clip with", + "type": "text" + }, + { + "bbox": [ + 469, + 711, + 480, + 721 + ], + "score": 0.84, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 706, + 507, + 725 + ], + "score": 1.0, + "content": "being", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 718, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 104, + 718, + 267, + 734 + ], + "score": 1.0, + "content": "a vector of acoustic features at time-step", + "type": "text" + }, + { + "bbox": [ + 267, + 722, + 272, + 730 + ], + "score": 0.77, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 718, + 293, + 734 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 293, + 721, + 347, + 732 + ], + "score": 0.93, + "content": "\\bar { Y = } \\{ y _ { t } \\} _ { t = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 718, + 506, + 734 + ], + "score": 1.0, + "content": "is a sequence of dance movements with", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 84, + 116, + 94 + ], + "score": 0.82, + "content": "y _ { t }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 116, + 82, + 158, + 95 + ], + "score": 1.0, + "content": "aligned to", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 159, + 84, + 169, + 93 + ], + "score": 0.84, + "content": "x _ { t }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 169, + 82, + 342, + 95 + ], + "score": 1.0, + "content": ". The goal is to estimate a generation model", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 342, + 82, + 359, + 95 + ], + "score": 0.9, + "content": "g ( \\cdot )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 359, + 82, + 381, + 95 + ], + "score": 1.0, + "content": "from", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 382, + 83, + 390, + 92 + ], + "score": 0.75, + "content": "\\mathcal { D }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 391, + 82, + 505, + 95 + ], + "score": 1.0, + "content": ", and thus given a new music", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 129, + 107 + ], + "score": 1.0, + "content": "input", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 129, + 94, + 139, + 104 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 140, + 93, + 269, + 107 + ], + "score": 1.0, + "content": ", the model can synthesize dance", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 269, + 94, + 279, + 104 + ], + "score": 0.73, + "content": "Y", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 279, + 93, + 316, + 107 + ], + "score": 1.0, + "content": "to music", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 316, + 94, + 326, + 104 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 327, + 93, + 365, + 107 + ], + "score": 1.0, + "content": "based on", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 365, + 94, + 387, + 105 + ], + "score": 0.91, + "content": "g ( X )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 388, + 93, + 505, + 107 + ], + "score": 1.0, + "content": ". We first present our seq2seq", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "architectures chosen for music-conditioned dance generation in the following section. Later in the", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 507, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 507, + 128 + ], + "score": 1.0, + "content": "experiments, we empirically justify the choice by comparing the architectures with other alternatives.", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 39.5, + "bbox_fs": [ + 104, + 706, + 507, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 84, + 116, + 94 + ], + "score": 0.82, + "content": "y _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 82, + 158, + 95 + ], + "score": 1.0, + "content": "aligned to", + "type": "text" + }, + { + "bbox": [ + 159, + 84, + 169, + 93 + ], + "score": 0.84, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 82, + 342, + 95 + ], + "score": 1.0, + "content": ". The goal is to estimate a generation model", + "type": "text" + }, + { + "bbox": [ + 342, + 82, + 359, + 95 + ], + "score": 0.9, + "content": "g ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 82, + 381, + 95 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 382, + 83, + 390, + 92 + ], + "score": 0.75, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 82, + 505, + 95 + ], + "score": 1.0, + "content": ", and thus given a new music", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 129, + 107 + ], + "score": 1.0, + "content": "input", + "type": "text" + }, + { + "bbox": [ + 129, + 94, + 139, + 104 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 93, + 269, + 107 + ], + "score": 1.0, + "content": ", the model can synthesize dance", + "type": "text" + }, + { + "bbox": [ + 269, + 94, + 279, + 104 + ], + "score": 0.73, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 93, + 316, + 107 + ], + "score": 1.0, + "content": "to music", + "type": "text" + }, + { + "bbox": [ + 316, + 94, + 326, + 104 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 93, + 365, + 107 + ], + "score": 1.0, + "content": "based on", + "type": "text" + }, + { + "bbox": [ + 365, + 94, + 387, + 105 + ], + "score": 0.91, + "content": "g ( X )", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 93, + 505, + 107 + ], + "score": 1.0, + "content": ". We first present our seq2seq", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "architectures chosen for music-conditioned dance generation in the following section. Later in the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 507, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 507, + 128 + ], + "score": 1.0, + "content": "experiments, we empirically justify the choice by comparing the architectures with other alternatives.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 107, + 140, + 235, + 151 + ], + "lines": [ + { + "bbox": [ + 106, + 140, + 235, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 235, + 151 + ], + "score": 1.0, + "content": "3.2 MODEL ARCHITECTURE", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 160, + 505, + 206 + ], + "lines": [ + { + "bbox": [ + 105, + 159, + 507, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 196, + 173 + ], + "score": 1.0, + "content": "In the architecture of", + "type": "text" + }, + { + "bbox": [ + 197, + 160, + 213, + 172 + ], + "score": 0.88, + "content": "g ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 159, + 354, + 173 + ], + "score": 1.0, + "content": ", a music encoder first transforms", + "type": "text" + }, + { + "bbox": [ + 355, + 160, + 427, + 172 + ], + "score": 0.81, + "content": "X = ( x _ { 1 } , . . . , x _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 160, + 476, + 172 + ], + "score": 0.71, + "content": "( x _ { i } \\in \\mathbb { R } ^ { d _ { x } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 159, + 507, + 173 + ], + "score": 1.0, + "content": "into a", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 171, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 104, + 171, + 177, + 185 + ], + "score": 1.0, + "content": "hidden sequence", + "type": "text" + }, + { + "bbox": [ + 177, + 172, + 243, + 184 + ], + "score": 0.65, + "content": "Z = ( z _ { 1 } , . . . , z _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 171, + 288, + 183 + ], + "score": 0.68, + "content": "( z _ { i } \\in \\mathbb { R } ^ { d _ { z } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 171, + 506, + 185 + ], + "score": 1.0, + "content": "using a local self-attention mechanism to reduce the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 194 + ], + "score": 1.0, + "content": "memory requirement for long sequence modeling, and then a dance decoder exploits a recurrent", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 194, + 446, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 303, + 206 + ], + "score": 1.0, + "content": "structure to autoregressively predicts movements", + "type": "text" + }, + { + "bbox": [ + 304, + 194, + 370, + 206 + ], + "score": 0.93, + "content": "Y = ( y _ { 1 } , . . . , y _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 194, + 434, + 206 + ], + "score": 1.0, + "content": "conditioned on", + "type": "text" + }, + { + "bbox": [ + 434, + 194, + 443, + 204 + ], + "score": 0.79, + "content": "Z", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 194, + 446, + 206 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 216, + 506, + 360 + ], + "lines": [ + { + "bbox": [ + 105, + 216, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 506, + 230 + ], + "score": 1.0, + "content": "Music Encoder. Encouraged by the compelling performance on music generation (Huang et al.,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 227, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 506, + 240 + ], + "score": 1.0, + "content": "2018), we define the music encoder with a transformer encoder structure. While the self-attention", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 239, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 505, + 251 + ], + "score": 1.0, + "content": "mechanism (Vaswani et al., 2017) in transformer can effectively represent the multi-scale structure of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 272, + 263 + ], + "score": 1.0, + "content": "music, the quadratic memory complexity", + "type": "text" + }, + { + "bbox": [ + 272, + 249, + 300, + 262 + ], + "score": 0.93, + "content": "\\scriptstyle { \\mathcal { O } } ( n ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 250, + 392, + 263 + ], + "score": 1.0, + "content": "about sequence length", + "type": "text" + }, + { + "bbox": [ + 392, + 252, + 399, + 260 + ], + "score": 0.74, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 250, + 506, + 263 + ], + "score": 1.0, + "content": "impedes its application to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 261, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 506, + 274 + ], + "score": 1.0, + "content": "long sequence modeling due to huge of GPU memory consumption (Child et al., 2019). To keep the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 272, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 506, + 284 + ], + "score": 1.0, + "content": "effectiveness of representation and control the cost, we introduce a local self-attention mechanism that", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 283, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 465, + 295 + ], + "score": 1.0, + "content": "modifies the receptive field of self-attention by restricting the element connections within", + "type": "text" + }, + { + "bbox": [ + 465, + 284, + 472, + 293 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 283, + 505, + 295 + ], + "score": 1.0, + "content": "-nearest", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 293, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 322, + 306 + ], + "score": 1.0, + "content": "neighbors. Thus, the memory complexity is reduced to", + "type": "text" + }, + { + "bbox": [ + 323, + 294, + 351, + 306 + ], + "score": 0.9, + "content": "\\mathcal { O } ( n k )", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 293, + 355, + 306 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 355, + 294, + 362, + 304 + ], + "score": 0.71, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 293, + 506, + 306 + ], + "score": 1.0, + "content": "could be small in our scenario since", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 304, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 506, + 317 + ], + "score": 1.0, + "content": "we only pursue an effective representation for a given music clip. Therefore, the local patterns of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 316, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 193, + 329 + ], + "score": 1.0, + "content": "music are encoded in", + "type": "text" + }, + { + "bbox": [ + 194, + 317, + 203, + 327 + ], + "score": 0.86, + "content": "z _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 316, + 399, + 329 + ], + "score": 1.0, + "content": "that is sufficient for the generation of movement", + "type": "text" + }, + { + "bbox": [ + 400, + 317, + 409, + 327 + ], + "score": 0.84, + "content": "y _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 316, + 460, + 329 + ], + "score": 1.0, + "content": "at time-step", + "type": "text" + }, + { + "bbox": [ + 460, + 317, + 465, + 326 + ], + "score": 0.7, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 316, + 506, + 329 + ], + "score": 1.0, + "content": ", which is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 326, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 505, + 340 + ], + "score": 1.0, + "content": "aligned with the common sense that the dance movement at certain time-step is highly influenced by", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 337, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 505, + 351 + ], + "score": 1.0, + "content": "the nearby clips of music. Yet we can handle long sequences of acoustic features, e.g., more than", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 347, + 342, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 342, + 362 + ], + "score": 1.0, + "content": "1000 elements, in an efficient and memory-economic way.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 503, + 388 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 216, + 378 + ], + "score": 1.0, + "content": "Specifically, we first embed", + "type": "text" + }, + { + "bbox": [ + 217, + 365, + 286, + 377 + ], + "score": 0.91, + "content": "X = ( x _ { 1 } , . . . , x _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 364, + 305, + 378 + ], + "score": 1.0, + "content": "into", + "type": "text" + }, + { + "bbox": [ + 305, + 365, + 373, + 378 + ], + "score": 0.91, + "content": "U = ( u _ { 1 } , . . . , u _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 364, + 505, + 378 + ], + "score": 1.0, + "content": "with a linear layer parameterized", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 373, + 372, + 390 + ], + "spans": [ + { + "bbox": [ + 104, + 373, + 119, + 390 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 119, + 376, + 180, + 387 + ], + "score": 0.92, + "content": "W ^ { E } \\in \\mathbf { \\bar { \\mathbb { R } } } ^ { d _ { x } \\times d _ { z } }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 373, + 208, + 390 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + }, + { + "bbox": [ + 208, + 377, + 268, + 389 + ], + "score": 0.91, + "content": "\\forall i \\in \\{ 1 , . . . , n \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 373, + 272, + 390 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 272, + 378, + 281, + 388 + ], + "score": 0.72, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 373, + 372, + 390 + ], + "score": 1.0, + "content": "can be formulated as:", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "interline_equation", + "bbox": [ + 186, + 393, + 424, + 430 + ], + "lines": [ + { + "bbox": [ + 186, + 393, + 424, + 430 + ], + "spans": [ + { + "bbox": [ + 186, + 393, + 424, + 430 + ], + "score": 0.94, + "content": "z _ { i } = \\mathcal { F } ( a _ { i } ) , \\quad a _ { i } = \\sum _ { j = i - \\lfloor k / 2 \\rfloor } ^ { i + \\lfloor k / 2 \\rfloor } \\alpha _ { i j } ( u _ { j } W _ { l } ^ { V } ) , \\quad U = X W ^ { E } ,", + "type": "interline_equation", + "image_path": "5dbdfb230aa9b6ecb8673ef1df3a2d5a1901bf5046f8c3547666140729add6dc.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 186, + 393, + 424, + 411.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 186, + 411.5, + 424, + 430.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 435, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 133, + 448 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 435, + 211, + 448 + ], + "score": 0.92, + "content": "\\mathcal { F } ( \\cdot ) : \\mathbb { R } ^ { d _ { v } } \\mathbb { R } ^ { d _ { z } }", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 434, + 376, + 448 + ], + "score": 1.0, + "content": "is a feed forward neural network. Each", + "type": "text" + }, + { + "bbox": [ + 377, + 438, + 388, + 448 + ], + "score": 0.86, + "content": "u _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 434, + 505, + 448 + ], + "score": 1.0, + "content": "is only allowed to attend its", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 107, + 448, + 113, + 457 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 447, + 283, + 459 + ], + "score": 1.0, + "content": "-nearest neighbors including itself, where", + "type": "text" + }, + { + "bbox": [ + 283, + 447, + 290, + 457 + ], + "score": 0.84, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "is a hyper-parameter referring to the sliding window", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 457, + 503, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 320, + 470 + ], + "score": 1.0, + "content": "size of the local self-attention. Then attention weight", + "type": "text" + }, + { + "bbox": [ + 320, + 459, + 334, + 470 + ], + "score": 0.89, + "content": "\\alpha _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 457, + 503, + 470 + ], + "score": 1.0, + "content": "is calculated using a softmax function as:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 188, + 474, + 422, + 509 + ], + "lines": [ + { + "bbox": [ + 188, + 474, + 422, + 509 + ], + "spans": [ + { + "bbox": [ + 188, + 474, + 422, + 509 + ], + "score": 0.94, + "content": "\\alpha _ { i j } = \\frac { \\exp { e _ { i j } } } { \\sum _ { t = j - \\lfloor k / 2 \\rfloor } ^ { j + \\lfloor k / 2 \\rfloor } \\exp { e _ { i t } } } , e _ { i j } = \\frac { ( u _ { i } W _ { l } ^ { Q } ) ( u _ { j } W _ { l } ^ { K } ) ^ { \\top } } { \\sqrt { d _ { k } } } ,", + "type": "interline_equation", + "image_path": "a0a63acb0667d3d33495087c124f59b9e1835abeaea8585f230fab1324b86adb.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 188, + 474, + 422, + 491.5 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 188, + 491.5, + 422, + 509.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 513, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 511, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 177, + 528 + ], + "score": 1.0, + "content": "where for the the", + "type": "text" + }, + { + "bbox": [ + 177, + 516, + 181, + 525 + ], + "score": 0.76, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 511, + 218, + 528 + ], + "score": 1.0, + "content": "-th head,", + "type": "text" + }, + { + "bbox": [ + 218, + 513, + 236, + 527 + ], + "score": 0.83, + "content": "W _ { l } ^ { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 511, + 240, + 528 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 241, + 513, + 303, + 527 + ], + "score": 0.84, + "content": "W _ { l } ^ { K } \\in \\mathbb { R } ^ { d _ { z } \\times d _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 511, + 321, + 528 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 322, + 514, + 383, + 527 + ], + "score": 0.93, + "content": "W _ { l } ^ { V } \\in \\mathbb { R } ^ { d _ { z } \\times d _ { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 511, + 505, + 528 + ], + "score": 1.0, + "content": "are parameters that transform", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 107, + 527, + 115, + 536 + ], + "score": 0.78, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 526, + 293, + 538 + ], + "score": 1.0, + "content": "into a query, a key, and a value respectively.", + "type": "text" + }, + { + "bbox": [ + 294, + 527, + 304, + 537 + ], + "score": 0.86, + "content": "d _ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 526, + 433, + 538 + ], + "score": 1.0, + "content": "is the dimension of hidden state", + "type": "text" + }, + { + "bbox": [ + 434, + 528, + 443, + 537 + ], + "score": 0.85, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 526, + 469, + 538 + ], + "score": 1.0, + "content": "while", + "type": "text" + }, + { + "bbox": [ + 469, + 527, + 480, + 537 + ], + "score": 0.88, + "content": "d _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 537, + 339, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 221, + 549 + ], + "score": 1.0, + "content": "dimension of query, key and", + "type": "text" + }, + { + "bbox": [ + 221, + 537, + 232, + 548 + ], + "score": 0.89, + "content": "d _ { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 537, + 339, + 549 + ], + "score": 1.0, + "content": "is the dimension of value.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 106, + 560, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 506, + 573 + ], + "score": 1.0, + "content": "Dance Decoder. We choose a recurrent neural network as the dance decoder in consideration of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 570, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 506, + 584 + ], + "score": 1.0, + "content": "two factors: (1) the chain structure can well capture the spatial-temporal dependency among human", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "movement dynamics, which has proven to be highly effective in the state-of-the-art methods for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 593, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 506, + 606 + ], + "score": 1.0, + "content": "human motion prediction (Li et al., 2018; Mao et al., 2019); (2) our proposed learning strategy is", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 603, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 506, + 617 + ], + "score": 1.0, + "content": "tailored for the autoregressive models like RNN, as will be described in the next section. Specifically,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 614, + 432, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 126, + 628 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 615, + 192, + 627 + ], + "score": 0.93, + "content": "Z = ( z _ { 1 } , . . . , z _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 614, + 284, + 628 + ], + "score": 1.0, + "content": ", the dance movements", + "type": "text" + }, + { + "bbox": [ + 285, + 615, + 352, + 627 + ], + "score": 0.93, + "content": "\\boldsymbol { Y } = ( \\hat { y } _ { 1 } , . . . , \\hat { y } _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 614, + 432, + 628 + ], + "score": 1.0, + "content": "are synthesized by:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5 + }, + { + "type": "interline_equation", + "bbox": [ + 255, + 631, + 354, + 662 + ], + "lines": [ + { + "bbox": [ + 255, + 631, + 354, + 662 + ], + "spans": [ + { + "bbox": [ + 255, + 631, + 354, + 662 + ], + "score": 0.91, + "content": "\\begin{array} { l } { \\hat { y } _ { i } = [ h _ { i } ; z _ { i } ] W ^ { S } + b , } \\\\ { h _ { i } = \\mathrm { R N N } ( h _ { i - 1 } , \\hat { y } _ { i - 1 } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "3f8ddabcf4f07733079df1da51a1bdcc825cb9c03f3726d9786db20821f1d134.jpg" + } + ] + } + ], + "index": 40.5, + "virtual_lines": [ + { + "bbox": [ + 255, + 631, + 354, + 646.5 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 255, + 646.5, + 354, + 662.0 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 663, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 664, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 133, + 677 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 665, + 144, + 676 + ], + "score": 0.88, + "content": "h _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 664, + 169, + 677 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 169, + 665, + 174, + 674 + ], + "score": 0.73, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 664, + 314, + 677 + ], + "score": 1.0, + "content": "-th hidden state of the decoder and", + "type": "text" + }, + { + "bbox": [ + 315, + 665, + 326, + 676 + ], + "score": 0.89, + "content": "h _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 664, + 505, + 677 + ], + "score": 1.0, + "content": "is initialized by sampling from the standard", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 675, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 381, + 688 + ], + "score": 1.0, + "content": "normal distribution to enhance variation of the generated sequences.", + "type": "text" + }, + { + "bbox": [ + 382, + 676, + 398, + 687 + ], + "score": 0.88, + "content": "[ \\cdot ; \\cdot ]", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 675, + 505, + 688 + ], + "score": 1.0, + "content": "denotes the concatenation", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 103, + 684, + 507, + 702 + ], + "spans": [ + { + "bbox": [ + 103, + 684, + 150, + 702 + ], + "score": 1.0, + "content": "operation.", + "type": "text" + }, + { + "bbox": [ + 151, + 686, + 230, + 698 + ], + "score": 0.91, + "content": "W ^ { S } \\in \\mathbb { R } ^ { ( d _ { s } + d _ { z } ) \\times d _ { y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 684, + 250, + 702 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 250, + 687, + 282, + 698 + ], + "score": 0.9, + "content": "b \\in \\mathbb { R } ^ { d _ { y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 684, + 374, + 702 + ], + "score": 1.0, + "content": "are parameters where", + "type": "text" + }, + { + "bbox": [ + 374, + 689, + 384, + 699 + ], + "score": 0.88, + "content": "d _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 684, + 403, + 702 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 403, + 688, + 414, + 700 + ], + "score": 0.89, + "content": "d _ { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 684, + 507, + 702 + ], + "score": 1.0, + "content": "are the dimensions of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 699, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 117, + 710 + ], + "score": 0.88, + "content": "h _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 699, + 134, + 712 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 135, + 699, + 144, + 710 + ], + "score": 0.87, + "content": "\\hat { y } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 699, + 226, + 712 + ], + "score": 1.0, + "content": ", respectively. At the", + "type": "text" + }, + { + "bbox": [ + 226, + 700, + 230, + 709 + ], + "score": 0.77, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 699, + 422, + 712 + ], + "score": 1.0, + "content": "-th time-step, the decoder predicts the movement", + "type": "text" + }, + { + "bbox": [ + 422, + 700, + 432, + 710 + ], + "score": 0.88, + "content": "\\hat { y } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 699, + 493, + 712 + ], + "score": 1.0, + "content": "conditioned on", + "type": "text" + }, + { + "bbox": [ + 494, + 700, + 504, + 710 + ], + "score": 0.86, + "content": "h _ { i }", + "type": "inline_equation" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 277, + 722 + ], + "score": 1.0, + "content": "as well as the latent feature representation", + "type": "text" + }, + { + "bbox": [ + 277, + 714, + 286, + 721 + ], + "score": 0.84, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 709, + 505, + 722 + ], + "score": 1.0, + "content": ", and thus can capture the fine-grained correspondence", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 721, + 366, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 366, + 734 + ], + "score": 1.0, + "content": "between music feature sequence and dance movement sequence.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 105, + 82, + 507, + 128 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 140, + 235, + 151 + ], + "lines": [ + { + "bbox": [ + 106, + 140, + 235, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 235, + 151 + ], + "score": 1.0, + "content": "3.2 MODEL ARCHITECTURE", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 160, + 505, + 206 + ], + "lines": [ + { + "bbox": [ + 105, + 159, + 507, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 196, + 173 + ], + "score": 1.0, + "content": "In the architecture of", + "type": "text" + }, + { + "bbox": [ + 197, + 160, + 213, + 172 + ], + "score": 0.88, + "content": "g ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 159, + 354, + 173 + ], + "score": 1.0, + "content": ", a music encoder first transforms", + "type": "text" + }, + { + "bbox": [ + 355, + 160, + 427, + 172 + ], + "score": 0.81, + "content": "X = ( x _ { 1 } , . . . , x _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 160, + 476, + 172 + ], + "score": 0.71, + "content": "( x _ { i } \\in \\mathbb { R } ^ { d _ { x } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 159, + 507, + 173 + ], + "score": 1.0, + "content": "into a", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 171, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 104, + 171, + 177, + 185 + ], + "score": 1.0, + "content": "hidden sequence", + "type": "text" + }, + { + "bbox": [ + 177, + 172, + 243, + 184 + ], + "score": 0.65, + "content": "Z = ( z _ { 1 } , . . . , z _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 171, + 288, + 183 + ], + "score": 0.68, + "content": "( z _ { i } \\in \\mathbb { R } ^ { d _ { z } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 171, + 506, + 185 + ], + "score": 1.0, + "content": "using a local self-attention mechanism to reduce the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 194 + ], + "score": 1.0, + "content": "memory requirement for long sequence modeling, and then a dance decoder exploits a recurrent", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 194, + 446, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 303, + 206 + ], + "score": 1.0, + "content": "structure to autoregressively predicts movements", + "type": "text" + }, + { + "bbox": [ + 304, + 194, + 370, + 206 + ], + "score": 0.93, + "content": "Y = ( y _ { 1 } , . . . , y _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 194, + 434, + 206 + ], + "score": 1.0, + "content": "conditioned on", + "type": "text" + }, + { + "bbox": [ + 434, + 194, + 443, + 204 + ], + "score": 0.79, + "content": "Z", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 194, + 446, + 206 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 104, + 159, + 507, + 206 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 216, + 506, + 360 + ], + "lines": [ + { + "bbox": [ + 105, + 216, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 506, + 230 + ], + "score": 1.0, + "content": "Music Encoder. Encouraged by the compelling performance on music generation (Huang et al.,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 227, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 506, + 240 + ], + "score": 1.0, + "content": "2018), we define the music encoder with a transformer encoder structure. While the self-attention", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 239, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 505, + 251 + ], + "score": 1.0, + "content": "mechanism (Vaswani et al., 2017) in transformer can effectively represent the multi-scale structure of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 272, + 263 + ], + "score": 1.0, + "content": "music, the quadratic memory complexity", + "type": "text" + }, + { + "bbox": [ + 272, + 249, + 300, + 262 + ], + "score": 0.93, + "content": "\\scriptstyle { \\mathcal { O } } ( n ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 250, + 392, + 263 + ], + "score": 1.0, + "content": "about sequence length", + "type": "text" + }, + { + "bbox": [ + 392, + 252, + 399, + 260 + ], + "score": 0.74, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 250, + 506, + 263 + ], + "score": 1.0, + "content": "impedes its application to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 261, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 506, + 274 + ], + "score": 1.0, + "content": "long sequence modeling due to huge of GPU memory consumption (Child et al., 2019). To keep the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 272, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 506, + 284 + ], + "score": 1.0, + "content": "effectiveness of representation and control the cost, we introduce a local self-attention mechanism that", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 283, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 465, + 295 + ], + "score": 1.0, + "content": "modifies the receptive field of self-attention by restricting the element connections within", + "type": "text" + }, + { + "bbox": [ + 465, + 284, + 472, + 293 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 283, + 505, + 295 + ], + "score": 1.0, + "content": "-nearest", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 293, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 322, + 306 + ], + "score": 1.0, + "content": "neighbors. Thus, the memory complexity is reduced to", + "type": "text" + }, + { + "bbox": [ + 323, + 294, + 351, + 306 + ], + "score": 0.9, + "content": "\\mathcal { O } ( n k )", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 293, + 355, + 306 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 355, + 294, + 362, + 304 + ], + "score": 0.71, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 293, + 506, + 306 + ], + "score": 1.0, + "content": "could be small in our scenario since", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 304, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 506, + 317 + ], + "score": 1.0, + "content": "we only pursue an effective representation for a given music clip. Therefore, the local patterns of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 316, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 193, + 329 + ], + "score": 1.0, + "content": "music are encoded in", + "type": "text" + }, + { + "bbox": [ + 194, + 317, + 203, + 327 + ], + "score": 0.86, + "content": "z _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 316, + 399, + 329 + ], + "score": 1.0, + "content": "that is sufficient for the generation of movement", + "type": "text" + }, + { + "bbox": [ + 400, + 317, + 409, + 327 + ], + "score": 0.84, + "content": "y _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 316, + 460, + 329 + ], + "score": 1.0, + "content": "at time-step", + "type": "text" + }, + { + "bbox": [ + 460, + 317, + 465, + 326 + ], + "score": 0.7, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 316, + 506, + 329 + ], + "score": 1.0, + "content": ", which is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 326, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 505, + 340 + ], + "score": 1.0, + "content": "aligned with the common sense that the dance movement at certain time-step is highly influenced by", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 337, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 505, + 351 + ], + "score": 1.0, + "content": "the nearby clips of music. Yet we can handle long sequences of acoustic features, e.g., more than", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 347, + 342, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 342, + 362 + ], + "score": 1.0, + "content": "1000 elements, in an efficient and memory-economic way.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 216, + 506, + 362 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 503, + 388 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 216, + 378 + ], + "score": 1.0, + "content": "Specifically, we first embed", + "type": "text" + }, + { + "bbox": [ + 217, + 365, + 286, + 377 + ], + "score": 0.91, + "content": "X = ( x _ { 1 } , . . . , x _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 364, + 305, + 378 + ], + "score": 1.0, + "content": "into", + "type": "text" + }, + { + "bbox": [ + 305, + 365, + 373, + 378 + ], + "score": 0.91, + "content": "U = ( u _ { 1 } , . . . , u _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 364, + 505, + 378 + ], + "score": 1.0, + "content": "with a linear layer parameterized", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 373, + 372, + 390 + ], + "spans": [ + { + "bbox": [ + 104, + 373, + 119, + 390 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 119, + 376, + 180, + 387 + ], + "score": 0.92, + "content": "W ^ { E } \\in \\mathbf { \\bar { \\mathbb { R } } } ^ { d _ { x } \\times d _ { z } }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 373, + 208, + 390 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + }, + { + "bbox": [ + 208, + 377, + 268, + 389 + ], + "score": 0.91, + "content": "\\forall i \\in \\{ 1 , . . . , n \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 373, + 272, + 390 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 272, + 378, + 281, + 388 + ], + "score": 0.72, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 373, + 372, + 390 + ], + "score": 1.0, + "content": "can be formulated as:", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 104, + 364, + 505, + 390 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 186, + 393, + 424, + 430 + ], + "lines": [ + { + "bbox": [ + 186, + 393, + 424, + 430 + ], + "spans": [ + { + "bbox": [ + 186, + 393, + 424, + 430 + ], + "score": 0.94, + "content": "z _ { i } = \\mathcal { F } ( a _ { i } ) , \\quad a _ { i } = \\sum _ { j = i - \\lfloor k / 2 \\rfloor } ^ { i + \\lfloor k / 2 \\rfloor } \\alpha _ { i j } ( u _ { j } W _ { l } ^ { V } ) , \\quad U = X W ^ { E } ,", + "type": "interline_equation", + "image_path": "5dbdfb230aa9b6ecb8673ef1df3a2d5a1901bf5046f8c3547666140729add6dc.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 186, + 393, + 424, + 411.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 186, + 411.5, + 424, + 430.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 435, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 133, + 448 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 435, + 211, + 448 + ], + "score": 0.92, + "content": "\\mathcal { F } ( \\cdot ) : \\mathbb { R } ^ { d _ { v } } \\mathbb { R } ^ { d _ { z } }", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 434, + 376, + 448 + ], + "score": 1.0, + "content": "is a feed forward neural network. Each", + "type": "text" + }, + { + "bbox": [ + 377, + 438, + 388, + 448 + ], + "score": 0.86, + "content": "u _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 434, + 505, + 448 + ], + "score": 1.0, + "content": "is only allowed to attend its", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 107, + 448, + 113, + 457 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 447, + 283, + 459 + ], + "score": 1.0, + "content": "-nearest neighbors including itself, where", + "type": "text" + }, + { + "bbox": [ + 283, + 447, + 290, + 457 + ], + "score": 0.84, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "is a hyper-parameter referring to the sliding window", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 457, + 503, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 320, + 470 + ], + "score": 1.0, + "content": "size of the local self-attention. Then attention weight", + "type": "text" + }, + { + "bbox": [ + 320, + 459, + 334, + 470 + ], + "score": 0.89, + "content": "\\alpha _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 457, + 503, + 470 + ], + "score": 1.0, + "content": "is calculated using a softmax function as:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 434, + 505, + 470 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 188, + 474, + 422, + 509 + ], + "lines": [ + { + "bbox": [ + 188, + 474, + 422, + 509 + ], + "spans": [ + { + "bbox": [ + 188, + 474, + 422, + 509 + ], + "score": 0.94, + "content": "\\alpha _ { i j } = \\frac { \\exp { e _ { i j } } } { \\sum _ { t = j - \\lfloor k / 2 \\rfloor } ^ { j + \\lfloor k / 2 \\rfloor } \\exp { e _ { i t } } } , e _ { i j } = \\frac { ( u _ { i } W _ { l } ^ { Q } ) ( u _ { j } W _ { l } ^ { K } ) ^ { \\top } } { \\sqrt { d _ { k } } } ,", + "type": "interline_equation", + "image_path": "a0a63acb0667d3d33495087c124f59b9e1835abeaea8585f230fab1324b86adb.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 188, + 474, + 422, + 491.5 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 188, + 491.5, + 422, + 509.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 513, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 511, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 177, + 528 + ], + "score": 1.0, + "content": "where for the the", + "type": "text" + }, + { + "bbox": [ + 177, + 516, + 181, + 525 + ], + "score": 0.76, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 511, + 218, + 528 + ], + "score": 1.0, + "content": "-th head,", + "type": "text" + }, + { + "bbox": [ + 218, + 513, + 236, + 527 + ], + "score": 0.83, + "content": "W _ { l } ^ { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 511, + 240, + 528 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 241, + 513, + 303, + 527 + ], + "score": 0.84, + "content": "W _ { l } ^ { K } \\in \\mathbb { R } ^ { d _ { z } \\times d _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 511, + 321, + 528 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 322, + 514, + 383, + 527 + ], + "score": 0.93, + "content": "W _ { l } ^ { V } \\in \\mathbb { R } ^ { d _ { z } \\times d _ { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 511, + 505, + 528 + ], + "score": 1.0, + "content": "are parameters that transform", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 107, + 527, + 115, + 536 + ], + "score": 0.78, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 526, + 293, + 538 + ], + "score": 1.0, + "content": "into a query, a key, and a value respectively.", + "type": "text" + }, + { + "bbox": [ + 294, + 527, + 304, + 537 + ], + "score": 0.86, + "content": "d _ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 526, + 433, + 538 + ], + "score": 1.0, + "content": "is the dimension of hidden state", + "type": "text" + }, + { + "bbox": [ + 434, + 528, + 443, + 537 + ], + "score": 0.85, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 526, + 469, + 538 + ], + "score": 1.0, + "content": "while", + "type": "text" + }, + { + "bbox": [ + 469, + 527, + 480, + 537 + ], + "score": 0.88, + "content": "d _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 537, + 339, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 221, + 549 + ], + "score": 1.0, + "content": "dimension of query, key and", + "type": "text" + }, + { + "bbox": [ + 221, + 537, + 232, + 548 + ], + "score": 0.89, + "content": "d _ { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 537, + 339, + 549 + ], + "score": 1.0, + "content": "is the dimension of value.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 511, + 505, + 549 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 106, + 560, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 506, + 573 + ], + "score": 1.0, + "content": "Dance Decoder. We choose a recurrent neural network as the dance decoder in consideration of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 570, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 506, + 584 + ], + "score": 1.0, + "content": "two factors: (1) the chain structure can well capture the spatial-temporal dependency among human", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "movement dynamics, which has proven to be highly effective in the state-of-the-art methods for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 593, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 506, + 606 + ], + "score": 1.0, + "content": "human motion prediction (Li et al., 2018; Mao et al., 2019); (2) our proposed learning strategy is", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 603, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 506, + 617 + ], + "score": 1.0, + "content": "tailored for the autoregressive models like RNN, as will be described in the next section. Specifically,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 614, + 432, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 126, + 628 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 615, + 192, + 627 + ], + "score": 0.93, + "content": "Z = ( z _ { 1 } , . . . , z _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 614, + 284, + 628 + ], + "score": 1.0, + "content": ", the dance movements", + "type": "text" + }, + { + "bbox": [ + 285, + 615, + 352, + 627 + ], + "score": 0.93, + "content": "\\boldsymbol { Y } = ( \\hat { y } _ { 1 } , . . . , \\hat { y } _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 614, + 432, + 628 + ], + "score": 1.0, + "content": "are synthesized by:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 560, + 506, + 628 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 255, + 631, + 354, + 662 + ], + "lines": [ + { + "bbox": [ + 255, + 631, + 354, + 662 + ], + "spans": [ + { + "bbox": [ + 255, + 631, + 354, + 662 + ], + "score": 0.91, + "content": "\\begin{array} { l } { \\hat { y } _ { i } = [ h _ { i } ; z _ { i } ] W ^ { S } + b , } \\\\ { h _ { i } = \\mathrm { R N N } ( h _ { i - 1 } , \\hat { y } _ { i - 1 } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "3f8ddabcf4f07733079df1da51a1bdcc825cb9c03f3726d9786db20821f1d134.jpg" + } + ] + } + ], + "index": 40.5, + "virtual_lines": [ + { + "bbox": [ + 255, + 631, + 354, + 646.5 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 255, + 646.5, + 354, + 662.0 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 663, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 664, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 133, + 677 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 665, + 144, + 676 + ], + "score": 0.88, + "content": "h _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 664, + 169, + 677 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 169, + 665, + 174, + 674 + ], + "score": 0.73, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 664, + 314, + 677 + ], + "score": 1.0, + "content": "-th hidden state of the decoder and", + "type": "text" + }, + { + "bbox": [ + 315, + 665, + 326, + 676 + ], + "score": 0.89, + "content": "h _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 664, + 505, + 677 + ], + "score": 1.0, + "content": "is initialized by sampling from the standard", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 675, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 381, + 688 + ], + "score": 1.0, + "content": "normal distribution to enhance variation of the generated sequences.", + "type": "text" + }, + { + "bbox": [ + 382, + 676, + 398, + 687 + ], + "score": 0.88, + "content": "[ \\cdot ; \\cdot ]", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 675, + 505, + 688 + ], + "score": 1.0, + "content": "denotes the concatenation", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 103, + 684, + 507, + 702 + ], + "spans": [ + { + "bbox": [ + 103, + 684, + 150, + 702 + ], + "score": 1.0, + "content": "operation.", + "type": "text" + }, + { + "bbox": [ + 151, + 686, + 230, + 698 + ], + "score": 0.91, + "content": "W ^ { S } \\in \\mathbb { R } ^ { ( d _ { s } + d _ { z } ) \\times d _ { y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 684, + 250, + 702 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 250, + 687, + 282, + 698 + ], + "score": 0.9, + "content": "b \\in \\mathbb { R } ^ { d _ { y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 684, + 374, + 702 + ], + "score": 1.0, + "content": "are parameters where", + "type": "text" + }, + { + "bbox": [ + 374, + 689, + 384, + 699 + ], + "score": 0.88, + "content": "d _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 684, + 403, + 702 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 403, + 688, + 414, + 700 + ], + "score": 0.89, + "content": "d _ { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 684, + 507, + 702 + ], + "score": 1.0, + "content": "are the dimensions of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 699, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 117, + 710 + ], + "score": 0.88, + "content": "h _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 699, + 134, + 712 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 135, + 699, + 144, + 710 + ], + "score": 0.87, + "content": "\\hat { y } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 699, + 226, + 712 + ], + "score": 1.0, + "content": ", respectively. At the", + "type": "text" + }, + { + "bbox": [ + 226, + 700, + 230, + 709 + ], + "score": 0.77, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 699, + 422, + 712 + ], + "score": 1.0, + "content": "-th time-step, the decoder predicts the movement", + "type": "text" + }, + { + "bbox": [ + 422, + 700, + 432, + 710 + ], + "score": 0.88, + "content": "\\hat { y } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 699, + 493, + 712 + ], + "score": 1.0, + "content": "conditioned on", + "type": "text" + }, + { + "bbox": [ + 494, + 700, + 504, + 710 + ], + "score": 0.86, + "content": "h _ { i }", + "type": "inline_equation" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 277, + 722 + ], + "score": 1.0, + "content": "as well as the latent feature representation", + "type": "text" + }, + { + "bbox": [ + 277, + 714, + 286, + 721 + ], + "score": 0.84, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 709, + 505, + 722 + ], + "score": 1.0, + "content": ", and thus can capture the fine-grained correspondence", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 721, + 366, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 366, + 734 + ], + "score": 1.0, + "content": "between music feature sequence and dance movement sequence.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44.5, + "bbox_fs": [ + 103, + 664, + 507, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 82, + 353, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 354, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 354, + 95 + ], + "score": 1.0, + "content": "3.3 DYNAMIC AUTO-CONDITION LEARNING APPROACH", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 102, + 506, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 102, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 505, + 116 + ], + "score": 1.0, + "content": "Exposure bias (Ranzato et al., 2015; Lamb et al., 2016; Zhang et al., 2019) is a notorious issue", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 127 + ], + "score": 1.0, + "content": "in natural language generation (NLG), which refers to the train-test discrepancy of autoregressive", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 125, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 506, + 137 + ], + "score": 1.0, + "content": "generative models. This kind of models are never exposed to their own prediction errors at the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 137, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 506, + 148 + ], + "score": 1.0, + "content": "training phase due to the conventional teacher-forcing training scheme. However, compared to NLG,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 147, + 506, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 506, + 159 + ], + "score": 1.0, + "content": "motion sequence generation suffers from the more severe exposure bias problem. In NLG, the bias of", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 158, + 507, + 171 + ], + "spans": [ + { + "bbox": [ + 104, + 158, + 507, + 171 + ], + "score": 1.0, + "content": "predicted probability distribution over vocabulary can be corrected by sampling strategy. For instance,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 169, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 506, + 181 + ], + "score": 1.0, + "content": "we can still generate the target token with index 2 by sampling on probability distribution [0.3, 0.3,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 178, + 506, + 193 + ], + "spans": [ + { + "bbox": [ + 104, + 178, + 506, + 193 + ], + "score": 1.0, + "content": "0.4] whose groundtruth is [0, 0, 1]. While any small biases of the predicted motion at each time-step", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 190, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 506, + 204 + ], + "score": 1.0, + "content": "will be accumulated and propagated to the future, since each generated motion is a real-valued vector", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "(rather than a discrete token) representing 2D or 3D body keyjoints in the continuous space. As a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 213, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 505, + 225 + ], + "score": 1.0, + "content": "result, the generated motion sequences tend to quickly freeze within a few seconds. This problem", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 223, + 503, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 503, + 236 + ], + "score": 1.0, + "content": "becomes even worse in the generation of long motion sequences, e.g., more than 1000 movements.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 240, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "Scheduled sampling (Bengio et al., 2015) is proposed to alleviate exposure bias in NLG, which does", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 251, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 265 + ], + "score": 1.0, + "content": "not work in long motion sequence generation. Since its sampling-based strategy would feed long", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 263, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 506, + 275 + ], + "score": 1.0, + "content": "predicted sub-sequences of high bias into the model at early stage, which causes gradient vanishing at", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 272, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 506, + 287 + ], + "score": 1.0, + "content": "training due to error accumulation of these high-biased sub-sequences. Motivated by this, we propose", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 284, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 505, + 298 + ], + "score": 1.0, + "content": "a dynamic auto-condition training method as a novel curriculum learning strategy to alleviate the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "error accumulation of autoregressive models in long motion sequence generation. Specifically, our", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "score": 1.0, + "content": "learning strategy gently changes the training process from a fully guided teacher-forcing scheme", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "using the previous ground-truth movements, towards a less guided autoregressive scheme mostly", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 327, + 487, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 487, + 343 + ], + "score": 1.0, + "content": "using the generated movements instead, which bridges the gap between training and inference.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 344, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 104, + 343, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 104, + 343, + 293, + 360 + ], + "score": 1.0, + "content": "As shown in Figure 1, The decoder predicts", + "type": "text" + }, + { + "bbox": [ + 293, + 345, + 480, + 357 + ], + "score": 0.87, + "content": "Y _ { t g t } = \\{ y _ { 1 } , . . . , y _ { p } , y _ { p + 1 } , . . . , y _ { p + q } , y _ { p + q + 1 } , . . . \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 343, + 506, + 360 + ], + "score": 1.0, + "content": "from", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 354, + 507, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 302, + 368 + ], + "score": 0.84, + "content": "Y _ { i n } = \\{ y _ { 0 } , \\hat { y _ { 1 } } , . . . , \\hat { y _ { p } } , y _ { p + 1 } , . . . , y _ { p + q } , \\hat { y } _ { p + q + 1 } , . . . \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 354, + 507, + 371 + ], + "score": 1.0, + "content": "which alternates the predicted sub-sequences with", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 366, + 504, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 133, + 381 + ], + "score": 1.0, + "content": "length", + "type": "text" + }, + { + "bbox": [ + 134, + 369, + 140, + 379 + ], + "score": 0.78, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 366, + 164, + 381 + ], + "score": 1.0, + "content": "(e.g.,", + "type": "text" + }, + { + "bbox": [ + 164, + 368, + 204, + 379 + ], + "score": 0.89, + "content": "\\hat { y } _ { 1 } , . . . , \\hat { y } _ { p } )", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 366, + 398, + 381 + ], + "score": 1.0, + "content": "and the ground-truth sub-sequences with length", + "type": "text" + }, + { + "bbox": [ + 398, + 369, + 405, + 379 + ], + "score": 0.78, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 366, + 428, + 381 + ], + "score": 1.0, + "content": "(e.g.,", + "type": "text" + }, + { + "bbox": [ + 428, + 368, + 489, + 380 + ], + "score": 0.88, + "content": "y _ { p + 1 } , . . . , y _ { { p + q } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 366, + 493, + 381 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 493, + 369, + 504, + 379 + ], + "score": 0.31, + "content": "y _ { 0 }", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 377, + 504, + 392 + ], + "spans": [ + { + "bbox": [ + 104, + 377, + 395, + 392 + ], + "score": 1.0, + "content": "is a predefined begin of pose (BOP). During the training phase, we fix", + "type": "text" + }, + { + "bbox": [ + 396, + 380, + 402, + 389 + ], + "score": 0.8, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 377, + 497, + 392 + ], + "score": 1.0, + "content": "and gradually increase", + "type": "text" + }, + { + "bbox": [ + 498, + 381, + 504, + 390 + ], + "score": 0.77, + "content": "p", + "type": "inline_equation" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 230, + 402 + ], + "score": 1.0, + "content": "according to a growth function", + "type": "text" + }, + { + "bbox": [ + 230, + 388, + 372, + 401 + ], + "score": 0.82, + "content": "f ( t ) \\in \\{ c o n s t , \\lfloor \\lambda t \\rfloor , \\lfloor \\lambda t ^ { 2 } \\rfloor , \\lfloor \\lambda e ^ { t } \\rfloor \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 388, + 399, + 402 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 399, + 390, + 405, + 399 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "is the number of training", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 153, + 412 + ], + "score": 1.0, + "content": "epochs and", + "type": "text" + }, + { + "bbox": [ + 154, + 400, + 179, + 410 + ], + "score": 0.9, + "content": "\\lambda < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 400, + 276, + 412 + ], + "score": 1.0, + "content": "controls how frequently", + "type": "text" + }, + { + "bbox": [ + 277, + 402, + 284, + 411 + ], + "score": 0.82, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 400, + 376, + 412 + ], + "score": 1.0, + "content": "is updated. We choose", + "type": "text" + }, + { + "bbox": [ + 376, + 400, + 426, + 412 + ], + "score": 0.92, + "content": "f ( t ) = \\lfloor \\lambda t \\rfloor", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 400, + 505, + 412 + ], + "score": 1.0, + "content": "in practice through", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 410, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 506, + 424 + ], + "score": 1.0, + "content": "the empirical comparison. Note that our learning strategy degenerates to the method proposed in Li", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 421, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 104, + 421, + 180, + 435 + ], + "score": 1.0, + "content": "et al. (2017) when", + "type": "text" + }, + { + "bbox": [ + 180, + 422, + 235, + 434 + ], + "score": 0.9, + "content": "f ( t ) = c o n s t", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 421, + 506, + 435 + ], + "score": 1.0, + "content": ". While the advantage of a dynamic strategy over a static strategy lies", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 432, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 506, + 446 + ], + "score": 1.0, + "content": "in that the former introduces the curriculum learning spirit to dynamically increase the difficulty of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 199, + 457 + ], + "score": 1.0, + "content": "curriculum (e.g., larger", + "type": "text" + }, + { + "bbox": [ + 199, + 446, + 205, + 455 + ], + "score": 0.71, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 444, + 505, + 457 + ], + "score": 1.0, + "content": ") as the model capacity improves during training, which can further improve", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 454, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 505, + 467 + ], + "score": 1.0, + "content": "the model capacity and alleviate the error accumulation of autoregressive predictions. Finally, we", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 466, + 347, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 199, + 478 + ], + "score": 1.0, + "content": "estimate parameters of", + "type": "text" + }, + { + "bbox": [ + 200, + 466, + 216, + 478 + ], + "score": 0.91, + "content": "g ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 466, + 292, + 478 + ], + "score": 1.0, + "content": "by minimizing the", + "type": "text" + }, + { + "bbox": [ + 293, + 466, + 303, + 477 + ], + "score": 0.87, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 466, + 334, + 478 + ], + "score": 1.0, + "content": "loss on", + "type": "text" + }, + { + "bbox": [ + 335, + 466, + 343, + 475 + ], + "score": 0.82, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 466, + 347, + 478 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 27.5 + }, + { + "type": "interline_equation", + "bbox": [ + 243, + 483, + 367, + 517 + ], + "lines": [ + { + "bbox": [ + 243, + 483, + 367, + 517 + ], + "spans": [ + { + "bbox": [ + 243, + 483, + 367, + 517 + ], + "score": 0.95, + "content": "\\ell _ { 1 } = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\left| \\left| g ( X _ { i } ) - Y _ { t g t } ^ { ( i ) } \\right| \\right| _ { 1 }", + "type": "interline_equation", + "image_path": "3096d0a2855285191c1b6add72da2f174905654e997f8cf923920c0b29a39611.jpg" + } + ] + } + ], + "index": 34.5, + "virtual_lines": [ + { + "bbox": [ + 243, + 483, + 367, + 500.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 243, + 500.0, + 367, + 517.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "title", + "bbox": [ + 108, + 529, + 244, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 528, + 245, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 245, + 545 + ], + "score": 1.0, + "content": "4 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "title", + "bbox": [ + 107, + 554, + 324, + 566 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 325, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 325, + 567 + ], + "score": 1.0, + "content": "4.1 DATASET COLLECTION AND PREPROCESSING", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 575, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "Although the dataset (Lee et al., 2019) contains about 70-hour video clips, the average length of each", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "score": 1.0, + "content": "clip is only about 6 seconds. Moreover, their dataset cannot be used by other methods to generate long-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 598, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 506, + 610 + ], + "score": 1.0, + "content": "term dance except for their specially designed decomposition-to-composition framework. To facilitate", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 609, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 506, + 622 + ], + "score": 1.0, + "content": "the task of long-term dance generation with music, we collect a high-quality dataset consisting of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "score": 1.0, + "content": "790 one-minute clips of pose data with 15 FPS, totaling about 12 hours. And there are three styles in", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 630, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 505, + 643 + ], + "score": 1.0, + "content": "the dataset including “Ballet”, “Hiphop” and “Japanese Pop”. Table 1 shows the statistics of our", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 642, + 502, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 364, + 654 + ], + "score": 1.0, + "content": "dataset. In the experiment, we randomly split the dataset into the", + "type": "text" + }, + { + "bbox": [ + 365, + 642, + 384, + 652 + ], + "score": 0.86, + "content": "90 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 642, + 449, + 654 + ], + "score": 1.0, + "content": "training set and", + "type": "text" + }, + { + "bbox": [ + 450, + 642, + 469, + 652 + ], + "score": 0.86, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 642, + 502, + 654 + ], + "score": 1.0, + "content": "test set.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 658, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 659, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 505, + 671 + ], + "score": 1.0, + "content": "Pose Preprocess. For the human pose estimation, we leverage OpenPose (Cao et al., 2017) to", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "score": 1.0, + "content": "extract 2D body keyjoints from videos. Each pose consists of 25 keyjoints2 and is represented by a", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 680, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 505, + 694 + ], + "score": 1.0, + "content": "50-dimension vector in the continuous space. In practice, we develop a linear interpolation algorithm", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 692, + 471, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 471, + 704 + ], + "score": 1.0, + "content": "to find the missing keyjoints from nearby frames to reduce the noise in extracted pose data.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 712, + 496, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 710, + 497, + 723 + ], + "spans": [ + { + "bbox": [ + 118, + 710, + 497, + 723 + ], + "score": 1.0, + "content": "2https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 721, + 326, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 326, + 733 + ], + "score": 1.0, + "content": "doc/output.md#pose-output-format-body_25", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 82, + 353, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 354, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 354, + 95 + ], + "score": 1.0, + "content": "3.3 DYNAMIC AUTO-CONDITION LEARNING APPROACH", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 102, + 506, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 102, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 505, + 116 + ], + "score": 1.0, + "content": "Exposure bias (Ranzato et al., 2015; Lamb et al., 2016; Zhang et al., 2019) is a notorious issue", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 127 + ], + "score": 1.0, + "content": "in natural language generation (NLG), which refers to the train-test discrepancy of autoregressive", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 125, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 506, + 137 + ], + "score": 1.0, + "content": "generative models. This kind of models are never exposed to their own prediction errors at the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 137, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 506, + 148 + ], + "score": 1.0, + "content": "training phase due to the conventional teacher-forcing training scheme. However, compared to NLG,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 147, + 506, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 506, + 159 + ], + "score": 1.0, + "content": "motion sequence generation suffers from the more severe exposure bias problem. In NLG, the bias of", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 158, + 507, + 171 + ], + "spans": [ + { + "bbox": [ + 104, + 158, + 507, + 171 + ], + "score": 1.0, + "content": "predicted probability distribution over vocabulary can be corrected by sampling strategy. For instance,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 169, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 506, + 181 + ], + "score": 1.0, + "content": "we can still generate the target token with index 2 by sampling on probability distribution [0.3, 0.3,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 178, + 506, + 193 + ], + "spans": [ + { + "bbox": [ + 104, + 178, + 506, + 193 + ], + "score": 1.0, + "content": "0.4] whose groundtruth is [0, 0, 1]. While any small biases of the predicted motion at each time-step", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 190, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 506, + 204 + ], + "score": 1.0, + "content": "will be accumulated and propagated to the future, since each generated motion is a real-valued vector", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "(rather than a discrete token) representing 2D or 3D body keyjoints in the continuous space. As a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 213, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 505, + 225 + ], + "score": 1.0, + "content": "result, the generated motion sequences tend to quickly freeze within a few seconds. This problem", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 223, + 503, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 503, + 236 + ], + "score": 1.0, + "content": "becomes even worse in the generation of long motion sequences, e.g., more than 1000 movements.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 6.5, + "bbox_fs": [ + 104, + 102, + 507, + 236 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 240, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "Scheduled sampling (Bengio et al., 2015) is proposed to alleviate exposure bias in NLG, which does", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 251, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 265 + ], + "score": 1.0, + "content": "not work in long motion sequence generation. Since its sampling-based strategy would feed long", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 263, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 506, + 275 + ], + "score": 1.0, + "content": "predicted sub-sequences of high bias into the model at early stage, which causes gradient vanishing at", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 272, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 506, + 287 + ], + "score": 1.0, + "content": "training due to error accumulation of these high-biased sub-sequences. Motivated by this, we propose", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 284, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 505, + 298 + ], + "score": 1.0, + "content": "a dynamic auto-condition training method as a novel curriculum learning strategy to alleviate the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "error accumulation of autoregressive models in long motion sequence generation. Specifically, our", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "score": 1.0, + "content": "learning strategy gently changes the training process from a fully guided teacher-forcing scheme", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "using the previous ground-truth movements, towards a less guided autoregressive scheme mostly", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 327, + 487, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 487, + 343 + ], + "score": 1.0, + "content": "using the generated movements instead, which bridges the gap between training and inference.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 240, + 506, + 343 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 344, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 104, + 343, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 104, + 343, + 293, + 360 + ], + "score": 1.0, + "content": "As shown in Figure 1, The decoder predicts", + "type": "text" + }, + { + "bbox": [ + 293, + 345, + 480, + 357 + ], + "score": 0.87, + "content": "Y _ { t g t } = \\{ y _ { 1 } , . . . , y _ { p } , y _ { p + 1 } , . . . , y _ { p + q } , y _ { p + q + 1 } , . . . \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 343, + 506, + 360 + ], + "score": 1.0, + "content": "from", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 354, + 507, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 302, + 368 + ], + "score": 0.84, + "content": "Y _ { i n } = \\{ y _ { 0 } , \\hat { y _ { 1 } } , . . . , \\hat { y _ { p } } , y _ { p + 1 } , . . . , y _ { p + q } , \\hat { y } _ { p + q + 1 } , . . . \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 354, + 507, + 371 + ], + "score": 1.0, + "content": "which alternates the predicted sub-sequences with", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 366, + 504, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 133, + 381 + ], + "score": 1.0, + "content": "length", + "type": "text" + }, + { + "bbox": [ + 134, + 369, + 140, + 379 + ], + "score": 0.78, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 366, + 164, + 381 + ], + "score": 1.0, + "content": "(e.g.,", + "type": "text" + }, + { + "bbox": [ + 164, + 368, + 204, + 379 + ], + "score": 0.89, + "content": "\\hat { y } _ { 1 } , . . . , \\hat { y } _ { p } )", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 366, + 398, + 381 + ], + "score": 1.0, + "content": "and the ground-truth sub-sequences with length", + "type": "text" + }, + { + "bbox": [ + 398, + 369, + 405, + 379 + ], + "score": 0.78, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 366, + 428, + 381 + ], + "score": 1.0, + "content": "(e.g.,", + "type": "text" + }, + { + "bbox": [ + 428, + 368, + 489, + 380 + ], + "score": 0.88, + "content": "y _ { p + 1 } , . . . , y _ { { p + q } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 366, + 493, + 381 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 493, + 369, + 504, + 379 + ], + "score": 0.31, + "content": "y _ { 0 }", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 377, + 504, + 392 + ], + "spans": [ + { + "bbox": [ + 104, + 377, + 395, + 392 + ], + "score": 1.0, + "content": "is a predefined begin of pose (BOP). During the training phase, we fix", + "type": "text" + }, + { + "bbox": [ + 396, + 380, + 402, + 389 + ], + "score": 0.8, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 377, + 497, + 392 + ], + "score": 1.0, + "content": "and gradually increase", + "type": "text" + }, + { + "bbox": [ + 498, + 381, + 504, + 390 + ], + "score": 0.77, + "content": "p", + "type": "inline_equation" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 230, + 402 + ], + "score": 1.0, + "content": "according to a growth function", + "type": "text" + }, + { + "bbox": [ + 230, + 388, + 372, + 401 + ], + "score": 0.82, + "content": "f ( t ) \\in \\{ c o n s t , \\lfloor \\lambda t \\rfloor , \\lfloor \\lambda t ^ { 2 } \\rfloor , \\lfloor \\lambda e ^ { t } \\rfloor \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 388, + 399, + 402 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 399, + 390, + 405, + 399 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "is the number of training", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 153, + 412 + ], + "score": 1.0, + "content": "epochs and", + "type": "text" + }, + { + "bbox": [ + 154, + 400, + 179, + 410 + ], + "score": 0.9, + "content": "\\lambda < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 400, + 276, + 412 + ], + "score": 1.0, + "content": "controls how frequently", + "type": "text" + }, + { + "bbox": [ + 277, + 402, + 284, + 411 + ], + "score": 0.82, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 400, + 376, + 412 + ], + "score": 1.0, + "content": "is updated. We choose", + "type": "text" + }, + { + "bbox": [ + 376, + 400, + 426, + 412 + ], + "score": 0.92, + "content": "f ( t ) = \\lfloor \\lambda t \\rfloor", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 400, + 505, + 412 + ], + "score": 1.0, + "content": "in practice through", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 410, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 506, + 424 + ], + "score": 1.0, + "content": "the empirical comparison. Note that our learning strategy degenerates to the method proposed in Li", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 421, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 104, + 421, + 180, + 435 + ], + "score": 1.0, + "content": "et al. (2017) when", + "type": "text" + }, + { + "bbox": [ + 180, + 422, + 235, + 434 + ], + "score": 0.9, + "content": "f ( t ) = c o n s t", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 421, + 506, + 435 + ], + "score": 1.0, + "content": ". While the advantage of a dynamic strategy over a static strategy lies", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 432, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 506, + 446 + ], + "score": 1.0, + "content": "in that the former introduces the curriculum learning spirit to dynamically increase the difficulty of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 199, + 457 + ], + "score": 1.0, + "content": "curriculum (e.g., larger", + "type": "text" + }, + { + "bbox": [ + 199, + 446, + 205, + 455 + ], + "score": 0.71, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 444, + 505, + 457 + ], + "score": 1.0, + "content": ") as the model capacity improves during training, which can further improve", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 454, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 505, + 467 + ], + "score": 1.0, + "content": "the model capacity and alleviate the error accumulation of autoregressive predictions. Finally, we", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 466, + 347, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 199, + 478 + ], + "score": 1.0, + "content": "estimate parameters of", + "type": "text" + }, + { + "bbox": [ + 200, + 466, + 216, + 478 + ], + "score": 0.91, + "content": "g ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 466, + 292, + 478 + ], + "score": 1.0, + "content": "by minimizing the", + "type": "text" + }, + { + "bbox": [ + 293, + 466, + 303, + 477 + ], + "score": 0.87, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 466, + 334, + 478 + ], + "score": 1.0, + "content": "loss on", + "type": "text" + }, + { + "bbox": [ + 335, + 466, + 343, + 475 + ], + "score": 0.82, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 466, + 347, + 478 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 27.5, + "bbox_fs": [ + 104, + 343, + 507, + 478 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 243, + 483, + 367, + 517 + ], + "lines": [ + { + "bbox": [ + 243, + 483, + 367, + 517 + ], + "spans": [ + { + "bbox": [ + 243, + 483, + 367, + 517 + ], + "score": 0.95, + "content": "\\ell _ { 1 } = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\left| \\left| g ( X _ { i } ) - Y _ { t g t } ^ { ( i ) } \\right| \\right| _ { 1 }", + "type": "interline_equation", + "image_path": "3096d0a2855285191c1b6add72da2f174905654e997f8cf923920c0b29a39611.jpg" + } + ] + } + ], + "index": 34.5, + "virtual_lines": [ + { + "bbox": [ + 243, + 483, + 367, + 500.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 243, + 500.0, + 367, + 517.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "title", + "bbox": [ + 108, + 529, + 244, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 528, + 245, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 245, + 545 + ], + "score": 1.0, + "content": "4 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "title", + "bbox": [ + 107, + 554, + 324, + 566 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 325, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 325, + 567 + ], + "score": 1.0, + "content": "4.1 DATASET COLLECTION AND PREPROCESSING", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 575, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "Although the dataset (Lee et al., 2019) contains about 70-hour video clips, the average length of each", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "score": 1.0, + "content": "clip is only about 6 seconds. Moreover, their dataset cannot be used by other methods to generate long-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 598, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 506, + 610 + ], + "score": 1.0, + "content": "term dance except for their specially designed decomposition-to-composition framework. To facilitate", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 609, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 506, + 622 + ], + "score": 1.0, + "content": "the task of long-term dance generation with music, we collect a high-quality dataset consisting of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "score": 1.0, + "content": "790 one-minute clips of pose data with 15 FPS, totaling about 12 hours. And there are three styles in", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 630, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 505, + 643 + ], + "score": 1.0, + "content": "the dataset including “Ballet”, “Hiphop” and “Japanese Pop”. Table 1 shows the statistics of our", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 642, + 502, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 364, + 654 + ], + "score": 1.0, + "content": "dataset. In the experiment, we randomly split the dataset into the", + "type": "text" + }, + { + "bbox": [ + 365, + 642, + 384, + 652 + ], + "score": 0.86, + "content": "90 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 642, + 449, + 654 + ], + "score": 1.0, + "content": "training set and", + "type": "text" + }, + { + "bbox": [ + 450, + 642, + 469, + 652 + ], + "score": 0.86, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 642, + 502, + 654 + ], + "score": 1.0, + "content": "test set.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 576, + 506, + 654 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 658, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 659, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 505, + 671 + ], + "score": 1.0, + "content": "Pose Preprocess. For the human pose estimation, we leverage OpenPose (Cao et al., 2017) to", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "score": 1.0, + "content": "extract 2D body keyjoints from videos. Each pose consists of 25 keyjoints2 and is represented by a", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 680, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 505, + 694 + ], + "score": 1.0, + "content": "50-dimension vector in the continuous space. In practice, we develop a linear interpolation algorithm", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 692, + 471, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 471, + 704 + ], + "score": 1.0, + "content": "to find the missing keyjoints from nearby frames to reduce the noise in extracted pose data.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 659, + 506, + 704 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 103, + 455, + 160 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 240, + 89, + 371, + 101 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 239, + 88, + 371, + 102 + ], + "spans": [ + { + "bbox": [ + 239, + 88, + 371, + 102 + ], + "score": 1.0, + "content": "Table 1: Statistics of the dataset.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 156, + 103, + 455, + 160 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 156, + 103, + 455, + 160 + ], + "spans": [ + { + "bbox": [ + 156, + 103, + 455, + 160 + ], + "score": 0.978, + "html": "
CategoryNum of ClipsClip Length (min)FPSResolution
Ballet136115720P
Hiphop298115720P
Japanese Pop356115720P
", + "type": "table", + "image_path": "2f2496b0f3b777149324c651c6b22b6b1a131d08728d71168ae9012dcb9e265e.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 156, + 103, + 455, + 122.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 156, + 122.0, + 455, + 141.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 156, + 141.0, + 455, + 160.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 107, + 181, + 505, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 180, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 505, + 196 + ], + "score": 1.0, + "content": "Audio Preprocess. Librosa (McFee et al., 2015) is a well-known audio and music analysis library", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 193, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 505, + 205 + ], + "score": 1.0, + "content": "in the music information retrieval, which provides flexible ways to extract the spectral and rhythm", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "features of audio data. Specifically, we extract the following features: mel frequency cepstral", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 214, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 281, + 228 + ], + "score": 1.0, + "content": "coefficients (MFCC), MFCC delta, constant-", + "type": "text" + }, + { + "bbox": [ + 281, + 215, + 290, + 226 + ], + "score": 0.38, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 214, + 506, + 228 + ], + "score": 1.0, + "content": "chromagram, tempogram and onset strength. To better", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 225, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 239 + ], + "score": 1.0, + "content": "capture the beat information of music, we convert onset strength into a one-hot vector as an additional", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 237, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 505, + 249 + ], + "score": 1.0, + "content": "feature to explicitly represent the beat sequence of music, i.e., beat one-hot. Finally, we concatenate", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 246, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 506, + 261 + ], + "score": 1.0, + "content": "these audio features as the representation of a music frame, as shown in Table 4 in Section 6.2. During", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 258, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 294, + 271 + ], + "score": 1.0, + "content": "extracting audio features, the sampling rate is", + "type": "text" + }, + { + "bbox": [ + 294, + 259, + 335, + 269 + ], + "score": 0.44, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 258, + 505, + 271 + ], + "score": 1.0, + "content": "and hop size is 1024. Hence, we have 15", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 269, + 389, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 389, + 282 + ], + "score": 1.0, + "content": "audio samples per second that is aligned with the 15 FPS of pose data.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 8 + }, + { + "type": "title", + "bbox": [ + 108, + 296, + 248, + 307 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 250, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 250, + 308 + ], + "score": 1.0, + "content": "4.2 IMPLEMENTATION DETAILS", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 317, + 506, + 405 + ], + "lines": [ + { + "bbox": [ + 106, + 317, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 265, + 329 + ], + "score": 1.0, + "content": "The music encoder consists of a stack of", + "type": "text" + }, + { + "bbox": [ + 265, + 318, + 294, + 327 + ], + "score": 0.9, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 317, + 506, + 329 + ], + "score": 1.0, + "content": "identical layers. Each layer has two sublayers: a local", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 329, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 218, + 340 + ], + "score": 1.0, + "content": "self-attention sublayer with", + "type": "text" + }, + { + "bbox": [ + 219, + 329, + 242, + 339 + ], + "score": 0.89, + "content": "l = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 329, + 505, + 340 + ], + "score": 1.0, + "content": "heads and a position-wise fully connected feed-forward sublayer", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 339, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 446, + 352 + ], + "score": 1.0, + "content": "with 1024 hidden units. Each head contains a scaled dot-product attention layer with", + "type": "text" + }, + { + "bbox": [ + 446, + 339, + 505, + 350 + ], + "score": 0.91, + "content": "d _ { k } = d _ { v } = 6 4", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 351, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 291, + 363 + ], + "score": 1.0, + "content": "and its receptive yield is restricted by setting", + "type": "text" + }, + { + "bbox": [ + 291, + 351, + 326, + 361 + ], + "score": 0.89, + "content": "k = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 351, + 466, + 363 + ], + "score": 1.0, + "content": ". Then we set the sequence length", + "type": "text" + }, + { + "bbox": [ + 466, + 351, + 502, + 361 + ], + "score": 0.87, + "content": "n = 9 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 351, + 506, + 363 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 360, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 255, + 374 + ], + "score": 1.0, + "content": "dimension of acoustic feature vector", + "type": "text" + }, + { + "bbox": [ + 255, + 362, + 295, + 372 + ], + "score": 0.91, + "content": "d _ { x } = 4 3 8", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 360, + 411, + 374 + ], + "score": 1.0, + "content": ", dimension of hidden vector", + "type": "text" + }, + { + "bbox": [ + 411, + 361, + 450, + 372 + ], + "score": 0.91, + "content": "d _ { z } = 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 360, + 506, + 374 + ], + "score": 1.0, + "content": ", respectively.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 371, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 284, + 384 + ], + "score": 1.0, + "content": "The dance decoder is a 3-layer LSTM with", + "type": "text" + }, + { + "bbox": [ + 284, + 372, + 329, + 383 + ], + "score": 0.92, + "content": "d _ { s } = 1 0 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 371, + 468, + 384 + ], + "score": 1.0, + "content": "and the dimension of pose vector", + "type": "text" + }, + { + "bbox": [ + 468, + 372, + 503, + 384 + ], + "score": 0.92, + "content": "d _ { y } = 5 0", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 371, + 506, + 384 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 135, + 396 + ], + "score": 1.0, + "content": "We set", + "type": "text" + }, + { + "bbox": [ + 135, + 383, + 173, + 394 + ], + "score": 0.9, + "content": "\\lambda = 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 383, + 190, + 396 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 191, + 384, + 220, + 394 + ], + "score": 0.91, + "content": "q = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 383, + 505, + 396 + ], + "score": 1.0, + "content": "for the proposed learning approach and train the model using the Adam", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 394, + 366, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 366, + 406 + ], + "score": 1.0, + "content": "optimizer with the learning rate 1e-4 on 2 NVIDIA V100 GPUs.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 107, + 421, + 180, + 432 + ], + "lines": [ + { + "bbox": [ + 105, + 420, + 181, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 181, + 433 + ], + "score": 1.0, + "content": "4.3 BASELINES", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 442, + 505, + 519 + ], + "lines": [ + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "The music-conditioned generation is an emerging task and there are few methods proposed to solve", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 452, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 466 + ], + "score": 1.0, + "content": "this problem. In our experiment, we compare our proposed approach to the following baselines:", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "score": 1.0, + "content": "(1) Dancing2Music. We use Dancing2Music (Lee et al., 2019) as the primary baseline that is the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "score": 1.0, + "content": "previous state-of-the-art; (2) Aud-MoCoGAN. Aud-MoCoGAN is an auxiliary baseline used in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "Dancing2Music, which is original from MoCoGAN (Tulyakov et al., 2018); (3) LSTM. Shlizerman", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "et al. (2018) propose a LSTM network to predict body dynamics from the audio signal. We modify", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 507, + 507, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 507, + 521 + ], + "score": 1.0, + "content": "their open-source code to take audio features of music as inputs and produce human pose sequences.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26 + }, + { + "type": "title", + "bbox": [ + 108, + 537, + 256, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 536, + 258, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 258, + 551 + ], + "score": 1.0, + "content": "5 EXPERIMENTAL RESULTS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 563, + 505, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 575 + ], + "score": 1.0, + "content": "In this section, we conduct extensive experiments to evaluate our approach and compare it to the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "aforementioned baselines. Note that we do not include comparison experiment on the dataset collect", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 585, + 338, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 338, + 597 + ], + "score": 1.0, + "content": "by Lee et al. (2019) due to the above issue in Section 4.1.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "title", + "bbox": [ + 107, + 611, + 342, + 623 + ], + "lines": [ + { + "bbox": [ + 105, + 611, + 343, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 343, + 624 + ], + "score": 1.0, + "content": "5.1 AUTOMATIC METRICS AND HUMAN EVALUATION", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "We evaluate the different methods by automatic metrics and conduct a human evaluation on motion", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "realism and smoothness of generated dance movements, as well as style consistency. Specifically, we", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "randomly sample 60 music clips from the test set to generate dances, then divide the dances generated", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "by three methods and the real dances into four pairs: (LSTM, ours), (Dancing2Music, ours), (real,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "ours) and (real, real). We invite 10 amateur dancers as the annotators to answer 3 questions for each", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "pair that is blind for them (use (A, B) pair instead): (1) Which dance is more realistic regardless of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "music? (2) Which dance is more smooth regardless of music? (3) Which dance matches the music", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "better with respect to style? Note that, we do not include Aud-MoCoGAN in the human evaluation,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 491, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 491, + 733 + ], + "score": 1.0, + "content": "since both it and Dancing2Music are GAN based methods while the latter is the state-of-the-art.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 39 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 103, + 455, + 160 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 240, + 89, + 371, + 101 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 239, + 88, + 371, + 102 + ], + "spans": [ + { + "bbox": [ + 239, + 88, + 371, + 102 + ], + "score": 1.0, + "content": "Table 1: Statistics of the dataset.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 156, + 103, + 455, + 160 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 156, + 103, + 455, + 160 + ], + "spans": [ + { + "bbox": [ + 156, + 103, + 455, + 160 + ], + "score": 0.978, + "html": "
CategoryNum of ClipsClip Length (min)FPSResolution
Ballet136115720P
Hiphop298115720P
Japanese Pop356115720P
", + "type": "table", + "image_path": "2f2496b0f3b777149324c651c6b22b6b1a131d08728d71168ae9012dcb9e265e.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 156, + 103, + 455, + 122.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 156, + 122.0, + 455, + 141.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 156, + 141.0, + 455, + 160.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 107, + 181, + 505, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 180, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 505, + 196 + ], + "score": 1.0, + "content": "Audio Preprocess. Librosa (McFee et al., 2015) is a well-known audio and music analysis library", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 193, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 505, + 205 + ], + "score": 1.0, + "content": "in the music information retrieval, which provides flexible ways to extract the spectral and rhythm", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "features of audio data. Specifically, we extract the following features: mel frequency cepstral", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 214, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 281, + 228 + ], + "score": 1.0, + "content": "coefficients (MFCC), MFCC delta, constant-", + "type": "text" + }, + { + "bbox": [ + 281, + 215, + 290, + 226 + ], + "score": 0.38, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 214, + 506, + 228 + ], + "score": 1.0, + "content": "chromagram, tempogram and onset strength. To better", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 225, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 239 + ], + "score": 1.0, + "content": "capture the beat information of music, we convert onset strength into a one-hot vector as an additional", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 237, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 505, + 249 + ], + "score": 1.0, + "content": "feature to explicitly represent the beat sequence of music, i.e., beat one-hot. Finally, we concatenate", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 246, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 506, + 261 + ], + "score": 1.0, + "content": "these audio features as the representation of a music frame, as shown in Table 4 in Section 6.2. During", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 258, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 294, + 271 + ], + "score": 1.0, + "content": "extracting audio features, the sampling rate is", + "type": "text" + }, + { + "bbox": [ + 294, + 259, + 335, + 269 + ], + "score": 0.44, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 258, + 505, + 271 + ], + "score": 1.0, + "content": "and hop size is 1024. Hence, we have 15", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 269, + 389, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 389, + 282 + ], + "score": 1.0, + "content": "audio samples per second that is aligned with the 15 FPS of pose data.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 180, + 506, + 282 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 296, + 248, + 307 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 250, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 250, + 308 + ], + "score": 1.0, + "content": "4.2 IMPLEMENTATION DETAILS", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 317, + 506, + 405 + ], + "lines": [ + { + "bbox": [ + 106, + 317, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 265, + 329 + ], + "score": 1.0, + "content": "The music encoder consists of a stack of", + "type": "text" + }, + { + "bbox": [ + 265, + 318, + 294, + 327 + ], + "score": 0.9, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 317, + 506, + 329 + ], + "score": 1.0, + "content": "identical layers. Each layer has two sublayers: a local", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 329, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 218, + 340 + ], + "score": 1.0, + "content": "self-attention sublayer with", + "type": "text" + }, + { + "bbox": [ + 219, + 329, + 242, + 339 + ], + "score": 0.89, + "content": "l = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 329, + 505, + 340 + ], + "score": 1.0, + "content": "heads and a position-wise fully connected feed-forward sublayer", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 339, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 446, + 352 + ], + "score": 1.0, + "content": "with 1024 hidden units. Each head contains a scaled dot-product attention layer with", + "type": "text" + }, + { + "bbox": [ + 446, + 339, + 505, + 350 + ], + "score": 0.91, + "content": "d _ { k } = d _ { v } = 6 4", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 351, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 291, + 363 + ], + "score": 1.0, + "content": "and its receptive yield is restricted by setting", + "type": "text" + }, + { + "bbox": [ + 291, + 351, + 326, + 361 + ], + "score": 0.89, + "content": "k = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 351, + 466, + 363 + ], + "score": 1.0, + "content": ". Then we set the sequence length", + "type": "text" + }, + { + "bbox": [ + 466, + 351, + 502, + 361 + ], + "score": 0.87, + "content": "n = 9 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 351, + 506, + 363 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 360, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 255, + 374 + ], + "score": 1.0, + "content": "dimension of acoustic feature vector", + "type": "text" + }, + { + "bbox": [ + 255, + 362, + 295, + 372 + ], + "score": 0.91, + "content": "d _ { x } = 4 3 8", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 360, + 411, + 374 + ], + "score": 1.0, + "content": ", dimension of hidden vector", + "type": "text" + }, + { + "bbox": [ + 411, + 361, + 450, + 372 + ], + "score": 0.91, + "content": "d _ { z } = 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 360, + 506, + 374 + ], + "score": 1.0, + "content": ", respectively.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 371, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 284, + 384 + ], + "score": 1.0, + "content": "The dance decoder is a 3-layer LSTM with", + "type": "text" + }, + { + "bbox": [ + 284, + 372, + 329, + 383 + ], + "score": 0.92, + "content": "d _ { s } = 1 0 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 371, + 468, + 384 + ], + "score": 1.0, + "content": "and the dimension of pose vector", + "type": "text" + }, + { + "bbox": [ + 468, + 372, + 503, + 384 + ], + "score": 0.92, + "content": "d _ { y } = 5 0", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 371, + 506, + 384 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 135, + 396 + ], + "score": 1.0, + "content": "We set", + "type": "text" + }, + { + "bbox": [ + 135, + 383, + 173, + 394 + ], + "score": 0.9, + "content": "\\lambda = 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 383, + 190, + 396 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 191, + 384, + 220, + 394 + ], + "score": 0.91, + "content": "q = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 383, + 505, + 396 + ], + "score": 1.0, + "content": "for the proposed learning approach and train the model using the Adam", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 394, + 366, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 366, + 406 + ], + "score": 1.0, + "content": "optimizer with the learning rate 1e-4 on 2 NVIDIA V100 GPUs.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 317, + 506, + 406 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 421, + 180, + 432 + ], + "lines": [ + { + "bbox": [ + 105, + 420, + 181, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 181, + 433 + ], + "score": 1.0, + "content": "4.3 BASELINES", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 442, + 505, + 519 + ], + "lines": [ + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "The music-conditioned generation is an emerging task and there are few methods proposed to solve", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 452, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 466 + ], + "score": 1.0, + "content": "this problem. In our experiment, we compare our proposed approach to the following baselines:", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "score": 1.0, + "content": "(1) Dancing2Music. We use Dancing2Music (Lee et al., 2019) as the primary baseline that is the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "score": 1.0, + "content": "previous state-of-the-art; (2) Aud-MoCoGAN. Aud-MoCoGAN is an auxiliary baseline used in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "Dancing2Music, which is original from MoCoGAN (Tulyakov et al., 2018); (3) LSTM. Shlizerman", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "et al. (2018) propose a LSTM network to predict body dynamics from the audio signal. We modify", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 507, + 507, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 507, + 521 + ], + "score": 1.0, + "content": "their open-source code to take audio features of music as inputs and produce human pose sequences.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 442, + 507, + 521 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 537, + 256, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 536, + 258, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 258, + 551 + ], + "score": 1.0, + "content": "5 EXPERIMENTAL RESULTS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 563, + 505, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 575 + ], + "score": 1.0, + "content": "In this section, we conduct extensive experiments to evaluate our approach and compare it to the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "aforementioned baselines. Note that we do not include comparison experiment on the dataset collect", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 585, + 338, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 338, + 597 + ], + "score": 1.0, + "content": "by Lee et al. (2019) due to the above issue in Section 4.1.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 563, + 505, + 597 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 611, + 342, + 623 + ], + "lines": [ + { + "bbox": [ + 105, + 611, + 343, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 343, + 624 + ], + "score": 1.0, + "content": "5.1 AUTOMATIC METRICS AND HUMAN EVALUATION", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "We evaluate the different methods by automatic metrics and conduct a human evaluation on motion", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "realism and smoothness of generated dance movements, as well as style consistency. Specifically, we", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "randomly sample 60 music clips from the test set to generate dances, then divide the dances generated", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "by three methods and the real dances into four pairs: (LSTM, ours), (Dancing2Music, ours), (real,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "ours) and (real, real). We invite 10 amateur dancers as the annotators to answer 3 questions for each", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "pair that is blind for them (use (A, B) pair instead): (1) Which dance is more realistic regardless of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "music? (2) Which dance is more smooth regardless of music? (3) Which dance matches the music", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "better with respect to style? Note that, we do not include Aud-MoCoGAN in the human evaluation,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 491, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 491, + 733 + ], + "score": 1.0, + "content": "since both it and Dancing2Music are GAN based methods while the latter is the state-of-the-art.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 633, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 175, + 505, + 378 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 105, + 175, + 505, + 378 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 175, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 378 + ], + "score": 0.495, + "type": "image", + "image_path": "160f9ed77699ffb31249fa04fe6af19e09624d27fa86094970947262a6f5633e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 105, + 175, + 505, + 242.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 105, + 242.66666666666669, + 505, + 310.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 105, + 310.33333333333337, + 505, + 378.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 387, + 505, + 421 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "Figure 2: Human evaluation results on motion realism, style consistency and smoothness. We", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "conduct a human evaluation to ask annotators to select the dances that are more realistic and more", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 409, + 472, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 472, + 423 + ], + "score": 1.0, + "content": "smooth regardless of music, more style-consistent with music through pairwise comparison.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 434, + 505, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "Motion Realism, Style Consistency and Smoothness. We evaluate the visual quality and realism", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 446, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 457 + ], + "score": 1.0, + "content": "of generated dances by Fréchet Inception Distance (FID) (Heusel et al., 2017), which is used to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "score": 1.0, + "content": "measure how close the distribution of generated dances is to the real. Similar to Lee et al. (2019), we", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 468, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 506, + 479 + ], + "score": 1.0, + "content": "train a style classifier on dance movement sequences of three styles and use it to extract features for", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "the given dances. Then we calculate FID between the synthesized dances and the real. As shown in", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "Table 2, our FID score is significantly lower than those of baselines and much closer to that of the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 500, + 507, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 507, + 513 + ], + "score": 1.0, + "content": "real, which means our generated dances are more motion-realistic and more likely to the real dances.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 511, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 506, + 523 + ], + "score": 1.0, + "content": "Besides, we use the same classifier to measure the style accuracy of generated dance to the music,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 522, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 213, + 535 + ], + "score": 1.0, + "content": "and our approach achieves", + "type": "text" + }, + { + "bbox": [ + 213, + 522, + 241, + 533 + ], + "score": 0.85, + "content": "7 7 . 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 522, + 476, + 535 + ], + "score": 1.0, + "content": "accuracy and significantly outperforms Dancing2Music by", + "type": "text" + }, + { + "bbox": [ + 477, + 522, + 503, + 533 + ], + "score": 0.84, + "content": "1 7 . 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 522, + 506, + 535 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 539, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "score": 1.0, + "content": "Human evaluation in Figure 2 consistently shows the superior performance of our approach, compared", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "to baselines on motion realism, style consistency and smoothness. We observe the dances generated", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "by LSTM have lots of floating movements and tend to quickly freeze within several seconds due", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "to the error accumulation, which results in lower preferences. While Dancing2Music can generate", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "the smooth dance units, the synthesized dances still have significant jumps where dance units are", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 595, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 506, + 606 + ], + "score": 1.0, + "content": "connected due to the inherent gap between them, which harm its overall performance in practice.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 439, + 618 + ], + "score": 1.0, + "content": "This is also reflected in the preference comparisons to our approach, in which only", + "type": "text" + }, + { + "bbox": [ + 440, + 605, + 460, + 615 + ], + "score": 0.86, + "content": "3 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "annotators", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 288, + 629 + ], + "score": 1.0, + "content": "prefer Dancing2Music on motion realism and", + "type": "text" + }, + { + "bbox": [ + 288, + 616, + 315, + 627 + ], + "score": 0.86, + "content": "2 1 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 616, + 506, + 629 + ], + "score": 1.0, + "content": "prefer Dancing2Music on smoothness. Besides,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "the result on style consistency shows our approach can generate more style-consistent dances with", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 639, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 505, + 650 + ], + "score": 1.0, + "content": "musics. Since our approach models the fine-grained correspondence between music and dance at the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "methodological level and also introduces more specific music features. However, Dancing2Music", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 660, + 492, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 492, + 673 + ], + "score": 1.0, + "content": "only considers a global style feature extracted by a pre-trained classifier in the dance generation.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 237, + 689 + ], + "score": 1.0, + "content": "In the comparison to real dances,", + "type": "text" + }, + { + "bbox": [ + 237, + 677, + 264, + 687 + ], + "score": 0.87, + "content": "4 1 . 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 676, + 478, + 689 + ], + "score": 1.0, + "content": "annotators prefer our method on motion realism while", + "type": "text" + }, + { + "bbox": [ + 478, + 677, + 505, + 687 + ], + "score": 0.86, + "content": "3 0 . 3 \\%", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 335, + 700 + ], + "score": 1.0, + "content": "prefer on style consistency. Additionally, we found that", + "type": "text" + }, + { + "bbox": [ + 335, + 688, + 363, + 698 + ], + "score": 0.88, + "content": "5 7 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "annotators prefer our approach on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "smoothness compared to real dances. Since the imperfect OpenPose (Cao et al., 2017) introduces the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "minor noise on pose data extracted from dance videos. On the other hand, in the preprocessing, we", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "develop an interpolation algorithm to find missing keyjoints and reduce the noise of training pose data.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 175, + 505, + 378 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 105, + 175, + 505, + 378 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 175, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 378 + ], + "score": 0.495, + "type": "image", + "image_path": "160f9ed77699ffb31249fa04fe6af19e09624d27fa86094970947262a6f5633e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 105, + 175, + 505, + 242.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 105, + 242.66666666666669, + 505, + 310.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 105, + 310.33333333333337, + 505, + 378.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 387, + 505, + 421 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "Figure 2: Human evaluation results on motion realism, style consistency and smoothness. We", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "conduct a human evaluation to ask annotators to select the dances that are more realistic and more", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 409, + 472, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 472, + 423 + ], + "score": 1.0, + "content": "smooth regardless of music, more style-consistent with music through pairwise comparison.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 434, + 505, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "Motion Realism, Style Consistency and Smoothness. We evaluate the visual quality and realism", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 446, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 457 + ], + "score": 1.0, + "content": "of generated dances by Fréchet Inception Distance (FID) (Heusel et al., 2017), which is used to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "score": 1.0, + "content": "measure how close the distribution of generated dances is to the real. Similar to Lee et al. (2019), we", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 468, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 506, + 479 + ], + "score": 1.0, + "content": "train a style classifier on dance movement sequences of three styles and use it to extract features for", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "the given dances. Then we calculate FID between the synthesized dances and the real. As shown in", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "Table 2, our FID score is significantly lower than those of baselines and much closer to that of the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 500, + 507, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 507, + 513 + ], + "score": 1.0, + "content": "real, which means our generated dances are more motion-realistic and more likely to the real dances.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 511, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 506, + 523 + ], + "score": 1.0, + "content": "Besides, we use the same classifier to measure the style accuracy of generated dance to the music,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 522, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 213, + 535 + ], + "score": 1.0, + "content": "and our approach achieves", + "type": "text" + }, + { + "bbox": [ + 213, + 522, + 241, + 533 + ], + "score": 0.85, + "content": "7 7 . 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 522, + 476, + 535 + ], + "score": 1.0, + "content": "accuracy and significantly outperforms Dancing2Music by", + "type": "text" + }, + { + "bbox": [ + 477, + 522, + 503, + 533 + ], + "score": 0.84, + "content": "1 7 . 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 522, + 506, + 535 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 434, + 507, + 535 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 539, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "score": 1.0, + "content": "Human evaluation in Figure 2 consistently shows the superior performance of our approach, compared", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "to baselines on motion realism, style consistency and smoothness. We observe the dances generated", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "by LSTM have lots of floating movements and tend to quickly freeze within several seconds due", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "to the error accumulation, which results in lower preferences. While Dancing2Music can generate", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "the smooth dance units, the synthesized dances still have significant jumps where dance units are", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 595, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 506, + 606 + ], + "score": 1.0, + "content": "connected due to the inherent gap between them, which harm its overall performance in practice.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 439, + 618 + ], + "score": 1.0, + "content": "This is also reflected in the preference comparisons to our approach, in which only", + "type": "text" + }, + { + "bbox": [ + 440, + 605, + 460, + 615 + ], + "score": 0.86, + "content": "3 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "annotators", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 288, + 629 + ], + "score": 1.0, + "content": "prefer Dancing2Music on motion realism and", + "type": "text" + }, + { + "bbox": [ + 288, + 616, + 315, + 627 + ], + "score": 0.86, + "content": "2 1 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 616, + 506, + 629 + ], + "score": 1.0, + "content": "prefer Dancing2Music on smoothness. Besides,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "the result on style consistency shows our approach can generate more style-consistent dances with", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 639, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 505, + 650 + ], + "score": 1.0, + "content": "musics. Since our approach models the fine-grained correspondence between music and dance at the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "methodological level and also introduces more specific music features. However, Dancing2Music", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 660, + 492, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 492, + 673 + ], + "score": 1.0, + "content": "only considers a global style feature extracted by a pre-trained classifier in the dance generation.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 539, + 506, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 237, + 689 + ], + "score": 1.0, + "content": "In the comparison to real dances,", + "type": "text" + }, + { + "bbox": [ + 237, + 677, + 264, + 687 + ], + "score": 0.87, + "content": "4 1 . 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 676, + 478, + 689 + ], + "score": 1.0, + "content": "annotators prefer our method on motion realism while", + "type": "text" + }, + { + "bbox": [ + 478, + 677, + 505, + 687 + ], + "score": 0.86, + "content": "3 0 . 3 \\%", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 335, + 700 + ], + "score": 1.0, + "content": "prefer on style consistency. Additionally, we found that", + "type": "text" + }, + { + "bbox": [ + 335, + 688, + 363, + 698 + ], + "score": 0.88, + "content": "5 7 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "annotators prefer our approach on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "smoothness compared to real dances. Since the imperfect OpenPose (Cao et al., 2017) introduces the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "minor noise on pose data extracted from dance videos. On the other hand, in the preprocessing, we", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "develop an interpolation algorithm to find missing keyjoints and reduce the noise of training pose data.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 676, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Besides, this also indicates the chain structure based decoder can well capture the spatial-temporal", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 436, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 436, + 106 + ], + "score": 1.0, + "content": "dependency among human motion dynamics and produce the smooth movements.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 110, + 505, + 264 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 505, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 505, + 122 + ], + "score": 1.0, + "content": "Beat Coverage and Hit Rate. We also evaluate the beat coverage and hit rate introduced in Lee", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 504, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 309, + 134 + ], + "score": 1.0, + "content": "et al. (2019), which defines the beat coverage as", + "type": "text" + }, + { + "bbox": [ + 309, + 121, + 343, + 133 + ], + "score": 0.93, + "content": "B _ { k } / B _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 120, + 426, + 134 + ], + "score": 1.0, + "content": "and beat hit rate as", + "type": "text" + }, + { + "bbox": [ + 426, + 121, + 457, + 133 + ], + "score": 0.92, + "content": "B _ { a } / B _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 120, + 490, + 134 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 490, + 122, + 504, + 132 + ], + "score": 0.88, + "content": "B _ { k }", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 249, + 145 + ], + "score": 1.0, + "content": "is the number of kinematic beats,", + "type": "text" + }, + { + "bbox": [ + 249, + 133, + 266, + 144 + ], + "score": 0.9, + "content": "B _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 132, + 417, + 145 + ], + "score": 1.0, + "content": "is the number of musical beats and", + "type": "text" + }, + { + "bbox": [ + 418, + 133, + 431, + 144 + ], + "score": 0.89, + "content": "B _ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "is the number of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "kinematic beats that are aligned with the musical beats. An early study (Ho et al., 2013) reveals that", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "kinematic beat frames occur when the direction of the movement changes. Therefore, similar to Yalta", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 177 + ], + "score": 1.0, + "content": "et al. (2019), we use the standard deviation (SD) of movement to detect the kinematic beats in our", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "experiment. The onset strength (Ellis, 2007) is a common way to detect musical beats. Figure 3", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "shows two short clips of motion SD curves and aligned musical beats. We observed that the kinematic", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 199, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 210 + ], + "score": 1.0, + "content": "beats occur where the musical beats occur, which is consistent with the common sense that dancers", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 506, + 222 + ], + "score": 1.0, + "content": "would step on musical beats during dancing but do not step on every musical beat. As we can see in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 218, + 506, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 506, + 233 + ], + "score": 1.0, + "content": "Table 2, LSTM and Aud-MoCoGAN generate dances with few kinematic beats and most of them do", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 409, + 244 + ], + "score": 1.0, + "content": "not match the musical beats. Our method outperforms Dancing2Music by", + "type": "text" + }, + { + "bbox": [ + 410, + 231, + 433, + 242 + ], + "score": 0.87, + "content": "6 . 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 230, + 505, + 244 + ], + "score": 1.0, + "content": "on beat coverage", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 242, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 123, + 254 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 242, + 146, + 253 + ], + "score": 0.86, + "content": "2 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 242, + 506, + 254 + ], + "score": 1.0, + "content": "on hit rate, which indicates that introducing the features about musical beat into model is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 253, + 405, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 405, + 265 + ], + "score": 1.0, + "content": "beneficial to the better beat alignment of generated dance and input music.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 8.5 + }, + { + "type": "image", + "bbox": [ + 124, + 275, + 483, + 364 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 275, + 483, + 364 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 275, + 483, + 364 + ], + "spans": [ + { + "bbox": [ + 124, + 275, + 483, + 364 + ], + "score": 0.97, + "type": "image", + "image_path": "167107ef7ff1cf1ce81b167922e4cc91ebfc7d4110451112c0a7a53b8c70d63b.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 124, + 275, + 483, + 304.6666666666667 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 124, + 304.6666666666667, + 483, + 334.33333333333337 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 124, + 334.33333333333337, + 483, + 364.00000000000006 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 375, + 505, + 408 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 375, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 505, + 388 + ], + "score": 1.0, + "content": "Figure 3: Beat tracking curves for music and dance by onset strength and motion standard deviation", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 386, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 505, + 399 + ], + "score": 1.0, + "content": "respectively. The left is a short clip of ballet while the right is a short clip of hip-hop. The red circles", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 397, + 348, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 348, + 410 + ], + "score": 1.0, + "content": "are kinematic beats and dash lines denote the musical beats.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 428, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "Diversity and Multimodality. Lee et al. (2019) introduce the diversity metric and the multimodality", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "metric. The diversity metric evaluates the variations among generated dances corresponding to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "various music, which reflects the generalization ability of the model and its dependency on the input", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 460, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 474 + ], + "score": 1.0, + "content": "music. Similarly, we use the average feature distance as the measurement and these features are", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 484 + ], + "score": 1.0, + "content": "extracted by the same classifier used in measuring FID. We randomly sample 60 music clips from", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "test set to generate dances and randomly pick 500 combinations to compute the average feature", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "distance. Results in Table 2 show the performance of our approach is superior to Dancing2Music on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "diversity. The reason is that Dancing2Music only considers a global music style feature in generation", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "while different music of the same style have the almost same style feature. Our approach models the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "fine-grained correspondence between music and dance at the methodological level, simultaneously", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "takes more specific music features into account. In other words, our approach is more dependent on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 549, + 491, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 491, + 561 + ], + "score": 1.0, + "content": "input music and thus can generate different samples for difference music clips of the same style.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 565, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 506, + 578 + ], + "score": 1.0, + "content": "Multimodality metric evaluates the variations of generated dances for the same music clip. Specifi-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 574, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 591 + ], + "score": 1.0, + "content": "cally, we generate 5 dances for each of randomly sampled 60 music clips and compute the average", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "score": 1.0, + "content": "feature distance. As we can see, our approach slightly underperforms Dancing2Music and Aud-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "score": 1.0, + "content": "MoCoGAN since it has more dependency on music inputs as aforementioned, which make the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 609, + 423, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 423, + 622 + ], + "score": 1.0, + "content": "generated dances have the limited variation patterns given the same music clip.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + }, + { + "type": "title", + "bbox": [ + 108, + 634, + 279, + 645 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 280, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 280, + 646 + ], + "score": 1.0, + "content": "5.2 LONG-TERM DANCE GENERATION", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "To further investigate the performance of different methods on long-term dance generation, we", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "evaluate FID of dances generated by different methods over time. Specifically, we first split a", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "generated dance with 1 minute into 15 four-second clips and measure FID of each clip. Figure 4 shows", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "FID scores of LSTM and Aud-MoCoGAN grow rapidly in the early stage due to error accumulation", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "and converge to the high FID, since the generated dances quickly become frozen. Dancing2Music", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 710, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 721 + ], + "score": 1.0, + "content": "maintains the relatively stable FID scores all the time, which benefits from its decomposition-to-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "composition method. While its curve still has subtle fluctuations since it synthesizes whole dances by", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Besides, this also indicates the chain structure based decoder can well capture the spatial-temporal", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 436, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 436, + 106 + ], + "score": 1.0, + "content": "dependency among human motion dynamics and produce the smooth movements.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 110, + 505, + 264 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 505, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 505, + 122 + ], + "score": 1.0, + "content": "Beat Coverage and Hit Rate. We also evaluate the beat coverage and hit rate introduced in Lee", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 504, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 309, + 134 + ], + "score": 1.0, + "content": "et al. (2019), which defines the beat coverage as", + "type": "text" + }, + { + "bbox": [ + 309, + 121, + 343, + 133 + ], + "score": 0.93, + "content": "B _ { k } / B _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 120, + 426, + 134 + ], + "score": 1.0, + "content": "and beat hit rate as", + "type": "text" + }, + { + "bbox": [ + 426, + 121, + 457, + 133 + ], + "score": 0.92, + "content": "B _ { a } / B _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 120, + 490, + 134 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 490, + 122, + 504, + 132 + ], + "score": 0.88, + "content": "B _ { k }", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 249, + 145 + ], + "score": 1.0, + "content": "is the number of kinematic beats,", + "type": "text" + }, + { + "bbox": [ + 249, + 133, + 266, + 144 + ], + "score": 0.9, + "content": "B _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 132, + 417, + 145 + ], + "score": 1.0, + "content": "is the number of musical beats and", + "type": "text" + }, + { + "bbox": [ + 418, + 133, + 431, + 144 + ], + "score": 0.89, + "content": "B _ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "is the number of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "kinematic beats that are aligned with the musical beats. An early study (Ho et al., 2013) reveals that", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "kinematic beat frames occur when the direction of the movement changes. Therefore, similar to Yalta", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 177 + ], + "score": 1.0, + "content": "et al. (2019), we use the standard deviation (SD) of movement to detect the kinematic beats in our", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "experiment. The onset strength (Ellis, 2007) is a common way to detect musical beats. Figure 3", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "shows two short clips of motion SD curves and aligned musical beats. We observed that the kinematic", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 199, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 210 + ], + "score": 1.0, + "content": "beats occur where the musical beats occur, which is consistent with the common sense that dancers", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 506, + 222 + ], + "score": 1.0, + "content": "would step on musical beats during dancing but do not step on every musical beat. As we can see in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 218, + 506, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 506, + 233 + ], + "score": 1.0, + "content": "Table 2, LSTM and Aud-MoCoGAN generate dances with few kinematic beats and most of them do", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 409, + 244 + ], + "score": 1.0, + "content": "not match the musical beats. Our method outperforms Dancing2Music by", + "type": "text" + }, + { + "bbox": [ + 410, + 231, + 433, + 242 + ], + "score": 0.87, + "content": "6 . 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 230, + 505, + 244 + ], + "score": 1.0, + "content": "on beat coverage", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 242, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 123, + 254 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 242, + 146, + 253 + ], + "score": 0.86, + "content": "2 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 242, + 506, + 254 + ], + "score": 1.0, + "content": "on hit rate, which indicates that introducing the features about musical beat into model is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 253, + 405, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 405, + 265 + ], + "score": 1.0, + "content": "beneficial to the better beat alignment of generated dance and input music.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 110, + 506, + 265 + ] + }, + { + "type": "image", + "bbox": [ + 124, + 275, + 483, + 364 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 275, + 483, + 364 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 275, + 483, + 364 + ], + "spans": [ + { + "bbox": [ + 124, + 275, + 483, + 364 + ], + "score": 0.97, + "type": "image", + "image_path": "167107ef7ff1cf1ce81b167922e4cc91ebfc7d4110451112c0a7a53b8c70d63b.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 124, + 275, + 483, + 304.6666666666667 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 124, + 304.6666666666667, + 483, + 334.33333333333337 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 124, + 334.33333333333337, + 483, + 364.00000000000006 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 375, + 505, + 408 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 375, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 505, + 388 + ], + "score": 1.0, + "content": "Figure 3: Beat tracking curves for music and dance by onset strength and motion standard deviation", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 386, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 505, + 399 + ], + "score": 1.0, + "content": "respectively. The left is a short clip of ballet while the right is a short clip of hip-hop. The red circles", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 397, + 348, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 348, + 410 + ], + "score": 1.0, + "content": "are kinematic beats and dash lines denote the musical beats.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 428, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "Diversity and Multimodality. Lee et al. (2019) introduce the diversity metric and the multimodality", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "metric. The diversity metric evaluates the variations among generated dances corresponding to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "various music, which reflects the generalization ability of the model and its dependency on the input", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 460, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 474 + ], + "score": 1.0, + "content": "music. Similarly, we use the average feature distance as the measurement and these features are", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 484 + ], + "score": 1.0, + "content": "extracted by the same classifier used in measuring FID. We randomly sample 60 music clips from", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "test set to generate dances and randomly pick 500 combinations to compute the average feature", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "distance. Results in Table 2 show the performance of our approach is superior to Dancing2Music on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "diversity. The reason is that Dancing2Music only considers a global music style feature in generation", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "while different music of the same style have the almost same style feature. Our approach models the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "fine-grained correspondence between music and dance at the methodological level, simultaneously", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "takes more specific music features into account. In other words, our approach is more dependent on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 549, + 491, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 491, + 561 + ], + "score": 1.0, + "content": "input music and thus can generate different samples for difference music clips of the same style.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 428, + 506, + 561 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 565, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 506, + 578 + ], + "score": 1.0, + "content": "Multimodality metric evaluates the variations of generated dances for the same music clip. Specifi-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 574, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 591 + ], + "score": 1.0, + "content": "cally, we generate 5 dances for each of randomly sampled 60 music clips and compute the average", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "score": 1.0, + "content": "feature distance. As we can see, our approach slightly underperforms Dancing2Music and Aud-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "score": 1.0, + "content": "MoCoGAN since it has more dependency on music inputs as aforementioned, which make the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 609, + 423, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 423, + 622 + ], + "score": 1.0, + "content": "generated dances have the limited variation patterns given the same music clip.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 565, + 506, + 622 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 634, + 279, + 645 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 280, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 280, + 646 + ], + "score": 1.0, + "content": "5.2 LONG-TERM DANCE GENERATION", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "To further investigate the performance of different methods on long-term dance generation, we", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "evaluate FID of dances generated by different methods over time. Specifically, we first split a", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "generated dance with 1 minute into 15 four-second clips and measure FID of each clip. Figure 4 shows", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "FID scores of LSTM and Aud-MoCoGAN grow rapidly in the early stage due to error accumulation", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "and converge to the high FID, since the generated dances quickly become frozen. Dancing2Music", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 710, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 721 + ], + "score": 1.0, + "content": "maintains the relatively stable FID scores all the time, which benefits from its decomposition-to-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "composition method. While its curve still has subtle fluctuations since it synthesizes whole dances by", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "composing the generated dance units. Compared to baselines, FID of our method are much lower and", + "type": "text", + "cross_page": true + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 368, + 504, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 504, + 381 + ], + "score": 1.0, + "content": "close to real dances, which validates the good performance of our method on long-term generation.", + "type": "text", + "cross_page": true + } + ], + "index": 22 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 654, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 207, + 82, + 400, + 207 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 207, + 82, + 400, + 207 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 207, + 82, + 400, + 207 + ], + "spans": [ + { + "bbox": [ + 207, + 82, + 400, + 207 + ], + "score": 0.966, + "type": "image", + "image_path": "b131c94c16a3461b63bf6187b0f782cc2dc65245cdd3adee1cfccc5add824bad.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 207, + 82, + 400, + 95.88888888888889 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 207, + 95.88888888888889, + 400, + 109.77777777777777 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 207, + 109.77777777777777, + 400, + 123.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 207, + 123.66666666666666, + 400, + 137.55555555555554 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 207, + 137.55555555555554, + 400, + 151.44444444444443 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 207, + 151.44444444444443, + 400, + 165.33333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 207, + 165.33333333333331, + 400, + 179.2222222222222 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 207, + 179.2222222222222, + 400, + 193.1111111111111 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 207, + 193.1111111111111, + 400, + 206.99999999999997 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 199, + 212, + 411, + 224 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 199, + 212, + 411, + 225 + ], + "spans": [ + { + "bbox": [ + 199, + 212, + 411, + 225 + ], + "score": 1.0, + "content": "Figure 4: FID curves of different methods over time.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + } + ], + "index": 6.5 + }, + { + "type": "table", + "bbox": [ + 124, + 269, + 285, + 333 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 124, + 269, + 285, + 333 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 269, + 285, + 333 + ], + "spans": [ + { + "bbox": [ + 124, + 269, + 285, + 333 + ], + "score": 0.966, + "html": "
Encoder Structure LSTMFIDACC (%)
7.945
ConvS2S encoder13.354
Global self-attention3.680
Local self-attention4.378.5
", + "type": "table", + "image_path": "ec6b92b6e4c9db67e824443b66fac70b206a85a4a2c1c09c95deafb57a1a89d9.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 124, + 269, + 285, + 285.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 124, + 285.0, + 285, + 301.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 124, + 301.0, + 285, + 317.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 124, + 317.0, + 285, + 333.0 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 13.5 + }, + { + "type": "table", + "bbox": [ + 307, + 264, + 475, + 337 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 232, + 505, + 256 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 232, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 505, + 245 + ], + "score": 1.0, + "content": "Table 3: Left: Automatic metrics of our model with different encoder structures. Right: Automatic", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 243, + 352, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 352, + 258 + ], + "score": 1.0, + "content": "metrics of our model training by different learning strategies.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "table_body", + "bbox": [ + 307, + 264, + 475, + 337 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 475, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 475, + 337 + ], + "score": 0.926, + "html": "
Learning ApproachFIDACC (%)
Teacher-forcing61.25.3
Auto-condition (con st)15.735
Ours ([λet J)9.869
Ours ([λt²])6.473
Ours ([λt])5.177.6
", + "type": "table", + "image_path": "93832bc96a22cc8bcbf625fa826d04271ee181ac45bbcbe9616453589df5ecae.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 307, + 264, + 475, + 278.6 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 307, + 278.6, + 475, + 293.20000000000005 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 307, + 293.20000000000005, + 475, + 307.80000000000007 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 307, + 307.80000000000007, + 475, + 322.4000000000001 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 307, + 322.4000000000001, + 475, + 337.0000000000001 + ], + "spans": [], + "index": 20 + } + ] + } + ], + "index": 14.25 + }, + { + "type": "text", + "bbox": [ + 108, + 358, + 504, + 380 + ], + "lines": [ + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "composing the generated dance units. Compared to baselines, FID of our method are much lower and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 368, + 504, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 504, + 381 + ], + "score": 1.0, + "content": "close to real dances, which validates the good performance of our method on long-term generation.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "title", + "bbox": [ + 108, + 393, + 209, + 405 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 210, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 210, + 406 + ], + "score": 1.0, + "content": "5.3 ABLATION STUDY", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 505, + 491 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "Comparison on Encoder Structures. We compare the different encoder structures in our model", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "with the same LSTM decoder and the same curriculum learning strategy, as shown in Table 3. The", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 435, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 449 + ], + "score": 1.0, + "content": "transformer encoders outperform LSTM and the encoder in ConvS2S (Gehring et al., 2017) on", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "both FID and style consistency, due to its superior performance on sequence modeling. Although", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 458, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 505, + 471 + ], + "score": 1.0, + "content": "the transformer encoder with our proposed local self-attention slightly underperforms the global", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 469, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 268, + 482 + ], + "score": 1.0, + "content": "self-attention by 0.7 higher in FID and", + "type": "text" + }, + { + "bbox": [ + 268, + 469, + 291, + 480 + ], + "score": 0.86, + "content": "1 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 469, + 506, + 482 + ], + "score": 1.0, + "content": "lower in ACC, it has much fewer parameters in our", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 479, + 485, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 485, + 494 + ], + "score": 1.0, + "content": "settings. This confirms the effectiveness of the local self-attention on long sequence modeling.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 496, + 505, + 585 + ], + "lines": [ + { + "bbox": [ + 106, + 496, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 510 + ], + "score": 1.0, + "content": "Comparison on Learning Strategies. To investigate the performance of our proposed learning", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "approach, we conduct an ablation study to compare the performances of our model using different", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "score": 1.0, + "content": "learning strategies. As we can see the right of in Table 3, all curriculum learning strategies significantly", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "outperform the static auto-condition strategy (Li et al., 2017). Among them, the linear growth function", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 541, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 107, + 541, + 156, + 553 + ], + "score": 0.92, + "content": "f ( t ) = \\lfloor \\lambda t \\rfloor", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 541, + 505, + 553 + ], + "score": 1.0, + "content": "performs best, which might be that increasing the difficulty of curriculum too fast would", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "score": 1.0, + "content": "result in the model degradation. Besides, the original teacher-forcing strategy has the highest FID", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 561, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 576 + ], + "score": 1.0, + "content": "and the lowest style accuracy due to severe error accumulation problem, which indicates that using", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 573, + 489, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 489, + 587 + ], + "score": 1.0, + "content": "our proposed curriculum learning strategy to train the model can effectively alleviate this issue.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5 + }, + { + "type": "title", + "bbox": [ + 107, + 601, + 195, + 614 + ], + "lines": [ + { + "bbox": [ + 105, + 598, + 198, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 198, + 617 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 106, + 626, + 505, + 725 + ], + "lines": [ + { + "bbox": [ + 106, + 626, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 506, + 638 + ], + "score": 1.0, + "content": "In this work, we propose a novel seq2seq architecture for long-term dance generation with music,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 104, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "e.g., about one-minute length. To efficiently process long sequences of music features, we introduce", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 646, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 104, + 646, + 506, + 662 + ], + "score": 1.0, + "content": "a local self-attention mechanism in the transformer encoder structure to reduce the quadratic memory", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "requirements to the linear in the sequence length. Besides, we also propose a dynamic auto-condition", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 669, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 683 + ], + "score": 1.0, + "content": "training method as a novel curriculum learning strategy to alleviate the error accumulation of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "autoregressive models in long motion sequence generation, thus facilitate the long-term dance", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 691, + 506, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 506, + 705 + ], + "score": 1.0, + "content": "generation. Extensive experiments have demonstrated the superior performance of our proposed", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 702, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 506, + 716 + ], + "score": 1.0, + "content": "approach on automatic metrics and human evaluation. Future works include the exploration on", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 714, + 491, + 726 + ], + "spans": [ + { + "bbox": [ + 106, + 714, + 491, + 726 + ], + "score": 1.0, + "content": "end-to-end methods that could work with raw audio data instead of preprocessed audio features.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 44 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 207, + 82, + 400, + 207 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 207, + 82, + 400, + 207 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 207, + 82, + 400, + 207 + ], + "spans": [ + { + "bbox": [ + 207, + 82, + 400, + 207 + ], + "score": 0.966, + "type": "image", + "image_path": "b131c94c16a3461b63bf6187b0f782cc2dc65245cdd3adee1cfccc5add824bad.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 207, + 82, + 400, + 95.88888888888889 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 207, + 95.88888888888889, + 400, + 109.77777777777777 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 207, + 109.77777777777777, + 400, + 123.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 207, + 123.66666666666666, + 400, + 137.55555555555554 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 207, + 137.55555555555554, + 400, + 151.44444444444443 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 207, + 151.44444444444443, + 400, + 165.33333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 207, + 165.33333333333331, + 400, + 179.2222222222222 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 207, + 179.2222222222222, + 400, + 193.1111111111111 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 207, + 193.1111111111111, + 400, + 206.99999999999997 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 199, + 212, + 411, + 224 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 199, + 212, + 411, + 225 + ], + "spans": [ + { + "bbox": [ + 199, + 212, + 411, + 225 + ], + "score": 1.0, + "content": "Figure 4: FID curves of different methods over time.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + } + ], + "index": 6.5 + }, + { + "type": "table", + "bbox": [ + 124, + 269, + 285, + 333 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 124, + 269, + 285, + 333 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 269, + 285, + 333 + ], + "spans": [ + { + "bbox": [ + 124, + 269, + 285, + 333 + ], + "score": 0.966, + "html": "
Encoder Structure LSTMFIDACC (%)
7.945
ConvS2S encoder13.354
Global self-attention3.680
Local self-attention4.378.5
", + "type": "table", + "image_path": "ec6b92b6e4c9db67e824443b66fac70b206a85a4a2c1c09c95deafb57a1a89d9.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 124, + 269, + 285, + 285.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 124, + 285.0, + 285, + 301.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 124, + 301.0, + 285, + 317.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 124, + 317.0, + 285, + 333.0 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 13.5 + }, + { + "type": "table", + "bbox": [ + 307, + 264, + 475, + 337 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 232, + 505, + 256 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 232, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 505, + 245 + ], + "score": 1.0, + "content": "Table 3: Left: Automatic metrics of our model with different encoder structures. Right: Automatic", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 243, + 352, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 352, + 258 + ], + "score": 1.0, + "content": "metrics of our model training by different learning strategies.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "table_body", + "bbox": [ + 307, + 264, + 475, + 337 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 475, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 475, + 337 + ], + "score": 0.926, + "html": "
Learning ApproachFIDACC (%)
Teacher-forcing61.25.3
Auto-condition (con st)15.735
Ours ([λet J)9.869
Ours ([λt²])6.473
Ours ([λt])5.177.6
", + "type": "table", + "image_path": "93832bc96a22cc8bcbf625fa826d04271ee181ac45bbcbe9616453589df5ecae.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 307, + 264, + 475, + 278.6 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 307, + 278.6, + 475, + 293.20000000000005 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 307, + 293.20000000000005, + 475, + 307.80000000000007 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 307, + 307.80000000000007, + 475, + 322.4000000000001 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 307, + 322.4000000000001, + 475, + 337.0000000000001 + ], + "spans": [], + "index": 20 + } + ] + } + ], + "index": 14.25 + }, + { + "type": "text", + "bbox": [ + 108, + 358, + 504, + 380 + ], + "lines": [], + "index": 21.5, + "bbox_fs": [ + 106, + 357, + 505, + 381 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 393, + 209, + 405 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 210, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 210, + 406 + ], + "score": 1.0, + "content": "5.3 ABLATION STUDY", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 505, + 491 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "Comparison on Encoder Structures. We compare the different encoder structures in our model", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "with the same LSTM decoder and the same curriculum learning strategy, as shown in Table 3. The", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 435, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 449 + ], + "score": 1.0, + "content": "transformer encoders outperform LSTM and the encoder in ConvS2S (Gehring et al., 2017) on", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "both FID and style consistency, due to its superior performance on sequence modeling. Although", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 458, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 505, + 471 + ], + "score": 1.0, + "content": "the transformer encoder with our proposed local self-attention slightly underperforms the global", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 469, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 268, + 482 + ], + "score": 1.0, + "content": "self-attention by 0.7 higher in FID and", + "type": "text" + }, + { + "bbox": [ + 268, + 469, + 291, + 480 + ], + "score": 0.86, + "content": "1 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 469, + 506, + 482 + ], + "score": 1.0, + "content": "lower in ACC, it has much fewer parameters in our", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 479, + 485, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 485, + 494 + ], + "score": 1.0, + "content": "settings. This confirms the effectiveness of the local self-attention on long sequence modeling.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 414, + 506, + 494 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 496, + 505, + 585 + ], + "lines": [ + { + "bbox": [ + 106, + 496, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 510 + ], + "score": 1.0, + "content": "Comparison on Learning Strategies. To investigate the performance of our proposed learning", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "approach, we conduct an ablation study to compare the performances of our model using different", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "score": 1.0, + "content": "learning strategies. As we can see the right of in Table 3, all curriculum learning strategies significantly", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "outperform the static auto-condition strategy (Li et al., 2017). Among them, the linear growth function", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 541, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 107, + 541, + 156, + 553 + ], + "score": 0.92, + "content": "f ( t ) = \\lfloor \\lambda t \\rfloor", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 541, + 505, + 553 + ], + "score": 1.0, + "content": "performs best, which might be that increasing the difficulty of curriculum too fast would", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "score": 1.0, + "content": "result in the model degradation. Besides, the original teacher-forcing strategy has the highest FID", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 561, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 576 + ], + "score": 1.0, + "content": "and the lowest style accuracy due to severe error accumulation problem, which indicates that using", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 573, + 489, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 489, + 587 + ], + "score": 1.0, + "content": "our proposed curriculum learning strategy to train the model can effectively alleviate this issue.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 496, + 505, + 587 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 601, + 195, + 614 + ], + "lines": [ + { + "bbox": [ + 105, + 598, + 198, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 198, + 617 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 106, + 626, + 505, + 725 + ], + "lines": [ + { + "bbox": [ + 106, + 626, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 506, + 638 + ], + "score": 1.0, + "content": "In this work, we propose a novel seq2seq architecture for long-term dance generation with music,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 104, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "e.g., about one-minute length. To efficiently process long sequences of music features, we introduce", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 646, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 104, + 646, + 506, + 662 + ], + "score": 1.0, + "content": "a local self-attention mechanism in the transformer encoder structure to reduce the quadratic memory", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "requirements to the linear in the sequence length. Besides, we also propose a dynamic auto-condition", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 669, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 683 + ], + "score": 1.0, + "content": "training method as a novel curriculum learning strategy to alleviate the error accumulation of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "autoregressive models in long motion sequence generation, thus facilitate the long-term dance", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 691, + 506, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 506, + 705 + ], + "score": 1.0, + "content": "generation. Extensive experiments have demonstrated the superior performance of our proposed", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 702, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 506, + 716 + ], + "score": 1.0, + "content": "approach on automatic metrics and human evaluation. Future works include the exploration on", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 714, + 491, + 726 + ], + "spans": [ + { + "bbox": [ + 106, + 714, + 491, + 726 + ], + "score": 1.0, + "content": "end-to-end methods that could work with raw audio data instead of preprocessed audio features.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 44, + "bbox_fs": [ + 104, + 626, + 506, + 726 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 176, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 104, + 100, + 480, + 112 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 480, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 480, + 113 + ], + "score": 1.0, + "content": "Janet Adshead-Lansdale and June Layson. Dance history: An introduction. Routledge, 2006.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 119, + 505, + 153 + ], + "lines": [ + { + "bbox": [ + 105, + 119, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 505, + 133 + ], + "score": 1.0, + "content": "Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 131, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 115, + 131, + 506, + 145 + ], + "score": 1.0, + "content": "prediction with recurrent neural networks. In Advances in Neural Information Processing Systems,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 142, + 207, + 154 + ], + "spans": [ + { + "bbox": [ + 114, + 142, + 207, + 154 + ], + "score": 1.0, + "content": "pp. 1171–1179, 2015.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 104, + 160, + 505, + 185 + ], + "lines": [ + { + "bbox": [ + 106, + 160, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 505, + 174 + ], + "score": 1.0, + "content": "Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 172, + 501, + 185 + ], + "spans": [ + { + "bbox": [ + 115, + 172, + 501, + 185 + ], + "score": 1.0, + "content": "Proceedings of the 26th annual international conference on machine learning, pp. 41–48, 2009.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 191, + 504, + 226 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "score": 1.0, + "content": "Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 203, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 116, + 203, + 506, + 216 + ], + "score": 1.0, + "content": "using part affinity fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 213, + 259, + 227 + ], + "spans": [ + { + "bbox": [ + 115, + 213, + 259, + 227 + ], + "score": 1.0, + "content": "Recognition, pp. 7291–7299, 2017.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 233, + 505, + 256 + ], + "lines": [ + { + "bbox": [ + 107, + 234, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 107, + 234, + 505, + 245 + ], + "score": 1.0, + "content": "Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A Efros. Everybody dance now. In", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 244, + 495, + 257 + ], + "spans": [ + { + "bbox": [ + 115, + 244, + 495, + 257 + ], + "score": 1.0, + "content": "Proceedings of the IEEE International Conference on Computer Vision, pp. 5933–5942, 2019.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 263, + 503, + 287 + ], + "lines": [ + { + "bbox": [ + 106, + 263, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 505, + 277 + ], + "score": 1.0, + "content": "Joyce L Chen, Virginia B Penhune, and Robert J Zatorre. Listening to musical rhythms recruits motor", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 275, + 372, + 287 + ], + "spans": [ + { + "bbox": [ + 115, + 275, + 372, + 287 + ], + "score": 1.0, + "content": "regions of the brain. Cerebral cortex, 18(12):2844–2854, 2008.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 294, + 503, + 317 + ], + "lines": [ + { + "bbox": [ + 105, + 293, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 308 + ], + "score": 1.0, + "content": "Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 305, + 335, + 318 + ], + "spans": [ + { + "bbox": [ + 115, + 305, + 335, + 318 + ], + "score": 1.0, + "content": "transformers. arXiv preprint arXiv:1904.10509, 2019.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 324, + 506, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 114, + 335, + 508, + 351 + ], + "spans": [ + { + "bbox": [ + 114, + 335, + 508, + 351 + ], + "score": 1.0, + "content": "pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pp.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 347, + 191, + 358 + ], + "spans": [ + { + "bbox": [ + 116, + 347, + 191, + 358 + ], + "score": 1.0, + "content": "2262–2271, 2019.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 366, + 504, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 365, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 506, + 380 + ], + "score": 1.0, + "content": "Daniel PW Ellis. Beat tracking by dynamic programming. Journal of New Music Research, 36(1):", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 378, + 172, + 389 + ], + "spans": [ + { + "bbox": [ + 116, + 378, + 172, + 389 + ], + "score": 1.0, + "content": "51–60, 2007.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 506, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 396, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 410 + ], + "score": 1.0, + "content": "Rukun Fan, Songhua Xu, and Weidong Geng. Example-based automatic music-driven conventional", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 408, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 116, + 408, + 506, + 421 + ], + "score": 1.0, + "content": "dance motion synthesis. IEEE transactions on visualization and computer graphics, 18(3):501–515,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 419, + 142, + 431 + ], + "spans": [ + { + "bbox": [ + 115, + 419, + 142, + 431 + ], + "score": 1.0, + "content": "2011.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 438, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 507, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 507, + 452 + ], + "score": 1.0, + "content": "Joao P Ferreira, Thiago M Coutinho, Thiago L Gomes, José F Neto, Rafael Azevedo, Renato Martins,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 116, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "and Erickson R Nascimento. Learning to dance: A graph convolutional adversarial network to", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 460, + 463, + 473 + ], + "spans": [ + { + "bbox": [ + 115, + 460, + 463, + 473 + ], + "score": 1.0, + "content": "generate realistic dance motions from audio. Computers & Graphics, 94:11–21, 2021.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 479, + 505, + 514 + ], + "lines": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 490, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 115, + 490, + 506, + 504 + ], + "score": 1.0, + "content": "for human dynamics. In Proceedings of the IEEE International Conference on Computer Vision,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 503, + 206, + 514 + ], + "spans": [ + { + "bbox": [ + 115, + 503, + 206, + 514 + ], + "score": 1.0, + "content": "pp. 4346–4354, 2015.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 521, + 506, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "score": 1.0, + "content": "Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 115, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "sequence to sequence learning. In Proceedings of the 34th International Conference on Machine", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 543, + 342, + 556 + ], + "spans": [ + { + "bbox": [ + 116, + 543, + 342, + 556 + ], + "score": 1.0, + "content": "Learning-Volume 70, pp. 1243–1252. JMLR. org, 2017.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 563, + 506, + 596 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 505, + 575 + ], + "score": 1.0, + "content": "Partha Ghosh, Jie Song, Emre Aksan, and Otmar Hilliges. Learning human motion models for", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 574, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 116, + 574, + 506, + 586 + ], + "score": 1.0, + "content": "long-term predictions. In 2017 International Conference on 3D Vision (3DV), pp. 458–466. IEEE,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 584, + 143, + 597 + ], + "spans": [ + { + "bbox": [ + 115, + 584, + 143, + 597 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 604, + 505, + 639 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 116, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in neural", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 627, + 337, + 639 + ], + "spans": [ + { + "bbox": [ + 116, + 627, + 337, + 639 + ], + "score": 1.0, + "content": "information processing systems, pp. 6626–6637, 2017.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 645, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 507, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 507, + 659 + ], + "score": 1.0, + "content": "Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 657, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 115, + 657, + 505, + 669 + ], + "score": 1.0, + "content": "Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 668, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 115, + 668, + 506, + 681 + ], + "score": 1.0, + "content": "for acoustic modeling in speech recognition: The shared views of four research groups. IEEE", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 678, + 312, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 678, + 312, + 692 + ], + "score": 1.0, + "content": "Signal processing magazine, 29(6):82–97, 2012.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Chieh Ho, Wei-Tze Tsai, Keng-Sheng Lin, and Homer H Chen. Extraction and alignment evaluation", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "of motion beats for street dance. In 2013 IEEE International Conference on Acoustics, Speech and", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 721, + 309, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 309, + 733 + ], + "score": 1.0, + "content": "Signal Processing, pp. 2429–2433. IEEE, 2013.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 176, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 104, + 100, + 480, + 112 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 480, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 480, + 113 + ], + "score": 1.0, + "content": "Janet Adshead-Lansdale and June Layson. Dance history: An introduction. Routledge, 2006.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 99, + 480, + 113 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 119, + 505, + 153 + ], + "lines": [ + { + "bbox": [ + 105, + 119, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 505, + 133 + ], + "score": 1.0, + "content": "Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 131, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 115, + 131, + 506, + 145 + ], + "score": 1.0, + "content": "prediction with recurrent neural networks. In Advances in Neural Information Processing Systems,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 142, + 207, + 154 + ], + "spans": [ + { + "bbox": [ + 114, + 142, + 207, + 154 + ], + "score": 1.0, + "content": "pp. 1171–1179, 2015.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 119, + 506, + 154 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 160, + 505, + 185 + ], + "lines": [ + { + "bbox": [ + 106, + 160, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 505, + 174 + ], + "score": 1.0, + "content": "Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 172, + 501, + 185 + ], + "spans": [ + { + "bbox": [ + 115, + 172, + 501, + 185 + ], + "score": 1.0, + "content": "Proceedings of the 26th annual international conference on machine learning, pp. 41–48, 2009.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 106, + 160, + 505, + 185 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 191, + 504, + 226 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "score": 1.0, + "content": "Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 203, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 116, + 203, + 506, + 216 + ], + "score": 1.0, + "content": "using part affinity fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 213, + 259, + 227 + ], + "spans": [ + { + "bbox": [ + 115, + 213, + 259, + 227 + ], + "score": 1.0, + "content": "Recognition, pp. 7291–7299, 2017.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 191, + 506, + 227 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 233, + 505, + 256 + ], + "lines": [ + { + "bbox": [ + 107, + 234, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 107, + 234, + 505, + 245 + ], + "score": 1.0, + "content": "Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A Efros. Everybody dance now. In", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 244, + 495, + 257 + ], + "spans": [ + { + "bbox": [ + 115, + 244, + 495, + 257 + ], + "score": 1.0, + "content": "Proceedings of the IEEE International Conference on Computer Vision, pp. 5933–5942, 2019.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 107, + 234, + 505, + 257 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 263, + 503, + 287 + ], + "lines": [ + { + "bbox": [ + 106, + 263, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 505, + 277 + ], + "score": 1.0, + "content": "Joyce L Chen, Virginia B Penhune, and Robert J Zatorre. Listening to musical rhythms recruits motor", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 275, + 372, + 287 + ], + "spans": [ + { + "bbox": [ + 115, + 275, + 372, + 287 + ], + "score": 1.0, + "content": "regions of the brain. Cerebral cortex, 18(12):2844–2854, 2008.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 106, + 263, + 505, + 287 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 294, + 503, + 317 + ], + "lines": [ + { + "bbox": [ + 105, + 293, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 308 + ], + "score": 1.0, + "content": "Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 305, + 335, + 318 + ], + "spans": [ + { + "bbox": [ + 115, + 305, + 335, + 318 + ], + "score": 1.0, + "content": "transformers. arXiv preprint arXiv:1904.10509, 2019.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 293, + 505, + 318 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 324, + 506, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 114, + 335, + 508, + 351 + ], + "spans": [ + { + "bbox": [ + 114, + 335, + 508, + 351 + ], + "score": 1.0, + "content": "pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pp.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 347, + 191, + 358 + ], + "spans": [ + { + "bbox": [ + 116, + 347, + 191, + 358 + ], + "score": 1.0, + "content": "2262–2271, 2019.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 325, + 508, + 358 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 366, + 504, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 365, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 506, + 380 + ], + "score": 1.0, + "content": "Daniel PW Ellis. Beat tracking by dynamic programming. Journal of New Music Research, 36(1):", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 378, + 172, + 389 + ], + "spans": [ + { + "bbox": [ + 116, + 378, + 172, + 389 + ], + "score": 1.0, + "content": "51–60, 2007.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 365, + 506, + 389 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 506, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 396, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 410 + ], + "score": 1.0, + "content": "Rukun Fan, Songhua Xu, and Weidong Geng. Example-based automatic music-driven conventional", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 408, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 116, + 408, + 506, + 421 + ], + "score": 1.0, + "content": "dance motion synthesis. IEEE transactions on visualization and computer graphics, 18(3):501–515,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 419, + 142, + 431 + ], + "spans": [ + { + "bbox": [ + 115, + 419, + 142, + 431 + ], + "score": 1.0, + "content": "2011.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 396, + 506, + 431 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 438, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 507, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 507, + 452 + ], + "score": 1.0, + "content": "Joao P Ferreira, Thiago M Coutinho, Thiago L Gomes, José F Neto, Rafael Azevedo, Renato Martins,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 116, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "and Erickson R Nascimento. Learning to dance: A graph convolutional adversarial network to", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 460, + 463, + 473 + ], + "spans": [ + { + "bbox": [ + 115, + 460, + 463, + 473 + ], + "score": 1.0, + "content": "generate realistic dance motions from audio. Computers & Graphics, 94:11–21, 2021.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 437, + 507, + 473 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 479, + 505, + 514 + ], + "lines": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 490, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 115, + 490, + 506, + 504 + ], + "score": 1.0, + "content": "for human dynamics. In Proceedings of the IEEE International Conference on Computer Vision,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 503, + 206, + 514 + ], + "spans": [ + { + "bbox": [ + 115, + 503, + 206, + 514 + ], + "score": 1.0, + "content": "pp. 4346–4354, 2015.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 106, + 480, + 506, + 514 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 521, + 506, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "score": 1.0, + "content": "Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 115, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "sequence to sequence learning. In Proceedings of the 34th International Conference on Machine", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 543, + 342, + 556 + ], + "spans": [ + { + "bbox": [ + 116, + 543, + 342, + 556 + ], + "score": 1.0, + "content": "Learning-Volume 70, pp. 1243–1252. JMLR. org, 2017.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 520, + 506, + 556 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 563, + 506, + 596 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 505, + 575 + ], + "score": 1.0, + "content": "Partha Ghosh, Jie Song, Emre Aksan, and Otmar Hilliges. Learning human motion models for", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 574, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 116, + 574, + 506, + 586 + ], + "score": 1.0, + "content": "long-term predictions. In 2017 International Conference on 3D Vision (3DV), pp. 458–466. IEEE,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 584, + 143, + 597 + ], + "spans": [ + { + "bbox": [ + 115, + 584, + 143, + 597 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 106, + 563, + 506, + 597 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 604, + 505, + 639 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 116, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in neural", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 627, + 337, + 639 + ], + "spans": [ + { + "bbox": [ + 116, + 627, + 337, + 639 + ], + "score": 1.0, + "content": "information processing systems, pp. 6626–6637, 2017.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 604, + 505, + 639 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 645, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 507, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 507, + 659 + ], + "score": 1.0, + "content": "Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 657, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 115, + 657, + 505, + 669 + ], + "score": 1.0, + "content": "Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 668, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 115, + 668, + 506, + 681 + ], + "score": 1.0, + "content": "for acoustic modeling in speech recognition: The shared views of four research groups. IEEE", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 678, + 312, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 678, + 312, + 692 + ], + "score": 1.0, + "content": "Signal processing magazine, 29(6):82–97, 2012.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 645, + 507, + 692 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Chieh Ho, Wei-Tze Tsai, Keng-Sheng Lin, and Homer H Chen. Extraction and alignment evaluation", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "of motion beats for street dance. In 2013 IEEE International Conference on Acoustics, Speech and", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 721, + 309, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 309, + 733 + ], + "score": 1.0, + "content": "Signal Processing, pp. 2429–2433. IEEE, 2013.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 106, + 699, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Cheng-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit, Noam Shazeer, Ian Simon, Curtis", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "Hawthorne, Andrew M Dai, Matthew D Hoffman, Monica Dinculescu, and Douglas Eck. Music", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 117, + 105, + 331, + 117 + ], + "spans": [ + { + "bbox": [ + 117, + 105, + 331, + 117 + ], + "score": 1.0, + "content": "transformer. arXiv preprint arXiv:1809.04281, 2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 108, + 122, + 504, + 156 + ], + "lines": [ + { + "bbox": [ + 105, + 121, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 136 + ], + "score": 1.0, + "content": "Ashesh Jain, Amir R Zamir, Silvio Savarese, and Ashutosh Saxena. Structural-rnn: Deep learning", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 134, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 134, + 505, + 146 + ], + "score": 1.0, + "content": "on spatio-temporal graphs. In Proceedings of the ieee conference on computer vision and pattern", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 145, + 257, + 156 + ], + "spans": [ + { + "bbox": [ + 115, + 145, + 257, + 156 + ], + "score": 1.0, + "content": "recognition, pp. 5308–5317, 2016.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 162, + 504, + 196 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 506, + 176 + ], + "score": 1.0, + "content": "Alex M Lamb, Anirudh Goyal Alias Parth Goyal, Ying Zhang, Saizheng Zhang, Aaron C Courville,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 173, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 115, + 173, + 505, + 186 + ], + "score": 1.0, + "content": "and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. In", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 114, + 184, + 420, + 197 + ], + "spans": [ + { + "bbox": [ + 114, + 184, + 420, + 197 + ], + "score": 1.0, + "content": "Advances In Neural Information Processing Systems, pp. 4601–4609, 2016.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 202, + 505, + 236 + ], + "lines": [ + { + "bbox": [ + 105, + 201, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 506, + 217 + ], + "score": 1.0, + "content": "Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 212, + 507, + 228 + ], + "spans": [ + { + "bbox": [ + 115, + 212, + 507, + 228 + ], + "score": 1.0, + "content": "and Jan Kautz. Dancing to music. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 117, + 225, + 191, + 236 + ], + "spans": [ + { + "bbox": [ + 117, + 225, + 191, + 236 + ], + "score": 1.0, + "content": "3581–3591, 2019.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 242, + 504, + 266 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 505, + 256 + ], + "score": 1.0, + "content": "Juheon Lee, Seohyun Kim, and Kyogu Lee. Listen to dance: Music-driven choreography generation", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 254, + 469, + 266 + ], + "spans": [ + { + "bbox": [ + 116, + 254, + 469, + 266 + ], + "score": 1.0, + "content": "using autoregressive encoder-decoder network. arXiv preprint arXiv:1811.00818, 2018.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 105, + 271, + 504, + 295 + ], + "lines": [ + { + "bbox": [ + 106, + 271, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 505, + 285 + ], + "score": 1.0, + "content": "Minho Lee, Kyogu Lee, and Jaeheung Park. Music similarity-based approach to generating dance", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 282, + 421, + 295 + ], + "spans": [ + { + "bbox": [ + 115, + 282, + 421, + 295 + ], + "score": 1.0, + "content": "motion sequence. Multimedia tools and applications, 62(3):895–912, 2013.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 108, + 300, + 504, + 335 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 504, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 504, + 312 + ], + "score": 1.0, + "content": "Andreas M Lehrmann, Peter V Gehler, and Sebastian Nowozin. Efficient nonlinear markov models", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 115, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "for human motion. In Proceedings of the IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 323, + 259, + 335 + ], + "spans": [ + { + "bbox": [ + 115, + 323, + 259, + 335 + ], + "score": 1.0, + "content": "Recognition, pp. 1314–1321, 2014.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 108, + 340, + 504, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 354 + ], + "score": 1.0, + "content": "Chen Li, Zhen Zhang, Wee Sun Lee, and Gim Hee Lee. Convolutional sequence to sequence model", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 352, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 115, + 352, + 506, + 365 + ], + "score": 1.0, + "content": "for human dynamics. In Proceedings of the IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 363, + 259, + 375 + ], + "spans": [ + { + "bbox": [ + 115, + 363, + 259, + 375 + ], + "score": 1.0, + "content": "Recognition, pp. 5226–5234, 2018.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 104, + 381, + 505, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "Zimo Li, Yi Zhou, Shuangjiu Xiao, Chong He, Zeng Huang, and Hao Li. Auto-conditioned recurrent", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 393, + 507, + 405 + ], + "spans": [ + { + "bbox": [ + 115, + 393, + 507, + 405 + ], + "score": 1.0, + "content": "networks for extended complex human motion synthesis. arXiv preprint arXiv:1707.05363, 2017.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 410, + 504, + 444 + ], + "lines": [ + { + "bbox": [ + 105, + 409, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 505, + 424 + ], + "score": 1.0, + "content": "Jiasen Lu, Caiming Xiong, Devi Parikh, and Richard Socher. Knowing when to look: Adaptive", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 422, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 116, + 422, + 505, + 434 + ], + "score": 1.0, + "content": "attention via a visual sentinel for image captioning. In Proceedings of the IEEE conference on", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 433, + 362, + 444 + ], + "spans": [ + { + "bbox": [ + 116, + 433, + 362, + 444 + ], + "score": 1.0, + "content": "computer vision and pattern recognition, pp. 375–383, 2017.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 504, + 484 + ], + "lines": [ + { + "bbox": [ + 106, + 450, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 463 + ], + "score": 1.0, + "content": "Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 115, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "for human motion prediction. In Proceedings of the IEEE International Conference on Computer", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 473, + 236, + 484 + ], + "spans": [ + { + "bbox": [ + 116, + 473, + 236, + 484 + ], + "score": 1.0, + "content": "Vision, pp. 9489–9497, 2019.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 490, + 505, + 524 + ], + "lines": [ + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "Brian McFee, Colin Raffel, Dawen Liang, Daniel PW Ellis, Matt McVicar, Eric Battenberg, and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 116, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "Oriol Nieto. librosa: Audio and music signal analysis in python. In Proceedings of the 14th python", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 513, + 274, + 524 + ], + "spans": [ + { + "bbox": [ + 116, + 513, + 274, + 524 + ], + "score": 1.0, + "content": "in science conference, volume 8, 2015.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 105, + 530, + 505, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 529, + 507, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 507, + 544 + ], + "score": 1.0, + "content": "Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 541, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 115, + 541, + 505, + 555 + ], + "score": 1.0, + "content": "Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 117, + 553, + 307, + 565 + ], + "spans": [ + { + "bbox": [ + 117, + 553, + 307, + 565 + ], + "score": 1.0, + "content": "audio. arXiv preprint arXiv:1609.03499, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 570, + 505, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 568, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 585 + ], + "score": 1.0, + "content": "Marc’Aurelio Ranzato, Sumit Chopra, Michael Auli, and Wojciech Zaremba. Sequence level training", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 582, + 407, + 594 + ], + "spans": [ + { + "bbox": [ + 116, + 582, + 407, + 594 + ], + "score": 1.0, + "content": "with recurrent neural networks. arXiv preprint arXiv:1511.06732, 2015.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 504, + 623 + ], + "lines": [ + { + "bbox": [ + 106, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 611, + 468, + 624 + ], + "spans": [ + { + "bbox": [ + 116, + 611, + 468, + 624 + ], + "score": 1.0, + "content": "Generative adversarial text to image synthesis. arXiv preprint arXiv:1605.05396, 2016.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 107, + 628, + 504, + 652 + ], + "lines": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "Xuanchi Ren, Haoran Li, Zijian Huang, and Qifeng Chen. Music-oriented dance video synthesis", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 640, + 385, + 653 + ], + "spans": [ + { + "bbox": [ + 116, + 640, + 385, + 653 + ], + "score": 1.0, + "content": "with pose perceptual loss. arXiv preprint arXiv:1912.06606, 2019.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 106, + 658, + 505, + 692 + ], + "lines": [ + { + "bbox": [ + 106, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "Eli Shlizerman, Lucio Dery, Hayden Schoen, and Ira Kemelmacher-Shlizerman. Audio to body", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 670, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 116, + 670, + 506, + 682 + ], + "score": 1.0, + "content": "dynamics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 680, + 207, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 207, + 692 + ], + "score": 1.0, + "content": "pp. 7574–7583, 2018.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 696, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 505, + 713 + ], + "score": 1.0, + "content": "Guofei Sun, Yongkang Wong, Zhiyong Cheng, Mohan S Kankanhalli, Weidong Geng, and Xiangdong", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Li. Deepdance: music-to-dance motion choreography with adversarial learning. IEEE Transactions", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 721, + 257, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 257, + 732 + ], + "score": 1.0, + "content": "on Multimedia, 23:497–509, 2020.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Cheng-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit, Noam Shazeer, Ian Simon, Curtis", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "Hawthorne, Andrew M Dai, Matthew D Hoffman, Monica Dinculescu, and Douglas Eck. Music", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 117, + 105, + 331, + 117 + ], + "spans": [ + { + "bbox": [ + 117, + 105, + 331, + 117 + ], + "score": 1.0, + "content": "transformer. arXiv preprint arXiv:1809.04281, 2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 82, + 505, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 122, + 504, + 156 + ], + "lines": [ + { + "bbox": [ + 105, + 121, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 136 + ], + "score": 1.0, + "content": "Ashesh Jain, Amir R Zamir, Silvio Savarese, and Ashutosh Saxena. Structural-rnn: Deep learning", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 134, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 134, + 505, + 146 + ], + "score": 1.0, + "content": "on spatio-temporal graphs. In Proceedings of the ieee conference on computer vision and pattern", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 145, + 257, + 156 + ], + "spans": [ + { + "bbox": [ + 115, + 145, + 257, + 156 + ], + "score": 1.0, + "content": "recognition, pp. 5308–5317, 2016.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 121, + 505, + 156 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 162, + 504, + 196 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 506, + 176 + ], + "score": 1.0, + "content": "Alex M Lamb, Anirudh Goyal Alias Parth Goyal, Ying Zhang, Saizheng Zhang, Aaron C Courville,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 173, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 115, + 173, + 505, + 186 + ], + "score": 1.0, + "content": "and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. In", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 114, + 184, + 420, + 197 + ], + "spans": [ + { + "bbox": [ + 114, + 184, + 420, + 197 + ], + "score": 1.0, + "content": "Advances In Neural Information Processing Systems, pp. 4601–4609, 2016.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 162, + 506, + 197 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 202, + 505, + 236 + ], + "lines": [ + { + "bbox": [ + 105, + 201, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 506, + 217 + ], + "score": 1.0, + "content": "Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 212, + 507, + 228 + ], + "spans": [ + { + "bbox": [ + 115, + 212, + 507, + 228 + ], + "score": 1.0, + "content": "and Jan Kautz. Dancing to music. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 117, + 225, + 191, + 236 + ], + "spans": [ + { + "bbox": [ + 117, + 225, + 191, + 236 + ], + "score": 1.0, + "content": "3581–3591, 2019.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 201, + 507, + 236 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 242, + 504, + 266 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 505, + 256 + ], + "score": 1.0, + "content": "Juheon Lee, Seohyun Kim, and Kyogu Lee. Listen to dance: Music-driven choreography generation", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 254, + 469, + 266 + ], + "spans": [ + { + "bbox": [ + 116, + 254, + 469, + 266 + ], + "score": 1.0, + "content": "using autoregressive encoder-decoder network. arXiv preprint arXiv:1811.00818, 2018.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 106, + 243, + 505, + 266 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 271, + 504, + 295 + ], + "lines": [ + { + "bbox": [ + 106, + 271, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 505, + 285 + ], + "score": 1.0, + "content": "Minho Lee, Kyogu Lee, and Jaeheung Park. Music similarity-based approach to generating dance", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 282, + 421, + 295 + ], + "spans": [ + { + "bbox": [ + 115, + 282, + 421, + 295 + ], + "score": 1.0, + "content": "motion sequence. Multimedia tools and applications, 62(3):895–912, 2013.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 271, + 505, + 295 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 300, + 504, + 335 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 504, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 504, + 312 + ], + "score": 1.0, + "content": "Andreas M Lehrmann, Peter V Gehler, and Sebastian Nowozin. Efficient nonlinear markov models", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 115, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "for human motion. In Proceedings of the IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 323, + 259, + 335 + ], + "spans": [ + { + "bbox": [ + 115, + 323, + 259, + 335 + ], + "score": 1.0, + "content": "Recognition, pp. 1314–1321, 2014.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 302, + 505, + 335 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 340, + 504, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 354 + ], + "score": 1.0, + "content": "Chen Li, Zhen Zhang, Wee Sun Lee, and Gim Hee Lee. Convolutional sequence to sequence model", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 352, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 115, + 352, + 506, + 365 + ], + "score": 1.0, + "content": "for human dynamics. In Proceedings of the IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 363, + 259, + 375 + ], + "spans": [ + { + "bbox": [ + 115, + 363, + 259, + 375 + ], + "score": 1.0, + "content": "Recognition, pp. 5226–5234, 2018.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 106, + 340, + 506, + 375 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 381, + 505, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "Zimo Li, Yi Zhou, Shuangjiu Xiao, Chong He, Zeng Huang, and Hao Li. Auto-conditioned recurrent", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 393, + 507, + 405 + ], + "spans": [ + { + "bbox": [ + 115, + 393, + 507, + 405 + ], + "score": 1.0, + "content": "networks for extended complex human motion synthesis. arXiv preprint arXiv:1707.05363, 2017.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 381, + 507, + 405 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 410, + 504, + 444 + ], + "lines": [ + { + "bbox": [ + 105, + 409, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 505, + 424 + ], + "score": 1.0, + "content": "Jiasen Lu, Caiming Xiong, Devi Parikh, and Richard Socher. Knowing when to look: Adaptive", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 422, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 116, + 422, + 505, + 434 + ], + "score": 1.0, + "content": "attention via a visual sentinel for image captioning. In Proceedings of the IEEE conference on", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 433, + 362, + 444 + ], + "spans": [ + { + "bbox": [ + 116, + 433, + 362, + 444 + ], + "score": 1.0, + "content": "computer vision and pattern recognition, pp. 375–383, 2017.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 409, + 505, + 444 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 504, + 484 + ], + "lines": [ + { + "bbox": [ + 106, + 450, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 463 + ], + "score": 1.0, + "content": "Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 115, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "for human motion prediction. In Proceedings of the IEEE International Conference on Computer", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 473, + 236, + 484 + ], + "spans": [ + { + "bbox": [ + 116, + 473, + 236, + 484 + ], + "score": 1.0, + "content": "Vision, pp. 9489–9497, 2019.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 106, + 450, + 505, + 484 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 490, + 505, + 524 + ], + "lines": [ + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "Brian McFee, Colin Raffel, Dawen Liang, Daniel PW Ellis, Matt McVicar, Eric Battenberg, and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 116, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "Oriol Nieto. librosa: Audio and music signal analysis in python. In Proceedings of the 14th python", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 513, + 274, + 524 + ], + "spans": [ + { + "bbox": [ + 116, + 513, + 274, + 524 + ], + "score": 1.0, + "content": "in science conference, volume 8, 2015.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 106, + 490, + 505, + 524 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 530, + 505, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 529, + 507, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 507, + 544 + ], + "score": 1.0, + "content": "Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 541, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 115, + 541, + 505, + 555 + ], + "score": 1.0, + "content": "Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 117, + 553, + 307, + 565 + ], + "spans": [ + { + "bbox": [ + 117, + 553, + 307, + 565 + ], + "score": 1.0, + "content": "audio. arXiv preprint arXiv:1609.03499, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 529, + 507, + 565 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 570, + 505, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 568, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 585 + ], + "score": 1.0, + "content": "Marc’Aurelio Ranzato, Sumit Chopra, Michael Auli, and Wojciech Zaremba. Sequence level training", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 582, + 407, + 594 + ], + "spans": [ + { + "bbox": [ + 116, + 582, + 407, + 594 + ], + "score": 1.0, + "content": "with recurrent neural networks. arXiv preprint arXiv:1511.06732, 2015.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 568, + 506, + 594 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 504, + 623 + ], + "lines": [ + { + "bbox": [ + 106, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 611, + 468, + 624 + ], + "spans": [ + { + "bbox": [ + 116, + 611, + 468, + 624 + ], + "score": 1.0, + "content": "Generative adversarial text to image synthesis. arXiv preprint arXiv:1605.05396, 2016.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 106, + 600, + 506, + 624 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 628, + 504, + 652 + ], + "lines": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "Xuanchi Ren, Haoran Li, Zijian Huang, and Qifeng Chen. Music-oriented dance video synthesis", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 640, + 385, + 653 + ], + "spans": [ + { + "bbox": [ + 116, + 640, + 385, + 653 + ], + "score": 1.0, + "content": "with pose perceptual loss. arXiv preprint arXiv:1912.06606, 2019.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 628, + 505, + 653 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 658, + 505, + 692 + ], + "lines": [ + { + "bbox": [ + 106, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "Eli Shlizerman, Lucio Dery, Hayden Schoen, and Ira Kemelmacher-Shlizerman. Audio to body", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 670, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 116, + 670, + 506, + 682 + ], + "score": 1.0, + "content": "dynamics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 680, + 207, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 207, + 692 + ], + "score": 1.0, + "content": "pp. 7574–7583, 2018.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 106, + 658, + 506, + 692 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 696, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 505, + 713 + ], + "score": 1.0, + "content": "Guofei Sun, Yongkang Wong, Zhiyong Cheng, Mohan S Kankanhalli, Weidong Geng, and Xiangdong", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Li. Deepdance: music-to-dance motion choreography with adversarial learning. IEEE Transactions", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 721, + 257, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 257, + 732 + ], + "score": 1.0, + "content": "on Multimedia, 23:497–509, 2020.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 696, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Taoran Tang, Jia Jia, and Hanyang Mao. Dance with melody: An lstm-autoencoder approach to", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "music-oriented dance synthesis. In Proceedings of the 26th ACM international conference on", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 257, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 257, + 117 + ], + "score": 1.0, + "content": "Multimedia, pp. 1598–1606, 2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 105, + 123, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "score": 1.0, + "content": "Graham W Taylor, Geoffrey E Hinton, and Sam T Roweis. Modeling human motion using binary", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 133, + 493, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 133, + 493, + 146 + ], + "score": 1.0, + "content": "latent variables. In Advances in neural information processing systems, pp. 1345–1352, 2007.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 108, + 153, + 504, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 152, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 506, + 167 + ], + "score": 1.0, + "content": "Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 116, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "and content for video generation. In Proceedings of the IEEE conference on computer vision and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 176, + 288, + 187 + ], + "spans": [ + { + "bbox": [ + 115, + 176, + 288, + 187 + ], + "score": 1.0, + "content": "pattern recognition, pp. 1526–1535, 2018.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 504, + 228 + ], + "lines": [ + { + "bbox": [ + 106, + 193, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 505, + 206 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 116, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 217, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 115, + 217, + 286, + 228 + ], + "score": 1.0, + "content": "processing systems, pp. 5998–6008, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 105, + 234, + 504, + 258 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 505, + 248 + ], + "score": 1.0, + "content": "Jack Wang, Aaron Hertzmann, and David J Fleet. Gaussian process dynamical models. In Advances", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 246, + 375, + 258 + ], + "spans": [ + { + "bbox": [ + 115, + 246, + 375, + 258 + ], + "score": 1.0, + "content": "in neural information processing systems, pp. 1441–1448, 2006.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 264, + 503, + 288 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 505, + 278 + ], + "score": 1.0, + "content": "Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 275, + 430, + 288 + ], + "spans": [ + { + "bbox": [ + 116, + 275, + 430, + 288 + ], + "score": 1.0, + "content": "Catanzaro. Video-to-video synthesis. arXiv preprint arXiv:1808.06601, 2018.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 294, + 504, + 328 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 306, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 116, + 306, + 505, + 318 + ], + "score": 1.0, + "content": "Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 316, + 447, + 330 + ], + "spans": [ + { + "bbox": [ + 116, + 316, + 447, + 330 + ], + "score": 1.0, + "content": "attention. In International conference on machine learning, pp. 2048–2057, 2015.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 335, + 504, + 369 + ], + "lines": [ + { + "bbox": [ + 105, + 334, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 349 + ], + "score": 1.0, + "content": "Nelson Yalta, Shinji Watanabe, Kazuhiro Nakadai, and Tetsuya Ogata. Weakly-supervised deep", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 347, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 115, + 347, + 505, + 359 + ], + "score": 1.0, + "content": "recurrent neural networks for basic dance step generation. In 2019 International Joint Conference", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 357, + 324, + 369 + ], + "spans": [ + { + "bbox": [ + 116, + 357, + 324, + 369 + ], + "score": 1.0, + "content": "on Neural Networks (IJCNN), pp. 1–8. IEEE, 2019.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 375, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 375, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 389 + ], + "score": 1.0, + "content": "Zijie Ye, Haozhe Wu, Jia Jia, Yaohua Bu, Wei Chen, Fanbo Meng, and Yanfeng Wang. Choreonet:", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 387, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 116, + 387, + 505, + 400 + ], + "score": 1.0, + "content": "Towards music to dance synthesis with choreographic action unit. In Proceedings of the 28th ACM", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 397, + 362, + 411 + ], + "spans": [ + { + "bbox": [ + 115, + 397, + 362, + 411 + ], + "score": 1.0, + "content": "International Conference on Multimedia, pp. 744–752, 2020.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 417, + 505, + 461 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "score": 1.0, + "content": "Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 428, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 116, + 428, + 505, + 440 + ], + "score": 1.0, + "content": "Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 440, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 115, + 440, + 506, + 451 + ], + "score": 1.0, + "content": "networks. In Proceedings of the IEEE international conference on computer vision, pp. 5907–5915,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 450, + 142, + 461 + ], + "spans": [ + { + "bbox": [ + 115, + 450, + 142, + 461 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 469, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 467, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 483 + ], + "score": 1.0, + "content": "Wen Zhang, Yang Feng, Fandong Meng, Di You, and Qun Liu. Bridging the gap between training", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 480, + 461, + 492 + ], + "spans": [ + { + "bbox": [ + 116, + 480, + 461, + 492 + ], + "score": 1.0, + "content": "and inference for neural machine translation. arXiv preprint arXiv:1906.02448, 2019.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Taoran Tang, Jia Jia, and Hanyang Mao. Dance with melody: An lstm-autoencoder approach to", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "music-oriented dance synthesis. In Proceedings of the 26th ACM international conference on", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 257, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 257, + 117 + ], + "score": 1.0, + "content": "Multimedia, pp. 1598–1606, 2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 81, + 505, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 123, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "score": 1.0, + "content": "Graham W Taylor, Geoffrey E Hinton, and Sam T Roweis. Modeling human motion using binary", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 133, + 493, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 133, + 493, + 146 + ], + "score": 1.0, + "content": "latent variables. In Advances in neural information processing systems, pp. 1345–1352, 2007.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 122, + 505, + 146 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 153, + 504, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 152, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 506, + 167 + ], + "score": 1.0, + "content": "Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 116, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "and content for video generation. In Proceedings of the IEEE conference on computer vision and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 176, + 288, + 187 + ], + "spans": [ + { + "bbox": [ + 115, + 176, + 288, + 187 + ], + "score": 1.0, + "content": "pattern recognition, pp. 1526–1535, 2018.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 152, + 506, + 187 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 504, + 228 + ], + "lines": [ + { + "bbox": [ + 106, + 193, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 505, + 206 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 116, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 217, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 115, + 217, + 286, + 228 + ], + "score": 1.0, + "content": "processing systems, pp. 5998–6008, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 193, + 505, + 228 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 234, + 504, + 258 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 505, + 248 + ], + "score": 1.0, + "content": "Jack Wang, Aaron Hertzmann, and David J Fleet. Gaussian process dynamical models. In Advances", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 246, + 375, + 258 + ], + "spans": [ + { + "bbox": [ + 115, + 246, + 375, + 258 + ], + "score": 1.0, + "content": "in neural information processing systems, pp. 1441–1448, 2006.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 234, + 505, + 258 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 264, + 503, + 288 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 505, + 278 + ], + "score": 1.0, + "content": "Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 275, + 430, + 288 + ], + "spans": [ + { + "bbox": [ + 116, + 275, + 430, + 288 + ], + "score": 1.0, + "content": "Catanzaro. Video-to-video synthesis. arXiv preprint arXiv:1808.06601, 2018.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 264, + 505, + 288 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 294, + 504, + 328 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 306, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 116, + 306, + 505, + 318 + ], + "score": 1.0, + "content": "Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 316, + 447, + 330 + ], + "spans": [ + { + "bbox": [ + 116, + 316, + 447, + 330 + ], + "score": 1.0, + "content": "attention. In International conference on machine learning, pp. 2048–2057, 2015.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 294, + 505, + 330 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 335, + 504, + 369 + ], + "lines": [ + { + "bbox": [ + 105, + 334, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 349 + ], + "score": 1.0, + "content": "Nelson Yalta, Shinji Watanabe, Kazuhiro Nakadai, and Tetsuya Ogata. Weakly-supervised deep", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 347, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 115, + 347, + 505, + 359 + ], + "score": 1.0, + "content": "recurrent neural networks for basic dance step generation. In 2019 International Joint Conference", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 357, + 324, + 369 + ], + "spans": [ + { + "bbox": [ + 116, + 357, + 324, + 369 + ], + "score": 1.0, + "content": "on Neural Networks (IJCNN), pp. 1–8. IEEE, 2019.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 334, + 505, + 369 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 375, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 375, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 389 + ], + "score": 1.0, + "content": "Zijie Ye, Haozhe Wu, Jia Jia, Yaohua Bu, Wei Chen, Fanbo Meng, and Yanfeng Wang. Choreonet:", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 387, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 116, + 387, + 505, + 400 + ], + "score": 1.0, + "content": "Towards music to dance synthesis with choreographic action unit. In Proceedings of the 28th ACM", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 397, + 362, + 411 + ], + "spans": [ + { + "bbox": [ + 115, + 397, + 362, + 411 + ], + "score": 1.0, + "content": "International Conference on Multimedia, pp. 744–752, 2020.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 375, + 506, + 411 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 417, + 505, + 461 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "score": 1.0, + "content": "Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 428, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 116, + 428, + 505, + 440 + ], + "score": 1.0, + "content": "Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 440, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 115, + 440, + 506, + 451 + ], + "score": 1.0, + "content": "networks. In Proceedings of the IEEE international conference on computer vision, pp. 5907–5915,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 450, + 142, + 461 + ], + "spans": [ + { + "bbox": [ + 115, + 450, + 142, + 461 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 417, + 506, + 461 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 469, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 467, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 483 + ], + "score": 1.0, + "content": "Wen Zhang, Yang Feng, Fandong Meng, Di You, and Qun Liu. Bridging the gap between training", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 480, + 461, + 492 + ], + "spans": [ + { + "bbox": [ + 116, + 480, + 461, + 492 + ], + "score": 1.0, + "content": "and inference for neural machine translation. arXiv preprint arXiv:1906.02448, 2019.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 467, + 506, + 492 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 161, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 163, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 163, + 96 + ], + "score": 1.0, + "content": "APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 105, + 316, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 317, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 317, + 118 + ], + "score": 1.0, + "content": "6.1 MORE DETAILS ABOUT DATA COLLECTION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 126, + 505, + 203 + ], + "lines": [ + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "The dance videos are collected from YouTube by crowd workers, which are all solo dance videos", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 137, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 149 + ], + "score": 1.0, + "content": "with 30FPS. We trim the beginning and ending several seconds for each of collected videos to remove", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "the silence parts, and split them into one-minute video clips. Then, we extract 2D pose data from", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "score": 1.0, + "content": "these video clips by OpenPose (Cao et al., 2017) with 15FPS setting and collect 790 one-minute clips", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 169, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 505, + 182 + ], + "score": 1.0, + "content": "of pose data with 15FPS, totaling about 13 hours. Finally, we extract the corresponding audio data", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 180, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 506, + 194 + ], + "score": 1.0, + "content": "from video clips by FFmpeg3. The code is implemented based on PyTorch framework. MIT License", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 192, + 280, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 280, + 204 + ], + "score": 1.0, + "content": "will be used for the released code and data.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 108, + 216, + 263, + 227 + ], + "lines": [ + { + "bbox": [ + 106, + 215, + 265, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 265, + 229 + ], + "score": 1.0, + "content": "6.2 EXTRACTED MUSIC FEATURES", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 237, + 506, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 507, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 507, + 250 + ], + "score": 1.0, + "content": "In this section, we detail the extracted features of music that are feeded into our model. Specifically,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 248, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 506, + 260 + ], + "score": 1.0, + "content": "we leverage the public Librosa (McFee et al., 2015) to extract the music features including: 20-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 257, + 507, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 507, + 272 + ], + "score": 1.0, + "content": "dim MFCC, 20-dim MFCC delta, 12-dim chroma, 384-dim tempogram, 1-dim onset strength (i.e.,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 270, + 330, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 330, + 281 + ], + "score": 1.0, + "content": "envelope) and 1-dim one-hot beat, as shown in Table 4.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "table", + "bbox": [ + 196, + 313, + 414, + 397 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 233, + 299, + 378, + 310 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 232, + 299, + 378, + 311 + ], + "spans": [ + { + "bbox": [ + 232, + 299, + 378, + 311 + ], + "score": 1.0, + "content": "Table 4: Acoustic features of music.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "table_body", + "bbox": [ + 196, + 313, + 414, + 397 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 196, + 313, + 414, + 397 + ], + "spans": [ + { + "bbox": [ + 196, + 313, + 414, + 397 + ], + "score": 0.978, + "html": "
FeatureCharacteristicDimension
MFCCPitch20
MFCC deltaPitch20
Constant-QchromagramPitch12
TempogramStrength384
Onset strengthStrength1
Beat one-hotBeat1
", + "type": "table", + "image_path": "b358b082044b77105e16eb9a4b95f99a56e408023a5b5130b4909fa107e10dbf.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 196, + 313, + 414, + 325.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 196, + 325.0, + 414, + 337.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 196, + 337.0, + 414, + 349.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 196, + 349.0, + 414, + 361.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 196, + 361.0, + 414, + 373.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 196, + 373.0, + 414, + 385.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 196, + 385.0, + 414, + 397.0 + ], + "spans": [], + "index": 21 + } + ] + } + ], + "index": 16.0 + }, + { + "type": "title", + "bbox": [ + 108, + 414, + 260, + 426 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 262, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 262, + 427 + ], + "score": 1.0, + "content": "6.3 DOWNSTREAM APPLICATIONS", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 434, + 506, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "score": 1.0, + "content": "The downstream applications of our proposed audio-conditioned dance generation method include:", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "score": 1.0, + "content": "(1) Our model can be used to help professional people choreograph new dances for a given song and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "score": 1.0, + "content": "teach human how to dance with regard to this song; (2) With the help of 3D human pose estimation", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "(Ci et al., 2019) and 3D animation driving techniques, our model can be used to drive the various 3D", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 478, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 492 + ], + "score": 1.0, + "content": "character models, such as the 3D model of Hatsune Miku (very popular virtual character in Japan).", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 490, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 506, + 502 + ], + "score": 1.0, + "content": "This technique has the great potential for the virtual advertisement video generation that could be", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 501, + 412, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 412, + 513 + ], + "score": 1.0, + "content": "used for the promotion events on social medias like TikTok, Twitter and etc.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26 + }, + { + "type": "title", + "bbox": [ + 107, + 525, + 386, + 536 + ], + "lines": [ + { + "bbox": [ + 106, + 524, + 387, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 387, + 538 + ], + "score": 1.0, + "content": "6.4 MUSICAL BEAT DETECTION UNDER LOW SAMPLING RATES", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 545, + 505, + 634 + ], + "lines": [ + { + "bbox": [ + 105, + 546, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 558 + ], + "score": 1.0, + "content": "In this section, we conduct an additional experiment to evaluate the performance of Librosa beat", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 557, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 162, + 569 + ], + "score": 1.0, + "content": "tracker under", + "type": "text" + }, + { + "bbox": [ + 163, + 557, + 203, + 567 + ], + "score": 0.4, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 557, + 505, + 569 + ], + "score": 1.0, + "content": "sampling rate. Specifically, we first utilize it to extract the onset beats from", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 568, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 170, + 579 + ], + "score": 1.0, + "content": "audio data with", + "type": "text" + }, + { + "bbox": [ + 171, + 568, + 210, + 578 + ], + "score": 0.68, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 568, + 227, + 579 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 228, + 568, + 268, + 578 + ], + "score": 0.71, + "content": "2 2 { , } 0 5 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 568, + 505, + 579 + ], + "score": 1.0, + "content": "respectively, then compare two groups of onset beats under", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 579, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 213, + 591 + ], + "score": 1.0, + "content": "the same time scale. Since", + "type": "text" + }, + { + "bbox": [ + 213, + 579, + 254, + 589 + ], + "score": 0.42, + "content": "2 2 , 0 5 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 579, + 506, + 591 + ], + "score": 1.0, + "content": "is a common sampling rate used in music information retrieval,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 590, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 255, + 602 + ], + "score": 1.0, + "content": "we define the beat alignment ratio as", + "type": "text" + }, + { + "bbox": [ + 255, + 590, + 285, + 601 + ], + "score": 0.92, + "content": "B _ { 2 } / B _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 590, + 316, + 602 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 316, + 590, + 329, + 600 + ], + "score": 0.89, + "content": "B _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 590, + 447, + 602 + ], + "score": 1.0, + "content": "is the number of beats under", + "type": "text" + }, + { + "bbox": [ + 447, + 590, + 487, + 600 + ], + "score": 0.75, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 590, + 506, + 602 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 107, + 601, + 120, + 612 + ], + "score": 0.88, + "content": "B _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 600, + 423, + 613 + ], + "score": 1.0, + "content": "is the number of beats under 15,400Hz that are aligned with the beats under", + "type": "text" + }, + { + "bbox": [ + 423, + 601, + 463, + 611 + ], + "score": 0.37, + "content": "2 2 { , } 0 5 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 600, + 506, + 613 + ], + "score": 1.0, + "content": ". One beat", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 611, + 504, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 172, + 624 + ], + "score": 1.0, + "content": "is counted when", + "type": "text" + }, + { + "bbox": [ + 173, + 611, + 248, + 623 + ], + "score": 0.89, + "content": "\\mid t _ { 1 } - t _ { 2 } \\mid \\leq \\Delta t . \\ t _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 612, + 430, + 624 + ], + "score": 1.0, + "content": "denotes the timestamp of a certain beat under", + "type": "text" + }, + { + "bbox": [ + 430, + 612, + 470, + 622 + ], + "score": 0.67, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 612, + 494, + 624 + ], + "score": 1.0, + "content": "while", + "type": "text" + }, + { + "bbox": [ + 495, + 612, + 504, + 622 + ], + "score": 0.84, + "content": "t _ { 2 }", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 623, + 461, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 292, + 634 + ], + "score": 1.0, + "content": "refers to the timestamp of a certain beat under", + "type": "text" + }, + { + "bbox": [ + 292, + 623, + 333, + 633 + ], + "score": 0.41, + "content": "2 2 { , } 0 5 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 623, + 336, + 634 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 336, + 623, + 349, + 632 + ], + "score": 0.76, + "content": "\\Delta t", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 623, + 461, + 634 + ], + "score": 1.0, + "content": "is the time offset threshold.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 639, + 505, + 716 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 211, + 651 + ], + "score": 1.0, + "content": "In the experiment, we set", + "type": "text" + }, + { + "bbox": [ + 211, + 639, + 262, + 651 + ], + "score": 0.92, + "content": "\\Delta t = 1 / 1 5 s", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 639, + 505, + 651 + ], + "score": 1.0, + "content": "(FPS is 15) and calculate the beat alignment ratio for audio", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "data of three styles respectively. We randomly sample 10 audio clips from each style to calculate", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 661, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 506, + 673 + ], + "score": 1.0, + "content": "the beat alignment ratio and do the sampling for 10 times to take the average. As shown in Table", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 672, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 232, + 685 + ], + "score": 1.0, + "content": "5, most of beats extracted with", + "type": "text" + }, + { + "bbox": [ + 233, + 672, + 273, + 683 + ], + "score": 0.57, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 672, + 425, + 685 + ], + "score": 1.0, + "content": "are aligned with those extracted with", + "type": "text" + }, + { + "bbox": [ + 425, + 672, + 466, + 683 + ], + "score": 0.35, + "content": "2 2 { , } 0 5 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 672, + 506, + 685 + ], + "score": 1.0, + "content": ". Besides,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 683, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 506, + 695 + ], + "score": 1.0, + "content": "we randomly sample an audio clip and visualize beat tracking curves of its first 20 seconds under", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 694, + 505, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 195, + 706 + ], + "score": 1.0, + "content": "the sampling rates of", + "type": "text" + }, + { + "bbox": [ + 195, + 694, + 236, + 705 + ], + "score": 0.47, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 694, + 254, + 706 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 254, + 694, + 295, + 705 + ], + "score": 0.52, + "content": "2 2 { , } 0 5 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 694, + 505, + 706 + ], + "score": 1.0, + "content": ". As we can see in Figure 5, most of the beats from", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 705, + 434, + 717 + ], + "spans": [ + { + "bbox": [ + 106, + 705, + 146, + 716 + ], + "score": 0.57, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 705, + 164, + 717 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 705, + 205, + 716 + ], + "score": 0.42, + "content": "2 2 { , } 0 5 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 705, + 379, + 717 + ], + "score": 1.0, + "content": "are aligned within the time offset threshold", + "type": "text" + }, + { + "bbox": [ + 380, + 705, + 430, + 717 + ], + "score": 0.92, + "content": "\\Delta t = 1 / 1 5 s", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 705, + 434, + 717 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 119, + 722, + 226, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 719, + 227, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 227, + 734 + ], + "score": 1.0, + "content": "3https://ffmpeg.org/", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 161, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 163, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 163, + 96 + ], + "score": 1.0, + "content": "APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 105, + 316, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 317, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 317, + 118 + ], + "score": 1.0, + "content": "6.1 MORE DETAILS ABOUT DATA COLLECTION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 126, + 505, + 203 + ], + "lines": [ + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "The dance videos are collected from YouTube by crowd workers, which are all solo dance videos", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 137, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 149 + ], + "score": 1.0, + "content": "with 30FPS. We trim the beginning and ending several seconds for each of collected videos to remove", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "the silence parts, and split them into one-minute video clips. Then, we extract 2D pose data from", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "score": 1.0, + "content": "these video clips by OpenPose (Cao et al., 2017) with 15FPS setting and collect 790 one-minute clips", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 169, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 505, + 182 + ], + "score": 1.0, + "content": "of pose data with 15FPS, totaling about 13 hours. Finally, we extract the corresponding audio data", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 180, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 506, + 194 + ], + "score": 1.0, + "content": "from video clips by FFmpeg3. The code is implemented based on PyTorch framework. MIT License", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 192, + 280, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 280, + 204 + ], + "score": 1.0, + "content": "will be used for the released code and data.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 126, + 506, + 204 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 216, + 263, + 227 + ], + "lines": [ + { + "bbox": [ + 106, + 215, + 265, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 265, + 229 + ], + "score": 1.0, + "content": "6.2 EXTRACTED MUSIC FEATURES", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 237, + 506, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 507, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 507, + 250 + ], + "score": 1.0, + "content": "In this section, we detail the extracted features of music that are feeded into our model. Specifically,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 248, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 506, + 260 + ], + "score": 1.0, + "content": "we leverage the public Librosa (McFee et al., 2015) to extract the music features including: 20-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 257, + 507, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 507, + 272 + ], + "score": 1.0, + "content": "dim MFCC, 20-dim MFCC delta, 12-dim chroma, 384-dim tempogram, 1-dim onset strength (i.e.,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 270, + 330, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 330, + 281 + ], + "score": 1.0, + "content": "envelope) and 1-dim one-hot beat, as shown in Table 4.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 236, + 507, + 281 + ] + }, + { + "type": "table", + "bbox": [ + 196, + 313, + 414, + 397 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 233, + 299, + 378, + 310 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 232, + 299, + 378, + 311 + ], + "spans": [ + { + "bbox": [ + 232, + 299, + 378, + 311 + ], + "score": 1.0, + "content": "Table 4: Acoustic features of music.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "table_body", + "bbox": [ + 196, + 313, + 414, + 397 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 196, + 313, + 414, + 397 + ], + "spans": [ + { + "bbox": [ + 196, + 313, + 414, + 397 + ], + "score": 0.978, + "html": "
FeatureCharacteristicDimension
MFCCPitch20
MFCC deltaPitch20
Constant-QchromagramPitch12
TempogramStrength384
Onset strengthStrength1
Beat one-hotBeat1
", + "type": "table", + "image_path": "b358b082044b77105e16eb9a4b95f99a56e408023a5b5130b4909fa107e10dbf.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 196, + 313, + 414, + 325.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 196, + 325.0, + 414, + 337.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 196, + 337.0, + 414, + 349.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 196, + 349.0, + 414, + 361.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 196, + 361.0, + 414, + 373.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 196, + 373.0, + 414, + 385.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 196, + 385.0, + 414, + 397.0 + ], + "spans": [], + "index": 21 + } + ] + } + ], + "index": 16.0 + }, + { + "type": "title", + "bbox": [ + 108, + 414, + 260, + 426 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 262, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 262, + 427 + ], + "score": 1.0, + "content": "6.3 DOWNSTREAM APPLICATIONS", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 434, + 506, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "score": 1.0, + "content": "The downstream applications of our proposed audio-conditioned dance generation method include:", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "score": 1.0, + "content": "(1) Our model can be used to help professional people choreograph new dances for a given song and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "score": 1.0, + "content": "teach human how to dance with regard to this song; (2) With the help of 3D human pose estimation", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "(Ci et al., 2019) and 3D animation driving techniques, our model can be used to drive the various 3D", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 478, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 492 + ], + "score": 1.0, + "content": "character models, such as the 3D model of Hatsune Miku (very popular virtual character in Japan).", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 490, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 506, + 502 + ], + "score": 1.0, + "content": "This technique has the great potential for the virtual advertisement video generation that could be", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 501, + 412, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 412, + 513 + ], + "score": 1.0, + "content": "used for the promotion events on social medias like TikTok, Twitter and etc.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 434, + 506, + 513 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 525, + 386, + 536 + ], + "lines": [ + { + "bbox": [ + 106, + 524, + 387, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 387, + 538 + ], + "score": 1.0, + "content": "6.4 MUSICAL BEAT DETECTION UNDER LOW SAMPLING RATES", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 545, + 505, + 634 + ], + "lines": [ + { + "bbox": [ + 105, + 546, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 558 + ], + "score": 1.0, + "content": "In this section, we conduct an additional experiment to evaluate the performance of Librosa beat", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 557, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 162, + 569 + ], + "score": 1.0, + "content": "tracker under", + "type": "text" + }, + { + "bbox": [ + 163, + 557, + 203, + 567 + ], + "score": 0.4, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 557, + 505, + 569 + ], + "score": 1.0, + "content": "sampling rate. Specifically, we first utilize it to extract the onset beats from", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 568, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 170, + 579 + ], + "score": 1.0, + "content": "audio data with", + "type": "text" + }, + { + "bbox": [ + 171, + 568, + 210, + 578 + ], + "score": 0.68, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 568, + 227, + 579 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 228, + 568, + 268, + 578 + ], + "score": 0.71, + "content": "2 2 { , } 0 5 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 568, + 505, + 579 + ], + "score": 1.0, + "content": "respectively, then compare two groups of onset beats under", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 579, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 213, + 591 + ], + "score": 1.0, + "content": "the same time scale. Since", + "type": "text" + }, + { + "bbox": [ + 213, + 579, + 254, + 589 + ], + "score": 0.42, + "content": "2 2 , 0 5 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 579, + 506, + 591 + ], + "score": 1.0, + "content": "is a common sampling rate used in music information retrieval,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 590, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 255, + 602 + ], + "score": 1.0, + "content": "we define the beat alignment ratio as", + "type": "text" + }, + { + "bbox": [ + 255, + 590, + 285, + 601 + ], + "score": 0.92, + "content": "B _ { 2 } / B _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 590, + 316, + 602 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 316, + 590, + 329, + 600 + ], + "score": 0.89, + "content": "B _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 590, + 447, + 602 + ], + "score": 1.0, + "content": "is the number of beats under", + "type": "text" + }, + { + "bbox": [ + 447, + 590, + 487, + 600 + ], + "score": 0.75, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 590, + 506, + 602 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 107, + 601, + 120, + 612 + ], + "score": 0.88, + "content": "B _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 600, + 423, + 613 + ], + "score": 1.0, + "content": "is the number of beats under 15,400Hz that are aligned with the beats under", + "type": "text" + }, + { + "bbox": [ + 423, + 601, + 463, + 611 + ], + "score": 0.37, + "content": "2 2 { , } 0 5 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 600, + 506, + 613 + ], + "score": 1.0, + "content": ". One beat", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 611, + 504, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 172, + 624 + ], + "score": 1.0, + "content": "is counted when", + "type": "text" + }, + { + "bbox": [ + 173, + 611, + 248, + 623 + ], + "score": 0.89, + "content": "\\mid t _ { 1 } - t _ { 2 } \\mid \\leq \\Delta t . \\ t _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 612, + 430, + 624 + ], + "score": 1.0, + "content": "denotes the timestamp of a certain beat under", + "type": "text" + }, + { + "bbox": [ + 430, + 612, + 470, + 622 + ], + "score": 0.67, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 612, + 494, + 624 + ], + "score": 1.0, + "content": "while", + "type": "text" + }, + { + "bbox": [ + 495, + 612, + 504, + 622 + ], + "score": 0.84, + "content": "t _ { 2 }", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 623, + 461, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 292, + 634 + ], + "score": 1.0, + "content": "refers to the timestamp of a certain beat under", + "type": "text" + }, + { + "bbox": [ + 292, + 623, + 333, + 633 + ], + "score": 0.41, + "content": "2 2 { , } 0 5 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 623, + 336, + 634 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 336, + 623, + 349, + 632 + ], + "score": 0.76, + "content": "\\Delta t", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 623, + 461, + 634 + ], + "score": 1.0, + "content": "is the time offset threshold.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 546, + 506, + 634 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 639, + 505, + 716 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 211, + 651 + ], + "score": 1.0, + "content": "In the experiment, we set", + "type": "text" + }, + { + "bbox": [ + 211, + 639, + 262, + 651 + ], + "score": 0.92, + "content": "\\Delta t = 1 / 1 5 s", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 639, + 505, + 651 + ], + "score": 1.0, + "content": "(FPS is 15) and calculate the beat alignment ratio for audio", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "data of three styles respectively. We randomly sample 10 audio clips from each style to calculate", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 661, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 506, + 673 + ], + "score": 1.0, + "content": "the beat alignment ratio and do the sampling for 10 times to take the average. As shown in Table", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 672, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 232, + 685 + ], + "score": 1.0, + "content": "5, most of beats extracted with", + "type": "text" + }, + { + "bbox": [ + 233, + 672, + 273, + 683 + ], + "score": 0.57, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 672, + 425, + 685 + ], + "score": 1.0, + "content": "are aligned with those extracted with", + "type": "text" + }, + { + "bbox": [ + 425, + 672, + 466, + 683 + ], + "score": 0.35, + "content": "2 2 { , } 0 5 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 672, + 506, + 685 + ], + "score": 1.0, + "content": ". Besides,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 683, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 506, + 695 + ], + "score": 1.0, + "content": "we randomly sample an audio clip and visualize beat tracking curves of its first 20 seconds under", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 694, + 505, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 195, + 706 + ], + "score": 1.0, + "content": "the sampling rates of", + "type": "text" + }, + { + "bbox": [ + 195, + 694, + 236, + 705 + ], + "score": 0.47, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 694, + 254, + 706 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 254, + 694, + 295, + 705 + ], + "score": 0.52, + "content": "2 2 { , } 0 5 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 694, + 505, + 706 + ], + "score": 1.0, + "content": ". As we can see in Figure 5, most of the beats from", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 705, + 434, + 717 + ], + "spans": [ + { + "bbox": [ + 106, + 705, + 146, + 716 + ], + "score": 0.57, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 705, + 164, + 717 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 705, + 205, + 716 + ], + "score": 0.42, + "content": "2 2 { , } 0 5 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 705, + 379, + 717 + ], + "score": 1.0, + "content": "are aligned within the time offset threshold", + "type": "text" + }, + { + "bbox": [ + 380, + 705, + 430, + 717 + ], + "score": 0.92, + "content": "\\Delta t = 1 / 1 5 s", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 705, + 434, + 717 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 639, + 506, + 717 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 216, + 106, + 396, + 164 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 221, + 90, + 389, + 102 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 220, + 88, + 390, + 103 + ], + "spans": [ + { + "bbox": [ + 220, + 88, + 356, + 103 + ], + "score": 1.0, + "content": "Table 5: The beat alignment ratio", + "type": "text" + }, + { + "bbox": [ + 357, + 90, + 386, + 102 + ], + "score": 0.91, + "content": "B _ { 2 } / B _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 88, + 390, + 103 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 216, + 106, + 396, + 164 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 216, + 106, + 396, + 164 + ], + "spans": [ + { + "bbox": [ + 216, + 106, + 396, + 164 + ], + "score": 0.972, + "html": "
CategoryBeat Alignment Ratio (%)
Ballet88.7
Hiphop93.2
Japanese Pop91.6
", + "type": "table", + "image_path": "c25230cbddf59843dade0c1962c09ff1911fd40378025f0a07c73d06bf2336f2.jpg" + } + ] + } + ], + "index": 2.5, + "virtual_lines": [ + { + "bbox": [ + 216, + 106, + 396, + 120.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 216, + 120.5, + 396, + 135.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 216, + 135.0, + 396, + 149.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 216, + 149.5, + 396, + 164.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.25 + }, + { + "type": "image", + "bbox": [ + 107, + 184, + 502, + 380 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 184, + 502, + 380 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 184, + 502, + 380 + ], + "spans": [ + { + "bbox": [ + 107, + 184, + 502, + 380 + ], + "score": 0.972, + "type": "image", + "image_path": "960e4e9bd2cd632b5f494cde13e98e1b266fa0bc90a5990a3085ee12ef8de7cc.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 107, + 184, + 502, + 249.33333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 249.33333333333331, + 502, + 314.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 107, + 314.66666666666663, + 502, + 379.99999999999994 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 388, + 505, + 423 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "Figure 5: Beat tracking curves for the first 20 seconds of a music clip randomly sampled from audio", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 239, + 412 + ], + "score": 1.0, + "content": "data under the sampling rates of", + "type": "text" + }, + { + "bbox": [ + 239, + 400, + 279, + 410 + ], + "score": 0.76, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 399, + 297, + 412 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 297, + 400, + 338, + 410 + ], + "score": 0.54, + "content": "2 2 { , } 0 5 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 399, + 505, + 412 + ], + "score": 1.0, + "content": ". Most of the beats are aligned within the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 411, + 246, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 191, + 423 + ], + "score": 1.0, + "content": "time offset threshold", + "type": "text" + }, + { + "bbox": [ + 191, + 411, + 242, + 423 + ], + "score": 0.92, + "content": "\\Delta t = 1 / 1 5 s", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 411, + 246, + 423 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + } + ], + "index": 7.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 216, + 106, + 396, + 164 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 221, + 90, + 389, + 102 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 220, + 88, + 390, + 103 + ], + "spans": [ + { + "bbox": [ + 220, + 88, + 356, + 103 + ], + "score": 1.0, + "content": "Table 5: The beat alignment ratio", + "type": "text" + }, + { + "bbox": [ + 357, + 90, + 386, + 102 + ], + "score": 0.91, + "content": "B _ { 2 } / B _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 88, + 390, + 103 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 216, + 106, + 396, + 164 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 216, + 106, + 396, + 164 + ], + "spans": [ + { + "bbox": [ + 216, + 106, + 396, + 164 + ], + "score": 0.972, + "html": "
CategoryBeat Alignment Ratio (%)
Ballet88.7
Hiphop93.2
Japanese Pop91.6
", + "type": "table", + "image_path": "c25230cbddf59843dade0c1962c09ff1911fd40378025f0a07c73d06bf2336f2.jpg" + } + ] + } + ], + "index": 2.5, + "virtual_lines": [ + { + "bbox": [ + 216, + 106, + 396, + 120.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 216, + 120.5, + 396, + 135.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 216, + 135.0, + 396, + 149.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 216, + 149.5, + 396, + 164.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.25 + }, + { + "type": "image", + "bbox": [ + 107, + 184, + 502, + 380 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 184, + 502, + 380 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 184, + 502, + 380 + ], + "spans": [ + { + "bbox": [ + 107, + 184, + 502, + 380 + ], + "score": 0.972, + "type": "image", + "image_path": "960e4e9bd2cd632b5f494cde13e98e1b266fa0bc90a5990a3085ee12ef8de7cc.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 107, + 184, + 502, + 249.33333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 249.33333333333331, + 502, + 314.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 107, + 314.66666666666663, + 502, + 379.99999999999994 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 388, + 505, + 423 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "Figure 5: Beat tracking curves for the first 20 seconds of a music clip randomly sampled from audio", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 239, + 412 + ], + "score": 1.0, + "content": "data under the sampling rates of", + "type": "text" + }, + { + "bbox": [ + 239, + 400, + 279, + 410 + ], + "score": 0.76, + "content": "1 5 { , } 4 0 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 399, + 297, + 412 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 297, + 400, + 338, + 410 + ], + "score": 0.54, + "content": "2 2 { , } 0 5 0 \\mathrm { H z }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 399, + 505, + 412 + ], + "score": 1.0, + "content": ". Most of the beats are aligned within the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 411, + 246, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 191, + 423 + ], + "score": 1.0, + "content": "time offset threshold", + "type": "text" + }, + { + "bbox": [ + 191, + 411, + 242, + 423 + ], + "score": 0.92, + "content": "\\Delta t = 1 / 1 5 s", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 411, + 246, + 423 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + } + ], + "index": 7.5 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/xGZG2kS5bFk/xGZG2kS5bFk_model.json b/parse/train/xGZG2kS5bFk/xGZG2kS5bFk_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9903932b6baed46238e3d7cc521c8a28eba7f8d5 --- /dev/null +++ b/parse/train/xGZG2kS5bFk/xGZG2kS5bFk_model.json @@ -0,0 +1,18787 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 648, + 1304, + 648, + 1304, + 1260, + 398, + 1260 + ], + "score": 0.986 + }, + { + "category_id": 1, + "poly": [ + 298, + 1389, + 1405, + 1389, + 1405, + 1725, + 298, + 1725 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1740, + 1404, + 1740, + 1404, + 1924, + 298, + 1924 + ], + "score": 0.977 + }, + { + "category_id": 0, + "poly": [ + 300, + 218, + 1407, + 218, + 1407, + 324, + 300, + 324 + ], + "score": 0.962 + }, + { + "category_id": 2, + "poly": [ + 330, + 1945, + 872, + 1945, + 872, + 2034, + 330, + 2034 + ], + "score": 0.953 + }, + { + "category_id": 0, + "poly": [ + 302, + 1319, + 573, + 1319, + 573, + 1354, + 302, + 1354 + ], + "score": 0.891 + }, + { + "category_id": 1, + "poly": [ + 312, + 371, + 1353, + 371, + 1353, + 501, + 312, + 501 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 773, + 581, + 927, + 581, + 927, + 615, + 773, + 615 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 104, + 299, + 104 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 856, + 2088, + 856, + 2112, + 840, + 2112 + ], + "score": 0.742 + }, + { + "category_id": 13, + "poly": [ + 617, + 374, + 684, + 374, + 684, + 406, + 617, + 406 + ], + "score": 0.85, + "latex": "\\mathbf { H } \\mathbf { u } ^ { 2 * }" + }, + { + "category_id": 13, + "poly": [ + 761, + 374, + 820, + 374, + 820, + 406, + 761, + 406 + ], + "score": 0.66, + "latex": "{ \\bf W } { \\bf u } ^ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 1048, + 1234, + 1067, + 1234, + 1067, + 1255, + 1048, + 1255 + ], + "score": 0.43, + "latex": "=" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 218.0, + 1410.0, + 218.0, + 1410.0, + 271.0, + 295.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 275.0, + 1295.0, + 275.0, + 1295.0, + 328.0, + 294.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1942.0, + 541.0, + 1942.0, + 541.0, + 1974.0, + 329.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1967.0, + 877.0, + 1967.0, + 877.0, + 2009.0, + 327.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1999.0, + 571.0, + 1999.0, + 571.0, + 2039.0, + 328.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1316.0, + 579.0, + 1316.0, + 579.0, + 1363.0, + 294.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 580.0, + 932.0, + 580.0, + 932.0, + 618.0, + 770.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 859.0, + 2087.0, + 859.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 651.0, + 1308.0, + 651.0, + 1308.0, + 684.0, + 394.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 682.0, + 1308.0, + 682.0, + 1308.0, + 718.0, + 394.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 710.0, + 1306.0, + 710.0, + 1306.0, + 748.0, + 393.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 740.0, + 1308.0, + 740.0, + 1308.0, + 780.0, + 392.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 774.0, + 1306.0, + 774.0, + 1306.0, + 807.0, + 394.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 804.0, + 1305.0, + 804.0, + 1305.0, + 836.0, + 394.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 833.0, + 1305.0, + 833.0, + 1305.0, + 868.0, + 393.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 863.0, + 1305.0, + 863.0, + 1305.0, + 899.0, + 394.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 895.0, + 1307.0, + 895.0, + 1307.0, + 930.0, + 394.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 926.0, + 1305.0, + 926.0, + 1305.0, + 958.0, + 393.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 955.0, + 1306.0, + 955.0, + 1306.0, + 991.0, + 394.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 985.0, + 1307.0, + 985.0, + 1307.0, + 1021.0, + 393.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1017.0, + 1306.0, + 1017.0, + 1306.0, + 1053.0, + 393.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1045.0, + 1306.0, + 1045.0, + 1306.0, + 1083.0, + 393.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1076.0, + 1309.0, + 1076.0, + 1309.0, + 1114.0, + 392.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1110.0, + 1306.0, + 1110.0, + 1306.0, + 1143.0, + 394.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1139.0, + 1304.0, + 1139.0, + 1304.0, + 1172.0, + 394.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1170.0, + 1304.0, + 1170.0, + 1304.0, + 1200.0, + 395.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1199.0, + 1306.0, + 1199.0, + 1306.0, + 1235.0, + 394.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1231.0, + 1047.0, + 1231.0, + 1047.0, + 1263.0, + 394.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1231.0, + 1262.0, + 1231.0, + 1262.0, + 1263.0, + 1068.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1387.0, + 1406.0, + 1387.0, + 1406.0, + 1427.0, + 292.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1420.0, + 1406.0, + 1420.0, + 1406.0, + 1454.0, + 295.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1449.0, + 1408.0, + 1449.0, + 1408.0, + 1484.0, + 293.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1476.0, + 1407.0, + 1476.0, + 1407.0, + 1519.0, + 292.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1509.0, + 1406.0, + 1509.0, + 1406.0, + 1543.0, + 295.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1541.0, + 1406.0, + 1541.0, + 1406.0, + 1576.0, + 293.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1571.0, + 1406.0, + 1571.0, + 1406.0, + 1608.0, + 292.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1603.0, + 1405.0, + 1603.0, + 1405.0, + 1638.0, + 295.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1633.0, + 1407.0, + 1633.0, + 1407.0, + 1667.0, + 293.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1661.0, + 1405.0, + 1661.0, + 1405.0, + 1701.0, + 293.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1693.0, + 1393.0, + 1693.0, + 1393.0, + 1728.0, + 295.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1737.0, + 1405.0, + 1737.0, + 1405.0, + 1778.0, + 292.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1767.0, + 1406.0, + 1767.0, + 1406.0, + 1809.0, + 292.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1800.0, + 1405.0, + 1800.0, + 1405.0, + 1837.0, + 293.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1832.0, + 1408.0, + 1832.0, + 1408.0, + 1868.0, + 293.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1864.0, + 1405.0, + 1864.0, + 1405.0, + 1895.0, + 296.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1893.0, + 1404.0, + 1893.0, + 1404.0, + 1925.0, + 296.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 369.0, + 616.0, + 369.0, + 616.0, + 413.0, + 308.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 369.0, + 760.0, + 369.0, + 760.0, + 413.0, + 685.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 369.0, + 1359.0, + 369.0, + 1359.0, + 413.0, + 821.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 404.0, + 1026.0, + 404.0, + 1026.0, + 441.0, + 308.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 436.0, + 1345.0, + 436.0, + 1345.0, + 476.0, + 311.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 468.0, + 1064.0, + 468.0, + 1064.0, + 504.0, + 311.0, + 504.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 308, + 1403, + 308, + 1403, + 643, + 298, + 643 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 660, + 1404, + 660, + 1404, + 1147, + 298, + 1147 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1331, + 1404, + 1331, + 1404, + 1607, + 298, + 1607 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1756, + 1405, + 1756, + 1405, + 1941, + 298, + 1941 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1162, + 1405, + 1162, + 1405, + 1315, + 298, + 1315 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 296, + 229, + 1400, + 229, + 1400, + 292, + 296, + 292 + ], + "score": 0.945 + }, + { + "category_id": 2, + "poly": [ + 331, + 2006, + 974, + 2006, + 974, + 2033, + 331, + 2033 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 301, + 1673, + 587, + 1673, + 587, + 1710, + 301, + 1710 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2112, + 841, + 2112 + ], + "score": 0.73 + }, + { + "category_id": 13, + "poly": [ + 297, + 844, + 316, + 844, + 316, + 870, + 297, + 870 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 2003.0, + 977.0, + 2003.0, + 977.0, + 2037.0, + 333.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1671.0, + 593.0, + 1671.0, + 593.0, + 1715.0, + 292.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2121.0, + 839.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 309.0, + 1403.0, + 309.0, + 1403.0, + 343.0, + 294.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 337.0, + 1405.0, + 337.0, + 1405.0, + 374.0, + 293.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 368.0, + 1405.0, + 368.0, + 1405.0, + 403.0, + 294.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 398.0, + 1405.0, + 398.0, + 1405.0, + 436.0, + 293.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 431.0, + 1405.0, + 431.0, + 1405.0, + 463.0, + 293.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 459.0, + 1407.0, + 459.0, + 1407.0, + 498.0, + 293.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 489.0, + 1407.0, + 489.0, + 1407.0, + 527.0, + 293.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 522.0, + 1405.0, + 522.0, + 1405.0, + 556.0, + 293.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 552.0, + 1405.0, + 552.0, + 1405.0, + 587.0, + 294.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 582.0, + 1405.0, + 582.0, + 1405.0, + 619.0, + 293.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 613.0, + 956.0, + 613.0, + 956.0, + 647.0, + 294.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 656.0, + 1405.0, + 656.0, + 1405.0, + 698.0, + 292.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 691.0, + 1404.0, + 691.0, + 1404.0, + 724.0, + 293.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 721.0, + 1405.0, + 721.0, + 1405.0, + 756.0, + 294.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 751.0, + 1405.0, + 751.0, + 1405.0, + 785.0, + 294.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 781.0, + 1405.0, + 781.0, + 1405.0, + 815.0, + 294.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 812.0, + 1405.0, + 812.0, + 1405.0, + 846.0, + 294.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 845.0, + 1402.0, + 845.0, + 1402.0, + 875.0, + 317.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 873.0, + 1404.0, + 873.0, + 1404.0, + 907.0, + 293.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 905.0, + 1405.0, + 905.0, + 1405.0, + 938.0, + 293.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 933.0, + 1408.0, + 933.0, + 1408.0, + 971.0, + 293.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 963.0, + 1405.0, + 963.0, + 1405.0, + 1001.0, + 293.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 990.0, + 1406.0, + 990.0, + 1406.0, + 1033.0, + 292.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1022.0, + 1405.0, + 1022.0, + 1405.0, + 1064.0, + 292.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1055.0, + 1402.0, + 1055.0, + 1402.0, + 1089.0, + 294.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1084.0, + 1404.0, + 1084.0, + 1404.0, + 1121.0, + 293.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1118.0, + 1408.0, + 1118.0, + 1408.0, + 1152.0, + 294.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1331.0, + 1405.0, + 1331.0, + 1405.0, + 1366.0, + 296.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1362.0, + 1404.0, + 1362.0, + 1404.0, + 1398.0, + 292.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1392.0, + 1406.0, + 1392.0, + 1406.0, + 1429.0, + 294.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1425.0, + 1404.0, + 1425.0, + 1404.0, + 1457.0, + 294.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1452.0, + 1405.0, + 1452.0, + 1405.0, + 1491.0, + 292.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1483.0, + 1405.0, + 1483.0, + 1405.0, + 1520.0, + 294.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1516.0, + 1404.0, + 1516.0, + 1404.0, + 1549.0, + 296.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1546.0, + 1406.0, + 1546.0, + 1406.0, + 1581.0, + 293.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1577.0, + 1298.0, + 1577.0, + 1298.0, + 1610.0, + 296.0, + 1610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1758.0, + 1403.0, + 1758.0, + 1403.0, + 1790.0, + 298.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1789.0, + 1408.0, + 1789.0, + 1408.0, + 1822.0, + 293.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1818.0, + 1406.0, + 1818.0, + 1406.0, + 1854.0, + 293.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1848.0, + 1406.0, + 1848.0, + 1406.0, + 1884.0, + 293.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1878.0, + 1406.0, + 1878.0, + 1406.0, + 1913.0, + 295.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1910.0, + 1197.0, + 1910.0, + 1197.0, + 1945.0, + 295.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1163.0, + 1406.0, + 1163.0, + 1406.0, + 1196.0, + 296.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1190.0, + 1406.0, + 1190.0, + 1406.0, + 1229.0, + 293.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1225.0, + 1403.0, + 1225.0, + 1403.0, + 1258.0, + 296.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1253.0, + 1408.0, + 1253.0, + 1408.0, + 1291.0, + 292.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1282.0, + 1363.0, + 1282.0, + 1363.0, + 1322.0, + 293.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 260.0, + 561.0, + 260.0, + 561.0, + 295.0, + 296.0, + 295.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 633, + 1405, + 633, + 1405, + 969, + 298, + 969 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1405, + 229, + 1405, + 596, + 297, + 596 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 299, + 1087, + 1404, + 1087, + 1404, + 1180, + 299, + 1180 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 346, + 1213, + 1343, + 1213, + 1343, + 1563, + 346, + 1563 + ], + "score": 0.973 + }, + { + "category_id": 4, + "poly": [ + 296, + 1593, + 1405, + 1593, + 1405, + 1839, + 296, + 1839 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 298, + 1969, + 1401, + 1969, + 1401, + 2037, + 298, + 2037 + ], + "score": 0.946 + }, + { + "category_id": 0, + "poly": [ + 299, + 1016, + 508, + 1016, + 508, + 1051, + 299, + 1051 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 813, + 76, + 813, + 104, + 299, + 104 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 301, + 1915, + 690, + 1915, + 690, + 1944, + 301, + 1944 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.581 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.532 + }, + { + "category_id": 13, + "poly": [ + 815, + 2003, + 964, + 2003, + 964, + 2036, + 815, + 2036 + ], + "score": 0.93, + "latex": "\\bar { Y = } \\{ y _ { t } \\} _ { t = 1 } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 930, + 1973, + 1085, + 1973, + 1085, + 2006, + 930, + 2006 + ], + "score": 0.93, + "latex": "X = \\{ x _ { t } \\} _ { t = 1 } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 629, + 1970, + 846, + 1970, + 846, + 2006, + 629, + 2006 + ], + "score": 0.91, + "latex": "\\mathcal { D } = \\{ ( X _ { i } , Y _ { i } ) \\} _ { i = 1 } ^ { N }" + }, + { + "category_id": 13, + "poly": [ + 1130, + 1686, + 1159, + 1686, + 1159, + 1716, + 1130, + 1716 + ], + "score": 0.88, + "latex": "h _ { i }" + }, + { + "category_id": 13, + "poly": [ + 516, + 1720, + 545, + 1720, + 545, + 1748, + 516, + 1748 + ], + "score": 0.85, + "latex": "y _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 371, + 1720, + 396, + 1720, + 396, + 1747, + 371, + 1747 + ], + "score": 0.84, + "latex": "z _ { i }" + }, + { + "category_id": 13, + "poly": [ + 606, + 1690, + 633, + 1690, + 633, + 1718, + 606, + 1718 + ], + "score": 0.84, + "latex": "y _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1305, + 1976, + 1334, + 1976, + 1334, + 2003, + 1305, + 2003 + ], + "score": 0.84, + "latex": "x _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1202, + 1596, + 1231, + 1596, + 1231, + 1622, + 1202, + 1622 + ], + "score": 0.77, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 743, + 2008, + 757, + 2008, + 757, + 2030, + 743, + 2030 + ], + "score": 0.77, + "latex": "t" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 1219.0, + 1329.0, + 1219.0, + 1329.0, + 1242.0, + 1280.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 1243.0, + 549.0, + 1243.0, + 549.0, + 1265.0, + 518.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 1243.0, + 607.0, + 1243.0, + 607.0, + 1265.0, + 578.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1243.0, + 639.0, + 1243.0, + 639.0, + 1265.0, + 610.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 1243.0, + 670.0, + 1243.0, + 670.0, + 1265.0, + 641.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1242.0, + 729.0, + 1242.0, + 729.0, + 1265.0, + 702.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 1240.0, + 805.0, + 1240.0, + 805.0, + 1268.0, + 775.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1263.0, + 536.0, + 1263.0, + 536.0, + 1271.0, + 522.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1263.0, + 568.0, + 1263.0, + 568.0, + 1272.0, + 553.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1261.0, + 598.0, + 1261.0, + 598.0, + 1272.0, + 586.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 1260.0, + 632.0, + 1260.0, + 632.0, + 1274.0, + 615.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 1263.0, + 659.0, + 1263.0, + 659.0, + 1272.0, + 646.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1263.0, + 689.0, + 1263.0, + 689.0, + 1273.0, + 678.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1260.0, + 723.0, + 1260.0, + 723.0, + 1274.0, + 707.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 1261.0, + 793.0, + 1261.0, + 793.0, + 1272.0, + 783.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 1267.0, + 767.0, + 1267.0, + 767.0, + 1283.0, + 737.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 1273.0, + 1334.0, + 1273.0, + 1334.0, + 1297.0, + 1273.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1300.0, + 766.0, + 1300.0, + 766.0, + 1322.0, + 553.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 1312.0, + 856.0, + 1312.0, + 856.0, + 1342.0, + 818.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 1334.0, + 768.0, + 1334.0, + 768.0, + 1356.0, + 550.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 1338.0, + 1345.0, + 1338.0, + 1345.0, + 1372.0, + 1266.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1372.0, + 565.0, + 1372.0, + 565.0, + 1392.0, + 552.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 1373.0, + 595.0, + 1373.0, + 595.0, + 1383.0, + 585.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 1373.0, + 626.0, + 1373.0, + 626.0, + 1382.0, + 616.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1373.0, + 687.0, + 1373.0, + 687.0, + 1383.0, + 678.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 1373.0, + 719.0, + 1373.0, + 719.0, + 1382.0, + 710.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1374.0, + 769.0, + 1374.0, + 769.0, + 1395.0, + 734.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 1373.0, + 792.0, + 1373.0, + 792.0, + 1382.0, + 783.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1390.0, + 486.0, + 1390.0, + 486.0, + 1416.0, + 347.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 1396.0, + 733.0, + 1396.0, + 733.0, + 1425.0, + 513.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 1397.0, + 806.0, + 1397.0, + 806.0, + 1425.0, + 773.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 1435.0, + 638.0, + 1435.0, + 638.0, + 1464.0, + 491.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1465.0, + 898.0, + 1465.0, + 898.0, + 1492.0, + 867.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1464.0, + 945.0, + 1464.0, + 945.0, + 1492.0, + 912.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1464.0, + 988.0, + 1464.0, + 988.0, + 1492.0, + 955.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 1464.0, + 1038.0, + 1464.0, + 1038.0, + 1492.0, + 1006.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 1464.0, + 1083.0, + 1464.0, + 1083.0, + 1494.0, + 1048.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 1467.0, + 1128.0, + 1467.0, + 1128.0, + 1493.0, + 1095.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 1467.0, + 1175.0, + 1467.0, + 1175.0, + 1493.0, + 1143.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1485.0, + 1215.0, + 1485.0, + 1215.0, + 1501.0, + 1185.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 1467.0, + 1263.0, + 1467.0, + 1263.0, + 1495.0, + 1231.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1505.0, + 499.0, + 1505.0, + 499.0, + 1531.0, + 351.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 1502.0, + 900.0, + 1502.0, + 900.0, + 1523.0, + 862.0, + 1523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1519.0, + 1025.0, + 1519.0, + 1025.0, + 1542.0, + 915.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 1518.0, + 1166.0, + 1518.0, + 1166.0, + 1544.0, + 1051.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1536.0, + 1033.0, + 1536.0, + 1033.0, + 1561.0, + 914.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 1536.0, + 1170.0, + 1536.0, + 1170.0, + 1561.0, + 1051.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 1241.0, + 574.0, + 1241.0, + 574.0, + 1264.5, + 547.0, + 1264.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 1241.5, + 696.0, + 1241.5, + 696.0, + 1264.0, + 670.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1595.0, + 1201.0, + 1595.0, + 1201.0, + 1628.0, + 295.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 1595.0, + 1403.0, + 1595.0, + 1403.0, + 1628.0, + 1232.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1625.0, + 1403.0, + 1625.0, + 1403.0, + 1657.0, + 294.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1654.0, + 1403.0, + 1654.0, + 1403.0, + 1688.0, + 294.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1686.0, + 605.0, + 1686.0, + 605.0, + 1720.0, + 295.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1686.0, + 1129.0, + 1686.0, + 1129.0, + 1720.0, + 634.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 1686.0, + 1403.0, + 1686.0, + 1403.0, + 1720.0, + 1160.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1716.0, + 370.0, + 1716.0, + 370.0, + 1750.0, + 295.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 1716.0, + 515.0, + 1716.0, + 515.0, + 1750.0, + 397.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 1716.0, + 1405.0, + 1716.0, + 1405.0, + 1750.0, + 546.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1744.0, + 1405.0, + 1744.0, + 1405.0, + 1785.0, + 291.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1776.0, + 1406.0, + 1776.0, + 1406.0, + 1812.0, + 294.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1809.0, + 1126.0, + 1809.0, + 1126.0, + 1841.0, + 295.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1011.0, + 514.0, + 1011.0, + 514.0, + 1059.0, + 291.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1911.0, + 695.0, + 1911.0, + 695.0, + 1949.0, + 294.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 631.0, + 1407.0, + 631.0, + 1407.0, + 668.0, + 293.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 665.0, + 1406.0, + 665.0, + 1406.0, + 699.0, + 293.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 694.0, + 1405.0, + 694.0, + 1405.0, + 729.0, + 295.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 724.0, + 1405.0, + 724.0, + 1405.0, + 759.0, + 293.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 755.0, + 1406.0, + 755.0, + 1406.0, + 789.0, + 292.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 785.0, + 1407.0, + 785.0, + 1407.0, + 820.0, + 292.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 815.0, + 1405.0, + 815.0, + 1405.0, + 851.0, + 293.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 844.0, + 1406.0, + 844.0, + 1406.0, + 885.0, + 292.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 878.0, + 1407.0, + 878.0, + 1407.0, + 912.0, + 295.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 909.0, + 1405.0, + 909.0, + 1405.0, + 943.0, + 293.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 938.0, + 907.0, + 938.0, + 907.0, + 973.0, + 295.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 226.0, + 1405.0, + 226.0, + 1405.0, + 267.0, + 292.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 260.0, + 1406.0, + 260.0, + 1406.0, + 297.0, + 291.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 292.0, + 1406.0, + 292.0, + 1406.0, + 326.0, + 295.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 319.0, + 1405.0, + 319.0, + 1405.0, + 358.0, + 294.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 353.0, + 1405.0, + 353.0, + 1405.0, + 387.0, + 295.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 383.0, + 1407.0, + 383.0, + 1407.0, + 417.0, + 295.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 412.0, + 1410.0, + 412.0, + 1410.0, + 450.0, + 294.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 443.0, + 1406.0, + 443.0, + 1406.0, + 476.0, + 294.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 475.0, + 1406.0, + 475.0, + 1406.0, + 509.0, + 294.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 505.0, + 1405.0, + 505.0, + 1405.0, + 539.0, + 294.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 535.0, + 1405.0, + 535.0, + 1405.0, + 569.0, + 292.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 565.0, + 1266.0, + 565.0, + 1266.0, + 599.0, + 295.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1086.0, + 1405.0, + 1086.0, + 1405.0, + 1124.0, + 293.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1119.0, + 1404.0, + 1119.0, + 1404.0, + 1153.0, + 295.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1148.0, + 1308.0, + 1148.0, + 1308.0, + 1184.0, + 294.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1962.0, + 628.0, + 1962.0, + 628.0, + 2016.0, + 291.0, + 2016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 1962.0, + 929.0, + 1962.0, + 929.0, + 2016.0, + 847.0, + 2016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 1962.0, + 1304.0, + 1962.0, + 1304.0, + 2016.0, + 1086.0, + 2016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 1962.0, + 1410.0, + 1962.0, + 1410.0, + 2016.0, + 1335.0, + 2016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1995.0, + 742.0, + 1995.0, + 742.0, + 2040.0, + 291.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 1995.0, + 814.0, + 1995.0, + 814.0, + 2040.0, + 758.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 1995.0, + 1406.0, + 1995.0, + 1406.0, + 2040.0, + 965.0, + 2040.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 602, + 1406, + 602, + 1406, + 1001, + 297, + 1001 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 297, + 1557, + 1405, + 1557, + 1405, + 1744, + 297, + 1744 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1844, + 1404, + 1844, + 1404, + 2035, + 297, + 2035 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 445, + 1404, + 445, + 1404, + 573, + 297, + 573 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 300, + 1210, + 1405, + 1210, + 1405, + 1305, + 300, + 1305 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 228, + 1404, + 228, + 1404, + 354, + 299, + 354 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1426, + 1403, + 1426, + 1403, + 1525, + 299, + 1525 + ], + "score": 0.977 + }, + { + "category_id": 8, + "poly": [ + 518, + 1094, + 1180, + 1094, + 1180, + 1196, + 518, + 1196 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 524, + 1319, + 1175, + 1319, + 1175, + 1414, + 524, + 1414 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 299, + 1014, + 1398, + 1014, + 1398, + 1080, + 299, + 1080 + ], + "score": 0.952 + }, + { + "category_id": 0, + "poly": [ + 299, + 390, + 655, + 390, + 655, + 422, + 299, + 422 + ], + "score": 0.908 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.904 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1128, + 1400, + 1128, + 1400, + 1158, + 1366, + 1158 + ], + "score": 0.887 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1347, + 1400, + 1347, + 1400, + 1377, + 1366, + 1377 + ], + "score": 0.886 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1763, + 1399, + 1763, + 1399, + 1792, + 1366, + 1792 + ], + "score": 0.878 + }, + { + "category_id": 9, + "poly": [ + 1367, + 1802, + 1399, + 1802, + 1399, + 1830, + 1367, + 1830 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 857, + 2089, + 857, + 2111, + 840, + 2111 + ], + "score": 0.797 + }, + { + "category_id": 8, + "poly": [ + 710, + 1756, + 989, + 1756, + 989, + 1835, + 710, + 1835 + ], + "score": 0.796 + }, + { + "category_id": 8, + "poly": [ + 712, + 1757, + 953, + 1757, + 953, + 1794, + 712, + 1794 + ], + "score": 0.356 + }, + { + "category_id": 8, + "poly": [ + 712, + 1798, + 987, + 1798, + 987, + 1834, + 712, + 1834 + ], + "score": 0.33 + }, + { + "category_id": 14, + "poly": [ + 524, + 1318, + 1173, + 1318, + 1173, + 1414, + 524, + 1414 + ], + "score": 0.94, + "latex": "\\alpha _ { i j } = \\frac { \\exp { e _ { i j } } } { \\sum _ { t = j - \\lfloor k / 2 \\rfloor } ^ { j + \\lfloor k / 2 \\rfloor } \\exp { e _ { i t } } } , e _ { i j } = \\frac { ( u _ { i } W _ { l } ^ { Q } ) ( u _ { j } W _ { l } ^ { K } ) ^ { \\top } } { \\sqrt { d _ { k } } } ," + }, + { + "category_id": 14, + "poly": [ + 517, + 1092, + 1180, + 1092, + 1180, + 1197, + 517, + 1197 + ], + "score": 0.94, + "latex": "z _ { i } = \\mathcal { F } ( a _ { i } ) , \\quad a _ { i } = \\sum _ { j = i - \\lfloor k / 2 \\rfloor } ^ { i + \\lfloor k / 2 \\rfloor } \\alpha _ { i j } ( u _ { j } W _ { l } ^ { V } ) , \\quad U = X W ^ { E } ," + }, + { + "category_id": 13, + "poly": [ + 845, + 539, + 1030, + 539, + 1030, + 573, + 845, + 573 + ], + "score": 0.93, + "latex": "Y = ( y _ { 1 } , . . . , y _ { n } )" + }, + { + "category_id": 13, + "poly": [ + 895, + 1428, + 1064, + 1428, + 1064, + 1466, + 895, + 1466 + ], + "score": 0.93, + "latex": "W _ { l } ^ { V } \\in \\mathbb { R } ^ { d _ { z } \\times d _ { v } }" + }, + { + "category_id": 13, + "poly": [ + 353, + 1711, + 535, + 1711, + 535, + 1744, + 353, + 1744 + ], + "score": 0.93, + "latex": "Z = ( z _ { 1 } , . . . , z _ { n } )" + }, + { + "category_id": 13, + "poly": [ + 792, + 1711, + 978, + 1711, + 978, + 1744, + 792, + 1744 + ], + "score": 0.93, + "latex": "\\boldsymbol { Y } = ( \\hat { y } _ { 1 } , . . . , \\hat { y } _ { n } )" + }, + { + "category_id": 13, + "poly": [ + 758, + 694, + 834, + 694, + 834, + 729, + 758, + 729 + ], + "score": 0.93, + "latex": "\\scriptstyle { \\mathcal { O } } ( n ^ { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 373, + 1210, + 588, + 1210, + 588, + 1246, + 373, + 1246 + ], + "score": 0.92, + "latex": "\\mathcal { F } ( \\cdot ) : \\mathbb { R } ^ { d _ { v } } \\mathbb { R } ^ { d _ { z } }" + }, + { + "category_id": 13, + "poly": [ + 332, + 1045, + 502, + 1045, + 502, + 1076, + 332, + 1076 + ], + "score": 0.92, + "latex": "W ^ { E } \\in \\mathbf { \\bar { \\mathbb { R } } } ^ { d _ { x } \\times d _ { z } }" + }, + { + "category_id": 13, + "poly": [ + 420, + 1907, + 641, + 1907, + 641, + 1940, + 420, + 1940 + ], + "score": 0.91, + "latex": "W ^ { S } \\in \\mathbb { R } ^ { ( d _ { s } + d _ { z } ) \\times d _ { y } }" + }, + { + "category_id": 13, + "poly": [ + 849, + 1016, + 1038, + 1016, + 1038, + 1050, + 849, + 1050 + ], + "score": 0.91, + "latex": "U = ( u _ { 1 } , . . . , u _ { n } )" + }, + { + "category_id": 13, + "poly": [ + 1015, + 262, + 1077, + 262, + 1077, + 294, + 1015, + 294 + ], + "score": 0.91, + "latex": "g ( X )" + }, + { + "category_id": 13, + "poly": [ + 579, + 1049, + 746, + 1049, + 746, + 1081, + 579, + 1081 + ], + "score": 0.91, + "latex": "\\forall i \\in \\{ 1 , . . . , n \\}" + }, + { + "category_id": 14, + "poly": [ + 710, + 1753, + 986, + 1753, + 986, + 1840, + 710, + 1840 + ], + "score": 0.91, + "latex": "\\begin{array} { l } { \\hat { y } _ { i } = [ h _ { i } ; z _ { i } ] W ^ { S } + b , } \\\\ { h _ { i } = \\mathrm { R N N } ( h _ { i - 1 } , \\hat { y } _ { i - 1 } ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 603, + 1016, + 795, + 1016, + 795, + 1049, + 603, + 1049 + ], + "score": 0.91, + "latex": "X = ( x _ { 1 } , . . . , x _ { n } )" + }, + { + "category_id": 13, + "poly": [ + 898, + 818, + 975, + 818, + 975, + 851, + 898, + 851 + ], + "score": 0.9, + "latex": "\\mathcal { O } ( n k )" + }, + { + "category_id": 13, + "poly": [ + 697, + 1911, + 786, + 1911, + 786, + 1940, + 697, + 1940 + ], + "score": 0.9, + "latex": "b \\in \\mathbb { R } ^ { d _ { y } }" + }, + { + "category_id": 13, + "poly": [ + 952, + 230, + 998, + 230, + 998, + 264, + 952, + 264 + ], + "score": 0.9, + "latex": "g ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 616, + 1494, + 646, + 1494, + 646, + 1523, + 616, + 1523 + ], + "score": 0.89, + "latex": "d _ { v }" + }, + { + "category_id": 13, + "poly": [ + 875, + 1848, + 906, + 1848, + 906, + 1878, + 875, + 1878 + ], + "score": 0.89, + "latex": "h _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 890, + 1277, + 930, + 1277, + 930, + 1308, + 890, + 1308 + ], + "score": 0.89, + "latex": "\\alpha _ { i j }" + }, + { + "category_id": 13, + "poly": [ + 1122, + 1912, + 1152, + 1912, + 1152, + 1945, + 1122, + 1945 + ], + "score": 0.89, + "latex": "d _ { y }" + }, + { + "category_id": 13, + "poly": [ + 371, + 1848, + 400, + 1848, + 400, + 1878, + 371, + 1878 + ], + "score": 0.88, + "latex": "h _ { i }" + }, + { + "category_id": 13, + "poly": [ + 548, + 447, + 593, + 447, + 593, + 480, + 548, + 480 + ], + "score": 0.88, + "latex": "g ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1041, + 1914, + 1069, + 1914, + 1069, + 1942, + 1041, + 1942 + ], + "score": 0.88, + "latex": "d _ { s }" + }, + { + "category_id": 13, + "poly": [ + 1062, + 1878, + 1107, + 1878, + 1107, + 1911, + 1062, + 1911 + ], + "score": 0.88, + "latex": "[ \\cdot ; \\cdot ]" + }, + { + "category_id": 13, + "poly": [ + 1174, + 1945, + 1200, + 1945, + 1200, + 1974, + 1174, + 1974 + ], + "score": 0.88, + "latex": "\\hat { y } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1304, + 1464, + 1335, + 1464, + 1335, + 1493, + 1304, + 1493 + ], + "score": 0.88, + "latex": "d _ { k }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1944, + 325, + 1944, + 325, + 1973, + 297, + 1973 + ], + "score": 0.88, + "latex": "h _ { i }" + }, + { + "category_id": 13, + "poly": [ + 375, + 1944, + 400, + 1944, + 400, + 1974, + 375, + 1974 + ], + "score": 0.87, + "latex": "\\hat { y } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 539, + 882, + 566, + 882, + 566, + 909, + 539, + 909 + ], + "score": 0.86, + "latex": "z _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1373, + 1946, + 1400, + 1946, + 1400, + 1973, + 1373, + 1973 + ], + "score": 0.86, + "latex": "h _ { i }" + }, + { + "category_id": 13, + "poly": [ + 817, + 1464, + 847, + 1464, + 847, + 1493, + 817, + 1493 + ], + "score": 0.86, + "latex": "d _ { z }" + }, + { + "category_id": 13, + "poly": [ + 1048, + 1217, + 1078, + 1217, + 1078, + 1246, + 1048, + 1246 + ], + "score": 0.86, + "latex": "u _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1206, + 1467, + 1232, + 1467, + 1232, + 1493, + 1206, + 1493 + ], + "score": 0.85, + "latex": "z _ { i }" + }, + { + "category_id": 13, + "poly": [ + 670, + 1427, + 842, + 1427, + 842, + 1466, + 670, + 1466 + ], + "score": 0.84, + "latex": "W _ { l } ^ { K } \\in \\mathbb { R } ^ { d _ { z } \\times d _ { k } }" + }, + { + "category_id": 13, + "poly": [ + 442, + 235, + 470, + 235, + 470, + 261, + 442, + 261 + ], + "score": 0.84, + "latex": "x _ { t }" + }, + { + "category_id": 13, + "poly": [ + 771, + 1986, + 795, + 1986, + 795, + 2003, + 771, + 2003 + ], + "score": 0.84, + "latex": "z _ { i }" + }, + { + "category_id": 13, + "poly": [ + 788, + 1244, + 806, + 1244, + 806, + 1270, + 788, + 1270 + ], + "score": 0.84, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1112, + 883, + 1138, + 883, + 1138, + 910, + 1112, + 910 + ], + "score": 0.84, + "latex": "y _ { t }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1245, + 316, + 1245, + 316, + 1270, + 298, + 1270 + ], + "score": 0.83, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 607, + 1426, + 657, + 1426, + 657, + 1466, + 607, + 1466 + ], + "score": 0.83, + "latex": "W _ { l } ^ { Q }" + }, + { + "category_id": 13, + "poly": [ + 1294, + 789, + 1313, + 789, + 1313, + 815, + 1294, + 815 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 360, + 263, + 388, + 263, + 388, + 289, + 360, + 289 + ], + "score": 0.82, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 879, + 262, + 908, + 262, + 908, + 290, + 879, + 290 + ], + "score": 0.82, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 297, + 236, + 323, + 236, + 323, + 262, + 297, + 262 + ], + "score": 0.82, + "latex": "y _ { t }" + }, + { + "category_id": 13, + "poly": [ + 987, + 447, + 1187, + 447, + 1187, + 480, + 987, + 480 + ], + "score": 0.81, + "latex": "X = ( x _ { 1 } , . . . , x _ { n } )" + }, + { + "category_id": 13, + "poly": [ + 1207, + 541, + 1231, + 541, + 1231, + 567, + 1207, + 567 + ], + "score": 0.79, + "latex": "Z" + }, + { + "category_id": 13, + "poly": [ + 298, + 1465, + 322, + 1465, + 322, + 1490, + 298, + 1490 + ], + "score": 0.78, + "latex": "U" + }, + { + "category_id": 13, + "poly": [ + 629, + 1947, + 641, + 1947, + 641, + 1970, + 629, + 1970 + ], + "score": 0.77, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 493, + 1434, + 505, + 1434, + 505, + 1460, + 493, + 1460 + ], + "score": 0.76, + "latex": "l" + }, + { + "category_id": 13, + "poly": [ + 1062, + 232, + 1086, + 232, + 1086, + 258, + 1062, + 258 + ], + "score": 0.75, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 1090, + 701, + 1111, + 701, + 1111, + 723, + 1090, + 723 + ], + "score": 0.74, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 471, + 1849, + 484, + 1849, + 484, + 1874, + 471, + 1874 + ], + "score": 0.73, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 749, + 263, + 776, + 263, + 776, + 290, + 749, + 290 + ], + "score": 0.73, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 757, + 1052, + 782, + 1052, + 782, + 1079, + 757, + 1079 + ], + "score": 0.72, + "latex": "z _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1196, + 445, + 1323, + 445, + 1323, + 479, + 1196, + 479 + ], + "score": 0.71, + "latex": "( x _ { i } \\in \\mathbb { R } ^ { d _ { x } } )" + }, + { + "category_id": 13, + "poly": [ + 988, + 819, + 1007, + 819, + 1007, + 845, + 988, + 845 + ], + "score": 0.71, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1280, + 882, + 1293, + 882, + 1293, + 906, + 1280, + 906 + ], + "score": 0.7, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 684, + 476, + 801, + 476, + 801, + 510, + 684, + 510 + ], + "score": 0.68, + "latex": "( z _ { i } \\in \\mathbb { R } ^ { d _ { z } } )" + }, + { + "category_id": 13, + "poly": [ + 493, + 479, + 676, + 479, + 676, + 512, + 493, + 512 + ], + "score": 0.65, + "latex": "Z = ( z _ { 1 } , . . . , z _ { n } )" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 390.0, + 655.0, + 390.0, + 655.0, + 422.0, + 296.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 601.0, + 1408.0, + 601.0, + 1408.0, + 640.0, + 294.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 631.0, + 1406.0, + 631.0, + 1406.0, + 669.0, + 294.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 665.0, + 1403.0, + 665.0, + 1403.0, + 698.0, + 295.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 695.0, + 757.0, + 695.0, + 757.0, + 731.0, + 294.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 695.0, + 1089.0, + 695.0, + 1089.0, + 731.0, + 835.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 695.0, + 1406.0, + 695.0, + 1406.0, + 731.0, + 1112.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 726.0, + 1406.0, + 726.0, + 1406.0, + 763.0, + 294.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 757.0, + 1406.0, + 757.0, + 1406.0, + 790.0, + 296.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 787.0, + 1293.0, + 787.0, + 1293.0, + 820.0, + 295.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 787.0, + 1404.0, + 787.0, + 1404.0, + 820.0, + 1314.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 816.0, + 897.0, + 816.0, + 897.0, + 852.0, + 294.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 816.0, + 987.0, + 816.0, + 987.0, + 852.0, + 976.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 816.0, + 1406.0, + 816.0, + 1406.0, + 852.0, + 1008.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 847.0, + 1407.0, + 847.0, + 1407.0, + 882.0, + 294.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 878.0, + 538.0, + 878.0, + 538.0, + 915.0, + 294.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 878.0, + 1111.0, + 878.0, + 1111.0, + 915.0, + 567.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 878.0, + 1279.0, + 878.0, + 1279.0, + 915.0, + 1139.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 878.0, + 1406.0, + 878.0, + 1406.0, + 915.0, + 1294.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 907.0, + 1404.0, + 907.0, + 1404.0, + 946.0, + 294.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 938.0, + 1404.0, + 938.0, + 1404.0, + 976.0, + 294.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 965.0, + 950.0, + 965.0, + 950.0, + 1008.0, + 293.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1557.0, + 1406.0, + 1557.0, + 1406.0, + 1593.0, + 295.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1586.0, + 1406.0, + 1586.0, + 1406.0, + 1624.0, + 294.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1621.0, + 1405.0, + 1621.0, + 1405.0, + 1653.0, + 295.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1648.0, + 1406.0, + 1648.0, + 1406.0, + 1685.0, + 292.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1677.0, + 1408.0, + 1677.0, + 1408.0, + 1716.0, + 294.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1708.0, + 352.0, + 1708.0, + 352.0, + 1747.0, + 292.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 1708.0, + 791.0, + 1708.0, + 791.0, + 1747.0, + 536.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 1708.0, + 1201.0, + 1708.0, + 1201.0, + 1747.0, + 979.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1845.0, + 370.0, + 1845.0, + 370.0, + 1881.0, + 295.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 1845.0, + 470.0, + 1845.0, + 470.0, + 1881.0, + 401.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 1845.0, + 874.0, + 1845.0, + 874.0, + 1881.0, + 485.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 1845.0, + 1405.0, + 1845.0, + 1405.0, + 1881.0, + 907.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1876.0, + 1061.0, + 1876.0, + 1061.0, + 1912.0, + 294.0, + 1912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 1876.0, + 1405.0, + 1876.0, + 1405.0, + 1912.0, + 1108.0, + 1912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1901.0, + 419.0, + 1901.0, + 419.0, + 1951.0, + 288.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1901.0, + 696.0, + 1901.0, + 696.0, + 1951.0, + 642.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 1901.0, + 1040.0, + 1901.0, + 1040.0, + 1951.0, + 787.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 1901.0, + 1121.0, + 1901.0, + 1121.0, + 1951.0, + 1070.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1901.0, + 1411.0, + 1901.0, + 1411.0, + 1951.0, + 1153.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1942.0, + 374.0, + 1942.0, + 374.0, + 1978.0, + 326.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 1942.0, + 628.0, + 1942.0, + 628.0, + 1978.0, + 401.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1942.0, + 1173.0, + 1942.0, + 1173.0, + 1978.0, + 642.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 1942.0, + 1372.0, + 1942.0, + 1372.0, + 1978.0, + 1201.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1978.0, + 1401.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 770.0, + 1972.0, + 770.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2008.0, + 796.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 1018.0, + 2003.0, + 1018.0, + 2040.0, + 295.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 442.0, + 547.0, + 442.0, + 547.0, + 483.0, + 292.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 442.0, + 986.0, + 442.0, + 986.0, + 483.0, + 594.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 442.0, + 1195.0, + 442.0, + 1195.0, + 483.0, + 1188.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1324.0, + 442.0, + 1409.0, + 442.0, + 1409.0, + 483.0, + 1324.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 475.0, + 492.0, + 475.0, + 492.0, + 514.0, + 291.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 475.0, + 683.0, + 475.0, + 683.0, + 514.0, + 677.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 802.0, + 475.0, + 1406.0, + 475.0, + 1406.0, + 514.0, + 802.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 510.0, + 1405.0, + 510.0, + 1405.0, + 541.0, + 293.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 540.0, + 844.0, + 540.0, + 844.0, + 574.0, + 292.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 540.0, + 1206.0, + 540.0, + 1206.0, + 574.0, + 1031.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 540.0, + 1241.0, + 540.0, + 1241.0, + 574.0, + 1232.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1206.0, + 372.0, + 1206.0, + 372.0, + 1247.0, + 294.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 1206.0, + 1047.0, + 1206.0, + 1047.0, + 1247.0, + 589.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1206.0, + 1405.0, + 1206.0, + 1405.0, + 1247.0, + 1079.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1243.0, + 787.0, + 1243.0, + 787.0, + 1277.0, + 317.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 1243.0, + 1404.0, + 1243.0, + 1404.0, + 1277.0, + 807.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1271.0, + 889.0, + 1271.0, + 889.0, + 1307.0, + 294.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 1271.0, + 1399.0, + 1271.0, + 1399.0, + 1307.0, + 931.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 230.0, + 296.0, + 230.0, + 296.0, + 265.0, + 293.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 230.0, + 441.0, + 230.0, + 441.0, + 265.0, + 324.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 230.0, + 951.0, + 230.0, + 951.0, + 265.0, + 471.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 230.0, + 1061.0, + 230.0, + 1061.0, + 265.0, + 999.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 230.0, + 1405.0, + 230.0, + 1405.0, + 265.0, + 1087.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 259.0, + 359.0, + 259.0, + 359.0, + 298.0, + 293.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 259.0, + 748.0, + 259.0, + 748.0, + 298.0, + 389.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 259.0, + 878.0, + 259.0, + 878.0, + 298.0, + 777.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 259.0, + 1014.0, + 259.0, + 1014.0, + 298.0, + 909.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 259.0, + 1405.0, + 259.0, + 1405.0, + 298.0, + 1078.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 293.0, + 1404.0, + 293.0, + 1404.0, + 326.0, + 295.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 322.0, + 1410.0, + 322.0, + 1410.0, + 357.0, + 293.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1422.0, + 492.0, + 1422.0, + 492.0, + 1469.0, + 292.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 1422.0, + 606.0, + 1422.0, + 606.0, + 1469.0, + 506.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 658.0, + 1422.0, + 669.0, + 1422.0, + 669.0, + 1469.0, + 658.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1422.0, + 894.0, + 1422.0, + 894.0, + 1469.0, + 843.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 1422.0, + 1405.0, + 1422.0, + 1405.0, + 1469.0, + 1065.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1462.0, + 816.0, + 1462.0, + 816.0, + 1497.0, + 323.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 1462.0, + 1205.0, + 1462.0, + 1205.0, + 1497.0, + 848.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 1462.0, + 1303.0, + 1462.0, + 1303.0, + 1497.0, + 1233.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1336.0, + 1462.0, + 1404.0, + 1462.0, + 1404.0, + 1497.0, + 1336.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1492.0, + 615.0, + 1492.0, + 615.0, + 1526.0, + 295.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 1492.0, + 944.0, + 1492.0, + 944.0, + 1526.0, + 647.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1012.0, + 602.0, + 1012.0, + 602.0, + 1052.0, + 294.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1012.0, + 848.0, + 1012.0, + 848.0, + 1052.0, + 796.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 1012.0, + 1404.0, + 1012.0, + 1404.0, + 1052.0, + 1039.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1038.0, + 331.0, + 1038.0, + 331.0, + 1085.0, + 290.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 1038.0, + 578.0, + 1038.0, + 578.0, + 1085.0, + 503.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1038.0, + 756.0, + 1038.0, + 756.0, + 1085.0, + 747.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 1038.0, + 1036.0, + 1038.0, + 1036.0, + 1085.0, + 783.0, + 1085.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 958, + 1404, + 958, + 1404, + 1327, + 297, + 1327 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 668, + 1404, + 668, + 1404, + 946, + 298, + 946 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 286, + 1406, + 286, + 1406, + 654, + 297, + 654 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1598, + 1405, + 1598, + 1405, + 1815, + 297, + 1815 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1830, + 1403, + 1830, + 1403, + 1954, + 299, + 1954 + ], + "score": 0.975 + }, + { + "category_id": 8, + "poly": [ + 679, + 1347, + 1019, + 1347, + 1019, + 1435, + 679, + 1435 + ], + "score": 0.953 + }, + { + "category_id": 2, + "poly": [ + 299, + 1979, + 1379, + 1979, + 1379, + 2034, + 299, + 2034 + ], + "score": 0.941 + }, + { + "category_id": 0, + "poly": [ + 300, + 1472, + 680, + 1472, + 680, + 1509, + 300, + 1509 + ], + "score": 0.914 + }, + { + "category_id": 0, + "poly": [ + 299, + 1541, + 902, + 1541, + 902, + 1574, + 299, + 1574 + ], + "score": 0.912 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1376, + 1400, + 1376, + 1400, + 1406, + 1366, + 1406 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 104, + 300, + 104 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.753 + }, + { + "category_id": 0, + "poly": [ + 297, + 229, + 982, + 229, + 982, + 261, + 297, + 261 + ], + "score": 0.704 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 982, + 229, + 982, + 261, + 297, + 261 + ], + "score": 0.197 + }, + { + "category_id": 14, + "poly": [ + 677, + 1342, + 1021, + 1342, + 1021, + 1438, + 677, + 1438 + ], + "score": 0.95, + "latex": "\\ell _ { 1 } = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\left| \\left| g ( X _ { i } ) - Y _ { t g t } ^ { ( i ) } \\right| \\right| _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1047, + 1112, + 1184, + 1112, + 1184, + 1146, + 1047, + 1146 + ], + "score": 0.92, + "latex": "f ( t ) = \\lfloor \\lambda t \\rfloor" + }, + { + "category_id": 13, + "poly": [ + 556, + 1295, + 601, + 1295, + 601, + 1329, + 556, + 1329 + ], + "score": 0.91, + "latex": "g ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 501, + 1173, + 653, + 1173, + 653, + 1207, + 501, + 1207 + ], + "score": 0.9, + "latex": "f ( t ) = c o n s t" + }, + { + "category_id": 13, + "poly": [ + 428, + 1113, + 498, + 1113, + 498, + 1141, + 428, + 1141 + ], + "score": 0.9, + "latex": "\\lambda < 1" + }, + { + "category_id": 13, + "poly": [ + 458, + 1024, + 569, + 1024, + 569, + 1055, + 458, + 1055 + ], + "score": 0.89, + "latex": "\\hat { y } _ { 1 } , . . . , \\hat { y } _ { p } )" + }, + { + "category_id": 13, + "poly": [ + 1191, + 1023, + 1360, + 1023, + 1360, + 1056, + 1191, + 1056 + ], + "score": 0.88, + "latex": "y _ { p + 1 } , . . . , y _ { { p + q } } )" + }, + { + "category_id": 13, + "poly": [ + 814, + 1296, + 842, + 1296, + 842, + 1326, + 814, + 1326 + ], + "score": 0.87, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 815, + 960, + 1336, + 960, + 1336, + 994, + 815, + 994 + ], + "score": 0.87, + "latex": "Y _ { t g t } = \\{ y _ { 1 } , . . . , y _ { p } , y _ { p + 1 } , . . . , y _ { p + q } , y _ { p + q + 1 } , . . . \\}" + }, + { + "category_id": 13, + "poly": [ + 1014, + 1784, + 1069, + 1784, + 1069, + 1813, + 1014, + 1813 + ], + "score": 0.86, + "latex": "90 \\%" + }, + { + "category_id": 13, + "poly": [ + 1250, + 1784, + 1303, + 1784, + 1303, + 1813, + 1250, + 1813 + ], + "score": 0.86, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 296, + 991, + 840, + 991, + 840, + 1024, + 296, + 1024 + ], + "score": 0.84, + "latex": "Y _ { i n } = \\{ y _ { 0 } , \\hat { y _ { 1 } } , . . . , \\hat { y _ { p } } , y _ { p + 1 } , . . . , y _ { p + q } , \\hat { y } _ { p + q + 1 } , . . . \\}" + }, + { + "category_id": 13, + "poly": [ + 931, + 1297, + 955, + 1297, + 955, + 1322, + 931, + 1322 + ], + "score": 0.82, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 640, + 1080, + 1035, + 1080, + 1035, + 1115, + 640, + 1115 + ], + "score": 0.82, + "latex": "f ( t ) \\in \\{ c o n s t , \\lfloor \\lambda t \\rfloor , \\lfloor \\lambda t ^ { 2 } \\rfloor , \\lfloor \\lambda e ^ { t } \\rfloor \\}" + }, + { + "category_id": 13, + "poly": [ + 770, + 1119, + 789, + 1119, + 789, + 1144, + 770, + 1144 + ], + "score": 0.82, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1057, + 1117, + 1057, + 1117, + 1083, + 1100, + 1083 + ], + "score": 0.8, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 1111, + 1085, + 1125, + 1085, + 1125, + 1109, + 1111, + 1109 + ], + "score": 0.78, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 373, + 1027, + 391, + 1027, + 391, + 1053, + 373, + 1053 + ], + "score": 0.78, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 1107, + 1026, + 1125, + 1026, + 1125, + 1054, + 1107, + 1054 + ], + "score": 0.78, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 1384, + 1059, + 1401, + 1059, + 1401, + 1084, + 1384, + 1084 + ], + "score": 0.77, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 555, + 1239, + 572, + 1239, + 572, + 1266, + 555, + 1266 + ], + "score": 0.71, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 1372, + 1026, + 1401, + 1026, + 1401, + 1055, + 1372, + 1055 + ], + "score": 0.31, + "latex": "y _ { 0 }" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1974.0, + 1382.0, + 1974.0, + 1382.0, + 2010.0, + 329.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2003.0, + 907.0, + 2003.0, + 907.0, + 2037.0, + 296.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1468.0, + 682.0, + 1468.0, + 682.0, + 1514.0, + 292.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1542.0, + 904.0, + 1542.0, + 904.0, + 1577.0, + 295.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 986.0, + 229.0, + 986.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 954.0, + 814.0, + 954.0, + 814.0, + 1002.0, + 290.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1337.0, + 954.0, + 1408.0, + 954.0, + 1408.0, + 1002.0, + 1337.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 985.0, + 295.0, + 985.0, + 295.0, + 1033.0, + 290.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 985.0, + 1409.0, + 985.0, + 1409.0, + 1033.0, + 841.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1018.0, + 372.0, + 1018.0, + 372.0, + 1061.0, + 292.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1018.0, + 457.0, + 1018.0, + 457.0, + 1061.0, + 392.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 1018.0, + 1106.0, + 1018.0, + 1106.0, + 1061.0, + 570.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 1018.0, + 1190.0, + 1018.0, + 1190.0, + 1061.0, + 1126.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1361.0, + 1018.0, + 1371.0, + 1018.0, + 1371.0, + 1061.0, + 1361.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1018.0, + 1405.0, + 1018.0, + 1405.0, + 1061.0, + 1402.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1048.0, + 1099.0, + 1048.0, + 1099.0, + 1090.0, + 291.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 1048.0, + 1383.0, + 1048.0, + 1383.0, + 1090.0, + 1118.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1048.0, + 1408.0, + 1048.0, + 1408.0, + 1090.0, + 1402.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1080.0, + 639.0, + 1080.0, + 639.0, + 1118.0, + 294.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 1080.0, + 1110.0, + 1080.0, + 1110.0, + 1118.0, + 1036.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 1080.0, + 1407.0, + 1080.0, + 1407.0, + 1118.0, + 1126.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1113.0, + 427.0, + 1113.0, + 427.0, + 1147.0, + 295.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 1113.0, + 769.0, + 1113.0, + 769.0, + 1147.0, + 499.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 1113.0, + 1046.0, + 1113.0, + 1046.0, + 1147.0, + 790.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1113.0, + 1404.0, + 1113.0, + 1404.0, + 1147.0, + 1185.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1140.0, + 1406.0, + 1140.0, + 1406.0, + 1178.0, + 292.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1171.0, + 500.0, + 1171.0, + 500.0, + 1211.0, + 291.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1171.0, + 1408.0, + 1171.0, + 1408.0, + 1211.0, + 654.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1202.0, + 1407.0, + 1202.0, + 1407.0, + 1240.0, + 292.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1235.0, + 554.0, + 1235.0, + 554.0, + 1270.0, + 295.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1235.0, + 1404.0, + 1235.0, + 1404.0, + 1270.0, + 573.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1263.0, + 1404.0, + 1263.0, + 1404.0, + 1299.0, + 294.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1295.0, + 555.0, + 1295.0, + 555.0, + 1329.0, + 295.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1295.0, + 813.0, + 1295.0, + 813.0, + 1329.0, + 602.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1295.0, + 930.0, + 1295.0, + 930.0, + 1329.0, + 843.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1295.0, + 966.0, + 1295.0, + 966.0, + 1329.0, + 956.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 669.0, + 1406.0, + 669.0, + 1406.0, + 706.0, + 296.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 699.0, + 1406.0, + 699.0, + 1406.0, + 737.0, + 292.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 731.0, + 1407.0, + 731.0, + 1407.0, + 766.0, + 292.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 757.0, + 1408.0, + 757.0, + 1408.0, + 799.0, + 292.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 791.0, + 1405.0, + 791.0, + 1405.0, + 828.0, + 293.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 820.0, + 1405.0, + 820.0, + 1405.0, + 858.0, + 293.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 853.0, + 1404.0, + 853.0, + 1404.0, + 890.0, + 294.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 882.0, + 1404.0, + 882.0, + 1404.0, + 919.0, + 296.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 911.0, + 1354.0, + 911.0, + 1354.0, + 953.0, + 293.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 286.0, + 1404.0, + 286.0, + 1404.0, + 323.0, + 295.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 319.0, + 1406.0, + 319.0, + 1406.0, + 353.0, + 294.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 348.0, + 1406.0, + 348.0, + 1406.0, + 383.0, + 292.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 382.0, + 1406.0, + 382.0, + 1406.0, + 413.0, + 296.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 410.0, + 1406.0, + 410.0, + 1406.0, + 444.0, + 294.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 440.0, + 1411.0, + 440.0, + 1411.0, + 476.0, + 291.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 471.0, + 1408.0, + 471.0, + 1408.0, + 505.0, + 294.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 497.0, + 1407.0, + 497.0, + 1407.0, + 538.0, + 291.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 528.0, + 1406.0, + 528.0, + 1406.0, + 568.0, + 292.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 560.0, + 1404.0, + 560.0, + 1404.0, + 594.0, + 295.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 592.0, + 1403.0, + 592.0, + 1403.0, + 626.0, + 294.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 622.0, + 1398.0, + 622.0, + 1398.0, + 656.0, + 295.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1600.0, + 1405.0, + 1600.0, + 1405.0, + 1635.0, + 295.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1630.0, + 1408.0, + 1630.0, + 1408.0, + 1669.0, + 292.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1662.0, + 1406.0, + 1662.0, + 1406.0, + 1697.0, + 294.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1693.0, + 1406.0, + 1693.0, + 1406.0, + 1728.0, + 295.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1720.0, + 1405.0, + 1720.0, + 1405.0, + 1756.0, + 292.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1752.0, + 1405.0, + 1752.0, + 1405.0, + 1787.0, + 295.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1785.0, + 1013.0, + 1785.0, + 1013.0, + 1819.0, + 295.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 1785.0, + 1249.0, + 1785.0, + 1249.0, + 1819.0, + 1070.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1304.0, + 1785.0, + 1397.0, + 1785.0, + 1397.0, + 1819.0, + 1304.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1831.0, + 1405.0, + 1831.0, + 1405.0, + 1864.0, + 295.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1859.0, + 1407.0, + 1859.0, + 1407.0, + 1897.0, + 293.0, + 1897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1890.0, + 1404.0, + 1890.0, + 1404.0, + 1928.0, + 293.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1924.0, + 1311.0, + 1924.0, + 1311.0, + 1956.0, + 295.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 986.0, + 229.0, + 986.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1758, + 1406, + 1758, + 1406, + 2034, + 297, + 2034 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 881, + 1406, + 881, + 1406, + 1127, + 298, + 1127 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 505, + 1404, + 505, + 1404, + 781, + 298, + 781 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1228, + 1405, + 1228, + 1405, + 1443, + 298, + 1443 + ], + "score": 0.98 + }, + { + "category_id": 5, + "poly": [ + 434, + 287, + 1264, + 287, + 1264, + 446, + 434, + 446 + ], + "score": 0.978, + "html": "
CategoryNum of ClipsClip Length (min)FPSResolution
Ballet136115720P
Hiphop298115720P
Japanese Pop356115720P
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1565, + 1403, + 1565, + 1403, + 1657, + 299, + 1657 + ], + "score": 0.972 + }, + { + "category_id": 6, + "poly": [ + 667, + 249, + 1031, + 249, + 1031, + 281, + 667, + 281 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 298, + 1699, + 952, + 1699, + 952, + 1732, + 298, + 1732 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 299, + 1170, + 501, + 1170, + 501, + 1201, + 299, + 1201 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 301, + 823, + 691, + 823, + 691, + 854, + 301, + 854 + ], + "score": 0.889 + }, + { + "category_id": 0, + "poly": [ + 302, + 1492, + 713, + 1492, + 713, + 1527, + 302, + 1527 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.779 + }, + { + "category_id": 13, + "poly": [ + 791, + 1035, + 916, + 1035, + 916, + 1066, + 791, + 1066 + ], + "score": 0.92, + "latex": "d _ { s } = 1 0 2 4" + }, + { + "category_id": 13, + "poly": [ + 1301, + 1035, + 1398, + 1035, + 1398, + 1069, + 1301, + 1069 + ], + "score": 0.92, + "latex": "d _ { y } = 5 0" + }, + { + "category_id": 13, + "poly": [ + 531, + 1067, + 613, + 1067, + 613, + 1097, + 531, + 1097 + ], + "score": 0.91, + "latex": "q = 1 0" + }, + { + "category_id": 13, + "poly": [ + 1144, + 1005, + 1252, + 1005, + 1252, + 1035, + 1144, + 1035 + ], + "score": 0.91, + "latex": "d _ { z } = 2 5 6" + }, + { + "category_id": 13, + "poly": [ + 1241, + 944, + 1403, + 944, + 1403, + 974, + 1241, + 974 + ], + "score": 0.91, + "latex": "d _ { k } = d _ { v } = 6 4" + }, + { + "category_id": 13, + "poly": [ + 711, + 1006, + 820, + 1006, + 820, + 1035, + 711, + 1035 + ], + "score": 0.91, + "latex": "d _ { x } = 4 3 8" + }, + { + "category_id": 13, + "poly": [ + 738, + 884, + 817, + 884, + 817, + 910, + 738, + 910 + ], + "score": 0.9, + "latex": "N = 2" + }, + { + "category_id": 13, + "poly": [ + 376, + 1066, + 481, + 1066, + 481, + 1095, + 376, + 1095 + ], + "score": 0.9, + "latex": "\\lambda = 0 . 0 1" + }, + { + "category_id": 13, + "poly": [ + 609, + 914, + 673, + 914, + 673, + 942, + 609, + 942 + ], + "score": 0.89, + "latex": "l = 8" + }, + { + "category_id": 13, + "poly": [ + 810, + 975, + 908, + 975, + 908, + 1003, + 810, + 1003 + ], + "score": 0.89, + "latex": "k = 1 0 0" + }, + { + "category_id": 13, + "poly": [ + 1297, + 976, + 1397, + 976, + 1397, + 1004, + 1297, + 1004 + ], + "score": 0.87, + "latex": "n = 9 0 0" + }, + { + "category_id": 13, + "poly": [ + 819, + 720, + 933, + 720, + 933, + 749, + 819, + 749 + ], + "score": 0.44, + "latex": "1 5 { , } 4 0 0 \\mathrm { H z }" + }, + { + "category_id": 13, + "poly": [ + 783, + 599, + 807, + 599, + 807, + 629, + 783, + 629 + ], + "score": 0.38, + "latex": "Q" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 245.0, + 1032.0, + 245.0, + 1032.0, + 284.0, + 666.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1699.0, + 955.0, + 1699.0, + 955.0, + 1735.0, + 294.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1168.0, + 504.0, + 1168.0, + 504.0, + 1204.0, + 294.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 820.0, + 696.0, + 820.0, + 696.0, + 858.0, + 294.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1489.0, + 717.0, + 1489.0, + 717.0, + 1532.0, + 293.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1760.0, + 1404.0, + 1760.0, + 1404.0, + 1793.0, + 296.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1789.0, + 1404.0, + 1789.0, + 1404.0, + 1824.0, + 294.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1820.0, + 1406.0, + 1820.0, + 1406.0, + 1857.0, + 294.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1852.0, + 1407.0, + 1852.0, + 1407.0, + 1884.0, + 296.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1883.0, + 1407.0, + 1883.0, + 1407.0, + 1916.0, + 295.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1912.0, + 1408.0, + 1912.0, + 1408.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1970.0, + 1408.0, + 1970.0, + 1408.0, + 2008.0, + 292.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1365.0, + 2002.0, + 1365.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 881.0, + 737.0, + 881.0, + 737.0, + 915.0, + 295.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 881.0, + 1406.0, + 881.0, + 1406.0, + 915.0, + 818.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 915.0, + 608.0, + 915.0, + 608.0, + 945.0, + 296.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 915.0, + 1404.0, + 915.0, + 1404.0, + 945.0, + 674.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 944.0, + 1240.0, + 944.0, + 1240.0, + 978.0, + 295.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 944.0, + 1407.0, + 944.0, + 1407.0, + 978.0, + 1404.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 975.0, + 809.0, + 975.0, + 809.0, + 1009.0, + 295.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 975.0, + 1296.0, + 975.0, + 1296.0, + 1009.0, + 909.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 975.0, + 1408.0, + 975.0, + 1408.0, + 1009.0, + 1398.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1000.0, + 710.0, + 1000.0, + 710.0, + 1041.0, + 293.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 1000.0, + 1143.0, + 1000.0, + 1143.0, + 1041.0, + 821.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1000.0, + 1408.0, + 1000.0, + 1408.0, + 1041.0, + 1253.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1033.0, + 790.0, + 1033.0, + 790.0, + 1069.0, + 293.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1033.0, + 1300.0, + 1033.0, + 1300.0, + 1069.0, + 917.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1033.0, + 1408.0, + 1033.0, + 1408.0, + 1069.0, + 1399.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1064.0, + 375.0, + 1064.0, + 375.0, + 1100.0, + 293.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1064.0, + 530.0, + 1064.0, + 530.0, + 1100.0, + 482.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 1064.0, + 1404.0, + 1064.0, + 1404.0, + 1100.0, + 614.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1097.0, + 1018.0, + 1097.0, + 1018.0, + 1128.0, + 294.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 501.0, + 1405.0, + 501.0, + 1405.0, + 545.0, + 292.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 537.0, + 1404.0, + 537.0, + 1404.0, + 570.0, + 294.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 566.0, + 1406.0, + 566.0, + 1406.0, + 603.0, + 294.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 597.0, + 782.0, + 597.0, + 782.0, + 634.0, + 294.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 597.0, + 1406.0, + 597.0, + 1406.0, + 634.0, + 808.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 627.0, + 1406.0, + 627.0, + 1406.0, + 664.0, + 294.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 660.0, + 1405.0, + 660.0, + 1405.0, + 693.0, + 296.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 686.0, + 1406.0, + 686.0, + 1406.0, + 725.0, + 293.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 719.0, + 818.0, + 719.0, + 818.0, + 754.0, + 293.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 719.0, + 1405.0, + 719.0, + 1405.0, + 754.0, + 934.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 749.0, + 1081.0, + 749.0, + 1081.0, + 785.0, + 294.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1230.0, + 1405.0, + 1230.0, + 1405.0, + 1264.0, + 295.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1257.0, + 1408.0, + 1257.0, + 1408.0, + 1296.0, + 292.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1290.0, + 1405.0, + 1290.0, + 1405.0, + 1324.0, + 295.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1320.0, + 1406.0, + 1320.0, + 1406.0, + 1356.0, + 292.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1350.0, + 1405.0, + 1350.0, + 1405.0, + 1385.0, + 295.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1382.0, + 1403.0, + 1382.0, + 1403.0, + 1417.0, + 295.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1411.0, + 1409.0, + 1411.0, + 1409.0, + 1448.0, + 293.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1565.0, + 1404.0, + 1565.0, + 1404.0, + 1599.0, + 294.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1595.0, + 1405.0, + 1595.0, + 1405.0, + 1629.0, + 295.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1627.0, + 940.0, + 1627.0, + 940.0, + 1660.0, + 295.0, + 1660.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1499, + 1405, + 1499, + 1405, + 1866, + 297, + 1866 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1207, + 1405, + 1207, + 1405, + 1485, + 298, + 1485 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1881, + 1403, + 1881, + 1403, + 2034, + 299, + 2034 + ], + "score": 0.977 + }, + { + "category_id": 4, + "poly": [ + 298, + 1076, + 1405, + 1076, + 1405, + 1171, + 298, + 1171 + ], + "score": 0.961 + }, + { + "category_id": 6, + "poly": [ + 296, + 248, + 1404, + 248, + 1404, + 465, + 296, + 465 + ], + "score": 0.933 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.746 + }, + { + "category_id": 3, + "poly": [ + 294, + 488, + 1403, + 488, + 1403, + 1052, + 294, + 1052 + ], + "score": 0.495 + }, + { + "category_id": 1, + "poly": [ + 296, + 248, + 1404, + 248, + 1404, + 465, + 296, + 465 + ], + "score": 0.104 + }, + { + "category_id": 13, + "poly": [ + 933, + 1912, + 1010, + 1912, + 1010, + 1941, + 933, + 1941 + ], + "score": 0.88, + "latex": "5 7 . 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 661, + 1882, + 735, + 1882, + 735, + 1911, + 661, + 1911 + ], + "score": 0.87, + "latex": "4 1 . 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 1223, + 1682, + 1278, + 1682, + 1278, + 1711, + 1223, + 1711 + ], + "score": 0.86, + "latex": "3 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 802, + 1712, + 876, + 1712, + 876, + 1743, + 802, + 1743 + ], + "score": 0.86, + "latex": "2 1 . 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 1329, + 1881, + 1404, + 1881, + 1404, + 1911, + 1329, + 1911 + ], + "score": 0.86, + "latex": "3 0 . 3 \\%" + }, + { + "category_id": 13, + "poly": [ + 594, + 1452, + 670, + 1452, + 670, + 1482, + 594, + 1482 + ], + "score": 0.85, + "latex": "7 7 . 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 1325, + 1452, + 1398, + 1452, + 1398, + 1483, + 1325, + 1483 + ], + "score": 0.84, + "latex": "1 7 . 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 503, + 312, + 564, + 312, + 564, + 339, + 503, + 339 + ], + "score": 0.27, + "latex": "A C C" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1077.0, + 1405.0, + 1077.0, + 1405.0, + 1111.0, + 296.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1106.0, + 1403.0, + 1106.0, + 1403.0, + 1141.0, + 295.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1138.0, + 1313.0, + 1138.0, + 1313.0, + 1175.0, + 292.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 249.0, + 1405.0, + 249.0, + 1405.0, + 283.0, + 294.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 282.0, + 1405.0, + 282.0, + 1405.0, + 313.0, + 295.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 309.0, + 502.0, + 309.0, + 502.0, + 345.0, + 293.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 309.0, + 1407.0, + 309.0, + 1407.0, + 345.0, + 565.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 340.0, + 1406.0, + 340.0, + 1406.0, + 375.0, + 295.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 371.0, + 1406.0, + 371.0, + 1406.0, + 405.0, + 293.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 401.0, + 1407.0, + 401.0, + 1407.0, + 436.0, + 295.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 432.0, + 1410.0, + 432.0, + 1410.0, + 465.0, + 294.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 495.0, + 398.0, + 495.0, + 398.0, + 529.0, + 307.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 494.0, + 560.0, + 494.0, + 560.0, + 530.0, + 502.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 494.0, + 684.0, + 494.0, + 684.0, + 531.0, + 574.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 494.0, + 900.0, + 494.0, + 900.0, + 531.0, + 696.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 494.0, + 1104.0, + 494.0, + 1104.0, + 531.0, + 910.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 494.0, + 1224.0, + 494.0, + 1224.0, + 531.0, + 1114.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 491.0, + 1393.0, + 491.0, + 1393.0, + 533.0, + 1233.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 534.0, + 445.0, + 534.0, + 445.0, + 568.0, + 307.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 531.0, + 556.0, + 531.0, + 556.0, + 570.0, + 503.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 533.0, + 660.0, + 533.0, + 660.0, + 570.0, + 598.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 533.0, + 827.0, + 533.0, + 827.0, + 570.0, + 767.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 533.0, + 1037.0, + 533.0, + 1037.0, + 570.0, + 977.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 533.0, + 1200.0, + 533.0, + 1200.0, + 570.0, + 1138.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1304.0, + 542.0, + 1325.0, + 542.0, + 1325.0, + 565.0, + 1304.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 572.0, + 388.0, + 572.0, + 388.0, + 607.0, + 306.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 572.0, + 561.0, + 572.0, + 561.0, + 608.0, + 501.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 572.0, + 657.0, + 572.0, + 657.0, + 608.0, + 598.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 570.0, + 822.0, + 570.0, + 822.0, + 608.0, + 770.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 570.0, + 1034.0, + 570.0, + 1034.0, + 608.0, + 982.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 570.0, + 1200.0, + 570.0, + 1200.0, + 608.0, + 1141.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1304.0, + 581.0, + 1325.0, + 581.0, + 1325.0, + 603.0, + 1304.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 599.0, + 486.0, + 599.0, + 486.0, + 633.0, + 308.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 599.0, + 561.0, + 599.0, + 561.0, + 635.0, + 500.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 599.0, + 658.0, + 599.0, + 658.0, + 635.0, + 598.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 599.0, + 826.0, + 599.0, + 826.0, + 635.0, + 769.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 599.0, + 1037.0, + 599.0, + 1037.0, + 635.0, + 977.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 599.0, + 1199.0, + 599.0, + 1199.0, + 635.0, + 1141.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 600.0, + 1343.0, + 600.0, + 1343.0, + 633.0, + 1287.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 625.0, + 480.0, + 625.0, + 480.0, + 665.0, + 305.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 625.0, + 561.0, + 625.0, + 561.0, + 662.0, + 501.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 626.0, + 658.0, + 626.0, + 658.0, + 662.0, + 598.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 626.0, + 827.0, + 626.0, + 827.0, + 662.0, + 768.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 627.0, + 1036.0, + 627.0, + 1036.0, + 660.0, + 975.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 623.0, + 1201.0, + 623.0, + 1201.0, + 664.0, + 1140.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 626.0, + 1345.0, + 626.0, + 1345.0, + 662.0, + 1284.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 651.0, + 372.0, + 651.0, + 372.0, + 691.0, + 307.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 651.0, + 556.0, + 651.0, + 556.0, + 690.0, + 506.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 652.0, + 658.0, + 652.0, + 658.0, + 688.0, + 598.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 652.0, + 827.0, + 652.0, + 827.0, + 688.0, + 767.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 652.0, + 1037.0, + 652.0, + 1037.0, + 688.0, + 977.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 652.0, + 1199.0, + 652.0, + 1199.0, + 688.0, + 1139.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 652.0, + 1345.0, + 652.0, + 1345.0, + 690.0, + 1286.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 734.0, + 616.0, + 734.0, + 616.0, + 765.0, + 364.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 735.0, + 988.0, + 735.0, + 988.0, + 767.0, + 725.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 734.0, + 1331.0, + 734.0, + 1331.0, + 766.0, + 1116.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 769.0, + 341.0, + 769.0, + 341.0, + 791.0, + 298.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 775.0, + 415.0, + 775.0, + 415.0, + 801.0, + 354.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 804.0, + 495.0, + 804.0, + 495.0, + 835.0, + 422.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 832.0, + 568.0, + 832.0, + 568.0, + 862.0, + 496.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 823.0, + 626.0, + 823.0, + 626.0, + 850.0, + 582.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 775.0, + 783.0, + 775.0, + 783.0, + 806.0, + 720.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 807.0, + 863.0, + 807.0, + 863.0, + 838.0, + 786.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 823.0, + 993.0, + 823.0, + 993.0, + 850.0, + 949.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 775.0, + 1157.0, + 775.0, + 1157.0, + 812.0, + 1084.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 789.0, + 1229.0, + 789.0, + 1229.0, + 821.0, + 1160.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 810.0, + 1302.0, + 810.0, + 1302.0, + 843.0, + 1232.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 822.0, + 1361.0, + 822.0, + 1361.0, + 850.0, + 1317.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 844.0, + 935.0, + 844.0, + 935.0, + 874.0, + 862.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 879.0, + 341.0, + 879.0, + 341.0, + 902.0, + 304.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 878.0, + 422.0, + 878.0, + 422.0, + 915.0, + 346.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 882.0, + 790.0, + 882.0, + 790.0, + 917.0, + 716.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 889.0, + 1155.0, + 889.0, + 1155.0, + 920.0, + 1081.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 897.0, + 1229.0, + 897.0, + 1229.0, + 931.0, + 1157.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 915.0, + 495.0, + 915.0, + 495.0, + 945.0, + 422.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 940.0, + 568.0, + 940.0, + 568.0, + 972.0, + 500.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 935.0, + 621.0, + 935.0, + 621.0, + 958.0, + 585.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 918.0, + 935.0, + 918.0, + 935.0, + 984.0, + 789.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 935.0, + 991.0, + 935.0, + 991.0, + 958.0, + 951.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 922.0, + 1303.0, + 922.0, + 1303.0, + 953.0, + 1231.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 932.0, + 1361.0, + 932.0, + 1361.0, + 961.0, + 1318.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 989.0, + 342.0, + 989.0, + 342.0, + 1014.0, + 311.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 1019.0, + 1063.0, + 1019.0, + 1063.0, + 1048.0, + 632.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1499.0, + 1406.0, + 1499.0, + 1406.0, + 1537.0, + 294.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1530.0, + 1405.0, + 1530.0, + 1405.0, + 1565.0, + 294.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1561.0, + 1405.0, + 1561.0, + 1405.0, + 1595.0, + 295.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1589.0, + 1405.0, + 1589.0, + 1405.0, + 1627.0, + 292.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1620.0, + 1405.0, + 1620.0, + 1405.0, + 1655.0, + 295.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1655.0, + 1406.0, + 1655.0, + 1406.0, + 1685.0, + 296.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1679.0, + 1222.0, + 1679.0, + 1222.0, + 1719.0, + 292.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 1679.0, + 1406.0, + 1679.0, + 1406.0, + 1719.0, + 1279.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1712.0, + 801.0, + 1712.0, + 801.0, + 1748.0, + 292.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1712.0, + 1408.0, + 1712.0, + 1408.0, + 1748.0, + 877.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1742.0, + 1406.0, + 1742.0, + 1406.0, + 1780.0, + 294.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1775.0, + 1405.0, + 1775.0, + 1405.0, + 1807.0, + 294.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1802.0, + 1405.0, + 1802.0, + 1405.0, + 1839.0, + 294.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1836.0, + 1367.0, + 1836.0, + 1367.0, + 1870.0, + 297.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1207.0, + 1405.0, + 1207.0, + 1405.0, + 1242.0, + 295.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1240.0, + 1405.0, + 1240.0, + 1405.0, + 1272.0, + 296.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1267.0, + 1406.0, + 1267.0, + 1406.0, + 1305.0, + 292.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1300.0, + 1406.0, + 1300.0, + 1406.0, + 1333.0, + 296.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1332.0, + 1405.0, + 1332.0, + 1405.0, + 1364.0, + 296.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1358.0, + 1405.0, + 1358.0, + 1405.0, + 1395.0, + 293.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1391.0, + 1410.0, + 1391.0, + 1410.0, + 1427.0, + 293.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1422.0, + 1407.0, + 1422.0, + 1407.0, + 1455.0, + 296.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1451.0, + 593.0, + 1451.0, + 593.0, + 1488.0, + 295.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 1451.0, + 1324.0, + 1451.0, + 1324.0, + 1488.0, + 671.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1451.0, + 1408.0, + 1451.0, + 1408.0, + 1488.0, + 1399.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 660.0, + 1880.0, + 660.0, + 1915.0, + 293.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1880.0, + 1328.0, + 1880.0, + 1328.0, + 1915.0, + 736.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 932.0, + 1910.0, + 932.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1947.0, + 1011.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1979.0, + 294.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1969.0, + 1405.0, + 1969.0, + 1405.0, + 2010.0, + 294.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1408.0, + 2002.0, + 1408.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 249.0, + 1405.0, + 249.0, + 1405.0, + 283.0, + 294.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 282.0, + 1405.0, + 282.0, + 1405.0, + 313.0, + 295.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 309.0, + 502.0, + 309.0, + 502.0, + 345.0, + 293.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 309.0, + 1407.0, + 309.0, + 1407.0, + 345.0, + 565.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 340.0, + 1406.0, + 340.0, + 1406.0, + 375.0, + 295.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 371.0, + 1406.0, + 371.0, + 1406.0, + 405.0, + 293.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 401.0, + 1407.0, + 401.0, + 1407.0, + 436.0, + 295.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 432.0, + 1410.0, + 432.0, + 1410.0, + 465.0, + 294.0, + 465.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1189, + 1404, + 1189, + 1404, + 1557, + 299, + 1557 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 308, + 1404, + 308, + 1404, + 734, + 297, + 734 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1819, + 1403, + 1819, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1572, + 1404, + 1572, + 1404, + 1725, + 299, + 1725 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 345, + 764, + 1342, + 764, + 1342, + 1013, + 345, + 1013 + ], + "score": 0.97 + }, + { + "category_id": 4, + "poly": [ + 299, + 1044, + 1405, + 1044, + 1405, + 1136, + 299, + 1136 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 295, + 230, + 1400, + 230, + 1400, + 293, + 295, + 293 + ], + "score": 0.954 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 302, + 1763, + 775, + 1763, + 775, + 1794, + 302, + 1794 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.789 + }, + { + "category_id": 13, + "poly": [ + 861, + 338, + 954, + 338, + 954, + 372, + 861, + 372 + ], + "score": 0.93, + "latex": "B _ { k } / B _ { m }" + }, + { + "category_id": 13, + "poly": [ + 1186, + 338, + 1271, + 338, + 1271, + 371, + 1186, + 371 + ], + "score": 0.92, + "latex": "B _ { a } / B _ { k }" + }, + { + "category_id": 13, + "poly": [ + 694, + 370, + 740, + 370, + 740, + 400, + 694, + 400 + ], + "score": 0.9, + "latex": "B _ { m }" + }, + { + "category_id": 13, + "poly": [ + 1162, + 371, + 1199, + 371, + 1199, + 400, + 1162, + 400 + ], + "score": 0.89, + "latex": "B _ { a }" + }, + { + "category_id": 13, + "poly": [ + 1363, + 339, + 1400, + 339, + 1400, + 369, + 1363, + 369 + ], + "score": 0.88, + "latex": "B _ { k }" + }, + { + "category_id": 13, + "poly": [ + 1139, + 643, + 1203, + 643, + 1203, + 673, + 1139, + 673 + ], + "score": 0.87, + "latex": "6 . 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 344, + 673, + 406, + 673, + 406, + 703, + 344, + 703 + ], + "score": 0.86, + "latex": "2 . 7 \\%" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 777.0, + 365.0, + 777.0, + 365.0, + 786.0, + 352.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 772.0, + 812.0, + 772.0, + 812.0, + 790.0, + 750.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 777.0, + 861.0, + 777.0, + 861.0, + 786.0, + 850.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 772.0, + 1310.0, + 772.0, + 1310.0, + 790.0, + 1247.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 787.0, + 838.0, + 787.0, + 838.0, + 808.0, + 749.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 785.0, + 1336.0, + 785.0, + 1336.0, + 809.0, + 1245.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 803.0, + 824.0, + 803.0, + 824.0, + 824.0, + 749.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 803.0, + 1320.0, + 803.0, + 1320.0, + 824.0, + 1246.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 820.0, + 365.0, + 820.0, + 365.0, + 829.0, + 352.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 818.0, + 862.0, + 818.0, + 862.0, + 830.0, + 846.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 943.0, + 773.0, + 943.0, + 773.0, + 1005.0, + 373.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1044.0, + 1403.0, + 1044.0, + 1403.0, + 1078.0, + 297.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1073.0, + 1405.0, + 1073.0, + 1405.0, + 1110.0, + 294.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1104.0, + 968.0, + 1104.0, + 968.0, + 1139.0, + 294.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1760.0, + 780.0, + 1760.0, + 780.0, + 1797.0, + 294.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1189.0, + 1404.0, + 1189.0, + 1404.0, + 1227.0, + 294.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1222.0, + 1405.0, + 1222.0, + 1405.0, + 1256.0, + 294.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1252.0, + 1404.0, + 1252.0, + 1404.0, + 1286.0, + 295.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1279.0, + 1406.0, + 1279.0, + 1406.0, + 1318.0, + 293.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1312.0, + 1403.0, + 1312.0, + 1403.0, + 1346.0, + 294.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1344.0, + 1404.0, + 1344.0, + 1404.0, + 1378.0, + 294.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1372.0, + 1404.0, + 1372.0, + 1404.0, + 1408.0, + 294.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1403.0, + 1404.0, + 1403.0, + 1404.0, + 1438.0, + 294.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1434.0, + 1404.0, + 1434.0, + 1404.0, + 1468.0, + 295.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1462.0, + 1405.0, + 1462.0, + 1405.0, + 1501.0, + 293.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1494.0, + 1405.0, + 1494.0, + 1405.0, + 1530.0, + 294.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1527.0, + 1366.0, + 1527.0, + 1366.0, + 1561.0, + 297.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 306.0, + 1405.0, + 306.0, + 1405.0, + 341.0, + 295.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 336.0, + 860.0, + 336.0, + 860.0, + 374.0, + 292.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 336.0, + 1185.0, + 336.0, + 1185.0, + 374.0, + 955.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 336.0, + 1362.0, + 336.0, + 1362.0, + 374.0, + 1272.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 336.0, + 1406.0, + 336.0, + 1406.0, + 374.0, + 1401.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 368.0, + 693.0, + 368.0, + 693.0, + 403.0, + 292.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 368.0, + 1161.0, + 368.0, + 1161.0, + 403.0, + 741.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 368.0, + 1406.0, + 368.0, + 1406.0, + 403.0, + 1200.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 398.0, + 1406.0, + 398.0, + 1406.0, + 434.0, + 294.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 429.0, + 1406.0, + 429.0, + 1406.0, + 464.0, + 294.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 460.0, + 1406.0, + 460.0, + 1406.0, + 494.0, + 294.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 491.0, + 1406.0, + 491.0, + 1406.0, + 525.0, + 292.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 522.0, + 1405.0, + 522.0, + 1405.0, + 554.0, + 295.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 553.0, + 1405.0, + 553.0, + 1405.0, + 584.0, + 295.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 583.0, + 1406.0, + 583.0, + 1406.0, + 618.0, + 294.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 608.0, + 1406.0, + 608.0, + 1406.0, + 649.0, + 292.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 640.0, + 1138.0, + 640.0, + 1138.0, + 679.0, + 292.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 640.0, + 1404.0, + 640.0, + 1404.0, + 679.0, + 1204.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 673.0, + 343.0, + 673.0, + 343.0, + 708.0, + 295.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 673.0, + 1406.0, + 673.0, + 1406.0, + 708.0, + 407.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 703.0, + 1126.0, + 703.0, + 1126.0, + 737.0, + 295.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1818.0, + 1405.0, + 1818.0, + 1405.0, + 1856.0, + 293.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1851.0, + 1407.0, + 1851.0, + 1407.0, + 1886.0, + 294.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1880.0, + 1406.0, + 1880.0, + 1406.0, + 1917.0, + 292.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1946.0, + 293.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1977.0, + 296.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1974.0, + 1407.0, + 1974.0, + 1407.0, + 2005.0, + 296.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2040.0, + 294.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1571.0, + 1408.0, + 1571.0, + 1408.0, + 1608.0, + 295.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1597.0, + 1405.0, + 1597.0, + 1405.0, + 1642.0, + 293.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1633.0, + 1408.0, + 1633.0, + 1408.0, + 1670.0, + 294.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1660.0, + 1406.0, + 1660.0, + 1406.0, + 1700.0, + 293.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1693.0, + 1175.0, + 1693.0, + 1175.0, + 1730.0, + 294.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 262.0, + 1212.0, + 262.0, + 1212.0, + 297.0, + 296.0, + 297.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1739, + 1405, + 1739, + 1405, + 2016, + 297, + 2016 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1380, + 1404, + 1380, + 1404, + 1627, + 298, + 1627 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1150, + 1404, + 1150, + 1404, + 1366, + 298, + 1366 + ], + "score": 0.981 + }, + { + "category_id": 3, + "poly": [ + 575, + 230, + 1112, + 230, + 1112, + 577, + 575, + 577 + ], + "score": 0.966 + }, + { + "category_id": 5, + "poly": [ + 347, + 748, + 792, + 748, + 792, + 927, + 347, + 927 + ], + "score": 0.966, + "html": "
Encoder Structure LSTMFIDACC (%)
7.945
ConvS2S encoder13.354
Global self-attention3.680
Local self-attention4.378.5
" + }, + { + "category_id": 1, + "poly": [ + 300, + 995, + 1402, + 995, + 1402, + 1057, + 300, + 1057 + ], + "score": 0.956 + }, + { + "category_id": 5, + "poly": [ + 855, + 736, + 1320, + 736, + 1320, + 938, + 855, + 938 + ], + "score": 0.926, + "html": "
Learning ApproachFIDACC (%)
Teacher-forcing61.25.3
Auto-condition (con st)15.735
Ours ([λet J)9.869
Ours ([λt²])6.473
Ours ([λt])5.177.6
" + }, + { + "category_id": 4, + "poly": [ + 554, + 591, + 1143, + 591, + 1143, + 623, + 554, + 623 + ], + "score": 0.922 + }, + { + "category_id": 0, + "poly": [ + 299, + 1670, + 544, + 1670, + 544, + 1706, + 299, + 1706 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 300, + 1094, + 581, + 1094, + 581, + 1125, + 300, + 1125 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.796 + }, + { + "category_id": 6, + "poly": [ + 293, + 647, + 1403, + 647, + 1403, + 712, + 293, + 712 + ], + "score": 0.561 + }, + { + "category_id": 1, + "poly": [ + 293, + 647, + 1403, + 647, + 1403, + 712, + 293, + 712 + ], + "score": 0.264 + }, + { + "category_id": 13, + "poly": [ + 298, + 1503, + 435, + 1503, + 435, + 1537, + 298, + 1537 + ], + "score": 0.92, + "latex": "f ( t ) = \\lfloor \\lambda t \\rfloor" + }, + { + "category_id": 13, + "poly": [ + 747, + 1305, + 809, + 1305, + 809, + 1334, + 747, + 1334 + ], + "score": 0.86, + "latex": "1 . 5 \\%" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 230.0, + 623.0, + 230.0, + 623.0, + 254.0, + 596.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 244.0, + 693.0, + 244.0, + 693.0, + 264.0, + 654.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 249.0, + 655.0, + 249.0, + 655.0, + 259.0, + 637.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 266.0, + 624.0, + 266.0, + 624.0, + 291.0, + 596.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 259.0, + 735.0, + 259.0, + 735.0, + 293.0, + 632.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 284.0, + 689.0, + 284.0, + 689.0, + 308.0, + 654.0, + 308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 287.0, + 658.0, + 287.0, + 658.0, + 305.0, + 631.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 304.0, + 624.0, + 304.0, + 624.0, + 328.0, + 596.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 301.0, + 716.0, + 301.0, + 716.0, + 321.0, + 630.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 319.0, + 750.0, + 319.0, + 750.0, + 327.0, + 738.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 341.0, + 624.0, + 341.0, + 624.0, + 366.0, + 596.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 344.0, + 782.0, + 344.0, + 782.0, + 352.0, + 770.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 369.0, + 602.0, + 369.0, + 602.0, + 408.0, + 576.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 378.0, + 623.0, + 378.0, + 623.0, + 402.0, + 595.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 407.0, + 655.0, + 407.0, + 655.0, + 419.0, + 641.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 413.0, + 625.0, + 413.0, + 625.0, + 441.0, + 595.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 451.0, + 625.0, + 451.0, + 625.0, + 478.0, + 594.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 489.0, + 623.0, + 489.0, + 623.0, + 513.0, + 597.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 515.0, + 682.0, + 515.0, + 682.0, + 522.0, + 672.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 516.0, + 752.0, + 516.0, + 752.0, + 523.0, + 734.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 513.0, + 1093.0, + 513.0, + 1093.0, + 524.0, + 1075.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 519.0, + 1067.0, + 519.0, + 1067.0, + 526.0, + 1053.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 539.0, + 654.0, + 539.0, + 654.0, + 557.0, + 637.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 536.0, + 723.0, + 536.0, + 723.0, + 560.0, + 696.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 535.0, + 786.0, + 535.0, + 786.0, + 559.0, + 759.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 535.0, + 849.0, + 535.0, + 849.0, + 559.0, + 823.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 535.0, + 913.0, + 535.0, + 913.0, + 559.0, + 885.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 536.0, + 977.0, + 536.0, + 977.0, + 559.0, + 949.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 536.0, + 1040.0, + 536.0, + 1040.0, + 559.0, + 1013.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 536.0, + 1103.0, + 536.0, + 1103.0, + 559.0, + 1076.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 549.0, + 930.0, + 549.0, + 930.0, + 581.0, + 806.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 589.0, + 1144.0, + 589.0, + 1144.0, + 626.0, + 555.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1663.0, + 550.0, + 1663.0, + 550.0, + 1715.0, + 292.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1091.0, + 586.0, + 1091.0, + 586.0, + 1130.0, + 294.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 647.0, + 1403.0, + 647.0, + 1403.0, + 683.0, + 293.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 677.0, + 980.0, + 677.0, + 980.0, + 717.0, + 294.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1741.0, + 1406.0, + 1741.0, + 1406.0, + 1774.0, + 295.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1770.0, + 1406.0, + 1770.0, + 1406.0, + 1808.0, + 291.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1796.0, + 1406.0, + 1796.0, + 1406.0, + 1841.0, + 290.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1832.0, + 1405.0, + 1832.0, + 1405.0, + 1868.0, + 294.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1861.0, + 1408.0, + 1861.0, + 1408.0, + 1898.0, + 292.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1892.0, + 1405.0, + 1892.0, + 1405.0, + 1929.0, + 295.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1922.0, + 1406.0, + 1922.0, + 1406.0, + 1961.0, + 292.0, + 1961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1952.0, + 1406.0, + 1952.0, + 1406.0, + 1991.0, + 292.0, + 1991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1984.0, + 1366.0, + 1984.0, + 1366.0, + 2019.0, + 295.0, + 2019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1378.0, + 1405.0, + 1378.0, + 1405.0, + 1419.0, + 295.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1413.0, + 1405.0, + 1413.0, + 1405.0, + 1446.0, + 294.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1443.0, + 1404.0, + 1443.0, + 1404.0, + 1477.0, + 293.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1473.0, + 1404.0, + 1473.0, + 1404.0, + 1507.0, + 294.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1504.0, + 297.0, + 1504.0, + 297.0, + 1537.0, + 294.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 1504.0, + 1404.0, + 1504.0, + 1404.0, + 1537.0, + 436.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1532.0, + 1405.0, + 1532.0, + 1405.0, + 1570.0, + 293.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1561.0, + 1405.0, + 1561.0, + 1405.0, + 1602.0, + 292.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1594.0, + 1359.0, + 1594.0, + 1359.0, + 1631.0, + 293.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1152.0, + 1405.0, + 1152.0, + 1405.0, + 1186.0, + 296.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1183.0, + 1404.0, + 1183.0, + 1404.0, + 1217.0, + 294.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1210.0, + 1404.0, + 1210.0, + 1404.0, + 1249.0, + 293.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1244.0, + 1405.0, + 1244.0, + 1405.0, + 1279.0, + 293.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1274.0, + 1405.0, + 1274.0, + 1405.0, + 1309.0, + 294.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1304.0, + 746.0, + 1304.0, + 746.0, + 1341.0, + 293.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 1304.0, + 1406.0, + 1304.0, + 1406.0, + 1341.0, + 810.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1333.0, + 1348.0, + 1333.0, + 1348.0, + 1373.0, + 292.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 994.0, + 1404.0, + 994.0, + 1404.0, + 1029.0, + 295.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1023.0, + 1401.0, + 1023.0, + 1401.0, + 1061.0, + 295.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 647.0, + 1403.0, + 647.0, + 1403.0, + 683.0, + 293.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 677.0, + 980.0, + 677.0, + 980.0, + 717.0, + 294.0, + 717.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 332, + 1405, + 332, + 1405, + 427, + 298, + 427 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.885 + }, + { + "category_id": 1, + "poly": [ + 291, + 279, + 1336, + 279, + 1336, + 313, + 291, + 313 + ], + "score": 0.882 + }, + { + "category_id": 1, + "poly": [ + 299, + 1942, + 1402, + 1942, + 1402, + 2035, + 299, + 2035 + ], + "score": 0.874 + }, + { + "category_id": 1, + "poly": [ + 297, + 1217, + 1404, + 1217, + 1404, + 1314, + 297, + 1314 + ], + "score": 0.871 + }, + { + "category_id": 1, + "poly": [ + 297, + 1448, + 1406, + 1448, + 1406, + 1545, + 297, + 1545 + ], + "score": 0.87 + }, + { + "category_id": 0, + "poly": [ + 300, + 227, + 489, + 227, + 489, + 262, + 300, + 262 + ], + "score": 0.867 + }, + { + "category_id": 1, + "poly": [ + 296, + 1333, + 1405, + 1333, + 1405, + 1428, + 296, + 1428 + ], + "score": 0.859 + }, + { + "category_id": 1, + "poly": [ + 296, + 1018, + 1401, + 1018, + 1401, + 1082, + 296, + 1082 + ], + "score": 0.859 + }, + { + "category_id": 1, + "poly": [ + 297, + 1564, + 1406, + 1564, + 1406, + 1658, + 297, + 1658 + ], + "score": 0.853 + }, + { + "category_id": 1, + "poly": [ + 291, + 447, + 1403, + 447, + 1403, + 514, + 291, + 514 + ], + "score": 0.849 + }, + { + "category_id": 1, + "poly": [ + 296, + 817, + 1399, + 817, + 1399, + 883, + 296, + 883 + ], + "score": 0.848 + }, + { + "category_id": 1, + "poly": [ + 297, + 1103, + 1408, + 1103, + 1408, + 1196, + 297, + 1196 + ], + "score": 0.847 + }, + { + "category_id": 1, + "poly": [ + 298, + 1794, + 1405, + 1794, + 1405, + 1920, + 298, + 1920 + ], + "score": 0.847 + }, + { + "category_id": 1, + "poly": [ + 297, + 648, + 1403, + 648, + 1403, + 713, + 297, + 713 + ], + "score": 0.836 + }, + { + "category_id": 1, + "poly": [ + 297, + 732, + 1398, + 732, + 1398, + 798, + 297, + 798 + ], + "score": 0.835 + }, + { + "category_id": 1, + "poly": [ + 297, + 902, + 1407, + 902, + 1407, + 997, + 297, + 997 + ], + "score": 0.832 + }, + { + "category_id": 1, + "poly": [ + 298, + 531, + 1402, + 531, + 1402, + 628, + 298, + 628 + ], + "score": 0.831 + }, + { + "category_id": 1, + "poly": [ + 296, + 1680, + 1404, + 1680, + 1404, + 1775, + 296, + 1775 + ], + "score": 0.821 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.806 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 490.0, + 230.0, + 490.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 868.0, + 2084.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 332.0, + 1405.0, + 332.0, + 1405.0, + 371.0, + 293.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 364.0, + 1408.0, + 364.0, + 1408.0, + 403.0, + 321.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 396.0, + 575.0, + 396.0, + 575.0, + 428.0, + 319.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 277.0, + 1336.0, + 277.0, + 1336.0, + 316.0, + 292.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1976.0, + 297.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2007.0, + 324.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2003.0, + 861.0, + 2003.0, + 861.0, + 2037.0, + 323.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1216.0, + 1409.0, + 1216.0, + 1409.0, + 1256.0, + 292.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1251.0, + 1405.0, + 1251.0, + 1405.0, + 1285.0, + 324.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1280.0, + 1287.0, + 1280.0, + 1287.0, + 1315.0, + 322.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1447.0, + 1406.0, + 1447.0, + 1406.0, + 1485.0, + 294.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1482.0, + 1404.0, + 1482.0, + 1404.0, + 1515.0, + 321.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1511.0, + 950.0, + 1511.0, + 950.0, + 1546.0, + 323.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1334.0, + 1405.0, + 1334.0, + 1405.0, + 1368.0, + 296.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1363.0, + 1408.0, + 1363.0, + 1408.0, + 1401.0, + 322.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1398.0, + 574.0, + 1398.0, + 574.0, + 1428.0, + 321.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1014.0, + 1407.0, + 1014.0, + 1407.0, + 1056.0, + 292.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1050.0, + 479.0, + 1050.0, + 479.0, + 1083.0, + 323.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1565.0, + 1404.0, + 1565.0, + 1404.0, + 1599.0, + 295.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1596.0, + 1408.0, + 1596.0, + 1408.0, + 1630.0, + 323.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1623.0, + 399.0, + 1623.0, + 399.0, + 1660.0, + 320.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 447.0, + 1404.0, + 447.0, + 1404.0, + 484.0, + 297.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 479.0, + 1392.0, + 479.0, + 1392.0, + 514.0, + 321.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 815.0, + 1404.0, + 815.0, + 1404.0, + 858.0, + 292.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 849.0, + 933.0, + 849.0, + 933.0, + 885.0, + 322.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1102.0, + 1406.0, + 1102.0, + 1406.0, + 1140.0, + 294.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1136.0, + 1408.0, + 1136.0, + 1408.0, + 1170.0, + 323.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1164.0, + 397.0, + 1164.0, + 397.0, + 1198.0, + 321.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1792.0, + 1409.0, + 1792.0, + 1409.0, + 1832.0, + 295.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1826.0, + 1405.0, + 1826.0, + 1405.0, + 1859.0, + 322.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1857.0, + 1406.0, + 1857.0, + 1406.0, + 1893.0, + 321.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1886.0, + 868.0, + 1886.0, + 868.0, + 1923.0, + 321.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 651.0, + 1404.0, + 651.0, + 1404.0, + 683.0, + 299.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 679.0, + 1375.0, + 679.0, + 1375.0, + 714.0, + 321.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 731.0, + 1403.0, + 731.0, + 1403.0, + 770.0, + 295.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 765.0, + 1034.0, + 765.0, + 1034.0, + 799.0, + 322.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 904.0, + 1403.0, + 904.0, + 1403.0, + 939.0, + 295.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 931.0, + 1412.0, + 931.0, + 1412.0, + 976.0, + 319.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 965.0, + 533.0, + 965.0, + 533.0, + 997.0, + 324.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 531.0, + 1403.0, + 531.0, + 1403.0, + 570.0, + 294.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 566.0, + 1406.0, + 566.0, + 1406.0, + 600.0, + 323.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 594.0, + 721.0, + 594.0, + 721.0, + 632.0, + 322.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1678.0, + 1405.0, + 1678.0, + 1405.0, + 1718.0, + 293.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1712.0, + 1405.0, + 1712.0, + 1405.0, + 1746.0, + 323.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1742.0, + 937.0, + 1742.0, + 937.0, + 1776.0, + 323.0, + 1776.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 861, + 2088, + 861, + 2113, + 835, + 2113 + ], + "score": 0.823 + }, + { + "category_id": 1, + "poly": [ + 296, + 1585, + 1403, + 1585, + 1403, + 1651, + 296, + 1651 + ], + "score": 0.625 + }, + { + "category_id": 1, + "poly": [ + 299, + 1667, + 1402, + 1667, + 1402, + 1733, + 299, + 1733 + ], + "score": 0.618 + }, + { + "category_id": 1, + "poly": [ + 296, + 1362, + 1403, + 1362, + 1403, + 1458, + 296, + 1458 + ], + "score": 0.603 + }, + { + "category_id": 1, + "poly": [ + 293, + 755, + 1400, + 755, + 1400, + 820, + 293, + 820 + ], + "score": 0.584 + }, + { + "category_id": 1, + "poly": [ + 293, + 1474, + 1403, + 1474, + 1403, + 1570, + 293, + 1570 + ], + "score": 0.583 + }, + { + "category_id": 1, + "poly": [ + 300, + 947, + 1402, + 947, + 1402, + 1043, + 300, + 1043 + ], + "score": 0.582 + }, + { + "category_id": 1, + "poly": [ + 298, + 1747, + 1400, + 1747, + 1400, + 1813, + 298, + 1813 + ], + "score": 0.582 + }, + { + "category_id": 1, + "poly": [ + 301, + 836, + 1401, + 836, + 1401, + 932, + 301, + 932 + ], + "score": 0.569 + }, + { + "category_id": 1, + "poly": [ + 296, + 674, + 1401, + 674, + 1401, + 739, + 296, + 739 + ], + "score": 0.562 + }, + { + "category_id": 1, + "poly": [ + 290, + 1059, + 1403, + 1059, + 1403, + 1124, + 290, + 1124 + ], + "score": 0.556 + }, + { + "category_id": 1, + "poly": [ + 300, + 229, + 1401, + 229, + 1401, + 324, + 300, + 324 + ], + "score": 0.555 + }, + { + "category_id": 1, + "poly": [ + 297, + 1139, + 1401, + 1139, + 1401, + 1236, + 297, + 1236 + ], + "score": 0.554 + }, + { + "category_id": 1, + "poly": [ + 298, + 562, + 1403, + 562, + 1403, + 657, + 298, + 657 + ], + "score": 0.549 + }, + { + "category_id": 1, + "poly": [ + 299, + 451, + 1400, + 451, + 1400, + 547, + 299, + 547 + ], + "score": 0.54 + }, + { + "category_id": 1, + "poly": [ + 299, + 1251, + 1401, + 1251, + 1401, + 1346, + 299, + 1346 + ], + "score": 0.535 + }, + { + "category_id": 1, + "poly": [ + 297, + 1828, + 1403, + 1828, + 1403, + 1923, + 297, + 1923 + ], + "score": 0.525 + }, + { + "category_id": 1, + "poly": [ + 301, + 339, + 1400, + 339, + 1400, + 435, + 301, + 435 + ], + "score": 0.485 + }, + { + "category_id": 1, + "poly": [ + 299, + 1941, + 1404, + 1941, + 1404, + 2034, + 299, + 2034 + ], + "score": 0.364 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1580.0, + 1407.0, + 1580.0, + 1407.0, + 1627.0, + 293.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1617.0, + 1131.0, + 1617.0, + 1131.0, + 1652.0, + 323.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1668.0, + 1407.0, + 1668.0, + 1407.0, + 1704.0, + 297.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1699.0, + 1300.0, + 1699.0, + 1300.0, + 1735.0, + 324.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1363.0, + 1405.0, + 1363.0, + 1405.0, + 1398.0, + 295.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1394.0, + 1404.0, + 1394.0, + 1404.0, + 1430.0, + 323.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1426.0, + 763.0, + 1426.0, + 763.0, + 1458.0, + 323.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 754.0, + 1404.0, + 754.0, + 1404.0, + 794.0, + 295.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 786.0, + 1172.0, + 786.0, + 1172.0, + 821.0, + 322.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1472.0, + 1409.0, + 1472.0, + 1409.0, + 1513.0, + 293.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1505.0, + 1405.0, + 1505.0, + 1405.0, + 1542.0, + 321.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1538.0, + 854.0, + 1538.0, + 854.0, + 1572.0, + 325.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 947.0, + 1403.0, + 947.0, + 1403.0, + 984.0, + 295.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 979.0, + 1406.0, + 979.0, + 1406.0, + 1016.0, + 321.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1010.0, + 721.0, + 1010.0, + 721.0, + 1044.0, + 321.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1746.0, + 1404.0, + 1746.0, + 1404.0, + 1785.0, + 294.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1780.0, + 1072.0, + 1780.0, + 1072.0, + 1814.0, + 323.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 839.0, + 1402.0, + 839.0, + 1402.0, + 869.0, + 297.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 867.0, + 1405.0, + 867.0, + 1405.0, + 902.0, + 321.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 899.0, + 721.0, + 899.0, + 721.0, + 933.0, + 322.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 676.0, + 1405.0, + 676.0, + 1405.0, + 712.0, + 295.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 707.0, + 1305.0, + 707.0, + 1305.0, + 740.0, + 323.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1061.0, + 1405.0, + 1061.0, + 1405.0, + 1097.0, + 294.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1092.0, + 1410.0, + 1092.0, + 1410.0, + 1125.0, + 321.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 1405.0, + 230.0, + 1405.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 259.0, + 1405.0, + 259.0, + 1405.0, + 296.0, + 321.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 292.0, + 921.0, + 292.0, + 921.0, + 326.0, + 325.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1138.0, + 1405.0, + 1138.0, + 1405.0, + 1178.0, + 293.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1173.0, + 1405.0, + 1173.0, + 1405.0, + 1207.0, + 324.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1205.0, + 1007.0, + 1205.0, + 1007.0, + 1235.0, + 323.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 560.0, + 1408.0, + 560.0, + 1408.0, + 603.0, + 293.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 591.0, + 1409.0, + 591.0, + 1409.0, + 634.0, + 321.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 625.0, + 533.0, + 625.0, + 533.0, + 657.0, + 325.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 451.0, + 1406.0, + 451.0, + 1406.0, + 489.0, + 295.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 483.0, + 1405.0, + 483.0, + 1405.0, + 519.0, + 320.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 512.0, + 1168.0, + 512.0, + 1168.0, + 549.0, + 319.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1251.0, + 1403.0, + 1251.0, + 1403.0, + 1288.0, + 295.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1281.0, + 1405.0, + 1281.0, + 1405.0, + 1319.0, + 322.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1314.0, + 656.0, + 1314.0, + 656.0, + 1347.0, + 323.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1828.0, + 1404.0, + 1828.0, + 1404.0, + 1865.0, + 295.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1862.0, + 1408.0, + 1862.0, + 1408.0, + 1896.0, + 324.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1890.0, + 576.0, + 1890.0, + 576.0, + 1924.0, + 320.0, + 1924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 337.0, + 1404.0, + 337.0, + 1404.0, + 378.0, + 294.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 374.0, + 1404.0, + 374.0, + 1404.0, + 408.0, + 323.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 403.0, + 714.0, + 403.0, + 714.0, + 436.0, + 322.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1936.0, + 1405.0, + 1936.0, + 1405.0, + 1983.0, + 292.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1972.0, + 1403.0, + 1972.0, + 1403.0, + 2008.0, + 321.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2003.0, + 716.0, + 2003.0, + 716.0, + 2036.0, + 322.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.834 + }, + { + "category_id": 1, + "poly": [ + 295, + 1303, + 1403, + 1303, + 1403, + 1367, + 295, + 1367 + ], + "score": 0.75 + }, + { + "category_id": 1, + "poly": [ + 301, + 229, + 1401, + 229, + 1401, + 323, + 301, + 323 + ], + "score": 0.606 + }, + { + "category_id": 1, + "poly": [ + 298, + 1044, + 1403, + 1044, + 1403, + 1140, + 298, + 1140 + ], + "score": 0.587 + }, + { + "category_id": 1, + "poly": [ + 296, + 932, + 1401, + 932, + 1401, + 1027, + 296, + 1027 + ], + "score": 0.547 + }, + { + "category_id": 1, + "poly": [ + 297, + 735, + 1398, + 735, + 1398, + 800, + 297, + 800 + ], + "score": 0.542 + }, + { + "category_id": 1, + "poly": [ + 298, + 818, + 1400, + 818, + 1400, + 913, + 298, + 913 + ], + "score": 0.529 + }, + { + "category_id": 1, + "poly": [ + 296, + 1159, + 1403, + 1159, + 1403, + 1282, + 296, + 1282 + ], + "score": 0.511 + }, + { + "category_id": 1, + "poly": [ + 294, + 651, + 1400, + 651, + 1400, + 717, + 294, + 717 + ], + "score": 0.502 + }, + { + "category_id": 1, + "poly": [ + 294, + 342, + 1400, + 342, + 1400, + 407, + 294, + 407 + ], + "score": 0.492 + }, + { + "category_id": 1, + "poly": [ + 300, + 425, + 1401, + 425, + 1401, + 520, + 300, + 520 + ], + "score": 0.439 + }, + { + "category_id": 1, + "poly": [ + 296, + 539, + 1402, + 539, + 1402, + 634, + 296, + 634 + ], + "score": 0.432 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1298.0, + 1406.0, + 1298.0, + 1406.0, + 1343.0, + 294.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1335.0, + 1283.0, + 1335.0, + 1283.0, + 1367.0, + 324.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 226.0, + 1404.0, + 226.0, + 1404.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 260.0, + 1405.0, + 260.0, + 1405.0, + 297.0, + 321.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 291.0, + 715.0, + 291.0, + 715.0, + 325.0, + 322.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1043.0, + 1408.0, + 1043.0, + 1408.0, + 1083.0, + 293.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1077.0, + 1405.0, + 1077.0, + 1405.0, + 1112.0, + 323.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1105.0, + 1007.0, + 1105.0, + 1007.0, + 1142.0, + 322.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 928.0, + 1405.0, + 928.0, + 1405.0, + 971.0, + 293.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 964.0, + 1405.0, + 964.0, + 1405.0, + 998.0, + 320.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 994.0, + 902.0, + 994.0, + 902.0, + 1027.0, + 323.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 735.0, + 1403.0, + 735.0, + 1403.0, + 773.0, + 296.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 766.0, + 1197.0, + 766.0, + 1197.0, + 802.0, + 323.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 818.0, + 1405.0, + 818.0, + 1405.0, + 855.0, + 293.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 851.0, + 1405.0, + 851.0, + 1405.0, + 885.0, + 323.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 880.0, + 1242.0, + 880.0, + 1242.0, + 918.0, + 323.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1159.0, + 1406.0, + 1159.0, + 1406.0, + 1195.0, + 294.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1191.0, + 1405.0, + 1191.0, + 1405.0, + 1224.0, + 323.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1223.0, + 1408.0, + 1223.0, + 1408.0, + 1255.0, + 321.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1251.0, + 397.0, + 1251.0, + 397.0, + 1283.0, + 321.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 650.0, + 1405.0, + 650.0, + 1405.0, + 689.0, + 294.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 685.0, + 1042.0, + 685.0, + 1042.0, + 718.0, + 321.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 340.0, + 1404.0, + 340.0, + 1404.0, + 381.0, + 294.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 372.0, + 1371.0, + 372.0, + 1371.0, + 408.0, + 322.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 424.0, + 1406.0, + 424.0, + 1406.0, + 465.0, + 294.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 458.0, + 1405.0, + 458.0, + 1405.0, + 492.0, + 323.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 490.0, + 800.0, + 490.0, + 800.0, + 521.0, + 321.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 538.0, + 1404.0, + 538.0, + 1404.0, + 574.0, + 295.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 569.0, + 1404.0, + 569.0, + 1404.0, + 605.0, + 323.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 604.0, + 797.0, + 604.0, + 797.0, + 634.0, + 321.0, + 634.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 350, + 1404, + 350, + 1404, + 565, + 297, + 565 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1208, + 1406, + 1208, + 1406, + 1424, + 297, + 1424 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1516, + 1404, + 1516, + 1404, + 1762, + 298, + 1762 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1776, + 1405, + 1776, + 1405, + 1991, + 297, + 1991 + ], + "score": 0.98 + }, + { + "category_id": 5, + "poly": [ + 546, + 871, + 1152, + 871, + 1152, + 1105, + 546, + 1105 + ], + "score": 0.978, + "html": "
FeatureCharacteristicDimension
MFCCPitch20
MFCC deltaPitch20
Constant-QchromagramPitch12
TempogramStrength384
Onset strengthStrength1
Beat one-hotBeat1
" + }, + { + "category_id": 1, + "poly": [ + 298, + 659, + 1407, + 659, + 1407, + 781, + 298, + 781 + ], + "score": 0.977 + }, + { + "category_id": 6, + "poly": [ + 648, + 832, + 1051, + 832, + 1051, + 863, + 648, + 863 + ], + "score": 0.913 + }, + { + "category_id": 0, + "poly": [ + 300, + 1152, + 724, + 1152, + 724, + 1184, + 300, + 1184 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 301, + 602, + 733, + 602, + 733, + 633, + 301, + 633 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 299, + 293, + 878, + 293, + 878, + 326, + 299, + 326 + ], + "score": 0.871 + }, + { + "category_id": 0, + "poly": [ + 301, + 228, + 449, + 228, + 449, + 261, + 301, + 261 + ], + "score": 0.865 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.852 + }, + { + "category_id": 2, + "poly": [ + 333, + 2006, + 628, + 2006, + 628, + 2034, + 333, + 2034 + ], + "score": 0.851 + }, + { + "category_id": 0, + "poly": [ + 299, + 1459, + 1073, + 1459, + 1073, + 1491, + 299, + 1491 + ], + "score": 0.839 + }, + { + "category_id": 13, + "poly": [ + 1056, + 1960, + 1197, + 1960, + 1197, + 1993, + 1056, + 1993 + ], + "score": 0.92, + "latex": "\\Delta t = 1 / 1 5 s" + }, + { + "category_id": 13, + "poly": [ + 710, + 1639, + 793, + 1639, + 793, + 1671, + 710, + 1671 + ], + "score": 0.92, + "latex": "B _ { 2 } / B _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 588, + 1777, + 730, + 1777, + 730, + 1810, + 588, + 1810 + ], + "score": 0.92, + "latex": "\\Delta t = 1 / 1 5 s" + }, + { + "category_id": 13, + "poly": [ + 481, + 1699, + 690, + 1699, + 690, + 1733, + 481, + 1733 + ], + "score": 0.89, + "latex": "\\mid t _ { 1 } - t _ { 2 } \\mid \\leq \\Delta t . \\ t _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 879, + 1640, + 915, + 1640, + 915, + 1669, + 879, + 1669 + ], + "score": 0.89, + "latex": "B _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1670, + 334, + 1670, + 334, + 1700, + 298, + 1700 + ], + "score": 0.88, + "latex": "B _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1375, + 1702, + 1401, + 1702, + 1401, + 1730, + 1375, + 1730 + ], + "score": 0.84, + "latex": "t _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 936, + 1731, + 972, + 1731, + 972, + 1758, + 936, + 1758 + ], + "score": 0.76, + "latex": "\\Delta t" + }, + { + "category_id": 13, + "poly": [ + 1243, + 1639, + 1355, + 1639, + 1355, + 1668, + 1243, + 1668 + ], + "score": 0.75, + "latex": "1 5 { , } 4 0 0 \\mathrm { H z }" + }, + { + "category_id": 13, + "poly": [ + 634, + 1578, + 746, + 1578, + 746, + 1607, + 634, + 1607 + ], + "score": 0.71, + "latex": "2 2 { , } 0 5 0 \\mathrm { H z }" + }, + { + "category_id": 13, + "poly": [ + 475, + 1578, + 586, + 1578, + 586, + 1607, + 475, + 1607 + ], + "score": 0.68, + "latex": "1 5 { , } 4 0 0 \\mathrm { H z }" + }, + { + "category_id": 13, + "poly": [ + 1196, + 1700, + 1306, + 1700, + 1306, + 1729, + 1196, + 1729 + ], + "score": 0.67, + "latex": "1 5 { , } 4 0 0 \\mathrm { H z }" + }, + { + "category_id": 13, + "poly": [ + 296, + 1960, + 408, + 1960, + 408, + 1990, + 296, + 1990 + ], + "score": 0.57, + "latex": "1 5 { , } 4 0 0 \\mathrm { H z }" + }, + { + "category_id": 13, + "poly": [ + 648, + 1869, + 760, + 1869, + 760, + 1898, + 648, + 1898 + ], + "score": 0.57, + "latex": "1 5 { , } 4 0 0 \\mathrm { H z }" + }, + { + "category_id": 13, + "poly": [ + 707, + 1930, + 821, + 1930, + 821, + 1959, + 707, + 1959 + ], + "score": 0.52, + "latex": "2 2 { , } 0 5 0 \\mathrm { H z }" + }, + { + "category_id": 13, + "poly": [ + 543, + 1930, + 656, + 1930, + 656, + 1959, + 543, + 1959 + ], + "score": 0.47, + "latex": "1 5 { , } 4 0 0 \\mathrm { H z }" + }, + { + "category_id": 13, + "poly": [ + 665, + 1701, + 692, + 1701, + 692, + 1731, + 665, + 1731 + ], + "score": 0.47, + "latex": "t _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 594, + 1609, + 706, + 1609, + 706, + 1638, + 594, + 1638 + ], + "score": 0.42, + "latex": "2 2 , 0 5 0 \\mathrm { H z }" + }, + { + "category_id": 13, + "poly": [ + 457, + 1960, + 570, + 1960, + 570, + 1990, + 457, + 1990 + ], + "score": 0.42, + "latex": "2 2 { , } 0 5 0 \\mathrm { H z }" + }, + { + "category_id": 13, + "poly": [ + 813, + 1731, + 925, + 1731, + 925, + 1760, + 813, + 1760 + ], + "score": 0.41, + "latex": "2 2 { , } 0 5 0 \\mathrm { H z }" + }, + { + "category_id": 13, + "poly": [ + 453, + 1548, + 564, + 1548, + 564, + 1577, + 453, + 1577 + ], + "score": 0.4, + "latex": "1 5 { , } 4 0 0 \\mathrm { H z }" + }, + { + "category_id": 13, + "poly": [ + 1176, + 1670, + 1287, + 1670, + 1287, + 1699, + 1176, + 1699 + ], + "score": 0.37, + "latex": "2 2 { , } 0 5 0 \\mathrm { H z }" + }, + { + "category_id": 13, + "poly": [ + 1182, + 1869, + 1295, + 1869, + 1295, + 1898, + 1182, + 1898 + ], + "score": 0.35, + "latex": "2 2 { , } 0 5 0 \\mathrm { H z }" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 831.0, + 1052.0, + 831.0, + 1052.0, + 864.0, + 646.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1150.0, + 729.0, + 1150.0, + 729.0, + 1187.0, + 296.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 599.0, + 737.0, + 599.0, + 737.0, + 638.0, + 295.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 292.0, + 882.0, + 292.0, + 882.0, + 329.0, + 295.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 225.0, + 454.0, + 225.0, + 454.0, + 268.0, + 296.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 867.0, + 2084.0, + 867.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1999.0, + 631.0, + 1999.0, + 631.0, + 2040.0, + 330.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1457.0, + 1077.0, + 1457.0, + 1077.0, + 1496.0, + 295.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 350.0, + 1405.0, + 350.0, + 1405.0, + 385.0, + 294.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 381.0, + 1405.0, + 381.0, + 1405.0, + 416.0, + 294.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 413.0, + 1405.0, + 413.0, + 1405.0, + 448.0, + 295.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 440.0, + 1404.0, + 440.0, + 1404.0, + 478.0, + 294.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 471.0, + 1405.0, + 471.0, + 1405.0, + 508.0, + 294.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 502.0, + 1406.0, + 502.0, + 1406.0, + 540.0, + 294.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 536.0, + 778.0, + 536.0, + 778.0, + 567.0, + 295.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1206.0, + 1408.0, + 1206.0, + 1408.0, + 1246.0, + 292.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1239.0, + 1406.0, + 1239.0, + 1406.0, + 1277.0, + 292.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1270.0, + 1406.0, + 1270.0, + 1406.0, + 1305.0, + 294.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1300.0, + 1404.0, + 1300.0, + 1404.0, + 1335.0, + 292.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1330.0, + 1407.0, + 1330.0, + 1407.0, + 1367.0, + 292.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1362.0, + 1406.0, + 1362.0, + 1406.0, + 1396.0, + 295.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1394.0, + 1145.0, + 1394.0, + 1145.0, + 1425.0, + 295.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1518.0, + 1406.0, + 1518.0, + 1406.0, + 1551.0, + 294.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1548.0, + 452.0, + 1548.0, + 452.0, + 1581.0, + 294.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 1548.0, + 1405.0, + 1548.0, + 1405.0, + 1581.0, + 565.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1578.0, + 474.0, + 1578.0, + 474.0, + 1611.0, + 294.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1578.0, + 633.0, + 1578.0, + 633.0, + 1611.0, + 587.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1578.0, + 1405.0, + 1578.0, + 1405.0, + 1611.0, + 747.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1609.0, + 593.0, + 1609.0, + 593.0, + 1643.0, + 294.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1609.0, + 1406.0, + 1609.0, + 1406.0, + 1643.0, + 707.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1639.0, + 709.0, + 1639.0, + 709.0, + 1673.0, + 294.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1639.0, + 878.0, + 1639.0, + 878.0, + 1673.0, + 794.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1639.0, + 1242.0, + 1639.0, + 1242.0, + 1673.0, + 916.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 1639.0, + 1406.0, + 1639.0, + 1406.0, + 1673.0, + 1356.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 1669.0, + 1175.0, + 1669.0, + 1175.0, + 1703.0, + 335.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1288.0, + 1669.0, + 1406.0, + 1669.0, + 1406.0, + 1703.0, + 1288.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1700.0, + 480.0, + 1700.0, + 480.0, + 1734.0, + 294.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 1700.0, + 1195.0, + 1700.0, + 1195.0, + 1734.0, + 693.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 1700.0, + 1374.0, + 1700.0, + 1374.0, + 1734.0, + 1307.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1700.0, + 1405.0, + 1700.0, + 1405.0, + 1734.0, + 1402.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1732.0, + 812.0, + 1732.0, + 812.0, + 1763.0, + 293.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 1732.0, + 935.0, + 1732.0, + 935.0, + 1763.0, + 926.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 1732.0, + 1282.0, + 1732.0, + 1282.0, + 1763.0, + 973.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1776.0, + 587.0, + 1776.0, + 587.0, + 1811.0, + 294.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 1776.0, + 1403.0, + 1776.0, + 1403.0, + 1811.0, + 731.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1807.0, + 1405.0, + 1807.0, + 1405.0, + 1842.0, + 294.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1837.0, + 1406.0, + 1837.0, + 1406.0, + 1872.0, + 294.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1867.0, + 647.0, + 1867.0, + 647.0, + 1903.0, + 292.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 1867.0, + 1181.0, + 1867.0, + 1181.0, + 1903.0, + 761.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 1867.0, + 1408.0, + 1867.0, + 1408.0, + 1903.0, + 1296.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1898.0, + 1407.0, + 1898.0, + 1407.0, + 1931.0, + 292.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1928.0, + 542.0, + 1928.0, + 542.0, + 1962.0, + 294.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 1928.0, + 706.0, + 1928.0, + 706.0, + 1962.0, + 657.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 1928.0, + 1405.0, + 1928.0, + 1405.0, + 1962.0, + 822.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 1961.0, + 456.0, + 1961.0, + 456.0, + 1992.0, + 409.0, + 1992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 1961.0, + 1055.0, + 1961.0, + 1055.0, + 1992.0, + 571.0, + 1992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 1961.0, + 1207.0, + 1961.0, + 1207.0, + 1992.0, + 1198.0, + 1992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 656.0, + 1409.0, + 656.0, + 1409.0, + 695.0, + 292.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 689.0, + 1408.0, + 689.0, + 1408.0, + 723.0, + 292.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 716.0, + 1409.0, + 716.0, + 1409.0, + 756.0, + 292.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 752.0, + 918.0, + 752.0, + 918.0, + 782.0, + 295.0, + 782.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 600, + 297, + 1101, + 297, + 1101, + 456, + 600, + 456 + ], + "score": 0.972, + "html": "
CategoryBeat Alignment Ratio (%)
Ballet88.7
Hiphop93.2
Japanese Pop91.6
" + }, + { + "category_id": 3, + "poly": [ + 299, + 512, + 1396, + 512, + 1396, + 1057, + 299, + 1057 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 297, + 1080, + 1403, + 1080, + 1403, + 1175, + 297, + 1175 + ], + "score": 0.94 + }, + { + "category_id": 6, + "poly": [ + 614, + 251, + 1083, + 251, + 1083, + 284, + 614, + 284 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 104, + 299, + 104 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 863, + 2088, + 863, + 2112, + 835, + 2112 + ], + "score": 0.845 + }, + { + "category_id": 4, + "poly": [ + 297, + 1080, + 1403, + 1080, + 1403, + 1175, + 297, + 1175 + ], + "score": 0.098 + }, + { + "category_id": 13, + "poly": [ + 533, + 1142, + 674, + 1142, + 674, + 1176, + 533, + 1176 + ], + "score": 0.92, + "latex": "\\Delta t = 1 / 1 5 s" + }, + { + "category_id": 13, + "poly": [ + 992, + 251, + 1074, + 251, + 1074, + 285, + 992, + 285 + ], + "score": 0.91, + "latex": "B _ { 2 } / B _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 665, + 1112, + 777, + 1112, + 777, + 1141, + 665, + 1141 + ], + "score": 0.76, + "latex": "1 5 { , } 4 0 0 \\mathrm { H z }" + }, + { + "category_id": 13, + "poly": [ + 827, + 1112, + 941, + 1112, + 941, + 1141, + 827, + 1141 + ], + "score": 0.54, + "latex": "2 2 { , } 0 5 0 \\mathrm { H z }" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 526.0, + 333.0, + 526.0, + 333.0, + 549.0, + 300.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 539.0, + 1099.0, + 539.0, + 1099.0, + 632.0, + 1088.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 516.0, + 1383.0, + 516.0, + 1383.0, + 596.0, + 1165.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 581.0, + 1371.0, + 581.0, + 1371.0, + 618.0, + 1166.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 621.0, + 334.0, + 621.0, + 334.0, + 645.0, + 300.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 715.0, + 334.0, + 715.0, + 334.0, + 739.0, + 300.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 811.0, + 333.0, + 811.0, + 333.0, + 835.0, + 300.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 907.0, + 334.0, + 907.0, + 334.0, + 929.0, + 300.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 1002.0, + 331.0, + 1002.0, + 331.0, + 1021.0, + 301.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1034.0, + 480.0, + 1034.0, + 480.0, + 1058.0, + 451.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 1036.0, + 604.0, + 1036.0, + 604.0, + 1055.0, + 589.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1034.0, + 743.0, + 1034.0, + 743.0, + 1058.0, + 715.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 1035.0, + 871.0, + 1035.0, + 871.0, + 1057.0, + 850.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 1035.0, + 1003.0, + 1035.0, + 1003.0, + 1057.0, + 982.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 1035.0, + 1136.0, + 1035.0, + 1136.0, + 1057.0, + 1114.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 1036.0, + 1268.0, + 1036.0, + 1268.0, + 1057.0, + 1246.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1379.0, + 1039.0, + 1397.0, + 1039.0, + 1397.0, + 1054.0, + 1379.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1081.0, + 1403.0, + 1081.0, + 1403.0, + 1115.0, + 296.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1110.0, + 664.0, + 1110.0, + 664.0, + 1146.0, + 294.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 1110.0, + 826.0, + 1110.0, + 826.0, + 1146.0, + 778.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 1110.0, + 1404.0, + 1110.0, + 1404.0, + 1146.0, + 942.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1142.0, + 532.0, + 1142.0, + 532.0, + 1176.0, + 296.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 1142.0, + 685.0, + 1142.0, + 685.0, + 1176.0, + 675.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 246.0, + 991.0, + 246.0, + 991.0, + 288.0, + 613.0, + 288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 246.0, + 1086.0, + 246.0, + 1086.0, + 288.0, + 1075.0, + 288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 869.0, + 2084.0, + 869.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1081.0, + 1403.0, + 1081.0, + 1403.0, + 1115.0, + 296.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1110.0, + 664.0, + 1110.0, + 664.0, + 1146.0, + 294.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 1110.0, + 826.0, + 1110.0, + 826.0, + 1146.0, + 778.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 1110.0, + 1404.0, + 1110.0, + 1404.0, + 1146.0, + 942.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1142.0, + 532.0, + 1142.0, + 532.0, + 1176.0, + 296.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 1142.0, + 685.0, + 1142.0, + 685.0, + 1176.0, + 675.0, + 1176.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/xQGYquca0gB/xQGYquca0gB.md b/parse/train/xQGYquca0gB/xQGYquca0gB.md new file mode 100644 index 0000000000000000000000000000000000000000..1824a7582a47b62ea5560ae3abe63574aaabf738 --- /dev/null +++ b/parse/train/xQGYquca0gB/xQGYquca0gB.md @@ -0,0 +1,297 @@ +# Neural Production Systems + +Aniket Didolkar\*, 1, Anirudh Goyal \*, 1, Nan Rosemary $\mathbf { K } \mathbf { e } ^ { 2 }$ , Charles Blundell 2, Philippe Beaudoin 3 Nicolas Heess 2, Michael Mozer 4, \*\*, Yoshua Bengio 1, \*\* + +# Abstract + +Visual environments are structured, consisting of distinct objects or entities. These entities have properties—visible or latent—that determine the manner in which they interact with one another. To partition images into entities, deep-learning researchers have proposed structural inductive biases such as slot-based architectures. To model interactions among entities, equivariant graph neural nets (GNNs) are used, but these are not particularly well suited to the task for two reasons. First, GNNs do not predispose interactions to be sparse, as relationships among independent entities are likely to be. Second, GNNs do not factorize knowledge about interactions in an entity-conditional manner. As an alternative, we take inspiration from cognitive science and resurrect a classic approach, production systems, which consist of a set of rule templates that are applied by binding placeholder variables in the rules to specific entities. Rules are scored on their match to entities, and the best fitting rules are applied to update entity properties. In a series of experiments, we demonstrate that this architecture achieves a flexible, dynamic flow of control and serves to factorize entity-specific and rule-based information. This disentangling of knowledge achieves robust future-state prediction in rich visual environments, outperforming state-of-the-art methods using GNNs, and allows for the extrapolation from simple (few object) environments to more complex environments. + +# 1 Introduction + +Despite never having taken a physics course, every child beyond a young age appreciates that pushing a plate off the dining table will cause the plate to break. The laws of physics accurately characterize the dynamics of our natural world, and although explicit knowledge of these laws is not necessary to reason, we can reason explicitly about objects interacting through these laws. Humans can verbalize knowledge in propositional expressions such as “If a plate drops from table height, it will break,” and “If a video-game opponent approaches from behind and they are carrying a weapon, they are likely to attack you.” Expressing propositional knowledge is not a strength of current deep learning methods for several reasons. First, propositions are discrete and independent from one another. Second, propositions must be quantified in the manner of first-order logic; for example, the video-game proposition applies to any $X$ for which $X$ is an opponent and has a weapon. Incorporating the ability to express and reason about propositions should improve generalization in deep learning methods because this knowledge is modular— propositions can be formulated independently of each other— and can therefore be acquired incrementally. Propositions can also be composed with each other and applied consistently to all entities that match, yielding a powerful form of systematic generalization. + +The classical AI literature from the 1980s can offer deep learning researchers a valuable perspective. In this era, reasoning, planning, and prediction were handled by architectures that performed proposi tional inference on symbolic knowledge representations. A simple example of such an architecture is the production system (Laird et al., 1986; Anderson, 1987), which expresses knowledge by conditionaction rules. The rules operate on a working memory: rule conditions are matched to entities in working memory inspired by cognitive science, and such a match can trigger computational actions that update working memory or external actions that operate on the outside world. + +![](images/efc8469bfcada4d38e5ed73c14c9b9bca7fe99729b314148655c694d81e11332.jpg) +Figure 1: In this figure we show a visual comparison between NPS and dense architectures like GNNs. In NPS, a rule is only applied when an interaction takes place and it is applied only to the slots affected by the interaction. NPS also uses different rules for different kinds of interactions, while in GNN a common rule is applied to all slots irrespective of whether an interaction takes place or not (because of parameter sharing). Note the dynamic nature of the interaction graph in NPS, while in GNN, the graph is static. + +Production systems were typically used to model high-level cognition, e.g., mathematical problem solving or procedure following; perception was not the focus of these models. It was assumed that the results of perception were placed into working memory in a symbolic form that could be operated on with the rules. In this article, we revisit production systems but from a deep learning perspective which naturally integrates perceptual processing and subsequent inference for visual reasoning problems. We describe an end-to-end deep learning model that constructs object-centric representations of entities in videos, and then operates on these entities with differentiable—and thus learnable—production rules. The essence of these rules, carried over from traditional symbolic system, is that they operate on variables that are bound, or linked, to the entities in the world. In the deep learning implementation, each production rule is represented by a distinct MLP with query-key attention mechanisms to specify the rule-entity binding and to determine when the rule should be triggered for a given entity. We are not the first to propose a neural instantiation of a production system architecture. Touretzky & Hinton (1988) gave a proof of principle that neural net hardware could be hardwired to implement a production system for symbolic reasoning; our work fundamentally differs from theirs in that (1) we focus on perceptual inference problems and (2) we use the architecture as an inductive bias for learning. + +# 1.1 Variables and entities + +What makes a rule general-purpose is that it incorporates placeholder variables that can be bound to arbitrary values or—the term we prefer in this article—entities. This notion of binding is familiar in functional programming languages, where these variables are called arguments. Analogously, the use of variables in the production rules we describe enable a model to reason about any set of entities that satisfy the selection criteria of the rule. + +Consider a simple function in C like int add(int a, int b). This function binds its two integer operands to variables $a$ and $b$ . The function does not apply if the operands are, say, character strings. The use of variables enables a programmer to reuse the same function to add any two integer values + +In order for rules to operate on entities, these entities must be represented explicitly. That is, the visual world needs to be parsed in a task-relevant manner, e.g., distinguishing the sprites in a video game or the vehicles and pedestrians approaching an autonomous vehicle. Only in the past few years have deep learning vision researchers developed methods for object-centric representation (Le Roux et al., 2011; Eslami et al., 2016; Greff et al., 2016; Raposo et al., 2017; Van Steenkiste et al., 2018; Kosiorek et al., 2018; Engelcke et al., 2019; Burgess et al., 2019; Greff et al., 2019; Locatello et al., 2020a; Ahmed et al., 2020; Goyal et al., 2019; Zablotskaia et al., 2020; Rahaman et al., 2020; Du et al., 2020; Ding et al., 2020; Goyal et al., 2020; Ke et al., 2021). These methods differ in details but share the notion of a fixed number of slots (see Figure 1 for example), also known as object files, each encapsulating information about a single object. Importantly, the slots are interchangeable, meaning that it doesn’t matter if a scene with an apple and an orange encodes the apple in slot 1 and orange in slot 2 or vice-versa. + +A model of visual reasoning must not only be able to represent entities but must also express knowledge about entity dynamics and interactions. To ensure systematic predictions, a model must be capable of applying knowledge to an entity regardless of the slot it is in and must be capable of applying the same knowledge to multiple instances of an entity. Several distinct approaches exist in the literature. The predominant approach uses graph neural networks to model slot-to-slot interactions (Scarselli et al., 2008; Bronstein et al., 2017; Watters et al., 2017; Van Steenkiste et al., 2018; Kipf et al., 2018; Battaglia et al., 2018; Tacchetti et al., 2018). To ensure systematicity, the GNN must share parameters among the edges. In a recent article, Goyal et al. (2020) developed a more general framework in which parameters are shared but slots can dynamically select which parameters to use in a state-dependent manner. Each set of parameters is referred to as a schema, and slots use a query-key attention mechanism to select which schema to apply at each time step. Multiple slots can select the same schema. In both GNNs and SCOFF, modeling dynamics involves each slot interacting with each other slot. In the work we describe in this article, we replace the direct slot-to-slot interactions with rules, which mediate sparse interactions among slots (See arrows in Figure 1). + +Thus our main contribution is that we introduce NPS, which offers a way to model dynamic and sparse interactions among the variables in a graph and also allows dynamic sharing of multiple sets of parameters among these interactions. Most architectures used for modelling interactions in the current literature use statically instantiated graph which model all possible interactions for a given variable at each step i.e. dense interactions. Also such dense architectures share a single set of parameters across all interactions which maybe quite restrictive in terms of representational capacity. A visual comparison between these two kinds of architectures is shown in Figure 1. Through our experiments we show the advantage of modeling interactions in the proposed manner using NPS in visually rich physical environments. We also show that our method results in an intuitive factorization of rules and entities. + +# 2 Production System + +Formally, our notion of a production system consists of a set of entities and a set of rules, along with a mechanism for selecting rules to apply on subsets of the entities. Implicit in a rule is a specification of the properties of relevant entities, e.g., a rule might apply to one type of sprite in a video game but not another. The control flow of a production system dynamically selects rules as well as bindings between rules and entities, allowing different rules to be chosen and different entities to be manipulated at each point in time. + +The neural production system we describe shares essential properties with traditional production system, particularly with regard to the compositionality and generality of the knowledge they embody. Lovett & Anderson (2005) describe four desirable properties commonly attributed to symbolic systems that apply to our work as well. + +Production rules are modular. Each production rule represents a unit of knowledge and are atomic such that any production rule can be intervened (added, modified or deleted) independently of other production rules in the system. + +Production rules are abstract. Production rules allow for generalization because their conditions may be represented as high-level abstract knowledge that match to a wide range of patterns. These conditions specify the attributes of relationship(s) between entities without specifying the entities themselves. The ability to represent abstract knowledge allows for the transfer of learning across different environments as long as they fit within the conditions of the given production rule. + +Production rules are sparse. In order that production rules have broad applicability, they involve only a subset of entities. This assumption imposes a strong prior that dependencies among entities are sparse. In the context of visual reasoning, we conjecture that this prior is superior to what has often been assumed in the past, particularly in the disentanglement literature—independence among entities Higgins et al. (2016); Chen et al. (2018). + +Production rules represent causal knowledge and are thus asymmetric. Each rule can be decomposed into a {condition, action} pair, where the action reflects a state change that is a causal consequence of the conditions being met. + +These four properties are sufficient conditions for knowledge to be expressed in production rule form. These properties specify how knowledge is represented, but not what knowledge is represented. The latter is inferred by learning mechanisms under the inductive bias provided by the form of production rules. + +# 3 Neural Production System: Slots and Sparse Rules + +The Neural Production System (NPS), illustrated in Figure 2, provides an architectural backbone that supports the detection and inference of entity (object) representations in an input sequence, and the underlying rules which govern the interactions between these entities in time and space. The input sequence indexed by time step $t$ , $\{ \pmb { x } ^ { 1 } , \dotsc , \pmb { x } ^ { t } , \dotsc , \pmb { x } ^ { T } \}$ , for instance the frames in a video, are processed by a neural encoder (Burgess et al., 2019; Greff et al., 2019; Goyal et al., 2019, 2020) applied to each $\mathbf { \boldsymbol { x } } ^ { t }$ , to obtain a set of $M$ entity representations $\{ V _ { 1 } ^ { t } , \ldots , \ldots \stackrel { \cdot } { , } V _ { M } ^ { t } \}$ , one for each of the $M$ slots. These representations describe an entity and are updated based on both the previous state, $V ^ { t - 1 }$ and the current input, $\mathbf { \boldsymbol { x } } ^ { t }$ . + +NPS consists of $N$ separately encoded rules, $\{ R _ { 1 } , R _ { 2 } , . . , R _ { N } \}$ . Each rule consists of two components, ${ \pmb R } _ { i } = ( \vec { R _ { i } } , M L P _ { i } )$ , where $\vec { R _ { i } }$ is a learned rule embedding vector, which can be thought of as a template defining the condition for when a rule applies; and $M L P _ { i }$ , which determines the action taken by a rule. Both $\vec { R _ { i } }$ and the parameters of $M L P _ { i }$ are learned along with the other parameters of the model using back-propagation on an objective optimized end-to-end. + +In the general form of the model, each slot selects a rule that will be applied to it to change its state. This can potentially be performed several times, with possibly different rules applied at each step. Rule selection is done using an attention mechanism described in detail below. Each rule specifies conditions and actions on a pair of slots. Therefore, while modifying the state of a slot using a rule, it can take the state of another slot into account. The slot which is being modified is called the primary slot and other is called the contextual slot. The contextual slot is also selected using an attention mechanism described in detail below. + +![](images/7621aad51e5596db4e7dfe45bfb3fb1b7959f7deac83d4c49f5f51470e8d77b1.jpg) +Figure 2: Rule and slot combinatorics. Condition-action rules specify how entities interact. Slots maintain the time-varying state of an entity. Every rule is matched to every pair of slots. Through key-value attention, a goodness of match is determined, and a rule is selected along with its binding to slots. + +# 3.1 Computational Steps in NPS + +In this section, we give a detailed description of the rule selection and application procedure for the slots. First, we will formalize the definitions of a few terms that we will use to explain our method. We use the term primary slot to refer to slot $V _ { p }$ whose state gets modified by a rule $\scriptstyle { \mathbf { } } R _ { r }$ . We use the term contextual slot to refer to the slot $V _ { c }$ that the rule $\scriptstyle { R _ { r } }$ takes into account while modifying the state of the primary slot $V _ { p }$ . + +Notation. We consider a set of $N$ rules $\{ R _ { 1 } , R _ { 2 } , \ldots , R _ { N } \}$ and a set of $_ { \mathbf { T } }$ input frames $\{ \pmb { x } ^ { 1 } , \pmb { x } ^ { 2 } , \dots , \pmb { x } ^ { T } \}$ . Each frame $\mathbf { \boldsymbol { x } } ^ { t }$ is encoded into a set of $M$ slots $\{ V _ { 1 } ^ { t } , V _ { 2 } ^ { t } , \ldots , \bar { V _ { M } ^ { t } } \}$ . In the following discussion, we omit the index over $\pmb { t }$ for simplicity. + +Step $\imath .$ is external to NPS and involves parsing an input image, $\mathbf { \boldsymbol { x } } ^ { t }$ , into slot-based entities conditioned on the previous state of the slot-based entities. Any of the methods proposed in the literature to obtain a slot-wise representation of entities can be used (Burgess et al., 2019; Greff et al., 2019; Goyal et al., 2019, 2020). The next three steps constitute the rule selection and application procedure. + +Step 2. For each primary slot $V _ { p }$ , we attend to a rule $\scriptstyle { R _ { r } }$ to be applied. Here, the queries come from the primary slot: $\begin{array} { r } { \pmb q _ { p } = \pmb V _ { p } W ^ { q } } \end{array}$ , and the keys come from the rules: $\pmb { k } _ { i } = \vec { R _ { i } } W ^ { k } \quad \forall i \in \{ 1 , \dots , N \}$ The rule is selected using a straight-through Gumbel softmax (Jang et al., 2016) to achieve a learnable hard decision: $r = \mathrm { a r g m a x } _ { i } ( q _ { p } k _ { i } + \gamma )$ , where $\gamma \sim \mathrm { G u m b e l } ( 0 , 1 )$ . This competition is a noisy version of rule matching and prioritization in traditional production systems. + +Step 3. For a given primary slot $V _ { p }$ and selected rule $\scriptstyle { R _ { r } }$ , a contextual slot $V _ { c }$ is selected using another attention mechanism. In this case the query comes from the primary slot: $q _ { p } = V _ { p } W ^ { q }$ , and the keys from all the slots: $\pmb { k } _ { j } = V _ { j } W ^ { q } \quad \forall j \in \{ 1 , \dots , M \}$ . The selection takes place using a straightthrough Gumbel softmax similar to step 2: $c = \mathrm { a r g m a x } _ { j } ( q _ { p } k _ { j } + \gamma )$ , where $\gamma \sim \mathrm { G u m b e l } ( 0 , 1 )$ Note that each rule application is sparse since it takes into account only 1 contextual slot for modifying a primary slot, while other methods like GNNs take into account all slots for modifying a primary slot. + +Step 4. Rule Application: the selected rule $\scriptstyle { R _ { r } }$ is applied to the primary slot $V _ { p }$ based on the rule and the current contents of the primary and contextual slots. The rule-specific $M L P _ { r }$ , takes as input the concatenated representation of the state of the primary and contextual slots, $V _ { p }$ and $V _ { c }$ , and produces an output, which is then used to change the state of the primary slot $V _ { p }$ by residual addition. + +# 3.2 Rule Application: Sequential vs Parallel Rule Application + +In the previous section, we have described how each rule application only considers another contextual slot for the given primary slot i.e., contextual sparsity. We can also consider application sparsity, wherein we use the rules to update the states of only a subset of the slots. In this scenario, only the selected slots would be primary slots. This setting will be helpful when there is an entity in an environment that is stationary, or it is following its own default dynamics unaffected by other entities. Therefore, it does not need to consider other entities to update its state. We explore two scenarios for enabling application sparsity. + +Parallel Rule Application. Each of the $M$ slots selects a rule to potentially change its state. To enable sparse changes, we provide an extra Null Rule in addition to the available $N$ rules. If a slot picks the null rule in step 2 of the above procedure, we do not update its state. + +Sequential Rule Application. In this setting, only one slot gets updated in each rule application step. Therefore, only one slot is selected as the primary slot. This can be facilitated by modifying step 2 above to select one {primary + +![](images/d2db80f1c75eadd39bbfb8c94bc9f20aca13ed711f1d46b4857e95be0ac7103f.jpg) +Figure 3: This figure demonstrates the sequential and parallel rule application. + +slot, rule} pair among $_ { N M }$ {rule, slot} pairs. The queries come from each slot: $\pmb { q } _ { j } = \pmb { V } _ { j } \pmb { W } ^ { q } \quad \forall j \in$ $\{ 1 , \ldots , M \}$ , the keys come from the rules: $\pmb { k } _ { i } = R _ { i } W ^ { k } \quad \forall i \in \{ 1 , \dots , N \}$ . The straight-through Gumbel softmax selects one (primary slot, rule) pair: $p , r = \mathrm { a r g m a x } _ { i , j } ( q _ { p } k _ { i } + \gamma )$ , where $\gamma \sim$ $\mathrm { G u m b e l } ( 0 , 1 )$ . In the sequential regime, we allow the rule application procedure (step 2, 3, 4 above) to be performed multiple times iteratively in $K$ rule application stages for each time-step $t$ . + +A pictorial demonstration of both rule application regimes can be found in Figure 3. We provide detailed algorithms for the sequential and parallel regimes in Appendix. + +# 4 Experiments + +We demonstrate the effectiveness of NPS on multiple tasks and compare to a comprehensive set of baselines. To show that NPS can learn intuitive rules from the data generating distribution, we design a couple of simple toy experiments with well-defined discrete operations. Results show that NPS can accurately recover each operation defined by the data and learn to represent each operation using a separate rule. We then move to a much more complicated and visually rich setting with abstract physical rules and show that factorization of knowledge into rules as offered by NPS does scale up to such settings. We study and compare the parallel and sequential rule application procedures and try to understand the settings which favour each. We then + +Table 1: This table shows the segregation of rules for the MNIST Transformation task. Each cell indicates the number of times the corresponding rule is used for the given operation. We can see that NPS automatically and perfectly learns a separate rule for each operation. + +
Rule 1Rule 2Rule3Rule 4
TranslateDownTranslate UpRotate RightRotate Left5039000
00495000
050300
0004981
+ +evaluate the benefits of reusable, dynamic and sparse interactions as offered by NPS in a wide variety of physical environments by comparing it against various baselines. We conduct ablation studies to assess the contribution of different components of NPS. Here we briefly outline the tasks considered and direct the reader to the Appendix for full details on each task and details on hyperparameter settings. + +Discussion of baselines. NPS is an interaction network, therefore we use other widely used interaction networks such as multihead attention and graph neural networks (Goyal et al. (2019), Goyal et al. (2020), Veerapaneni et al. (2019), Kipf et al. (2019)) for comparison. Goyal et al. (2019) and Goyal et al. (2020) use an attention based interaction network to capture interactions between the slots, while Veerapaneni et al. (2019) and Kipf et al. (2019) use a GNN based interaction network. We also consider the recently introduced convolutional interaction network (CIN) (Qi et al., 2021) which captures dense pairwise interactions like GNN but uses a convolutional network instead of MLPs to better utilize spatial information. The proposed method, similar to other interaction networks, is agnostic to the encoder backbone used to encode the input image into slots, therefore we compare NPS to other interaction networks across a wide-variety of encoder backbones. + +# 4.1 Learning intuitive rules with NPS: Toy Simulations + +We designed a couple of simple tasks with well-defined discrete rules to show that NPS can learn intuitive and interpretable rules. We also show the efficiency and effectiveness of the selection procedure (step 2 and step 3 in section 3.1) by comparing against a baseline with many more parameters. Both tasks require a single modification of only one of the available entities, therefore the use of sequential or parallel rule application would not make a difference here since parallel rule application in which all-but-one slots select the null rule is similar to sequential rule application with 1 rule application step. To simplify the presentation, we describe the setup for both tasks using the sequential rule application procedure. + +MNIST Transformation. We test whether NPS can learn simple rules for performing transformations on MNIST digits. We generate data with four transformations: {Translate Up, Translate Down, Rotate Right, Rotate Left}. We feed the input image $( X )$ and the transformation $( o )$ to be performed as a one-hot vector to the model. The detailed setup is described in Appendix. For this task, we evaluate whether NPS can learn to use a unique rule for each transformation. + +We use 4 rules corresponding to the 4 transformations with the hope that the correct transformations are recovered. Indeed, we observe that NPS successfully learns to represent each transformation using a separate rule as shown in Table 1. Our model achieves an MSE of 0.02. A visualization of the outputs from our model and further details can be found in Appendix C. + +Coordinate Arithmetic Task. The model is tasked with performing arithmetic operations on 2D coordinates. Given $( X _ { 0 } , Y _ { 0 } )$ and $( X _ { 1 } , Y _ { 1 } )$ , we can apply the following operations: $\{ \mathbf { X } \mathbf { A d } \mathbf { - }$ dition: $\bar { ( X _ { r } , Y _ { r } ) } \ = \ ( X _ { 0 } \bar { + } X _ { 1 } , Y _ { 0 } )$ , $\mathbf { X }$ Subtraction: $( X _ { r } , Y _ { r } ) \ : = \ : ( X _ { 0 } - X _ { 1 } , Y _ { 0 } )$ , Y Addition: $( X _ { r } , Y _ { r } ) = ( X _ { 0 } , Y _ { 0 } + Y _ { 1 } )$ , Y Subtraction: $X _ { r } , Y _ { r } = ( X _ { 0 } , Y _ { 0 } - Y _ { 1 } ) \}$ , where $( X _ { r } , Y _ { r } )$ is the resultant coordinate. + +In this task, the model is given 2 input coordinates $X = [ ( x _ { i } , y _ { i } ) , ( x _ { j } , \bar { y _ { j } } ) ]$ and the expected output coordinates $Y = [ ( \hat { x _ { i } } , \hat { y _ { i } } ) , ( \hat { x _ { j } } , \hat { y _ { j } } ) ] .$ . The model is supposed to infer the correct rule to produce the correct output coordinates. During data collection, the true output is obtained by performing a random transformation on a randomly selected coordinate in $X$ (primary coordinate), taking another randomly selected coordinate from $X$ (contextual coordinate) into account. The detailed setup is described in Appendix D. We use an NPS model with 4 rules for this task. We use the the selection procedure in step 2 and step 3 of algorithm 1 to select the primary coordinate, contextual coordinate, and the rule. For the baseline we replace the selection procedure in NPS (i.e. step 2 and step 3 in algorithm 1) with a routing MLP similar to Fedus et al. (2021). + +![](images/93fa4fe136835a329ff05316b00960119a6e62b649e79c974d4d8c5f78adba3a.jpg) +Figure 4: Coordinate Arithmetic Task. Here, we compare NPS to the baseline model in terms of segregation of rules as the training progresses. X-axis shows the epochs and Y-axis shows the frequency with which Rule $i$ is used for the given operation. We can see that NPS disentangles the operations perfectly as training progresses with a unique rule specializing to every operation while the baseline model fails to do so. + +This routing MLP has 3 heads (one each for selecting the primary coordinate, contextual coordinate, and the rule). The baseline has 4 times more parameters than NPS. The final output is produced by the rule MLP which does not have access to the true output, hence the model cannot simply copy the true output to produce the actual output. Unlike the MNIST transformation task, we do not provide the operation to be performed as a one-hot vector input to the model, therefore it needs to infer the available operations from the data demonstrations. + +We show the segregation of rules for NPS and the baseline in Figure 4. We can see that NPS learns to use a unique rule for each operation while the baseline struggles to disentangle the underlying operations properly. NPS also outperforms the baseline in terms of MSE achieving an MSE of $0 . 0 1 _ { \pm 0 . 0 0 1 }$ while the baseline achieves an MSE of $0 . 0 4 _ { \pm 0 . 0 0 8 }$ . To further confirm that NPS learns all the available operations correctly from raw data demonstrations, we use an NPS model with 5 rules. We expect that in this case NPS should utilize only 4 rules since the data describes only 4 unique operations and indeed we observe that NPS ends up mostly utilizing 4 of the available 5 rules as shown in Table 2. + +Table 2: This table shows segregation of rules when we use NPS with 5 rules but the data generation distributions describes only 4 possible operations. We can see that only 4 rules get majorly utilized thus confirming that NPS successfully recovers all possible operations described by the data. + +
Rule 1Rule 2Rule 3Rule 4Rule 5
X Addition3609945130
X Subtraction0482010
Y Subtraction03945320
Y Addition0571599335
+ +# 4.2 Parallel vs Sequential Rule Application + +We compare the parallel and sequential rule application procedures, to understand the settings that favour one or the other, over two tasks: (1) Bouncing Balls, (2) Shapes Stack. We use the term PNPS to refer to parallel rule application and SNPS to refer to sequential rule application. + +Shapes Stack. We use the shapes stack dataset introduced by Groth et al. (2018). This dataset consists of objects stacked on top of each other as shown in Figure 5. These objects fall under the influence of gravity. For our experiments, We follow the same setup as Qi et al. (2021). In this task, given the first frame, the model is tasked with predicting the object bounding boxes for the next $\pmb { t }$ timesteps. The first frame is encoded using a convolutional network followed by RoIPooling (Girshick (2015)) to extract object-centric visual features. The object-centric features are then passed to the dynamics model to predict object bounding boxes of the next $\pmb { t }$ steps. Qi et al. (2021) propose a Region Proposal Interaction Network (RPIN) to solve this task. The dynamics model in RPIN consists of an Interaction Network proposed in Battaglia et al. (2016). + +
Model NameTestTransfer
RPIN (Qi et al. (2021))1.254±0.0086.377±0.325
PNPSSNPS1.250±0.0075.411±0.45
1.68±0.025.80±0.15
+ +Table 3: Prediction error of the compared models on the shapes stack dataset (lower is better) for the test as well as transfer setting. In the test setting the number of rollout steps $\pmb { t }$ is set to 15 and in the transfer setting it is set to 30. We can see that PNPS outperforms the RPIN baseline in both the test and transfer setting while SNPS fails to do so. Results across 15 seeds. + +To better utilize spatial information, Qi et al. (2021) propose an extension of the interaction operators in interaction net to operate on 3D tensors. This is achieved by replacing the MLP operations in the original interaction networks with convolutions. They call this new network Convolutional Interaction Network (CIN). For the proposed model, we replace this CIN in RPIN by NPS. To ensure a fair comparison to CIN, we use CNNs to represent rules in NPS instead of MLPs. CIN captures all pairwise interactions between objects using a convolutional network. In NPS, we capture sparse interactions (contextual sparsity) as compared to dense pairwise interactions captured by CIN. Also, in NPS we update only a few subset of slots per step instead of all slots (application sparsity). + +We consider two evaluation settings. (1) Test setting: The number of rollout timesteps is same as that seen during training (i.e. $t = 1 5$ ); (2) Transfer Setting: The number of rollout timesteps is higher than that seen during training (i.e. $t = 3 0$ ). + +We present our results on the shapes stack dataset in Table 3. We can see that both PNPS and SNPS outperform the baseline RPIN in the transfer setting, while only PNPS outperforms the baseline in the test setting and SNPS fails to do so. We can see that PNPS outperforms SNPS. We attribute this to the reduced application sparsity with PNPS, i.e., it is more likely that the state of a slot gets updated in PNPS as compared to SNPS. For instance, consider an NPS model with $N$ uniformly chosen rules and $M$ slots. The probability that the state of a slot gets updated in PNPS is $P _ { P N P S } \dot { = } N - 1 / N$ (since 1 rule is the null rule), while the same probability for SNPS is $P _ { S N P S } = 1 / M$ (since only 1 slot gets updated per rule application step). + +For this task, we run both PNPS and SNPS for $N = \{ 1 , 2 , 4 , 6 \}$ rules and $M = 3$ . For any given $_ N$ , we observe that $P _ { P N P S } > P _ { S N P S }$ . Even when we have multiple rule application steps in SNPS, it might end up selecting the same slot to be updated in more than one of these steps. We report the best performance obtained for PNPS and SNPS across all $N$ , which is $N = \{ 2 + 1$ Null Rule} for PNPS and $N = 4$ for SNPS, in Table 3. Shapes stack is a dataset that would prefer a model with less application sparsity since all the objects are tightly bound to each other (objects are placed on top of each other), therefore all objects spend the majority of their time interacting with the objects directly above or below them. We attribute the higher performance of PNPS compared to RPIN to the higher contextual sparsity of PNPS. Each example in the shapes stack task consists of 3 objects. Even though the blocks are tightly bound to each other, each block is only affected by the objects it is in direct contact with. For example, the top-most object is only affected by the object directly below it. The contextual sparsity offered by PNPS is a strong inductive bias to model such sparse interactions while RPIN models all pairwise interactions between the objects. Figure 5 shows an intuitive illustration of the PNPS model for the shapes stack dataset. In the figure, Rule 2 actually refers to the Null Rule, while Rule 1 refers to all the other non-null rules. The bottom-most block picks the Null Rule most times, as the bottom-most block generally does not move. + +Bouncing Balls. We consider a bouncingballs environment in which multiple balls move with billiard-ball dynamics. We validate our model on a colored version of this dataset. This is a next-step prediction task in which the model is tasked with predicting the final binary mask of each ball. We compare the following methods: (a) SCOFF (Goyal et al., 2020): factorization of knowledge in terms of slots (object properties) and schemata, the latter capturing object dynamics; (b) ${ \mathrm { S C O F F } } { + } { + }$ : we extend SCOFF by using the idea of iterative competition as proposed in slot attention (SA) (Locatello et al., 2020a); SCOFF $^ +$ PNPS/SNPS: We replace pairwise slot-to-slot interaction in ${ \mathrm { S C O F F } } { + } { + }$ with parallel or sequential rule application. For comparing different methods, we use the Adjusted Rand Index or ARI (Rand, 1971). To investigate how the factorization in the form of rules allows for extrapolating knowledge from fewer to more objects, we increase the number of objects from 4 during training to 6-8 during testing. + +![](images/0bafd38ea5e1df32a31ff88f73b7b52d92be55ef3f1eb31c0228c553cd29402e.jpg) +Figure 5: Here we show the rule selection statistics from the proposed model for all entities in the shapes stack dataset across all examples. Each example contains 3 entities as shown above. Each cell in the table shows the probability with which the given rule is triggered for the corresponding entity. We can see that the bottom-most entity triggers rule 2 most of the time while the other 2 entities trigger rule 1 most often. This is quite intuitive as, for most examples, the bottom-most entity remains static and does not move at all while the upper entities fall. Therefore, rule 2 captures information which is relevant to static entities, while rule 1 captures physical rules relevant to the inter + +actions and motion of the upper entities.Table 4. Contrary to the shapes stack task, we see that SNPS outperforms PNPS for the bouncing balls task. The balls are not tightly bound together into a single tower as in the shapes stack. Most of the time, a single ball follows its own dynamics, only occasionally interacting with another ball. Rules in NPS capture interaction dynamics between entities, hence they would only be required to change the state of an entity when it interacts with another entity. In the case of bouncing balls, this interaction takes place through a collision between multiple balls. Since for a single ball, such collisions are rare, SNPS, which has higher application sparsity (less probability of modifying the state of an entity), performs better as compared to PNPS (lower application sparsity). Also note that, SNPS has the ability to compose multiple rules together by virtue of having multiple rule application stages. A visualization of the rule and entity selections by the proposed algorithm can be found in Appendix Figure 9. + +Given the analysis in this section, we can conclude that PNPS is expected to work better when interactions among entities are more frequent while SNPS is expected to work better when interactions are rare and most of the time, each entity follows its own dynamics. Note that, for both SNPS and PNPS, the rule application considers only 1 other entity as context. Therefore, both approaches have equal contextual sparsity while the baselines that we consider (SCOFF and RPIN) capture dense pairwise interactions. We discuss the benefits of contextual sparsity in more detail in the next section. More details regarding our setup for the above experiments can be found in Appendix. + +# 4.3 Benefits of Sparse Interactions Offered by NPS + +In NPS, one can view the computational graph as a dynamically constructed GNN resulting from applying dynamically selected rules, where the states of the slots are represented on the different nodes of the graph, and different rules dynamically instantiate an hyper-edge between a set of slots (the primary slot and the contextual slot). It is important to emphasize that the topology of the graph induced in NPS is dynamic and sparse (only a few nodes affected), while in most GNNs the topology is fixed and dense (all nodes affected). In this section, through a thorough set of experiments, we show that learning sparse and dynamic interactions using NPS indeed works better for the problems we consider than learning dense interactions using GNNs. We consider two types of tasks: (1) Learning Action Conditioned World Models (2) Physical Reasoning. We use SNPS for all these experiments since in the environments that we consider here, interactions among entities are rare. + +Learning Action-Conditioned World Models. For learning action-conditioned world models, we follow the same experimental setup as Kipf et al. (2019). Therefore, all the tasks in this section are next- $K$ step $K =$ $\{ 1 , 5 , 1 0 \} $ ) prediction tasks, given the intermediate actions, and with the predictions being performed in the latent space. We use the Hits at Rank 1 $( \mathrm { H } @ 1 )$ metrics described by Kipf et al. (2019) for evaluation. $\mathrm { H @ 1 }$ is 1 for a particular example if the predicted state representation is nearest to the encoded true observation and 0 otherwise. We report the average of this score over the test set (higher is better). + +
Model NameTestTransfer
SCOFF0.280.15
SCOFF++0.84370.2632
PNPS (10 Rules+1 Null Rule)0.78130.1997
SNPS (10 Rules)0.85180.3553
+ +Table 4: Here we show the ARI achieved by the models on the bouncing balls dataset (higher is better). We can see that SNPS outperforms SCOFF and ${ \mathrm { S C O F F + + } }$ while PNPS has a poor performance in this task. Results average across 2 seeds. + +Physics Environment. The physics environment (Ke et al., 2021) simulates a simple physical world. It consists of blocks of unique but unknown weights. The dynamics for the interaction between blocks is that the movement of heavier blocks pushes lighter blocks on their path. This rule creates an acyclic causal graph between the blocks. For an accurate world model, the learner needs to infer the correct weights through demonstrations. Interactions in this environment are sparse and only involve two blocks at a time, therefore we expect NPS to outperform dense architectures like GNNs. This environment is demonstrated in Appendix Fig 11. + +We follow the same setup as Kipf et al. (2019). We use their C-SWM model as baseline. For the proposed model, we only replace the GNN from C-SWM by NPS. GNNs generally share parameters across edges, but in NPS each rule has separate parameters. For a fair comparison to GNN, we use an NPS model with 1 rule. Note that this setting is still different from GNNs as in GNNs at each step every slot is updated by instantiating edges between all pairs of slots, while in NPS an edge is dynamically instantiated between a single pair of slots and only the state of the selected slot (i.e., primary slot) gets updated. + +![](images/6eae7884c2ab3f56bf47fca18f661df64ab49fc50960444321e063fbe05b8b8b.jpg) +Figure 6: Action-Conditioned World Models, with number of future steps to be predicted for the world-model on the horizontal axes. (a) Here we show a comparison between GNNs and the proposed NPS on the physics environment using the $\mathrm { H @ 1 }$ metric (higher is better). (b) Comparison of average $\mathrm { H @ 1 }$ scores across 5 Atari games for the proposed model NPS and GNN. + +The results of our experiments are presented in Figure 6(a). We can see that NPS outperforms GNNs for all rollouts. Multi-step settings are more difficult to model as errors may get compounded over time steps. The sparsity of NPS (only a single slot affected per step) reduces compounding of errors and enhances symmetry-breaking in the assignment of transformations to rules, while in the case of GNNs, since all entities are affected per step, there is a higher possibility of errors getting compounded. We can see that even with a single rule, we significantly outperform GNNs thus proving the effectiveness of dynamically instantiating edges between entities. + +Atari Games. We also test the proposed model in the more complicated setting of Atari. Atari games also have sparse interactions between entities. For instance, in Pong, any interaction involves only 2 entities: (1) paddle and ball or (2) ball and the wall. Therefore, we expect sparse interactions captured by NPS to outperform GNNs here as well. + +We follow the same setup as for the physics environment described in the previous section. We present the results for the Atari experiments in Figure 6(b), showing the average $\mathrm { H @ 1 }$ score across 5 games: Pong, Space Invaders, Freeway, Breakout, and QBert. As expected, we can see that the proposed model achieves a higher score than the GNN-based C-SWM. The results for the Atari experiments reinforce the claim that NPS is especially good at learning sparse interactions. + +Learning Rules for Physical Reasoning. To show the effectiveness of the proposed approach for physical reasoning tasks, we evaluate NPS on another dataset: Sprites-MOT (He et al., 2018). The Sprites-MOT dataset was introduced by He et al. (2018). The dataset contains a set of moving objects of various shapes. This dataset aims to test whether a model can handle occlusions correctly. Each frame has consistent bounding boxes which may cause the objects to appear or disappear from the scene. A model which performs well should be able to track the motion of all objects irrespective of whether they are occluded or not. We follow the same setup as Weis et al. (2020). We use the OP3 model (Veerapaneni et al., 2019) as our baseline. To test the proposed model, we replace the GNN-based transition model in OP3 with the proposed NPS. + +We use the same evaluation protocol as followed by Weis et al. (2020) which is based on the MOT (Multi-object tracking) challenge (Milan et al., 2016). The results on the MOTA and MOTP metrics for this task are presented in Table 5. The results on the other metrics are presented in appendix Table 10. We ask the reader to refer to appendix F.1 for more details about these metrics. We can see that for almost all metrics, NPS outperforms the OP3 baseline. Although this dataset does not contain physical interactions between the objects, sparse rule application should still be useful in dealing with occlusions. At any time step, only a single object is affected by occlusions i.e., it may get occluded due to another object or due to a prespecified bounding box, while the other objects follow their default dynamics. Therefore, a rule should be applied to only the object (or entity) affected (i.e., not visible) due to occlusion and may take into account any other object or entity that is responsible for the occlusion. + +Table 5: Sprites-MOT. Comparison between the proposed NPS and the baseline OP3 using the MOTA and MOTP metrics on the sprites-MOT dataset ( $\uparrow$ : higher is better). Average over 3 random seeds. + +
ModelMOTA↑MOTP↑
OP3NPS89.1±5.178.4±2.4
90.72±5.1579.91±0.3
+ +# 5 Discussion and Conclusion + +For AI agents such as robots trying to make sense of their environment, the only observables are low-level variables like pixels in images. To generalize well, an agent must induce high-level entities as well as discover and disentangle the rules that govern how these entities actually interact with each other. Here we have focused on perceptual inference problems and proposed NPS, a neural instantiation of production systems by introducing an important inductive bias in the architecture following the proposals of Bengio (2017); Goyal & Bengio (2020); Ke et al. (2021). + +Limitations & Looking Forward. Our experiments highlight the advantages brought by the factorization of knowledge into a small set of entities and sparse sequentially applied rules. Immediate future work would investigate how to take advantage of these inductive biases for more complex physical environments (Ahmed et al., 2020) and novel planning methods, which might be more sample efficient than standard ones (Schrittwieser et al., 2020). + +We also find that Sequential and Parallel NPS have different properties suited towards different domains. Future work should explore how to effectively combine these two approaches. We discuss this in more detail in Appendix section E.3. + +# 6 Acknowledgements + +The authors would like to thank Matthew Botvinick for useful discussions. The authors would also like to thank Alex Lamb, Stefan Bauer, Nicolas Chapados, Danilo Rezende and Kelsey Allen for brainstorming sessions. We are also thankful to Dianbo Liu, Damjan Kalajdzievski and Osama Ahmed for proofreading. We would like to thank Samsung Electronics Co. Ltd. and CIFAR for funding this research. We would also like to thank Google for providing Google cloud credits used in this work. + +# References + +Ahmed, O., Träuble, F., Goyal, A., Neitz, A., Wüthrich, M., Bengio, Y., Schölkopf, B., and Bauer, S. Causalworld: A robotic manipulation benchmark for causal structure and transfer learning. arXiv preprint arXiv:2010.04296, 2020. +Anderson, J. R. Skill acquisition: Compilation of weak-method problem situations. Psychological review, 94(2):192, 1987. +Andreas, J., Rohrbach, M., Darrell, T., and Klein, D. Neural module networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 39–48, 2016. +Bahdanau, D., Cho, K., and Bengio, Y. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. +Battaglia, P. W., Pascanu, R., Lai, M., Rezende, D. J., and Kavukcuoglu, K. Interaction networks for learning about objects, relations and physics. CoRR, abs/1612.00222, 2016. URL http: //arxiv.org/abs/1612.00222. +Battaglia, P. W., Hamrick, J. B., Bapst, V., Sanchez-Gonzalez, A., Zambaldi, V., Malinowski, M., Tacchetti, A., Raposo, D., Santoro, A., Faulkner, R., et al. Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261, 2018. +Bengio, Y. The consciousness prior. arXiv preprint arXiv:1709.08568, 2017. +Bottou, L. and Gallinari, P. A framework for the cooperation of learning algorithms. In Advances in neural information processing systems, pp. 781–788, 1991. +Bronstein, M. M., Bruna, J., LeCun, Y., Szlam, A., and Vandergheynst, P. Geometric deep learning: going beyond euclidean data. IEEE Signal Processing Magazine, 34(4):18–42, 2017. +Bunel, R., Hausknecht, M., Devlin, J., Singh, R., and Kohli, P. Leveraging grammar and reinforcement learning for neural program synthesis. arXiv preprint arXiv:1805.04276, 2018. +Burgess, C. P., Matthey, L., Watters, N., Kabra, R., Higgins, I., Botvinick, M., and Lerchner, A. Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390, 2019. +Cai, J., Shin, R., and Song, D. Making neural programming architectures generalize via recursion. arXiv preprint arXiv:1704.06611, 2017. +Chen, R. T., Li, X., Grosse, R., and Duvenaud, D. Isolating sources of disentanglement in variational autoencoders. arXiv preprint arXiv:1802.04942, 2018. +Ding, D., Hill, F., Santoro, A., and Botvinick, M. Object-based attention for spatio-temporal reasoning: Outperforming neuro-symbolic models with flexible distributed architectures. arXiv preprint arXiv:2012.08508, 2020. +Du, Y., Smith, K., Ulman, T., Tenenbaum, J., and Wu, J. Unsupervised discovery of 3d physical objects from video. arXiv preprint arXiv:2007.12348, 2020. +Engelcke, M., Kosiorek, A. R., Jones, O. P., and Posner, I. Genesis: Generative scene inference and sampling with object-centric latent representations. arXiv preprint arXiv:1907.13052, 2019. +Eslami, S., Heess, N., Weber, T., Tassa, Y., Szepesvari, D., Kavukcuoglu, K., and Hinton, G. E. Attend, infer, repeat: Fast scene understanding with generative models. arXiv preprint arXiv:1603.08575, 2016. +Evans, R., Hernández-Orallo, J., Welbl, J., Kohli, P., and Sergot, M. Making sense of sensory input. Artificial Intelligence, pp. 103438, 2019. +Fedus, W., Zoph, B., and Shazeer, N. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. CoRR, abs/2101.03961, 2021. URL https://arxiv.org/abs/ 2101.03961. +Fernando, C., Banarse, D., Blundell, C., Zwols, Y., Ha, D., Rusu, A. A., Pritzel, A., and Wierstra, D. Pathnet: Evolution channels gradient descent in super neural networks. arXiv preprint arXiv:1701.08734, 2017. +Girshick, R. B. Fast R-CNN. CoRR, abs/1504.08083, 2015. URL http://arxiv.org/abs/1504. 08083. +Goyal, A. and Bengio, Y. Inductive biases for deep learning of higher-level cognition. arXiv preprint arXiv:2011.15091, 2020. +Goyal, A., Lamb, A., Hoffmann, J., Sodhani, S., Levine, S., Bengio, Y., and Schölkopf, B. Recurrent independent mechanisms, 2019. +Goyal, A., Lamb, A., Gampa, P., Beaudoin, P., Levine, S., Blundell, C., Bengio, Y., and Mozer, M. Object files and schemata: Factorizing declarative and procedural knowledge in dynamical systems. arXiv preprint arXiv:2006.16225, 2020. +Graves, A., Wayne, G., and Danihelka, I. Neural turing machines. arXiv preprint arXiv:1410.5401, 2014. +Greff, K., Rasmus, A., Berglund, M., Hao, T. H., Schmidhuber, J., and Valpola, H. Tagger: Deep unsupervised perceptual grouping. arXiv preprint arXiv:1606.06724, 2016. +Greff, K., Kaufman, R. L., Kabra, R., Watters, N., Burgess, C., Zoran, D., Matthey, L., Botvinick, M., and Lerchner, A. Multi-object representation learning with iterative variational inference. arXiv preprint arXiv:1903.00450, 2019. +Groth, O., Fuchs, F., Posner, I., and Vedaldi, A. Shapestacks: Learning vision-based physical intuition for generalised object stacking. CoRR, abs/1804.08018, 2018. URL http://arxiv.org/abs/ 1804.08018. +He, Z., Li, J., Liu, D., He, H., and Barber, D. Tracking by animation: Unsupervised learning of multi-object attentive trackers. CoRR, abs/1809.03137, 2018. URL http://arxiv.org/abs/ 1809.03137. +Higgins, I., Matthey, L., Pal, A., Burgess, C., Glorot, X., Botvinick, M., Mohamed, S., and Lerchner, A. beta-vae: Learning basic visual concepts with a constrained variational framework. 2016. +Jacobs, R. A., Jordan, M. I., Nowlan, S. J., Hinton, G. E., et al. Adaptive mixtures of local experts. Neural computation, 3(1):79–87, 1991. +Jang, E., Gu, S., and Poole, B. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. +Ke, N. R., Didolkar, A. R., Mittal, S., Goyal, A., Lajoie, G., Bauer, S., Rezende, D. J., Mozer, M. C., Bengio, Y., and Pal, C. Systematic evaluation of causal discovery in visual model based reinforcement learning. 2021. +Kipf, T., Fetaya, E., Wang, K.-C., Welling, M., and Zemel, R. Neural relational inference for interacting systems. arXiv preprint arXiv:1802.04687, 2018. +Kipf, T., van der Pol, E., and Welling, M. Contrastive learning of structured world models. arXiv preprint arXiv:1911.12247, 2019. +Kirsch, L., Kunze, J., and Barber, D. Modular networks: Learning to decompose neural computation. In Advances in Neural Information Processing Systems, pp. 2408–2418, 2018. +Kosiorek, A., Kim, H., Teh, Y. W., and Posner, I. Sequential attend, infer, repeat: Generative modelling of moving objects. Advances in Neural Information Processing Systems, 31:8606–8616, 2018. +Laird, J. E., Rosenbloom, P. S., and Newell, A. Chunking in soar: The anatomy of a general learning mechanism. Machine learning, 1(1):11–46, 1986. +Lamb, A., Goyal, A., Słowik, A., Mozer, M., Beaudoin, P., and Bengio, Y. Neural function modules with sparse arguments: A dynamic approach to integrating information across layers. arXiv preprint arXiv:2010.08012, 2020. +Le Roux, N., Heess, N., Shotton, J., and Winn, J. Learning a generative model of images by factoring appearance and shape. Neural Computation, 23(3):593–650, 2011. +Li, Y., Gimeno, F., Kohli, P., and Vinyals, O. Strong generalization and efficiency in neural programs. arXiv preprint arXiv:2007.03629, 2020. +Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., and Kipf, T. Object-centric learning with slot attention. arXiv preprint arXiv:2006.15055, 2020a. +Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., and Kipf, T. Object-centric learning with slot attention, 2020b. +Lovett, M. C. and Anderson, J. R. Thinking as a production system. The Cambridge handbook of thinking and reasoning, pp. 401–429, 2005. +McMillan, C., Mozer, M. C., and Smolensky, P. The connectionist scientist game: rule extraction and refinement in a neural network. In Proceedings of the 13th Annual Conference of the Cognitive Science Society, pp. 424–430, 1991. +Milan, A., Leal-Taixé, L., Reid, I. D., Roth, S., and Schindler, K. MOT16: A benchmark for multiobject tracking. CoRR, abs/1603.00831, 2016. URL http://arxiv.org/abs/1603.00831. +Neelakantan, A., Le, Q. V., and Sutskever, I. Neural programmer: Inducing latent programs with gradient descent. arXiv preprint arXiv:1511.04834, 2015. +Qi, H., Wang, X., Pathak, D., Ma, Y., and Malik, J. Learning long-term visual dynamics with region proposal interaction networks. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=_X_4Akcd8Re. +Rahaman, N., Goyal, A., Gondal, M. W., Wuthrich, M., Bauer, S., Sharma, Y., Bengio, Y., and Schölkopf, B. S2rms: Spatially structured recurrent modules. arXiv preprint arXiv:2007.06533, 2020. +Rand, W. M. Objective criteria for the evaluation of clustering methods. Journal of the American Statistical association, 66(336):846–850, 1971. +Raposo, D., Santoro, A., Barrett, D., Pascanu, R., Lillicrap, T., and Battaglia, P. Discovering objects and their relations from entangled scene representations. arXiv preprint arXiv:1702.05068, 2017. +Reed, S. and De Freitas, N. Neural programmer-interpreters. arXiv preprint arXiv:1511.06279, 2015. +Ronco, E., Gollee, H., and Gawthrop, P. J. Modular neural networks and self-decomposition. Technical Report CSC-96012, 1997. +Rosenbaum, C., Klinger, T., and Riemer, M. Routing networks: Adaptive selection of non-linear functions for multi-task learning. arXiv preprint arXiv:1711.01239, 2017. +Rosenbaum, C., Cases, I., Riemer, M., and Klinger, T. Routing networks and the challenges of modular and compositional computation. arXiv preprint arXiv:1904.12774, 2019. +Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M., and Monfardini, G. The graph neural network model. IEEE Transactions on Neural Networks, 20(1):61–80, 2008. +Schrittwieser, J., Antonoglou, I., Hubert, T., Simonyan, K., Sifre, L., Schmitt, S., Guez, A., Lockhart, E., Hassabis, D., Graepel, T., et al. Mastering atari, go, chess and shogi by planning with a learned model. Nature, 588(7839):604–609, 2020. +Shazeer, N., Mirhoseini, A., Maziarz, K., Davis, A., Le, Q., Hinton, G., and Dean, J. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. +Tacchetti, A., Song, H. F., Mediano, P. A., Zambaldi, V., Rabinowitz, N. C., Graepel, T., Botvinick, M., and Battaglia, P. W. Relational forward models for multi-agent learning. arXiv preprint arXiv:1809.11044, 2018. +Touretzky, D. S. and Hinton, G. E. A distributed connectionist production system. Cognitive science, 12(3):423–466, 1988. +Trask, A., Hill, F., Reed, S. E., Rae, J., Dyer, C., and Blunsom, P. Neural arithmetic logic units. In Advances in Neural Information Processing Systems, pp. 8035–8044, 2018. +Van Steenkiste, S., Chang, M., Greff, K., and Schmidhuber, J. Relational neural expectation maximization: Unsupervised discovery of objects and their interactions. arXiv preprint arXiv:1802.10353, 2018. +Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. Attention is all you need, 2017. +Veerapaneni, R., Co-Reyes, J. D., Chang, M., Janner, M., Finn, C., Wu, J., Tenenbaum, J. B., and Levine, S. Entity abstraction in visual model-based reinforcement learning. CoRR, abs/1910.12827, 2019. URL http://arxiv.org/abs/1910.12827. +Veerapaneni, R., Co-Reyes, J. D., Chang, M., Janner, M., Finn, C., Wu, J., Tenenbaum, J., and Levine, S. Entity abstraction in visual model-based reinforcement learning. In Conference on Robot Learning, pp. 1439–1456. PMLR, 2020. +Watters, N., Zoran, D., Weber, T., Battaglia, P., Pascanu, R., and Tacchetti, A. Visual interaction networks: Learning a physics simulator from video. In Advances in neural information processing systems, pp. 4539–4547, 2017. +Weis, M. A., Chitta, K., Sharma, Y., Brendel, W., Bethge, M., Geiger, A., and Ecker, A. S. Unmasking the inductive biases of unsupervised object representations for video sequences, 2020. +Xu, D., Nair, S., Zhu, Y., Gao, J., Garg, A., Fei-Fei, L., and Savarese, S. Neural task programming: Learning to generalize across hierarchical tasks. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 1–8. IEEE, 2018. +Zablotskaia, P., Dominici, E. A., Sigal, L., and Lehrmann, A. M. Unsupervised video decomposition using spatio-temporal iterative inference. arXiv preprint arXiv:2006.14727, 2020. + +# Checklist + +1. For all authors... + +(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] +(b) Did you describe the limitations of your work? [Yes] +(c) Did you discuss any potential negative societal impacts of your work? [Yes] +(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] + +2. If you are including theoretical results... + +(a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A] + +3. If you ran experiments... + +(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] +(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] +(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] See section 4.3 +(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] + +4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... + +(a) If your work uses existing assets, did you cite the creators? [Yes] +(b) Did you mention the license of the assets? [No] All datasets we used are open-sourced +(c) Did you include any new assets either in the supplemental material or as a URL? [Yes] We include the code for reproducing our experiments +(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A] All datasets we use are synthetic and created using simulation hence do not include any personal or user data. +(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A] We do not use any dataset that includes any kind of identifying information. + +5. If you used crowdsourcing or conducted research with human subjects... + +(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A] +(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A] +(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A] \ No newline at end of file diff --git a/parse/train/xQGYquca0gB/xQGYquca0gB_content_list.json b/parse/train/xQGYquca0gB/xQGYquca0gB_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8b5f88ee9bff981d44079f75a76c858a24299a26 --- /dev/null +++ b/parse/train/xQGYquca0gB/xQGYquca0gB_content_list.json @@ -0,0 +1,1346 @@ +[ + { + "type": "text", + "text": "Neural Production Systems ", + "text_level": 1, + "bbox": [ + 333, + 122, + 665, + 147 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aniket Didolkar\\*, 1, Anirudh Goyal \\*, 1, Nan Rosemary $\\mathbf { K } \\mathbf { e } ^ { 2 }$ , Charles Blundell 2, Philippe Beaudoin 3 Nicolas Heess 2, Michael Mozer 4, \\*\\*, Yoshua Bengio 1, \\*\\* ", + "bbox": [ + 223, + 199, + 777, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract ", + "text_level": 1, + "bbox": [ + 462, + 265, + 535, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual environments are structured, consisting of distinct objects or entities. These entities have properties—visible or latent—that determine the manner in which they interact with one another. To partition images into entities, deep-learning researchers have proposed structural inductive biases such as slot-based architectures. To model interactions among entities, equivariant graph neural nets (GNNs) are used, but these are not particularly well suited to the task for two reasons. First, GNNs do not predispose interactions to be sparse, as relationships among independent entities are likely to be. Second, GNNs do not factorize knowledge about interactions in an entity-conditional manner. As an alternative, we take inspiration from cognitive science and resurrect a classic approach, production systems, which consist of a set of rule templates that are applied by binding placeholder variables in the rules to specific entities. Rules are scored on their match to entities, and the best fitting rules are applied to update entity properties. In a series of experiments, we demonstrate that this architecture achieves a flexible, dynamic flow of control and serves to factorize entity-specific and rule-based information. This disentangling of knowledge achieves robust future-state prediction in rich visual environments, outperforming state-of-the-art methods using GNNs, and allows for the extrapolation from simple (few object) environments to more complex environments. ", + "bbox": [ + 233, + 297, + 766, + 546 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction ", + "text_level": 1, + "bbox": [ + 174, + 574, + 310, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite never having taken a physics course, every child beyond a young age appreciates that pushing a plate off the dining table will cause the plate to break. The laws of physics accurately characterize the dynamics of our natural world, and although explicit knowledge of these laws is not necessary to reason, we can reason explicitly about objects interacting through these laws. Humans can verbalize knowledge in propositional expressions such as “If a plate drops from table height, it will break,” and “If a video-game opponent approaches from behind and they are carrying a weapon, they are likely to attack you.” Expressing propositional knowledge is not a strength of current deep learning methods for several reasons. First, propositions are discrete and independent from one another. Second, propositions must be quantified in the manner of first-order logic; for example, the video-game proposition applies to any $X$ for which $X$ is an opponent and has a weapon. Incorporating the ability to express and reason about propositions should improve generalization in deep learning methods because this knowledge is modular— propositions can be formulated independently of each other— and can therefore be acquired incrementally. Propositions can also be composed with each other and applied consistently to all entities that match, yielding a powerful form of systematic generalization. ", + "bbox": [ + 174, + 606, + 825, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The classical AI literature from the 1980s can offer deep learning researchers a valuable perspective. In this era, reasoning, planning, and prediction were handled by architectures that performed proposi tional inference on symbolic knowledge representations. A simple example of such an architecture is the production system (Laird et al., 1986; Anderson, 1987), which expresses knowledge by conditionaction rules. The rules operate on a working memory: rule conditions are matched to entities in working memory inspired by cognitive science, and such a match can trigger computational actions that update working memory or external actions that operate on the outside world. ", + "bbox": [ + 176, + 806, + 821, + 848 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/efc8469bfcada4d38e5ed73c14c9b9bca7fe99729b314148655c694d81e11332.jpg", + "image_caption": [ + "Figure 1: In this figure we show a visual comparison between NPS and dense architectures like GNNs. In NPS, a rule is only applied when an interaction takes place and it is applied only to the slots affected by the interaction. NPS also uses different rules for different kinds of interactions, while in GNN a common rule is applied to all slots irrespective of whether an interaction takes place or not (because of parameter sharing). Note the dynamic nature of the interaction graph in NPS, while in GNN, the graph is static. " + ], + "image_footnote": [], + "bbox": [ + 174, + 85, + 823, + 188 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 285, + 825, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Production systems were typically used to model high-level cognition, e.g., mathematical problem solving or procedure following; perception was not the focus of these models. It was assumed that the results of perception were placed into working memory in a symbolic form that could be operated on with the rules. In this article, we revisit production systems but from a deep learning perspective which naturally integrates perceptual processing and subsequent inference for visual reasoning problems. We describe an end-to-end deep learning model that constructs object-centric representations of entities in videos, and then operates on these entities with differentiable—and thus learnable—production rules. The essence of these rules, carried over from traditional symbolic system, is that they operate on variables that are bound, or linked, to the entities in the world. In the deep learning implementation, each production rule is represented by a distinct MLP with query-key attention mechanisms to specify the rule-entity binding and to determine when the rule should be triggered for a given entity. We are not the first to propose a neural instantiation of a production system architecture. Touretzky & Hinton (1988) gave a proof of principle that neural net hardware could be hardwired to implement a production system for symbolic reasoning; our work fundamentally differs from theirs in that (1) we focus on perceptual inference problems and (2) we use the architecture as an inductive bias for learning. ", + "bbox": [ + 174, + 347, + 825, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.1 Variables and entities ", + "text_level": 1, + "bbox": [ + 176, + 592, + 362, + 607 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "What makes a rule general-purpose is that it incorporates placeholder variables that can be bound to arbitrary values or—the term we prefer in this article—entities. This notion of binding is familiar in functional programming languages, where these variables are called arguments. Analogously, the use of variables in the production rules we describe enable a model to reason about any set of entities that satisfy the selection criteria of the rule. ", + "bbox": [ + 174, + 621, + 825, + 690 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Consider a simple function in C like int add(int a, int b). This function binds its two integer operands to variables $a$ and $b$ . The function does not apply if the operands are, say, character strings. The use of variables enables a programmer to reuse the same function to add any two integer values ", + "bbox": [ + 174, + 696, + 825, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order for rules to operate on entities, these entities must be represented explicitly. That is, the visual world needs to be parsed in a task-relevant manner, e.g., distinguishing the sprites in a video game or the vehicles and pedestrians approaching an autonomous vehicle. Only in the past few years have deep learning vision researchers developed methods for object-centric representation (Le Roux et al., 2011; Eslami et al., 2016; Greff et al., 2016; Raposo et al., 2017; Van Steenkiste et al., 2018; Kosiorek et al., 2018; Engelcke et al., 2019; Burgess et al., 2019; Greff et al., 2019; Locatello et al., 2020a; Ahmed et al., 2020; Goyal et al., 2019; Zablotskaia et al., 2020; Rahaman et al., 2020; Du et al., 2020; Ding et al., 2020; Goyal et al., 2020; Ke et al., 2021). These methods differ in details but share the notion of a fixed number of slots (see Figure 1 for example), also known as object files, each encapsulating information about a single object. Importantly, the slots are interchangeable, meaning that it doesn’t matter if a scene with an apple and an orange encodes the apple in slot 1 and orange in slot 2 or vice-versa. ", + "bbox": [ + 174, + 746, + 825, + 911 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A model of visual reasoning must not only be able to represent entities but must also express knowledge about entity dynamics and interactions. To ensure systematic predictions, a model must be capable of applying knowledge to an entity regardless of the slot it is in and must be capable of applying the same knowledge to multiple instances of an entity. Several distinct approaches exist in the literature. The predominant approach uses graph neural networks to model slot-to-slot interactions (Scarselli et al., 2008; Bronstein et al., 2017; Watters et al., 2017; Van Steenkiste et al., 2018; Kipf et al., 2018; Battaglia et al., 2018; Tacchetti et al., 2018). To ensure systematicity, the GNN must share parameters among the edges. In a recent article, Goyal et al. (2020) developed a more general framework in which parameters are shared but slots can dynamically select which parameters to use in a state-dependent manner. Each set of parameters is referred to as a schema, and slots use a query-key attention mechanism to select which schema to apply at each time step. Multiple slots can select the same schema. In both GNNs and SCOFF, modeling dynamics involves each slot interacting with each other slot. In the work we describe in this article, we replace the direct slot-to-slot interactions with rules, which mediate sparse interactions among slots (See arrows in Figure 1). ", + "bbox": [ + 173, + 90, + 825, + 285 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thus our main contribution is that we introduce NPS, which offers a way to model dynamic and sparse interactions among the variables in a graph and also allows dynamic sharing of multiple sets of parameters among these interactions. Most architectures used for modelling interactions in the current literature use statically instantiated graph which model all possible interactions for a given variable at each step i.e. dense interactions. Also such dense architectures share a single set of parameters across all interactions which maybe quite restrictive in terms of representational capacity. A visual comparison between these two kinds of architectures is shown in Figure 1. Through our experiments we show the advantage of modeling interactions in the proposed manner using NPS in visually rich physical environments. We also show that our method results in an intuitive factorization of rules and entities. ", + "bbox": [ + 174, + 290, + 825, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Production System ", + "text_level": 1, + "bbox": [ + 176, + 446, + 362, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, our notion of a production system consists of a set of entities and a set of rules, along with a mechanism for selecting rules to apply on subsets of the entities. Implicit in a rule is a specification of the properties of relevant entities, e.g., a rule might apply to one type of sprite in a video game but not another. The control flow of a production system dynamically selects rules as well as bindings between rules and entities, allowing different rules to be chosen and different entities to be manipulated at each point in time. ", + "bbox": [ + 174, + 483, + 825, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The neural production system we describe shares essential properties with traditional production system, particularly with regard to the compositionality and generality of the knowledge they embody. Lovett & Anderson (2005) describe four desirable properties commonly attributed to symbolic systems that apply to our work as well. ", + "bbox": [ + 174, + 573, + 825, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Production rules are modular. Each production rule represents a unit of knowledge and are atomic such that any production rule can be intervened (added, modified or deleted) independently of other production rules in the system. ", + "bbox": [ + 174, + 635, + 825, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Production rules are abstract. Production rules allow for generalization because their conditions may be represented as high-level abstract knowledge that match to a wide range of patterns. These conditions specify the attributes of relationship(s) between entities without specifying the entities themselves. The ability to represent abstract knowledge allows for the transfer of learning across different environments as long as they fit within the conditions of the given production rule. ", + "bbox": [ + 174, + 683, + 823, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Production rules are sparse. In order that production rules have broad applicability, they involve only a subset of entities. This assumption imposes a strong prior that dependencies among entities are sparse. In the context of visual reasoning, we conjecture that this prior is superior to what has often been assumed in the past, particularly in the disentanglement literature—independence among entities Higgins et al. (2016); Chen et al. (2018). ", + "bbox": [ + 174, + 758, + 825, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Production rules represent causal knowledge and are thus asymmetric. Each rule can be decomposed into a {condition, action} pair, where the action reflects a state change that is a causal consequence of the conditions being met. ", + "bbox": [ + 176, + 835, + 825, + 877 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These four properties are sufficient conditions for knowledge to be expressed in production rule form. These properties specify how knowledge is represented, but not what knowledge is represented. The latter is inferred by learning mechanisms under the inductive bias provided by the form of production rules. ", + "bbox": [ + 174, + 882, + 821, + 911 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 90, + 823, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Neural Production System: Slots and Sparse Rules ", + "text_level": 1, + "bbox": [ + 173, + 132, + 629, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Neural Production System (NPS), illustrated in Figure 2, provides an architectural backbone that supports the detection and inference of entity (object) representations in an input sequence, and the underlying rules which govern the interactions between these entities in time and space. The input sequence indexed by time step $t$ , $\\{ \\pmb { x } ^ { 1 } , \\dotsc , \\pmb { x } ^ { t } , \\dotsc , \\pmb { x } ^ { T } \\}$ , for instance the frames in a video, are processed by a neural encoder (Burgess et al., 2019; Greff et al., 2019; Goyal et al., 2019, 2020) applied to each $\\mathbf { \\boldsymbol { x } } ^ { t }$ , to obtain a set of $M$ entity representations $\\{ V _ { 1 } ^ { t } , \\ldots , \\ldots \\stackrel { \\cdot } { , } V _ { M } ^ { t } \\}$ , one for each of the $M$ slots. These representations describe an entity and are updated based on both the previous state, $V ^ { t - 1 }$ and the current input, $\\mathbf { \\boldsymbol { x } } ^ { t }$ . ", + "bbox": [ + 173, + 154, + 825, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NPS consists of $N$ separately encoded rules, $\\{ R _ { 1 } , R _ { 2 } , . . , R _ { N } \\}$ . Each rule consists of two components, ${ \\pmb R } _ { i } = ( \\vec { R _ { i } } , M L P _ { i } )$ , where $\\vec { R _ { i } }$ is a learned rule embedding vector, which can be thought of as a template defining the condition for when a rule applies; and $M L P _ { i }$ , which determines the action taken by a rule. Both $\\vec { R _ { i } }$ and the parameters of $M L P _ { i }$ are learned along with the other parameters of the model using back-propagation on an objective optimized end-to-end. ", + "bbox": [ + 173, + 270, + 826, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the general form of the model, each slot selects a rule that will be applied to it to change its state. This can potentially be performed several times, with possibly different rules applied at each step. Rule selection is done using an attention mechanism described in detail below. Each rule specifies conditions and actions on a pair of slots. Therefore, while modifying the state of a slot using a rule, it can take the state of another slot into account. The slot which is being modified is called the primary slot and other is called the contextual slot. The contextual slot is also selected using an attention mechanism described in detail below. ", + "bbox": [ + 174, + 352, + 485, + 545 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7621aad51e5596db4e7dfe45bfb3fb1b7959f7deac83d4c49f5f51470e8d77b1.jpg", + "image_caption": [ + "Figure 2: Rule and slot combinatorics. Condition-action rules specify how entities interact. Slots maintain the time-varying state of an entity. Every rule is matched to every pair of slots. Through key-value attention, a goodness of match is determined, and a rule is selected along with its binding to slots. " + ], + "image_footnote": [], + "bbox": [ + 578, + 348, + 743, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Computational Steps in NPS ", + "text_level": 1, + "bbox": [ + 174, + 559, + 413, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we give a detailed description of the rule selection and application procedure for the slots. First, we will formalize the definitions of a few terms that we will use to explain our method. We use the term primary slot to refer to slot $V _ { p }$ whose state gets modified by a rule $\\scriptstyle { \\mathbf { } } R _ { r }$ . We use the term contextual slot to refer to the slot $V _ { c }$ that the rule $\\scriptstyle { R _ { r } }$ takes into account while modifying the state of the primary slot $V _ { p }$ . ", + "bbox": [ + 173, + 577, + 825, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notation. We consider a set of $N$ rules $\\{ R _ { 1 } , R _ { 2 } , \\ldots , R _ { N } \\}$ and a set of $_ { \\mathbf { T } }$ input frames $\\{ \\pmb { x } ^ { 1 } , \\pmb { x } ^ { 2 } , \\dots , \\pmb { x } ^ { T } \\}$ . Each frame $\\mathbf { \\boldsymbol { x } } ^ { t }$ is encoded into a set of $M$ slots $\\{ V _ { 1 } ^ { t } , V _ { 2 } ^ { t } , \\ldots , \\bar { V _ { M } ^ { t } } \\}$ . In the following discussion, we omit the index over $\\pmb { t }$ for simplicity. ", + "bbox": [ + 174, + 651, + 821, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step $\\imath .$ is external to NPS and involves parsing an input image, $\\mathbf { \\boldsymbol { x } } ^ { t }$ , into slot-based entities conditioned on the previous state of the slot-based entities. Any of the methods proposed in the literature to obtain a slot-wise representation of entities can be used (Burgess et al., 2019; Greff et al., 2019; Goyal et al., 2019, 2020). The next three steps constitute the rule selection and application procedure. ", + "bbox": [ + 173, + 700, + 825, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 2. For each primary slot $V _ { p }$ , we attend to a rule $\\scriptstyle { R _ { r } }$ to be applied. Here, the queries come from the primary slot: $\\begin{array} { r } { \\pmb q _ { p } = \\pmb V _ { p } W ^ { q } } \\end{array}$ , and the keys come from the rules: $\\pmb { k } _ { i } = \\vec { R _ { i } } W ^ { k } \\quad \\forall i \\in \\{ 1 , \\dots , N \\}$ The rule is selected using a straight-through Gumbel softmax (Jang et al., 2016) to achieve a learnable hard decision: $r = \\mathrm { a r g m a x } _ { i } ( q _ { p } k _ { i } + \\gamma )$ , where $\\gamma \\sim \\mathrm { G u m b e l } ( 0 , 1 )$ . This competition is a noisy version of rule matching and prioritization in traditional production systems. ", + "bbox": [ + 173, + 762, + 825, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 3. For a given primary slot $V _ { p }$ and selected rule $\\scriptstyle { R _ { r } }$ , a contextual slot $V _ { c }$ is selected using another attention mechanism. In this case the query comes from the primary slot: $q _ { p } = V _ { p } W ^ { q }$ , and the keys from all the slots: $\\pmb { k } _ { j } = V _ { j } W ^ { q } \\quad \\forall j \\in \\{ 1 , \\dots , M \\}$ . The selection takes place using a straightthrough Gumbel softmax similar to step 2: $c = \\mathrm { a r g m a x } _ { j } ( q _ { p } k _ { j } + \\gamma )$ , where $\\gamma \\sim \\mathrm { G u m b e l } ( 0 , 1 )$ Note that each rule application is sparse since it takes into account only 1 contextual slot for modifying a primary slot, while other methods like GNNs take into account all slots for modifying a primary slot. ", + "bbox": [ + 173, + 842, + 825, + 911 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 90, + 823, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 4. Rule Application: the selected rule $\\scriptstyle { R _ { r } }$ is applied to the primary slot $V _ { p }$ based on the rule and the current contents of the primary and contextual slots. The rule-specific $M L P _ { r }$ , takes as input the concatenated representation of the state of the primary and contextual slots, $V _ { p }$ and $V _ { c }$ , and produces an output, which is then used to change the state of the primary slot $V _ { p }$ by residual addition. ", + "bbox": [ + 173, + 126, + 825, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Rule Application: Sequential vs Parallel Rule Application ", + "text_level": 1, + "bbox": [ + 173, + 189, + 614, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the previous section, we have described how each rule application only considers another contextual slot for the given primary slot i.e., contextual sparsity. We can also consider application sparsity, wherein we use the rules to update the states of only a subset of the slots. In this scenario, only the selected slots would be primary slots. This setting will be helpful when there is an entity in an environment that is stationary, or it is following its own default dynamics unaffected by other entities. Therefore, it does not need to consider other entities to update its state. We explore two scenarios for enabling application sparsity. ", + "bbox": [ + 173, + 205, + 826, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Parallel Rule Application. Each of the $M$ slots selects a rule to potentially change its state. To enable sparse changes, we provide an extra Null Rule in addition to the available $N$ rules. If a slot picks the null rule in step 2 of the above procedure, we do not update its state. ", + "bbox": [ + 173, + 308, + 485, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sequential Rule Application. In this setting, only one slot gets updated in each rule application step. Therefore, only one slot is selected as the primary slot. This can be facilitated by modifying step 2 above to select one {primary ", + "bbox": [ + 173, + 398, + 485, + 468 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d2db80f1c75eadd39bbfb8c94bc9f20aca13ed711f1d46b4857e95be0ac7103f.jpg", + "image_caption": [ + "Figure 3: This figure demonstrates the sequential and parallel rule application. " + ], + "image_footnote": [], + "bbox": [ + 506, + 311, + 813, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "slot, rule} pair among $_ { N M }$ {rule, slot} pairs. The queries come from each slot: $\\pmb { q } _ { j } = \\pmb { V } _ { j } \\pmb { W } ^ { q } \\quad \\forall j \\in$ $\\{ 1 , \\ldots , M \\}$ , the keys come from the rules: $\\pmb { k } _ { i } = R _ { i } W ^ { k } \\quad \\forall i \\in \\{ 1 , \\dots , N \\}$ . The straight-through Gumbel softmax selects one (primary slot, rule) pair: $p , r = \\mathrm { a r g m a x } _ { i , j } ( q _ { p } k _ { i } + \\gamma )$ , where $\\gamma \\sim$ $\\mathrm { G u m b e l } ( 0 , 1 )$ . In the sequential regime, we allow the rule application procedure (step 2, 3, 4 above) to be performed multiple times iteratively in $K$ rule application stages for each time-step $t$ . ", + "bbox": [ + 174, + 468, + 825, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A pictorial demonstration of both rule application regimes can be found in Figure 3. We provide detailed algorithms for the sequential and parallel regimes in Appendix. ", + "bbox": [ + 173, + 546, + 823, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Experiments ", + "text_level": 1, + "bbox": [ + 174, + 585, + 312, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of NPS on multiple tasks and compare to a comprehensive set of baselines. To show that NPS can learn intuitive rules from the data generating distribution, we design a couple of simple toy experiments with well-defined discrete operations. Results show that NPS can accurately recover each operation defined by the data and learn to represent each operation using a separate rule. We then move to a much more complicated and visually rich setting with abstract physical rules and show that factorization of knowledge into rules as offered by NPS does scale up to such settings. We study and compare the parallel and sequential rule application procedures and try to understand the settings which favour each. We then ", + "bbox": [ + 174, + 614, + 552, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: This table shows the segregation of rules for the MNIST Transformation task. Each cell indicates the number of times the corresponding rule is used for the given operation. We can see that NPS automatically and perfectly learns a separate rule for each operation. ", + "bbox": [ + 566, + 617, + 825, + 713 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ee74d8c57f03621d2b258b5ec0498a5041292fee11a2f8e0e03478c8a8528093.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Rule 1Rule 2Rule3Rule 4
TranslateDownTranslate UpRotate RightRotate Left5039000
00495000
050300
0004981
", + "bbox": [ + 584, + 723, + 802, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "evaluate the benefits of reusable, dynamic and sparse interactions as offered by NPS in a wide variety of physical environments by comparing it against various baselines. We conduct ablation studies to assess the contribution of different components of NPS. Here we briefly outline the tasks considered and direct the reader to the Appendix for full details on each task and details on hyperparameter settings. ", + "bbox": [ + 174, + 794, + 825, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion of baselines. NPS is an interaction network, therefore we use other widely used interaction networks such as multihead attention and graph neural networks (Goyal et al. (2019), Goyal et al. (2020), Veerapaneni et al. (2019), Kipf et al. (2019)) for comparison. Goyal et al. (2019) and Goyal et al. (2020) use an attention based interaction network to capture interactions between the slots, while Veerapaneni et al. (2019) and Kipf et al. (2019) use a GNN based interaction network. We also consider the recently introduced convolutional interaction network (CIN) (Qi et al., 2021) which captures dense pairwise interactions like GNN but uses a convolutional network instead of MLPs to better utilize spatial information. The proposed method, similar to other interaction networks, is agnostic to the encoder backbone used to encode the input image into slots, therefore we compare NPS to other interaction networks across a wide-variety of encoder backbones. ", + "bbox": [ + 174, + 869, + 825, + 911 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 92, + 825, + 189 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Learning intuitive rules with NPS: Toy Simulations ", + "text_level": 1, + "bbox": [ + 173, + 202, + 568, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We designed a couple of simple tasks with well-defined discrete rules to show that NPS can learn intuitive and interpretable rules. We also show the efficiency and effectiveness of the selection procedure (step 2 and step 3 in section 3.1) by comparing against a baseline with many more parameters. Both tasks require a single modification of only one of the available entities, therefore the use of sequential or parallel rule application would not make a difference here since parallel rule application in which all-but-one slots select the null rule is similar to sequential rule application with 1 rule application step. To simplify the presentation, we describe the setup for both tasks using the sequential rule application procedure. ", + "bbox": [ + 173, + 228, + 825, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MNIST Transformation. We test whether NPS can learn simple rules for performing transformations on MNIST digits. We generate data with four transformations: {Translate Up, Translate Down, Rotate Right, Rotate Left}. We feed the input image $( X )$ and the transformation $( o )$ to be performed as a one-hot vector to the model. The detailed setup is described in Appendix. For this task, we evaluate whether NPS can learn to use a unique rule for each transformation. ", + "bbox": [ + 174, + 345, + 825, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use 4 rules corresponding to the 4 transformations with the hope that the correct transformations are recovered. Indeed, we observe that NPS successfully learns to represent each transformation using a separate rule as shown in Table 1. Our model achieves an MSE of 0.02. A visualization of the outputs from our model and further details can be found in Appendix C. ", + "bbox": [ + 173, + 421, + 826, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Coordinate Arithmetic Task. The model is tasked with performing arithmetic operations on 2D coordinates. Given $( X _ { 0 } , Y _ { 0 } )$ and $( X _ { 1 } , Y _ { 1 } )$ , we can apply the following operations: $\\{ \\mathbf { X } \\mathbf { A d } \\mathbf { - }$ dition: $\\bar { ( X _ { r } , Y _ { r } ) } \\ = \\ ( X _ { 0 } \\bar { + } X _ { 1 } , Y _ { 0 } )$ , $\\mathbf { X }$ Subtraction: $( X _ { r } , Y _ { r } ) \\ : = \\ : ( X _ { 0 } - X _ { 1 } , Y _ { 0 } )$ , Y Addition: $( X _ { r } , Y _ { r } ) = ( X _ { 0 } , Y _ { 0 } + Y _ { 1 } )$ , Y Subtraction: $X _ { r } , Y _ { r } = ( X _ { 0 } , Y _ { 0 } - Y _ { 1 } ) \\}$ , where $( X _ { r } , Y _ { r } )$ is the resultant coordinate. ", + "bbox": [ + 174, + 497, + 486, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this task, the model is given 2 input coordinates $X = [ ( x _ { i } , y _ { i } ) , ( x _ { j } , \\bar { y _ { j } } ) ]$ and the expected output coordinates $Y = [ ( \\hat { x _ { i } } , \\hat { y _ { i } } ) , ( \\hat { x _ { j } } , \\hat { y _ { j } } ) ] .$ . The model is supposed to infer the correct rule to produce the correct output coordinates. During data collection, the true output is obtained by performing a random transformation on a randomly selected coordinate in $X$ (primary coordinate), taking another randomly selected coordinate from $X$ (contextual coordinate) into account. The detailed setup is described in Appendix D. We use an NPS model with 4 rules for this task. We use the the selection procedure in step 2 and step 3 of algorithm 1 to select the primary coordinate, contextual coordinate, and the rule. For the baseline we replace the selection procedure in NPS (i.e. step 2 and step 3 in algorithm 1) with a routing MLP similar to Fedus et al. (2021). ", + "bbox": [ + 174, + 628, + 485, + 876 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/93fa4fe136835a329ff05316b00960119a6e62b649e79c974d4d8c5f78adba3a.jpg", + "image_caption": [ + "Figure 4: Coordinate Arithmetic Task. Here, we compare NPS to the baseline model in terms of segregation of rules as the training progresses. X-axis shows the epochs and Y-axis shows the frequency with which Rule $i$ is used for the given operation. We can see that NPS disentangles the operations perfectly as training progresses with a unique rule specializing to every operation while the baseline model fails to do so. " + ], + "image_footnote": [], + "bbox": [ + 503, + 484, + 816, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 863, + 588, + 876 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This routing MLP has 3 heads (one each for selecting the primary coordinate, contextual coordinate, and the rule). The baseline has 4 times more parameters than NPS. The final output is produced by the rule MLP which does not have access to the true output, hence the model cannot simply copy the true output to produce the actual output. Unlike the MNIST transformation task, we do not provide the operation to be performed as a one-hot vector input to the model, therefore it needs to infer the available operations from the data demonstrations. ", + "bbox": [ + 173, + 883, + 825, + 911 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 92, + 825, + 147 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show the segregation of rules for NPS and the baseline in Figure 4. We can see that NPS learns to use a unique rule for each operation while the baseline struggles to disentangle the underlying operations properly. NPS also outperforms the baseline in terms of MSE achieving an MSE of $0 . 0 1 _ { \\pm 0 . 0 0 1 }$ while the baseline achieves an MSE of $0 . 0 4 _ { \\pm 0 . 0 0 8 }$ . To further confirm that NPS learns all the available operations correctly from raw data demonstrations, we use an NPS model with 5 rules. We expect that in this case NPS should utilize only 4 rules since the data describes only 4 unique operations and indeed we observe that NPS ends up mostly utilizing 4 of the available 5 rules as shown in Table 2. ", + "bbox": [ + 174, + 154, + 553, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: This table shows segregation of rules when we use NPS with 5 rules but the data generation distributions describes only 4 possible operations. We can see that only 4 rules get majorly utilized thus confirming that NPS successfully recovers all possible operations described by the data. ", + "bbox": [ + 566, + 152, + 825, + 263 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a8f84dde9bca14de9d9314267a57d24b53ba78f2f3b4666f783cd97b86554942.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Rule 1Rule 2Rule 3Rule 4Rule 5
X Addition3609945130
X Subtraction0482010
Y Subtraction03945320
Y Addition0571599335
", + "bbox": [ + 570, + 272, + 816, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Parallel vs Sequential Rule Application ", + "text_level": 1, + "bbox": [ + 176, + 351, + 485, + 366 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the parallel and sequential rule application procedures, to understand the settings that favour one or the other, over two tasks: (1) Bouncing Balls, (2) Shapes Stack. We use the term PNPS to refer to parallel rule application and SNPS to refer to sequential rule application. ", + "bbox": [ + 176, + 377, + 821, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Shapes Stack. We use the shapes stack dataset introduced by Groth et al. (2018). This dataset consists of objects stacked on top of each other as shown in Figure 5. These objects fall under the influence of gravity. For our experiments, We follow the same setup as Qi et al. (2021). In this task, given the first frame, the model is tasked with predicting the object bounding boxes for the next $\\pmb { t }$ timesteps. The first frame is encoded using a convolutional network followed by RoIPooling (Girshick (2015)) to extract object-centric visual features. The object-centric features are then passed to the dynamics model to predict object bounding boxes of the next $\\pmb { t }$ steps. Qi et al. (2021) propose a Region Proposal Interaction Network (RPIN) to solve this task. The dynamics model in RPIN consists of an Interaction Network proposed in Battaglia et al. (2016). ", + "bbox": [ + 174, + 430, + 553, + 635 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6667bcdf1c3ef24b7d961678d2db241d0c4a6cfee87cc246cf897591d179148d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model NameTestTransfer
RPIN (Qi et al. (2021))1.254±0.0086.377±0.325
PNPSSNPS1.250±0.0075.411±0.45
1.68±0.025.80±0.15
", + "bbox": [ + 571, + 431, + 816, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: Prediction error of the compared models on the shapes stack dataset (lower is better) for the test as well as transfer setting. In the test setting the number of rollout steps $\\pmb { t }$ is set to 15 and in the transfer setting it is set to 30. We can see that PNPS outperforms the RPIN baseline in both the test and transfer setting while SNPS fails to do so. Results across 15 seeds. ", + "bbox": [ + 566, + 497, + 825, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To better utilize spatial information, Qi et al. (2021) propose an extension of the interaction operators in interaction net to operate on 3D tensors. This is achieved by replacing the MLP operations in the original interaction networks with convolutions. They call this new network Convolutional Interaction Network (CIN). For the proposed model, we replace this CIN in RPIN by NPS. To ensure a fair comparison to CIN, we use CNNs to represent rules in NPS instead of MLPs. CIN captures all pairwise interactions between objects using a convolutional network. In NPS, we capture sparse interactions (contextual sparsity) as compared to dense pairwise interactions captured by CIN. Also, in NPS we update only a few subset of slots per step instead of all slots (application sparsity). ", + "bbox": [ + 174, + 636, + 825, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We consider two evaluation settings. (1) Test setting: The number of rollout timesteps is same as that seen during training (i.e. $t = 1 5$ ); (2) Transfer Setting: The number of rollout timesteps is higher than that seen during training (i.e. $t = 3 0$ ). ", + "bbox": [ + 176, + 752, + 820, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present our results on the shapes stack dataset in Table 3. We can see that both PNPS and SNPS outperform the baseline RPIN in the transfer setting, while only PNPS outperforms the baseline in the test setting and SNPS fails to do so. We can see that PNPS outperforms SNPS. We attribute this to the reduced application sparsity with PNPS, i.e., it is more likely that the state of a slot gets updated in PNPS as compared to SNPS. For instance, consider an NPS model with $N$ uniformly chosen rules and $M$ slots. The probability that the state of a slot gets updated in PNPS is $P _ { P N P S } \\dot { = } N - 1 / N$ (since 1 rule is the null rule), while the same probability for SNPS is $P _ { S N P S } = 1 / M$ (since only 1 slot gets updated per rule application step). ", + "bbox": [ + 173, + 800, + 825, + 911 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For this task, we run both PNPS and SNPS for $N = \\{ 1 , 2 , 4 , 6 \\}$ rules and $M = 3$ . For any given $_ N$ , we observe that $P _ { P N P S } > P _ { S N P S }$ . Even when we have multiple rule application steps in SNPS, it might end up selecting the same slot to be updated in more than one of these steps. We report the best performance obtained for PNPS and SNPS across all $N$ , which is $N = \\{ 2 + 1$ Null Rule} for PNPS and $N = 4$ for SNPS, in Table 3. Shapes stack is a dataset that would prefer a model with less application sparsity since all the objects are tightly bound to each other (objects are placed on top of each other), therefore all objects spend the majority of their time interacting with the objects directly above or below them. We attribute the higher performance of PNPS compared to RPIN to the higher contextual sparsity of PNPS. Each example in the shapes stack task consists of 3 objects. Even though the blocks are tightly bound to each other, each block is only affected by the objects it is in direct contact with. For example, the top-most object is only affected by the object directly below it. The contextual sparsity offered by PNPS is a strong inductive bias to model such sparse interactions while RPIN models all pairwise interactions between the objects. Figure 5 shows an intuitive illustration of the PNPS model for the shapes stack dataset. In the figure, Rule 2 actually refers to the Null Rule, while Rule 1 refers to all the other non-null rules. The bottom-most block picks the Null Rule most times, as the bottom-most block generally does not move. ", + "bbox": [ + 174, + 90, + 825, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Bouncing Balls. We consider a bouncingballs environment in which multiple balls move with billiard-ball dynamics. We validate our model on a colored version of this dataset. This is a next-step prediction task in which the model is tasked with predicting the final binary mask of each ball. We compare the following methods: (a) SCOFF (Goyal et al., 2020): factorization of knowledge in terms of slots (object properties) and schemata, the latter capturing object dynamics; (b) ${ \\mathrm { S C O F F } } { + } { + }$ : we extend SCOFF by using the idea of iterative competition as proposed in slot attention (SA) (Locatello et al., 2020a); SCOFF $^ +$ PNPS/SNPS: We replace pairwise slot-to-slot interaction in ${ \\mathrm { S C O F F } } { + } { + }$ with parallel or sequential rule application. For comparing different methods, we use the Adjusted Rand Index or ARI (Rand, 1971). To investigate how the factorization in the form of rules allows for extrapolating knowledge from fewer to more objects, we increase the number of objects from 4 during training to 6-8 during testing. ", + "bbox": [ + 174, + 332, + 485, + 636 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0bafd38ea5e1df32a31ff88f73b7b52d92be55ef3f1eb31c0228c553cd29402e.jpg", + "image_caption": [ + "Figure 5: Here we show the rule selection statistics from the proposed model for all entities in the shapes stack dataset across all examples. Each example contains 3 entities as shown above. Each cell in the table shows the probability with which the given rule is triggered for the corresponding entity. We can see that the bottom-most entity triggers rule 2 most of the time while the other 2 entities trigger rule 1 most often. This is quite intuitive as, for most examples, the bottom-most entity remains static and does not move at all while the upper entities fall. Therefore, rule 2 captures information which is relevant to static entities, while rule 1 captures physical rules relevant to the inter" + ], + "image_footnote": [], + "bbox": [ + 498, + 315, + 790, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "actions and motion of the upper entities.Table 4. Contrary to the shapes stack task, we see that SNPS outperforms PNPS for the bouncing balls task. The balls are not tightly bound together into a single tower as in the shapes stack. Most of the time, a single ball follows its own dynamics, only occasionally interacting with another ball. Rules in NPS capture interaction dynamics between entities, hence they would only be required to change the state of an entity when it interacts with another entity. In the case of bouncing balls, this interaction takes place through a collision between multiple balls. Since for a single ball, such collisions are rare, SNPS, which has higher application sparsity (less probability of modifying the state of an entity), performs better as compared to PNPS (lower application sparsity). Also note that, SNPS has the ability to compose multiple rules together by virtue of having multiple rule application stages. A visualization of the rule and entity selections by the proposed algorithm can be found in Appendix Figure 9. ", + "bbox": [ + 173, + 662, + 825, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given the analysis in this section, we can conclude that PNPS is expected to work better when interactions among entities are more frequent while SNPS is expected to work better when interactions are rare and most of the time, each entity follows its own dynamics. Note that, for both SNPS and PNPS, the rule application considers only 1 other entity as context. Therefore, both approaches have equal contextual sparsity while the baselines that we consider (SCOFF and RPIN) capture dense pairwise interactions. We discuss the benefits of contextual sparsity in more detail in the next section. More details regarding our setup for the above experiments can be found in Appendix. ", + "bbox": [ + 174, + 814, + 825, + 911 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Benefits of Sparse Interactions Offered by NPS ", + "text_level": 1, + "bbox": [ + 173, + 90, + 540, + 106 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In NPS, one can view the computational graph as a dynamically constructed GNN resulting from applying dynamically selected rules, where the states of the slots are represented on the different nodes of the graph, and different rules dynamically instantiate an hyper-edge between a set of slots (the primary slot and the contextual slot). It is important to emphasize that the topology of the graph induced in NPS is dynamic and sparse (only a few nodes affected), while in most GNNs the topology is fixed and dense (all nodes affected). In this section, through a thorough set of experiments, we show that learning sparse and dynamic interactions using NPS indeed works better for the problems we consider than learning dense interactions using GNNs. We consider two types of tasks: (1) Learning Action Conditioned World Models (2) Physical Reasoning. We use SNPS for all these experiments since in the environments that we consider here, interactions among entities are rare. ", + "bbox": [ + 173, + 116, + 825, + 256 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Learning Action-Conditioned World Models. For learning action-conditioned world models, we follow the same experimental setup as Kipf et al. (2019). Therefore, all the tasks in this section are next- $K$ step $K =$ $\\{ 1 , 5 , 1 0 \\} $ ) prediction tasks, given the intermediate actions, and with the predictions being performed in the latent space. We use the Hits at Rank 1 $( \\mathrm { H } @ 1 )$ metrics described by Kipf et al. (2019) for evaluation. $\\mathrm { H @ 1 }$ is 1 for a particular example if the predicted state representation is nearest to the encoded true observation and 0 otherwise. We report the average of this score over the test set (higher is better). ", + "bbox": [ + 174, + 282, + 553, + 435 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/930cf3ea314d168c1dc25a6a6f25081ea7d31aebfed64db5ef0370e7e94372f0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model NameTestTransfer
SCOFF0.280.15
SCOFF++0.84370.2632
PNPS (10 Rules+1 Null Rule)0.78130.1997
SNPS (10 Rules)0.85180.3553
", + "bbox": [ + 576, + 271, + 813, + 342 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 4: Here we show the ARI achieved by the models on the bouncing balls dataset (higher is better). We can see that SNPS outperforms SCOFF and ${ \\mathrm { S C O F F + + } }$ while PNPS has a poor performance in this task. Results average across 2 seeds. ", + "bbox": [ + 566, + 348, + 825, + 444 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Physics Environment. The physics environment (Ke et al., 2021) simulates a simple physical world. It consists of blocks of unique but unknown weights. The dynamics for the interaction between blocks is that the movement of heavier blocks pushes lighter blocks on their path. This rule creates an acyclic causal graph between the blocks. For an accurate world model, the learner needs to infer the correct weights through demonstrations. Interactions in this environment are sparse and only involve two blocks at a time, therefore we expect NPS to outperform dense architectures like GNNs. This environment is demonstrated in Appendix Fig 11. ", + "bbox": [ + 173, + 449, + 825, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We follow the same setup as Kipf et al. (2019). We use their C-SWM model as baseline. For the proposed model, we only replace the GNN from C-SWM by NPS. GNNs generally share parameters across edges, but in NPS each rule has separate parameters. For a fair comparison to GNN, we use an NPS model with 1 rule. Note that this setting is still different from GNNs as in GNNs at each step every slot is updated by instantiating edges between all pairs of slots, while in NPS an edge is dynamically instantiated between a single pair of slots and only the state of the selected slot (i.e., primary slot) gets updated. ", + "bbox": [ + 173, + 553, + 825, + 650 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6eae7884c2ab3f56bf47fca18f661df64ab49fc50960444321e063fbe05b8b8b.jpg", + "image_caption": [ + "Figure 6: Action-Conditioned World Models, with number of future steps to be predicted for the world-model on the horizontal axes. (a) Here we show a comparison between GNNs and the proposed NPS on the physics environment using the $\\mathrm { H @ 1 }$ metric (higher is better). (b) Comparison of average $\\mathrm { H @ 1 }$ scores across 5 Atari games for the proposed model NPS and GNN. " + ], + "image_footnote": [], + "bbox": [ + 223, + 659, + 766, + 786 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The results of our experiments are presented in Figure 6(a). We can see that NPS outperforms GNNs for all rollouts. Multi-step settings are more difficult to model as errors may get compounded over time steps. The sparsity of NPS (only a single slot affected per step) reduces compounding of errors and enhances symmetry-breaking in the assignment of transformations to rules, while in the case of GNNs, since all entities are affected per step, there is a higher possibility of errors getting compounded. We can see that even with a single rule, we significantly outperform GNNs thus proving the effectiveness of dynamically instantiating edges between entities. ", + "bbox": [ + 174, + 856, + 825, + 911 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 92, + 823, + 133 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Atari Games. We also test the proposed model in the more complicated setting of Atari. Atari games also have sparse interactions between entities. For instance, in Pong, any interaction involves only 2 entities: (1) paddle and ball or (2) ball and the wall. Therefore, we expect sparse interactions captured by NPS to outperform GNNs here as well. ", + "bbox": [ + 174, + 151, + 825, + 207 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We follow the same setup as for the physics environment described in the previous section. We present the results for the Atari experiments in Figure 6(b), showing the average $\\mathrm { H @ 1 }$ score across 5 games: Pong, Space Invaders, Freeway, Breakout, and QBert. As expected, we can see that the proposed model achieves a higher score than the GNN-based C-SWM. The results for the Atari experiments reinforce the claim that NPS is especially good at learning sparse interactions. ", + "bbox": [ + 174, + 213, + 825, + 282 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Learning Rules for Physical Reasoning. To show the effectiveness of the proposed approach for physical reasoning tasks, we evaluate NPS on another dataset: Sprites-MOT (He et al., 2018). The Sprites-MOT dataset was introduced by He et al. (2018). The dataset contains a set of moving objects of various shapes. This dataset aims to test whether a model can handle occlusions correctly. Each frame has consistent bounding boxes which may cause the objects to appear or disappear from the scene. A model which performs well should be able to track the motion of all objects irrespective of whether they are occluded or not. We follow the same setup as Weis et al. (2020). We use the OP3 model (Veerapaneni et al., 2019) as our baseline. To test the proposed model, we replace the GNN-based transition model in OP3 with the proposed NPS. ", + "bbox": [ + 173, + 299, + 825, + 425 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We use the same evaluation protocol as followed by Weis et al. (2020) which is based on the MOT (Multi-object tracking) challenge (Milan et al., 2016). The results on the MOTA and MOTP metrics for this task are presented in Table 5. The results on the other metrics are presented in appendix Table 10. We ask the reader to refer to appendix F.1 for more details about these metrics. We can see that for almost all metrics, NPS outperforms the OP3 baseline. Although this dataset does not contain physical interactions between the objects, sparse rule application should still be useful in dealing with occlusions. At any time step, only a single object is affected by occlusions i.e., it may get occluded due to another object or due to a prespecified bounding box, while the other objects follow their default dynamics. Therefore, a rule should be applied to only the object (or entity) affected (i.e., not visible) due to occlusion and may take into account any other object or entity that is responsible for the occlusion. ", + "bbox": [ + 174, + 431, + 553, + 597 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/c8b9c1fb0031e32c3375d2a376992d3dee1d9313854b01ff0e732e92e5576ab7.jpg", + "table_caption": [ + "Table 5: Sprites-MOT. Comparison between the proposed NPS and the baseline OP3 using the MOTA and MOTP metrics on the sprites-MOT dataset ( $\\uparrow$ : higher is better). Average over 3 random seeds. " + ], + "table_footnote": [], + "table_body": "
ModelMOTA↑MOTP↑
OP3NPS89.1±5.178.4±2.4
90.72±5.1579.91±0.3
", + "bbox": [ + 609, + 449, + 779, + 501 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 597, + 825, + 651 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Discussion and Conclusion ", + "text_level": 1, + "bbox": [ + 176, + 664, + 428, + 681 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For AI agents such as robots trying to make sense of their environment, the only observables are low-level variables like pixels in images. To generalize well, an agent must induce high-level entities as well as discover and disentangle the rules that govern how these entities actually interact with each other. Here we have focused on perceptual inference problems and proposed NPS, a neural instantiation of production systems by introducing an important inductive bias in the architecture following the proposals of Bengio (2017); Goyal & Bengio (2020); Ke et al. (2021). ", + "bbox": [ + 173, + 693, + 825, + 776 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Limitations & Looking Forward. Our experiments highlight the advantages brought by the factorization of knowledge into a small set of entities and sparse sequentially applied rules. Immediate future work would investigate how to take advantage of these inductive biases for more complex physical environments (Ahmed et al., 2020) and novel planning methods, which might be more sample efficient than standard ones (Schrittwieser et al., 2020). ", + "bbox": [ + 174, + 794, + 825, + 863 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We also find that Sequential and Parallel NPS have different properties suited towards different domains. Future work should explore how to effectively combine these two approaches. We discuss this in more detail in Appendix section E.3. ", + "bbox": [ + 176, + 869, + 823, + 911 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6 Acknowledgements ", + "text_level": 1, + "bbox": [ + 174, + 89, + 367, + 107 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The authors would like to thank Matthew Botvinick for useful discussions. The authors would also like to thank Alex Lamb, Stefan Bauer, Nicolas Chapados, Danilo Rezende and Kelsey Allen for brainstorming sessions. We are also thankful to Dianbo Liu, Damjan Kalajdzievski and Osama Ahmed for proofreading. We would like to thank Samsung Electronics Co. Ltd. and CIFAR for funding this research. We would also like to thank Google for providing Google cloud credits used in this work. ", + "bbox": [ + 173, + 121, + 825, + 204 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "References ", + "text_level": 1, + "bbox": [ + 174, + 224, + 266, + 241 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ahmed, O., Träuble, F., Goyal, A., Neitz, A., Wüthrich, M., Bengio, Y., Schölkopf, B., and Bauer, S. Causalworld: A robotic manipulation benchmark for causal structure and transfer learning. arXiv preprint arXiv:2010.04296, 2020. \nAnderson, J. R. Skill acquisition: Compilation of weak-method problem situations. Psychological review, 94(2):192, 1987. \nAndreas, J., Rohrbach, M., Darrell, T., and Klein, D. Neural module networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 39–48, 2016. \nBahdanau, D., Cho, K., and Bengio, Y. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. \nBattaglia, P. W., Pascanu, R., Lai, M., Rezende, D. J., and Kavukcuoglu, K. Interaction networks for learning about objects, relations and physics. CoRR, abs/1612.00222, 2016. URL http: //arxiv.org/abs/1612.00222. \nBattaglia, P. W., Hamrick, J. B., Bapst, V., Sanchez-Gonzalez, A., Zambaldi, V., Malinowski, M., Tacchetti, A., Raposo, D., Santoro, A., Faulkner, R., et al. Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261, 2018. \nBengio, Y. The consciousness prior. arXiv preprint arXiv:1709.08568, 2017. \nBottou, L. and Gallinari, P. A framework for the cooperation of learning algorithms. In Advances in neural information processing systems, pp. 781–788, 1991. \nBronstein, M. M., Bruna, J., LeCun, Y., Szlam, A., and Vandergheynst, P. Geometric deep learning: going beyond euclidean data. IEEE Signal Processing Magazine, 34(4):18–42, 2017. \nBunel, R., Hausknecht, M., Devlin, J., Singh, R., and Kohli, P. Leveraging grammar and reinforcement learning for neural program synthesis. arXiv preprint arXiv:1805.04276, 2018. \nBurgess, C. P., Matthey, L., Watters, N., Kabra, R., Higgins, I., Botvinick, M., and Lerchner, A. Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390, 2019. \nCai, J., Shin, R., and Song, D. Making neural programming architectures generalize via recursion. arXiv preprint arXiv:1704.06611, 2017. \nChen, R. T., Li, X., Grosse, R., and Duvenaud, D. Isolating sources of disentanglement in variational autoencoders. arXiv preprint arXiv:1802.04942, 2018. \nDing, D., Hill, F., Santoro, A., and Botvinick, M. Object-based attention for spatio-temporal reasoning: Outperforming neuro-symbolic models with flexible distributed architectures. arXiv preprint arXiv:2012.08508, 2020. \nDu, Y., Smith, K., Ulman, T., Tenenbaum, J., and Wu, J. Unsupervised discovery of 3d physical objects from video. arXiv preprint arXiv:2007.12348, 2020. \nEngelcke, M., Kosiorek, A. R., Jones, O. P., and Posner, I. Genesis: Generative scene inference and sampling with object-centric latent representations. arXiv preprint arXiv:1907.13052, 2019. \nEslami, S., Heess, N., Weber, T., Tassa, Y., Szepesvari, D., Kavukcuoglu, K., and Hinton, G. E. Attend, infer, repeat: Fast scene understanding with generative models. arXiv preprint arXiv:1603.08575, 2016. \nEvans, R., Hernández-Orallo, J., Welbl, J., Kohli, P., and Sergot, M. Making sense of sensory input. Artificial Intelligence, pp. 103438, 2019. \nFedus, W., Zoph, B., and Shazeer, N. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. CoRR, abs/2101.03961, 2021. URL https://arxiv.org/abs/ 2101.03961. \nFernando, C., Banarse, D., Blundell, C., Zwols, Y., Ha, D., Rusu, A. A., Pritzel, A., and Wierstra, D. Pathnet: Evolution channels gradient descent in super neural networks. arXiv preprint arXiv:1701.08734, 2017. \nGirshick, R. B. Fast R-CNN. CoRR, abs/1504.08083, 2015. URL http://arxiv.org/abs/1504. 08083. \nGoyal, A. and Bengio, Y. Inductive biases for deep learning of higher-level cognition. arXiv preprint arXiv:2011.15091, 2020. \nGoyal, A., Lamb, A., Hoffmann, J., Sodhani, S., Levine, S., Bengio, Y., and Schölkopf, B. Recurrent independent mechanisms, 2019. \nGoyal, A., Lamb, A., Gampa, P., Beaudoin, P., Levine, S., Blundell, C., Bengio, Y., and Mozer, M. Object files and schemata: Factorizing declarative and procedural knowledge in dynamical systems. arXiv preprint arXiv:2006.16225, 2020. \nGraves, A., Wayne, G., and Danihelka, I. Neural turing machines. arXiv preprint arXiv:1410.5401, 2014. \nGreff, K., Rasmus, A., Berglund, M., Hao, T. H., Schmidhuber, J., and Valpola, H. Tagger: Deep unsupervised perceptual grouping. arXiv preprint arXiv:1606.06724, 2016. \nGreff, K., Kaufman, R. L., Kabra, R., Watters, N., Burgess, C., Zoran, D., Matthey, L., Botvinick, M., and Lerchner, A. Multi-object representation learning with iterative variational inference. arXiv preprint arXiv:1903.00450, 2019. \nGroth, O., Fuchs, F., Posner, I., and Vedaldi, A. Shapestacks: Learning vision-based physical intuition for generalised object stacking. CoRR, abs/1804.08018, 2018. URL http://arxiv.org/abs/ 1804.08018. \nHe, Z., Li, J., Liu, D., He, H., and Barber, D. Tracking by animation: Unsupervised learning of multi-object attentive trackers. CoRR, abs/1809.03137, 2018. URL http://arxiv.org/abs/ 1809.03137. \nHiggins, I., Matthey, L., Pal, A., Burgess, C., Glorot, X., Botvinick, M., Mohamed, S., and Lerchner, A. beta-vae: Learning basic visual concepts with a constrained variational framework. 2016. \nJacobs, R. A., Jordan, M. I., Nowlan, S. J., Hinton, G. E., et al. Adaptive mixtures of local experts. Neural computation, 3(1):79–87, 1991. \nJang, E., Gu, S., and Poole, B. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. \nKe, N. R., Didolkar, A. R., Mittal, S., Goyal, A., Lajoie, G., Bauer, S., Rezende, D. J., Mozer, M. C., Bengio, Y., and Pal, C. Systematic evaluation of causal discovery in visual model based reinforcement learning. 2021. \nKipf, T., Fetaya, E., Wang, K.-C., Welling, M., and Zemel, R. Neural relational inference for interacting systems. arXiv preprint arXiv:1802.04687, 2018. \nKipf, T., van der Pol, E., and Welling, M. Contrastive learning of structured world models. arXiv preprint arXiv:1911.12247, 2019. \nKirsch, L., Kunze, J., and Barber, D. Modular networks: Learning to decompose neural computation. In Advances in Neural Information Processing Systems, pp. 2408–2418, 2018. \nKosiorek, A., Kim, H., Teh, Y. W., and Posner, I. Sequential attend, infer, repeat: Generative modelling of moving objects. Advances in Neural Information Processing Systems, 31:8606–8616, 2018. \nLaird, J. E., Rosenbloom, P. S., and Newell, A. Chunking in soar: The anatomy of a general learning mechanism. Machine learning, 1(1):11–46, 1986. \nLamb, A., Goyal, A., Słowik, A., Mozer, M., Beaudoin, P., and Bengio, Y. Neural function modules with sparse arguments: A dynamic approach to integrating information across layers. arXiv preprint arXiv:2010.08012, 2020. \nLe Roux, N., Heess, N., Shotton, J., and Winn, J. Learning a generative model of images by factoring appearance and shape. Neural Computation, 23(3):593–650, 2011. \nLi, Y., Gimeno, F., Kohli, P., and Vinyals, O. Strong generalization and efficiency in neural programs. arXiv preprint arXiv:2007.03629, 2020. \nLocatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., and Kipf, T. Object-centric learning with slot attention. arXiv preprint arXiv:2006.15055, 2020a. \nLocatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., and Kipf, T. Object-centric learning with slot attention, 2020b. \nLovett, M. C. and Anderson, J. R. Thinking as a production system. The Cambridge handbook of thinking and reasoning, pp. 401–429, 2005. \nMcMillan, C., Mozer, M. C., and Smolensky, P. The connectionist scientist game: rule extraction and refinement in a neural network. In Proceedings of the 13th Annual Conference of the Cognitive Science Society, pp. 424–430, 1991. \nMilan, A., Leal-Taixé, L., Reid, I. D., Roth, S., and Schindler, K. MOT16: A benchmark for multiobject tracking. CoRR, abs/1603.00831, 2016. URL http://arxiv.org/abs/1603.00831. \nNeelakantan, A., Le, Q. V., and Sutskever, I. Neural programmer: Inducing latent programs with gradient descent. arXiv preprint arXiv:1511.04834, 2015. \nQi, H., Wang, X., Pathak, D., Ma, Y., and Malik, J. Learning long-term visual dynamics with region proposal interaction networks. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=_X_4Akcd8Re. \nRahaman, N., Goyal, A., Gondal, M. W., Wuthrich, M., Bauer, S., Sharma, Y., Bengio, Y., and Schölkopf, B. S2rms: Spatially structured recurrent modules. arXiv preprint arXiv:2007.06533, 2020. \nRand, W. M. Objective criteria for the evaluation of clustering methods. Journal of the American Statistical association, 66(336):846–850, 1971. \nRaposo, D., Santoro, A., Barrett, D., Pascanu, R., Lillicrap, T., and Battaglia, P. Discovering objects and their relations from entangled scene representations. arXiv preprint arXiv:1702.05068, 2017. \nReed, S. and De Freitas, N. Neural programmer-interpreters. arXiv preprint arXiv:1511.06279, 2015. \nRonco, E., Gollee, H., and Gawthrop, P. J. Modular neural networks and self-decomposition. Technical Report CSC-96012, 1997. \nRosenbaum, C., Klinger, T., and Riemer, M. Routing networks: Adaptive selection of non-linear functions for multi-task learning. arXiv preprint arXiv:1711.01239, 2017. \nRosenbaum, C., Cases, I., Riemer, M., and Klinger, T. Routing networks and the challenges of modular and compositional computation. arXiv preprint arXiv:1904.12774, 2019. \nScarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M., and Monfardini, G. The graph neural network model. IEEE Transactions on Neural Networks, 20(1):61–80, 2008. \nSchrittwieser, J., Antonoglou, I., Hubert, T., Simonyan, K., Sifre, L., Schmitt, S., Guez, A., Lockhart, E., Hassabis, D., Graepel, T., et al. Mastering atari, go, chess and shogi by planning with a learned model. Nature, 588(7839):604–609, 2020. \nShazeer, N., Mirhoseini, A., Maziarz, K., Davis, A., Le, Q., Hinton, G., and Dean, J. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. \nTacchetti, A., Song, H. F., Mediano, P. A., Zambaldi, V., Rabinowitz, N. C., Graepel, T., Botvinick, M., and Battaglia, P. W. Relational forward models for multi-agent learning. arXiv preprint arXiv:1809.11044, 2018. \nTouretzky, D. S. and Hinton, G. E. A distributed connectionist production system. Cognitive science, 12(3):423–466, 1988. \nTrask, A., Hill, F., Reed, S. E., Rae, J., Dyer, C., and Blunsom, P. Neural arithmetic logic units. In Advances in Neural Information Processing Systems, pp. 8035–8044, 2018. \nVan Steenkiste, S., Chang, M., Greff, K., and Schmidhuber, J. Relational neural expectation maximization: Unsupervised discovery of objects and their interactions. arXiv preprint arXiv:1802.10353, 2018. \nVaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. Attention is all you need, 2017. \nVeerapaneni, R., Co-Reyes, J. D., Chang, M., Janner, M., Finn, C., Wu, J., Tenenbaum, J. B., and Levine, S. Entity abstraction in visual model-based reinforcement learning. CoRR, abs/1910.12827, 2019. URL http://arxiv.org/abs/1910.12827. \nVeerapaneni, R., Co-Reyes, J. D., Chang, M., Janner, M., Finn, C., Wu, J., Tenenbaum, J., and Levine, S. Entity abstraction in visual model-based reinforcement learning. In Conference on Robot Learning, pp. 1439–1456. PMLR, 2020. \nWatters, N., Zoran, D., Weber, T., Battaglia, P., Pascanu, R., and Tacchetti, A. Visual interaction networks: Learning a physics simulator from video. In Advances in neural information processing systems, pp. 4539–4547, 2017. \nWeis, M. A., Chitta, K., Sharma, Y., Brendel, W., Bethge, M., Geiger, A., and Ecker, A. S. Unmasking the inductive biases of unsupervised object representations for video sequences, 2020. \nXu, D., Nair, S., Zhu, Y., Gao, J., Garg, A., Fei-Fei, L., and Savarese, S. Neural task programming: Learning to generalize across hierarchical tasks. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 1–8. IEEE, 2018. \nZablotskaia, P., Dominici, E. A., Sigal, L., and Lehrmann, A. M. Unsupervised video decomposition using spatio-temporal iterative inference. arXiv preprint arXiv:2006.14727, 2020. ", + "bbox": [ + 171, + 248, + 828, + 915 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "", + "bbox": [ + 169, + 79, + 828, + 921 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "", + "bbox": [ + 169, + 61, + 828, + 917 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 90, + 828, + 717 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Checklist ", + "text_level": 1, + "bbox": [ + 174, + 738, + 254, + 755 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "1. For all authors... ", + "bbox": [ + 214, + 765, + 339, + 780 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] \n(b) Did you describe the limitations of your work? [Yes] \n(c) Did you discuss any potential negative societal impacts of your work? [Yes] \n(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] ", + "bbox": [ + 238, + 784, + 825, + 875 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "2. If you are including theoretical results... ", + "bbox": [ + 214, + 878, + 493, + 893 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "(a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A] ", + "bbox": [ + 238, + 897, + 738, + 911 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "", + "bbox": [ + 236, + 92, + 691, + 106 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "3. If you ran experiments... ", + "bbox": [ + 212, + 111, + 393, + 125 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] \n(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] \n(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] See section 4.3 \n(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] ", + "bbox": [ + 238, + 128, + 825, + 247 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... ", + "bbox": [ + 217, + 252, + 823, + 267 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "(a) If your work uses existing assets, did you cite the creators? [Yes] \n(b) Did you mention the license of the assets? [No] All datasets we used are open-sourced \n(c) Did you include any new assets either in the supplemental material or as a URL? [Yes] We include the code for reproducing our experiments \n(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A] All datasets we use are synthetic and created using simulation hence do not include any personal or user data. \n(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A] We do not use any dataset that includes any kind of identifying information. ", + "bbox": [ + 238, + 270, + 825, + 419 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "5. If you used crowdsourcing or conducted research with human subjects... ", + "bbox": [ + 214, + 422, + 705, + 438 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A] \n(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A] \n(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A] ", + "bbox": [ + 238, + 443, + 825, + 531 + ], + "page_idx": 14 + } +] \ No newline at end of file diff --git a/parse/train/xQGYquca0gB/xQGYquca0gB_middle.json b/parse/train/xQGYquca0gB/xQGYquca0gB_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..0be49c241db09085d736645681990181fc6527fe --- /dev/null +++ b/parse/train/xQGYquca0gB/xQGYquca0gB_middle.json @@ -0,0 +1,45659 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 204, + 97, + 407, + 117 + ], + "lines": [ + { + "bbox": [ + 202, + 95, + 408, + 119 + ], + "spans": [ + { + "bbox": [ + 202, + 95, + 408, + 119 + ], + "score": 1.0, + "content": "Neural Production Systems", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 137, + 158, + 476, + 182 + ], + "lines": [ + { + "bbox": [ + 136, + 157, + 479, + 172 + ], + "spans": [ + { + "bbox": [ + 136, + 157, + 370, + 172 + ], + "score": 1.0, + "content": "Aniket Didolkar*, 1, Anirudh Goyal *, 1, Nan Rosemary", + "type": "text" + }, + { + "bbox": [ + 371, + 158, + 390, + 170 + ], + "score": 0.42, + "content": "\\mathbf { K } \\mathbf { e } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 157, + 479, + 172 + ], + "score": 1.0, + "content": ", Charles Blundell 2,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 139, + 167, + 471, + 185 + ], + "spans": [ + { + "bbox": [ + 139, + 167, + 471, + 185 + ], + "score": 1.0, + "content": "Philippe Beaudoin 3 Nicolas Heess 2, Michael Mozer 4, **, Yoshua Bengio 1, **", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 283, + 210, + 328, + 224 + ], + "lines": [ + { + "bbox": [ + 282, + 210, + 330, + 225 + ], + "spans": [ + { + "bbox": [ + 282, + 210, + 330, + 225 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 143, + 236, + 469, + 433 + ], + "lines": [ + { + "bbox": [ + 142, + 236, + 469, + 249 + ], + "spans": [ + { + "bbox": [ + 142, + 236, + 469, + 249 + ], + "score": 1.0, + "content": "Visual environments are structured, consisting of distinct objects or entities. These", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 248, + 469, + 259 + ], + "spans": [ + { + "bbox": [ + 142, + 248, + 469, + 259 + ], + "score": 1.0, + "content": "entities have properties—visible or latent—that determine the manner in which", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 258, + 470, + 272 + ], + "spans": [ + { + "bbox": [ + 141, + 258, + 470, + 272 + ], + "score": 1.0, + "content": "they interact with one another. To partition images into entities, deep-learning re-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 270, + 471, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 270, + 471, + 281 + ], + "score": 1.0, + "content": "searchers have proposed structural inductive biases such as slot-based architectures.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 280, + 470, + 293 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 470, + 293 + ], + "score": 1.0, + "content": "To model interactions among entities, equivariant graph neural nets (GNNs) are", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 291, + 471, + 303 + ], + "spans": [ + { + "bbox": [ + 142, + 291, + 471, + 303 + ], + "score": 1.0, + "content": "used, but these are not particularly well suited to the task for two reasons. First,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 303, + 470, + 315 + ], + "spans": [ + { + "bbox": [ + 142, + 303, + 470, + 315 + ], + "score": 1.0, + "content": "GNNs do not predispose interactions to be sparse, as relationships among inde-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 313, + 469, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 313, + 469, + 325 + ], + "score": 1.0, + "content": "pendent entities are likely to be. Second, GNNs do not factorize knowledge about", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 324, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 142, + 324, + 469, + 335 + ], + "score": 1.0, + "content": "interactions in an entity-conditional manner. As an alternative, we take inspiration", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 335, + 469, + 347 + ], + "spans": [ + { + "bbox": [ + 141, + 335, + 469, + 347 + ], + "score": 1.0, + "content": "from cognitive science and resurrect a classic approach, production systems, which", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 346, + 469, + 358 + ], + "spans": [ + { + "bbox": [ + 141, + 346, + 469, + 358 + ], + "score": 1.0, + "content": "consist of a set of rule templates that are applied by binding placeholder variables in", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 357, + 469, + 368 + ], + "spans": [ + { + "bbox": [ + 142, + 357, + 469, + 368 + ], + "score": 1.0, + "content": "the rules to specific entities. Rules are scored on their match to entities, and the best", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 367, + 470, + 381 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 470, + 381 + ], + "score": 1.0, + "content": "fitting rules are applied to update entity properties. In a series of experiments, we", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 379, + 470, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 379, + 470, + 390 + ], + "score": 1.0, + "content": "demonstrate that this architecture achieves a flexible, dynamic flow of control and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 140, + 389, + 470, + 402 + ], + "spans": [ + { + "bbox": [ + 140, + 389, + 470, + 402 + ], + "score": 1.0, + "content": "serves to factorize entity-specific and rule-based information. This disentangling of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 400, + 470, + 413 + ], + "spans": [ + { + "bbox": [ + 141, + 400, + 470, + 413 + ], + "score": 1.0, + "content": "knowledge achieves robust future-state prediction in rich visual environments, out-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 412, + 469, + 423 + ], + "spans": [ + { + "bbox": [ + 141, + 412, + 469, + 423 + ], + "score": 1.0, + "content": "performing state-of-the-art methods using GNNs, and allows for the extrapolation", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 422, + 429, + 434 + ], + "spans": [ + { + "bbox": [ + 141, + 422, + 429, + 434 + ], + "score": 1.0, + "content": "from simple (few object) environments to more complex environments.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 107, + 455, + 190, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 192, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 192, + 470 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 480, + 505, + 633 + ], + "lines": [ + { + "bbox": [ + 105, + 480, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 493 + ], + "score": 1.0, + "content": "Despite never having taken a physics course, every child beyond a young age appreciates that pushing", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 492, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 504 + ], + "score": 1.0, + "content": "a plate off the dining table will cause the plate to break. The laws of physics accurately characterize", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "score": 1.0, + "content": "the dynamics of our natural world, and although explicit knowledge of these laws is not necessary to", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 104, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "reason, we can reason explicitly about objects interacting through these laws. Humans can verbalize", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 524, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 506, + 537 + ], + "score": 1.0, + "content": "knowledge in propositional expressions such as “If a plate drops from table height, it will break,” and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 534, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 104, + 534, + 506, + 549 + ], + "score": 1.0, + "content": "“If a video-game opponent approaches from behind and they are carrying a weapon, they are likely to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "attack you.” Expressing propositional knowledge is not a strength of current deep learning methods", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "for several reasons. First, propositions are discrete and independent from one another. Second,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 568, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 506, + 581 + ], + "score": 1.0, + "content": "propositions must be quantified in the manner of first-order logic; for example, the video-game", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 578, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 210, + 592 + ], + "score": 1.0, + "content": "proposition applies to any", + "type": "text" + }, + { + "bbox": [ + 210, + 579, + 221, + 588 + ], + "score": 0.83, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 578, + 262, + 592 + ], + "score": 1.0, + "content": "for which", + "type": "text" + }, + { + "bbox": [ + 262, + 579, + 272, + 588 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 578, + 506, + 592 + ], + "score": 1.0, + "content": "is an opponent and has a weapon. Incorporating the ability", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "score": 1.0, + "content": "to express and reason about propositions should improve generalization in deep learning methods", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 601, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 613 + ], + "score": 1.0, + "content": "because this knowledge is modular— propositions can be formulated independently of each other—", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 612, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 506, + 624 + ], + "score": 1.0, + "content": "and can therefore be acquired incrementally. Propositions can also be composed with each other and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 622, + 507, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 507, + 636 + ], + "score": 1.0, + "content": "applied consistently to all entities that match, yielding a powerful form of systematic generalization.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 108, + 639, + 503, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 638, + 504, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 504, + 651 + ], + "score": 1.0, + "content": "The classical AI literature from the 1980s can offer deep learning researchers a valuable perspective.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 649, + 504, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 504, + 662 + ], + "score": 1.0, + "content": "In this era, reasoning, planning, and prediction were handled by architectures that performed proposi", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 661, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 661, + 506, + 673 + ], + "score": 1.0, + "content": "tional inference on symbolic knowledge representations. A simple example of such an architecture is", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 681, + 505, + 712 + ], + "lines": [ + { + "bbox": [ + 122, + 679, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 122, + 679, + 505, + 694 + ], + "score": 1.0, + "content": "* Equal Contribution, ** Equal Advising 1 Mila, University of Montreal, 2 Google Deepmind, 3", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 690, + 507, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 507, + 704 + ], + "score": 1.0, + "content": "Waverly, 4 Google Research, Brain Team. Corresponding authors: anirudhgoyal9119@gmail.com,", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 701, + 213, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 701, + 213, + 713 + ], + "score": 1.0, + "content": "adidolkar123@gmail.com", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 731, + 385, + 742 + ], + "lines": [ + { + "bbox": [ + 105, + 730, + 386, + 743 + ], + "spans": [ + { + "bbox": [ + 105, + 730, + 386, + 743 + ], + "score": 1.0, + "content": "35th Conference on Neural Information Processing Systems (NeurIPS 2021).", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 204, + 97, + 407, + 117 + ], + "lines": [ + { + "bbox": [ + 202, + 95, + 408, + 119 + ], + "spans": [ + { + "bbox": [ + 202, + 95, + 408, + 119 + ], + "score": 1.0, + "content": "Neural Production Systems", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 137, + 158, + 476, + 182 + ], + "lines": [ + { + "bbox": [ + 136, + 157, + 479, + 172 + ], + "spans": [ + { + "bbox": [ + 136, + 157, + 370, + 172 + ], + "score": 1.0, + "content": "Aniket Didolkar*, 1, Anirudh Goyal *, 1, Nan Rosemary", + "type": "text" + }, + { + "bbox": [ + 371, + 158, + 390, + 170 + ], + "score": 0.42, + "content": "\\mathbf { K } \\mathbf { e } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 157, + 479, + 172 + ], + "score": 1.0, + "content": ", Charles Blundell 2,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 139, + 167, + 471, + 185 + ], + "spans": [ + { + "bbox": [ + 139, + 167, + 471, + 185 + ], + "score": 1.0, + "content": "Philippe Beaudoin 3 Nicolas Heess 2, Michael Mozer 4, **, Yoshua Bengio 1, **", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 136, + 157, + 479, + 185 + ] + }, + { + "type": "title", + "bbox": [ + 283, + 210, + 328, + 224 + ], + "lines": [ + { + "bbox": [ + 282, + 210, + 330, + 225 + ], + "spans": [ + { + "bbox": [ + 282, + 210, + 330, + 225 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 143, + 236, + 469, + 433 + ], + "lines": [ + { + "bbox": [ + 142, + 236, + 469, + 249 + ], + "spans": [ + { + "bbox": [ + 142, + 236, + 469, + 249 + ], + "score": 1.0, + "content": "Visual environments are structured, consisting of distinct objects or entities. These", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 248, + 469, + 259 + ], + "spans": [ + { + "bbox": [ + 142, + 248, + 469, + 259 + ], + "score": 1.0, + "content": "entities have properties—visible or latent—that determine the manner in which", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 258, + 470, + 272 + ], + "spans": [ + { + "bbox": [ + 141, + 258, + 470, + 272 + ], + "score": 1.0, + "content": "they interact with one another. To partition images into entities, deep-learning re-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 270, + 471, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 270, + 471, + 281 + ], + "score": 1.0, + "content": "searchers have proposed structural inductive biases such as slot-based architectures.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 280, + 470, + 293 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 470, + 293 + ], + "score": 1.0, + "content": "To model interactions among entities, equivariant graph neural nets (GNNs) are", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 291, + 471, + 303 + ], + "spans": [ + { + "bbox": [ + 142, + 291, + 471, + 303 + ], + "score": 1.0, + "content": "used, but these are not particularly well suited to the task for two reasons. First,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 303, + 470, + 315 + ], + "spans": [ + { + "bbox": [ + 142, + 303, + 470, + 315 + ], + "score": 1.0, + "content": "GNNs do not predispose interactions to be sparse, as relationships among inde-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 313, + 469, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 313, + 469, + 325 + ], + "score": 1.0, + "content": "pendent entities are likely to be. Second, GNNs do not factorize knowledge about", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 324, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 142, + 324, + 469, + 335 + ], + "score": 1.0, + "content": "interactions in an entity-conditional manner. As an alternative, we take inspiration", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 335, + 469, + 347 + ], + "spans": [ + { + "bbox": [ + 141, + 335, + 469, + 347 + ], + "score": 1.0, + "content": "from cognitive science and resurrect a classic approach, production systems, which", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 346, + 469, + 358 + ], + "spans": [ + { + "bbox": [ + 141, + 346, + 469, + 358 + ], + "score": 1.0, + "content": "consist of a set of rule templates that are applied by binding placeholder variables in", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 357, + 469, + 368 + ], + "spans": [ + { + "bbox": [ + 142, + 357, + 469, + 368 + ], + "score": 1.0, + "content": "the rules to specific entities. Rules are scored on their match to entities, and the best", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 367, + 470, + 381 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 470, + 381 + ], + "score": 1.0, + "content": "fitting rules are applied to update entity properties. In a series of experiments, we", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 379, + 470, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 379, + 470, + 390 + ], + "score": 1.0, + "content": "demonstrate that this architecture achieves a flexible, dynamic flow of control and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 140, + 389, + 470, + 402 + ], + "spans": [ + { + "bbox": [ + 140, + 389, + 470, + 402 + ], + "score": 1.0, + "content": "serves to factorize entity-specific and rule-based information. This disentangling of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 400, + 470, + 413 + ], + "spans": [ + { + "bbox": [ + 141, + 400, + 470, + 413 + ], + "score": 1.0, + "content": "knowledge achieves robust future-state prediction in rich visual environments, out-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 412, + 469, + 423 + ], + "spans": [ + { + "bbox": [ + 141, + 412, + 469, + 423 + ], + "score": 1.0, + "content": "performing state-of-the-art methods using GNNs, and allows for the extrapolation", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 422, + 429, + 434 + ], + "spans": [ + { + "bbox": [ + 141, + 422, + 429, + 434 + ], + "score": 1.0, + "content": "from simple (few object) environments to more complex environments.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 12.5, + "bbox_fs": [ + 140, + 236, + 471, + 434 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 455, + 190, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 192, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 192, + 470 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 480, + 505, + 633 + ], + "lines": [ + { + "bbox": [ + 105, + 480, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 493 + ], + "score": 1.0, + "content": "Despite never having taken a physics course, every child beyond a young age appreciates that pushing", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 492, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 504 + ], + "score": 1.0, + "content": "a plate off the dining table will cause the plate to break. The laws of physics accurately characterize", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "score": 1.0, + "content": "the dynamics of our natural world, and although explicit knowledge of these laws is not necessary to", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 104, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "reason, we can reason explicitly about objects interacting through these laws. Humans can verbalize", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 524, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 506, + 537 + ], + "score": 1.0, + "content": "knowledge in propositional expressions such as “If a plate drops from table height, it will break,” and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 534, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 104, + 534, + 506, + 549 + ], + "score": 1.0, + "content": "“If a video-game opponent approaches from behind and they are carrying a weapon, they are likely to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "attack you.” Expressing propositional knowledge is not a strength of current deep learning methods", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "for several reasons. First, propositions are discrete and independent from one another. Second,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 568, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 506, + 581 + ], + "score": 1.0, + "content": "propositions must be quantified in the manner of first-order logic; for example, the video-game", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 578, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 210, + 592 + ], + "score": 1.0, + "content": "proposition applies to any", + "type": "text" + }, + { + "bbox": [ + 210, + 579, + 221, + 588 + ], + "score": 0.83, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 578, + 262, + 592 + ], + "score": 1.0, + "content": "for which", + "type": "text" + }, + { + "bbox": [ + 262, + 579, + 272, + 588 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 578, + 506, + 592 + ], + "score": 1.0, + "content": "is an opponent and has a weapon. Incorporating the ability", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "score": 1.0, + "content": "to express and reason about propositions should improve generalization in deep learning methods", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 601, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 613 + ], + "score": 1.0, + "content": "because this knowledge is modular— propositions can be formulated independently of each other—", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 612, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 506, + 624 + ], + "score": 1.0, + "content": "and can therefore be acquired incrementally. Propositions can also be composed with each other and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 622, + 507, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 507, + 636 + ], + "score": 1.0, + "content": "applied consistently to all entities that match, yielding a powerful form of systematic generalization.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 29.5, + "bbox_fs": [ + 104, + 480, + 507, + 636 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 639, + 503, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 638, + 504, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 504, + 651 + ], + "score": 1.0, + "content": "The classical AI literature from the 1980s can offer deep learning researchers a valuable perspective.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 649, + 504, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 504, + 662 + ], + "score": 1.0, + "content": "In this era, reasoning, planning, and prediction were handled by architectures that performed proposi", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 661, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 661, + 506, + 673 + ], + "score": 1.0, + "content": "tional inference on symbolic knowledge representations. A simple example of such an architecture is", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 226, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 506, + 238 + ], + "score": 1.0, + "content": "the production system (Laird et al., 1986; Anderson, 1987), which expresses knowledge by condition-", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "action rules. The rules operate on a working memory: rule conditions are matched to entities in", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "working memory inspired by cognitive science, and such a match can trigger computational actions", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 259, + 436, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 436, + 271 + ], + "score": 1.0, + "content": "that update working memory or external actions that operate on the outside world.", + "type": "text", + "cross_page": true + } + ], + "index": 12 + } + ], + "index": 38, + "bbox_fs": [ + 106, + 638, + 506, + 673 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 68, + 504, + 149 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 68, + 504, + 149 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 68, + 504, + 149 + ], + "spans": [ + { + "bbox": [ + 107, + 68, + 504, + 149 + ], + "score": 0.929, + "type": "image", + "image_path": "efc8469bfcada4d38e5ed73c14c9b9bca7fe99729b314148655c694d81e11332.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 68, + 504, + 95.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 95.0, + 504, + 122.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 122.0, + 504, + 149.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 154, + 505, + 220 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "Figure 1: In this figure we show a visual comparison between NPS and dense architectures like", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "score": 1.0, + "content": "GNNs. In NPS, a rule is only applied when an interaction takes place and it is applied only to the slots", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "affected by the interaction. NPS also uses different rules for different kinds of interactions, while in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 186, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 506, + 200 + ], + "score": 1.0, + "content": "GNN a common rule is applied to all slots irrespective of whether an interaction takes place or not", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "(because of parameter sharing). Note the dynamic nature of the interaction graph in NPS, while in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 209, + 208, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 208, + 222 + ], + "score": 1.0, + "content": "GNN, the graph is static.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 226, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 506, + 238 + ], + "score": 1.0, + "content": "the production system (Laird et al., 1986; Anderson, 1987), which expresses knowledge by condition-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "action rules. The rules operate on a working memory: rule conditions are matched to entities in", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "working memory inspired by cognitive science, and such a match can trigger computational actions", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 259, + 436, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 436, + 271 + ], + "score": 1.0, + "content": "that update working memory or external actions that operate on the outside world.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 275, + 505, + 450 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 288 + ], + "score": 1.0, + "content": "Production systems were typically used to model high-level cognition, e.g., mathematical problem", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "solving or procedure following; perception was not the focus of these models. It was assumed", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "that the results of perception were placed into working memory in a symbolic form that could be", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 307, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 104, + 307, + 506, + 322 + ], + "score": 1.0, + "content": "operated on with the rules. In this article, we revisit production systems but from a deep learning", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 319, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 505, + 331 + ], + "score": 1.0, + "content": "perspective which naturally integrates perceptual processing and subsequent inference for visual", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "reasoning problems. We describe an end-to-end deep learning model that constructs object-centric", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 341, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 505, + 353 + ], + "score": 1.0, + "content": "representations of entities in videos, and then operates on these entities with differentiable—and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "thus learnable—production rules. The essence of these rules, carried over from traditional symbolic", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 363, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 375 + ], + "score": 1.0, + "content": "system, is that they operate on variables that are bound, or linked, to the entities in the world. In the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "score": 1.0, + "content": "deep learning implementation, each production rule is represented by a distinct MLP with query-key", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "attention mechanisms to specify the rule-entity binding and to determine when the rule should be", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "score": 1.0, + "content": "triggered for a given entity. We are not the first to propose a neural instantiation of a production system", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "architecture. Touretzky & Hinton (1988) gave a proof of principle that neural net hardware could be", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 417, + 504, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 504, + 429 + ], + "score": 1.0, + "content": "hardwired to implement a production system for symbolic reasoning; our work fundamentally differs", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 427, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 506, + 442 + ], + "score": 1.0, + "content": "from theirs in that (1) we focus on perceptual inference problems and (2) we use the architecture as", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 438, + 227, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 227, + 452 + ], + "score": 1.0, + "content": "an inductive bias for learning.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 20.5 + }, + { + "type": "title", + "bbox": [ + 108, + 469, + 222, + 481 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 223, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 223, + 483 + ], + "score": 1.0, + "content": "1.1 Variables and entities", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 492, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 491, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 505, + 505 + ], + "score": 1.0, + "content": "What makes a rule general-purpose is that it incorporates placeholder variables that can be bound to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 503, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 515 + ], + "score": 1.0, + "content": "arbitrary values or—the term we prefer in this article—entities. This notion of binding is familiar in", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "score": 1.0, + "content": "functional programming languages, where these variables are called arguments. Analogously, the use", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 524, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 538 + ], + "score": 1.0, + "content": "of variables in the production rules we describe enable a model to reason about any set of entities that", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 536, + 264, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 264, + 549 + ], + "score": 1.0, + "content": "satisfy the selection criteria of the rule.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 552, + 505, + 586 + ], + "lines": [ + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "score": 1.0, + "content": "Consider a simple function in C like int add(int a, int b). This function binds its two integer", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 192, + 576 + ], + "score": 1.0, + "content": "operands to variables", + "type": "text" + }, + { + "bbox": [ + 193, + 565, + 200, + 573 + ], + "score": 0.7, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 563, + 217, + 576 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 217, + 564, + 223, + 573 + ], + "score": 0.56, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 563, + 506, + 576 + ], + "score": 1.0, + "content": ". The function does not apply if the operands are, say, character strings.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 573, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 505, + 588 + ], + "score": 1.0, + "content": "The use of variables enables a programmer to reuse the same function to add any two integer values", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 591, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "In order for rules to operate on entities, these entities must be represented explicitly. That is, the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "score": 1.0, + "content": "visual world needs to be parsed in a task-relevant manner, e.g., distinguishing the sprites in a video", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 612, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 626 + ], + "score": 1.0, + "content": "game or the vehicles and pedestrians approaching an autonomous vehicle. Only in the past few years", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 624, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 505, + 635 + ], + "score": 1.0, + "content": "have deep learning vision researchers developed methods for object-centric representation (Le Roux", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 635, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 506, + 646 + ], + "score": 1.0, + "content": "et al., 2011; Eslami et al., 2016; Greff et al., 2016; Raposo et al., 2017; Van Steenkiste et al., 2018;", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 644, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 658 + ], + "score": 1.0, + "content": "Kosiorek et al., 2018; Engelcke et al., 2019; Burgess et al., 2019; Greff et al., 2019; Locatello et al.,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "2020a; Ahmed et al., 2020; Goyal et al., 2019; Zablotskaia et al., 2020; Rahaman et al., 2020; Du", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 667, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 679 + ], + "score": 1.0, + "content": "et al., 2020; Ding et al., 2020; Goyal et al., 2020; Ke et al., 2021). These methods differ in details but", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "share the notion of a fixed number of slots (see Figure 1 for example), also known as object files, each", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 104, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "encapsulating information about a single object. Importantly, the slots are interchangeable, meaning", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "that it doesn’t matter if a scene with an apple and an orange encodes the apple in slot 1 and orange in", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 710, + 187, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 187, + 723 + ], + "score": 1.0, + "content": "slot 2 or vice-versa.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 43.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 753 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 8 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 68, + 504, + 149 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 68, + 504, + 149 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 68, + 504, + 149 + ], + "spans": [ + { + "bbox": [ + 107, + 68, + 504, + 149 + ], + "score": 0.929, + "type": "image", + "image_path": "efc8469bfcada4d38e5ed73c14c9b9bca7fe99729b314148655c694d81e11332.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 68, + 504, + 95.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 95.0, + 504, + 122.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 122.0, + 504, + 149.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 154, + 505, + 220 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "Figure 1: In this figure we show a visual comparison between NPS and dense architectures like", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "score": 1.0, + "content": "GNNs. In NPS, a rule is only applied when an interaction takes place and it is applied only to the slots", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "affected by the interaction. NPS also uses different rules for different kinds of interactions, while in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 186, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 506, + 200 + ], + "score": 1.0, + "content": "GNN a common rule is applied to all slots irrespective of whether an interaction takes place or not", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "(because of parameter sharing). Note the dynamic nature of the interaction graph in NPS, while in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 209, + 208, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 208, + 222 + ], + "score": 1.0, + "content": "GNN, the graph is static.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 270 + ], + "lines": [], + "index": 10.5, + "bbox_fs": [ + 105, + 226, + 506, + 271 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 275, + 505, + 450 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 288 + ], + "score": 1.0, + "content": "Production systems were typically used to model high-level cognition, e.g., mathematical problem", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "solving or procedure following; perception was not the focus of these models. It was assumed", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "that the results of perception were placed into working memory in a symbolic form that could be", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 307, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 104, + 307, + 506, + 322 + ], + "score": 1.0, + "content": "operated on with the rules. In this article, we revisit production systems but from a deep learning", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 319, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 505, + 331 + ], + "score": 1.0, + "content": "perspective which naturally integrates perceptual processing and subsequent inference for visual", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "reasoning problems. We describe an end-to-end deep learning model that constructs object-centric", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 341, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 505, + 353 + ], + "score": 1.0, + "content": "representations of entities in videos, and then operates on these entities with differentiable—and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "thus learnable—production rules. The essence of these rules, carried over from traditional symbolic", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 363, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 375 + ], + "score": 1.0, + "content": "system, is that they operate on variables that are bound, or linked, to the entities in the world. In the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "score": 1.0, + "content": "deep learning implementation, each production rule is represented by a distinct MLP with query-key", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "attention mechanisms to specify the rule-entity binding and to determine when the rule should be", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "score": 1.0, + "content": "triggered for a given entity. We are not the first to propose a neural instantiation of a production system", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "architecture. Touretzky & Hinton (1988) gave a proof of principle that neural net hardware could be", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 417, + 504, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 504, + 429 + ], + "score": 1.0, + "content": "hardwired to implement a production system for symbolic reasoning; our work fundamentally differs", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 427, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 506, + 442 + ], + "score": 1.0, + "content": "from theirs in that (1) we focus on perceptual inference problems and (2) we use the architecture as", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 438, + 227, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 227, + 452 + ], + "score": 1.0, + "content": "an inductive bias for learning.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 20.5, + "bbox_fs": [ + 104, + 276, + 506, + 452 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 469, + 222, + 481 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 223, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 223, + 483 + ], + "score": 1.0, + "content": "1.1 Variables and entities", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 492, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 491, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 505, + 505 + ], + "score": 1.0, + "content": "What makes a rule general-purpose is that it incorporates placeholder variables that can be bound to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 503, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 515 + ], + "score": 1.0, + "content": "arbitrary values or—the term we prefer in this article—entities. This notion of binding is familiar in", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "score": 1.0, + "content": "functional programming languages, where these variables are called arguments. Analogously, the use", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 524, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 538 + ], + "score": 1.0, + "content": "of variables in the production rules we describe enable a model to reason about any set of entities that", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 536, + 264, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 264, + 549 + ], + "score": 1.0, + "content": "satisfy the selection criteria of the rule.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 491, + 505, + 549 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 552, + 505, + 586 + ], + "lines": [ + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "score": 1.0, + "content": "Consider a simple function in C like int add(int a, int b). This function binds its two integer", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 192, + 576 + ], + "score": 1.0, + "content": "operands to variables", + "type": "text" + }, + { + "bbox": [ + 193, + 565, + 200, + 573 + ], + "score": 0.7, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 563, + 217, + 576 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 217, + 564, + 223, + 573 + ], + "score": 0.56, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 563, + 506, + 576 + ], + "score": 1.0, + "content": ". The function does not apply if the operands are, say, character strings.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 573, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 505, + 588 + ], + "score": 1.0, + "content": "The use of variables enables a programmer to reuse the same function to add any two integer values", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 551, + 506, + 588 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 591, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "In order for rules to operate on entities, these entities must be represented explicitly. That is, the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "score": 1.0, + "content": "visual world needs to be parsed in a task-relevant manner, e.g., distinguishing the sprites in a video", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 612, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 626 + ], + "score": 1.0, + "content": "game or the vehicles and pedestrians approaching an autonomous vehicle. Only in the past few years", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 624, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 505, + 635 + ], + "score": 1.0, + "content": "have deep learning vision researchers developed methods for object-centric representation (Le Roux", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 635, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 506, + 646 + ], + "score": 1.0, + "content": "et al., 2011; Eslami et al., 2016; Greff et al., 2016; Raposo et al., 2017; Van Steenkiste et al., 2018;", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 644, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 658 + ], + "score": 1.0, + "content": "Kosiorek et al., 2018; Engelcke et al., 2019; Burgess et al., 2019; Greff et al., 2019; Locatello et al.,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "2020a; Ahmed et al., 2020; Goyal et al., 2019; Zablotskaia et al., 2020; Rahaman et al., 2020; Du", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 667, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 679 + ], + "score": 1.0, + "content": "et al., 2020; Ding et al., 2020; Goyal et al., 2020; Ke et al., 2021). These methods differ in details but", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "share the notion of a fixed number of slots (see Figure 1 for example), also known as object files, each", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 104, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "encapsulating information about a single object. Importantly, the slots are interchangeable, meaning", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "that it doesn’t matter if a scene with an apple and an orange encodes the apple in slot 1 and orange in", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 710, + 187, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 187, + 723 + ], + "score": 1.0, + "content": "slot 2 or vice-versa.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 43.5, + "bbox_fs": [ + 104, + 591, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 226 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "score": 1.0, + "content": "A model of visual reasoning must not only be able to represent entities but must also express", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "knowledge about entity dynamics and interactions. To ensure systematic predictions, a model must", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 506, + 107 + ], + "score": 1.0, + "content": "be capable of applying knowledge to an entity regardless of the slot it is in and must be capable of", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 506, + 118 + ], + "score": 1.0, + "content": "applying the same knowledge to multiple instances of an entity. Several distinct approaches exist in", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 116, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 506, + 129 + ], + "score": 1.0, + "content": "the literature. The predominant approach uses graph neural networks to model slot-to-slot interactions", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "(Scarselli et al., 2008; Bronstein et al., 2017; Watters et al., 2017; Van Steenkiste et al., 2018; Kipf", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "et al., 2018; Battaglia et al., 2018; Tacchetti et al., 2018). To ensure systematicity, the GNN must", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "score": 1.0, + "content": "share parameters among the edges. In a recent article, Goyal et al. (2020) developed a more general", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 160, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 505, + 172 + ], + "score": 1.0, + "content": "framework in which parameters are shared but slots can dynamically select which parameters to use in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 169, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 104, + 169, + 506, + 185 + ], + "score": 1.0, + "content": "a state-dependent manner. Each set of parameters is referred to as a schema, and slots use a query-key", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "attention mechanism to select which schema to apply at each time step. Multiple slots can select the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 192, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 205 + ], + "score": 1.0, + "content": "same schema. In both GNNs and SCOFF, modeling dynamics involves each slot interacting with", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 204, + 505, + 215 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 505, + 215 + ], + "score": 1.0, + "content": "each other slot. In the work we describe in this article, we replace the direct slot-to-slot interactions", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 215, + 439, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 439, + 227 + ], + "score": 1.0, + "content": "with rules, which mediate sparse interactions among slots (See arrows in Figure 1).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 230, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 231, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 506, + 243 + ], + "score": 1.0, + "content": "Thus our main contribution is that we introduce NPS, which offers a way to model dynamic and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 242, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 506, + 254 + ], + "score": 1.0, + "content": "sparse interactions among the variables in a graph and also allows dynamic sharing of multiple sets of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "parameters among these interactions. Most architectures used for modelling interactions in the current", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "literature use statically instantiated graph which model all possible interactions for a given variable", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 275, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 286 + ], + "score": 1.0, + "content": "at each step i.e. dense interactions. Also such dense architectures share a single set of parameters", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 505, + 298 + ], + "score": 1.0, + "content": "across all interactions which maybe quite restrictive in terms of representational capacity. A visual", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 297, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 308 + ], + "score": 1.0, + "content": "comparison between these two kinds of architectures is shown in Figure 1. Through our experiments", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "score": 1.0, + "content": "we show the advantage of modeling interactions in the proposed manner using NPS in visually rich", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 318, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 506, + 330 + ], + "score": 1.0, + "content": "physical environments. We also show that our method results in an intuitive factorization of rules and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 329, + 141, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 141, + 341 + ], + "score": 1.0, + "content": "entities.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 108, + 354, + 222, + 368 + ], + "lines": [ + { + "bbox": [ + 104, + 352, + 224, + 371 + ], + "spans": [ + { + "bbox": [ + 104, + 352, + 224, + 371 + ], + "score": 1.0, + "content": "2 Production System", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 383, + 505, + 449 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 506, + 397 + ], + "score": 1.0, + "content": "Formally, our notion of a production system consists of a set of entities and a set of rules, along", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 393, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 506, + 407 + ], + "score": 1.0, + "content": "with a mechanism for selecting rules to apply on subsets of the entities. Implicit in a rule is a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 405, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 506, + 418 + ], + "score": 1.0, + "content": "specification of the properties of relevant entities, e.g., a rule might apply to one type of sprite in a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 416, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 428 + ], + "score": 1.0, + "content": "video game but not another. The control flow of a production system dynamically selects rules as", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 427, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 438 + ], + "score": 1.0, + "content": "well as bindings between rules and entities, allowing different rules to be chosen and different entities", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 438, + 267, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 267, + 450 + ], + "score": 1.0, + "content": "to be manipulated at each point in time.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 454, + 505, + 498 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 506, + 468 + ], + "score": 1.0, + "content": "The neural production system we describe shares essential properties with traditional production", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 465, + 507, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 507, + 478 + ], + "score": 1.0, + "content": "system, particularly with regard to the compositionality and generality of the knowledge they embody.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 476, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 488 + ], + "score": 1.0, + "content": "Lovett & Anderson (2005) describe four desirable properties commonly attributed to symbolic", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 487, + 264, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 264, + 499 + ], + "score": 1.0, + "content": "systems that apply to our work as well.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 503, + 505, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "score": 1.0, + "content": "Production rules are modular. Each production rule represents a unit of knowledge and are atomic", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "such that any production rule can be intervened (added, modified or deleted) independently of other", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 526, + 231, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 231, + 538 + ], + "score": 1.0, + "content": "production rules in the system.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 541, + 504, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "score": 1.0, + "content": "Production rules are abstract. Production rules allow for generalization because their conditions", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 552, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 506, + 565 + ], + "score": 1.0, + "content": "may be represented as high-level abstract knowledge that match to a wide range of patterns. These", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "conditions specify the attributes of relationship(s) between entities without specifying the entities", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "score": 1.0, + "content": "themselves. The ability to represent abstract knowledge allows for the transfer of learning across", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 585, + 474, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 474, + 598 + ], + "score": 1.0, + "content": "different environments as long as they fit within the conditions of the given production rule.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 601, + 505, + 657 + ], + "lines": [ + { + "bbox": [ + 105, + 600, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 614 + ], + "score": 1.0, + "content": "Production rules are sparse. In order that production rules have broad applicability, they involve", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 613, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 505, + 625 + ], + "score": 1.0, + "content": "only a subset of entities. This assumption imposes a strong prior that dependencies among entities", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "score": 1.0, + "content": "are sparse. In the context of visual reasoning, we conjecture that this prior is superior to what has", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 632, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 649 + ], + "score": 1.0, + "content": "often been assumed in the past, particularly in the disentanglement literature—independence among", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 646, + 302, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 302, + 658 + ], + "score": 1.0, + "content": "entities Higgins et al. (2016); Chen et al. (2018).", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 108, + 662, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 105, + 661, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 505, + 675 + ], + "score": 1.0, + "content": "Production rules represent causal knowledge and are thus asymmetric. Each rule can be decomposed", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 104, + 672, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 672, + 506, + 685 + ], + "score": 1.0, + "content": "into a {condition, action} pair, where the action reflects a state change that is a causal consequence of", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 683, + 210, + 696 + ], + "spans": [ + { + "bbox": [ + 106, + 683, + 210, + 696 + ], + "score": 1.0, + "content": "the conditions being met.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 503, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 700, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 504, + 712 + ], + "score": 1.0, + "content": "These four properties are sufficient conditions for knowledge to be expressed in production rule form.", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "These properties specify how knowledge is represented, but not what knowledge is represented. The", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 51.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 309, + 752 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 226 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "score": 1.0, + "content": "A model of visual reasoning must not only be able to represent entities but must also express", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "knowledge about entity dynamics and interactions. To ensure systematic predictions, a model must", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 506, + 107 + ], + "score": 1.0, + "content": "be capable of applying knowledge to an entity regardless of the slot it is in and must be capable of", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 506, + 118 + ], + "score": 1.0, + "content": "applying the same knowledge to multiple instances of an entity. Several distinct approaches exist in", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 116, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 506, + 129 + ], + "score": 1.0, + "content": "the literature. The predominant approach uses graph neural networks to model slot-to-slot interactions", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "(Scarselli et al., 2008; Bronstein et al., 2017; Watters et al., 2017; Van Steenkiste et al., 2018; Kipf", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "et al., 2018; Battaglia et al., 2018; Tacchetti et al., 2018). To ensure systematicity, the GNN must", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "score": 1.0, + "content": "share parameters among the edges. In a recent article, Goyal et al. (2020) developed a more general", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 160, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 505, + 172 + ], + "score": 1.0, + "content": "framework in which parameters are shared but slots can dynamically select which parameters to use in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 169, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 104, + 169, + 506, + 185 + ], + "score": 1.0, + "content": "a state-dependent manner. Each set of parameters is referred to as a schema, and slots use a query-key", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "attention mechanism to select which schema to apply at each time step. Multiple slots can select the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 192, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 205 + ], + "score": 1.0, + "content": "same schema. In both GNNs and SCOFF, modeling dynamics involves each slot interacting with", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 204, + 505, + 215 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 505, + 215 + ], + "score": 1.0, + "content": "each other slot. In the work we describe in this article, we replace the direct slot-to-slot interactions", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 215, + 439, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 439, + 227 + ], + "score": 1.0, + "content": "with rules, which mediate sparse interactions among slots (See arrows in Figure 1).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 6.5, + "bbox_fs": [ + 104, + 72, + 506, + 227 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 230, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 231, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 506, + 243 + ], + "score": 1.0, + "content": "Thus our main contribution is that we introduce NPS, which offers a way to model dynamic and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 242, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 506, + 254 + ], + "score": 1.0, + "content": "sparse interactions among the variables in a graph and also allows dynamic sharing of multiple sets of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "parameters among these interactions. Most architectures used for modelling interactions in the current", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "literature use statically instantiated graph which model all possible interactions for a given variable", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 275, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 286 + ], + "score": 1.0, + "content": "at each step i.e. dense interactions. Also such dense architectures share a single set of parameters", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 505, + 298 + ], + "score": 1.0, + "content": "across all interactions which maybe quite restrictive in terms of representational capacity. A visual", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 297, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 308 + ], + "score": 1.0, + "content": "comparison between these two kinds of architectures is shown in Figure 1. Through our experiments", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "score": 1.0, + "content": "we show the advantage of modeling interactions in the proposed manner using NPS in visually rich", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 318, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 506, + 330 + ], + "score": 1.0, + "content": "physical environments. We also show that our method results in an intuitive factorization of rules and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 329, + 141, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 141, + 341 + ], + "score": 1.0, + "content": "entities.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 231, + 506, + 341 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 354, + 222, + 368 + ], + "lines": [ + { + "bbox": [ + 104, + 352, + 224, + 371 + ], + "spans": [ + { + "bbox": [ + 104, + 352, + 224, + 371 + ], + "score": 1.0, + "content": "2 Production System", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 383, + 505, + 449 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 506, + 397 + ], + "score": 1.0, + "content": "Formally, our notion of a production system consists of a set of entities and a set of rules, along", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 393, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 506, + 407 + ], + "score": 1.0, + "content": "with a mechanism for selecting rules to apply on subsets of the entities. Implicit in a rule is a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 405, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 506, + 418 + ], + "score": 1.0, + "content": "specification of the properties of relevant entities, e.g., a rule might apply to one type of sprite in a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 416, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 428 + ], + "score": 1.0, + "content": "video game but not another. The control flow of a production system dynamically selects rules as", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 427, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 438 + ], + "score": 1.0, + "content": "well as bindings between rules and entities, allowing different rules to be chosen and different entities", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 438, + 267, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 267, + 450 + ], + "score": 1.0, + "content": "to be manipulated at each point in time.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 381, + 506, + 450 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 454, + 505, + 498 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 506, + 468 + ], + "score": 1.0, + "content": "The neural production system we describe shares essential properties with traditional production", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 465, + 507, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 507, + 478 + ], + "score": 1.0, + "content": "system, particularly with regard to the compositionality and generality of the knowledge they embody.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 476, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 488 + ], + "score": 1.0, + "content": "Lovett & Anderson (2005) describe four desirable properties commonly attributed to symbolic", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 487, + 264, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 264, + 499 + ], + "score": 1.0, + "content": "systems that apply to our work as well.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 453, + 507, + 499 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 503, + 505, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "score": 1.0, + "content": "Production rules are modular. Each production rule represents a unit of knowledge and are atomic", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "such that any production rule can be intervened (added, modified or deleted) independently of other", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 526, + 231, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 231, + 538 + ], + "score": 1.0, + "content": "production rules in the system.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 502, + 505, + 538 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 541, + 504, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "score": 1.0, + "content": "Production rules are abstract. Production rules allow for generalization because their conditions", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 552, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 506, + 565 + ], + "score": 1.0, + "content": "may be represented as high-level abstract knowledge that match to a wide range of patterns. These", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "conditions specify the attributes of relationship(s) between entities without specifying the entities", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "score": 1.0, + "content": "themselves. The ability to represent abstract knowledge allows for the transfer of learning across", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 585, + 474, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 474, + 598 + ], + "score": 1.0, + "content": "different environments as long as they fit within the conditions of the given production rule.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 541, + 506, + 598 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 601, + 505, + 657 + ], + "lines": [ + { + "bbox": [ + 105, + 600, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 614 + ], + "score": 1.0, + "content": "Production rules are sparse. In order that production rules have broad applicability, they involve", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 613, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 505, + 625 + ], + "score": 1.0, + "content": "only a subset of entities. This assumption imposes a strong prior that dependencies among entities", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "score": 1.0, + "content": "are sparse. In the context of visual reasoning, we conjecture that this prior is superior to what has", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 632, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 649 + ], + "score": 1.0, + "content": "often been assumed in the past, particularly in the disentanglement literature—independence among", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 646, + 302, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 302, + 658 + ], + "score": 1.0, + "content": "entities Higgins et al. (2016); Chen et al. (2018).", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 600, + 505, + 658 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 662, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 105, + 661, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 505, + 675 + ], + "score": 1.0, + "content": "Production rules represent causal knowledge and are thus asymmetric. Each rule can be decomposed", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 104, + 672, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 672, + 506, + 685 + ], + "score": 1.0, + "content": "into a {condition, action} pair, where the action reflects a state change that is a causal consequence of", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 683, + 210, + 696 + ], + "spans": [ + { + "bbox": [ + 106, + 683, + 210, + 696 + ], + "score": 1.0, + "content": "the conditions being met.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49, + "bbox_fs": [ + 104, + 661, + 506, + 696 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 503, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 700, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 504, + 712 + ], + "score": 1.0, + "content": "These four properties are sufficient conditions for knowledge to be expressed in production rule form.", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "These properties specify how knowledge is represented, but not what knowledge is represented. The", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "score": 1.0, + "content": "latter is inferred by learning mechanisms under the inductive bias provided by the form of production", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 81, + 132, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 81, + 132, + 97 + ], + "score": 1.0, + "content": "rules.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 51.5, + "bbox_fs": [ + 106, + 700, + 505, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 504, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "score": 1.0, + "content": "latter is inferred by learning mechanisms under the inductive bias provided by the form of production", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 81, + 132, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 81, + 132, + 97 + ], + "score": 1.0, + "content": "rules.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 106, + 105, + 385, + 119 + ], + "lines": [ + { + "bbox": [ + 104, + 103, + 386, + 122 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 386, + 122 + ], + "score": 1.0, + "content": "3 Neural Production System: Slots and Sparse Rules", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 122, + 505, + 210 + ], + "lines": [ + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "score": 1.0, + "content": "The Neural Production System (NPS), illustrated in Figure 2, provides an architectural backbone", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 133, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 133, + 505, + 145 + ], + "score": 1.0, + "content": "that supports the detection and inference of entity (object) representations in an input sequence, and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "the underlying rules which govern the interactions between these entities in time and space. The", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 252, + 168 + ], + "score": 1.0, + "content": "input sequence indexed by time step", + "type": "text" + }, + { + "bbox": [ + 252, + 156, + 257, + 165 + ], + "score": 0.58, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 154, + 261, + 168 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 261, + 155, + 350, + 167 + ], + "score": 0.9, + "content": "\\{ \\pmb { x } ^ { 1 } , \\dotsc , \\pmb { x } ^ { t } , \\dotsc , \\pmb { x } ^ { T } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 154, + 506, + 168 + ], + "score": 1.0, + "content": ", for instance the frames in a video, are", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 166, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 506, + 178 + ], + "score": 1.0, + "content": "processed by a neural encoder (Burgess et al., 2019; Greff et al., 2019; Goyal et al., 2019, 2020)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 173, + 507, + 192 + ], + "spans": [ + { + "bbox": [ + 104, + 173, + 170, + 192 + ], + "score": 1.0, + "content": "applied to each", + "type": "text" + }, + { + "bbox": [ + 171, + 177, + 181, + 187 + ], + "score": 0.87, + "content": "\\mathbf { \\boldsymbol { x } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 173, + 255, + 192 + ], + "score": 1.0, + "content": ", to obtain a set of", + "type": "text" + }, + { + "bbox": [ + 256, + 177, + 268, + 187 + ], + "score": 0.83, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 173, + 358, + 192 + ], + "score": 1.0, + "content": "entity representations", + "type": "text" + }, + { + "bbox": [ + 359, + 177, + 437, + 189 + ], + "score": 0.92, + "content": "\\{ V _ { 1 } ^ { t } , \\ldots , \\ldots \\stackrel { \\cdot } { , } V _ { M } ^ { t } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 173, + 507, + 192 + ], + "score": 1.0, + "content": ", one for each of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 121, + 201 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 122, + 188, + 133, + 198 + ], + "score": 0.82, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 186, + 506, + 201 + ], + "score": 1.0, + "content": "slots. These representations describe an entity and are updated based on both the previous", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 258, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 130, + 211 + ], + "score": 1.0, + "content": "state,", + "type": "text" + }, + { + "bbox": [ + 130, + 198, + 154, + 209 + ], + "score": 0.89, + "content": "V ^ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 198, + 243, + 211 + ], + "score": 1.0, + "content": "and the current input,", + "type": "text" + }, + { + "bbox": [ + 243, + 199, + 254, + 209 + ], + "score": 0.86, + "content": "\\mathbf { \\boldsymbol { x } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 198, + 258, + 211 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 214, + 506, + 274 + ], + "lines": [ + { + "bbox": [ + 105, + 214, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 175, + 228 + ], + "score": 1.0, + "content": "NPS consists of", + "type": "text" + }, + { + "bbox": [ + 175, + 215, + 185, + 225 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 214, + 292, + 228 + ], + "score": 1.0, + "content": "separately encoded rules,", + "type": "text" + }, + { + "bbox": [ + 292, + 215, + 364, + 227 + ], + "score": 0.93, + "content": "\\{ R _ { 1 } , R _ { 2 } , . . , R _ { N } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 214, + 506, + 228 + ], + "score": 1.0, + "content": ". Each rule consists of two compo-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 226, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 132, + 241 + ], + "score": 1.0, + "content": "nents,", + "type": "text" + }, + { + "bbox": [ + 133, + 226, + 210, + 240 + ], + "score": 0.93, + "content": "{ \\pmb R } _ { i } = ( \\vec { R _ { i } } , M L P _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 227, + 241, + 241 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 241, + 226, + 254, + 239 + ], + "score": 0.91, + "content": "\\vec { R _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 227, + 506, + 241 + ], + "score": 1.0, + "content": "is a learned rule embedding vector, which can be thought of as", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 239, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 357, + 251 + ], + "score": 1.0, + "content": "a template defining the condition for when a rule applies; and", + "type": "text" + }, + { + "bbox": [ + 357, + 239, + 385, + 250 + ], + "score": 0.91, + "content": "M L P _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 239, + 505, + 251 + ], + "score": 1.0, + "content": ", which determines the action", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 250, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 191, + 264 + ], + "score": 1.0, + "content": "taken by a rule. Both", + "type": "text" + }, + { + "bbox": [ + 191, + 250, + 204, + 263 + ], + "score": 0.9, + "content": "\\vec { R _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 252, + 293, + 264 + ], + "score": 1.0, + "content": "and the parameters of", + "type": "text" + }, + { + "bbox": [ + 293, + 252, + 321, + 263 + ], + "score": 0.91, + "content": "M L P _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 252, + 506, + 264 + ], + "score": 1.0, + "content": "are learned along with the other parameters of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 263, + 398, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 398, + 275 + ], + "score": 1.0, + "content": "the model using back-propagation on an objective optimized end-to-end.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 279, + 297, + 432 + ], + "lines": [ + { + "bbox": [ + 106, + 279, + 298, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 298, + 291 + ], + "score": 1.0, + "content": "In the general form of the model, each slot se-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 290, + 297, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 297, + 302 + ], + "score": 1.0, + "content": "lects a rule that will be applied to it to change its", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 301, + 297, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 297, + 312 + ], + "score": 1.0, + "content": "state. This can potentially be performed several", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 311, + 298, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 298, + 324 + ], + "score": 1.0, + "content": "times, with possibly different rules applied at", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 323, + 298, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 298, + 334 + ], + "score": 1.0, + "content": "each step. Rule selection is done using an atten-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 334, + 297, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 297, + 345 + ], + "score": 1.0, + "content": "tion mechanism described in detail below. Each", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 344, + 298, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 298, + 356 + ], + "score": 1.0, + "content": "rule specifies conditions and actions on a pair", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 356, + 298, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 298, + 367 + ], + "score": 1.0, + "content": "of slots. Therefore, while modifying the state", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 366, + 298, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 298, + 378 + ], + "score": 1.0, + "content": "of a slot using a rule, it can take the state of an-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 377, + 297, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 297, + 389 + ], + "score": 1.0, + "content": "other slot into account. The slot which is being", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 388, + 297, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 297, + 400 + ], + "score": 1.0, + "content": "modified is called the primary slot and other is", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 399, + 297, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 297, + 411 + ], + "score": 1.0, + "content": "called the contextual slot. The contextual slot", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 409, + 297, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 297, + 422 + ], + "score": 1.0, + "content": "is also selected using an attention mechanism", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 420, + 210, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 210, + 433 + ], + "score": 1.0, + "content": "described in detail below.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 23.5 + }, + { + "type": "image", + "bbox": [ + 354, + 276, + 455, + 369 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 354, + 276, + 455, + 369 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 354, + 276, + 455, + 369 + ], + "spans": [ + { + "bbox": [ + 354, + 276, + 455, + 369 + ], + "score": 0.958, + "type": "image", + "image_path": "7621aad51e5596db4e7dfe45bfb3fb1b7959f7deac83d4c49f5f51470e8d77b1.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 354, + 276, + 455, + 322.5 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 354, + 322.5, + 455, + 369.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 374, + 506, + 451 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 303, + 373, + 507, + 387 + ], + "spans": [ + { + "bbox": [ + 303, + 373, + 507, + 387 + ], + "score": 1.0, + "content": "Figure 2: Rule and slot combinatorics.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 384, + 507, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 507, + 397 + ], + "score": 1.0, + "content": "Condition-action rules specify how entities inter-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 303, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 303, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "act. Slots maintain the time-varying state of an", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 303, + 407, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 303, + 407, + 506, + 418 + ], + "score": 1.0, + "content": "entity. Every rule is matched to every pair of slots.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 303, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 303, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "Through key-value attention, a goodness of match", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 303, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 303, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "is determined, and a rule is selected along with its", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 303, + 439, + 371, + 452 + ], + "spans": [ + { + "bbox": [ + 303, + 439, + 371, + 452 + ], + "score": 1.0, + "content": "binding to slots.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35 + } + ], + "index": 30.25 + }, + { + "type": "title", + "bbox": [ + 107, + 443, + 253, + 455 + ], + "lines": [ + { + "bbox": [ + 104, + 442, + 254, + 456 + ], + "spans": [ + { + "bbox": [ + 104, + 442, + 254, + 456 + ], + "score": 1.0, + "content": "3.1 Computational Steps in NPS", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 106, + 457, + 505, + 513 + ], + "lines": [ + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "In this section, we give a detailed description of the rule selection and application procedure for the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 468, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 506, + 480 + ], + "score": 1.0, + "content": "slots. First, we will formalize the definitions of a few terms that we will use to explain our method.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 286, + 491 + ], + "score": 1.0, + "content": "We use the term primary slot to refer to slot", + "type": "text" + }, + { + "bbox": [ + 286, + 479, + 298, + 491 + ], + "score": 0.89, + "content": "V _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 479, + 441, + 491 + ], + "score": 1.0, + "content": "whose state gets modified by a rule", + "type": "text" + }, + { + "bbox": [ + 441, + 479, + 455, + 490 + ], + "score": 0.89, + "content": "\\scriptstyle { \\mathbf { } } R _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 479, + 505, + 491 + ], + "score": 1.0, + "content": ". We use the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 490, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 266, + 502 + ], + "score": 1.0, + "content": "term contextual slot to refer to the slot", + "type": "text" + }, + { + "bbox": [ + 266, + 490, + 278, + 501 + ], + "score": 0.88, + "content": "V _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 490, + 330, + 502 + ], + "score": 1.0, + "content": "that the rule", + "type": "text" + }, + { + "bbox": [ + 330, + 490, + 344, + 501 + ], + "score": 0.91, + "content": "\\scriptstyle { R _ { r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 490, + 505, + 502 + ], + "score": 1.0, + "content": "takes into account while modifying the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 501, + 220, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 204, + 515 + ], + "score": 1.0, + "content": "state of the primary slot", + "type": "text" + }, + { + "bbox": [ + 204, + 501, + 216, + 513 + ], + "score": 0.88, + "content": "V _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 501, + 220, + 515 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 516, + 503, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 253, + 529 + ], + "score": 1.0, + "content": "Notation. We consider a set of", + "type": "text" + }, + { + "bbox": [ + 253, + 517, + 265, + 527 + ], + "score": 0.79, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 516, + 293, + 529 + ], + "score": 1.0, + "content": "rules", + "type": "text" + }, + { + "bbox": [ + 294, + 517, + 374, + 529 + ], + "score": 0.92, + "content": "\\{ R _ { 1 } , R _ { 2 } , \\ldots , R _ { N } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 516, + 436, + 529 + ], + "score": 1.0, + "content": "and a set of", + "type": "text" + }, + { + "bbox": [ + 436, + 518, + 446, + 527 + ], + "score": 0.81, + "content": "_ { \\mathbf { T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "input frames", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 107, + 524, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 107, + 527, + 179, + 540 + ], + "score": 0.92, + "content": "\\{ \\pmb { x } ^ { 1 } , \\pmb { x } ^ { 2 } , \\dots , \\pmb { x } ^ { T } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 524, + 237, + 543 + ], + "score": 1.0, + "content": ". Each frame", + "type": "text" + }, + { + "bbox": [ + 238, + 528, + 249, + 538 + ], + "score": 0.87, + "content": "\\mathbf { \\boldsymbol { x } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 524, + 353, + 543 + ], + "score": 1.0, + "content": "is encoded into a set of", + "type": "text" + }, + { + "bbox": [ + 354, + 529, + 367, + 538 + ], + "score": 0.76, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 524, + 392, + 543 + ], + "score": 1.0, + "content": "slots", + "type": "text" + }, + { + "bbox": [ + 392, + 528, + 470, + 541 + ], + "score": 0.93, + "content": "\\{ V _ { 1 } ^ { t } , V _ { 2 } ^ { t } , \\ldots , \\bar { V _ { M } ^ { t } } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 524, + 506, + 543 + ], + "score": 1.0, + "content": ". In the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 538, + 353, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 288, + 552 + ], + "score": 1.0, + "content": "following discussion, we omit the index over", + "type": "text" + }, + { + "bbox": [ + 288, + 540, + 293, + 549 + ], + "score": 0.69, + "content": "\\pmb { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 538, + 353, + 552 + ], + "score": 1.0, + "content": "for simplicity.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 106, + 555, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 126, + 568 + ], + "score": 1.0, + "content": "Step", + "type": "text" + }, + { + "bbox": [ + 126, + 556, + 133, + 565 + ], + "score": 0.29, + "content": "\\imath .", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 555, + 352, + 568 + ], + "score": 1.0, + "content": "is external to NPS and involves parsing an input image,", + "type": "text" + }, + { + "bbox": [ + 353, + 555, + 363, + 565 + ], + "score": 0.86, + "content": "\\mathbf { \\boldsymbol { x } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 555, + 506, + 568 + ], + "score": 1.0, + "content": ", into slot-based entities conditioned", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 567, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 578 + ], + "score": 1.0, + "content": "on the previous state of the slot-based entities. Any of the methods proposed in the literature to obtain", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 577, + 507, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 507, + 590 + ], + "score": 1.0, + "content": "a slot-wise representation of entities can be used (Burgess et al., 2019; Greff et al., 2019; Goyal et al.,", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 587, + 462, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 462, + 600 + ], + "score": 1.0, + "content": "2019, 2020). The next three steps constitute the rule selection and application procedure.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 49.5 + }, + { + "type": "text", + "bbox": [ + 106, + 604, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 225, + 617 + ], + "score": 1.0, + "content": "Step 2. For each primary slot", + "type": "text" + }, + { + "bbox": [ + 225, + 605, + 237, + 617 + ], + "score": 0.89, + "content": "V _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 603, + 317, + 617 + ], + "score": 1.0, + "content": ", we attend to a rule", + "type": "text" + }, + { + "bbox": [ + 317, + 605, + 331, + 615 + ], + "score": 0.89, + "content": "\\scriptstyle { R _ { r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 603, + 505, + 617 + ], + "score": 1.0, + "content": "to be applied. Here, the queries come from", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 104, + 615, + 503, + 632 + ], + "spans": [ + { + "bbox": [ + 104, + 615, + 177, + 632 + ], + "score": 1.0, + "content": "the primary slot:", + "type": "text" + }, + { + "bbox": [ + 177, + 618, + 228, + 631 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\pmb q _ { p } = \\pmb V _ { p } W ^ { q } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 615, + 372, + 632 + ], + "score": 1.0, + "content": ", and the keys come from the rules:", + "type": "text" + }, + { + "bbox": [ + 373, + 616, + 503, + 630 + ], + "score": 0.85, + "content": "\\pmb { k } _ { i } = \\vec { R _ { i } } W ^ { k } \\quad \\forall i \\in \\{ 1 , \\dots , N \\}", + "type": "inline_equation" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "The rule is selected using a straight-through Gumbel softmax (Jang et al., 2016) to achieve a learnable", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 168, + 653 + ], + "score": 1.0, + "content": "hard decision:", + "type": "text" + }, + { + "bbox": [ + 169, + 640, + 272, + 652 + ], + "score": 0.92, + "content": "r = \\mathrm { a r g m a x } _ { i } ( q _ { p } k _ { i } + \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 639, + 305, + 653 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 306, + 640, + 384, + 651 + ], + "score": 0.91, + "content": "\\gamma \\sim \\mathrm { G u m b e l } ( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 639, + 505, + 653 + ], + "score": 1.0, + "content": ". This competition is a noisy", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 650, + 413, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 413, + 663 + ], + "score": 1.0, + "content": "version of rule matching and prioritization in traditional production systems.", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 54 + }, + { + "type": "text", + "bbox": [ + 106, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 231, + 679 + ], + "score": 1.0, + "content": "Step 3. For a given primary slot", + "type": "text" + }, + { + "bbox": [ + 232, + 667, + 244, + 679 + ], + "score": 0.87, + "content": "V _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 667, + 312, + 679 + ], + "score": 1.0, + "content": "and selected rule", + "type": "text" + }, + { + "bbox": [ + 312, + 668, + 326, + 678 + ], + "score": 0.86, + "content": "\\scriptstyle { R _ { r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 667, + 395, + 679 + ], + "score": 1.0, + "content": ", a contextual slot", + "type": "text" + }, + { + "bbox": [ + 396, + 668, + 407, + 678 + ], + "score": 0.87, + "content": "V _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 667, + 506, + 679 + ], + "score": 1.0, + "content": "is selected using another", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 678, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 415, + 691 + ], + "score": 1.0, + "content": "attention mechanism. In this case the query comes from the primary slot:", + "type": "text" + }, + { + "bbox": [ + 415, + 678, + 467, + 690 + ], + "score": 0.92, + "content": "q _ { p } = V _ { p } W ^ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 678, + 506, + 691 + ], + "score": 1.0, + "content": ", and the", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 200, + 702 + ], + "score": 1.0, + "content": "keys from all the slots:", + "type": "text" + }, + { + "bbox": [ + 200, + 689, + 333, + 702 + ], + "score": 0.9, + "content": "\\pmb { k } _ { j } = V _ { j } W ^ { q } \\quad \\forall j \\in \\{ 1 , \\dots , M \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 688, + 506, + 702 + ], + "score": 1.0, + "content": ". The selection takes place using a straight-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 699, + 502, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 287, + 714 + ], + "score": 1.0, + "content": "through Gumbel softmax similar to step 2:", + "type": "text" + }, + { + "bbox": [ + 287, + 700, + 391, + 713 + ], + "score": 0.87, + "content": "c = \\mathrm { a r g m a x } _ { j } ( q _ { p } k _ { j } + \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 699, + 424, + 714 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 424, + 700, + 502, + 712 + ], + "score": 0.9, + "content": "\\gamma \\sim \\mathrm { G u m b e l } ( 0 , 1 )", + "type": "inline_equation" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 709, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 725 + ], + "score": 1.0, + "content": "Note that each rule application is sparse since it takes into account only 1 contextual slot for modifying", + "type": "text" + } + ], + "index": 61 + } + ], + "index": 59 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 11, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 504, + 95 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 104, + 72, + 505, + 97 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 106, + 105, + 385, + 119 + ], + "lines": [ + { + "bbox": [ + 104, + 103, + 386, + 122 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 386, + 122 + ], + "score": 1.0, + "content": "3 Neural Production System: Slots and Sparse Rules", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 122, + 505, + 210 + ], + "lines": [ + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "score": 1.0, + "content": "The Neural Production System (NPS), illustrated in Figure 2, provides an architectural backbone", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 133, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 133, + 505, + 145 + ], + "score": 1.0, + "content": "that supports the detection and inference of entity (object) representations in an input sequence, and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "the underlying rules which govern the interactions between these entities in time and space. The", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 252, + 168 + ], + "score": 1.0, + "content": "input sequence indexed by time step", + "type": "text" + }, + { + "bbox": [ + 252, + 156, + 257, + 165 + ], + "score": 0.58, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 154, + 261, + 168 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 261, + 155, + 350, + 167 + ], + "score": 0.9, + "content": "\\{ \\pmb { x } ^ { 1 } , \\dotsc , \\pmb { x } ^ { t } , \\dotsc , \\pmb { x } ^ { T } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 154, + 506, + 168 + ], + "score": 1.0, + "content": ", for instance the frames in a video, are", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 166, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 506, + 178 + ], + "score": 1.0, + "content": "processed by a neural encoder (Burgess et al., 2019; Greff et al., 2019; Goyal et al., 2019, 2020)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 173, + 507, + 192 + ], + "spans": [ + { + "bbox": [ + 104, + 173, + 170, + 192 + ], + "score": 1.0, + "content": "applied to each", + "type": "text" + }, + { + "bbox": [ + 171, + 177, + 181, + 187 + ], + "score": 0.87, + "content": "\\mathbf { \\boldsymbol { x } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 173, + 255, + 192 + ], + "score": 1.0, + "content": ", to obtain a set of", + "type": "text" + }, + { + "bbox": [ + 256, + 177, + 268, + 187 + ], + "score": 0.83, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 173, + 358, + 192 + ], + "score": 1.0, + "content": "entity representations", + "type": "text" + }, + { + "bbox": [ + 359, + 177, + 437, + 189 + ], + "score": 0.92, + "content": "\\{ V _ { 1 } ^ { t } , \\ldots , \\ldots \\stackrel { \\cdot } { , } V _ { M } ^ { t } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 173, + 507, + 192 + ], + "score": 1.0, + "content": ", one for each of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 121, + 201 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 122, + 188, + 133, + 198 + ], + "score": 0.82, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 186, + 506, + 201 + ], + "score": 1.0, + "content": "slots. These representations describe an entity and are updated based on both the previous", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 258, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 130, + 211 + ], + "score": 1.0, + "content": "state,", + "type": "text" + }, + { + "bbox": [ + 130, + 198, + 154, + 209 + ], + "score": 0.89, + "content": "V ^ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 198, + 243, + 211 + ], + "score": 1.0, + "content": "and the current input,", + "type": "text" + }, + { + "bbox": [ + 243, + 199, + 254, + 209 + ], + "score": 0.86, + "content": "\\mathbf { \\boldsymbol { x } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 198, + 258, + 211 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5, + "bbox_fs": [ + 104, + 121, + 507, + 211 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 214, + 506, + 274 + ], + "lines": [ + { + "bbox": [ + 105, + 214, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 175, + 228 + ], + "score": 1.0, + "content": "NPS consists of", + "type": "text" + }, + { + "bbox": [ + 175, + 215, + 185, + 225 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 214, + 292, + 228 + ], + "score": 1.0, + "content": "separately encoded rules,", + "type": "text" + }, + { + "bbox": [ + 292, + 215, + 364, + 227 + ], + "score": 0.93, + "content": "\\{ R _ { 1 } , R _ { 2 } , . . , R _ { N } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 214, + 506, + 228 + ], + "score": 1.0, + "content": ". Each rule consists of two compo-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 226, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 132, + 241 + ], + "score": 1.0, + "content": "nents,", + "type": "text" + }, + { + "bbox": [ + 133, + 226, + 210, + 240 + ], + "score": 0.93, + "content": "{ \\pmb R } _ { i } = ( \\vec { R _ { i } } , M L P _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 227, + 241, + 241 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 241, + 226, + 254, + 239 + ], + "score": 0.91, + "content": "\\vec { R _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 227, + 506, + 241 + ], + "score": 1.0, + "content": "is a learned rule embedding vector, which can be thought of as", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 239, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 357, + 251 + ], + "score": 1.0, + "content": "a template defining the condition for when a rule applies; and", + "type": "text" + }, + { + "bbox": [ + 357, + 239, + 385, + 250 + ], + "score": 0.91, + "content": "M L P _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 239, + 505, + 251 + ], + "score": 1.0, + "content": ", which determines the action", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 250, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 191, + 264 + ], + "score": 1.0, + "content": "taken by a rule. Both", + "type": "text" + }, + { + "bbox": [ + 191, + 250, + 204, + 263 + ], + "score": 0.9, + "content": "\\vec { R _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 252, + 293, + 264 + ], + "score": 1.0, + "content": "and the parameters of", + "type": "text" + }, + { + "bbox": [ + 293, + 252, + 321, + 263 + ], + "score": 0.91, + "content": "M L P _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 252, + 506, + 264 + ], + "score": 1.0, + "content": "are learned along with the other parameters of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 263, + 398, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 398, + 275 + ], + "score": 1.0, + "content": "the model using back-propagation on an objective optimized end-to-end.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 214, + 506, + 275 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 279, + 297, + 432 + ], + "lines": [ + { + "bbox": [ + 106, + 279, + 298, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 298, + 291 + ], + "score": 1.0, + "content": "In the general form of the model, each slot se-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 290, + 297, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 297, + 302 + ], + "score": 1.0, + "content": "lects a rule that will be applied to it to change its", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 301, + 297, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 297, + 312 + ], + "score": 1.0, + "content": "state. This can potentially be performed several", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 311, + 298, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 298, + 324 + ], + "score": 1.0, + "content": "times, with possibly different rules applied at", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 323, + 298, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 298, + 334 + ], + "score": 1.0, + "content": "each step. Rule selection is done using an atten-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 334, + 297, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 297, + 345 + ], + "score": 1.0, + "content": "tion mechanism described in detail below. Each", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 344, + 298, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 298, + 356 + ], + "score": 1.0, + "content": "rule specifies conditions and actions on a pair", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 356, + 298, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 298, + 367 + ], + "score": 1.0, + "content": "of slots. Therefore, while modifying the state", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 366, + 298, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 298, + 378 + ], + "score": 1.0, + "content": "of a slot using a rule, it can take the state of an-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 377, + 297, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 297, + 389 + ], + "score": 1.0, + "content": "other slot into account. The slot which is being", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 388, + 297, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 297, + 400 + ], + "score": 1.0, + "content": "modified is called the primary slot and other is", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 399, + 297, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 297, + 411 + ], + "score": 1.0, + "content": "called the contextual slot. The contextual slot", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 409, + 297, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 297, + 422 + ], + "score": 1.0, + "content": "is also selected using an attention mechanism", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 420, + 210, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 210, + 433 + ], + "score": 1.0, + "content": "described in detail below.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 279, + 298, + 433 + ] + }, + { + "type": "image", + "bbox": [ + 354, + 276, + 455, + 369 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 354, + 276, + 455, + 369 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 354, + 276, + 455, + 369 + ], + "spans": [ + { + "bbox": [ + 354, + 276, + 455, + 369 + ], + "score": 0.958, + "type": "image", + "image_path": "7621aad51e5596db4e7dfe45bfb3fb1b7959f7deac83d4c49f5f51470e8d77b1.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 354, + 276, + 455, + 322.5 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 354, + 322.5, + 455, + 369.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 374, + 506, + 451 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 303, + 373, + 507, + 387 + ], + "spans": [ + { + "bbox": [ + 303, + 373, + 507, + 387 + ], + "score": 1.0, + "content": "Figure 2: Rule and slot combinatorics.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 384, + 507, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 507, + 397 + ], + "score": 1.0, + "content": "Condition-action rules specify how entities inter-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 303, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 303, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "act. Slots maintain the time-varying state of an", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 303, + 407, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 303, + 407, + 506, + 418 + ], + "score": 1.0, + "content": "entity. Every rule is matched to every pair of slots.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 303, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 303, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "Through key-value attention, a goodness of match", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 303, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 303, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "is determined, and a rule is selected along with its", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 303, + 439, + 371, + 452 + ], + "spans": [ + { + "bbox": [ + 303, + 439, + 371, + 452 + ], + "score": 1.0, + "content": "binding to slots.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35 + } + ], + "index": 30.25 + }, + { + "type": "title", + "bbox": [ + 107, + 443, + 253, + 455 + ], + "lines": [ + { + "bbox": [ + 104, + 442, + 254, + 456 + ], + "spans": [ + { + "bbox": [ + 104, + 442, + 254, + 456 + ], + "score": 1.0, + "content": "3.1 Computational Steps in NPS", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 106, + 457, + 505, + 513 + ], + "lines": [ + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "In this section, we give a detailed description of the rule selection and application procedure for the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 468, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 506, + 480 + ], + "score": 1.0, + "content": "slots. First, we will formalize the definitions of a few terms that we will use to explain our method.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 286, + 491 + ], + "score": 1.0, + "content": "We use the term primary slot to refer to slot", + "type": "text" + }, + { + "bbox": [ + 286, + 479, + 298, + 491 + ], + "score": 0.89, + "content": "V _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 479, + 441, + 491 + ], + "score": 1.0, + "content": "whose state gets modified by a rule", + "type": "text" + }, + { + "bbox": [ + 441, + 479, + 455, + 490 + ], + "score": 0.89, + "content": "\\scriptstyle { \\mathbf { } } R _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 479, + 505, + 491 + ], + "score": 1.0, + "content": ". We use the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 490, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 266, + 502 + ], + "score": 1.0, + "content": "term contextual slot to refer to the slot", + "type": "text" + }, + { + "bbox": [ + 266, + 490, + 278, + 501 + ], + "score": 0.88, + "content": "V _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 490, + 330, + 502 + ], + "score": 1.0, + "content": "that the rule", + "type": "text" + }, + { + "bbox": [ + 330, + 490, + 344, + 501 + ], + "score": 0.91, + "content": "\\scriptstyle { R _ { r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 490, + 505, + 502 + ], + "score": 1.0, + "content": "takes into account while modifying the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 501, + 220, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 204, + 515 + ], + "score": 1.0, + "content": "state of the primary slot", + "type": "text" + }, + { + "bbox": [ + 204, + 501, + 216, + 513 + ], + "score": 0.88, + "content": "V _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 501, + 220, + 515 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 457, + 506, + 515 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 516, + 503, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 253, + 529 + ], + "score": 1.0, + "content": "Notation. We consider a set of", + "type": "text" + }, + { + "bbox": [ + 253, + 517, + 265, + 527 + ], + "score": 0.79, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 516, + 293, + 529 + ], + "score": 1.0, + "content": "rules", + "type": "text" + }, + { + "bbox": [ + 294, + 517, + 374, + 529 + ], + "score": 0.92, + "content": "\\{ R _ { 1 } , R _ { 2 } , \\ldots , R _ { N } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 516, + 436, + 529 + ], + "score": 1.0, + "content": "and a set of", + "type": "text" + }, + { + "bbox": [ + 436, + 518, + 446, + 527 + ], + "score": 0.81, + "content": "_ { \\mathbf { T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "input frames", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 107, + 524, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 107, + 527, + 179, + 540 + ], + "score": 0.92, + "content": "\\{ \\pmb { x } ^ { 1 } , \\pmb { x } ^ { 2 } , \\dots , \\pmb { x } ^ { T } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 524, + 237, + 543 + ], + "score": 1.0, + "content": ". Each frame", + "type": "text" + }, + { + "bbox": [ + 238, + 528, + 249, + 538 + ], + "score": 0.87, + "content": "\\mathbf { \\boldsymbol { x } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 524, + 353, + 543 + ], + "score": 1.0, + "content": "is encoded into a set of", + "type": "text" + }, + { + "bbox": [ + 354, + 529, + 367, + 538 + ], + "score": 0.76, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 524, + 392, + 543 + ], + "score": 1.0, + "content": "slots", + "type": "text" + }, + { + "bbox": [ + 392, + 528, + 470, + 541 + ], + "score": 0.93, + "content": "\\{ V _ { 1 } ^ { t } , V _ { 2 } ^ { t } , \\ldots , \\bar { V _ { M } ^ { t } } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 524, + 506, + 543 + ], + "score": 1.0, + "content": ". In the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 538, + 353, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 288, + 552 + ], + "score": 1.0, + "content": "following discussion, we omit the index over", + "type": "text" + }, + { + "bbox": [ + 288, + 540, + 293, + 549 + ], + "score": 0.69, + "content": "\\pmb { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 538, + 353, + 552 + ], + "score": 1.0, + "content": "for simplicity.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 516, + 506, + 552 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 555, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 126, + 568 + ], + "score": 1.0, + "content": "Step", + "type": "text" + }, + { + "bbox": [ + 126, + 556, + 133, + 565 + ], + "score": 0.29, + "content": "\\imath .", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 555, + 352, + 568 + ], + "score": 1.0, + "content": "is external to NPS and involves parsing an input image,", + "type": "text" + }, + { + "bbox": [ + 353, + 555, + 363, + 565 + ], + "score": 0.86, + "content": "\\mathbf { \\boldsymbol { x } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 555, + 506, + 568 + ], + "score": 1.0, + "content": ", into slot-based entities conditioned", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 567, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 578 + ], + "score": 1.0, + "content": "on the previous state of the slot-based entities. Any of the methods proposed in the literature to obtain", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 577, + 507, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 507, + 590 + ], + "score": 1.0, + "content": "a slot-wise representation of entities can be used (Burgess et al., 2019; Greff et al., 2019; Goyal et al.,", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 587, + 462, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 462, + 600 + ], + "score": 1.0, + "content": "2019, 2020). The next three steps constitute the rule selection and application procedure.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 49.5, + "bbox_fs": [ + 105, + 555, + 507, + 600 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 604, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 225, + 617 + ], + "score": 1.0, + "content": "Step 2. For each primary slot", + "type": "text" + }, + { + "bbox": [ + 225, + 605, + 237, + 617 + ], + "score": 0.89, + "content": "V _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 603, + 317, + 617 + ], + "score": 1.0, + "content": ", we attend to a rule", + "type": "text" + }, + { + "bbox": [ + 317, + 605, + 331, + 615 + ], + "score": 0.89, + "content": "\\scriptstyle { R _ { r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 603, + 505, + 617 + ], + "score": 1.0, + "content": "to be applied. Here, the queries come from", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 104, + 615, + 503, + 632 + ], + "spans": [ + { + "bbox": [ + 104, + 615, + 177, + 632 + ], + "score": 1.0, + "content": "the primary slot:", + "type": "text" + }, + { + "bbox": [ + 177, + 618, + 228, + 631 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\pmb q _ { p } = \\pmb V _ { p } W ^ { q } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 615, + 372, + 632 + ], + "score": 1.0, + "content": ", and the keys come from the rules:", + "type": "text" + }, + { + "bbox": [ + 373, + 616, + 503, + 630 + ], + "score": 0.85, + "content": "\\pmb { k } _ { i } = \\vec { R _ { i } } W ^ { k } \\quad \\forall i \\in \\{ 1 , \\dots , N \\}", + "type": "inline_equation" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "The rule is selected using a straight-through Gumbel softmax (Jang et al., 2016) to achieve a learnable", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 168, + 653 + ], + "score": 1.0, + "content": "hard decision:", + "type": "text" + }, + { + "bbox": [ + 169, + 640, + 272, + 652 + ], + "score": 0.92, + "content": "r = \\mathrm { a r g m a x } _ { i } ( q _ { p } k _ { i } + \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 639, + 305, + 653 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 306, + 640, + 384, + 651 + ], + "score": 0.91, + "content": "\\gamma \\sim \\mathrm { G u m b e l } ( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 639, + 505, + 653 + ], + "score": 1.0, + "content": ". This competition is a noisy", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 650, + 413, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 413, + 663 + ], + "score": 1.0, + "content": "version of rule matching and prioritization in traditional production systems.", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 54, + "bbox_fs": [ + 104, + 603, + 505, + 663 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 231, + 679 + ], + "score": 1.0, + "content": "Step 3. For a given primary slot", + "type": "text" + }, + { + "bbox": [ + 232, + 667, + 244, + 679 + ], + "score": 0.87, + "content": "V _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 667, + 312, + 679 + ], + "score": 1.0, + "content": "and selected rule", + "type": "text" + }, + { + "bbox": [ + 312, + 668, + 326, + 678 + ], + "score": 0.86, + "content": "\\scriptstyle { R _ { r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 667, + 395, + 679 + ], + "score": 1.0, + "content": ", a contextual slot", + "type": "text" + }, + { + "bbox": [ + 396, + 668, + 407, + 678 + ], + "score": 0.87, + "content": "V _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 667, + 506, + 679 + ], + "score": 1.0, + "content": "is selected using another", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 678, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 415, + 691 + ], + "score": 1.0, + "content": "attention mechanism. In this case the query comes from the primary slot:", + "type": "text" + }, + { + "bbox": [ + 415, + 678, + 467, + 690 + ], + "score": 0.92, + "content": "q _ { p } = V _ { p } W ^ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 678, + 506, + 691 + ], + "score": 1.0, + "content": ", and the", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 200, + 702 + ], + "score": 1.0, + "content": "keys from all the slots:", + "type": "text" + }, + { + "bbox": [ + 200, + 689, + 333, + 702 + ], + "score": 0.9, + "content": "\\pmb { k } _ { j } = V _ { j } W ^ { q } \\quad \\forall j \\in \\{ 1 , \\dots , M \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 688, + 506, + 702 + ], + "score": 1.0, + "content": ". The selection takes place using a straight-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 699, + 502, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 287, + 714 + ], + "score": 1.0, + "content": "through Gumbel softmax similar to step 2:", + "type": "text" + }, + { + "bbox": [ + 287, + 700, + 391, + 713 + ], + "score": 0.87, + "content": "c = \\mathrm { a r g m a x } _ { j } ( q _ { p } k _ { j } + \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 699, + 424, + 714 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 424, + 700, + 502, + 712 + ], + "score": 0.9, + "content": "\\gamma \\sim \\mathrm { G u m b e l } ( 0 , 1 )", + "type": "inline_equation" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 709, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 725 + ], + "score": 1.0, + "content": "Note that each rule application is sparse since it takes into account only 1 contextual slot for modifying", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 104, + 71, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 104, + 71, + 506, + 86 + ], + "score": 1.0, + "content": "a primary slot, while other methods like GNNs take into account all slots for modifying a primary", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 82, + 127, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 82, + 127, + 96 + ], + "score": 1.0, + "content": "slot.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 59, + "bbox_fs": [ + 105, + 667, + 506, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 504, + 95 + ], + "lines": [ + { + "bbox": [ + 104, + 71, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 104, + 71, + 506, + 86 + ], + "score": 1.0, + "content": "a primary slot, while other methods like GNNs take into account all slots for modifying a primary", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 82, + 127, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 82, + 127, + 96 + ], + "score": 1.0, + "content": "slot.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 505, + 144 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 506, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 277, + 113 + ], + "score": 1.0, + "content": "Step 4. Rule Application: the selected rule", + "type": "text" + }, + { + "bbox": [ + 278, + 101, + 292, + 111 + ], + "score": 0.88, + "content": "\\scriptstyle { R _ { r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 100, + 407, + 113 + ], + "score": 1.0, + "content": "is applied to the primary slot", + "type": "text" + }, + { + "bbox": [ + 407, + 101, + 419, + 111 + ], + "score": 0.84, + "content": "V _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 100, + 506, + 113 + ], + "score": 1.0, + "content": "based on the rule and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 111, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 400, + 124 + ], + "score": 1.0, + "content": "the current contents of the primary and contextual slots. The rule-specific", + "type": "text" + }, + { + "bbox": [ + 401, + 111, + 430, + 122 + ], + "score": 0.9, + "content": "M L P _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 111, + 505, + 124 + ], + "score": 1.0, + "content": ", takes as input the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 122, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 404, + 135 + ], + "score": 1.0, + "content": "concatenated representation of the state of the primary and contextual slots,", + "type": "text" + }, + { + "bbox": [ + 405, + 122, + 417, + 134 + ], + "score": 0.88, + "content": "V _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 122, + 435, + 135 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 435, + 122, + 447, + 133 + ], + "score": 0.88, + "content": "V _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 122, + 505, + 135 + ], + "score": 1.0, + "content": ", and produces", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 133, + 475, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 378, + 146 + ], + "score": 1.0, + "content": "an output, which is then used to change the state of the primary slot", + "type": "text" + }, + { + "bbox": [ + 379, + 133, + 390, + 145 + ], + "score": 0.89, + "content": "V _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 133, + 475, + 146 + ], + "score": 1.0, + "content": "by residual addition.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 106, + 150, + 376, + 162 + ], + "lines": [ + { + "bbox": [ + 106, + 150, + 375, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 375, + 164 + ], + "score": 1.0, + "content": "3.2 Rule Application: Sequential vs Parallel Rule Application", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 163, + 506, + 240 + ], + "lines": [ + { + "bbox": [ + 106, + 163, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 505, + 174 + ], + "score": 1.0, + "content": "In the previous section, we have described how each rule application only considers another contextual", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 174, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 506, + 186 + ], + "score": 1.0, + "content": "slot for the given primary slot i.e., contextual sparsity. We can also consider application sparsity,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 183, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 104, + 183, + 506, + 198 + ], + "score": 1.0, + "content": "wherein we use the rules to update the states of only a subset of the slots. In this scenario, only", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 195, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 506, + 209 + ], + "score": 1.0, + "content": "the selected slots would be primary slots. This setting will be helpful when there is an entity in an", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 206, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 506, + 219 + ], + "score": 1.0, + "content": "environment that is stationary, or it is following its own default dynamics unaffected by other entities.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "score": 1.0, + "content": "Therefore, it does not need to consider other entities to update its state. We explore two scenarios for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 228, + 225, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 225, + 241 + ], + "score": 1.0, + "content": "enabling application sparsity.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 244, + 297, + 311 + ], + "lines": [ + { + "bbox": [ + 106, + 244, + 297, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 264, + 255 + ], + "score": 1.0, + "content": "Parallel Rule Application. Each of the", + "type": "text" + }, + { + "bbox": [ + 264, + 245, + 276, + 254 + ], + "score": 0.56, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 244, + 297, + 255 + ], + "score": 1.0, + "content": "slots", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 255, + 297, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 297, + 267 + ], + "score": 1.0, + "content": "selects a rule to potentially change its state. To", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 266, + 297, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 297, + 277 + ], + "score": 1.0, + "content": "enable sparse changes, we provide an extra Null", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 277, + 298, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 241, + 289 + ], + "score": 1.0, + "content": "Rule in addition to the available", + "type": "text" + }, + { + "bbox": [ + 242, + 278, + 252, + 287 + ], + "score": 0.74, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 277, + 298, + 289 + ], + "score": 1.0, + "content": "rules. If a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 288, + 297, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 297, + 300 + ], + "score": 1.0, + "content": "slot picks the null rule in step 2 of the above", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 299, + 258, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 258, + 312 + ], + "score": 1.0, + "content": "procedure, we do not update its state.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 316, + 297, + 371 + ], + "lines": [ + { + "bbox": [ + 106, + 314, + 299, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 299, + 329 + ], + "score": 1.0, + "content": "Sequential Rule Application. In this setting,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 327, + 298, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 298, + 338 + ], + "score": 1.0, + "content": "only one slot gets updated in each rule applica-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 338, + 298, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 298, + 349 + ], + "score": 1.0, + "content": "tion step. Therefore, only one slot is selected", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 348, + 297, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 297, + 360 + ], + "score": 1.0, + "content": "as the primary slot. This can be facilitated by", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 359, + 298, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 298, + 373 + ], + "score": 1.0, + "content": "modifying step 2 above to select one {primary", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 310, + 247, + 498, + 338 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 310, + 247, + 498, + 338 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 310, + 247, + 498, + 338 + ], + "spans": [ + { + "bbox": [ + 310, + 247, + 498, + 338 + ], + "score": 0.968, + "type": "image", + "image_path": "d2db80f1c75eadd39bbfb8c94bc9f20aca13ed711f1d46b4857e95be0ac7103f.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 310, + 247, + 498, + 260.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 310, + 260.0, + 498, + 273.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 310, + 273.0, + 498, + 286.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 310, + 286.0, + 498, + 299.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 310, + 299.0, + 498, + 312.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 310, + 312.0, + 498, + 325.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 310, + 325.0, + 498, + 338.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 344, + 504, + 367 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 343, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 343, + 505, + 357 + ], + "score": 1.0, + "content": "Figure 3: This figure demonstrates the sequential", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 356, + 420, + 367 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 420, + 367 + ], + "score": 1.0, + "content": "and parallel rule application.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + } + ], + "index": 30.25 + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 428 + ], + "lines": [ + { + "bbox": [ + 106, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 194, + 383 + ], + "score": 1.0, + "content": "slot, rule} pair among", + "type": "text" + }, + { + "bbox": [ + 195, + 370, + 219, + 381 + ], + "score": 0.5, + "content": "_ { N M }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 370, + 423, + 383 + ], + "score": 1.0, + "content": "{rule, slot} pairs. The queries come from each slot:", + "type": "text" + }, + { + "bbox": [ + 424, + 370, + 505, + 383 + ], + "score": 0.77, + "content": "\\pmb { q } _ { j } = \\pmb { V } _ { j } \\pmb { W } ^ { q } \\quad \\forall j \\in", + "type": "inline_equation" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 380, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 107, + 382, + 157, + 394 + ], + "score": 0.91, + "content": "\\{ 1 , \\ldots , M \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 380, + 284, + 397 + ], + "score": 1.0, + "content": ", the keys come from the rules:", + "type": "text" + }, + { + "bbox": [ + 284, + 381, + 415, + 394 + ], + "score": 0.81, + "content": "\\pmb { k } _ { i } = R _ { i } W ^ { k } \\quad \\forall i \\in \\{ 1 , \\dots , N \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 380, + 506, + 397 + ], + "score": 1.0, + "content": ". The straight-through", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 392, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 331, + 408 + ], + "score": 1.0, + "content": "Gumbel softmax selects one (primary slot, rule) pair:", + "type": "text" + }, + { + "bbox": [ + 332, + 394, + 453, + 406 + ], + "score": 0.88, + "content": "p , r = \\mathrm { a r g m a x } _ { i , j } ( q _ { p } k _ { i } + \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 392, + 484, + 408 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 485, + 395, + 505, + 405 + ], + "score": 0.86, + "content": "\\gamma \\sim", + "type": "inline_equation" + } + ], + "index": 36 + }, + { + "bbox": [ + 107, + 403, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 107, + 405, + 164, + 417 + ], + "score": 0.42, + "content": "\\mathrm { G u m b e l } ( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 403, + 506, + 419 + ], + "score": 1.0, + "content": ". In the sequential regime, we allow the rule application procedure (step 2, 3, 4 above)", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 416, + 472, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 285, + 430 + ], + "score": 1.0, + "content": "to be performed multiple times iteratively in", + "type": "text" + }, + { + "bbox": [ + 285, + 417, + 296, + 426 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 416, + 462, + 430 + ], + "score": 1.0, + "content": "rule application stages for each time-step", + "type": "text" + }, + { + "bbox": [ + 463, + 418, + 468, + 426 + ], + "score": 0.69, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 416, + 472, + 430 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 433, + 504, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "A pictorial demonstration of both rule application regimes can be found in Figure 3. We provide", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 444, + 395, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 395, + 457 + ], + "score": 1.0, + "content": "detailed algorithms for the sequential and parallel regimes in Appendix.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "title", + "bbox": [ + 107, + 464, + 191, + 477 + ], + "lines": [ + { + "bbox": [ + 104, + 462, + 193, + 480 + ], + "spans": [ + { + "bbox": [ + 104, + 462, + 193, + 480 + ], + "score": 1.0, + "content": "4 Experiments", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 487, + 338, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 339, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 339, + 498 + ], + "score": 1.0, + "content": "We demonstrate the effectiveness of NPS on multiple tasks", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 498, + 339, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 339, + 510 + ], + "score": 1.0, + "content": "and compare to a comprehensive set of baselines. To show", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 507, + 339, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 339, + 522 + ], + "score": 1.0, + "content": "that NPS can learn intuitive rules from the data generating", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 520, + 339, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 339, + 532 + ], + "score": 1.0, + "content": "distribution, we design a couple of simple toy experiments", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 531, + 339, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 339, + 542 + ], + "score": 1.0, + "content": "with well-defined discrete operations. Results show that", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 541, + 339, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 339, + 554 + ], + "score": 1.0, + "content": "NPS can accurately recover each operation defined by the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 552, + 339, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 339, + 565 + ], + "score": 1.0, + "content": "data and learn to represent each operation using a separate", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 564, + 340, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 340, + 576 + ], + "score": 1.0, + "content": "rule. We then move to a much more complicated and", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 575, + 340, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 340, + 586 + ], + "score": 1.0, + "content": "visually rich setting with abstract physical rules and show", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 585, + 339, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 339, + 598 + ], + "score": 1.0, + "content": "that factorization of knowledge into rules as offered by", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 596, + 339, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 339, + 608 + ], + "score": 1.0, + "content": "NPS does scale up to such settings. We study and compare", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 608, + 339, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 339, + 619 + ], + "score": 1.0, + "content": "the parallel and sequential rule application procedures and", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 618, + 339, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 339, + 630 + ], + "score": 1.0, + "content": "try to understand the settings which favour each. We then", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 48 + }, + { + "type": "text", + "bbox": [ + 347, + 489, + 505, + 565 + ], + "lines": [ + { + "bbox": [ + 345, + 487, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 345, + 487, + 506, + 501 + ], + "score": 1.0, + "content": "Table 1: This table shows the segrega-", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 346, + 500, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 346, + 500, + 506, + 511 + ], + "score": 1.0, + "content": "tion of rules for the MNIST Transforma-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 346, + 511, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 346, + 511, + 505, + 522 + ], + "score": 1.0, + "content": "tion task. Each cell indicates the number", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 346, + 522, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 346, + 522, + 505, + 533 + ], + "score": 1.0, + "content": "of times the corresponding rule is used", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 346, + 532, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 346, + 532, + 505, + 544 + ], + "score": 1.0, + "content": "for the given operation. We can see that", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 345, + 543, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 345, + 543, + 505, + 556 + ], + "score": 1.0, + "content": "NPS automatically and perfectly learns", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 346, + 555, + 484, + 567 + ], + "spans": [ + { + "bbox": [ + 346, + 555, + 484, + 567 + ], + "score": 1.0, + "content": "a separate rule for each operation.", + "type": "text" + } + ], + "index": 61 + } + ], + "index": 58 + }, + { + "type": "table", + "bbox": [ + 358, + 573, + 491, + 615 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 358, + 573, + 491, + 615 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 358, + 573, + 491, + 615 + ], + "spans": [ + { + "bbox": [ + 358, + 573, + 491, + 615 + ], + "score": 0.97, + "html": "
Rule 1Rule 2Rule3Rule 4
TranslateDownTranslate UpRotate RightRotate Left5039000
00495000
050300
0004981
", + "type": "table", + "image_path": "ee74d8c57f03621d2b258b5ec0498a5041292fee11a2f8e0e03478c8a8528093.jpg" + } + ] + } + ], + "index": 62.5, + "virtual_lines": [ + { + "bbox": [ + 358, + 573, + 491, + 594.0 + ], + "spans": [], + "index": 62 + }, + { + "bbox": [ + 358, + 594.0, + 491, + 615.0 + ], + "spans": [], + "index": 63 + } + ] + } + ], + "index": 62.5 + }, + { + "type": "text", + "bbox": [ + 107, + 629, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "evaluate the benefits of reusable, dynamic and sparse interactions as offered by NPS in a wide variety", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 106, + 640, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 652 + ], + "score": 1.0, + "content": "of physical environments by comparing it against various baselines. We conduct ablation studies to", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 105, + 650, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 662 + ], + "score": 1.0, + "content": "assess the contribution of different components of NPS. Here we briefly outline the tasks considered", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 105, + 661, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 506, + 675 + ], + "score": 1.0, + "content": "and direct the reader to the Appendix for full details on each task and details on hyperparameter", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 104, + 671, + 143, + 686 + ], + "spans": [ + { + "bbox": [ + 104, + 671, + 143, + 686 + ], + "score": 1.0, + "content": "settings.", + "type": "text" + } + ], + "index": 68 + } + ], + "index": 66 + }, + { + "type": "text", + "bbox": [ + 107, + 689, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "Discussion of baselines. NPS is an interaction network, therefore we use other widely used interac-", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "score": 1.0, + "content": "tion networks such as multihead attention and graph neural networks (Goyal et al. (2019), Goyal et al.", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 106, + 711, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 506, + 724 + ], + "score": 1.0, + "content": "(2020), Veerapaneni et al. (2019), Kipf et al. (2019)) for comparison. Goyal et al. (2019) and Goyal", + "type": "text" + } + ], + "index": 71 + } + ], + "index": 70 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 504, + 95 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 104, + 71, + 506, + 96 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 505, + 144 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 506, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 277, + 113 + ], + "score": 1.0, + "content": "Step 4. Rule Application: the selected rule", + "type": "text" + }, + { + "bbox": [ + 278, + 101, + 292, + 111 + ], + "score": 0.88, + "content": "\\scriptstyle { R _ { r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 100, + 407, + 113 + ], + "score": 1.0, + "content": "is applied to the primary slot", + "type": "text" + }, + { + "bbox": [ + 407, + 101, + 419, + 111 + ], + "score": 0.84, + "content": "V _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 100, + 506, + 113 + ], + "score": 1.0, + "content": "based on the rule and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 111, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 400, + 124 + ], + "score": 1.0, + "content": "the current contents of the primary and contextual slots. The rule-specific", + "type": "text" + }, + { + "bbox": [ + 401, + 111, + 430, + 122 + ], + "score": 0.9, + "content": "M L P _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 111, + 505, + 124 + ], + "score": 1.0, + "content": ", takes as input the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 122, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 404, + 135 + ], + "score": 1.0, + "content": "concatenated representation of the state of the primary and contextual slots,", + "type": "text" + }, + { + "bbox": [ + 405, + 122, + 417, + 134 + ], + "score": 0.88, + "content": "V _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 122, + 435, + 135 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 435, + 122, + 447, + 133 + ], + "score": 0.88, + "content": "V _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 122, + 505, + 135 + ], + "score": 1.0, + "content": ", and produces", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 133, + 475, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 378, + 146 + ], + "score": 1.0, + "content": "an output, which is then used to change the state of the primary slot", + "type": "text" + }, + { + "bbox": [ + 379, + 133, + 390, + 145 + ], + "score": 0.89, + "content": "V _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 133, + 475, + 146 + ], + "score": 1.0, + "content": "by residual addition.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 100, + 506, + 146 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 150, + 376, + 162 + ], + "lines": [ + { + "bbox": [ + 106, + 150, + 375, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 375, + 164 + ], + "score": 1.0, + "content": "3.2 Rule Application: Sequential vs Parallel Rule Application", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 163, + 506, + 240 + ], + "lines": [ + { + "bbox": [ + 106, + 163, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 505, + 174 + ], + "score": 1.0, + "content": "In the previous section, we have described how each rule application only considers another contextual", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 174, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 506, + 186 + ], + "score": 1.0, + "content": "slot for the given primary slot i.e., contextual sparsity. We can also consider application sparsity,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 183, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 104, + 183, + 506, + 198 + ], + "score": 1.0, + "content": "wherein we use the rules to update the states of only a subset of the slots. In this scenario, only", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 195, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 506, + 209 + ], + "score": 1.0, + "content": "the selected slots would be primary slots. This setting will be helpful when there is an entity in an", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 206, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 506, + 219 + ], + "score": 1.0, + "content": "environment that is stationary, or it is following its own default dynamics unaffected by other entities.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "score": 1.0, + "content": "Therefore, it does not need to consider other entities to update its state. We explore two scenarios for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 228, + 225, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 225, + 241 + ], + "score": 1.0, + "content": "enabling application sparsity.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10, + "bbox_fs": [ + 104, + 163, + 506, + 241 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 244, + 297, + 311 + ], + "lines": [ + { + "bbox": [ + 106, + 244, + 297, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 264, + 255 + ], + "score": 1.0, + "content": "Parallel Rule Application. Each of the", + "type": "text" + }, + { + "bbox": [ + 264, + 245, + 276, + 254 + ], + "score": 0.56, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 244, + 297, + 255 + ], + "score": 1.0, + "content": "slots", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 255, + 297, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 297, + 267 + ], + "score": 1.0, + "content": "selects a rule to potentially change its state. To", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 266, + 297, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 297, + 277 + ], + "score": 1.0, + "content": "enable sparse changes, we provide an extra Null", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 277, + 298, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 241, + 289 + ], + "score": 1.0, + "content": "Rule in addition to the available", + "type": "text" + }, + { + "bbox": [ + 242, + 278, + 252, + 287 + ], + "score": 0.74, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 277, + 298, + 289 + ], + "score": 1.0, + "content": "rules. If a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 288, + 297, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 297, + 300 + ], + "score": 1.0, + "content": "slot picks the null rule in step 2 of the above", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 299, + 258, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 258, + 312 + ], + "score": 1.0, + "content": "procedure, we do not update its state.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 244, + 298, + 312 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 316, + 297, + 371 + ], + "lines": [ + { + "bbox": [ + 106, + 314, + 299, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 299, + 329 + ], + "score": 1.0, + "content": "Sequential Rule Application. In this setting,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 327, + 298, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 298, + 338 + ], + "score": 1.0, + "content": "only one slot gets updated in each rule applica-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 338, + 298, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 298, + 349 + ], + "score": 1.0, + "content": "tion step. Therefore, only one slot is selected", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 348, + 297, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 297, + 360 + ], + "score": 1.0, + "content": "as the primary slot. This can be facilitated by", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 359, + 298, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 298, + 373 + ], + "score": 1.0, + "content": "modifying step 2 above to select one {primary", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 314, + 299, + 373 + ] + }, + { + "type": "image", + "bbox": [ + 310, + 247, + 498, + 338 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 310, + 247, + 498, + 338 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 310, + 247, + 498, + 338 + ], + "spans": [ + { + "bbox": [ + 310, + 247, + 498, + 338 + ], + "score": 0.968, + "type": "image", + "image_path": "d2db80f1c75eadd39bbfb8c94bc9f20aca13ed711f1d46b4857e95be0ac7103f.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 310, + 247, + 498, + 260.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 310, + 260.0, + 498, + 273.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 310, + 273.0, + 498, + 286.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 310, + 286.0, + 498, + 299.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 310, + 299.0, + 498, + 312.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 310, + 312.0, + 498, + 325.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 310, + 325.0, + 498, + 338.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 344, + 504, + 367 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 343, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 343, + 505, + 357 + ], + "score": 1.0, + "content": "Figure 3: This figure demonstrates the sequential", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 356, + 420, + 367 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 420, + 367 + ], + "score": 1.0, + "content": "and parallel rule application.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + } + ], + "index": 30.25 + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 428 + ], + "lines": [ + { + "bbox": [ + 106, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 194, + 383 + ], + "score": 1.0, + "content": "slot, rule} pair among", + "type": "text" + }, + { + "bbox": [ + 195, + 370, + 219, + 381 + ], + "score": 0.5, + "content": "_ { N M }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 370, + 423, + 383 + ], + "score": 1.0, + "content": "{rule, slot} pairs. The queries come from each slot:", + "type": "text" + }, + { + "bbox": [ + 424, + 370, + 505, + 383 + ], + "score": 0.77, + "content": "\\pmb { q } _ { j } = \\pmb { V } _ { j } \\pmb { W } ^ { q } \\quad \\forall j \\in", + "type": "inline_equation" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 380, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 107, + 382, + 157, + 394 + ], + "score": 0.91, + "content": "\\{ 1 , \\ldots , M \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 380, + 284, + 397 + ], + "score": 1.0, + "content": ", the keys come from the rules:", + "type": "text" + }, + { + "bbox": [ + 284, + 381, + 415, + 394 + ], + "score": 0.81, + "content": "\\pmb { k } _ { i } = R _ { i } W ^ { k } \\quad \\forall i \\in \\{ 1 , \\dots , N \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 380, + 506, + 397 + ], + "score": 1.0, + "content": ". The straight-through", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 392, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 331, + 408 + ], + "score": 1.0, + "content": "Gumbel softmax selects one (primary slot, rule) pair:", + "type": "text" + }, + { + "bbox": [ + 332, + 394, + 453, + 406 + ], + "score": 0.88, + "content": "p , r = \\mathrm { a r g m a x } _ { i , j } ( q _ { p } k _ { i } + \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 392, + 484, + 408 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 485, + 395, + 505, + 405 + ], + "score": 0.86, + "content": "\\gamma \\sim", + "type": "inline_equation" + } + ], + "index": 36 + }, + { + "bbox": [ + 107, + 403, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 107, + 405, + 164, + 417 + ], + "score": 0.42, + "content": "\\mathrm { G u m b e l } ( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 403, + 506, + 419 + ], + "score": 1.0, + "content": ". In the sequential regime, we allow the rule application procedure (step 2, 3, 4 above)", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 416, + 472, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 285, + 430 + ], + "score": 1.0, + "content": "to be performed multiple times iteratively in", + "type": "text" + }, + { + "bbox": [ + 285, + 417, + 296, + 426 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 416, + 462, + 430 + ], + "score": 1.0, + "content": "rule application stages for each time-step", + "type": "text" + }, + { + "bbox": [ + 463, + 418, + 468, + 426 + ], + "score": 0.69, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 416, + 472, + 430 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 370, + 506, + 430 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 433, + 504, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "A pictorial demonstration of both rule application regimes can be found in Figure 3. We provide", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 444, + 395, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 395, + 457 + ], + "score": 1.0, + "content": "detailed algorithms for the sequential and parallel regimes in Appendix.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 432, + 505, + 457 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 464, + 191, + 477 + ], + "lines": [ + { + "bbox": [ + 104, + 462, + 193, + 480 + ], + "spans": [ + { + "bbox": [ + 104, + 462, + 193, + 480 + ], + "score": 1.0, + "content": "4 Experiments", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 487, + 338, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 339, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 339, + 498 + ], + "score": 1.0, + "content": "We demonstrate the effectiveness of NPS on multiple tasks", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 498, + 339, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 339, + 510 + ], + "score": 1.0, + "content": "and compare to a comprehensive set of baselines. To show", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 507, + 339, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 339, + 522 + ], + "score": 1.0, + "content": "that NPS can learn intuitive rules from the data generating", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 520, + 339, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 339, + 532 + ], + "score": 1.0, + "content": "distribution, we design a couple of simple toy experiments", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 531, + 339, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 339, + 542 + ], + "score": 1.0, + "content": "with well-defined discrete operations. Results show that", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 541, + 339, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 339, + 554 + ], + "score": 1.0, + "content": "NPS can accurately recover each operation defined by the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 552, + 339, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 339, + 565 + ], + "score": 1.0, + "content": "data and learn to represent each operation using a separate", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 564, + 340, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 340, + 576 + ], + "score": 1.0, + "content": "rule. We then move to a much more complicated and", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 575, + 340, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 340, + 586 + ], + "score": 1.0, + "content": "visually rich setting with abstract physical rules and show", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 585, + 339, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 339, + 598 + ], + "score": 1.0, + "content": "that factorization of knowledge into rules as offered by", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 596, + 339, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 339, + 608 + ], + "score": 1.0, + "content": "NPS does scale up to such settings. We study and compare", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 608, + 339, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 339, + 619 + ], + "score": 1.0, + "content": "the parallel and sequential rule application procedures and", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 618, + 339, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 339, + 630 + ], + "score": 1.0, + "content": "try to understand the settings which favour each. We then", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 48, + "bbox_fs": [ + 105, + 487, + 340, + 630 + ] + }, + { + "type": "text", + "bbox": [ + 347, + 489, + 505, + 565 + ], + "lines": [ + { + "bbox": [ + 345, + 487, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 345, + 487, + 506, + 501 + ], + "score": 1.0, + "content": "Table 1: This table shows the segrega-", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 346, + 500, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 346, + 500, + 506, + 511 + ], + "score": 1.0, + "content": "tion of rules for the MNIST Transforma-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 346, + 511, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 346, + 511, + 505, + 522 + ], + "score": 1.0, + "content": "tion task. Each cell indicates the number", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 346, + 522, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 346, + 522, + 505, + 533 + ], + "score": 1.0, + "content": "of times the corresponding rule is used", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 346, + 532, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 346, + 532, + 505, + 544 + ], + "score": 1.0, + "content": "for the given operation. We can see that", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 345, + 543, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 345, + 543, + 505, + 556 + ], + "score": 1.0, + "content": "NPS automatically and perfectly learns", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 346, + 555, + 484, + 567 + ], + "spans": [ + { + "bbox": [ + 346, + 555, + 484, + 567 + ], + "score": 1.0, + "content": "a separate rule for each operation.", + "type": "text" + } + ], + "index": 61 + } + ], + "index": 58, + "bbox_fs": [ + 345, + 487, + 506, + 567 + ] + }, + { + "type": "table", + "bbox": [ + 358, + 573, + 491, + 615 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 358, + 573, + 491, + 615 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 358, + 573, + 491, + 615 + ], + "spans": [ + { + "bbox": [ + 358, + 573, + 491, + 615 + ], + "score": 0.97, + "html": "
Rule 1Rule 2Rule3Rule 4
TranslateDownTranslate UpRotate RightRotate Left5039000
00495000
050300
0004981
", + "type": "table", + "image_path": "ee74d8c57f03621d2b258b5ec0498a5041292fee11a2f8e0e03478c8a8528093.jpg" + } + ] + } + ], + "index": 62.5, + "virtual_lines": [ + { + "bbox": [ + 358, + 573, + 491, + 594.0 + ], + "spans": [], + "index": 62 + }, + { + "bbox": [ + 358, + 594.0, + 491, + 615.0 + ], + "spans": [], + "index": 63 + } + ] + } + ], + "index": 62.5 + }, + { + "type": "text", + "bbox": [ + 107, + 629, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "evaluate the benefits of reusable, dynamic and sparse interactions as offered by NPS in a wide variety", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 106, + 640, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 652 + ], + "score": 1.0, + "content": "of physical environments by comparing it against various baselines. We conduct ablation studies to", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 105, + 650, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 662 + ], + "score": 1.0, + "content": "assess the contribution of different components of NPS. Here we briefly outline the tasks considered", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 105, + 661, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 506, + 675 + ], + "score": 1.0, + "content": "and direct the reader to the Appendix for full details on each task and details on hyperparameter", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 104, + 671, + 143, + 686 + ], + "spans": [ + { + "bbox": [ + 104, + 671, + 143, + 686 + ], + "score": 1.0, + "content": "settings.", + "type": "text" + } + ], + "index": 68 + } + ], + "index": 66, + "bbox_fs": [ + 104, + 628, + 506, + 686 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 689, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "Discussion of baselines. NPS is an interaction network, therefore we use other widely used interac-", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "score": 1.0, + "content": "tion networks such as multihead attention and graph neural networks (Goyal et al. (2019), Goyal et al.", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 106, + 711, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 506, + 724 + ], + "score": 1.0, + "content": "(2020), Veerapaneni et al. (2019), Kipf et al. (2019)) for comparison. Goyal et al. (2019) and Goyal", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 105, + 73, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 506, + 85 + ], + "score": 1.0, + "content": "et al. (2020) use an attention based interaction network to capture interactions between the slots,", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 84, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 84, + 505, + 95 + ], + "score": 1.0, + "content": "while Veerapaneni et al. (2019) and Kipf et al. (2019) use a GNN based interaction network. We", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 95, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 95, + 505, + 106 + ], + "score": 1.0, + "content": "also consider the recently introduced convolutional interaction network (CIN) (Qi et al., 2021) which", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 106, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 505, + 117 + ], + "score": 1.0, + "content": "captures dense pairwise interactions like GNN but uses a convolutional network instead of MLPs", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 505, + 128 + ], + "score": 1.0, + "content": "to better utilize spatial information. The proposed method, similar to other interaction networks, is", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "agnostic to the encoder backbone used to encode the input image into slots, therefore we compare", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 137, + 424, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 424, + 150 + ], + "score": 1.0, + "content": "NPS to other interaction networks across a wide-variety of encoder backbones.", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 70, + "bbox_fs": [ + 106, + 689, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 73, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 105, + 73, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 506, + 85 + ], + "score": 1.0, + "content": "et al. (2020) use an attention based interaction network to capture interactions between the slots,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 84, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 84, + 505, + 95 + ], + "score": 1.0, + "content": "while Veerapaneni et al. (2019) and Kipf et al. (2019) use a GNN based interaction network. We", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 95, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 95, + 505, + 106 + ], + "score": 1.0, + "content": "also consider the recently introduced convolutional interaction network (CIN) (Qi et al., 2021) which", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 106, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 505, + 117 + ], + "score": 1.0, + "content": "captures dense pairwise interactions like GNN but uses a convolutional network instead of MLPs", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 505, + 128 + ], + "score": 1.0, + "content": "to better utilize spatial information. The proposed method, similar to other interaction networks, is", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "agnostic to the encoder backbone used to encode the input image into slots, therefore we compare", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 137, + 424, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 424, + 150 + ], + "score": 1.0, + "content": "NPS to other interaction networks across a wide-variety of encoder backbones.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 106, + 160, + 348, + 172 + ], + "lines": [ + { + "bbox": [ + 105, + 160, + 349, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 349, + 175 + ], + "score": 1.0, + "content": "4.1 Learning intuitive rules with NPS: Toy Simulations", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 181, + 505, + 269 + ], + "lines": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "We designed a couple of simple tasks with well-defined discrete rules to show that NPS can learn", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 193, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 505, + 205 + ], + "score": 1.0, + "content": "intuitive and interpretable rules. We also show the efficiency and effectiveness of the selection", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 204, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 506, + 217 + ], + "score": 1.0, + "content": "procedure (step 2 and step 3 in section 3.1) by comparing against a baseline with many more", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 104, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "parameters. Both tasks require a single modification of only one of the available entities, therefore", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 225, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 237 + ], + "score": 1.0, + "content": "the use of sequential or parallel rule application would not make a difference here since parallel rule", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 235, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 505, + 248 + ], + "score": 1.0, + "content": "application in which all-but-one slots select the null rule is similar to sequential rule application with", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 248, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 259 + ], + "score": 1.0, + "content": "1 rule application step. To simplify the presentation, we describe the setup for both tasks using the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 258, + 258, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 258, + 270 + ], + "score": 1.0, + "content": "sequential rule application procedure.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 274, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 106, + 274, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 506, + 286 + ], + "score": 1.0, + "content": "MNIST Transformation. We test whether NPS can learn simple rules for performing transforma-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 284, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 506, + 298 + ], + "score": 1.0, + "content": "tions on MNIST digits. We generate data with four transformations: {Translate Up, Translate Down,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 296, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 315, + 308 + ], + "score": 1.0, + "content": "Rotate Right, Rotate Left}. We feed the input image", + "type": "text" + }, + { + "bbox": [ + 315, + 296, + 332, + 307 + ], + "score": 0.68, + "content": "( X )", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 297, + 425, + 308 + ], + "score": 1.0, + "content": "and the transformation", + "type": "text" + }, + { + "bbox": [ + 425, + 297, + 438, + 307 + ], + "score": 0.61, + "content": "( o )", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 297, + 505, + 308 + ], + "score": 1.0, + "content": "to be performed", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 306, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 104, + 306, + 506, + 321 + ], + "score": 1.0, + "content": "as a one-hot vector to the model. The detailed setup is described in Appendix. For this task, we", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 317, + 414, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 414, + 331 + ], + "score": 1.0, + "content": "evaluate whether NPS can learn to use a unique rule for each transformation.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 334, + 506, + 378 + ], + "lines": [ + { + "bbox": [ + 105, + 334, + 507, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 507, + 347 + ], + "score": 1.0, + "content": "We use 4 rules corresponding to the 4 transformations with the hope that the correct transforma-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 345, + 507, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 507, + 357 + ], + "score": 1.0, + "content": "tions are recovered. Indeed, we observe that NPS successfully learns to represent each trans-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 356, + 507, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 507, + 368 + ], + "score": 1.0, + "content": "formation using a separate rule as shown in Table 1. Our model achieves an MSE of 0.02.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 367, + 507, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 507, + 380 + ], + "score": 1.0, + "content": "A visualization of the outputs from our model and further details can be found in Appendix C.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 394, + 298, + 492 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 297, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 297, + 406 + ], + "score": 1.0, + "content": "Coordinate Arithmetic Task. The model is", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 405, + 298, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 298, + 417 + ], + "score": 1.0, + "content": "tasked with performing arithmetic operations on", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 416, + 298, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 203, + 428 + ], + "score": 1.0, + "content": "2D coordinates. Given", + "type": "text" + }, + { + "bbox": [ + 203, + 416, + 239, + 428 + ], + "score": 0.93, + "content": "( X _ { 0 } , Y _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 416, + 258, + 428 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 259, + 416, + 294, + 428 + ], + "score": 0.92, + "content": "( X _ { 1 } , Y _ { 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 416, + 298, + 428 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 427, + 298, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 266, + 439 + ], + "score": 1.0, + "content": "we can apply the following operations:", + "type": "text" + }, + { + "bbox": [ + 266, + 428, + 298, + 438 + ], + "score": 0.44, + "content": "\\{ \\mathbf { X } \\mathbf { A d } \\mathbf { - }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 437, + 299, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 141, + 450 + ], + "score": 1.0, + "content": "dition:", + "type": "text" + }, + { + "bbox": [ + 141, + 438, + 258, + 449 + ], + "score": 0.85, + "content": "\\bar { ( X _ { r } , Y _ { r } ) } \\ = \\ ( X _ { 0 } \\bar { + } X _ { 1 } , Y _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 437, + 264, + 450 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 265, + 438, + 274, + 448 + ], + "score": 0.4, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 437, + 299, + 450 + ], + "score": 1.0, + "content": "Sub-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 448, + 299, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 146, + 461 + ], + "score": 1.0, + "content": "traction:", + "type": "text" + }, + { + "bbox": [ + 146, + 450, + 262, + 460 + ], + "score": 0.8, + "content": "( X _ { r } , Y _ { r } ) \\ : = \\ : ( X _ { 0 } - X _ { 1 } , Y _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 448, + 299, + 461 + ], + "score": 1.0, + "content": ", Y Ad-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 459, + 299, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 138, + 472 + ], + "score": 1.0, + "content": "dition:", + "type": "text" + }, + { + "bbox": [ + 138, + 461, + 245, + 471 + ], + "score": 0.86, + "content": "( X _ { r } , Y _ { r } ) = ( X _ { 0 } , Y _ { 0 } + Y _ { 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 459, + 299, + 472 + ], + "score": 1.0, + "content": ", Y Subtrac-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 469, + 297, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 129, + 484 + ], + "score": 1.0, + "content": "tion:", + "type": "text" + }, + { + "bbox": [ + 129, + 471, + 231, + 483 + ], + "score": 0.87, + "content": "X _ { r } , Y _ { r } = ( X _ { 0 } , Y _ { 0 } - Y _ { 1 } ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 469, + 261, + 484 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 261, + 471, + 297, + 482 + ], + "score": 0.91, + "content": "( X _ { r } , Y _ { r } )", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 480, + 215, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 215, + 493 + ], + "score": 1.0, + "content": "is the resultant coordinate.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 498, + 297, + 694 + ], + "lines": [ + { + "bbox": [ + 106, + 497, + 298, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 298, + 510 + ], + "score": 1.0, + "content": "In this task, the model is given 2 input coordi-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 507, + 297, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 130, + 522 + ], + "score": 1.0, + "content": "nates", + "type": "text" + }, + { + "bbox": [ + 131, + 509, + 225, + 521 + ], + "score": 0.9, + "content": "X = [ ( x _ { i } , y _ { i } ) , ( x _ { j } , \\bar { y _ { j } } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 507, + 297, + 522 + ], + "score": 1.0, + "content": "and the expected", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 519, + 297, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 180, + 532 + ], + "score": 1.0, + "content": "output coordinates", + "type": "text" + }, + { + "bbox": [ + 181, + 520, + 276, + 532 + ], + "score": 0.88, + "content": "Y = [ ( \\hat { x _ { i } } , \\hat { y _ { i } } ) , ( \\hat { x _ { j } } , \\hat { y _ { j } } ) ] .", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 519, + 297, + 532 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 531, + 297, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 297, + 543 + ], + "score": 1.0, + "content": "model is supposed to infer the correct rule to", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 542, + 298, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 298, + 553 + ], + "score": 1.0, + "content": "produce the correct output coordinates. Dur-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 553, + 297, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 297, + 564 + ], + "score": 1.0, + "content": "ing data collection, the true output is obtained", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 563, + 298, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 298, + 576 + ], + "score": 1.0, + "content": "by performing a random transformation on a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 574, + 298, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 235, + 586 + ], + "score": 1.0, + "content": "randomly selected coordinate in", + "type": "text" + }, + { + "bbox": [ + 235, + 575, + 245, + 584 + ], + "score": 0.78, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 574, + 298, + 586 + ], + "score": 1.0, + "content": "(primary co-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 585, + 298, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 298, + 596 + ], + "score": 1.0, + "content": "ordinate), taking another randomly selected co-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 596, + 298, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 167, + 608 + ], + "score": 1.0, + "content": "ordinate from", + "type": "text" + }, + { + "bbox": [ + 167, + 597, + 177, + 606 + ], + "score": 0.79, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 596, + 298, + 608 + ], + "score": 1.0, + "content": "(contextual coordinate) into", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 606, + 298, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 298, + 619 + ], + "score": 1.0, + "content": "account. The detailed setup is described in Ap-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 618, + 297, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 297, + 629 + ], + "score": 1.0, + "content": "pendix D. We use an NPS model with 4 rules", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 629, + 297, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 297, + 641 + ], + "score": 1.0, + "content": "for this task. We use the the selection procedure", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 640, + 297, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 297, + 651 + ], + "score": 1.0, + "content": "in step 2 and step 3 of algorithm 1 to select the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 651, + 297, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 297, + 662 + ], + "score": 1.0, + "content": "primary coordinate, contextual coordinate, and", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 662, + 298, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 298, + 673 + ], + "score": 1.0, + "content": "the rule. For the baseline we replace the selec-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 672, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 672, + 297, + 685 + ], + "score": 1.0, + "content": "tion procedure in NPS (i.e. step 2 and step 3 in", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 682, + 360, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 360, + 696 + ], + "score": 1.0, + "content": "algorithm 1) with a routing MLP similar to Fedus et al. (2021).", + "type": "text" + } + ], + "index": 76 + } + ], + "index": 42.5 + }, + { + "type": "image", + "bbox": [ + 308, + 384, + 500, + 575 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 308, + 384, + 500, + 575 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 500, + 575 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 500, + 575 + ], + "score": 0.966, + "type": "image", + "image_path": "93fa4fe136835a329ff05316b00960119a6e62b649e79c974d4d8c5f78adba3a.jpg" + } + ] + } + ], + "index": 58, + "virtual_lines": [ + { + "bbox": [ + 308, + 384, + 500, + 396.73333333333335 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 308, + 396.73333333333335, + 500, + 409.4666666666667 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 308, + 409.4666666666667, + 500, + 422.20000000000005 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 308, + 422.20000000000005, + 500, + 434.9333333333334 + ], + "spans": [], + "index": 54 + }, + { + "bbox": [ + 308, + 434.9333333333334, + 500, + 447.66666666666674 + ], + "spans": [], + "index": 55 + }, + { + "bbox": [ + 308, + 447.66666666666674, + 500, + 460.4000000000001 + ], + "spans": [], + "index": 56 + }, + { + "bbox": [ + 308, + 460.4000000000001, + 500, + 473.13333333333344 + ], + "spans": [], + "index": 57 + }, + { + "bbox": [ + 308, + 473.13333333333344, + 500, + 485.8666666666668 + ], + "spans": [], + "index": 58 + }, + { + "bbox": [ + 308, + 485.8666666666668, + 500, + 498.60000000000014 + ], + "spans": [], + "index": 59 + }, + { + "bbox": [ + 308, + 498.60000000000014, + 500, + 511.3333333333335 + ], + "spans": [], + "index": 60 + }, + { + "bbox": [ + 308, + 511.3333333333335, + 500, + 524.0666666666668 + ], + "spans": [], + "index": 61 + }, + { + "bbox": [ + 308, + 524.0666666666668, + 500, + 536.8000000000002 + ], + "spans": [], + "index": 62 + }, + { + "bbox": [ + 308, + 536.8000000000002, + 500, + 549.5333333333335 + ], + "spans": [], + "index": 63 + }, + { + "bbox": [ + 308, + 549.5333333333335, + 500, + 562.2666666666669 + ], + "spans": [], + "index": 64 + }, + { + "bbox": [ + 308, + 562.2666666666669, + 500, + 575.0000000000002 + ], + "spans": [], + "index": 65 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 585, + 505, + 683 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 303, + 584, + 507, + 597 + ], + "spans": [ + { + "bbox": [ + 303, + 584, + 507, + 597 + ], + "score": 1.0, + "content": "Figure 4: Coordinate Arithmetic Task. Here,", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 303, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 303, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "we compare NPS to the baseline model in terms", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 303, + 606, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 303, + 606, + 506, + 619 + ], + "score": 1.0, + "content": "of segregation of rules as the training progresses.", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 303, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 303, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "X-axis shows the epochs and Y-axis shows the", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 303, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 303, + 628, + 415, + 641 + ], + "score": 1.0, + "content": "frequency with which Rule", + "type": "text" + }, + { + "bbox": [ + 416, + 630, + 421, + 639 + ], + "score": 0.71, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "is used for the given", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 303, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 303, + 639, + 505, + 651 + ], + "score": 1.0, + "content": "operation. We can see that NPS disentangles the", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 303, + 651, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 303, + 651, + 505, + 662 + ], + "score": 1.0, + "content": "operations perfectly as training progresses with a", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 303, + 662, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 303, + 662, + 505, + 673 + ], + "score": 1.0, + "content": "unique rule specializing to every operation while", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 304, + 672, + 438, + 684 + ], + "spans": [ + { + "bbox": [ + 304, + 672, + 438, + 684 + ], + "score": 1.0, + "content": "the baseline model fails to do so.", + "type": "text" + } + ], + "index": 74 + } + ], + "index": 70 + } + ], + "index": 64.0 + }, + { + "type": "text", + "bbox": [ + 108, + 684, + 360, + 694 + ], + "lines": [ + { + "bbox": [ + 108, + 684, + 360, + 694 + ], + "spans": [], + "index": 75 + } + ], + "index": 75 + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "This routing MLP has 3 heads (one each for selecting the primary coordinate, contextual coordinate,", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "and the rule). The baseline has 4 times more parameters than NPS. The final output is produced by", + "type": "text" + } + ], + "index": 78 + } + ], + "index": 77.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 73, + 505, + 150 + ], + "lines": [], + "index": 3, + "bbox_fs": [ + 105, + 73, + 506, + 150 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 106, + 160, + 348, + 172 + ], + "lines": [ + { + "bbox": [ + 105, + 160, + 349, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 349, + 175 + ], + "score": 1.0, + "content": "4.1 Learning intuitive rules with NPS: Toy Simulations", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 181, + 505, + 269 + ], + "lines": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "We designed a couple of simple tasks with well-defined discrete rules to show that NPS can learn", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 193, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 505, + 205 + ], + "score": 1.0, + "content": "intuitive and interpretable rules. We also show the efficiency and effectiveness of the selection", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 204, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 506, + 217 + ], + "score": 1.0, + "content": "procedure (step 2 and step 3 in section 3.1) by comparing against a baseline with many more", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 104, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "parameters. Both tasks require a single modification of only one of the available entities, therefore", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 225, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 237 + ], + "score": 1.0, + "content": "the use of sequential or parallel rule application would not make a difference here since parallel rule", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 235, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 505, + 248 + ], + "score": 1.0, + "content": "application in which all-but-one slots select the null rule is similar to sequential rule application with", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 248, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 259 + ], + "score": 1.0, + "content": "1 rule application step. To simplify the presentation, we describe the setup for both tasks using the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 258, + 258, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 258, + 270 + ], + "score": 1.0, + "content": "sequential rule application procedure.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5, + "bbox_fs": [ + 104, + 182, + 506, + 270 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 274, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 106, + 274, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 506, + 286 + ], + "score": 1.0, + "content": "MNIST Transformation. We test whether NPS can learn simple rules for performing transforma-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 284, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 506, + 298 + ], + "score": 1.0, + "content": "tions on MNIST digits. We generate data with four transformations: {Translate Up, Translate Down,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 296, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 315, + 308 + ], + "score": 1.0, + "content": "Rotate Right, Rotate Left}. We feed the input image", + "type": "text" + }, + { + "bbox": [ + 315, + 296, + 332, + 307 + ], + "score": 0.68, + "content": "( X )", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 297, + 425, + 308 + ], + "score": 1.0, + "content": "and the transformation", + "type": "text" + }, + { + "bbox": [ + 425, + 297, + 438, + 307 + ], + "score": 0.61, + "content": "( o )", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 297, + 505, + 308 + ], + "score": 1.0, + "content": "to be performed", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 306, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 104, + 306, + 506, + 321 + ], + "score": 1.0, + "content": "as a one-hot vector to the model. The detailed setup is described in Appendix. For this task, we", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 317, + 414, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 414, + 331 + ], + "score": 1.0, + "content": "evaluate whether NPS can learn to use a unique rule for each transformation.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18, + "bbox_fs": [ + 104, + 274, + 506, + 331 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 334, + 506, + 378 + ], + "lines": [ + { + "bbox": [ + 105, + 334, + 507, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 507, + 347 + ], + "score": 1.0, + "content": "We use 4 rules corresponding to the 4 transformations with the hope that the correct transforma-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 345, + 507, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 507, + 357 + ], + "score": 1.0, + "content": "tions are recovered. Indeed, we observe that NPS successfully learns to represent each trans-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 356, + 507, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 507, + 368 + ], + "score": 1.0, + "content": "formation using a separate rule as shown in Table 1. Our model achieves an MSE of 0.02.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 367, + 507, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 507, + 380 + ], + "score": 1.0, + "content": "A visualization of the outputs from our model and further details can be found in Appendix C.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 334, + 507, + 380 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 394, + 298, + 492 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 297, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 297, + 406 + ], + "score": 1.0, + "content": "Coordinate Arithmetic Task. The model is", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 405, + 298, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 298, + 417 + ], + "score": 1.0, + "content": "tasked with performing arithmetic operations on", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 416, + 298, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 203, + 428 + ], + "score": 1.0, + "content": "2D coordinates. Given", + "type": "text" + }, + { + "bbox": [ + 203, + 416, + 239, + 428 + ], + "score": 0.93, + "content": "( X _ { 0 } , Y _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 416, + 258, + 428 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 259, + 416, + 294, + 428 + ], + "score": 0.92, + "content": "( X _ { 1 } , Y _ { 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 416, + 298, + 428 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 427, + 298, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 266, + 439 + ], + "score": 1.0, + "content": "we can apply the following operations:", + "type": "text" + }, + { + "bbox": [ + 266, + 428, + 298, + 438 + ], + "score": 0.44, + "content": "\\{ \\mathbf { X } \\mathbf { A d } \\mathbf { - }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 437, + 299, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 141, + 450 + ], + "score": 1.0, + "content": "dition:", + "type": "text" + }, + { + "bbox": [ + 141, + 438, + 258, + 449 + ], + "score": 0.85, + "content": "\\bar { ( X _ { r } , Y _ { r } ) } \\ = \\ ( X _ { 0 } \\bar { + } X _ { 1 } , Y _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 437, + 264, + 450 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 265, + 438, + 274, + 448 + ], + "score": 0.4, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 437, + 299, + 450 + ], + "score": 1.0, + "content": "Sub-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 448, + 299, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 146, + 461 + ], + "score": 1.0, + "content": "traction:", + "type": "text" + }, + { + "bbox": [ + 146, + 450, + 262, + 460 + ], + "score": 0.8, + "content": "( X _ { r } , Y _ { r } ) \\ : = \\ : ( X _ { 0 } - X _ { 1 } , Y _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 448, + 299, + 461 + ], + "score": 1.0, + "content": ", Y Ad-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 459, + 299, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 138, + 472 + ], + "score": 1.0, + "content": "dition:", + "type": "text" + }, + { + "bbox": [ + 138, + 461, + 245, + 471 + ], + "score": 0.86, + "content": "( X _ { r } , Y _ { r } ) = ( X _ { 0 } , Y _ { 0 } + Y _ { 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 459, + 299, + 472 + ], + "score": 1.0, + "content": ", Y Subtrac-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 469, + 297, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 129, + 484 + ], + "score": 1.0, + "content": "tion:", + "type": "text" + }, + { + "bbox": [ + 129, + 471, + 231, + 483 + ], + "score": 0.87, + "content": "X _ { r } , Y _ { r } = ( X _ { 0 } , Y _ { 0 } - Y _ { 1 } ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 469, + 261, + 484 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 261, + 471, + 297, + 482 + ], + "score": 0.91, + "content": "( X _ { r } , Y _ { r } )", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 480, + 215, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 215, + 493 + ], + "score": 1.0, + "content": "is the resultant coordinate.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 393, + 299, + 493 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 498, + 297, + 694 + ], + "lines": [ + { + "bbox": [ + 106, + 497, + 298, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 298, + 510 + ], + "score": 1.0, + "content": "In this task, the model is given 2 input coordi-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 507, + 297, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 130, + 522 + ], + "score": 1.0, + "content": "nates", + "type": "text" + }, + { + "bbox": [ + 131, + 509, + 225, + 521 + ], + "score": 0.9, + "content": "X = [ ( x _ { i } , y _ { i } ) , ( x _ { j } , \\bar { y _ { j } } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 507, + 297, + 522 + ], + "score": 1.0, + "content": "and the expected", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 519, + 297, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 180, + 532 + ], + "score": 1.0, + "content": "output coordinates", + "type": "text" + }, + { + "bbox": [ + 181, + 520, + 276, + 532 + ], + "score": 0.88, + "content": "Y = [ ( \\hat { x _ { i } } , \\hat { y _ { i } } ) , ( \\hat { x _ { j } } , \\hat { y _ { j } } ) ] .", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 519, + 297, + 532 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 531, + 297, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 297, + 543 + ], + "score": 1.0, + "content": "model is supposed to infer the correct rule to", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 542, + 298, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 298, + 553 + ], + "score": 1.0, + "content": "produce the correct output coordinates. Dur-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 553, + 297, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 297, + 564 + ], + "score": 1.0, + "content": "ing data collection, the true output is obtained", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 563, + 298, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 298, + 576 + ], + "score": 1.0, + "content": "by performing a random transformation on a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 574, + 298, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 235, + 586 + ], + "score": 1.0, + "content": "randomly selected coordinate in", + "type": "text" + }, + { + "bbox": [ + 235, + 575, + 245, + 584 + ], + "score": 0.78, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 574, + 298, + 586 + ], + "score": 1.0, + "content": "(primary co-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 585, + 298, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 298, + 596 + ], + "score": 1.0, + "content": "ordinate), taking another randomly selected co-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 596, + 298, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 167, + 608 + ], + "score": 1.0, + "content": "ordinate from", + "type": "text" + }, + { + "bbox": [ + 167, + 597, + 177, + 606 + ], + "score": 0.79, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 596, + 298, + 608 + ], + "score": 1.0, + "content": "(contextual coordinate) into", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 606, + 298, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 298, + 619 + ], + "score": 1.0, + "content": "account. The detailed setup is described in Ap-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 618, + 297, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 297, + 629 + ], + "score": 1.0, + "content": "pendix D. We use an NPS model with 4 rules", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 629, + 297, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 297, + 641 + ], + "score": 1.0, + "content": "for this task. We use the the selection procedure", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 640, + 297, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 297, + 651 + ], + "score": 1.0, + "content": "in step 2 and step 3 of algorithm 1 to select the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 651, + 297, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 297, + 662 + ], + "score": 1.0, + "content": "primary coordinate, contextual coordinate, and", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 662, + 298, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 298, + 673 + ], + "score": 1.0, + "content": "the rule. For the baseline we replace the selec-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 672, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 672, + 297, + 685 + ], + "score": 1.0, + "content": "tion procedure in NPS (i.e. step 2 and step 3 in", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 682, + 360, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 360, + 696 + ], + "score": 1.0, + "content": "algorithm 1) with a routing MLP similar to Fedus et al. (2021).", + "type": "text" + } + ], + "index": 76 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 497, + 360, + 696 + ] + }, + { + "type": "image", + "bbox": [ + 308, + 384, + 500, + 575 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 308, + 384, + 500, + 575 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 500, + 575 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 500, + 575 + ], + "score": 0.966, + "type": "image", + "image_path": "93fa4fe136835a329ff05316b00960119a6e62b649e79c974d4d8c5f78adba3a.jpg" + } + ] + } + ], + "index": 58, + "virtual_lines": [ + { + "bbox": [ + 308, + 384, + 500, + 396.73333333333335 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 308, + 396.73333333333335, + 500, + 409.4666666666667 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 308, + 409.4666666666667, + 500, + 422.20000000000005 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 308, + 422.20000000000005, + 500, + 434.9333333333334 + ], + "spans": [], + "index": 54 + }, + { + "bbox": [ + 308, + 434.9333333333334, + 500, + 447.66666666666674 + ], + "spans": [], + "index": 55 + }, + { + "bbox": [ + 308, + 447.66666666666674, + 500, + 460.4000000000001 + ], + "spans": [], + "index": 56 + }, + { + "bbox": [ + 308, + 460.4000000000001, + 500, + 473.13333333333344 + ], + "spans": [], + "index": 57 + }, + { + "bbox": [ + 308, + 473.13333333333344, + 500, + 485.8666666666668 + ], + "spans": [], + "index": 58 + }, + { + "bbox": [ + 308, + 485.8666666666668, + 500, + 498.60000000000014 + ], + "spans": [], + "index": 59 + }, + { + "bbox": [ + 308, + 498.60000000000014, + 500, + 511.3333333333335 + ], + "spans": [], + "index": 60 + }, + { + "bbox": [ + 308, + 511.3333333333335, + 500, + 524.0666666666668 + ], + "spans": [], + "index": 61 + }, + { + "bbox": [ + 308, + 524.0666666666668, + 500, + 536.8000000000002 + ], + "spans": [], + "index": 62 + }, + { + "bbox": [ + 308, + 536.8000000000002, + 500, + 549.5333333333335 + ], + "spans": [], + "index": 63 + }, + { + "bbox": [ + 308, + 549.5333333333335, + 500, + 562.2666666666669 + ], + "spans": [], + "index": 64 + }, + { + "bbox": [ + 308, + 562.2666666666669, + 500, + 575.0000000000002 + ], + "spans": [], + "index": 65 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 585, + 505, + 683 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 303, + 584, + 507, + 597 + ], + "spans": [ + { + "bbox": [ + 303, + 584, + 507, + 597 + ], + "score": 1.0, + "content": "Figure 4: Coordinate Arithmetic Task. Here,", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 303, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 303, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "we compare NPS to the baseline model in terms", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 303, + 606, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 303, + 606, + 506, + 619 + ], + "score": 1.0, + "content": "of segregation of rules as the training progresses.", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 303, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 303, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "X-axis shows the epochs and Y-axis shows the", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 303, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 303, + 628, + 415, + 641 + ], + "score": 1.0, + "content": "frequency with which Rule", + "type": "text" + }, + { + "bbox": [ + 416, + 630, + 421, + 639 + ], + "score": 0.71, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "is used for the given", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 303, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 303, + 639, + 505, + 651 + ], + "score": 1.0, + "content": "operation. We can see that NPS disentangles the", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 303, + 651, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 303, + 651, + 505, + 662 + ], + "score": 1.0, + "content": "operations perfectly as training progresses with a", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 303, + 662, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 303, + 662, + 505, + 673 + ], + "score": 1.0, + "content": "unique rule specializing to every operation while", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 304, + 672, + 438, + 684 + ], + "spans": [ + { + "bbox": [ + 304, + 672, + 438, + 684 + ], + "score": 1.0, + "content": "the baseline model fails to do so.", + "type": "text" + } + ], + "index": 74 + } + ], + "index": 70 + } + ], + "index": 64.0 + }, + { + "type": "text", + "bbox": [ + 108, + 684, + 360, + 694 + ], + "lines": [ + { + "bbox": [ + 108, + 684, + 360, + 694 + ], + "spans": [], + "index": 75 + } + ], + "index": 75, + "bbox_fs": [ + 108, + 684, + 360, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "This routing MLP has 3 heads (one each for selecting the primary coordinate, contextual coordinate,", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "and the rule). The baseline has 4 times more parameters than NPS. The final output is produced by", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "score": 1.0, + "content": "the rule MLP which does not have access to the true output, hence the model cannot simply copy the", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 84, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 84, + 505, + 95 + ], + "score": 1.0, + "content": "true output to produce the actual output. Unlike the MNIST transformation task, we do not provide", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 107 + ], + "score": 1.0, + "content": "the operation to be performed as a one-hot vector input to the model, therefore it needs to infer the", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 106, + 310, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 310, + 118 + ], + "score": 1.0, + "content": "available operations from the data demonstrations.", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 77.5, + "bbox_fs": [ + 105, + 699, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 73, + 505, + 117 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "score": 1.0, + "content": "the rule MLP which does not have access to the true output, hence the model cannot simply copy the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 84, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 84, + 505, + 95 + ], + "score": 1.0, + "content": "true output to produce the actual output. Unlike the MNIST transformation task, we do not provide", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 107 + ], + "score": 1.0, + "content": "the operation to be performed as a one-hot vector input to the model, therefore it needs to infer the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 106, + 310, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 310, + 118 + ], + "score": 1.0, + "content": "available operations from the data demonstrations.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 107, + 122, + 339, + 264 + ], + "lines": [ + { + "bbox": [ + 106, + 121, + 339, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 121, + 339, + 134 + ], + "score": 1.0, + "content": "We show the segregation of rules for NPS and the baseline", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 131, + 340, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 340, + 146 + ], + "score": 1.0, + "content": "in Figure 4. We can see that NPS learns to use a unique", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 143, + 339, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 339, + 156 + ], + "score": 1.0, + "content": "rule for each operation while the baseline struggles to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 154, + 339, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 339, + 167 + ], + "score": 1.0, + "content": "disentangle the underlying operations properly. NPS", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 166, + 340, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 340, + 177 + ], + "score": 1.0, + "content": "also outperforms the baseline in terms of MSE achiev-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 177, + 339, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 169, + 189 + ], + "score": 1.0, + "content": "ing an MSE of", + "type": "text" + }, + { + "bbox": [ + 170, + 177, + 211, + 188 + ], + "score": 0.9, + "content": "0 . 0 1 _ { \\pm 0 . 0 0 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 177, + 339, + 189 + ], + "score": 1.0, + "content": "while the baseline achieves an", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 186, + 340, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 140, + 200 + ], + "score": 1.0, + "content": "MSE of", + "type": "text" + }, + { + "bbox": [ + 140, + 187, + 181, + 199 + ], + "score": 0.9, + "content": "0 . 0 4 _ { \\pm 0 . 0 0 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 186, + 340, + 200 + ], + "score": 1.0, + "content": ". To further confirm that NPS learns all", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 198, + 340, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 340, + 210 + ], + "score": 1.0, + "content": "the available operations correctly from raw data demon-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 209, + 339, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 339, + 221 + ], + "score": 1.0, + "content": "strations, we use an NPS model with 5 rules. We expect", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 219, + 339, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 339, + 232 + ], + "score": 1.0, + "content": "that in this case NPS should utilize only 4 rules since", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 231, + 339, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 339, + 243 + ], + "score": 1.0, + "content": "the data describes only 4 unique operations and indeed", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 242, + 339, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 339, + 254 + ], + "score": 1.0, + "content": "we observe that NPS ends up mostly utilizing 4 of the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 253, + 262, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 262, + 264 + ], + "score": 1.0, + "content": "available 5 rules as shown in Table 2.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 347, + 121, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 345, + 119, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 345, + 119, + 505, + 133 + ], + "score": 1.0, + "content": "Table 2: This table shows segregation", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 346, + 131, + 505, + 143 + ], + "spans": [ + { + "bbox": [ + 346, + 131, + 505, + 143 + ], + "score": 1.0, + "content": "of rules when we use NPS with 5 rules", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 345, + 142, + 506, + 154 + ], + "spans": [ + { + "bbox": [ + 345, + 142, + 506, + 154 + ], + "score": 1.0, + "content": "but the data generation distributions de-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 346, + 154, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 346, + 154, + 505, + 165 + ], + "score": 1.0, + "content": "scribes only 4 possible operations. We", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 346, + 164, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 346, + 164, + 506, + 176 + ], + "score": 1.0, + "content": "can see that only 4 rules get majorly uti-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 345, + 174, + 507, + 187 + ], + "spans": [ + { + "bbox": [ + 345, + 174, + 507, + 187 + ], + "score": 1.0, + "content": "lized thus confirming that NPS success-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 346, + 186, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 346, + 186, + 506, + 199 + ], + "score": 1.0, + "content": "fully recovers all possible operations de-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 347, + 198, + 427, + 209 + ], + "spans": [ + { + "bbox": [ + 347, + 198, + 427, + 209 + ], + "score": 1.0, + "content": "scribed by the data.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20.5 + }, + { + "type": "table", + "bbox": [ + 349, + 216, + 500, + 270 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 349, + 216, + 500, + 270 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 349, + 216, + 500, + 270 + ], + "spans": [ + { + "bbox": [ + 349, + 216, + 500, + 270 + ], + "score": 0.974, + "html": "
Rule 1Rule 2Rule 3Rule 4Rule 5
X Addition3609945130
X Subtraction0482010
Y Subtraction03945320
Y Addition0571599335
", + "type": "table", + "image_path": "a8f84dde9bca14de9d9314267a57d24b53ba78f2f3b4666f783cd97b86554942.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 349, + 216, + 500, + 243.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 349, + 243.0, + 500, + 270.0 + ], + "spans": [], + "index": 26 + } + ] + } + ], + "index": 25.5 + }, + { + "type": "title", + "bbox": [ + 108, + 278, + 297, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 298, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 298, + 294 + ], + "score": 1.0, + "content": "4.2 Parallel vs Sequential Rule Application", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 108, + 299, + 503, + 332 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "We compare the parallel and sequential rule application procedures, to understand the settings that", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 310, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 505, + 322 + ], + "score": 1.0, + "content": "favour one or the other, over two tasks: (1) Bouncing Balls, (2) Shapes Stack. We use the term PNPS", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 321, + 443, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 443, + 334 + ], + "score": 1.0, + "content": "to refer to parallel rule application and SNPS to refer to sequential rule application.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 341, + 339, + 503 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 340, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 340, + 352 + ], + "score": 1.0, + "content": "Shapes Stack. We use the shapes stack dataset intro-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 351, + 340, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 340, + 363 + ], + "score": 1.0, + "content": "duced by Groth et al. (2018). This dataset consists of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 362, + 339, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 339, + 374 + ], + "score": 1.0, + "content": "objects stacked on top of each other as shown in Figure", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 372, + 340, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 340, + 385 + ], + "score": 1.0, + "content": "5. These objects fall under the influence of gravity. For", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 384, + 340, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 340, + 395 + ], + "score": 1.0, + "content": "our experiments, We follow the same setup as Qi et al.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 394, + 340, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 340, + 407 + ], + "score": 1.0, + "content": "(2021). In this task, given the first frame, the model is", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 405, + 340, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 340, + 417 + ], + "score": 1.0, + "content": "tasked with predicting the object bounding boxes for the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 417, + 341, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 125, + 428 + ], + "score": 1.0, + "content": "next", + "type": "text" + }, + { + "bbox": [ + 126, + 417, + 132, + 426 + ], + "score": 0.51, + "content": "\\pmb { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 417, + 341, + 428 + ], + "score": 1.0, + "content": "timesteps. The first frame is encoded using a convo-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 427, + 340, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 340, + 439 + ], + "score": 1.0, + "content": "lutional network followed by RoIPooling (Girshick (2015))", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 439, + 339, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 339, + 450 + ], + "score": 1.0, + "content": "to extract object-centric visual features. The object-centric", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 449, + 339, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 339, + 461 + ], + "score": 1.0, + "content": "features are then passed to the dynamics model to predict", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 460, + 340, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 244, + 472 + ], + "score": 1.0, + "content": "object bounding boxes of the next", + "type": "text" + }, + { + "bbox": [ + 244, + 461, + 250, + 470 + ], + "score": 0.37, + "content": "\\pmb { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 460, + 340, + 472 + ], + "score": 1.0, + "content": "steps. Qi et al. (2021)", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 471, + 340, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 340, + 483 + ], + "score": 1.0, + "content": "propose a Region Proposal Interaction Network (RPIN) to", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 482, + 340, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 340, + 493 + ], + "score": 1.0, + "content": "solve this task. The dynamics model in RPIN consists of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 492, + 341, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 341, + 505 + ], + "score": 1.0, + "content": "an Interaction Network proposed in Battaglia et al. (2016).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 38 + }, + { + "type": "table", + "bbox": [ + 350, + 342, + 500, + 389 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 350, + 342, + 500, + 389 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 350, + 342, + 500, + 389 + ], + "spans": [ + { + "bbox": [ + 350, + 342, + 500, + 389 + ], + "score": 0.971, + "html": "
Model NameTestTransfer
RPIN (Qi et al. (2021))1.254±0.0086.377±0.325
PNPSSNPS1.250±0.0075.411±0.45
1.68±0.025.80±0.15
", + "type": "table", + "image_path": "6667bcdf1c3ef24b7d961678d2db241d0c4a6cfee87cc246cf897591d179148d.jpg" + } + ] + } + ], + "index": 46.5, + "virtual_lines": [ + { + "bbox": [ + 350, + 342, + 500, + 365.5 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 350, + 365.5, + 500, + 389.0 + ], + "spans": [], + "index": 47 + } + ] + } + ], + "index": 46.5 + }, + { + "type": "text", + "bbox": [ + 347, + 394, + 505, + 503 + ], + "lines": [ + { + "bbox": [ + 346, + 393, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 346, + 393, + 506, + 405 + ], + "score": 1.0, + "content": "Table 3: Prediction error of the com-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 345, + 405, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 345, + 405, + 505, + 416 + ], + "score": 1.0, + "content": "pared models on the shapes stack", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 346, + 415, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 346, + 415, + 505, + 427 + ], + "score": 1.0, + "content": "dataset (lower is better) for the test as", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 346, + 426, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 346, + 426, + 506, + 438 + ], + "score": 1.0, + "content": "well as transfer setting. In the test set-", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 346, + 437, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 346, + 437, + 475, + 449 + ], + "score": 1.0, + "content": "ting the number of rollout steps", + "type": "text" + }, + { + "bbox": [ + 475, + 438, + 481, + 447 + ], + "score": 0.71, + "content": "\\pmb { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 437, + 505, + 449 + ], + "score": 1.0, + "content": "is set", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 345, + 448, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 345, + 448, + 505, + 460 + ], + "score": 1.0, + "content": "to 15 and in the transfer setting it is set", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 345, + 459, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 345, + 459, + 506, + 471 + ], + "score": 1.0, + "content": "to 30. We can see that PNPS outper-", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 345, + 470, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 345, + 470, + 505, + 482 + ], + "score": 1.0, + "content": "forms the RPIN baseline in both the test", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 346, + 481, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 346, + 481, + 505, + 493 + ], + "score": 1.0, + "content": "and transfer setting while SNPS fails to", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 347, + 492, + 471, + 503 + ], + "spans": [ + { + "bbox": [ + 347, + 492, + 471, + 503 + ], + "score": 1.0, + "content": "do so. Results across 15 seeds.", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 52.5 + }, + { + "type": "text", + "bbox": [ + 107, + 504, + 505, + 591 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 505, + 517 + ], + "score": 1.0, + "content": "To better utilize spatial information, Qi et al. (2021) propose an extension of the interaction operators", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 515, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 526 + ], + "score": 1.0, + "content": "in interaction net to operate on 3D tensors. This is achieved by replacing the MLP operations in the", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 525, + 504, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 504, + 538 + ], + "score": 1.0, + "content": "original interaction networks with convolutions. They call this new network Convolutional Interaction", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 536, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 506, + 549 + ], + "score": 1.0, + "content": "Network (CIN). For the proposed model, we replace this CIN in RPIN by NPS. To ensure a fair", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "comparison to CIN, we use CNNs to represent rules in NPS instead of MLPs. CIN captures all", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "score": 1.0, + "content": "pairwise interactions between objects using a convolutional network. In NPS, we capture sparse", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 105, + 569, + 507, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 507, + 582 + ], + "score": 1.0, + "content": "interactions (contextual sparsity) as compared to dense pairwise interactions captured by CIN. Also,", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 580, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 481, + 592 + ], + "score": 1.0, + "content": "in NPS we update only a few subset of slots per step instead of all slots (application sparsity).", + "type": "text" + } + ], + "index": 65 + } + ], + "index": 61.5 + }, + { + "type": "text", + "bbox": [ + 108, + 596, + 502, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 609 + ], + "score": 1.0, + "content": "We consider two evaluation settings. (1) Test setting: The number of rollout timesteps is same as that", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 105, + 607, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 208, + 620 + ], + "score": 1.0, + "content": "seen during training (i.e.", + "type": "text" + }, + { + "bbox": [ + 209, + 608, + 238, + 618 + ], + "score": 0.88, + "content": "t = 1 5", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 607, + 505, + 620 + ], + "score": 1.0, + "content": "); (2) Transfer Setting: The number of rollout timesteps is higher", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 106, + 618, + 279, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 244, + 630 + ], + "score": 1.0, + "content": "than that seen during training (i.e.", + "type": "text" + }, + { + "bbox": [ + 244, + 618, + 273, + 628 + ], + "score": 0.87, + "content": "t = 3 0", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 618, + 279, + 630 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 68 + } + ], + "index": 67 + }, + { + "type": "text", + "bbox": [ + 106, + 634, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 634, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 505, + 646 + ], + "score": 1.0, + "content": "We present our results on the shapes stack dataset in Table 3. We can see that both PNPS and SNPS", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 106, + 645, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 505, + 657 + ], + "score": 1.0, + "content": "outperform the baseline RPIN in the transfer setting, while only PNPS outperforms the baseline in the", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "test setting and SNPS fails to do so. We can see that PNPS outperforms SNPS. We attribute this to", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 106, + 668, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 506, + 680 + ], + "score": 1.0, + "content": "the reduced application sparsity with PNPS, i.e., it is more likely that the state of a slot gets updated", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 105, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 400, + 690 + ], + "score": 1.0, + "content": "in PNPS as compared to SNPS. For instance, consider an NPS model with", + "type": "text" + }, + { + "bbox": [ + 400, + 678, + 412, + 688 + ], + "score": 0.76, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "uniformly chosen rules", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 104, + 688, + 504, + 703 + ], + "spans": [ + { + "bbox": [ + 104, + 688, + 123, + 703 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 689, + 137, + 699 + ], + "score": 0.62, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 688, + 414, + 703 + ], + "score": 1.0, + "content": "slots. The probability that the state of a slot gets updated in PNPS is", + "type": "text" + }, + { + "bbox": [ + 415, + 689, + 504, + 700 + ], + "score": 0.92, + "content": "P _ { P N P S } \\dot { = } N - 1 / N", + "type": "inline_equation" + } + ], + "index": 74 + }, + { + "bbox": [ + 104, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 104, + 699, + 383, + 713 + ], + "score": 1.0, + "content": "(since 1 rule is the null rule), while the same probability for SNPS is", + "type": "text" + }, + { + "bbox": [ + 383, + 700, + 451, + 712 + ], + "score": 0.92, + "content": "P _ { S N P S } = 1 / M", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "(since only 1", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 105, + 711, + 280, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 280, + 723 + ], + "score": 1.0, + "content": "slot gets updated per rule application step).", + "type": "text" + } + ], + "index": 76 + } + ], + "index": 72.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 73, + 505, + 117 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 105, + 72, + 505, + 118 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 122, + 339, + 264 + ], + "lines": [ + { + "bbox": [ + 106, + 121, + 339, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 121, + 339, + 134 + ], + "score": 1.0, + "content": "We show the segregation of rules for NPS and the baseline", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 131, + 340, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 340, + 146 + ], + "score": 1.0, + "content": "in Figure 4. We can see that NPS learns to use a unique", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 143, + 339, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 339, + 156 + ], + "score": 1.0, + "content": "rule for each operation while the baseline struggles to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 154, + 339, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 339, + 167 + ], + "score": 1.0, + "content": "disentangle the underlying operations properly. NPS", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 166, + 340, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 340, + 177 + ], + "score": 1.0, + "content": "also outperforms the baseline in terms of MSE achiev-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 177, + 339, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 169, + 189 + ], + "score": 1.0, + "content": "ing an MSE of", + "type": "text" + }, + { + "bbox": [ + 170, + 177, + 211, + 188 + ], + "score": 0.9, + "content": "0 . 0 1 _ { \\pm 0 . 0 0 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 177, + 339, + 189 + ], + "score": 1.0, + "content": "while the baseline achieves an", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 186, + 340, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 140, + 200 + ], + "score": 1.0, + "content": "MSE of", + "type": "text" + }, + { + "bbox": [ + 140, + 187, + 181, + 199 + ], + "score": 0.9, + "content": "0 . 0 4 _ { \\pm 0 . 0 0 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 186, + 340, + 200 + ], + "score": 1.0, + "content": ". To further confirm that NPS learns all", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 198, + 340, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 340, + 210 + ], + "score": 1.0, + "content": "the available operations correctly from raw data demon-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 209, + 339, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 339, + 221 + ], + "score": 1.0, + "content": "strations, we use an NPS model with 5 rules. We expect", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 219, + 339, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 339, + 232 + ], + "score": 1.0, + "content": "that in this case NPS should utilize only 4 rules since", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 231, + 339, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 339, + 243 + ], + "score": 1.0, + "content": "the data describes only 4 unique operations and indeed", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 242, + 339, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 339, + 254 + ], + "score": 1.0, + "content": "we observe that NPS ends up mostly utilizing 4 of the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 253, + 262, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 262, + 264 + ], + "score": 1.0, + "content": "available 5 rules as shown in Table 2.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 121, + 340, + 264 + ] + }, + { + "type": "text", + "bbox": [ + 347, + 121, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 345, + 119, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 345, + 119, + 505, + 133 + ], + "score": 1.0, + "content": "Table 2: This table shows segregation", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 346, + 131, + 505, + 143 + ], + "spans": [ + { + "bbox": [ + 346, + 131, + 505, + 143 + ], + "score": 1.0, + "content": "of rules when we use NPS with 5 rules", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 345, + 142, + 506, + 154 + ], + "spans": [ + { + "bbox": [ + 345, + 142, + 506, + 154 + ], + "score": 1.0, + "content": "but the data generation distributions de-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 346, + 154, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 346, + 154, + 505, + 165 + ], + "score": 1.0, + "content": "scribes only 4 possible operations. We", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 346, + 164, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 346, + 164, + 506, + 176 + ], + "score": 1.0, + "content": "can see that only 4 rules get majorly uti-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 345, + 174, + 507, + 187 + ], + "spans": [ + { + "bbox": [ + 345, + 174, + 507, + 187 + ], + "score": 1.0, + "content": "lized thus confirming that NPS success-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 346, + 186, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 346, + 186, + 506, + 199 + ], + "score": 1.0, + "content": "fully recovers all possible operations de-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 347, + 198, + 427, + 209 + ], + "spans": [ + { + "bbox": [ + 347, + 198, + 427, + 209 + ], + "score": 1.0, + "content": "scribed by the data.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20.5, + "bbox_fs": [ + 345, + 119, + 507, + 209 + ] + }, + { + "type": "table", + "bbox": [ + 349, + 216, + 500, + 270 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 349, + 216, + 500, + 270 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 349, + 216, + 500, + 270 + ], + "spans": [ + { + "bbox": [ + 349, + 216, + 500, + 270 + ], + "score": 0.974, + "html": "
Rule 1Rule 2Rule 3Rule 4Rule 5
X Addition3609945130
X Subtraction0482010
Y Subtraction03945320
Y Addition0571599335
", + "type": "table", + "image_path": "a8f84dde9bca14de9d9314267a57d24b53ba78f2f3b4666f783cd97b86554942.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 349, + 216, + 500, + 243.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 349, + 243.0, + 500, + 270.0 + ], + "spans": [], + "index": 26 + } + ] + } + ], + "index": 25.5 + }, + { + "type": "title", + "bbox": [ + 108, + 278, + 297, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 298, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 298, + 294 + ], + "score": 1.0, + "content": "4.2 Parallel vs Sequential Rule Application", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 108, + 299, + 503, + 332 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "We compare the parallel and sequential rule application procedures, to understand the settings that", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 310, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 505, + 322 + ], + "score": 1.0, + "content": "favour one or the other, over two tasks: (1) Bouncing Balls, (2) Shapes Stack. We use the term PNPS", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 321, + 443, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 443, + 334 + ], + "score": 1.0, + "content": "to refer to parallel rule application and SNPS to refer to sequential rule application.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 299, + 506, + 334 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 341, + 339, + 503 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 340, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 340, + 352 + ], + "score": 1.0, + "content": "Shapes Stack. We use the shapes stack dataset intro-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 351, + 340, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 340, + 363 + ], + "score": 1.0, + "content": "duced by Groth et al. (2018). This dataset consists of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 362, + 339, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 339, + 374 + ], + "score": 1.0, + "content": "objects stacked on top of each other as shown in Figure", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 372, + 340, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 340, + 385 + ], + "score": 1.0, + "content": "5. These objects fall under the influence of gravity. For", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 384, + 340, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 340, + 395 + ], + "score": 1.0, + "content": "our experiments, We follow the same setup as Qi et al.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 394, + 340, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 340, + 407 + ], + "score": 1.0, + "content": "(2021). In this task, given the first frame, the model is", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 405, + 340, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 340, + 417 + ], + "score": 1.0, + "content": "tasked with predicting the object bounding boxes for the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 417, + 341, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 125, + 428 + ], + "score": 1.0, + "content": "next", + "type": "text" + }, + { + "bbox": [ + 126, + 417, + 132, + 426 + ], + "score": 0.51, + "content": "\\pmb { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 417, + 341, + 428 + ], + "score": 1.0, + "content": "timesteps. The first frame is encoded using a convo-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 427, + 340, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 340, + 439 + ], + "score": 1.0, + "content": "lutional network followed by RoIPooling (Girshick (2015))", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 439, + 339, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 339, + 450 + ], + "score": 1.0, + "content": "to extract object-centric visual features. The object-centric", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 449, + 339, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 339, + 461 + ], + "score": 1.0, + "content": "features are then passed to the dynamics model to predict", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 460, + 340, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 244, + 472 + ], + "score": 1.0, + "content": "object bounding boxes of the next", + "type": "text" + }, + { + "bbox": [ + 244, + 461, + 250, + 470 + ], + "score": 0.37, + "content": "\\pmb { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 460, + 340, + 472 + ], + "score": 1.0, + "content": "steps. Qi et al. (2021)", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 471, + 340, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 340, + 483 + ], + "score": 1.0, + "content": "propose a Region Proposal Interaction Network (RPIN) to", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 482, + 340, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 340, + 493 + ], + "score": 1.0, + "content": "solve this task. The dynamics model in RPIN consists of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 492, + 341, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 341, + 505 + ], + "score": 1.0, + "content": "an Interaction Network proposed in Battaglia et al. (2016).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 340, + 341, + 505 + ] + }, + { + "type": "table", + "bbox": [ + 350, + 342, + 500, + 389 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 350, + 342, + 500, + 389 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 350, + 342, + 500, + 389 + ], + "spans": [ + { + "bbox": [ + 350, + 342, + 500, + 389 + ], + "score": 0.971, + "html": "
Model NameTestTransfer
RPIN (Qi et al. (2021))1.254±0.0086.377±0.325
PNPSSNPS1.250±0.0075.411±0.45
1.68±0.025.80±0.15
", + "type": "table", + "image_path": "6667bcdf1c3ef24b7d961678d2db241d0c4a6cfee87cc246cf897591d179148d.jpg" + } + ] + } + ], + "index": 46.5, + "virtual_lines": [ + { + "bbox": [ + 350, + 342, + 500, + 365.5 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 350, + 365.5, + 500, + 389.0 + ], + "spans": [], + "index": 47 + } + ] + } + ], + "index": 46.5 + }, + { + "type": "text", + "bbox": [ + 347, + 394, + 505, + 503 + ], + "lines": [ + { + "bbox": [ + 346, + 393, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 346, + 393, + 506, + 405 + ], + "score": 1.0, + "content": "Table 3: Prediction error of the com-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 345, + 405, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 345, + 405, + 505, + 416 + ], + "score": 1.0, + "content": "pared models on the shapes stack", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 346, + 415, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 346, + 415, + 505, + 427 + ], + "score": 1.0, + "content": "dataset (lower is better) for the test as", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 346, + 426, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 346, + 426, + 506, + 438 + ], + "score": 1.0, + "content": "well as transfer setting. In the test set-", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 346, + 437, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 346, + 437, + 475, + 449 + ], + "score": 1.0, + "content": "ting the number of rollout steps", + "type": "text" + }, + { + "bbox": [ + 475, + 438, + 481, + 447 + ], + "score": 0.71, + "content": "\\pmb { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 437, + 505, + 449 + ], + "score": 1.0, + "content": "is set", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 345, + 448, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 345, + 448, + 505, + 460 + ], + "score": 1.0, + "content": "to 15 and in the transfer setting it is set", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 345, + 459, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 345, + 459, + 506, + 471 + ], + "score": 1.0, + "content": "to 30. We can see that PNPS outper-", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 345, + 470, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 345, + 470, + 505, + 482 + ], + "score": 1.0, + "content": "forms the RPIN baseline in both the test", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 346, + 481, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 346, + 481, + 505, + 493 + ], + "score": 1.0, + "content": "and transfer setting while SNPS fails to", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 347, + 492, + 471, + 503 + ], + "spans": [ + { + "bbox": [ + 347, + 492, + 471, + 503 + ], + "score": 1.0, + "content": "do so. Results across 15 seeds.", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 52.5, + "bbox_fs": [ + 345, + 393, + 506, + 503 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 504, + 505, + 591 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 505, + 517 + ], + "score": 1.0, + "content": "To better utilize spatial information, Qi et al. (2021) propose an extension of the interaction operators", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 515, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 526 + ], + "score": 1.0, + "content": "in interaction net to operate on 3D tensors. This is achieved by replacing the MLP operations in the", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 525, + 504, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 504, + 538 + ], + "score": 1.0, + "content": "original interaction networks with convolutions. They call this new network Convolutional Interaction", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 536, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 506, + 549 + ], + "score": 1.0, + "content": "Network (CIN). For the proposed model, we replace this CIN in RPIN by NPS. To ensure a fair", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "comparison to CIN, we use CNNs to represent rules in NPS instead of MLPs. CIN captures all", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "score": 1.0, + "content": "pairwise interactions between objects using a convolutional network. In NPS, we capture sparse", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 105, + 569, + 507, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 507, + 582 + ], + "score": 1.0, + "content": "interactions (contextual sparsity) as compared to dense pairwise interactions captured by CIN. Also,", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 580, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 481, + 592 + ], + "score": 1.0, + "content": "in NPS we update only a few subset of slots per step instead of all slots (application sparsity).", + "type": "text" + } + ], + "index": 65 + } + ], + "index": 61.5, + "bbox_fs": [ + 105, + 502, + 507, + 592 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 596, + 502, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 609 + ], + "score": 1.0, + "content": "We consider two evaluation settings. (1) Test setting: The number of rollout timesteps is same as that", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 105, + 607, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 208, + 620 + ], + "score": 1.0, + "content": "seen during training (i.e.", + "type": "text" + }, + { + "bbox": [ + 209, + 608, + 238, + 618 + ], + "score": 0.88, + "content": "t = 1 5", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 607, + 505, + 620 + ], + "score": 1.0, + "content": "); (2) Transfer Setting: The number of rollout timesteps is higher", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 106, + 618, + 279, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 244, + 630 + ], + "score": 1.0, + "content": "than that seen during training (i.e.", + "type": "text" + }, + { + "bbox": [ + 244, + 618, + 273, + 628 + ], + "score": 0.87, + "content": "t = 3 0", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 618, + 279, + 630 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 68 + } + ], + "index": 67, + "bbox_fs": [ + 105, + 595, + 505, + 630 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 634, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 634, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 505, + 646 + ], + "score": 1.0, + "content": "We present our results on the shapes stack dataset in Table 3. We can see that both PNPS and SNPS", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 106, + 645, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 505, + 657 + ], + "score": 1.0, + "content": "outperform the baseline RPIN in the transfer setting, while only PNPS outperforms the baseline in the", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "test setting and SNPS fails to do so. We can see that PNPS outperforms SNPS. We attribute this to", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 106, + 668, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 506, + 680 + ], + "score": 1.0, + "content": "the reduced application sparsity with PNPS, i.e., it is more likely that the state of a slot gets updated", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 105, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 400, + 690 + ], + "score": 1.0, + "content": "in PNPS as compared to SNPS. For instance, consider an NPS model with", + "type": "text" + }, + { + "bbox": [ + 400, + 678, + 412, + 688 + ], + "score": 0.76, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "uniformly chosen rules", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 104, + 688, + 504, + 703 + ], + "spans": [ + { + "bbox": [ + 104, + 688, + 123, + 703 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 689, + 137, + 699 + ], + "score": 0.62, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 688, + 414, + 703 + ], + "score": 1.0, + "content": "slots. The probability that the state of a slot gets updated in PNPS is", + "type": "text" + }, + { + "bbox": [ + 415, + 689, + 504, + 700 + ], + "score": 0.92, + "content": "P _ { P N P S } \\dot { = } N - 1 / N", + "type": "inline_equation" + } + ], + "index": 74 + }, + { + "bbox": [ + 104, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 104, + 699, + 383, + 713 + ], + "score": 1.0, + "content": "(since 1 rule is the null rule), while the same probability for SNPS is", + "type": "text" + }, + { + "bbox": [ + 383, + 700, + 451, + 712 + ], + "score": 0.92, + "content": "P _ { S N P S } = 1 / M", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "(since only 1", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 105, + 711, + 280, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 280, + 723 + ], + "score": 1.0, + "content": "slot gets updated per rule application step).", + "type": "text" + } + ], + "index": 76 + } + ], + "index": 72.5, + "bbox_fs": [ + 104, + 634, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 247 + ], + "lines": [ + { + "bbox": [ + 105, + 71, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 71, + 301, + 86 + ], + "score": 1.0, + "content": "For this task, we run both PNPS and SNPS for", + "type": "text" + }, + { + "bbox": [ + 302, + 73, + 370, + 85 + ], + "score": 0.93, + "content": "N = \\{ 1 , 2 , 4 , 6 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 71, + 410, + 86 + ], + "score": 1.0, + "content": "rules and", + "type": "text" + }, + { + "bbox": [ + 411, + 73, + 443, + 83 + ], + "score": 0.89, + "content": "M = 3", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 71, + 506, + 86 + ], + "score": 1.0, + "content": ". For any given", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 107, + 84, + 118, + 94 + ], + "score": 0.76, + "content": "_ N", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 83, + 185, + 97 + ], + "score": 1.0, + "content": ", we observe that", + "type": "text" + }, + { + "bbox": [ + 186, + 84, + 262, + 95 + ], + "score": 0.91, + "content": "P _ { P N P S } > P _ { S N P S }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 83, + 506, + 97 + ], + "score": 1.0, + "content": ". Even when we have multiple rule application steps in SNPS,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 505, + 107 + ], + "score": 1.0, + "content": "it might end up selecting the same slot to be updated in more than one of these steps. We report the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 345, + 119 + ], + "score": 1.0, + "content": "best performance obtained for PNPS and SNPS across all", + "type": "text" + }, + { + "bbox": [ + 346, + 106, + 357, + 116 + ], + "score": 0.76, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 105, + 397, + 119 + ], + "score": 1.0, + "content": ", which is", + "type": "text" + }, + { + "bbox": [ + 398, + 105, + 450, + 118 + ], + "score": 0.83, + "content": "N = \\{ 2 + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 105, + 505, + 119 + ], + "score": 1.0, + "content": "Null Rule}", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 115, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 167, + 129 + ], + "score": 1.0, + "content": "for PNPS and", + "type": "text" + }, + { + "bbox": [ + 168, + 117, + 199, + 127 + ], + "score": 0.9, + "content": "N = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 115, + 506, + 129 + ], + "score": 1.0, + "content": "for SNPS, in Table 3. Shapes stack is a dataset that would prefer a model", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 127, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 506, + 140 + ], + "score": 1.0, + "content": "with less application sparsity since all the objects are tightly bound to each other (objects are placed", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "on top of each other), therefore all objects spend the majority of their time interacting with the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 149, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 161 + ], + "score": 1.0, + "content": "objects directly above or below them. We attribute the higher performance of PNPS compared to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "RPIN to the higher contextual sparsity of PNPS. Each example in the shapes stack task consists", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "of 3 objects. Even though the blocks are tightly bound to each other, each block is only affected", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "by the objects it is in direct contact with. For example, the top-most object is only affected by", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "the object directly below it. The contextual sparsity offered by PNPS is a strong inductive bias to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 204, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 506, + 216 + ], + "score": 1.0, + "content": "model such sparse interactions while RPIN models all pairwise interactions between the objects.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 215, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 505, + 226 + ], + "score": 1.0, + "content": "Figure 5 shows an intuitive illustration of the PNPS model for the shapes stack dataset. In the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "figure, Rule 2 actually refers to the Null Rule, while Rule 1 refers to all the other non-null rules. The", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 235, + 507, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 507, + 250 + ], + "score": 1.0, + "content": "bottom-most block picks the Null Rule most times, as the bottom-most block generally does not move.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 297, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 298, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 298, + 276 + ], + "score": 1.0, + "content": "Bouncing Balls. We consider a bouncing-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 274, + 297, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 297, + 286 + ], + "score": 1.0, + "content": "balls environment in which multiple balls move", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 285, + 297, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 297, + 296 + ], + "score": 1.0, + "content": "with billiard-ball dynamics. We validate our", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 296, + 297, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 297, + 307 + ], + "score": 1.0, + "content": "model on a colored version of this dataset. This", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 307, + 297, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 297, + 318 + ], + "score": 1.0, + "content": "is a next-step prediction task in which the model", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 318, + 298, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 298, + 330 + ], + "score": 1.0, + "content": "is tasked with predicting the final binary mask of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 329, + 298, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 298, + 341 + ], + "score": 1.0, + "content": "each ball. We compare the following methods:", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 340, + 297, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 297, + 352 + ], + "score": 1.0, + "content": "(a) SCOFF (Goyal et al., 2020): factorization", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 350, + 298, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 298, + 363 + ], + "score": 1.0, + "content": "of knowledge in terms of slots (object proper-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 361, + 298, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 298, + 374 + ], + "score": 1.0, + "content": "ties) and schemata, the latter capturing object", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 372, + 298, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 163, + 385 + ], + "score": 1.0, + "content": "dynamics; (b)", + "type": "text" + }, + { + "bbox": [ + 163, + 372, + 205, + 383 + ], + "score": 0.78, + "content": "{ \\mathrm { S C O F F } } { + } { + }", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 372, + 298, + 385 + ], + "score": 1.0, + "content": ": we extend SCOFF by", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 384, + 299, + 395 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 299, + 395 + ], + "score": 1.0, + "content": "using the idea of iterative competition as pro-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 394, + 298, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 298, + 406 + ], + "score": 1.0, + "content": "posed in slot attention (SA) (Locatello et al.,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 404, + 298, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 170, + 417 + ], + "score": 1.0, + "content": "2020a); SCOFF", + "type": "text" + }, + { + "bbox": [ + 170, + 406, + 178, + 415 + ], + "score": 0.53, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 404, + 298, + 417 + ], + "score": 1.0, + "content": "PNPS/SNPS: We replace pair-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 416, + 297, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 231, + 427 + ], + "score": 1.0, + "content": "wise slot-to-slot interaction in", + "type": "text" + }, + { + "bbox": [ + 231, + 416, + 276, + 427 + ], + "score": 0.86, + "content": "{ \\mathrm { S C O F F } } { + } { + }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 416, + 297, + 427 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 427, + 298, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 298, + 439 + ], + "score": 1.0, + "content": "parallel or sequential rule application. For com-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 438, + 297, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 297, + 450 + ], + "score": 1.0, + "content": "paring different methods, we use the Adjusted", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 448, + 297, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 297, + 461 + ], + "score": 1.0, + "content": "Rand Index or ARI (Rand, 1971). To investigate", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 460, + 297, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 297, + 471 + ], + "score": 1.0, + "content": "how the factorization in the form of rules allows", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 471, + 297, + 482 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 297, + 482 + ], + "score": 1.0, + "content": "for extrapolating knowledge from fewer to more", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 482, + 297, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 297, + 493 + ], + "score": 1.0, + "content": "objects, we increase the number of objects from", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 492, + 261, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 261, + 506 + ], + "score": 1.0, + "content": "4 during training to 6-8 during testing.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 26.5 + }, + { + "type": "image", + "bbox": [ + 305, + 250, + 484, + 361 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 305, + 250, + 484, + 361 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 305, + 250, + 484, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 250, + 484, + 361 + ], + "score": 0.966, + "type": "image", + "image_path": "0bafd38ea5e1df32a31ff88f73b7b52d92be55ef3f1eb31c0228c553cd29402e.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 305, + 250, + 484, + 262.3333333333333 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 305, + 262.3333333333333, + 484, + 274.66666666666663 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 305, + 274.66666666666663, + 484, + 286.99999999999994 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 305, + 286.99999999999994, + 484, + 299.33333333333326 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 305, + 299.33333333333326, + 484, + 311.6666666666666 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 305, + 311.6666666666666, + 484, + 323.9999999999999 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 305, + 323.9999999999999, + 484, + 336.3333333333332 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 305, + 336.3333333333332, + 484, + 348.6666666666665 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 305, + 348.6666666666665, + 484, + 360.99999999999983 + ], + "spans": [], + "index": 46 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 305, + 367, + 505, + 530 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 303, + 366, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 303, + 366, + 506, + 379 + ], + "score": 1.0, + "content": "Figure 5: Here we show the rule selection statis-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "tics from the proposed model for all entities in the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 304, + 389, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 506, + 400 + ], + "score": 1.0, + "content": "shapes stack dataset across all examples. Each ex-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 399, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 505, + 410 + ], + "score": 1.0, + "content": "ample contains 3 entities as shown above. Each", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 304, + 411, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 505, + 422 + ], + "score": 1.0, + "content": "cell in the table shows the probability with which", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 303, + 421, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 303, + 421, + 505, + 434 + ], + "score": 1.0, + "content": "the given rule is triggered for the corresponding", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 433, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 505, + 444 + ], + "score": 1.0, + "content": "entity. We can see that the bottom-most entity", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 303, + 443, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 303, + 443, + 506, + 455 + ], + "score": 1.0, + "content": "triggers rule 2 most of the time while the other 2", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 303, + 454, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 303, + 454, + 506, + 466 + ], + "score": 1.0, + "content": "entities trigger rule 1 most often. This is quite intu-", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 303, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 303, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "itive as, for most examples, the bottom-most entity", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 303, + 476, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 303, + 476, + 505, + 487 + ], + "score": 1.0, + "content": "remains static and does not move at all while the", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 303, + 487, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 303, + 487, + 506, + 499 + ], + "score": 1.0, + "content": "upper entities fall. Therefore, rule 2 captures in-", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 304, + 498, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 505, + 509 + ], + "score": 1.0, + "content": "formation which is relevant to static entities, while", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 303, + 509, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 303, + 509, + 506, + 520 + ], + "score": 1.0, + "content": "rule 1 captures physical rules relevant to the inter-", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 53.5 + } + ], + "index": 47.75 + }, + { + "type": "text", + "bbox": [ + 106, + 525, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 107, + 522, + 467, + 532 + ], + "spans": [ + { + "bbox": [ + 107, + 522, + 467, + 532 + ], + "score": 1.0, + "content": "actions and motion of the upper entities.Table 4. Contrary to the shapes stack task, we", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "see that SNPS outperforms PNPS for the bouncing balls task. The balls are not tightly bound", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 106, + 542, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 554 + ], + "score": 1.0, + "content": "together into a single tower as in the shapes stack. Most of the time, a single ball follows its own", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 106, + 552, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 506, + 566 + ], + "score": 1.0, + "content": "dynamics, only occasionally interacting with another ball. Rules in NPS capture interaction dynamics", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "score": 1.0, + "content": "between entities, hence they would only be required to change the state of an entity when it interacts", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 106, + 575, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 505, + 586 + ], + "score": 1.0, + "content": "with another entity. In the case of bouncing balls, this interaction takes place through a collision", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 104, + 584, + 506, + 598 + ], + "spans": [ + { + "bbox": [ + 104, + 584, + 506, + 598 + ], + "score": 1.0, + "content": "between multiple balls. Since for a single ball, such collisions are rare, SNPS, which has higher", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 105, + 595, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 610 + ], + "score": 1.0, + "content": "application sparsity (less probability of modifying the state of an entity), performs better as compared", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "score": 1.0, + "content": "to PNPS (lower application sparsity). Also note that, SNPS has the ability to compose multiple rules", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 105, + 617, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 632 + ], + "score": 1.0, + "content": "together by virtue of having multiple rule application stages. A visualization of the rule and entity", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 106, + 628, + 400, + 642 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 400, + 642 + ], + "score": 1.0, + "content": "selections by the proposed algorithm can be found in Appendix Figure 9.", + "type": "text" + } + ], + "index": 71 + } + ], + "index": 66 + }, + { + "type": "text", + "bbox": [ + 107, + 645, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "Given the analysis in this section, we can conclude that PNPS is expected to work better when", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "interactions among entities are more frequent while SNPS is expected to work better when interactions", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "are rare and most of the time, each entity follows its own dynamics. Note that, for both SNPS and", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "PNPS, the rule application considers only 1 other entity as context. Therefore, both approaches have", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "equal contextual sparsity while the baselines that we consider (SCOFF and RPIN) capture dense", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 105, + 700, + 507, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 507, + 712 + ], + "score": 1.0, + "content": "pairwise interactions. We discuss the benefits of contextual sparsity in more detail in the next section.", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 105, + 711, + 453, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 453, + 724 + ], + "score": 1.0, + "content": "More details regarding our setup for the above experiments can be found in Appendix.", + "type": "text" + } + ], + "index": 78 + } + ], + "index": 75 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 247 + ], + "lines": [ + { + "bbox": [ + 105, + 71, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 71, + 301, + 86 + ], + "score": 1.0, + "content": "For this task, we run both PNPS and SNPS for", + "type": "text" + }, + { + "bbox": [ + 302, + 73, + 370, + 85 + ], + "score": 0.93, + "content": "N = \\{ 1 , 2 , 4 , 6 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 71, + 410, + 86 + ], + "score": 1.0, + "content": "rules and", + "type": "text" + }, + { + "bbox": [ + 411, + 73, + 443, + 83 + ], + "score": 0.89, + "content": "M = 3", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 71, + 506, + 86 + ], + "score": 1.0, + "content": ". For any given", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 107, + 84, + 118, + 94 + ], + "score": 0.76, + "content": "_ N", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 83, + 185, + 97 + ], + "score": 1.0, + "content": ", we observe that", + "type": "text" + }, + { + "bbox": [ + 186, + 84, + 262, + 95 + ], + "score": 0.91, + "content": "P _ { P N P S } > P _ { S N P S }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 83, + 506, + 97 + ], + "score": 1.0, + "content": ". Even when we have multiple rule application steps in SNPS,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 505, + 107 + ], + "score": 1.0, + "content": "it might end up selecting the same slot to be updated in more than one of these steps. We report the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 345, + 119 + ], + "score": 1.0, + "content": "best performance obtained for PNPS and SNPS across all", + "type": "text" + }, + { + "bbox": [ + 346, + 106, + 357, + 116 + ], + "score": 0.76, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 105, + 397, + 119 + ], + "score": 1.0, + "content": ", which is", + "type": "text" + }, + { + "bbox": [ + 398, + 105, + 450, + 118 + ], + "score": 0.83, + "content": "N = \\{ 2 + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 105, + 505, + 119 + ], + "score": 1.0, + "content": "Null Rule}", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 115, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 167, + 129 + ], + "score": 1.0, + "content": "for PNPS and", + "type": "text" + }, + { + "bbox": [ + 168, + 117, + 199, + 127 + ], + "score": 0.9, + "content": "N = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 115, + 506, + 129 + ], + "score": 1.0, + "content": "for SNPS, in Table 3. Shapes stack is a dataset that would prefer a model", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 127, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 506, + 140 + ], + "score": 1.0, + "content": "with less application sparsity since all the objects are tightly bound to each other (objects are placed", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "on top of each other), therefore all objects spend the majority of their time interacting with the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 149, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 161 + ], + "score": 1.0, + "content": "objects directly above or below them. We attribute the higher performance of PNPS compared to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "RPIN to the higher contextual sparsity of PNPS. Each example in the shapes stack task consists", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "of 3 objects. Even though the blocks are tightly bound to each other, each block is only affected", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "by the objects it is in direct contact with. For example, the top-most object is only affected by", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "the object directly below it. The contextual sparsity offered by PNPS is a strong inductive bias to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 204, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 506, + 216 + ], + "score": 1.0, + "content": "model such sparse interactions while RPIN models all pairwise interactions between the objects.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 215, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 505, + 226 + ], + "score": 1.0, + "content": "Figure 5 shows an intuitive illustration of the PNPS model for the shapes stack dataset. In the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "figure, Rule 2 actually refers to the Null Rule, while Rule 1 refers to all the other non-null rules. The", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 235, + 507, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 507, + 250 + ], + "score": 1.0, + "content": "bottom-most block picks the Null Rule most times, as the bottom-most block generally does not move.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 71, + 507, + 250 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 297, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 298, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 298, + 276 + ], + "score": 1.0, + "content": "Bouncing Balls. We consider a bouncing-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 274, + 297, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 297, + 286 + ], + "score": 1.0, + "content": "balls environment in which multiple balls move", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 285, + 297, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 297, + 296 + ], + "score": 1.0, + "content": "with billiard-ball dynamics. We validate our", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 296, + 297, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 297, + 307 + ], + "score": 1.0, + "content": "model on a colored version of this dataset. This", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 307, + 297, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 297, + 318 + ], + "score": 1.0, + "content": "is a next-step prediction task in which the model", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 318, + 298, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 298, + 330 + ], + "score": 1.0, + "content": "is tasked with predicting the final binary mask of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 329, + 298, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 298, + 341 + ], + "score": 1.0, + "content": "each ball. We compare the following methods:", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 340, + 297, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 297, + 352 + ], + "score": 1.0, + "content": "(a) SCOFF (Goyal et al., 2020): factorization", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 350, + 298, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 298, + 363 + ], + "score": 1.0, + "content": "of knowledge in terms of slots (object proper-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 361, + 298, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 298, + 374 + ], + "score": 1.0, + "content": "ties) and schemata, the latter capturing object", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 372, + 298, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 163, + 385 + ], + "score": 1.0, + "content": "dynamics; (b)", + "type": "text" + }, + { + "bbox": [ + 163, + 372, + 205, + 383 + ], + "score": 0.78, + "content": "{ \\mathrm { S C O F F } } { + } { + }", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 372, + 298, + 385 + ], + "score": 1.0, + "content": ": we extend SCOFF by", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 384, + 299, + 395 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 299, + 395 + ], + "score": 1.0, + "content": "using the idea of iterative competition as pro-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 394, + 298, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 298, + 406 + ], + "score": 1.0, + "content": "posed in slot attention (SA) (Locatello et al.,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 404, + 298, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 170, + 417 + ], + "score": 1.0, + "content": "2020a); SCOFF", + "type": "text" + }, + { + "bbox": [ + 170, + 406, + 178, + 415 + ], + "score": 0.53, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 404, + 298, + 417 + ], + "score": 1.0, + "content": "PNPS/SNPS: We replace pair-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 416, + 297, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 231, + 427 + ], + "score": 1.0, + "content": "wise slot-to-slot interaction in", + "type": "text" + }, + { + "bbox": [ + 231, + 416, + 276, + 427 + ], + "score": 0.86, + "content": "{ \\mathrm { S C O F F } } { + } { + }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 416, + 297, + 427 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 427, + 298, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 298, + 439 + ], + "score": 1.0, + "content": "parallel or sequential rule application. For com-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 438, + 297, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 297, + 450 + ], + "score": 1.0, + "content": "paring different methods, we use the Adjusted", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 448, + 297, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 297, + 461 + ], + "score": 1.0, + "content": "Rand Index or ARI (Rand, 1971). To investigate", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 460, + 297, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 297, + 471 + ], + "score": 1.0, + "content": "how the factorization in the form of rules allows", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 471, + 297, + 482 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 297, + 482 + ], + "score": 1.0, + "content": "for extrapolating knowledge from fewer to more", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 482, + 297, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 297, + 493 + ], + "score": 1.0, + "content": "objects, we increase the number of objects from", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 492, + 261, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 261, + 506 + ], + "score": 1.0, + "content": "4 during training to 6-8 during testing.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 262, + 299, + 506 + ] + }, + { + "type": "image", + "bbox": [ + 305, + 250, + 484, + 361 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 305, + 250, + 484, + 361 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 305, + 250, + 484, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 250, + 484, + 361 + ], + "score": 0.966, + "type": "image", + "image_path": "0bafd38ea5e1df32a31ff88f73b7b52d92be55ef3f1eb31c0228c553cd29402e.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 305, + 250, + 484, + 262.3333333333333 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 305, + 262.3333333333333, + 484, + 274.66666666666663 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 305, + 274.66666666666663, + 484, + 286.99999999999994 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 305, + 286.99999999999994, + 484, + 299.33333333333326 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 305, + 299.33333333333326, + 484, + 311.6666666666666 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 305, + 311.6666666666666, + 484, + 323.9999999999999 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 305, + 323.9999999999999, + 484, + 336.3333333333332 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 305, + 336.3333333333332, + 484, + 348.6666666666665 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 305, + 348.6666666666665, + 484, + 360.99999999999983 + ], + "spans": [], + "index": 46 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 305, + 367, + 505, + 530 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 303, + 366, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 303, + 366, + 506, + 379 + ], + "score": 1.0, + "content": "Figure 5: Here we show the rule selection statis-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "tics from the proposed model for all entities in the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 304, + 389, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 506, + 400 + ], + "score": 1.0, + "content": "shapes stack dataset across all examples. Each ex-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 399, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 505, + 410 + ], + "score": 1.0, + "content": "ample contains 3 entities as shown above. Each", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 304, + 411, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 505, + 422 + ], + "score": 1.0, + "content": "cell in the table shows the probability with which", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 303, + 421, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 303, + 421, + 505, + 434 + ], + "score": 1.0, + "content": "the given rule is triggered for the corresponding", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 433, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 505, + 444 + ], + "score": 1.0, + "content": "entity. We can see that the bottom-most entity", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 303, + 443, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 303, + 443, + 506, + 455 + ], + "score": 1.0, + "content": "triggers rule 2 most of the time while the other 2", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 303, + 454, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 303, + 454, + 506, + 466 + ], + "score": 1.0, + "content": "entities trigger rule 1 most often. This is quite intu-", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 303, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 303, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "itive as, for most examples, the bottom-most entity", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 303, + 476, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 303, + 476, + 505, + 487 + ], + "score": 1.0, + "content": "remains static and does not move at all while the", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 303, + 487, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 303, + 487, + 506, + 499 + ], + "score": 1.0, + "content": "upper entities fall. Therefore, rule 2 captures in-", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 304, + 498, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 505, + 509 + ], + "score": 1.0, + "content": "formation which is relevant to static entities, while", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 303, + 509, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 303, + 509, + 506, + 520 + ], + "score": 1.0, + "content": "rule 1 captures physical rules relevant to the inter-", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 53.5 + } + ], + "index": 47.75 + }, + { + "type": "text", + "bbox": [ + 106, + 525, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 107, + 522, + 467, + 532 + ], + "spans": [ + { + "bbox": [ + 107, + 522, + 467, + 532 + ], + "score": 1.0, + "content": "actions and motion of the upper entities.Table 4. Contrary to the shapes stack task, we", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "see that SNPS outperforms PNPS for the bouncing balls task. The balls are not tightly bound", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 106, + 542, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 554 + ], + "score": 1.0, + "content": "together into a single tower as in the shapes stack. Most of the time, a single ball follows its own", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 106, + 552, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 506, + 566 + ], + "score": 1.0, + "content": "dynamics, only occasionally interacting with another ball. Rules in NPS capture interaction dynamics", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "score": 1.0, + "content": "between entities, hence they would only be required to change the state of an entity when it interacts", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 106, + 575, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 505, + 586 + ], + "score": 1.0, + "content": "with another entity. In the case of bouncing balls, this interaction takes place through a collision", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 104, + 584, + 506, + 598 + ], + "spans": [ + { + "bbox": [ + 104, + 584, + 506, + 598 + ], + "score": 1.0, + "content": "between multiple balls. Since for a single ball, such collisions are rare, SNPS, which has higher", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 105, + 595, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 610 + ], + "score": 1.0, + "content": "application sparsity (less probability of modifying the state of an entity), performs better as compared", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "score": 1.0, + "content": "to PNPS (lower application sparsity). Also note that, SNPS has the ability to compose multiple rules", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 105, + 617, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 632 + ], + "score": 1.0, + "content": "together by virtue of having multiple rule application stages. A visualization of the rule and entity", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 106, + 628, + 400, + 642 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 400, + 642 + ], + "score": 1.0, + "content": "selections by the proposed algorithm can be found in Appendix Figure 9.", + "type": "text" + } + ], + "index": 71 + } + ], + "index": 66, + "bbox_fs": [ + 104, + 522, + 506, + 642 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 645, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "Given the analysis in this section, we can conclude that PNPS is expected to work better when", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "interactions among entities are more frequent while SNPS is expected to work better when interactions", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "are rare and most of the time, each entity follows its own dynamics. Note that, for both SNPS and", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "PNPS, the rule application considers only 1 other entity as context. Therefore, both approaches have", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "equal contextual sparsity while the baselines that we consider (SCOFF and RPIN) capture dense", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 105, + 700, + 507, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 507, + 712 + ], + "score": 1.0, + "content": "pairwise interactions. We discuss the benefits of contextual sparsity in more detail in the next section.", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 105, + 711, + 453, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 453, + 724 + ], + "score": 1.0, + "content": "More details regarding our setup for the above experiments can be found in Appendix.", + "type": "text" + } + ], + "index": 78 + } + ], + "index": 75, + "bbox_fs": [ + 105, + 645, + 507, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 72, + 331, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 331, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 331, + 87 + ], + "score": 1.0, + "content": "4.3 Benefits of Sparse Interactions Offered by NPS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 92, + 505, + 203 + ], + "lines": [ + { + "bbox": [ + 105, + 92, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 505, + 105 + ], + "score": 1.0, + "content": "In NPS, one can view the computational graph as a dynamically constructed GNN resulting from", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 506, + 116 + ], + "score": 1.0, + "content": "applying dynamically selected rules, where the states of the slots are represented on the different", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "nodes of the graph, and different rules dynamically instantiate an hyper-edge between a set of slots", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 505, + 138 + ], + "score": 1.0, + "content": "(the primary slot and the contextual slot). It is important to emphasize that the topology of the graph", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 135, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 104, + 135, + 506, + 150 + ], + "score": 1.0, + "content": "induced in NPS is dynamic and sparse (only a few nodes affected), while in most GNNs the topology", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 146, + 506, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 506, + 160 + ], + "score": 1.0, + "content": "is fixed and dense (all nodes affected). In this section, through a thorough set of experiments, we", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 505, + 170 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 505, + 170 + ], + "score": 1.0, + "content": "show that learning sparse and dynamic interactions using NPS indeed works better for the problems", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 169, + 506, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 506, + 182 + ], + "score": 1.0, + "content": "we consider than learning dense interactions using GNNs. We consider two types of tasks: (1)", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 180, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 505, + 192 + ], + "score": 1.0, + "content": "Learning Action Conditioned World Models (2) Physical Reasoning. We use SNPS for all these", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 190, + 507, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 507, + 204 + ], + "score": 1.0, + "content": "experiments since in the environments that we consider here, interactions among entities are rare.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 224, + 339, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 225, + 340, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 340, + 236 + ], + "score": 1.0, + "content": "Learning Action-Conditioned World Models. For", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 235, + 339, + 247 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 339, + 247 + ], + "score": 1.0, + "content": "learning action-conditioned world models, we follow the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 246, + 340, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 340, + 258 + ], + "score": 1.0, + "content": "same experimental setup as Kipf et al. (2019). There-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 257, + 339, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 280, + 269 + ], + "score": 1.0, + "content": "fore, all the tasks in this section are next-", + "type": "text" + }, + { + "bbox": [ + 280, + 258, + 290, + 268 + ], + "score": 0.79, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 257, + 315, + 269 + ], + "score": 1.0, + "content": "step", + "type": "text" + }, + { + "bbox": [ + 315, + 258, + 339, + 268 + ], + "score": 0.84, + "content": "K =", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 107, + 268, + 340, + 280 + ], + "spans": [ + { + "bbox": [ + 107, + 268, + 147, + 280 + ], + "score": 0.9, + "content": "\\{ 1 , 5 , 1 0 \\} ", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 268, + 340, + 280 + ], + "score": 1.0, + "content": ") prediction tasks, given the intermediate actions,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 279, + 341, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 341, + 291 + ], + "score": 1.0, + "content": "and with the predictions being performed in the latent space.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 289, + 340, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 207, + 302 + ], + "score": 1.0, + "content": "We use the Hits at Rank 1", + "type": "text" + }, + { + "bbox": [ + 208, + 291, + 235, + 301 + ], + "score": 0.85, + "content": "( \\mathrm { H } @ 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 289, + 340, + 302 + ], + "score": 1.0, + "content": "metrics described by Kipf", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 301, + 341, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 221, + 313 + ], + "score": 1.0, + "content": "et al. (2019) for evaluation.", + "type": "text" + }, + { + "bbox": [ + 221, + 302, + 244, + 311 + ], + "score": 0.83, + "content": "\\mathrm { H @ 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 301, + 341, + 313 + ], + "score": 1.0, + "content": "is 1 for a particular ex-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 312, + 339, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 339, + 324 + ], + "score": 1.0, + "content": "ample if the predicted state representation is nearest to the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 323, + 339, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 339, + 335 + ], + "score": 1.0, + "content": "encoded true observation and 0 otherwise. We report the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 334, + 326, + 346 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 326, + 346 + ], + "score": 1.0, + "content": "average of this score over the test set (higher is better).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 353, + 215, + 498, + 271 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 353, + 215, + 498, + 271 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 353, + 215, + 498, + 271 + ], + "spans": [ + { + "bbox": [ + 353, + 215, + 498, + 271 + ], + "score": 0.973, + "html": "
Model NameTestTransfer
SCOFF0.280.15
SCOFF++0.84370.2632
PNPS (10 Rules+1 Null Rule)0.78130.1997
SNPS (10 Rules)0.85180.3553
", + "type": "table", + "image_path": "930cf3ea314d168c1dc25a6a6f25081ea7d31aebfed64db5ef0370e7e94372f0.jpg" + } + ] + } + ], + "index": 22.5, + "virtual_lines": [ + { + "bbox": [ + 353, + 215, + 498, + 243.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 353, + 243.0, + 498, + 271.0 + ], + "spans": [], + "index": 23 + } + ] + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 347, + 276, + 505, + 352 + ], + "lines": [ + { + "bbox": [ + 346, + 276, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 346, + 276, + 505, + 286 + ], + "score": 1.0, + "content": "Table 4: Here we show the ARI", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 346, + 285, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 346, + 285, + 506, + 299 + ], + "score": 1.0, + "content": "achieved by the models on the bouncing", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 346, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 346, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "balls dataset (higher is better). We can", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 346, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 346, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "see that SNPS outperforms SCOFF and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 347, + 318, + 507, + 332 + ], + "spans": [ + { + "bbox": [ + 347, + 320, + 391, + 330 + ], + "score": 0.66, + "content": "{ \\mathrm { S C O F F + + } }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 318, + 507, + 332 + ], + "score": 1.0, + "content": "while PNPS has a poor per-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 345, + 329, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 345, + 329, + 505, + 343 + ], + "score": 1.0, + "content": "formance in this task. Results average", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 345, + 342, + 408, + 352 + ], + "spans": [ + { + "bbox": [ + 345, + 342, + 408, + 352 + ], + "score": 1.0, + "content": "across 2 seeds.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 356, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "Physics Environment. The physics environment (Ke et al., 2021) simulates a simple physical", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "score": 1.0, + "content": "world. It consists of blocks of unique but unknown weights. The dynamics for the interaction between", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 378, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 506, + 391 + ], + "score": 1.0, + "content": "blocks is that the movement of heavier blocks pushes lighter blocks on their path. This rule creates an", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "acyclic causal graph between the blocks. For an accurate world model, the learner needs to infer the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 401, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 412 + ], + "score": 1.0, + "content": "correct weights through demonstrations. Interactions in this environment are sparse and only involve", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "two blocks at a time, therefore we expect NPS to outperform dense architectures like GNNs. This", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 423, + 307, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 307, + 435 + ], + "score": 1.0, + "content": "environment is demonstrated in Appendix Fig 11.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 438, + 505, + 515 + ], + "lines": [ + { + "bbox": [ + 105, + 438, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 505, + 451 + ], + "score": 1.0, + "content": "We follow the same setup as Kipf et al. (2019). We use their C-SWM model as baseline. For the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "score": 1.0, + "content": "proposed model, we only replace the GNN from C-SWM by NPS. GNNs generally share parameters", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "score": 1.0, + "content": "across edges, but in NPS each rule has separate parameters. For a fair comparison to GNN, we use", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 472, + 504, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 504, + 483 + ], + "score": 1.0, + "content": "an NPS model with 1 rule. Note that this setting is still different from GNNs as in GNNs at each", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 482, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 495 + ], + "score": 1.0, + "content": "step every slot is updated by instantiating edges between all pairs of slots, while in NPS an edge is", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 493, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 506, + 506 + ], + "score": 1.0, + "content": "dynamically instantiated between a single pair of slots and only the state of the selected slot (i.e.,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 505, + 216, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 216, + 516 + ], + "score": 1.0, + "content": "primary slot) gets updated.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 137, + 522, + 469, + 623 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 137, + 522, + 469, + 623 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 137, + 522, + 469, + 623 + ], + "spans": [ + { + "bbox": [ + 137, + 522, + 469, + 623 + ], + "score": 0.968, + "type": "image", + "image_path": "6eae7884c2ab3f56bf47fca18f661df64ab49fc50960444321e063fbe05b8b8b.jpg" + } + ] + } + ], + "index": 46, + "virtual_lines": [ + { + "bbox": [ + 137, + 522, + 469, + 555.6666666666666 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 137, + 555.6666666666666, + 469, + 589.3333333333333 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 137, + 589.3333333333333, + 469, + 622.9999999999999 + ], + "spans": [], + "index": 47 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 627, + 505, + 672 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "Figure 6: Action-Conditioned World Models, with number of future steps to be predicted for the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 639, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 505, + 650 + ], + "score": 1.0, + "content": "world-model on the horizontal axes. (a) Here we show a comparison between GNNs and the proposed", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 104, + 648, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 104, + 648, + 276, + 663 + ], + "score": 1.0, + "content": "NPS on the physics environment using the", + "type": "text" + }, + { + "bbox": [ + 276, + 649, + 299, + 659 + ], + "score": 0.81, + "content": "\\mathrm { H @ 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 648, + 505, + 663 + ], + "score": 1.0, + "content": "metric (higher is better). (b) Comparison of average", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 660, + 406, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 129, + 670 + ], + "score": 0.69, + "content": "\\mathrm { H @ 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 660, + 406, + 673 + ], + "score": 1.0, + "content": "scores across 5 Atari games for the proposed model NPS and GNN.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 49.5 + } + ], + "index": 47.75 + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "The results of our experiments are presented in Figure 6(a). We can see that NPS outperforms", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "GNNs for all rollouts. Multi-step settings are more difficult to model as errors may get compounded", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "over time steps. The sparsity of NPS (only a single slot affected per step) reduces compounding of", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 711, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 505, + 722 + ], + "score": 1.0, + "content": "errors and enhances symmetry-breaking in the assignment of transformations to rules, while in the", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 53.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 72, + 331, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 331, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 331, + 87 + ], + "score": 1.0, + "content": "4.3 Benefits of Sparse Interactions Offered by NPS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 92, + 505, + 203 + ], + "lines": [ + { + "bbox": [ + 105, + 92, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 505, + 105 + ], + "score": 1.0, + "content": "In NPS, one can view the computational graph as a dynamically constructed GNN resulting from", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 506, + 116 + ], + "score": 1.0, + "content": "applying dynamically selected rules, where the states of the slots are represented on the different", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "nodes of the graph, and different rules dynamically instantiate an hyper-edge between a set of slots", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 505, + 138 + ], + "score": 1.0, + "content": "(the primary slot and the contextual slot). It is important to emphasize that the topology of the graph", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 135, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 104, + 135, + 506, + 150 + ], + "score": 1.0, + "content": "induced in NPS is dynamic and sparse (only a few nodes affected), while in most GNNs the topology", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 146, + 506, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 506, + 160 + ], + "score": 1.0, + "content": "is fixed and dense (all nodes affected). In this section, through a thorough set of experiments, we", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 505, + 170 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 505, + 170 + ], + "score": 1.0, + "content": "show that learning sparse and dynamic interactions using NPS indeed works better for the problems", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 169, + 506, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 506, + 182 + ], + "score": 1.0, + "content": "we consider than learning dense interactions using GNNs. We consider two types of tasks: (1)", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 180, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 505, + 192 + ], + "score": 1.0, + "content": "Learning Action Conditioned World Models (2) Physical Reasoning. We use SNPS for all these", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 190, + 507, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 507, + 204 + ], + "score": 1.0, + "content": "experiments since in the environments that we consider here, interactions among entities are rare.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5.5, + "bbox_fs": [ + 104, + 92, + 507, + 204 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 224, + 339, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 225, + 340, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 340, + 236 + ], + "score": 1.0, + "content": "Learning Action-Conditioned World Models. For", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 235, + 339, + 247 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 339, + 247 + ], + "score": 1.0, + "content": "learning action-conditioned world models, we follow the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 246, + 340, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 340, + 258 + ], + "score": 1.0, + "content": "same experimental setup as Kipf et al. (2019). There-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 257, + 339, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 280, + 269 + ], + "score": 1.0, + "content": "fore, all the tasks in this section are next-", + "type": "text" + }, + { + "bbox": [ + 280, + 258, + 290, + 268 + ], + "score": 0.79, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 257, + 315, + 269 + ], + "score": 1.0, + "content": "step", + "type": "text" + }, + { + "bbox": [ + 315, + 258, + 339, + 268 + ], + "score": 0.84, + "content": "K =", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 107, + 268, + 340, + 280 + ], + "spans": [ + { + "bbox": [ + 107, + 268, + 147, + 280 + ], + "score": 0.9, + "content": "\\{ 1 , 5 , 1 0 \\} ", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 268, + 340, + 280 + ], + "score": 1.0, + "content": ") prediction tasks, given the intermediate actions,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 279, + 341, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 341, + 291 + ], + "score": 1.0, + "content": "and with the predictions being performed in the latent space.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 289, + 340, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 207, + 302 + ], + "score": 1.0, + "content": "We use the Hits at Rank 1", + "type": "text" + }, + { + "bbox": [ + 208, + 291, + 235, + 301 + ], + "score": 0.85, + "content": "( \\mathrm { H } @ 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 289, + 340, + 302 + ], + "score": 1.0, + "content": "metrics described by Kipf", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 301, + 341, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 221, + 313 + ], + "score": 1.0, + "content": "et al. (2019) for evaluation.", + "type": "text" + }, + { + "bbox": [ + 221, + 302, + 244, + 311 + ], + "score": 0.83, + "content": "\\mathrm { H @ 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 301, + 341, + 313 + ], + "score": 1.0, + "content": "is 1 for a particular ex-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 312, + 339, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 339, + 324 + ], + "score": 1.0, + "content": "ample if the predicted state representation is nearest to the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 323, + 339, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 339, + 335 + ], + "score": 1.0, + "content": "encoded true observation and 0 otherwise. We report the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 334, + 326, + 346 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 326, + 346 + ], + "score": 1.0, + "content": "average of this score over the test set (higher is better).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 225, + 341, + 346 + ] + }, + { + "type": "table", + "bbox": [ + 353, + 215, + 498, + 271 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 353, + 215, + 498, + 271 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 353, + 215, + 498, + 271 + ], + "spans": [ + { + "bbox": [ + 353, + 215, + 498, + 271 + ], + "score": 0.973, + "html": "
Model NameTestTransfer
SCOFF0.280.15
SCOFF++0.84370.2632
PNPS (10 Rules+1 Null Rule)0.78130.1997
SNPS (10 Rules)0.85180.3553
", + "type": "table", + "image_path": "930cf3ea314d168c1dc25a6a6f25081ea7d31aebfed64db5ef0370e7e94372f0.jpg" + } + ] + } + ], + "index": 22.5, + "virtual_lines": [ + { + "bbox": [ + 353, + 215, + 498, + 243.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 353, + 243.0, + 498, + 271.0 + ], + "spans": [], + "index": 23 + } + ] + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 347, + 276, + 505, + 352 + ], + "lines": [ + { + "bbox": [ + 346, + 276, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 346, + 276, + 505, + 286 + ], + "score": 1.0, + "content": "Table 4: Here we show the ARI", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 346, + 285, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 346, + 285, + 506, + 299 + ], + "score": 1.0, + "content": "achieved by the models on the bouncing", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 346, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 346, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "balls dataset (higher is better). We can", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 346, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 346, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "see that SNPS outperforms SCOFF and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 347, + 318, + 507, + 332 + ], + "spans": [ + { + "bbox": [ + 347, + 320, + 391, + 330 + ], + "score": 0.66, + "content": "{ \\mathrm { S C O F F + + } }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 318, + 507, + 332 + ], + "score": 1.0, + "content": "while PNPS has a poor per-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 345, + 329, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 345, + 329, + 505, + 343 + ], + "score": 1.0, + "content": "formance in this task. Results average", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 345, + 342, + 408, + 352 + ], + "spans": [ + { + "bbox": [ + 345, + 342, + 408, + 352 + ], + "score": 1.0, + "content": "across 2 seeds.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27, + "bbox_fs": [ + 345, + 276, + 507, + 352 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 356, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "Physics Environment. The physics environment (Ke et al., 2021) simulates a simple physical", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "score": 1.0, + "content": "world. It consists of blocks of unique but unknown weights. The dynamics for the interaction between", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 378, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 506, + 391 + ], + "score": 1.0, + "content": "blocks is that the movement of heavier blocks pushes lighter blocks on their path. This rule creates an", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "acyclic causal graph between the blocks. For an accurate world model, the learner needs to infer the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 401, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 412 + ], + "score": 1.0, + "content": "correct weights through demonstrations. Interactions in this environment are sparse and only involve", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "two blocks at a time, therefore we expect NPS to outperform dense architectures like GNNs. This", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 423, + 307, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 307, + 435 + ], + "score": 1.0, + "content": "environment is demonstrated in Appendix Fig 11.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 357, + 506, + 435 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 438, + 505, + 515 + ], + "lines": [ + { + "bbox": [ + 105, + 438, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 505, + 451 + ], + "score": 1.0, + "content": "We follow the same setup as Kipf et al. (2019). We use their C-SWM model as baseline. For the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "score": 1.0, + "content": "proposed model, we only replace the GNN from C-SWM by NPS. GNNs generally share parameters", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "score": 1.0, + "content": "across edges, but in NPS each rule has separate parameters. For a fair comparison to GNN, we use", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 472, + 504, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 504, + 483 + ], + "score": 1.0, + "content": "an NPS model with 1 rule. Note that this setting is still different from GNNs as in GNNs at each", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 482, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 495 + ], + "score": 1.0, + "content": "step every slot is updated by instantiating edges between all pairs of slots, while in NPS an edge is", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 493, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 506, + 506 + ], + "score": 1.0, + "content": "dynamically instantiated between a single pair of slots and only the state of the selected slot (i.e.,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 505, + 216, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 216, + 516 + ], + "score": 1.0, + "content": "primary slot) gets updated.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 438, + 506, + 516 + ] + }, + { + "type": "image", + "bbox": [ + 137, + 522, + 469, + 623 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 137, + 522, + 469, + 623 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 137, + 522, + 469, + 623 + ], + "spans": [ + { + "bbox": [ + 137, + 522, + 469, + 623 + ], + "score": 0.968, + "type": "image", + "image_path": "6eae7884c2ab3f56bf47fca18f661df64ab49fc50960444321e063fbe05b8b8b.jpg" + } + ] + } + ], + "index": 46, + "virtual_lines": [ + { + "bbox": [ + 137, + 522, + 469, + 555.6666666666666 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 137, + 555.6666666666666, + 469, + 589.3333333333333 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 137, + 589.3333333333333, + 469, + 622.9999999999999 + ], + "spans": [], + "index": 47 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 627, + 505, + 672 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "Figure 6: Action-Conditioned World Models, with number of future steps to be predicted for the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 639, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 505, + 650 + ], + "score": 1.0, + "content": "world-model on the horizontal axes. (a) Here we show a comparison between GNNs and the proposed", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 104, + 648, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 104, + 648, + 276, + 663 + ], + "score": 1.0, + "content": "NPS on the physics environment using the", + "type": "text" + }, + { + "bbox": [ + 276, + 649, + 299, + 659 + ], + "score": 0.81, + "content": "\\mathrm { H @ 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 648, + 505, + 663 + ], + "score": 1.0, + "content": "metric (higher is better). (b) Comparison of average", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 660, + 406, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 129, + 670 + ], + "score": 0.69, + "content": "\\mathrm { H @ 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 660, + 406, + 673 + ], + "score": 1.0, + "content": "scores across 5 Atari games for the proposed model NPS and GNN.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 49.5 + } + ], + "index": 47.75 + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "The results of our experiments are presented in Figure 6(a). We can see that NPS outperforms", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "GNNs for all rollouts. Multi-step settings are more difficult to model as errors may get compounded", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "over time steps. The sparsity of NPS (only a single slot affected per step) reduces compounding of", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 711, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 505, + 722 + ], + "score": 1.0, + "content": "errors and enhances symmetry-breaking in the assignment of transformations to rules, while in the", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "score": 1.0, + "content": "case of GNNs, since all entities are affected per step, there is a higher possibility of errors getting", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 82, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 97 + ], + "score": 1.0, + "content": "compounded. We can see that even with a single rule, we significantly outperform GNNs thus proving", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 384, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 384, + 107 + ], + "score": 1.0, + "content": "the effectiveness of dynamically instantiating edges between entities.", + "type": "text", + "cross_page": true + } + ], + "index": 2 + } + ], + "index": 53.5, + "bbox_fs": [ + 105, + 678, + 506, + 722 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 73, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "score": 1.0, + "content": "case of GNNs, since all entities are affected per step, there is a higher possibility of errors getting", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 82, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 97 + ], + "score": 1.0, + "content": "compounded. We can see that even with a single rule, we significantly outperform GNNs thus proving", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 384, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 384, + 107 + ], + "score": 1.0, + "content": "the effectiveness of dynamically instantiating edges between entities.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 120, + 505, + 164 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 505, + 132 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 505, + 132 + ], + "score": 1.0, + "content": "Atari Games. We also test the proposed model in the more complicated setting of Atari. Atari", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 506, + 144 + ], + "score": 1.0, + "content": "games also have sparse interactions between entities. For instance, in Pong, any interaction involves", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 142, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 505, + 154 + ], + "score": 1.0, + "content": "only 2 entities: (1) paddle and ball or (2) ball and the wall. Therefore, we expect sparse interactions", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 316, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 316, + 165 + ], + "score": 1.0, + "content": "captured by NPS to outperform GNNs here as well.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 169, + 505, + 224 + ], + "lines": [ + { + "bbox": [ + 105, + 168, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 505, + 182 + ], + "score": 1.0, + "content": "We follow the same setup as for the physics environment described in the previous section. We", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 180, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 429, + 193 + ], + "score": 1.0, + "content": "present the results for the Atari experiments in Figure 6(b), showing the average", + "type": "text" + }, + { + "bbox": [ + 430, + 180, + 453, + 191 + ], + "score": 0.81, + "content": "\\mathrm { H @ 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 180, + 505, + 193 + ], + "score": 1.0, + "content": "score across", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 191, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 104, + 191, + 505, + 204 + ], + "score": 1.0, + "content": "5 games: Pong, Space Invaders, Freeway, Breakout, and QBert. As expected, we can see that the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 201, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 506, + 214 + ], + "score": 1.0, + "content": "proposed model achieves a higher score than the GNN-based C-SWM. The results for the Atari", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 213, + 473, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 473, + 226 + ], + "score": 1.0, + "content": "experiments reinforce the claim that NPS is especially good at learning sparse interactions.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 505, + 337 + ], + "lines": [ + { + "bbox": [ + 106, + 238, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 505, + 250 + ], + "score": 1.0, + "content": "Learning Rules for Physical Reasoning. To show the effectiveness of the proposed approach for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 250, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 505, + 261 + ], + "score": 1.0, + "content": "physical reasoning tasks, we evaluate NPS on another dataset: Sprites-MOT (He et al., 2018). The", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 259, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 506, + 273 + ], + "score": 1.0, + "content": "Sprites-MOT dataset was introduced by He et al. (2018). The dataset contains a set of moving objects", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "score": 1.0, + "content": "of various shapes. This dataset aims to test whether a model can handle occlusions correctly. Each", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "frame has consistent bounding boxes which may cause the objects to appear or disappear from the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 292, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 307 + ], + "score": 1.0, + "content": "scene. A model which performs well should be able to track the motion of all objects irrespective", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "score": 1.0, + "content": "of whether they are occluded or not. We follow the same setup as Weis et al. (2020). We use the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "score": 1.0, + "content": "OP3 model (Veerapaneni et al., 2019) as our baseline. To test the proposed model, we replace the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 325, + 354, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 354, + 338 + ], + "score": 1.0, + "content": "GNN-based transition model in OP3 with the proposed NPS.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 342, + 339, + 473 + ], + "lines": [ + { + "bbox": [ + 106, + 341, + 340, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 340, + 354 + ], + "score": 1.0, + "content": "We use the same evaluation protocol as followed by Weis", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 353, + 339, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 339, + 365 + ], + "score": 1.0, + "content": "et al. (2020) which is based on the MOT (Multi-object", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 363, + 339, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 339, + 376 + ], + "score": 1.0, + "content": "tracking) challenge (Milan et al., 2016). The results on the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 374, + 339, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 339, + 387 + ], + "score": 1.0, + "content": "MOTA and MOTP metrics for this task are presented in", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 385, + 340, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 340, + 398 + ], + "score": 1.0, + "content": "Table 5. The results on the other metrics are presented in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 397, + 340, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 340, + 408 + ], + "score": 1.0, + "content": "appendix Table 10. We ask the reader to refer to appendix", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 407, + 340, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 340, + 419 + ], + "score": 1.0, + "content": "F.1 for more details about these metrics. We can see that", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 418, + 340, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 340, + 430 + ], + "score": 1.0, + "content": "for almost all metrics, NPS outperforms the OP3 baseline.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 429, + 339, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 339, + 441 + ], + "score": 1.0, + "content": "Although this dataset does not contain physical interactions", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 440, + 339, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 339, + 452 + ], + "score": 1.0, + "content": "between the objects, sparse rule application should still be", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 451, + 339, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 339, + 463 + ], + "score": 1.0, + "content": "useful in dealing with occlusions. At any time step, only", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 461, + 340, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 340, + 475 + ], + "score": 1.0, + "content": "a single object is affected by occlusions i.e., it may get", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 26.5 + }, + { + "type": "table", + "bbox": [ + 373, + 356, + 477, + 397 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 373, + 356, + 477, + 397 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 373, + 356, + 477, + 397 + ], + "spans": [ + { + "bbox": [ + 373, + 356, + 477, + 397 + ], + "score": 0.968, + "html": "
ModelMOTA↑MOTP↑
OP3NPS89.1±5.178.4±2.4
90.72±5.1579.91±0.3
", + "type": "table", + "image_path": "c8b9c1fb0031e32c3375d2a376992d3dee1d9313854b01ff0e732e92e5576ab7.jpg" + } + ] + } + ], + "index": 33.5, + "virtual_lines": [ + { + "bbox": [ + 373, + 356, + 477, + 376.5 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 373, + 376.5, + 477, + 397.0 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 346, + 402, + 505, + 468 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 345, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 345, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "Table 5: Sprites-MOT. Comparison be-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 346, + 414, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 346, + 414, + 506, + 425 + ], + "score": 1.0, + "content": "tween the proposed NPS and the base-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 346, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 346, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "line OP3 using the MOTA and MOTP", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 345, + 435, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 345, + 435, + 496, + 447 + ], + "score": 1.0, + "content": "metrics on the sprites-MOT dataset (", + "type": "text" + }, + { + "bbox": [ + 496, + 436, + 502, + 447 + ], + "score": 0.36, + "content": "\\uparrow", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 435, + 506, + 447 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 346, + 446, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 346, + 446, + 505, + 458 + ], + "score": 1.0, + "content": "higher is better). Average over 3 random", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 345, + 457, + 373, + 468 + ], + "spans": [ + { + "bbox": [ + 345, + 457, + 373, + 468 + ], + "score": 1.0, + "content": "seeds.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 473, + 505, + 516 + ], + "lines": [ + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "score": 1.0, + "content": "occluded due to another object or due to a prespecified bounding box, while the other objects follow", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "score": 1.0, + "content": "their default dynamics. Therefore, a rule should be applied to only the object (or entity) affected (i.e.,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "not visible) due to occlusion and may take into account any other object or entity that is responsible", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 505, + 179, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 179, + 518 + ], + "score": 1.0, + "content": "for the occlusion.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + }, + { + "type": "title", + "bbox": [ + 108, + 526, + 262, + 540 + ], + "lines": [ + { + "bbox": [ + 105, + 525, + 263, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 263, + 541 + ], + "score": 1.0, + "content": "5 Discussion and Conclusion", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 106, + 549, + 505, + 615 + ], + "lines": [ + { + "bbox": [ + 105, + 549, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 562 + ], + "score": 1.0, + "content": "For AI agents such as robots trying to make sense of their environment, the only observables are", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 560, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 506, + 573 + ], + "score": 1.0, + "content": "low-level variables like pixels in images. To generalize well, an agent must induce high-level entities", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 572, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 583 + ], + "score": 1.0, + "content": "as well as discover and disentangle the rules that govern how these entities actually interact with", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "each other. Here we have focused on perceptual inference problems and proposed NPS, a neural", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "score": 1.0, + "content": "instantiation of production systems by introducing an important inductive bias in the architecture", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 603, + 444, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 444, + 617 + ], + "score": 1.0, + "content": "following the proposals of Bengio (2017); Goyal & Bengio (2020); Ke et al. (2021).", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 48.5 + }, + { + "type": "text", + "bbox": [ + 107, + 629, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 105, + 628, + 507, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 507, + 642 + ], + "score": 1.0, + "content": "Limitations & Looking Forward. Our experiments highlight the advantages brought by the factor-", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 640, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 652 + ], + "score": 1.0, + "content": "ization of knowledge into a small set of entities and sparse sequentially applied rules. Immediate", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "future work would investigate how to take advantage of these inductive biases for more complex", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 662, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 505, + 675 + ], + "score": 1.0, + "content": "physical environments (Ahmed et al., 2020) and novel planning methods, which might be more", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 672, + 359, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 359, + 685 + ], + "score": 1.0, + "content": "sample efficient than standard ones (Schrittwieser et al., 2020).", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 54 + }, + { + "type": "text", + "bbox": [ + 108, + 689, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 506, + 701 + ], + "score": 1.0, + "content": "We also find that Sequential and Parallel NPS have different properties suited towards different", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "domains. Future work should explore how to effectively combine these two approaches. We discuss", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 711, + 283, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 283, + 723 + ], + "score": 1.0, + "content": "this in more detail in Appendix section E.3.", + "type": "text" + } + ], + "index": 59 + } + ], + "index": 58 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 301, + 742, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 73, + 504, + 106 + ], + "lines": [], + "index": 1, + "bbox_fs": [ + 105, + 72, + 506, + 107 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 120, + 505, + 164 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 505, + 132 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 505, + 132 + ], + "score": 1.0, + "content": "Atari Games. We also test the proposed model in the more complicated setting of Atari. Atari", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 506, + 144 + ], + "score": 1.0, + "content": "games also have sparse interactions between entities. For instance, in Pong, any interaction involves", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 142, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 505, + 154 + ], + "score": 1.0, + "content": "only 2 entities: (1) paddle and ball or (2) ball and the wall. Therefore, we expect sparse interactions", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 316, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 316, + 165 + ], + "score": 1.0, + "content": "captured by NPS to outperform GNNs here as well.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 120, + 506, + 165 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 169, + 505, + 224 + ], + "lines": [ + { + "bbox": [ + 105, + 168, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 505, + 182 + ], + "score": 1.0, + "content": "We follow the same setup as for the physics environment described in the previous section. We", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 180, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 429, + 193 + ], + "score": 1.0, + "content": "present the results for the Atari experiments in Figure 6(b), showing the average", + "type": "text" + }, + { + "bbox": [ + 430, + 180, + 453, + 191 + ], + "score": 0.81, + "content": "\\mathrm { H @ 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 180, + 505, + 193 + ], + "score": 1.0, + "content": "score across", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 191, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 104, + 191, + 505, + 204 + ], + "score": 1.0, + "content": "5 games: Pong, Space Invaders, Freeway, Breakout, and QBert. As expected, we can see that the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 201, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 506, + 214 + ], + "score": 1.0, + "content": "proposed model achieves a higher score than the GNN-based C-SWM. The results for the Atari", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 213, + 473, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 473, + 226 + ], + "score": 1.0, + "content": "experiments reinforce the claim that NPS is especially good at learning sparse interactions.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 104, + 168, + 506, + 226 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 505, + 337 + ], + "lines": [ + { + "bbox": [ + 106, + 238, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 505, + 250 + ], + "score": 1.0, + "content": "Learning Rules for Physical Reasoning. To show the effectiveness of the proposed approach for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 250, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 505, + 261 + ], + "score": 1.0, + "content": "physical reasoning tasks, we evaluate NPS on another dataset: Sprites-MOT (He et al., 2018). The", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 259, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 506, + 273 + ], + "score": 1.0, + "content": "Sprites-MOT dataset was introduced by He et al. (2018). The dataset contains a set of moving objects", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "score": 1.0, + "content": "of various shapes. This dataset aims to test whether a model can handle occlusions correctly. Each", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "frame has consistent bounding boxes which may cause the objects to appear or disappear from the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 292, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 307 + ], + "score": 1.0, + "content": "scene. A model which performs well should be able to track the motion of all objects irrespective", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "score": 1.0, + "content": "of whether they are occluded or not. We follow the same setup as Weis et al. (2020). We use the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "score": 1.0, + "content": "OP3 model (Veerapaneni et al., 2019) as our baseline. To test the proposed model, we replace the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 325, + 354, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 354, + 338 + ], + "score": 1.0, + "content": "GNN-based transition model in OP3 with the proposed NPS.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 238, + 506, + 338 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 342, + 339, + 473 + ], + "lines": [ + { + "bbox": [ + 106, + 341, + 340, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 340, + 354 + ], + "score": 1.0, + "content": "We use the same evaluation protocol as followed by Weis", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 353, + 339, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 339, + 365 + ], + "score": 1.0, + "content": "et al. (2020) which is based on the MOT (Multi-object", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 363, + 339, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 339, + 376 + ], + "score": 1.0, + "content": "tracking) challenge (Milan et al., 2016). The results on the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 374, + 339, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 339, + 387 + ], + "score": 1.0, + "content": "MOTA and MOTP metrics for this task are presented in", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 385, + 340, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 340, + 398 + ], + "score": 1.0, + "content": "Table 5. The results on the other metrics are presented in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 397, + 340, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 340, + 408 + ], + "score": 1.0, + "content": "appendix Table 10. We ask the reader to refer to appendix", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 407, + 340, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 340, + 419 + ], + "score": 1.0, + "content": "F.1 for more details about these metrics. We can see that", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 418, + 340, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 340, + 430 + ], + "score": 1.0, + "content": "for almost all metrics, NPS outperforms the OP3 baseline.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 429, + 339, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 339, + 441 + ], + "score": 1.0, + "content": "Although this dataset does not contain physical interactions", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 440, + 339, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 339, + 452 + ], + "score": 1.0, + "content": "between the objects, sparse rule application should still be", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 451, + 339, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 339, + 463 + ], + "score": 1.0, + "content": "useful in dealing with occlusions. At any time step, only", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 461, + 340, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 340, + 475 + ], + "score": 1.0, + "content": "a single object is affected by occlusions i.e., it may get", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "score": 1.0, + "content": "occluded due to another object or due to a prespecified bounding box, while the other objects follow", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "score": 1.0, + "content": "their default dynamics. Therefore, a rule should be applied to only the object (or entity) affected (i.e.,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "not visible) due to occlusion and may take into account any other object or entity that is responsible", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 505, + 179, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 179, + 518 + ], + "score": 1.0, + "content": "for the occlusion.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 341, + 340, + 475 + ] + }, + { + "type": "table", + "bbox": [ + 373, + 356, + 477, + 397 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 373, + 356, + 477, + 397 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 373, + 356, + 477, + 397 + ], + "spans": [ + { + "bbox": [ + 373, + 356, + 477, + 397 + ], + "score": 0.968, + "html": "
ModelMOTA↑MOTP↑
OP3NPS89.1±5.178.4±2.4
90.72±5.1579.91±0.3
", + "type": "table", + "image_path": "c8b9c1fb0031e32c3375d2a376992d3dee1d9313854b01ff0e732e92e5576ab7.jpg" + } + ] + } + ], + "index": 33.5, + "virtual_lines": [ + { + "bbox": [ + 373, + 356, + 477, + 376.5 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 373, + 376.5, + 477, + 397.0 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 346, + 402, + 505, + 468 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 345, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 345, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "Table 5: Sprites-MOT. Comparison be-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 346, + 414, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 346, + 414, + 506, + 425 + ], + "score": 1.0, + "content": "tween the proposed NPS and the base-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 346, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 346, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "line OP3 using the MOTA and MOTP", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 345, + 435, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 345, + 435, + 496, + 447 + ], + "score": 1.0, + "content": "metrics on the sprites-MOT dataset (", + "type": "text" + }, + { + "bbox": [ + 496, + 436, + 502, + 447 + ], + "score": 0.36, + "content": "\\uparrow", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 435, + 506, + 447 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 346, + 446, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 346, + 446, + 505, + 458 + ], + "score": 1.0, + "content": "higher is better). Average over 3 random", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 345, + 457, + 373, + 468 + ], + "spans": [ + { + "bbox": [ + 345, + 457, + 373, + 468 + ], + "score": 1.0, + "content": "seeds.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 473, + 505, + 516 + ], + "lines": [], + "index": 42.5, + "bbox_fs": [ + 105, + 472, + 506, + 518 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 526, + 262, + 540 + ], + "lines": [ + { + "bbox": [ + 105, + 525, + 263, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 263, + 541 + ], + "score": 1.0, + "content": "5 Discussion and Conclusion", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 106, + 549, + 505, + 615 + ], + "lines": [ + { + "bbox": [ + 105, + 549, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 562 + ], + "score": 1.0, + "content": "For AI agents such as robots trying to make sense of their environment, the only observables are", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 560, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 506, + 573 + ], + "score": 1.0, + "content": "low-level variables like pixels in images. To generalize well, an agent must induce high-level entities", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 572, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 583 + ], + "score": 1.0, + "content": "as well as discover and disentangle the rules that govern how these entities actually interact with", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "each other. Here we have focused on perceptual inference problems and proposed NPS, a neural", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "score": 1.0, + "content": "instantiation of production systems by introducing an important inductive bias in the architecture", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 603, + 444, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 444, + 617 + ], + "score": 1.0, + "content": "following the proposals of Bengio (2017); Goyal & Bengio (2020); Ke et al. (2021).", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 48.5, + "bbox_fs": [ + 105, + 549, + 506, + 617 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 629, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 105, + 628, + 507, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 507, + 642 + ], + "score": 1.0, + "content": "Limitations & Looking Forward. Our experiments highlight the advantages brought by the factor-", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 640, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 652 + ], + "score": 1.0, + "content": "ization of knowledge into a small set of entities and sparse sequentially applied rules. Immediate", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "future work would investigate how to take advantage of these inductive biases for more complex", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 662, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 505, + 675 + ], + "score": 1.0, + "content": "physical environments (Ahmed et al., 2020) and novel planning methods, which might be more", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 672, + 359, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 359, + 685 + ], + "score": 1.0, + "content": "sample efficient than standard ones (Schrittwieser et al., 2020).", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 54, + "bbox_fs": [ + 105, + 628, + 507, + 685 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 689, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 506, + 701 + ], + "score": 1.0, + "content": "We also find that Sequential and Parallel NPS have different properties suited towards different", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "domains. Future work should explore how to effectively combine these two approaches. We discuss", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 711, + 283, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 283, + 723 + ], + "score": 1.0, + "content": "this in more detail in Appendix section E.3.", + "type": "text" + } + ], + "index": 59 + } + ], + "index": 58, + "bbox_fs": [ + 106, + 688, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 71, + 225, + 85 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 226, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 226, + 87 + ], + "score": 1.0, + "content": "6 Acknowledgements", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 96, + 505, + 162 + ], + "lines": [ + { + "bbox": [ + 105, + 95, + 505, + 109 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 505, + 109 + ], + "score": 1.0, + "content": "The authors would like to thank Matthew Botvinick for useful discussions. The authors would also", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 120 + ], + "score": 1.0, + "content": "like to thank Alex Lamb, Stefan Bauer, Nicolas Chapados, Danilo Rezende and Kelsey Allen for", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 116, + 506, + 131 + ], + "spans": [ + { + "bbox": [ + 104, + 116, + 506, + 131 + ], + "score": 1.0, + "content": "brainstorming sessions. We are also thankful to Dianbo Liu, Damjan Kalajdzievski and Osama", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 128, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 505, + 141 + ], + "score": 1.0, + "content": "Ahmed for proofreading. We would like to thank Samsung Electronics Co. Ltd. and CIFAR for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 139, + 505, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 505, + 152 + ], + "score": 1.0, + "content": "funding this research. We would also like to thank Google for providing Google cloud credits used in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 150, + 149, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 149, + 162 + ], + "score": 1.0, + "content": "this work.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 107, + 178, + 163, + 191 + ], + "lines": [ + { + "bbox": [ + 106, + 176, + 165, + 193 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 165, + 193 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 105, + 197, + 507, + 725 + ], + "lines": [ + { + "bbox": [ + 105, + 198, + 507, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 507, + 210 + ], + "score": 1.0, + "content": "Ahmed, O., Träuble, F., Goyal, A., Neitz, A., Wüthrich, M., Bengio, Y., Schölkopf, B., and Bauer, S.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 115, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "Causalworld: A robotic manipulation benchmark for causal structure and transfer learning. arXiv", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 114, + 220, + 254, + 232 + ], + "spans": [ + { + "bbox": [ + 114, + 220, + 254, + 232 + ], + "score": 1.0, + "content": "preprint arXiv:2010.04296, 2020.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 237, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 104, + 237, + 506, + 253 + ], + "score": 1.0, + "content": "Anderson, J. R. Skill acquisition: Compilation of weak-method problem situations. Psychological", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 250, + 216, + 263 + ], + "spans": [ + { + "bbox": [ + 116, + 250, + 216, + 263 + ], + "score": 1.0, + "content": "review, 94(2):192, 1987.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 269, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 506, + 284 + ], + "score": 1.0, + "content": "Andreas, J., Rohrbach, M., Darrell, T., and Klein, D. Neural module networks. In Proceedings of the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 281, + 442, + 294 + ], + "spans": [ + { + "bbox": [ + 115, + 281, + 442, + 294 + ], + "score": 1.0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition, pp. 39–48, 2016.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 299, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 104, + 299, + 506, + 314 + ], + "score": 1.0, + "content": "Bahdanau, D., Cho, K., and Bengio, Y. Neural machine translation by jointly learning to align and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 311, + 315, + 325 + ], + "spans": [ + { + "bbox": [ + 114, + 311, + 315, + 325 + ], + "score": 1.0, + "content": "translate. arXiv preprint arXiv:1409.0473, 2014.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "Battaglia, P. W., Pascanu, R., Lai, M., Rezende, D. J., and Kavukcuoglu, K. Interaction networks", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 114, + 342, + 507, + 356 + ], + "spans": [ + { + "bbox": [ + 114, + 342, + 507, + 356 + ], + "score": 1.0, + "content": "for learning about objects, relations and physics. CoRR, abs/1612.00222, 2016. URL http:", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 114, + 353, + 257, + 367 + ], + "spans": [ + { + "bbox": [ + 114, + 353, + 257, + 367 + ], + "score": 1.0, + "content": "//arxiv.org/abs/1612.00222.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 371, + 507, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 507, + 386 + ], + "score": 1.0, + "content": "Battaglia, P. W., Hamrick, J. B., Bapst, V., Sanchez-Gonzalez, A., Zambaldi, V., Malinowski, M.,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 114, + 381, + 507, + 399 + ], + "spans": [ + { + "bbox": [ + 114, + 381, + 507, + 399 + ], + "score": 1.0, + "content": "Tacchetti, A., Raposo, D., Santoro, A., Faulkner, R., et al. Relational inductive biases, deep", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 394, + 401, + 408 + ], + "spans": [ + { + "bbox": [ + 115, + 394, + 401, + 408 + ], + "score": 1.0, + "content": "learning, and graph networks. arXiv preprint arXiv:1806.01261, 2018.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 414, + 417, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 417, + 428 + ], + "score": 1.0, + "content": "Bengio, Y. The consciousness prior. arXiv preprint arXiv:1709.08568, 2017.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 432, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 104, + 432, + 506, + 448 + ], + "score": 1.0, + "content": "Bottou, L. and Gallinari, P. A framework for the cooperation of learning algorithms. In Advances in", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 114, + 444, + 356, + 458 + ], + "spans": [ + { + "bbox": [ + 114, + 444, + 356, + 458 + ], + "score": 1.0, + "content": "neural information processing systems, pp. 781–788, 1991.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 462, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 479 + ], + "score": 1.0, + "content": "Bronstein, M. M., Bruna, J., LeCun, Y., Szlam, A., and Vandergheynst, P. Geometric deep learning:", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 114, + 474, + 460, + 490 + ], + "spans": [ + { + "bbox": [ + 114, + 474, + 460, + 490 + ], + "score": 1.0, + "content": "going beyond euclidean data. IEEE Signal Processing Magazine, 34(4):18–42, 2017.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 493, + 507, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 507, + 509 + ], + "score": 1.0, + "content": "Bunel, R., Hausknecht, M., Devlin, J., Singh, R., and Kohli, P. Leveraging grammar and reinforcement", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 114, + 505, + 435, + 519 + ], + "spans": [ + { + "bbox": [ + 114, + 505, + 435, + 519 + ], + "score": 1.0, + "content": "learning for neural program synthesis. arXiv preprint arXiv:1805.04276, 2018.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "Burgess, C. P., Matthey, L., Watters, N., Kabra, R., Higgins, I., Botvinick, M., and Lerchner, A.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 114, + 535, + 507, + 551 + ], + "spans": [ + { + "bbox": [ + 114, + 535, + 507, + 551 + ], + "score": 1.0, + "content": "Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 546, + 142, + 560 + ], + "spans": [ + { + "bbox": [ + 115, + 546, + 142, + 560 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 566, + 507, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 507, + 581 + ], + "score": 1.0, + "content": "Cai, J., Shin, R., and Song, D. Making neural programming architectures generalize via recursion.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 114, + 577, + 280, + 592 + ], + "spans": [ + { + "bbox": [ + 114, + 577, + 280, + 592 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1704.06611, 2017.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "Chen, R. T., Li, X., Grosse, R., and Duvenaud, D. Isolating sources of disentanglement in variational", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 114, + 608, + 339, + 622 + ], + "spans": [ + { + "bbox": [ + 114, + 608, + 339, + 622 + ], + "score": 1.0, + "content": "autoencoders. arXiv preprint arXiv:1802.04942, 2018.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "Ding, D., Hill, F., Santoro, A., and Botvinick, M. Object-based attention for spatio-temporal", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "score": 1.0, + "content": "reasoning: Outperforming neuro-symbolic models with flexible distributed architectures. arXiv", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 114, + 650, + 255, + 663 + ], + "spans": [ + { + "bbox": [ + 114, + 650, + 255, + 663 + ], + "score": 1.0, + "content": "preprint arXiv:2012.08508, 2020.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 666, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 666, + 506, + 685 + ], + "score": 1.0, + "content": "Du, Y., Smith, K., Ulman, T., Tenenbaum, J., and Wu, J. Unsupervised discovery of 3d physical", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 680, + 360, + 695 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 360, + 695 + ], + "score": 1.0, + "content": "objects from video. arXiv preprint arXiv:2007.12348, 2020.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "Engelcke, M., Kosiorek, A. R., Jones, O. P., and Posner, I. Genesis: Generative scene inference and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 114, + 711, + 486, + 725 + ], + "spans": [ + { + "bbox": [ + 114, + 711, + 486, + 725 + ], + "score": 1.0, + "content": "sampling with object-centric latent representations. arXiv preprint arXiv:1907.13052, 2019.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 25.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 310, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 755 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 755 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 71, + 225, + 85 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 226, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 226, + 87 + ], + "score": 1.0, + "content": "6 Acknowledgements", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 96, + 505, + 162 + ], + "lines": [ + { + "bbox": [ + 105, + 95, + 505, + 109 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 505, + 109 + ], + "score": 1.0, + "content": "The authors would like to thank Matthew Botvinick for useful discussions. The authors would also", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 120 + ], + "score": 1.0, + "content": "like to thank Alex Lamb, Stefan Bauer, Nicolas Chapados, Danilo Rezende and Kelsey Allen for", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 116, + 506, + 131 + ], + "spans": [ + { + "bbox": [ + 104, + 116, + 506, + 131 + ], + "score": 1.0, + "content": "brainstorming sessions. We are also thankful to Dianbo Liu, Damjan Kalajdzievski and Osama", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 128, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 505, + 141 + ], + "score": 1.0, + "content": "Ahmed for proofreading. We would like to thank Samsung Electronics Co. Ltd. and CIFAR for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 139, + 505, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 505, + 152 + ], + "score": 1.0, + "content": "funding this research. We would also like to thank Google for providing Google cloud credits used in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 150, + 149, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 149, + 162 + ], + "score": 1.0, + "content": "this work.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3.5, + "bbox_fs": [ + 104, + 95, + 506, + 162 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 178, + 163, + 191 + ], + "lines": [ + { + "bbox": [ + 106, + 176, + 165, + 193 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 165, + 193 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "list", + "bbox": [ + 105, + 197, + 507, + 725 + ], + "lines": [ + { + "bbox": [ + 105, + 198, + 507, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 507, + 210 + ], + "score": 1.0, + "content": "Ahmed, O., Träuble, F., Goyal, A., Neitz, A., Wüthrich, M., Bengio, Y., Schölkopf, B., and Bauer, S.", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 115, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "Causalworld: A robotic manipulation benchmark for causal structure and transfer learning. arXiv", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 114, + 220, + 254, + 232 + ], + "spans": [ + { + "bbox": [ + 114, + 220, + 254, + 232 + ], + "score": 1.0, + "content": "preprint arXiv:2010.04296, 2020.", + "type": "text" + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 237, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 104, + 237, + 506, + 253 + ], + "score": 1.0, + "content": "Anderson, J. R. Skill acquisition: Compilation of weak-method problem situations. Psychological", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 250, + 216, + 263 + ], + "spans": [ + { + "bbox": [ + 116, + 250, + 216, + 263 + ], + "score": 1.0, + "content": "review, 94(2):192, 1987.", + "type": "text" + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 269, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 506, + 284 + ], + "score": 1.0, + "content": "Andreas, J., Rohrbach, M., Darrell, T., and Klein, D. Neural module networks. In Proceedings of the", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 281, + 442, + 294 + ], + "spans": [ + { + "bbox": [ + 115, + 281, + 442, + 294 + ], + "score": 1.0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition, pp. 39–48, 2016.", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 299, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 104, + 299, + 506, + 314 + ], + "score": 1.0, + "content": "Bahdanau, D., Cho, K., and Bengio, Y. Neural machine translation by jointly learning to align and", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 311, + 315, + 325 + ], + "spans": [ + { + "bbox": [ + 114, + 311, + 315, + 325 + ], + "score": 1.0, + "content": "translate. arXiv preprint arXiv:1409.0473, 2014.", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "Battaglia, P. W., Pascanu, R., Lai, M., Rezende, D. J., and Kavukcuoglu, K. Interaction networks", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 342, + 507, + 356 + ], + "spans": [ + { + "bbox": [ + 114, + 342, + 507, + 356 + ], + "score": 1.0, + "content": "for learning about objects, relations and physics. CoRR, abs/1612.00222, 2016. URL http:", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 114, + 353, + 257, + 367 + ], + "spans": [ + { + "bbox": [ + 114, + 353, + 257, + 367 + ], + "score": 1.0, + "content": "//arxiv.org/abs/1612.00222.", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 371, + 507, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 507, + 386 + ], + "score": 1.0, + "content": "Battaglia, P. W., Hamrick, J. B., Bapst, V., Sanchez-Gonzalez, A., Zambaldi, V., Malinowski, M.,", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 381, + 507, + 399 + ], + "spans": [ + { + "bbox": [ + 114, + 381, + 507, + 399 + ], + "score": 1.0, + "content": "Tacchetti, A., Raposo, D., Santoro, A., Faulkner, R., et al. Relational inductive biases, deep", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 394, + 401, + 408 + ], + "spans": [ + { + "bbox": [ + 115, + 394, + 401, + 408 + ], + "score": 1.0, + "content": "learning, and graph networks. arXiv preprint arXiv:1806.01261, 2018.", + "type": "text" + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 414, + 417, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 417, + 428 + ], + "score": 1.0, + "content": "Bengio, Y. The consciousness prior. arXiv preprint arXiv:1709.08568, 2017.", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 432, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 104, + 432, + 506, + 448 + ], + "score": 1.0, + "content": "Bottou, L. and Gallinari, P. A framework for the cooperation of learning algorithms. In Advances in", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 444, + 356, + 458 + ], + "spans": [ + { + "bbox": [ + 114, + 444, + 356, + 458 + ], + "score": 1.0, + "content": "neural information processing systems, pp. 781–788, 1991.", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 462, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 479 + ], + "score": 1.0, + "content": "Bronstein, M. M., Bruna, J., LeCun, Y., Szlam, A., and Vandergheynst, P. Geometric deep learning:", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 474, + 460, + 490 + ], + "spans": [ + { + "bbox": [ + 114, + 474, + 460, + 490 + ], + "score": 1.0, + "content": "going beyond euclidean data. IEEE Signal Processing Magazine, 34(4):18–42, 2017.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 493, + 507, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 507, + 509 + ], + "score": 1.0, + "content": "Bunel, R., Hausknecht, M., Devlin, J., Singh, R., and Kohli, P. Leveraging grammar and reinforcement", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 505, + 435, + 519 + ], + "spans": [ + { + "bbox": [ + 114, + 505, + 435, + 519 + ], + "score": 1.0, + "content": "learning for neural program synthesis. arXiv preprint arXiv:1805.04276, 2018.", + "type": "text" + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "Burgess, C. P., Matthey, L., Watters, N., Kabra, R., Higgins, I., Botvinick, M., and Lerchner, A.", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 535, + 507, + 551 + ], + "spans": [ + { + "bbox": [ + 114, + 535, + 507, + 551 + ], + "score": 1.0, + "content": "Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 546, + 142, + 560 + ], + "spans": [ + { + "bbox": [ + 115, + 546, + 142, + 560 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 566, + 507, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 507, + 581 + ], + "score": 1.0, + "content": "Cai, J., Shin, R., and Song, D. Making neural programming architectures generalize via recursion.", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 577, + 280, + 592 + ], + "spans": [ + { + "bbox": [ + 114, + 577, + 280, + 592 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1704.06611, 2017.", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "Chen, R. T., Li, X., Grosse, R., and Duvenaud, D. Isolating sources of disentanglement in variational", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 608, + 339, + 622 + ], + "spans": [ + { + "bbox": [ + 114, + 608, + 339, + 622 + ], + "score": 1.0, + "content": "autoencoders. arXiv preprint arXiv:1802.04942, 2018.", + "type": "text" + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "Ding, D., Hill, F., Santoro, A., and Botvinick, M. Object-based attention for spatio-temporal", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "score": 1.0, + "content": "reasoning: Outperforming neuro-symbolic models with flexible distributed architectures. arXiv", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 114, + 650, + 255, + 663 + ], + "spans": [ + { + "bbox": [ + 114, + 650, + 255, + 663 + ], + "score": 1.0, + "content": "preprint arXiv:2012.08508, 2020.", + "type": "text" + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 666, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 666, + 506, + 685 + ], + "score": 1.0, + "content": "Du, Y., Smith, K., Ulman, T., Tenenbaum, J., and Wu, J. Unsupervised discovery of 3d physical", + "type": "text" + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 680, + 360, + 695 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 360, + 695 + ], + "score": 1.0, + "content": "objects from video. arXiv preprint arXiv:2007.12348, 2020.", + "type": "text" + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "Engelcke, M., Kosiorek, A. R., Jones, O. P., and Posner, I. Genesis: Generative scene inference and", + "type": "text" + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 711, + 486, + 725 + ], + "spans": [ + { + "bbox": [ + 114, + 711, + 486, + 725 + ], + "score": 1.0, + "content": "sampling with object-centric latent representations. arXiv preprint arXiv:1907.13052, 2019.", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 73, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 506, + 85 + ], + "score": 1.0, + "content": "Eslami, S., Heess, N., Weber, T., Tassa, Y., Szepesvari, D., Kavukcuoglu, K., and Hinton, G. E. Attend,", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 84, + 507, + 97 + ], + "spans": [ + { + "bbox": [ + 114, + 84, + 507, + 97 + ], + "score": 1.0, + "content": "infer, repeat: Fast scene understanding with generative models. arXiv preprint arXiv:1603.08575,", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 94, + 143, + 109 + ], + "spans": [ + { + "bbox": [ + 114, + 94, + 143, + 109 + ], + "score": 1.0, + "content": "2016.", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 111, + 507, + 128 + ], + "spans": [ + { + "bbox": [ + 104, + 111, + 507, + 128 + ], + "score": 1.0, + "content": "Evans, R., Hernández-Orallo, J., Welbl, J., Kohli, P., and Sergot, M. Making sense of sensory input.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 125, + 281, + 138 + ], + "spans": [ + { + "bbox": [ + 114, + 125, + 281, + 138 + ], + "score": 1.0, + "content": "Artificial Intelligence, pp. 103438, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 104, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "Fedus, W., Zoph, B., and Shazeer, N. Switch transformers: Scaling to trillion parameter models with", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 153, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 114, + 153, + 506, + 169 + ], + "score": 1.0, + "content": "simple and efficient sparsity. CoRR, abs/2101.03961, 2021. URL https://arxiv.org/abs/", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 167, + 174, + 178 + ], + "spans": [ + { + "bbox": [ + 116, + 167, + 174, + 178 + ], + "score": 1.0, + "content": "2101.03961.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 183, + 507, + 198 + ], + "spans": [ + { + "bbox": [ + 104, + 183, + 507, + 198 + ], + "score": 1.0, + "content": "Fernando, C., Banarse, D., Blundell, C., Zwols, Y., Ha, D., Rusu, A. A., Pritzel, A., and Wierstra,", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 194, + 507, + 210 + ], + "spans": [ + { + "bbox": [ + 113, + 194, + 507, + 210 + ], + "score": 1.0, + "content": "D. Pathnet: Evolution channels gradient descent in super neural networks. arXiv preprint", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 206, + 220, + 220 + ], + "spans": [ + { + "bbox": [ + 115, + 206, + 220, + 220 + ], + "score": 1.0, + "content": "arXiv:1701.08734, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 225, + 507, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 507, + 239 + ], + "score": 1.0, + "content": "Girshick, R. B. Fast R-CNN. CoRR, abs/1504.08083, 2015. URL http://arxiv.org/abs/1504.", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 237, + 149, + 249 + ], + "spans": [ + { + "bbox": [ + 115, + 237, + 149, + 249 + ], + "score": 1.0, + "content": "08083.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "Goyal, A. and Bengio, Y. Inductive biases for deep learning of higher-level cognition. arXiv preprint", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 267, + 219, + 279 + ], + "spans": [ + { + "bbox": [ + 115, + 267, + 219, + 279 + ], + "score": 1.0, + "content": "arXiv:2011.15091, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "score": 1.0, + "content": "Goyal, A., Lamb, A., Hoffmann, J., Sodhani, S., Levine, S., Bengio, Y., and Schölkopf, B. Recurrent", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 298, + 248, + 309 + ], + "spans": [ + { + "bbox": [ + 116, + 298, + 248, + 309 + ], + "score": 1.0, + "content": "independent mechanisms, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 315, + 507, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 507, + 329 + ], + "score": 1.0, + "content": "Goyal, A., Lamb, A., Gampa, P., Beaudoin, P., Levine, S., Blundell, C., Bengio, Y., and Mozer, M.", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 326, + 507, + 340 + ], + "spans": [ + { + "bbox": [ + 115, + 326, + 507, + 340 + ], + "score": 1.0, + "content": "Object files and schemata: Factorizing declarative and procedural knowledge in dynamical systems.", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 337, + 279, + 351 + ], + "spans": [ + { + "bbox": [ + 115, + 337, + 279, + 351 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2006.16225, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 356, + 507, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 507, + 370 + ], + "score": 1.0, + "content": "Graves, A., Wayne, G., and Danihelka, I. Neural turing machines. arXiv preprint arXiv:1410.5401,", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 366, + 144, + 381 + ], + "spans": [ + { + "bbox": [ + 114, + 366, + 144, + 381 + ], + "score": 1.0, + "content": "2014.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 384, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 104, + 384, + 506, + 402 + ], + "score": 1.0, + "content": "Greff, K., Rasmus, A., Berglund, M., Hao, T. H., Schmidhuber, J., and Valpola, H. Tagger: Deep", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 397, + 420, + 411 + ], + "spans": [ + { + "bbox": [ + 115, + 397, + 420, + 411 + ], + "score": 1.0, + "content": "unsupervised perceptual grouping. arXiv preprint arXiv:1606.06724, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 417, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 506, + 429 + ], + "score": 1.0, + "content": "Greff, K., Kaufman, R. L., Kabra, R., Watters, N., Burgess, C., Zoran, D., Matthey, L., Botvinick, M.,", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 427, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 115, + 427, + 505, + 441 + ], + "score": 1.0, + "content": "and Lerchner, A. Multi-object representation learning with iterative variational inference. arXiv", + "type": "text", + "cross_page": true + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 438, + 254, + 451 + ], + "spans": [ + { + "bbox": [ + 114, + 438, + 254, + 451 + ], + "score": 1.0, + "content": "preprint arXiv:1903.00450, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 457, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 471 + ], + "score": 1.0, + "content": "Groth, O., Fuchs, F., Posner, I., and Vedaldi, A. Shapestacks: Learning vision-based physical intuition", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 468, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 115, + 468, + 505, + 482 + ], + "score": 1.0, + "content": "for generalised object stacking. CoRR, abs/1804.08018, 2018. URL http://arxiv.org/abs/", + "type": "text", + "cross_page": true + } + ], + "index": 28 + }, + { + "bbox": [ + 117, + 480, + 174, + 492 + ], + "spans": [ + { + "bbox": [ + 117, + 480, + 174, + 492 + ], + "score": 1.0, + "content": "1804.08018.", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 497, + 507, + 511 + ], + "spans": [ + { + "bbox": [ + 104, + 497, + 507, + 511 + ], + "score": 1.0, + "content": "He, Z., Li, J., Liu, D., He, H., and Barber, D. Tracking by animation: Unsupervised learning of", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 509, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 115, + 509, + 505, + 523 + ], + "score": 1.0, + "content": "multi-object attentive trackers. CoRR, abs/1809.03137, 2018. URL http://arxiv.org/abs/", + "type": "text", + "cross_page": true + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 519, + 174, + 533 + ], + "spans": [ + { + "bbox": [ + 115, + 519, + 174, + 533 + ], + "score": 1.0, + "content": "1809.03137.", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 537, + 507, + 554 + ], + "spans": [ + { + "bbox": [ + 104, + 537, + 507, + 554 + ], + "score": 1.0, + "content": "Higgins, I., Matthey, L., Pal, A., Burgess, C., Glorot, X., Botvinick, M., Mohamed, S., and Lerchner,", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 549, + 489, + 563 + ], + "spans": [ + { + "bbox": [ + 114, + 549, + 489, + 563 + ], + "score": 1.0, + "content": "A. beta-vae: Learning basic visual concepts with a constrained variational framework. 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 568, + 507, + 582 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 507, + 582 + ], + "score": 1.0, + "content": "Jacobs, R. A., Jordan, M. I., Nowlan, S. J., Hinton, G. E., et al. Adaptive mixtures of local experts.", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 579, + 277, + 593 + ], + "spans": [ + { + "bbox": [ + 114, + 579, + 277, + 593 + ], + "score": 1.0, + "content": "Neural computation, 3(1):79–87, 1991.", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "Jang, E., Gu, S., and Poole, B. Categorical reparameterization with gumbel-softmax. arXiv preprint", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 610, + 219, + 623 + ], + "spans": [ + { + "bbox": [ + 114, + 610, + 219, + 623 + ], + "score": 1.0, + "content": "arXiv:1611.01144, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 628, + 507, + 642 + ], + "spans": [ + { + "bbox": [ + 104, + 628, + 507, + 642 + ], + "score": 1.0, + "content": "Ke, N. R., Didolkar, A. R., Mittal, S., Goyal, A., Lajoie, G., Bauer, S., Rezende, D. J., Mozer,", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 640, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 114, + 640, + 506, + 654 + ], + "score": 1.0, + "content": "M. C., Bengio, Y., and Pal, C. Systematic evaluation of causal discovery in visual model based", + "type": "text", + "cross_page": true + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 650, + 240, + 664 + ], + "spans": [ + { + "bbox": [ + 115, + 650, + 240, + 664 + ], + "score": 1.0, + "content": "reinforcement learning. 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "Kipf, T., Fetaya, E., Wang, K.-C., Welling, M., and Zemel, R. Neural relational inference for", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 681, + 362, + 694 + ], + "spans": [ + { + "bbox": [ + 114, + 681, + 362, + 694 + ], + "score": 1.0, + "content": "interacting systems. arXiv preprint arXiv:1802.04687, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "Kipf, T., van der Pol, E., and Welling, M. Contrastive learning of structured world models. arXiv", + "type": "text", + "cross_page": true + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 710, + 254, + 725 + ], + "spans": [ + { + "bbox": [ + 114, + 710, + 254, + 725 + ], + "score": 1.0, + "content": "preprint arXiv:1911.12247, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 45, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 71, + 507, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 71, + 507, + 87 + ], + "score": 1.0, + "content": "Kirsch, L., Kunze, J., and Barber, D. Modular networks: Learning to decompose neural computation.", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 84, + 432, + 98 + ], + "spans": [ + { + "bbox": [ + 114, + 84, + 432, + 98 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems, pp. 2408–2418, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 101, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 506, + 117 + ], + "score": 1.0, + "content": "Kosiorek, A., Kim, H., Teh, Y. W., and Posner, I. Sequential attend, infer, repeat: Generative", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 115, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 115, + 115, + 506, + 127 + ], + "score": 1.0, + "content": "modelling of moving objects. Advances in Neural Information Processing Systems, 31:8606–8616,", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 124, + 144, + 139 + ], + "spans": [ + { + "bbox": [ + 114, + 124, + 144, + 139 + ], + "score": 1.0, + "content": "2018.", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 142, + 506, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 159 + ], + "score": 1.0, + "content": "Laird, J. E., Rosenbloom, P. S., and Newell, A. Chunking in soar: The anatomy of a general learning", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 154, + 318, + 168 + ], + "spans": [ + { + "bbox": [ + 114, + 154, + 318, + 168 + ], + "score": 1.0, + "content": "mechanism. Machine learning, 1(1):11–46, 1986.", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 174, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 505, + 186 + ], + "score": 1.0, + "content": "Lamb, A., Goyal, A., Słowik, A., Mozer, M., Beaudoin, P., and Bengio, Y. Neural function modules", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 185, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 114, + 185, + 506, + 199 + ], + "score": 1.0, + "content": "with sparse arguments: A dynamic approach to integrating information across layers. arXiv", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 196, + 254, + 209 + ], + "spans": [ + { + "bbox": [ + 114, + 196, + 254, + 209 + ], + "score": 1.0, + "content": "preprint arXiv:2010.08012, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 214, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 104, + 214, + 506, + 230 + ], + "score": 1.0, + "content": "Le Roux, N., Heess, N., Shotton, J., and Winn, J. Learning a generative model of images by factoring", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 225, + 387, + 240 + ], + "spans": [ + { + "bbox": [ + 114, + 225, + 387, + 240 + ], + "score": 1.0, + "content": "appearance and shape. Neural Computation, 23(3):593–650, 2011.", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 245, + 507, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 507, + 259 + ], + "score": 1.0, + "content": "Li, Y., Gimeno, F., Kohli, P., and Vinyals, O. Strong generalization and efficiency in neural programs.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 256, + 280, + 270 + ], + "spans": [ + { + "bbox": [ + 113, + 256, + 280, + 270 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2007.03629, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 274, + 507, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 507, + 290 + ], + "score": 1.0, + "content": "Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy,", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 285, + 507, + 300 + ], + "spans": [ + { + "bbox": [ + 113, + 285, + 507, + 300 + ], + "score": 1.0, + "content": "A., and Kipf, T. Object-centric learning with slot attention. arXiv preprint arXiv:2006.15055,", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 295, + 148, + 312 + ], + "spans": [ + { + "bbox": [ + 114, + 295, + 148, + 312 + ], + "score": 1.0, + "content": "2020a.", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 315, + 507, + 331 + ], + "spans": [ + { + "bbox": [ + 104, + 315, + 507, + 331 + ], + "score": 1.0, + "content": "Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy,", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 327, + 385, + 340 + ], + "spans": [ + { + "bbox": [ + 113, + 327, + 385, + 340 + ], + "score": 1.0, + "content": "A., and Kipf, T. Object-centric learning with slot attention, 2020b.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 346, + 507, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 507, + 361 + ], + "score": 1.0, + "content": "Lovett, M. C. and Anderson, J. R. Thinking as a production system. The Cambridge handbook of", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 357, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 114, + 357, + 294, + 371 + ], + "score": 1.0, + "content": "thinking and reasoning, pp. 401–429, 2005.", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 375, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 104, + 375, + 506, + 390 + ], + "score": 1.0, + "content": "McMillan, C., Mozer, M. C., and Smolensky, P. The connectionist scientist game: rule extraction and", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 387, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 114, + 387, + 506, + 401 + ], + "score": 1.0, + "content": "refinement in a neural network. In Proceedings of the 13th Annual Conference of the Cognitive", + "type": "text", + "cross_page": true + } + ], + "index": 22 + }, + { + "bbox": [ + 114, + 398, + 264, + 412 + ], + "spans": [ + { + "bbox": [ + 114, + 398, + 264, + 412 + ], + "score": 1.0, + "content": "Science Society, pp. 424–430, 1991.", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 416, + 507, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 507, + 431 + ], + "score": 1.0, + "content": "Milan, A., Leal-Taixé, L., Reid, I. D., Roth, S., and Schindler, K. MOT16: A benchmark for multi-", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 428, + 495, + 442 + ], + "spans": [ + { + "bbox": [ + 114, + 428, + 495, + 442 + ], + "score": 1.0, + "content": "object tracking. CoRR, abs/1603.00831, 2016. URL http://arxiv.org/abs/1603.00831.", + "type": "text", + "cross_page": true + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 447, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 506, + 462 + ], + "score": 1.0, + "content": "Neelakantan, A., Le, Q. V., and Sutskever, I. Neural programmer: Inducing latent programs with", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 459, + 350, + 472 + ], + "spans": [ + { + "bbox": [ + 114, + 459, + 350, + 472 + ], + "score": 1.0, + "content": "gradient descent. arXiv preprint arXiv:1511.04834, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "Qi, H., Wang, X., Pathak, D., Ma, Y., and Malik, J. Learning long-term visual dynamics with region", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 489, + 507, + 502 + ], + "spans": [ + { + "bbox": [ + 114, + 489, + 507, + 502 + ], + "score": 1.0, + "content": "proposal interaction networks. In International Conference on Learning Representations, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 29 + }, + { + "bbox": [ + 114, + 499, + 370, + 514 + ], + "spans": [ + { + "bbox": [ + 114, + 499, + 370, + 514 + ], + "score": 1.0, + "content": "URL https://openreview.net/forum?id=_X_4Akcd8Re.", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "score": 1.0, + "content": "Rahaman, N., Goyal, A., Gondal, M. W., Wuthrich, M., Bauer, S., Sharma, Y., Bengio, Y., and", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 529, + 507, + 544 + ], + "spans": [ + { + "bbox": [ + 114, + 529, + 507, + 544 + ], + "score": 1.0, + "content": "Schölkopf, B. S2rms: Spatially structured recurrent modules. arXiv preprint arXiv:2007.06533,", + "type": "text", + "cross_page": true + } + ], + "index": 32 + }, + { + "bbox": [ + 114, + 540, + 144, + 555 + ], + "spans": [ + { + "bbox": [ + 114, + 540, + 144, + 555 + ], + "score": 1.0, + "content": "2020.", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 559, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 506, + 574 + ], + "score": 1.0, + "content": "Rand, W. M. Objective criteria for the evaluation of clustering methods. Journal of the American", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 571, + 309, + 583 + ], + "spans": [ + { + "bbox": [ + 114, + 571, + 309, + 583 + ], + "score": 1.0, + "content": "Statistical association, 66(336):846–850, 1971.", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "Raposo, D., Santoro, A., Barrett, D., Pascanu, R., Lillicrap, T., and Battaglia, P. Discovering objects", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 601, + 507, + 615 + ], + "spans": [ + { + "bbox": [ + 114, + 601, + 507, + 615 + ], + "score": 1.0, + "content": "and their relations from entangled scene representations. arXiv preprint arXiv:1702.05068, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 621, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 506, + 633 + ], + "score": 1.0, + "content": "Reed, S. and De Freitas, N. Neural programmer-interpreters. arXiv preprint arXiv:1511.06279, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 104, + 637, + 507, + 654 + ], + "spans": [ + { + "bbox": [ + 104, + 637, + 507, + 654 + ], + "score": 1.0, + "content": "Ronco, E., Gollee, H., and Gawthrop, P. J. Modular neural networks and self-decomposition.", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 651, + 264, + 663 + ], + "spans": [ + { + "bbox": [ + 115, + 651, + 264, + 663 + ], + "score": 1.0, + "content": "Technical Report CSC-96012, 1997.", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 669, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 683 + ], + "score": 1.0, + "content": "Rosenbaum, C., Klinger, T., and Riemer, M. Routing networks: Adaptive selection of non-linear", + "type": "text", + "cross_page": true + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 680, + 414, + 694 + ], + "spans": [ + { + "bbox": [ + 114, + 680, + 414, + 694 + ], + "score": 1.0, + "content": "functions for multi-task learning. arXiv preprint arXiv:1711.01239, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 700, + 507, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 507, + 714 + ], + "score": 1.0, + "content": "Rosenbaum, C., Cases, I., Riemer, M., and Klinger, T. Routing networks and the challenges of", + "type": "text", + "cross_page": true + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 711, + 445, + 725 + ], + "spans": [ + { + "bbox": [ + 114, + 711, + 445, + 725 + ], + "score": 1.0, + "content": "modular and compositional computation. arXiv preprint arXiv:1904.12774, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 73, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 506, + 86 + ], + "score": 1.0, + "content": "Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M., and Monfardini, G. The graph neural network", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 84, + 389, + 96 + ], + "spans": [ + { + "bbox": [ + 115, + 84, + 389, + 96 + ], + "score": 1.0, + "content": "model. IEEE Transactions on Neural Networks, 20(1):61–80, 2008.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 101, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 506, + 116 + ], + "score": 1.0, + "content": "Schrittwieser, J., Antonoglou, I., Hubert, T., Simonyan, K., Sifre, L., Schmitt, S., Guez, A., Lockhart,", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 113, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 115, + 113, + 506, + 127 + ], + "score": 1.0, + "content": "E., Hassabis, D., Graepel, T., et al. Mastering atari, go, chess and shogi by planning with a learned", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 124, + 289, + 138 + ], + "spans": [ + { + "bbox": [ + 115, + 124, + 289, + 138 + ], + "score": 1.0, + "content": "model. Nature, 588(7839):604–609, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "Shazeer, N., Mirhoseini, A., Maziarz, K., Davis, A., Le, Q., Hinton, G., and Dean, J. Outra-", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 153, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 114, + 153, + 506, + 168 + ], + "score": 1.0, + "content": "geously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 164, + 221, + 177 + ], + "spans": [ + { + "bbox": [ + 114, + 164, + 221, + 177 + ], + "score": 1.0, + "content": "arXiv:1701.06538, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 182, + 507, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 507, + 196 + ], + "score": 1.0, + "content": "Tacchetti, A., Song, H. F., Mediano, P. A., Zambaldi, V., Rabinowitz, N. C., Graepel, T., Botvinick,", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 193, + 506, + 208 + ], + "spans": [ + { + "bbox": [ + 114, + 193, + 506, + 208 + ], + "score": 1.0, + "content": "M., and Battaglia, P. W. Relational forward models for multi-agent learning. arXiv preprint", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 205, + 219, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 205, + 219, + 217 + ], + "score": 1.0, + "content": "arXiv:1809.11044, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 222, + 507, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 507, + 237 + ], + "score": 1.0, + "content": "Touretzky, D. S. and Hinton, G. E. A distributed connectionist production system. Cognitive science,", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 234, + 205, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 234, + 205, + 248 + ], + "score": 1.0, + "content": "12(3):423–466, 1988.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "Trask, A., Hill, F., Reed, S. E., Rae, J., Dyer, C., and Blunsom, P. Neural arithmetic logic units. In", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 263, + 420, + 277 + ], + "spans": [ + { + "bbox": [ + 113, + 263, + 420, + 277 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems, pp. 8035–8044, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 282, + 507, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 507, + 296 + ], + "score": 1.0, + "content": "Van Steenkiste, S., Chang, M., Greff, K., and Schmidhuber, J. Relational neural expectation maximiza-", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 292, + 507, + 307 + ], + "spans": [ + { + "bbox": [ + 114, + 292, + 507, + 307 + ], + "score": 1.0, + "content": "tion: Unsupervised discovery of objects and their interactions. arXiv preprint arXiv:1802.10353,", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 303, + 142, + 318 + ], + "spans": [ + { + "bbox": [ + 115, + 303, + 142, + 318 + ], + "score": 1.0, + "content": "2018.", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 322, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 506, + 336 + ], + "score": 1.0, + "content": "Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 334, + 304, + 346 + ], + "spans": [ + { + "bbox": [ + 115, + 334, + 304, + 346 + ], + "score": 1.0, + "content": "Polosukhin, I. Attention is all you need, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 352, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 506, + 366 + ], + "score": 1.0, + "content": "Veerapaneni, R., Co-Reyes, J. D., Chang, M., Janner, M., Finn, C., Wu, J., Tenenbaum, J. B., and", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 363, + 507, + 377 + ], + "spans": [ + { + "bbox": [ + 115, + 363, + 507, + 377 + ], + "score": 1.0, + "content": "Levine, S. Entity abstraction in visual model-based reinforcement learning. CoRR, abs/1910.12827,", + "type": "text", + "cross_page": true + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 373, + 333, + 388 + ], + "spans": [ + { + "bbox": [ + 115, + 373, + 333, + 388 + ], + "score": 1.0, + "content": "2019. URL http://arxiv.org/abs/1910.12827.", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 392, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 505, + 406 + ], + "score": 1.0, + "content": "Veerapaneni, R., Co-Reyes, J. D., Chang, M., Janner, M., Finn, C., Wu, J., Tenenbaum, J., and", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 404, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 115, + 404, + 506, + 418 + ], + "score": 1.0, + "content": "Levine, S. Entity abstraction in visual model-based reinforcement learning. In Conference on", + "type": "text", + "cross_page": true + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 414, + 306, + 428 + ], + "spans": [ + { + "bbox": [ + 115, + 414, + 306, + 428 + ], + "score": 1.0, + "content": "Robot Learning, pp. 1439–1456. PMLR, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 433, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 506, + 447 + ], + "score": 1.0, + "content": "Watters, N., Zoran, D., Weber, T., Battaglia, P., Pascanu, R., and Tacchetti, A. Visual interaction", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 443, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 114, + 443, + 506, + 459 + ], + "score": 1.0, + "content": "networks: Learning a physics simulator from video. In Advances in neural information processing", + "type": "text", + "cross_page": true + } + ], + "index": 27 + }, + { + "bbox": [ + 114, + 455, + 242, + 468 + ], + "spans": [ + { + "bbox": [ + 114, + 455, + 242, + 468 + ], + "score": 1.0, + "content": "systems, pp. 4539–4547, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 473, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 506, + 488 + ], + "score": 1.0, + "content": "Weis, M. A., Chitta, K., Sharma, Y., Brendel, W., Bethge, M., Geiger, A., and Ecker, A. S. Unmasking", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 484, + 462, + 498 + ], + "spans": [ + { + "bbox": [ + 115, + 484, + 462, + 498 + ], + "score": 1.0, + "content": "the inductive biases of unsupervised object representations for video sequences, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 501, + 507, + 518 + ], + "spans": [ + { + "bbox": [ + 104, + 501, + 507, + 518 + ], + "score": 1.0, + "content": "Xu, D., Nair, S., Zhu, Y., Gao, J., Garg, A., Fei-Fei, L., and Savarese, S. Neural task programming:", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 513, + 507, + 529 + ], + "spans": [ + { + "bbox": [ + 114, + 513, + 507, + 529 + ], + "score": 1.0, + "content": "Learning to generalize across hierarchical tasks. In 2018 IEEE International Conference on", + "type": "text", + "cross_page": true + } + ], + "index": 32 + }, + { + "bbox": [ + 114, + 524, + 341, + 538 + ], + "spans": [ + { + "bbox": [ + 114, + 524, + 341, + 538 + ], + "score": 1.0, + "content": "Robotics and Automation (ICRA), pp. 1–8. IEEE, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "score": 1.0, + "content": "Zablotskaia, P., Dominici, E. A., Sigal, L., and Lehrmann, A. M. Unsupervised video decomposition", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 555, + 445, + 568 + ], + "spans": [ + { + "bbox": [ + 114, + 555, + 445, + 568 + ], + "score": 1.0, + "content": "using spatio-temporal iterative inference. arXiv preprint arXiv:2006.14727, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_end_line": true + } + ], + "index": 25.5, + "bbox_fs": [ + 104, + 198, + 507, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 63, + 507, + 730 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 506, + 85 + ], + "score": 1.0, + "content": "Eslami, S., Heess, N., Weber, T., Tassa, Y., Szepesvari, D., Kavukcuoglu, K., and Hinton, G. E. Attend,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 114, + 84, + 507, + 97 + ], + "spans": [ + { + "bbox": [ + 114, + 84, + 507, + 97 + ], + "score": 1.0, + "content": "infer, repeat: Fast scene understanding with generative models. arXiv preprint arXiv:1603.08575,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 94, + 143, + 109 + ], + "spans": [ + { + "bbox": [ + 114, + 94, + 143, + 109 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 111, + 507, + 128 + ], + "spans": [ + { + "bbox": [ + 104, + 111, + 507, + 128 + ], + "score": 1.0, + "content": "Evans, R., Hernández-Orallo, J., Welbl, J., Kohli, P., and Sergot, M. Making sense of sensory input.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 125, + 281, + 138 + ], + "spans": [ + { + "bbox": [ + 114, + 125, + 281, + 138 + ], + "score": 1.0, + "content": "Artificial Intelligence, pp. 103438, 2019.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 104, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "Fedus, W., Zoph, B., and Shazeer, N. Switch transformers: Scaling to trillion parameter models with", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 153, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 114, + 153, + 506, + 169 + ], + "score": 1.0, + "content": "simple and efficient sparsity. CoRR, abs/2101.03961, 2021. URL https://arxiv.org/abs/", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 167, + 174, + 178 + ], + "spans": [ + { + "bbox": [ + 116, + 167, + 174, + 178 + ], + "score": 1.0, + "content": "2101.03961.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 183, + 507, + 198 + ], + "spans": [ + { + "bbox": [ + 104, + 183, + 507, + 198 + ], + "score": 1.0, + "content": "Fernando, C., Banarse, D., Blundell, C., Zwols, Y., Ha, D., Rusu, A. A., Pritzel, A., and Wierstra,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 113, + 194, + 507, + 210 + ], + "spans": [ + { + "bbox": [ + 113, + 194, + 507, + 210 + ], + "score": 1.0, + "content": "D. Pathnet: Evolution channels gradient descent in super neural networks. arXiv preprint", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 206, + 220, + 220 + ], + "spans": [ + { + "bbox": [ + 115, + 206, + 220, + 220 + ], + "score": 1.0, + "content": "arXiv:1701.08734, 2017.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 225, + 507, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 507, + 239 + ], + "score": 1.0, + "content": "Girshick, R. B. Fast R-CNN. CoRR, abs/1504.08083, 2015. URL http://arxiv.org/abs/1504.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 237, + 149, + 249 + ], + "spans": [ + { + "bbox": [ + 115, + 237, + 149, + 249 + ], + "score": 1.0, + "content": "08083.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "Goyal, A. and Bengio, Y. Inductive biases for deep learning of higher-level cognition. arXiv preprint", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 267, + 219, + 279 + ], + "spans": [ + { + "bbox": [ + 115, + 267, + 219, + 279 + ], + "score": 1.0, + "content": "arXiv:2011.15091, 2020.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "score": 1.0, + "content": "Goyal, A., Lamb, A., Hoffmann, J., Sodhani, S., Levine, S., Bengio, Y., and Schölkopf, B. Recurrent", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 298, + 248, + 309 + ], + "spans": [ + { + "bbox": [ + 116, + 298, + 248, + 309 + ], + "score": 1.0, + "content": "independent mechanisms, 2019.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 315, + 507, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 507, + 329 + ], + "score": 1.0, + "content": "Goyal, A., Lamb, A., Gampa, P., Beaudoin, P., Levine, S., Blundell, C., Bengio, Y., and Mozer, M.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 326, + 507, + 340 + ], + "spans": [ + { + "bbox": [ + 115, + 326, + 507, + 340 + ], + "score": 1.0, + "content": "Object files and schemata: Factorizing declarative and procedural knowledge in dynamical systems.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 337, + 279, + 351 + ], + "spans": [ + { + "bbox": [ + 115, + 337, + 279, + 351 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2006.16225, 2020.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 356, + 507, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 507, + 370 + ], + "score": 1.0, + "content": "Graves, A., Wayne, G., and Danihelka, I. Neural turing machines. arXiv preprint arXiv:1410.5401,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 114, + 366, + 144, + 381 + ], + "spans": [ + { + "bbox": [ + 114, + 366, + 144, + 381 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 384, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 104, + 384, + 506, + 402 + ], + "score": 1.0, + "content": "Greff, K., Rasmus, A., Berglund, M., Hao, T. H., Schmidhuber, J., and Valpola, H. Tagger: Deep", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 397, + 420, + 411 + ], + "spans": [ + { + "bbox": [ + 115, + 397, + 420, + 411 + ], + "score": 1.0, + "content": "unsupervised perceptual grouping. arXiv preprint arXiv:1606.06724, 2016.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 417, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 506, + 429 + ], + "score": 1.0, + "content": "Greff, K., Kaufman, R. L., Kabra, R., Watters, N., Burgess, C., Zoran, D., Matthey, L., Botvinick, M.,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 427, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 115, + 427, + 505, + 441 + ], + "score": 1.0, + "content": "and Lerchner, A. Multi-object representation learning with iterative variational inference. arXiv", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 438, + 254, + 451 + ], + "spans": [ + { + "bbox": [ + 114, + 438, + 254, + 451 + ], + "score": 1.0, + "content": "preprint arXiv:1903.00450, 2019.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 457, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 471 + ], + "score": 1.0, + "content": "Groth, O., Fuchs, F., Posner, I., and Vedaldi, A. Shapestacks: Learning vision-based physical intuition", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 468, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 115, + 468, + 505, + 482 + ], + "score": 1.0, + "content": "for generalised object stacking. CoRR, abs/1804.08018, 2018. URL http://arxiv.org/abs/", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 117, + 480, + 174, + 492 + ], + "spans": [ + { + "bbox": [ + 117, + 480, + 174, + 492 + ], + "score": 1.0, + "content": "1804.08018.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 497, + 507, + 511 + ], + "spans": [ + { + "bbox": [ + 104, + 497, + 507, + 511 + ], + "score": 1.0, + "content": "He, Z., Li, J., Liu, D., He, H., and Barber, D. Tracking by animation: Unsupervised learning of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 509, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 115, + 509, + 505, + 523 + ], + "score": 1.0, + "content": "multi-object attentive trackers. CoRR, abs/1809.03137, 2018. URL http://arxiv.org/abs/", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 519, + 174, + 533 + ], + "spans": [ + { + "bbox": [ + 115, + 519, + 174, + 533 + ], + "score": 1.0, + "content": "1809.03137.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 537, + 507, + 554 + ], + "spans": [ + { + "bbox": [ + 104, + 537, + 507, + 554 + ], + "score": 1.0, + "content": "Higgins, I., Matthey, L., Pal, A., Burgess, C., Glorot, X., Botvinick, M., Mohamed, S., and Lerchner,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 114, + 549, + 489, + 563 + ], + "spans": [ + { + "bbox": [ + 114, + 549, + 489, + 563 + ], + "score": 1.0, + "content": "A. beta-vae: Learning basic visual concepts with a constrained variational framework. 2016.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 568, + 507, + 582 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 507, + 582 + ], + "score": 1.0, + "content": "Jacobs, R. A., Jordan, M. I., Nowlan, S. J., Hinton, G. E., et al. Adaptive mixtures of local experts.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 114, + 579, + 277, + 593 + ], + "spans": [ + { + "bbox": [ + 114, + 579, + 277, + 593 + ], + "score": 1.0, + "content": "Neural computation, 3(1):79–87, 1991.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "Jang, E., Gu, S., and Poole, B. Categorical reparameterization with gumbel-softmax. arXiv preprint", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 114, + 610, + 219, + 623 + ], + "spans": [ + { + "bbox": [ + 114, + 610, + 219, + 623 + ], + "score": 1.0, + "content": "arXiv:1611.01144, 2016.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 628, + 507, + 642 + ], + "spans": [ + { + "bbox": [ + 104, + 628, + 507, + 642 + ], + "score": 1.0, + "content": "Ke, N. R., Didolkar, A. R., Mittal, S., Goyal, A., Lajoie, G., Bauer, S., Rezende, D. J., Mozer,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 114, + 640, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 114, + 640, + 506, + 654 + ], + "score": 1.0, + "content": "M. C., Bengio, Y., and Pal, C. Systematic evaluation of causal discovery in visual model based", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 650, + 240, + 664 + ], + "spans": [ + { + "bbox": [ + 115, + 650, + 240, + 664 + ], + "score": 1.0, + "content": "reinforcement learning. 2021.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "Kipf, T., Fetaya, E., Wang, K.-C., Welling, M., and Zemel, R. Neural relational inference for", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 114, + 681, + 362, + 694 + ], + "spans": [ + { + "bbox": [ + 114, + 681, + 362, + 694 + ], + "score": 1.0, + "content": "interacting systems. arXiv preprint arXiv:1802.04687, 2018.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "Kipf, T., van der Pol, E., and Welling, M. Contrastive learning of structured world models. arXiv", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 114, + 710, + 254, + 725 + ], + "spans": [ + { + "bbox": [ + 114, + 710, + 254, + 725 + ], + "score": 1.0, + "content": "preprint arXiv:1911.12247, 2019.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 22.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 751 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 104, + 63, + 507, + 730 + ], + "lines": [], + "index": 22.5, + "bbox_fs": [ + 104, + 73, + 507, + 725 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 49, + 507, + 727 + ], + "lines": [ + { + "bbox": [ + 105, + 71, + 507, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 71, + 507, + 87 + ], + "score": 1.0, + "content": "Kirsch, L., Kunze, J., and Barber, D. Modular networks: Learning to decompose neural computation.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 114, + 84, + 432, + 98 + ], + "spans": [ + { + "bbox": [ + 114, + 84, + 432, + 98 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems, pp. 2408–2418, 2018.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 101, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 506, + 117 + ], + "score": 1.0, + "content": "Kosiorek, A., Kim, H., Teh, Y. W., and Posner, I. Sequential attend, infer, repeat: Generative", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 115, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 115, + 115, + 506, + 127 + ], + "score": 1.0, + "content": "modelling of moving objects. Advances in Neural Information Processing Systems, 31:8606–8616,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 124, + 144, + 139 + ], + "spans": [ + { + "bbox": [ + 114, + 124, + 144, + 139 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 506, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 159 + ], + "score": 1.0, + "content": "Laird, J. E., Rosenbloom, P. S., and Newell, A. Chunking in soar: The anatomy of a general learning", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 154, + 318, + 168 + ], + "spans": [ + { + "bbox": [ + 114, + 154, + 318, + 168 + ], + "score": 1.0, + "content": "mechanism. Machine learning, 1(1):11–46, 1986.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 174, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 505, + 186 + ], + "score": 1.0, + "content": "Lamb, A., Goyal, A., Słowik, A., Mozer, M., Beaudoin, P., and Bengio, Y. Neural function modules", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 114, + 185, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 114, + 185, + 506, + 199 + ], + "score": 1.0, + "content": "with sparse arguments: A dynamic approach to integrating information across layers. arXiv", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 196, + 254, + 209 + ], + "spans": [ + { + "bbox": [ + 114, + 196, + 254, + 209 + ], + "score": 1.0, + "content": "preprint arXiv:2010.08012, 2020.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 214, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 104, + 214, + 506, + 230 + ], + "score": 1.0, + "content": "Le Roux, N., Heess, N., Shotton, J., and Winn, J. Learning a generative model of images by factoring", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 114, + 225, + 387, + 240 + ], + "spans": [ + { + "bbox": [ + 114, + 225, + 387, + 240 + ], + "score": 1.0, + "content": "appearance and shape. Neural Computation, 23(3):593–650, 2011.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 245, + 507, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 507, + 259 + ], + "score": 1.0, + "content": "Li, Y., Gimeno, F., Kohli, P., and Vinyals, O. Strong generalization and efficiency in neural programs.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 113, + 256, + 280, + 270 + ], + "spans": [ + { + "bbox": [ + 113, + 256, + 280, + 270 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2007.03629, 2020.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 274, + 507, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 507, + 290 + ], + "score": 1.0, + "content": "Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 113, + 285, + 507, + 300 + ], + "spans": [ + { + "bbox": [ + 113, + 285, + 507, + 300 + ], + "score": 1.0, + "content": "A., and Kipf, T. Object-centric learning with slot attention. arXiv preprint arXiv:2006.15055,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 295, + 148, + 312 + ], + "spans": [ + { + "bbox": [ + 114, + 295, + 148, + 312 + ], + "score": 1.0, + "content": "2020a.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 315, + 507, + 331 + ], + "spans": [ + { + "bbox": [ + 104, + 315, + 507, + 331 + ], + "score": 1.0, + "content": "Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 113, + 327, + 385, + 340 + ], + "spans": [ + { + "bbox": [ + 113, + 327, + 385, + 340 + ], + "score": 1.0, + "content": "A., and Kipf, T. Object-centric learning with slot attention, 2020b.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 346, + 507, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 507, + 361 + ], + "score": 1.0, + "content": "Lovett, M. C. and Anderson, J. R. Thinking as a production system. The Cambridge handbook of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 357, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 114, + 357, + 294, + 371 + ], + "score": 1.0, + "content": "thinking and reasoning, pp. 401–429, 2005.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 375, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 104, + 375, + 506, + 390 + ], + "score": 1.0, + "content": "McMillan, C., Mozer, M. C., and Smolensky, P. The connectionist scientist game: rule extraction and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 114, + 387, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 114, + 387, + 506, + 401 + ], + "score": 1.0, + "content": "refinement in a neural network. In Proceedings of the 13th Annual Conference of the Cognitive", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 114, + 398, + 264, + 412 + ], + "spans": [ + { + "bbox": [ + 114, + 398, + 264, + 412 + ], + "score": 1.0, + "content": "Science Society, pp. 424–430, 1991.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 416, + 507, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 507, + 431 + ], + "score": 1.0, + "content": "Milan, A., Leal-Taixé, L., Reid, I. D., Roth, S., and Schindler, K. MOT16: A benchmark for multi-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 114, + 428, + 495, + 442 + ], + "spans": [ + { + "bbox": [ + 114, + 428, + 495, + 442 + ], + "score": 1.0, + "content": "object tracking. CoRR, abs/1603.00831, 2016. URL http://arxiv.org/abs/1603.00831.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 447, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 506, + 462 + ], + "score": 1.0, + "content": "Neelakantan, A., Le, Q. V., and Sutskever, I. Neural programmer: Inducing latent programs with", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 114, + 459, + 350, + 472 + ], + "spans": [ + { + "bbox": [ + 114, + 459, + 350, + 472 + ], + "score": 1.0, + "content": "gradient descent. arXiv preprint arXiv:1511.04834, 2015.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "Qi, H., Wang, X., Pathak, D., Ma, Y., and Malik, J. Learning long-term visual dynamics with region", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 114, + 489, + 507, + 502 + ], + "spans": [ + { + "bbox": [ + 114, + 489, + 507, + 502 + ], + "score": 1.0, + "content": "proposal interaction networks. In International Conference on Learning Representations, 2021.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 114, + 499, + 370, + 514 + ], + "spans": [ + { + "bbox": [ + 114, + 499, + 370, + 514 + ], + "score": 1.0, + "content": "URL https://openreview.net/forum?id=_X_4Akcd8Re.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "score": 1.0, + "content": "Rahaman, N., Goyal, A., Gondal, M. W., Wuthrich, M., Bauer, S., Sharma, Y., Bengio, Y., and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 529, + 507, + 544 + ], + "spans": [ + { + "bbox": [ + 114, + 529, + 507, + 544 + ], + "score": 1.0, + "content": "Schölkopf, B. S2rms: Spatially structured recurrent modules. arXiv preprint arXiv:2007.06533,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 114, + 540, + 144, + 555 + ], + "spans": [ + { + "bbox": [ + 114, + 540, + 144, + 555 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 559, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 506, + 574 + ], + "score": 1.0, + "content": "Rand, W. M. Objective criteria for the evaluation of clustering methods. Journal of the American", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 114, + 571, + 309, + 583 + ], + "spans": [ + { + "bbox": [ + 114, + 571, + 309, + 583 + ], + "score": 1.0, + "content": "Statistical association, 66(336):846–850, 1971.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "Raposo, D., Santoro, A., Barrett, D., Pascanu, R., Lillicrap, T., and Battaglia, P. Discovering objects", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 601, + 507, + 615 + ], + "spans": [ + { + "bbox": [ + 114, + 601, + 507, + 615 + ], + "score": 1.0, + "content": "and their relations from entangled scene representations. arXiv preprint arXiv:1702.05068, 2017.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 621, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 506, + 633 + ], + "score": 1.0, + "content": "Reed, S. and De Freitas, N. Neural programmer-interpreters. arXiv preprint arXiv:1511.06279, 2015.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 637, + 507, + 654 + ], + "spans": [ + { + "bbox": [ + 104, + 637, + 507, + 654 + ], + "score": 1.0, + "content": "Ronco, E., Gollee, H., and Gawthrop, P. J. Modular neural networks and self-decomposition.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 651, + 264, + 663 + ], + "spans": [ + { + "bbox": [ + 115, + 651, + 264, + 663 + ], + "score": 1.0, + "content": "Technical Report CSC-96012, 1997.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 669, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 683 + ], + "score": 1.0, + "content": "Rosenbaum, C., Klinger, T., and Riemer, M. Routing networks: Adaptive selection of non-linear", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 114, + 680, + 414, + 694 + ], + "spans": [ + { + "bbox": [ + 114, + 680, + 414, + 694 + ], + "score": 1.0, + "content": "functions for multi-task learning. arXiv preprint arXiv:1711.01239, 2017.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 700, + 507, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 507, + 714 + ], + "score": 1.0, + "content": "Rosenbaum, C., Cases, I., Riemer, M., and Klinger, T. Routing networks and the challenges of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 114, + 711, + 445, + 725 + ], + "spans": [ + { + "bbox": [ + 114, + 711, + 445, + 725 + ], + "score": 1.0, + "content": "modular and compositional computation. arXiv preprint arXiv:1904.12774, 2019.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 22 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 751 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 104, + 49, + 507, + 727 + ], + "lines": [], + "index": 22, + "bbox_fs": [ + 104, + 71, + 507, + 725 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 72, + 507, + 568 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 506, + 86 + ], + "score": 1.0, + "content": "Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M., and Monfardini, G. The graph neural network", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 84, + 389, + 96 + ], + "spans": [ + { + "bbox": [ + 115, + 84, + 389, + 96 + ], + "score": 1.0, + "content": "model. IEEE Transactions on Neural Networks, 20(1):61–80, 2008.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 101, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 506, + 116 + ], + "score": 1.0, + "content": "Schrittwieser, J., Antonoglou, I., Hubert, T., Simonyan, K., Sifre, L., Schmitt, S., Guez, A., Lockhart,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 113, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 115, + 113, + 506, + 127 + ], + "score": 1.0, + "content": "E., Hassabis, D., Graepel, T., et al. Mastering atari, go, chess and shogi by planning with a learned", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 124, + 289, + 138 + ], + "spans": [ + { + "bbox": [ + 115, + 124, + 289, + 138 + ], + "score": 1.0, + "content": "model. Nature, 588(7839):604–609, 2020.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "Shazeer, N., Mirhoseini, A., Maziarz, K., Davis, A., Le, Q., Hinton, G., and Dean, J. Outra-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 153, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 114, + 153, + 506, + 168 + ], + "score": 1.0, + "content": "geously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 164, + 221, + 177 + ], + "spans": [ + { + "bbox": [ + 114, + 164, + 221, + 177 + ], + "score": 1.0, + "content": "arXiv:1701.06538, 2017.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 182, + 507, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 507, + 196 + ], + "score": 1.0, + "content": "Tacchetti, A., Song, H. F., Mediano, P. A., Zambaldi, V., Rabinowitz, N. C., Graepel, T., Botvinick,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 193, + 506, + 208 + ], + "spans": [ + { + "bbox": [ + 114, + 193, + 506, + 208 + ], + "score": 1.0, + "content": "M., and Battaglia, P. W. Relational forward models for multi-agent learning. arXiv preprint", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 205, + 219, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 205, + 219, + 217 + ], + "score": 1.0, + "content": "arXiv:1809.11044, 2018.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 222, + 507, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 507, + 237 + ], + "score": 1.0, + "content": "Touretzky, D. S. and Hinton, G. E. A distributed connectionist production system. Cognitive science,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 234, + 205, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 234, + 205, + 248 + ], + "score": 1.0, + "content": "12(3):423–466, 1988.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "Trask, A., Hill, F., Reed, S. E., Rae, J., Dyer, C., and Blunsom, P. Neural arithmetic logic units. In", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 113, + 263, + 420, + 277 + ], + "spans": [ + { + "bbox": [ + 113, + 263, + 420, + 277 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems, pp. 8035–8044, 2018.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 282, + 507, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 507, + 296 + ], + "score": 1.0, + "content": "Van Steenkiste, S., Chang, M., Greff, K., and Schmidhuber, J. Relational neural expectation maximiza-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 292, + 507, + 307 + ], + "spans": [ + { + "bbox": [ + 114, + 292, + 507, + 307 + ], + "score": 1.0, + "content": "tion: Unsupervised discovery of objects and their interactions. arXiv preprint arXiv:1802.10353,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 303, + 142, + 318 + ], + "spans": [ + { + "bbox": [ + 115, + 303, + 142, + 318 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 322, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 506, + 336 + ], + "score": 1.0, + "content": "Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 334, + 304, + 346 + ], + "spans": [ + { + "bbox": [ + 115, + 334, + 304, + 346 + ], + "score": 1.0, + "content": "Polosukhin, I. Attention is all you need, 2017.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 352, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 506, + 366 + ], + "score": 1.0, + "content": "Veerapaneni, R., Co-Reyes, J. D., Chang, M., Janner, M., Finn, C., Wu, J., Tenenbaum, J. B., and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 363, + 507, + 377 + ], + "spans": [ + { + "bbox": [ + 115, + 363, + 507, + 377 + ], + "score": 1.0, + "content": "Levine, S. Entity abstraction in visual model-based reinforcement learning. CoRR, abs/1910.12827,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 373, + 333, + 388 + ], + "spans": [ + { + "bbox": [ + 115, + 373, + 333, + 388 + ], + "score": 1.0, + "content": "2019. URL http://arxiv.org/abs/1910.12827.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 392, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 505, + 406 + ], + "score": 1.0, + "content": "Veerapaneni, R., Co-Reyes, J. D., Chang, M., Janner, M., Finn, C., Wu, J., Tenenbaum, J., and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 404, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 115, + 404, + 506, + 418 + ], + "score": 1.0, + "content": "Levine, S. Entity abstraction in visual model-based reinforcement learning. In Conference on", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 414, + 306, + 428 + ], + "spans": [ + { + "bbox": [ + 115, + 414, + 306, + 428 + ], + "score": 1.0, + "content": "Robot Learning, pp. 1439–1456. PMLR, 2020.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 433, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 506, + 447 + ], + "score": 1.0, + "content": "Watters, N., Zoran, D., Weber, T., Battaglia, P., Pascanu, R., and Tacchetti, A. Visual interaction", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 114, + 443, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 114, + 443, + 506, + 459 + ], + "score": 1.0, + "content": "networks: Learning a physics simulator from video. In Advances in neural information processing", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 114, + 455, + 242, + 468 + ], + "spans": [ + { + "bbox": [ + 114, + 455, + 242, + 468 + ], + "score": 1.0, + "content": "systems, pp. 4539–4547, 2017.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 473, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 506, + 488 + ], + "score": 1.0, + "content": "Weis, M. A., Chitta, K., Sharma, Y., Brendel, W., Bethge, M., Geiger, A., and Ecker, A. S. Unmasking", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 484, + 462, + 498 + ], + "spans": [ + { + "bbox": [ + 115, + 484, + 462, + 498 + ], + "score": 1.0, + "content": "the inductive biases of unsupervised object representations for video sequences, 2020.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 501, + 507, + 518 + ], + "spans": [ + { + "bbox": [ + 104, + 501, + 507, + 518 + ], + "score": 1.0, + "content": "Xu, D., Nair, S., Zhu, Y., Gao, J., Garg, A., Fei-Fei, L., and Savarese, S. Neural task programming:", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 513, + 507, + 529 + ], + "spans": [ + { + "bbox": [ + 114, + 513, + 507, + 529 + ], + "score": 1.0, + "content": "Learning to generalize across hierarchical tasks. In 2018 IEEE International Conference on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 114, + 524, + 341, + 538 + ], + "spans": [ + { + "bbox": [ + 114, + 524, + 341, + 538 + ], + "score": 1.0, + "content": "Robotics and Automation (ICRA), pp. 1–8. IEEE, 2018.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "score": 1.0, + "content": "Zablotskaia, P., Dominici, E. A., Sigal, L., and Lehrmann, A. M. Unsupervised video decomposition", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 114, + 555, + 445, + 568 + ], + "spans": [ + { + "bbox": [ + 114, + 555, + 445, + 568 + ], + "score": 1.0, + "content": "using spatio-temporal iterative inference. arXiv preprint arXiv:2006.14727, 2020.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 107, + 585, + 156, + 598 + ], + "lines": [ + { + "bbox": [ + 106, + 585, + 158, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 158, + 599 + ], + "score": 1.0, + "content": "Checklist", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 131, + 606, + 208, + 618 + ], + "lines": [ + { + "bbox": [ + 129, + 605, + 210, + 619 + ], + "spans": [ + { + "bbox": [ + 129, + 605, + 210, + 619 + ], + "score": 1.0, + "content": "1. For all authors...", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 146, + 621, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 145, + 620, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 145, + 620, + 505, + 634 + ], + "score": 1.0, + "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 162, + 632, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 162, + 632, + 288, + 645 + ], + "score": 1.0, + "content": "contributions and scope? [Yes]", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 145, + 645, + 376, + 657 + ], + "spans": [ + { + "bbox": [ + 145, + 645, + 376, + 657 + ], + "score": 1.0, + "content": "(b) Did you describe the limitations of your work? [Yes]", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 145, + 657, + 468, + 670 + ], + "spans": [ + { + "bbox": [ + 145, + 657, + 468, + 670 + ], + "score": 1.0, + "content": "(c) Did you discuss any potential negative societal impacts of your work? [Yes]", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 145, + 669, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 145, + 669, + 505, + 683 + ], + "score": 1.0, + "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 161, + 681, + 214, + 694 + ], + "spans": [ + { + "bbox": [ + 161, + 681, + 214, + 694 + ], + "score": 1.0, + "content": "them? [Yes]", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 131, + 696, + 302, + 708 + ], + "lines": [ + { + "bbox": [ + 129, + 695, + 304, + 709 + ], + "spans": [ + { + "bbox": [ + 129, + 695, + 304, + 709 + ], + "score": 1.0, + "content": "2. If you are including theoretical results...", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 146, + 711, + 452, + 722 + ], + "lines": [ + { + "bbox": [ + 145, + 710, + 453, + 724 + ], + "spans": [ + { + "bbox": [ + 145, + 710, + 453, + 724 + ], + "score": 1.0, + "content": "(a) Did you state the full set of assumptions of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 72, + 507, + 568 + ], + "lines": [], + "index": 17.5, + "bbox_fs": [ + 104, + 73, + 507, + 568 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 585, + 156, + 598 + ], + "lines": [ + { + "bbox": [ + 106, + 585, + 158, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 158, + 599 + ], + "score": 1.0, + "content": "Checklist", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 131, + 606, + 208, + 618 + ], + "lines": [ + { + "bbox": [ + 129, + 605, + 210, + 619 + ], + "spans": [ + { + "bbox": [ + 129, + 605, + 210, + 619 + ], + "score": 1.0, + "content": "1. For all authors...", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37, + "bbox_fs": [ + 129, + 605, + 210, + 619 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 621, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 145, + 620, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 145, + 620, + 505, + 634 + ], + "score": 1.0, + "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s", + "type": "text" + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 632, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 162, + 632, + 288, + 645 + ], + "score": 1.0, + "content": "contributions and scope? [Yes]", + "type": "text" + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 645, + 376, + 657 + ], + "spans": [ + { + "bbox": [ + 145, + 645, + 376, + 657 + ], + "score": 1.0, + "content": "(b) Did you describe the limitations of your work? [Yes]", + "type": "text" + } + ], + "index": 40, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 657, + 468, + 670 + ], + "spans": [ + { + "bbox": [ + 145, + 657, + 468, + 670 + ], + "score": 1.0, + "content": "(c) Did you discuss any potential negative societal impacts of your work? [Yes]", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 669, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 145, + 669, + 505, + 683 + ], + "score": 1.0, + "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to", + "type": "text" + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 681, + 214, + 694 + ], + "spans": [ + { + "bbox": [ + 161, + 681, + 214, + 694 + ], + "score": 1.0, + "content": "them? [Yes]", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + } + ], + "index": 40.5, + "bbox_fs": [ + 145, + 620, + 505, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 131, + 696, + 302, + 708 + ], + "lines": [ + { + "bbox": [ + 129, + 695, + 304, + 709 + ], + "spans": [ + { + "bbox": [ + 129, + 695, + 304, + 709 + ], + "score": 1.0, + "content": "2. If you are including theoretical results...", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44, + "bbox_fs": [ + 129, + 695, + 304, + 709 + ] + }, + { + "type": "text", + "bbox": [ + 146, + 711, + 452, + 722 + ], + "lines": [ + { + "bbox": [ + 145, + 710, + 453, + 724 + ], + "spans": [ + { + "bbox": [ + 145, + 710, + 453, + 724 + ], + "score": 1.0, + "content": "(a) Did you state the full set of assumptions of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 145, + 72, + 424, + 86 + ], + "spans": [ + { + "bbox": [ + 145, + 72, + 424, + 86 + ], + "score": 1.0, + "content": "(b) Did you include complete proofs of all theoretical results? [N/A]", + "type": "text", + "cross_page": true + } + ], + "index": 0 + } + ], + "index": 45, + "bbox_fs": [ + 145, + 710, + 453, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 145, + 73, + 423, + 84 + ], + "lines": [ + { + "bbox": [ + 145, + 72, + 424, + 86 + ], + "spans": [ + { + "bbox": [ + 145, + 72, + 424, + 86 + ], + "score": 1.0, + "content": "(b) Did you include complete proofs of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 130, + 88, + 241, + 99 + ], + "lines": [ + { + "bbox": [ + 128, + 86, + 243, + 101 + ], + "spans": [ + { + "bbox": [ + 128, + 86, + 243, + 101 + ], + "score": 1.0, + "content": "3. If you ran experiments...", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 146, + 102, + 505, + 196 + ], + "lines": [ + { + "bbox": [ + 145, + 101, + 507, + 115 + ], + "spans": [ + { + "bbox": [ + 145, + 101, + 507, + 115 + ], + "score": 1.0, + "content": "(a) Did you include the code, data, and instructions needed to reproduce the main experi-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 113, + 447, + 126 + ], + "spans": [ + { + "bbox": [ + 162, + 113, + 447, + 126 + ], + "score": 1.0, + "content": "mental results (either in the supplemental material or as a URL)? [Yes]", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 145, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 145, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 162, + 137, + 246, + 149 + ], + "spans": [ + { + "bbox": [ + 162, + 137, + 246, + 149 + ], + "score": 1.0, + "content": "were chosen)? [Yes]", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 150, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 146, + 151, + 160, + 162 + ], + "score": 1.0, + "content": "(c)", + "type": "text" + }, + { + "bbox": [ + 160, + 150, + 506, + 163 + ], + "score": 1.0, + "content": "Did you report error bars (e.g., with respect to the random seed after running experi-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 162, + 162, + 343, + 173 + ], + "spans": [ + { + "bbox": [ + 162, + 162, + 343, + 173 + ], + "score": 1.0, + "content": "ments multiple times)? [Yes] See section 4.3", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 146, + 172, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 146, + 172, + 505, + 188 + ], + "score": 1.0, + "content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 162, + 185, + 370, + 198 + ], + "spans": [ + { + "bbox": [ + 162, + 185, + 370, + 198 + ], + "score": 1.0, + "content": "of GPUs, internal cluster, or cloud provider)? [Yes]", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 133, + 200, + 504, + 212 + ], + "lines": [ + { + "bbox": [ + 129, + 198, + 507, + 214 + ], + "spans": [ + { + "bbox": [ + 129, + 198, + 507, + 214 + ], + "score": 1.0, + "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 146, + 214, + 505, + 332 + ], + "lines": [ + { + "bbox": [ + 145, + 214, + 424, + 228 + ], + "spans": [ + { + "bbox": [ + 145, + 214, + 424, + 228 + ], + "score": 1.0, + "content": "(a) If your work uses existing assets, did you cite the creators? [Yes]", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 145, + 227, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 145, + 227, + 506, + 241 + ], + "score": 1.0, + "content": "(b) Did you mention the license of the assets? [No] All datasets we used are open-sourced", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 146, + 240, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 146, + 240, + 505, + 253 + ], + "score": 1.0, + "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [Yes]", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 162, + 250, + 377, + 264 + ], + "spans": [ + { + "bbox": [ + 162, + 250, + 377, + 264 + ], + "score": 1.0, + "content": "We include the code for reproducing our experiments", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 146, + 264, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 146, + 264, + 505, + 277 + ], + "score": 1.0, + "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 161, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 161, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "using/curating? [N/A] All datasets we use are synthetic and created using simulation", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 161, + 286, + 352, + 299 + ], + "spans": [ + { + "bbox": [ + 161, + 286, + 352, + 299 + ], + "score": 1.0, + "content": "hence do not include any personal or user data.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 146, + 299, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 146, + 299, + 505, + 311 + ], + "score": 1.0, + "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 162, + 308, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 162, + 308, + 505, + 323 + ], + "score": 1.0, + "content": "information or offensive content? [N/A] We do not use any dataset that includes any", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 162, + 321, + 291, + 333 + ], + "spans": [ + { + "bbox": [ + 162, + 321, + 291, + 333 + ], + "score": 1.0, + "content": "kind of identifying information.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 131, + 335, + 432, + 347 + ], + "lines": [ + { + "bbox": [ + 128, + 334, + 433, + 349 + ], + "spans": [ + { + "bbox": [ + 128, + 334, + 433, + 349 + ], + "score": 1.0, + "content": "5. If you used crowdsourcing or conducted research with human subjects...", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 146, + 351, + 505, + 421 + ], + "lines": [ + { + "bbox": [ + 145, + 349, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 145, + 349, + 506, + 363 + ], + "score": 1.0, + "content": "(a) Did you include the full text of instructions given to participants and screenshots, if", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 161, + 361, + 237, + 374 + ], + "spans": [ + { + "bbox": [ + 161, + 361, + 237, + 374 + ], + "score": 1.0, + "content": "applicable? [N/A]", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 145, + 374, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 145, + 374, + 505, + 387 + ], + "score": 1.0, + "content": "(b) Did you describe any potential participant risks, with links to Institutional Review", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 161, + 385, + 342, + 398 + ], + "spans": [ + { + "bbox": [ + 161, + 385, + 342, + 398 + ], + "score": 1.0, + "content": "Board (IRB) approvals, if applicable? [N/A]", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 147, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 147, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 162, + 409, + 333, + 422 + ], + "spans": [ + { + "bbox": [ + 162, + 409, + 333, + 422 + ], + "score": 1.0, + "content": "spent on participant compensation? [N/A]", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 145, + 73, + 423, + 84 + ], + "lines": [], + "index": 0, + "bbox_fs": [ + 145, + 72, + 424, + 86 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 130, + 88, + 241, + 99 + ], + "lines": [ + { + "bbox": [ + 128, + 86, + 243, + 101 + ], + "spans": [ + { + "bbox": [ + 128, + 86, + 243, + 101 + ], + "score": 1.0, + "content": "3. If you ran experiments...", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 128, + 86, + 243, + 101 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 102, + 505, + 196 + ], + "lines": [ + { + "bbox": [ + 145, + 101, + 507, + 115 + ], + "spans": [ + { + "bbox": [ + 145, + 101, + 507, + 115 + ], + "score": 1.0, + "content": "(a) Did you include the code, data, and instructions needed to reproduce the main experi-", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 113, + 447, + 126 + ], + "spans": [ + { + "bbox": [ + 162, + 113, + 447, + 126 + ], + "score": 1.0, + "content": "mental results (either in the supplemental material or as a URL)? [Yes]", + "type": "text" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 145, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 137, + 246, + 149 + ], + "spans": [ + { + "bbox": [ + 162, + 137, + 246, + 149 + ], + "score": 1.0, + "content": "were chosen)? [Yes]", + "type": "text" + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 150, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 146, + 151, + 160, + 162 + ], + "score": 1.0, + "content": "(c)", + "type": "text" + }, + { + "bbox": [ + 160, + 150, + 506, + 163 + ], + "score": 1.0, + "content": "Did you report error bars (e.g., with respect to the random seed after running experi-", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 162, + 343, + 173 + ], + "spans": [ + { + "bbox": [ + 162, + 162, + 343, + 173 + ], + "score": 1.0, + "content": "ments multiple times)? [Yes] See section 4.3", + "type": "text" + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 172, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 146, + 172, + 505, + 188 + ], + "score": 1.0, + "content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 185, + 370, + 198 + ], + "spans": [ + { + "bbox": [ + 162, + 185, + 370, + 198 + ], + "score": 1.0, + "content": "of GPUs, internal cluster, or cloud provider)? [Yes]", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + } + ], + "index": 5.5, + "bbox_fs": [ + 145, + 101, + 507, + 198 + ] + }, + { + "type": "text", + "bbox": [ + 133, + 200, + 504, + 212 + ], + "lines": [ + { + "bbox": [ + 129, + 198, + 507, + 214 + ], + "spans": [ + { + "bbox": [ + 129, + 198, + 507, + 214 + ], + "score": 1.0, + "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 129, + 198, + 507, + 214 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 214, + 505, + 332 + ], + "lines": [ + { + "bbox": [ + 145, + 214, + 424, + 228 + ], + "spans": [ + { + "bbox": [ + 145, + 214, + 424, + 228 + ], + "score": 1.0, + "content": "(a) If your work uses existing assets, did you cite the creators? [Yes]", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 227, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 145, + 227, + 506, + 241 + ], + "score": 1.0, + "content": "(b) Did you mention the license of the assets? [No] All datasets we used are open-sourced", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 146, + 240, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 146, + 240, + 505, + 253 + ], + "score": 1.0, + "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [Yes]", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 250, + 377, + 264 + ], + "spans": [ + { + "bbox": [ + 162, + 250, + 377, + 264 + ], + "score": 1.0, + "content": "We include the code for reproducing our experiments", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 264, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 146, + 264, + 505, + 277 + ], + "score": 1.0, + "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 161, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "using/curating? [N/A] All datasets we use are synthetic and created using simulation", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 161, + 286, + 352, + 299 + ], + "spans": [ + { + "bbox": [ + 161, + 286, + 352, + 299 + ], + "score": 1.0, + "content": "hence do not include any personal or user data.", + "type": "text" + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 299, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 146, + 299, + 505, + 311 + ], + "score": 1.0, + "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 308, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 162, + 308, + 505, + 323 + ], + "score": 1.0, + "content": "information or offensive content? [N/A] We do not use any dataset that includes any", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 162, + 321, + 291, + 333 + ], + "spans": [ + { + "bbox": [ + 162, + 321, + 291, + 333 + ], + "score": 1.0, + "content": "kind of identifying information.", + "type": "text" + } + ], + "index": 20, + "is_list_end_line": true + } + ], + "index": 15.5, + "bbox_fs": [ + 145, + 214, + 506, + 333 + ] + }, + { + "type": "text", + "bbox": [ + 131, + 335, + 432, + 347 + ], + "lines": [ + { + "bbox": [ + 128, + 334, + 433, + 349 + ], + "spans": [ + { + "bbox": [ + 128, + 334, + 433, + 349 + ], + "score": 1.0, + "content": "5. If you used crowdsourcing or conducted research with human subjects...", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 128, + 334, + 433, + 349 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 351, + 505, + 421 + ], + "lines": [ + { + "bbox": [ + 145, + 349, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 145, + 349, + 506, + 363 + ], + "score": 1.0, + "content": "(a) Did you include the full text of instructions given to participants and screenshots, if", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 361, + 237, + 374 + ], + "spans": [ + { + "bbox": [ + 161, + 361, + 237, + 374 + ], + "score": 1.0, + "content": "applicable? [N/A]", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 374, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 145, + 374, + 505, + 387 + ], + "score": 1.0, + "content": "(b) Did you describe any potential participant risks, with links to Institutional Review", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 385, + 342, + 398 + ], + "spans": [ + { + "bbox": [ + 161, + 385, + 342, + 398 + ], + "score": 1.0, + "content": "Board (IRB) approvals, if applicable? [N/A]", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 147, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 147, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 409, + 333, + 422 + ], + "spans": [ + { + "bbox": [ + 162, + 409, + 333, + 422 + ], + "score": 1.0, + "content": "spent on participant compensation? [N/A]", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + } + ], + "index": 24.5, + "bbox_fs": [ + 145, + 349, + 506, + 422 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/xQGYquca0gB/xQGYquca0gB_model.json b/parse/train/xQGYquca0gB/xQGYquca0gB_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9adad3dd733d8b29d43c6f625f33d504aaf5cb34 --- /dev/null +++ b/parse/train/xQGYquca0gB/xQGYquca0gB_model.json @@ -0,0 +1,19602 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 658, + 1304, + 658, + 1304, + 1204, + 398, + 1204 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1336, + 1405, + 1336, + 1405, + 1761, + 298, + 1761 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 301, + 1775, + 1398, + 1775, + 1398, + 1867, + 301, + 1867 + ], + "score": 0.954 + }, + { + "category_id": 2, + "poly": [ + 298, + 1892, + 1405, + 1892, + 1405, + 1980, + 298, + 1980 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 382, + 440, + 1323, + 440, + 1323, + 508, + 382, + 508 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 299, + 1265, + 530, + 1265, + 530, + 1302, + 299, + 1302 + ], + "score": 0.908 + }, + { + "category_id": 2, + "poly": [ + 297, + 2033, + 1071, + 2033, + 1071, + 2062, + 297, + 2062 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 788, + 586, + 912, + 586, + 912, + 623, + 788, + 623 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 567, + 272, + 1132, + 272, + 1132, + 325, + 567, + 325 + ], + "score": 0.857 + }, + { + "category_id": 0, + "poly": [ + 566, + 272, + 1132, + 272, + 1132, + 326, + 566, + 326 + ], + "score": 0.099 + }, + { + "category_id": 13, + "poly": [ + 586, + 1610, + 614, + 1610, + 614, + 1636, + 586, + 1636 + ], + "score": 0.83, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 729, + 1610, + 757, + 1610, + 757, + 1636, + 729, + 1636 + ], + "score": 0.82, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1031, + 441, + 1085, + 441, + 1085, + 473, + 1031, + 473 + ], + "score": 0.42, + "latex": "\\mathbf { K } \\mathbf { e } ^ { 2 }" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1888.0, + 1405.0, + 1888.0, + 1405.0, + 1928.0, + 339.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1917.0, + 1409.0, + 1917.0, + 1409.0, + 1957.0, + 292.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1948.0, + 594.0, + 1948.0, + 594.0, + 1983.0, + 295.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1261.0, + 536.0, + 1261.0, + 536.0, + 1308.0, + 292.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2028.0, + 1073.0, + 2028.0, + 1073.0, + 2066.0, + 294.0, + 2066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 586.0, + 917.0, + 586.0, + 917.0, + 625.0, + 784.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 266.0, + 1136.0, + 266.0, + 1136.0, + 332.0, + 562.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 266.0, + 1137.0, + 266.0, + 1137.0, + 332.0, + 563.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 658.0, + 1305.0, + 658.0, + 1305.0, + 692.0, + 396.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 691.0, + 1303.0, + 691.0, + 1303.0, + 721.0, + 396.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 718.0, + 1308.0, + 718.0, + 1308.0, + 756.0, + 394.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 750.0, + 1309.0, + 750.0, + 1309.0, + 783.0, + 393.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 779.0, + 1306.0, + 779.0, + 1306.0, + 814.0, + 393.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 811.0, + 1309.0, + 811.0, + 1309.0, + 844.0, + 395.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 842.0, + 1308.0, + 842.0, + 1308.0, + 875.0, + 396.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 872.0, + 1305.0, + 872.0, + 1305.0, + 904.0, + 394.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 902.0, + 1303.0, + 902.0, + 1303.0, + 932.0, + 395.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 931.0, + 1305.0, + 931.0, + 1305.0, + 966.0, + 393.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 963.0, + 1305.0, + 963.0, + 1305.0, + 996.0, + 394.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 994.0, + 1304.0, + 994.0, + 1304.0, + 1023.0, + 396.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1020.0, + 1306.0, + 1020.0, + 1306.0, + 1060.0, + 392.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1053.0, + 1306.0, + 1053.0, + 1306.0, + 1086.0, + 394.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1083.0, + 1308.0, + 1083.0, + 1308.0, + 1117.0, + 391.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1112.0, + 1308.0, + 1112.0, + 1308.0, + 1148.0, + 393.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1145.0, + 1305.0, + 1145.0, + 1305.0, + 1177.0, + 393.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1173.0, + 1193.0, + 1173.0, + 1193.0, + 1208.0, + 393.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1334.0, + 1405.0, + 1334.0, + 1405.0, + 1372.0, + 293.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1367.0, + 1406.0, + 1367.0, + 1406.0, + 1402.0, + 293.0, + 1402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1396.0, + 1406.0, + 1396.0, + 1406.0, + 1432.0, + 293.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1426.0, + 1405.0, + 1426.0, + 1405.0, + 1461.0, + 291.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1456.0, + 1406.0, + 1456.0, + 1406.0, + 1492.0, + 292.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1484.0, + 1407.0, + 1484.0, + 1407.0, + 1525.0, + 291.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1518.0, + 1405.0, + 1518.0, + 1405.0, + 1554.0, + 295.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1547.0, + 1408.0, + 1547.0, + 1408.0, + 1583.0, + 295.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1578.0, + 1407.0, + 1578.0, + 1407.0, + 1615.0, + 291.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1608.0, + 585.0, + 1608.0, + 585.0, + 1647.0, + 292.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 1608.0, + 728.0, + 1608.0, + 728.0, + 1647.0, + 615.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 1608.0, + 1406.0, + 1608.0, + 1406.0, + 1647.0, + 758.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1639.0, + 1405.0, + 1639.0, + 1405.0, + 1675.0, + 293.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1670.0, + 1407.0, + 1670.0, + 1407.0, + 1705.0, + 293.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1700.0, + 1406.0, + 1700.0, + 1406.0, + 1735.0, + 295.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1729.0, + 1410.0, + 1729.0, + 1410.0, + 1767.0, + 293.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1773.0, + 1402.0, + 1773.0, + 1402.0, + 1810.0, + 296.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1805.0, + 1402.0, + 1805.0, + 1402.0, + 1839.0, + 295.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1837.0, + 1406.0, + 1837.0, + 1406.0, + 1870.0, + 297.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 437.0, + 1030.0, + 437.0, + 1030.0, + 478.0, + 378.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 437.0, + 1331.0, + 437.0, + 1331.0, + 478.0, + 1086.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 464.0, + 1311.0, + 464.0, + 1311.0, + 516.0, + 388.0, + 516.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 766, + 1404, + 766, + 1404, + 1252, + 298, + 1252 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 298, + 1642, + 1405, + 1642, + 1405, + 2006, + 298, + 2006 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1369, + 1403, + 1369, + 1403, + 1521, + 299, + 1521 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 628, + 1404, + 628, + 1404, + 751, + 299, + 751 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1536, + 1405, + 1536, + 1405, + 1628, + 298, + 1628 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 296, + 430, + 1405, + 430, + 1405, + 613, + 296, + 613 + ], + "score": 0.964 + }, + { + "category_id": 3, + "poly": [ + 299, + 189, + 1402, + 189, + 1402, + 416, + 299, + 416 + ], + "score": 0.929 + }, + { + "category_id": 0, + "poly": [ + 300, + 1305, + 617, + 1305, + 617, + 1337, + 300, + 1337 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 841, + 2062, + 858, + 2062, + 858, + 2085, + 841, + 2085 + ], + "score": 0.735 + }, + { + "category_id": 13, + "poly": [ + 537, + 1572, + 556, + 1572, + 556, + 1594, + 537, + 1594 + ], + "score": 0.7, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 605, + 1568, + 620, + 1568, + 620, + 1594, + 605, + 1594 + ], + "score": 0.56, + "latex": "b" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 429.0, + 1406.0, + 429.0, + 1406.0, + 465.0, + 294.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 458.0, + 1405.0, + 458.0, + 1405.0, + 495.0, + 293.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 490.0, + 1405.0, + 490.0, + 1405.0, + 527.0, + 294.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 519.0, + 1406.0, + 519.0, + 1406.0, + 556.0, + 294.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 551.0, + 1405.0, + 551.0, + 1405.0, + 586.0, + 294.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 581.0, + 580.0, + 581.0, + 580.0, + 619.0, + 294.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 223.0, + 1372.0, + 223.0, + 1372.0, + 240.0, + 1253.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 250.0, + 1374.0, + 250.0, + 1374.0, + 267.0, + 1253.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 267.0, + 1135.0, + 267.0, + 1135.0, + 285.0, + 1119.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 276.0, + 1398.0, + 276.0, + 1398.0, + 292.0, + 1250.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 300.0, + 458.0, + 300.0, + 458.0, + 310.0, + 446.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 311.0, + 339.0, + 311.0, + 339.0, + 334.0, + 302.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 311.0, + 1302.0, + 311.0, + 1302.0, + 332.0, + 1231.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 337.0, + 1313.0, + 337.0, + 1313.0, + 357.0, + 1233.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 363.0, + 1310.0, + 363.0, + 1310.0, + 383.0, + 1233.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 378.0, + 340.0, + 378.0, + 340.0, + 400.0, + 300.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 389.0, + 1361.0, + 389.0, + 1361.0, + 410.0, + 1233.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 310.5, + 590.0, + 310.5, + 590.0, + 331.5, + 584.0, + 331.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1301.0, + 622.0, + 1301.0, + 622.0, + 1343.0, + 293.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 862.0, + 2059.0, + 862.0, + 2094.0, + 839.0, + 2094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 767.0, + 1404.0, + 767.0, + 1404.0, + 801.0, + 294.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 797.0, + 1405.0, + 797.0, + 1405.0, + 831.0, + 293.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 827.0, + 1404.0, + 827.0, + 1404.0, + 861.0, + 294.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 854.0, + 1406.0, + 854.0, + 1406.0, + 897.0, + 291.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 887.0, + 1405.0, + 887.0, + 1405.0, + 922.0, + 292.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 919.0, + 1405.0, + 919.0, + 1405.0, + 953.0, + 293.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 949.0, + 1405.0, + 949.0, + 1405.0, + 981.0, + 293.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 978.0, + 1404.0, + 978.0, + 1404.0, + 1013.0, + 294.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1009.0, + 1404.0, + 1009.0, + 1404.0, + 1042.0, + 293.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1040.0, + 1404.0, + 1040.0, + 1404.0, + 1074.0, + 294.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1070.0, + 1405.0, + 1070.0, + 1405.0, + 1103.0, + 293.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1098.0, + 1406.0, + 1098.0, + 1406.0, + 1136.0, + 293.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1130.0, + 1404.0, + 1130.0, + 1404.0, + 1164.0, + 294.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1159.0, + 1402.0, + 1159.0, + 1402.0, + 1193.0, + 294.0, + 1193.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1188.0, + 1406.0, + 1188.0, + 1406.0, + 1228.0, + 292.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1219.0, + 633.0, + 1219.0, + 633.0, + 1257.0, + 293.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1643.0, + 1403.0, + 1643.0, + 1403.0, + 1677.0, + 295.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1672.0, + 1406.0, + 1672.0, + 1406.0, + 1709.0, + 293.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1702.0, + 1405.0, + 1702.0, + 1405.0, + 1739.0, + 293.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1736.0, + 1403.0, + 1736.0, + 1403.0, + 1766.0, + 296.0, + 1766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1764.0, + 1408.0, + 1764.0, + 1408.0, + 1797.0, + 293.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1791.0, + 1408.0, + 1791.0, + 1408.0, + 1828.0, + 293.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1822.0, + 1406.0, + 1822.0, + 1406.0, + 1857.0, + 293.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1854.0, + 1406.0, + 1854.0, + 1406.0, + 1888.0, + 293.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1884.0, + 1405.0, + 1884.0, + 1405.0, + 1918.0, + 295.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1912.0, + 1406.0, + 1912.0, + 1406.0, + 1952.0, + 291.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1944.0, + 1405.0, + 1944.0, + 1405.0, + 1981.0, + 293.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1974.0, + 522.0, + 1974.0, + 522.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1366.0, + 1405.0, + 1366.0, + 1405.0, + 1404.0, + 294.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1398.0, + 1404.0, + 1398.0, + 1404.0, + 1433.0, + 295.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1429.0, + 1405.0, + 1429.0, + 1405.0, + 1466.0, + 294.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1458.0, + 1405.0, + 1458.0, + 1405.0, + 1496.0, + 294.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1490.0, + 734.0, + 1490.0, + 734.0, + 1525.0, + 293.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 630.0, + 1406.0, + 630.0, + 1406.0, + 663.0, + 295.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 659.0, + 1405.0, + 659.0, + 1405.0, + 695.0, + 294.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 691.0, + 1404.0, + 691.0, + 1404.0, + 726.0, + 294.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 721.0, + 1212.0, + 721.0, + 1212.0, + 753.0, + 294.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1532.0, + 1407.0, + 1532.0, + 1407.0, + 1572.0, + 293.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1564.0, + 536.0, + 1564.0, + 536.0, + 1601.0, + 293.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 1564.0, + 604.0, + 1564.0, + 604.0, + 1601.0, + 557.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1564.0, + 1407.0, + 1564.0, + 1407.0, + 1601.0, + 621.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1593.0, + 1405.0, + 1593.0, + 1405.0, + 1634.0, + 292.0, + 1634.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 202, + 1405, + 202, + 1405, + 629, + 297, + 629 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 641, + 1405, + 641, + 1405, + 946, + 298, + 946 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1064, + 1403, + 1064, + 1403, + 1249, + 298, + 1249 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1505, + 1402, + 1505, + 1402, + 1658, + 299, + 1658 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1672, + 1403, + 1672, + 1403, + 1825, + 299, + 1825 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 1263, + 1404, + 1263, + 1404, + 1384, + 298, + 1384 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1399, + 1403, + 1399, + 1403, + 1491, + 299, + 1491 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 300, + 1839, + 1404, + 1839, + 1404, + 1931, + 300, + 1931 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 299, + 1944, + 1398, + 1944, + 1398, + 2007, + 299, + 2007 + ], + "score": 0.943 + }, + { + "category_id": 0, + "poly": [ + 300, + 985, + 617, + 985, + 617, + 1023, + 300, + 1023 + ], + "score": 0.912 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 858, + 2061, + 858, + 2084, + 841, + 2084 + ], + "score": 0.627 + }, + { + "category_id": 2, + "poly": [ + 842, + 2061, + 859, + 2061, + 859, + 2084, + 842, + 2084 + ], + "score": 0.358 + }, + { + "category_id": 15, + "poly": [ + 290.0, + 979.0, + 623.0, + 979.0, + 623.0, + 1031.0, + 290.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 861.0, + 2058.0, + 861.0, + 2091.0, + 838.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2090.0, + 839.0, + 2090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 201.0, + 1406.0, + 201.0, + 1406.0, + 238.0, + 292.0, + 238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 232.0, + 1407.0, + 232.0, + 1407.0, + 269.0, + 292.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 263.0, + 1406.0, + 263.0, + 1406.0, + 299.0, + 294.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 294.0, + 1407.0, + 294.0, + 1407.0, + 329.0, + 292.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 323.0, + 1406.0, + 323.0, + 1406.0, + 360.0, + 292.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 352.0, + 1405.0, + 352.0, + 1405.0, + 390.0, + 294.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 383.0, + 1405.0, + 383.0, + 1405.0, + 418.0, + 294.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 416.0, + 1406.0, + 416.0, + 1406.0, + 451.0, + 294.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 447.0, + 1405.0, + 447.0, + 1405.0, + 479.0, + 296.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 472.0, + 1406.0, + 472.0, + 1406.0, + 514.0, + 291.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 508.0, + 1403.0, + 508.0, + 1403.0, + 540.0, + 295.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 536.0, + 1405.0, + 536.0, + 1405.0, + 571.0, + 294.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 568.0, + 1403.0, + 568.0, + 1403.0, + 599.0, + 296.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 598.0, + 1222.0, + 598.0, + 1222.0, + 633.0, + 294.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 642.0, + 1406.0, + 642.0, + 1406.0, + 677.0, + 295.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 673.0, + 1406.0, + 673.0, + 1406.0, + 707.0, + 294.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 704.0, + 1405.0, + 704.0, + 1405.0, + 738.0, + 293.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 735.0, + 1403.0, + 735.0, + 1403.0, + 767.0, + 296.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 766.0, + 1405.0, + 766.0, + 1405.0, + 797.0, + 296.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 795.0, + 1403.0, + 795.0, + 1403.0, + 829.0, + 293.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 826.0, + 1403.0, + 826.0, + 1403.0, + 858.0, + 296.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 854.0, + 1403.0, + 854.0, + 1403.0, + 890.0, + 293.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 886.0, + 1406.0, + 886.0, + 1406.0, + 919.0, + 292.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 915.0, + 393.0, + 915.0, + 393.0, + 949.0, + 293.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1061.0, + 1407.0, + 1061.0, + 1407.0, + 1103.0, + 292.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1094.0, + 1408.0, + 1094.0, + 1408.0, + 1132.0, + 293.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1126.0, + 1407.0, + 1126.0, + 1407.0, + 1162.0, + 294.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1158.0, + 1405.0, + 1158.0, + 1405.0, + 1190.0, + 296.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1187.0, + 1404.0, + 1187.0, + 1404.0, + 1219.0, + 296.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1219.0, + 743.0, + 1219.0, + 743.0, + 1251.0, + 296.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1504.0, + 1404.0, + 1504.0, + 1404.0, + 1539.0, + 294.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1534.0, + 1406.0, + 1534.0, + 1406.0, + 1571.0, + 294.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1566.0, + 1406.0, + 1566.0, + 1406.0, + 1601.0, + 293.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1593.0, + 1406.0, + 1593.0, + 1406.0, + 1633.0, + 294.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1626.0, + 1318.0, + 1626.0, + 1318.0, + 1663.0, + 294.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1669.0, + 1404.0, + 1669.0, + 1404.0, + 1708.0, + 293.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1704.0, + 1404.0, + 1704.0, + 1404.0, + 1737.0, + 297.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1734.0, + 1405.0, + 1734.0, + 1405.0, + 1767.0, + 294.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1756.0, + 1405.0, + 1756.0, + 1405.0, + 1804.0, + 292.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1795.0, + 840.0, + 1795.0, + 840.0, + 1828.0, + 297.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1261.0, + 1406.0, + 1261.0, + 1406.0, + 1300.0, + 294.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1292.0, + 1409.0, + 1292.0, + 1409.0, + 1330.0, + 294.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1323.0, + 1404.0, + 1323.0, + 1404.0, + 1358.0, + 296.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1354.0, + 735.0, + 1354.0, + 735.0, + 1388.0, + 294.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1396.0, + 1405.0, + 1396.0, + 1405.0, + 1435.0, + 292.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1431.0, + 1405.0, + 1431.0, + 1405.0, + 1464.0, + 295.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1462.0, + 643.0, + 1462.0, + 643.0, + 1495.0, + 294.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1838.0, + 1405.0, + 1838.0, + 1405.0, + 1875.0, + 294.0, + 1875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1868.0, + 1406.0, + 1868.0, + 1406.0, + 1904.0, + 291.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1898.0, + 585.0, + 1898.0, + 585.0, + 1934.0, + 295.0, + 1934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1946.0, + 1401.0, + 1946.0, + 1401.0, + 1978.0, + 296.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1976.0, + 1403.0, + 1976.0, + 1403.0, + 2011.0, + 295.0, + 2011.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 340, + 1405, + 340, + 1405, + 586, + 297, + 586 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 777, + 827, + 777, + 827, + 1201, + 298, + 1201 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 597, + 1406, + 597, + 1406, + 763, + 297, + 763 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1270, + 1405, + 1270, + 1405, + 1426, + 297, + 1426 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 296, + 1679, + 1404, + 1679, + 1404, + 1841, + 296, + 1841 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1853, + 1405, + 1853, + 1405, + 2008, + 297, + 2008 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 297, + 1542, + 1405, + 1542, + 1405, + 1666, + 297, + 1666 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 846, + 1039, + 1406, + 1039, + 1406, + 1254, + 846, + 1254 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 298, + 1435, + 1398, + 1435, + 1398, + 1530, + 298, + 1530 + ], + "score": 0.962 + }, + { + "category_id": 3, + "poly": [ + 984, + 769, + 1265, + 769, + 1265, + 1027, + 984, + 1027 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 297, + 202, + 1402, + 202, + 1402, + 265, + 297, + 265 + ], + "score": 0.932 + }, + { + "category_id": 0, + "poly": [ + 297, + 293, + 1071, + 293, + 1071, + 331, + 297, + 331 + ], + "score": 0.93 + }, + { + "category_id": 0, + "poly": [ + 299, + 1232, + 705, + 1232, + 705, + 1264, + 299, + 1264 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 841, + 2062, + 858, + 2062, + 858, + 2084, + 841, + 2084 + ], + "score": 0.773 + }, + { + "category_id": 13, + "poly": [ + 813, + 598, + 1012, + 598, + 1012, + 633, + 813, + 633 + ], + "score": 0.93, + "latex": "\\{ R _ { 1 } , R _ { 2 } , . . , R _ { N } \\}" + }, + { + "category_id": 13, + "poly": [ + 1090, + 1467, + 1307, + 1467, + 1307, + 1503, + 1090, + 1503 + ], + "score": 0.93, + "latex": "\\{ V _ { 1 } ^ { t } , V _ { 2 } ^ { t } , \\ldots , \\bar { V _ { M } ^ { t } } \\}" + }, + { + "category_id": 13, + "poly": [ + 370, + 630, + 584, + 630, + 584, + 668, + 370, + 668 + ], + "score": 0.93, + "latex": "{ \\pmb R } _ { i } = ( \\vec { R _ { i } } , M L P _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 998, + 492, + 1216, + 492, + 1216, + 527, + 998, + 527 + ], + "score": 0.92, + "latex": "\\{ V _ { 1 } ^ { t } , \\ldots , \\ldots \\stackrel { \\cdot } { , } V _ { M } ^ { t } \\}" + }, + { + "category_id": 13, + "poly": [ + 493, + 1717, + 634, + 1717, + 634, + 1753, + 493, + 1753 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\pmb q _ { p } = \\pmb V _ { p } W ^ { q } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1154, + 1885, + 1299, + 1885, + 1299, + 1919, + 1154, + 1919 + ], + "score": 0.92, + "latex": "q _ { p } = V _ { p } W ^ { q }" + }, + { + "category_id": 13, + "poly": [ + 817, + 1437, + 1040, + 1437, + 1040, + 1470, + 817, + 1470 + ], + "score": 0.92, + "latex": "\\{ R _ { 1 } , R _ { 2 } , \\ldots , R _ { N } \\}" + }, + { + "category_id": 13, + "poly": [ + 299, + 1466, + 498, + 1466, + 498, + 1502, + 299, + 1502 + ], + "score": 0.92, + "latex": "\\{ \\pmb { x } ^ { 1 } , \\pmb { x } ^ { 2 } , \\dots , \\pmb { x } ^ { T } \\}" + }, + { + "category_id": 13, + "poly": [ + 470, + 1779, + 758, + 1779, + 758, + 1813, + 470, + 1813 + ], + "score": 0.92, + "latex": "r = \\mathrm { a r g m a x } _ { i } ( q _ { p } k _ { i } + \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 815, + 701, + 894, + 701, + 894, + 732, + 815, + 732 + ], + "score": 0.91, + "latex": "M L P _ { i }" + }, + { + "category_id": 13, + "poly": [ + 993, + 666, + 1070, + 666, + 1070, + 696, + 993, + 696 + ], + "score": 0.91, + "latex": "M L P _ { i }" + }, + { + "category_id": 13, + "poly": [ + 850, + 1779, + 1067, + 1779, + 1067, + 1811, + 850, + 1811 + ], + "score": 0.91, + "latex": "\\gamma \\sim \\mathrm { G u m b e l } ( 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 671, + 630, + 707, + 630, + 707, + 666, + 671, + 666 + ], + "score": 0.91, + "latex": "\\vec { R _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 918, + 1363, + 957, + 1363, + 957, + 1393, + 918, + 1393 + ], + "score": 0.91, + "latex": "\\scriptstyle { R _ { r } }" + }, + { + "category_id": 13, + "poly": [ + 533, + 696, + 569, + 696, + 569, + 732, + 533, + 732 + ], + "score": 0.9, + "latex": "\\vec { R _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 727, + 431, + 973, + 431, + 973, + 464, + 727, + 464 + ], + "score": 0.9, + "latex": "\\{ \\pmb { x } ^ { 1 } , \\dotsc , \\pmb { x } ^ { t } , \\dotsc , \\pmb { x } ^ { T } \\}" + }, + { + "category_id": 13, + "poly": [ + 557, + 1915, + 927, + 1915, + 927, + 1950, + 557, + 1950 + ], + "score": 0.9, + "latex": "\\pmb { k } _ { j } = V _ { j } W ^ { q } \\quad \\forall j \\in \\{ 1 , \\dots , M \\}" + }, + { + "category_id": 13, + "poly": [ + 1180, + 1946, + 1397, + 1946, + 1397, + 1978, + 1180, + 1978 + ], + "score": 0.9, + "latex": "\\gamma \\sim \\mathrm { G u m b e l } ( 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 882, + 1682, + 920, + 1682, + 920, + 1711, + 882, + 1711 + ], + "score": 0.89, + "latex": "\\scriptstyle { R _ { r } }" + }, + { + "category_id": 13, + "poly": [ + 1226, + 1332, + 1264, + 1332, + 1264, + 1362, + 1226, + 1362 + ], + "score": 0.89, + "latex": "\\scriptstyle { \\mathbf { } } R _ { r }" + }, + { + "category_id": 13, + "poly": [ + 363, + 552, + 428, + 552, + 428, + 582, + 363, + 582 + ], + "score": 0.89, + "latex": "V ^ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 626, + 1681, + 659, + 1681, + 659, + 1715, + 626, + 1715 + ], + "score": 0.89, + "latex": "V _ { p }" + }, + { + "category_id": 13, + "poly": [ + 796, + 1332, + 830, + 1332, + 830, + 1365, + 796, + 1365 + ], + "score": 0.89, + "latex": "V _ { p }" + }, + { + "category_id": 13, + "poly": [ + 568, + 1393, + 601, + 1393, + 601, + 1427, + 568, + 1427 + ], + "score": 0.88, + "latex": "V _ { p }" + }, + { + "category_id": 13, + "poly": [ + 740, + 1363, + 774, + 1363, + 774, + 1393, + 740, + 1393 + ], + "score": 0.88, + "latex": "V _ { c }" + }, + { + "category_id": 13, + "poly": [ + 662, + 1468, + 693, + 1468, + 693, + 1496, + 662, + 1496 + ], + "score": 0.87, + "latex": "\\mathbf { \\boldsymbol { x } } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1856, + 1132, + 1856, + 1132, + 1885, + 1100, + 1885 + ], + "score": 0.87, + "latex": "V _ { c }" + }, + { + "category_id": 13, + "poly": [ + 475, + 492, + 505, + 492, + 505, + 521, + 475, + 521 + ], + "score": 0.87, + "latex": "\\mathbf { \\boldsymbol { x } } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 645, + 1855, + 678, + 1855, + 678, + 1888, + 645, + 1888 + ], + "score": 0.87, + "latex": "V _ { p }" + }, + { + "category_id": 13, + "poly": [ + 799, + 1946, + 1088, + 1946, + 1088, + 1981, + 799, + 1981 + ], + "score": 0.87, + "latex": "c = \\mathrm { a r g m a x } _ { j } ( q _ { p } k _ { j } + \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 869, + 1856, + 906, + 1856, + 906, + 1885, + 869, + 1885 + ], + "score": 0.86, + "latex": "\\scriptstyle { R _ { r } }" + }, + { + "category_id": 13, + "poly": [ + 677, + 553, + 707, + 553, + 707, + 581, + 677, + 581 + ], + "score": 0.86, + "latex": "\\mathbf { \\boldsymbol { x } } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 981, + 1543, + 1011, + 1543, + 1011, + 1572, + 981, + 1572 + ], + "score": 0.86, + "latex": "\\mathbf { \\boldsymbol { x } } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 1037, + 1712, + 1399, + 1712, + 1399, + 1751, + 1037, + 1751 + ], + "score": 0.85, + "latex": "\\pmb { k } _ { i } = \\vec { R _ { i } } W ^ { k } \\quad \\forall i \\in \\{ 1 , \\dots , N \\}" + }, + { + "category_id": 13, + "poly": [ + 712, + 493, + 745, + 493, + 745, + 521, + 712, + 521 + ], + "score": 0.83, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 488, + 599, + 516, + 599, + 516, + 626, + 488, + 626 + ], + "score": 0.83, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 339, + 524, + 372, + 524, + 372, + 550, + 339, + 550 + ], + "score": 0.82, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 1213, + 1439, + 1239, + 1439, + 1239, + 1465, + 1213, + 1465 + ], + "score": 0.81, + "latex": "_ { \\mathbf { T } }" + }, + { + "category_id": 13, + "poly": [ + 705, + 1438, + 738, + 1438, + 738, + 1466, + 705, + 1466 + ], + "score": 0.79, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 984, + 1471, + 1022, + 1471, + 1022, + 1497, + 984, + 1497 + ], + "score": 0.76, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 801, + 1501, + 816, + 1501, + 816, + 1526, + 801, + 1526 + ], + "score": 0.69, + "latex": "\\pmb { t }" + }, + { + "category_id": 13, + "poly": [ + 702, + 434, + 715, + 434, + 715, + 459, + 702, + 459 + ], + "score": 0.58, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1037, + 1711, + 1178, + 1711, + 1178, + 1749, + 1037, + 1749 + ], + "score": 0.31, + "latex": "\\pmb { k } _ { i } = \\vec { R _ { i } } W ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 351, + 1545, + 372, + 1545, + 372, + 1572, + 351, + 1572 + ], + "score": 0.29, + "latex": "\\imath ." + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1037.0, + 1409.0, + 1037.0, + 1409.0, + 1075.0, + 844.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1068.0, + 1409.0, + 1068.0, + 1409.0, + 1104.0, + 845.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1101.0, + 1405.0, + 1101.0, + 1405.0, + 1135.0, + 844.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1132.0, + 1408.0, + 1132.0, + 1408.0, + 1163.0, + 844.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1160.0, + 1405.0, + 1160.0, + 1405.0, + 1195.0, + 844.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1189.0, + 1403.0, + 1189.0, + 1403.0, + 1225.0, + 843.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1220.0, + 1033.0, + 1220.0, + 1033.0, + 1258.0, + 844.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 770.0, + 1234.0, + 770.0, + 1234.0, + 790.0, + 1193.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 784.0, + 1016.0, + 784.0, + 1016.0, + 817.0, + 994.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 790.0, + 1252.0, + 790.0, + 1252.0, + 812.0, + 1210.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 805.0, + 1261.0, + 805.0, + 1261.0, + 827.0, + 1220.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 837.0, + 1016.0, + 837.0, + 1016.0, + 869.0, + 994.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 890.0, + 1017.0, + 890.0, + 1017.0, + 921.0, + 996.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 946.0, + 1014.0, + 946.0, + 1014.0, + 973.0, + 997.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 1000.0, + 1110.0, + 1000.0, + 1110.0, + 1021.0, + 1078.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 998.0, + 1215.0, + 998.0, + 1215.0, + 1019.0, + 1185.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 1000.0, + 1054.0, + 1000.0, + 1054.0, + 1018.5, + 1027.0, + 1018.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 999.0, + 1161.0, + 999.0, + 1161.0, + 1018.0, + 1133.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 287.0, + 1073.0, + 287.0, + 1073.0, + 340.0, + 290.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1228.0, + 707.0, + 1228.0, + 707.0, + 1269.0, + 291.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2059.0, + 862.0, + 2059.0, + 862.0, + 2091.0, + 838.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 338.0, + 1406.0, + 338.0, + 1406.0, + 377.0, + 292.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 372.0, + 1405.0, + 372.0, + 1405.0, + 405.0, + 295.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 402.0, + 1405.0, + 402.0, + 1405.0, + 435.0, + 295.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 428.0, + 701.0, + 428.0, + 701.0, + 469.0, + 292.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 428.0, + 726.0, + 428.0, + 726.0, + 469.0, + 716.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 428.0, + 1408.0, + 428.0, + 1408.0, + 469.0, + 974.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 462.0, + 1407.0, + 462.0, + 1407.0, + 497.0, + 292.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 483.0, + 474.0, + 483.0, + 474.0, + 535.0, + 289.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 483.0, + 711.0, + 483.0, + 711.0, + 535.0, + 506.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 483.0, + 997.0, + 483.0, + 997.0, + 535.0, + 746.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 483.0, + 1411.0, + 483.0, + 1411.0, + 535.0, + 1217.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 519.0, + 338.0, + 519.0, + 338.0, + 559.0, + 292.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 519.0, + 1406.0, + 519.0, + 1406.0, + 559.0, + 373.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 550.0, + 362.0, + 550.0, + 362.0, + 588.0, + 292.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 550.0, + 676.0, + 550.0, + 676.0, + 588.0, + 429.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 550.0, + 719.0, + 550.0, + 719.0, + 588.0, + 708.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 775.0, + 830.0, + 775.0, + 830.0, + 810.0, + 295.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 807.0, + 826.0, + 807.0, + 826.0, + 841.0, + 295.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 838.0, + 826.0, + 838.0, + 826.0, + 869.0, + 294.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 866.0, + 829.0, + 866.0, + 829.0, + 901.0, + 295.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 898.0, + 829.0, + 898.0, + 829.0, + 930.0, + 296.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 929.0, + 827.0, + 929.0, + 827.0, + 960.0, + 295.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 958.0, + 828.0, + 958.0, + 828.0, + 991.0, + 294.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 989.0, + 828.0, + 989.0, + 828.0, + 1022.0, + 294.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1019.0, + 830.0, + 1019.0, + 830.0, + 1052.0, + 294.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1048.0, + 827.0, + 1048.0, + 827.0, + 1083.0, + 295.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1080.0, + 827.0, + 1080.0, + 827.0, + 1112.0, + 295.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1110.0, + 826.0, + 1110.0, + 826.0, + 1143.0, + 296.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1138.0, + 827.0, + 1138.0, + 827.0, + 1173.0, + 294.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1168.0, + 586.0, + 1168.0, + 586.0, + 1203.0, + 295.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 596.0, + 487.0, + 596.0, + 487.0, + 635.0, + 294.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 596.0, + 812.0, + 596.0, + 812.0, + 635.0, + 517.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 596.0, + 1407.0, + 596.0, + 1407.0, + 635.0, + 1013.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 631.0, + 369.0, + 631.0, + 369.0, + 671.0, + 292.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 631.0, + 670.0, + 631.0, + 670.0, + 671.0, + 585.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 631.0, + 1407.0, + 631.0, + 1407.0, + 671.0, + 708.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 666.0, + 992.0, + 666.0, + 992.0, + 699.0, + 294.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 666.0, + 1404.0, + 666.0, + 1404.0, + 699.0, + 1071.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 701.0, + 532.0, + 701.0, + 532.0, + 735.0, + 295.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 701.0, + 814.0, + 701.0, + 814.0, + 735.0, + 570.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 701.0, + 1406.0, + 701.0, + 1406.0, + 735.0, + 895.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 732.0, + 1107.0, + 732.0, + 1107.0, + 766.0, + 296.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1272.0, + 1403.0, + 1272.0, + 1403.0, + 1305.0, + 295.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1301.0, + 1406.0, + 1301.0, + 1406.0, + 1335.0, + 295.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1332.0, + 795.0, + 1332.0, + 795.0, + 1365.0, + 295.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 1332.0, + 1225.0, + 1332.0, + 1225.0, + 1365.0, + 831.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 1332.0, + 1405.0, + 1332.0, + 1405.0, + 1365.0, + 1265.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1363.0, + 739.0, + 1363.0, + 739.0, + 1396.0, + 295.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 1363.0, + 917.0, + 1363.0, + 917.0, + 1396.0, + 775.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 1363.0, + 1403.0, + 1363.0, + 1403.0, + 1396.0, + 958.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1392.0, + 567.0, + 1392.0, + 567.0, + 1431.0, + 294.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1392.0, + 613.0, + 1392.0, + 613.0, + 1431.0, + 602.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1677.0, + 625.0, + 1677.0, + 625.0, + 1716.0, + 293.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1677.0, + 881.0, + 1677.0, + 881.0, + 1716.0, + 660.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1677.0, + 1405.0, + 1677.0, + 1405.0, + 1716.0, + 921.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1710.0, + 492.0, + 1710.0, + 492.0, + 1756.0, + 291.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1710.0, + 1036.0, + 1710.0, + 1036.0, + 1756.0, + 635.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1400.0, + 1710.0, + 1410.0, + 1710.0, + 1410.0, + 1756.0, + 1400.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1746.0, + 1405.0, + 1746.0, + 1405.0, + 1783.0, + 294.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1775.0, + 469.0, + 1775.0, + 469.0, + 1816.0, + 293.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 1775.0, + 849.0, + 1775.0, + 849.0, + 1816.0, + 759.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1775.0, + 1404.0, + 1775.0, + 1404.0, + 1816.0, + 1068.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1808.0, + 1148.0, + 1808.0, + 1148.0, + 1842.0, + 295.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1855.0, + 644.0, + 1855.0, + 644.0, + 1888.0, + 295.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 1855.0, + 868.0, + 1855.0, + 868.0, + 1888.0, + 679.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 1855.0, + 1099.0, + 1855.0, + 1099.0, + 1888.0, + 907.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1855.0, + 1406.0, + 1855.0, + 1406.0, + 1888.0, + 1133.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1884.0, + 1153.0, + 1884.0, + 1153.0, + 1921.0, + 294.0, + 1921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1300.0, + 1884.0, + 1406.0, + 1884.0, + 1406.0, + 1921.0, + 1300.0, + 1921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1913.0, + 556.0, + 1913.0, + 556.0, + 1950.0, + 295.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1913.0, + 1408.0, + 1913.0, + 1408.0, + 1950.0, + 928.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 798.0, + 1942.0, + 798.0, + 1984.0, + 294.0, + 1984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 1942.0, + 1179.0, + 1942.0, + 1179.0, + 1984.0, + 1089.0, + 1984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1942.0, + 1410.0, + 1942.0, + 1410.0, + 1984.0, + 1398.0, + 1984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1970.0, + 1406.0, + 1970.0, + 1406.0, + 2014.0, + 292.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1542.0, + 350.0, + 1542.0, + 350.0, + 1578.0, + 294.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1542.0, + 980.0, + 1542.0, + 980.0, + 1578.0, + 373.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1542.0, + 1406.0, + 1542.0, + 1406.0, + 1578.0, + 1012.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1575.0, + 1405.0, + 1575.0, + 1405.0, + 1607.0, + 294.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1604.0, + 1410.0, + 1604.0, + 1410.0, + 1640.0, + 292.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1633.0, + 1286.0, + 1633.0, + 1286.0, + 1669.0, + 294.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1434.0, + 704.0, + 1434.0, + 704.0, + 1472.0, + 293.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 1434.0, + 816.0, + 1434.0, + 816.0, + 1472.0, + 739.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1434.0, + 1212.0, + 1434.0, + 1212.0, + 1472.0, + 1041.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 1434.0, + 1404.0, + 1434.0, + 1404.0, + 1472.0, + 1240.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1458.0, + 298.0, + 1458.0, + 298.0, + 1509.0, + 293.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 1458.0, + 661.0, + 1458.0, + 661.0, + 1509.0, + 499.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 1458.0, + 983.0, + 1458.0, + 983.0, + 1509.0, + 694.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 1458.0, + 1089.0, + 1458.0, + 1089.0, + 1509.0, + 1023.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 1458.0, + 1407.0, + 1458.0, + 1407.0, + 1509.0, + 1308.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1495.0, + 800.0, + 1495.0, + 800.0, + 1535.0, + 293.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1495.0, + 983.0, + 1495.0, + 983.0, + 1535.0, + 817.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 200.0, + 1404.0, + 200.0, + 1404.0, + 239.0, + 293.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 227.0, + 368.0, + 227.0, + 368.0, + 270.0, + 291.0, + 270.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 455, + 1406, + 455, + 1406, + 667, + 296, + 667 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1354, + 941, + 1354, + 941, + 1747, + 299, + 1747 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1031, + 1405, + 1031, + 1405, + 1191, + 298, + 1191 + ], + "score": 0.971 + }, + { + "category_id": 5, + "poly": [ + 995, + 1592, + 1365, + 1592, + 1365, + 1710, + 995, + 1710 + ], + "score": 0.97, + "html": "
Rule 1Rule 2Rule3Rule 4
TranslateDownTranslate UpRotate RightRotate Left5039000
00495000
050300
0004981
" + }, + { + "category_id": 1, + "poly": [ + 297, + 279, + 1404, + 279, + 1404, + 402, + 297, + 402 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 863, + 688, + 1386, + 688, + 1386, + 939, + 863, + 939 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 298, + 1915, + 1405, + 1915, + 1405, + 2008, + 298, + 2008 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 298, + 1749, + 1404, + 1749, + 1404, + 1901, + 298, + 1901 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 297, + 878, + 827, + 878, + 827, + 1031, + 297, + 1031 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 964, + 1359, + 1405, + 1359, + 1405, + 1572, + 964, + 1572 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 297, + 680, + 826, + 680, + 826, + 864, + 297, + 864 + ], + "score": 0.961 + }, + { + "category_id": 4, + "poly": [ + 845, + 958, + 1402, + 958, + 1402, + 1022, + 845, + 1022 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 297, + 1203, + 1402, + 1203, + 1402, + 1267, + 297, + 1267 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 297, + 202, + 1402, + 202, + 1402, + 264, + 297, + 264 + ], + "score": 0.937 + }, + { + "category_id": 0, + "poly": [ + 297, + 419, + 1045, + 419, + 1045, + 451, + 297, + 451 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 298, + 1289, + 532, + 1289, + 532, + 1327, + 298, + 1327 + ], + "score": 0.909 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 858, + 2061, + 858, + 2085, + 841, + 2085 + ], + "score": 0.709 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.095 + }, + { + "category_id": 13, + "poly": [ + 298, + 1063, + 438, + 1063, + 438, + 1097, + 298, + 1097 + ], + "score": 0.91, + "latex": "\\{ 1 , \\ldots , M \\}" + }, + { + "category_id": 13, + "poly": [ + 1114, + 311, + 1195, + 311, + 1195, + 339, + 1114, + 339 + ], + "score": 0.9, + "latex": "M L P _ { r }" + }, + { + "category_id": 13, + "poly": [ + 1053, + 371, + 1086, + 371, + 1086, + 405, + 1053, + 405 + ], + "score": 0.89, + "latex": "V _ { p }" + }, + { + "category_id": 13, + "poly": [ + 923, + 1095, + 1259, + 1095, + 1259, + 1130, + 923, + 1130 + ], + "score": 0.88, + "latex": "p , r = \\mathrm { a r g m a x } _ { i , j } ( q _ { p } k _ { i } + \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 1125, + 341, + 1159, + 341, + 1159, + 374, + 1125, + 374 + ], + "score": 0.88, + "latex": "V _ { p }" + }, + { + "category_id": 13, + "poly": [ + 773, + 281, + 812, + 281, + 812, + 310, + 773, + 310 + ], + "score": 0.88, + "latex": "\\scriptstyle { R _ { r } }" + }, + { + "category_id": 13, + "poly": [ + 1210, + 340, + 1242, + 340, + 1242, + 370, + 1210, + 370 + ], + "score": 0.88, + "latex": "V _ { c }" + }, + { + "category_id": 13, + "poly": [ + 1348, + 1098, + 1403, + 1098, + 1403, + 1126, + 1348, + 1126 + ], + "score": 0.86, + "latex": "\\gamma \\sim" + }, + { + "category_id": 13, + "poly": [ + 1132, + 281, + 1165, + 281, + 1165, + 311, + 1132, + 311 + ], + "score": 0.84, + "latex": "V _ { p }" + }, + { + "category_id": 13, + "poly": [ + 791, + 1061, + 1153, + 1061, + 1153, + 1096, + 791, + 1096 + ], + "score": 0.81, + "latex": "\\pmb { k } _ { i } = R _ { i } W ^ { k } \\quad \\forall i \\in \\{ 1 , \\dots , N \\}" + }, + { + "category_id": 13, + "poly": [ + 793, + 1160, + 823, + 1160, + 823, + 1186, + 793, + 1186 + ], + "score": 0.8, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1178, + 1029, + 1403, + 1029, + 1403, + 1064, + 1178, + 1064 + ], + "score": 0.77, + "latex": "\\pmb { q } _ { j } = \\pmb { V } _ { j } \\pmb { W } ^ { q } \\quad \\forall j \\in" + }, + { + "category_id": 13, + "poly": [ + 673, + 773, + 702, + 773, + 702, + 799, + 673, + 799 + ], + "score": 0.74, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1287, + 1162, + 1300, + 1162, + 1300, + 1186, + 1287, + 1186 + ], + "score": 0.69, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 735, + 681, + 768, + 681, + 768, + 708, + 735, + 708 + ], + "score": 0.56, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 395, + 1127, + 456, + 1127, + 456, + 1161, + 395, + 1161 + ], + "score": 0.54, + "latex": "( 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 542, + 1030, + 609, + 1030, + 609, + 1060, + 542, + 1060 + ], + "score": 0.5, + "latex": "_ { N M }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1126, + 458, + 1126, + 458, + 1161, + 298, + 1161 + ], + "score": 0.42, + "latex": "\\mathrm { G u m b e l } ( 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 1338, + 1030, + 1404, + 1030, + 1404, + 1063, + 1338, + 1063 + ], + "score": 0.28, + "latex": "\\forall j \\in" + }, + { + "category_id": 13, + "poly": [ + 792, + 1061, + 932, + 1061, + 932, + 1094, + 792, + 1094 + ], + "score": 0.25, + "latex": "\\pmb { k } _ { i } = R _ { i } W ^ { k }" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 689.0, + 1137.0, + 689.0, + 1137.0, + 709.0, + 1097.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 712.0, + 1308.0, + 712.0, + 1308.0, + 735.0, + 1217.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 732.0, + 1326.0, + 732.0, + 1326.0, + 755.0, + 1216.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 756.0, + 1087.0, + 756.0, + 1087.0, + 775.0, + 1014.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1288.0, + 755.0, + 1343.0, + 755.0, + 1343.0, + 775.0, + 1288.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 788.0, + 1159.0, + 788.0, + 1159.0, + 795.0, + 1142.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 795.0, + 1004.0, + 795.0, + 1004.0, + 807.0, + 990.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 795.0, + 1074.0, + 795.0, + 1074.0, + 807.0, + 1061.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 797.0, + 1144.0, + 797.0, + 1144.0, + 806.0, + 1135.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 837.0, + 1002.0, + 837.0, + 1002.0, + 845.0, + 991.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 838.0, + 1144.0, + 838.0, + 1144.0, + 845.0, + 1135.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 853.0, + 901.0, + 853.0, + 901.0, + 870.0, + 864.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 853.0, + 1250.0, + 853.0, + 1250.0, + 870.0, + 1211.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 876.0, + 1074.0, + 876.0, + 1074.0, + 887.0, + 1059.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 878.0, + 1144.0, + 878.0, + 1144.0, + 886.0, + 1136.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 918.0, + 1002.0, + 918.0, + 1002.0, + 927.0, + 991.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 919.0, + 1073.0, + 919.0, + 1073.0, + 927.0, + 1064.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 955.0, + 1403.0, + 955.0, + 1403.0, + 992.0, + 845.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 989.0, + 1169.0, + 989.0, + 1169.0, + 1022.0, + 847.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 417.0, + 1044.0, + 417.0, + 1044.0, + 457.0, + 295.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1284.0, + 537.0, + 1284.0, + 537.0, + 1335.0, + 289.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2058.0, + 861.0, + 2058.0, + 861.0, + 2093.0, + 839.0, + 2093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2092.0, + 838.0, + 2092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 455.0, + 1403.0, + 455.0, + 1403.0, + 485.0, + 295.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 484.0, + 1406.0, + 484.0, + 1406.0, + 518.0, + 294.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 511.0, + 1406.0, + 511.0, + 1406.0, + 551.0, + 291.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 543.0, + 1406.0, + 543.0, + 1406.0, + 581.0, + 294.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 573.0, + 1407.0, + 573.0, + 1407.0, + 610.0, + 293.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 604.0, + 1406.0, + 604.0, + 1406.0, + 640.0, + 294.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 634.0, + 627.0, + 634.0, + 627.0, + 672.0, + 295.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1354.0, + 943.0, + 1354.0, + 943.0, + 1386.0, + 297.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1385.0, + 943.0, + 1385.0, + 943.0, + 1417.0, + 296.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1411.0, + 944.0, + 1411.0, + 944.0, + 1451.0, + 295.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1445.0, + 944.0, + 1445.0, + 944.0, + 1480.0, + 295.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1477.0, + 944.0, + 1477.0, + 944.0, + 1508.0, + 296.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1503.0, + 944.0, + 1503.0, + 944.0, + 1540.0, + 293.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1536.0, + 944.0, + 1536.0, + 944.0, + 1571.0, + 294.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1568.0, + 945.0, + 1568.0, + 945.0, + 1600.0, + 294.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1598.0, + 945.0, + 1598.0, + 945.0, + 1630.0, + 296.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1625.0, + 944.0, + 1625.0, + 944.0, + 1662.0, + 293.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1656.0, + 944.0, + 1656.0, + 944.0, + 1691.0, + 294.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1689.0, + 944.0, + 1689.0, + 944.0, + 1721.0, + 296.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1718.0, + 944.0, + 1718.0, + 944.0, + 1750.0, + 295.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1028.0, + 541.0, + 1028.0, + 541.0, + 1066.0, + 295.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1028.0, + 1177.0, + 1028.0, + 1177.0, + 1066.0, + 610.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1057.0, + 297.0, + 1057.0, + 297.0, + 1103.0, + 293.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 1057.0, + 790.0, + 1057.0, + 790.0, + 1103.0, + 439.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 1057.0, + 1406.0, + 1057.0, + 1406.0, + 1103.0, + 1154.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1089.0, + 922.0, + 1089.0, + 922.0, + 1134.0, + 292.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 1089.0, + 1347.0, + 1089.0, + 1347.0, + 1134.0, + 1260.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1122.0, + 297.0, + 1122.0, + 297.0, + 1165.0, + 292.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 1122.0, + 1408.0, + 1122.0, + 1408.0, + 1165.0, + 459.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1158.0, + 792.0, + 1158.0, + 792.0, + 1196.0, + 295.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 1158.0, + 1286.0, + 1158.0, + 1286.0, + 1196.0, + 824.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1301.0, + 1158.0, + 1312.0, + 1158.0, + 1312.0, + 1196.0, + 1301.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 279.0, + 772.0, + 279.0, + 772.0, + 315.0, + 294.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 279.0, + 1131.0, + 279.0, + 1131.0, + 315.0, + 813.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 279.0, + 1406.0, + 279.0, + 1406.0, + 315.0, + 1166.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 309.0, + 1113.0, + 309.0, + 1113.0, + 345.0, + 294.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 309.0, + 1405.0, + 309.0, + 1405.0, + 345.0, + 1196.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 339.0, + 1124.0, + 339.0, + 1124.0, + 375.0, + 294.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 339.0, + 1209.0, + 339.0, + 1209.0, + 375.0, + 1160.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 339.0, + 1405.0, + 339.0, + 1405.0, + 375.0, + 1243.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 370.0, + 1052.0, + 370.0, + 1052.0, + 406.0, + 294.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 370.0, + 1322.0, + 370.0, + 1322.0, + 406.0, + 1087.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1914.0, + 1407.0, + 1914.0, + 1407.0, + 1948.0, + 296.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1946.0, + 1408.0, + 1946.0, + 1408.0, + 1980.0, + 295.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1975.0, + 1406.0, + 1975.0, + 1406.0, + 2012.0, + 295.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1745.0, + 1405.0, + 1745.0, + 1405.0, + 1784.0, + 293.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1779.0, + 1405.0, + 1779.0, + 1405.0, + 1812.0, + 296.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1808.0, + 1405.0, + 1808.0, + 1405.0, + 1841.0, + 293.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1837.0, + 1406.0, + 1837.0, + 1406.0, + 1875.0, + 292.0, + 1875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1865.0, + 398.0, + 1865.0, + 398.0, + 1908.0, + 291.0, + 1908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 874.0, + 831.0, + 874.0, + 831.0, + 914.0, + 295.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 910.0, + 829.0, + 910.0, + 829.0, + 941.0, + 297.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 939.0, + 828.0, + 939.0, + 828.0, + 970.0, + 294.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 968.0, + 827.0, + 968.0, + 827.0, + 1002.0, + 294.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 998.0, + 828.0, + 998.0, + 828.0, + 1037.0, + 294.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 1355.0, + 1407.0, + 1355.0, + 1407.0, + 1394.0, + 960.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1389.0, + 1407.0, + 1389.0, + 1407.0, + 1422.0, + 963.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1420.0, + 1405.0, + 1420.0, + 1405.0, + 1452.0, + 962.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1450.0, + 1405.0, + 1450.0, + 1405.0, + 1483.0, + 962.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1480.0, + 1405.0, + 1480.0, + 1405.0, + 1513.0, + 962.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 1509.0, + 1404.0, + 1509.0, + 1404.0, + 1545.0, + 961.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1544.0, + 1345.0, + 1544.0, + 1345.0, + 1575.0, + 962.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 680.0, + 734.0, + 680.0, + 734.0, + 711.0, + 297.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 680.0, + 826.0, + 680.0, + 826.0, + 711.0, + 769.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 711.0, + 827.0, + 711.0, + 827.0, + 744.0, + 296.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 741.0, + 826.0, + 741.0, + 826.0, + 772.0, + 295.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 770.0, + 672.0, + 770.0, + 672.0, + 804.0, + 295.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 770.0, + 828.0, + 770.0, + 828.0, + 804.0, + 703.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 801.0, + 827.0, + 801.0, + 827.0, + 836.0, + 294.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 833.0, + 717.0, + 833.0, + 717.0, + 867.0, + 294.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1200.0, + 1404.0, + 1200.0, + 1404.0, + 1239.0, + 293.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1234.0, + 1098.0, + 1234.0, + 1098.0, + 1270.0, + 295.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 198.0, + 1406.0, + 198.0, + 1406.0, + 241.0, + 290.0, + 241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 230.0, + 354.0, + 230.0, + 354.0, + 267.0, + 291.0, + 267.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 504, + 1404, + 504, + 1404, + 749, + 297, + 749 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 203, + 1405, + 203, + 1405, + 417, + 297, + 417 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1386, + 826, + 1386, + 826, + 1930, + 299, + 1930 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1096, + 828, + 1096, + 828, + 1369, + 298, + 1369 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 763, + 1405, + 763, + 1405, + 915, + 298, + 915 + ], + "score": 0.976 + }, + { + "category_id": 4, + "poly": [ + 847, + 1626, + 1404, + 1626, + 1404, + 1898, + 847, + 1898 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 297, + 929, + 1407, + 929, + 1407, + 1052, + 297, + 1052 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 858, + 1067, + 1390, + 1067, + 1390, + 1598, + 858, + 1598 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 296, + 1945, + 1405, + 1945, + 1405, + 2007, + 296, + 2007 + ], + "score": 0.941 + }, + { + "category_id": 0, + "poly": [ + 297, + 446, + 968, + 446, + 968, + 480, + 297, + 480 + ], + "score": 0.927 + }, + { + "category_id": 2, + "poly": [ + 840, + 2062, + 858, + 2062, + 858, + 2085, + 840, + 2085 + ], + "score": 0.795 + }, + { + "category_id": 1, + "poly": [ + 300, + 1901, + 1000, + 1901, + 1000, + 1930, + 300, + 1930 + ], + "score": 0.416 + }, + { + "category_id": 13, + "poly": [ + 566, + 1157, + 665, + 1157, + 665, + 1190, + 566, + 1190 + ], + "score": 0.93, + "latex": "( X _ { 0 } , Y _ { 0 } )" + }, + { + "category_id": 13, + "poly": [ + 720, + 1157, + 818, + 1157, + 818, + 1190, + 720, + 1190 + ], + "score": 0.92, + "latex": "( X _ { 1 } , Y _ { 1 } )" + }, + { + "category_id": 13, + "poly": [ + 727, + 1309, + 825, + 1309, + 825, + 1341, + 727, + 1341 + ], + "score": 0.91, + "latex": "( X _ { r } , Y _ { r } )" + }, + { + "category_id": 13, + "poly": [ + 364, + 1414, + 625, + 1414, + 625, + 1448, + 364, + 1448 + ], + "score": 0.9, + "latex": "X = [ ( x _ { i } , y _ { i } ) , ( x _ { j } , \\bar { y _ { j } } ) ]" + }, + { + "category_id": 13, + "poly": [ + 503, + 1445, + 768, + 1445, + 768, + 1480, + 503, + 1480 + ], + "score": 0.88, + "latex": "Y = [ ( \\hat { x _ { i } } , \\hat { y _ { i } } ) , ( \\hat { x _ { j } } , \\hat { y _ { j } } ) ] ." + }, + { + "category_id": 13, + "poly": [ + 360, + 1311, + 642, + 1311, + 642, + 1342, + 360, + 1342 + ], + "score": 0.87, + "latex": "X _ { r } , Y _ { r } = ( X _ { 0 } , Y _ { 0 } - Y _ { 1 } ) \\}" + }, + { + "category_id": 13, + "poly": [ + 385, + 1281, + 683, + 1281, + 683, + 1310, + 385, + 1310 + ], + "score": 0.86, + "latex": "( X _ { r } , Y _ { r } ) = ( X _ { 0 } , Y _ { 0 } + Y _ { 1 } )" + }, + { + "category_id": 13, + "poly": [ + 393, + 1217, + 718, + 1217, + 718, + 1249, + 393, + 1249 + ], + "score": 0.85, + "latex": "\\bar { ( X _ { r } , Y _ { r } ) } \\ = \\ ( X _ { 0 } \\bar { + } X _ { 1 } , Y _ { 0 } )" + }, + { + "category_id": 13, + "poly": [ + 408, + 1250, + 730, + 1250, + 730, + 1279, + 408, + 1279 + ], + "score": 0.8, + "latex": "( X _ { r } , Y _ { r } ) \\ : = \\ : ( X _ { 0 } - X _ { 1 } , Y _ { 0 } )" + }, + { + "category_id": 13, + "poly": [ + 465, + 1659, + 494, + 1659, + 494, + 1685, + 465, + 1685 + ], + "score": 0.79, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 655, + 1598, + 683, + 1598, + 683, + 1624, + 655, + 1624 + ], + "score": 0.78, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1156, + 1750, + 1170, + 1750, + 1170, + 1775, + 1156, + 1775 + ], + "score": 0.71, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 877, + 824, + 924, + 824, + 924, + 854, + 877, + 854 + ], + "score": 0.68, + "latex": "( X )" + }, + { + "category_id": 13, + "poly": [ + 1183, + 826, + 1219, + 826, + 1219, + 855, + 1183, + 855 + ], + "score": 0.61, + "latex": "( o )" + }, + { + "category_id": 13, + "poly": [ + 741, + 1190, + 829, + 1190, + 829, + 1218, + 741, + 1218 + ], + "score": 0.44, + "latex": "\\{ \\mathbf { X } \\mathbf { A d } \\mathbf { - }" + }, + { + "category_id": 13, + "poly": [ + 737, + 1219, + 762, + 1219, + 762, + 1245, + 737, + 1245 + ], + "score": 0.4, + "latex": "\\mathbf { X }" + }, + { + "category_id": 13, + "poly": [ + 741, + 1190, + 777, + 1190, + 777, + 1218, + 741, + 1218 + ], + "score": 0.39, + "latex": "\\{ \\mathbf { X }" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1623.0, + 1409.0, + 1623.0, + 1409.0, + 1661.0, + 844.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1656.0, + 1404.0, + 1656.0, + 1404.0, + 1689.0, + 844.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1684.0, + 1408.0, + 1684.0, + 1408.0, + 1722.0, + 844.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1715.0, + 1404.0, + 1715.0, + 1404.0, + 1750.0, + 843.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1747.0, + 1155.0, + 1747.0, + 1155.0, + 1781.0, + 844.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 1747.0, + 1404.0, + 1747.0, + 1404.0, + 1781.0, + 1171.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1777.0, + 1405.0, + 1777.0, + 1405.0, + 1811.0, + 843.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1809.0, + 1405.0, + 1809.0, + 1405.0, + 1841.0, + 844.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1839.0, + 1404.0, + 1839.0, + 1404.0, + 1871.0, + 844.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1868.0, + 1217.0, + 1868.0, + 1217.0, + 1901.0, + 845.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 1061.0, + 1176.0, + 1061.0, + 1176.0, + 1077.0, + 1116.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 1079.0, + 1045.0, + 1079.0, + 1045.0, + 1099.0, + 988.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 1077.0, + 1141.0, + 1077.0, + 1141.0, + 1100.0, + 1089.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 1079.0, + 1236.0, + 1079.0, + 1236.0, + 1099.0, + 1184.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 1079.0, + 1327.0, + 1079.0, + 1327.0, + 1099.0, + 1260.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 1117.0, + 1040.0, + 1117.0, + 1040.0, + 1139.0, + 1002.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 1115.0, + 1305.0, + 1115.0, + 1305.0, + 1141.0, + 1239.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 1159.0, + 912.0, + 1159.0, + 912.0, + 1242.0, + 859.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 1159.0, + 1161.0, + 1159.0, + 1161.0, + 1184.0, + 1125.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1367.0, + 1151.0, + 1390.0, + 1151.0, + 1390.0, + 1229.0, + 1367.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 1190.0, + 913.0, + 1190.0, + 913.0, + 1215.0, + 879.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 1190.0, + 1162.0, + 1190.0, + 1162.0, + 1215.0, + 1125.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1225.0, + 912.0, + 1225.0, + 912.0, + 1242.0, + 897.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 1224.0, + 1160.0, + 1224.0, + 1160.0, + 1243.0, + 1143.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1261.0, + 912.0, + 1261.0, + 912.0, + 1338.0, + 861.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 1260.0, + 1161.0, + 1260.0, + 1161.0, + 1284.0, + 1125.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1367.0, + 1245.0, + 1390.0, + 1245.0, + 1390.0, + 1343.0, + 1367.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 1295.0, + 912.0, + 1295.0, + 912.0, + 1318.0, + 880.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 1294.0, + 1162.0, + 1294.0, + 1162.0, + 1320.0, + 1125.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1332.0, + 912.0, + 1332.0, + 912.0, + 1348.0, + 897.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 1331.0, + 1160.0, + 1331.0, + 1160.0, + 1349.0, + 1143.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1369.0, + 913.0, + 1369.0, + 913.0, + 1449.0, + 860.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 1369.0, + 1162.0, + 1369.0, + 1162.0, + 1427.0, + 1125.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1365.0, + 1364.0, + 1388.0, + 1364.0, + 1388.0, + 1446.0, + 1365.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 1440.0, + 911.0, + 1440.0, + 911.0, + 1454.0, + 898.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 1438.0, + 1160.0, + 1438.0, + 1160.0, + 1456.0, + 1144.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1477.0, + 913.0, + 1477.0, + 913.0, + 1559.0, + 861.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 1477.0, + 1161.0, + 1477.0, + 1161.0, + 1535.0, + 1125.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1365.0, + 1458.0, + 1388.0, + 1458.0, + 1388.0, + 1557.0, + 1365.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1545.0, + 913.0, + 1545.0, + 913.0, + 1562.0, + 897.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 1542.0, + 973.0, + 1542.0, + 973.0, + 1552.0, + 960.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 1544.0, + 1160.0, + 1544.0, + 1160.0, + 1563.0, + 1143.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 1537.0, + 1208.0, + 1537.0, + 1208.0, + 1554.0, + 1184.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1561.0, + 932.0, + 1561.0, + 932.0, + 1579.0, + 916.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 1558.0, + 1003.0, + 1558.0, + 1003.0, + 1581.0, + 970.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 1558.0, + 1066.0, + 1558.0, + 1066.0, + 1581.0, + 1033.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 1558.0, + 1128.0, + 1558.0, + 1128.0, + 1581.0, + 1095.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 1560.0, + 1179.0, + 1560.0, + 1179.0, + 1579.0, + 1163.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1557.0, + 1251.0, + 1557.0, + 1251.0, + 1582.0, + 1216.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 1557.0, + 1315.0, + 1557.0, + 1315.0, + 1582.0, + 1279.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1341.0, + 1559.0, + 1373.0, + 1559.0, + 1373.0, + 1579.0, + 1341.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 1576.0, + 1050.0, + 1576.0, + 1050.0, + 1599.0, + 992.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 1575.0, + 1302.0, + 1575.0, + 1302.0, + 1598.0, + 1242.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 445.0, + 971.0, + 445.0, + 971.0, + 488.0, + 293.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2060.0, + 861.0, + 2060.0, + 861.0, + 2091.0, + 840.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 506.0, + 1404.0, + 506.0, + 1404.0, + 539.0, + 295.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 537.0, + 1405.0, + 537.0, + 1405.0, + 570.0, + 295.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 567.0, + 1406.0, + 567.0, + 1406.0, + 603.0, + 292.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 597.0, + 1405.0, + 597.0, + 1405.0, + 631.0, + 291.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 627.0, + 1404.0, + 627.0, + 1404.0, + 659.0, + 294.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 655.0, + 1405.0, + 655.0, + 1405.0, + 691.0, + 292.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 689.0, + 1404.0, + 689.0, + 1404.0, + 722.0, + 294.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 719.0, + 718.0, + 719.0, + 718.0, + 752.0, + 293.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 203.0, + 1408.0, + 203.0, + 1408.0, + 238.0, + 294.0, + 238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 235.0, + 1403.0, + 235.0, + 1403.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 265.0, + 1403.0, + 265.0, + 1403.0, + 295.0, + 296.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 295.0, + 1405.0, + 295.0, + 1405.0, + 327.0, + 294.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 323.0, + 1405.0, + 323.0, + 1405.0, + 358.0, + 294.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 355.0, + 1405.0, + 355.0, + 1405.0, + 390.0, + 295.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 382.0, + 1178.0, + 382.0, + 1178.0, + 419.0, + 292.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1381.0, + 829.0, + 1381.0, + 829.0, + 1418.0, + 295.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1411.0, + 363.0, + 1411.0, + 363.0, + 1450.0, + 294.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 1411.0, + 827.0, + 1411.0, + 827.0, + 1450.0, + 626.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1444.0, + 502.0, + 1444.0, + 502.0, + 1480.0, + 294.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 1444.0, + 827.0, + 1444.0, + 827.0, + 1480.0, + 769.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1475.0, + 827.0, + 1475.0, + 827.0, + 1509.0, + 295.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1506.0, + 829.0, + 1506.0, + 829.0, + 1538.0, + 295.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1537.0, + 827.0, + 1537.0, + 827.0, + 1568.0, + 295.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1566.0, + 829.0, + 1566.0, + 829.0, + 1601.0, + 295.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1596.0, + 654.0, + 1596.0, + 654.0, + 1630.0, + 295.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1596.0, + 830.0, + 1596.0, + 830.0, + 1630.0, + 684.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1627.0, + 829.0, + 1627.0, + 829.0, + 1658.0, + 296.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1658.0, + 464.0, + 1658.0, + 464.0, + 1689.0, + 296.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1658.0, + 828.0, + 1658.0, + 828.0, + 1689.0, + 495.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1686.0, + 830.0, + 1686.0, + 830.0, + 1722.0, + 294.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1718.0, + 827.0, + 1718.0, + 827.0, + 1749.0, + 295.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1748.0, + 826.0, + 1748.0, + 826.0, + 1781.0, + 295.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1780.0, + 826.0, + 1780.0, + 826.0, + 1810.0, + 295.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1811.0, + 827.0, + 1811.0, + 827.0, + 1840.0, + 295.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1840.0, + 829.0, + 1840.0, + 829.0, + 1871.0, + 296.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1869.0, + 827.0, + 1869.0, + 827.0, + 1904.0, + 295.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1899.0, + 830.0, + 1899.0, + 830.0, + 1933.0, + 295.0, + 1933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1094.0, + 827.0, + 1094.0, + 827.0, + 1129.0, + 296.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1125.0, + 828.0, + 1125.0, + 828.0, + 1160.0, + 294.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1157.0, + 565.0, + 1157.0, + 565.0, + 1191.0, + 294.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 1157.0, + 719.0, + 1157.0, + 719.0, + 1191.0, + 666.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 1157.0, + 828.0, + 1157.0, + 828.0, + 1191.0, + 819.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1187.0, + 740.0, + 1187.0, + 740.0, + 1220.0, + 293.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1215.0, + 392.0, + 1215.0, + 392.0, + 1252.0, + 294.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 1215.0, + 736.0, + 1215.0, + 736.0, + 1252.0, + 719.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 1215.0, + 832.0, + 1215.0, + 832.0, + 1252.0, + 763.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1246.0, + 407.0, + 1246.0, + 407.0, + 1283.0, + 294.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 1246.0, + 832.0, + 1246.0, + 832.0, + 1283.0, + 731.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1275.0, + 384.0, + 1275.0, + 384.0, + 1313.0, + 293.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1275.0, + 833.0, + 1275.0, + 833.0, + 1313.0, + 684.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1303.0, + 359.0, + 1303.0, + 359.0, + 1345.0, + 292.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 1303.0, + 726.0, + 1303.0, + 726.0, + 1345.0, + 643.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 1303.0, + 831.0, + 1303.0, + 831.0, + 1345.0, + 826.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1335.0, + 598.0, + 1335.0, + 598.0, + 1371.0, + 294.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 762.0, + 1407.0, + 762.0, + 1407.0, + 797.0, + 295.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 790.0, + 1407.0, + 790.0, + 1407.0, + 830.0, + 293.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 825.0, + 876.0, + 825.0, + 876.0, + 857.0, + 296.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 825.0, + 1182.0, + 825.0, + 1182.0, + 857.0, + 925.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 825.0, + 1405.0, + 825.0, + 1405.0, + 857.0, + 1220.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 851.0, + 1407.0, + 851.0, + 1407.0, + 893.0, + 291.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 882.0, + 1152.0, + 882.0, + 1152.0, + 920.0, + 292.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 928.0, + 1409.0, + 928.0, + 1409.0, + 965.0, + 294.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 961.0, + 1410.0, + 961.0, + 1410.0, + 994.0, + 295.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 991.0, + 1409.0, + 991.0, + 1409.0, + 1023.0, + 294.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1020.0, + 1409.0, + 1020.0, + 1409.0, + 1056.0, + 292.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1943.0, + 1407.0, + 1943.0, + 1407.0, + 1981.0, + 294.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2011.0, + 294.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1896.0, + 1001.0, + 1896.0, + 1001.0, + 1936.0, + 294.0, + 1936.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 340, + 942, + 340, + 942, + 734, + 298, + 734 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1763, + 1405, + 1763, + 1405, + 2008, + 297, + 2008 + ], + "score": 0.979 + }, + { + "category_id": 5, + "poly": [ + 971, + 600, + 1390, + 600, + 1390, + 751, + 971, + 751 + ], + "score": 0.974, + "html": "
Rule 1Rule 2Rule 3Rule 4Rule 5
X Addition3609945130
X Subtraction0482010
Y Subtraction03945320
Y Addition0571599335
" + }, + { + "category_id": 1, + "poly": [ + 297, + 203, + 1403, + 203, + 1403, + 325, + 297, + 325 + ], + "score": 0.973 + }, + { + "category_id": 5, + "poly": [ + 974, + 951, + 1391, + 951, + 1391, + 1081, + 974, + 1081 + ], + "score": 0.971, + "html": "
Model NameTestTransfer
RPIN (Qi et al. (2021))1.254±0.0086.377±0.325
PNPSSNPS1.250±0.0075.411±0.45
1.68±0.025.80±0.15
" + }, + { + "category_id": 1, + "poly": [ + 301, + 832, + 1399, + 832, + 1399, + 924, + 301, + 924 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 301, + 1657, + 1397, + 1657, + 1397, + 1750, + 301, + 1750 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 964, + 1095, + 1403, + 1095, + 1403, + 1398, + 964, + 1398 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 299, + 948, + 944, + 948, + 944, + 1398, + 299, + 1398 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 964, + 337, + 1403, + 337, + 1403, + 581, + 964, + 581 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 298, + 1400, + 1404, + 1400, + 1404, + 1644, + 298, + 1644 + ], + "score": 0.921 + }, + { + "category_id": 0, + "poly": [ + 301, + 774, + 825, + 774, + 825, + 807, + 301, + 807 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 858, + 2061, + 858, + 2084, + 841, + 2084 + ], + "score": 0.711 + }, + { + "category_id": 13, + "poly": [ + 1066, + 1947, + 1253, + 1947, + 1253, + 1979, + 1066, + 1979 + ], + "score": 0.92, + "latex": "P _ { S N P S } = 1 / M" + }, + { + "category_id": 13, + "poly": [ + 1153, + 1914, + 1400, + 1914, + 1400, + 1947, + 1153, + 1947 + ], + "score": 0.92, + "latex": "P _ { P N P S } \\dot { = } N - 1 / N" + }, + { + "category_id": 13, + "poly": [ + 473, + 492, + 588, + 492, + 588, + 523, + 473, + 523 + ], + "score": 0.9, + "latex": "0 . 0 1 _ { \\pm 0 . 0 0 1 }" + }, + { + "category_id": 13, + "poly": [ + 391, + 522, + 504, + 522, + 504, + 554, + 391, + 554 + ], + "score": 0.9, + "latex": "0 . 0 4 _ { \\pm 0 . 0 0 8 }" + }, + { + "category_id": 13, + "poly": [ + 581, + 1689, + 662, + 1689, + 662, + 1717, + 581, + 1717 + ], + "score": 0.88, + "latex": "t = 1 5" + }, + { + "category_id": 13, + "poly": [ + 680, + 1719, + 759, + 1719, + 759, + 1746, + 680, + 1746 + ], + "score": 0.87, + "latex": "t = 3 0" + }, + { + "category_id": 13, + "poly": [ + 1296, + 1053, + 1374, + 1053, + 1374, + 1076, + 1296, + 1076 + ], + "score": 0.8, + "latex": "5 . 8 0 { \\scriptstyle \\pm 0 . 1 5 }" + }, + { + "category_id": 13, + "poly": [ + 1292, + 1020, + 1378, + 1020, + 1378, + 1043, + 1292, + 1043 + ], + "score": 0.77, + "latex": "5 . 4 1 1 { \\scriptstyle \\pm 0 . 4 5 }" + }, + { + "category_id": 13, + "poly": [ + 1288, + 987, + 1380, + 987, + 1380, + 1012, + 1288, + 1012 + ], + "score": 0.77, + "latex": "6 . 3 7 7 { \\scriptstyle \\pm 0 . 3 2 5 }" + }, + { + "category_id": 13, + "poly": [ + 1113, + 1885, + 1146, + 1885, + 1146, + 1912, + 1113, + 1912 + ], + "score": 0.76, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1183, + 1053, + 1261, + 1053, + 1261, + 1076, + 1183, + 1076 + ], + "score": 0.72, + "latex": "1 . 6 8 { \\scriptstyle \\pm 0 . 0 2 }" + }, + { + "category_id": 13, + "poly": [ + 1322, + 1219, + 1338, + 1219, + 1338, + 1244, + 1322, + 1244 + ], + "score": 0.71, + "latex": "\\pmb { t }" + }, + { + "category_id": 13, + "poly": [ + 1176, + 1019, + 1268, + 1019, + 1268, + 1043, + 1176, + 1043 + ], + "score": 0.68, + "latex": "1 . 2 5 0 { \\scriptstyle \\pm 0 . 0 0 7 }" + }, + { + "category_id": 13, + "poly": [ + 345, + 1915, + 382, + 1915, + 382, + 1943, + 345, + 1943 + ], + "score": 0.62, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 1175, + 987, + 1268, + 987, + 1268, + 1012, + 1175, + 1012 + ], + "score": 0.56, + "latex": "1 . 2 5 4 { \\scriptstyle \\pm 0 . 0 0 8 }" + }, + { + "category_id": 13, + "poly": [ + 350, + 1160, + 367, + 1160, + 367, + 1185, + 350, + 1185 + ], + "score": 0.51, + "latex": "\\pmb { t }" + }, + { + "category_id": 13, + "poly": [ + 679, + 1281, + 696, + 1281, + 696, + 1306, + 679, + 1306 + ], + "score": 0.37, + "latex": "\\pmb { t }" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 768.0, + 828.0, + 768.0, + 828.0, + 817.0, + 294.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2091.0, + 839.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 338.0, + 944.0, + 338.0, + 944.0, + 373.0, + 296.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 365.0, + 945.0, + 365.0, + 945.0, + 406.0, + 293.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 398.0, + 943.0, + 398.0, + 943.0, + 435.0, + 295.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 430.0, + 944.0, + 430.0, + 944.0, + 466.0, + 296.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 462.0, + 946.0, + 462.0, + 946.0, + 492.0, + 294.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 493.0, + 472.0, + 493.0, + 472.0, + 525.0, + 294.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 493.0, + 944.0, + 493.0, + 944.0, + 525.0, + 589.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 518.0, + 390.0, + 518.0, + 390.0, + 556.0, + 294.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 518.0, + 946.0, + 518.0, + 946.0, + 556.0, + 505.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 551.0, + 947.0, + 551.0, + 947.0, + 585.0, + 295.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 582.0, + 943.0, + 582.0, + 943.0, + 614.0, + 296.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 610.0, + 944.0, + 610.0, + 944.0, + 645.0, + 295.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 642.0, + 944.0, + 642.0, + 944.0, + 677.0, + 294.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 674.0, + 943.0, + 674.0, + 943.0, + 706.0, + 294.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 704.0, + 728.0, + 704.0, + 728.0, + 736.0, + 297.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1763.0, + 1405.0, + 1763.0, + 1405.0, + 1797.0, + 295.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1793.0, + 1405.0, + 1793.0, + 1405.0, + 1827.0, + 295.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1823.0, + 1405.0, + 1823.0, + 1405.0, + 1857.0, + 295.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1856.0, + 1406.0, + 1856.0, + 1406.0, + 1889.0, + 295.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1884.0, + 1112.0, + 1884.0, + 1112.0, + 1918.0, + 294.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1884.0, + 1405.0, + 1884.0, + 1405.0, + 1918.0, + 1147.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1912.0, + 344.0, + 1912.0, + 344.0, + 1953.0, + 291.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 1912.0, + 1152.0, + 1912.0, + 1152.0, + 1953.0, + 383.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1912.0, + 1406.0, + 1912.0, + 1406.0, + 1953.0, + 1401.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1943.0, + 1065.0, + 1943.0, + 1065.0, + 1983.0, + 291.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 1943.0, + 1406.0, + 1943.0, + 1406.0, + 1983.0, + 1254.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1975.0, + 779.0, + 1975.0, + 779.0, + 2011.0, + 294.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 200.0, + 1403.0, + 200.0, + 1403.0, + 239.0, + 294.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 234.0, + 1404.0, + 234.0, + 1404.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 263.0, + 1404.0, + 263.0, + 1404.0, + 298.0, + 295.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 295.0, + 863.0, + 295.0, + 863.0, + 328.0, + 293.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 831.0, + 1407.0, + 831.0, + 1407.0, + 868.0, + 295.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 862.0, + 1403.0, + 862.0, + 1403.0, + 896.0, + 293.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 892.0, + 1231.0, + 892.0, + 1231.0, + 930.0, + 295.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1655.0, + 1405.0, + 1655.0, + 1405.0, + 1693.0, + 296.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1687.0, + 580.0, + 1687.0, + 580.0, + 1723.0, + 293.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1687.0, + 1403.0, + 1687.0, + 1403.0, + 1723.0, + 663.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1718.0, + 679.0, + 1718.0, + 679.0, + 1752.0, + 297.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 1718.0, + 777.0, + 1718.0, + 777.0, + 1752.0, + 760.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1093.0, + 1407.0, + 1093.0, + 1407.0, + 1126.0, + 962.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 1125.0, + 1405.0, + 1125.0, + 1405.0, + 1158.0, + 960.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1155.0, + 1405.0, + 1155.0, + 1405.0, + 1188.0, + 963.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1185.0, + 1407.0, + 1185.0, + 1407.0, + 1217.0, + 962.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1216.0, + 1321.0, + 1216.0, + 1321.0, + 1248.0, + 962.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1339.0, + 1216.0, + 1405.0, + 1216.0, + 1405.0, + 1248.0, + 1339.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 1246.0, + 1405.0, + 1246.0, + 1405.0, + 1278.0, + 961.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 1277.0, + 1408.0, + 1277.0, + 1408.0, + 1309.0, + 961.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 1307.0, + 1405.0, + 1307.0, + 1405.0, + 1339.0, + 961.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1337.0, + 1405.0, + 1337.0, + 1405.0, + 1370.0, + 962.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 1369.0, + 1311.0, + 1369.0, + 1311.0, + 1399.0, + 964.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 945.0, + 946.0, + 945.0, + 946.0, + 979.0, + 296.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 975.0, + 946.0, + 975.0, + 946.0, + 1009.0, + 296.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1006.0, + 943.0, + 1006.0, + 943.0, + 1041.0, + 296.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1036.0, + 945.0, + 1036.0, + 945.0, + 1070.0, + 295.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1068.0, + 946.0, + 1068.0, + 946.0, + 1098.0, + 294.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1096.0, + 945.0, + 1096.0, + 945.0, + 1131.0, + 295.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1126.0, + 945.0, + 1126.0, + 945.0, + 1161.0, + 295.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1159.0, + 349.0, + 1159.0, + 349.0, + 1190.0, + 295.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1159.0, + 948.0, + 1159.0, + 948.0, + 1190.0, + 368.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1187.0, + 945.0, + 1187.0, + 945.0, + 1222.0, + 294.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1220.0, + 944.0, + 1220.0, + 944.0, + 1251.0, + 295.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1249.0, + 944.0, + 1249.0, + 944.0, + 1283.0, + 295.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1278.0, + 678.0, + 1278.0, + 678.0, + 1313.0, + 297.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 1278.0, + 945.0, + 1278.0, + 945.0, + 1313.0, + 697.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1309.0, + 946.0, + 1309.0, + 946.0, + 1343.0, + 292.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1339.0, + 945.0, + 1339.0, + 945.0, + 1370.0, + 295.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1369.0, + 948.0, + 1369.0, + 948.0, + 1404.0, + 296.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 333.0, + 1404.0, + 333.0, + 1404.0, + 371.0, + 960.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 366.0, + 1405.0, + 366.0, + 1405.0, + 399.0, + 963.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 396.0, + 1408.0, + 396.0, + 1408.0, + 430.0, + 961.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 429.0, + 1405.0, + 429.0, + 1405.0, + 460.0, + 962.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 457.0, + 1406.0, + 457.0, + 1406.0, + 490.0, + 962.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 486.0, + 1409.0, + 486.0, + 1409.0, + 522.0, + 961.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 518.0, + 1408.0, + 518.0, + 1408.0, + 554.0, + 962.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 551.0, + 1187.0, + 551.0, + 1187.0, + 582.0, + 964.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1397.0, + 1405.0, + 1397.0, + 1405.0, + 1437.0, + 292.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1432.0, + 1404.0, + 1432.0, + 1404.0, + 1462.0, + 294.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1461.0, + 1402.0, + 1461.0, + 1402.0, + 1495.0, + 294.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1490.0, + 1406.0, + 1490.0, + 1406.0, + 1527.0, + 293.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1518.0, + 1405.0, + 1518.0, + 1405.0, + 1555.0, + 293.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1552.0, + 1406.0, + 1552.0, + 1406.0, + 1589.0, + 293.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1582.0, + 1409.0, + 1582.0, + 1409.0, + 1619.0, + 293.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1612.0, + 1338.0, + 1612.0, + 1338.0, + 1647.0, + 293.0, + 1647.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 201, + 1405, + 201, + 1405, + 688, + 298, + 688 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1793, + 1404, + 1793, + 1404, + 2007, + 298, + 2007 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 731, + 826, + 731, + 826, + 1401, + 298, + 1401 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 848, + 695, + 1345, + 695, + 1345, + 1004, + 848, + 1004 + ], + "score": 0.966 + }, + { + "category_id": 4, + "poly": [ + 848, + 1021, + 1403, + 1021, + 1403, + 1474, + 848, + 1474 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 297, + 1460, + 1404, + 1460, + 1404, + 1780, + 297, + 1780 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 841, + 2062, + 858, + 2062, + 858, + 2085, + 841, + 2085 + ], + "score": 0.795 + }, + { + "category_id": 13, + "poly": [ + 839, + 203, + 1028, + 203, + 1028, + 237, + 839, + 237 + ], + "score": 0.93, + "latex": "N = \\{ 1 , 2 , 4 , 6 \\}" + }, + { + "category_id": 13, + "poly": [ + 517, + 234, + 729, + 234, + 729, + 265, + 517, + 265 + ], + "score": 0.91, + "latex": "P _ { P N P S } > P _ { S N P S }" + }, + { + "category_id": 13, + "poly": [ + 467, + 325, + 554, + 325, + 554, + 353, + 467, + 353 + ], + "score": 0.9, + "latex": "N = 4" + }, + { + "category_id": 13, + "poly": [ + 1142, + 204, + 1231, + 204, + 1231, + 232, + 1142, + 232 + ], + "score": 0.89, + "latex": "M = 3" + }, + { + "category_id": 13, + "poly": [ + 644, + 1158, + 767, + 1158, + 767, + 1187, + 644, + 1187 + ], + "score": 0.86, + "latex": "{ \\mathrm { S C O F F } } { + } { + }" + }, + { + "category_id": 13, + "poly": [ + 1106, + 294, + 1251, + 294, + 1251, + 328, + 1106, + 328 + ], + "score": 0.83, + "latex": "N = \\{ 2 + 1" + }, + { + "category_id": 13, + "poly": [ + 455, + 1036, + 572, + 1036, + 572, + 1065, + 455, + 1065 + ], + "score": 0.78, + "latex": "{ \\mathrm { S C O F F } } { + } { + }" + }, + { + "category_id": 13, + "poly": [ + 962, + 295, + 993, + 295, + 993, + 323, + 962, + 323 + ], + "score": 0.76, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 298, + 234, + 330, + 234, + 330, + 262, + 298, + 262 + ], + "score": 0.76, + "latex": "_ N" + }, + { + "category_id": 13, + "poly": [ + 474, + 1128, + 497, + 1128, + 497, + 1154, + 474, + 1154 + ], + "score": 0.53, + "latex": "^ +" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 875.0, + 1082.0, + 875.0, + 1082.0, + 899.0, + 1033.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 875.0, + 1153.0, + 875.0, + 1153.0, + 898.0, + 1102.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 902.0, + 1075.0, + 902.0, + 1075.0, + 922.0, + 1040.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 901.0, + 1146.0, + 901.0, + 1146.0, + 922.0, + 1110.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 938.0, + 1076.0, + 938.0, + 1076.0, + 963.0, + 1039.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 938.0, + 1147.0, + 938.0, + 1147.0, + 963.0, + 1109.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 976.0, + 1075.0, + 976.0, + 1075.0, + 997.0, + 1040.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 976.0, + 1145.0, + 976.0, + 1145.0, + 998.0, + 1110.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1019.0, + 1407.0, + 1019.0, + 1407.0, + 1053.0, + 844.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1049.0, + 1404.0, + 1049.0, + 1404.0, + 1083.0, + 845.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1082.0, + 1406.0, + 1082.0, + 1406.0, + 1113.0, + 846.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1111.0, + 1405.0, + 1111.0, + 1405.0, + 1141.0, + 845.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1142.0, + 1404.0, + 1142.0, + 1404.0, + 1173.0, + 846.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1170.0, + 1405.0, + 1170.0, + 1405.0, + 1207.0, + 844.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1203.0, + 1404.0, + 1203.0, + 1404.0, + 1234.0, + 845.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1233.0, + 1406.0, + 1233.0, + 1406.0, + 1265.0, + 843.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1263.0, + 1408.0, + 1263.0, + 1408.0, + 1297.0, + 844.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1292.0, + 1405.0, + 1292.0, + 1405.0, + 1326.0, + 844.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1324.0, + 1404.0, + 1324.0, + 1404.0, + 1354.0, + 843.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1353.0, + 1408.0, + 1353.0, + 1408.0, + 1387.0, + 843.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1386.0, + 1403.0, + 1386.0, + 1403.0, + 1414.0, + 846.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1415.0, + 1408.0, + 1415.0, + 1408.0, + 1445.0, + 844.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1445.0, + 1298.0, + 1445.0, + 1298.0, + 1478.0, + 845.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2089.0, + 839.0, + 2089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 199.0, + 838.0, + 199.0, + 838.0, + 240.0, + 292.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 199.0, + 1141.0, + 199.0, + 1141.0, + 240.0, + 1029.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 199.0, + 1406.0, + 199.0, + 1406.0, + 240.0, + 1232.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 232.0, + 297.0, + 232.0, + 297.0, + 270.0, + 293.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 232.0, + 516.0, + 232.0, + 516.0, + 270.0, + 331.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 232.0, + 1408.0, + 232.0, + 1408.0, + 270.0, + 730.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 264.0, + 1405.0, + 264.0, + 1405.0, + 298.0, + 293.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 292.0, + 961.0, + 292.0, + 961.0, + 331.0, + 292.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 292.0, + 1105.0, + 292.0, + 1105.0, + 331.0, + 994.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 292.0, + 1405.0, + 292.0, + 1405.0, + 331.0, + 1252.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 321.0, + 466.0, + 321.0, + 466.0, + 360.0, + 292.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 321.0, + 1406.0, + 321.0, + 1406.0, + 360.0, + 555.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 355.0, + 1406.0, + 355.0, + 1406.0, + 389.0, + 295.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 385.0, + 1405.0, + 385.0, + 1405.0, + 419.0, + 293.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 415.0, + 1405.0, + 415.0, + 1405.0, + 449.0, + 295.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 443.0, + 1405.0, + 443.0, + 1405.0, + 480.0, + 293.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 475.0, + 1406.0, + 475.0, + 1406.0, + 509.0, + 295.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 507.0, + 1405.0, + 507.0, + 1405.0, + 541.0, + 295.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 535.0, + 1406.0, + 535.0, + 1406.0, + 572.0, + 293.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 567.0, + 1408.0, + 567.0, + 1408.0, + 601.0, + 293.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 599.0, + 1405.0, + 599.0, + 1405.0, + 629.0, + 296.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 628.0, + 1405.0, + 628.0, + 1405.0, + 662.0, + 295.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 655.0, + 1410.0, + 655.0, + 1410.0, + 695.0, + 292.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1793.0, + 1404.0, + 1793.0, + 1404.0, + 1828.0, + 296.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1824.0, + 1405.0, + 1824.0, + 1405.0, + 1857.0, + 293.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1855.0, + 1405.0, + 1855.0, + 1405.0, + 1889.0, + 293.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1883.0, + 1405.0, + 1883.0, + 1405.0, + 1920.0, + 293.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1914.0, + 1405.0, + 1914.0, + 1405.0, + 1950.0, + 293.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1947.0, + 1409.0, + 1947.0, + 1409.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1975.0, + 1259.0, + 1975.0, + 1259.0, + 2012.0, + 293.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 728.0, + 830.0, + 728.0, + 830.0, + 769.0, + 294.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 762.0, + 826.0, + 762.0, + 826.0, + 795.0, + 295.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 793.0, + 827.0, + 793.0, + 827.0, + 824.0, + 295.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 823.0, + 825.0, + 823.0, + 825.0, + 853.0, + 295.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 855.0, + 826.0, + 855.0, + 826.0, + 886.0, + 294.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 884.0, + 828.0, + 884.0, + 828.0, + 918.0, + 293.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 914.0, + 829.0, + 914.0, + 829.0, + 948.0, + 295.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 945.0, + 827.0, + 945.0, + 827.0, + 978.0, + 294.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 974.0, + 830.0, + 974.0, + 830.0, + 1010.0, + 294.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1004.0, + 829.0, + 1004.0, + 829.0, + 1040.0, + 294.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1036.0, + 454.0, + 1036.0, + 454.0, + 1070.0, + 295.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1036.0, + 828.0, + 1036.0, + 828.0, + 1070.0, + 573.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1068.0, + 832.0, + 1068.0, + 832.0, + 1099.0, + 295.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1097.0, + 830.0, + 1097.0, + 830.0, + 1129.0, + 293.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1124.0, + 473.0, + 1124.0, + 473.0, + 1160.0, + 295.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 1124.0, + 830.0, + 1124.0, + 830.0, + 1160.0, + 498.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1157.0, + 643.0, + 1157.0, + 643.0, + 1188.0, + 294.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1157.0, + 826.0, + 1157.0, + 826.0, + 1188.0, + 768.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1188.0, + 830.0, + 1188.0, + 830.0, + 1220.0, + 294.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1217.0, + 827.0, + 1217.0, + 827.0, + 1250.0, + 293.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1246.0, + 826.0, + 1246.0, + 826.0, + 1282.0, + 294.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1278.0, + 826.0, + 1278.0, + 826.0, + 1311.0, + 294.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1310.0, + 825.0, + 1310.0, + 825.0, + 1341.0, + 296.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1340.0, + 825.0, + 1340.0, + 825.0, + 1371.0, + 296.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1368.0, + 727.0, + 1368.0, + 727.0, + 1406.0, + 293.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1451.0, + 1298.0, + 1451.0, + 1298.0, + 1480.0, + 299.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1477.0, + 1404.0, + 1477.0, + 1404.0, + 1510.0, + 296.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1507.0, + 1405.0, + 1507.0, + 1405.0, + 1540.0, + 296.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1536.0, + 1406.0, + 1536.0, + 1406.0, + 1573.0, + 295.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1566.0, + 1406.0, + 1566.0, + 1406.0, + 1603.0, + 294.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1598.0, + 1404.0, + 1598.0, + 1404.0, + 1630.0, + 295.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1624.0, + 1407.0, + 1624.0, + 1407.0, + 1663.0, + 291.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1655.0, + 1406.0, + 1655.0, + 1406.0, + 1696.0, + 294.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1687.0, + 1406.0, + 1687.0, + 1406.0, + 1724.0, + 294.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1716.0, + 1405.0, + 1716.0, + 1405.0, + 1756.0, + 294.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1747.0, + 1112.0, + 1747.0, + 1112.0, + 1784.0, + 295.0, + 1784.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 257, + 1405, + 257, + 1405, + 564, + 296, + 564 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 991, + 1404, + 991, + 1404, + 1207, + 297, + 1207 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1218, + 1405, + 1218, + 1405, + 1433, + 297, + 1433 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 624, + 944, + 624, + 944, + 960, + 298, + 960 + ], + "score": 0.979 + }, + { + "category_id": 5, + "poly": [ + 981, + 599, + 1386, + 599, + 1386, + 754, + 981, + 754 + ], + "score": 0.973, + "html": "
Model NameTestTransfer
SCOFF0.280.15
SCOFF++0.84370.2632
PNPS (10 Rules+1 Null Rule)0.78130.1997
SNPS (10 Rules)0.85180.3553
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1884, + 1403, + 1884, + 1403, + 2007, + 298, + 2007 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 383, + 1451, + 1304, + 1451, + 1304, + 1732, + 383, + 1732 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 296, + 1742, + 1405, + 1742, + 1405, + 1867, + 296, + 1867 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 964, + 767, + 1403, + 767, + 1403, + 979, + 964, + 979 + ], + "score": 0.913 + }, + { + "category_id": 0, + "poly": [ + 297, + 201, + 920, + 201, + 920, + 236, + 297, + 236 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 840, + 2061, + 858, + 2061, + 858, + 2084, + 840, + 2084 + ], + "score": 0.799 + }, + { + "category_id": 13, + "poly": [ + 298, + 746, + 411, + 746, + 411, + 780, + 298, + 780 + ], + "score": 0.9, + "latex": "\\{ 1 , 5 , 1 0 \\} " + }, + { + "category_id": 13, + "poly": [ + 578, + 809, + 655, + 809, + 655, + 838, + 578, + 838 + ], + "score": 0.85, + "latex": "( \\mathrm { H } @ 1 )" + }, + { + "category_id": 13, + "poly": [ + 877, + 717, + 943, + 717, + 943, + 746, + 877, + 746 + ], + "score": 0.84, + "latex": "K =" + }, + { + "category_id": 13, + "poly": [ + 615, + 839, + 679, + 839, + 679, + 866, + 615, + 866 + ], + "score": 0.83, + "latex": "\\mathrm { H @ 1 }" + }, + { + "category_id": 13, + "poly": [ + 768, + 1805, + 831, + 1805, + 831, + 1833, + 768, + 1833 + ], + "score": 0.81, + "latex": "\\mathrm { H @ 1 }" + }, + { + "category_id": 13, + "poly": [ + 779, + 718, + 808, + 718, + 808, + 745, + 779, + 745 + ], + "score": 0.79, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 296, + 1835, + 360, + 1835, + 360, + 1863, + 296, + 1863 + ], + "score": 0.69, + "latex": "\\mathrm { H @ 1 }" + }, + { + "category_id": 13, + "poly": [ + 964, + 889, + 1088, + 889, + 1088, + 919, + 964, + 919 + ], + "score": 0.66, + "latex": "{ \\mathrm { S C O F F + + } }" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1468.0, + 791.0, + 1468.0, + 791.0, + 1480.0, + 768.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1468.0, + 915.0, + 1468.0, + 915.0, + 1481.0, + 900.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1457.0, + 1293.0, + 1457.0, + 1293.0, + 1487.0, + 1253.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1479.0, + 794.0, + 1479.0, + 794.0, + 1494.0, + 767.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 1481.0, + 1297.0, + 1481.0, + 1297.0, + 1499.0, + 1267.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 1495.0, + 412.0, + 1495.0, + 412.0, + 1508.0, + 397.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1492.0, + 915.0, + 1492.0, + 915.0, + 1505.0, + 899.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 1514.0, + 917.0, + 1514.0, + 917.0, + 1530.0, + 898.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 1536.0, + 413.0, + 1536.0, + 413.0, + 1579.0, + 385.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 1537.0, + 921.0, + 1537.0, + 921.0, + 1583.0, + 883.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 1577.0, + 412.0, + 1577.0, + 412.0, + 1590.0, + 397.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 1585.0, + 916.0, + 1585.0, + 916.0, + 1601.0, + 898.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 1609.0, + 916.0, + 1609.0, + 916.0, + 1625.0, + 898.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 1619.0, + 412.0, + 1619.0, + 412.0, + 1632.0, + 397.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1667.0, + 623.0, + 1667.0, + 623.0, + 1692.0, + 587.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 1666.0, + 741.0, + 1666.0, + 741.0, + 1679.0, + 727.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 1670.0, + 1125.0, + 1670.0, + 1125.0, + 1692.0, + 1089.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 1665.0, + 1243.0, + 1665.0, + 1243.0, + 1677.0, + 1229.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 1699.0, + 688.0, + 1699.0, + 688.0, + 1735.0, + 509.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 1699.0, + 1196.0, + 1699.0, + 1196.0, + 1737.0, + 1007.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1742.0, + 1405.0, + 1742.0, + 1405.0, + 1779.0, + 294.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1775.0, + 1405.0, + 1775.0, + 1405.0, + 1808.0, + 294.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1802.0, + 767.0, + 1802.0, + 767.0, + 1842.0, + 291.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 1802.0, + 1405.0, + 1802.0, + 1405.0, + 1842.0, + 832.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1834.0, + 1128.0, + 1834.0, + 1128.0, + 1870.0, + 361.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 200.0, + 921.0, + 200.0, + 921.0, + 243.0, + 294.0, + 243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2060.0, + 860.0, + 2060.0, + 860.0, + 2090.0, + 839.0, + 2090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 258.0, + 1405.0, + 258.0, + 1405.0, + 293.0, + 294.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 287.0, + 1407.0, + 287.0, + 1407.0, + 324.0, + 294.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 319.0, + 1405.0, + 319.0, + 1405.0, + 355.0, + 294.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 349.0, + 1405.0, + 349.0, + 1405.0, + 385.0, + 294.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 375.0, + 1406.0, + 375.0, + 1406.0, + 418.0, + 291.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 408.0, + 1406.0, + 408.0, + 1406.0, + 447.0, + 293.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 442.0, + 1403.0, + 442.0, + 1403.0, + 474.0, + 295.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 470.0, + 1407.0, + 470.0, + 1407.0, + 506.0, + 294.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 501.0, + 1405.0, + 501.0, + 1405.0, + 536.0, + 294.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 530.0, + 1410.0, + 530.0, + 1410.0, + 568.0, + 294.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 992.0, + 1405.0, + 992.0, + 1405.0, + 1026.0, + 295.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1021.0, + 1406.0, + 1021.0, + 1406.0, + 1057.0, + 292.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1050.0, + 1406.0, + 1050.0, + 1406.0, + 1088.0, + 294.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1084.0, + 1404.0, + 1084.0, + 1404.0, + 1118.0, + 294.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1114.0, + 1404.0, + 1114.0, + 1404.0, + 1147.0, + 294.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1143.0, + 1405.0, + 1143.0, + 1405.0, + 1177.0, + 294.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1176.0, + 854.0, + 1176.0, + 854.0, + 1210.0, + 295.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1217.0, + 1405.0, + 1217.0, + 1405.0, + 1254.0, + 294.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1247.0, + 1405.0, + 1247.0, + 1405.0, + 1285.0, + 294.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1279.0, + 1406.0, + 1279.0, + 1406.0, + 1316.0, + 292.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1312.0, + 1402.0, + 1312.0, + 1402.0, + 1343.0, + 295.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1340.0, + 1406.0, + 1340.0, + 1406.0, + 1375.0, + 292.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1372.0, + 1408.0, + 1372.0, + 1408.0, + 1407.0, + 295.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1403.0, + 602.0, + 1403.0, + 602.0, + 1436.0, + 293.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 625.0, + 945.0, + 625.0, + 945.0, + 658.0, + 295.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 655.0, + 943.0, + 655.0, + 943.0, + 688.0, + 296.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 686.0, + 946.0, + 686.0, + 946.0, + 719.0, + 294.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 714.0, + 778.0, + 714.0, + 778.0, + 749.0, + 295.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 714.0, + 876.0, + 714.0, + 876.0, + 749.0, + 809.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 745.0, + 946.0, + 745.0, + 946.0, + 779.0, + 412.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 776.0, + 948.0, + 776.0, + 948.0, + 810.0, + 296.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 805.0, + 577.0, + 805.0, + 577.0, + 841.0, + 294.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 805.0, + 946.0, + 805.0, + 946.0, + 841.0, + 656.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 837.0, + 614.0, + 837.0, + 614.0, + 870.0, + 295.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 837.0, + 948.0, + 837.0, + 948.0, + 870.0, + 680.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 868.0, + 944.0, + 868.0, + 944.0, + 901.0, + 295.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 898.0, + 943.0, + 898.0, + 943.0, + 931.0, + 295.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 929.0, + 908.0, + 929.0, + 908.0, + 962.0, + 295.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1884.0, + 1404.0, + 1884.0, + 1404.0, + 1920.0, + 293.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1914.0, + 1405.0, + 1914.0, + 1405.0, + 1950.0, + 294.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1945.0, + 1407.0, + 1945.0, + 1407.0, + 1982.0, + 293.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1977.0, + 1404.0, + 1977.0, + 1404.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 767.0, + 1404.0, + 767.0, + 1404.0, + 797.0, + 962.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 794.0, + 1406.0, + 794.0, + 1406.0, + 831.0, + 962.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 827.0, + 1404.0, + 827.0, + 1404.0, + 859.0, + 962.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 858.0, + 1404.0, + 858.0, + 1404.0, + 889.0, + 962.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 885.0, + 1410.0, + 885.0, + 1410.0, + 923.0, + 1089.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 916.0, + 1405.0, + 916.0, + 1405.0, + 954.0, + 961.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 950.0, + 1135.0, + 950.0, + 1135.0, + 980.0, + 960.0, + 980.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 661, + 1404, + 661, + 1404, + 937, + 297, + 937 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1527, + 1404, + 1527, + 1404, + 1711, + 297, + 1711 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 951, + 942, + 951, + 942, + 1315, + 299, + 1315 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 470, + 1403, + 470, + 1403, + 624, + 298, + 624 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 334, + 1403, + 334, + 1403, + 456, + 298, + 456 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 1748, + 1404, + 1748, + 1404, + 1901, + 298, + 1901 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 300, + 1914, + 1402, + 1914, + 1402, + 2007, + 300, + 2007 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 299, + 1316, + 1404, + 1316, + 1404, + 1436, + 299, + 1436 + ], + "score": 0.968 + }, + { + "category_id": 5, + "poly": [ + 1038, + 989, + 1327, + 989, + 1327, + 1104, + 1038, + 1104 + ], + "score": 0.968, + "html": "
ModelMOTA↑MOTP↑
OP3NPS89.1±5.178.4±2.4
90.72±5.1579.91±0.3
" + }, + { + "category_id": 1, + "poly": [ + 298, + 203, + 1401, + 203, + 1401, + 295, + 298, + 295 + ], + "score": 0.961 + }, + { + "category_id": 0, + "poly": [ + 300, + 1462, + 728, + 1462, + 728, + 1500, + 300, + 1500 + ], + "score": 0.911 + }, + { + "category_id": 2, + "poly": [ + 837, + 2062, + 865, + 2062, + 865, + 2085, + 837, + 2085 + ], + "score": 0.849 + }, + { + "category_id": 6, + "poly": [ + 963, + 1119, + 1405, + 1119, + 1405, + 1301, + 963, + 1301 + ], + "score": 0.779 + }, + { + "category_id": 1, + "poly": [ + 963, + 1119, + 1405, + 1119, + 1405, + 1301, + 963, + 1301 + ], + "score": 0.152 + }, + { + "category_id": 13, + "poly": [ + 1195, + 502, + 1259, + 502, + 1259, + 531, + 1195, + 531 + ], + "score": 0.81, + "latex": "\\mathrm { H @ 1 }" + }, + { + "category_id": 13, + "poly": [ + 1227, + 1072, + 1306, + 1072, + 1306, + 1096, + 1227, + 1096 + ], + "score": 0.74, + "latex": "7 9 . 9 1 _ { \\pm 0 . 3 }" + }, + { + "category_id": 13, + "poly": [ + 1234, + 1041, + 1304, + 1041, + 1304, + 1060, + 1234, + 1060 + ], + "score": 0.73, + "latex": "7 8 . 4 \\pm 2 . 4" + }, + { + "category_id": 13, + "poly": [ + 1123, + 1038, + 1193, + 1038, + 1193, + 1060, + 1123, + 1060 + ], + "score": 0.54, + "latex": "8 9 . 1 \\pm 5 . 1" + }, + { + "category_id": 13, + "poly": [ + 1115, + 1071, + 1201, + 1071, + 1201, + 1096, + 1115, + 1096 + ], + "score": 0.5, + "latex": "9 0 . 7 2 _ { \\pm 5 . 1 5 }" + }, + { + "category_id": 13, + "poly": [ + 1380, + 1213, + 1397, + 1213, + 1397, + 1242, + 1380, + 1242 + ], + "score": 0.36, + "latex": "\\uparrow" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1461.0, + 733.0, + 1461.0, + 733.0, + 1505.0, + 292.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 1118.0, + 1407.0, + 1118.0, + 1407.0, + 1153.0, + 961.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1150.0, + 1408.0, + 1150.0, + 1408.0, + 1181.0, + 962.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1180.0, + 1405.0, + 1180.0, + 1405.0, + 1213.0, + 962.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 1209.0, + 1379.0, + 1209.0, + 1379.0, + 1244.0, + 961.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1209.0, + 1407.0, + 1209.0, + 1407.0, + 1244.0, + 1398.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1240.0, + 1404.0, + 1240.0, + 1404.0, + 1274.0, + 962.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 1271.0, + 1038.0, + 1271.0, + 1038.0, + 1302.0, + 959.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 663.0, + 1405.0, + 663.0, + 1405.0, + 696.0, + 295.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 695.0, + 1404.0, + 695.0, + 1404.0, + 727.0, + 296.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 722.0, + 1406.0, + 722.0, + 1406.0, + 760.0, + 292.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 752.0, + 1404.0, + 752.0, + 1404.0, + 787.0, + 294.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 785.0, + 1405.0, + 785.0, + 1405.0, + 821.0, + 294.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 812.0, + 1406.0, + 812.0, + 1406.0, + 853.0, + 292.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 843.0, + 1405.0, + 843.0, + 1405.0, + 879.0, + 294.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 872.0, + 1404.0, + 872.0, + 1404.0, + 911.0, + 294.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 905.0, + 984.0, + 905.0, + 984.0, + 941.0, + 295.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1525.0, + 1405.0, + 1525.0, + 1405.0, + 1562.0, + 294.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1556.0, + 1406.0, + 1556.0, + 1406.0, + 1592.0, + 292.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1589.0, + 1404.0, + 1589.0, + 1404.0, + 1622.0, + 292.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1618.0, + 1405.0, + 1618.0, + 1405.0, + 1653.0, + 295.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1647.0, + 1405.0, + 1647.0, + 1405.0, + 1682.0, + 294.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1677.0, + 1235.0, + 1677.0, + 1235.0, + 1714.0, + 294.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 948.0, + 945.0, + 948.0, + 945.0, + 986.0, + 296.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 981.0, + 944.0, + 981.0, + 944.0, + 1016.0, + 295.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1011.0, + 942.0, + 1011.0, + 942.0, + 1045.0, + 294.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1041.0, + 943.0, + 1041.0, + 943.0, + 1075.0, + 293.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1070.0, + 945.0, + 1070.0, + 945.0, + 1107.0, + 292.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1103.0, + 945.0, + 1103.0, + 945.0, + 1135.0, + 294.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1131.0, + 945.0, + 1131.0, + 945.0, + 1164.0, + 293.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1162.0, + 947.0, + 1162.0, + 947.0, + 1196.0, + 294.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1192.0, + 944.0, + 1192.0, + 944.0, + 1227.0, + 296.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1224.0, + 944.0, + 1224.0, + 944.0, + 1256.0, + 294.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1253.0, + 943.0, + 1253.0, + 943.0, + 1288.0, + 294.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1283.0, + 947.0, + 1283.0, + 947.0, + 1321.0, + 293.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 468.0, + 1403.0, + 468.0, + 1403.0, + 506.0, + 294.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 501.0, + 1194.0, + 501.0, + 1194.0, + 538.0, + 293.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 501.0, + 1405.0, + 501.0, + 1405.0, + 538.0, + 1260.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 531.0, + 1405.0, + 531.0, + 1405.0, + 567.0, + 291.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 560.0, + 1406.0, + 560.0, + 1406.0, + 597.0, + 292.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 592.0, + 1315.0, + 592.0, + 1315.0, + 629.0, + 296.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 335.0, + 1405.0, + 335.0, + 1405.0, + 367.0, + 296.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 365.0, + 1406.0, + 365.0, + 1406.0, + 400.0, + 292.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 397.0, + 1403.0, + 397.0, + 1403.0, + 430.0, + 296.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 426.0, + 880.0, + 426.0, + 880.0, + 459.0, + 293.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1745.0, + 1409.0, + 1745.0, + 1409.0, + 1784.0, + 293.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1780.0, + 1404.0, + 1780.0, + 1404.0, + 1813.0, + 296.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1807.0, + 1405.0, + 1807.0, + 1405.0, + 1844.0, + 294.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1839.0, + 1405.0, + 1839.0, + 1405.0, + 1876.0, + 296.0, + 1876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1868.0, + 999.0, + 1868.0, + 999.0, + 1905.0, + 293.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1912.0, + 1406.0, + 1912.0, + 1406.0, + 1948.0, + 295.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1946.0, + 1404.0, + 1946.0, + 1404.0, + 1980.0, + 296.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1976.0, + 787.0, + 1976.0, + 787.0, + 2010.0, + 295.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1313.0, + 1405.0, + 1313.0, + 1405.0, + 1351.0, + 294.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1342.0, + 1408.0, + 1342.0, + 1408.0, + 1382.0, + 293.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1376.0, + 1404.0, + 1376.0, + 1404.0, + 1410.0, + 295.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1403.0, + 498.0, + 1403.0, + 498.0, + 1441.0, + 293.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 200.0, + 1405.0, + 200.0, + 1405.0, + 241.0, + 292.0, + 241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 230.0, + 1406.0, + 230.0, + 1406.0, + 272.0, + 292.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 262.0, + 1068.0, + 262.0, + 1068.0, + 298.0, + 294.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 1118.0, + 1407.0, + 1118.0, + 1407.0, + 1153.0, + 961.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1150.0, + 1408.0, + 1150.0, + 1408.0, + 1181.0, + 962.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1180.0, + 1405.0, + 1180.0, + 1405.0, + 1213.0, + 962.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 1209.0, + 1379.0, + 1209.0, + 1379.0, + 1244.0, + 961.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1209.0, + 1407.0, + 1209.0, + 1407.0, + 1244.0, + 1398.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1240.0, + 1404.0, + 1240.0, + 1404.0, + 1274.0, + 962.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 1271.0, + 1038.0, + 1271.0, + 1038.0, + 1302.0, + 959.0, + 1302.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 267, + 1405, + 267, + 1405, + 450, + 297, + 450 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 293, + 549, + 1410, + 549, + 1410, + 2016, + 293, + 2016 + ], + "score": 0.925 + }, + { + "category_id": 0, + "poly": [ + 299, + 198, + 625, + 198, + 625, + 237, + 299, + 237 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 299, + 495, + 455, + 495, + 455, + 531, + 299, + 531 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 863, + 2061, + 863, + 2086, + 836, + 2086 + ], + "score": 0.637 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 863, + 2061, + 863, + 2086, + 836, + 2086 + ], + "score": 0.44 + }, + { + "category_id": 15, + "poly": [ + 292.0, + 195.0, + 629.0, + 195.0, + 629.0, + 243.0, + 292.0, + 243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 491.0, + 460.0, + 491.0, + 460.0, + 537.0, + 295.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 868.0, + 2058.0, + 868.0, + 2098.0, + 832.0, + 2098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 868.0, + 2058.0, + 868.0, + 2098.0, + 832.0, + 2098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 266.0, + 1405.0, + 266.0, + 1405.0, + 303.0, + 294.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 294.0, + 1405.0, + 294.0, + 1405.0, + 335.0, + 294.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 324.0, + 1406.0, + 324.0, + 1406.0, + 366.0, + 291.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 357.0, + 1405.0, + 357.0, + 1405.0, + 394.0, + 294.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 388.0, + 1405.0, + 388.0, + 1405.0, + 424.0, + 292.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 418.0, + 414.0, + 418.0, + 414.0, + 451.0, + 294.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 550.0, + 1409.0, + 550.0, + 1409.0, + 586.0, + 294.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 581.0, + 1404.0, + 581.0, + 1404.0, + 617.0, + 321.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 612.0, + 707.0, + 612.0, + 707.0, + 646.0, + 317.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 661.0, + 1407.0, + 661.0, + 1407.0, + 703.0, + 291.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 697.0, + 602.0, + 697.0, + 602.0, + 732.0, + 324.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 749.0, + 1407.0, + 749.0, + 1407.0, + 790.0, + 294.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 781.0, + 1230.0, + 781.0, + 1230.0, + 817.0, + 322.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 832.0, + 1407.0, + 832.0, + 1407.0, + 874.0, + 291.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 865.0, + 875.0, + 865.0, + 875.0, + 904.0, + 319.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 918.0, + 1404.0, + 918.0, + 1404.0, + 954.0, + 296.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 950.0, + 1410.0, + 950.0, + 1410.0, + 990.0, + 319.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 981.0, + 716.0, + 981.0, + 716.0, + 1021.0, + 319.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1033.0, + 1410.0, + 1033.0, + 1410.0, + 1074.0, + 294.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1060.0, + 1409.0, + 1060.0, + 1409.0, + 1110.0, + 317.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1095.0, + 1116.0, + 1095.0, + 1116.0, + 1136.0, + 321.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1150.0, + 1159.0, + 1150.0, + 1159.0, + 1191.0, + 293.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1201.0, + 1407.0, + 1201.0, + 1407.0, + 1245.0, + 291.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1235.0, + 989.0, + 1235.0, + 989.0, + 1274.0, + 319.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1286.0, + 1407.0, + 1286.0, + 1407.0, + 1332.0, + 293.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1318.0, + 1280.0, + 1318.0, + 1280.0, + 1362.0, + 317.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1371.0, + 1409.0, + 1371.0, + 1409.0, + 1415.0, + 293.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1405.0, + 1209.0, + 1405.0, + 1209.0, + 1444.0, + 319.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1459.0, + 1407.0, + 1459.0, + 1407.0, + 1495.0, + 294.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1488.0, + 1410.0, + 1488.0, + 1410.0, + 1531.0, + 317.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1519.0, + 397.0, + 1519.0, + 397.0, + 1558.0, + 321.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1573.0, + 1409.0, + 1573.0, + 1409.0, + 1616.0, + 293.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1604.0, + 779.0, + 1604.0, + 779.0, + 1645.0, + 319.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1661.0, + 1405.0, + 1661.0, + 1405.0, + 1697.0, + 296.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1691.0, + 942.0, + 1691.0, + 942.0, + 1730.0, + 319.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1746.0, + 1405.0, + 1746.0, + 1405.0, + 1782.0, + 294.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1777.0, + 1405.0, + 1777.0, + 1405.0, + 1813.0, + 321.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1807.0, + 709.0, + 1807.0, + 709.0, + 1842.0, + 317.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1852.0, + 1407.0, + 1852.0, + 1407.0, + 1904.0, + 289.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1890.0, + 1000.0, + 1890.0, + 1000.0, + 1931.0, + 321.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1947.0, + 1405.0, + 1947.0, + 1405.0, + 1983.0, + 296.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1975.0, + 1352.0, + 1975.0, + 1352.0, + 2014.0, + 319.0, + 2014.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 289, + 176, + 1411, + 176, + 1411, + 2029, + 289, + 2029 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 835, + 2060, + 866, + 2060, + 866, + 2087, + 835, + 2087 + ], + "score": 0.834 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 205.0, + 1407.0, + 205.0, + 1407.0, + 238.0, + 295.0, + 238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 234.0, + 1409.0, + 234.0, + 1409.0, + 272.0, + 319.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 262.0, + 398.0, + 262.0, + 398.0, + 303.0, + 319.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 311.0, + 1409.0, + 311.0, + 1409.0, + 358.0, + 291.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 348.0, + 782.0, + 348.0, + 782.0, + 386.0, + 317.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 399.0, + 1407.0, + 399.0, + 1407.0, + 435.0, + 291.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 425.0, + 1407.0, + 425.0, + 1407.0, + 472.0, + 317.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 464.0, + 484.0, + 464.0, + 484.0, + 496.0, + 323.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 510.0, + 1409.0, + 510.0, + 1409.0, + 551.0, + 289.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 539.0, + 1409.0, + 539.0, + 1409.0, + 586.0, + 315.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 574.0, + 613.0, + 574.0, + 613.0, + 612.0, + 321.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 626.0, + 1411.0, + 626.0, + 1411.0, + 665.0, + 293.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 659.0, + 414.0, + 659.0, + 414.0, + 694.0, + 321.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 710.0, + 1407.0, + 710.0, + 1407.0, + 749.0, + 293.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 742.0, + 611.0, + 742.0, + 611.0, + 775.0, + 321.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 793.0, + 1407.0, + 793.0, + 1407.0, + 832.0, + 293.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 828.0, + 689.0, + 828.0, + 689.0, + 860.0, + 323.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 877.0, + 1409.0, + 877.0, + 1409.0, + 915.0, + 293.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 907.0, + 1411.0, + 907.0, + 1411.0, + 946.0, + 321.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 938.0, + 776.0, + 938.0, + 776.0, + 976.0, + 321.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 991.0, + 1409.0, + 991.0, + 1409.0, + 1029.0, + 293.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1019.0, + 402.0, + 1019.0, + 402.0, + 1060.0, + 319.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1068.0, + 1407.0, + 1068.0, + 1407.0, + 1117.0, + 289.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1105.0, + 1168.0, + 1105.0, + 1168.0, + 1143.0, + 321.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1159.0, + 1407.0, + 1159.0, + 1407.0, + 1192.0, + 297.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1188.0, + 1405.0, + 1188.0, + 1405.0, + 1227.0, + 321.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1219.0, + 707.0, + 1219.0, + 707.0, + 1255.0, + 317.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1271.0, + 1407.0, + 1271.0, + 1407.0, + 1310.0, + 293.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1302.0, + 1405.0, + 1302.0, + 1405.0, + 1341.0, + 321.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1334.0, + 484.0, + 1334.0, + 484.0, + 1367.0, + 325.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1383.0, + 1409.0, + 1383.0, + 1409.0, + 1422.0, + 291.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1416.0, + 1405.0, + 1416.0, + 1405.0, + 1454.0, + 321.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1444.0, + 486.0, + 1444.0, + 486.0, + 1483.0, + 321.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1493.0, + 1409.0, + 1493.0, + 1409.0, + 1540.0, + 289.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1526.0, + 1361.0, + 1526.0, + 1361.0, + 1566.0, + 317.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1579.0, + 1409.0, + 1579.0, + 1409.0, + 1619.0, + 291.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1611.0, + 770.0, + 1611.0, + 770.0, + 1648.0, + 317.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1664.0, + 1407.0, + 1664.0, + 1407.0, + 1703.0, + 293.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1695.0, + 611.0, + 1695.0, + 611.0, + 1733.0, + 319.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1745.0, + 1409.0, + 1745.0, + 1409.0, + 1786.0, + 291.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1778.0, + 1407.0, + 1778.0, + 1407.0, + 1817.0, + 319.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1808.0, + 669.0, + 1808.0, + 669.0, + 1847.0, + 321.0, + 1847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1863.0, + 1405.0, + 1863.0, + 1405.0, + 1896.0, + 297.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1892.0, + 1007.0, + 1892.0, + 1007.0, + 1930.0, + 319.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1945.0, + 1405.0, + 1945.0, + 1405.0, + 1983.0, + 293.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1973.0, + 707.0, + 1973.0, + 707.0, + 2014.0, + 317.0, + 2014.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 835, + 2060, + 865, + 2060, + 865, + 2087, + 835, + 2087 + ], + "score": 0.83 + }, + { + "category_id": 1, + "poly": [ + 291, + 137, + 1410, + 137, + 1410, + 2021, + 291, + 2021 + ], + "score": 0.751 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 869.0, + 2058.0, + 869.0, + 2096.0, + 832.0, + 2096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 199.0, + 1411.0, + 199.0, + 1411.0, + 242.0, + 292.0, + 242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 234.0, + 1200.0, + 234.0, + 1200.0, + 273.0, + 319.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 283.0, + 1407.0, + 283.0, + 1407.0, + 325.0, + 292.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 321.0, + 1407.0, + 321.0, + 1407.0, + 354.0, + 321.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 345.0, + 400.0, + 345.0, + 400.0, + 387.0, + 319.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 395.0, + 1407.0, + 395.0, + 1407.0, + 442.0, + 292.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 430.0, + 886.0, + 430.0, + 886.0, + 467.0, + 317.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 486.0, + 1405.0, + 486.0, + 1405.0, + 519.0, + 294.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 515.0, + 1407.0, + 515.0, + 1407.0, + 554.0, + 319.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 546.0, + 708.0, + 546.0, + 708.0, + 583.0, + 317.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 595.0, + 1407.0, + 595.0, + 1407.0, + 639.0, + 290.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 626.0, + 1075.0, + 626.0, + 1075.0, + 668.0, + 317.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 682.0, + 1409.0, + 682.0, + 1409.0, + 721.0, + 292.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 713.0, + 778.0, + 713.0, + 778.0, + 750.0, + 315.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 763.0, + 1409.0, + 763.0, + 1409.0, + 806.0, + 292.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 794.0, + 1409.0, + 794.0, + 1409.0, + 835.0, + 315.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 820.0, + 413.0, + 820.0, + 413.0, + 867.0, + 319.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 876.0, + 1409.0, + 876.0, + 1409.0, + 920.0, + 290.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 910.0, + 1071.0, + 910.0, + 1071.0, + 947.0, + 315.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 963.0, + 1409.0, + 963.0, + 1409.0, + 1003.0, + 294.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 994.0, + 819.0, + 994.0, + 819.0, + 1031.0, + 317.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1044.0, + 1407.0, + 1044.0, + 1407.0, + 1085.0, + 290.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1077.0, + 1407.0, + 1077.0, + 1407.0, + 1116.0, + 319.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1108.0, + 734.0, + 1108.0, + 734.0, + 1147.0, + 319.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1158.0, + 1409.0, + 1158.0, + 1409.0, + 1199.0, + 292.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1189.0, + 1377.0, + 1189.0, + 1377.0, + 1228.0, + 319.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1242.0, + 1407.0, + 1242.0, + 1407.0, + 1284.0, + 292.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1275.0, + 973.0, + 1275.0, + 973.0, + 1312.0, + 317.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1331.0, + 1405.0, + 1331.0, + 1405.0, + 1364.0, + 296.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1360.0, + 1409.0, + 1360.0, + 1409.0, + 1397.0, + 317.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1387.0, + 1030.0, + 1387.0, + 1030.0, + 1428.0, + 319.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1439.0, + 1407.0, + 1439.0, + 1407.0, + 1480.0, + 292.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1470.0, + 1411.0, + 1470.0, + 1411.0, + 1513.0, + 317.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1501.0, + 400.0, + 1501.0, + 400.0, + 1542.0, + 319.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1554.0, + 1407.0, + 1554.0, + 1407.0, + 1596.0, + 292.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1587.0, + 859.0, + 1587.0, + 859.0, + 1620.0, + 319.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1643.0, + 1405.0, + 1643.0, + 1405.0, + 1676.0, + 296.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1670.0, + 1411.0, + 1670.0, + 1411.0, + 1709.0, + 319.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1726.0, + 1407.0, + 1726.0, + 1407.0, + 1759.0, + 296.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1771.0, + 1411.0, + 1771.0, + 1411.0, + 1819.0, + 290.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1811.0, + 734.0, + 1811.0, + 734.0, + 1844.0, + 321.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1860.0, + 1407.0, + 1860.0, + 1407.0, + 1899.0, + 294.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1891.0, + 1151.0, + 1891.0, + 1151.0, + 1930.0, + 319.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1945.0, + 1409.0, + 1945.0, + 1409.0, + 1984.0, + 294.0, + 1984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1976.0, + 1238.0, + 1976.0, + 1238.0, + 2015.0, + 319.0, + 2015.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 407, + 1727, + 1404, + 1727, + 1404, + 1926, + 407, + 1926 + ], + "score": 0.956 + }, + { + "category_id": 0, + "poly": [ + 299, + 1627, + 434, + 1627, + 434, + 1663, + 299, + 1663 + ], + "score": 0.904 + }, + { + "category_id": 1, + "poly": [ + 364, + 1935, + 840, + 1935, + 840, + 1967, + 364, + 1967 + ], + "score": 0.9 + }, + { + "category_id": 1, + "poly": [ + 406, + 1975, + 1256, + 1975, + 1256, + 2008, + 406, + 2008 + ], + "score": 0.876 + }, + { + "category_id": 1, + "poly": [ + 365, + 1686, + 580, + 1686, + 580, + 1717, + 365, + 1717 + ], + "score": 0.863 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2085, + 836, + 2085 + ], + "score": 0.854 + }, + { + "category_id": 1, + "poly": [ + 293, + 201, + 1409, + 201, + 1409, + 1580, + 293, + 1580 + ], + "score": 0.783 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1625.0, + 440.0, + 1625.0, + 440.0, + 1666.0, + 296.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2057.0, + 871.0, + 2057.0, + 871.0, + 2095.0, + 832.0, + 2095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1723.0, + 1405.0, + 1723.0, + 1405.0, + 1762.0, + 404.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1756.0, + 802.0, + 1756.0, + 802.0, + 1793.0, + 450.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 1793.0, + 1045.0, + 1793.0, + 1045.0, + 1827.0, + 405.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 1827.0, + 1300.0, + 1827.0, + 1300.0, + 1863.0, + 405.0, + 1863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 1860.0, + 1405.0, + 1860.0, + 1405.0, + 1899.0, + 405.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1892.0, + 596.0, + 1892.0, + 596.0, + 1929.0, + 448.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1932.0, + 846.0, + 1932.0, + 846.0, + 1972.0, + 359.0, + 1972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1973.0, + 1260.0, + 1973.0, + 1260.0, + 2012.0, + 404.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1681.0, + 586.0, + 1681.0, + 586.0, + 1721.0, + 361.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 205.0, + 1406.0, + 205.0, + 1406.0, + 239.0, + 296.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 234.0, + 1082.0, + 234.0, + 1082.0, + 268.0, + 322.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 283.0, + 1407.0, + 283.0, + 1407.0, + 324.0, + 292.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 316.0, + 1406.0, + 316.0, + 1406.0, + 354.0, + 322.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 345.0, + 804.0, + 345.0, + 804.0, + 384.0, + 322.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 396.0, + 1407.0, + 396.0, + 1407.0, + 434.0, + 295.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 427.0, + 1407.0, + 427.0, + 1407.0, + 468.0, + 317.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 458.0, + 614.0, + 458.0, + 614.0, + 493.0, + 319.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 507.0, + 1409.0, + 507.0, + 1409.0, + 547.0, + 292.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 538.0, + 1407.0, + 538.0, + 1407.0, + 578.0, + 319.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 572.0, + 611.0, + 572.0, + 611.0, + 605.0, + 322.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 619.0, + 1409.0, + 619.0, + 1409.0, + 661.0, + 293.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 650.0, + 570.0, + 650.0, + 570.0, + 689.0, + 321.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 701.0, + 1406.0, + 701.0, + 1406.0, + 741.0, + 292.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 733.0, + 1167.0, + 733.0, + 1167.0, + 770.0, + 316.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 784.0, + 1409.0, + 784.0, + 1409.0, + 824.0, + 295.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 813.0, + 1409.0, + 813.0, + 1409.0, + 854.0, + 318.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 842.0, + 397.0, + 842.0, + 397.0, + 885.0, + 320.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 897.0, + 1406.0, + 897.0, + 1406.0, + 935.0, + 295.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 929.0, + 845.0, + 929.0, + 845.0, + 963.0, + 322.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 980.0, + 1406.0, + 980.0, + 1406.0, + 1018.0, + 295.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1009.0, + 1409.0, + 1009.0, + 1409.0, + 1048.0, + 321.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1038.0, + 926.0, + 1038.0, + 926.0, + 1078.0, + 321.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1089.0, + 1404.0, + 1089.0, + 1404.0, + 1128.0, + 295.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1123.0, + 1407.0, + 1123.0, + 1407.0, + 1162.0, + 321.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1152.0, + 851.0, + 1152.0, + 851.0, + 1191.0, + 321.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1203.0, + 1406.0, + 1203.0, + 1406.0, + 1242.0, + 295.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1231.0, + 1407.0, + 1231.0, + 1407.0, + 1277.0, + 318.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1265.0, + 674.0, + 1265.0, + 674.0, + 1300.0, + 319.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1314.0, + 1406.0, + 1314.0, + 1406.0, + 1356.0, + 293.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1347.0, + 1284.0, + 1347.0, + 1284.0, + 1385.0, + 322.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1394.0, + 1410.0, + 1394.0, + 1410.0, + 1441.0, + 290.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1425.0, + 1409.0, + 1425.0, + 1409.0, + 1470.0, + 318.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1456.0, + 948.0, + 1456.0, + 948.0, + 1496.0, + 319.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1510.0, + 1406.0, + 1510.0, + 1406.0, + 1548.0, + 292.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1542.0, + 1237.0, + 1542.0, + 1237.0, + 1578.0, + 319.0, + 1578.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 407, + 286, + 1405, + 286, + 1405, + 547, + 407, + 547 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 406, + 975, + 1405, + 975, + 1405, + 1171, + 406, + 1171 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 407, + 597, + 1403, + 597, + 1403, + 924, + 407, + 924 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 363, + 246, + 671, + 246, + 671, + 277, + 363, + 277 + ], + "score": 0.909 + }, + { + "category_id": 1, + "poly": [ + 404, + 203, + 1177, + 203, + 1177, + 236, + 404, + 236 + ], + "score": 0.893 + }, + { + "category_id": 1, + "poly": [ + 366, + 933, + 1200, + 933, + 1200, + 966, + 366, + 966 + ], + "score": 0.868 + }, + { + "category_id": 1, + "poly": [ + 371, + 556, + 1402, + 556, + 1402, + 590, + 371, + 590 + ], + "score": 0.856 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2085, + 836, + 2085 + ], + "score": 0.828 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 283.0, + 1409.0, + 283.0, + 1409.0, + 321.0, + 404.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 315.0, + 1242.0, + 315.0, + 1242.0, + 350.0, + 450.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 351.0, + 1405.0, + 351.0, + 1405.0, + 388.0, + 405.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 383.0, + 684.0, + 383.0, + 684.0, + 415.0, + 450.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 421.0, + 446.0, + 421.0, + 446.0, + 451.0, + 407.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 419.0, + 1408.0, + 419.0, + 1408.0, + 454.0, + 447.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 450.0, + 954.0, + 450.0, + 954.0, + 481.0, + 450.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 480.0, + 1405.0, + 480.0, + 1405.0, + 523.0, + 408.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 515.0, + 1030.0, + 515.0, + 1030.0, + 550.0, + 451.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 972.0, + 1407.0, + 972.0, + 1407.0, + 1011.0, + 404.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1005.0, + 661.0, + 1005.0, + 661.0, + 1039.0, + 449.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 1041.0, + 1403.0, + 1041.0, + 1403.0, + 1076.0, + 405.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1071.0, + 951.0, + 1071.0, + 951.0, + 1107.0, + 449.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 1105.0, + 1403.0, + 1105.0, + 1403.0, + 1141.0, + 409.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1138.0, + 926.0, + 1138.0, + 926.0, + 1174.0, + 450.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 596.0, + 1180.0, + 596.0, + 1180.0, + 634.0, + 404.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 633.0, + 1406.0, + 633.0, + 1406.0, + 670.0, + 403.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 667.0, + 1405.0, + 667.0, + 1405.0, + 705.0, + 407.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 697.0, + 1048.0, + 697.0, + 1048.0, + 735.0, + 451.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 734.0, + 1404.0, + 734.0, + 1404.0, + 771.0, + 407.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 766.0, + 1405.0, + 766.0, + 1405.0, + 801.0, + 449.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 795.0, + 978.0, + 795.0, + 978.0, + 831.0, + 448.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 831.0, + 1404.0, + 831.0, + 1404.0, + 866.0, + 407.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 858.0, + 1404.0, + 858.0, + 1404.0, + 898.0, + 450.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 893.0, + 809.0, + 893.0, + 809.0, + 926.0, + 450.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 240.0, + 675.0, + 240.0, + 675.0, + 282.0, + 358.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 200.0, + 1180.0, + 200.0, + 1180.0, + 240.0, + 403.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 929.0, + 1205.0, + 929.0, + 1205.0, + 972.0, + 357.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 552.0, + 1410.0, + 552.0, + 1410.0, + 596.0, + 360.0, + 596.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/vlm/dev/3uwj8QZROL/0.png b/vlm/dev/3uwj8QZROL/0.png new file mode 100644 index 0000000000000000000000000000000000000000..db011754cf4603519bd101ea8bd499251e0d7ed3 --- /dev/null +++ b/vlm/dev/3uwj8QZROL/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e7736d63f2dea7ab613ccf69088f1c2be07c84740ac3bcdfc05fdffc82c3291 +size 606917 diff --git a/vlm/dev/3uwj8QZROL/1.png b/vlm/dev/3uwj8QZROL/1.png new file mode 100644 index 0000000000000000000000000000000000000000..fa16e8244d2d188bc3a93b434e0530d91af50dbd --- /dev/null +++ b/vlm/dev/3uwj8QZROL/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:131e4ca928af675e361777a11e5e718565d3ad449277af3591a1279632040858 +size 673418 diff --git a/vlm/dev/3uwj8QZROL/10.png b/vlm/dev/3uwj8QZROL/10.png new file mode 100644 index 0000000000000000000000000000000000000000..b1def2123b952b72898748638f33f0005b254539 --- /dev/null +++ b/vlm/dev/3uwj8QZROL/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd34b7d2dbe26a7e50401a64c349e1b7d4736317368616e75e8e3c1045b19d7c +size 584303 diff --git a/vlm/dev/3uwj8QZROL/11.png b/vlm/dev/3uwj8QZROL/11.png new file mode 100644 index 0000000000000000000000000000000000000000..475a0a4f53b86ca09f4865b3c29438697710c12c --- /dev/null +++ b/vlm/dev/3uwj8QZROL/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db7cd6976a5f32b5c40174f53af4fae773fcef024267ae8569b76863cb457445 +size 354398 diff --git a/vlm/dev/3uwj8QZROL/2.png b/vlm/dev/3uwj8QZROL/2.png new file mode 100644 index 0000000000000000000000000000000000000000..4fb7568af829ff5e9914bdf879c8fdad15c23717 --- /dev/null +++ b/vlm/dev/3uwj8QZROL/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b457e64fd602e8e7b2040cffd8c2510021e75f059d1c59d46e1f8c287926431a +size 796419 diff --git a/vlm/dev/3uwj8QZROL/3.png b/vlm/dev/3uwj8QZROL/3.png new file mode 100644 index 0000000000000000000000000000000000000000..e97721b568ae9eecf984cc642c2d5447454c142a --- /dev/null +++ b/vlm/dev/3uwj8QZROL/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6c944bc80f9882cad47cbd03ee4df7d51a4795c6013e97a49d02df7a7d96534 +size 732558 diff --git a/vlm/dev/3uwj8QZROL/4.png b/vlm/dev/3uwj8QZROL/4.png new file mode 100644 index 0000000000000000000000000000000000000000..b01538e86ab0bdde0f211ca8074a5c02ee2a6e18 --- /dev/null +++ b/vlm/dev/3uwj8QZROL/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17d172a6d7b0e67b7ca842d2a13f41e365bb9e7578fb8e0874b3e2151a824129 +size 712723 diff --git a/vlm/dev/3uwj8QZROL/5.png b/vlm/dev/3uwj8QZROL/5.png new file mode 100644 index 0000000000000000000000000000000000000000..cd381d2f41aa87ebb777cc97b4286dc6d2c9c41f --- /dev/null +++ b/vlm/dev/3uwj8QZROL/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:368361a3861ffd28735f6d971bad5e225cd01ccd6189dcd6909e3c40faea7afe +size 994496 diff --git a/vlm/dev/3uwj8QZROL/6.png b/vlm/dev/3uwj8QZROL/6.png new file mode 100644 index 0000000000000000000000000000000000000000..bbaea99755519870f7d455cefd0b32b5868f2f80 --- /dev/null +++ b/vlm/dev/3uwj8QZROL/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30aa0d373c58f09b479b22354d9255811f413be5cc273326d573862b5b018958 +size 687425 diff --git a/vlm/dev/3uwj8QZROL/7.png b/vlm/dev/3uwj8QZROL/7.png new file mode 100644 index 0000000000000000000000000000000000000000..1929960b40471903d1d70161a87a4386d4f884e3 --- /dev/null +++ b/vlm/dev/3uwj8QZROL/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:108ba4c3b2a97e3266deba05e4a032a9acc46badcfd1e85f0de40363904e2d62 +size 631604 diff --git a/vlm/dev/3uwj8QZROL/8.png b/vlm/dev/3uwj8QZROL/8.png new file mode 100644 index 0000000000000000000000000000000000000000..8e2e028c16ca4518d58f1a1b1502a2708295598d --- /dev/null +++ b/vlm/dev/3uwj8QZROL/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4d6e045926c9d88cb31eb01f61b51265345ab90cfb99887a8bc08582f6a9f2a +size 587690 diff --git a/vlm/dev/3uwj8QZROL/9.png b/vlm/dev/3uwj8QZROL/9.png new file mode 100644 index 0000000000000000000000000000000000000000..4503732337d241e44a4c937edff01019fa639113 --- /dev/null +++ b/vlm/dev/3uwj8QZROL/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bbe3205995a8dd3c213b63e4073477511e49f3afc2d0feca53b3612316a06a6 +size 563755 diff --git a/vlm/dev/5hLP5JY9S2d/0.png b/vlm/dev/5hLP5JY9S2d/0.png new file mode 100644 index 0000000000000000000000000000000000000000..1bdd62e9667700e042f021d3c8cf63dd6e6d88a7 --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59233213ff7c696741ce0e0cfce87078fd04b7ea2ffbed18094b970b2c929c4d +size 511593 diff --git a/vlm/dev/5hLP5JY9S2d/1.png b/vlm/dev/5hLP5JY9S2d/1.png new file mode 100644 index 0000000000000000000000000000000000000000..c118ccd24de546ad5e8ff3cb08f29daf3a272e53 --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:253df4d666647ebd55ce584a0069e4c169eb1460d91f2dcf6bf118bcead8e16c +size 762890 diff --git a/vlm/dev/5hLP5JY9S2d/10.png b/vlm/dev/5hLP5JY9S2d/10.png new file mode 100644 index 0000000000000000000000000000000000000000..b9fbbd73f73be185704f4b1ba67dfb0347f7de34 --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48079dfdcfdb88f400fe2664cdd478960db0fec4cf756b409bb431e6c2a3fecd +size 475779 diff --git a/vlm/dev/5hLP5JY9S2d/11.png b/vlm/dev/5hLP5JY9S2d/11.png new file mode 100644 index 0000000000000000000000000000000000000000..5b7a89d83013f20b70cc52cc331a6af7c56f17fc --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b652502e61db12cd19c88ae39e2f7da31158789fc3ff9f43b6f3358e4b8ae382 +size 473067 diff --git a/vlm/dev/5hLP5JY9S2d/12.png b/vlm/dev/5hLP5JY9S2d/12.png new file mode 100644 index 0000000000000000000000000000000000000000..264ff809d48107777bf3f4659d7c4c5b441b0e5f --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:afbe210e2490fa7a8bdc8578b45e0815230f8ed3771d03a444463b60b6118933 +size 522386 diff --git a/vlm/dev/5hLP5JY9S2d/13.png b/vlm/dev/5hLP5JY9S2d/13.png new file mode 100644 index 0000000000000000000000000000000000000000..077b2660c2ec9abbe67f02f027ea5e447b9a290a --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ef949e935f0d325ab92f84fcbab9e657d746bb1053c794148522532a739feb9 +size 64790 diff --git a/vlm/dev/5hLP5JY9S2d/14.png b/vlm/dev/5hLP5JY9S2d/14.png new file mode 100644 index 0000000000000000000000000000000000000000..7f3ffd06932d53411b8099805c1e2f034d7a53a6 --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad01758a811515050193308728d4a01bce0f80ff31371e981a18beb39a00847f +size 472626 diff --git a/vlm/dev/5hLP5JY9S2d/15.png b/vlm/dev/5hLP5JY9S2d/15.png new file mode 100644 index 0000000000000000000000000000000000000000..43a7bf6d1e31796f92354b42985f03f40b192b22 --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c345b36c49137fb157f517dc322d5db83e74db835a63ac1054ae4a3d61224c2b +size 681521 diff --git a/vlm/dev/5hLP5JY9S2d/16.png b/vlm/dev/5hLP5JY9S2d/16.png new file mode 100644 index 0000000000000000000000000000000000000000..95d0150876aa0b5da78afdf57134af3ac8270446 --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3a5e476fa9ec967467121c5e2dbc2cba35b3847fb98ebba487d22beab511505 +size 585046 diff --git a/vlm/dev/5hLP5JY9S2d/17.png b/vlm/dev/5hLP5JY9S2d/17.png new file mode 100644 index 0000000000000000000000000000000000000000..5f3a6eebe2480313143e797c4cc6879684eb2f8b --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c07b25853cb123ec7ac4d54fb3e63b0f937eb9f1e6dfb3307d62b505186888ff +size 565635 diff --git a/vlm/dev/5hLP5JY9S2d/18.png b/vlm/dev/5hLP5JY9S2d/18.png new file mode 100644 index 0000000000000000000000000000000000000000..f3079bd0b9301b4d0f3a8190e422ae4bab9aed4c --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9286ac4cdc1fa06b91b71082df9431313b5010711f881990a87133190ae39262 +size 617120 diff --git a/vlm/dev/5hLP5JY9S2d/19.png b/vlm/dev/5hLP5JY9S2d/19.png new file mode 100644 index 0000000000000000000000000000000000000000..ac699fa08e21017be672e71265d3b90aea146fec --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84de31a6e46daff6757e2a398b0c19a7fa691a53aa0dabff14b48415ed39b5fd +size 525204 diff --git a/vlm/dev/5hLP5JY9S2d/2.png b/vlm/dev/5hLP5JY9S2d/2.png new file mode 100644 index 0000000000000000000000000000000000000000..5ccac46651f354c00e82de0457f4be29b601401c --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42569c53370d6dec5c860c41e011ba394e58bcc4974b3255e4c3404fdb623c3b +size 669006 diff --git a/vlm/dev/5hLP5JY9S2d/20.png b/vlm/dev/5hLP5JY9S2d/20.png new file mode 100644 index 0000000000000000000000000000000000000000..88c24bc3021e25aa7491355f6055d4eea48fbd98 --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c437a83f5c5b3f147360ac624049632f828d4bd98b24da96da81de0094e7404 +size 556313 diff --git a/vlm/dev/5hLP5JY9S2d/21.png b/vlm/dev/5hLP5JY9S2d/21.png new file mode 100644 index 0000000000000000000000000000000000000000..4a45154255a7e87db1d6bcd04992005f1afe405f --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f14892319aff9e304e83a29fd7109f96c14ad146e557e87a9ab63115470b4f7d +size 578902 diff --git a/vlm/dev/5hLP5JY9S2d/22.png b/vlm/dev/5hLP5JY9S2d/22.png new file mode 100644 index 0000000000000000000000000000000000000000..689613ba90568f2bb356b60f7aa45b7cec38ddbf --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7aacdaeb8f48b8537a79205cdda24f4520fc52d01e7f66e4d6b2afa6ed736fb7 +size 1549937 diff --git a/vlm/dev/5hLP5JY9S2d/23.png b/vlm/dev/5hLP5JY9S2d/23.png new file mode 100644 index 0000000000000000000000000000000000000000..6edd1c503ea936dfee5e1488053e1412fa6aec75 --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:254cd5e09d9b024645b976cb0d3ae4309c65f9b22b7b7b6c4c5884f9603dd373 +size 1539595 diff --git a/vlm/dev/5hLP5JY9S2d/24.png b/vlm/dev/5hLP5JY9S2d/24.png new file mode 100644 index 0000000000000000000000000000000000000000..aa7769524b825c17acf7f7cc40db11d34d814311 --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:341780fc9094e592bac5cebe066cb72a8e858c1a684d6b5bbb557649a4ddd697 +size 1216472 diff --git a/vlm/dev/5hLP5JY9S2d/25.png b/vlm/dev/5hLP5JY9S2d/25.png new file mode 100644 index 0000000000000000000000000000000000000000..26dbcbe49c7a39ad265171ea7d31efab030bbbc4 --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc5a7214a960b899fecd25b40a8f2e8cea1ce98ef85701dc9cea38ca68a41901 +size 134807 diff --git a/vlm/dev/5hLP5JY9S2d/3.png b/vlm/dev/5hLP5JY9S2d/3.png new file mode 100644 index 0000000000000000000000000000000000000000..eecf53cdf25d9745a2fdddeaa0ec3c0e22d34a77 --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6a9fa14c6b41308e7aef6c1f0bca7feb51c14aee84d2c80844e5892d06d0653 +size 684514 diff --git a/vlm/dev/5hLP5JY9S2d/4.png b/vlm/dev/5hLP5JY9S2d/4.png new file mode 100644 index 0000000000000000000000000000000000000000..f5d1f6c8c852877e94371178c1e481a45143f4a5 --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:583d1b469f7c6e93b0f481dba736c1596bc261b8a2118549700146f03261fa37 +size 565393 diff --git a/vlm/dev/5hLP5JY9S2d/5.png b/vlm/dev/5hLP5JY9S2d/5.png new file mode 100644 index 0000000000000000000000000000000000000000..dc9df710a0d22a6f5580852c7cc7aaaab2975366 --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be824383b2821c0fe68bb8f7a8ab17f5bf84177bd8f4088b274c9b9035c26882 +size 634158 diff --git a/vlm/dev/5hLP5JY9S2d/6.png b/vlm/dev/5hLP5JY9S2d/6.png new file mode 100644 index 0000000000000000000000000000000000000000..0aae7a26585368ed366ad2dbd4b5c6ac97858445 --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0ec92e3cd936f4dfe61b25712ffd26a1a6e260d05f8de00a4967beb0f48f592 +size 632118 diff --git a/vlm/dev/5hLP5JY9S2d/7.png b/vlm/dev/5hLP5JY9S2d/7.png new file mode 100644 index 0000000000000000000000000000000000000000..24e7adc4f6248d06aaa010ebfab2c1d70cbd7a3b --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:626d17b78e5afa8b8c22d067c83cced7d2fc3a7e0d488db3a32d7684e4328d33 +size 773903 diff --git a/vlm/dev/5hLP5JY9S2d/8.png b/vlm/dev/5hLP5JY9S2d/8.png new file mode 100644 index 0000000000000000000000000000000000000000..92361acfbc90775987ea908b4be9018afe882e08 --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71283071bbf38a4f50153e98668ba5c34c62df80b0bc9c83a1c4be617d808e94 +size 601347 diff --git a/vlm/dev/5hLP5JY9S2d/9.png b/vlm/dev/5hLP5JY9S2d/9.png new file mode 100644 index 0000000000000000000000000000000000000000..d8aac943feeb72895cea9cb25fed6cd0668ae9c8 --- /dev/null +++ b/vlm/dev/5hLP5JY9S2d/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d23a1583c82a307446de0d8a9b9910a8d774eb90ab70b5509ba4642f679d8bc +size 467153 diff --git a/vlm/dev/68n2s9ZJWF8/0.png b/vlm/dev/68n2s9ZJWF8/0.png new file mode 100644 index 0000000000000000000000000000000000000000..04b8db5b35a45a1e347df086b068524eceacb594 --- /dev/null +++ b/vlm/dev/68n2s9ZJWF8/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14b4ecaa7330c58730f5d5cf2491a778b19586859a42595343d1ece6293c3fe5 +size 488463 diff --git a/vlm/dev/68n2s9ZJWF8/1.png b/vlm/dev/68n2s9ZJWF8/1.png new file mode 100644 index 0000000000000000000000000000000000000000..d42bd3d000180cfcaf31833e9de8132ada495d5d --- /dev/null +++ b/vlm/dev/68n2s9ZJWF8/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73ebc3e6dfc9c278b1af90ca85532bbe0ffa7652118f51427999c049278af319 +size 621642 diff --git a/vlm/dev/68n2s9ZJWF8/10.png b/vlm/dev/68n2s9ZJWF8/10.png new file mode 100644 index 0000000000000000000000000000000000000000..95eb5667a5163ac562be6dd9628302f80c07ca6d --- /dev/null +++ b/vlm/dev/68n2s9ZJWF8/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19107a26f300d5f700f9dacebdc9c75f21adde69824950cb84e4b6a50a560070 +size 504349 diff --git a/vlm/dev/68n2s9ZJWF8/2.png b/vlm/dev/68n2s9ZJWF8/2.png new file mode 100644 index 0000000000000000000000000000000000000000..835c56475a09e61e2ff16c5974742d675acf4136 --- /dev/null +++ b/vlm/dev/68n2s9ZJWF8/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17838a0e272556f431070577c771008ed6a84f3ef69574d2c02195c99d44d68f +size 561737 diff --git a/vlm/dev/68n2s9ZJWF8/3.png b/vlm/dev/68n2s9ZJWF8/3.png new file mode 100644 index 0000000000000000000000000000000000000000..83052129f17b6f019db9a031864fc3278917ec36 --- /dev/null +++ b/vlm/dev/68n2s9ZJWF8/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10d41ca6b9a91182e8234238580a1b0db7d7e4d92b7a9f579c8468c81763dfa7 +size 549115 diff --git a/vlm/dev/68n2s9ZJWF8/4.png b/vlm/dev/68n2s9ZJWF8/4.png new file mode 100644 index 0000000000000000000000000000000000000000..beba5b8a2a8b303c8b8e37fd5e1751bc38fc39a8 --- /dev/null +++ b/vlm/dev/68n2s9ZJWF8/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7be9f12a93e7e996b78229783603428a92f37aa1d96490d3981479c5fee20b1c +size 534162 diff --git a/vlm/dev/68n2s9ZJWF8/5.png b/vlm/dev/68n2s9ZJWF8/5.png new file mode 100644 index 0000000000000000000000000000000000000000..76acf45da1fe4188ef5a5695cb9725142b722287 --- /dev/null +++ b/vlm/dev/68n2s9ZJWF8/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adaf3290c602be8a8ae0dde694b0f156de53fea3e7539ae3fa7236196a1870b0 +size 476186 diff --git a/vlm/dev/68n2s9ZJWF8/6.png b/vlm/dev/68n2s9ZJWF8/6.png new file mode 100644 index 0000000000000000000000000000000000000000..3d67a5d66a868069a5d72959178a211a7ab20619 --- /dev/null +++ b/vlm/dev/68n2s9ZJWF8/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2556d7b66ffe65c34cc7da6c16f2e7dc7c5cbc9bffcf1bc1b87fd9f628cf048 +size 547824 diff --git a/vlm/dev/68n2s9ZJWF8/7.png b/vlm/dev/68n2s9ZJWF8/7.png new file mode 100644 index 0000000000000000000000000000000000000000..08658354f85310c61476faa3a3d610a8d0cd8b10 --- /dev/null +++ b/vlm/dev/68n2s9ZJWF8/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5489f15efc97e0ecc56c9ca77d7cd458bec31ea22e92458369e1cdf715b2a766 +size 659093 diff --git a/vlm/dev/68n2s9ZJWF8/8.png b/vlm/dev/68n2s9ZJWF8/8.png new file mode 100644 index 0000000000000000000000000000000000000000..720228c05e2c25e56c901e4cc8f5c40f317657bf --- /dev/null +++ b/vlm/dev/68n2s9ZJWF8/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:839048d89a6668be8006367eb9235bf4aaa94ae4f5b15edf6aa4bfffe326dafd +size 639127 diff --git a/vlm/dev/68n2s9ZJWF8/9.png b/vlm/dev/68n2s9ZJWF8/9.png new file mode 100644 index 0000000000000000000000000000000000000000..fa933cb60620b2788045deb7f5a88949b348ad28 --- /dev/null +++ b/vlm/dev/68n2s9ZJWF8/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12f530eb51ebd75c0bff2679e1c2bc1c5aab2160b62d4d82f08c75a411daba91 +size 509943 diff --git a/vlm/dev/8aHzds2uUyB/0.png b/vlm/dev/8aHzds2uUyB/0.png new file mode 100644 index 0000000000000000000000000000000000000000..7bd4f8cdf68ef11b5a8fbacafc0ceb71ae0f4a7e --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93aa717b65b1b07dc6167012cd43dea47c620321111e9d6523112b01c58d26c7 +size 496839 diff --git a/vlm/dev/8aHzds2uUyB/1.png b/vlm/dev/8aHzds2uUyB/1.png new file mode 100644 index 0000000000000000000000000000000000000000..812b4cc665169ba1e49041c36d684fa73939d5be --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd6c9776458105b4fbaf2de67f793a78b0099b61f03acb1b05d26233d9be267a +size 615089 diff --git a/vlm/dev/8aHzds2uUyB/10.png b/vlm/dev/8aHzds2uUyB/10.png new file mode 100644 index 0000000000000000000000000000000000000000..5bc7fafacb8615489d7989ba7073d28a1e98717d --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c14dc57e83c5d1855f1c58af81c70e3ff40820d9bc57e8d9073af3720abf0cc8 +size 542374 diff --git a/vlm/dev/8aHzds2uUyB/11.png b/vlm/dev/8aHzds2uUyB/11.png new file mode 100644 index 0000000000000000000000000000000000000000..8b0bb6d6de96dbd922644292c0b8da15895ccfe5 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca49dd754305877aa0e0002a3bbbf4d18858c0c0fd5b47e744f3d75d7edc15ae +size 612112 diff --git a/vlm/dev/8aHzds2uUyB/12.png b/vlm/dev/8aHzds2uUyB/12.png new file mode 100644 index 0000000000000000000000000000000000000000..5ea6c787f06ec1631bfa8ae6327a4b615a1615f2 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6a99dc86895e2f16c0f3bd74347df88eb1f781fbbe6fccc75b57c677cead66b +size 589536 diff --git a/vlm/dev/8aHzds2uUyB/14.png b/vlm/dev/8aHzds2uUyB/14.png new file mode 100644 index 0000000000000000000000000000000000000000..2ed8c1ad4bfa9cd04979153dd4ba3ce52b2af47c --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe5279f14020d46150ba493244f3b57ca04d8b36edff10cccd36fceafb36e915 +size 346456 diff --git a/vlm/dev/8aHzds2uUyB/15.png b/vlm/dev/8aHzds2uUyB/15.png new file mode 100644 index 0000000000000000000000000000000000000000..2c0d045ff95ee24bf8f571c0942d4573b0ba6b1f --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f9a45670d142715788b6294e3856051b638aeb7cf4a33194e1cf74d4cf9d65e +size 245308 diff --git a/vlm/dev/8aHzds2uUyB/17.png b/vlm/dev/8aHzds2uUyB/17.png new file mode 100644 index 0000000000000000000000000000000000000000..34607a01ce174908ece3956448832ed5117c808d --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1119d062824f0843476b6a28506a5f1a50bfadd77258e98ae9d634a16625132 +size 271770 diff --git a/vlm/dev/8aHzds2uUyB/2.png b/vlm/dev/8aHzds2uUyB/2.png new file mode 100644 index 0000000000000000000000000000000000000000..a6b833efd43bd37e7b8ab8d85bbbc687e97ab5f0 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b79c23a3245a6668725d23fd75dbfe0519f9724d5f23b0fd2abe7864a84f72f +size 649966 diff --git a/vlm/dev/8aHzds2uUyB/20.png b/vlm/dev/8aHzds2uUyB/20.png new file mode 100644 index 0000000000000000000000000000000000000000..67ff4b56e7bd2f61ea399956670b47cd39d48796 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6cc8da58eec9bb27d89545c3ad8b976a2b34bbc44824a5b2d58579415588264e +size 544959 diff --git a/vlm/dev/8aHzds2uUyB/21.png b/vlm/dev/8aHzds2uUyB/21.png new file mode 100644 index 0000000000000000000000000000000000000000..3a93129352ee26f3d31215c64b568aaeb1974e62 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba3a3e69d5c607882e05b3c919b3b68925c8271695653b0c44e87cdaae0cc506 +size 528658 diff --git a/vlm/dev/8aHzds2uUyB/22.png b/vlm/dev/8aHzds2uUyB/22.png new file mode 100644 index 0000000000000000000000000000000000000000..3859b24e54d7605b05326a490350bb5898c93920 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b97628be7febccda02aca3bc1958ae9af1faf9516392a65c2c1e3557cae0ab01 +size 379221 diff --git a/vlm/dev/8aHzds2uUyB/23.png b/vlm/dev/8aHzds2uUyB/23.png new file mode 100644 index 0000000000000000000000000000000000000000..cb9fca8ef7663402928af62904ae52b97fa55c17 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32c79b59a15d3fcfb8d63618e5e86c561f3e572b19bb55843af6d4802b2ea8ba +size 534380 diff --git a/vlm/dev/8aHzds2uUyB/24.png b/vlm/dev/8aHzds2uUyB/24.png new file mode 100644 index 0000000000000000000000000000000000000000..742b47f9a1aed25dc8cffae6e008d3a7c8806f01 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5aa952ad0d4ca376b7d82f4e79708cdb4d644e07c1d0dfcc80865037dd9cf21 +size 389944 diff --git a/vlm/dev/8aHzds2uUyB/25.png b/vlm/dev/8aHzds2uUyB/25.png new file mode 100644 index 0000000000000000000000000000000000000000..33582e84a84c92bc4741e599fd0881ea44e8cc1a --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e5553e8a3542a1f5f88382bb57e88ed2a2c998a7d3dff4e7eb85f54b3d6670 +size 588764 diff --git a/vlm/dev/8aHzds2uUyB/27.png b/vlm/dev/8aHzds2uUyB/27.png new file mode 100644 index 0000000000000000000000000000000000000000..21f5d695060da94c882e77ad61ea16a484ddb815 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/27.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74033959eef260b9dcd9cb181c046543b6270f4b21972be64e094abe7b1a313b +size 360296 diff --git a/vlm/dev/8aHzds2uUyB/28.png b/vlm/dev/8aHzds2uUyB/28.png new file mode 100644 index 0000000000000000000000000000000000000000..7a40ea30d63b12750a74768136997739512bbb4c --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/28.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa270920308a519590fe53705801237ae989c77940e08fa5aa37a6163a92ff8d +size 323908 diff --git a/vlm/dev/8aHzds2uUyB/29.png b/vlm/dev/8aHzds2uUyB/29.png new file mode 100644 index 0000000000000000000000000000000000000000..09c8d95a12ef1a5515e885420ba1406cafa3ac7a --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/29.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:345e417dde8f4ea80a0f6bfb0997b049b67c879a9052f1e12e7961490ca5c53f +size 273008 diff --git a/vlm/dev/8aHzds2uUyB/3.png b/vlm/dev/8aHzds2uUyB/3.png new file mode 100644 index 0000000000000000000000000000000000000000..c4d585ebf947644526c16884830cc78917315cbb --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c62483b63634fb12e81ae60ec5aba140de80c7abe7566452aaa13715030ecbd9 +size 631817 diff --git a/vlm/dev/8aHzds2uUyB/30.png b/vlm/dev/8aHzds2uUyB/30.png new file mode 100644 index 0000000000000000000000000000000000000000..b2ce43c64955d9ef65e13321b111a1fdb9a9531b --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/30.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:793e2caf02f476437f7981503e77107cd6569bdf7e529e0f0dd8a6da563c7387 +size 316708 diff --git a/vlm/dev/8aHzds2uUyB/31.png b/vlm/dev/8aHzds2uUyB/31.png new file mode 100644 index 0000000000000000000000000000000000000000..76dd8291dbba40fff1002ad75ac1b5088b119c2c --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/31.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b59dfa47c6798ae0d3026bf829896812d44a7f58bfbbced677e97c221db5ae39 +size 522816 diff --git a/vlm/dev/8aHzds2uUyB/33.png b/vlm/dev/8aHzds2uUyB/33.png new file mode 100644 index 0000000000000000000000000000000000000000..6818e2fdcb4bea23602c08506ed09d7a035df2c4 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/33.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4670de96d0e90d33e758525e3c43a700d5463ed82c09fdc99da11b5fe387bd27 +size 154170 diff --git a/vlm/dev/8aHzds2uUyB/36.png b/vlm/dev/8aHzds2uUyB/36.png new file mode 100644 index 0000000000000000000000000000000000000000..0861d5e3b92036de46646673ccb43513f0582798 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/36.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:497193e6e3b8f00081d839e73a8cbd62141d8db78a9f2a070c39adeb1a83871e +size 297522 diff --git a/vlm/dev/8aHzds2uUyB/37.png b/vlm/dev/8aHzds2uUyB/37.png new file mode 100644 index 0000000000000000000000000000000000000000..504d5ceefd06d79a2caad847db363c8e2fa60cdc --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/37.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2d3490e77e713f73710f01033df7e1d8c482fe93eb55c17e6bd633cc44c729e +size 369200 diff --git a/vlm/dev/8aHzds2uUyB/38.png b/vlm/dev/8aHzds2uUyB/38.png new file mode 100644 index 0000000000000000000000000000000000000000..92689dec1140bceccdd97e25a0c04e14db95d3c2 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/38.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f9793efecdd8fcf2898b58fdedb9aa31ca114925b4f88fbb9ae3dc8e49ba9b3 +size 673627 diff --git a/vlm/dev/8aHzds2uUyB/40.png b/vlm/dev/8aHzds2uUyB/40.png new file mode 100644 index 0000000000000000000000000000000000000000..8869b74223a3f291a7e2aa21ceea35e3e79d8b4d --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/40.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7899413f70d8328ea1a37156bc396b3daa05ff3763d99621e31311d44b3c883 +size 689643 diff --git a/vlm/dev/8aHzds2uUyB/41.png b/vlm/dev/8aHzds2uUyB/41.png new file mode 100644 index 0000000000000000000000000000000000000000..d49bcbf064abadf3b62dd64451608a4acd6d463b --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/41.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9945e3cdf1f49d0cd316224d2248f86090c8893316b3bf0165da59a2de05e88a +size 587788 diff --git a/vlm/dev/8aHzds2uUyB/42.png b/vlm/dev/8aHzds2uUyB/42.png new file mode 100644 index 0000000000000000000000000000000000000000..be3b089c3d011145617379995aa3bd5c84be2b58 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/42.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:344b792a81a8779cbe2b7abcea73a73218e6d206572e250b5b963b707fc9a32e +size 305760 diff --git a/vlm/dev/8aHzds2uUyB/43.png b/vlm/dev/8aHzds2uUyB/43.png new file mode 100644 index 0000000000000000000000000000000000000000..37a6aa07280e2672ff5fc442b5128f8ab83922a2 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/43.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:539bea11d009bae7f1031a704fc4aedc3916ce9e512049d2e6b466b4c0be61fe +size 263805 diff --git a/vlm/dev/8aHzds2uUyB/45.png b/vlm/dev/8aHzds2uUyB/45.png new file mode 100644 index 0000000000000000000000000000000000000000..7bd2decbf01636686aca269d34a1f1f20d47361c --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/45.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1969b761deca7b65d8e1e5471a02320811cfd7e1d477ef2f36792faeaa063f5b +size 322031 diff --git a/vlm/dev/8aHzds2uUyB/46.png b/vlm/dev/8aHzds2uUyB/46.png new file mode 100644 index 0000000000000000000000000000000000000000..56306b5b7ec6b7b0cfd1c3e96509d1c7dc4386f5 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/46.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fb26fa8a0e3a646b96225ef879a73325e7dc04e00be25d4fd5ea44dc9c38447 +size 646214 diff --git a/vlm/dev/8aHzds2uUyB/49.png b/vlm/dev/8aHzds2uUyB/49.png new file mode 100644 index 0000000000000000000000000000000000000000..f4fbe1fc64b1a1afc276f9aa0eb4e92f3aad73af --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/49.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c66e64726eb36ccf74eec5906655504be9dd50f502b68083bca206453b47fe5 +size 162370 diff --git a/vlm/dev/8aHzds2uUyB/5.png b/vlm/dev/8aHzds2uUyB/5.png new file mode 100644 index 0000000000000000000000000000000000000000..56c5c7b070e8ff403903a8cd2a078ff90901169c --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57f070c66e0aa45532be0ae6b0984277b12dc2e7fd7f4d4fe2f25b53441a2bcd +size 600583 diff --git a/vlm/dev/8aHzds2uUyB/52.png b/vlm/dev/8aHzds2uUyB/52.png new file mode 100644 index 0000000000000000000000000000000000000000..2174cd10078f7f0efaa5b8599df2dca61364d534 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/52.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c18da3bc70b649078a624b8a481fd668111568508bd58f05a40b668b98a7d20 +size 265158 diff --git a/vlm/dev/8aHzds2uUyB/53.png b/vlm/dev/8aHzds2uUyB/53.png new file mode 100644 index 0000000000000000000000000000000000000000..ea21780eab237e48a77c90a11266be0dbd0624a0 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/53.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9cde24482e79478cd1b95da75d3f5984827e199ce505b8e403082c74a2f59ce +size 312661 diff --git a/vlm/dev/8aHzds2uUyB/54.png b/vlm/dev/8aHzds2uUyB/54.png new file mode 100644 index 0000000000000000000000000000000000000000..e3a72a15d73204d9334eaf6c8d4dc31f092b2fd6 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/54.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40fbb4e93078d738317efd03f8ab839720bdbd46bf45175f77f218cac0903e8c +size 297446 diff --git a/vlm/dev/8aHzds2uUyB/55.png b/vlm/dev/8aHzds2uUyB/55.png new file mode 100644 index 0000000000000000000000000000000000000000..4f557e2deebac893371e22de4aa1dad168c02f7b --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/55.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f007ddbf041e6c02d755af89b1e278a5f58376a32d42ed0b9d041000a035f4b4 +size 399026 diff --git a/vlm/dev/8aHzds2uUyB/56.png b/vlm/dev/8aHzds2uUyB/56.png new file mode 100644 index 0000000000000000000000000000000000000000..ddb4ec08ac6ba2bb4018623cffcca5d646941240 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/56.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff7f261857f836f7de00836dc1f6ae5f725f4d450540f01b98b7f11f740121b4 +size 341823 diff --git a/vlm/dev/8aHzds2uUyB/57.png b/vlm/dev/8aHzds2uUyB/57.png new file mode 100644 index 0000000000000000000000000000000000000000..74292e3971c560bc6d10d7edc2c6a8ddad5501d2 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/57.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ddbd02c32593419604bfb513ad65b6a3cd5224b00752cc71e8acf6e2e593d7b +size 126971 diff --git a/vlm/dev/8aHzds2uUyB/58.png b/vlm/dev/8aHzds2uUyB/58.png new file mode 100644 index 0000000000000000000000000000000000000000..85eef7367489834d22eb6da62cc75dbfe5ae2934 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/58.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:beee117391d30f2de55d701a62cea63660e9d3cbae7db969601bdac28d012dd2 +size 405671 diff --git a/vlm/dev/8aHzds2uUyB/59.png b/vlm/dev/8aHzds2uUyB/59.png new file mode 100644 index 0000000000000000000000000000000000000000..bafec1e81dc2eb9dda5a0c43df5e66974e58727a --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/59.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:308ed654c2c60edca49527cae6fc56edd0c61e20663f436b1594c95014f511c9 +size 527259 diff --git a/vlm/dev/8aHzds2uUyB/6.png b/vlm/dev/8aHzds2uUyB/6.png new file mode 100644 index 0000000000000000000000000000000000000000..5ad9b75530fbfebb8ff9a8738b31b1d300bbef86 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed17f4566969ed8df90aa1b4414ef5d84e0f179882c2eb51142203cf81bfe208 +size 603995 diff --git a/vlm/dev/8aHzds2uUyB/60.png b/vlm/dev/8aHzds2uUyB/60.png new file mode 100644 index 0000000000000000000000000000000000000000..86918c81adc3f294b23b0821efe60266644acefa --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/60.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef158add3a07c98af8010964f8396baa515156f036d9c375b722c56b4c68eef4 +size 207578 diff --git a/vlm/dev/8aHzds2uUyB/7.png b/vlm/dev/8aHzds2uUyB/7.png new file mode 100644 index 0000000000000000000000000000000000000000..ae9b8ca15d8b1196ebc9cbf3411159463612179a --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a8a88484d1f5311fa9286624754a877158e91a82ddbea8ed1d78f4d32b7dc45 +size 661126 diff --git a/vlm/dev/8aHzds2uUyB/8.png b/vlm/dev/8aHzds2uUyB/8.png new file mode 100644 index 0000000000000000000000000000000000000000..7f14269cceb11985c29c2c03a7bfa019e84495ed --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fde39f88246c30596f96c1f67c097987a3134cd43c625209c18981541ca5f386 +size 589910 diff --git a/vlm/dev/8aHzds2uUyB/9.png b/vlm/dev/8aHzds2uUyB/9.png new file mode 100644 index 0000000000000000000000000000000000000000..87c5c1082a859ab4180405274efe2b72ef6a80c7 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b36a01cdf661627fde5742c0ab3338dff52d9397058c99e9f02128901f961c6f +size 568818 diff --git a/vlm/dev/CQsmMYmlP5T/0.png b/vlm/dev/CQsmMYmlP5T/0.png new file mode 100644 index 0000000000000000000000000000000000000000..64f0f1c9ce79bf90f458b799ffa9f22e7641cd8a --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:923f0a10ce0e4af23c90a2102324539347dc2d7ecc8536828c20f6e3c431ee13 +size 476688 diff --git a/vlm/dev/CQsmMYmlP5T/1.png b/vlm/dev/CQsmMYmlP5T/1.png new file mode 100644 index 0000000000000000000000000000000000000000..b52a707a35383bc4b28d2f9fb1671b462deb3dc0 --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53227928033290d2e520e5afc511278e474edd89782d742acaea42bf90abe6c5 +size 482189 diff --git a/vlm/dev/CQsmMYmlP5T/12.png b/vlm/dev/CQsmMYmlP5T/12.png new file mode 100644 index 0000000000000000000000000000000000000000..8126518ae9a7a07d88c755be84844ae18898901f --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd74d31d8f83f18831ec265a74961fd46ab84ced4e385ff3eb5a619f88dd37e1 +size 642174 diff --git a/vlm/dev/CQsmMYmlP5T/13.png b/vlm/dev/CQsmMYmlP5T/13.png new file mode 100644 index 0000000000000000000000000000000000000000..a959fb050bbe6bc419cad95ffbcc03777606d4ae --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fd4483a4f1175d08410b03c100fb2361351c4d57a983cf8e1710f71ab61a967 +size 594210 diff --git a/vlm/dev/CQsmMYmlP5T/16.png b/vlm/dev/CQsmMYmlP5T/16.png new file mode 100644 index 0000000000000000000000000000000000000000..e2d7cf70383207ac052f05503b4c7c98763f3687 --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e09babe0ecccb3fbf5ced08ba386dcecf310b5c41bbafb5f4b41b93bf643cdd +size 639809 diff --git a/vlm/dev/CQsmMYmlP5T/18.png b/vlm/dev/CQsmMYmlP5T/18.png new file mode 100644 index 0000000000000000000000000000000000000000..f0bea799daf787d8cf5f8084a7dbb44d34c68dcf --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47de91aaa33cb5958fe220ef4cf3a7a3a726223f3387b885413051276ac103c4 +size 500271 diff --git a/vlm/dev/CQsmMYmlP5T/19.png b/vlm/dev/CQsmMYmlP5T/19.png new file mode 100644 index 0000000000000000000000000000000000000000..41abad5f6edd6a6aef692c4c2b693a2cd74d8412 --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97cff966f3a5dc946226b644825e9c2f3c93734e5778670b1c1cfdf7446785e5 +size 574176 diff --git a/vlm/dev/CQsmMYmlP5T/20.png b/vlm/dev/CQsmMYmlP5T/20.png new file mode 100644 index 0000000000000000000000000000000000000000..39b1dd277c495f3dce8f8fe8d0edf2d6a8fa801f --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5c27fdfdca615ddc9b996e750ed1b4db9b7aace4cbcd5afe619ed8b09c24754 +size 453792 diff --git a/vlm/dev/CQsmMYmlP5T/22.png b/vlm/dev/CQsmMYmlP5T/22.png new file mode 100644 index 0000000000000000000000000000000000000000..70193664f4ec5cf71313513dd2430cf7e0a654fd --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:176bf3b1c069d32c8e4764f5264994f9460fb779de9fc18b65821c682c567118 +size 548508 diff --git a/vlm/dev/CQsmMYmlP5T/25.png b/vlm/dev/CQsmMYmlP5T/25.png new file mode 100644 index 0000000000000000000000000000000000000000..ffa2bf3e043722f3cb23f9f41e52d5a1a69eeb81 --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e24b1379af8597f8eb860c59762cde235b4e532836144a11b0f52561a833981 +size 347162 diff --git a/vlm/dev/CQsmMYmlP5T/26.png b/vlm/dev/CQsmMYmlP5T/26.png new file mode 100644 index 0000000000000000000000000000000000000000..bd77c64cdb7e5a4964a14848b90da514249e483b --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/26.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6d8b727ce8be1729bc22bb5ef33feaf8f0f009a5873a335099e9555dc918f69 +size 380705 diff --git a/vlm/dev/CQsmMYmlP5T/27.png b/vlm/dev/CQsmMYmlP5T/27.png new file mode 100644 index 0000000000000000000000000000000000000000..beee78f3b66b276ffddd7da26e97620cdacf0069 --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/27.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3ba227dd627f614a2336d592ebdf96be8850b6f2639b427a6fa4830bb75a99f +size 419547 diff --git a/vlm/dev/CQsmMYmlP5T/28.png b/vlm/dev/CQsmMYmlP5T/28.png new file mode 100644 index 0000000000000000000000000000000000000000..ee32e6595a6d3e9757ed42b438f8417b3c741eb2 --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/28.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93a48e68b87d05610f4150374d2911738642a12b7613902d99c1cd3065a6bc92 +size 225239 diff --git a/vlm/dev/CQsmMYmlP5T/4.png b/vlm/dev/CQsmMYmlP5T/4.png new file mode 100644 index 0000000000000000000000000000000000000000..1e27e207bb3a5165573a016a0e10b05e0e113206 --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:126ae9bced6d075782610d618a6f876600e8d3fde9bdc3384719e3246ffeb938 +size 472495 diff --git a/vlm/dev/CQsmMYmlP5T/5.png b/vlm/dev/CQsmMYmlP5T/5.png new file mode 100644 index 0000000000000000000000000000000000000000..3e71ffabedee9f2a7b230512a69637a968a7623b --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d661b5cfec72efd13d38b2706c86b66b756b241f1b6b6ee921cc4f468c5febae +size 502638 diff --git a/vlm/dev/CQsmMYmlP5T/7.png b/vlm/dev/CQsmMYmlP5T/7.png new file mode 100644 index 0000000000000000000000000000000000000000..f9d3238872f46051f27b44076e349c7095874ac9 --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd395e11f9948832fbb01a5d3fe8f734e02e5bbad6d1417b29c7b2165539854d +size 582983 diff --git a/vlm/dev/CQsmMYmlP5T/8.png b/vlm/dev/CQsmMYmlP5T/8.png new file mode 100644 index 0000000000000000000000000000000000000000..15fe049e51e1d8fd4709ca1cd536a4c313d4838e --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fdd3157ae8709ffa07172a06f2cae39c41f9a5763924f3c9287f0913dd1e3561 +size 598936 diff --git a/vlm/dev/I3mLa12s_H/0.png b/vlm/dev/I3mLa12s_H/0.png new file mode 100644 index 0000000000000000000000000000000000000000..8d1b3112bc0a59b5b2ed285be83cc78d030fc1aa --- /dev/null +++ b/vlm/dev/I3mLa12s_H/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c88a9b2efa383f73140192f96ec9349024ba83872155c47e99c04841fff47598 +size 455784 diff --git a/vlm/dev/I3mLa12s_H/1.png b/vlm/dev/I3mLa12s_H/1.png new file mode 100644 index 0000000000000000000000000000000000000000..d87ce7dc05bca43f355d3a42f80999ebd4d8864c --- /dev/null +++ b/vlm/dev/I3mLa12s_H/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:323a68b2e11267943e2ca9b81e1993b7e888b54f71a074b74e3ae4cbc2038fca +size 546962 diff --git a/vlm/dev/I3mLa12s_H/10.png b/vlm/dev/I3mLa12s_H/10.png new file mode 100644 index 0000000000000000000000000000000000000000..c48ca6d4abf93cb83516af412c0498f2f2cc7078 --- /dev/null +++ b/vlm/dev/I3mLa12s_H/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c64a123f546c50b568bcd30a891ced18f1af6b0238ffccb7b5de0af9b55b4b2b +size 480848 diff --git a/vlm/dev/I3mLa12s_H/11.png b/vlm/dev/I3mLa12s_H/11.png new file mode 100644 index 0000000000000000000000000000000000000000..a8cfaf83c9466c7c95abceccc2a781800ebdf868 --- /dev/null +++ b/vlm/dev/I3mLa12s_H/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4605684a9744e2130b1fd1e0d3d5334eccea87fdb9de1534c8e268bbc90f9ca9 +size 496460 diff --git a/vlm/dev/I3mLa12s_H/12.png b/vlm/dev/I3mLa12s_H/12.png new file mode 100644 index 0000000000000000000000000000000000000000..cd268b416f6749e5863cf3c37cf958a178e69c1e --- /dev/null +++ b/vlm/dev/I3mLa12s_H/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2901748ecfa6badf0085e02772394abf913a258bdc11544663be66f3021db1f +size 184506 diff --git a/vlm/dev/I3mLa12s_H/2.png b/vlm/dev/I3mLa12s_H/2.png new file mode 100644 index 0000000000000000000000000000000000000000..3ecf65e89e9e5c2fea8505ba361f86ff8e384821 --- /dev/null +++ b/vlm/dev/I3mLa12s_H/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3da25e8d3897b4bd5d21fad4049c5d66df6af138c804670bb84d985ae388426a +size 446446 diff --git a/vlm/dev/I3mLa12s_H/3.png b/vlm/dev/I3mLa12s_H/3.png new file mode 100644 index 0000000000000000000000000000000000000000..8d2b8ec37ef452407f4e8632af0786a43b1dbc04 --- /dev/null +++ b/vlm/dev/I3mLa12s_H/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c871f551ce98882792f88f4af152a9b5b5a4a0ef1f2d6e1c4bc4e29ef2d2f5ed +size 563195 diff --git a/vlm/dev/I3mLa12s_H/4.png b/vlm/dev/I3mLa12s_H/4.png new file mode 100644 index 0000000000000000000000000000000000000000..d963a9ef9fe417fec822ae2d638c1f99fcedd603 --- /dev/null +++ b/vlm/dev/I3mLa12s_H/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96dac9e186effb0a90e9c5452ceb19f70bc97c17bd6634ed14090d659480f288 +size 413581 diff --git a/vlm/dev/I3mLa12s_H/5.png b/vlm/dev/I3mLa12s_H/5.png new file mode 100644 index 0000000000000000000000000000000000000000..30027bcef2e386ffd24795c3282b7a0aae95129d --- /dev/null +++ b/vlm/dev/I3mLa12s_H/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b4968514e049f14d16876fdf149c445fb5c0f386280a99d82682e4090402fe2 +size 498698 diff --git a/vlm/dev/I3mLa12s_H/6.png b/vlm/dev/I3mLa12s_H/6.png new file mode 100644 index 0000000000000000000000000000000000000000..84aea88e41d08508ffe3fbc3543d9a1b031bd0d6 --- /dev/null +++ b/vlm/dev/I3mLa12s_H/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:535b92d709435e519dabea253b4aeb42ad115b96af6dd9698322d07bb065fb3f +size 523889 diff --git a/vlm/dev/I3mLa12s_H/7.png b/vlm/dev/I3mLa12s_H/7.png new file mode 100644 index 0000000000000000000000000000000000000000..e6e4a858d7e7d881f9798295d9fd208bdff8c6cf --- /dev/null +++ b/vlm/dev/I3mLa12s_H/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27ce8c1bdd4b3bdf3a03f01cd2c31affd50f64ca96590d66ac37d46e3fcc9c27 +size 971097 diff --git a/vlm/dev/I3mLa12s_H/8.png b/vlm/dev/I3mLa12s_H/8.png new file mode 100644 index 0000000000000000000000000000000000000000..191695e31d64d75dfa343b3ae1dbcf9d7a375d94 --- /dev/null +++ b/vlm/dev/I3mLa12s_H/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6a18412a0e0c6dc8b02fb6267534b4c216c8898f1e2f322b101b47080b28c0f +size 507130 diff --git a/vlm/dev/I3mLa12s_H/9.png b/vlm/dev/I3mLa12s_H/9.png new file mode 100644 index 0000000000000000000000000000000000000000..192c48853e51fcb662c1b2d9799efd52d43f5a11 --- /dev/null +++ b/vlm/dev/I3mLa12s_H/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6daa0ee27e5eb33ba2e12fa80f290d1e4bfa295230ed828edc2a4777a9dfed9b +size 520010 diff --git a/vlm/dev/LQnyIk5dUA/0.png b/vlm/dev/LQnyIk5dUA/0.png new file mode 100644 index 0000000000000000000000000000000000000000..f4d3f1620957637fcc3c5b8d7061a4f3f44a105b --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:025e52330d81e3967f5d214424be375e2d60df3d6cec40a8e8ea35b2da9d9347 +size 421579 diff --git a/vlm/dev/LQnyIk5dUA/1.png b/vlm/dev/LQnyIk5dUA/1.png new file mode 100644 index 0000000000000000000000000000000000000000..a717c8f75fc03a8e063e6f91283320805f444879 --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23241ee0e62f2a5af215d62eae32feddd27099b9e61fb01807785f1b28ff3128 +size 654893 diff --git a/vlm/dev/LQnyIk5dUA/10.png b/vlm/dev/LQnyIk5dUA/10.png new file mode 100644 index 0000000000000000000000000000000000000000..d21e437734fdbeca3791ca71f9d31642e0b509aa --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:352d82eab9aa821e72c3f4e1970edd349bcfbeedb0e488a4690da6c2ff888cd4 +size 284773 diff --git a/vlm/dev/LQnyIk5dUA/11.png b/vlm/dev/LQnyIk5dUA/11.png new file mode 100644 index 0000000000000000000000000000000000000000..c49de12a2352dd8835f720fff532cc964550e887 --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8c4576d3a7cc4439d50f4814b8a38faaffdd9a8bf8bc11cf0a140cc94a2d27d +size 343396 diff --git a/vlm/dev/LQnyIk5dUA/12.png b/vlm/dev/LQnyIk5dUA/12.png new file mode 100644 index 0000000000000000000000000000000000000000..73300320ceec31e088ac14231c4f70be2ec11b1a --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bba9d99f3252a1440f900afafcce32fae8c4db3ac2dce44f3653ff23062b01c4 +size 534520 diff --git a/vlm/dev/LQnyIk5dUA/13.png b/vlm/dev/LQnyIk5dUA/13.png new file mode 100644 index 0000000000000000000000000000000000000000..39eb5725e23b119059cbfedd1f3d8f131308c728 --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba86486d76425d119685f1eccb3e9012a6d8213b692bd87c077add4f1d991561 +size 307229 diff --git a/vlm/dev/LQnyIk5dUA/14.png b/vlm/dev/LQnyIk5dUA/14.png new file mode 100644 index 0000000000000000000000000000000000000000..51ba259e3e6eae91354aadfd1c8f42ecc4f342f6 --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00a52bd1922a19eae8abe46325aa592299b2fedd85ed9e384355a7a09a488505 +size 303525 diff --git a/vlm/dev/LQnyIk5dUA/15.png b/vlm/dev/LQnyIk5dUA/15.png new file mode 100644 index 0000000000000000000000000000000000000000..b58e45b5998b99e9b2a8d889637c8d59da611e94 --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c58c597bf63e32ea20c3af2e03ff6c728ecba1b90cf51b01ddf34b79204c19a +size 288274 diff --git a/vlm/dev/LQnyIk5dUA/16.png b/vlm/dev/LQnyIk5dUA/16.png new file mode 100644 index 0000000000000000000000000000000000000000..d2ba111171b781c263421b7d3f5b9c10358112e9 --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efa495cffd3b5ef60b9e857de967396a7d82b63ec9c086f843e9e3b8c23500a2 +size 377991 diff --git a/vlm/dev/LQnyIk5dUA/17.png b/vlm/dev/LQnyIk5dUA/17.png new file mode 100644 index 0000000000000000000000000000000000000000..703df31d995e0ba1b504b6c0be9746c312deb43d --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9145299379089c6087f54ec56a42f998e5962f5289e09b2740462a69ba3b4487 +size 350221 diff --git a/vlm/dev/LQnyIk5dUA/18.png b/vlm/dev/LQnyIk5dUA/18.png new file mode 100644 index 0000000000000000000000000000000000000000..531a76b33b4ce9b6ad1657f78944e8706c565d0d --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eaf4134261c583e4fdcdd42685d135428c075ad958e82a6cf58a4cf6e5bc3bfc +size 314758 diff --git a/vlm/dev/LQnyIk5dUA/19.png b/vlm/dev/LQnyIk5dUA/19.png new file mode 100644 index 0000000000000000000000000000000000000000..7e14b6fdea8a3d7caf3f4499f5082971e50e6e3f --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d7590a7cbb0cdc151f041402fe246cbfd55c09ce625e2136a3ef713038c4bfa +size 298760 diff --git a/vlm/dev/LQnyIk5dUA/2.png b/vlm/dev/LQnyIk5dUA/2.png new file mode 100644 index 0000000000000000000000000000000000000000..92dfbfb393a426ac3817e174e9492d19ce338719 --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f21d62bdcf46615c76ec7dcde938554e3ff92ad7758b553f94312a7f6b39674 +size 463928 diff --git a/vlm/dev/LQnyIk5dUA/20.png b/vlm/dev/LQnyIk5dUA/20.png new file mode 100644 index 0000000000000000000000000000000000000000..fae4544aadc5db0db92dfee9ad6750db41fefd6a --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:698e91c128cfac88974fffc9d97c21fb19feaf53d66801085a4a0fcad780d2f3 +size 358987 diff --git a/vlm/dev/LQnyIk5dUA/21.png b/vlm/dev/LQnyIk5dUA/21.png new file mode 100644 index 0000000000000000000000000000000000000000..b8894ef6687b2a84a0b85d977499538fd3626ed1 --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1d69a09253f8a49d9c133d9756f7c944d63d8ea0bb92a21f3f273fc568cce60 +size 440158 diff --git a/vlm/dev/LQnyIk5dUA/22.png b/vlm/dev/LQnyIk5dUA/22.png new file mode 100644 index 0000000000000000000000000000000000000000..668ee0411c7d3de9bf44ff588ab69c67c02272ce --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2beac7adf7d1dc608b30f69c8e3bd8c1f7d0ad087785e5fcfd2bfd264d6e316 +size 402862 diff --git a/vlm/dev/LQnyIk5dUA/23.png b/vlm/dev/LQnyIk5dUA/23.png new file mode 100644 index 0000000000000000000000000000000000000000..35ed1953214057c52a0a909739a11ad1ec03a8ec --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:654e78ca02495ddbf8a8ed1bb71ab2d8e7947656c0d7be1961e08709d8b7a431 +size 342177 diff --git a/vlm/dev/LQnyIk5dUA/3.png b/vlm/dev/LQnyIk5dUA/3.png new file mode 100644 index 0000000000000000000000000000000000000000..22cacefb9a420aef0b749efda64b4be774c3ea9d --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10dad8d7040e567ad124ceb999eb5217e4c21ddf3a4358589467947d4da56a8d +size 451953 diff --git a/vlm/dev/LQnyIk5dUA/4.png b/vlm/dev/LQnyIk5dUA/4.png new file mode 100644 index 0000000000000000000000000000000000000000..4eada627095f56f41039d221c47f134781df60e4 --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a577d74e0aa7c2fb78f4157fed51018a884f57eaa03deca191b577052cbdb12 +size 392142 diff --git a/vlm/dev/LQnyIk5dUA/5.png b/vlm/dev/LQnyIk5dUA/5.png new file mode 100644 index 0000000000000000000000000000000000000000..55bb3f3b4d36481f882d26bd91da3e7433e147df --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d103bcf4afdecc74134644ed5db5723d831154e8b178a12a4da4fcfb46bf0831 +size 468829 diff --git a/vlm/dev/LQnyIk5dUA/6.png b/vlm/dev/LQnyIk5dUA/6.png new file mode 100644 index 0000000000000000000000000000000000000000..f743e1a3451cb5cac7c68717b0fc4936a8b66c7f --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b931460ec567c311634861a084eccaf5bdc8191a05d710e686791c6fc3fae19 +size 524651 diff --git a/vlm/dev/LQnyIk5dUA/7.png b/vlm/dev/LQnyIk5dUA/7.png new file mode 100644 index 0000000000000000000000000000000000000000..329983b34006282f6c47bd0a252dd9b9236e8508 --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9c4b67cefbdc197759f816c5af9efaa63d87c5692520f14abd17848058f64dc +size 428685 diff --git a/vlm/dev/LQnyIk5dUA/8.png b/vlm/dev/LQnyIk5dUA/8.png new file mode 100644 index 0000000000000000000000000000000000000000..8b8cd8a302b864ab99f56fd6ba1d2f79f2fcc9e3 --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9351d3e8cccaf6420442b2c7b2320cf45e9a7035a59ddadfc7fe7fb1c7f57f58 +size 587849 diff --git a/vlm/dev/LQnyIk5dUA/9.png b/vlm/dev/LQnyIk5dUA/9.png new file mode 100644 index 0000000000000000000000000000000000000000..e438fd07a5b05d28bd8bb2541d672c1e134c5d20 --- /dev/null +++ b/vlm/dev/LQnyIk5dUA/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7f265585bf97776c1e3bc437b5a0c8ba6975db79fa3c2f339fe932301dae73e +size 541733 diff --git a/vlm/dev/Nayau9fwXU/0.png b/vlm/dev/Nayau9fwXU/0.png new file mode 100644 index 0000000000000000000000000000000000000000..411ca4a0946d0185b8deec5eab7c1697be2a653e --- /dev/null +++ b/vlm/dev/Nayau9fwXU/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17551fb7385c7159ce5004961f0d00bd12052f93a063bdfbce9d0b4b7f653d1d +size 690596 diff --git a/vlm/dev/Nayau9fwXU/1.png b/vlm/dev/Nayau9fwXU/1.png new file mode 100644 index 0000000000000000000000000000000000000000..9eac463fa5d34fb0f296f06fe0bb62b0991b84ce --- /dev/null +++ b/vlm/dev/Nayau9fwXU/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bea876588c04bde75cb0f1f6274ac86e000fc8c28ac33c9b30970d803285b0a +size 595106 diff --git a/vlm/dev/Nayau9fwXU/10.png b/vlm/dev/Nayau9fwXU/10.png new file mode 100644 index 0000000000000000000000000000000000000000..b3f0fc2aafd603bb0aab913c4ecd3ba102c3cb4a --- /dev/null +++ b/vlm/dev/Nayau9fwXU/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4fe56cf294d12fba873fd2fb662e39efea21d543ede57221b9038b704954a4b +size 556646 diff --git a/vlm/dev/Nayau9fwXU/11.png b/vlm/dev/Nayau9fwXU/11.png new file mode 100644 index 0000000000000000000000000000000000000000..4e8193043eb9dc2b77b521e337c1dcc3284f160d --- /dev/null +++ b/vlm/dev/Nayau9fwXU/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1744d5bb49b5442b6c118d0e8e106deef487e6318631e31c31149768f901ffd4 +size 554707 diff --git a/vlm/dev/Nayau9fwXU/12.png b/vlm/dev/Nayau9fwXU/12.png new file mode 100644 index 0000000000000000000000000000000000000000..2aa0e349ae82f7d12f1c25921c05dbbbb2dc7f09 --- /dev/null +++ b/vlm/dev/Nayau9fwXU/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1807538a5b419f3dcbe0fde8386aab9f8434c7330faa60b0e79c2f8e46bc38dd +size 361597 diff --git a/vlm/dev/Nayau9fwXU/13.png b/vlm/dev/Nayau9fwXU/13.png new file mode 100644 index 0000000000000000000000000000000000000000..f8c566e0f1e2d5e8cc4c5dae3eac45231ff00f7c --- /dev/null +++ b/vlm/dev/Nayau9fwXU/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee1271309844200678139274143e0e8bde56b878397c67c514b0eaf251610517 +size 537264 diff --git a/vlm/dev/Nayau9fwXU/14.png b/vlm/dev/Nayau9fwXU/14.png new file mode 100644 index 0000000000000000000000000000000000000000..881109d6a7bc564746ab477e8a6ac5388b57529c --- /dev/null +++ b/vlm/dev/Nayau9fwXU/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b3f99ffc0f7dfb68a297e3a85d7dd6882fa56707df65a4ec158298723858c21 +size 534169 diff --git a/vlm/dev/Nayau9fwXU/15.png b/vlm/dev/Nayau9fwXU/15.png new file mode 100644 index 0000000000000000000000000000000000000000..436bbef9b23cde83f020f7039157d9d30fb300fe --- /dev/null +++ b/vlm/dev/Nayau9fwXU/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97ac62132e8dfe22acf8dded04ffd48979d2366ed29db7739a7e302bf80cb3d3 +size 365069 diff --git a/vlm/dev/Nayau9fwXU/16.png b/vlm/dev/Nayau9fwXU/16.png new file mode 100644 index 0000000000000000000000000000000000000000..880626017bdd317989da58760f48541446ea36f9 --- /dev/null +++ b/vlm/dev/Nayau9fwXU/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c681dfb9cc524565dcf86d27251f3a121bf64f4f940fa9e79fe8703200dca310 +size 1743142 diff --git a/vlm/dev/Nayau9fwXU/17.png b/vlm/dev/Nayau9fwXU/17.png new file mode 100644 index 0000000000000000000000000000000000000000..fb044dea3b8daf97cf9a4eea8586010765caaf77 --- /dev/null +++ b/vlm/dev/Nayau9fwXU/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82b5c386aff0a4d3b568f084f2e9d753c5f6eb41883b070f24779a1c2b734768 +size 818779 diff --git a/vlm/dev/Nayau9fwXU/18.png b/vlm/dev/Nayau9fwXU/18.png new file mode 100644 index 0000000000000000000000000000000000000000..06f4c5423373a7b5ec5534603932fee9d85be044 --- /dev/null +++ b/vlm/dev/Nayau9fwXU/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18cc5c21924573d8192ebd8880db8868b62c6da5b99d4cb46759727beaa33369 +size 1570533 diff --git a/vlm/dev/Nayau9fwXU/19.png b/vlm/dev/Nayau9fwXU/19.png new file mode 100644 index 0000000000000000000000000000000000000000..e2361f209d28e5c6890b5028a3770f7e1640b2d4 --- /dev/null +++ b/vlm/dev/Nayau9fwXU/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45911c394768bcbded7a671ee25c09c165242d0cac12127dbd005c0b6ffccf6e +size 700534 diff --git a/vlm/dev/Nayau9fwXU/2.png b/vlm/dev/Nayau9fwXU/2.png new file mode 100644 index 0000000000000000000000000000000000000000..2c774b9087a1120696c1688d266b647fff2d0424 --- /dev/null +++ b/vlm/dev/Nayau9fwXU/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca962367cbd63142d435f9d7bac9fbb96a9cd3568a747653d2cc58415b0056a4 +size 485674 diff --git a/vlm/dev/Nayau9fwXU/20.png b/vlm/dev/Nayau9fwXU/20.png new file mode 100644 index 0000000000000000000000000000000000000000..9d8dc51fa05d10c7a21b1c10dc541492582ca8b9 --- /dev/null +++ b/vlm/dev/Nayau9fwXU/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:414abdc58f37b31dba3415c71575688f0b6800e764c95e0b1f1b9899496bc585 +size 1828323 diff --git a/vlm/dev/Nayau9fwXU/21.png b/vlm/dev/Nayau9fwXU/21.png new file mode 100644 index 0000000000000000000000000000000000000000..c662d6a831026a5bb2ce277cf301613a18acd7ec --- /dev/null +++ b/vlm/dev/Nayau9fwXU/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1290824ea9f97ce4c82e4759d8509a9a7201b3a0d113633ef549625900ea4350 +size 2061458 diff --git a/vlm/dev/Nayau9fwXU/3.png b/vlm/dev/Nayau9fwXU/3.png new file mode 100644 index 0000000000000000000000000000000000000000..49f2b4b2d2cf6d88b1c8bc3bfc018ce47f1b62ec --- /dev/null +++ b/vlm/dev/Nayau9fwXU/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b32ed26194e69de55f719c29ec37b9e98e7b07627ca8675370d62d4c4b6a70ab +size 658484 diff --git a/vlm/dev/Nayau9fwXU/4.png b/vlm/dev/Nayau9fwXU/4.png new file mode 100644 index 0000000000000000000000000000000000000000..3d6995ed01440ef2d18d1af3e6f570670ef9d8e7 --- /dev/null +++ b/vlm/dev/Nayau9fwXU/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff5efb17db15572182d2ee813803615efb584303f8896b7b8179aa3f097a75ca +size 560695 diff --git a/vlm/dev/Nayau9fwXU/5.png b/vlm/dev/Nayau9fwXU/5.png new file mode 100644 index 0000000000000000000000000000000000000000..694f041d3dcba2cccd7b8dc8d2f8f4c8e73df8f9 --- /dev/null +++ b/vlm/dev/Nayau9fwXU/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97db196424a8a1e324af8200fa557a9bf88834ec7daa2c14fff916a582fb9d1a +size 514882 diff --git a/vlm/dev/Nayau9fwXU/6.png b/vlm/dev/Nayau9fwXU/6.png new file mode 100644 index 0000000000000000000000000000000000000000..2617abcfe78cc2427ae332338500e8310fe5cb8e --- /dev/null +++ b/vlm/dev/Nayau9fwXU/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a038801914f686e7b362673d0752fe378dd9b065b5a3504dbcacb8588faa43c +size 2296283 diff --git a/vlm/dev/Nayau9fwXU/7.png b/vlm/dev/Nayau9fwXU/7.png new file mode 100644 index 0000000000000000000000000000000000000000..b163ad1a2af8b68a88b3e4d982df33824bb9d17c --- /dev/null +++ b/vlm/dev/Nayau9fwXU/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d90ea9b39d4767d062a3d66a207171eeb485a520c2c11726d57a3350ca02165 +size 602768 diff --git a/vlm/dev/Nayau9fwXU/8.png b/vlm/dev/Nayau9fwXU/8.png new file mode 100644 index 0000000000000000000000000000000000000000..4bc7d97428888cc3f3610f186a5fe66a7f17f987 --- /dev/null +++ b/vlm/dev/Nayau9fwXU/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e75d162e88056ad06b7e1d2aa29549c3376c5323d2869669bb5c91d3f6f1f5ab +size 1725504 diff --git a/vlm/dev/Nayau9fwXU/9.png b/vlm/dev/Nayau9fwXU/9.png new file mode 100644 index 0000000000000000000000000000000000000000..9c56ed0be2a910461365aef1b53ce764b7d5e3f6 --- /dev/null +++ b/vlm/dev/Nayau9fwXU/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:880c83b6a0d7a5b297587999419926233e71eeb96be6a16d7234420aaeb7ed8a +size 555896 diff --git a/vlm/dev/QYvFUlF19n/13.png b/vlm/dev/QYvFUlF19n/13.png new file mode 100644 index 0000000000000000000000000000000000000000..a96c35b15725cdc9850eceae71f5d62f104aca22 --- /dev/null +++ b/vlm/dev/QYvFUlF19n/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:beccfd26d0eba4dd9e10ec8a0aadf1a34da23c2b8a0c1714a38b58a4dae5aabe +size 388047 diff --git a/vlm/dev/QYvFUlF19n/4.png b/vlm/dev/QYvFUlF19n/4.png new file mode 100644 index 0000000000000000000000000000000000000000..ea3c6e0ad270c33edb9913695fc3404078bd8985 --- /dev/null +++ b/vlm/dev/QYvFUlF19n/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0911fd9842fa4e0b4a45c972a345096df5a459bcb87da52f2c3e4ac1cd074658 +size 715431 diff --git a/vlm/dev/QYvFUlF19n/8.png b/vlm/dev/QYvFUlF19n/8.png new file mode 100644 index 0000000000000000000000000000000000000000..f02d14cc9ee2485c34640fb92f299f3eb5c190bd --- /dev/null +++ b/vlm/dev/QYvFUlF19n/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:add3d5522b7a8b2654738833532674e04108a303486cff7720125160df5c267e +size 244621 diff --git a/vlm/dev/Qx8lUU8CzQ/0.png b/vlm/dev/Qx8lUU8CzQ/0.png new file mode 100644 index 0000000000000000000000000000000000000000..7a7a62ca3a076a438da41b2f4eb33ac8ab981da6 --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2c9e9b81edb94c7bb98c6c57092b2fa79ae8b73ae9b3e68d1fd038f2bb82aa0 +size 487273 diff --git a/vlm/dev/Qx8lUU8CzQ/1.png b/vlm/dev/Qx8lUU8CzQ/1.png new file mode 100644 index 0000000000000000000000000000000000000000..79c868c20c1e8004082ee9f97566cc720eed0498 --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83353f7de52cf107b0a71bd771ef6f768efa45505745433a27c2bd64a76a26a4 +size 655732 diff --git a/vlm/dev/Qx8lUU8CzQ/10.png b/vlm/dev/Qx8lUU8CzQ/10.png new file mode 100644 index 0000000000000000000000000000000000000000..0f8124ffb7b13c314f75a9dee61f34a7c938e055 --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:493416f20d5b99b1fcbba1f06bb4b8f0f683f67a23f1233bbee5b5fd0e1c9c86 +size 574707 diff --git a/vlm/dev/Qx8lUU8CzQ/11.png b/vlm/dev/Qx8lUU8CzQ/11.png new file mode 100644 index 0000000000000000000000000000000000000000..1f3385960e6dfb4c361ff3b47f955ec94a920fc2 --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7529f9cd38a3a9bcb332768355b2183ec31681c8e4fb67c78bfee4b8dd43ab94 +size 546877 diff --git a/vlm/dev/Qx8lUU8CzQ/12.png b/vlm/dev/Qx8lUU8CzQ/12.png new file mode 100644 index 0000000000000000000000000000000000000000..2a2c65e07eab12e4805db07b086edd4c4844b9d3 --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c8ad274f6396eddd7fc9259f2be58ed9b4fffb64f78dc0988be1cdb7a224c1c +size 582286 diff --git a/vlm/dev/Qx8lUU8CzQ/13.png b/vlm/dev/Qx8lUU8CzQ/13.png new file mode 100644 index 0000000000000000000000000000000000000000..a4aa9f1249f281b15027970cfd2e98bbb44fdcb0 --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0310b912683ea4187d1a1123746dffaa2b2c62ef94f93b64a73fd113ff272e16 +size 115875 diff --git a/vlm/dev/Qx8lUU8CzQ/14.png b/vlm/dev/Qx8lUU8CzQ/14.png new file mode 100644 index 0000000000000000000000000000000000000000..4d5caecb6c4b9a068cbf904d18712559b5786e42 --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6544c220b65d27051b43670381acc4238ef53952f87c51161e9996f4feef980e +size 497330 diff --git a/vlm/dev/Qx8lUU8CzQ/15.png b/vlm/dev/Qx8lUU8CzQ/15.png new file mode 100644 index 0000000000000000000000000000000000000000..d3d406d8a78f02f8dbfcf72318bec62194c237b8 --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7233850e02471afe02bd69e3c998d7e133196368292596d634a61ca8644d105c +size 346435 diff --git a/vlm/dev/Qx8lUU8CzQ/16.png b/vlm/dev/Qx8lUU8CzQ/16.png new file mode 100644 index 0000000000000000000000000000000000000000..82f39307f44b36c1771cfad24482d25554344dda --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b27061a5aa18aa7eb0e923e1c70eaf960a469fb7c38186b700d25d3d2040e837 +size 1202444 diff --git a/vlm/dev/Qx8lUU8CzQ/17.png b/vlm/dev/Qx8lUU8CzQ/17.png new file mode 100644 index 0000000000000000000000000000000000000000..ba6c357d067e4cf3a60c29556fafa80af77a9a34 --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9df4ccfed9103007aa36f44b520a77346f5f8fb615279fb1d176a6d522b15350 +size 756441 diff --git a/vlm/dev/Qx8lUU8CzQ/18.png b/vlm/dev/Qx8lUU8CzQ/18.png new file mode 100644 index 0000000000000000000000000000000000000000..11f4ab2e2aad718c59e66414da799b459113368e --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15dc84580500ca7e790a916f4be35ec1a2a41177a477d0119d31b49d46edea99 +size 432198 diff --git a/vlm/dev/Qx8lUU8CzQ/2.png b/vlm/dev/Qx8lUU8CzQ/2.png new file mode 100644 index 0000000000000000000000000000000000000000..b680f416ce66fce62436b2287a684ca8c1d2f677 --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea1d87989a6f44ca656561f187a50848f73ae57d19ceb5f9cd92fe13fb2c1fb4 +size 558393 diff --git a/vlm/dev/Qx8lUU8CzQ/3.png b/vlm/dev/Qx8lUU8CzQ/3.png new file mode 100644 index 0000000000000000000000000000000000000000..3faad11375724de9ca7d33b71539d207ef136773 --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b13d62b97b5f06506db5d1efb37c43e61b158dcb3bd461c1a244f10b08e473ec +size 541459 diff --git a/vlm/dev/Qx8lUU8CzQ/4.png b/vlm/dev/Qx8lUU8CzQ/4.png new file mode 100644 index 0000000000000000000000000000000000000000..9cf93991d1b5fd1146d6510a493ee378d6ba043d --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d4c31ac63415e8164f778f5ea491ac6d70258df787bfada2bf12c03e8a933cd +size 514076 diff --git a/vlm/dev/Qx8lUU8CzQ/5.png b/vlm/dev/Qx8lUU8CzQ/5.png new file mode 100644 index 0000000000000000000000000000000000000000..5288c6cc9f99e2e9e19404ea9b5f194297162acb --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb0a11793d6bc13509ef077541ce0e32e58d6156677cfa6ebe664adc5d67b2cb +size 515103 diff --git a/vlm/dev/Qx8lUU8CzQ/6.png b/vlm/dev/Qx8lUU8CzQ/6.png new file mode 100644 index 0000000000000000000000000000000000000000..3eb2c0f434dfe328a9c9660d46b618b99fccf03c --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a0d36142b5e892f7a0571fabec372ccf362bdbc217cc758b25d3a9b8b54ae7a +size 781750 diff --git a/vlm/dev/Qx8lUU8CzQ/7.png b/vlm/dev/Qx8lUU8CzQ/7.png new file mode 100644 index 0000000000000000000000000000000000000000..ae682d034238d53a2357a5a5c5ed6a9973fbbea1 --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c006bda18d584ad288a7cbec28301dda0abb7342f09ce523270f5a8ec212d3a +size 683468 diff --git a/vlm/dev/Qx8lUU8CzQ/8.png b/vlm/dev/Qx8lUU8CzQ/8.png new file mode 100644 index 0000000000000000000000000000000000000000..9db4d9a1605c2b713df028e5ed5896da0d873854 --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c43450cd044d2d3d77a6687156309617ba8583d5775665b2c6e61d8a91d9e3e +size 642526 diff --git a/vlm/dev/Qx8lUU8CzQ/9.png b/vlm/dev/Qx8lUU8CzQ/9.png new file mode 100644 index 0000000000000000000000000000000000000000..a2942362e2f476ca298e628e559ddd01781151cb --- /dev/null +++ b/vlm/dev/Qx8lUU8CzQ/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38839029b71e61cfd7ff8b5f9ad0d29b5cde6b71afa95ee480f9416b984715ff +size 491314 diff --git a/vlm/dev/dMK7EwoTYp/0.png b/vlm/dev/dMK7EwoTYp/0.png new file mode 100644 index 0000000000000000000000000000000000000000..6d80656bb555231f86eaff7e82da4b9fb00cc443 --- /dev/null +++ b/vlm/dev/dMK7EwoTYp/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0a90e2c36fa8d1bc1131357248223fd2853d595b67cbbcdab4dd25634138948 +size 456704 diff --git a/vlm/dev/dMK7EwoTYp/1.png b/vlm/dev/dMK7EwoTYp/1.png new file mode 100644 index 0000000000000000000000000000000000000000..d7277ebd684a82e724a6ff0d30fde6506f245233 --- /dev/null +++ b/vlm/dev/dMK7EwoTYp/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c38f621f73f3c94b49a90d420bfb8c0de017bb41ce942e283c26680dc98bbefd +size 830102 diff --git a/vlm/dev/dMK7EwoTYp/10.png b/vlm/dev/dMK7EwoTYp/10.png new file mode 100644 index 0000000000000000000000000000000000000000..2fc06fea58367ad5d1c225b6b557c9d35b7fb237 --- /dev/null +++ b/vlm/dev/dMK7EwoTYp/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94b741e9c2b825ee318f6000a50a7ec4ff95949fb3d94e8af260cf870c77ada2 +size 641020 diff --git a/vlm/dev/dMK7EwoTYp/11.png b/vlm/dev/dMK7EwoTYp/11.png new file mode 100644 index 0000000000000000000000000000000000000000..2ebec35910594ad0d6f9034df37ebb429788aa15 --- /dev/null +++ b/vlm/dev/dMK7EwoTYp/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43e13014e4c7ed06e9cf83143463d68f9c651990ce4b5942af189684440241ea +size 682997 diff --git a/vlm/dev/dMK7EwoTYp/12.png b/vlm/dev/dMK7EwoTYp/12.png new file mode 100644 index 0000000000000000000000000000000000000000..83ca736d0190e7cbd7766eccf00eac3f10038da3 --- /dev/null +++ b/vlm/dev/dMK7EwoTYp/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07e15223659b34bfde4064ed9d2b5e628e4cdafd616e05ec3dd4f0b364bd564c +size 668590 diff --git a/vlm/dev/dMK7EwoTYp/13.png b/vlm/dev/dMK7EwoTYp/13.png new file mode 100644 index 0000000000000000000000000000000000000000..72cc009495de7a88f164559959afebe221a1b336 --- /dev/null +++ b/vlm/dev/dMK7EwoTYp/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c34062133440ad72d7eb5b3e4d32c0fdd65d28cb318423c7ab886f292ccdce6 +size 610769 diff --git a/vlm/dev/dMK7EwoTYp/14.png b/vlm/dev/dMK7EwoTYp/14.png new file mode 100644 index 0000000000000000000000000000000000000000..5263c426a7ee42dfa549980dbe4f2d40c75e8600 --- /dev/null +++ b/vlm/dev/dMK7EwoTYp/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efc95b9a70d59c166eb0cf066a574dda0abb4e9b7c0bc05a61afa3ad8d9c8a69 +size 245652 diff --git a/vlm/dev/dMK7EwoTYp/2.png b/vlm/dev/dMK7EwoTYp/2.png new file mode 100644 index 0000000000000000000000000000000000000000..55574d8323a1864616ba376d4041193428e0ae25 --- /dev/null +++ b/vlm/dev/dMK7EwoTYp/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c85cfae1fc26cb64041230352916e633fe921d14aae5cb002b9445d0e9debac7 +size 586230 diff --git a/vlm/dev/dMK7EwoTYp/3.png b/vlm/dev/dMK7EwoTYp/3.png new file mode 100644 index 0000000000000000000000000000000000000000..ce032d89e324329bb3c373073d74deb8cae2ebaf --- /dev/null +++ b/vlm/dev/dMK7EwoTYp/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3cfb8271e02d4c4ed786b39ee20f374669e0df0c22f1b1c31a34a712c73b32a4 +size 588093 diff --git a/vlm/dev/dMK7EwoTYp/4.png b/vlm/dev/dMK7EwoTYp/4.png new file mode 100644 index 0000000000000000000000000000000000000000..e959cff521c771ca13d14f86063b00372528effc --- /dev/null +++ b/vlm/dev/dMK7EwoTYp/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6671c5ff4d7d4f73e81e64631b0f0152ea60f4128828c71cc68e768b94bdef8d +size 453370 diff --git a/vlm/dev/dMK7EwoTYp/5.png b/vlm/dev/dMK7EwoTYp/5.png new file mode 100644 index 0000000000000000000000000000000000000000..5b96828928c5a02cd45d3ff70d9b599ee9fc85fe --- /dev/null +++ b/vlm/dev/dMK7EwoTYp/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9be88501c96d56bdf9a037076b4204f3ba199bd553bea95788677918b853cc9 +size 470434 diff --git a/vlm/dev/dMK7EwoTYp/6.png b/vlm/dev/dMK7EwoTYp/6.png new file mode 100644 index 0000000000000000000000000000000000000000..ce88402f645c737c9699e2033c38537a559d08a6 --- /dev/null +++ b/vlm/dev/dMK7EwoTYp/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:92ac1a0293c5a012a81b8e926d993c0055c0cc423d8e731198bd9469644c4876 +size 657803 diff --git a/vlm/dev/dMK7EwoTYp/7.png b/vlm/dev/dMK7EwoTYp/7.png new file mode 100644 index 0000000000000000000000000000000000000000..721f1697e3866db02d43e0fe187d04ce7cbde8b2 --- /dev/null +++ b/vlm/dev/dMK7EwoTYp/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13924dffb934ba0d983b95da770e8b6ecd9c8437a1fad6e593f109a051857873 +size 637315 diff --git a/vlm/dev/dMK7EwoTYp/8.png b/vlm/dev/dMK7EwoTYp/8.png new file mode 100644 index 0000000000000000000000000000000000000000..221f78cc51714bcfbe82945943c2ae3017139e7e --- /dev/null +++ b/vlm/dev/dMK7EwoTYp/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6619b07a614092a84ea22ca7096ed7ad97619a1bf34bae7eafb7baf99ee50688 +size 677180 diff --git a/vlm/dev/dMK7EwoTYp/9.png b/vlm/dev/dMK7EwoTYp/9.png new file mode 100644 index 0000000000000000000000000000000000000000..d265af314ac8ba319769480a9d20df7279a6c477 --- /dev/null +++ b/vlm/dev/dMK7EwoTYp/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed2ccccb43f58de26f167ea3788a43e281f4a4ce7295db72275b0de2aa339cdb +size 669501 diff --git a/vlm/dev/gUL6zYN4Uaf/10.png b/vlm/dev/gUL6zYN4Uaf/10.png new file mode 100644 index 0000000000000000000000000000000000000000..2f2a47094eefe4836003f0bd5b3479809ce228fc --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81053ee584ba8fc62af992d70739679661de7bbf177a62290090ed844abbd35e +size 681578 diff --git a/vlm/dev/gUL6zYN4Uaf/6.png b/vlm/dev/gUL6zYN4Uaf/6.png new file mode 100644 index 0000000000000000000000000000000000000000..a2df3f663205f698c018918c455db0545ae4ca43 --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6cf1a6962971feff6b5ee20a3b371e01a48123a2b891de49e1273abbdb08aa9 +size 615546 diff --git a/vlm/dev/hQwb-lbM6EL/0.png b/vlm/dev/hQwb-lbM6EL/0.png new file mode 100644 index 0000000000000000000000000000000000000000..128c53b075c6f37484999f31bc3c61240d666317 --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c3f040710b20eda8d1f5ec4a8c8b236b8fff2f9861b159646c6dcb2a9a5e4a8 +size 500927 diff --git a/vlm/dev/hQwb-lbM6EL/1.png b/vlm/dev/hQwb-lbM6EL/1.png new file mode 100644 index 0000000000000000000000000000000000000000..9b22cf066dd3bf308170946cd3cb8cc3fa960d9c --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d0824b0b585d816f3d1ac6689ff509297a06ebbf50b966d1cfdf28c73a81682 +size 435578 diff --git a/vlm/dev/hQwb-lbM6EL/10.png b/vlm/dev/hQwb-lbM6EL/10.png new file mode 100644 index 0000000000000000000000000000000000000000..8aa2370c4969cc3481be8878465bf8f17e543c66 --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6a46e9afa64eb1c72c4f220beb62d085a7f3072ff108fe3968e53b9ec185502 +size 540580 diff --git a/vlm/dev/hQwb-lbM6EL/11.png b/vlm/dev/hQwb-lbM6EL/11.png new file mode 100644 index 0000000000000000000000000000000000000000..ee38976c133fe0b069c66e0b1fb885c10f667e1a --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7111585b4284f6b7b3043fcb4911e19df3f6efc77897dab9e1040232b619b2c +size 533410 diff --git a/vlm/dev/hQwb-lbM6EL/12.png b/vlm/dev/hQwb-lbM6EL/12.png new file mode 100644 index 0000000000000000000000000000000000000000..1ec4bd2f8359eebf3934d370d5516c4a074fcbfa --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f80b46e0032ee3c2df0908dc37225f1f0c915d614d98c2b2dabb655f61e18bfa +size 517263 diff --git a/vlm/dev/hQwb-lbM6EL/13.png b/vlm/dev/hQwb-lbM6EL/13.png new file mode 100644 index 0000000000000000000000000000000000000000..9859fe39801b94fd6914ec37ecfdcfe7a731e2ec --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9fc52d1bc4af48d0185755d0aac73c15a50398d07342c3a245181a74e65d1cf0 +size 507735 diff --git a/vlm/dev/hQwb-lbM6EL/14.png b/vlm/dev/hQwb-lbM6EL/14.png new file mode 100644 index 0000000000000000000000000000000000000000..036d6f0b91e47b3095bae4f0702dfb443fb87a52 --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24f128781580b2a001294e54ab53d412a64a01a90fe38bfc25acff1397e5d1f2 +size 416253 diff --git a/vlm/dev/hQwb-lbM6EL/15.png b/vlm/dev/hQwb-lbM6EL/15.png new file mode 100644 index 0000000000000000000000000000000000000000..7acc8728866ba7a946a08bb82c8492df33a9fd61 --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffd767cf347dc9f3378ec953024fb974438bff0ca0bc295a0fb172f2e415c6f0 +size 319571 diff --git a/vlm/dev/hQwb-lbM6EL/16.png b/vlm/dev/hQwb-lbM6EL/16.png new file mode 100644 index 0000000000000000000000000000000000000000..8efd9f9de728768e446969cdfb8e577db91bf1cb --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26e18e573949aa924566668ed5e19be56a6f87940bcda279db6010427bf171ca +size 471579 diff --git a/vlm/dev/hQwb-lbM6EL/17.png b/vlm/dev/hQwb-lbM6EL/17.png new file mode 100644 index 0000000000000000000000000000000000000000..054aba9bdb12a8184d4b68e07d9aad9e649b0739 --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5b3908aeb2d87c755a6b10af691935ae4f032446678ade9462dea726055a660 +size 464300 diff --git a/vlm/dev/hQwb-lbM6EL/18.png b/vlm/dev/hQwb-lbM6EL/18.png new file mode 100644 index 0000000000000000000000000000000000000000..550ec36764317b8e51e011c98d483ccf6db645d7 --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b2898a270869954f6643fa53e20911123056899ad694cf202beffb2fa7ba2a2 +size 444409 diff --git a/vlm/dev/hQwb-lbM6EL/19.png b/vlm/dev/hQwb-lbM6EL/19.png new file mode 100644 index 0000000000000000000000000000000000000000..600a61900d33f7f65cdfe492f3586326f994b054 --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56570a0e7ae663131cc8a0d8f3d0b97efe8b7fba3545932a1baf9fba8304a73d +size 494934 diff --git a/vlm/dev/hQwb-lbM6EL/2.png b/vlm/dev/hQwb-lbM6EL/2.png new file mode 100644 index 0000000000000000000000000000000000000000..c4ee0bc3154812d4be40b5429533d9ed8922378b --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37ba29cab6218f00726a04ab0a46b93677da826d4b30c0d1ef975bb2a8e0ace9 +size 542268 diff --git a/vlm/dev/hQwb-lbM6EL/20.png b/vlm/dev/hQwb-lbM6EL/20.png new file mode 100644 index 0000000000000000000000000000000000000000..0d5cdb912111c9bcb46324e5d67fe3cf0e263cdf --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14f40fa6170024d37eed7e54d166603749955072828896ed61cc7cb886651781 +size 435160 diff --git a/vlm/dev/hQwb-lbM6EL/21.png b/vlm/dev/hQwb-lbM6EL/21.png new file mode 100644 index 0000000000000000000000000000000000000000..0a1b7cc44c640e8af2413ec409e54345d409d1e6 --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c64b3f56ce19c467a9c8800509841a9a0cfed235429050787a1305190c4f8c30 +size 260037 diff --git a/vlm/dev/hQwb-lbM6EL/22.png b/vlm/dev/hQwb-lbM6EL/22.png new file mode 100644 index 0000000000000000000000000000000000000000..1f5bc63fafd146fb3f0f31911dfb40de9132c32c --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ceac49e11b8aea6703218aad119a59c625b6f14bc8a15c0c99a8a7f9a178e28 +size 257506 diff --git a/vlm/dev/hQwb-lbM6EL/23.png b/vlm/dev/hQwb-lbM6EL/23.png new file mode 100644 index 0000000000000000000000000000000000000000..15ddf1032d1a37a363c12a85d1a7151b51a3fb1b --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:091c5d153429a4a94a04eab6cca96dfa29e240b7844f6c572815fef503bf3bd7 +size 218454 diff --git a/vlm/dev/hQwb-lbM6EL/24.png b/vlm/dev/hQwb-lbM6EL/24.png new file mode 100644 index 0000000000000000000000000000000000000000..15e3a034621617209a5d71fadcd0a070ec88f7fe --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75c51137bca6e9a64e8c6e9b61a8005512a5965d0314e234228752b565de0b91 +size 289749 diff --git a/vlm/dev/hQwb-lbM6EL/25.png b/vlm/dev/hQwb-lbM6EL/25.png new file mode 100644 index 0000000000000000000000000000000000000000..37aea7969718a582114089dbefb340a1b2035d86 --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a503b79b4877ebe0f717ea1befbebf278503594d3443ce2e9af3788da8120a82 +size 323083 diff --git a/vlm/dev/hQwb-lbM6EL/3.png b/vlm/dev/hQwb-lbM6EL/3.png new file mode 100644 index 0000000000000000000000000000000000000000..63b987a2eb35bd1259c3cd67cb3dc9094baea35b --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a833248762764bedc6074b28869924367d239485e25904ca66be567b7fd2a2b +size 570916 diff --git a/vlm/dev/hQwb-lbM6EL/4.png b/vlm/dev/hQwb-lbM6EL/4.png new file mode 100644 index 0000000000000000000000000000000000000000..682c63da7ac6c649043a72e182e56b1a23afd1db --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7181fd6f481603dabc2d8d99ab29639ef7aa35e0e245ce5e0b8b09e865f67f77 +size 485667 diff --git a/vlm/dev/hQwb-lbM6EL/5.png b/vlm/dev/hQwb-lbM6EL/5.png new file mode 100644 index 0000000000000000000000000000000000000000..7ec00a6aaff2b76509170ca3477041e973979266 --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c86bd8246248b061738c2aa6ba8d46cf836a5c8a566883894324d6d71e447710 +size 508335 diff --git a/vlm/dev/hQwb-lbM6EL/6.png b/vlm/dev/hQwb-lbM6EL/6.png new file mode 100644 index 0000000000000000000000000000000000000000..a9103ca679513e0ca846e2e80a6850862c96698f --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:905b029e65d1f460652bf02d9edc6da2854648aeae310f321147b6c1fbaeec5f +size 565214 diff --git a/vlm/dev/hQwb-lbM6EL/7.png b/vlm/dev/hQwb-lbM6EL/7.png new file mode 100644 index 0000000000000000000000000000000000000000..158717bea120ea1f08d090cf7b51d98e7bd0388f --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c59b31d8b56c24a01678ac5b7c26bad1391cac46d29d98bf3bf3baa11efa9cc +size 498235 diff --git a/vlm/dev/hQwb-lbM6EL/8.png b/vlm/dev/hQwb-lbM6EL/8.png new file mode 100644 index 0000000000000000000000000000000000000000..5e713a0f0e74d46834f1ed20ce42d309cce702d0 --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ad3fd1a99c6fcb6ba7e1c87b4c88a31485a868224b52b05dc6d339eea9b0707 +size 607792 diff --git a/vlm/dev/hQwb-lbM6EL/9.png b/vlm/dev/hQwb-lbM6EL/9.png new file mode 100644 index 0000000000000000000000000000000000000000..9cfb7785f22c0b71f16473a886e878b867f29e90 --- /dev/null +++ b/vlm/dev/hQwb-lbM6EL/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:699dcd9b863156ef7117a677e512c1fc4b3d62b7397bf7005ef94f1da57780f3 +size 543381 diff --git a/vlm/dev/nO5caZwFwYu/0.png b/vlm/dev/nO5caZwFwYu/0.png new file mode 100644 index 0000000000000000000000000000000000000000..056a6eed30d071d1c98f7671154462e55075ff6b --- /dev/null +++ b/vlm/dev/nO5caZwFwYu/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:936b5d3fcc978a989455acaab796e91e9b5687a485c4d7ba49ed65f237a89ec6 +size 467566 diff --git a/vlm/dev/nO5caZwFwYu/1.png b/vlm/dev/nO5caZwFwYu/1.png new file mode 100644 index 0000000000000000000000000000000000000000..ab0b5205be12e1abb711d43b61852ab62f5d3f97 --- /dev/null +++ b/vlm/dev/nO5caZwFwYu/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41840a838b7890098ccb47b138c2ee7d4062bd77988b249d4c90adf7a28ceb81 +size 608306 diff --git a/vlm/dev/nO5caZwFwYu/10.png b/vlm/dev/nO5caZwFwYu/10.png new file mode 100644 index 0000000000000000000000000000000000000000..d865ab82f76ecbbcc03b5438590b4557470c0333 --- /dev/null +++ b/vlm/dev/nO5caZwFwYu/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7bdee600057eb96ced4bb4c78f69d4f938a6a5d5c50cf8769dd8ce81ae772c62 +size 546683 diff --git a/vlm/dev/nO5caZwFwYu/11.png b/vlm/dev/nO5caZwFwYu/11.png new file mode 100644 index 0000000000000000000000000000000000000000..8cf54a8277ca611d5826cf9c954a2bb03e8ef370 --- /dev/null +++ b/vlm/dev/nO5caZwFwYu/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c202d09987fa0b59917a5f0ebb1cc65f69b1145d15a718007762c41218fc69ff +size 588316 diff --git a/vlm/dev/nO5caZwFwYu/12.png b/vlm/dev/nO5caZwFwYu/12.png new file mode 100644 index 0000000000000000000000000000000000000000..41c42eb1e9fea9a4028747fd905eea2d9600e91a --- /dev/null +++ b/vlm/dev/nO5caZwFwYu/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:250f6ec3378abbeb4ffc8a01ec03622cb23764d70799faf1a0a764ec49a43713 +size 215767 diff --git a/vlm/dev/nO5caZwFwYu/2.png b/vlm/dev/nO5caZwFwYu/2.png new file mode 100644 index 0000000000000000000000000000000000000000..9cda35ae982bca3e5d7be5d08c37ad77609a01e5 --- /dev/null +++ b/vlm/dev/nO5caZwFwYu/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73907346d6caf59ea5cdfb737bcd0b7e641df268f4b0c30c6c3dc689493d15da +size 616153 diff --git a/vlm/dev/nO5caZwFwYu/3.png b/vlm/dev/nO5caZwFwYu/3.png new file mode 100644 index 0000000000000000000000000000000000000000..57ec334f06f447ebfd52274843fda20d9b94ed03 --- /dev/null +++ b/vlm/dev/nO5caZwFwYu/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6bc93feae5b2984fd79021ff8084f5bfb4439f921f6c78380dc043daa3a6a904 +size 450465 diff --git a/vlm/dev/nO5caZwFwYu/4.png b/vlm/dev/nO5caZwFwYu/4.png new file mode 100644 index 0000000000000000000000000000000000000000..2ff26b09520382f5d3c2321cf5bcbd09a28a508e --- /dev/null +++ b/vlm/dev/nO5caZwFwYu/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e53b8d81d11f371a447d094df7041f148975af88fe3d288752a55dd46b0faa91 +size 562808 diff --git a/vlm/dev/nO5caZwFwYu/5.png b/vlm/dev/nO5caZwFwYu/5.png new file mode 100644 index 0000000000000000000000000000000000000000..7333f03181442f8a0951ebc31e5ae2c3a8e703c8 --- /dev/null +++ b/vlm/dev/nO5caZwFwYu/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9951cd93a4ec12d4156737d9f870fad9fd2210095f01011cc89956360b339c84 +size 603161 diff --git a/vlm/dev/nO5caZwFwYu/6.png b/vlm/dev/nO5caZwFwYu/6.png new file mode 100644 index 0000000000000000000000000000000000000000..d70d618b82077edd6b62e392c049b579d8983b6a --- /dev/null +++ b/vlm/dev/nO5caZwFwYu/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34a04f95e2d2af0e675decd721f188fbb7a2445d3abfade63f851aadfd3608d0 +size 600437 diff --git a/vlm/dev/nO5caZwFwYu/7.png b/vlm/dev/nO5caZwFwYu/7.png new file mode 100644 index 0000000000000000000000000000000000000000..f8d7d935c750c7e1d18d76e7d3a45d0e1b276ede --- /dev/null +++ b/vlm/dev/nO5caZwFwYu/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be04fdc339ad29b3ec6304c1963c34cb97d5d3fb6edc337148f2419295edcc25 +size 570434 diff --git a/vlm/dev/nO5caZwFwYu/8.png b/vlm/dev/nO5caZwFwYu/8.png new file mode 100644 index 0000000000000000000000000000000000000000..709be926fdd090b7dca59f98b40387e2eeace4b9 --- /dev/null +++ b/vlm/dev/nO5caZwFwYu/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a80f9d44e12671aef9e40b88cca0142862ce199478e0bddc6a78bcde4e7087a1 +size 511976 diff --git a/vlm/dev/nO5caZwFwYu/9.png b/vlm/dev/nO5caZwFwYu/9.png new file mode 100644 index 0000000000000000000000000000000000000000..96bfe5d5b629b93a6a724e5448b41de610c63816 --- /dev/null +++ b/vlm/dev/nO5caZwFwYu/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eced9acb23ea5b6b640ac87be577e7f0af58c2cf9d25f96cc6701a17292e966f +size 524921 diff --git a/vlm/dev/p4RvNzlJX7W/0.png b/vlm/dev/p4RvNzlJX7W/0.png new file mode 100644 index 0000000000000000000000000000000000000000..205a63ac3cdf59c7379d459430cb01d8aae4f8df --- /dev/null +++ b/vlm/dev/p4RvNzlJX7W/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a020ba472919ef69f6d44d2e574a9e76c99a00d27f07f938b74563905e1d03d +size 538418 diff --git a/vlm/dev/p4RvNzlJX7W/10.png b/vlm/dev/p4RvNzlJX7W/10.png new file mode 100644 index 0000000000000000000000000000000000000000..c7e6ae16a5a0f333fc5cd2d012afbf32d8d3af07 --- /dev/null +++ b/vlm/dev/p4RvNzlJX7W/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41d3f31cb77d0c7843e29d21a1da4a5e6c05d5d6bc5bb1eef6d712cdfb36e7f2 +size 538712 diff --git a/vlm/dev/p4RvNzlJX7W/11.png b/vlm/dev/p4RvNzlJX7W/11.png new file mode 100644 index 0000000000000000000000000000000000000000..7260a5928121723872463f7148216838db492873 --- /dev/null +++ b/vlm/dev/p4RvNzlJX7W/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59f08a8332603b10ff0c6d71f5f8dc4e52996f4c9ae8b339bfb6bfe5627b0090 +size 62205 diff --git a/vlm/dev/p4RvNzlJX7W/14.png b/vlm/dev/p4RvNzlJX7W/14.png new file mode 100644 index 0000000000000000000000000000000000000000..d404d7b5a5ddcda4a86efa06d925b2d3f98dfd90 --- /dev/null +++ b/vlm/dev/p4RvNzlJX7W/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24e9d45ddf797f1d5130ee3c558e812fc1a528fe92a035a3c24c702ea8732e8a +size 337349 diff --git a/vlm/dev/p4RvNzlJX7W/15.png b/vlm/dev/p4RvNzlJX7W/15.png new file mode 100644 index 0000000000000000000000000000000000000000..f1366b838e10742764a6574927cebea554a5a466 --- /dev/null +++ b/vlm/dev/p4RvNzlJX7W/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd685dbd8eddb74cd70c5b99e46fe09865fc6051d94e8fbe167990230b6da691 +size 250945 diff --git a/vlm/dev/p4RvNzlJX7W/17.png b/vlm/dev/p4RvNzlJX7W/17.png new file mode 100644 index 0000000000000000000000000000000000000000..c8deceb9fa90b3e201edf7c978d1a353b7cf8f62 --- /dev/null +++ b/vlm/dev/p4RvNzlJX7W/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b94dd33b6c27d39bb34176b50e7f9c5d4bfc1ba252ec8d3923f032ab246aef46 +size 372710 diff --git a/vlm/dev/p4RvNzlJX7W/2.png b/vlm/dev/p4RvNzlJX7W/2.png new file mode 100644 index 0000000000000000000000000000000000000000..0c532bce82e7ae186a8afa1280bd7cb0088725c5 --- /dev/null +++ b/vlm/dev/p4RvNzlJX7W/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2eb87f3e91561e28f0d07448ddf744fc9ab1d2565b3c8ede03f97716a9fa59e +size 618061 diff --git a/vlm/dev/p4RvNzlJX7W/3.png b/vlm/dev/p4RvNzlJX7W/3.png new file mode 100644 index 0000000000000000000000000000000000000000..9e049b1493a955f1882c45101a0f018c21ce852b --- /dev/null +++ b/vlm/dev/p4RvNzlJX7W/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca897d4baddc9617a63e5e4779bb133c729452c0726d46c1f4024422141f4277 +size 594648 diff --git a/vlm/dev/p4RvNzlJX7W/9.png b/vlm/dev/p4RvNzlJX7W/9.png new file mode 100644 index 0000000000000000000000000000000000000000..3f96812c13e56e78c9b3f13fffac6a79c863d8ce --- /dev/null +++ b/vlm/dev/p4RvNzlJX7W/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71084eb92c009b3a1d4524742b2c4efb142ceab21274731e6e849d7b935255cc +size 551729 diff --git a/vlm/dev/wlEOsQ917F/0.png b/vlm/dev/wlEOsQ917F/0.png new file mode 100644 index 0000000000000000000000000000000000000000..5ffa5b5adb199de0e85e54e587ec7bd35fa33db3 --- /dev/null +++ b/vlm/dev/wlEOsQ917F/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5eb04ac6a2b0e0921edec74ca19a6b90f30881dce9de657a9896d29dc8f13627 +size 413086 diff --git a/vlm/dev/wlEOsQ917F/1.png b/vlm/dev/wlEOsQ917F/1.png new file mode 100644 index 0000000000000000000000000000000000000000..0d3285401aa3cdce58f857d35d07a6f678f8958c --- /dev/null +++ b/vlm/dev/wlEOsQ917F/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1bfcedf9c841b358350e08ff087dee8809e5af1c0ce177cfb26aeb9dc0603492 +size 634152 diff --git a/vlm/dev/wlEOsQ917F/10.png b/vlm/dev/wlEOsQ917F/10.png new file mode 100644 index 0000000000000000000000000000000000000000..40707dd889b456ff5cdc36e0554c7678c1f32511 --- /dev/null +++ b/vlm/dev/wlEOsQ917F/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9b5eee40e488731c3bc7eead0808198dfa914ff8f7367d7069068d9e215ea3b +size 556036 diff --git a/vlm/dev/wlEOsQ917F/11.png b/vlm/dev/wlEOsQ917F/11.png new file mode 100644 index 0000000000000000000000000000000000000000..5c4d9a6a44df39695e112b8b6e5ccc2fd7322ad3 --- /dev/null +++ b/vlm/dev/wlEOsQ917F/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57cc93e03d8ab9d693f8e03ad957b1c679b5e11efb34abb4e7f98a53bcc4c779 +size 420082 diff --git a/vlm/dev/wlEOsQ917F/12.png b/vlm/dev/wlEOsQ917F/12.png new file mode 100644 index 0000000000000000000000000000000000000000..5143a6698c86b9358bf012e255e54a379440d3cd --- /dev/null +++ b/vlm/dev/wlEOsQ917F/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ea77f402251bb5e54e5827e011ffc151a83998d2c3a81ddea3c5f3ba809929d +size 327743 diff --git a/vlm/dev/wlEOsQ917F/2.png b/vlm/dev/wlEOsQ917F/2.png new file mode 100644 index 0000000000000000000000000000000000000000..100634dc850a1d493064e59d148781315f7e1775 --- /dev/null +++ b/vlm/dev/wlEOsQ917F/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4385c10ba47128b14dbf860e3cccb784d337bf063168e3c29295b9d487b47985 +size 628167 diff --git a/vlm/dev/wlEOsQ917F/3.png b/vlm/dev/wlEOsQ917F/3.png new file mode 100644 index 0000000000000000000000000000000000000000..54b28f7cb0bc03a225301048be55c07c96e4dc7b --- /dev/null +++ b/vlm/dev/wlEOsQ917F/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76743b13b56b1c7419da6bcb632138ed6bf8495776c2e502ca7a0e38f44f9857 +size 570199 diff --git a/vlm/dev/wlEOsQ917F/4.png b/vlm/dev/wlEOsQ917F/4.png new file mode 100644 index 0000000000000000000000000000000000000000..4191718a332d33f128133f82100e706680806f39 --- /dev/null +++ b/vlm/dev/wlEOsQ917F/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ed0ac1863d62830d0f72e0ef663e990863c01f50e582a28a99aacf799b8e080 +size 568441 diff --git a/vlm/dev/wlEOsQ917F/5.png b/vlm/dev/wlEOsQ917F/5.png new file mode 100644 index 0000000000000000000000000000000000000000..7c0a4d03cabfb7240c77dfa9caee7c0a295ef649 --- /dev/null +++ b/vlm/dev/wlEOsQ917F/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b89b5f24da8e9b8ec0ac1ebbe390876c73c98f15f9bedadeb6fba57f0e87ba98 +size 549069 diff --git a/vlm/dev/wlEOsQ917F/6.png b/vlm/dev/wlEOsQ917F/6.png new file mode 100644 index 0000000000000000000000000000000000000000..83fff1601dc3c64c431f86d8b5f8c1f9f7fc7926 --- /dev/null +++ b/vlm/dev/wlEOsQ917F/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f72408fcb9373ca91526f9bdd8c13315bf5134ea57049fa80ade96e1e06d4b2b +size 464736 diff --git a/vlm/dev/wlEOsQ917F/7.png b/vlm/dev/wlEOsQ917F/7.png new file mode 100644 index 0000000000000000000000000000000000000000..4791cd8dfe76c27157b423faa1d39b4e1015b857 --- /dev/null +++ b/vlm/dev/wlEOsQ917F/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28686721d48bf8bea763b9af188a2978f9f4b23f37681065d435b7905996f857 +size 577235 diff --git a/vlm/dev/wlEOsQ917F/8.png b/vlm/dev/wlEOsQ917F/8.png new file mode 100644 index 0000000000000000000000000000000000000000..3a11e18a28c85ac981f2fc84f9a9dfb063c878f8 --- /dev/null +++ b/vlm/dev/wlEOsQ917F/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bce375d2324d32c8701501920affa916760faaa3beebfe361e639dd8f3f7a0ae +size 564602 diff --git a/vlm/dev/wlEOsQ917F/9.png b/vlm/dev/wlEOsQ917F/9.png new file mode 100644 index 0000000000000000000000000000000000000000..1ca6c75a81c92d3e351dcd682ddd8a8a234aa42a --- /dev/null +++ b/vlm/dev/wlEOsQ917F/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a0332e8bd062c4fe7a4d81c5d6eb3722a7b6b52ab96a6a9b4fb8e71c9a06428 +size 577705 diff --git a/vlm/dev/wv79UiY5U7/0.png b/vlm/dev/wv79UiY5U7/0.png new file mode 100644 index 0000000000000000000000000000000000000000..15f6a3d6489523661764fb261ea3ab80c15742a2 --- /dev/null +++ b/vlm/dev/wv79UiY5U7/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3537fa76a9afd38b552d3b29a7090a44c21b7d8708b411329eef2899b2405b76 +size 445356 diff --git a/vlm/dev/wv79UiY5U7/1.png b/vlm/dev/wv79UiY5U7/1.png new file mode 100644 index 0000000000000000000000000000000000000000..706a8dda8b125ba8adc383154cc18ba68f99d1f4 --- /dev/null +++ b/vlm/dev/wv79UiY5U7/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c26fbb2d6d1e302ca548ec1ae609cda8b972db1fa45e024fceec6fce2dc6d924 +size 694990 diff --git a/vlm/dev/wv79UiY5U7/10.png b/vlm/dev/wv79UiY5U7/10.png new file mode 100644 index 0000000000000000000000000000000000000000..cfe3ee5f41346ea58acbb4293f0d8fbfe3e423b0 --- /dev/null +++ b/vlm/dev/wv79UiY5U7/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4389b5aa0df70ee785b824bf076dafcdfeb103f77471a412eca0e9da1e6e0273 +size 591506 diff --git a/vlm/dev/wv79UiY5U7/11.png b/vlm/dev/wv79UiY5U7/11.png new file mode 100644 index 0000000000000000000000000000000000000000..ee0019767e36e64f430c86b2a092ee366f6d0f55 --- /dev/null +++ b/vlm/dev/wv79UiY5U7/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:453d6222d5ba170e3212ae083f8d9c767760ca8e0788e35ccc056b3098d40fd0 +size 609849 diff --git a/vlm/dev/wv79UiY5U7/12.png b/vlm/dev/wv79UiY5U7/12.png new file mode 100644 index 0000000000000000000000000000000000000000..6055a83da76493b4f524b6305199019c8757dc1b --- /dev/null +++ b/vlm/dev/wv79UiY5U7/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2f0400dc097a31eba0c3b1fb535af5f7aba54f668e60dba9ca94afe369dabd7 +size 590943 diff --git a/vlm/dev/wv79UiY5U7/2.png b/vlm/dev/wv79UiY5U7/2.png new file mode 100644 index 0000000000000000000000000000000000000000..b01dcaf1499f97a6b964343e00d0c5c007fb1a69 --- /dev/null +++ b/vlm/dev/wv79UiY5U7/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d3079c30a43ece1ad42a68443d0db3f14a8cdecaead8761d0d6e69afcd973d2 +size 566677 diff --git a/vlm/dev/wv79UiY5U7/3.png b/vlm/dev/wv79UiY5U7/3.png new file mode 100644 index 0000000000000000000000000000000000000000..b916b415c1a65a9b07abccda124324ca6a78630d --- /dev/null +++ b/vlm/dev/wv79UiY5U7/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b13cac6c00854f579006a72720232abb65d3ab69ec8327e61a39e8e4b2d4667a +size 674368 diff --git a/vlm/dev/wv79UiY5U7/4.png b/vlm/dev/wv79UiY5U7/4.png new file mode 100644 index 0000000000000000000000000000000000000000..99e1177dea9ff6b926b0df6af1d5227c2cfcec34 --- /dev/null +++ b/vlm/dev/wv79UiY5U7/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3bbdc9593784a3e6a26851e97655e5cec5af7abbfdaf17b90a03f9a6cb94d52 +size 657584 diff --git a/vlm/dev/wv79UiY5U7/5.png b/vlm/dev/wv79UiY5U7/5.png new file mode 100644 index 0000000000000000000000000000000000000000..e5b6cefb1d0a718d4b63ba5b27ad05f761ab55de --- /dev/null +++ b/vlm/dev/wv79UiY5U7/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72ccd5da7acf1737cc01574f7c5272c6b7a9d9e941a5b44829a85bebf77a3fd1 +size 545917 diff --git a/vlm/dev/wv79UiY5U7/6.png b/vlm/dev/wv79UiY5U7/6.png new file mode 100644 index 0000000000000000000000000000000000000000..8f39abf7782573542dfa71bc74e3497c425fff08 --- /dev/null +++ b/vlm/dev/wv79UiY5U7/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45c17056f5cf048ca945415aec9586c8d83ea9ea6a0343f0bbb0165e8a0e1c77 +size 449403 diff --git a/vlm/dev/wv79UiY5U7/7.png b/vlm/dev/wv79UiY5U7/7.png new file mode 100644 index 0000000000000000000000000000000000000000..b47fd64a9038e3090e15f9564f80ef612ebf8d0a --- /dev/null +++ b/vlm/dev/wv79UiY5U7/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6bfa112d404b298fcd40d0ad67a6ba968e522d81f31fab2619701c520b4847b +size 703946 diff --git a/vlm/dev/wv79UiY5U7/8.png b/vlm/dev/wv79UiY5U7/8.png new file mode 100644 index 0000000000000000000000000000000000000000..d9bfca726ab2c5bec17e79be6e0452c6e27c6deb --- /dev/null +++ b/vlm/dev/wv79UiY5U7/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:501ad4b67fc6af3afb98b69e48e014a14129a71e03c9e7cb417213e78c701a86 +size 509694 diff --git a/vlm/dev/wv79UiY5U7/9.png b/vlm/dev/wv79UiY5U7/9.png new file mode 100644 index 0000000000000000000000000000000000000000..0a85edd212d1dbe6e7e11b5b63a3a69ab944a8b4 --- /dev/null +++ b/vlm/dev/wv79UiY5U7/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:343b102a6182a6eb276e858b78565cc13c64e685919cc230574f6b472ca6ef58 +size 594239 diff --git a/vlm/dev/xpjsOQtKqx/0.png b/vlm/dev/xpjsOQtKqx/0.png new file mode 100644 index 0000000000000000000000000000000000000000..3860d6b416cd32296c9c88eaab8beb2b6e99d449 --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47789dbbc192a146cabcc56c406659a99659992b8e9e4b7e69b1cd46ed3669fc +size 529196 diff --git a/vlm/dev/xpjsOQtKqx/1.png b/vlm/dev/xpjsOQtKqx/1.png new file mode 100644 index 0000000000000000000000000000000000000000..a1ba12002ffe93793f1e1982a04e4625b54f5895 --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04b8679c6281bccac90650075f74f38f7038612c6522d56787c3b68a913d565a +size 584208 diff --git a/vlm/dev/xpjsOQtKqx/10.png b/vlm/dev/xpjsOQtKqx/10.png new file mode 100644 index 0000000000000000000000000000000000000000..3216db243115148216ff7c98e5d199976754dda1 --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07a512cab391e4b0060025305c4e940c28574d807bdf4cd7d2c094bfa4200fd8 +size 491750 diff --git a/vlm/dev/xpjsOQtKqx/11.png b/vlm/dev/xpjsOQtKqx/11.png new file mode 100644 index 0000000000000000000000000000000000000000..a461e6059f39ba4b4a96e0f7ec0374c39b30c6b0 --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:daca306cc0fe0fa93ac93ca943e4e13aa482b754683af04cba33c654d398215e +size 510968 diff --git a/vlm/dev/xpjsOQtKqx/12.png b/vlm/dev/xpjsOQtKqx/12.png new file mode 100644 index 0000000000000000000000000000000000000000..4a8075e9ad18e9e1b58a23df7d19c36ea78c5a70 --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc3a3324acc156c0782aae03122c3c7946ee1fbb3a8371e4410e9ca788abe5e1 +size 495676 diff --git a/vlm/dev/xpjsOQtKqx/13.png b/vlm/dev/xpjsOQtKqx/13.png new file mode 100644 index 0000000000000000000000000000000000000000..1415ad53bdb722254a5baaf480cdfd35d60ba582 --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13bc3ff4bf840045a421ea8b61c177e33205b96b137d7512bf4cae594816236e +size 538342 diff --git a/vlm/dev/xpjsOQtKqx/14.png b/vlm/dev/xpjsOQtKqx/14.png new file mode 100644 index 0000000000000000000000000000000000000000..93ec4b42ab24afb7b4074298b82e59e1184f7869 --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8dba16f12206a3512d5d3bc6b919886f2912ad0f36a96856b5c2a58d4d453f0 +size 486074 diff --git a/vlm/dev/xpjsOQtKqx/15.png b/vlm/dev/xpjsOQtKqx/15.png new file mode 100644 index 0000000000000000000000000000000000000000..a3a1adffccd7b067a13cb6bca21bd739efd7dbd0 --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46fa95d6b9d8d5fb24a04f84a1783c53c89146ae4415844001097a934350bbd8 +size 431821 diff --git a/vlm/dev/xpjsOQtKqx/16.png b/vlm/dev/xpjsOQtKqx/16.png new file mode 100644 index 0000000000000000000000000000000000000000..cec6f6d975da5e73d6860913495af0438bd54eff --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d776bdad18ab247d45f2232b07eda851de9d3e08157c86880b0742da5bf5fc0b +size 325315 diff --git a/vlm/dev/xpjsOQtKqx/17.png b/vlm/dev/xpjsOQtKqx/17.png new file mode 100644 index 0000000000000000000000000000000000000000..4d9fec2e3fb641c46e7b8d4824fd5a7760895534 --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f196b843fa0af95bbee4d96ec02e821d460fd6e4f3f177795c1f08f256c9f94 +size 403257 diff --git a/vlm/dev/xpjsOQtKqx/18.png b/vlm/dev/xpjsOQtKqx/18.png new file mode 100644 index 0000000000000000000000000000000000000000..4d9e15006004505bc13f811994b570f6f5324ea6 --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e4e89271604e4f82b99214c877f5fe35a1962dacc1372e3020e3ad4870599ef +size 413973 diff --git a/vlm/dev/xpjsOQtKqx/19.png b/vlm/dev/xpjsOQtKqx/19.png new file mode 100644 index 0000000000000000000000000000000000000000..534684b98db01adc44319790b782bc8db4f89d17 --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:211a745d87ee03457840a44dc47a976f37e596d3cd699480820c76aa27357afc +size 422173 diff --git a/vlm/dev/xpjsOQtKqx/2.png b/vlm/dev/xpjsOQtKqx/2.png new file mode 100644 index 0000000000000000000000000000000000000000..d57e41ae70c18e05918f8e66545cd7f652a33615 --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e175c5529bbf613f2b3fb92d60706872654be86852d76a8e76e8a57bb329d524 +size 566844 diff --git a/vlm/dev/xpjsOQtKqx/20.png b/vlm/dev/xpjsOQtKqx/20.png new file mode 100644 index 0000000000000000000000000000000000000000..6c8b58486c971e2e395335c43facff5cbb1a856c --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8392e0d16a8b26582c10b92c10ae1e5695abc0b65d922bf53e23f36a022dd213 +size 906799 diff --git a/vlm/dev/xpjsOQtKqx/3.png b/vlm/dev/xpjsOQtKqx/3.png new file mode 100644 index 0000000000000000000000000000000000000000..016aed6f8169e8e5e797265d684eb9c2897ab7c9 --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e62ceca5b0119233b80aab367e30b917ca33299be01b76ac20d2673e33d652d +size 489755 diff --git a/vlm/dev/xpjsOQtKqx/4.png b/vlm/dev/xpjsOQtKqx/4.png new file mode 100644 index 0000000000000000000000000000000000000000..6718249270f243a04599980920c8c39b8bc06b82 --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a8c04debafdaeed605f796e4fdcda8cb58a00d830bc00ab8b434f8767dbd88e +size 493731 diff --git a/vlm/dev/xpjsOQtKqx/5.png b/vlm/dev/xpjsOQtKqx/5.png new file mode 100644 index 0000000000000000000000000000000000000000..41f6a594eb757450ef39c580da545c383b4e7062 --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:235645143543c10ac9053173af501beb3738094278fedc0bc6e8008b4f31fd0d +size 570824 diff --git a/vlm/dev/xpjsOQtKqx/6.png b/vlm/dev/xpjsOQtKqx/6.png new file mode 100644 index 0000000000000000000000000000000000000000..05027289f8ad461ac864a5e8abc0d6a5f5c8dc6d --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5147b1dc77e29769733618c909f5bc9b5c4c70c2f129743be941173a0da50bf +size 487243 diff --git a/vlm/dev/xpjsOQtKqx/7.png b/vlm/dev/xpjsOQtKqx/7.png new file mode 100644 index 0000000000000000000000000000000000000000..8c5a2ccca6e5aaa5deecfdaef6e1de7e9f360880 --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7324d260a01bf5b4ce1c3fd6a26c7fb373bf800885bad729416a4266df1079f6 +size 526029 diff --git a/vlm/dev/xpjsOQtKqx/8.png b/vlm/dev/xpjsOQtKqx/8.png new file mode 100644 index 0000000000000000000000000000000000000000..a5dc85a9f877c5e816c855ca06762ca6941f0d4a --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc3fc23c7b88a236c59dad66dbb67e618573bbf5f85d8e707826e56848fb884f +size 722332 diff --git a/vlm/dev/xpjsOQtKqx/9.png b/vlm/dev/xpjsOQtKqx/9.png new file mode 100644 index 0000000000000000000000000000000000000000..fdf9e04c0a5a74ad08e8bcb8c87712b7c4a80369 --- /dev/null +++ b/vlm/dev/xpjsOQtKqx/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58b2e4652d4f80e7d21453ace481adb20ad23728d1934324a5c4fbe228bbb7db +size 605495 diff --git a/vlm/dev/zf_Ll3HZWgy/0.png b/vlm/dev/zf_Ll3HZWgy/0.png new file mode 100644 index 0000000000000000000000000000000000000000..a3305fc9655a58fac8620e9182c1b3083f6aa0db --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bcf9691195facd3d7b97ad1a9bc6d43b497ecb7cccae793728f9410ecd9aa218 +size 511796 diff --git a/vlm/dev/zf_Ll3HZWgy/1.png b/vlm/dev/zf_Ll3HZWgy/1.png new file mode 100644 index 0000000000000000000000000000000000000000..01ec97f2949b696e31b190f4a039a0bcfa4e7833 --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:882ef175403fe21aeb97836b8eb652c87befb46224fc2f1e740e790caed23d61 +size 631623 diff --git a/vlm/dev/zf_Ll3HZWgy/10.png b/vlm/dev/zf_Ll3HZWgy/10.png new file mode 100644 index 0000000000000000000000000000000000000000..569f504904aff3f16ad8101f5ef312a2953469c2 --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94255a0384f7f6f619c67865d66f3f5239b0318128a66b2ffb2d583ebc3d042d +size 546180 diff --git a/vlm/dev/zf_Ll3HZWgy/11.png b/vlm/dev/zf_Ll3HZWgy/11.png new file mode 100644 index 0000000000000000000000000000000000000000..eb74edb1654c2a8b2b7325fc13648b2b8410434d --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8f8d68498cac5a4782cd35db30ccae75db37445393ab51adb2f2da46f06b609 +size 548974 diff --git a/vlm/dev/zf_Ll3HZWgy/12.png b/vlm/dev/zf_Ll3HZWgy/12.png new file mode 100644 index 0000000000000000000000000000000000000000..0ae3ae9a5026c64359aefd2a05e46067cd463e97 --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c527c5d8c918e9c8054c0a391ee853046bcee3d144bc76332d98964b512e65a +size 582141 diff --git a/vlm/dev/zf_Ll3HZWgy/13.png b/vlm/dev/zf_Ll3HZWgy/13.png new file mode 100644 index 0000000000000000000000000000000000000000..7970669a5fd442c6a64f7f3eced96f15f336574d --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1588b0d402bdd5ec0dfb613139670c7a26abd478473afc6f3ee3196a3370735b +size 564201 diff --git a/vlm/dev/zf_Ll3HZWgy/14.png b/vlm/dev/zf_Ll3HZWgy/14.png new file mode 100644 index 0000000000000000000000000000000000000000..7d4b8dc8493c046c61e48d4af7f900f52178173d --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd243788f886030ee70d77221473d4119c9590000f6e2bbff39c9bf884e03b8b +size 498612 diff --git a/vlm/dev/zf_Ll3HZWgy/15.png b/vlm/dev/zf_Ll3HZWgy/15.png new file mode 100644 index 0000000000000000000000000000000000000000..99d3cb4b81e5c3c532f2e593732a0bd6b377fcb7 --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1053c9e9a1d055871777cc0a03fef5faa32c7eb43b99f4a4da210565e7eaf4a3 +size 546275 diff --git a/vlm/dev/zf_Ll3HZWgy/16.png b/vlm/dev/zf_Ll3HZWgy/16.png new file mode 100644 index 0000000000000000000000000000000000000000..936f242b7c83d92b0c38d72028cee46257bc337a --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea308b96f4727a7bcd5a37b4499f2598a36156fadbddbe8a00902c2ad6f6a10e +size 874033 diff --git a/vlm/dev/zf_Ll3HZWgy/17.png b/vlm/dev/zf_Ll3HZWgy/17.png new file mode 100644 index 0000000000000000000000000000000000000000..3f906b724017c7cec820520bd10afd6c599c970c --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:223d30845a4a2d1f3b5cadab6bc69dc5419dceb7eba7cb0b87b192b071cd30bb +size 869310 diff --git a/vlm/dev/zf_Ll3HZWgy/2.png b/vlm/dev/zf_Ll3HZWgy/2.png new file mode 100644 index 0000000000000000000000000000000000000000..2e9868a5a034054ba6c50d7e91389783b74786c9 --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9acf655d410a7969f7f69bde1e12afc26e0d5bbbd369885abd7587474795b798 +size 582751 diff --git a/vlm/dev/zf_Ll3HZWgy/3.png b/vlm/dev/zf_Ll3HZWgy/3.png new file mode 100644 index 0000000000000000000000000000000000000000..b45e6362a6c11067c001f5cf5cc92eedbf616b13 --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e9b1ec262a21a530751624511efd75957106b2b8d78a85b936dc66cc5171efb +size 572399 diff --git a/vlm/dev/zf_Ll3HZWgy/4.png b/vlm/dev/zf_Ll3HZWgy/4.png new file mode 100644 index 0000000000000000000000000000000000000000..06437c8cfa7757b3d0a4163e4d822c2bddbc79bb --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38cfa4634877bf813b1ca3c49f08b611f46e5415f2f47bb4de9cedbfb7946b46 +size 620335 diff --git a/vlm/dev/zf_Ll3HZWgy/5.png b/vlm/dev/zf_Ll3HZWgy/5.png new file mode 100644 index 0000000000000000000000000000000000000000..a161d2d32cc17566172eb5977bce0d85ab18b564 --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e9a58e15f2d180feea0f042e2431675fda2305decfc15ddf7ec29dd01747285 +size 563031 diff --git a/vlm/dev/zf_Ll3HZWgy/6.png b/vlm/dev/zf_Ll3HZWgy/6.png new file mode 100644 index 0000000000000000000000000000000000000000..907fa6856023b98ef8d59df8c23dcc54ade38e5f --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a287013d7b335285b1074c923b518f3b53e7c2cbfb1abec75fd7ea27ec83e5ee +size 555646 diff --git a/vlm/dev/zf_Ll3HZWgy/7.png b/vlm/dev/zf_Ll3HZWgy/7.png new file mode 100644 index 0000000000000000000000000000000000000000..d47748717552fc59aa77d186a42a0c02bacecb5f --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bc883ed0df11856c76fee8592d3f3e193e3337dc2eb343c8bc92dddfa64cc41 +size 562383 diff --git a/vlm/dev/zf_Ll3HZWgy/8.png b/vlm/dev/zf_Ll3HZWgy/8.png new file mode 100644 index 0000000000000000000000000000000000000000..546316c3e9a27f7ba3f2589ccd01204125609fa8 --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd28ed4b74b0244b9ff6b977bb7c7679cc3db142872f7616606e73ab2382efff +size 935081 diff --git a/vlm/dev/zf_Ll3HZWgy/9.png b/vlm/dev/zf_Ll3HZWgy/9.png new file mode 100644 index 0000000000000000000000000000000000000000..3d37166cdff7490c8ce0a1de6ce2292835ed22eb --- /dev/null +++ b/vlm/dev/zf_Ll3HZWgy/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58655057586edf385fed798c3c96d7d42a1e12edcbe85819223ba7d8a9200169 +size 528749 diff --git a/vlm/dev/zq1iJkNk3uN/0.png b/vlm/dev/zq1iJkNk3uN/0.png new file mode 100644 index 0000000000000000000000000000000000000000..6986f7faa0efd1d436b93a0541d1336ab8a89b7d --- /dev/null +++ b/vlm/dev/zq1iJkNk3uN/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:432d76ebff06d33023c3dd0ba826b7d360dfc25cb41d9bc9b7d56148d48ebff4 +size 500297 diff --git a/vlm/dev/zq1iJkNk3uN/1.png b/vlm/dev/zq1iJkNk3uN/1.png new file mode 100644 index 0000000000000000000000000000000000000000..c91e2805fcf515ab10a20174b4824abd26e376ed --- /dev/null +++ b/vlm/dev/zq1iJkNk3uN/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8357a8dc75783f9ebc9806476a06758c81cc6eb246c1dc245d5c8c7ff0939945 +size 664492 diff --git a/vlm/dev/zq1iJkNk3uN/10.png b/vlm/dev/zq1iJkNk3uN/10.png new file mode 100644 index 0000000000000000000000000000000000000000..6df8e7a85b109700cc0ab74c7df9b030e4528aa7 --- /dev/null +++ b/vlm/dev/zq1iJkNk3uN/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0ce63d6bf325e3d5d96fec9bafe4b08f79c7a353c1be330997845ae509f43f8 +size 547417 diff --git a/vlm/dev/zq1iJkNk3uN/11.png b/vlm/dev/zq1iJkNk3uN/11.png new file mode 100644 index 0000000000000000000000000000000000000000..3eadbba163d99220712b9900e3a82f75745425d7 --- /dev/null +++ b/vlm/dev/zq1iJkNk3uN/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5a0ab82399cbcc6334223c00640cd32db5a889066a0cf4eefc09224a9604692 +size 569902 diff --git a/vlm/dev/zq1iJkNk3uN/12.png b/vlm/dev/zq1iJkNk3uN/12.png new file mode 100644 index 0000000000000000000000000000000000000000..b7c1763bf96ef11aa49aea650fa2a64c088c8454 --- /dev/null +++ b/vlm/dev/zq1iJkNk3uN/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec7a7f27b40df68ef34199328ea2c119b87dabacdbc92edb98b7173275966f22 +size 428023 diff --git a/vlm/dev/zq1iJkNk3uN/13.png b/vlm/dev/zq1iJkNk3uN/13.png new file mode 100644 index 0000000000000000000000000000000000000000..4b2f694f6d4090d5e297b251c2e740c538d63b85 --- /dev/null +++ b/vlm/dev/zq1iJkNk3uN/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c549678c67ad1873336339b5b1efd79ce4739ee44913e4c2566fc621e09cb469 +size 500082 diff --git a/vlm/dev/zq1iJkNk3uN/14.png b/vlm/dev/zq1iJkNk3uN/14.png new file mode 100644 index 0000000000000000000000000000000000000000..5e722c28ac5599abf430d2ad921321b21787d0c2 --- /dev/null +++ b/vlm/dev/zq1iJkNk3uN/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7aa48b4bd5d22fdfd2df27fab750903775b990e19b902fa5810e1fbbdca5021a +size 1042973 diff --git a/vlm/dev/zq1iJkNk3uN/15.png b/vlm/dev/zq1iJkNk3uN/15.png new file mode 100644 index 0000000000000000000000000000000000000000..3ccefcfa97d6d390dda27f5134e5ad51752441ba --- /dev/null +++ b/vlm/dev/zq1iJkNk3uN/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c52d6f1a0009285a11054d545decf0bd0e2a799bda5a7fa61c6dfded7c48d91f +size 484962 diff --git a/vlm/dev/zq1iJkNk3uN/16.png b/vlm/dev/zq1iJkNk3uN/16.png new file mode 100644 index 0000000000000000000000000000000000000000..8dccce644da4c48dde711eb9719a6d3604ffc6bd --- /dev/null +++ b/vlm/dev/zq1iJkNk3uN/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5e7a7db344e3da63114f3e708c786557c208eefc83166cb880319983762d6dd +size 268222 diff --git a/vlm/dev/zq1iJkNk3uN/2.png b/vlm/dev/zq1iJkNk3uN/2.png new file mode 100644 index 0000000000000000000000000000000000000000..a6ab72377ef6f4ba5c276856eb0d411daac368de --- /dev/null +++ b/vlm/dev/zq1iJkNk3uN/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47edaba11a4691a7fc5c1045611cbec5c08284fdf30049a650a0e179eeaabde6 +size 562370 diff --git a/vlm/dev/zq1iJkNk3uN/3.png b/vlm/dev/zq1iJkNk3uN/3.png new file mode 100644 index 0000000000000000000000000000000000000000..936268259f60b04cfe71a3fd07d7650b9eec5df5 --- /dev/null +++ b/vlm/dev/zq1iJkNk3uN/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:956cf49c3b11a77d302a6482fb799228b524d4e65e695b34f9820d6f3d0f6aad +size 558330 diff --git a/vlm/dev/zq1iJkNk3uN/4.png b/vlm/dev/zq1iJkNk3uN/4.png new file mode 100644 index 0000000000000000000000000000000000000000..1ba15e5eebb1caf6858d14689066fd8b318732a2 --- /dev/null +++ b/vlm/dev/zq1iJkNk3uN/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ab5c4b43363c88429e53813a15d87e6623b7a89bfcafb89c7a2dad1e63a38f2 +size 629221 diff --git a/vlm/dev/zq1iJkNk3uN/5.png b/vlm/dev/zq1iJkNk3uN/5.png new file mode 100644 index 0000000000000000000000000000000000000000..d41e514645dac562cc258cdee7b525947cc3432c --- /dev/null +++ b/vlm/dev/zq1iJkNk3uN/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6acc4cdd858aa6707d5d106ac55a1586f337809e8756a86055664801421b431e +size 604824 diff --git a/vlm/dev/zq1iJkNk3uN/6.png b/vlm/dev/zq1iJkNk3uN/6.png new file mode 100644 index 0000000000000000000000000000000000000000..0ab65aa10d4fe0453b0382680ae77e4ac72952d0 --- /dev/null +++ b/vlm/dev/zq1iJkNk3uN/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:183ac65231a1cf6b0575198cabafb8f471ade6b27c508902b47ae957c4ffc288 +size 543706 diff --git a/vlm/dev/zq1iJkNk3uN/7.png b/vlm/dev/zq1iJkNk3uN/7.png new file mode 100644 index 0000000000000000000000000000000000000000..736e0677f1ac901c1d39864646e932a0991bb03e --- /dev/null +++ b/vlm/dev/zq1iJkNk3uN/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5ad7ae360cbf565c2fc3b933c435abfdd0bc10ca6a65dc589b0ea77f4ec301e +size 522059 diff --git a/vlm/dev/zq1iJkNk3uN/8.png b/vlm/dev/zq1iJkNk3uN/8.png new file mode 100644 index 0000000000000000000000000000000000000000..f1b21d73d2aa4054acdb60fcd7317b17bd9362e1 --- /dev/null +++ b/vlm/dev/zq1iJkNk3uN/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b99b918bf63f522105aafece379fdc5cb6d171ef6060e9b800b3848119a5b2f +size 1012570 diff --git a/vlm/dev/zq1iJkNk3uN/9.png b/vlm/dev/zq1iJkNk3uN/9.png new file mode 100644 index 0000000000000000000000000000000000000000..3b1730dd6021c75f641b31935051cf0d48b52037 --- /dev/null +++ b/vlm/dev/zq1iJkNk3uN/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4248706d6f9ba430b8fbf2e1804e0e38ae66ea41c237ece27fd092b29199941d +size 537936